Wasserstein Regression with Empirical Measures and Density Estimation for Sparse Data
Extracted source text · arXiv 2308.12540
\documentclass[12pt]{article}
%\pdfoutput=1
\input{macros}
%% Useful packages
\usepackage{tabularx}% Tabulars with adjustable-width columns
\usepackage{longtable}% allow tables to flow over page boundaries
%\usepackage[colorlinks=true, allcolors=blue]{hyperref}
\usepackage{tikz-cd}% commutative diagrams with TikZ
% http://ctan.math.washington.edu/tex-archive/graphics/pgf/contrib/tikz-cd/tikz-cd-doc.pdf
%\usepackage{optidef}% formulate optimization problem
\usepackage[linesnumbered,ruled,vlined]{algorithm2e}
\def\T{{ \mathrm{\scriptscriptstyle T} }}
\title{Wasserstein Regression with Empirical Measures and Density Estimation for Sparse Data}
\author{Yidong Zhou and Hans-Georg M\"{u}ller\\
Department of Statistics, University of California, Davis\\
Davis, CA 95616, USA}
\date{July 27, 2023}
\begin{document}
\maketitle
\begin{abstract}
The problem of modeling the relationship between univariate distributions and one or more explanatory variables lately has found increasing interest. Traditional functional data methods cannot be applied directly to distributional data because of their inherent constraints. Modeling distributions as elements of the Wasserstein space, a geodesic metric space equipped with the Wasserstein metric that is related to optimal transport, is attractive for statistical applications. Existing approaches proceed by substituting proxy estimated distributions for the typically unknown response distributions. These estimates are obtained from available data but are problematic when for some of the distributions only few data are available. Such situations are common in practice and cannot be addressed with currently available approaches, especially when one aims at density estimates. We show how this and other problems associated with density estimation such as tuning parameter selection and bias issues can be side-stepped when covariates are available. We also introduce a novel version of distribution-response regression that is based on empirical measures. By avoiding the preprocessing step of recovering complete individual response distributions, the proposed approach is applicable when the sample size available for each distribution varies and especially when it is small for some of the distributions but large for others. In this case, one can still obtain consistent distribution estimates even for distributions with only few data by gaining strength across the entire sample of distributions, while traditional approaches where distributions or densities are estimated individually fail, since sparsely sampled densities cannot be consistently estimated. The proposed model is demonstrated to outperform existing approaches through simulations. Its efficacy is corroborated in two case studies on Environmental Influences on Child Health Outcomes (ECHO) data and eBay auction data.
\bigskip
\noindent KEY WORDS: distributional data analysis, \f mean, multi-cohort study, optimal transport, sample of distributions, Wasserstein distance.
\end{abstract}
\double
\section{Introduction}
Data consisting of samples of univariate probability distributions are increasingly prevalent across various research areas. Data that contain distributions as a basic unit of observation are encountered in the analysis of mortality \citep{chen:23:1,ghod:21}, population pyramids \citep{hron:16, bigo:17}, brain connectivity \citep{mull:16:1}, and financial returns \citep{zhan:22, mull:23:2}, among many other applications. This has led to the emerging field of distributional data analysis \citep{pete:22}. Distributions, represented as probability density functions, cumulative distribution functions or quantile functions, come with inherent constraints that are not present in traditional functional data. The space where distributions are situated is not a vector space, and linear methods developed for functional data cannot be directly applied. Various approaches have been proposed to address this challenge, including global transformations of univariate distributions to a Hilbert space \citep{hron:16, mull:16:1}, which however do not take the geometry of the space of distributions into account and are not isometric. % and lead to distortions that change distances between pairs of distributions.
More recently, attention has focused on directly modeling distributions as elements of the Wasserstein space, a geodesic metric space related to optimal transport \citep{vill:03,pana:20}. The pseudo-Riemannian structure of this space can be used to construct isometric exponential maps from the space to tangent bundles, where linear operations can be deployed \citep{bigo:17, chen:23:1}; however the inverse log map is not defined on the entire tangent space, which causes substantial difficulties and requires an ad-hoc constraint or projection step \citep{flet:13, pego:22}.
A commonly encountered problem is to model the relationship between distributions and one or more explanatory variables. Distribution-response regression problems arise in many modern data analysis settings. Examples include neuroimaging, where varying patterns of brain connectivity distributions across age and other clinical covariates are of interest \citep{mull:16:1}, metabolomics, where the aim is to model the dependence of a metabolite distribution on birth weight \citep{tals:18}, and also mortality where one is interested in the dependence of age-at-death distributions on economic indicators \citep{pete:22}. Using the Wasserstein metric, distribution-response regression can be implemented as a special case of \f regression \citep{mull:19:3}, which models the relationship between random objects that lie in a generic metric space as responses and scalar or vector covariates as predictors.
In line with all local and global transformation methods, \f regression requires that each distribution is observed as a distributional object. This is however usually unrealistic in practice as one rarely observes data samples where the atoms of the samples are entire distributions. Instead, one usually has samples of independent data $\{Y_{ij}\}_{j=1}^{N_i}$ that are generated by the distribution $\nu_i$ for each $i$; see Figure~\ref{fig:diagram} in Section \ref{subsec:imp} for a demonstration. To address this, current approaches for virtually all distributional data analysis methods include a prior distribution estimation step where one substitutes a smooth estimate for the unobservable distributional objects. Commonly this is done through a density estimate such as a kernel estimate or smoothed histogram that is obtained from data $\{Y_{ij}\}_{j=1}^{N_i}$ \citep{pana:16, bigo:18, mull:19:3} or a quantile function obtained by smoothing the empirical quantile function \citep{pete:21, mull:21:10}. %In the framework of Cox point process regression, \citet{mull:21:10} adopts \f regression to estimate the shape part of the conditional intensity function.
In the current literature, %either the smoothing bias is not taken into account in the asymptotic theory or
strong assumptions are used to ensure uniform convergence of density estimates, uniform across all random distributions \citep{pana:16,mull:16:1, pete:21, mull:21:10}. This preliminary smoothing step %, in particular, necessitates
requires the minimum number of observations $m=m(n)$ that are generated by each random distribution, where $n$ is the number of random distributions, to increase to infinity at a fast rate, typically faster than $n$. This entails convergence of the density estimation step at a rate that is faster than that of the subsequent \f regression or other operation so that the preliminary estimation step can be ignored in the overall asymptotic analysis. However, the assumption of a fast increase in the number of observations made for each of the distributions across the board is often unrealistic and the need to choose a tuning parameter for each distribution is another downside.
This state of affairs motivates the approach proposed in this paper, namely to use empirical measures $\hat{\nu}_i=(1/N_i)\sum_{j=1}^{N_i}\delta_{Y_{ij}}$, where $N_i \ge 1$ and $\delta_{Y_{ij}}$ denotes the Dirac measure at $Y_{ij}$, i.e., using the observations made for each random distribution directly, rather than first forming density estimates or other distributional estimates when conducting \f regression. Besides avoiding the preliminary distribution estimation step with all its downsides, we will show that this has the further major benefit that it leads to a consistent density estimate for a distribution for which one has only very few observations, where the number of observations may not even increase with $n$ and density estimation would normally be a hopeless enterprise. This becomes possible by harnessing the data across all distributions and exploiting the assumed smooth dependency of the response distributions on the predictors.%, with details below.
The proposed regression model, implemented through least common multiples and supported with asymptotic theory, avoids smoothing bias and tuning parameter choice in the pre-smoothing step and hence is more broadly applicable in practice, especially when the available sample size for each distribution greatly varies \citep{qiu:22:2}. A typical example is provided by multi-cohort studies \citep{ocon:22}, where the availability and cost of observations highly varies across cohorts \citep{bone:14}. We illustrate such a scenario with Environmental influences on Child Health Outcomes (ECHO) data \citep{knap:23}, where the dependence of body mass index distributions for children across cohorts on demographic covariates is of interest. A second example is provided by point processes, where distributions of event arrival times are inherently discretely observed with heterogeneous sample sizes. This scenario is demonstrated with eBay auction data \citep{jank:10}, where we investigate the relationship between the bid time distribution and the opening bid.
The main innovations and strengths of the proposed approach are as follows. First, we illustrate that the universally employed pre-smoothing step is superfluous for distribution-response \f regression; omitting this step avoids initial smoothing bias and tuning parameter selection and also avoids the requirement of a rapidly increasing number of observations per distribution. Second, we obtain consistent density estimates for densities with sparse number of observations by borrowing strength across the sample. Third, the proposed regression model is shown to outperform the existing smoothing approach in finite sample situations. Fourth, the proposed methods are supported by theory, including pointwise and uniform rates of convergence. Fifth, we demonstrate the utility and flexibility of the proposed distribution-response regression model.
The rest of the paper is organized as follows. In Section \ref{sec:pre}, we introduce some basic background and notations for the Wasserstein space of distributions. The proposed regression model for empirical measure responses and vector covariates is introduced in Section \ref{sec:emr}. Pointwise and uniform rates of convergence for the estimators are established in Section \ref{sec:th}. Computational details and simulation results are presented in Section 5. The proposed framework is illustrated in Section 6 using the ECHO and eBay auction data. %Finally, we conclude with a brief discussion presented in Section 7.
Detailed theoretical proofs are in the Appendix.
\section{Preliminaries}
\label{sec:pre}
For a closed interval $\Omega$ %be a closed interval of $\mathbb{R}$ and denote the
with Borel $\sigma$-algebra $\mathcal{B}(\Omega)$, we denote the set of probability measures, also referred to as measures or distributions, $\mu$ with domain $\Omega$ as $\mathcal{P}(\Omega)$ and define the space of measures with finite second moments as %, which are equivalently referred to simply as measures or distributions, is denoted as %by
%$\mathcal{W}=\mathcal{W}_2(\Omega)$,
\begin{equation*}
\mathcal{W}:=\left\{\mu\in \mathcal{P}(\Omega); \int_\Omega x^2\mu(dx)<\infty\right\}.
\end{equation*}
%Here $\mathcal{P}(\Omega)$ is the space of probability measures over $\{\Omega, \mathcal{B}(\Omega)\}$.
For any measure $\mu\in\mathcal{W}$ with cumulative distribution function $F_{\mu}$, we consider the quantile function $F_{\mu}^{-1}$ to be the left continuous inverse of $F_{\mu}$, i.e., $F_{\mu}^{-1}(\alpha)=\inf\{x\in\Omega: F_{\mu}(x)\geq\alpha\}$, for $\alpha\in(0, 1)$. The space $\mathcal{W}$ is a metric space with the 2-Wasserstein, or simply Wasserstein distance between two measures $\mu_1, \mu_2\in\mathcal{W}$ defined as
\[d_\mathcal{W}^2(\mu_1, \mu_2):=\inf_{\pi\in\Pi(\mu_1, \mu_2)}\int_{\Omega\times\Omega}|x-y|^2d\pi(x, y),\]
where $\Pi(\mu_1, \mu_2)$ is the set of joint measures on $\Omega\times\Omega$ with marginals $\mu_1$ and $\mu_2$ \citep{kant:42}. It is well known %By Theorem 2.18 of
\citep{vill:03} that the Wasserstein distance can be expressed as the $L^2$ distance between the corresponding quantile functions,
\begin{equation}
\label{eq:dw}
d_\mathcal{W}^2(\mu_1, \mu_2)=\int_0^1\{F_{\mu_1}^{-1}(\alpha)-F_{\mu_2}^{-1}(\alpha)\}^2d\alpha.
\end{equation}
It can be shown that $\mathcal{W}$ endowed with $d_{\mathcal{W}}$ is a complete and separable metric space, the Wasserstein space \citep{pana:20, vill:03}. % for more details about Wasserstein space and its connection to optimal transport theory.
Consider a random element $\nu$ taking values in the Wasserstein space $\mathcal{W}$, assumed to be square integrable in the sense that $E\{d_{\mathcal{W}}^2(\nu, \mu)\}<\infty$ for %some (and thus
all $\mu\in\mathcal{W}$. The \f mean of $\nu$ \citep{frec:48}, extending the usual notion of mean, is
\[\nu_{\oplus}=\argmin{\mu\in\mathcal{W}}E\{d_{\mathcal{W}}^2(\nu, \mu)\},\]
which is well-defined and unique as the Wasserstein space $\mathcal{W}$ is a Hadamard space \citep{kloe:10}. It follows from \eqref{eq:dw} that the quantile function of the \f mean $\nu_{\oplus}$ is
$F^{-1}_{\nu_{\oplus}}(\cdot)=E\{F_{\nu}^{-1}(\cdot)\}.$
In the following, the notation $\mu$ refers to fixed measures in $\mathcal{W}$, $\nu$ to random measures, $a\lesssim b$ means that there exists a positive constant $C$ such that $a\leq Cb$ and $a\asymp b$ that $a\lesssim b$ and $b\lesssim a$. The Euclidean norm in $\mathbb{R}^p$ is denoted by $\|\cdot\|_E$.
\section{Wasserstein Regression with Empirical Measures}
\label{sec:emr}
Let $(Z, \nu)$ be a random pair with joint distribution $\mathcal{F}$ on the product space $\mathbb{R}^p\times\mathcal{W}$. Denote the mean and variance of $Z$ by $\theta=E(Z)$ and $\Sigma=\Var(Z)$, with $\Sigma$ strictly positive definite. To model the regression relation between random measures $\nu$ and vector covariates $Z$, we adopt the framework of \f regression, a version of conditional \f means designed for the regression of metric space-valued responses on Euclidean predictors \citep{mull:19:3, mull:22:8}. \f regression targets the conditional \f mean of $\nu$ given $Z=z$,
\begin{equation*}
m(z)=\argmin{\mu\in\mathcal{W}}E\{d_{\mathcal{W}}^2(\nu, \mu)|Z=z\}.
