Estimating cluster masses: a comparative study between machine learning and maximum likelihood
Extracted source text · arXiv 2507.21876
%% Beginning of file 'sample631.tex'
%%
%\documentclass[twocolumn, linenumbers]{aastex631}
\documentclass[twocolumn]{aastex631}
%\documentclass[twocolumn,fleqn,usenatbib]{aastex631}
\UseRawInputEncoding
%\pdfoutput=1
\usepackage{enumitem}
\usepackage{appendix}
\usepackage{array} % for p{..} column spec
\usepackage{tabularx} % auto-wrapping tables
\usepackage{amsfonts}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{multirow}
\usepackage{bbm}
\usepackage{cleveref}
\usepackage{pgfplots}
\usepackage{placeins}
\usepackage{tikz}
\usepackage{upgreek}
\usepackage{booktabs}
\usepackage[normalem]{ulem}
\usepackage{physics}
%\usepackage{caption}
%\usepackage{subcaption}
%\usepackage{aas_macros}
\usepackage{ulem}
\usepackage{cancel}
\usepackage{graphicx}
\usepackage{color}
%%%%%
\def\fB{f_\textrm{B}}
\def\hmpc{\, h^{-1}\,{\rm Mpc}}
\def\mpc{\, \rm Mpc}
\def\msol{\, {\rm M}_\odot}
\def\hmsol{\, h^{-1} \, {\rm M}_\odot}
\def\myr{\, {\rm Myr}}
\def\gyr{\, {\rm Gyr}}
\def\hgpc{h^{-1}\,{\rm Gpc}}
\def\st{{\tilde s}}
\def\Rt{{\tilde R}}
\def\rt{{\tilde r}}
%%\def\typii{\textit{Type II}}
%\def\typiii{\textit{Type III}}
\def\mtrue{M^\text{true}}
\def\mhat{\hat M}
\def\dest{d^\text{obs}}
\def\dtrue{d^\text{true}}
\def\vest{v^\text{obs}}
\newcommand{\pcolor}{{grey}}
% \newcommand{\typi}{\textit{Type I}}
% \newcommand{\typii}{\textit{Type II}}
% \newcommand{\typiii}{\textit{Type III}}
\newcommand{\cah}{CAHs}
\newcommand{\typi}{$N(s)$}
\newcommand{\typii}{$\vest (s)$}
\newcommand{\typiii}{$\vest(\dest)$}
%% The default is a single spaced, 10 point font, single spaced article.
%% There are 5 other style options available via an optional argument. They
%% can be invoked like this:
%%
%% \documentclass[arguments]{aastex631}
%%
%% where the layout options are:
%%
%% twocolumn : two text columns, 10 point font, single spaced article.
%% This is the most compact and represent the final published
%% derived PDF copy of the accepted manuscript from the publisher
%% manuscript : one text column, 12 point font, double spaced article.
%% preprint : one text column, 12 point font, single spaced article.
%% preprint2 : two text columns, 12 point font, single spaced article.
%% modern : a stylish, single text column, 12 point font, article with
%% wider left and right margins. This uses the Daniel
%% e" in hyperref
%% to improve the accessibility of the compiled pdf file.
%%
%% If you want to create your own macros, you can do so
%% using \newcommand. Your macros should appear before
%% the \begin{document} command.
%%
\usepackage{xcolor}
% Define a new command for Adi's comments
\newcommand{\adi}[1]{\textcolor{red}{Adi:``#1"}}
\newcommand{\adiblu}[1]{\textcolor{blue}{Adi:``#1"}}
\newcolumntype{Y}{>{\raggedright\arraybackslash}X}
\newcommand{\vdag}{(v)^\dagger}
\newcommand\aastex{AAS\TeX}
\newcommand\latex{La\TeX}
\def\rv{r_\mathrm{v}}
\def\Mv{M_\mathrm{v}}
\def\mv{m_\mathrm{v}}
\def\rhv{\hat{r}_\mathrm{v}}
\creflabelformat{equation}{#2#1#3} % removes parentheses around equation numbers
\crefname{equation}{eq.}{eqs.}
\Crefname{equation}{Eq.}{Eqs.}
\begin{document}
\title{Estimating cluster masses: a comparative study between machine learning and maximum likelihood }
\author[0009-0006-4297-2678]{Raeed Mundow}
\email{raeed7sm@gmail.com}
\affiliation{ Department of Physics and the Asher Space Research Institute\\ Israel Institute of Technology Technion, Haifa 32000, Israel}
\author[0000-0002-8272-4779]{Adi Nusser}
\email{adi@physics.technion.ac.il}
\affiliation{ Department of Physics and the Asher Space Research Institute\\ Israel Institute of Technology Technion, Haifa 32000, Israel}
\begin{abstract}
We compare an autoencoder convolutional neural network (AE-CNN) with a conventional maximum-likelihood estimator (MLE) for inferring cluster virial masses, $\Mv$, directly from the galaxy distribution around clusters, without identifying members or interlopers.
The AE-CNN is trained on mock galaxy catalogues, whereas the MLE assumes that clusters of similar mass share the same phase-space galaxy profile.
Conceptually, the MLE returns an unbiased estimate of $\log \Mv$ at fixed true mass, whereas the AE-CNN approximates the posterior mean, so the true $\log \Mv$ is unbiased at fixed estimate.
Using MDPL2 mock clusters with redshift space number density as input, the AE-CNN attains an rms scatter of $0.10\,\textrm{dex}$ between predicted and true $\log \Mv$, compared with $0.16\,\textrm{dex}$ for the MLE.
With inputs based on mean peculiar velocities, binned in redshift space or observed distance, the AE-CNN achieves scatters of $0.12\,\textrm{dex}$ and $0.16\,\textrm{dex}$, respectively, despite strong inhomogeneous Malmquist bias.
\end{abstract}
%% Keywords should appear after the \end{abstract} command.
%% The AAS Journals now uses Unified Astronomy Thesaurus concepts:
%% https://astrothesaurus.org
%% You will be asked to selected these concepts during the submission process
%% but this old "keyword" functionality is maintained in case authors want
%% to include these concepts in their preprints.
\keywords{large-scale structure of Universe --- methods: data analysis --- methods: numerical --- methods: convolutional neural networks --- galaxies: clusters}
%% From the front matter, we move on to the body of the paper.
%% Sections are demarcated by \section and \subsection, respectively.
%% Observe the use of the LaTeX \label
%% command after the \subsection to give a symbolic KEY to the
%% subsection for cross-referencing in a \ref command.
%% You can use LaTeX's \ref and \label commands to keep track of
%% cross-references to sections, equations, tables, and figures.
%% That way, if you change the order of any elements, LaTeX will
%% automatically renumber them.
%%
%% We recommend that authors also use the natbib \citep
%% and \citet commands to identify citations. The citations are
%% tied to the reference list via symbolic KEYs. The KEY corresponds
%% to the KEY in the \bibitem in the reference list below.
\section{Introduction} \label{sec:intro}
Galaxy clusters, the most massive gravitationally virialized objects in the universe, comprise galaxies, hot intracluster gas (ICM), and dark matter (DM), serving as valuable laboratories for studying cosmology and astrophysics. Determining their abundance as a function of mass provides essential constraints on the matter and energy content of the Universe as well as the amplitude of mass fluctuations. Various techniques have been developed to estimate cluster masses, including gravitational lensing, which infers the projected mass distribution from the distortion of background galaxy images \citep[e.g.][]{LotzLensing}, and X-ray spectroscopy, which derives mass profiles based on the ICM temperature and density under the assumption of hydrostatic equilibrium. Additionally, the Sunyaev-Zel'dovich (SZ) effect, which results from the distortion of cosmic microwave background radiation by the hot ICM, offers an alternative mass estimation method when combined with X-ray data \citep{allingham}.
