Improved walker population control for full configuration interaction quantum Monte Carlo

TL;DR

This paper introduces an improved walker population control method for FCIQMC that ensures stable, rapid equilibration at a set walker number, reducing memory use and maintaining energy estimation accuracy.

Contribution

It presents a novel population control protocol modeled as a damped harmonic oscillator, enabling precise walker number regulation and detection of sign problem overcoming plateaus.

Findings

Achieves stable walker population at a pre-defined level

Demonstrates rapid equilibration with minimal code changes

Maintains accuracy of ground state energy estimates

Abstract

Full configuration interaction quantum Monte Carlo (FCIQMC) is a stochastic approach for finding the ground state of a quantum many-body Hamiltonian. It is based on the dynamical evolution of a walker population in Hilbert space, which samples the ground state configuration vector over many iterations. Here we present a modification of the original protocol for walker population control of Booth et al. JCP 131, 054106 (2009) in order to achieve equilibration at a pre-defined average walker number and to avoid walker number overshoots. The dynamics of the walker population is described by a noisy damped harmonic oscillator and controlled by two parameters responsible for damping and forcing, respectively, for which reasonable values are suggested. We further introduce a population growth witness that can be used to detect annihilation plateaus related to overcoming the FCIQMC sign problem. Features of the new population control procedure such as precise walker number control and fast equilibration are demonstrated. The standard error of the shift estimator for the ground state energy as well as the population control bias are found to be unaffected by the population control procedure or its parameters. The improved control of the walker number, and thereby memory consumption, is a desirable feature required for automating FCIQMC calculations and requires minimal modifications to existing code.