Thermodynamic integration methods, infinite swapping, and the calculation of generalized averages

J. D. Doll, Paul Dupuis, Pierre NyquistThe Journal of Chemical Physics 2017 · DOI · arXiv

Abstract

In the present paper we examine the risk-sensitive and sampling issues associated with the problem of calculating generalized averages. By combining thermodynamic integration, stationary phase Monte Carlo, and infinite swapping techniques, we develop an approach for such problems and explore its utility for a prototypical class of applications.

Introduction

Calculating generalized averages, such as moment generating functions of the form $\langle e^{bx} \rangle$, is a fundamental task in quantum statistical mechanics and chemical physics. However, these calculations are frequently stymied by two significant numerical hurdles: risk-sensitivity and the sign problem. In risk-sensitive regimes, where the parameter $b$ is large, the average is dominated by the rare events in the tails of the probability distribution. Standard Monte Carlo (MC) methods suffer from statistical errors that grow exponentially with $b$, rendering them computationally intractable. Furthermore, in quantum systems, the presence of complex-valued weights leads to rapid phase oscillations where positive and negative regions nearly cancel, burying the signal in overwhelming numerical noise.

This paper introduces a robust computational framework designed to transform these exponential scaling errors into manageable polynomial ones. By integrating Kirkwood-style thermodynamic integration, Stationary Phase Monte Carlo (SPMC), and Infinite Swapping (INS), we provide a method to accurately sample isolated stationary phase regions even in highly oscillatory environments.

Methodology

The proposed approach rests on three synergistic elements that stabilize the calculation of generalized averages:

1. Kirkwood-Style Thermodynamic Integration

Rather than attempting to calculate the average $\langle e^{bx} \rangle$ directly, the authors employ a Kirkwood-style identity to recast the problem as the exponential of an integral: $$\langle e^{bx} \rangle = \exp\left(\int_0^b \langle x \rangle_\lambda d\lambda\right)$$ This reformulation is the cornerstone of the method's stability. While the error in direct sampling grows exponentially with the parameter $b$, the error in this thermodynamic integration approach scales only as $\sqrt{b/N}$. This shift in the fundamental "error physics" allows for precise calculations in regimes previously considered unreachable.

2. Stationary Phase Monte Carlo (SPMC)

To address the sign problem caused by oscillatory integrands, the framework utilizes SPMC. This technique involves "pre-averaging" or smoothing the integrand over a small length scale, denoted as $\epsilon$. This smoothing acts as a high-frequency filter, removing oscillations that contribute zero to the final integral but otherwise generate immense statistical noise. By focusing on the stationary phase regions, the calculation becomes significantly more efficient.

3. Infinite Swapping (INS) across Smoothing Scales

Even when stationary phase regions are identified, they are often spatially isolated, causing standard random walk samplers to become trapped in local modes. To ensure ergodic sampling, the authors implement Infinite Swapping—a high-performance variant of parallel tempering. In a novel departure from traditional temperature-based swapping, this method swaps between different SPMC smoothing scales ($\epsilon$). The sampler uses "blurred" versions of the distribution (large $\epsilon$) to move freely between isolated peaks, which are then resolved in the "sharp" version (small $\epsilon$).

Experimental Validation: The Gauss/Airy Model

To evaluate the methodology, the authors utilized a complex-valued "Gauss/Airy" model system. This model combines a quadratic (Gaussian) term with a cubic (Airy-like) term. While simple enough to allow for an analytical ground-truth solution, the model successfully replicates the severe phase oscillations and sparse sampling challenges inherent in real-world quantum dynamics.

Key Insights

  • Formulation over Brute Force: The research emphasizes that mathematical formulation is more critical than raw sample size. No amount of "brute-force" sampling can overcome the exponential error growth of direct Monte Carlo; a change in the integral's structure is required.
  • Entropy as a Diagnostic: The authors demonstrate that plotting the information entropy of the sampling density against the smoothing parameter $\epsilon$ serves as a diagnostic map. Peaks and valleys in entropy reveal where the distribution bifurcates into stationary phase regions, guiding the selection of the simulation ensemble.

Results and Discussion

The combined INS/SPMC framework demonstrated remarkable precision in the "hard" limit of the Gauss/Airy model. In regimes where direct Monte Carlo was completely overwhelmed by noise—failing by orders of magnitude—the proposed method tracked the exact analytical solution perfectly.

Furthermore, the efficiency of the method was notable; isolated stationary phase regions were resolved using approximately $10^6$ Monte Carlo points. This is a relatively modest computational cost given the complexity of the oscillatory, complex-valued problem being solved.

Conclusion and Future Work

This paper provides a powerful alternative to direct sampling for risk-sensitive and oscillatory problems. While the current demonstrations are one-dimensional, the authors suggest that the framework is a prime candidate for calculating equilibrium time correlation functions in higher-dimensional molecular systems. Future research will focus on addressing the "curse of dimensionality" in the Kirkwood integration grid and automating the selection of the optimal range for the smoothing parameter $\epsilon$.

Summary updated

Abstract

In the present paper we examine the risk-sensitive and sampling issues associated with the problem of calculating generalized averages. By combining thermodynamic integration, stationary phase Monte Carlo, and infinite swapping techniques, we develop an approach for such problems and explore its utility for a prototypical class of applications.
BibTeX
@article{doll2017thermodynamic,
  author = {J. D. Doll and Paul Dupuis and Pierre Nyquist},
  title = {{Thermodynamic integration methods, infinite swapping, and the calculation of generalized averages}},
  journal = {The Journal of Chemical Physics},
  year = {2017},
  doi = {10.1063/1.4979493},
}
Figure