The current study is concerned with estimation of cluster mass through observations of the kinematics of galaxies bound to clusters and galaxies in the vicinity of clusters.
Most algorithms for identifying members of groups and clusters in observations of galaxies in redshift surveys are based on Friends-of-Friends (FoF) type of algorithms \citep[e.g.,][]{Tempel2016,Marini2025}.
By measuring the line-of-sight velocities of cluster members, the velocity dispersion can be determined. Through the Jeans equation, the observed kinematics of a spherical system can be related to its mass profile \citep[e.g.][]{Carlberg97,Mamon2013}. This approach assumes the system is virialized and may be affected by interlopers—galaxies that are not gravitationally bound to the cluster.
In contrast, the caustic \citep{1999MNRAS.309..610D,2024A&A...682A..80P} technique estimates cluster mass profiles by measuring galaxy escape velocities as a function of cluster-centric distance. Unlike virial methods, it does not require dynamical equilibrium assumptions and performs well in cluster outskirts, while naturally filtering out interlopers through its velocity threshold criteria. This approach has been validated using hydrodynamical simulations \citep{geller2023}, demonstrating its complementary role to virial-based estimates.
Here we focus on comparing Neural Networks (NNs) and Maximum-Likelihood Estimation (MLE) approaches for determining the mass based on galaxy redshift surveys, particularly by analyzing the distribution and motions of galaxies within clusters and their surroundings. The large virialized motions of galaxies bound to a cluster's gravitational potential cause the observed distribution of these galaxies to appear elongated along the line-of-sight in redshift space, a distortion known as the Finger of God (FoG) effect. The methods employed in this study account for all galaxies within the FoG region, irrespective of whether they are actual cluster members or merely lie in the vicinity of the cluster.
The MLE approach involves positing a probability distribution (likelihood function) for the observed data given some unknown parameters, $\theta$. The goal of MLE is to find the parameter values that make the observed data most probable under this assumed model. In practice, this means finding the $\theta$ that maximizes it the likelihood function $L(\theta) = P(\text{data} \vert \theta)$.
Given a robust assumed model, MLE provides interpretable results in terms of the fitted parameters and their corresponding confidence level.
Machine learning has opened a new avenue for studying and analyzing cosmological observations, including clusters of galaxies.
A key advantage of NNs over MLE is their ability to identify complex relationships between observables and the physical properties of a system. NNs learn patterns and mappings directly from training data, incorporating a wide range of physical effects and observational limitations. Many of these effects are too intricate for simplified models to capture accurately or too computationally expensive to evaluate during inference in an MLE framework. As a result, recent years have seen a growing number of studies utilizing NNs. For instance, \citep{Luisa2022} trained a deep learning model to predict the mass profile of halo clusters using their mass accretion history as input, aiming to elucidate the contributions of the
accretion history and initial density fluctuations to the final halo mass buildup.
NNs have also been used for reconstructing cluster properties from limited projected observations \citep[e.g.,][]{Iqbal2023,Caro2024}.
Despite their advantages, NNs do not explicitly exploit well-known physical properties of a system the way MLE-based methods do.
Capturing these properties with NNs is possible, but requires that a sufficiently large and reliable training dataset is available.
A goal of this work is to present a comparative study of NN and MLE approaches for estimating cluster masses from a fair sample of the redshift space distribution of tracers (galaxies) in the vicinity of clusters.
In the MLE approach here, we estimate the virial mass of a cluster only from the redshift distribution of ``galaxies" in and around clusters.
On the NNs side, in addition to training on redshift space distributions, we will also consider NNs trained on ``observed" line-of-sight peculiar velocities of galaxies.
A major challenge in this context is whether NNs can mitigate the effects of Inhomogeneous Malmquist Bias resulting from placing galaxies at either their redshift coordinate or observed noisy distances rather than true positions.
The paper is organized as follows. In \S\ref{sec:simulation}, we describe the simulation data used in this work.
Our implementation of the MLE approach, along with its underlying assumptions, is presented in \S\ref{sec:method-chi2}.
In \S\ref{sec:method-CNN}, we outline the architecture of the neural network (NN), detail the training and validation datasets, and report the model’s performance based on the loss function.
In \S\ref{sec:interp}, we provide a clarification of a fundamental difference between the MLE and NN approaches, focusing on how their outputs should be interpreted.
The results of applying MLE and NNs are presented in \S\ref{sec:results}. We end with a discussion in \S\ref{sec:discus}.
\section{Simulation data}
\label{sec:simulation}
Assessment of the MLE and NN estimation of cluster masses from kinematical data, will rely on the data extracted from the MultiDark Planck 2 (MDPL2) N-body simulation \citep{mdpl...793..127V}, provided by the CosmoSim database.
The dark matter only simulation adopts the Planck cosmological parameters and contains $3840^3$ particles in a cubic box of $1000\hmpc$ on the side.
Halo catalogs generated using the Rockstar halo finder \citep{Behroozi2013} are also available with halo mass and radius according to various definitions. We adopt the definition of a halo virial radius as the radius within which the mean density equals to $\Delta_{\mathrm{v}}(z)\rho_\mathrm{m}$ where $\Delta_\mathrm{v} \approx 360$ at $z=0$ \citep{Bryan1998} and $\rho_\mathrm{m}$ is the mean background density. Accordingly, the virial mass is approximated by
\begin{equation}
M_{\mathrm{v}} = \frac{4\pi}{3}\Delta_{\mathrm{v}}(z)\rho_{\mathrm{m}}r_{\mathrm{v}}^3 .
\label{eq:virdef}
\end{equation}
For both MLE and NN analyses, we limit our sample to cluster halos with $M_\mathrm{v} > 10^{14}\hmsol$, yielding approximately $26,000$ clusters in the simulation volume.
\Cref{table:cluster_count} lists the number of clusters from the simulation in four mass ranges (In units of $10^{14}\,h^{-1}\,\textrm{M}_\odot$). For each cluster, we assemble its \textit{Cluster-Associated Halos} (hereafter \cah ) population by first collecting all of its sub-halos and then including any additional distinct halos whose three-dimensional positions lie within a cube of side length $68\,h^{-1}\,\mathrm{Mpc}$ centered on the cluster. This cube size is sufficient to encompass the maximum FoG elongations produced by our full range of cluster virial masses.
\begin{table}[h]
\centering
\caption{}
\label{tab:sim_clusters}
\begin{tabular}{lc}
\hline
Mass range & Clusters count \\
\hline
$1<M_\mathrm{v14}<2.5$ & 21925 \\
$2.5<M_\mathrm{v14}<5$ & 4401 \\
$5<M_\mathrm{v14}<7.5$ & 821 \\
$M_\mathrm{v14}>7.5$ & 391 \\
\hline
\end{tabular}
\label{table:cluster_count}
\end{table}
We work in the distant observer limit and define a Cartesian coordinate system centered on the cluster, in which the $x_1-x_2$ plane is perpendicular to line-of-sight axis $x_3$.
