betaFeedback

WorksPhysics and Astronomy

Interpretable and Equation-Free Response Theory for Complex Systems

Abstract

Response theory provides a pathway for understanding the sensitivity of a system and for predicting how its statistical properties change when a perturbation is applied. In the case of complex and multiscale systems, to achieve enhanced practical applicability, response theory should be interpretable, capable of focusing on relevant timescales, and amenable to data-driven and equation-agnostic implementations. Along these lines, in the spirit of Markov state modelling, we present linear and nonlinear response formulas for Markov chains. We obtain simple and easily implementable expressions that can be used to predict the response of observables as well as of higher-order correlations. The methodology proposed here can be implemented in a purely data-driven setting and even if the underlying evolution equations are unknown. The use of algebraic expansions inspired by Koopmanism allows to elucidate the role of different time scales and modes of variability, and to find explicit and interpretable expressions for the Green's functions at all orders. This is a major advantage of the framework proposed here. We illustrate our methodology in a very simple yet instructive metastable system. Finally, our results provide a dynamical foundation for the Prony method, which is commonly used for the statistical analysis of discrete time signals.

© The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited.

Introduction

Response theory in statistical mechanics constitutes a powerful framework for analyzing the behavior of a large variety of systems subjected to external perturbations. It provides a powerful and unifying paradigm for connecting the microscopic dynamics and reference statistical properties of a system to its macroscopic response under external influences. Its foundations have been widely discussed in the mathematical literature and its applications permeate various domains of physics, chemistry, biology, materials science, and quantitative social sciences [1, 2, 3]. At the core of response theory lies the formulation of response functions, which quantify the system’s reaction to acting forcings. In the case of systems at thermodynamic equilibrium and considering for the moment only the linear approximation to the response, the fluctuation-dissipation theorem (FDT) establishes that such functions are expressed in terms of time-lagged correlation between suitably defined observables on the unperturbed state [4].

However, response theory is not limited to equilibrium systems [5]. For nonequilibrium systems possessing smooth invariant measure with respect to Lebesgue, as in the case of stochastic dynamical systems forced by sufficiently non-degenerate noise, one can still express the response formulas in terms of time-lagged correlations, which implies the existence of a clear correspondence between forced and free fluctuations of the system [6, 7]. Things become more problematic in the case of dissipative chaotic systems. Here, as a result of the singularity of the invariant measure with respect to Lebesgue, the FDT does not apply and there is no full equivalence between forced and free fluctuations. Yet, making suitable assumptions on the dynamics, it is possible to establish a response theory also in this case. The original results proposed by Ruelle, which required fairly restrictive conditions of uniform hyperbolicity [8, 9], have then been clarified and extended using functional analytical techniques [10, 11, 12].

When the perturbation is large or when the system exhibits a very amplified response, the linear approximation linking the amplitudes of the forcing and of the response breaks down, so that higher-order terms in the system’s response need to be considered [13, 14, 15]. Nonlinear response operators have a convoluted structure and depend on multiple time variables. The nonlinear response describes more complex interplay between internal feedbacks and acting forcings and, if more than one forcing is present, accounts for the interplay - which can be synergistic or antagonistic - of the various acting forcings. Nonlinear effects become quantitatively dominant in the proximity of critical transitions, which are associated with the divergence of the response of the system. However, the occurrence of such divergent behavior can be captured simply by looking at the linear response of the system [16, 17, 18].

A key difficulty of response formulas is that they are based on expressions that do not provide a clear imprint of the dominant modes of variability of the system. A way forward in this direction is provided by Koopmanism [19, 20, 21], which, roughly speaking, transforms nonlinear dynamics into a linear dynamical framework in an infinite dimensional space of observables, where the key information is contained in the eigenfunction and modes of the Koopman (or Kolmogorov in the case of stochastic dynamics) operator. From now on, with an abuse of language, we will use the expression Koopman operator also in the stochastic case. In practice, one needs to approximate the Koopman operator using data-driven methods such as the extended dynamic mode decomposition (eDMD) [22]. One considers a finite dictionary of observables, and each eigenvector of the finite-dimensional approximation of the Koopman operator is obtained as a linear combination of such observables [23]. See [24] for a comprehensive review of eDMD methods.

By linking response theory with Koopmanism, one can derive interpretable representations of the system’s response to perturbations. The use of Koopmanism enables the decomposition of the response into contributions from distinct modes of natural variability of the system [17, 18]. Recently, we have been successful in merging algorithmically response theory and Koopmanism [25, 26] and in showing that Koopmanism provides a pathway for extending response theory to the very relevant yet so far unexplored case where the stochasticity includes jump processes [27].

While these results are extremely encouraging, a nontrivial hurdle that still needs to be overcome is the applicability of this methodology to high-dimensional systems.

A Pipeline for an Interpretable and Equation-free Response Theory

A finite-state, discrete-time Markov chain is a stochastic process that describes the sequence of possible events, chosen among a finite set, evolving, at discrete time, according to a specified probabilistic rule. The probability of occurrence of the future state depends only on the current state and not on past states. Markov chains can be associated with directed graphs, whereby the states correspond to the nodes of the network, and the entries of the Markov matrix give the weight of the directed edge between two nodes [28, 29]. Markov chains are extremely relevant to understand the properties of general dynamical systems [30, 31]. The so-called Ulam method [32] approximates the Perron-Frobenius operator, which pushes forward the probability distributions of a dynamical system [33], by a Markov matrix whose entries represent transition probabilities between partition elements occurring for finite time horizons. The Ulam Conjecture states that the invariant measure of the finite Markov matrix converges to that of the true dynamical system as one considers finer and finer partitions [34].

While the Ulam method is by itself essentially a brute-force approximation, and its convergence is usually slow [35], it is possible to use it very effectively. Specifically, Markov state modeling (MSM) is a smart Ulam method that is particularly effective for studying systems with complex dynamics that evolve across multiple timescales. In this framework, the continuous phase space of a system is optimally discretized into a finite number of so-called (micro)states. Each state corresponds to a cell of the Voronoï tessellation [36] constructed following k-means clustering [37, 38] of the data. The transitions between such states are governed by a Markov chain [39, 40, 41, 42], which contains all the information needed to describe the statistics and dynamics of the system at the coarse-grained level. In this case, the slowest time scales are associated with the relaxation between the main metastable states, and metastable regions can be identified by studying the level sets of the dominant modes of the Koopman operator [43]. Instead, the dynamical processes occurring within each metastable region are associated with faster timescales. A maximally reduced version of MSM targets directly the metastable states and studies exclusively the transition rates between such states [44].

The Mori-Zwanzig theory [45, 46] has emerged as possibly the key statistical mechanical paradigm for extracting accurate coarse-grained models from multiscale systems; see discussion in e.g. [47, 48, 49, 50, 51]. Nonetheless, it has recently become apparent that one might desire to cast the problem of deriving the coarse-grained model not necessarily as an approximate or exact analytical exercise, but rather in purely algorithmic terms [52, 53].

Along these lines, MSM has the great advantage of being a) equation-agnostic: it is a data-driven method that can be deployed on observed or modelling data and is oblivious to the underlying evolution equations; b) constructed in such a way that memory effects are effectively neglected; and c) (possibly) able to beat the curse of dimensionality, because the geometric and dynamical complexity of the original system is bypassed once one is able to define smartly the basis of states associated with the markovian dynamics. Points b) and c) are facilitated if one is able to define reaction coordinates, i.e. collective variables able to provide a low dimensional description of the key macroscopic features of the system. Defining reaction coordinates for general complex systems that are far from equilibrium is highly nontrivial, as symmetry or thermodynamic arguments used for equilibrium dynamics cannot be simply adapted [39, 54, 55]; see discussion in [56, 57]. Applying the Ulam method to a dynamical systems amounts to considering a Koopman dictionary comprising of the characteristic function of all cells of the tessellation. Whilst considering a dictionary of discontinuous functions can cause headaches, the use of characteristic functions has also some clear advantages; see at this regard the recently proposed multiplicative DMD algorithm [58].

Comprehensive and easily implementable response formulas for Markov chains are - apart from their intrinsic interest - of great practical utility in the analysis of a complex system because they can be directly applied to its coarse-grained representation constructed according to MSM protocols and thus bypassing (and obliviously to) the underlying evolution equations.

We discussed elsewhere how the invariant measure of a Markov chain responds to a time-independent perturbation, presenting explicit bounds for the validity of classic perturbative approach and providing explicit formulas for linear and nonlinear terms, including renormalized results [59, 60]. Independent results that delve more deeply in the physical interpretation of the response formulas have been recently reported [61, 62].

Treating carefully the transition from the microscopic description of a system to its heavily coarse-grained representation as a discrete Markov process and linking the properties of fluctuations and response across scales is an extremely challenging task, see [63, 64, 65]. Here we set ourselves in a much simplified setting and we assume that upstream of our work someone has carefully constructed a coarse-grained representation of the system as a discrete Markov chain, e.g. by applying MSM to a system in a reference state and in a slightly perturbed state. Both states are characterised by autonomous dynamics. Our goal is to predict how different time modulations of the forcing impacts the statistical properties of the coarse-grained system. Hence, by construction, we will neglect the subscale processes.

We will derive formulas that are able to predict the linear as well as higher order response of the coarse-grained system for general, time-dependent perturbation via simple matrix relations. Our ability to treat explicitly time-independent perturbation is, as far as we know, novel, and goes in the direction of analysis of entropy production for non autonomous systems [66]. Response formulas can be derived for observables as well as for lagged correlations between observables, thus allowing for predicting how the forcing impacts the variability of the system. The latter had been attempted in a previous work but only in the case of static forcings [67]. We will also provide a simple but possibly very instructive novel way of expressing the linear and nonlinear response operators for Markov chains by taking advantage of the Koopman formalism for finite-state processes that clarifies the roles of the time scales that are intrinsic to the system.

The rest of the paper is structured as follows. The derivation and discussion of response formulas is presented in Sect. 2 for observables in in Sect. 3 for correlations. In order to illustrate some of our findings, we will provide in Sect. 4 a proof-of-concept application of some of our key results on a simple yet instructive two-dimensional (2D) Langevin equation closely related to an example provided in [68] which is characterized by nontrivial metastability properties. In Sect. 5 we present a discussion of our results, including a comment on their implications for providing a dynamical foundation for the time-series analysis Prony method [69, 70, 71, 72], as well as perspectives for future investigations. Additionally App. A, App. B, and App. § 8 provide general formulas for the nonlinear response of observables, the linear response of correlation functions, and the linear response of entropy production to time-dependent perturbations, respectively.

Response Theory for Markov Chains: Time-Dependent Perturbations

We set ourselves in the same framework described in [59, 60]. Let us consider a mixing NN-state (NN is finite) Markov process defined by the matrix M∈RN×NM \in\mathbb{R}^{N \times N}, Mij≥0M_{ij} \ge0 is a stochastic matrix that measures the probability of reaching the state ii at time nn given that at time n−1n - 1 the system is in the state jj. Since the process is mixing we can reach any state ii starting from any state jj is we wait a sufficiently long time, or, more specifically ∃p≥1∣Mijp>0\exists p \ge1|M^p_{ij} > 0. We consider the eigenvalue problem Mv=λvM v = \lambda v. For the Perron–Frobenius theorem, there is a unique invariant measure, i.e. ∃!νinv∣Mνinv=νinv\exists! \nu_{\mathrm{inv}}|M\nu_{\mathrm{inv}} = \nu_{\mathrm{inv}}, so that νinv∈RN×1\nu_{\mathrm{inv}} \in\mathbb{R}^{N \times1} defines the invariant measure associated with unitary eigenvalue [73]. We also have pairs {λj,vj}\{\lambda_j, v_j\}, such that Mvj=λjvjM v_j = \lambda_j v_j, with ∣λj∣<1|\lambda_j| < 1 and vj∈RN×1v_j \in\mathbb{R}^{N \times1} for j=2,…,Nj = 2,\ldots,N. Customarily, we label the eigenvalues in such a way that they are ordered by their magnitude, i.e. ∣λi∣≥∣λj∣|\lambda_i| \ge|\lambda_j| if i<ji < j. Additionally, we have that ∑i=1N(νinv)i=1\sum_{i=1}^{N}(\nu_{inv})_i = 1 and ∑i=1N(νj)i=0\sum_{i=1}^{N}(\nu_j)_i = 0, j>1j > 1.

Let us now consider a perturbation of the form M→Mϵ,n=M+ϵf(n)m\mathcal{M} \to\mathcal{M}_{\epsilon,n} = \mathcal{M} + \epsilon f(n)m, where f:N→Rf:\mathbb{N} \to\mathbb{R} defines a time-dependent modulation with ∣f(n)∣<1|f(n)| < 1, m∈RN×Nm \in\mathbb{R}^{N \times N}, and ϵ\epsilon is a small real number. We impose that Mϵ,n\mathcal{M}_{\epsilon,n} is at all times a stochastic matrix. Hence, ∑imij=0\sum_i m_{ij} = 0. The perturbed Markov chain evolves according to the following law:

ν(n)=Mϵ,nν(n−1)=(M+ϵf(n−1)m)ν(n−1)(1)\nu(n) = \mathcal{M}_{\epsilon,n}\nu(n-1) = \left(\mathcal{M} + \epsilon f(n-1)m\right)\nu(n-1) \tag*{(1)}

We plug ν(n)=νinv+ϵν(1)(n)+h.o.t.\nu(n) = \nu_{inv} + \epsilon\nu^{(1)}(n) + h.o.t. in the equation above and collect the terms proportional to ϵ\epsilon. The conditions behind the applicability of the perturbative approach are discussed in detail in [59, 60] and will not be repeated here; see also [74] and recent review devoted to continuous time Markov chains [75]. It suffices here to say that if the operator M\mathcal{M} has a finite spectral gap, we can define an ϵmax>0\epsilon_{max} > 0 such that ∀ϵ\forall\epsilon with ∣ϵ∣<ϵmax|\epsilon| < \epsilon_{max} the perturbative expansions converge. We obtain:

ν(1)(n)=Mν(1)(n−1)+f(n−1)mνinv(2)\nu^{(1)}(n) = \mathcal{M}\nu^{(1)}(n-1) + f(n-1)m\nu_{inv} \tag*{(2)}

By applying recursively the relationship above and considering that lim⁡n→∞∣Mnν(1)(n)∣=0\lim_{n \to\infty}|\mathcal{M}^{n}\nu^{(1)}(n)| = 0, we have:

ν(1)(n)=∑k=0∞Mkmνinvf(n−k−1)(3)\nu^{(1)}(n) = \sum_{k=0}^{\infty} \mathcal{M}^{k}m\nu_{inv}f(n-k-1) \tag*{(3)}
=∑k=−∞∞Θ(k)Mkmνinvf(n−k−1)(4)= \sum_{k=-\infty}^{\infty} \Theta(k)\mathcal{M}^{k}m\nu_{inv}f(n-k-1) \tag*{(4)}

where Θ(k)=1\Theta(k) = 1 if k≥0k \ge0 and Θ(k)=0\Theta(k) = 0 if k<0k < 0.

