betaFeedback

WorksPhysics and Astronomy

Efficient simulation of the generalized Langevin equation using a basis of dynamically relevant histories for memory and noise

Abstract

The Generalized Langevin Equation (GLE) is a cornerstone for modeling non-Markovian dynamics, but its numerical solution remains challenging due to the history-dependent convolution and the need to generate colored noise with a prescribed correlation function. We introduce a forward construction that addresses both challenges within a single framework. We represent an arbitrary trajectory as sum of relevant and irrelevant parts, where relevant part can be decomposed into a small number of characteristic history patterns that make the main contribution to future dynamics. It allows to reduce numerical cost to O(d2)O(d^2) for dd chosen patterns instead of full history. The same construction applies to colored noise generation, where the causal filter plays the role of the memory kernel. We demonstrate the accuracy and efficiency of the method for several representative kernels, including long-range kernels with power-law tails.

Introduction

The Generalized Langevin Equation (GLE) stands as a cornerstone of non-equilibrium statistical mechanics, providing a rigorous framework for describing systems with memory effects. Originally derived through projection operator techniques [1, 2, 3, 4, 5, 6, 7, 8], the GLE takes the form:

mv˙(t)=F(x(t))−∫−∞tM(t−τ)v(τ)dτ+ξ(t)(1)m\dot{v}(t) = F(x(t)) - \int_{-\infty}^{t} M(t-\tau)v(\tau)\mathrm{d}\tau+ \xi(t) \tag*{(1)}

where v(t)=x˙(t)v(t)=\dot{x}(t) is a velocity of a particle with mass mm, F(x)F(x) is external force, M(t)M(t) is the memory kernel encoding temporal correlations, and ξ(t)\xi(t) is a colored noise with the correlation function ⟨ξ(t)ξ(s)⟩=N(t−s)\langle\xi(t)\xi(s)\rangle= N(t-s).

This formalism finds applications in a wide range of physical problems, including anomalous diffusion and aging in complex environments [9, 10, 11], subdiffusive transport in viscoelastic media and Brownian motion with hydrodynamic memory [12, 13], anomalous polymer dynamics and non-Markovian coarse-grained models of macromolecules [14, 15, 16, 17], as well as colored-noise thermostats [18, 19]. Related applications also arise in quantum dissipative systems [20, 21, 22, 23, 24].

A central challenge is the numerical solution of the GLE, which involves the history-dependent convolution in Eq. (1) and the generation of colored noise with a determined correlation function. Traditional numerical strategies navigate a fundamental trade-off: simplifying the non-Markovian dynamics to make them tractable while attempting to retain essential memory effects.

One of the most straightforward strategies is the direct numerical integration of the GLE. In this approach, the history-dependent convolution integral is evaluated explicitly at each time step, often using a simple quadrature rule and a time cutoff for the memory kernel [25, 26].

While conceptually simple and applicable to arbitrary memory kernels, the direct evaluation of the convolution scales quadratically with the number of time steps (O(N2)O(N^2)) and requires storing the full history of the trajectory, which becomes prohibitively expensive for long-time simulations. Significant efforts have been made to optimize this direct approach [27, 28, 29]. However, direct methods still face challenges in efficiently and accurately handling strongly non-Markovian systems, especially those with long-range or power-law memory kernels like anomalous diffusion and low-temperature Brownian motion.

A common approach is the Markovian embedding paradigm, where the dynamics are reformulated as a high-dimensional system of coupled Markovian stochastic differential equations. For general kernels, this reformulation is not exact and requires a separate approximation step: the memory kernel must first be represented in a tractable class, typically as a positive Prony series (a sum of exponentials) [30, 31, 32, 33]. An alternative route to a tractable kernel representation is to reconstruct it from trajectory data, i.e., to solve the inverse problem. Several data-driven approaches have been developed in this direction, ranging from rational approximations in the Laplace domain [34] and iterative reconstruction schemes [29, 35] to deep-learning-based kernel extraction [36, 37]. Beyond homogeneous kernels, recent work also addresses state-dependent memory [38] and ab initio parametrizations [39]. These methods extract the memory kernel (or its parameters) from molecular dynamics simulations and then embed the resulting kernel into a Markovian system, so they still inherit the two-step structure—reconstruction followed by embedding. Recent refinements of this concept focus on deriving systematic memory corrections within this embedded framework. For instance, the effective mass approach yields a Markovian Langevin equation with a renormalized mass, capturing leading-order memory effects for short correlation times and systematically improvable via higher-order corrections [40, 41]. While valuable, this methodology exemplifies the prevalent two-step, prescriptive paradigm: first, the memory kernel is approximated, fitted, or reconstructed to a specific class; second, high-performance integrators are applied to the resulting approximate, quasi-Markovian system [27]. Such a separation requires a high-quality initial fit of the kernel—or a high-quality reconstruction stage—which is particularly problematic for kernels of a power-law form.

Furthermore, the need to generate the correlated noise ξ(t)\xi(t) presents a parallel challenge, often addressed by algorithms designed for specific correlation forms or assuming underlying auxiliary noise dynamics [42, 43, 44]. In cases where the classical FDT is not satisfied, the application of standard approaches becomes challenging or even impossible [45, 46].

In this paper, we propose a different approach to Markovian embedding. Instead of approximating the kernel or introducing additional variables, we analyze which patterns of past history contribute most significantly to the memory-dependent convolution. Based on this, we construct an efficient, low-dimensional basis of relevant historical patterns. Consequently, rather than needing to store the entire history, we only need to retain a few decomposition components, which are sufficient to compute the convolution to a specified level of accuracy. Furthermore, we require these decomposition coefficients to evolve in a self-consistent manner, rendering them natural auxiliary Markovian variables.

Moreover, the same construction applies to the generation of colored noise. For a stationary correlation function whose covariance matrix is positive definite, a realization of the colored noise is obtained by convolving white noise with a causal filter whose autocorrelation reproduces the target correlation function. In this representation, the sequence of independent, normally distributed random variables plays exactly the same role as the velocity trajectory, while the causal filter acts as the memory kernel.

