Equilibrium Free Energies from Nonequilibrium Measurements Using Maximum-Likelihood Methods

Michael R. Shirts, Eric Bair, Giles Hooker, Vijay S. PandePhysical 2003 · DOI

TL;DR

This paper reinterprets the widely used Bennett acceptance ratio method for calculating free energy differences as a maximum likelihood estimator. This provides a rigorous statistical foundation, demonstrating it's the optimal unbiased estimator for free energy calculations from nonequilibrium work measurements, and offers a new way to estimate its variance.

What Problem Does This Paper Address?

Calculating free energy differences between different states of a physical system is crucial across many scientific disciplines, from drug design to quantum chromodynamics. A common approach involves performing nonequilibrium simulations and analyzing the "work" done during the transition. However, extracting accurate free energy estimates from these work distributions can be challenging due to statistical noise and potential biases, especially with limited data. While methods like the Bennett acceptance ratio (BAR) are known to be effective, their statistical underpinnings and optimal usage weren't fully elucidated. This paper aims to provide a rigorous statistical derivation for BAR, establishing its optimality and offering a clear framework for understanding its variance.

Methods

The authors reframe the Bennett acceptance ratio method through the lens of maximum likelihood estimation (MLE). They leverage the connection between the probability distributions of forward and reverse work measurements and logistic regression, a standard statistical technique. By treating the free energy difference as a parameter to be estimated, they derive the BAR formula by maximizing the likelihood function of the observed work data. This MLE approach allows them to rigorously prove that BAR is an asymptotically unbiased estimator with the minimum possible variance. Additionally, they derive a formula for the variance of the MLE, which is directly applicable to the BAR method.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and statistical properties of the method, rather than its application to specific simulation datasets.

Key Insights and Ideas

  • BAR as the "Best" Estimator: The paper rigorously demonstrates that the Bennett acceptance ratio method is the maximum likelihood estimator for free energy differences from nonequilibrium work data. This means, under certain conditions, it's the statistically optimal choice for an unbiased estimator, offering the lowest possible variance.
  • Logistic Regression Connection: Framing the problem as logistic regression provides a powerful statistical analogy, allowing the authors to borrow well-established theoretical results from that field to understand the properties of BAR.
  • New Variance Formula: A novel, simple formula for the variance of free energy estimates obtained via BAR is derived. This is crucial for researchers to accurately quantify the uncertainty in their free energy calculations.
  • Efficiency of Bidirectional Sampling: The paper reinforces the idea that sampling work in both forward and reverse directions is significantly more efficient than using only one direction, especially when the number of samples in each direction is balanced.

Key Results

  • The Bennett acceptance ratio method is mathematically shown to be the maximum likelihood estimator for free energy differences derived from nonequilibrium work measurements.
  • This MLE property implies that BAR is the minimum variance asymptotically unbiased estimator.
  • A new, explicit formula for the variance of the free energy estimate obtained using BAR is derived.
  • The paper confirms that BAR is generally superior to simpler methods like exponential averaging, particularly for smaller datasets or when sample distributions are uneven.

Limitations and Future Work

The authors acknowledge that the MLE is "asymptotically unbiased," meaning it becomes unbiased as the number of measurements approaches infinity. For finite datasets, there can still be a small bias, though it's typically proportional to $n^{-1}$ and thus less significant than the statistical uncertainty ($n^{-1/2}$). They also mention that while MLE is optimal in terms of variance, other estimators (potentially asymptotically biased) might offer comparable or even better mean squared errors for finite samples, suggesting avenues for future research such as smoothing probability distributions. The paper also notes that while the derivation is general for work distributions, practical applications often involve specific energy differences.

Summary updated

Abstract

We present a maximum likelihood argument for the Bennett acceptance ratio method, and derive a simple formula for the variance of free energy estimates generated using this method. This derivation of the acceptance ratio method, using a form of logistic regression, a common statistical technique, allows us to shed additional light on the underlying physical and statistical properties of the method. For example, we demonstrate that the acceptance ratio method yields the lowest variance for any estimator of the free energy which is unbiased in the limit of large numbers of measurements.
BibTeX
@inproceedings{shirts2003equilibrium,
  author = {Michael R. Shirts and Eric Bair and Giles Hooker and Vijay S. Pande},
  title = {{Equilibrium Free Energies from Nonequilibrium Measurements Using Maximum-Likelihood Methods}},
  booktitle = {Physical},
  year = {2003},
  doi = {10.1103/physrevlett.91.140601},
}

Loading related content...

Figure