The projected distances of \cah\ is computed as $R=(x_1^2+x_2^2)^{1/2}$. At low redshifts, the redshift space coordinate, $s=cz/H_0$, of a galaxy is related to the true distance coordinate \( x_3 \) by,
\begin{equation}
\label{eq:s}
s = x_3 + v_3\; ,
\end{equation}
where \( v_3 \) is the line-of-sight component of the galaxy's physical peculiar velocity relative to the cluster center, expressed in distance units.
\section{The Maximum Likelihood approach}
\label{sec:method-chi2}
MLE requires a robust model for the observed distribution of
redshifts and projected distance coordinates \((s, R)\) of \cah\ relative to the centers of their respective parent clusters.
We assume a universal spatial and kinematic pattern of \cah . By universality we mean here that, after appropriate scaling and normalization, spherically averaged profiles of number density, $\mathcal{N}(r)$ and moments of the peculiar velocity PDF, $P(\vb{v}|r)$, remain consistent across different clusters. Individual clusters deviate from this pattern due to non-sphericity, mergers, or other dynamical effects, but these deviations are assumed to be minor compared to the universal profiles obtained by stacking many clusters. In \S\ref{sec:Universality} we will demonstrate the universality in the MDPL2 simulation. However, as we argue in \S\ref{sec:discus}, the general framework that we describe here can easily accommodate a ``weak" breaking of this universality such as
an explicit dependence on the cluster mass, provided that individual cluster kinematic properties remain close to the mean profile of clusters with similar mass.
Given a value for \( \rv \), or equivalently \( \Mv \), we introduce the joint probability distribution function (PDF) of having a galaxy at
line-of-sight true distance coordinate \( x_3 \), projected radial coordinate \( R \), and observed line-of-sight redshift space
coordinate \( s \) as,
\begin{equation}
P(s, x_3, R; \rv) = P(s| x_3, R; \rv) P(x_3, R; \rv)\; .
\label{eq:thepdf}
\end{equation}
The conditional probability \( P(s| x_3, R; \rv) \) describes the mapping between $x_3$, and $s$, which, by \cref{eq:s}, is actually the PDF of $v_3$ at point $(x_3,R)$.
The term \( P(x_3, R; \rv) \dd x_3 \dd R \) represents the
probability that a galaxy has coordinates \( s \) and \( R \) within
the intervals \([x_3, x_3+\dd x_3]\) and \([R, R+\dd R]\), respectively.
The PDF of the observed coordinates \( (s_i, R_i) \) is then,
\begin{equation}
\label{eq:PsR}
\prod_{i=1}^N P(s_i, R_i; \rv) = \prod_{i=1}^N \int \dd x_3 \, P(s_i, x_3, R_i; \rv)\; .
\end{equation}
The universality assumption implies that $P(s,R;\rv) $ depends on $\rv$ only through the scaled coordinates,
\begin{equation}
\st = s/\rv \quad \textrm{and} \quad \Rt = R/\rv\; .
\end{equation}
Instead of seeking parametric forms to approximate $\mathcal{N}(r) $ and moments of $P(\vb|r)$ and then compute $P(s,R; \rv)$, we simply use the distribution of $(\st,\Rt) $ as evaluated directly from the simulation. This is done as follows.
For each cluster we compute $(\st,\Rt)$ of its \cah . Universality implies that $P(\st, \Rt)$ is approximately independent of the cluster virial radius $\rv$. Therefore, we construct a reference distribution, $P_\textrm{ref}(\st, \Rt)$, by stacking all $\st$ and $\Rt$ values from all clusters.
The reference distribution $P_\textrm{ref}(\st,\Rt)$ is defined on a two-dimensional (2D) grid in $\st$ and $\Rt$. Each bin value of $P_\textrm{ref}$ is proportional to the total number of points $(\st,\Rt)$ from all clusters that fall within that bin. The distribution $P_\textrm{ref}$ is then normalized so that the sum over all bins, weighted by their respective bin areas, equals unity.
For a given cluster, $\alpha$, with a total number of $N^\alpha$ of \cah\ having coordinates $(s^\alpha_i, R^\alpha_i)$ for $i = 1, \dots, N^\alpha$, an estimate of its virial radius is determined as follows.
Assuming a trial value \(\hat{r}_\mathrm{v}\), we scale the cluster dataset to
\begin{equation}
\label{eq:srtilde}
(\hat{{\tilde s}}^\alpha, \hat{{\tilde R}}^\alpha)=\left(\frac{s^\alpha_i}{\hat{r}_\mathrm{v}}, \frac{R^\alpha_i}{\hat{r}_\mathrm{v}}\right),
\end{equation}
and bin these points onto the same 2D grid used for $P_\textrm{ref}$. Let \(m^\alpha(b)\) be the number of points in bin \(b\), so that
\begin{equation}
N^\alpha = \sum_{b=1}^{B} m^\alpha(b).
\end{equation}
We then compute the Chi-Square statistic
\begin{equation}
\label{eq:chisq}
\chi^2\left(\rhv\right)
= \sum_{b=1}^{B}
\frac{\bigl[m^\alpha_b - E_b(\rhv)\bigr]^2}{E_b(\rhv)},
\end{equation}
where \(E_b(\rhv)\propto P_\textrm{ref}(\rhv) \) is the expected count in bin \(b\),
satisfying
\begin{equation}
\label{eq:Ebnorm}
\sum_b E_b = N^\alpha \; .
\end{equation}
Minimizing \(\chi^2\) over \(\rhv\) then provides the best-fit virial radius for cluster \(\alpha\).
Specifically the minimization leads to
\begin{equation}
\pdv{\chi^2}{\rhv}=-\sum_b \pdv{\ln E_b}{\rhv} \frac{N_b}{E_b}=0
\end{equation}
where we have used \cref{eq:Ebnorm}. In the limit of large $E_b$, we have $N_b \approx E_b$ and hence the minimization of $\chi^2 $ becomes equivalent to maximizing the log-likelihood function
given by
\begin{equation}
\label{eq:mathcalL}
\log {\cal L}(\rhv)=\sum_{i=1}^{N^\alpha}\ln P_{\textrm{ref}}({\hat {\tilde s}}^\alpha_i, {\hat {\tilde R}}^\alpha_i)\; ,
\end{equation}
where the dependence of $P_{\textrm{ref}}$ on $\rhv$ is via \cref{eq:srtilde}.
\subsection{Approximate Universal profiles in number density and kinematic profiles}
\label{sec:Universality}
\begin{figure}[ht!]
\centering
\includegraphics[width=1.0\columnwidth]{density_of_combined_clusters.png}
\caption{\textit{Top panel:} Normalized number density profiles of \cah\ of virial masses $\mv>10^{10}\hmsol$, around clusters, for three cluster mass bins, as indicated in the figure
(\(M_\mathrm{v14}\equiv \Mv/10^{14}\hmsol\)). Solid curves represent stacked profiles, while dashed lines are examples of individual clusters, demonstrating deviations from the universal pattern.
\textit{Bottom panel:} Profiles
separated by mass of \cah\ \(\mv\) (in §
units of \(10^{11}\hmsol \)), for all clusters. The differences do not affect the MLE, since
the same $\mv$ cut is adopted.
