Wasserstein mirror gradient flows and the low temperature limit of the Sinkhorn algorithm

Soumik Pal, University of Washington
-
SMI 405

The Sinkhorn or the Iterated Proportional Fitting Procedure (IPFP) algorithm is a fundamental
iterative algorithm with many uses in statistics and other areas of applications. For example, it
can be used to approximately compute an optimal transport coupling between two probability
measures. However, much of its behavior is still shrouded in mystery. We will talk about limit of
the iterates, as the temperature goes to zero, as an absolutely continuous curve on the
Wasserstein space that has three equivalent descriptions. One, it is a gradient flow of relative
entropy for a modified Wasserstein geometry. Two, it is the family of marginal distributions of a
novel family of diffusions. And, three, the family of measures can be generated by solutions of
the parabolic Monge-Ampere PDE. We will introduce this novel family of flows and related
stochastic processes and PDEs and talk about their properties including their rates to
equilibrium.


Based on joint work with Nabarun Deb, Young-Heon Kim and Geoff Schiebinger

Event Type
Event Subcalendar