\end{equation*}
A brief description of \f regression is as follows; for more details see \citet{mull:19:3}. Suppose that $\{(Z_i, \nu_i)\}_{i=1}^n$ are $n$ independent realizations of $(Z, \nu)$ with
\[\bar{Z}=\frac{1}{n}\sum_{i=1}^nZ_i,\quad \hat{\Sigma}=\frac{1}{n}\sum_{i=1}^n(Z_i-\bar{Z})(Z_i-\bar{Z})^\T.\]
Global \f regression can be considered as a generalization of classical multiple linear regression for real-valued responses and targets
\begin{equation}
\label{eq:mg}
m_G(z)=\argmin{\mu\in\mathcal{W}}E\{s_G(z)d_{\mathcal{W}}^2(\nu, \mu)\},
\end{equation}
with weight function $s_G(z)=1+(Z-\theta)^\T\Sigma^{-1}(z-\theta)$ and empirical version
\begin{equation}
\label{eq:hatmg}
\tilde{m}_G(z)=\argmin{\mu\in\mathcal{W}}\frac{1}{n}\sum_{i=1}^ns_{iG}(z)d^2_{\mathcal{W}}(\nu_i, \mu),
\end{equation}
for a sample $\{(Z_i, \nu_i)\}_{i=1}^n$ , where $s_{iG}(z)=1+(Z_i-\bar{Z})^\T\hat{\Sigma}^{-1}(z-\bar{Z})$.
Analogously, local \f regression extends local linear regression to metric space-valued responses. For the special case of a scalar predictor $Z\in\mathbb{R}$ it targets
\begin{equation}
\label{eq:ml}
m_{L, h}(z)=\argmin{\mu\in\mathcal{W}}E\{s_{L}(z, h)d_{\mathcal{W}}^2(\nu, \mu)\},
\end{equation}
where $s_L(z, h)=K_h(Z-z)[u_2-u_1(Z-z)]/\sigma_0^2$, $u_j=E[K_h(Z-z)(Z-z)^j]$ for $j=0, 1, 2$, $\sigma_0^2=u_0u_2-u_1^2$, and $K_h(\cdot)=h^{-1}K(\cdot/h)$ with $K(\cdot)$ a continuous symmetric probability density function on $[-1, 1]$ and $h$ a bandwidth and with empirical version
\begin{equation}
\label{eq:hatml}
\tilde{m}_{L, h}(z)=\argmin{\mu\in\mathcal{W}}\frac{1}{n}\sum_{i=1}^ns_{iL}(z, h)d_{\mathcal{W}}^2(\nu_i, \mu).
\end{equation}
Here $s_{iL}(z, h)=K_h(Z_i-z)[\hat{u}_2-\hat{u}_1(Z_i-z)]/\hat{\sigma}_0^2$, where $\hat{u}_j=n^{-1}\sum_{i=1}^nK_h(Z_i-z)(Z_i-z)^j$ for $j=0, 1, 2$ and $\hat{\sigma}_0^2=\hat{u}_0\hat{u}_2-\hat{u}_1^2$.
In previous work on Fr\'echet regression \citep{mull:19:3}, response distributions in the Wasserstein space $\mathcal{W}$ served as one of the key examples but it has been typically assumed that the measures are fully observed while at the same time they may be randomly perturbed in analogy to usual additive noise models \citep{mull:22:8}. However none of this applies in the more realistic situation where one has data $\{Y_{ij}\}_{j=1}^{N_i}$ that are generated from each distribution $\nu_i$. The stopgap solution applied previously is a preprocessing density estimation step \citep{pana:16,mull:16:1}, where it is assumed that the number of data available for all measures increases at a faster rate than the available number of observation points $(Z_i,\nu_i)$ in order to preserve asymptotic convergence rates.
However the intermediate kernel density estimates will be biased and inconsistent if some of the measures generate only very few observations, which then renders this approach infeasible. This %motivates of the two-step approach a one-step approach holds promise,
% In their example and also theory, each measure $\nu_i$ is assumed to be fully observed, which is usually unrealistic in practice. In practical applications, one generally only have access to random observations $\{Y_{ij}\}_{j=1}^{N_i}$ sampled from each distribution $\nu_i$. While developing the Cox point process regression model, \citet{mull:21:10} applied \f regression to estimate the shape part of the conditional intensity function, where the smoothed empirical quantile function was used to replace the unobservable distribution. However, the bias introduced by smoothing was not considered in the theory, though the smoothed empirical quantile function was adopted during implementation. Moreover, a strong assumption was imposed on the quantile function $F_{\mu}^{-1}$ where
%\begin{equation}
% \label{eq:cox}
% M|x-y|\leq|F_{\mu}^{-1}(x)-F_{\mu}^{-1}(y)|\leq L|x-y|\quad\text{for all }x, y\in(0, 1)
%\end{equation}
%with $0<M<L<\infty$.
provides the motivation to replace the unobservable measures $\nu_i$ with the empirical measure $\hat{\nu}_i=(1/{N_i})\sum_{j=1}^{N_i}\delta_{Y_{ij}}$. This circumvents the pre-smoothing step and hence eliminates the corresponding tuning parameter selection and smoothing bias. Using empirical measures $\hat{\nu}_i$ in lieu of the unobservable measures $\nu_i$ as responses, the approaches proposed here are the global and local Regression with Empirical Measures (REM), replacing \eqref{eq:hatmg} and \eqref{eq:hatml} by
\begin{align}
&\hat{m}_G(z)=\argmin{\mu\in\mathcal{W}}\frac{1}{n}\sum_{i=1}^ns_{iG}(z)d^2_{\mathcal{W}}(\hat{\nu}_i, \mu),\label{eq:tildemg}\\&
\hat{m}_{L, h}(z)=\argmin{\mu\in\mathcal{W}}\frac{1}{n}\sum_{i=1}^ns_{iL}(z, h)d_{\mathcal{W}}^2(\hat{\nu}_i, \mu).\label{eq:tildeml}
\end{align}
%An obstacle that needs to be overcome %to using the empirical measure $\hat{\nu}_i$ during implementation
The computation of the minimizers in \eqref{eq:tildemg} and \eqref{eq:tildeml} is not straightforward, as some of the weights in these minimization problems are inherently negative. We show in Section \ref{subsec:imp} that adopting least common multiples leads to a simple and efficient algorithm. % can be devised to implement the proposed regression model. %; see Section \ref{subsec:imp}.
%\red{Comments: (1) Much more intuition is needed about the averaging of the empirical measure, perhaps with a worked out example for the barycenter and a second example when a 1-d predictor is at an endpoint, the example can be based on small samples with $n=3$ and $N_i=10,15,20$ or even smaller and could include a graphical schematic where also it is clear that there are negative weights. (2) It is also not clear that the optimization problem gives a discrete measure while the algorithm certainly does and whether that is the final product or a smoothing step should be applied since the true measure is assumed to be a.c. But if smoothing step we are back to bandwidth choice etc.(3) Also in Fig 4a one could add a comparison with a direct density estimation for the sparsely observed data.}
\section{Asymptotic Properties}
\label{sec:th}
%In this section, we investigate asymptotic behaviors of the proposed regression model, where both
We establish pointwise and uniform rates of convergence under the framework of M-estimation. Denote by $\mathcal{W}^{\text{ac}}$ the set of measures $\mu\in\mathcal{W}$ that are absolutely continuous with respect to the Lebesgue measure $dx$ on $\mathbb{R}$. To establish rates of convergence for estimates \eqref{eq:tildemg} and \eqref{eq:tildeml}, we require the following conditions.
\bem[label=(C\arabic*), leftmargin=1cm]
%\setcounter{enumi}{-1}
\item The random measure $\nu$ satisfies $\nu\in\mathcal{W}^{\text{ac}}$ and its probability density function $f_\nu$ is bounded below by a positive constant.\label{itm:A1}
\item The sample size $N_i$ is independently Poisson distributed with parameter $c_i\lambda_n$ where $c_i\geq c>0$ are constants and $0<\lambda_n/\log n\to\infty$ as $n\to\infty$. \label{itm:A2}%where $\lambda_n/\log n\to\infty$ as $n\to\infty$.
\eem
Condition \ref{itm:A1} ensures the random measure $\nu$ is well-behaved. %a stronger condition for $\lambda_n$ is required if omitted.
Since measures $\nu_i$ are discretely observed, rates of convergence will be affected by the rate at which the sample size $N_i$ increases to infinity. Instead of directly requiring the sample size $N_i$ to diverge to infinity, Condition \ref{itm:A2} imposes a distributional condition on $N_i$. % whose mean diverges to infinity. Such a condition not only makes it possible to derive the rate of convergence but also takes into account
This set-up reflects the case of heterogeneous sample sizes, as it allows some sample sizes $N_i$ to be very small and even $0$ for some of the observed distributions, but also % Additionally, the condition that $\lambda_n/\log n\to\infty$ as $n\to\infty$
ensures that the $N_i$ are positive almost surely for large enough $n$; compare Lemma 3 in \citet{pana:16}. The following result formalizes the consistency of the proposed global REM estimates and provides rates of convergence.
\begin{thm}
\label{thm:1}
Under Conditions \ref{itm:A1} and \ref{itm:A2}, the global REM estimate defined in \eqref{eq:tildemg} satisfies
\begin{equation*}
d_\mathcal{W}\{\hat{m}_G(z), m_G(z)\}=O_p(n^{-1/2}+\lambda_n^{-1/2}).
\end{equation*}
Furthermore, for any constant $B$ it holds that for any $\varepsilon>0$,
\[\sup_{\|z\|_E\leq B}d_{\mathcal{W}}\{\hat{m}_G(z), m_G(z)\}=O_p(n^{-1/\{2(1+\varepsilon)\}}+\lambda_n^{-1/2}).\]
\end{thm}
All proofs are in the Appendix. The pointwise rate of convergence is $O_p(n^{-1/2})$ as long as the sample size $N_i$, on average, grows at the same rate as $n$, i.e., $\lambda_n\asymp n$, which is the same as the well-known optimal rate for multiple linear regression. Similarly we obtain the following result for local REM, where the kernel and distributional assumptions \ref{itm:lp0}--\ref{itm:lu1} listed in the Appendix A are standard for local regression.
\begin{thm}
\label{thm:2}
Under Conditions \ref{itm:A1} and \ref{itm:A2}, the local REM estimate defined in \eqref{eq:tildeml} satisfies
\begin{equation*}
d_\mathcal{W}\{\hat{m}_{L, h}(z), m(z)\}=O_p(n^{-2/5}+\lambda_n^{-1/2})
\end{equation*}
for $h\sim n^{-1/5}$ if Assupmtions \ref{itm:lp0} and \ref{itm:lp1} hold. Furthermore, for a closed interval $\mathcal{T}\subset\mathbb{R}$,
\[\sup_{z\in\mathcal{T}}d_\mathcal{W}\{\hat{m}_{L, h}(z), m(z)\}=O_p(n^{-1/(3+\varepsilon)}+\lambda_n^{-1/2})\]
for $h\sim n^{-1/(6+2\varepsilon)}$ and any $\varepsilon>0$ if Assumptions \ref{itm:lu0} and \ref{itm:lu1} hold.
\end{thm}
As long as $\lambda_n\asymp n^{4/5}$, the local REM estimate achieves the pointwise rate $O_p(n^{-2/5})$, corresponding to the well-known optimal rate for standard local linear regression. % with real valued responses.
%In their work on the Cox point process regression, \citet{mull:21:10} obtained consistency results similar to ours. However, our results have three advantages. Firstly, we do not rely on the assumption in \eqref{eq:cox} for the quantile function, which is a strong assumption. Secondly, the convergence rate for the second term involving $\lambda_n$ is twice as fast as previous results, which implies that the same convergence rates can be achieved with weaker conditions on the sample size $N_i$. Thirdly, in addition to pointwise rates of convergence, we also derive uniform rates of convergence, providing a unified theory for distribution-response regression.
\iffalse
\subsection{Weighted empirical Wasserstein barycenter}
Assume $\Omega=[0, 1]$ and the number of observations per measure is Poisson distributed with parameter $\lambda_n$ diverging with the sample size $n$. Specifically, $N_i\sim\mr{Pois}(\lambda_n)$ for each $i=1, \ldots, n$. Consider the weighted empirical Wasserstein barycenter with weights $N_i/N$ where $N=\sum_{i=1}^n N_i$ defined by
\be
\hat{\bs{\nu}}_n^{\oplus}=\argmin{\mu\in\cW}\sum_{i=1}^n\frac{N_i}{N}d_{\cW}^2(\hat{\bs{\mc{\nu}}}_i, \mu),
\ee
where $\hat{\bs{\mc{\nu}}}_i=\frac{1}{N_i}\sum_{j=1}^{N_i}\delta_{X_{ij}}$ is the empirical measure. By the triangle inequality, we have
\be
d_{\cW}(\hat{\bs{\nu}}_n^{\oplus}, \nu_0)\leq d_{\cW}(\hat{\bs{\nu}}_n^{\oplus}, \tilde{\bs{\nu}}_n^{\oplus}) + d_{\cW}(\tilde{\bs{\nu}}_n^{\oplus}, \nu_0),
\ee
where
\be \tilde{\bs{\nu}}_n^{\oplus}=\argmin{\mu\in\cW}\sum_{i=1}^n\frac{N_i}{N}d_{\cW}^2(\bs{\mc{\nu}}_i, \mu)
\ee
is the weighted Wasserstein barycenter.
Again by triangle inequality, it follows that
\begin{align}
d_{\cW}(\hat{\bs{\nu}}_n^{\oplus}, \tilde{\bs{\nu}}_n^{\oplus})&\leq\sum_{i=1}^n\frac{N_i}{N}d_{\cW}(\hat{\bs{\mc{\nu}}}_i, \bs{\mc{\nu}}_i)\\&=\sum_{i=1}^n\frac{N_i}{N}d_{\cW}(\hat{\bs{\mc{\nu}}}_i, \bs{\mc{\nu}}_i)1_{N_i>0} + \sum_{i=1}^n\frac{N_i}{N}d_{\cW}(\hat{\bs{\mc{\nu}}}_i, \bs{\mc{\nu}}_i)1_{N_i=0}\\&\leq\sum_{i=1}^n\frac{N_i}{N}d_{\cW}(\hat{\bs{\mc{\nu}}}_i, \bs{\mc{\nu}}_i)1_{N_i>0} + \sum_{i=1}^n\frac{N_i}{N}1_{N_i=0},
\end{align}
where the last inequality follows from the convention that $\hat{\bs{\nu}}_i$ is the Lebesgue measure when $N_i=0$ and the distance between any two measures is no larger than one.
