English

Deformations of Boltzmann Distributions

High Energy Physics - Lattice 2022-11-16 v3 Statistical Mechanics Machine Learning

Abstract

Consider a one-parameter family of Boltzmann distributions pt(x)=1ZteSt(x)p_t(x) = \tfrac{1}{Z_t}e^{-S_t(x)}. This work studies the problem of sampling from pt0p_{t_0} by first sampling from pt1p_{t_1} and then applying a transformation Ψt1t0\Psi_{t_1}^{t_0} so that the transformed samples follow pt0p_{t_0}. We derive an equation relating Ψ\Psi and the corresponding family of unnormalized log-likelihoods StS_t. The utility of this idea is demonstrated on the ϕ4\phi^4 lattice field theory by extending its defining action S0S_0 to a family of actions StS_t and finding a τ\tau such that normalizing flows perform better at learning the Boltzmann distribution pτp_\tau than at learning p0p_0.

Keywords

Cite

@article{arxiv.2210.13772,
  title  = {Deformations of Boltzmann Distributions},
  author = {Bálint Máté and François Fleuret},
  journal= {arXiv preprint arXiv:2210.13772},
  year   = {2022}
}

Comments

Machine Learning for the Physical Sciences Workshop at NeurIPS '22

R2 v1 2026-06-28T04:26:04.340Z