Cost of the Generalised Hybrid Monte Carlo Algorithm for Free Field Theory

TL;DR

This paper analytically evaluates the computational cost of the Generalised Hybrid Monte Carlo algorithm for free field theory, optimizing parameters and comparing efficiency with related algorithms.

Contribution

It provides a detailed cost analysis of GHMC, including optimal parameters and comparison with HMC and L2MC algorithms.

Findings

Long trajectories are optimal for GHMC.

HMC is more efficient than L2MC in certain cases.

HMC and L2MC have the same volume dependence but different dynamical critical exponents.

Abstract

We study analytically the computational cost of the Generalised Hybrid Monte Carlo (GHMC) algorithm for free field theory. We calculate the Metropolis acceptance probability for leapfrog and higher-order discretisations of the Molecular Dynamics (MD) equations of motion. We show how to calculate autocorrelation functions of arbitrary polynomial operators, and use these to optimise the GHMC momentum mixing angle, the trajectory length, and the integration stepsize for the special cases of linear and quadratic operators. We show that long trajectories are optimal for GHMC, and that standard HMC is more efficient than algorithms based on Second Order Langevin Monte Carlo (L2MC), sometimes known as Kramers Equation. We show that contrary to naive expectations HMC and L2MC have the same volume dependence, but their dynamical critical exponents are z = 1 and z = 3/2 respectively.