We demonstrate that this formalism accurately reproduces results from direct integration benchmarks for several representative kernels, including long-range kernels with power-law tails relevant to anomalous diffusion, where traditional Prony series approximations converge slowly.

The paper is organized as follows. Section II introduces the model and formulates the problem for both the memory force and the colored noise. Section III presents the recursive basis method, covering the construction and convergence of the basis, its application to memory and noise, and the resulting integration scheme for the GLE. Section IV reports numerical validation for representative kernels. We conclude in Section V.

Model and problem formulation

We consider a particle described by the GLE, Eq. (1), where M(t)M(t) is the memory kernel characterizing the time-nonlocal friction and ξ(t)\xi(t) is a stationary Gaussian noise with mean ⟨ξ(t)⟩=0\langle\xi(t)\rangle=0, and correlation function ⟨ξ(t)ξ(s)⟩=N(t−s)\langle\xi(t)\xi(s)\rangle=N(t-s); in general, N(t)N(t) is not related to M(t)M(t).

For example, for quantum Brownian motion at zero temperature [24] these functions are M(t)∝ϵ/(ϵ2+t2)M(t)\propto\epsilon/(\epsilon^2+t^2) and N(t)∝(ϵ2−t2)/(ϵ2+t2)2N(t)\propto(\epsilon^2-t^2)/(\epsilon^2+t^2)^2. Noise correlation function has slow negative algebraic tails and does not satisfy the classical FDT.

The numerical challenge is evaluating the memory force and generating colored noise. We begin with the memory force which can be rewritten as

Fmem(t)=∫0∞M(s)v(t−s) ds.(2)F_{\mathrm{mem}}(t)=\int_{0}^{\infty}M(s)v(t-s)\,ds. \tag*{(2)}

For its direct computation we must use whole past history, which is numerically expensive. We simplify this problem by the next way. Represent history as

v(t−s)=vrel(t−s)+δv(t−s).(3)v(t-s)=v_{\mathrm{rel}}(t-s)+\delta v(t-s). \tag*{(3)}

Let us assume that there exist characteristic patterns vi(s)v_i(s) such that vrel(t−s)=∑i=1rai(t)ui(s)v_{\mathrm{rel}}(t-s)=\sum_{i=1}^{r}a_i(t)u_i(s), where rr is a sufficiently small number (∼5\sim5–1010) and ai(t)a_i(t) are the expansion coefficients. It is assumed that the functions ui(s)u_i(s) are known; thus, the memory force is uniquely determined solely by the coefficients ai(t)a_i(t). We are looking for a representation in which the full memory dynamics are governed exclusively by the self-consistent dynamics of these coefficients.

Naturally, it is necessary to require that the irrelevant part δv(t−s)\delta v(t-s) makes a controllably negligible contribution to the memory force, and that it can be disregarded without introducing errors at future times.

Problem statement

We discretize the dynamics with a uniform time step Δt\Delta t, using a past window tpast=pΔtt_{\mathrm{past}}=p\Delta t and a future window tfuture=fΔtt_{\mathrm{future}}=f\Delta t, with positive integers pp and ff. Both windows must be much larger than the memory time τM\tau_M, so that the contribution to the integral beyond them is negligible.

At time step nn, let xn∈Rp\mathbf{x}_n\in\mathbb{R}^{p} be the velocity history, (xn)j=v(tn−jΔt)(x_n)_j=v(t_n-j\Delta t) for j=0,…,p−1j=0,\ldots,p-1. The memory kernel is discretized as mi=Δt M(iΔt)m_i=\Delta t\,M(i\Delta t) for i=0,…,p−1i=0,\ldots,p-1, and mi=0m_i=0 for i≥pi\ge p, reflecting the truncation of the kernel beyond the past window. The memory force at time tnt_n is then

Fnmem=−mTxn=−∑j=0p−1mj(xn)j.(4)F_n^{\mathrm{mem}}=-\mathbf{m}^{\mathsf{T}}\mathbf{x}_n=-\sum_{j=0}^{p-1}m_j(x_n)_j. \tag*{(4)}

At each time step, the memory force requires evaluating the dot product mTxn\mathbf{m}^{\mathsf{T}}\mathbf{x}_n, which costs pp operations. We are looking for a low-dimensional subspace V⊂Rp\mathcal{V} \subset\mathbb{R}^{p} of dimension r≪pr \ll p that contains nearly all the information about any possible history in its expansion coefficients. Formally, let V∈Rp×rV \in\mathbb{R}^{p \times r} be an orthonormal matrix, VTV=IrV^{\mathsf{T}}V = I_r, and let V=range⁡(V)\mathcal{V} = \operatorname{range}(V). The coordinate map is

an=VTxn∈Rr.(5)\mathbf{a}_n = V^{\mathsf{T}}\mathbf{x}_n \in\mathbb{R}^{r}. \tag*{(5)}

We want to find VV and the dynamics operator D∈Rr×(r+1)D \in\mathbb{R}^{r \times(r+1)} such that

Fn=−mTxn≈−mrelTan,(6)F_n = -\mathbf{m}^{\mathsf{T}}\mathbf{x}_n \approx-\mathbf{m}_{\mathrm{rel}}^{\mathsf{T}}\mathbf{a}_n, \tag*{(6)}
an+1=D(vn+1an),(7)\mathbf{a}_{n+1} = D \begin{pmatrix} v_{n+1} \\ \mathbf{a}_n \end{pmatrix}, \tag*{(7)}

where mrel:=VTm\mathbf{m}_{\mathrm{rel}} := V^{\mathsf{T}}\mathbf{m}. The first requirement ensures a correct calculation of the memory force: an\mathbf{a}_n contains all the necessary information about the full history. This is a static condition for the relevance of the selected subspace. The second ensures dynamical consistency: the evolution of an\mathbf{a}_n is determined only by an\mathbf{a}_n and the most recent velocity vn+1v_{n+1}.

