Generalized Wasserstein Barycenters

Abstract: We propose a generalization, where negative weights are allowed, of the Wasserstein barycenter of $n$ probability measures. The barycenter is found, as usual, as a minimum of a functional. In this paper, we prove existence of a minimizer for probability measures on a separable Hilbert space and uniqueness in the case of one positive coefficient and $n-1$ negative ones. In the one-dimensional case, we characterize the quantile function of the unique minimum as the orthogonal projection of the $L^2$-barycenter of the quantiles on the cone of nonincreasing functions in $L^2(0,1)$. Further, we provide a stability estimate in dimension one and a counterexample to uniqueness in $\mathbb{R}^2$.