31
5

Weak convergence of empirical Wasserstein type distances

Abstract

We estimate contrasts 01ρ(F1(u)G1(u))du\int_0 ^1 \rho(F^{-1}(u)-G^{-1}(u))du between two continuous distributions FF and GG on R\mathbb R such that the set {F=G}\{F=G\} is a finite union of intervals, possibly empty or R\mathbb{R}. The non-negative convex cost function ρ\rho is not necessarily symmetric and the sample may come from any joint distribution HH on R2\mathbb{R}^2 with marginals FF and GG having light enough tails with respect to ρ\rho. The rates of weak convergence and the limiting distributions are derived in a wide class of situations including the classical Wasserstein distances W1W_1 and W2W_2. The new phenomenon we describe in the case F=GF=G involves the behavior of ρ\rho near 00, which we assume to be regularly varying with index ranging from 11 to 22 and to satisfy a key relation with the behavior of ρ\rho near \infty through the common tails. Rates are then also regularly varying with powers ranging from 1/21/2 to 11 also affecting the limiting distribution, in addition to HH.

View on arXiv
Comments on this paper