To formalize these requirements, we split the history as

xn=Van+δxn,δxn:=(I−VVT)xn,(8)\mathbf{x}_n = V\mathbf{a}_n + \delta\mathbf{x}_n,\qquad\delta\mathbf{x}_n := (I - VV^{\mathsf{T}})\mathbf{x}_n, \tag*{(8)}

so that the memory force becomes

Fn=−mTxn=−mTVan−mTδxn.(9)F_n = -\mathbf{m}^{\mathsf{T}}\mathbf{x}_n = -\mathbf{m}^{\mathsf{T}}V\mathbf{a}_n - \mathbf{m}^{\mathsf{T}}\delta\mathbf{x}_n. \tag*{(9)}

Let’s examine both requirements in turn.

Relevant subspace and memory intensity operator

The requirement (6) is therefore equivalent to mTδxn≈0\mathbf{m}^{\mathsf{T}}\delta\mathbf{x}_n \approx0: discarded, irrelevant part does not contribute into the memory force. However, we must also ensure that it does not contribute in the future. The full history evolves as

xn+1=Sxn+vn+1ein,(10)\mathbf{x}_{n+1} = S\mathbf{x}_n + v_{n+1}\mathbf{e}_{\mathrm{in}}, \tag*{(10)}

where SS is the shift operator and ein=[1,0,…,0]T\mathbf{e}_{\mathrm{in}} = [1,0,\ldots,0]^{\mathsf{T}} is a unit vector representing an injection concentrated at the most recent past. Substituting the split (8) into (10) and iterating ss times, we obtain

xn+s=Ssxn+∑k=1svn+kSs−kein,(11)\mathbf{x}_{n+s} = S^{s}\mathbf{x}_n + \sum_{k=1}^{s}v_{n+k}S^{s-k}\mathbf{e}_{\mathrm{in}}, \tag*{(11)}

and hence

Fn+s=−(m(s))Txn−∑k=1sms−kvn+k,(12)F_{n+s} = -(\mathbf{m}^{(s)})^{\mathsf{T}}\mathbf{x}_n - \sum_{k=1}^{s}m_{s-k}v_{n+k}, \tag*{(12)}

where mj(s)=ms+j\mathbf{m}^{(s)}_j = m_{s+j}, with the convention mi=0m_i = 0 for i≥pi \ge p. The residual contribution to the future force is (m(s))Trn(\mathbf{m}^{(s)})^{\mathsf{T}}\mathbf{r}_n, so the requirement mTrn≈0\mathbf{m}^{\mathsf{T}}\mathbf{r}_n \approx0 extends to

∣(m(s))Trn∣≈0for all s=0,…,f−1.(13)\left|(\mathbf{m}^{(s)})^{\mathsf{T}}\mathbf{r}_n\right| \approx0 \quad\text{for all } s=0,\ldots,f-1. \tag*{(13)}

Combining these requirements, we obtain final static condition for relevant basis

E(V)=∑s=0f−1∥m(s)−VVTm(s)∥2⟶min⁡.(14)E(V) = \sum_{s=0}^{f-1}\left\|\mathbf{m}^{(s)} - VV^{\mathsf{T}}\mathbf{m}^{(s)}\right\|^{2} \longrightarrow\min. \tag*{(14)}

We introduce the memory intensity operator R∈Rp×pR \in\mathbb{R}^{p \times p}, which aggregates the influence of the history over all future shifts within the considered window:

R:=∑s=0f−1m(s)m(s)T.(15)R := \sum_{s=0}^{f-1}\mathbf{m}^{(s)}\mathbf{m}^{(s)\mathsf{T}}. \tag*{(15)}

In component form, Rij=∑s=0f−1ms+ims+jR_{ij} = \sum_{s=0}^{f-1}m_{s+i}m_{s+j}, with the convention mi=0m_i = 0 for i≥pi \ge p. For a unit-norm history pattern x\mathbf{x}, the quadratic form ⟨x,Rx⟩\langle\mathbf{x},R\mathbf{x}\rangle measures the total squared force exerted on the future by this pattern over the future window. Thus, (14) is equivalent to

E(V)=∑s=0f−1∥m(s)∥2−tr⁡(VTRV).(16)E(V) = \sum_{s=0}^{f-1}\left\|\mathbf{m}^{(s)}\right\|^{2} - \operatorname{tr}(V^{\mathsf{T}}RV). \tag*{(16)}

Dynamical consistency

Except for a basis V∈Rp×rV \in\mathbb{R}^{p \times r} that approximately satisfies the static requirement (16) we also need the dynamics operator with the dynamical consistency condition introduced below.

The simulating dynamics requires us to track the system’s position within a subspace as new information arrives at each time step. Consider the system at a discrete time step nn. We maintain a set of rr orthonormal basis vectors, stored as columns in a matrix V∈Rp×rV \in\mathbb{R}^{p \times r}. Each column represents a memory pattern that is currently relevant.

As time advances by one step Δt\Delta t, two operations occur. Firstly, aging — all current memory patterns shift one step into the past. Mathematically, this is equivalent to shifting the basis, S(V)S(V). Second, injection — a new piece of information enters the system: the velocity vn+1v_{n+1} at the next time step. This corresponds to the unit vector ein\mathbf{e}_{\mathrm{in}} in the history space.