The following elementary result is stated without proof; see \citet[p. 39]{pana:20} for details.
\begin{lem}
\label{lem:pq}
For $p\geq q\geq 1$ and two probability measures $\mu, \nu\in\cW_p(\Omega)$, it holds that
\be
W_p^p(\mu, \nu)\leq W_q^q(\mu, \nu)\{\mr{diam}(\Omega)\}^{p-q},
\ee
where $W_p(\mu, \nu)$ and $W_q(\mu, \nu)$ are the $p$- and $q$-Wasserstein distances between $\mu$ and $\nu$, respectively.
\end{lem}
Observe that
\begin{align}
E[\{\frac{N_i}{N}d_{\cW}(\hat{\bs{\mc{\nu}}}_i, \bs{\mc{\nu}}_i)1_{N_i>0}\}^2]&=E\{\frac{N_i^2}{N^2}d_{\cW}^2(\hat{\bs{\mc{\nu}}}_i, \bs{\mc{\nu}}_i)1_{N_i>0}\}\\&\leq E\{\frac{N_i^2}{N^2}\int_0^1|\hat{F}_i(x)-F_i(x)|1_{N_i>0}dx\}\\&=\int_0^1E\{\frac{N_i^2}{N^2}|\hat{F}_i(x)-F_i(x)|1_{N_i>0}\}dx.
\end{align}
for $i=1, \ldots, n$, where the inequality is followed by taking $p=2, q=1$ in Lemma \ref{lem:pq} and the last equality is due to Fubini's theorem. For simplicity, define the following random variables
\be
Z_{ni}=\frac{N_i^2}{N^2}\{\hat{F}_i(x)-F_i(x)\}.
\ee
It follows that
\be
Z_{ni}|N_i, N, F_i(x)\sim \frac{N_i^2}{N^2}[\frac{B\{N_i, F_i(x)\}}{N_i}-F_i(x)],
\ee
where $B\{N_i, F_i(x)\}$ is the binomial distribution with parameters $N_i$ and $F_i(x)$. Therefore
\begin{align}
E\{Z_{ni}^2|N_i, N, F_i(x)\}&=\Var\{Z_{ni}|N_i, N, F_i(x)\}+[E\{Z_{ni}|N_i, N, F_i(x)\}]^2\\&=\frac{N_i^4}{N^4}\frac{N_iF_i(x)\{1-F_i(x)\}}{N_i^2}\\&=\frac{N_i^3}{N^4}F_i(x)\{1-F_i(x)\}\\&\leq\frac{N_i^3}{4N^4}.
\end{align}
Since the bound is independent of $F_i(x)$, one has
\be
E(Z_{ni}^2|N_i, N)\leq\frac{N_i^3}{4N^4}
\ee
Note that $N_i|N\sim B(N, \frac{1}{n})$, taking expectation with respect to $N$ leads to
\be
E(Z_{ni}|N_i)=
\ee
\fi
\section{Implementation and Simulations}
\subsection{Implementation details}
\label{subsec:imp}
To implement the proposed methods, one needs to solve minimization problems \eqref{eq:tildemg} and \eqref{eq:tildeml}. By standard properties of the $L^2(0, 1)$ inner product, \eqref{eq:tildemg} and \eqref{eq:tildeml} can be simplified to
\begin{align}
&\hat{m}_G(z)=\argmin{\mu\in \mathcal{W}}d_{L^2}^2\{\hat{B}_G(z), F_{\mu}^{-1}\},\label{eq:bg}\\
&\hat{m}_{L, h}(z)=\argmin{\mu\in \mathcal{W}}d_{L^2}^2\{\hat{B}_{L, h}(z), F_{\mu}^{-1}\},\nonumber
\end{align}
where $\hat{B}_G(z)=n^{-1}\sum_{i=1}^ns_{iG}(z)F_{\hat{\nu}_i}^{-1}$, $\hat{B}_{L, h}(z)=n^{-1}\sum_{i=1}^ns_{iL}(z, h)F_{\hat{\nu}_i}^{-1}$ and $d_{L^2}(\cdot, \cdot)$ denotes the $L^2$ distance; see the proof of Theorem \ref{thm:1} for details. The minimizers $\hat{m}_G(z)$ and $\hat{m}_{L, h}(z)$, viewed as projections onto $\mathcal{W}$, exist and are unique for any $z$ by convexity and closedness of $\mathcal{W}$. To tackle the challenge of substantial variation in the number of jumps for empirical quantile functions $F_{\hat{\nu}_i}^{-1}$ due to varying sample sizes $N_i$, we propose an algorithm based on the least common multiple, which allows us to compute $\hat{B}_G(z)$ and $\hat{B}_{L, h}(z)$ as straightforward weighted averages. The algorithm for global REM is outlined in Algorithm \ref{alg:lcm}. The only difference for local REM is the substitution of the weight function $s_{iL}(z, h)$ for $s_{iG}(z)$.
% \vspace{.3cm}
\begin{algorithm}[tb]
%\DontPrintSemicolon
\SetAlgoLined
\KwIn{data $\{(Z_i, \{Y_{ij}\}_{j=1}^{N_i})\}_{i=1}^n$, and a new predictor level $z$.}
\KwOut{prediction $\hat{m}_G(z)$.}
$M\longleftarrow$ the least common multiple of $\{N_i\}_{i=1}^n$\;
$\{\mathbb{V}_i\}_{i=1}^n\longleftarrow$ for each $i=1, \ldots, n$, stretch $(Y_{i1}, \ldots, Y_{iN_i})^\T$ to a vector of length $M$, $\mathbb{V}_i=(V_{i1}, \ldots, V_{iM})^\T$, by repeating each element $\frac{M}{N_i}$ times and then arranging in ascending order\;
$s_{iG}(z)\longleftarrow1+(Z_i-\bar{Z})^\T\hat{\Sigma}^{-1}(z-\bar{Z})$ where $\bar{Z}=\frac{1}{n}\sum_{i=1}^n Z_i$ and $\hat{\Sigma}=\frac{1}{n}\sum_{i=1}^n(Z_i-\bar{Z})(Z_i-\bar{Z})^\T$ are the sample mean and variance of $\{Z_i\}_{i=1}^n$, respectively\;
$\bar{\mathbb{V}}=(\bar{V}_1, \ldots, \bar{V}_M)^\T\longleftarrow\frac{1}{n}\sum_{i=1}^ns_{iG}(z)\mathbb{V}_i$, the element-wise weighted average of $\{\mathbb{V}_i\}_{i=1}^n$\;
$\hat{m}_G(z)\longleftarrow\argmin{\mu\in\mathcal{W}}d_{L^2}^2(\hat{B}_G(z), F_{\mu}^{-1})$, where $\hat{B}_G(z)=\sum_{m=1}^M\bar{V}_m\mathbf{1}_{q\in(\frac{m-1}{M}, \frac{m}{M}]}$ is a step function defined on $q\in[0, 1]$.
\caption{Global REM}
\label{alg:lcm}
\end{algorithm}
%\vspace{.3cm}
In the first step of Algorithm \ref{alg:lcm}, one needs to calculate $M=\mathrm{lcm}(\{N_i\}_{i=1}^n)$, the least common multiple of the ensemble $\{N_i\}_{i=1}^n$. %, denoted by $\mathrm{lcm}(\{N_i\}_{i=1}^n)$.
To control the size of $\mathrm{lcm}(\{N_i\}_{i=1}^n)$ especially when $n$ is large and $N_i$ greatly varies, we incorporate a threshold $M=\min\{\mathrm{lcm}(\{N_i\}_{i=1}^n), M_0\}$, where $M_0$ is a pre-specified constant, say $5,000$. The fifth step of Algorithm \ref{alg:lcm} involves a minimization problem to obtain the projection onto $\mathcal{W}$. For this step, we use a Riemann sum approximation of the $L^2$ distance. This leads to the following convex quadratic optimization problem,
\begin{align}
\begin{split}
&\text{minimize}\quad \sum_{m=1}^M(q_m-\bar{V}_m)^2\\
&\text{subject to}\quad q_1\leq q_2\leq\cdots\leq q_{M-1}\leq q_M,\\
&\hspace{0.87in}q_j\in\Omega,\quad j=1, \ldots, M.
\end{split}
\label{eq:osqp}
\end{align}
The solution $q^*$ represents a discretized version of the predicted quantile function. We use the \texttt{osqp} package \citep{stel:20} in R to solve the optimization problem presented in \eqref{eq:osqp}.
An illustrative example is in Figure~\ref{fig:diagram} to demonstrate the proposed approach and algorithm for the case of global REM. In the top left panel of Figure~\ref{fig:diagram}, we depict large variation in sample sizes $N_i$ of the observations available for distributions $\nu_i$, where the smallest $N_i$ is $N_i=1.$ A simple calculation shows that the global weight function is $s_{iG}(z)=1+(2i-5)(z-5)/5$ for $i=1, 2, 3, 4$. The predicted distribution at predictor level $z$ is obtained by calculating $\hat{B}_G(z)$ in \eqref{eq:bg}, a weighted average of empirical quantile functions $F_{\hat{\nu}_i}^{-1}$, see the top right panel of Figure~\ref{fig:diagram}, utilizing weights $s_{iG}(z)$, followed by a projection onto $\mathcal{W}$. The weighted average is implemented with Algorithm~\ref{alg:lcm}. The predicted distribution $\hat{m}_G(z)$ obtained in Algorithm~\ref{alg:lcm} is represented as an empirical quantile function, which can subsequently be converted into a density function for improved visualization. We adopted the qf2pdf function in the \texttt{frechet} package \citep{chen:20} to construct densities from quantile functions in this example and for our data applications; there are also alternative options \citep{nile:22}.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{img/diagram.pdf}
\caption{Illustrating global REM, where $N_i$ observations $\{Y_{ij}\}_{j=1}^{N_i}$ are sampled from distributions $\nu_i=N(0, Z_i^2)$ for $i=1, 2, 3, 4$, with sample sizes $N_1=1, N_2=2, N_3=5, N_4=10$ and scalar covariates $Z_1=2, Z_2=4, Z_3=6, Z_4=8$. Top left: Visualization of $Y_{ij}$ versus covariate levels $Z_i$, along with the underlying true densities. Top right: Empirical quantile functions corresponding to the empirical measures $\hat{\nu}_i=(1/{N_i})\sum_{j=1}^{N_i}\delta_{Y_{ij}}$. Bottom: Predicted quantile functions (left) and corresponding densities (right), obtained with global REM for predictor levels $z=3$ and $z=5$.}
\label{fig:diagram}
\end{figure}
The bottom two panels of Figure~\ref{fig:diagram} illustrate the predicted distributions using global REM for predictor levels $z=3$ and $z=5$. For $z=\bar{Z}=5$, one obtains the Wasserstein barycenter of the four empirical measures as for this case the global weight function is constant, i.e., $s_{iG}(5)=1$ for all $i$. Since $z=3$ is close to the left endpoint, the estimation requires negative weights where $s_{4G}(3)=-1/5$. In the presence of negative weights, $\hat{B}_G(3)$ no longer resides in $\mathcal{W}$, necessitating the additional projection step as per \eqref{eq:bg}.
\subsection{Simulations}
To assess the finite sample performance of the proposed methods, we construct a generative model that produces random responses $\nu$ along with an Euclidean predictor $Z\in\mathbb{R}$. Consider the true regression function $m(Z)$, represented as quantile function,
\[m(Z)=E(\eta|Z)+E(\sigma|Z)\Phi^{-1}(\cdot),\]
which corresponds to a Gaussian distribution with mean and standard deviation depending on $Z$.
The distribution parameters of the true regression function $m(Z)$ are generated conditionally on $Z$, where the mean and the standard deviation are assumed to follow a normal distribution and a Gamma distribution, respectively. Four different simulation settings are examined as summarized in Table~\ref{tab:simusetting}. In Settings I and II, the response is generated, on average, as a Gaussian distribution with parameters depending on $Z$. Setting I corresponds to a global scenario where the response $\nu$ depends linearly on the predictor $Z$ as $E(\eta|Z)=\eta_0+\alpha Z$ and $E(\sigma|Z)=\sigma_0+\beta Z$. In Setting II, the nonlinear relationships %between $\nu$ and $Z$ where
$E(\eta|Z)=\eta_0+\alpha\sin(\pi Z)$ and $E(\sigma|Z)=\sigma_0+\beta\sin(\pi Z)$ is considered for the true underlying regression model. % to examine the performance of local measure regression.
