Simple Approximative Algorithms for Free-Support Wasserstein Barycenters

TL;DR

This paper analyzes simple, fast algorithms for approximating Wasserstein barycenters in the free-support setting, providing theoretical error bounds and demonstrating practical efficiency and accuracy in data science applications.

Contribution

It introduces and analyzes a straightforward framework for approximating Wasserstein-$p$ barycenters, especially for $p=1$ and $p=2$, with theoretical guarantees and practical effectiveness.

Findings

Algorithms produce sparse support solutions.

Error bounds are at most a few percent in practice.

Methods are fast, memory-efficient, and easy to implement.

Abstract

Computing Wasserstein barycenters of discrete measures has recently attracted considerable attention due to its wide variety of applications in data science. In general, this problem is NP-hard, calling for practical approximative algorithms. In this paper, we analyze a well-known simple framework for approximating Wasserstein-$p$ barycenters, where we mainly consider the most common case $p=2$ and $p=1$, which is not as well discussed. The framework produces sparse support solutions and shows good numerical results in the free-support setting. Depending on the desired level of accuracy, this requires only $N-1$ or $N(N-1)/2$ standard two-marginal optimal transport (OT) computations between the $N$ input measures, respectively, which is fast, memory-efficient and easy to implement using any OT solver as a black box. What is more, these methods yield a relative error of at most $N$ and $2$, respectively, for both $p=1, 2$. We show that these bounds are practically sharp. In light of the hardness of the problem, it is not surprising that such guarantees cannot be close to optimality in general. Nevertheless, these error bounds usually turn out to be drastically lower for a given particular problem in practice and can be evaluated with almost no computational overhead, in particular without knowledge of the optimal solution. In our numerical experiments, this guaranteed errors of at most a few percent.