\label{fig:density_profiles}}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=1.0\columnwidth]{r_vel_of_combined_clusters.png}
\caption{\textit{Top panels:} Moments of radial (outward from the cluster halo center) peculiar velocity of \cah\ with \(m_\mathrm{v}>10^{10}\hmsol \) normalized by $H_0\rv$. In the left panel, stacked profiles of the mean radial velocity are plotted for three cluster mass bins. In the panel to the right, the curves of the standard deviation of the radial velocity dispersion are shown. \textit{Bottom panels:} Same quantities as the top panels, but showing dependence of profiles on the mass of \cah\ in all clusters used. For reference, by \cref{eq:virdef}, the circular velocity is $V_\mathrm{c}\approx 0.14 \rv$ at $z=0$.
\label{fig:radial_velocities}}
\end{figure}
To test the universality assumption, we divide the simulated cluster halos into three bins in virial mass, stack them within each bin, and analyze their number density and velocity profiles. The top panel of \Cref{fig:density_profiles} shows, as solid lines, the stacked number density, \(\mathcal{N}\), of \cah\ with main cluster halos as a function of their distance \(r\) from the cluster centers, in units of \(\rv\). All profiles are normalized by the total number of \cah\ in each cluster.
The profiles exhibit remarkable consistency across the $\Mv$ ranges.
The bottom panel illustrates that while the stacked profiles show minimal dependence on cluster halo masses, they are somewhat sensitive to the \cah\ masses, $\mv$. The $m_\mathrm{v11}<5$ and $m_\mathrm{v11}>5$ profiles share a similar slope over most of the region within the cluster virial radius, $r/\rv \lesssim 1$, and nearly overlap well outside the virial radius. However, their overall amplitudes differ at $r/\rv \lesssim 1$. Nonetheless, MLE will be applied using $P_\textrm{ref}$ calibrated with the same $\mv$ threshold, and is therefore unaffected by this dependence on $\mv$.
In principle, however, accuracy could be improved by incorporating an explicit dependence on $\mv$.
\Cref{fig:radial_velocities} illustrates the universality of peculiar velocity moment profiles.
The stacked mean and standard deviation of the radial peculiar velocity, scaled by $H_0 \rv$ for each cluster, vary only slightly across the three cluster mass bins.
At $r/\rv \approx 1$, the mean scaled peculiar velocity is close to $-1$, reflecting how the mean physical velocity, $V_\textrm{physical} = H_0 r + \langle v_r \rangle$, approaches zero at the virial radius.
The curves reach their minima at $\log r/\rv \approx 0.4$ ($r/\rv \approx 2.5)$, indicating the radius of maximum infall velocity.
The turn-around radius, marking the transition between outwardly expanding matter and infall onto the cluster halo, appears at $\log r/\rv \approx 0.6$ ($r/\rv \approx 4$).
Well inside the virial radius, $\langle v_r \rangle$ remains around $-1$, implying that $V_\textrm{physical}$ does not vanish entirely but instead hovers around $-0.6 \, H_0 \, \rv \approx -0.08 \, V_\mathrm{c}$.
The velocity dispersion, as measured by the standard deviation in the right panels, increases as we approach the central region of the cluster halos. There is a bump in the dispersion around he same location as the dip of the mean velocity curves.
%Further assessment of the approximate universality is presented in the appendix.
\section{Estimation via Convolutional Neural Networks}
\label{sec:method-CNN}
Independently of the MLE approach, we develop a deep learning framework to infer the virial mass of cluster halos. We will assume that we have measurements of projected distance $R$ and redshift space coordinate $s$, for \cah\ within a sufficiently large volume centered on the cluster, as will be described in \S\ref{sec:trainvalid}. Further, for a smaller sample of \cah , we assume we have observations of their distances, $\dest$,
and peculiar velocities, $ \vest$.
To account for observational uncertainties, the observed distances $\dest$ are modeled as lognormally distributed with mean values equal to the true distances $\dtrue$ with a root-mean-square (rms) scatter that scales linearly with \(\dtrue\). Specifically,
\begin{equation}
\label{eq:dest}
\dest = \dtrue \left(e^{\sigma \epsilon -\sigma^2/2}-1\right)\; ,
\end{equation}
where $\epsilon\sim \mathcal{N}(0,1)$. To first order in $\sigma$, this implies an rms
distance error of $\sigma \dtrue$. For distance indicators like the Tully-Fisher relation \citep{brent_tully2015}, \(\sigma\approx 0.1-0.2\).
Given accurately measured redshifts, the ``observed" line-of-sight peculiar velocity is then
% \begin{equation}
% cz = H_0 d^{\text{est}} + v^{\text{est}}_p\; ,
% \end{equation}
% yielding
\begin{equation}
\vest = cz-H_0\dest \; .
\label{eq:vel_est}
\end{equation}
\begin{figure}[ht!]
\centering
\includegraphics[width=1.0\columnwidth]{errored_velocity_vs_distance.png}
\caption{Illustration of inhomogeneous Malmquist bias. Mean line-of-sight velocity profiles \cah\ for a simulated cluster with $M_{\mathrm{v}} = 5\times 10^{14}\,h^{-1}\,\textrm{M}_\odot$. The line-of-sight passes through the cluster center, with the observer located to the left.
The red curve corresponds to true velocity binned by true distance, $\dtrue$;
the orange curve represents the observed velocity, $\vest$, binned by observed distance, $\dest$;
and the blue curve is the mean $\vest$ as a function of the redshift space coordinate $s$.
The cluster center is positioned at \(\dtrue = 50\,h^{-1}\,\mathrm{Mpc}\), and a relative distance error of $20\%$ is assumed in computing $\dest$ and $\vest$.
\label{fig:errored_vel_plot}}
\end{figure}
We train three deep neural networks, each using a different input derived from the catalog of \cah :
\begin{itemize}
\item \textbf{Input \typi:}
The number density of \cah\ on a 2D grid defined by the projected radius $R$ and the redshift space coordinate $s = cz/H_0$.
\item \textbf{Input \typii:}
The mean ``observed'' peculiar velocity $\vest$ of \cah , computed in bins of $R$ and $s$.
\item \textbf{Input \typiii:}
The mean $\vest$ of \cah , computed in bins of $R$ and ``observed'' comoving distance $\dest$.
\end{itemize}
The velocity input fields \typii\ and \typiii\ are affected by the inhomogeneous Malmquist bias (IMB) \cite{malmquist1920,Lynden-Bell1988,Strauss1995}.
On large scales, where the redshift coordinate differs from the true distance only by small-amplitude, incoherent motions, the IMB induced by placing galaxies at their redshifts is negligible \citep{a82,DN10}. However, this is not the case for clusters, where incoherent virial motions are significant, producing prominent FoG distortions and resulting in a strong IMB in the \typii\ input field. The field
\typiii\ is also strongly susceptible to IMB due to the large errors in the distance estimation in particular for the more distant clusters.
\Cref{fig:errored_vel_plot} illustrates the IMB for velocities along a line-of-sight through the center of a simulated cluster.
The mean \emph{true} velocity in bins of \emph{true} distance (red curve) exhibits the familiar infall pattern, with positive (negative) mean velocity on the near (far) side of the cluster.
When probed by \typiii\ (yellow curve), this infall pattern appears severely biased.
The input \typii\ (blue curve) also departs from the true infall pattern; in fact, it flips to an apparent \emph{outflow}.
This inversion is readily understood: galaxies within FoGs that have negative redshifts relative to the cluster center typically have large negative peculiar velocities, skewing the mean toward an outflow signature.
\begin{figure}[ht!]
