Source-linked AI summary
Data-driven parameterization of the generalized Langevin equation
Huan Lei, Nathan Baker, Xiantao Li
TL;DR
Generalized Langevin equations require costly history-dependent memory terms. This work parameterizes the kernel hierarchically and constructs extended stochastic models with exact second fluctuation-dissipation compliance, recovering standard stochastic processes at low orders.
Problem
Practical generalized Langevin equation implementations require specifying memory terms, whose history dependence is costly and whose kernel computation involves an integral equation.
Method
The method hierarchically parameterizes the kernel and introduces random noise so the resulting extended stochastic model satisfies the second fluctuation-dissipation theorem exactly.
Results
Zero- and first-order approximations recover Langevin and Ornstein-Uhlenbeck stochastic processes, respectively, while higher-order approximations are constructed from simulation time-series data.
Takeaways & Limitations
The approach avoids expensive history-dependent memory calculations and supports arbitrarily high-order approximations for generalized Langevin dynamics.
Abstract
from arXiv · showhide
We present a data-driven approach to determine the memory kernel and random noise in generalized Langevin equations. To facilitate practical implementations, we parameterize the kernel function in the Laplace domain by a rational function, with coefficients directly linked to the equilibrium statistics of the coarse-grain variables. We show that such an approximation can be constructed to arbitrarily high order and the resulting generalized Langevin dynamics can be embedded in an extended stochastic model without explicit memory. We demonstrate how to introduce the stochastic noise so that the second fluctuation-dissipation theorem is exactly satisfied. Results from several numerical tests are presented to demonstrate the effectiveness of the proposed method.
I. INTRODUCTION
GLE implementations face costly memory evaluation, difficult kernel estimation, and nontrivial noise generation, motivating a data-driven alternative. The proposed approach parameterizes the Laplace-domain kernel rationally, supports arbitrarily high-order models, and removes explicit memory and colored-noise sampling through an extended stochastic dynamics.
- Motivation: GLE implementations are costly because memory terms require coarse-grained history and numerical integration, while random noise must satisfy the second fluctuation-dissipation theorem.Kernel estimation can also involve high-dimensional calculations, and direct derivations may be numerically unstable.
- Contribution: The paper presents a hierarchical data-driven approach that obtains GLE kernel functions from simulation data for complex systems whose full dynamics may be unavailable.The central approximation parameterizes the Laplace transform of the kernel with a rational function.
- Contribution: The rational-approximation hierarchy can be constructed to arbitrarily high order, whereas its lowest approximations recover Markovian and Ornstein-Uhlenbeck descriptions that may be insufficient for dynamics properties.The approximation parameters are computed from statistical properties of coarse-grained variables.
- Contribution: Auxiliary variables embed the generalized Langevin dynamics in an extended model, eliminating explicit memory terms and enabling inexpensive white-noise forcing that exactly satisfies the second FDT.This avoids evaluating history-dependent memory terms and sampling colored noise.
II. NUMERICAL METHOD
The method estimates the generalized Langevin memory kernel from equilibrium trajectory statistics by fitting its Laplace transform with a rational function. This approximation supports memory-free extended dynamics whose noise satisfies the second fluctuation-dissipation theorem exactly.
- Kernel estimation: The method estimates the memory kernel from equilibrium simulation data by converting the associated integral equation into a Laplace-domain rational approximation.The time-domain integral equation is first-kind and ill-posed, so the kernel is parameterized through its Laplace transform and fitted using limiting values extracted from data.
- Kernel estimation: Matching Laplace-transform limits determines approximation coefficients through a linear system solvable analytically for small orders or numerically for large orders.The construction uses interpolation information including λ = 0 and λ = +∞, with additional interpolation points permitted.
- Approximation hierarchy: The rational approximation can reproduce Markovian damping at zero order and represent memory functions with exponential or oscillatory behavior at higher orders.At zero order, Θ(λ) is constant and yields a Langevin model with damping coefficient γ = θ0; higher-order behavior depends on the eigenvalues of B0.
- Extended dynamics: Auxiliary equations remove the need to evaluate the memory integral at every simulation step, producing an extended stochastic model without explicit memory.The extended formulation is available to arbitrarily high order and uses matrices such as B, Q, and Z to match the rational kernel.
- Extended dynamics: Properly chosen white noise and auxiliary-variable initial conditions generate stationary colored noise that satisfies the second fluctuation-dissipation theorem exactly.The construction also provides an invariant distribution for the auxiliary variables and preserves the fluctuation-dissipation structure in the extended system.
III. NUMERICAL RESULTS
Numerical tests show that rational approximations reproduce memory kernels and coarse-grained dynamics increasingly well as their order rises, especially when non-Markovian effects are pronounced. In a double-well system, third- and fourth-order approximations recover molecular-dynamics transition behavior more accurately than zero-order dynamics.
- Harmonic-chain test: As the approximation order increases to third order, the rational approximation of the harmonic-chain memory kernel agrees well with the exact formulation.The harmonic-chain test uses a tagged particle in a chain of N = 1000 particles with K, m, and β set to unity.
- Particle-bath simulations: For case (II), the memory kernel has a pronounced peak near λ = 0.2, corresponding to significant time-domain oscillations in the kernel.This contrasts with case (I), whose plateau region resembles a Markovian delta-kernel approximation.
- Particle-bath simulations: Case (I) reproduces the system dynamics well from second order onward, while case (II) requires third- and fourth-order approximations because second order introduces artificial oscillations.The differing requirements reflect the absence of apparent time-scale separation between velocity and memory-kernel dynamics in case (II).
- Double-well transition dynamics: Third- and fourth-order rational approximations agree well with full molecular-dynamics transition-flux correlations, whereas zero-order Langevin dynamics deviates.The zero-order model also overestimates the reaction rate; agreement with full molecular dynamics is approached using third- and fourth-order kernels.
IV. SUMMARY
The paper develops a data-driven rational approximation of GLE memory kernels in the Laplace domain, linking coefficients to equilibrium coarse-grained statistics. The approximation supports memory-free extended dynamics, systematic higher-order constructions, and exact satisfaction of the second fluctuation-dissipation theorem.
- The rational-function approximation connects GLE kernel coefficients to equilibrium statistics obtainable from simulation time series.This establishes the method’s data-driven character in the Laplace transform domain.
- Higher-order approximations are systematically derived, while zero- and first-order cases recover Langevin and Ornstein-Uhlenbeck processes.
- Auxiliary variables embed the rationally approximated GLE in linear stochastic dynamics without explicit history-dependent memory terms.This makes numerical simulation more convenient than direct time-domain kernel representations.
- The approach automatically satisfies the second fluctuation-dissipation theorem.
- The method was tested on simple systems and is presented as applicable to biological and material systems with pronounced memory effects.Examples include sub-diffusion in single-molecule measurements and transition dynamics of chemical and biological reactions.
- Rational approximations enable transition-dynamics analysis in an augmented phase space when direct GLE analysis is difficult or inaccessible.