Dynamical reweighting for estimation of fluctuation formulas

Raphaël Gastaldello, Gabriel Stoltz, Urbain Vaes · 2025

Read Dynamical reweighting for estimation of fluctuation formulas by Raphaël Gastaldello, Gabriel Stoltz, Urbain Vaes (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-11968.txt.

Introduction

Raphaël Gastaldello, Gabriel Stoltz, and Urbain Vaes reduced statistical noise in molecular simulations in a 2025 paper. Transport coefficients, like how heat or momentum spreads in a fluid, are computed from fluctuation formulas applied to long simulated trajectories, and the estimates wobble a lot, forcing expensive runs. Their method biases the dynamics slightly, using a Girsanov reweighting identity to correct the bias in the final average.

One tuning knob sets the size of the added push, and the authors analyze long-time behavior fully to predict how much variance falls. Theory and numerics agree: variance does drop, though modestly in most situations. Honest accounting of a small win matters in a field where simulation hours are precious.

Catalog record

CREATORRaphaël Gastaldello, Gabriel Stoltz, Urbain Vaes
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE99 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-11968

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Dynamical reweighting for estimation of fluctuation formulas Raphaël Gastaldello, Gabriel Stoltz, Urbain Vaes Abstract We propose a variance reduction method for calculating transport coefficients in molecular dynamics using an importance sampling method via Girsanov’s theorem applied to Green–Kubo’s formula. We optimize the magnitude of the perturbation applied to the reference dynamics by means of a scalar parameter α \alpha and propose an asymptotic analysis to fully characterize the long-time behavior in order to evaluate the possible variance reduction. Theoretical results corroborated by numerical results show that this method allows for some reduction in variance, although rather modest in most situations. keywords Variance reduction, Green–Kubo formula, Girsanov theorem † † runningheads: Dynamical reweighting for estimation of fluctuation formulas / R. Gastaldello, G. Stoltz, and U. Vaes AMS 65C40, 82C31, 82C05, 82C70 1 Introduction Molecular dynamics has been developing rapidly for over 70 years [ 4 ] . It relies on statistical physics, which offers a path to understand the macroscopic behavior of matter from its microscopic description. Despite severe constraints in terms of time and space scales, it has proved to be an extremely powerful tool for carrying out in silico experiments, testing physical theories on a computer and also helping to make up for the absence of analytical theories. Nowadays, numerical simulations are essential for understanding the macroscopic behaviour of matter, particularly when experiments are impossible or costly, such as in high-pressure or high-temperature regimes, or when studying as yet unknown or unsynthesised materials especially for electronic and biological systems on the nanometric scale. Another major objective of molecular simulation is to calculate macroscopic quantities and thermodynamic properties, providing quantitative information about molecular systems. We refer for instance to [ 1 , 18 , 43 ] for a large panel of numerical methods in statistical physics and molecular simulation and [ 25 , 27 ] for a mathematical perspective. One of the aims of molecular dynamics is to understand the dynamical properties of matter. Transport coefficients such as those governing particle diffusion (self-diffusion), heat conduction (thermal conductivity), or momentum transport (shear viscosity), are examples of such properties. They determine the response of a system to an external perturbation; see [ 15 , 10 ] for an general overview and [ 26 , Section 5] for a mathematical perspective on this subject. Once these coefficients have been estimated through molecular simulation, they can be fed to macroscopic evolution equations like the Navier–Stokes equation. A standard method for computing transport coefficients, based on molecular dynamics, utilizes the well-known Green–Kubo formula [ 20 , 24 ] . This formula expresses transport coefficients as integrated time-correlations of relevant observables for the system at equilibrium. Another approach, relying on linear response theory, involves directly simulating non-equilibrium perturbations and measuring the resulting average response, a method known as non-equilibrium molecular dynamics (NEMD). However, estimators derived from these methods often suffer from significant statistical errors, as discussed in [ 42 ] . Achieving convergence often demands