(a) Linear Response

Let us define ⟨Ψ,μ⟩=∑i=1NΨiμi\langle\Psi,\mu\rangle= \sum_{i=1}^{N}\Psi_i\mu_i, the expectation value of an observable Ψ∈R1×N\Psi\in\mathbb{R}^{1 \times N} according to the measure μ\mu. We then have ⟨Ψ,ν(n)⟩=⟨Ψ,νinv⟩+ϵ⟨Ψ,ν(1)(n)⟩+h.o.t.\langle\Psi,\nu(n)\rangle= \langle\Psi,\nu_{inv}\rangle+ \epsilon\langle\Psi,\nu^{(1)}(n)\rangle+ h.o.t.. We then have:

d⟨Ψ,ν(n)⟩dϵ∣ϵ=0=⟨Ψ,ν(1)(n)⟩=∑k=−∞∞Θ(k)⟨mT(MT)kΨ,νinv⟩f(n−k−1)=(Gm,Ψ(1)⋆f)(n),Gm,Ψ(1)(k)=Θ(k)⟨mT(MT)kΨ,νinv⟩(5)\begin{aligned} \left.\frac{\mathrm{d}\langle\Psi,\nu(n)\rangle}{\mathrm{d}\epsilon}\right|_{\epsilon=0} = \langle\Psi,\nu^{(1)}(n)\rangle\\ &= \sum_{k=-\infty}^{\infty}\Theta(k)\langle m^{T}(\mathcal{M}^{T})^{k}\Psi,\nu_{inv}\rangle f(n-k-1) \\ &= (G^{(1)}_{m,\Psi} \star f)(n), \qquad G^{(1)}_{m,\Psi}(k) = \Theta(k)\langle m^{T}(\mathcal{M}^{T})^{k}\Psi,\nu_{inv}\rangle \tag*{(5)} \end{aligned}

where Gm,Ψ(1)(k)G^{(1)}_{m,\Psi}(k) is the (causal) first order Green’s function and ⋆\star indicates the convolution product. We define the Koopman operator K=MT\mathcal{K} = \mathcal{M}^{T}. We assume the absence of degeneracies and we define Λ∈RN×N=diag⁡(λ1,…,λN)\Lambda\in\mathbb{R}^{N \times N} = \operatorname{diag}(\lambda_1,\ldots,\lambda_N). Hence, K=VΛV−1\mathcal{K} = V\Lambda V^{-1}, with V∈RN×NV \in\mathbb{R}^{N \times N}. We have that

Km=∑i=1NλimviwiT=∑i=1NλimΠi(6)\mathcal{K}^{m} = \sum_{i=1}^{N}\lambda_i^{m}v_iw_i^{T} = \sum_{i=1}^{N}\lambda_i^{m}\Pi_i \tag*{(6)}

where viv_i is the ithi^{\mathrm{th}} right eigenvector, wiw_i is the ithi^{\mathrm{th}} left eigenvector of K\mathcal{K} and Πi\Pi_i defines the projector on the ithi^{\mathrm{th}} eigenmode of K\mathcal{K}. 1 By inserting the previous expression in the definition of the Green’s function we obtain:

Gm,Ψ(1)(k)=Θ(k)⟨mT∑i=2NλikviwiTΨ,νinv⟩=Θ(k)⟨mT∑i=2NλikΠiΨ,νinv⟩=∑i=1NGm,Ψ,i(1)(k)=∑i=2NGm,Ψ,i(1)(k)(7)\begin{aligned} \mathcal{G}^{(1)}_{m,\Psi}(k)=\Theta(k)\left\langle m^{T}\sum_{i=2}^{N}\lambda_i^{k}v_iw_i^{T}\Psi,\nu_{\mathrm{inv}}\right\rangle\\ &=\Theta(k)\left\langle m^{T}\sum_{i=2}^{N}\lambda_i^{k}\Pi_i\Psi,\nu_{\mathrm{inv}}\right\rangle=\sum_{i=1}^{N}\mathcal{G}^{(1)}_{m,\Psi,i}(k)=\sum_{i=2}^{N}\mathcal{G}^{(1)}_{m,\Psi,i}(k) \tag*{(7)} \end{aligned}

where

Gm,Ψ,i(1)(k)=Θ(k)αiλikαi=⟨mTΠiΨ,νinv⟩(8)\mathcal{G}^{(1)}_{m,\Psi,i}(k)=\Theta(k)\alpha_i\lambda_i^{k}\qquad\alpha_i=\left\langle m^{T}\Pi_i\Psi,\nu_{\mathrm{inv}}\right\rangle \tag*{(8)}

where we have broken down the Green’s function into the sum of N−1N-1 terms, each associated with a specific mode of variability of the system. Note that the first term i=1i=1 in the summation given in Eq. (7) vanishes because it can be proved that Π1Tm=0\Pi_1^{T}m=0 [59]. Since λik=exp⁡(kβi)\lambda_i^{k}=\exp(k\beta_i), with ℜ[βi]<0\Re[\beta_i]<0, the previous expansion provides a specific statistical model - the sum of exponentials - for fitting a Green’s function from data. We will comment on this matter in Sect. 5. The results presented in Eqs. (7)-(8) correspond, in the case of a finite-state Markov chain, to the key findings shown in [17] for a general Langevin equation and a general Koopman dictionary. The derivation is much simpler in the case presented here whilst very little is lost at conceptual level, as we are indeed considering, as discussed in Subsect. 1(a), the case where the dictionary is given by the characteristic functions of the elements of the tessellation in phase space defining the states of Markov chain.

Finally, by inserting the previous expression in the linear response formula above and by rearranging terms, we have:

d⟨Ψ,ν(n)⟩dϵ∣ϵ=0=∑i=1N(Gm,Ψ,i(1)⋆f)(n),(9)\left.\frac{\mathrm{d}\langle\Psi,\nu(n)\rangle}{\mathrm{d}\epsilon}\right|_{\epsilon=0}=\sum_{i=1}^{N}\left(\mathcal{G}^{(1)}_{m,\Psi,i}\star f\right)(n), \tag*{(9)}

which separates the linear response formula into NN distinct contributions. Each of this contributions can be computed from the knowledge of mm, M\mathcal{M}, Ψ\Psi, and ff.

(b) Second order response

Let’s now consider the full perturbative expansion ν(n)=νinv+∑l=1∞ϵlν(l)(n)\nu(n)=\nu_{\mathrm{inv}}+\sum_{l=1}^{\infty}\epsilon^{l}\nu^{(l)}(n). By equating terms proportional to powers of ϵ\epsilon larger than one, we obtain::

ν(l)(n)=Mν(l)(n−1)+f(n−1)mν(l−1)(n),l>1(10)\nu^{(l)}(n)=\mathcal{M}\nu^{(l)}(n-1)+f(n-1)m\nu^{(l-1)}(n),\qquad l>1 \tag*{(10)}

By applying recursively over times the relationship above and considering that lim⁡n→∞∣Mnν(l)(n)∣=0\lim_{n\to\infty}\lvert\mathcal{M}^{n}\nu^{(l)}(n)\rvert=0, we have:

ν(l)(n)=∑k=0∞Mkmν(l−1)(n−k−1)f(n−k−1)(11)\nu^{(l)}(n)=\sum_{k=0}^{\infty}\mathcal{M}^{k}m\nu^{(l-1)}(n-k-1)f(n-k-1) \tag*{(11)}
=∑k=−∞∞Θ(k)Mkmν(l−1)(n−k−1)f(n−k−1)(12)=\sum_{k=-\infty}^{\infty}\Theta(k)\mathcal{M}^{k}m\nu^{(l-1)}(n-k-1)f(n-k-1) \tag*{(12)}

Let’s now consider the second order term ν(2)(n)\nu^{(2)}(n). We have

ν(2)(n)=∑k=0∞Mkmν(1)(n−k−1)f(n−k−1)(13)\nu^{(2)}(n)=\sum_{k=0}^{\infty}\mathcal{M}^{k}m\nu^{(1)}(n-k-1)f(n-k-1) \tag*{(13)}
=∑k=−∞∞∑p=−∞∞Θ(k)Θ(p)MkmMpmνinvf(n−k−p−2)f(n−k−1)(14)=\sum_{k=-\infty}^{\infty}\sum_{p=-\infty}^{\infty}\Theta(k)\Theta(p)\mathcal{M}^{k}m\mathcal{M}^{p}m\nu_{\mathrm{inv}}f(n-k-p-2)f(n-k-1) \tag*{(14)}

where we have used the expression for ν(1)\nu^{(1)} above. Note that the Θ\Theta’s ensure a correctly time-ordered consideration of the acting perturbation.

Let’s now consider the second order response of a generic observable Ψ\Psi. We have:

12d2⟨Ψ,ν(n)⟩dϵ2∣ϵ=0=⟨Ψ,ν(2)(n)⟩(15)\frac{1}{2}\left.\frac{\mathrm{d}^{2}\langle\Psi,\nu(n)\rangle}{\mathrm{d}\epsilon^{2}}\right|_{\epsilon=0}=\langle\Psi,\nu^{(2)}(n)\rangle \tag*{(15)}
=∑k=−∞∞Θ(k)⟨mT(MT)kΨ,ν(1)(n−k−1)⟩f(n−k−1)(16)=\sum_{k=-\infty}^{\infty}\Theta(k)\langle m^{T}(\mathcal{M}^{T})^{k}\Psi,\nu^{(1)}(n-k-1)\rangle f(n-k-1) \tag*{(16)}
=∑k=−∞∞∑p=−∞∞Θ(k)Θ(p)⟨mT(MT)pmT(MT)kΨ,νinv⟩f(n−k−p−2)f(n−k−1)(17)=\sum_{k=-\infty}^{\infty}\sum_{p=-\infty}^{\infty}\Theta(k)\Theta(p)\langle m^{T}(\mathcal{M}^{T})^{p}m^{T}(\mathcal{M}^{T})^{k}\Psi,\nu_{\mathrm{inv}}\rangle f(n-k-p-2)f(n-k-1) \tag*{(17)}
=(Gm,Ψ(2)⋆f)(n)(18)=(\mathcal{G}^{(2)}_{m,\Psi}\star f)(n) \tag*{(18)}

where Gm,Ψ(2)(k,p)=Θ(k)Θ(p)⟨mT(MT)pmT(MT)kΨ,νinv⟩\mathcal{G}^{(2)}_{m,\Psi}(k,p)=\Theta(k)\Theta(p)\langle m^{T}(\mathcal{M}^{T})^{p}m^{T}(\mathcal{M}^{T})^{k}\Psi,\nu_{\mathrm{inv}}\rangle and where ⋆\star indicates here a double convolution sum. By using Km=(MT)m=∑i=1NλimΠi\mathcal{K}^{m}=(\mathcal{M}^{T})^{m}=\sum_{i=1}^{N}\lambda_{i}^{m}\Pi_{i} in the expression of the second order Green’s function above, one can express it as a double sum of terms, each describing the contribution to the nonlinear response coming from a specific pair of Koopman modes. Indeed, we have:

Gm,Ψ(2)(k,p)=Θ(k)Θ(p)∑i,j=2Nαijλikλjp,αij=⟨mTΠjmTΠiΨ,νinv⟩(19)\mathcal{G}^{(2)}_{m,\Psi}(k,p)=\Theta(k)\Theta(p)\sum_{i,j=2}^{N}\alpha_{ij}\lambda_{i}^{k}\lambda_{j}^{p},\qquad\alpha_{ij}=\langle m^{T}\Pi_{j}m^{T}\Pi_{i}\Psi,\nu_{\mathrm{inv}}\rangle \tag*{(19)}

Note again that our summation excludes the term corresponding to the invariant measure of the system. The results above can easily be extended at all orders of perturbations, see Appendix A. This implies that the full nonlinear time-dependent response can be obtained from the knowledge of mm, M\mathcal{M}, and ff for any observable of the system. Assuming f=1f=1 and taking the n→∞n\to\infty limit, the results presented in [59] can be easily recovered.

Response theory for Correlations

In the vast majority of cases, response theory has been used to compute the change of the measure of a system resulting from applied forcings. Yet, in many practical cases, it is relevant to study the impact of the perturbation on the temporal correlation properties of the system. Taking the example of climate science to illustrate this point, response theory applied to observables describes the change in the state of the climate at a certain time horizon with respect to a reference climatology, whilst response theory applied to correlation describes how climate variability and so-called teleconnections like the North Atlantic Oscillation or El-Niño-Southern Oscillation are impacted by the applied perturbation [76, 77]. Some preliminary contributions to the development of response formulas for correlations have been presented in [67], which provides a reference for the results presented below.

