Nonstationary Distribution Estimation via Wasserstein Probability Flows

TL;DR

This paper introduces Wasserstein Probability Flows, a novel method for estimating evolving probability distributions from data, effectively capturing nonstationary changes with computational efficiency.

Contribution

The paper proposes a new nonparametric approach that penalizes distribution changes using Wasserstein distance, reducing the problem to a network-flow model for efficient estimation.

Findings

Effective estimation of nonstationary distributions demonstrated.

Method outperforms traditional approaches in numerical tests.

Applicable to nonstationary stochastic optimization tasks.

Abstract

We study the problem of estimating a sequence of evolving probability distributions from historical data, where the underlying distribution changes over time in a nonstationary and nonparametric manner. To capture gradual changes, we introduce a model that penalises large deviations between consecutive distributions using the Wasserstein distance. This leads to a method in which we estimate the underlying series of distributions by maximizing the log-likelihood of the observations with a penalty applied to the sum of the Wasserstein distances between consecutive distributions. We show how this can be reduced to a simple network-flow problem enabling efficient computation. We call this the Wasserstein Probability Flow method. We derive some properties of the optimal solutions and carry out numerical tests in different settings. Our results suggest that the Wasserstein Probability Flow method is a promising tool for applications such as nonstationary stochastic optimization.