Abstract

Contraction in Wasserstein 1-distance with explicit rates is established forgeneralized Hamiltonian Monte Carlo with stochastic gradients under possiblynonconvex conditions. The algorithms considered include splitting schemes ofkinetic Langevin diffusion commonly used in molecular dynamics simulations. Toaccommodate the degenerate noise structure corresponding to inertia existing inthe chain, a characteristically discrete-in-time coupling and contraction proofis devised. As consequence, quantitative Gaussian concentration bounds areprovided for empirical averages. Convergence in Wasserstein 2-distance andtotal variation are also given, together with numerical bias estimates.

 

Quick Read (beta)