Asymptotic distribution and convergence rates of stochastic algorithms for entropic optimal transportation between probability measures (Q2039796)

Consider the estimation of a regularized Wasserstein distance \(W_\varepsilon(\mu,\nu)\) (also known as a Sinkhorn divergence) between two probability measures \(\mu\) and \(\nu\) in the case where \(\nu\) is discrete with finite support. This can be tackled with a stochastic Robbins-Monro algorithm using a sequence of observations from \(\mu\). In this paper, the authors establish asymptotic properties of such algorithms. In particular, their main results are almost sure convergence and asymptotic normality of the resulting estimator of the Sinkhorn divergence, as well as analysis of the expected excess risk of the estimator. Further, the authors also discuss the unregularized case (corresponding to the choice \(\varepsilon=0\)) in which case the Sinkhorn divergence becomes a Wasserstein distance. Numerical experiments illustrate the results of the paper.