Let us define Cl(Ψ,Φ)=⟨(MT)lΨ∘Φ,νinv⟩−⟨Ψ,νinv⟩⟨Φ,νinv⟩C_l(\Psi,\Phi)=\langle(\mathcal{M}^{T})^{l}\Psi\circ\Phi,\nu_{\mathrm{inv}}\rangle-\langle\Psi,\nu_{\mathrm{inv}}\rangle\langle\Phi,\nu_{\mathrm{inv}}\rangle the unperturbed ll-lagged correlation between the two observables Ψ\Psi and Φ\Phi, where ∘\circ indicates the Hadamard product. We first consider the case of a static perturbation to the Markov process of the form M→Mϵ=M+ϵm\mathcal{M}\to\mathcal{M}_{\epsilon}=\mathcal{M}+\epsilon m, which has not yet been explicitly treated in the literature up to our knowledge. We then have Clϵ(Ψ,Φ)=⟨(MT+ϵmT)lΨ∘Φ,νϵ⟩−⟨Ψ,νϵ⟩⟨Φ,νϵ⟩C_{l}^{\epsilon}(\Psi,\Phi)=\langle(\mathcal{M}^{T}+\epsilon m^{T})^{l}\Psi\circ\Phi,\nu_{\epsilon}\rangle-\langle\Psi,\nu_{\epsilon}\rangle\langle\Phi,\nu_{\epsilon}\rangle. Now we write down the terms up to first order in ϵ\epsilon in the expression of the correlation. For the first term, we obtain:

⟨(MT+ϵmT)lΨ∘Φ,νinv+∑p=1∞ϵpν(p)⟩=⟨(MT)lΨ∘Φ,νinv⟩+ϵ∑q=0l−1⟨(Ml−q−1)TmT(Mq)TΨ∘Φ,νinv⟩+ϵ∑k=0∞⟨(MT)lΨ∘Φ,Mkmνinv⟩+o(ϵ)(20)\begin{aligned} \left\langle(\mathcal{M}^{T}+\epsilon m^{T})^{l}\Psi\circ\Phi,\nu_{\mathrm{inv}}+\sum_{p=1}^{\infty}\epsilon^{p}\nu^{(p)}\right\rangle =\left\langle(\mathcal{M}^{T})^{l}\Psi\circ\Phi,\nu_{\mathrm{inv}}\right\rangle\\ &+\epsilon\sum_{q=0}^{l-1}\left\langle(\mathcal{M}^{l-q-1})^{T}m^{T}(\mathcal{M}^{q})^{T}\Psi\circ\Phi,\nu_{\mathrm{inv}}\right\rangle+\epsilon\sum_{k=0}^{\infty}\left\langle(\mathcal{M}^{T})^{l}\Psi\circ\Phi,\mathcal{M}^{k}m\nu_{\mathrm{inv}}\right\rangle+o(\epsilon) \tag*{(20)} \end{aligned}

For the second term, we have

⟨Ψ,νinv+∑p=1∞ϵpν(p)⟩⟨Φ,νinv+∑p=1∞ϵpν(p)⟩=⟨Ψ,νinv⟩⟨Φ,νinv⟩+ϵ∑k=0∞⟨Ψ,Mkminv⟩⟨Φ,νinv⟩+ϵ⟨Ψ,νinv⟩∑k=0∞⟨Φ,Mkminv⟩+o(ϵ).(21)\begin{aligned} \left\langle\Psi,\nu_{inv}+\sum_{p=1}^{\infty}\epsilon^{p}\nu^{(p)}\right\rangle\left\langle\Phi,\nu_{inv}+\sum_{p=1}^{\infty}\epsilon^{p}\nu^{(p)}\right\rangle=\langle\Psi,\nu_{inv}\rangle\langle\Phi,\nu_{inv}\rangle\\ &+\epsilon\sum_{k=0}^{\infty}\langle\Psi,\mathcal{M}^{k}m_{inv}\rangle\langle\Phi,\nu_{inv}\rangle+\epsilon\langle\Psi,\nu_{inv}\rangle\sum_{k=0}^{\infty}\langle\Phi,\mathcal{M}^{k}m_{inv}\rangle+o(\epsilon). \tag*{(21)} \end{aligned}

Keeping in mind that ∑k=0∞Mk=(1−M+Q1)−1=Z\sum_{k=0}^{\infty}\mathcal{M}^{k}=(1-\mathcal{M}+Q_{1})^{-1}=\mathcal{Z}, we obtain:

dClϵ(Ψ,Φ)dϵ∣ϵ=0=∑q=0l−1⟨(Ml−q−1)TmT(Mq)TΨ∘Φ,νinv⟩+⟨mTZT(MT)Ψ∘Φ,νinv⟩−⟨Φ,νinv⟩⟨mTZTΨ,νinv⟩−⟨Ψ,νinv⟩⟨mTZTΦ,νinv⟩(22)\left.\frac{\mathrm{d}C_{l}^{\epsilon}(\Psi,\Phi)}{\mathrm{d}\epsilon}\right|_{\epsilon=0}=\sum_{q=0}^{l-1}\left\langle(\mathcal{M}^{l-q-1})^{T}m^{T}(\mathcal{M}^{q})^{T}\Psi\circ\Phi,\nu_{inv}\right\rangle+\left\langle m^{T}\mathcal{Z}^{T}(\mathcal{M}^{T})\Psi\circ\Phi,\nu_{inv}\right\rangle-\langle\Phi,\nu_{inv}\rangle\left\langle m^{T}\mathcal{Z}^{T}\Psi,\nu_{inv}\right\rangle-\langle\Psi,\nu_{inv}\rangle\left\langle m^{T}\mathcal{Z}^{T}\Phi,\nu_{inv}\right\rangle \tag*{(22)}

Hence, the sensitivity of a time lagged correlation between Ψ\Psi and Φ\Phi is nontrivial and can be broken up into four terms. The first term δa,ϵ(1)(Ψ(l),Φ)\delta_{a,\epsilon}^{(1)}(\Psi(l),\Phi) is associated with the impact of perturbation on the evolution law of the system in the time interval of length ll. The second term δb,ϵ(1)(Ψ(l),Φ)\delta_{b,\epsilon}^{(1)}(\Psi(l),\Phi) describes the change in the expectation value of the product Ψ(l),Φ\Psi(l),\Phi due to the variation of the invariant measure. The last two terms associated with the change in the expectation value of the two observables. Note that the matrix expression provided here is much simpler than the corresponding functional expression given in [67]. We remark that by performing the spectral expansion of the MT\mathcal{M}^{T} matrix we are able to disentangle the contributions coming from the various modes of the Koopman operator.

It is indeed possible to extend the response theory for correlation to the case where ff has a non-trivial time dependence. The results are reported in Appendix B. It is important to note that the resulting formulas allow us to define how correlations behave in a non-autonomous system. In this case, correlations at time tt need to be interpreted as integrals performed on the measure of the snapshot attractor at time tt (which is a slice of the pullback attractor [78, 79]); see discussion in [80, 81, 82].

A Simple Example

We wish to provide here a proof of concept to test the usefulness of the framework proposed earlier. We consider the following 2-dimensional (2D) Langevin equation:

dx=Fx(x,y)+σdW1=−∂xV(x,y)+σdW1(23)\mathrm{d}x=F_{x}(x,y)+\sigma\mathrm{d}W_{1}=-\partial_{x}V(x,y)+\sigma\mathrm{d}W_{1} \tag*{(23)}
dy=Fy(x,y)+σdW2=−∂yV(x,y)+σdW2(24)\mathrm{d}y=F_{y}(x,y)+\sigma\mathrm{d}W_{2}=-\partial_{y}V(x,y)+\sigma\mathrm{d}W_{2} \tag*{(24)}

where dW1\mathrm{d}W_{1} and dW2\mathrm{d}W_{2} are increments of independent Wiener processes, and the drift is defined as minus the gradient of the following potential:

V(x)=V(x,y)=3exp⁡(−x2−(y−1/3)2)−3exp⁡(−x2−(y−5/3)2)V(\mathbf{x})=V(x,y)=3\exp\left(-x^{2}-(y-1/3)^{2}\right)-3\exp\left(-x^{2}-(y-5/3)^{2}\right)
−5exp⁡(−(x+1)2−y2)−5exp⁡(−(x−1)2−y2)+1/5x4+1/5(y−1/3)4−y.(25)\begin{aligned} -5\exp\left(-(x+1)^{2}-y^{2}\right)-5\exp\left(-(x-1)^{2}-y^{2}\right) \\ &+1/5x^{4}+1/5(y-1/3)^{4}-y. \tag*{(25)} \end{aligned}

Given the choice of the drift and of the noise law, we are considering here an equilibrium system obeying detailed balance [7]. The potential is depicted in Figure 1a and features three local minima, located at (x1,y1)≈(0,1.55)(x_{1},y_{1})\approx(0,1.55), (x2,y2)=(−1,0)(x_{2},y_{2})=(-1,0), and (x3,y3)=(1,0)(x_{3},y_{3})=(1,0). The potential we consider here corresponds almost exactly to a case study presented in [68]. We have included an additional term −y-y in the definition of the potential in order to reduce the value of VV at the local minimum in (x1,y1)(x_1,y_1) to a value closer to the absolute minimum that is realized at (x2,y2)(x_2,y_2) and (x3,y3)(x_3,y_3). The Fokker-Planck equation associated with the previous Langevin equation reads as [7, 83]:

Six contour plots showing the potential function, invariant measure, and subdominant Koopman and Perron–Frobenius modes

Figure 1. (a) Potential function V(x,y)V(x,y) with approximate indication of the three quasi-invariant regions surrounding the minima of the VV. (b) Invariant Measure ρ0∝exp⁡(−2V(x,y)/σ2)\rho_0 \propto\exp(-2V(x,y)/\sigma^2). (c). First subdominant Koopman mode, λ1=0.9916\lambda_1 = 0.9916. (d) First subdominant mode of the Perron-Frobenius Operator. (e) Second subdominant Koopman model λ2=0.9655\lambda_2 = 0.9655. (f) Second subdominant mode of the Perron-Frobenius Operator.

∂tρ(x,t)=L0ρ(x,t)=∂x(∂xV(x)ρ(x,t))+∂y(∂yV(x)ρ(x,t))+σ22(∂x2+∂y2)ρ(x,t)(26)\partial_t\rho(x,t) = \mathcal{L}_0\rho(x,t) = \partial_x(\partial_xV(x)\rho(x,t)) + \partial_y(\partial_yV(x)\rho(x,t)) + \frac{\sigma^2}{2}(\partial_x^2+\partial_y^2)\rho(x,t) \tag*{(26)}

The solution can be expressed as ρ(x,t)=exp⁡(L0t)ρin(x)\rho(x,t) = \exp(\mathcal{L}_0t)\rho_{\mathrm{in}}(x) where ρin(x)\rho_{\mathrm{in}}(x) is the initial condition and exp⁡(L0t)\exp(\mathcal{L}_0t) is the strongly continuous semigroup generated by L0\mathcal{L}_0, which is densely defined in a suitable Banach space [7, 16]. The Perron-Frobenius operator Pt,0=exp⁡(L0t)P^{t,0} = \exp(\mathcal{L}_0t) describes the evolution of measures for a time tt. Since we have a confining potential for the drift and non-degenerate noise, we have a unique invariant measure that is absolutely continuous with respect to Lebesgue, so that it can be written as ρ0dx\rho_0dx, where L0ρ0=0\mathcal{L}_0\rho_0 = 0 [7, 84].

We obtain ρ0(x)=Z−1exp⁡(−2V(x)/σ2)\rho_0(\mathbf{x}) = Z^{-1}\exp(-2V(\mathbf{x})/\sigma^2), where ZZ is the normalisation factor ensuring that ∫dxρ0(x)=1\int\mathrm{d}\mathbf{x}\rho_0(\mathbf{x}) = 1. As a result, the natural functional space for our problem is given by Lρ02L^2_{\rho_0}. The Koopman operator Kt,0K^{t,0} on measurable observables and is defined as the adjoint of the Perron-Frobenius operator: ∫Kt,0Ψ(x)ρin(x)dx=def∫dxΨ(x)Pt,0ρin(x)\int K^{t,0}\Psi(\mathbf{x})\rho_{\mathrm{in}}(\mathbf{x})\mathrm{d}\mathbf{x} \mathrel{\overset{\mathrm{def}}{=}} \int\mathrm{d}\mathbf{x}\Psi(\mathbf{x})P^{t,0}\rho_{\mathrm{in}}(\mathbf{x}), where Kt,0=(Pt,0)T=exp⁡(L0Tt)K^{t,0} = (P^{t,0})^{T} = \exp(\mathcal{L}^{T}_{0}t) and L0TΨ(x)=−∂xV(x)∂xΨ(x)−∂yV(x)∂yΨ(x)+σ22(∂x2+∂y2)Ψ(x)\mathcal{L}^{T}_{0}\Psi(\mathbf{x}) = -\partial_xV(\mathbf{x})\partial_x\Psi(\mathbf{x})-\partial_yV(\mathbf{x})\partial_y\Psi(\mathbf{x})+\frac{\sigma^2}{2}(\partial_x^2+\partial_y^2)\Psi(\mathbf{x}).

In what follows, we assume σ=0.7\sigma= 0.7. Our numerical simulations are performed using the standard Euler-Maruyama scheme [85] with dt=0.05\mathrm{d}t = 0.05. We sample our output every 20 time steps, so that our reference time scale is τ=1\tau= 1. The dynamics given by Eqs. (23)–(26) describes an equilibrium system.

To approximate Pτ,0P^{\tau,0}, we discretize the phase space of the system into NN disjoint subsets {B1,B2,…,BN}\{B_1,B_2,\ldots,B_N\}, forming a partition of the space. The Ulam transfer operator P{N},ijτ,0P^{\tau,0}_{\{N\},ij}, a finite-dimensional stochastic matrix describes the probability that the orbit of the system is at time t+τt+\tau in the subset BiB_i at time tt it was in the subset BjB_j. In our case, the BjB_j’s are given by the 625 squares with side 0.2 centred around the origin. Whilst in principle one would need to cover the entire R2\mathbb{R}^2 the confining potential makes all regions beyond those we consider here entirely irrelevant unless one consider extremely long time scales.

Following [68], we populate each cube with 1000 ensemble members distributed uniformly according to the Lebesgue measure. Each member evolves for τ=1\tau= 1 time unit. We then construct an estimate of P{N}τ,0P^{\tau,0}_{\{N\}} by counting the transitions. We repeat the operation 20 times and by averaging we obtain our best estimate of P{N}τ,0P^{\tau,0}_{\{N\}}, which constitutes our reference discretized stochastic matrix, so that M=P{N}τ,0\mathcal{M}=P^{\tau,0}_{\{N\}}. The solution to Mνinv=νinv\mathcal{M}\nu_{\mathrm{inv}}=\nu_{\mathrm{inv}} gives an accurate approximation in the gridded domain defined by the BjB_j’s of the true invariant measure ρ0\rho_0.

