22
2

On the Wasserstein median of probability measures

Abstract

Measures of central tendency such as mean and median are a primary way to summarize a given collection of random objects. In the field of optimal transport, the Wasserstein barycenter corresponds to Fr\'{e}chet or geometric mean of a set of probability measures, which is defined as a minimizer of the sum of squared distances to each element in a given set when the order is 2. We present the Wasserstein median, an equivalent of Fr\'{e}chet median under the 2-Wasserstein metric, as a robust alternative to the Wasserstein barycenter. We first establish existence and consistency of the Wasserstein median. We also propose a generic algorithm that makes use of any established routine for the Wasserstein barycenter in an iterative manner and prove its convergence. Our proposal is validated with simulated and real data examples when the objects of interest are univariate distributions, centered Gaussian distributions, and discrete measures on regular grids.

View on arXiv
Comments on this paper