🌀 nonlinear sciences

Interpretable and Equation-Free Response Theory for Complex Systems

This paper presents an interpretable, equation-free response theory for complex systems based on Markov chains and Koopman-inspired algebraic expansions, enabling data-driven prediction of observable responses and higher-order correlations while elucidating the roles of different timescales and providing a dynamical foundation for the Prony method.

Original authors: Valerio Lucarini

Published 2026-06-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Valerio Lucarini

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand how a giant, complex machine—like a weather system, a chemical reaction, or even a social network—reacts when you give it a little nudge. You want to know: if I push it here, how will it move there? And if I push it hard, does it break or change its personality?

This paper is a recipe book for answering those questions, but with a twist: it doesn’t require you to know the blueprints of the machine.

Here is the breakdown of what the authors, led by Valerio Lucarini, have done, explained through everyday analogies.

1. The Problem: The "Black Box" Dilemma

Usually, to predict how a complex system reacts to a change (a "perturbation"), scientists need the exact mathematical equations that govern the system. Think of it like trying to predict how a car will handle a turn without knowing how the engine, suspension, or tires work. For many complex systems (like climate or biological networks), we simply don’t have those exact equations, or they are too messy to use.

The authors want a way to predict the system’s response using only data—like watching the system move and recording what happens—without needing the underlying "engine manual."

2. The Solution: The "Markov State Model" (The Map of Rooms)

The authors use a technique called Markov State Modeling (MSM). Imagine the complex system is a huge, foggy mansion. You can’t see every corner, but you can divide the mansion into a few key rooms (states).

The States: These are the "rooms." For example, in a weather system, one room might be "Sunny," another "Rainy," and another "Stormy."
The Markov Chain: This is a map of how likely you are to move from one room to another. If you are in the "Sunny" room, what are the odds you’ll end up in the "Rainy" room next?

The beauty of this approach is that it simplifies the chaos. Instead of tracking every single air molecule, you just track which "room" the system is in. This is an equation-free method because you build the map just by observing where the system goes, not by solving physics equations.

3. The "Koopman" Lens: Seeing the Hidden Rhythms

Now, how do you predict what happens when you push the system? The authors use a mathematical concept called Koopmanism.

Think of the system’s behavior like a song. A song isn’t just noise; it’s made of specific notes (frequencies) and rhythms. Koopmanism allows the authors to break down the system’s movement into these distinct "notes" or modes.

Some modes are slow (like a bass drum), representing long-term changes.
Some modes are fast (like a hi-hat), representing quick, jittery movements.

By looking at the system through this lens, the authors can see exactly which "notes" are responsible for the system’s reaction. This makes the math interpretable. Instead of getting a confusing blob of numbers, you get a clear answer: "The system reacts this way because of its slow 'bass' mode."

4. The Recipe: Predicting the Response

The paper provides specific mathematical formulas (which are just matrix operations, easy for computers to handle) to calculate two things:

Linear Response: What happens if you give the system a tiny nudge? The authors show that you can predict this by looking at how the "rooms" (states) connect and which "notes" (Koopman modes) are active.
Nonlinear Response: What happens if you give the system a big shove? The system might react in complex, unexpected ways. The authors provide formulas for this too, showing how different "notes" interact with each other when the push is strong.

They also extend this to correlations. In science, we often care not just about the average state (e.g., "Is it raining?") but about how things are linked over time (e.g., "If it rains today, is it likely to rain tomorrow?"). The paper shows how to predict how these links change when you perturb the system.

5. The Proof: The Three-Well Potential

To prove their method works, the authors tested it on a simple but tricky system: a ball rolling in a landscape with three valleys (minima).

The ball spends most of its time sitting in one of the three valleys (these are the "metastable states" or "rooms").
Occasionally, noise (random kicks) pushes the ball over a hill into another valley.

The authors built a simplified 3-room map of this system. Then, they applied a force (a push) to the ball.

Result: Their simplified 3-room model predicted the ball’s reaction almost perfectly, matching the results of a much more complex, high-resolution simulation.
Insight: They found that even though the complex simulation had hundreds of "notes" (modes), only one or two specific notes were actually responsible for the response. The simplified model captured exactly those important notes.

6. The Bonus: The Prony Method Connection

Finally, the authors connect their work to a statistical tool called the Prony Method, which is used to analyze signals (like sound waves or heartbeats). The Prony Method tries to fit a signal as a sum of exponential decays (like a ringing bell fading out).

The authors show that their mathematical framework explains why the Prony Method works. The "exponentials" in the Prony Method are actually the "Koopman modes" (the notes) of the system. This gives a solid physical foundation to a statistical tool that was previously just a useful trick.

Summary

In short, this paper provides a data-driven, easy-to-use toolkit for predicting how complex systems react to changes.

It’s Equation-Free: You don’t need the underlying physics equations; you just need data.
It’s Interpretable: It breaks down the response into understandable "modes" or "notes."
It’s Accurate: Even a very simplified model (few states) can capture the essential response of a complex system.
It’s General: It works for small nudges and big pushes, and for both average behaviors and time-linked correlations.

It’s like having a universal translator that lets you ask a complex system, "What will you do if I push you?" and getting a clear, understandable answer based on its natural rhythms, without needing to know how it’s built.