\centering
\includegraphics[width=1.0\columnwidth]{AE.jpeg}
\caption{Illustration of the Auto-encoder architecture. The numbers inside the blue rectangles are the number of filters in each layer while the numbers outside represent the grids dimension as it passes through the network.
\label{fig:AE}}
\end{figure}
The model architecture is based on a convolutional
autoencoder (AE--CNN), designed to directly map the
two-dimensional input fields to their corresponding virial
mass estimates.
The AE--CNN architecture specifically consists of two key components:
an encoder and a fully connected neural network acting as a bottleneck
layer (illustrated in \cref{fig:AE}). The encoder reduces the dimensionality
of the input data progressively via a sequence of 2D convolutional
layers, interspersed with max pooling operations.
This procedure systematically halves the spatial
resolution while simultaneously increasing the number of feature maps. The resulting low-dimensional features are then processed by
the bottleneck network, which ultimately outputs the estimated virial mass.
The underlying structure of the model is a directed acyclic neural
network composed of an input layer, multiple hidden layers, and an
output layer, which collectively ensure a unidirectional flow of
information. Each node within the hidden layers computes a
nonlinear transformation of a weighted linear combination of nodes from the preceding layer, as described by
\begin{equation}
N_i^{(n)} = A_i^{(n)}\left(\sum_j w_{ij}^{(n,n-1)}N_j^{(n-1)} + b_i^{(n)}\right),
\label{eq:NN}
\end{equation}
where $w_{ij}^{(n,n-1)}$ represents the weights connecting
consecutive layers, and $b_i^{(n)}$ are the bias terms. Given
an input $I$, the network produces a scalar output estimate $\hat{T}$ (in our case it is $\hat{M}_\mathrm{v}$).
Our NN represented in \cref{fig:AE} comprises $1.1\times 10^6$ %1,139,489 free
parameters ($w_{ij}$ and $b_i$ in \cref{eq:NN}.
During training, the model parameters (weights and biases) are
optimized by minimizing a mean squared error (MSE) \emph{loss function}. Training samples are grouped into mini-batches, and the
parameters are iteratively updated based on gradients computed from the loss function averaged over these mini-batches. This optimization
procedure is carried out across multiple epochs, with samples shuffled into new mini-batches at each epoch until convergence is
achieved.
\subsection{Training/validation data}
We use all cluster halos with
\(M \ge 10^{14}\hmsol\),
yielding roughly \(2.6\times 10^{4}\) objects.
Because massive halos are strongly under-represented relative
to their low-mass counterparts (see \cref{table:cluster_count}),
we augment the training set preferentially at the high-mass
end by rotating each cluster $n_\text{rot} $ times about a random axis, where
\[
n_{\text{rot}}
\;=\;
5 \;+\;
5\,
\frac{M-10^{14}}{10^{15}-10^{14}}\; .
\]
This augmentation expands the training set to
\(\simeq 1.3\times 10^{5}\) clusters and improves AE-CNN training
and convergence.
The trade-off is that the augmented sample no longer follows
the true cosmological mass function; the most massive halos are
over-sampled relative to their physical abundance.
We tolerate this bias because the resulting improvement in
training stability and reduction of high-mass errors outweigh
the departure from a realistic prior.
For the redshift space inputs, \typi\ and \typii, we retain only those \cah\ satisfying the criterion
\begin{equation}
\label{eq:eppipss}
\left(\frac{R}{9\,h^{-1}\,\mathrm{Mpc}}\right)^2+\left(\frac{s}{34\,h^{-1}\,\mathrm{Mpc}}\right)^2\leq 1,
\end{equation}
while for observed real space input, \typiii , we only keep \cah\ with
\begin{equation}
\label{eq:ellipsd}
\left(\frac{R}{9\,h^{-1}\,\mathrm{Mpc}}\right)^2+\left(\frac{\dest}{60\,h^{-1}\,\mathrm{Mpc}}\right)^2\leq 1.
\end{equation}
The scale-length \( 34\,h^{-1}\mathrm{Mpc}\) in \cref{eq:eppipss}
is motivated as follows.
Assuming a typical FoG redshift space elongation of \(\sim 10\,\rv\), we require \(\lvert s\rvert \le 11\,\rv\).
For the largest halo in the sample, with \(\rv \simeq 3\hmpc\), this becomes \(\lvert s\rvert \le 34\hmpc\); hence the scale-length $34\hmpc$ adopted in \cref{eq:eppipss}.
By choosing this fixed line-of-sight range, we guarantee that even the most extended FoG distortions in our dataset are fully captured, automatically encompassing any smaller elongations as well, and we do so without imposing any cluster-by-cluster prior on the virial radius.
In deriving \typii\ and \typiii\ inputs we adopt $\sigma =0.2$ in \cref{eq:dest} for the observation distance uncertainty,
corresponding to a relative error of $20\% $. For these inputs, we assume that clusters in the training/validation datasets are centered at a $100\hmpc$ true distance from the observer.
In line with this, for the \typiii\ input, we adopt a larger scale-length of $60\hmpc$ in \cref{eq:ellipsd}, spanning approximately three times the distance error.
For either \typii\ or \typiii\
we randomly retain only \(10\%\) of the \cah .
This yields \(\approx 10\!-\!15\) \cah\ of
low-mass clusters
and
\(\approx 300\!-\!350\) in the case of the most massive clusters.
Each dataset corresponding to three input fields, is split into 80\%, 10\% and 10\%, respectively for training, validation and testing.
\label{sec:trainvalid}
\subsection{Model training and validation}
The MSEs for training and validation across all three models—each corresponding to one of the three data types are shown as functions of training epoch in \cref{fig:CNN_training}. It is encouraging that all AE-CNN have clearly converged. The MSE for training and validation are consistent showing only minor evidence for overfitting only for the case of \typii\ input. We select the final AE‑CNN model parameters at the epochs where the validation loss plateaus or begins to diverge from the training loss, signaling the onset of overfitting, in the figure these chosen epochs are marked with vertical dashed lines.
\begin{figure}[ht!]
\centering
\includegraphics[width=1.0\columnwidth]{training_mse_final.png}
\caption{Training and validation mean squared error (MSE) versus training epoch, for the three neural network models. Solid, thicker lines denote the training MSE, while thinner lines indicate the validation MSE. Vertical dashed lines mark the epochs at which the best-performing models are saved.
\label{fig:CNN_training}}
\end{figure}
\section{True vs inferred \textit{versus} inferred vs true}
\label{sec:interp}
Before presenting the results, we
highlight the fundamental difference between
the quantities returned by MLE and AE-CNN.
The MLE gives the value which maximizes the PDF of data given the underlying parameters. In our case,
\begin{equation}
\log \mhat^{\text{MLE}}
= \operatorname*{arg\,max}_{\log M}
\mathcal{L}\!\bigl[\rv(M),\text{data}\bigr],
\end{equation}
where $\mathcal{L}$ and $\rv(M)$ are defined in \cref{eq:mathcalL,eq:virdef}.
In the limit of a large galaxy sample, the MLE estimator is asymptotically unbiased: across an sub-sample of data sets corresponding to the same true parameter, the average of the inferred MLE values converges to that true value \citep{CoxHinkley1974}.