Dynamical consistency requires that if xn∈V\mathbf{x}_n \in\mathcal{V} (so xn≈Van\mathbf{x}_n \approx V\mathbf{a}_n), then the updated history xn+1=SVan+vn+1ein\mathbf{x}_{n+1} = SV\mathbf{a}_n + v_{n+1}\mathbf{e}_{\mathrm{in}} also remains in V\mathcal{V}. In matrix form, the ideal target is

U(vn+1an)≈VD(vn+1an),(17)U \begin{pmatrix} v_{n+1} \\ \mathbf{a}_n \end{pmatrix} \approx VD \begin{pmatrix} v_{n+1} \\ \mathbf{a}_n \end{pmatrix}, \tag*{(17)}

where U=(ein, SV)U = (\mathbf{e}_{\mathrm{in}},\, SV). Since the pair (vn+1,an)(\mathbf{v}_{n+1},\mathbf{a}_n) is arbitrary, this yields the dynamical condition

U≈VD.(18)U \approx VD. \tag*{(18)}

Equation (18) for r<pr < p cannot hold exactly. However, we do not require this equality to hold strictly. It suffices for the difference between the left- and right-hand sides of the equation to lie within an irrelevant space. Mathematically, this is equivalent to the projection of the residual onto the memory intensity matrix RR is minimal. Denote δU=U−VD∈Rp×(r+1)\delta U = U - VD \in\mathbb{R}^{p \times(r+1)}, so, by analogy with (16), we obtain the second functional

tr⁡(δUTRδU) is minimized.(19)\operatorname{tr}(\delta U^{\mathsf{T}} R \delta U)\ \text{is minimized}. \tag*{(19)}

Simultaneously, the static requirement must be satisfied:

tr⁡(VTRV) is maximized.(20)\operatorname{tr}(V^{\mathsf{T}} R V)\ \text{is maximized}. \tag*{(20)}

In summary, we are looking for a subspace that is memory-intensity-maximizing and the evolution rule which maintain this subspace approximately shift-invariant in the sense of memory intensity. We choose a basis in the history space such that each history splits into relevant and irrelevant parts, where the relevant part contains all the information necessary to calculate the memory force and its own evolution (19), while the irrelevant part makes no contribution to future (20).

Colored noise: from correlation function to causal filter

To generate noise with a specified correlation N(t)N(t), we are looking for a causal operation that maps an uncorrelated Gaussian white noise sequence {ζn}\{\zeta_n\} into the desired colored noise {ξn}\{\xi_n\}. Specifically, we want to find a causal filter c=[c0,c1,…,cp−1]T\mathbf{c} = [c_0,c_1,\ldots,c_{p-1}]^{\mathsf{T}} such that the convolution

ξn=∑j=0p−1cjζn−j(21)\xi_n = \sum_{j=0}^{p-1} c_j \zeta_{n-j} \tag*{(21)}

reproduces the target correlations:

⟨ξnξn+k⟩=N(kΔt)for k=0,1,…,p−1.(22)\langle\xi_n \xi_{n+k}\rangle= N(k\Delta t) \quad\text{for } k = 0,1,\ldots,p-1. \tag*{(22)}

The filter coefficients must satisfy

∑j=0p−k−1cjcj+k=N(kΔt)for k=0,1,…,p−1.(23)\sum_{j=0}^{p-k-1} c_j c_{j+k} = N(k\Delta t) \quad\text{for } k = 0,1,\ldots,p-1. \tag*{(23)}

Thus the filter encodes the same correlation structure, represented in a causal form suitable for sequential noise generation.

Finding this filter is equivalent to computing the square root of the noise covariance matrix TT, where

Tij=N(∣i−j∣Δt);T_{ij} = N(|i-j|\Delta t);

we assume TT is positive definite. The Cholesky decomposition T=LLTT = LL^{\mathsf{T}} provides such a factorization, with LL lower-triangular. For a stationary correlation N(t)N(t), the columns of LL become approximately translationally invariant far from the matrix boundaries:

Li+k,j+k≈Li,jL_{i+k,j+k} \approx L_{i,j}

for sufficiently large kk.

The causal filter c\mathbf{c} can be treated as a memory kernel acting on past white noise, so the problem of generating colored noise reduces entirely to the problem of calculating the memory force.

Basis construction and GLE integration

In this section we discuss explicit and recursive approaches to construction of the basis that satisfies the given requirements (19) and (20).

Explicit construction of relevant subspace

Let’s consider the spectral decomposition of RR:

Rui=λiui,ui∈Rp,λ1≥λ2≥⋯≥λp≥0.(24)R\mathbf{u}_i = \lambda_i \mathbf{u}_i,\qquad\mathbf{u}_i \in\mathbb{R}^p,\qquad\lambda_1 \ge\lambda_2 \ge\cdots\ge\lambda_p \ge0. \tag*{(24)}

The eigenvectors ui\mathbf{u}_i form an orthonormal basis for the space of past histories. Their corresponding eigenvalues λi\lambda_i quantify the importance of each eigenvector: a history aligned with ui\mathbf{u}_i couples to the future with an intensity proportional to λi\lambda_i.

By Ky Fan’s maximum principle [47], tr⁡(VTRV)\operatorname{tr}(V^{\mathsf{T}}RV) is maximized when the columns of VV are the eigenvectors of RR corresponding to the rr largest eigenvalues: V=(u1…ur)∈Rp×rV = (\mathbf{u}_1 \ldots\mathbf{u}_r) \in\mathbb{R}^{p \times r}. Thus, the maximally-static-optimal basis for a fixed dimension rr is

Srel={u1…ur},(25)\mathcal{S}_{\mathrm{rel}} = \{\mathbf{u}_1 \ldots\mathbf{u}_r\}, \tag*{(25)}

The span of Srel\mathcal{S}_{\mathrm{rel}} constitutes the relevant subspace V⊂Rp\mathcal{V} \subset\mathbb{R}^p, a low-dimensional subspace that captures the dominant pathways through which the past influences the future for the given kernel M(t)M(t).

The spectrum of RR for a representative kernel is shown in Figure 1: the eigenvalues decay rapidly with the mode index, indicating that only the first few modes carry significant weight in the action of the kernel. These few modes are also shown in the bottom plot.

Figure 1. Spectrum and first eigenfunctions of the memory intensity operator RR for the memory function M(t)=τ/(t2+τ2)M(t)=\tau/(t^2+\tau^2). The eigenvalues λi\lambda_i decay rapidly with the mode index ii, indicating that only the first few modes carry significant weight in the action of the kernel.

