Multistate Bennett Acceptance Ratio

MBAR is the minimum-variance estimator for free energy differences given the sampled data. Unlike older methods that solve pairwise equations sequentially, MBAR constructs a global likelihood function incorporating all configurations from every state simultaneously. This ensures no information is discarded and the statistical error is minimized. The method asymptotically achieves the Cramér-Rao lower bound, making it the most efficient unbiased estimator available for alchemical free energy calculations.

MBAR solves a set of coupled nonlinear equations to determine the dimensionless free energies of all states:

• Each state's free energy is expressed as a function of the reduced potential energies of all sampled configurations
• The equations are solved iteratively until self-consistency is achieved
• The algorithm naturally weights configurations from states where they are most probable
• Convergence is typically rapid, requiring only a few iterations with good overlap between adjacent states

MBAR provides a unified framework for reweighting expectations to unsampled states. Once the free energies are determined, any equilibrium observable at any thermodynamic state can be estimated by reweighting the existing samples. This enables:

• Calculation of potential energy distributions at unvisited temperatures
• Estimation of structural properties at intermediate alchemical states
• Direct computation of free energy differences between any pair of states, not just adjacent ones
• Seamless integration with thermodynamic integration and perturbation theory

MBAR naturally quantifies the reliability of free energy estimates through overlap matrices. The method computes the statistical uncertainty by analyzing the effective sample size contributed by each state. Key diagnostics include:

• The Kish effective sample size for each state's contribution
• Pairwise overlap metrics that identify poorly connected states
• Asymptotic covariance estimates for all free energy differences
• These diagnostics guide simulation design, indicating where additional sampling or intermediate states are needed to ensure robust convergence

MBAR generalizes and improves upon earlier free energy estimators:

• BAR (Bennett Acceptance Ratio): The optimal two-state estimator; MBAR extends this to an arbitrary number of states
• WHAM (Weighted Histogram Analysis Method): Requires binning data along a reaction coordinate; MBAR operates directly on energies without discretization error
• MBAR reduces to BAR for two states and to WHAM when data is binned, but provides lower variance in the general case
• The method is the recommended standard for modern alchemical free energy calculations in packages like pymbar and alchemlyb

Effective use of MBAR requires careful simulation design:

• Intermediate states must be spaced to ensure sufficient phase space overlap between neighbors
• Reduced potential energies for every sampled configuration must be evaluated at all states, not just the state where it was sampled
• The computational cost scales as O(K²N) where K is the number of states and N is the number of samples
• Modern GPU-accelerated implementations handle thousands of states efficiently
• The pymbar Python library provides a robust, well-tested reference implementation used in production free energy workflows