Technical Summary: Interpretable and Equation-Free Response Theory for Complex Systems

Problem Statement
Response theory provides a framework for analyzing how the statistical properties of a system change under external perturbations. While established for equilibrium systems via the Fluctuation-Dissipation Theorem (FDT) and extended to certain non-equilibrium systems, applying response theory to complex, multiscale systems presents significant challenges. Specifically, standard response formulas often lack interpretability regarding the dominant modes of variability and are difficult to implement for high-dimensional systems where underlying evolution equations may be unknown or intractable. Furthermore, while Koopmanism offers a linear framework for nonlinear dynamics, integrating it with response theory to handle time-dependent perturbations and provide equation-free, data-driven implementations remains a nontrivial hurdle.

Methodology
The authors propose a pipeline that combines Markov State Modelling (MSM) with Koopman operator theory to derive interpretable, equation-free response formulas. The methodology proceeds as follows:

Coarse-Graining via MSM: The continuous phase space of a complex system is discretized into a finite number of states using clustering algorithms (e.g., k-means), resulting in a discrete-time, finite-state Markov chain. This approach is "equation-agnostic," relying solely on observed data to construct the transition matrix $M$ , thereby bypassing the need for explicit evolution equations.
Perturbation Framework: The authors consider a perturbed Markov matrix $M_{\epsilon,n} = M + \epsilon f(n)m$ , where $f(n)$ is a time-dependent modulation and $m$ is a perturbation matrix. They derive perturbative expansions for the invariant measure $\nu(n)$ up to arbitrary order in $\epsilon$ .
Koopman Spectral Decomposition: By defining the Koopman operator $K = M^T$ and utilizing its spectral decomposition ( $K = V \Lambda V^{-1}$ ), the authors express the response operators in terms of the eigenvalues ( $\lambda_i$ ) and eigenmodes ( $\Pi_i$ ) of the system. This allows the Green’s functions to be decomposed into sums of exponential terms associated with specific time scales and modes of variability.
Derivation of Response Formulas:
- Observables: Linear and nonlinear response formulas for observables are derived as convolution sums involving the perturbation function and Green’s functions expressed via Koopman modes.
- Correlations: The authors extend the theory to time-lagged correlations, deriving formulas that separate the response into four distinct components: changes in the evolution law, changes in the invariant measure, and changes in the expectation values of the observables.
- Entropy Production: A linear response formula for total entropy production in time-dependent Markov chains is provided in the appendix.

Key Contributions

Explicit, Interpretable Formulas: The paper provides simple, matrix-based expressions for linear and nonlinear response functions for Markov chains. Unlike general functional analytic approaches, these formulas explicitly separate contributions from different Koopman modes, clarifying the role of distinct time scales.
Equation-Free Implementation: The framework is designed for purely data-driven settings. It requires only the construction of a Markov transition matrix from data, making it applicable even when the underlying dynamical equations are unknown.
Response of Correlations: The authors provide novel formulas for the sensitivity of time-lagged correlations to perturbations, a less explored area compared to the response of mean observables.
Dynamical Foundation for the Prony Method: The spectral expansion of the linear Green’s function as a sum of exponentials provides a theoretical justification for the Prony method, commonly used in signal processing. The number of exponentials in the Prony fit corresponds to the number of states in the Markov model.

Results
The methodology is validated on a 2D Langevin system with a potential featuring three metastable minima.

Model Reduction: The authors demonstrate that a highly reduced Markov model (RMM) with only three states (corresponding to the metastable regions) captures the essential dynamics. The eigenvalues of the RMM closely match the subdominant eigenvalues of the high-resolution Ulam discretization (625 states).
Linear Response: The linear Green’s functions for observables $x$ and $y$ are computed using both the FDT (via direct numerical simulation) and the proposed Markov chain formulas. The Markov-based estimates show good agreement with FDT results for times $t \ge 1$ . Crucially, the Koopman decomposition reveals that the response is dominated by a single mode for each forcing direction, demonstrating the interpretability of the framework.
Correlation Response: The sensitivity of time-lagged correlations $C_\tau(x,x)$ and $C_\tau(y,y)$ to perturbations is analyzed. The results from the 3-state RMM agree well with direct numerical simulations. The decomposition of the correlation response into four terms highlights the distinct dynamical processes involved, particularly the term $\delta^{(1)}_{a,\epsilon}$ , which captures the impact of perturbed dynamics over the lag time.
Nonlinear Response: Second-order Green’s functions are computed for the RMM. The results confirm that the nonlinear response is also dominated by a small number of Koopman mode interactions, supporting the efficiency of the model reduction.

Significance
The paper claims to provide a robust, interpretable, and computationally efficient framework for response theory in complex systems. By linking response theory with Koopmanism and MSM, the authors offer a pathway to analyze sensitivity and predict responses in a data-driven manner without requiring explicit knowledge of the system's equations. The ability to decompose responses into interpretable modes associated with specific time scales is highlighted as a major advantage. Additionally, the work provides a dynamical foundation for the Prony method, linking statistical signal analysis techniques to the underlying spectral properties of the system's dynamics. The authors suggest this approach is particularly useful for systems where traditional response theory is intractable due to high dimensionality or lack of analytical tractability.