Coarse-Grained Dynamics with Spatial Disorder and Non-Markovian Memory
Chuyi Liu, Yifeng Guan, Jingyuan Li, Mao Su · 2026
Read Coarse-Grained Dynamics with Spatial Disorder and Non-Markovian Memory by Chuyi Liu, Yifeng Guan, Jingyuan Li, Mao Su (2026) 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-2604-16825.txt.
Introduction
Liu and colleagues open the spatial messy-wide Langevin equation, a data-driven method for coarse-grained dynamics in heterogeneous systems. Unlike mean-field ways, it uses a variational Bayes framework with a random field prior to split static spatial disorder from stretchy friction. Number results show standard rules fail where the new method extends long-time motion from short paths. It captures the odd diffusion switch and regains group statistical traits of messy systems.
Catalog record
| CREATOR | Chuyi Liu, Yifeng Guan, Jingyuan Li, Mao Su |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 21748 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.16825
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Coarse-Grained Dynamics with Spatial Disorder and Non-Markovian Memory Chuyi Liu, Yifeng Guan, Jingyuan Li, Mao Su Abstract We introduce the spatial disorder-generalized Langevin equation (SD-GLE), a data-driven method for constructing coarse-grained (CG) dynamics in heterogeneous systems. Unlike conventional CG approaches that rely on a mean-field potential, SD-GLE utilizes a variational Bayesian framework with a random field prior to explicitly disentangle static spatial disorder from viscoelastic friction. Numerical results demonstrate the limits of standard GLEs, whereas SD-GLE accurately extrapolates long-time dynamics to capture the anomalous diffusion crossover from short trajectories and recover the ensemble statistical properties inherent to the disordered nature of these systems. Accurately predicting the long-term dynamics of particles in complex heterogeneous systems—such as living cells 45 ; 35 ; 40 , cross-linked polymer networks 12 ; 29 ; 47 , and glass-forming liquids 21 ; 2 —remains a central challenge in non-equilibrium statistical mechanics. In these systems, anomalous transport typically emerges from the coupling of two distinct physical mechanisms 48 ; 31 ; 15 : the viscoelastic response of the environment, which results in a non-Markovian friction, and static spatial disorder, which creates a rugged potential leading to localized trapping. Traditional phenomenological models, such as fractional Brownian motion (FBM) or continuous-time random walks (CTRW), generally assume a single underlying mechanism and thus struggle to quantitatively reproduce the observed dynamics 32 ; 30 . Data-driven methods have been introduced to accurately reconstruct the coupled dynamics directly from observations. Although existing studies have addressed the distinction between these two mechanisms of anomalous diffusion from observations 43 ; 33 ; 41 ; 34 ; 24 ; 5 , these efforts have primarily focused on a qualitative identification of their physical origins. Recent data-driven frameworks based on the generalized Langevin equation (GLE) 42 ; 25 ; 11 ; 46 ; 27 ; 49 ; 20 ; 14 ; 28 offer a systematic approach, enabling the joint inference of the effective potential energy surface and the viscoelastic memory kernel. However, in highly heterogeneous environments, these approaches encounter a fundamental limitation. Rugged potential energy surface induces long-lived trapping and weak ergodicity breaking 26 ; 19 , such that individual trajectories probe only local environments, leading to substantial inter-trajectory dispersion over finite-time ensembles 1 . Standard GLE methods rely on a global mean-field approximation of the effective potential, which averages out local spatial disorder. As a result, unresolved spatial heterogeneity is absorbed into the inferred memory effect, leading to an overestimation of temporal correlations and compromising the extrapolation of long-time transport properties 22 . Thus, developing a framework that quantitatively disentangles the respective contributions of spatial disorder and temporal memory from limited trajectory data remains an unresolved problem. In this letter, we introduce the spatial disorder-generalized Langevin equation (SD-GLE) framework, which employs a non-parametric variational Bayesian inference approach to explicitly decouple static spatial correlations from viscoelastic memory. The heterogeneous potential energy surface is modeled as a Gaussian random field. By maximizing the marginal likelihood over all trajectories, SD-GLE leverages single-trajectory variational posteriors to adaptively capture local spatial fluctuations, while simultaneously extracting the viscoelastic friction kernel and the spatial correlation function of the potential on a global scale. Through validation in both ideal systems and simulated glass-forming liquids, we demonstrate that the proposed model accurately reproduces the disordered characteristics of the system. Furthermore, by incorporating the underlying spatial heterogeneity, simulations based on this framework correctly extrapolate long-time anomalous diffusion and naturally recover the essential features of disordered transport, including transient non-Gaussian displacement distributions and the massive inter-trajectory dispersion induced by weak ergodicity breaking. We consider the dynamics of a low-dimensional coarse-grained (CG) coordinate Q(t) with conjugate