Therefore,
\begin{equation}
\bigl\langle \log \hat{M}^{\text{MLE}}
\mid \log \mtrue\bigr\rangle =\log \mtrue \; ,
\end{equation}
where here averaging is over sets with similar $\mtrue$.
In general
\begin{equation}
\bigl\langle \log \mtrue\mid \log \mhat^{\text{MLE}}
\bigr\rangle \ne \log \mhat^{\text{MLE}}\; ,
\end{equation}
and special care is required in relating $\log \mhat^{\text{MLE}}$ to the underlying true value.
By contrast, the AE-CNN approximates the posterior mean
\citep[e.g.,][]{Goodfellow2016,veena_large-scale_2023},
\begin{equation}
\log \hat{M}^{\text{AE-CNN}}
= \bigl\langle \log \Mv \mid \text{data} \bigr\rangle \; .
\end{equation}
This implies,
\begin{equation}
\label{eq:nninterp}
\bigl\langle \log \mtrue \mid \log \hat{M}^{\text{AE-CNN}}\bigr\rangle =\log \hat{M}^{\text{AE-CNN}}\; ,
\end{equation}
where the averaging is over all data sets that have approximately the same inferred $\log \hat{M}^{\text{AE-CNN}}$.
In other words, \cref{eq:nninterp} implies that $\log \mtrue $ equals $ \log \hat{M}^{\text{AE-CNN}}$ plus a random noise.
\section{Results}
\label{sec:results}
We have applied the MLE and AE-CNN methodologies to the simulated-cluster data.
The MLE is applied on all clusters with mass exceeding
$2.5\times10^{14}\hmsol$, however to reduce CPU time, only a random half of the halos in the lowest-mass bin, $(1-2.5)\times10^{14}\hmsol$, are considered.
The three trained AE-CNN models corresponding to the \typi, \typii, and \typiii\ inputs have then applied to their respective testing data sets.
\Cref{fig:scatters} summarizes the results by means of a comparison of the logarithm of the true cluster masses,
$\log M^\text{true}$, versus the logarithm of the inferred masses, $\log \hat{M}_\mathrm{v}$.
In each panel, the \pcolor\ points correspond individual
clusters.
In the top row, both the MLE and AE-CNN utilize the redshift space distribution of
\cah\ and can therefore be directly compared.
The behaviour of the
linear best-fit regression lines (red and blue in each panel) follows the interpretation given in \S\ref{sec:interp}.
For the MLE (top-left panel), the regression of
$\log \mtrue$ on
$\log \mhat$ (red line) tilts noticeably
away from the one-to-one line (dashed).
By contrast, for the AE-CNN with \typi\ input (top-right panel),
the regression of
$\log \mtrue$ on
$\log \mhat$ (red line) lies almost exactly on the one-to-one line.
The inverse relation shows the opposite pattern: the
regression of $\log \mhat$ on $\log \mtrue$
(blue line) is nearly unity for the MLE, whereas for the AE-CNN it departs appreciably from the diagonal.
The rms scatter of the deviations from the linear regression models is represented by the shaded areas attached to each regression line.
Since the blue band represents scatter in $\log \mhat$ given $\log \mtrue$, it is the \textit{horizontal} thickness of the band that corresponds to the rms scatter.
Conversely, it is the \textit{vertical} thickness of the red band that reflects the rms scatter in $\log \mtrue$ given $\log \mhat$.
The scatter, both vertical and horizontal, is significantly larger in the MLE than in the AE-CNN (top-right panel) by $\sim30\%$.
The performance of the AE-CNN model with \typi\ input is impressive, with an rms scatter of approximately $0.1\,\textrm{dex}$ from the red line across all of the mass range, which lies almost exactly along the one-to-one mapping.
The tight correlation between $\log \mtrue$ and $\log \mhat$ leads to only a small deviation from the
blue line representing the linear regression of $\log \mhat$
on $\log \mtrue$, though the scatter is somewhat larger as seen by comparing the red and blue shaded areas.
%%%
The bottom two panels of the figure display the AE-CNN results for the \typii\ and \typiii\ inputs, as indicated in the panels.
In both panels, the linear regressions of $\log \mtrue$ on $\log \mhat$ (red lines) are close to the diagonal, as it should be.
As seen by the widths of the red shaded areas in both panels, the
scatter in the \typii\ model (left panel) ranges from $0.7\,\textrm{dex}$ to $0.12\,\textrm{dex}$ and is roughly $\sim30\%$ larger than the \typii\ model (right panel) across all mass bins.
\Cref{fig:errored_vel_plot} sheds light on the reason: in the \typiii\ case, the Malmquist bias in the $\vest$-$\dest$ relation (yellow curve) \emph{amplifies} the genuine infall signal, whereas in the \typii\ input the bias in $\vest$ versus $s$ (blue curve) \emph{weakens} the signal and even reverses it, producing an apparent outflow. Thus, effectively the ``signal-to-noise" ratio is significantly reduced in the \typii\ input.
% This quantity is plotted computed in bins of versus $\mtrue $ \cref{fig:average_and_error} \adiblu{This is given versus $\mtrue $ but the binning should be on $\log \mhat$ to be consistent with the scatter figure.}
% The error bars in the figure represent the rms scatter deviation from this mean. The MLE method clearly has the largest uncertainties among all four approaches. This is expected, as the universality assumption underlying MLE presumes that all clusters share similar shapes and dynamical structures. In reality, most clusters deviate from spherical symmetry---due to elongation, substructure, or projection effects---leading to systematic errors when minimizing the likelihood function based on this oversimplified model.
Apart from the left bottom panel which shows only a single point with $\log \mhat <14$, the remaining AE-CNN models in the right column yield $\log \mhat > 14$.
This is an evidence of the high quality of these models which where trained on clusters with $\log \mtrue >14$ training data. There is no apriori guarantee that they would satisfy this limit when applied on the testing data or actually even on the validation data itself.
\begin{figure}[ht!]
\centering
\includegraphics[width=1.0\columnwidth]{scatters_linear_reg_with_shaded.png}
\caption{
Comparison of $\log \mtrue$ versus $\log \mhat$ inferred from the MLE and AE-CNN models trained on \typi, \typii, and \typiii\ inputs. The scattered points correspond to individual clusters. The solid red line is the best‑fit linear regression line of $\log \mtrue$ against $\log \mhat$ (true versus inferred), and the solid blue line is the best‑fit line for $\log \mhat$ against $\log \mtrue$ (inferred versus true).
Shaded area represent the scatter rms around these lines.
The one-to-one mapping is plotted as the dashed black line.
\label{fig:scatters}}
\end{figure}
\section{Discussion and Conclusions}
\label{sec:discus}
The application of the MLE presented here assumes that the distribution and kinematics of \cah\ (halos around and inside clusters) follows a universal form when scaled appropriately by the virial radius. This scaling is clearly approximate, with deviations from universality more evident in galaxy distributions than in kinematics, reflecting a stronger mass dependence in the former.
Nonetheless, the method can readily accommodate a much weaker assumption where the mean distributions and kinematics of galaxies associated with a cluster depend on its mass (or virial radius), as long as deviations of individual clusters (of the same mass) are relatively small.
The method can be applied similarly to what has been done here so far. Naturally, hydrodynamical simulations that incorporate galaxy formation can be used. Alternatively, dark matter-only simulations implemented with semi-analytic galaxy formation models can be utilized \citep[e.g.,][]{KNS,Benson2012}.