To take non-Gaussian distributions into consideration, we apply an additional transportation to the random response in Settings III and IV. Specifically, after sampling the distribution parameters as in the previous settings, the resulting distribution is transported in Wasserstein space via a random transport map $T$, uniformly sampled from the collection of maps $T_k(\alpha)=\alpha-\sin(k\alpha)/|k|$ for $k\in\{-2, -1, 1, 2\}$ \citep{pana:16}. Such a transportation significantly complicates Settings III and IV and makes the response distribution non-Gaussian. One can show that the true regression function remains the same after this random transportation.
\begin{table}[tb]
\caption{Four simulation settings, where $F_{\nu}^{-1}$ represents the quantile function of the generated random response. The distribution parameters of the random response depend on the predictor $Z$ as indicated for Settings I--IV. An additional transport map $T$, uniformly sampled from the collection of maps $T_k(\alpha)=\alpha-\sin(k\alpha)/|k|$ for $k\in\{-2, -1, 1, 2\}$, is applied to the resulting random responses in Settings III and IV.}
\centering
\btab{l|ll}
\hline
& Type & Setting
\\\hline\hline
\shortstack[l]{I\\\vspace{1em}} &
\shortstack[l]{global\\model\\\vspace{0.5em}} &
\shortstack[l]{\\$F_{\nu}^{-1}(\cdot)=\eta+\sigma\Phi^{-1}(\cdot)$,\\where $\eta|Z\sim N(\eta_0+\alpha Z, \tau^2)$ and \\$\sigma|Z\sim \mathrm{Gamma}[\{\sigma_0+\beta Z\}^2/\kappa, \kappa/\{\sigma_0+\beta Z\}]$}
\\\hline
\shortstack[l]{II\\\vspace{1em}} &
\shortstack[l]{local\\model \\\vspace{0.5em}} &
\shortstack[l]{\\$F_{\nu}^{-1}(\cdot)=\eta+\sigma\Phi^{-1}(\cdot)$,\\where $\eta|Z\sim N(\eta_0+\alpha\sin(\pi Z), \tau^2)$ and \\$\sigma|Z\sim \mathrm{Gamma}[\{\sigma_0+\beta\sin(\pi Z)\}^2/\kappa, \kappa/\{\sigma_0+\beta\sin(\pi Z)\}]$}
\\\hline
\shortstack[l]{III\\\vspace{1em}} &
\shortstack[l]{global model\\with random transport\\\vspace{0.5em}} &
\shortstack[l]{\\$F_{\nu}^{-1}(\cdot)=T\circ\{\eta+\sigma\Phi^{-1}(\cdot)\}$,\\where $\eta|Z\sim N(\eta_0+\alpha Z, \tau^2)$ and \\$\sigma|Z\sim \mathrm{Gamma}[\{\sigma_0+\beta Z\}^2/\kappa, \kappa/\{\sigma_0+\beta Z\}]$}
\\\hline
\shortstack[l]{IV\\\vspace{1em}} &
\shortstack[l]{local model\\with random transport\\\vspace{0.5em}} &
\shortstack[l]{\\$F_{\nu}^{-1}(\cdot)=T\circ\{\eta+\sigma\Phi^{-1}(\cdot)\}$,\\where $\eta|Z\sim N(\eta_0+\alpha\sin(\pi Z), \tau^2)$ and \\$\sigma|Z\sim \mathrm{Gamma}[\{\sigma_0+\beta\sin(\pi Z)\}^2/\kappa, \kappa/\{\sigma_0+\beta\sin(\pi Z)\}]$}
\\\hline
\etab
\label{tab:simusetting}
\end{table}
We consider $n=50, 100, 200, 500, 1000$, with $Q=1000$ Monte Carlo runs for each of the four simulation settings. For each $n$, the sample size $N_i$ is independently sampled from the Poisson distribution with parameter $0.25n$. In each Monte Carlo run, predictors $\{Z_i\}_{i=1}^n$ are independently sampled from $U(-1, 1)$. For each $i$, mean and standard deviation of $\nu_i$ are generated conditionally on $Z_i$ as described in Table~\ref{tab:simusetting}, with $\eta_0=0, \sigma_0=3, \alpha=3, \beta=0.5, \tau = 0.5$, and $\kappa = 1$. $N_i$ random observations $\{Y_{ij}\}_{j=1}^{N_i}$ are then independently sampled from $\nu_i$. For the $q$th Monte Carlo run for a given setting, the quality of the estimation is quantified by the integrated squared error (ISE)
\[\mathrm{ISE}_q=\int_{-1}^1d_{\mathcal{W}}^2\{\hat{m}_q(z), m(z)\}dz,\]
where $m(\cdot)$ is the true regression function and $\hat{m}_q(z)$ is the fitted regression function. %and the average quality of the estimation over the Q = 1000 Monte Carlo runs is assessed by \red{the mean integrated squared error}
%\[\mr{MISE}=\frac{1}{Q}\sum_{q=1}^Q\int_{-1}^1d_{\mathcal{W}}^2\{\tilde{m}_q(z), m(z)\}dz.\]
The bandwidths for the local REM in the second simulation setting are chosen as $n^{-1/5}$.
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{img/ise.pdf}
\caption{Boxplots of integrated square errors (ISE) for $Q=1000$ Monte Carlo runs and different simulation settings using the proposed REM (red) and the two-step approaches that were adopted previously (blue).}
\label{fig:ise}
\end{figure}
We also include comparisons with the previous two-step procedures that included a prior kernel smoothing step \citep{pana:16, mull:19:3} to estimate the densities of the $\nu_i$ from the available discrete observations. The integrated squared errors (ISE) for all Monte Carlo runs and different $n$ under the four simulation settings using the proposed methods and the comparison methods are summarized in the boxplots in Figure~\ref{fig:ise}. With increasing sample size, ISE is seen to decrease for all simulation settings, demonstrating the convergence of the proposed methods to the target. Importantly, the proposed REM outperforms the comparison methods under all simulation settings. Indeed, the preliminary smoothing step required in the previous literature involved a unnecessary tuning parameter selection and fails when the data for some measures are very sparse. %introduces an additional bias to the final estimate, while the proposed methods operate directly on the empirical measure, with no smoothing involved.
\section{Data Applications}
\subsection{Cohort-specific BMI distribution for US preschool children}
The Environmental influences on Child Health Outcomes (ECHO) program is an NIH-funded nationwide consortium of multiple cohort studies across the United States designed to investigate the effects of early life exposures on child health and development \citep{gill:18}. The ECHO program combines existing prenatal and pediatric data collected via cohort-specific protocols with a standardized ECHO-wide protocol that was established in 2019 \citep{knap:23}. The de-identified data on participants contributing extant and new data can be accessed through the National Institute of Child Health and Human Development (NICHD) Data and Specimen Hub (DASH); the version we use here has been made available on August 31, 2021 \citep{gill:22}. As a multi-cohort study, ECHO brings separate cohorts together so that researchers can access information from a large and diverse population of children followed from the prenatal period through adolescence.
It is of interest to study the role of demographic factors in child development, measured in terms of body mass index (BMI), calculated as weight in kilograms divided by height in meters squared. We extracted weight and height measurements of preschool children aged approximately 4 years for 17 cohorts from ECHO, along with demographic information for each cohort, aiming to shed light on how the distribution of BMI of preschool children varies across different cohorts in relation to cohort-specific demographic characteristics. The responses specifically are the cohort-specific distributions of BMI for 4-year-old children in the respective cohort. Cohort-specific covariates that reflect important demographic characteristics of each cohort include average BMI of mothers, average parental education and proportion of Asians.
Due to differences in the cost and accessibility of visits for different cohorts, there is a significant variation in sample size for each cohort. The amount of accessible weight and height measurements for each cohort varies substantially as a result; see Table \ref{tab:echo} for further information. Specifically, there is only one weight and height measurement for girls in AGA01 cohort, while 160 measurements are available for boys in AAV01 cohort. We applied global REM to boys and girls separately. The fitted BMI densities for two cohorts with the most and least weight and height measurements (AAV01, AGA01) are demonstrated in Figure~\ref{fig:echoBmi}(a), along with the corresponding BMI measurements shown as ticks. Although only one measurement is available for girls in the AGA01 cohort, the proposed regression model is able to borrow information from cohorts with more measurements to construct a reasonable density estimate. We also include kernel density estimates in Figure~\ref{fig:echoBmi}(a) for comparison, where the covariate information is not taken into account. Note that kernel density estimation is infeasible for girls in the AGA01 cohort since only one observation is available.
%\red{the density estimation step from the discrete measure is incorporated here, so a tuning parameter is needed and we should not make claims that are not correct such as the proposed method does not need a tuning parameter etc. It would perhaps be more accurate to phrase it from the beginning as: previous method, smoothing then weighting; new method, weighting, then smoothing. This is a delicate issue and needs to be brought out in much more detail}
\begin{table}[tb]
\caption{Number of weight and height measurements for each cohort.}
\centering
\btab{lllllllll}
\hline
Cohort & AAA01 & AAD01 & AAE01 & AAF01 & AAJ01 & AAU01 & AAV01 & AAW02
\\\hline
Boys & 139 & 70 & 77 & 35 & 9 & 10 & 160 & 12 \\
Girls & 101 & 70 & 81 & 15 & 7 & 8 & 145 & 8 \\
Total & 240 & 140 & 158 & 50 & 16 & 18 & 305 & 20 \\\hline
\etab
\\
\btab{lllllllll}
\hline
AAX04 & AAX06 & ADA01 & AGA01 & AJA02 & AJA03 & AKA01 & AKA02 & ALA01
\\\hline
83 & 124 & 8 & 3 & 44 & 6 & 7 & 76 & 134 \\
67 & 110 & 7 & 1 & 38 & 14 & 3 & 71 & 128 \\
150 & 234 & 15 & 4 & 82 & 20 & 10 & 147 & 262 \\\hline
\etab
\label{tab:echo}
\end{table}
\begin{figure}[p]
\begin{minipage}{1\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{img/echoBmi.pdf}\\
(a)
\end{minipage}\hfill
\begin{minipage}{1\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{img/echoBmiDensity.pdf}\\
(b)
\end{minipage}\\
\begin{minipage}{\textwidth}
\caption{(a) Fitted densities of BMI distributions (solid) of US preschool boys and girls for AAV01 and AGA01 cohorts, along with direct kernel density estimates (dashed). The corresponding BMI measurements are shown as ticks. (b) Predicted BMI densities of US preschool boys and girls at different predictor levels.}
\label{fig:echoBmi}
\end{minipage}
\end{figure}
To further investigate the effects of different demographic factors, predicted BMI densities at different predictor levels are shown in Figure~\ref{fig:echoBmi}(b). Separately for each predictor, the predictor level was varied from the first to the third quantile of the sample, while the other two predictors were held fixed at their median level. We observe that the average BMI of mothers is the most influential predictor, where with increasing average BMI of mothers, the BMI distribution for both 4-year-old boys and girls flattens out and becomes less concentrated. For the predictor average parental education, higher values are associated with heavier tails of the BMI distribution for both 4-year-old boys and girls. For higher average parental education, the BMI distribution of boys also shifts a little to the right. The proportion of Asians in the population seems to have little effect on the BMI distribution of preschool children.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{img/echoBmiObesity.pdf}
\caption{Prevalence of obesity for US preschool boys and girls at different predictor levels.}
\label{fig:echoObesity}
\end{figure}
Childhood obesity is continuing to rise in the US, and currently about 13.7 million children are considered to be overweight/obese. For the ECHO data, we can compute the probability of obesity from the BMI distribution to investigate demographic disparities in early childhood obesity. Obesity is defined as a BMI at or above the 95th percentile for children of the same age and sex. According to the sex-specific BMI-for-age 2000 CDC Growth Charts, the BMI threshold for a 4-year-old boy to be obese is 17.8, while for girl the value is 18. Based on the predicted BMI distributions (see Figure~\ref{fig:echoBmi}), we calculated the corresponding probabilities of obesity at different predictor levels, with results illustrated in Figure~\ref{fig:echoObesity}. The average BMI of mothers is found to be positively correlated with the prevalence of obesity among four-year-old children, while average parental education and proportion of Asians are negatively associated with obesity. None of these associations
establishes a causal effect, but these findings are in agreement with the current literature on obesity, where parental overweight and low socioeconomic status were found to be strong risk factors of obesity in children \citep{dani:04, vazq:20}, while obesity is less common in Asian children \citep{ande:09}.
\subsection{Bid time distribution for eBay auction data}
Electronic commerce and in particular eBay, with the prevalence of online auction, is a rich data source for the study of bidding behavior and strategies \citep{wang:08:2, mull:09:1, liu:19}. Here we implemented the proposed regression model to analyze eBay auction data consisting of 93 7-day auctions for an Xbox game console. The data are publicly available at \url{https://www.modelingonlineauctions.com/datasets} \citep{jank:10}. For each auction, the opening bid set by the seller serves as predictor and the time stamps where bids were placed as responses. We aim to investigate the dependence of bid time distribution on the opening bid. The number of bids for each auction varies from 2 to 75, necessitating the application of the proposed regression model, for which we used local REM. % was applied with the opening bid being the predictor and the discretely observed bid time distribution being the response.
\begin{figure}[p]
\begin{minipage}{1\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{img/bidTime.pdf}\\
(a)
\end{minipage}\hfill
\begin{minipage}{1\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{img/bidTimeDensity.pdf}\\
(b)
\end{minipage}\\
\begin{minipage}{\textwidth}
\caption{(a) Fitted bid time densities (solid) for four different auctions of an Xbox game console with the proposed method, along with auction-specific kernel density estimates (dashed). The corresponding bid times are shown as ticks. (b) Predicted bid time densities for different opening bids.}
\label{fig:bidTime}
\end{minipage}
\end{figure}
The fitted bid time densities for four auctions including the two with the most and least bids are illustrated in Figure~\ref{fig:bidTime}(a). The corresponding bid times are shown as ticks for comparison. We find that the proposed regression model is robust to the strong heterogeneity inherent in the auction data, where for example the bid time density for an auction with only two bids (top left panel in Figure~\ref{fig:bidTime}(a)) is reasonably well recovered. In contrast, standard kernel density estimation implemented separately for each auction lacks the ability to borrow strength across the entire sample of distributions and, as expected, fails to reconstruct the underlying bid time density for the auction with only two bids. To further investigate the effect of the opening bid amount on the bid time distribution, we show in Figure~\ref{fig:bidTime}(b) the predicted bid time densities for opening bids varying from 1 to 100 dollars. Bidders tend to bid late in the auction if the opening bid set by the seller is high. In contrast, if the opening bid is low, bid time is more evenly distributed over the 7-day period. Regardless of the opening bid, bid sniping is prevalent, where the bidding becomes frenzied near the end of an auction. A lower opening bid is associated with less bid sniping in the sense that bidding is more spread out over the entire time range of the auction.
\iffalse
\section{Discussion}
In this article, we propose a flexible and robust regression model for the analysis of discretely observed distributions with heterogenous sample sizes. The use of empirical measures circumvents the preliminary smoothing step adopted in existing literature and hence avoid the unnecessary smoothing bias. The proposed regression model is supported with theoretical justifications including pointwise and uniform rates of convergence, where the former are optimal for both global and local versions in the sense that they correspond to the known optimal rates in the case of Euclidean responses. Both numerical simulations and real data applications including ECHO data and eBay auction data substantiate the effectiveness and usefulness of the proposed regression model.