If we choose VV by this way, we can also explicitly obtain dynamics operator. For fixed VV, the condition (19) is least squares problem and it has solution [48]

D=(VTRV)−1VTRU(26)D = (V^{\mathsf{T}}RV)^{-1}V^{\mathsf{T}}RU \tag*{(26)}

For VV which composed from eigenvectors of RR this formula is simplified to

D=VTU(27)D = V^{\mathsf{T}}U \tag*{(27)}

Recursive construction

Other approach is recursive construction. We start from static-optimal V0=(u1…ur)V_0=(\mathbf{u}_1 \ldots\mathbf{u}_r). Suppose, that on the step nn we have any VnV_n. So on the step n+1n+1 we construct an augmented basis Un+1U_{n+1} by concatenating the new information with all the aged patterns from the previous step:

Un+1=(ein SVn)∈Rp×(r+1).(28)U_{n+1}=(\mathbf{e}_{in}\,S V_n)\in\mathbb{R}^{p\times(r+1)}. \tag*{(28)}

This collection of (r+1)(r+1) vectors spans a space that contains all possible linear combinations of the fresh data and aged memories. The dimensionality has temporarily increased by one.

The augmented basis Un+1U_{n+1} is overcomplete. To maintain a fixed, minimal dimension rr, we must compress this information back into the rr-dimensional relevant subspace V\mathcal{V} defined by the memory kernel. This is achieved by projecting the memory intensity operator RR onto the augmented basis:

ρn+1=Un+1TRUn+1∈R(r+1)×(r+1).(29)\rho_{n+1}=U_{n+1}^{\mathsf{T}}R U_{n+1}\in\mathbb{R}^{(r+1)\times(r+1)}. \tag*{(29)}

The matrix ρn+1\rho_{n+1} encodes how strongly each vector in Un+1U_{n+1} (and their combinations) couples to future dynamics through the memory kernel.

We then perform an eigendecomposition of ρn+1\rho_{n+1}:

ρn+1U~n+1=U~n+1Πn+1,(30)\rho_{n+1}\widetilde{U}_{n+1}=\widetilde{U}_{n+1}\Pi_{n+1}, \tag*{(30)}

where Πn+1\Pi_{n+1} is a diagonal matrix of eigenvalues π1≥π2≥⋯≥πr+1≥0\pi_1\ge\pi_2\ge\cdots\ge\pi_{r+1}\ge0, and U~n+1\widetilde{U}_{n+1} contains the corresponding orthonormal eigenvectors. These eigenvectors represent the optimal linear combinations of the fresh and aged data, ordered from most to least important in terms of their future influence.

The optimal combinations are mapped back to the original history space to form a new, full set of candidate patterns:

Vn+1(full)=Un+1U~n+1∈Rp×(r+1).(31)V_{n+1}^{(\mathrm{full})}=U_{n+1}\widetilde{U}_{n+1}\in\mathbb{R}^{p\times(r+1)}. \tag*{(31)}

The columns of Vn+1(full)V_{n+1}^{(\mathrm{full})} are orthonormal and ranked by importance. To complete the recursion and maintain a fixed state dimension, we discard the least important pattern—the one corresponding to the smallest eigenvalue πr+1\pi_{r+1}:

Vn+1=Vn+1(full)(Ir  0)T∈Rp×r.(32)V_{n+1}=V_{n+1}^{(\mathrm{full})}(I_r\ \ 0)^{\mathsf{T}}\in\mathbb{R}^{p\times r}. \tag*{(32)}

The matrix Vn+1V_{n+1} now contains our updated orthonormal basis for time step n+1n+1. Iterating this procedure from a suitable initial basis V0V_0 (e.g., the dominant eigenvectors of RR) leads to convergence. The sequence {Vn}\{V_n\} typically reaches a steady state V∞V_\infty, where the update becomes a pure rotation within the relevant subspace V\mathcal{V}. The associated transformation matrix U~∞\widetilde{U}_\infty is used in the dynamical integration:

D:=(Ir  0)U~∞T.(33)D:=(I_r\ \ 0)\widetilde{U}_\infty^{\mathsf{T}}. \tag*{(33)}

Colored Noise Generation

By Sec. II, we have reduced the problem of generating colored noise to the problem of finding a retarded force. We proceed exactly as in Sec. II and Sec. III, but now with cc playing the role of the discretized kernel. We construct the memory intensity operator RCR_C for the filter:

RC=1f∑s=0f−1c(s)(c(s))T,(34)R_C=\frac{1}{f}\sum_{s=0}^{f-1}\mathbf{c}^{(s)}\left(\mathbf{c}^{(s)}\right)^{\mathsf{T}}, \tag*{(34)}

where c(s)=[cs,…,cs+p−1]T\mathbf{c}^{(s)}=[c_s,\ldots,c_{s+p-1}]^{\mathsf{T}} is a segment of the filter. Repeating the procedure in its entirety for the memory we obtain the relevant basis and dynamics operator for the noise process.

Solving the GLE with the Recursive Basis

The final algorithm combines both tasks. We maintain two state vectors in the relevant subspaces VM\mathcal{V}_M and VN\mathcal{V}_N: an\mathbf{a}_n – the deterministic memory state, compressing past velocities and bn\mathbf{b}_n – the stochastic noise state, compressing past white noise inputs.

Recall the discretized GLE for a particle of unit mass, driven by a colored noise ξn\xi_n:

vn+1−vn=Δt(F(xn)+Fmem,n+ξn).(35)v_{n+1}-v_n=\Delta t\left(F(x_n)+F_{\mathrm{mem},n}+\xi_n\right). \tag*{(35)}

Let us describe the complete algorithm. During the preliminary computation stage, we calculate a causal filter for the noise and, using either an explicit or a recursive construction, form the corresponding bases VV and V(N)V^{(N)}, as well as the dynamics operators DD and D(N)D^{(N)} — for both memory and noise.

