16
10

Hamiltonian Monte Carlo for efficient Gaussian sampling: long and random steps

Abstract

Hamiltonian Monte Carlo (HMC) is a Markov chain algorithm for sampling from a high-dimensional distribution with density ef(x)e^{-f(x)}, given access to the gradient of ff. A particular case of interest is that of a dd-dimensional Gaussian distribution with covariance matrix Σ\Sigma, in which case f(x)=xΣ1xf(x) = x^\top \Sigma^{-1} x. We show that HMC can sample from a distribution that is ε\varepsilon-close in total variation distance using O~(κd1/4log(1/ε))\widetilde{O}(\sqrt{\kappa} d^{1/4} \log(1/\varepsilon)) gradient queries, where κ\kappa is the condition number of Σ\Sigma. Our algorithm uses long and random integration times for the Hamiltonian dynamics. This contrasts with (and was motivated by) recent results that give an Ω~(κd1/2)\widetilde\Omega(\kappa d^{1/2}) query lower bound for HMC with fixed integration times, even for the Gaussian case.

View on arXiv
Comments on this paper