%For the case of a time series of distributions, if one adopts an autoregressive model both the response and the predictor are distribution objects; see for example \citet{zhan:22}. This case is beyond the scope of the present paper. One potential approach is to use geodesics in the Wasserstein space. The relationship between response and predictor distributions can then be studied by relating movements along geodesics.
%Multivariate ditributions.
\fi
\section*{Acknowledgements}
\label{sec:akd}
We acknowledge support from NSF grants DMS-20146260, DMS-2310450 and IOS-2102953.
\appendix
\section*{Appendix A. Assumptions for Theorem \ref{thm:1} and Theorem \ref{thm:2}}
In the following, $f_Z(\cdot)$ and $f_{Z\mid\nu}(\cdot, \mu)$ stand for the marginal density of $Z$ and the conditional density of $Z$ given $\nu=\mu$, respectively, and $\mathcal{T}$ is a closed interval in $\mathbb{R}$ with interior $\mathcal{T}^o$.
\bem[label=(A\arabic*), leftmargin=1cm]
%\setcounter{enumi}{-1}
\item The kernel $K(\cdot)$ is a probability density function, symmetric around zero. Furthermore, defining $K_{kj}=\int_{\mathbb{R}}K^k(u)u^jdu$, $|K_{14}|$ and $|K_{26}|$ are both finite.\label{itm:lp0}
\item $f_Z(\cdot)$ and $f_{Z\mid\nu}(\cdot, \mu)$ both exist and are twice continuously differentiable, the latter for all $\mu\in\mathcal{W}$, and $\sup_{z, \mu}|(\partial^2 f_{Z\mid\nu}/\partial z^2)(z, \mu)|<\infty$. Additionally, for any open set $U\subset\mathcal{W}$, $\int_UdF_{\nu\mid Z}(z, \mu)$ is continuous as a function of $z$.\label{itm:lp1}
\item The kernel $K(\cdot)$ is a probability density function, symmetric around zero, and uniformly continuous on $\mathbb{R}$. Furthermore, defining $K_{jk}=\int_{\mathbb{R}}K(u)^ju^kdu$ for $j, k\in\mathbb{N}$, $|K_{14}|$ and $|K_{26}|$ are both finite. The derivative $K'$ exists and is bounded on the support of $K$, i.e., $\sup_{K(x)>0}|K'(x)|<\infty$; additionally, $\int_{\mathbb{R}}x^2|K'(x)|(|x\log|x||)^{1/2}dx<\infty$.\label{itm:lu0}
\item $f_Z(\cdot)$ and $f_{Z\mid\nu}(\cdot, \mu)$ both exist and are continuous on $\mathcal{T}$ and twice continuously differentiable on $\mathcal{T}^o$, the latter for all $\mu\in\mathcal{W}$. The marginal density $f_Z(\cdot)$ is bounded away from zero on $\mathcal{T}$, $\inf_{z\in\mathcal{T}}f_Z(z)>0$. The second-order derivative $f''_Z$ is bounded, $\sup_{z\in\mathcal{T}^o}|f''_Z(z)|<\infty$. The second-order partial derivatives $(\partial^2 f_{Z\mid\nu}/\partial z^2)(\cdot, \mu)$ are uniformly bounded, $\sup_{z\in\mathcal{T}^o, \mu\in\mathcal{W}}|(\partial^2 f_{Z\mid\nu}/\partial z^2)(z, \mu)|<\infty$. Additionally, for any open set $U\subset\mathcal{W}$, $\int_UdF_{\nu\mid Z}(z, \mu)$ is continuous as a function of $z$; for any $z\in\mathcal{T},$ $M(\mu, z)=E\{d_{\mathcal{W}}^2(\nu, \mu)|Z=z\}$ is equicontinuous, i.e.,
\[\limsup_{x\to z}\sup_{\mu\in\mathcal{W}}|M(\mu, x)-M(\mu, z)|=0.\]\label{itm:lu1}
\eem
\iffalse
\section*{Appendix B. Proofs}
%\newpage
%\section*{Supplementary Material}
%\setcounter{subsection}{0}
%\renewcommand{\thesubsection}{S.\arabic{subsection}}
%\subsection{Assumptions}
%\label{supp:assumption}
%\subsection{Proofs}
\begin{proof}[Proof of Theorem 1]
We first establish the pointwise rate of convergence for the global REM estimate. By the triangle inequality, we have
\begin{equation}
\label{eq:tildemm}
d_{\mathcal{W}}\{\hat{m}_G(z), m_G(z)\}\leq d_{\mathcal{W}}\{\hat{m}_G(z), \tilde{m}_G(z)\} + d_{\mathcal{W}}\{\tilde{m}_G(z), m_G(z)\}.
\end{equation}
The second term on the right hand side, corresponding to the pointwise rate of convergence for global \f regression with fully observed measures, is $O_p(n^{-1/2})$ by Proposition 1 and Theorem 2 in \citet{mull:19:3}.
To analyze the first term, we will utilize simplified expressions of $\tilde{m}_G(z)$ and $\hat{m}_G(z)$. Let $\langle\cdot, \cdot\rangle_{L^2}$, $\|\cdot\|_{L^2}$, and $d_{L^2}(\cdot, \cdot)$ be the inner product, norm and distance on the Hilbert space $L^2(0, 1)$. For any $\mu\in\mathcal{W}$, the map $Q:\mu\mapsto F_{\mu}^{-1}$ is an isometry from $\mathcal{W}$ to the subset of $L^2(0, 1)$ formed by equivalence classes of left-continuous nondecreasing functions on $(0, 1)$. The Wasserstein space $\mathcal{W}$ can thus be viewed as a subset of $L^2(0, 1)$, which has been shown to be convex and closed \citep{bigo:17}.
Define
\[\tilde{B}_{G}(z)=\frac{1}{n}\sum_{i=1}^ns_{iG}(z)F_{\nu_i}^{-1},\quad\hat{B}_{G}(z)=\frac{1}{n}\sum_{i=1}^ns_{iG}(z)F_{\hat{\nu}_i}^{-1},\]
where $F_{\nu_i}^{-1}$ and $F_{\hat{\nu_i}}^{-1}$ are the quantile functions of $\nu_i$ and $\hat{\nu}_i$, respectively. Since $n^{-1}\sum_{i=1}^ns_{iG}(z)=1$, it follows that
\begin{align*}
\frac{1}{n}\sum_{i=1}^ns_{iG}(z)d^2_{\mathcal{W}}(\nu_i, \mu)&=\frac{1}{n}\sum_{i=1}^ns_{iG}(z)[d_{L^2}^2\{F_{\nu_i}^{-1}, \tilde{B}_G(z)\}+d_{L^2}^2\{\tilde{B}_G(z), F_{\mu}^{-1}\}\\&\;\;\;\;+2\langle F_{\nu_i}^{-1}-\tilde{B}_G(z), \tilde{B}_G(z)-F_{\mu}^{-1}\rangle_{L^2}]\\&=\frac{1}{n}\sum_{i=1}^ns_{iG}(z)d_{L^2}^2\{F_{\nu_i}^{-1}, \tilde{B}_G(z)\}+d_{L^2}^2\{\tilde{B}_G(z), F_{\mu}^{-1}\}\\&\;\;\;\;+2\cdot\frac{1}{n}\sum_{i=1}^ns_{iG}(z)\langle F_{\nu_i}^{-1}-\tilde{B}_G(z), \tilde{B}_G(z)-F_{\mu}^{-1}\rangle_{L^2},
\end{align*}
where the last term
\begin{align*}
\frac{1}{n}\sum_{i=1}^ns_{iG}(z)\langle F_{\nu_i}^{-1}-\tilde{B}_G(z), \tilde{B}_G(z)-F_{\mu}^{-1}\rangle_{L^2}&=\langle \frac{1}{n}\sum_{i=1}^ns_{iG}(z)F_{\nu_i}^{-1}-\tilde{B}_G(z), \tilde{B}_G(z)-F_{\mu}^{-1}\rangle_{L^2}\\&=\langle \tilde{B}_G(z)-\tilde{B}_G(z), \tilde{B}_G(z)-F_{\mu}^{-1}\rangle_{L^2}\\&=0.
\end{align*}
Therefore,
\[\frac{1}{n}\sum_{i=1}^ns_{iG}(z)d^2_{\mathcal{W}}(\nu_i, \mu)=\frac{1}{n}\sum_{i=1}^ns_{iG}(z)d_{L^2}^2\{F_{\nu_i}^{-1}, \tilde{B}_G(z)\}+d_{L^2}^2\{\tilde{B}_G(z), F_{\mu}^{-1}\},\]
whence
\begin{align*}
\tilde{m}_G(z)&=\argmin{\mu\in \mathcal{W}}\frac{1}{n}\sum_{i=1}^ns_{iG}(z)d^2_{\mathcal{W}}(\nu_i, \mu)\\&=\argmin{\mu\in \mathcal{W}}d_{L^2}^2\{\tilde{B}_G(z), F_{\mu}^{-1}\}.
\end{align*}
One can similarly show that
\[\hat{m}_G(z)=\argmin{\mu\in \mathcal{W}}d_{L^2}^2\{\hat{B}_G(z), F_{\mu}^{-1}\}.\]
By the convexity and closedness of $\mathcal{W}$, the minimizers $\tilde{m}_G(z)$ and $\hat{m}_G(z)$, viewed as projections onto $\mathcal{W}$, exist and are unique for any $z\in\mathbb{R}^p$ \citep[chap.~3]{deut:01}.
Now consider the first term in \eqref{eq:tildemm}. The contractive property of the projection onto a closed and convex subset in the Hilbert space $L^2(0, 1)$ implies that
\begin{align*}
d_{\mathcal{W}}\{\hat{m}_G(z), \tilde{m}_G(z)\}&=d_{\mathcal{W}}[\argmin{\mu\in \mathcal{W}}d_{L^2}^2\{\hat{B}_G(z), F_{\mu}^{-1}\}, \argmin{\mu\in \mathcal{W}}d_{L^2}^2\{\tilde{B}_G(z), F_{\mu}^{-1}\}]\\&\leq d_{L^2}\{\hat{B}_G(z), \tilde{B}_G(z)\}\\&=d_{L^2}\{\frac{1}{n}\sum_{i=1}^ns_{iG}(z)F_{\hat{\nu}_i}^{-1}, \frac{1}{n}\sum_{i=1}^ns_{iG}(z)F_{\nu_i}^{-1}\}.
\end{align*}
By the triangle inequality, we have that
\begin{align*}
d_{\mathcal{W}}\{\hat{m}_G(z), \tilde{m}_G(z)\}&\leq\frac{1}{n}\sum_{i=1}^n|s_{iG}(z)|d_{L^2}(F_{\hat{\nu}_i}^{-1}, F_{\nu_i}^{-1})\\&=\frac{1}{n}\sum_{i=1}^n|s_{iG}(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i).
\end{align*}
Note that $s_{iG}(z)=s_i(z)+W_{0n}(z)+Z_i^\T W_{1n}(z)$, where
\begin{align*}
&s_i(z)=1+(Z_i-\theta)^\T\Sigma^{-1}(z-\theta),\\&W_{0n}(z)=\theta^\T\Sigma^{-1}(z-\theta)-\bar{Z}\hat{\Sigma}^{-1}(z-\bar{Z}),\\&W_{1n}(z)=\hat{\Sigma}^{-1}(z-\bar{Z})-\Sigma^{-1}(z-\theta).
\end{align*}
Both $W_{0n}(z)$ and $\|W_{1n}(z)\|_E$ are $O_p(n^{-1/2})$ by the central limit theorem. It follows that
\begin{align}
d_{\mathcal{W}}\{\hat{m}_G(z), \tilde{m}_G(z)\}&\leq\frac{1}{n}\sum_{i=1}^n|s_i(z)+W_{0n}(z)+Z_i^\T W_{1n}(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)\nonumber\\&\leq\frac{1}{n}\sum_{i=1}^n|s_i(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)\nonumber\\&\;\;\;\;+\frac{1}{n}\sum_{i=1}^n|W_{0n}(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)+\frac{1}{n}\sum_{i=1}^n|Z_i^\T W_{1n}(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)\label{eq:sww}.
\end{align}
By the Cauchy-Schwarz inequality, for the first term of \eqref{eq:sww} we have
\begin{align}
E\{\frac{1}{n}\sum_{i=1}^n|s_i(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)\}&=\frac{1}{n}\sum_{i=1}^nE\{|s_i(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)\}\nonumber\\&\leq\frac{1}{n}\sum_{i=1}^n[E\{|s_i(z)|^2\}]^{1/2}[E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\}]^{1/2}\label{eq:sdw}.
\end{align}
According to Lemma 3 in \citet{pana:16}, under Condition 2 one has
\[\liminf_{n\to\infty}\frac{\min_{1\leq i\leq n}N_i}{\lambda_n}\geq\frac{c(1-e^{-1})}{2}\quad\text{a.s.},\]
which implies that for any $i=1, \ldots, n$, one has $N_i>0$ almost surely for large enough $n$.
Note that
\[d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)=d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i\geq1}+d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i=0}.\]
The second term $d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i=0}$ hence equals $0$ almost surely for large enough $n$. Since $d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i=0}$ is dominated by $d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)<\infty$, it follows that $E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i=0}\}=0$ for large enough $n$. For the first term, the following upper bound holds by Theorem 5.1 in \citet{bobk:19},
\[E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i\geq1}|N_i\}\leq\frac{2}{N_i+1}E\{J_2(\nu)\},\]
where the so-called $J_2$-functional is defined as $J_2:\mathcal{W}^{\text{ac}}\mapsto\mathbb{R}^+\cup\{+\infty\}$, with
\[J_2(\nu)=\int_\Omega\frac{F_{\nu}(x)\{1-F_{\nu}(x)\}}{f_{\nu}(x)}dx.\]
%As pointed out in \citet{bigo:18}, $E\{J_2(\nu)\}$ is finite under Condition \ref{itm:A1}.