At each time step we calculate the memory force as a simple low-dimensional dot product

Fmem,n≈mrelTan,wheremrel=VTm.(36)F_{\mathrm{mem},n}\approx\mathbf{m}_{\mathrm{rel}}^{\mathsf{T}}\mathbf{a}_n,\quad\text{where}\quad\mathbf{m}_{\mathrm{rel}}=V^{\mathsf{T}}\mathbf{m}. \tag*{(36)}

The noise ξn\xi_n defined by correlation function C(t)C(t) is similarly constructed as

ξn≈crelTbn,wherecrel=V(N)Tc(37)\xi_n\approx\mathbf{c}_{\mathrm{rel}}^{\mathsf{T}}\mathbf{b}_n,\quad\text{where}\quad\mathbf{c}_{\mathrm{rel}}=V^{(N)\mathsf{T}}\mathbf{c} \tag*{(37)}

For the next time step, we update the relevant vectors

an+1=D(vn+1an)(38)\mathbf{a}_{n+1}=D\begin{pmatrix}v_{n+1}\\\mathbf{a}_n\end{pmatrix} \tag*{(38)}
bn+1=D(N)(ξn+1bn)(39)\mathbf{b}_{n+1}=D^{(N)}\begin{pmatrix}\xi_{n+1}\\\mathbf{b}_n\end{pmatrix} \tag*{(39)}

This scheme is visualized in Figure 2.

Schematic of the relevant histories approach, showing the conventional view with full past trajectory $v(s)$$, memory kernel $M(t-s)$, convolution, force $F_{n+1}$, and the relevant histories approach with relevant $\mathcal{V}$, injection, dynamics operator, dot product, and force $F_{n+1}$

Figure 2. Schematic of the relevant histories approach.

Results

In this section we provide the applicability of our method for some interest cases which especially can be difficult for other established approaches. We compare the recursive method with these approaches.

Logarithmic diffusion of quantum Brownian particle at zero temperature

We compare the numerical simulation by our method with analytical result for quantum Brownian particle in the Ohmic bath at T=0T=0 [23, 24].

M(t)=2πεε2+t2(40)M(t)=\frac{2}{\pi}\frac{\varepsilon}{\varepsilon^2+t^2} \tag*{(40)}
N(t)=1πε2−t2(ε2+t2)2(41)N(t)=\frac{1}{\pi}\frac{\varepsilon^2-t^2}{(\varepsilon^2+t^2)^2} \tag*{(41)}

In Figure 3 the equilibrium mean square displacement is shown for ε=0.5\varepsilon=0.5 for different rr. It is essentially non-Markovian dynamics with long algebraic tails, but already 8 numbers for both of memory and noise instead of whole history provide almost true result with error ∼0.3%\sim0.3\%.

Figure 3. Comparison of the simulated equilibrium mean squared displacement (MSD) ⟨(x(t)−x(0))2⟩\langle(x(t)-x(0))^2\rangle with the direct GLE integration for different rr for quantum Brownian motion at T=0T=0. Parameters: ε=0.5\varepsilon=0.5, m=1m=1, dt=0.001dt=0.001, Ntraj=2⋅105N_{\mathrm{traj}}=2\cdot10^5.

Subdiffusion

Another interesting problem is classical subdiffusion (N(t)=M(t)=1/πt)(N(t)=M(t)=1/\sqrt{\pi t}) [32, 49]. This case is quite challenging to investigate using standard methods. Figure 4 shows the velocity autocorrelation function (VACF) for various values of rr. When using 8+88+8 relevant modes, the maximum relative error is approximately 0.4%0.4\%.

Figure 4. Comparison of the simulated velocity autocorrelation function (VACF) ⟨v(t)v(0)⟩\langle v(t)v(0)\rangle with the direct GLE integration for different rr for subdiffusion. Parameters: m=1m = 1, dt=0.01dt = 0.01, Ntraj=2⋅105N_{\mathrm{traj}} = 2 \cdot10^{5}.

Comparison with other Markovian embedding approaches

A comparison was made with the Markovian embedding method [30], which utilizes a positive Prony representation. For a subdiffusive kernel (M(t)=1/πt)(M(t)=1/\sqrt{\pi t}), the computational efficiency is similar to that of the relevant histories method: given a comparable number of auxiliary variables, both approaches yield a similar level of accuracy for the VACF. However, the accuracy—and, consequently, the exact number of auxiliary modes required—depends on the specific method used to approximate the memory function, a matter that warrants further analysis. The relevant histories method does not require such an analysis.

In the case of quantum Brownian motion at T=0T = 0, the original algorithm cannot be applied directly, as it relies on the FDT and a positive Prony representation—conditions that are not met here, given the specified memory and noise correlation.

Conclusion

Numerical simulation based on the generalized Langevin equation (GLE) requires the computation of a history-dependent memory term, as well as the generation of colored noise with a specified correlation function—both of which are highly computationally intensive tasks. While existing approaches have significantly alleviated these challenges, they still have inherent limitations. Advanced direct integration methods accelerate the calculation of the convolution; however, because they remain direct integration techniques, the computational cost increases substantially as the time step is reduced or when dealing with long-memory kernels. Markovian embedding methods, such as those based on Prony series, allow for a representation of the system in a lower-dimensional space; yet, they exhibit slow convergence for kernels with algebraic tails and cannot account for kernels that are not positive-definite. Furthermore, such methods are often strictly tied to the classical FDT, which limits their applicability to scenarios where the noise does not satisfy this theorem. Consequently, there is a need for a method capable of handling kernels with algebraic tails (without relying on exponential approximation), accommodating kernels that are not positive-definite, generating noise independently of the FDT, and maintaining linear computational complexity with respect to the number of simulation steps.

We propose a method that addresses these limitations by analyzing which patterns of the past history contribute most significantly to the memory convolution. Instead of approximating the memory kernel, we represent the history in a low-dimensional basis of relevant patterns, selected so as to capture the dominant contribution to the convolution. By additionally requiring dynamical consistency—that the expansion coefficients evolve in a closed, self-contained manner—we obtain a natural Markovian embedding: the coefficients become auxiliary variables whose evolution is determined solely by themselves and the current velocity, and whose cost per step is far below that of direct convolution.

