Central limit theorems for entropy-regularized optimal transport on finite spaces and statistical applications (Q2283574)

The notion of entropy-regularized optimal transport (known as Sinkhorn divergence) makes feasible the use of smoothed optimal transportation distances for data analysis. The Sinkhorn divergence allows the fast computation of an entropically regularized Wasserstein distance between two probability distributions supported on a finite metric space of (possibly) high-dimension. The goal of the paper under review is to analyze the potential benefits of the Sinkhorn divergence and its centered version for statistical inference from empirical probability measures. The authors derive results on the asymptotic distribution of such divergences for data sampled from (unknown) distributions supported on a finite metric space. These results allow building new test statistics for measuring the discrepancies between multivariate probability distributions. The paper is organized as follows. In Section 2, the optimal transport problem between probability measures is recalled, and the notions of Sinkhorn divergence and Sinkhorn loss are introduced. Next, the asymptotic distributions for the empirical Sinkhorn divergence and the empirical Sinkhorn loss are derived. Also, the behavior of such statistics (when the regularization parameter tends to zero at a rate depending on the number of available observations) is analyzed. A bootstrap procedure to estimate unknown quantities of interest on the distribution of these statistics (such as their non-asymptotic variance and quantiles) is discussed in Section 3. The proposed approach is illustrated by simulated and real datasets in Section 4 and in Section 5, respectively. The authors write: ``Our work is inspired by [\textit{M. Sommerfeld} and \textit{A. Munk}, J. R. Stat. Soc., Ser. B, Stat. Methodol. 80, No. 1, 219--238 (2018; Zbl 1380.62121)].''