Despite adopting explicit assumptions verified by simulations, the MLE produced significantly larger scatter between true and estimated $\log \Mv $ than the AE-CNN on the same data set, namely the distribution of \cah\ in redshift space.
We have also trained AE-CNN models using observed velocity data for only $10\%$ of \cah. This constitutes a challenging test due to the presence of Malmquist bias in the flow pattern.
In this case, only velocity information is used, without incorporating the spatial distribution of the halos. Moreover, the input to the network consists solely of mean velocities in bins (in either redshift space or observed distance space), meaning the NNs receive no information about the degree of virial motions within clusters.
The model trained on distribution of \cah\ in redshift space achieving the best overall performance. It shows the smallest scatter and tightest error bars across the mass range. The other two NN models, trained on ``observed" velocities in redshift space and in observed distance-space, also yield good results but with larger errors.
We do not apply either of the methods described here on actual observational data.
We point out however, that an application observational data, we will assume that a catalog of groups/clusters is already compiled \citep[e.g.,][]{Tempel2016}. The methods utilize the distribution of all galaxies in the vicinity of the cluster. Since the goal of the methods is actually to estimate the virial radius and mass, we propose to include all galaxies around clusters out to fixed projected distance and line-of-sight extent, without the need for explicit identification of cluster members or interlopers.
Comparing the implementation of MLE and NN approaches can be viewed as a trade-off between prior knowledge and flexibility. MLE exploits an explicit statistical model to describe the data \cite[c.f.,][]{2014MNRAS.442.1887F,old2014}, it is computationally efficient, and provides transparent uncertainty estimates, but it is only as good as the assumed model. NNs offer immense flexibility to learn any relationship from the data, excelling in scenarios with complex patterns and large datasets; they require heavy computation and careful training.
We have also clarified a fundamental distinction between the MLE and AE-CNN estimates. Although both are unbiased, they are so in different senses.
The AE-CNN estimate corresponds to the mean of the posterior—i.e., the mean of all true values of the target quantity (in our case, $\log M$) conditioned on the estimate.
In contrast, the MLE returns the mean of the estimates conditioned on the true values.
Since the true values are not available in practice, we regard the AE-CNN as the more relevant approach.
While it is possible, in principle, to augment the MLE with a prior PDF to convert it into a posterior estimator, the key advantage of the AE-CNN is that it does not require an explicit form of the posterior.
Instead, it learns to extract the relevant information directly from training data that reflect the full complexity of real observations.
%\section{Software and third party data repository citations} \label{sec:cite}
\section{Acknowledgments}
We thank Punyakoti Ganeshaiah Veena for useful discussions. This research has been supported by a grant from the Israel Science Foundation and the Asher Space Research Institute.
The CosmoSim database used in this paper is a service by the Leibniz-Institute for Astrophysics Potsdam (AIP).
The MultiDark database was developed in cooperation with the Spanish MultiDark Consolider Project CSD2009-00064.
% \vspace{5mm}
% \facilities{HST(STIS), Swift(XRT and UVOT), AAVSO, CTIO:1.3m,
% CTIO:1.5m,CXO}
%% Similar to \facility{}, there is the optional \software command to allow
%% authors a place to specify which programs were used during the creation of
%% the manuscript. Authors should list each code and include either a
%% citation or url to the code inside ()s when available.
%\software{astropy \citep{2013A&A...558A..33A,2018AJ....156..123A}}
\bibliography{sample631}{}
\bibliographystyle{aasjournal}
%% This command is needed to show the entire author+affiliation list when
%% the collaboration and author truncation commands are used. It has to
%% go at the end of the manuscript.
%\allauthors
%% Include this line if you are using the \added, \replaced, \deleted
%% commands to see a summary list of all changes at the end of the article.
%\listofchanges
\end{document}
% End of file `sample631.tex'.
\begin{thebibliography}{}
\expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
\providecommand{\url}[1]{\href{#1}{#1}}
\providecommand{\dodoi}[1]{doi:~\href{http://doi.org/#1}{\nolinkurl{#1}}}
\providecommand{\doeprint}[1]{\href{http://ascl.net/#1}{\nolinkurl{http://ascl.net/#1}}}
\providecommand{\doarXiv}[1]{\href{https://arxiv.org/abs/#1}{\nolinkurl{https://arxiv.org/abs/#1}}}
\bibitem[{Aaronson {et~al.}(1982)Aaronson, Huchra, Mould, Schechter, \&
Tully}]{a82}
Aaronson, M., Huchra, J., Mould, J., Schechter, P.~L., \& Tully, R.~B. 1982,
The Astrophysical Journal, 258, 64, \dodoi{10.1086/160053}
\bibitem[{{Allingham} {et~al.}(2024){Allingham}, {B{\oe}hm}, {Eckert},
{Jauzac}, {Lagattuta}, {Mahler}, {Hilton}, {Lewis}, \& {Ettori}}]{allingham}
{Allingham}, J. F.~V., {B{\oe}hm}, C., {Eckert}, D., {et~al.} 2024, \mnras,
528, 1711, \dodoi{10.1093/mnras/stae107}
\bibitem[{Behroozi {et~al.}(2013)Behroozi, Wechsler, \& Wu}]{Behroozi2013}
Behroozi, P.~S., Wechsler, R.~H., \& Wu, H.-Y. 2013, The Astrophysical Journal,
762, 109, \dodoi{10.1088/0004-637X/762/2/109}
\bibitem[{Benson(2012)}]{Benson2012}
Benson, A.~J. 2012, New Astronomy, 17, 175,
\dodoi{10.1016/j.newast.2011.07.004}