Furthermore, we represent noise using a causal filter, which allows the problem of noise generation to be reduced entirely to the calculation of the memory strength. This approach does not require the assumption of positivity or the classical FDT, enabling the modeling of a much broader class of noise—a capability that is particularly essential for non-equilibrium molecular dynamics or low-temperature quantum Brownian motion.

This approach offers several key advantages:

  • It provides a direct construction that avoids kernel fitting and the introduction of auxiliary variables, while keeping the full, non-Markovian GLE as the starting point.

  • It delivers a controlled compression of the memory: for a given tolerance, the relevant basis retains the dominant spectral components of the memory operator and provides a direct measure of the effective dimensionality of the memory process.

  • It reduces the cost of history integration from O(N2)O(N^{2}) to O(d2N)O(d^{2}N) for a dd-dimensional basis, so that the cost scales linearly with the number of simulation steps for a fixed kernel.

  • It accommodates kernels that are not positive definite—including alternating-sign kernels, for which a positive Prony representation is not available—without requiring a different algorithmic structure, and it applies to noise correlation functions that are not tied to the kernel by the classical fluctuation-dissipation theorem.

We test our method on challenging benchmarks—subdiffusion and quantum Brownian motion at T=0T = 0—where the Prony representation is inapplicable and the classical FDT does not hold. Nevertheless, in these cases, a small number of relevant modes provide good agreement with direct integration. In summary, we propose a natural Markovian embedding method, in which the auxiliary variables are defined directly from the structure of the kernel. This approach allows for the efficient calculation of memory strength and the generation of colored noise with a specified correlation, which significantly facilitates the simulation of GLE with challenging kernels.

