Free Energy Sequential Monte Carlo Application to Mixture Modelling

We introduce a new class of Sequential Monte Carlo (SMC) methods, whichwe call free energy SMC. This class is inspired by free energy methods, whichoriginate from Physics, and where one samples from a biased distribution suchthat a given function !(") of the state " is forced to be uniformly distributedover a given interval. From an initial sequence of distributions (#t) of interest,and a particular choice of !("), a free energy SMC sampler computes sequentiallya sequence of biased distributions (˜#t) with the following properties: (a)the marginal distribution of !(") with respect to ˜#t is approximatively uniformover a specified interval, and (b) ˜#t and #t have the same conditional distributionwith respect to !. We apply our methodology to mixture posteriordistributions, which are highly multimodal. In the mixture context, forcingcertain hyper-parameters to higher values greatly faciliates mode swapping,and makes it possible to recover a symetric output. We illustrate our approachwith univariate and bivariate Gaussian mixtures and two real-world datasets.