🔬 condensed matter

Data-driven analysis of metastability in a stochastic bistable system

This paper presents a data-driven methodology using the Koopman operator to analyze metastability in stochastic bistable systems by tracking subdominant modes to accurately predict escape statistics, reconstruct basins of attraction, and characterize multi-scale dynamics in both equilibrium and nonequilibrium conditions without relying on trajectory following.

Original authors: Ankan Banerjee, Manuel Santos Gutierrez, John Moroney, Valerio Lucarini

Published 2026-07-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ankan Banerjee, Manuel Santos Gutierrez, John Moroney, 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 a world where everything is constantly jiggling, like a cup of coffee on a bumpy bus. Sometimes, that coffee sits still in a cup; other times, a sudden bump sends it splashing over the edge. In the vast universe of science, there's a field called stochastic dynamics that studies exactly this kind of "jiggly" behavior. It looks at systems that have two or more stable places they want to rest in (like the coffee sitting in the cup), but random noise (the bumps) can push them to jump to a different spot. Scientists call these stable spots "metastable states."

The big question researchers ask is: How long will it take for the system to jump? and What path does it take to get there? Usually, to answer this, you need to know the exact rules of the game—the precise equations that describe how the system moves. But in the real world, like in weather patterns or the human brain, we often don't have those perfect rules. We only have a messy stream of data, like a video recording of the coffee splashing. This paper tackles the challenge of understanding these "jumpy" systems using only that messy data, without needing to know the secret recipe of the physics behind it.

The Great Escape: Tracking Jumps with a Magic Lens

Imagine you are watching a marble rolling around in a landscape with two deep valleys separated by a high hill. The marble likes to sit at the bottom of a valley. But every now and then, a random gust of wind (noise) blows, and the marble might get pushed up the hill and roll down into the other valley. This is a bistable system: it has two "homes."

The authors of this paper wanted to figure out how to predict when the marble will jump from one valley to the other, and how long it usually takes, without needing to know the exact shape of the hills or the strength of the wind. Instead of trying to follow the marble's wobbly path step-by-step (which is hard because the path is chaotic), they used a mathematical tool called the Koopman operator.

Think of the Koopman operator as a magic lens or a special pair of glasses. Instead of looking at the marble itself, this lens looks at the patterns of the marble's movement. It turns the messy, non-linear, wobbly motion into a clean, straight-line story. When you look through this lens, you can see "modes" or "vibrations" of the system. Some of these vibrations happen very fast (like the marble shaking back and forth inside a valley), and some happen very slowly (like the rare moment the marble decides to cross the hill).

The Slow and the Fast

The researchers focused on a specific "vibration" called the subdominant mode. In their magic lens view, this mode acts like a giant switch. On one side of the switch, the value is negative (representing Valley A); on the other side, it's positive (representing Valley B). Right in the middle, where the value is zero, is the invisible boundary line between the two valleys.

By watching this single number (the value of the subdominant mode) change over time, the team could tell exactly when the system was about to jump. When the number crossed from negative to positive (or vice versa), they knew a transition had happened. It's like having a smoke alarm that goes off the moment the marble leaves the valley, without you needing to see the marble at all.

What They Found

The team tested this method on a computer model of a "double-well" system. They added different amounts of "noise" (wind) and even twisted the landscape so the valleys weren't perfectly symmetrical (creating a "nonequilibrium" system).

It Works Like a Charm: They found that the time it took for the system to jump, calculated by watching this magic switch, matched perfectly with the predictions made by advanced theories (called Large Deviation Theory). This theory says that in a weak wind, the time it takes to jump grows exponentially as the wind gets weaker. Their data-driven method confirmed this, even when the landscape was twisted and messy.
Mapping the Invisible: They could draw a map of the "basins of attraction" (the safe zones) just by looking at the data. The zero-line of their magic switch lined up perfectly with the actual boundary where the marble would fall into the other valley. This is huge because, in real life, we often don't know where these boundaries are.
The "Saddle" Secret: They also discovered something new about the "saddle point"—the very top of the hill between the valleys. Usually, theories ignore the time the marble spends wobbling right at the top. But by looking deeper into the Koopman spectrum (finding even more specific vibrations), they found a mode that describes exactly how long the marble lingers at the top before falling down. This helps explain the "fast" part of the escape process that was previously hard to catch.

Why This Matters

The most exciting part is that this method is data-driven. You don't need to know the equations of motion. You just need a recording of the system's behavior. The authors showed that by using this "magic lens," you can figure out how stable a system is, where the danger zones are, and how long it takes to tip over, even if the system is complex, twisted, or out of balance.

While this study was done on a simple computer model, the authors suggest this approach could be the key to understanding real-world tipping points—like climate change shifts or brain states—where we only have data and no perfect equations. They didn't solve every problem in the universe, but they built a very promising new tool for looking at the invisible boundaries that keep our world stable.