simulating long trajectories, incurring substantial computational costs. Various variance reduction techniques can be used for systems at equilibrium, in particular importance sampling [ 8 ] and control variates methods [ 31 , 32 ] . Most variance reduction techniques cannot be used as such for nonequilibrium systems, and therefore require a dedicated treatment, see for instance [ 40 , 41 , 6 , 35 ] . Concerning importance sampling, one has to resort to dynamical versions where a bias is introduced into the dynamics, and corrected for by some exponential weight based on Girsanov’s formula. This is known as “dynamical reweighting” or “Girsanov reweightings” in the theoretical chemistry literature. For example, in [ 9 ] this method is used to compute time-correlation functions, and the references [ 11 , 12 , 13 ] present how to implement the reweighting and use it to simulate proteins, or accelerate the exploration of the phase space. In all cases, the aim is to improve the computation of path ensemble averages. Objectives of this work The present work aims to study and assess the efficiency of the particular method described in [ 11 , 12 ] when used as a variance reduction technique for the calculation of transport coefficients. Mathematically, these coefficients are expressed as integrated (auto)correlations in the Green–Kubo formalism and can therefore be reformulated as ensemble averages with respect to the path measure associated with the stochastic dynamics. In the context of non-degenerate diffusive dynamics, the key idea of the method we study is to add a drift term in order to bias the dynamics: \displaystyle\mathrm{d}q_{t}^{\alpha}=b(q_{t}^{\alpha})\,\mathrm{d}t+\alpha\sigma(q_{t}^{\alpha})u(q_{t}^{\alpha})\,\mathrm{d}t+\sigma(q_{t}^{\alpha})\,\mathrm{d}W_{t}. (1.1) This modified dynamics induces a path measure that is different from that associated with the reference dynamics ( \alpha=0 ). Therefore, for consistency, statistical estimators based on the modified dynamics ( 1.1 ) must include Girsanov weights, which represent the likelihood ratio between the reference and modified path probability measures. In this work, we wish to understand whether employing a perturbed dynamics of the type ( 1.1 ) instead of the reference dynamics can be useful for the calculation of transport coefficients through Green–Kubo’s formula. Our contributions The paper makes the following contributions. • We first consider the following stochastic differential equation (SDE), \displaystyle\mathrm{d}q_{t}=b(q_{t})\,\mathrm{d}t+\sigma(q_{t})\,\mathrm{d}W_{t}. (1.2) In order to motivate this study, we prove for this reference dynamics that the variance of the Green–Kubo estimator is proportional to the integration time T , in the limit as T tends to infinity (see Proposition 2.2 ). • We then consider the dynamics ( 1.1 ) incorporating an additional drift term \alpha\sigma u compared to ( 1.2 ), with \alpha a scalar and u a vector field. After defining a weighted Green–Kubo-type estimator in terms of the solution to this equation, we show that, for any fixed time T , there exists a unique \alpha that minimizes the variance of the weighted Green–Kubo’s estimator (see Proposition 3.3 ). • Finally, we characterize the asymptotic behavior, as the integration time T tends to +\infty , of the optimal value of \alpha and of the amount of variance reduction obtained with this asymptotically optimal value. We emphasize that the main contribution of this work is the asymptotic analysis of the new estimator for the Green–Kubo formula. We however also explore the non-asymptotic regime through numerical simulations. The so-obtained results are quantitatively similar to those predicted by the asymptotic analysis, therefore achieving only a modest variance reduction. The development of numerical methods based on Girsanov’s theorem seems to be more effective for the Markov state models considered in [ 11 ] . Outline of the paper In Section 2 , we present the general mathematical context of the computation of transport coefficients through the Green–Kubo estimator, and study the scaling of the variance of this estimator with respect to the integration time. In Section 3 , we present and study the dynamical reweighting approach. Specifically, we apply Girsanov’s theorem in order to construct a new Green–Kubo-like estimator based on the biased dynamics ( 1.1 ) and incorporating Girsanov weights, depending on the parameter \alpha . We show that there exists a unique value of this parameter which minimizes the estimator variance, and give its scaling with respect to the integration time. In Section 4 we illustrate our main results, in particular those

Read the complete text