Free energy Sequential Monte Carlo, application to mixture modelling

TL;DR

This paper introduces free energy Sequential Monte Carlo methods inspired by physics, enabling better sampling of multimodal mixture posteriors by biasing distributions to improve mode exploration and recovery.

Contribution

The paper proposes a novel free energy SMC framework that biases distributions to facilitate mode switching in complex mixture models.

Findings

Effective mode swapping in multimodal mixtures

Successful application to Gaussian mixtures and real datasets

Improved sampling of symmetric and multimodal posteriors

Abstract

We introduce a new class of Sequential Monte Carlo (SMC) methods, which we call free energy SMC. This class is inspired by free energy methods, which originate from Physics, and where one samples from a biased distribution such that a given function $\xi(\theta)$ of the state $\theta$ is forced to be uniformly distributed over a given interval. From an initial sequence of distributions $(\pi_t)$ of interest, and a particular choice of $\xi(\theta)$, a free energy SMC sampler computes sequentially a sequence of biased distributions $(\tilde{\pi}_{t})$ with the following properties: (a) the marginal distribution of $\xi(\theta)$ with respect to $\tilde{\pi}_{t}$ is approximatively uniform over a specified interval, and (b) $\tilde{\pi}_{t}$ and $\pi_{t}$ have the same conditional distribution with respect to $\xi$. We apply our methodology to mixture posterior distributions, which are highly multimodal. In the mixture context, forcing certain hyper-parameters to higher values greatly faciliates mode swapping, and makes it possible to recover a symetric output. We illustrate our approach with univariate and bivariate Gaussian mixtures and two real-world datasets.