It is possible to filter out the dynamics associated with spurious non-equilibrium modes due to the finite sampling by proceeding as follows. First, we compute the matrix Nij=Mjiνinv,iνinv,jN_{ij}=\mathcal{M}_{ji}\frac{\nu_{\mathrm{inv},i}}{\nu_{\mathrm{inv},j}}. Then we update our estimate of M\mathcal{M} as follows: M→M+N2\mathcal{M}\to\frac{\mathcal{M}+N}{2}, so that detailed balance is enforced in the finite-state representation of the system.2

Whilst the first eigenvector of M\mathcal{M} is shown in Figure 1b, the first right eigenvector of the Koopman operator MT\mathcal{M}^{T}, is constant everywhere. We obtain 624 additional eigenvalue-eigenvector pairs for the matrix M\mathcal{M}. The two subdominant eigenvectors are depicted in Figure 1d and Figure 1f. The first one describes the transitions between the neighbourhoods of (x2,y2)(x_2,y_2) and (x3,y3)(x_3,y_3). The second one describes the transitions between the neighbourhoods of (x1,y1)(x_1,y_1) and either (x2,y2)(x_2,y_2) or (x3,y3)(x_3,y_3). The corresponding subdominant eigenvectors of the Koopman operator are depicted in Figure 1c and Figure 1e. There is a very large spectral gap between the three dominant modes (corresponding to λ=1\lambda=1, λ≈0.9917(1)\lambda\approx0.9917(1), λ≈0.9655(1)\lambda\approx0.9655(1)) the rest of the spectral components (λ4≈0.050(1)\lambda_4\approx0.050(1)), which indicates that the system can safely undergo a model reduction procedure. This could be algebraically achieved by substituting K→∑i=13λiΠi\mathcal{K}\to\sum_{i=1}^{3}\lambda_i\Pi_i, and M→∑i=13λiΠiT\mathcal{M}\to\sum_{i=1}^{3}\lambda_i\Pi_i^{T}. Additionally, the level sets of the two subdominant Koopman modes can be used to identify the three metastable regimes of the system, which are indicated as (1), (2) and (3) in Figure 1a and correspond to the basins of attraction of the three local minima in the case the stochastic forcing is switched off. Performing k-means clustering [37, 38] of the data choosing 3 as a number of clusters leads to the same partition of the data space, whereby the three regions (1), (2), (3) are the Voronoï cells [36] associated with the cluster centers.

We use such a geometrical partition of the phase space to introduce a separate reduced-order representation of the dynamics, whereby the phase space of the system is partitioned into three states, corresponding to the regions (1), (2), and (3). In the simple case described here, this corresponds to MSM. This amounts to neglecting entirely the intrawell dynamics, which, in the Ulam description above, is mostly captured by the Koopman modes with index ≥4\geq4. We estimate the reduced Markov model (RMM) discretized transfer operator MM by performing a single run lasting 10710^7 time units (after disregarding a small transient) and counting the transitions between the 3 states described above. We then enforce detailed balance as above. The obtained estimates for MM and its eigenvectors are given below, where all the numbers indicated below have an approximate uncertainty of 1 in the last digit we have written out.

M~=(0.97010.00950.00950.01490.99040.000060.01490.000060.9904)νinv=(0.2400.3800.380)ν2=(0.0000.701−0.701)ν3=(−0.8100.4050.405)\widetilde{M} = \begin{pmatrix} 0.9701 & 0.0095 & 0.0095 \\ 0.0149 & 0.9904 & 0.00006 \\ 0.0149 & 0.00006 & 0.9904 \end{pmatrix} \qquad \nu_{\mathrm{inv}} = \begin{pmatrix} 0.240 \\ 0.380 \\ 0.380 \end{pmatrix} \qquad \nu_2 = \begin{pmatrix} 0.000 \\ 0.701 \\ -0.701 \end{pmatrix} \qquad \nu_3 = \begin{pmatrix} -0.810 \\ 0.405 \\ 0.405 \end{pmatrix}

The two nontrivial eigenvalues λ2=0.9904(1)\lambda_2 = 0.9904(1) and λ3=0.9606(1)\lambda_3 = 0.9606(1) correspond closely to the first two subdominant eigenvalues obtained using the Ulam method, and the corresponding eigenvectors provide a coarse grained version of the figures provided in Figure 1d and Figure 1f.

(a) Applying Extra Forcings

We now consider the following perturbation to the drift term: (Fx(x,y),Fy(x,y))→(Fx(x,y)+ϵ1f1(t),Fy(x,y)+ϵ2f2(t))(F_x(x,y), F_y(x,y)) \to(F_x(x,y) + \epsilon_1 f_1(t), F_y(x,y) + \epsilon_2 f_2(t)), where f1(t)f_1(t) and f2(t)f_2(t) give the time modulation of the forcing with ∣f1(t)∣,∣f2(t)∣≤1|f_1(t)|, |f_2(t)| \le1 and ϵ1,ϵ2\epsilon_1,\epsilon_2 are the bookkeeping parameters controlling the intensity of the applied perturbation. We set ourselves in the regime of linear response, so that it is natural to assume ∣ϵ1∣,∣ϵ2∣≪1|\epsilon_1|, |\epsilon_2| \ll1. The time-dependent expectation value of a general observable Ψ(x,y)\Psi(x,y) can be written as

⟨Ψ⟩(t)=⟨Ψ⟩0+ϵ1∫−∞∞dt1Gx,Ψ(1)(t−t1)f1(t1)+ϵ2∫−∞∞dt1Gy,Ψ(1)(t−t1)f2(t1)+h.o.t.(27)\langle\Psi\rangle(t) = \langle\Psi\rangle_0 + \epsilon_1 \int_{-\infty}^{\infty} \mathrm{d}t_1 G^{(1)}_{x,\Psi}(t-t_1)f_1(t_1) + \epsilon_2 \int_{-\infty}^{\infty} \mathrm{d}t_1 G^{(1)}_{y,\Psi}(t-t_1)f_2(t_1) + \mathrm{h.o.t.} \tag*{(27)}

where ⟨Ψ⟩0=∫dxdyρ0(x,y)Ψ(x,y)\langle\Psi\rangle_0 = \int\mathrm{d}x\mathrm{d}y \rho_0(x,y)\Psi(x,y) is the expectation value of Ψ\Psi in the unperturbed state, whilst

Gx/y,Ψ(1)=−Θ(t)⟨∂x/ylog⁡(ρ0)Ψ(t)⟩0=2σ2Θ(t)⟨∂x/yVΨ(t)⟩0(28)G^{(1)}_{x/y,\Psi} = -\Theta(t)\langle\partial_{x/y}\log(\rho_0)\Psi(t)\rangle_0 = \frac{2}{\sigma^2}\Theta(t)\langle\partial_{x/y}V\Psi(t)\rangle_0 \tag*{(28)}

are the causal Green’s functions for the observable Ψ\Psi associated with the perturbations acting along the xx and yy directions, respectively. These formulas can be readily derived from the general version of the FDT [6, 17].

It is possible to associate the applied perturbation to the vector field to changes in the discretized Perron-Frobenius operators constructed according to the protocols above. We define mxm_x (mym_y) the perturbation matrix associated with the extra push in the xx (yy) direction, so that

M→M+ϵ1mxf1(n)+ϵ2myf2(n).\mathcal{M} \to\mathcal{M} + \epsilon_1 m_x f_1(n) + \epsilon_2 m_y f_2(n).

i Linear Response - Observables

In order to estimate the matrices mxm_x for the Ulam discretization, we repeat the same protocol considering the perturbed dynamics realised by choosing ϵ1=0.05\epsilon_1 = 0.05 and f=1f = 1. We derive P{N}+P^{+}_{\{N\}}. We repeat the experiment by choosing ϵ1=−0.05\epsilon_1 = -0.05 and f=1f = 1, and derive P{N}−P^{-}_{\{N\}}. We estimate mx=(P{N}+−P{N}−)/(2ϵ1)m_x = (P^{+}_{\{N\}} - P^{-}_{\{N\}})/(2\epsilon_1). Note that using centred differences ensures high precisions when studying linear response [86]. In order to estimate mym_y, we repeat the same procedure described above by considering ϵ2=−0.05\epsilon_2 = -0.05 and ϵ2=−0.05\epsilon_2 = -0.05, respectively.

Similarly, we estimate mxm_x and mym_y for 3-state Markov model by performing long simulations of duration 10710^7 time units with perturbed dynamics, by estimating the Perron-Frobenius operator in each case, and by taking the centred differences. We use the same value ∣ϵ1∣=∣ϵ2∣=0.05|\epsilon_1| = |\epsilon_2| = 0.05. We have verified in all cases that this is accurately within the regime of linearity of the system’s response.

We choose as observables Ψ1=x\Psi_1 = x and Ψ2=y\Psi_2 = y. We first note that because of the symmetry of the system Gx,y(1)(τ)=Gy,x(1)(τ)=0G^{(1)}_{x,y}(\tau) = G^{(1)}_{y,x}(\tau) = 0. We then focus on the case where we force the system in the jj direction and use jj as observable, j=x,yj = x,y. We then estimate Gx/y,x/y(1)G^{(1)}_{x/y,x/y} through formula 4.6 by collecting statistics for 500 independent ensemble runs of the unperturbed system each lasting 10610^6 time units. Note that this is 50 times as many data as those used for constructing the Markov chains.

From the knowledge of mxm_x, mym_y, it is straightforward to compute the Green’s functions for Ψ1=x\Psi_1=x and Ψ2=y\Psi_2=y using Eq. (5).

The estimates we obtain for the Green’s functions of interest are shown in Figure 2. When considering RMM, we clearly see that Gmx,x(1)(τ)∝λ2τG_{m_x,x}^{(1)}(\tau) \propto\lambda_2^\tau and Gmy,y(1)(τ)∝λ3τG_{m_y,y}^{(1)}(\tau) \propto\lambda_3^\tau: as a result of the choice of forcing/observable pair, only one Koopman mode is retained in the expansion given in Eqs. (7)–(8). Indeed, it is clear that if we force along the xx-direction and choose xx as observable, the second Koopman mode (or only the second Perron-Frobenius mode) is the only one that given a nonvanishing contribution. Similarly, the third mode is the only one retained in the spectral expansion of the Green’s function when forcing along yy-direction and choosing yy as observable.

Linear Green’s function for the x and y observables

Figure 2. Linear Green’s function for the xx and yy observables for additive forcing acting on xx (index mxm_x) or yy (index mym_y) direction. Results are shown for the FDT estimate and the estimates obtained using Markov models constructed with Ulam’s discretization and the 3-state RMM. The inset emphasizes the exponential decay of the Green’s functions.

Remarkably, also when considering the much higher complexity Markov chain constructed through the Ulam method, to a very good approximation we have Gmx,x(1)(τ)∝λ2τG_{m_x,x}^{(1)}(\tau) \propto\lambda_2^\tau and Gmy,y(1)(τ)∝λ3τG_{m_y,y}^{(1)}(\tau) \propto\lambda_3^\tau (note that the corresponding λ\lambda’s are slightly larger than in the RMM case). This means that by and large only one of the 625 natural modes of variability of the system matters in defining the response to perturbation, at least in the cases we consider here. This shows that the Koopman decomposition provides the much desired property of interpretability of the response. It is apparent that the estimates of the two Green’s functions obtained using the FDT are relatively noisy, despite the use of a much larger dataset than what has been used for constructing the Gree’s function via Markov chains. A side remark is that, by construction, lim⁡t→0+Gmx,x(1)(t)=lim⁡t→0+Gmy,y(1)(t)=1\lim_{t\to0+}G_{m_x,x}^{(1)}(t)=\lim_{t\to0+}G_{m_y,y}^{(1)}(t)=1. Nonetheless, the function collapses to much lower values within one time unit because of the very rapid decay of correlation due to the rapidly decaying Kolmogorov modes of the continuum system. As soon as we consider t≥1t\ge1, a rather good agreement is found with the estimates obtained via Markov chains.

Linear Response - Correlations

Next we venture into the analysis of correlations and of their response to perturbations for the full system and for its discretized representation via Markov chain. We consider the time-lagged correlations Cτ(x,x)C_\tau(x,x), Cτ(y,x)C_\tau(y,x), and Cτ(y,y)C_\tau(y,y). Our results are shown in Figure 3a)–Figure 3f). Note that since the potential VV of the unperturbed system and the observable xx have opposite symmetry with respect to the exchange x→−xx \to-x, we derive that Cτ(y,x)∣ϵ2=ϵ1=0=0C_\tau(y,x)|_{\epsilon_2=\epsilon_1=0}=0, ∂Cτ(y,x)/∂ϵ2∣ϵ2=ϵ1=0=0\partial C_\tau(y,x)/\partial\epsilon_2|_{\epsilon_2=\epsilon_1=0}=0, ∂Cτ(x,x)/∂ϵ1∣ϵ2=ϵ1=0=0\partial C_\tau(x,x)/\partial\epsilon_1|_{\epsilon_2=\epsilon_1=0}=0, and ∂Cτ(y,y)/∂ϵ1∣ϵ2=ϵ1=0=0\partial C_\tau(y,y)/\partial\epsilon_1|_{\epsilon_2=\epsilon_1=0}=0.

Six-panel plots showing estimates and decompositions of linear response

Figure 3. (a) Estimate of Cτ(x,x)C_{\tau}(x,x) via direct numerical simulations (DNS), RMM, and Ulam method. Red lines: Reference state. Black lines: sensitivity with respect to ϵ2\epsilon_2. (b) Decomposition of the linear response in (a) in the four terms discussed in Eq. (22) (RMM). (c) Same as (a), but for Cτ(y,x)C_{\tau}(y,x) and its sensitivity with respect to ϵ1\epsilon_1. (d) Same as (b), in reference to the linear response shown in (c). (e) Same as (a), but for Cτ(y,y)C_{\tau}(y,y) and its sensitivity with respect to ϵ2\epsilon_2. (f). Same as (b), in reference to the linear response shown in (e).