Note that $E\{J_2(\nu)\}$ is finite under Condition 1. Taking expectation with respect to $N_i$ yields
\[E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i\geq1}\}\leq E(\frac{2}{N_i+1})E\{J_2(\nu)\}.\]
Now by Condition 2 $N_i$ follows a Poisson distribution with parameter $c_i\lambda_n$. One has
\begin{align*}
E(\frac{2}{N_i+1})&=\sum_{k=0}^\infty\frac{2}{k+1}e^{-c_i\lambda_n}\frac{(c_i\lambda_n)^k}{k!}\\&=2\sum_{k=0}^\infty e^{-c_i\lambda_n}\frac{(c_i\lambda_n)^k}{(k+1)!}\\&=\frac{2}{c_i\lambda_n}\sum_{k=0}^\infty e^{-c_i\lambda_n}\frac{(c_i\lambda_n)^{k+1}}{(k+1)!}\\&=\frac{2}{c_i\lambda_n}\sum_{k=1}^\infty e^{-c_i\lambda_n}\frac{(c_i\lambda_n)^k}{k!}\\&=\frac{2}{c_i\lambda_n}(1-e^{-c_i\lambda_n})\\&\leq\frac{2}{c_i\lambda_n}.
\end{align*}
It follows that
\[E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i\geq1}\}\leq\frac{2E\{J_2(\nu)\}}{c_i}\lambda_n^{-1}\]
and hence for large enough $n$,
\[E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\}=E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i\geq1}\}+E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\mathbf{1}_{N_i=0}\}\leq\frac{2E\{J_2(\nu)\}}{c_i}\lambda_n^{-1},\]
which implies that
\[[E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\}]^{1/2}\leq\sqrt{\frac{2E\{J_2(\nu)\}}{c_i}}\cdot\lambda_n^{-1/2}\]
for large enough $n$. Since $[E\{|s_i(z)|^2\}]^{1/2}$ are finite and $c_i$ are uniformly bounded below by $c>0$, it follows from \eqref{eq:sdw} that
\[E\{\frac{1}{n}\sum_{i=1}^n|s_i(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)\}=O(\lambda_n^{-1/2}),\]
which implies
\[\frac{1}{n}\sum_{i=1}^n|s_i(z)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)=O_p(\lambda_n^{-1/2}).\]
By \eqref{eq:sww} and the fact that $W_{0n}(z)$ and $\|W_{1n}(z)\|_E$ are $O_p(n^{-1/2})$, one has
\[d_{\mathcal{W}}\{\hat{m}_G(z), \tilde{m}_G(z)\}=O_p(\lambda_n^{-1/2}).\]
Combined with \eqref{eq:tildemm}, we conclude that
\[d_{\mathcal{W}}\{\hat{m}_G(z), m_G(z)\}=O_p(n^{-1/2}+\lambda_n^{-1/2}).\]
For the uniform result over $\|z\|_E\leq B$, use the fact that $W_{0n}(z)$, $\|W_{1n}(z)\|_E$ are both $O_p(n^{-1/2})$, and $[E\{|s_i(z)|^2\}]^{1/2}$ is $O(1)$, uniformly over $\|z\|_E\leq B$. Applying similar arguments leads to
\[\sup_{\|z\|_E\leq B}d_{\mathcal{W}}\{\hat{m}_G(z), \tilde{m}_G(z)\}=O_p(\lambda_n^{-1/2}).\]
By Theorem 2 in \citet{mull:19:3}, one has
\[\sup_{\|z\|_E\leq B}d_{\mathcal{W}}\{\tilde{m}_G(z), m_G(z)\}=O_p(n^{-1/\{2(1+\varepsilon)\}})\]
for any $\varepsilon>0$. Again by the triangle inequality, we conclude that
\[\sup_{\|z\|_E\leq B}d_{\mathcal{W}}\{\hat{m}_G(z), m_G(z)\}=O_p(n^{-1/\{2(1+\varepsilon)\}}+\lambda_n^{-1/2}).\]
\end{proof}
\begin{proof}[Proof of Theorem 2]
We first establish the pointwise rate of convergence for the local REM estimate. By the triangle inequality, we have
\begin{equation}
\label{eq:tildemml}
d_{\mathcal{W}}\{\hat{m}_{L, h}(z), m(z)\}\leq d_{\mathcal{W}}\{\hat{m}_{L, h}(z), \tilde{m}_{L, h}(z)\} + d_{\mathcal{W}}\{\tilde{m}_{L, h}(z), m(z)\}.
\end{equation}
The second term on the right hand side, corresponding to the pointwise rate of convergence for local \f regression with fully observed measures, is $O_p(n^{-2/5})$ under Conditions A1 and A2 by Proposition 1 and Corollary 1 in \citet{mull:19:3}.
To analyze the first term, we will utilize simplified expressions of $\tilde{m}_{L, h}(z)$ and $\hat{m}_{L, h}(z)$. By similar arguments as in the proof of Theorem 1 and the fact that $n^{-1}\sum_{i=1}^ns_{iL}(z, h)=1$, one has
\begin{align*}&\tilde{m}_{L, h}(z)=\argmin{\mu\in \mathcal{W}}d_{L^2}^2\{\tilde{B}_{L, h}(z), F_{\mu}^{-1}\},\\&\hat{m}_{L, h}(z)=\argmin{\mu\in \mathcal{W}}d_{L^2}^2\{\hat{B}_{L, h}(z), F_{\mu}^{-1}\},
\end{align*}
where
\[\tilde{B}_{L, h}(z)=\frac{1}{n}\sum_{i=1}^ns_{iL}(z, h)F_{\nu_i}^{-1},\quad\hat{B}_{L, h}(z)=\frac{1}{n}\sum_{i=1}^ns_{iL}(z, h)F_{\hat{\nu}_i}^{-1},\]
with $F_{\nu_i}^{-1}$ and $F_{\hat{\nu_i}}^{-1}$ being the quantile functions of $\nu_i$ and $\hat{\nu}_i$, respectively. By the convexity and closedness of $\mathcal{W}$, the minimizers $\tilde{m}_{L, h}(z)$ and $\hat{m}_{L, h}(z)$, viewed as projections onto $\mathcal{W}$, exist and are unique for any $z\in\mathbb{R}$ \citep[chap.~3]{deut:01}.
Note that $s_{iL}(z, h)=s_i(z, h)+W_{0n}K_h(Z_i-z)+W_{1n}K_h(Z_i-z)(Z_i-z)$, where
\begin{align*}
&s_i(z, h)=\sigma_0^{-2}K_h(Z_i-z)\{\mu_2-\mu_1(Z_i-z)\},\\&W_{0n}=\frac{\hat{\mu}_2}{\hat{\sigma}_0^2}-\frac{\mu_2}{\sigma_0^2}, W_{1n}=\frac{\hat{\mu}_1}{\hat{\sigma}_0^2}-\frac{\mu_1}{\sigma_0^2}.
\end{align*}
From the proof of Lemma 2 in \citet{mull:19:3}, we have $s_i(z, h)=O_p(1)$, $W_{0n}=O_p\{(nh)^{-1/2}\}$, and $W_{1n}=O_p\{(nh^3)^{-1/2}\}$. Similar to the proof of Theorem 1 one can show that
\begin{align}
d_{\mathcal{W}}\{\hat{m}_{L, h}(z), \tilde{m}_{L, h}(z)\}&\leq\frac{1}{n}\sum_{i=1}^n|s_i(z, h)|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)+|W_{0n}|\cdot\frac{1}{n}\sum_{i=1}^nK_h(Z_i-z)d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)\nonumber\\&\;\;\;\;+|W_{1n}|\cdot\frac{1}{n}\sum_{i=1}^nK_h(Z_i-z)|Z_i-z|d_{\mathcal{W}}(\hat{\nu_i}, \nu_i)\label{eq:swwl}.
\end{align}
In the proof of Theorem 1, we have shown that
\[[E\{d_{\mathcal{W}}^2(\hat{\nu_i}, \nu_i)\}]^{1/2}\leq\sqrt{\frac{2E\{J_2(\nu)\}}{c_i}}\cdot\lambda_n^{-1/2}\]
for large enough $n$. Under Conditions 1 and 2, $E\{J_2(\nu)\}$ is finite and $c_i$ are uniformly bounded below by $c>0$. It follows that
\[d_{\mathcal{W}}(\hat{\nu}_i, \nu_i)=O_p(\lambda_n^{-1/2}).\]
We hence have
\[\frac{1}{n}\sum_{i=1}^nK_h(Z_i-z)|Z_i-z|^kd_{\mathcal{W}}(\hat{\nu_i}, \nu_i)=O_p(h^k\lambda_n^{-1/2})\]
for $k=0, 1$. Since $s_i(z, h)=O_p(1)$, $W_{0n}=O_p\{(nh)^{-1/2}\}$, and $W_{1n}=O_p\{(nh^3)^{-1/2}\}$, it follows from \eqref{eq:swwl} that
\[d_{\mathcal{W}}\{\hat{m}_{L, h}(z), \tilde{m}_{L, h}(z)\}=O_p(\lambda_n^{-1/2}).\]
With \eqref{eq:tildemml}, we conclude that
\[d_{\mathcal{W}}\{\hat{m}_{L, h}(z), m_{L, h}(z)\}=O_p(n^{-2/5}+\lambda_n^{-1/2}).\]
For the uniform result over $z\in\mathcal{T}$, we use the fact that $s_i(z, h)=O_p(1)$, $W_{0n}=O_p\{(nh)^{-1/2}\}$, and $W_{1n}=O_p\{(nh^3)^{-1/2}\}$, uniformly over $z\in\mathcal{T}$. Applying similar arguments leads to
\[\sup_{z\in\mathcal{T}}d_{\mathcal{W}}\{\hat{m}_{L, h}(z), \tilde{m}_{L, h}(z)\}=O_p(\lambda_n^{-1/2}).\]
By Theorem 1 in \cite{mull:22:8}, under Conditions A3 and A4 one has
\[\sup_{z\in\mathcal{T}}d_{\mathcal{W}}\{\tilde{m}_{L, h}(z), m(z)\}=O_p(n^{-1/(3+\varepsilon)})\]
for any $\varepsilon>0$. Again by the triangle inequality, we conclude that
\[\sup_{z\in\mathcal{T}}d_{\mathcal{W}}\{\hat{m}_{L, h}(z), m(z)\}=O_p(n^{-1/(3+\varepsilon)}+\lambda_n^{-1/2}).\]
\end{proof}
\fi
\single
% \bibliographystyle{dcu}% abbrvnat, agsm
\bibliographystyle{rss}
\bibliography{collection}
\end{document}
\usepackage[pdftex]{graphicx}
\usepackage{epstopdf}
\usepackage{natbib}
\usepackage{geometry,rotate,lscape,epsfig,booktabs,%setst,
colortbl,xspace,sectsty,amsthm,float,xy,xr,enumitem,%enumerate
amssymb,amsfonts,amsbsy,amsmath,mathtools,delarray,mathtools,
bm,color,multirow,latexsym,fancyhdr,fontenc,anysize,url,caption,subcaption,%subfigure,
rotating,ifthen,fancybox,fullpage,amscd,pifont,setspace,xargs}
\usepackage[mathscr]{eucal} %\command: \mathscr{A} gives script A
\usepackage{cleveref}
\usepackage{wrapfig}
%\usepackage[colorlinks=true, allcolors=blue]{hyperref}
%\renewcommand{\itemautorefname}{\hspace{-.4em}}
%\renewcommand{\sectionautorefname}{Section}
%\renewcommand{\subsectionautorefname}{Subsection}
%\newcommand{\Theoremautorefname}{Theorem}
%\newcommand{\Lemmaautorefname}{Lemma}
%\newcommand{\Corollaryautorefname}{Corollary}
%\usepackage{sectsty}
\usepackage{titlesec}
\titleformat*{\section}{\large\bfseries\sffamily}
\titleformat*{\subsection}{\small\bfseries\sffamily}
\titleformat*{\subsubsection}{\footnotesize\bfseries\sffamily}
\geometry{letterpaper,left=1.2in,right=1.2in,top=1.2in,bottom=1.1in}
\bibpunct{(}{)}{;}{a}{}{,}
\tolerance=500 \allowdisplaybreaks
%\externaldocument{Petersen_Mueller_AOS1624}
%\newgeometry{left=2cm,right=2cm,top=2cm,bottom=.1cm}
%graphicspath{{figures/}}
%\input{/Users/hans/Documents/hanstex/macros/macros1}
%\setlength{\parindent}{0pt} no paragraph indentation
%\cite[see, e.g.,][]{rice:91} citation with text in brackets
%\setcounter{page}{1}
%\setcounter{enumi}{171}
%\setcounter{section}{8}
%\pagenumbering{gobble}
%\def\references{\bibliography{/Users/hans/Documents/hanstex/macros/bib/1-1-18}}
%note: if several bib files, list with comma between them and no extra space
%\bibliographystyle{/Users/hans/Documents/hanstex/macros/style/jasa}
%\bibliographystyle{acm}
%\newcommandx*\coord[3][1=1, 3=n]{(#2_{#1},\ldots,#2_{#3})}
%Examples: $\coord{x}$(x1,...,xn)
%$\coord[0]{y}$(y0,...,yn)
%$\coord{z}[m]$(z1,...,zm)
%Letters:
%Letterhead: /Users/hans/Documents/hanstex/macros/style/letterhead-14-1.pdf
%Signature: /Users/hans/Documents/hanstex/macros/style/signature-eps-converted-to.pdf
%\geometry{letterpaper,left=1.5in,right=1.5in,top=.75in,bottom=1in}
%\newgeometry{left=2cm,right=2cm,top=2cm,bottom=.1cm}-- replaces previous geometry, useful for letters
%the \geometry framework with a new one in the middle of the page
%\renewcommand{\familydefault}{\sfdefault}
%\usepackage[scaled=0.96]{helvet} %Helvetica
%\setlength{\parindent}{0pt}
%\setlength{\parskip}{1.1ex plus 0.5ex minus 0.3ex}
%\renewcommand{\sffamily}{\sfdefault}
%\sffamily\rmfamily\ttfamily
%\newcommand{\single}{\renewcommand{\baselinestretch}{1.2}\normalsize}
%\newcommand{\double}{\renewcommand{\baselinestretch}{1.63}\normalsize}
%\begin{enumerate}[(i)]
%\pagestyle{empty} -- no page no.
