Small Sample Limit of the Bennett Acceptance Ratio Method and the Variationally Derived Intermediates

TL;DR

This paper analyzes the performance of Bennett Acceptance Ratio (BAR) and Variationally Derived Intermediates (VI) methods in free energy calculations with very few samples, revealing their limitations and providing guidelines for estimator choice.

Contribution

It demonstrates that BAR and VI are suboptimal with small sample sizes and derives a universal relation between sample number and state overlap, guiding estimator selection.

Findings

VI becomes optimal with fewer than seven samples per state

BAR requires increasing samples as state overlap decreases

Universal inverse power law relates sample number to state overlap

Abstract

Free energy calculations based on atomistic Hamiltonians provide microscopic insight into the thermodynamic driving forces of biophysical or condensed matter systems. Many approaches use intermediate Hamiltonians interpolating between the two states for which the free energy difference is calculated. The Bennett Acceptance Ratio (BAR) and Variationally Derived Intermediates (VI) methods are optimal estimator and intermediate states in that the mean-squared error of free energy calculations based on independent sampling is minimized. However, BAR and VI have been derived based on several approximations that do not hold for very few sample points. Analyzing one-dimensional test systems we show that in such cases BAR and VI are suboptimal and that established uncertainty estimates are inaccurate. Whereas for VI to become optimal less than seven samples per state suffice in all cases, for BAR the required number increases unboundedly with decreasing configuration space densities overlap of the end states. We show that for BAR the required number of samples is related to the overlap through an inverse power law. Because this relation seems to hold universally and almost independent of other system properties, these findings can guide the proper choice of estimator for free energy calculations.