Quantum Monte Carlo with very large multideterminant wavefunctions

TL;DR

This paper introduces an efficient algorithm for computing derivatives of large multideterminant wavefunctions in quantum Monte Carlo, significantly reducing computational costs and enabling larger expansions for accurate atomic energy calculations.

Contribution

The paper presents a novel, scalable algorithm for handling very large multideterminant wavefunctions in quantum Monte Carlo, with improved efficiency and a new truncation scheme.

Findings

Algorithm scales with the number of unique spin-specific determinants.

Feasible calculations with about 750,000 determinants.

Achieved a 400-fold increase in computational cost over single-determinant methods.

Abstract

An algorithm to compute efficiently the first two derivatives of (very) large multideterminant wavefunctions for quantum Monte Carlo calculations is presented. The calculation of determinants and their derivatives is performed using the Sherman-Morrison formula for updating the inverse Slater matrix. An improved implementation based on the reduction of the number of column substitutions and on a very efficient implementation of the calculation of the scalar products involved is presented. It is emphasized that multideterminant expansions contain in general a large number of identical spin-specific determinants: for typical configuration interaction-type wavefunctions the number of unique spin-specific determinants $N_{\rm det}^\sigma$ ($\sigma=\uparrow,\downarrow$) with a non-negligible weight in the expansion is of order ${\cal O}(\sqrt{N_{\rm det}})$. We show that a careful implementation of the calculation of the $N_{\rm det}$-dependent contributions can make this step negligible enough so that in practice the algorithm scales as the total number of unique spin-specific determinants, $\; N_{\rm det}^\uparrow + N_{\rm det}^\downarrow$, over a wide range of total number of determinants (here, $N_{\rm det}$ up to about one million), thus greatly reducing the total computational cost. Finally, a new truncation scheme for the multideterminant expansion is proposed so that larger expansions can be considered without increasing the computational time. The algorithm is illustrated with all-electron Fixed-Node Diffusion Monte Carlo calculations of the total energy of the chlorine atom. Calculations using a trial wavefunction including about 750 000 determinants with a computational increase of $\sim$ 400 compared to a single-determinant calculation are shown to be feasible.