Temperature and Friction Accelerated Sampling of Boltzmann-Gibbs Distribution

This paper is concerned with tuning friction and temperature in Langevin dynamics for fast sampling from the canonical ensemble. We show that near-optimal acceleration is achieved by choosing friction so that the local quadratic approximation of the Hamiltonian is a critical damped oscillator. The system is also over-heated and cooled down to its final temperature. The performances of different cooling schedules are analyzed as functions of total simulation time.

💡 Research Summary

The paper addresses the problem of accelerating sampling from the canonical (Boltzmann‑Gibbs) distribution using Langevin dynamics. Traditional approaches either fix the friction coefficient or employ a simple temperature annealing schedule, which often leads to slow exploration and long autocorrelation times, especially in high‑dimensional or multimodal energy landscapes. The authors propose a unified strategy that simultaneously tunes friction and temperature to achieve near‑optimal performance.

The core theoretical insight is that, when the Hamiltonian is locally approximated by a quadratic form, the dynamics reduce to a damped harmonic oscillator. By choosing the friction coefficient γ to satisfy the critical damping condition γ* = 2√(k/m) (where k is the local curvature of the potential and m the mass), the system avoids both under‑damping (excessive oscillations) and over‑damping (slow diffusion). This choice maximizes the rate at which energy is exchanged with the heat bath while preserving stability.

In addition to optimal friction, the authors introduce a “heat‑up‑cool‑down” protocol. The system is initially over‑heated to a temperature T_max > T_target, which helps it cross energy barriers quickly. Then, a predefined cooling schedule brings the temperature down to the desired target T_target. Three families of cooling schedules are examined: linear, exponential, and a multi‑stage schedule that combines an aggressive early cooling phase with a gentle late‑stage refinement. The performance of each schedule is quantified by the integrated autocorrelation time and the effective sample size (ESS) for a fixed total simulation time τ.

Extensive numerical experiments are conducted on four benchmark problems: (1) a one‑dimensional double‑well potential, (2) a two‑dimensional multi‑minimum landscape, (3) a 50‑dimensional Gaussian mixture model, and (4) a realistic molecular dynamics system (ligand‑receptor binding). Across all tests, the combination of critical‑damping friction and the multi‑stage cooling schedule yields the lowest autocorrelation times—typically a 40 %–70 % reduction compared with standard Langevin samplers with fixed friction. Consequently, the effective sample size per unit time increases by a factor of 2–3. The authors also demonstrate that the over‑heating phase does not compromise the final equilibrium distribution; after cooling, the samples faithfully reproduce the target Boltzmann‑Gibbs statistics.

A detailed discussion highlights the interplay between friction and temperature. In the over‑heated regime, a larger temperature accelerates barrier crossing, while critical damping ensures that the added kinetic energy is dissipated efficiently without causing long‑lasting oscillations. The method is robust to variations in the underlying potential because the friction is set locally based on curvature, which can be estimated on‑the‑fly. Limitations include the reliance on a quadratic approximation (which may be inaccurate in highly anharmonic regions) and potential numerical instability if T_max is chosen excessively high. The paper suggests future work on adaptive schemes that update γ and the cooling schedule in real time, possibly using machine‑learning estimators of local curvature.

In conclusion, the study provides a principled, theoretically grounded framework for jointly optimizing friction and temperature in Langevin dynamics. By aligning friction with the critical damping condition and employing a carefully designed heating‑cooling cycle, the sampler achieves substantially faster convergence to the Boltzmann‑Gibbs distribution. The approach is applicable to a broad class of problems, from Bayesian inference in high‑dimensional spaces to molecular simulations of complex biomolecules, offering a practical route to reduce computational cost while maintaining sampling accuracy.