momentum P(t) , resulting from the projection of a many-body system onto a reduced subspace. Within the Mori-Zwanzig formalism 18 , the effective dynamics takes the form: \displaystyle\dot{Q}(t)=M^{-1}P(t), (1) \displaystyle\dot{P}(t)=-\nabla U(Q(t))-\int_{0}^{t}K(t-s)V(s)\,ds+R(t). where M is the mass matrix, V=M^{-1}P is the velocity, and K(t) and R(t) denote the viscoelastic memory kernel and the fluctuating force respectively. To model the spatial disorder, we employ a non-parametric Bayesian approach, treating the local effective potential U(Q) as a realization of a Gaussian random field prior. This choice is motivated by the central limit theorem 6 for unresolved degrees of freedom, which universally governs the emergence of a Gaussian energy disorder in systems ranging from supercooled liquids 3 ; 17 to regulatory proteins on DNA tracks 15 ; 23 ; 13 ; 39 ; 4 ; 38 . Specifically, U(Q)\sim\mathcal{GP}(U_{\mathrm{mean}}(Q),C(Q,Q^{\prime})) , where U_{\mathrm{mean}} refers to the mean field and C(Q,Q^{\prime}) refers to the spatial correlation function. We consider a discrete trajectory X_{0:N}^{(\alpha)}=\{Q_{n}^{(\alpha)},P_{n}^{(\alpha)}\}_{n=0}^{N} for time step \Delta t and total time T=N\Delta t , where \alpha indexes a specific realization of the local potential. The likelihood of this trajectory is then obtained by marginalizing over all possible configurations of U^{(\alpha)} p(X_{0:N}^{(\alpha)};K,C)=\int\mathcal{D}[U^{(\alpha)}]p(X_{0:N}^{(\alpha)}\mid U^{(\alpha)};K)p(U^{(\alpha)};C). (2) As this high-dimensional path integral is computationally intractable, we employ variational inference to maximize a lower bound on the log-likelihood (the evidence lower bound, ELBO). By introducing an approximate posterior distribution q(U^{(\alpha)}) , the log-likelihood is strictly bounded by: \displaystyle\ln p(X_{0:N}^{(\alpha)};K,C)\geq \displaystyle\mathbb{E}_{q(U^{(\alpha)})}[\ln p(X_{0:N}^{(\alpha)}\mid U^{(\alpha)};K)] (3) \displaystyle-D_{\text{KL}}(q(U^{(\alpha)})\parallel p(U^{(\alpha)};C)). The direct parameterization of q(U^{(\alpha)}) is still unfeasible due to the continuous nature of the function space. However, evaluating the trajectory likelihood only requires the potential values at the discrete spatial coordinates visited in X_{0:N}^{(\alpha)} . To construct a tractable finite-dimensional representation, we introduce a set of M\ll N learnable inducing variables \mathbf{u}^{(\alpha)} to serve as a low-dimensional backbone for the local landscape 44 . By defining the joint variational distribution as q(U^{(\alpha)},\mathbf{u}^{(\alpha)})=p(U^{(\alpha)}\mid\mathbf{u}^{(\alpha)})q(\mathbf{u}^{(\alpha)}) , where q(\mathbf{u}^{(\alpha)})=\mathcal{N}(\bm{\mu},\mathbf{S}) is a Gaussian variational distribution, the marginal distribution over the potential field at the trajectory data locations analytically resolves to: \displaystyle q(U^{(\alpha)}) \displaystyle=\int q(\mathbf{u}^{(\alpha)})\,p(U^{(\alpha)}|\mathbf{u}^{(\alpha)})\,d\mathbf{u}^{(\alpha)} (4) \displaystyle=\mathcal{N}\!\left(U^{(\alpha)}\mid\mathbf{W}\bm{\mu},\;\mathbf{\Sigma}_{nn}+\mathbf{W}(\mathbf{S}-\mathbf{\Sigma}_{mm})\mathbf{W}^{\mathsf{T}}\right), where \mathbf{W}=\mathbf{\Sigma}_{nm}\mathbf{\Sigma}_{mm}^{-1} acts as a weight matrix. Here, the subscripts n and m index the trajectory data coordinates and the inducing point locations, such that \mathbf{\Sigma}_{nn} , \mathbf{\Sigma}_{mm} , and \mathbf{\Sigma}_{nm} denote the covariance matrices evaluated among these point sets (see Supplemental Material for details). Consequently, the intractable KL divergence in the ELBO strictly reduces to D_{\text{KL}}(q(\mathbf{u}^{(\alpha)})\parallel p(\mathbf{u}^{(\alpha)};C)) , yielding a fully differentiable objective function. Beyond the spatial disorder, computing the exact trajectory probability p(X_{0:N}^{(\alpha)}\mid U^{(\alpha)};K) also requires integrating out the viscoelastic memory. To this end, we embed the non-Markovian friction into an extended phase space by introducing auxiliary variables \bm{\zeta} , which represent the degrees of freedom of the heat bath coupled to the CG coordinates 9 ; 8 . The resulting Markovian process [Q,P,\bm{\zeta}] is given by: \displaystyle\dot{Q} \displaystyle=P, (5) \displaystyle\dot{P} \displaystyle=M^{-1}(-\nabla U(Q)+\mathbf{a}_{p}^{\mathsf{T}}\bm{\zeta}) \displaystyle\dot{\bm{\zeta}} \displaystyle=-\mathbf{A}\,\bm{\zeta}-\mathbf{a}_{p}V+\mathbf{B}\,\bm{\xi}(t), where \bm{\xi}(t) refers to the independent standard Gaussian noise. \{\mathbf{a}_{p},\mathbf{A}\} parameterize the memory kernel K(t)=\mathbf{a}_{p}^{\mathsf{T}}e^{-\lvert t\rvert\mathbf{A}}\mathbf{a}_{p} . For equilibrium processes, the fluctuation-dissipation theorem \mathbf{B}\mathbf{B}^{\mathsf{T}}=k_{B}T(\mathbf{A}+\mathbf{A}^{\mathsf{T}}) should be satisfied 25 . Conditioned on a specific local potential U^{(\alpha)} , evaluating the exact trajectory likelihood p(X^{(\alpha)}_{0:N}\mid U^{(\alpha)};K) requires marginalizing over the unobservable auxiliary variables \bm{\zeta}^{(\alpha)}_{0:N} . Due to the conditionally linear and Markovian nature of Eq. 5 , this high-dimensional path integral can be evaluated exactly. Taking advantage of the explicit form of the transition probabilities, we apply an iterative predictor–corrector approach (also known as the Kalman filter) 37 ; 46 . Starting from the trajectory up to step k-1 , we determine the probability of the hidden variable \bm{\zeta}^{(\alpha)}_{k}