In the case of the full system, we adopt a simple strategy of direct numerical simulation (DNS). The correlations have been computed by collecting statistics along the same simulation used to estimate the Green’s functions above. In order to evaluate their response to perturbations, we have run two additional simulations with f1=1f_1=1, f2=0f_2=0 and ϵ1=±0.05\epsilon_1=\pm0.05 plus two additional simulations with f1=0f_1=0, f2=1f_2=1 and ϵ2=±0.05\epsilon_2=\pm0.05 and taken centred differences to estimate the linear response to perturbation in the xx and yy directions.

In the case of the Markov chain, we have used the expression of correlations and the linear response formulas to static forcings provided in Sect. 3.

We obtain that - see Figure 3a), Figure 3a)Figure 3c), and Figure 3e) - there is a very good agreement between the correlations computed on the full system and those estimated via Markov chains. Comparing with Figure 2, it is apparent that Cτ(x,x)∣ϵ2=ϵ1=0∝exp⁡(λ2t)C_\tau(x,x)|_{\epsilon_2=\epsilon_1=0}\propto\exp(\lambda_2t) and Cτ(y,y)∣ϵ2=ϵ1=0∝exp⁡(λ3t)C_\tau(y,y)|_{\epsilon_2=\epsilon_1=0}\propto\exp(\lambda_3t).

Similarly, a good agreement is found when considering the sensitivities with respect to ϵ1\epsilon_1 and ϵ2\epsilon_2. In all cases it is apparent that a little fraction of the signal is lost when performing the coarse graining. The good performance obtained even in the case of the 3-state system indicates the effectiveness of the reduced order modelling strategy. Using the decomposition presented in Eq. (22), we are able to separate the change in the correlation function in four components, which are associated with fundamentally different dynamical processes. Two terms come from the sensitivity in the expectation value of the two observables we are considering. An additional term - indicated by δb,ϵ(1)\delta^{(1)}_{b,\epsilon} comes from the change in the expectation value of the lagged product of the observable due exclusively to the change in the measure (where instead the evolution occurs according to the unperturbed dynamics). The most interesting term is undoubtedly δa,ϵ(1)\delta^{(1)}_{a,\epsilon}, which measures the impact of the change in the dynamics occurring up to the considered time lag. Indeed, this terms vanishes as τ→0\tau\to0 and is, as already observed in [67], a specific element of response formulas for correlations. The interplay between the two terms δa,ϵ(1)\delta^{(1)}_{a,\epsilon} and δb,ϵ(1)\delta^{(1)}_{b,\epsilon} is nontrivial.

Nonlinear Response

Whilst explicit formulas for nonlinear Green’s functions exist [13, 14], their numerical implementation is extremely challenging because of the convoluted structure of differential operators acting at different times. The nonlinear Green functions can be formally seen as Volterra kernels [87] and can in principle be constructed using neural networks [88].

Using the formalism developed here, we derive easily implementable and easily interpretable formulas for the second (see Subsect. 2(b)) and well as the arbitrary order nonlinear response (see App. A), whereby the way different intrinsic time scales of the system and the corresponding modes interact with each other and with the forcing is extremely clear. Hence, as a final step proof-of-concept analysis of our system we have computed the second order Green’s functions for our system.

Indeed, we consider here a slightly more complicated setting than what has been earlier presented in Subsect. 2(b). We consider the case where both forcings described above are applied. It is easy to derive that the second-order response can be written as:

δ(2)Ψ(n)=∑i,j=x,y∑k=−∞∞∑p=−∞∞ϵiϵjΘ(k)Θ(p)⟨miT(MT)pmjT(MT)kΨ,νinv⟩×fi(n−k−p−2)fj(n−k−1)+h.o.t.(29)\delta^{(2)}\Psi(n)=\sum_{i,j=x,y}\sum_{k=-\infty}^{\infty}\sum_{p=-\infty}^{\infty}\epsilon_i\epsilon_j\Theta(k)\Theta(p)\langle m_i^T(\mathcal{M}^T)^p m_j^T(\mathcal{M}^T)^k\Psi,\nu_{\mathrm{inv}}\rangle \times f_i(n-k-p-2)f_j(n-k-1)+\mathrm{h.o.t.} \tag*{(29)}

where we can define the following Green’s functions

Gmi,mj,Ψ(2)=Θ(k)Θ(p)⟨miT(MT)pmjT(MT)kΨ,νinv⟩(30)G^{(2)}_{m_i,m_j,\Psi}=\Theta(k)\Theta(p)\langle m_i^T(\mathcal{M}^T)^p m_j^T(\mathcal{M}^T)^k\Psi,\nu_{\mathrm{inv}}\rangle \tag*{(30)}

describes the combined effect of first applying the perturbation described by mjm_j and then of the perturbation described by mim_i. One should note that in general Gmi,mj,Ψ(2)≠Gmj,mi,Ψ(2)G^{(2)}_{m_i,m_j,\Psi} \ne G^{(2)}_{m_j,m_i,\Psi} if mj≠mim_j \ne m_i, because the time ordering matters. In our case, the second order response depends in general on all of these four terms.

Given the symmetry properties of the system, if we choose xx as observable Gmx,mx,x(2)=Gmy,my,x(2)=0G^{(2)}_{m_x,m_x,x} = G^{(2)}_{m_y,m_y,x} = 0, whilst Gmx,my,x(2)G^{(2)}_{m_x,m_y,x} and Gmy,mx,x(2)G^{(2)}_{m_y,m_x,x} are in general non-vanishing. This implies that if we do not apply forcings in both direction, the second order response of xx vanishes. Instead, if we choose yy as observable, we have Gmx,my,y(2)=Gmy,mx,y(2)=0G^{(2)}_{m_x,m_y,y} = G^{(2)}_{m_y,m_x,y} = 0, whilst Gmx,mx,y(2)G^{(2)}_{m_x,m_x,y} and Gmy,my,y(2)=0G^{(2)}_{m_y,m_y,y} = 0 are non-vanishing.

The non-vanishing second-order Green’s functions computed for RMM are reported in Figure 4. As we see, they follow closely the functional dependence derived in Eq. (19), where in this case we obtain simple monomials because only one of the factors αij=⟨mpTΠjmqΠiΨ,νinv⟩\alpha_{ij} = \langle m_p^{T}\Pi_jm_q\Pi_i\Psi,\nu_{\mathrm{inv}}\rangle is non-vanishing, because of symmetry, for a choice p,q=x,yp,q=x,y; i,j=2,3i,j=2,3; and Ψ=x,y\Psi=x,y. It is extremely encouraging to observe that, just as for the linear case, if we repeat the same analysis using the high-resolution Ulam discretization the results are basically unchanged, compare Figure 4-Figure 5 despite the presence of hundreds of Koopman modes, and hence of hundreds of thousands of αij\alpha_{ij} factors, only one monomial appears to contribute to the second order Green’s function, thus supporting the efficiency of the model reduction attained with the RMM.

Second order Green’s functions

Figure 4. Second order Green’s functions. Note that decay rates are controlled by a suitable combination of the two eigenvalues of the Koopman operator. Results obtained using RMM. See Eqs. (29)-(30).

Four contour plots showing response formulas obtained using the Ulam discretization

Figure 5. Same as in Figure 4, but here the the results have been obtained using the Ulam discretization.

Conclusions

Constructing accurate and efficient response operators for complex systems is a problem of both theoretical and practical relevance across multiple fields in quantitative sciences [2, 6, 9, 89]. In the case of systems possessing smooth invariant measures, it is possible to resort to non-standard formulations of the FDT to recover such response operators. By combining such a formalism with Koopmanism [20], one gains the important property of interpretability, as it is possible to decompose the Green’s functions of interest into a sum of terms, each associated with a specific mode of variability of the system [17, 25]. The use of Koopmanism is instrumental for establishing response formulas valid also in the case the stochastic component of a system includes jump processes [27].

A key problem in the use of response theory is that one usually needs full knowledge of the evolution equations in order to construct the Green functions and translate them in usable objects at algorithmic level. The latter task is extremely daunting especially when one deals with systems obeying deterministic evolution laws [90, 91, 92, 93]. Recently, it has been proposed to derive response operators by deploying fairly sophisticated machine learning methods based on generative score model [94]. Such a strategy, despite its great potential, is not a silver bullet for cracking the problem of constructing response operators for high dimensional models because of the need to train properly and extensively the surrogate model and issues with out-of-sample performance.

It is indeed possible to devise workarounds to derive Green’s functions even in the case of extremely high dimensional systems by performing suitably defined perturbation experiments, as done in the case of climate models, where the direct evaluation of response operators seems an insurmountable task [95, 96, 97]. The flip side is that one can indeed construct useful and accurate black-box-like objects that translate forcings into predicted response, but lack the ability to disentangle and possibly organise hierarchically the impact of the multiple ongoing physical processes. Hence, the level of interpretability of the response operators, despite their skills, can again be disappointing.

Towards Equation-Free and Interpretable Response Formulas

Here we have developed a rather comprehensive set of response formulas - linear and nonlinear, for observables and for correlations, for static as well as for general time dependent perturbations - for Markov chains possessing a unique ergodic invariant measure, which are of primary importance for dynamical systems theory and statistical mechanics as a whole. See Appendix A and Appendix B for the general formulas. Appendix C presents a linear response formulas for the total entropy production of a Markov chain undergoing time-dependent perturbation. All of the matrix equations reported in this paper are pseudo-codes that can be seamlessly translated into functioning routines using e.g. software like MATLAB [98] (used here by the author) or open source equivalents like Julia [99], Python [100], and Octave [101], among others.

Hopefully these results can be useful for advancing our understanding of the sensitivity of Markov chains to perturbations and, in particular, of the response near criticality, associated with the closure of the spectral gap of the unperturbed transfer operator [16, 17, 59, 102, 103, 104]. We will focus on this specific and extremely important problem in a separate study. As for future investigations, it is also tempting to explore the case of absorbing Markov chains, which describe processes where there is a hole (or a trap) in the reference state space, so that the process is eventually killed [105]. In many cases it is possible to define and prove the existence of quasi-stationary and quasi-ergodic measures, which are constructed by adapting the usual notions to this specific case where one needs to take into account of the continuous leak of mass occurring in the state space [106]. Since the existence of such measures requires, roughly speaking, the presence of a spectral gap of the transfer operator, it seems interesting to explore whether response formulas could be developed also for such Markov chains.

Going back to applications, the main idea of this paper is to delineate a methodological pipeline for developing simple response formulas that

(i) can be used in a purely data-driven environment, or even if we do not know the evolution equations of the system;

(ii) can be cast as simple algebraic operations performed with matrices, thus taking full advantage of the outstanding development occurred in the last decades in numerical linear algebra and the vast availability of dedicated software environments;

(iii) allow for a clear interpretation of the response operator thanks to the use of Koopman formalism in the finite state space of the reduced order model.

We have thoroughly explored the efficacy of our strategy on a simple 2D gradient flow forced by additive and diagonal gaussian noise. The system we have investigated features three competing metastable states. We have been able to construct linear and nonlinear response operators that have allow us to define sensitivity and explore response to time-dependent perturbations for observables as well as for correlations of the system. We have shown that even considering a very severely reduced discrete representation of the system, we are able to obtain high-quality information on its response to perturbations and to associated the response to specific modes of unperturbed variability.

The framework above foresees the use of suitable methods of reduction of complexity of a system via MSM before the response theory developed here is used. MSM is very effective in creating a surrogate representation of a possibly multiscale, many-particle system in a moderate number of states [42]. What we propose here is to extend MSM in such a way that fairly general perturbations can be dealt with. All one needs is a reference dataset plus few extra datasets produced with a slightly perturbed dynamics, thus allowing to obtain an estimate of the unperturbed Perron-Frobenius operator and of its perturbation in the basis defined by the MSM. Indeed, one can then study the response of the system directly at the desired coarse grained level, bypassing the need to look into all the intricacies of the system in the original resolution. The wide availability of software tools facilitating the construction of MSM [107, 108, 109], the wide range of areas of applications for MSM [40, 42], recently extended to also to climate applications [110], and the growing evidence of the efficacy of MSM in capturing the correlation properties of the full system [111] support our strategy.

As mentioned above, constructing a discrete-time, finite-state Markov chain representation of a system amounts to applying the Koopman formalism where the elements of the dictionary are the characteristic functions of the Voronoi cells of the phase space. Hence, the results presented here can be seen as linking our attempts at combining of response theory and Koopmanism [17, 25, 26] with the multiplicative DMD algorithm [58]. The multiplicative DMD could be improved by using bisecting k-means to create the centroids of the Voronoi tessellation, as this methodology is better suited for detecting multiscale features and adapts well to high-dimensional datasets [112]. Increased accuracy in the estimate of response operators via Koopman methods is likely to be obtained by adopting more sophisticated DMD methods that avoid the problem of spectral pollution [113].

Concluding, it is worth noting that in many real-life applications, markovianity is a convenient, useful lie and the lack thereof is the honest, uncomfortable truth. Hence, in the future efforts should be directed at extending the results presented here to the case of Markov chains with memory [114] space and hidden Markov models [115].

A Final Comment on the Prony Method

In [77, 116] we explored the fact that response theory provides solid foundations for a key statistical analysis method used in climate science, the optimal fingerprinting method for detection and attribution of climate change [117, 118, 119, 120]. Also the results presented in this contribution seem to provide some clarification for a widely used statistical analysis method used, in this case, in many signal processing applications. In Eqs. (7)–(8) we have shown that the Koopman operator-based expansion of the linear Green’s functions allows to decompose it in a sum of exponential terms with weighting factors that depend on the chosen observables and on the applied forcings, whilst the exponential decay rates depend exclusively on the properties of the unperturbed system. Indeed, this functional representation points directly to the popular Prony method, which aims at representing the (in general, multivariate) response collected at discrete times of a general system to instantaneous perturbations as a weighted sum of exponentials [69, 70, 71, 72]. Usually, the number of exponentials one needs to use is a free and uncertain metaparameter of the statistical method. Our approach provides an interpretation of such a metaparameter, which corresponds to the number of discrete states (plus one) we consider in a hypothetical Markov state representation of the system. The well-known uncertainty in defining the optimal value for the metaparameter in the presence of strong noise and/or limited amount of available data can be linked to the difficulty in constructing an accurate Markov model in such conditions. Our observation is de facto dual to the findings presented in [121], where it was shown that a multivariate version of the Prony method can be used to perform DMD analysis of a nonlinear dynamical system.