%\varepsilon_{i}\stackrel{\mbox{i.i.d.}}{\sim} N(0, 1)$
%\def\stackunder#1#2{\mathrel{\mathop{#2}\limits_{#1}}}
%\stackunder{x \in A} {\text{argmax}} H(x)
%\underset{below}{\rightarrow}
%\overset{above}{\rightarrow}
%floating figures and tables
%\usepackage{float}
%\begin{table}[H]
%to force position
%\bfseries \mathversion{bold}
%\small\footnotesize\scriptsize\tiny\large\Large\LARGE\huge\Huge
%\renewcommand{\rmdefault}{ptm} switch font to Times
%\thispagestyle{empty}
%fontsizes: tiny scriptsize footnotesize small normalsize large Large LARGE huge Huge
%\graphicspath{{Figs140312/blsa130815_03/},{Figs140312/sim130129/}}
%\graphicspath{{subdir1/}{subdir2/}{subdir3/}...{subdirn/}} important: / at the end!
%sizing figures: \scalebox{0.5}{\includegraphics{figure}}
%\includegraphics[scale=0.195]{figure} \includegraphics[width=0.5in, height=0.3in]{figure}
%figures side-by-side:
%\begin{figure}[ht]\centering\begin{minipage}[b]{0.35\textwidth}\subcaptionbox{Normal Group}{\includegraphics[scale=0.75]{Figures/normal_corr1}}
%\label{fig:minipage1}\end{minipage}\quad\begin{minipage}[b]{0.35\textwidth}
%\subcaptionbox{Alzheimer's Group}{\includegraphics[scale=0.75]{Figures/dem_corr1}}\label{fig:minipage2}
%\end{minipage}\end{figure}
%\graphicspath{{fig-blsa-150331/}}
%The extsizes classes (extarticle, extreport etc.) provide support for sizes eight, nine, ten,
%eleven, twelve, fourteen, seventeen and twenty points.
%For supplements: \externaldocument{filename-maindoc} to reference eq no
%\be \label{eq1}
%\begin{split} \bsp
%A & = \frac{\pi r^2}{2} \\
%& = \frac{1}{2} \pi r^2
%\end{split} \esp
%\ee
%\setlength{\parindent}{0pt}
%\setlength{\parskip}{1.3ex plus 0.5ex minus 0.3ex}
%Special for Windows:
%\usepackage{C:/hanstex/macros/style/hypernat}
%mathptmx for formulas set in Times
%pstricks,psfrag, graphicx,graphics,
%\RequirePackage{C:/hanstex/macros/style/hypernat}
%\input{C:/hanstex/macros/macros1}
%\include command does not work anymore
%\usepackage{html}
%\def\references{\bibliography{C:/hanstex/macros/bib/11-1-11}}
%\bibliographystyle{C:/hanstex/macros/style/jasa}
%\bibliographystyle{asa}
%\renewcommand\refname{} allows to access bib file in same folder
% bibfile: "and others" at end of name list -> et al.
% word: remove properties file - tools - remove properties
%force old version of adobe for JASA sub:
%\pdfminorversion=4
%remove figures/tables:
%\usepackage{environ}
%\RenewEnviron{figure}{}
%\RenewEnviron{table}{}
%\def\note{\textcolor{gray}}
%\def\note#1{\iffalse #1 \fi}
\crefformat{equation}{(#2#1#3)}
\crefrangeformat{equation}{(#3#1#4) to~(#5#2#6)}
\crefname{equation}{}{}
\Crefname{Equation}{}{}
%cite eq. by \cref{label} in text or \Cref{label} at the beginning of a sentence
\def\jasaformat{\setlength{\textheight}{8.5in} \setlength{\textwidth}{6.3in}
\setlength{\topmargin}{0.2in} \setlength{\oddsidemargin}{0.1in}
\setlength{\evensidemargin}{0.12in} \tolerance=500 \hoffset 0in
\marginparsep 0in \marginparwidth 0in
\voffset 0in %-0.3in
\headsep 0in %0.5
\footskip 28pt \setlength{\baselineskip}{22pt}
\setlength{\parskip}{0.0in}
\renewcommand{\baselinestretch}{1.6}
\renewcommand{\arraystretch}{1.0}}
\definecolor{DarkBlue}{rgb}{0,.08,.45}
\definecolor{DarkRed}{rgb}{.7,0,.4}
\def\bla{\textcolor{black}}
\def\blu{\textcolor{blue}}
\def\red{\textcolor{red}}
\def\hg #1 {\texcolor{cyan}{{\it Hans:} #1}}
\def\rred{\textcolor{DarkRed}}
\def\bblu{\textcolor{DarkBlue}}
\def\br{\begin{\red}}
\def\er{\end{\red}}
\def\bb{\begin{\blu}}
\def\eb{\end{\blu}}
\newcommand{\rsa}{\rightsquigarrow}
\newcommand{\ct}{\lincmmt}%followed by {}, will b assigned red color
\newcommand{\ig}{\includegraphics}
\newcommand{\hd}{\hangindent=0.5cm\hangafter=1\noindent}
%\newcommand{\bea}{\begin{eqnarray*}}
%\newcommand{\eea}{\end{eqnarray*}}
%\newcommand{\be}{\begin{eqnarray}}
%\newcommand{\ee}{\end{eqnarray}}
\newcommand{\be}{\begin{equation}}
\newcommand{\ee}{\end{equation}}
\newcommand{\ben}{\begin{equation}}
\newcommand{\een}{\end{equation}}
\newcommand{\ba}{\begin{align}}
\newcommand{\ea}{\end{align}}
\newcommand{\ban}{\begin{align*}}
\newcommand{\ean}{\end{align*}}
\newcommand{\bsp}{\begin{split}}
\newcommand{\esp}{\end{split}}
\newcommand{\ed}{\end{document}}
\newcommand{\no}{\noindent}
\newcommand{\et}{\textit{et al. }}
\newcommand{\btab}{\begin{tabular}}
\newcommand{\etab}{\end{tabular}}
\newcommand{\bc}{\begin{center}}
\newcommand{\ec}{\end{center}}
\newcommand{\np}{\newpage}
\newcommand{\la}{\label}
\newcommand{\bi}{\begin{itemize}}
\newcommand{\ei}{\end{itemize}}
\newcommand{\bfi}{\begin{figure}}
\newcommand{\efi}{\end{figure}}
\newcommand{\bem}{\begin{enumerate}}
\newcommand{\eem}{\end{enumerate}}
\newcommand{\bdes}{\begin{description}}
\newcommand{\edes}{\end{description}}
\newcommand{\bay}{\begin{array}}
\newcommand{\eay}{\end{array}}
\newcommand{\bs}{\boldsymbol}
\newcommand{\nn}{\nonumber}
%\newcommand{\nl}{\newline}%conflict with algorithm2e
\newtheorem{thm}{Theorem}
\newtheorem{lem}{Lemma}
\newtheorem{crl}{Corollary}
\newtheorem{prop}{Proposition}
\newtheorem{rem}{Remark}
\newtheorem{exm}{Example}
\newtheorem{asmp}{Assumption}
%\newcommand{\bass}{\begin{assumption}}
%\newcommand{\eass}{\end{assumption}}
%\newcommand{\bthm}{\begin{theorem}}
%\newcommand{\ethm}{\end{theorem}}
%\newcommand{\blem}{\begin{lemma}}
%\newcommand{\elem}{\end{lemma}}
\newcommand{\ip}[2]{\langle #1,#2\rangle}
\newcommand{\scp}[2]{\langle #1,#2\rangle}
\newcommand{\ve}[1]{{\mbox{\boldmath ${#1}$}}}
%bring expression 1 into a bolded vector form
\def\ss{\subsection*}
\def\sss{\subsubsection*}
\newcommand{\single}{\renewcommand{\baselinestretch}{1.2}\normalsize}
\newcommand{\double}{\renewcommand{\baselinestretch}{1.6}\normalsize}
\def\stackunder#1#2{\mathrel{\mathop{#2}\limits_{#1}}}
%\stackunder{x \in A} {\text{argmax}} H(x)
\def\bco{\iffalse}
\def\cov{{\rm cov}}
\def\Cov{{\rm Cov}}
\def\var{{\rm var}}
\def\Var{{\rm Var}}
\def\corr{{\rm corr}}
\def\diag{{\rm diag}}
\def\trace{{\rm trace}}
\def\logit{\hbox{logit}}
\def\expit{\hbox{expit}}
\def\ci{\cite}
\def\cp{\citep}
%\cp[e.g.,][]{\mull:09:1}
%\def\F{Fr\'echet}
%\def\FF{Fr\'echet }
\def\f{Fr\'echet }
\def\rs{\vspace{-.24cm}}
\def\sm{\vspace{.2cm}}
\def\vm{\vspace{-1cm}}
\def\ssd{\vspace{-.1cm}}
\def\sd{\vspace{-.12in}}
\def\hs{\hspace{0.2cm}}
\def\vsp{\vspace{0.3cm}}
\def\vs{\vspace{.125cm}}
\def\vvs{\vspace{0.1cm}}
\def\vbs{\vspace{.7cm}}
\def\vs{\vspace{.125cm}}
\def\sn{\sum_{i=1}^n}
\def\xbar{\bar{X}}
\def\inv{^{-1}}
\def\xk{A_k}
\def\xik{A_{ik}}
\def\hxk{\hat{A}_k}
\def\hxik{\hat{A}_{ik}}
\def\tX{\tilde{X}}
\def\tY{\tilde{Y}}
\def\tij{t_{ij}}
\def\Yij{Y_{ij}}
\def\Xij{X_{ij}}
\def\hphi{{\widehat\phi}}
\def\hpsi{{\widehat\psi}}
\def\ij{_{ij}}
\def\i1{_{i1}}
\def\ik{_{ik}}
\def\ki{_{ki}}
\def\mhf{^{-1/2}}
\def\mo{^{-1}}
\def\ij{_{ij}}
\def\ijk{_{ijk}}
\def\ik{_{ik}}
\def\jj{_{jj}}
\def\jk{_{jk}}
\def\half{^{1/2}}
\def\eps{\epsilon}
\def\refmark{\par\vskip 2mm\noindent\refhg}
\def\gotolimit#1{\mathrel{\mathop{\longrightarrow}\limits^{#1}}}
%\def\argmin#1{\mathrel{\mathop{\arg\min}\limits_{#1}}}
%\def\argmax#1{\mathrel{\mathop{\arg\max}\limits_{#1}}}
\DeclareMathOperator*{\argmax}{argmax}
\DeclareMathOperator*{\argmin}{argmin}
%\def\ao{\stackunder{\om \in \Om} {\text{argmin}}}
\def\eyx{E(Y|X=x)}
\def\inv{^{-1}}
\def\Ltwoip#1#2{\langle #1, #2\rangle_{L^2}}
\def\Ltwonorm#1{\lVert #1 \rVert_{L^2}}
%\def\Ltwo#1#2{d_{L^2}(#1, #2)}
\def\Froip#1#2{\langle #1, #2\rangle_{F}}
\def\Fronorm#1{\lVert #1 \rVert_F}
\def\Fro#1#2{d_F(#1,#2)}
%\DeclareMathOperator*{\argmax}{arg\,max}
%\DeclareMathOperator*{\argmin}{arg\,min}
\def\argmin#1{\mathrel{\mathop{\arg\min}\limits_{#1}}}
\def\argmax#1{\mathrel{\mathop{\arg\max}\limits_{#1}}}
%\def\argmax{{\rm argmax}\,}Map time: Wed Mar 20 2019 15:25:30 GMT-0700 (Pacific Daylight Time)
\def\as{^*}
\newcommand*{\cA}{\mathcal{A}}
\newcommand*{\cH}{\mathcal{H}}
\newcommand*{\cE}{\mathcal{E}}
\newcommand*{\cF}{\mathcal{F}}
\def\cB{\mathcal{B}}
\def\cI{\mathcal{I}}
\def\cP{\mathcal{P}}
\def\cR{\mathcal{R}}
\def\cS{\mathcal{S}}
\def\cT{\mathcal{T}}
\def\cW{\mathcal{W}}
\def\bR{\mathbf{R}}
\def\pk{\phi_k}
\def\povr{\buildrel p\over\longrightarrow}
\def\pr{\hbox{pr}}
\def\wh{\widehat}
\def\wt{\widetilde}
\def\lra{\longrightarrow}
\def\ra{\rightarrow}
\def\E{\hbox{E}}
\def\det{\hbox{det}}
\def\foot#1#2{{\baselineskip=9pt\sevenrm\footnote{$^{#1}$}{#2}}}
\def\nh{\noindent\hangindent=1.5truecm\hangafter=1}
\newcommand{\tinT}{t\in \cT}
\newcommand{\supt}{\sup_{t\in\cT}}
\newcommand{\supst}{\sup_{s,t\in\cT}}
\newcommand{\almostsure}{\text{a.s.}}
\newcommand{\quadas}{\quad \almostsure}
\newcommand{\bpmt}{\begin{pmatrix}}
\newcommand{\epmt}{\end{pmatrix}}
\def\hy{\hat{y}}
\def\hyi{\hat{y}^{(-i)}}
\def\pal{\partial}
%\allowdisplaybreaks
\def\mc{\mathcal}
\def\mx{\mbox}
\def\mr{\mathrm}
\def\mb{\mathbf}
\def\mbb{\mathbb}
\def\real{\mathbb{R}}
\def\argmax{{\rm argmax}\,}
\def\half{^{1/2}}
\def\s1n{\sum_{i=1}^n}
\def\p1n{\prod_{i=1}^n}
\def\i01{\int_0^1}
\def\hp{http://anson.ucdavis.edu/$\sim$}
\newcommand{\Ltwo}[1]{\lVert #1 \rVert_2}
%L2 norm of [1]
\newcommand{\unif}[1]{\lVert #1 \rVert_\infty}
%sup norm of [1]
\newenvironment{lyx}[1]
{\begin{list}{}
{\settowidth{\labelwidth}{#1}
\setlength{\leftmargin}{\labelwidth}
\addtolength{\leftmargin}{\labelsep}
\renewcommand{\makelabel}[1]{##1\hfil}}}
{\end{list}}
% \renewcommand{\theequation}{A.\arabic{equation}}
%% redefine the command that creates the equation no.
