Improved Langevin Monte Carlo for stochastic optimization via landscape modification

Given a target function to minimize or a target Gibbs distribution to sample from in the low temperature, in this paper we propose and analyze Langevin Monte Carlo (LMC) algorithms that run on an alternative landscape as specified by and target a modified Gibbs distribution , where the landscape of is a transformed version of that of which depends on the parameters and . While the original Log-Sobolev constant affiliated with exhibits exponential dependence on both and the energy barrier in the low temperature regime, with appropriate tuning of these parameters and subject to assumptions on , we prove that the energy barrier of the transformed landscape is reduced which consequently leads to polynomial dependence on both and in the modified Log-Sobolev constant associated with . This yield improved total variation mixing time bounds and improved convergence toward a global minimum of . We stress that the technique developed in this paper is not only limited to LMC and is broadly applicable to other gradient-based optimization or sampling algorithms.
View on arXiv