Acknowledgements. VL wishes to thank M. Colbrook, I. Mezić, A. Souza, and N. Zagli and three anonymous reviewers for having provided very insightful comments on the first version of this paper.

Data Accessibility. The data needed to reproduce the figures included in the paper and the key components of the code are available here: https://figshare.com/s/60cc002c081fdcbadabc.

Competing Interests. The author declares no competing interests.

Funding. This research has been partially supported by the Horizon Europe Project ClimTIP (Grant No. 100018693), by the ARIA SCOP-PR01-P003 - Advancing Tipping Point Early Warning AdvanTip project, by the European Space Agency project PREDICT (contract 4000146344/24/I-LR), by the EPSRC project LINK (Grant No. EP/Y026675/1), and by the CROPS RETF project funded by the University of Reading.

General formulas for arbitrary order Green’s function

It is relatively straightforward to prove by induction the following result for the jthj^{\mathrm{th}} order response of a generic observable Ψ\Psi. We have:

1n!dn⟨Ψ,ν(n)⟩dεn∣ε=0=⟨Ψ,ν(j)(n)⟩=∑k1=−∞∞Θ(k1)⟨mT(MT)k1Ψ,ν(j−1)(n−k1−1)⟩f(n−k1−1)=∑k1,…,kj=−∞∞Θ(k1)…Θ(kn)⟨mT(MT)kn…mT(MT)k1Ψ,νinv⟩×f(n−∑p=1jkp−j)…f(n−k1−1)=(Gm,Ψ(n)∗f)(n)(31)\begin{aligned} \left.\frac{1}{n!}\frac{\mathrm{d}^{n}\langle\Psi,\nu(n)\rangle}{\mathrm{d}\varepsilon^{n}}\right|_{\varepsilon=0} &=\langle\Psi,\nu^{(j)}(n)\rangle\\ &=\sum_{k_{1}=-\infty}^{\infty}\Theta(k_{1})\langle m^{T}(\mathcal{M}^{T})^{k_{1}}\Psi,\nu^{(j-1)}(n-k_{1}-1)\rangle f(n-k_{1}-1) \\ &=\sum_{k_{1},\ldots,k_{j}=-\infty}^{\infty}\Theta(k_{1})\ldots\Theta(k_{n})\langle m^{T}(\mathcal{M}^{T})^{k_{n}}\ldots m^{T}(\mathcal{M}^{T})^{k_{1}}\Psi,\nu_{\mathrm{inv}}\rangle\times\\ &\qquad f\left(n-\sum_{p=1}^{j}k_{p}-j\right)\ldots f(n-k_{1}-1) \\ &=\left(\mathcal{G}^{(n)}_{m,\Psi}*f\right)(n) \tag*{(31)} \end{aligned}

where Gm,Ψ(n)(k1,…,kj)=Θ(k1)…Θ(kn)⟨mT(MT)kn…mT(MT)k1Ψ,νinv⟩\mathcal{G}^{(n)}_{m,\Psi}(k_{1},\ldots,k_{j})=\Theta(k_{1})\ldots\Theta(k_{n})\langle m^{T}(\mathcal{M}^{T})^{k_{n}}\ldots m^{T}(\mathcal{M}^{T})^{k_{1}}\Psi,\nu_{\mathrm{inv}}\rangle and where ∗* indicates here a nn-uple convolution sum.

Using the spectral expansion of the M\mathcal{M} operator in Kolmogorov modes, we have:

Gm,Ψ(n)(k1,…,kn)=Θ(k1)…Θ(kn)∑i1,…,in=2Nαi1,…,inλi1k1…λinkn,αi1,…,in=⟨mTΠin…mTΠi1Ψ,νinv⟩(32)\mathcal{G}^{(n)}_{m,\Psi}(k_1,\ldots,k_n)=\Theta(k_1)\ldots\Theta(k_n)\sum_{i_1,\ldots,i_n=2}^{N}\alpha_{i_1,\ldots,i_n}\lambda_{i_1}^{k_1}\ldots\lambda_{i_n}^{k_n},\qquad\alpha_{i_1,\ldots,i_n}=\langle m^T\Pi_{i_n}\ldots m^T\Pi_{i_1}\Psi,\nu_{\mathrm{inv}}\rangle \tag*{(32)}

Hence, we derive that the Green’s function of order nn is a function of nn variables that can be written as a sum of exponentials that decrease with time, considering that λik=exp⁡(kνi)\lambda_i^k=\exp(k\nu_i), with νi=log⁡(λi)<0\nu_i=\log(\lambda_i)<0. Note that the previous result provides also a general statistical model one can use to fit experimental or model generated data.

B Dynamic Response of Correlations

We extend here the results presented in Sect. 3 to the more general case M→Mϵ,n=M+ϵf(n)m\mathcal{M}\to\mathcal{M}_{\epsilon,n}=\mathcal{M}+\epsilon f(n)m. Since the system has explicit time-dependent dynamics, so that the statistical properties are indexed by the observation time nn. We have:

Cn,lϵ(Ψ,Φ)=⟨∏p=0l−1(MT+ϵf(n+p)mT)lΨ∘Φ,νinv+∑p=1∞ϵpν(p)(n)⟩−⟨Ψ,νϵ(n)⟩⟨Φ,νϵ(n)⟩(33)C^\epsilon_{n,l}(\Psi,\Phi)=\left\langle\prod_{p=0}^{l-1}\left(\mathcal{M}^T+\epsilon f(n+p)m^T\right)^l\Psi\circ\Phi,\nu_{\mathrm{inv}}+\sum_{p=1}^{\infty}\epsilon^p\nu^{(p)}(n)\right\rangle-\langle\Psi,\nu_\epsilon(n)\rangle\langle\Phi,\nu_\epsilon(n)\rangle \tag*{(33)}

If we now collect the terms up to first order in ϵ\epsilon, we obtain:

⟨∏q=0l−1(MT+ϵmTf(n+q))Ψ∘Φ,νinv+∑p=1∞ϵpν(p)(n)⟩+⟨(MT)lΨ∘Φ,νinv⟩+ϵ∑q=0l−1⟨(Ml−q−1)TmTf(n+q)(Mq)TΨ∘Φ,νinv⟩+ϵ∑k=−∞∞⟨(MT)lΨ∘Φ,Θ(k)Mkmνinvf(n−k−1)⟩+o(ϵ)(34)\begin{aligned} &\left\langle\prod_{q=0}^{l-1}\left(\mathcal{M}^T+\epsilon m^T f(n+q)\right)\Psi\circ\Phi,\nu_{\mathrm{inv}}+\sum_{p=1}^{\infty}\epsilon^p\nu^{(p)}(n)\right\rangle\\ &+\left\langle(\mathcal{M}^T)^l\Psi\circ\Phi,\nu_{\mathrm{inv}}\right\rangle\\ &+\epsilon\sum_{q=0}^{l-1}\left\langle(\mathcal{M}^{l-q-1})^T m^T f(n+q)(\mathcal{M}^q)^T\Psi\circ\Phi,\nu_{\mathrm{inv}}\right\rangle\\ &+\epsilon\sum_{k=-\infty}^{\infty}\left\langle(\mathcal{M}^T)^l\Psi\circ\Phi,\Theta(k)\mathcal{M}^k m\nu_{\mathrm{inv}}f(n-k-1)\right\rangle+o(\epsilon) \tag*{(34)} \end{aligned}

Additionally, up to first order in ϵ\epsilon, we have

⟨Ψ,νinv+∑p=1∞ϵpν(p)(n)⟩⟨Φ,νinv+∑p=1∞ϵpν(p)(n)⟩=⟨Ψ,νinv⟩⟨Φ,νinv⟩+ϵ∑k=−∞∞⟨Ψ,Θ(k)Mkmνinvf(n−k−1)⟩⟨Φ,νinv⟩+ϵ⟨Ψ,νinv⟩∑k=−∞∞⟨Φ,Θ(k)Mkmνinvf(n−k−1)⟩+o(ϵ)(35)\begin{aligned} &\left\langle\Psi,\nu_{\mathrm{inv}}+\sum_{p=1}^{\infty}\epsilon^p\nu^{(p)}(n)\right\rangle\left\langle\Phi,\nu_{\mathrm{inv}}+\sum_{p=1}^{\infty}\epsilon^p\nu^{(p)}(n)\right\rangle\\ &=\langle\Psi,\nu_{\mathrm{inv}}\rangle\langle\Phi,\nu_{\mathrm{inv}}\rangle\\ &+\epsilon\sum_{k=-\infty}^{\infty}\langle\Psi,\Theta(k)\mathcal{M}^k m\nu_{\mathrm{inv}}f(n-k-1)\rangle\langle\Phi,\nu_{\mathrm{inv}}\rangle\\ &+\epsilon\langle\Psi,\nu_{\mathrm{inv}}\rangle\sum_{k=-\infty}^{\infty}\langle\Phi,\Theta(k)\mathcal{M}^k m\nu_{\mathrm{inv}}f(n-k-1)\rangle+o(\epsilon) \tag*{(35)} \end{aligned}

We derive our final result:

dCl,nϵ(Ψ,Φ)dϵ∣ϵ=0=∑q=0l−1⟨(Ml−q−1)TmTf(n+q)(Mq)TΨ∘Φ,νinv⟩+∑k=−∞∞Θ(k)⟨mT(Mk)T((Ml)TΨ∘Φ−Ψ⟨Φ,νinv⟩−Φ⟨Ψ,νinv⟩),νinv⟩f(n−k−1),(36)\begin{aligned} \left.\frac{\mathrm{d}C^\epsilon_{l,n}(\Psi,\Phi)}{\mathrm{d}\epsilon}\right|_{\epsilon=0} =\sum_{q=0}^{l-1}\left\langle(\mathcal{M}^{l-q-1})^T m^T f(n+q)(\mathcal{M}^q)^T\Psi\circ\Phi,\nu_{\mathrm{inv}}\right\rangle\\ &+\sum_{k=-\infty}^{\infty}\Theta(k)\left\langle m^T(\mathcal{M}^k)^T\left((\mathcal{M}^l)^T\Psi\circ\Phi-\Psi\langle\Phi,\nu_{\mathrm{inv}}\rangle-\Phi\langle\Psi,\nu_{\mathrm{inv}}\rangle\right),\nu_{\mathrm{inv}}\right\rangle f(n-k-1), \tag*{(36)} \end{aligned}

which clearly agrees with the case of static perturbation shown before if one assumes f=1f=1 and n→∞n \to\infty.

Entropy Production

The total entropy production at time nn for a time-dependent Markov chain can be written as:

Stot(n)=∑i,jMϵ(n)ijν(n)jln⁡(Mϵ(n)ijν(n)jMϵ(n)jiν(n)i)=12∑i,j(Mϵ(n)ijν(n)j−Mϵ(n)jiν(n)i)ln⁡(Mϵ(n)ijν(n)jMϵ(n)jiν(n)i).(37)\begin{aligned} S_{\mathrm{tot}}(n) &= \sum_{i,j} \mathcal{M}_{\epsilon}(n)_{ij}\nu(n)_{j}\ln\left(\frac{\mathcal{M}_{\epsilon}(n)_{ij}\nu(n)_{j}}{\mathcal{M}_{\epsilon}(n)_{ji}\nu(n)_{i}}\right) \\ &= \frac{1}{2}\sum_{i,j}\left(\mathcal{M}_{\epsilon}(n)_{ij}\nu(n)_{j}-\mathcal{M}_{\epsilon}(n)_{ji}\nu(n)_{i}\right)\ln\left(\frac{\mathcal{M}_{\epsilon}(n)_{ij}\nu(n)_{j}}{\mathcal{M}_{\epsilon}(n)_{ji}\nu(n)_{i}}\right). \tag*{(37)} \end{aligned}

where we have adapted Seifert’s results [66] to the case of discrete time dynamics. We want to expand the previous expression up to first order in ϵ\epsilon. We obtain after lengthy but straightforward calculations:

Stot(n)=Stot(0)+∂Stot(n)∂ϵ∣ϵ=0ϵ+h.o.t.(38)S_{\mathrm{tot}}(n)=S_{\mathrm{tot}}^{(0)}+\left.\frac{\partial S_{\mathrm{tot}}(n)}{\partial\epsilon}\right|_{\epsilon=0}\epsilon+\mathrm{h.o.t.} \tag*{(38)}

where

Stot(0)=∑i,jMij(νinv)jln⁡(Mij(νinv)jMji(νinv)i)(39)S_{\mathrm{tot}}^{(0)}=\sum_{i,j}\mathcal{M}_{ij}(\nu_{\mathrm{inv}})_{j}\ln\left(\frac{\mathcal{M}_{ij}(\nu_{\mathrm{inv}})_{j}}{\mathcal{M}_{ji}(\nu_{\mathrm{inv}})_{i}}\right) \tag*{(39)}

and the linear response formula for the entropy production is given by

∂Stot(n)∂ϵ∣ϵ=0=∑i,j(mijf(n)(νinv)j+Mijνj(1)(n))ln⁡(Mij(νinv)jMji(νinv)i)(40)\left.\frac{\partial S_{\mathrm{tot}}(n)}{\partial\epsilon}\right|_{\epsilon=0}=\sum_{i,j}\left(m_{ij}f(n)(\nu_{\mathrm{inv}})_{j}+\mathcal{M}_{ij}\nu_{j}^{(1)}(n)\right)\ln\left(\frac{\mathcal{M}_{ij}(\nu_{\mathrm{inv}})_{j}}{\mathcal{M}_{ji}(\nu_{\mathrm{inv}})_{i}}\right) \tag*{(40)}

where, following Eq. (3)

νj(1)(n)=∑k=−∞∞∑iΘ(k)(Mkm)ji(νinv)if(n−k−1).(41)\nu_{j}^{(1)}(n)=\sum_{k=-\infty}^{\infty}\sum_{i}\Theta(k)(\mathcal{M}^{k}m)_{ji}(\nu_{\mathrm{inv}})_{i}f(n-k-1). \tag*{(41)}

Note that if the unperturbed system obeys detailed balance (Mji(νinv)i=Mij(νinv)j ∀i,j\mathcal{M}_{ji}(\nu_{\mathrm{inv}})_{i}=\mathcal{M}_{ij}(\nu_{\mathrm{inv}})_{j}\ \forall i,j), the entropy production of the unperturbed state as well as its sensitivity with respect to ϵ\epsilon vanish.