% \setcounter{equation}{0} % reset counter
%% \section*{APPENDIX} % use *-form to suppress numbering
%\newcounter{Bcount2}
% set the "default" label to print counter as A1.1 etc
%\vspace{-0.1in}
%\begin{list}{(B\arabic{Bcount2})}
% {\usecounter{Bcount2}
% \setlength{\rightmargin}{\leftmargin}}
% \item
%\end{list}
%\begin{list}{(B\arabic{Bcount2})}
% {\usecounter{Bcount2}
% \addtocounter{Bcount2}{1}
% \setlength{\rightmargin}{\leftmargin}}
% \item ....
%\end{list}
\begin{thebibliography}{38}
\expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
\expandafter\ifx\csname url\endcsname\relax
\def\url#1{\texttt{#1}}\fi
\expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL: }\fi
\bibitem[{Anderson and Whitaker(2009)}]{ande:09}
Anderson, S.~E. and Whitaker, R.~C. (2009) Prevalence of obesity among {US}
preschool children in different racial and ethnic groups.
\newblock \textit{Archives of Pediatrics \& Adolescent Medicine}, \textbf{163},
344--348.
\bibitem[{Bigot et~al.(2017)Bigot, Gouet, Klein and L{\'o}pez}]{bigo:17}
Bigot, J., Gouet, R., Klein, T. and L{\'o}pez, A. (2017) {Geodesic {PCA} in the
{W}asserstein space by convex {PCA}}.
\newblock \textit{Annales de l{'}Institut Henri Poincar\'{e} B: Probability and
Statistics}, \textbf{53}, 1--26.
\bibitem[{Bigot et~al.(2018)Bigot, Gouet, Klein and Lopez}]{bigo:18}
Bigot, J., Gouet, R., Klein, T. and Lopez, A. (2018) Upper and lower risk
bounds for estimating the {W}asserstein barycenter of random measures on the
real line.
\newblock \textit{Electronic Journal of Statistics}, \textbf{12}, 2253--2289.
\bibitem[{Bonevski et~al.(2014)Bonevski, Randell, Paul, Chapman, Twyman,
Bryant, Brozek and Hughes}]{bone:14}
Bonevski, B., Randell, M., Paul, C., Chapman, K., Twyman, L., Bryant, J.,
Brozek, I. and Hughes, C. (2014) Reaching the hard-to-reach: a systematic
review of strategies for improving health and medical research with socially
disadvantaged groups.
\newblock \textit{BMC Medical Research Methodology}, \textbf{14}, 1--29.
\bibitem[{Chen et~al.(2020)Chen, Gajardo, Fan, Zhong, Dubey, Bhattacharjee, Han
and M{\"u}ller}]{chen:20}
Chen, Y., Gajardo, A., Fan, J., Zhong, Q., Dubey, P., Bhattacharjee, S., Han,
K. and M{\"u}ller, H. (2020) \textit{frechet: statistical analysis for random
objects and non-{Eu}clidean data}.
\newblock R package version 0.2. 0.
\bibitem[{Chen et~al.(2023)Chen, Lin and M{\"u}ller}]{chen:23:1}
Chen, Y., Lin, Z. and M{\"u}ller, H.-G. (2023) {{Wasserstein} regression}.
\newblock \textit{Journal of the American Statistical Association},
\textbf{118}, 869--882.
\bibitem[{Chen and M{\"u}ller(2022)}]{mull:22:8}
Chen, Y. and M{\"u}ller, H.-G. (2022) {Uniform convergence of local {F}r\'echet
regression, with applications to locating extrema and time warping for
metric-space valued trajectories}.
\newblock \textit{Annals of Statistics}, \textbf{50}, 1573--1592.
\bibitem[{Danielzik et~al.(2004)Danielzik, Czerwinski-Mast, Langn{\"a}se, Dilba
and M{\"u}ller}]{dani:04}
Danielzik, S., Czerwinski-Mast, M., Langn{\"a}se, K., Dilba, B. and M{\"u}ller,
M. (2004) Parental overweight, socioeconomic status and high birth weight are
the major determinants of overweight and obesity in 5--7 y-old children:
baseline data of the {Kiel Obesity Prevention Study (KOPS)}.
\newblock \textit{International Journal of Obesity}, \textbf{28}, 1494--1502.
\bibitem[{Fletcher(2013)}]{flet:13}
Fletcher, P.~T. (2013) {Geodesic regression and the theory of least squares on
{R}iemannian manifolds}.
\newblock \textit{International Journal of Computer Vision}, \textbf{105},
171--185.
\bibitem[{Fr\'{e}chet(1948)}]{frec:48}
Fr\'{e}chet, M. (1948) {Les \'{e}l\'{e}ments al\'{e}atoires de nature
quelconque dans un espace distanci\'{e}}.
\newblock \textit{Annales de l'Institut Henri Poincar\'{e}}, \textbf{10},
215--310.
\bibitem[{Gajardo and M{\"u}ller(2021)}]{mull:21:10}
Gajardo, {\'A}. and M{\"u}ller, H.-G. (2021) {Cox} point process regression.
\newblock \textit{IEEE Transactions on Information Theory}, \textbf{68},
1133--1156.
\bibitem[{Ghodrati and Panaretos(2022)}]{ghod:21}
Ghodrati, L. and Panaretos, V.~M. (2022) Distribution-on-distribution
regression via optimal transport maps.
\newblock \textit{Biometrika}, \textbf{109}, 957--974.
\bibitem[{Gillman(2022)}]{gill:22}
Gillman, M. (2022) {Environmental Influences on Child Health Outcomes
(ECHO)-wide Cohort. NICHD Data and Specimen Hub.
https://dash.nichd.nih.gov/study/417122}.
\newblock \urlprefix\url{https://dash.nichd.nih.gov/study/417122}.
\bibitem[{Gillman and Blaisdell(2018)}]{gill:18}
Gillman, M.~W. and Blaisdell, C.~J. (2018) {Environmental influences on Child
Health Outcomes, a research program of the NIH}.
\newblock \textit{Current Opinion in Pediatrics}, \textbf{30}, 260.
\bibitem[{Hron et~al.(2016)Hron, Menafoglio, Templ, Hruzova and
Filzmoser}]{hron:16}
Hron, K., Menafoglio, A., Templ, M., Hruzova, K. and Filzmoser, P. (2016)
{Simplicial principal component analysis for density functions in Bayes
spaces}.
\newblock \textit{Computational Statistics and Data Analysis}, \textbf{94},
330--350.
\bibitem[{Jank and Shmueli(2010)}]{jank:10}
Jank, W. and Shmueli, G. (2010) \textit{Modeling Online Auctions}.
\newblock John Wiley \& Sons.
\bibitem[{Kantorovich(1942)}]{kant:42}
Kantorovich, L.~V. (1942) {On the translocation of masses}.
\newblock \textit{Dokl. Akad. Nauk SSSR (translated version in Journal of
Mathematical Sciences, 133, 1381-1382, 2006)}, \textbf{37}, 227--229.
\bibitem[{Kloeckner(2010)}]{kloe:10}
Kloeckner, B. (2010) {A geometric study of {W}asserstein spaces: Euclidean
spaces}.
\newblock \textit{Annali della Scuola Normale Superiore di Pisa-Classe di
Scienze}, \textbf{9}, 297--323.
\bibitem[{Knapp et~al.(2023)Knapp, Kress, Parker, Page, McArthur, Gachigi,
Alshawabkeh, Aschner, Bastain, Breton et~al.}]{knap:23}
Knapp, E.~A., Kress, A.~M., Parker, C.~B., Page, G.~P., McArthur, K., Gachigi,
K.~K., Alshawabkeh, A.~N., Aschner, J.~L., Bastain, T.~M., Breton, C.~V.
et~al. (2023) {The Environmental influences on Child Health Outcomes
(ECHO)-wide Cohort}.
\newblock \textit{American Journal of Epidemiology}, in press.
\bibitem[{Liu and M\"{u}ller(2009)}]{mull:09:1}
Liu, B. and M\"{u}ller, H.-G. (2009) {Estimating derivatives for samples of
sparsely observed functions, with application to on-line auction dynamics.}
\newblock \textit{Journal of the American Statistical Association},
\textbf{104}, 704--714.
\bibitem[{Liu et~al.(2019)Liu, Liu and Chan}]{liu:19}
Liu, W.~W., Liu, Y. and Chan, N.~H. (2019) Modeling {eBay} price using
stochastic differential equations.
\newblock \textit{Journal of Forecasting}, \textbf{38}, 63--72.
\bibitem[{Niles-Weed and Berthet(2022)}]{nile:22}
Niles-Weed, J. and Berthet, Q. (2022) Minimax estimation of smooth densities in
{W}asserstein distance.
\newblock \textit{Annals of Applied Statistics}, \textbf{50}, 1519--1540.
\bibitem[{O’Connor et~al.(2022)O’Connor, Spry, Patton, Moreno-Betancur,
Arnup, Downes, Goldfeld, Burgner and Olsson}]{ocon:22}
O’Connor, M., Spry, E., Patton, G., Moreno-Betancur, M., Arnup, S., Downes,
M., Goldfeld, S., Burgner, D. and Olsson, C.~A. (2022) Better together:
Advancing life course research through multi-cohort analytic approaches.
\newblock \textit{Advances in Life Course Research}, 100499.
\bibitem[{Panaretos and Zemel(2016)}]{pana:16}
Panaretos, V.~M. and Zemel, Y. (2016) {Amplitude and phase variation of point
processes}.
\newblock \textit{Annals of Statistics}, \textbf{44}, 771--812.
\bibitem[{Panaretos and Zemel(2020)}]{pana:20}
--- (2020) \textit{{An Invitation to Statistics in {W}asserstein Space}}.
\newblock Springer New York.
\bibitem[{Pegoraro and Beraha(2022)}]{pego:22}
Pegoraro, M. and Beraha, M. (2022) Projected statistical methods for
distributional data on the real line with the {W}asserstein metric.
\newblock \textit{Journal of Machine Learning Research}, \textbf{23}, 37--1.
\bibitem[{Petersen et~al.(2021)Petersen, Liu and Divani}]{pete:21}
Petersen, A., Liu, X. and Divani, A.~A. (2021) {{Wasserstein} $ F $-tests and
confidence bands for the Fr\'{e}chet regression of density response curves}.
\newblock \textit{Annals of Statistics}, \textbf{49}, 590--611.
\bibitem[{Petersen and M\"{u}ller(2016)}]{mull:16:1}
Petersen, A. and M\"{u}ller, H.-G. (2016) {Functional data analysis for density
functions by transformation to a {H}ilbert space}.
\newblock \textit{Annals of Statistics}, \textbf{44}, 183--218.
\bibitem[{Petersen and M\"{u}ller(2019)}]{mull:19:3}
--- (2019) {Fr\'{e}chet regression for random objects with {E}uclidean
predictors}.
\newblock \textit{Annals of Statistics}, \textbf{47}, 691--719.
\bibitem[{Petersen et~al.(2022)Petersen, Zhang and Kokoszka}]{pete:22}
Petersen, A., Zhang, C. and Kokoszka, P. (2022) {Modeling probability density
functions as data objects}.
\newblock \textit{Econometrics and Statistics}, \textbf{21}, 159--178.
\bibitem[{Qiu et~al.(2022)Qiu, Dai and Zhu}]{qiu:22:2}
Qiu, J., Dai, X. and Zhu, Z. (2022) Nonparametric estimation of repeated
densities with heterogeneous sample sizes.
\newblock \textit{Journal of the American Statistical Association}, 1--13.
\bibitem[{Stellato et~al.(2020)Stellato, Banjac, Goulart, Bemporad and
Boyd}]{stel:20}
Stellato, B., Banjac, G., Goulart, P., Bemporad, A. and Boyd, S. (2020) {OSQP}:
An operator splitting solver for quadratic programs.
\newblock \textit{Mathematical Programming Computation}, \textbf{12}, 637--672.
\bibitem[{Talsk{\'a} et~al.(2018)Talsk{\'a}, Menafoglio, Machalov{\'a}, Hron
and Fi{\v{s}}erov{\'a}}]{tals:18}
Talsk{\'a}, R., Menafoglio, A., Machalov{\'a}, J., Hron, K. and
Fi{\v{s}}erov{\'a}, E. (2018) {Compositional regression with functional
response}.
\newblock \textit{Computational Statistics \& Data Analysis}, \textbf{123},
66--85.
\bibitem[{Vazquez and Cubbin(2020)}]{vazq:20}
Vazquez, C.~E. and Cubbin, C. (2020) Socioeconomic status and childhood
obesity: a review of literature from the past decade to inform intervention
research.
\newblock \textit{Current Obesity Reports}, \textbf{9}, 562--570.
\bibitem[{Villani(2003)}]{vill:03}
Villani, C. (2003) \textit{{Topics in Optimal Transportation}}.
\newblock American Mathematical Society.
\bibitem[{Wang et~al.(2008)Wang, Jank, Shmueli and Smith}]{wang:08:2}
Wang, S., Jank, W., Shmueli, G. and Smith, P. (2008) Modeling price dynamics in
{eBay} auctions using principal differential analysis.
\newblock \textit{Journal of the American Statistical Association},
\textbf{103}, 1100--1118.
\bibitem[{Zhang et~al.(2022)Zhang, Kokoszka and Petersen}]{zhan:22}
Zhang, C., Kokoszka, P. and Petersen, A. (2022) {Wasserstein} autoregressive
models for density time series.
\newblock \textit{Journal of Time Series Analysis}, \textbf{43}, 30--52.
\bibitem[{Zhu and M{\"u}ller(2023)}]{mull:23:2}
Zhu, C. and M{\"u}ller, H.-G. (2023) Autoregressive optimal transport models.
\newblock \textit{Journal of the Royal Statistical Society Series B:
Statistical Methodology}, \textbf{85}, 1012--1033.
\end{thebibliography}