Identifying and estimating bias in overlap-sampling free-energy calculations

We examine several methods for detecting or quantifying bias in free energy calculations performed using Bennett's acceptance ratio (BAR) method for combining data from molecular simulations of two systems. The metrics for bias detection studied here are one based on the emergence, at the BAR estimate, of a zero-slope inflection of the free energy with respect to its optimisation parameter, and another that makes use of calculation of the dissipated work. Bias estimation methods studied here are two approaches based on consideration of the maximum possible contribution from a configuration that may be rarely sampled, and another that appeared recently in the literature that is derived explicitly using the neglected-tail ansatz. All approaches are tested as a function of sampling amount in application to a model multiharmonic system. The dissipated-work method provides the more reliable indicator of the presence of bias, and the maximal sample techniques yield the tightest bounds on the bias, among the methods tested.