References

  1. [1]Hänggi P, Thomas H. 1982 Stochastic processes: Time evolution, symmetries and linear response. Physics Reports 88, 207–319. (https://doi.org/10.1016/0370-1573(82)90045-X)
  2. [2]Baiesi M, Maes C. 2013 An update on the nonequilibrium linear response. New Journal of Physics 15, 013004.DOI
  3. [3]Sarracino A, Vulpiani A. 2019 On the fluctuation-dissipation relation in non-equilibrium and non-Hamiltonian systems. Chaos 29, 083132.DOI
  4. [4]Kubo R. 1966 The fluctuation-dissipation theorem. Reports on Progress in Physics 29, 255–284.
  5. [5]Harrer M, Majda AJ. 2010 A simple framework to justify linear response theory. Nonlinearity 23, 909–922. (10.1088/0951-7715/23/4/008)DOI
  6. [6]Marconi UMB, Puglisi A, Rondoni L, Vulpiani A. 2008 Fluctuation–dissipation: response theory in statistical physics. Physics reports 461, 111–195.DOI
  7. [7]Pavliotis GA. 2014 Stochastic Processes and Applications vol. 60. Springer, New York. (10.1007/978-1-4939-1323-7)
  8. [8]Ruelle D. 1998 General linear response formula in statistical mechanics, and the fluctuation-dissipation theorem far from equilibrium. Physics Letters, Section A: General, Atomic and Solid State Physics 245, 220–224. (10.1016/S0375-9601(98)00419-8)DOI
  9. [9]Ruelle D. 2009 A review of linear response theory for general differentiable dynamical systems. Nonlinearity 22, 855–870.DOI
  10. [10]Liverani C, Gouëzel S. 2006 Banach spaces adapted to Anosov systems. Ergodic Theory and Dynamical Systems 26, 189–217.DOI
  11. [11]Baladi V. 2014 Linear response, or else. Proceedings of the International Congress of Mathematicians-Seoul 3, 525–545.arxiv.org/abs/1408.2937
  12. [12]Baladi V, Kuna T, Lucarini V. 2017 Linear and fractional response for the SRB measure of smooth hyperbolic attractors and discontinuous observables. Nonlinearity 30, 1204–1220. (10.1088/1361-6544/aa5b13)DOI
  13. [13]Ruelle D. 1998 Nonequilibrium statistical mechanics near equilibrium: computing higher-order terms. Nonlinearity 11, 5–18.DOI
  14. [14]Lucarini V. 2008 Response theory for equilibrium and non-equilibrium statistical mechanics: Causality and generalized kramers-kronig relations. Journal of Statistical Physics 131, 543–558. (10.1007/s10955-008-9498-y)DOI
  15. [15]Lucarini V. 2009 Evidence of dispersion relations for the nonlinear response of the Lorenz 63 system. J. Stat. Phys. 134, 381–400. (10.1007/s10955-008-9675-z)DOI
  16. [16]Chekroun M, Tantet A, Dijkstra H, Neelin JD. 2020 Ruelle-Pollicott Resonances of Stochastic Systems in Reduced State Space. Part I: Theory. J. Stat. Phys. 179, 1366–1402.DOI
  17. [17]Santos Gutiérrez M, Lucarini V. 2022 On some aspects of the response to stochastic and deterministic forcings. Journal of Physics A: Mathematical and Theoretical 55, 425002. (10.1088/1751-8121/ac90fd)DOI
  18. [18]Lucarini V, Chekroun MD. 2023 Theoretical tools for understanding the climate crisis from Hasselmann’s programme and beyond. Nature Reviews Physics. (10.1038/s42254-023-00650-8)DOI
  19. [19]Mezić I. 2005 Spectral Properties of Dynamical Systems, Model Reduction and Decompositions. Nonlinear Dynamics 41, 309–325. (10.1007/s11071-005-2824-x)DOI
  20. [20]Budišić M, Mohr R, Mezić I. 2012 Applied Koopmanism. Chaos 22, 047510, 33. (10.1063/1.4772195)DOI
  21. [21]Kutz J, Brunton SL, Brunton B, Proctor J. 2016 Dynamic Mode Decomposition: Data-Driven Modeling of Complex Systems. Other Titles in Applied Mathematics. Philadelphia: Society for Industrial and Applied Mathematics.
  22. [22]Brunton SL, Kutz JN. 2022 Data-driven Science and Engineering: Machine Learning, Dynamical Systems, and Control. Cambridge University Press.
  23. [23]Klus S, Nü̈ske F, Koltai P, Wu H, Kevrekidis I, Schütte C, Noé F. 2018 Data-driven model reduction and transfer operator approximation. J. Nonlinear Sci. 28, 985–1010. (10.1007/s00332-017-9437-7)DOI
  24. [24]Colbrook MJ. 2024 Chapter 4 - The multiverse of dynamic mode decomposition algorithms. In Mishra S, Townsend A, editors, Numerical Analysis Meets Machine Learning, Handbook of Numerical Analysis, vol. 25, pp. 127–230. Elsevier. (https://doi.org/10.1016/bs.hna.2024.05.004)
  25. [25]Zagli N, Colbrook M, Lucarini V, Mezić I, Moroney J. 2024 Bridging the Gap between Koopmanism and Response Theory: Using Natural Variability to Predict Forced Response. arXiv:2410.01622. (https://arxiv.org/abs/2410.01622)
  26. [26]Lucarini V, Gutierrez MS, Moroney J, Zagli N. 2025 General Framework for Linking Free and Forced Fluctuations via Koopmanism. arXiv:2506.16446. (https://arxiv.org/abs/arXiv:2506.16446)
  27. [27]Chekroun MD, Zagli N, Lucarini V. 2024 Kolmogorov Modes and Linear Response of Jump-Diffusion Models: Applications to Stochastic Excitation of the ENSO Recharge Oscillator. arXiv preprint arXiv:2411.14769.DOI
  28. [28]Norris JR. 1998 Markov Chains. Cambridge University Press.
  29. [29]Behrends E. 2000 Introduction to Markov Chains. Vieweg+Teubner Verlag. (10.1007/978-3-663-11603-3)
  30. [30]Bowen R. 1970 Markov Partitions for Axiom A Diffeomorphisms. American Journal of Mathematics 92, 725–747. (10.2307/2373370)
  31. [31]Attal S. 2010 Markov Chains and Dynamical Systems: The Open System Point of View. Communications on Stochastic Analysis 4, 447–466. (10.31390/cosa.4.4.05)arxiv.org/abs/1010.2894
  32. [32]Ulam S. 1960 A Collection of Mathematical Problems. Interscience tracts in pure and applied mathematics. Interscience Publishers.
  33. [33]Baladi V. 2000 Positive Transfer Operators and Decay of Correlations vol. 16. World scientific.
  34. [34]Froyland G. 1998 Approximating physical invariant measures of mixing dynamical systems in higher dimensions. Nonlinear Analysis: Theory, Methods & Applications 32, 831–860. (https://doi.org/10.1016/S0362-546X(97)00527-0)
  35. [35]Ding J, Li TY, Zhou A. 2002 Finite approximations of Markov operators. Journal of Computational and Applied Mathematics 147, 137–152. (10.1016/S0377-0427(02)00429-6)DOI
  36. [36]Aurenhammer F. 1991 Voronoi diagrams—a survey of a fundamental geometric data structure. ACM Comput. Surv. 23, 345–405. (10.1145/116873.116880)DOI
  37. [37]Forgy E. 1965 Cluster Analysis of Multivariate Data: Efficiency versus Interpretability of Classification. Biometrics 21, 768–769.
  38. [38]Lloyd S. 1982 Least squares quantization in PCM. IEEE Transactions on Information Theory 28, 129–137. (10.1109/TIT.1982.1056489)DOI
  39. [39]Bussi G, Laio A, Parrinello M. 2006 Equilibrium Free Energies from Nonequilibrium Metadynamics. Phys. Rev. Lett. 96, 090601. (10.1103/PhysRevLett.96.090601)pubmed.ncbi.nlm.nih.gov/16606249
  40. [40]Pande VS, Beauchamp K, Bowman GR. 2010 Everything you wanted to know about Markov State Models but were afraid to ask. Methods 52, 99–105. (10.1016/j.ymeth.2010.06.002)DOI
  41. [41]Bowman GR, Pande V, Noé F. 2014 An introduction to Markov State Models and their application to long timescale molecular simulation. Advances in Experimental Medicine and Biology, volume 797. Dordrecht: Springer.
  42. [42]Husic BE, Pande VS. 2018 Markov State Models: From an Art to a Science. Journal of the American Chemical Society 140, 2386–2396. doi: 10.1021/jacs.7b12191 (10.1021/jacs.7b12191)DOI
  43. [43]Froyland G, Gottwald GA, Hammerlindl A. 2014 A Computational Method to Extract Macroscopic Variables and Their Dynamics in Multiscale Systems. SIAM Journal on Applied Dynamical Systems 13, 1816–1846. (10.1137/130943637)arxiv.org/abs/1310.8001
  44. [44]Bittracher A, Koltai P, Klus S, Banisch R, Dellnitz M, Schütte C. 2018 Transition Manifolds of Complex Metastable Systems. Journal of Nonlinear Science 28, 471–512. (10.1007/s00332-017-9415-0)DOI
  45. [45]Morii H. 1965 Transport, Collective Motion, and Brownian Motion. Progress of Theoretical Physics 33, 423–455.DOI
  46. [46]Zwanzig R. 1961 Memory Effects in Irreversible Thermodynamics. Physical Review 124, 983–992.
  47. [47]Kalliadassis S, Krumscheid S, Pavliotis G. 2015 A new framework for extracting coarse-grained models from time series with multiscale structure. Journal of Computational Physics 296, 314–328. (https://doi.org/10.1016/j.jcp.2015.05.002)
  48. [48]Chekroun MD, Liu H, Wang S. 2015 Stochastic Parameterizing Manifolds and Non-Markovian Reduced Equations. SpringerBriefs in Mathematics. Cham: Springer International Publishing.
  49. [49]Santos Gutiérrez M, Lucarini V, Chekroun MD, Ghil M. 2021 Reduced-order models for coupled dynamical systems: Data-driven methods and the Koopman operator. Chaos 31, 053116.
  50. [50]Chekroun MD, Liu H, McWilliams JC. 2021 Stochastic rectification of fast oscillations on slow manifold closures. Proc. Natl. Acad. Sci. USA 118, Paper No. e2113650118, 9. (10.1073/pnas.2113650118)DOI
  51. [51]Chekroun MD, Liu H, McWilliams JC. 2025 Non-Markovian reduced models to unravel transitions in non-equilibrium systems. Journal of Physics A: Mathematical and Theoretical 58, 045204. (10.1088/1751-8121/ada7ad)arxiv.org/abs/2408.13428
  52. [52]Kevrekidis IG, Gear CW, Hyman JM, Kevrekidis PG, , Runborg O, Theodoropoulos C. 2003 Equation-Free, Coarse-Grained Multiscale Computation: Enabling Macroscopic Simulators to Perform System-Level Analysis. Communications in Mathematical Sciences 1, 718–762. (10.4310/CMS.2003.v1.n4.a5)DOI
  53. [53]Kevrekidis IG, Samaey G. 2009 Equation-Free Multiscale Computation: Algorithms and Applications. Annual Review of Physical Chemistry 60, 321–344. (https://doi.org/10.1146/annurev.physchem.59.032007.093610)
  54. [54]Ma A, Dinner AR. 2005 Automatic Method for Identifying Reaction Coordinates in Complex Systems. The Journal of Physical Chemistry B 109, 6769–6779. (10.1021/jp045546c)pubmed.ncbi.nlm.nih.gov/16851762
  55. [55]Rogal J. 2021 Reaction coordinates in complex systems–a perspective. The European Physical Journal B 94, 223. (10.1140/epjb/s10051-021-0233-5)DOI
  56. [56]Zagli N, Lucarini V, Pavliotis GA. 2021 Spectroscopy of phase transitions for multiagent systems. Chaos: An Interdisciplinary Journal of Nonlinear Science 31, 061103. (10.1063/5.0053558)DOI
  57. [57]Zagli N, Lucarini V, Pavliotis GA. 2024 Response theory identifies reaction coordinates and explains critical phenomena in noisy interacting systems. J. Phys. A 34, 325004. (10.1088/1751-8121/ad6068)DOI
  58. [58]Boullé N, Colbrook MJ. 2024 Multiplicative Dynamic Mode Decomposition. arXiv:2405.05334. (https://arxiv.org/abs/2405.05334)
  59. [59]Lucarini V. 2016 Response Operators for Markov Processes in a Finite State Space: Radius of Convergence and Link to the Response Theory for Axiom A Systems. Journal of Statistical Physics 162, 312–333. (10.1007/s10955-015-1409-4)DOI
  60. [60]Santos Gutiérrez M, Lucarini V. 2020 Response and Sensitivity Using Markov Chains. Journal of Statistical Physics 179, 157–1593. (10.1007/s10955-020-02504-4)
  61. [61]Aslyamov T, Esposito M. 2024a Nonequilibrium Response for Markov Jump Processes: Exact Results and Tight Bounds. Phys. Rev. Lett. 132, 037101. (10.1103/PhysRevLett.132.037101)arxiv.org/abs/2308.04260
  62. [62]Aslyamov T, Esposito M. 2024b General Theory of Static Response for Markov Jump Processes. Phys. Rev. Lett. 133, 107103. (10.1103/PhysRevLett.133.107103)arxiv.org/abs/2402.13990
  63. [63]Zhang XJ, Qian H, Qian M. 2012 Stochastic theory of nonequilibrium steady states and its applications. Part I. Physics Reports 510, 1–86. Stochastic Theory of Nonequilibrium Steady States and Its Applications: Part I (https://doi.org/10.1016/j.physrep.2011.09.002)