References

References

  1. [1]P. Langevin, “Sur la théorie du mouvement brownien,” C. R. Acad. Sci. Paris 146, 530–533 (1908).
  2. [2]H. Mori, “Transport, Collective Motion, and Brownian Motion,” Prog. Theor. Phys. 33, 423–455 (1965).DOI
  3. [3]R. Zwanzig, “Nonlinear generalized Langevin equations,” J. Stat. Phys. 9, 215–220 (1973).DOI
  4. [4]P. Hänngi, “Path Integral Solution for Nonlinear Generalized Langevin Equations,” in Path Integrals from meV to MeV: Tutzing ’92, edited by H. Grabert, A. Inomata, L. Schulmann, and U. Weiss (World Scientific, Singapore, 1993), pp. 289–301.
  5. [5]P. Hänngi, P. Talkner, and M. Borkovec, “Reaction-rate theory: fifty years after Kramers,” Rev. Mod. Phys. 62, 251–341 (1990).
  6. [6]L. Ma, X. Li, and C. Liu, “The derivation and approximation of coarse-grained dynamics from Langevin dynamics,” J. Chem. Phys. 145, 204117 (2016).DOI
  7. [7]R. Zwanzig, “Statistical mechanics of irreversibility,” in Lectures in Theoretical Physics, Vol. 3, pp. 106–141, Interscience, New York (1961).
  8. [8]Y. Zhu and D. Venturi, “Faber approximation of the Mori–Zwanzig equation,” Journal of Computational Physics 372, 694–718 (2018).DOI
  9. [9]E. Lutz, “Fractional Langevin equation,” Phys. Rev. E 64, 051106 (2001).
  10. [10]J. M. Porrà, K.-G. Wang, and J. Masoliver, “Generalized Langevin equations: Anomalous diffusion and probability distributions,” Phys. Rev. E 53, 5872–5881 (1996).pubmed.ncbi.nlm.nih.gov/9964946
  11. [11]N. Pottier, “Aging properties of an anomalously diffusing particle,” Physica A 317, 371–382 (2003).DOI
  12. [12]V. Kharchenko and I. Goychuk, “Flashing subdiffusive ratchets in viscoelastic media,” New J. Phys. 14, 043042 (2012).DOI
  13. [13]T. Franosch, M. Grimm, M. Belushkin, F. M. Mor, G. Foffi, L. Forró, and S. Jeney, “Resonances arising from hydrodynamic memory in Brownian motion,” Nature 478, 85–88 (2011).
  14. [14]D. Panja, “Generalized Langevin equation formulation for anomalous polymer dynamics,” J. Stat. Mech. (2010), L02001.DOI
  15. [15]Z. Li, H. S. Lee, E. Darve, and G. E. Karniadakis, “Computing the non-Markovian coarse-grained interactions derived from the Mori–Zwanzig formalism in molecular systems: Application to polymer melts,” J. Chem. Phys. 146, 011404 (2017).DOI
  16. [16]M. Chen, X. Li, and C. Liu, “Computation of the memory functions in the generalized Langevin models for collective dynamics of macromolecules,” J. Chem. Phys. 141, 064112 (2014).DOI
  17. [17]V. Klippenstein, M. Tripathy, G. Jung, F. Schmid, and N. F. A. van der Vegt, “Introducing memory in coarse-grained molecular simulations,” J. Phys. Chem. B 125, 4931–4954 (2021).DOI
  18. [18]M. Ceriotti, G. Bussi, and M. Parrinello, “Langevin equation with colored noise for constant-temperature molecular dynamics simulations,” Phys. Rev. Lett. 102, 020601 (2009).pubmed.ncbi.nlm.nih.gov/19257259
  19. [19]M. Ceriotti, G. Bussi, and M. Parrinello, “Colored-noise thermostats à la carte,” J. Chem. Theory Comput. 6, 1170–1180 (2010).DOI
  20. [20]U. Weiss, Quantum Dissipative Systems, 2nd ed., World Scientific, Singapore (1999).
  21. [21]C. W. Gardiner and P. Zoller, Quantum Noise (Springer, Berlin, 2004).
  22. [22]H.-P. Breuer and F. Petruccione, The Theory of Open Quantum Systems (Oxford University Press, 2002).
  23. [23]H. Grabert, P. Schramm, and G.-L. Ingold, “Quantum Brownian Motion: The Functional Integral Approach,” Phys. Rep. 168, 115–207 (1988).
  24. [24]D. Kondaurov and E. Polyakov, “Quantum Brownian motion as a classical stochastic process in phase space,” Phys. Rev. A 114, 012123 (2026).
  25. [25]D. L. Ermak and H. Buckholz, “Numerical integration of the Langevin equation: Monte Carlo simulation,” Journal of Computational Physics 35, 169–182 (1980).DOI
  26. [26]D. E. Smith and C. B. Harris, “Generalized Brownian dynamics. I. Numerical integration of the generalized Langevin equation through autoregressive modeling of the memory function,” The Journal of Chemical Physics 92, 1304–1311 (1990).
  27. [27]B. Leimkuhler and M. Sachs, “Efficient Numerical Algorithms for the Generalized Langevin Equation,” SIAM J. Sci. Comput. 44, A364–A388 (2022).arxiv.org/abs/2012.04245
  28. [28]M. López-Fernández, C. Lubich, and A. Schädle, “Adaptive, Fast and Oblivious Convolution in Evolution Equations with Memory,” SIAM J. Sci. Comput. 28, 421–438 (2006).
  29. [29]G. Jung, M. Hanke, and F. Schmid, “Iterative reconstruction of memory kernels,” Journal of Chemical Theory and Computation 13, 2481–2488 (2017).DOI
  30. [30]A. D. Baczewski and S. D. Bond, “Numerical Integration of the Extended Variable Generalized Langevin Equation with a Positive Prony Representable Memory Kernel,” J. Chem. Phys. 139, 044107 (2013).
  31. [31]N. Bockius, M. Braun, K. Hofmann, F. Schmid, and M. Hanke, “Determining extended Markov parameterizations for vector-valued generalized Langevin equations,” Z. Naturforsch. A 81, 157–171 (2026).
  32. [32]N. Bockius, J. Shea, G. Jung, F. Schmid, and M. Hanke, “Model reduction techniques for the computation of extended Markov parameterizations for generalized Langevin equations,” J. Phys.: Condens. Matter 33, 214003 (2021).DOI
  33. [33]N. J. Bockius, “A Lanczos-based method for computing Ornstein-Uhlenbeck representations for the generalized Langevin equation,” (2025).DOI
  34. [34]H. Lei, N. A. Baker, and X. Li, “Data-driven parameterization of the generalized Langevin equation,” Proceedings of the National Academy of Sciences 113, 14183–14188 (2016).DOI
  35. [35]V. Klippenstein, N. Wolf, and N. F. A. van der Vegt, “A Gauss–Newton method for iterative optimization of memory kernels for generalized Langevin thermostats in coarse-grained molecular dynamics simulations,” The Journal of Chemical Physics 160(20), (2024).
  36. [36]M. Kerr Winter, I. Pihlamäa, V. E. Debets, and L. Janssen, “A deep learning approach to the measurement of long-lived memory kernels from generalized Langevin dynamics,” The Journal of Chemical Physics 158(24), (2023).DOI
  37. [37]A. Russo, M. A. Durán-Olivencia, I. G. Kevrekidis, and S. Kalliadiasis, “Machine learning memory kernels as closure for non-Markovian stochastic processes,” IEEE Transactions on Neural Networks and Learning Systems 35, 6531–6543 (2022).DOI
  38. [38]P. Ge, Z. Zhang, and H. Lei, “Data-driven learning of the generalized Langevin equation with state-dependent memory,” Physical Review Letters 133, 077301 (2024).DOI
  39. [39]P. Xie, R. Car, and W. E, “Ab initio generalized Langevin equation,” Proceedings of the National Academy of Sciences 121, e2308668121 (2024).DOI
  40. [40]M. Wiśniewski, J. Luczka, and J. Spiechowicz, “Effective mass approach to memory in non-Markovian systems,” Phys. Rev. E 109, 044116 (2024).arxiv.org/abs/2402.16673
  41. [41]M. Wiśniewski, J. Luczka, and J. Spiechowicz, “Memory Corrections to Markovian Langevin Dynamics,” Entropy 26, 425 (2024).arxiv.org/abs/2405.11370
  42. [42]R. Mannella and V. Palleschi, “Fast algorithm for the simulation of stochastic differential equations with colored noise,” Phys. Rev. A 40, 3381–3386 (1989).
  43. [43]J. García-Ojalvo and J. M. Sancho, Noise in Spatially Extended Systems, Springer-Verlag, New York (1999).
  44. [44]M. Hanke, “Stochastic modeling of stationary scalar Gaussian processes in continuous time from autocorrelation data,” Advances in Computational Mathematics 50, 60 (2024).DOI
  45. [45]R. R. Netz, “Derivation of the non-equilibrium generalized Langevin equation from a generic time-dependent Hamiltonian,” arXiv:2310.00748 (2023).arxiv.org/abs/2310.00748
  46. [46]L. Stella, C. D. Lorenz, and L. Kantorovich, “Generalized Langevin equation: An efficient approach to nonequilibrium molecular dynamics of open systems,” Physical Review B 89, 134303 (2014).DOI
  47. [47]K. Fan, “On a theorem of Weyl concerning eigenvalues of linear transformations I,” Proceedings of the National Academy of Sciences 35, 652–655 (1949).pubmed.ncbi.nlm.nih.gov/16578320
  48. [48]M. Wei and A. R. De Pierro, “Upper perturbation bounds of weighted projections, weighted and constrained least squares problems,” SIAM Journal on Matrix Analysis and Applications 21, 931–951 (2000).DOI
  49. [49]A. D. Viñales and M. A. Desposito, “Anomalous diffusion: Exact solution of the generalized Langevin equation for harmonically bounded particle,” Phys. Rev. E 73, 016111 (2006).pubmed.ncbi.nlm.nih.gov/16486220

Paper details

vlnc:W_yfp8cb4n
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents