Empirical measures and random walks on compact spaces in the quadratic Wasserstein metric

TL;DR

This paper establishes sharp bounds for the convergence rate of empirical measures in the quadratic Wasserstein metric on compact manifolds, extending classical results and analyzing random walks on Lie groups.

Contribution

It provides new asymptotic and nonasymptotic bounds for empirical measures in Wasserstein distance on manifolds, including nonstationary samples like random walks on Lie groups.

Findings

Bounds match classical optimal matching rates on the cube

Random walks on semisimple groups achieve near-optimal rates without spectral gaps

Fourier analysis and Berry-Esseen inequalities are key tools

Abstract

Estimating the rate of convergence of the empirical measure of an i.i.d. sample to the reference measure is a classical problem in probability theory. Extending recent results of Ambrosio, Stra and Trevisan on 2-dimensional manifolds, in this paper we prove sharp asymptotic and nonasymptotic upper bounds for the mean rate in the quadratic Wasserstein metric $W_2$ on a $d$-dimensional compact Riemannian manifold. Under a smoothness assumption on the reference measure, our bounds match the classical rate in the optimal matching problem on the unit cube due to Ajtai, Koml\'os, Tusn\'ady and Talagrand. The i.i.d. condition is relaxed to stationary samples with a mixing condition. As an example of a nonstationary sample, we also consider the empirical measure of a random walk on a compact Lie group. Surprisingly, on semisimple groups random walks attain almost optimal rates even without a spectral gap assumption. The proofs are based on Fourier analysis, and in particular on a Berry-Esseen smoothing inequality for $W_2$ on compact manifolds, a result of independent interest with a wide range of applications.