  64. [64]Koltai P, Schütte C. 2018 A Multiscale Perturbation Expansion Approach for Markov State Modeling of Nonstationary Molecular Dynamics. Multiscale Modeling & Simulation 16, 1455– 1485. (10.1137/17M1146403)DOI
  65. [65]Falasco G, Esposito M. 2025 Macroscopic stochastic thermodynamics. Rev. Mod. Phys. 97, 015002. (10.1103/RevModPhys.97.015002)DOI
  66. [66]Seifert U. 2005 Entropy Production along a Stochastic Trajectory and an Integral Fluctuation Theorem. Phys. Rev. Lett. 95, 040602. (10.1103/PhysRevLett.95.040602)DOI
  67. [67]Lucarini V, Wouters J. 2017 Response formulae for n-point correlations in statistical mechanical systems and application to a problem of coarse graining. Journal of Physics A: Mathematical and Theoretical 50, 355003. (10.1088/1751-8121/aa812c)DOI
  68. [68]Klus S, Koltai P, Schütte C. 2016 On the numerical approximation of the Perron-Frobenius and Koopman operator. Journal of Computational Dynamics 3, 51–79. (10.3934/jcd.2016003)DOI
  69. [69]Hua Y, Sarkar T. 1990 Matrix pencil method for estimating parameters of exponentially damped/undamped sinusoids in noise. IEEE Transactions on Acoustics, Speech, and Signal Processing 38, 814–824. (10.1109/29.56027)DOI
  70. [70]Park S, Schapery R. 1999 Methods of interconversion between linear viscoelastic material functions. Part I—a numerical method based on Prony series. International Journal of Solids and Structures 36, 1653–1675. (https://doi.org/10.1016/S0020-7683(98)00055-9)
  71. [71]Kunis S, Peter T, Römer T, von der Ohe U. 2016 A multivariate generalization of Prony’s method. Linear Algebra and its Applications 490, 31–47. (https://doi.org/10.1016/j.laa.2015.10.023)
  72. [72]Fernández Rodríguez A, de Santiago Rodrigo L, López Guillén E, Rodríguez Ascariz JM, Miguel Jiménez JM, Boquete L. 2018 Coding Prony’s method in MATLAB and applying it to biomedical signal filtering. BMC Bioinformatics 19, 451. (10.1186/s12859-018-2473-y)DOI
  73. [73]Seneta E. 1973 Non-negative matrices. George Allen and Unwin.
  74. [74]Abbas K, Berkhout J, Heidergott B. 2016 A Critical Account of Perturbation Analysis of Markov Chains. Markov Processes and Related Fields 22, 227–266.arxiv.org/abs/1609.04138
  75. [75]Mitrophanov AY. 2024 The Arsenal of Perturbation Bounds for Finite Continuous-Time Markov Chains: A Perspective. Mathematics 12. (10.3390/math12111608)DOI
  76. [76]Ghil M, Lucarini V. 2020 The physics of climate variability and climate change. Rev. Mod. Phys. 92, 035002. (10.1103/RevModPhys.92.035002)DOI
  77. [77]Lucarini V, Chekroun MD. 2023 Theoretical tools for understanding the climate crisis from Hasselmann’s programme and beyond. Nature Reviews Physics 5, 744–765. (10.1038/s42254- 023-00650-8)DOI
  78. [78]Ghil M, Chekroun MD, Simmonet E. 2008 Climate dynamics and fluid mechanics: Natural variability and related uncertainties. Physica D 237, 2111–2126.DOI
  79. [79]Chekroun MD, Simmonet E, Ghil M. 2011 Stochastic climate dynamics: Random attractors and time-dependent invariant measures. Physica D: Nonlinear Phenomena 240, 1685–1700. (10.1016/j.physd.2011.06.005)
  80. [80]Bódai T, Karolyi G, Tél T. 2011 A chaotically driven model climate: Extreme events and snapshot attractors. Nonlinear Processes in Geophysics 18, 573–580. (10.5194/npg-18-573-2011)DOI
  81. [81]Tél T, Bódai T, Drótos G, Haszpra T, Herein M, Kaszás B, Vincze M. 2020 The Theory of Parallel Climate Realizations. Journal of Statistical Physics 179, 1496–1530. (10.1007/s10955-019-02445-7)DOI
  82. [82]Bódai T, Drótos G, Herein M, Lunkeit F, Lucarini V. 2020 The Forced Response of the El Niño–Southern Oscillation–Indian Monsoon Connection in Ensembles of Earth System Models. Journal of Climate 33, 2163–2182. (10.1175/JCLI-D-19-0341.1)
  83. [83]Risken H. 1989 The Fokker-Planck Equation. Springer second edition.
  84. [84]Bolley F, Gentil I, Guillin A. 2012 Convergence to equilibrium in Wasserstein distance for Fokker–Planck equations. Journal of Functional Analysis 263, 2430–2457. (https://doi.org/10.1016/j.jfa.2012.07.007)
  85. [85]Kloeden PE, Platen E. 1992 Numerical solution of stochastic differential equations / Peter E. Kloeden, Eckhard Platen. Springer-Verlag Berlin ; New York.
  86. [86]Gritsun A, Lucarini V. 2017 Fluctuations, response, and resonances in a simple atmospheric model. Physica D: Nonlinear Phenomena 349, 62–76. (10.1016/j.physd.2017.02.015)DOI
  87. [87]Orcioni S. 2014 Improving the approximation ability of Volterra series identified with a cross-correlation method. Nonlinear Dynamics 78, 2861–2869. (10.1007/s11071-014-1631-7)
  88. [88]Wray J, Green GGR. 1994 Calculation of the Volterra kernels of non-linear dynamic systems using an artificial neural network. Biological Cybernetics 71, 187–195. (10.1007/BF00202758)DOI
  89. [89]Lucarini V, Paviotis GA, Zagli N. 2020 Response theory and phase transitions for the thermodynamic limit of interacting identical systems. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 476, 20200688. (10.1098/rspa.2020.0688)DOI
  90. [90]Wang Q. 2013 Forward and adjoint sensitivity computation of chaotic dynamical systems. Journal of Computational Physics 235, 1–13. (http://dx.doi.org/10.1016/j.jcp.2012.09.007)
  91. [91]Chandramoorthy N, Wang Q. 2020 A computable realization of Ruelle’s formula for linear response of statistics in chaotic systems. arXiv e-prints p. arXiv:2002.04117.arxiv.org/abs/2002.04117
  92. [92]Ni A. 2023 Fast Adjoint Algorithm for Linear Responses of Hyperbolic Chaos. SIAM Journal on Applied Dynamical Systems 22, 2792–2824. (10.1137/22M1522383)DOI
  93. [93]Ni A, Tong Y. 2024 Equivariant Divergence Formula for Hyperbolic Chaotic Flows. Journal of Statistical Physics 191. (10.1007/s10955-024-03329-1)arxiv.org/abs/2312.12171
  94. [94]Giorgini LT, Deck K, Bischoff T, Souza A. 2024 Response Theory via Generative Score Modeling. Phys. Rev. Lett. 133, 267302. (10.1103/PhysRevLett.133.267302)arxiv.org/abs/2402.01029
  95. [95]Ragone F, Lucarini V, Lunkeit F. 2016 A new framework for climate sensitivity and prediction: a modelling perspective. Climate Dynamics 46, 1459–1471. (10.1007/s00382-015-2657-3)DOI
  96. [96]Lucarini V, Ragone F, Lunkeit F. 2017 Predicting Climate Change Using Response Theory: Global Averages and Spatial Patterns. Journal of Statistical Physics 166, 1036–1064. (10.1007/s10955-016-1506-z)DOI
  97. [97]Lembo V, Lucarini V, Ragone F. 2020 Beyond Forcing Scenarios: Predicting Climate Change through Response Operators in a Coupled General Circulation Model. Scientific Reports 10, 8668. (10.1038/s41598-020-65297-2)
  98. [98]The Mathworks, Inc.. 2024 MATLAB version 24.1.0.2537033 (R2024a). Natick, Massachusetts.
  99. [99]Bezanson J, Edelman A, Karpinski S, Shah VB. 2017 Julia: A Fresh Approach to Numerical Computing. SIAM Review 59, 65–98. (10.1137/141000671)arxiv.org/abs/1411.1607
  100. [100]Python Software Foundation. 2023 Python 3.12.1 Documentation.
  101. [101]Eaton JW, Bateman D, Hauberg S, Webhring R. 2014 GNU Octave version 3.8.1 manual: a high-level interactive language for numerical computations. CreateSpace Independent Publishing Platform.
  102. [102]Tantet A, Lucarini V, Dijkstra HA. 2018a Resonances in a Chaotic Attractor Crisis of the Lorenz Flow. Journal of Statistical Physics 170, 584–616. (10.1007/s10955-017-1938-0)DOI
  103. [103]Tantet A, Lucarini V, Lunkeit F, Dijkstra HA. 2018 Crisis of the chaotic attractor of a climate model: A transfer operator approach. Nonlinearity 31, 2221–2251. (10.1088/1361-6544/aaaf42)DOI
  104. [104]Chekroun MD, Neelin JD, Kondrashov D, McWilliams JC, Ghil M. 2014 Rough parameter dependence in climate models: The role of Ruelle-Pollicott resonances. Proc. Natl. Acad. Sci USA 111, 1684–1690. (10.1073/pnas.1321618111)DOI
  105. [105]Collet P, Martínez S, San Martín J. 2013 Quasi-Stationary Distributions: Markov Chains, Diffusions and Dynamical Systems. Probability and Its Applications. Springer. (10.1007/978-3-642-33312-2)DOI
  106. [106]Castro MM, Lamb JS, Olicón-Méndez G, Rasmussen M. 2024 Existence and uniqueness of quasi-stationary and quasi-ergodic measures for absorbing Markov chains: A Banach lattice approach. Stochastic Processes and their Applications 173, 104364. (https://doi.org/10.1016/j.spa.2024.104364)
  107. [107]Senne M, Trendelkamp-Schroer B, Mey ASJS, Schütte C, Noé F. 2012 EMMA: A Software Package for Markov Model Building and Analysis. Journal of Chemical Theory and Computation 8, 2223–2238. (10.1021/ct300274u)DOI
  108. [108]Scherer MK, Trendelkamp-Schroer B, Paul F, Pérez-Hernández G, Hoffmann M, Plattner N, Wehmeyer C, Prinz JH, Noé F. 2015 PyEMMA 2: A Software Package for Estimation, Validation, and Analysis of Markov Models. Journal of Chemical Theory and Computation 11, 5525–5542. (10.1021/acs.jctc.5b00743)DOI
  109. [109]Harrigan MP, Sultan MM, Hernández CX, Husic BE, Eastman P, Schwantes CR, Beauchamp KA, McGibbon RT, Pande VS. 2017 MSMBuilder: Statistical Models for Biomolecular Dynamics. Biophysical Journal 112, 10–15. (10.1016/j.bpj.2016.10.042)DOI
  110. [110]Springer S, Laio A, Galfi VM, Lucarini V. 2024 Unsupervised detection of large-scale weather patterns in the northern hemisphere via Markov State Modelling: from blockings to teleconnections. npj Climate and Atmospheric Science 7, 105. (10.1038/s41612-024-00659-5)
  111. [111]Suárez E, Wiewiora RP, Wehmeyer C, Noé F, Chodera JD, Zuckerman DM. 2021 What Markov State Models Can and Cannot Do: Correlation versus Path-Based Observables in Protein-Folding Models. Journal of Chemical Theory and Computation 17, 3119–3133. (10.1021/acs.jctc.0c01154)DOI
  112. [112]Souza AN, Silvestri S. 2024 A Modified Bisecting K-Means for Approximating Transfer Operators: Application to the Lorenz Equations. .arxiv.org/abs/2412.03734
  113. [113]Colbrook MJ, Townsend A. 2024 Rigorous data-driven computation of spectral properties of Koopman operators for dynamical systems. Communications on Pure and Applied Mathematics 77, 221–283. (https://doi.org/10.1002/cpa.22125)
  114. [114]Wu SJ, Chu MT. 2017 Markov chains with memory, tensor formulation, and the dynamics of power iteration. Applied Mathematics and Computation 303, 226–239. (https://doi.org/10.1016/j.amc.2017.01.030)
  115. [115]Mor B, Garhwal S, Kumar A. 2021 A Systematic Review of Hidden Markov Models and Their Applications. Archives of Computational Methods in Engineering 28, 1429–1448. (10.1007/s11831-020-09422-4)DOI
  116. [116]Lucarini V, Chekroun MD. 2024 Detecting and Attributing Change in Climate and Complex Systems: Foundations, Green’s Functions, and Nonlinear Fingerprints. Phys. Rev. Lett. 133, 244201. (10.1103/PhysRevLett.133.244201)DOI
  117. [117]Hasselmann K. 1997 Multi-pattern fingerprint method for detection and attribution of climate change. Climate Dynamics 13, 601–611. (10.1007/s003820050185)DOI
  118. [118]Allen M, Tett S. 1999 Checking for model consistency in optimal fingerprinting. Climate Dynamics 15, 419–434. (10.1007/s003820050291)
  119. [119]Hegerl G, Zwiers F. 2011 Use of models in detection and attribution of climate change. WIREs Climate Change 2, 570–591. (https://doi.org/10.1002/wcc.121)
  120. [120]Hannart A, Ribes A, Naveau P. 2014 Optimal fingerprinting under multiple sources of uncertainty. Geophysical Research Letters 41, 1261–1268. (https://doi.org/10.1002/2013GL058653)
  121. [121]Susuki Y, Mezić I. 2015 A provable approximation of Koopman Mode Decomposition. In 2015 54th IEEE Conference on Decision and Control (CDC) pp. 7022–7027. (10.1109/CDC.2015.7403326)DOI

Paper details

vlnc:W_hm72gd7r
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