
Wasserstein barycenters in the manifold of all positive definite matrices
Author(s) -
Elham Nobari,
Bijan Ahmadi Kakavandi
Publication year - 2019
Publication title -
quarterly of applied mathematics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.603
H-Index - 41
eISSN - 1552-4485
pISSN - 0033-569X
DOI - 10.1090/qam/1535
Subject(s) - algorithm , annotation , artificial intelligence , computer science , mathematics
In this paper, we study the Wasserstein barycenter of finitely many Borel probability measures on P n \mathbb {P}_{n} , the Riemannian manifold of all n × n n\times n real positive definite matrices as well as its associated dual problem, namely the optimal transport problem. Our results generalize some results of Agueh and Carlier on R n \mathbb {R}^{n} to P n \mathbb {P}_{n} . We show the existence of the optimal solutions and the Wasserstein barycenter measure. Furthermore, via a discretization approach and using the BFGS (Broyden-Fletcher-Goldfarb-Shanno) method for nonsmooth convex optimization, we propose a numerical method for computing the potential functions of the optimal transport problem. Also, thanks to the so-called optimal transport Jacobian on Riemannian manifolds of Cordero-Erausquin, McCann, and Schmuckenschläger, we show that the density of the Wasserstein barycenter measure can be approximated numerically. The paper concludes with some numerical experiments.