Graph-theoretic perspectives on splitting methods for sparse optimal transport

Jacob Lindbäck, Mikael Johansson

Keywords: optimization, optimal transport, splitting methods, local guarantees

TL;DR: We relate the local behavior of a sparse OT solver with graph theoretic properties of the solution

Abstract:

We study the local behavior of splitting methods for sparse optimal transport. By leveraging
finite-time identification properties, we relate the algorithms’ local convergence behavior to graph-
theoretic properties of the solution. This connection offers insights into suitable stepsize choices,
which we use to design a simple stepsize heuristic. We demonstrate the efficiency and robustness
of the heuristic on a range of experiments.

Submission Number: 65