\bibitem[{Brent~Tully(2015)}]{brent_tully2015}
Brent~Tully, R. 2015, The Astronomical Journal, 149, 54,
\dodoi{10.1088/0004-6256/149/2/54}
\bibitem[{Bryan \& Norman(1998)}]{Bryan1998}
Bryan, G.~L., \& Norman, M.~L. 1998, The Astrophysical Journal, 495, 80,
\dodoi{10.1086/305262}
\bibitem[{{Carlberg} {et~al.}(1997){Carlberg}, {Yee}, {Ellingson}, {Morris},
{Abraham}, {Gravel}, {Pritchet}, {Smecker-Hane}, {Hartwick}, {Hesser},
{Hutchings}, \& {Oke}}]{Carlberg97}
{Carlberg}, R.~G., {Yee}, H.~K.~C., {Ellingson}, E., {et~al.} 1997, \apjl, 485,
L13, \dodoi{10.1086/310801}
\bibitem[{{Caro} {et~al.}(2024){Caro}, {de Andres}, {Cui}, {Yepes}, {De
Petris}, {Ferragamo}, {Schiltz}, \& {Nef}}]{Caro2024}
{Caro}, A., {de Andres}, D., {Cui}, W., {et~al.} 2024, arXiv e-prints,
arXiv:2410.04229, \dodoi{10.48550/arXiv.2410.04229}
\bibitem[{Cox \& Hinkley(1974)}]{CoxHinkley1974}
Cox, D.~R., \& Hinkley, D.~V. 1974, Theoretical Statistics, 1st edn. (New York:
Chapman and Hall/CRC), 528, \dodoi{10.1201/b14832}
\bibitem[{Davis {et~al.}(2011)Davis, Nusser, Masters, Springob, Huchra, \&
Lemson}]{DN10}
Davis, M., Nusser, A., Masters, K.~L., {et~al.} 2011, MNRAS, 413, 2906,
\dodoi{10.1111/j.1365-2966.2011.18362.x}
\bibitem[{{Diaferio}(1999)}]{1999MNRAS.309..610D}
{Diaferio}, A. 1999, \mnras, 309, 610, \dodoi{10.1046/j.1365-8711.1999.02864.x}
\bibitem[{{Falco} {et~al.}(2014){Falco}, {Hansen}, {Wojtak}, {Brinckmann},
{Lindholmer}, \& {Pandolfi}}]{2014MNRAS.442.1887F}
{Falco}, M., {Hansen}, S.~H., {Wojtak}, R., {et~al.} 2014, \mnras, 442, 1887,
\dodoi{10.1093/mnras/stu971}
\bibitem[{Goodfellow {et~al.}(2016)Goodfellow, Bengio, \&
Courville}]{Goodfellow2016}
Goodfellow, I., Bengio, Y., \& Courville, A. 2016, Deep Learning (MIT Press)
\bibitem[{{Iqbal} {et~al.}(2023){Iqbal}, {Pratt}, {Bobin}, {Arnaud}, {Rasia},
{Rossetti}, {Duffy}, {Bartalucci}, {Bourdin}, {De Luca}, {De Petris},
{Donahue}, {Eckert}, {Ettori}, {Ferragamo}, {Gaspari}, {Gastaldello},
{Gavazzi}, {Ghizzardi}, {Lovisari}, {Mazzotta}, {Maughan}, {Pointecouteau},
\& {Sereno}}]{Iqbal2023}
{Iqbal}, A., {Pratt}, G.~W., {Bobin}, J., {et~al.} 2023, \aap, 679, A51,
\dodoi{10.1051/0004-6361/202347234}
\bibitem[{Kauffmann {et~al.}(1995)Kauffmann, Nusser, \& Steinmetz}]{KNS}
Kauffmann, G., Nusser, A., \& Steinmetz, M. 1995, MNRAS, 286, 795,
\dodoi{10.1093/mnras/286.4.795}
\bibitem[{{Lotz} {et~al.}(2017){Lotz}, {Koekemoer}, {Coe}, {Grogin}, {Capak},
{Mack}, {Anderson}, {Avila}, {Barker}, {Borncamp}, {Brammer}, {Durbin},
{Gunning}, {Hilbert}, {Jenkner}, {Khandrika}, {Levay}, {Lucas}, {MacKenty},
{Ogaz}, {Porterfield}, {Reid}, {Robberto}, {Royle}, {Smith},
{Storrie-Lombardi}, {Sunnquist}, {Surace}, {Taylor}, {Williams}, {Bullock},
{Dickinson}, {Finkelstein}, {Natarajan}, {Richard}, {Robertson}, {Tumlinson},
{Zitrin}, {Flanagan}, {Sembach}, {Soifer}, \& {Mountain}}]{LotzLensing}
{Lotz}, J.~M., {Koekemoer}, A., {Coe}, D., {et~al.} 2017, \apj, 837, 97,
\dodoi{10.3847/1538-4357/837/1/97}
\bibitem[{Lucie-Smith {et~al.}(2022)Lucie-Smith, Adhikari, \&
Wechsler}]{Luisa2022}
Lucie-Smith, L., Adhikari, S., \& Wechsler, R.~H. 2022, Monthly Notices of the
Royal Astronomical Society, 515, 2164,
\dodoi{https://doi.org/10.1093/mnras/stac1833}
\bibitem[{Lynden-Bell {et~al.}(1988)Lynden-Bell, Faber, Burstein, Davies,
Dressler, Terlevich, \& Wegner}]{Lynden-Bell1988}
Lynden-Bell, D., Faber, S.~M., Burstein, D., {et~al.} 1988, ApJ, 326, 19,
\dodoi{10.1086/166066}
\bibitem[{Malmquist(1920)}]{malmquist1920}
Malmquist, K.~G. 1920, Meddelanden fran Lunds Astronomiska Observatorium, Serie
II, 22
\bibitem[{Mamon {et~al.}(2013)Mamon, Biviano, \& Boué}]{Mamon2013}
Mamon, G.~A., Biviano, A., \& Boué, G. 2013, MNRAS, 429, 3079,
\dodoi{10.1093/mnras/sts565}
\bibitem[{{Marini} {et~al.}(2025){Marini}, {Popesso}, {Dolag}, {Bravo},
{Robotham}, {Tempel}, {Li}, {Yang}, {Csizi}, {Behroozi}, {Biffi}, {Biviano},
{Lamer}, {Malavasi}, {Mazengo}, \& {Toptun}}]{Marini2025}
{Marini}, I., {Popesso}, P., {Dolag}, K., {et~al.} 2025, A\&A, 694, A207,
\dodoi{10.1051/0004-6361/202452028}
\bibitem[{Old {et~al.}(2014)Old, Skibba, Pearce, Croton, Muldrew,
Muñoz-Cuartas, \& Wojtak}]{old2014}
Old, L., Skibba, R.~A., Pearce, F.~R., {et~al.} 2014, Monthly Notices of the
Royal Astronomical Society, 441, 1513, \dodoi{10.1093/mnras/stu708}
\bibitem[{{Pizzardo} {et~al.}(2024){Pizzardo}, {Diaferio}, \&
{Rines}}]{2024A&A...682A..80P}
{Pizzardo}, M., {Diaferio}, A., \& {Rines}, K.~J. 2024, \aap, 682, A80,
\dodoi{10.1051/0004-6361/202244448}
\bibitem[{Pizzardo {et~al.}(2023)Pizzardo, Geller, Kenyon, Damjanov, \&
Diaferio}]{geller2023}
Pizzardo, M., Geller, M.~J., Kenyon, S.~J., Damjanov, I., \& Diaferio, A. 2023,
Astronomy \& Astrophysics, 675, A56, \dodoi{10.1051/0004-6361/202346545}
\bibitem[{Riebe {et~al.}(2013)Riebe, Kristina, Partl, M., Enke, Harrya,
Forero-Romero, Jaimea, Gottlober, Stefan, Klypin, Anatoly, Lemson, Gerard,
Prada, Francisco, Primack, R, Steinmetz, Matthiasa, Turchaninov, \&
Victore}]{mdpl...793..127V}
Riebe, Kristina, Partl, {et~al.} 2013, Astronomische Nachrichten, 334, 691,
\dodoi{10.1002/asna.201211900}
\bibitem[{Strauss \& Willick(1995)}]{Strauss1995}
Strauss, M., \& Willick, J. 1995, Physics Reports, 261, 271,
\dodoi{10.1016/0370-1573(95)00013-7}
\bibitem[{{Tempel} {et~al.}(2016){Tempel}, {Kipper}, {Tamm}, {Gramann},
{Einasto}, {Sepp}, \& {Tuvikene}}]{Tempel2016}
{Tempel}, E., {Kipper}, R., {Tamm}, A., {et~al.} 2016, \aap, 588, A14,
\dodoi{10.1051/0004-6361/201527755}
\bibitem[{Veena {et~al.}(2023)Veena, Lilow, \& Nusser}]{veena_large-scale_2023}
Veena, P.~G., Lilow, R., \& Nusser, A. 2023, Monthly Notices of the Royal
Astronomical Society, 522, 5291, \dodoi{10.1093/mnras/stad1222}
\end{thebibliography}