Technical Summary: Data-driven analysis of metastability in a stochastic bistable system

Problem Statement
Metastability, characterized by the coexistence of multiple competing states in complex systems, is a ubiquitous phenomenon with significant implications in climate science, ecology, and engineering. A critical challenge in understanding these systems is quantifying the stability of metastable states and predicting "noise-induced tipping"—stochastically driven transitions between these states. While rigorous mathematical frameworks like large deviation theory (LDT) exist to describe mean exit times and escape paths (instantons), they typically require full knowledge of the underlying dynamical equations and the geometry of the phase space (e.g., the quasipotential). For real-world complex systems where only partial time-series data of observables is available, these geometry-dependent methods face significant limitations. The authors address the need for a purely data-driven methodology to analyze metastability, specifically focusing on identifying slow inter-well transitions and fast saddle escape processes without prior knowledge of the system's governing equations or basin boundary geometry.

Methodology
The authors propose a framework based on the Koopman operator formalism, which shifts the analysis from the nonlinear evolution of state variables to the linear evolution of observables in an infinite-dimensional space. The study utilizes a prototypical two-dimensional stochastic bistable system (a rotating double-well potential) that can be tuned to represent both equilibrium ( $\mu=0$ ) and nonequilibrium ( $\mu \neq 0$ ) conditions.

The core methodology involves:

Data Generation: Numerical simulation of the stochastic differential equations using the Euler-Maruyama scheme to generate ensembles of time series under varying noise intensities ( $\alpha$ ) and rotational parameters ( $\mu$ ).
Koopman Approximation: Application of the Extended Dynamic Mode Decomposition (EDMD) algorithm. A dictionary of polynomial functions is used to approximate the Koopman operator and its generator from the time-series data.
Spectral Analysis: Identification of the eigenvalues and eigenfunctions (modes) of the Koopman operator. The study focuses on:
- The first subdominant mode ( $\phi_2$ ), associated with the slowest decay rate, which governs transitions between metastable states.
- Deeper subdominant modes, associated with faster decay rates, which capture intra-well dynamics and saddle-to-attractor relaxation.
Validation: Comparison of data-driven estimates against theoretical predictions derived from Large Deviation Theory (specifically Kramers' law and its extensions for irreversible systems) and the Friedlin-Wentzell theory.

Key Contributions and Results

Geometry-Agnostic Basin Identification: The first subdominant Koopman mode ( $\phi_2$ ) is shown to partition the phase space into two almost-invariant sets corresponding to the competing basins of attraction. The zero-level set (isosurface) of $\phi_2$ accurately reconstructs the deterministic basin boundary, even in the presence of noise and without explicit knowledge of the system's equations. This holds true for both equilibrium and nonequilibrium (rotational) cases.
Estimation of Mean Exit Times: The authors demonstrate two data-driven approaches to estimate mean exit times:
1. Time-Series Tracking: By monitoring the scalar time series of $\phi_2(x(t))$ , transitions are identified when the mode crosses zero. The statistics of these crossing times yield mean exit times that agree well with LDT predictions, including both the exponential scaling and the subexponential prefactor.
2. Spectral Estimation: The inverse of the absolute value of the second eigenvalue ( $1/|\lambda_2|$ ) of the Koopman generator is shown to equal the mean exit time from an attractor. This relationship holds for both reversible and irreversible systems, confirming recent theoretical results.
Decomposition of Escape Processes: By analyzing deeper modes in the Koopman spectrum, the authors successfully distinguish between:
- Inter-well mixing: Governed by $\phi_2$ .
- Intra-well variability: Associated with complex eigenvalues near the stable fixed points.
- Saddle escape: A specific mode is identified that captures the rapid relaxation of trajectories from the saddle (edge state) towards the attractors. The decay rate of this mode approximates the average time spent near the saddle, a quantity often neglected in standard LDT approximations.
Impact of Nonequilibrium Conditions: The study quantifies how the rotational component ( $\mu$ ) affects metastability. Increasing $\mu$ strengthens the instability of the system, leading to more frequent transitions. While the theoretical scaling of exit times with noise intensity remains valid, the geometric spread of crossing points increases for high $\mu$ , as instantonic trajectories become tangent to the basin boundary over a wider region, causing some escapes to occur further from the saddle.

Significance and Claims
The paper claims to provide a comprehensive, multi-scale framework for studying high-dimensional metastable systems using purely data-driven techniques. Its primary significance lies in:

Equation- and Geometry-Agnostic Analysis: The ability to identify basin boundaries and estimate transition statistics without requiring the explicit form of the drift vector field or the quasipotential.
Unified Treatment of Equilibrium and Nonequilibrium: The methodology successfully applies to both reversible and irreversible stochastic systems, bridging the gap between standard LDT results and data-driven spectral analysis.
Novel Insight into Edge States: The identification of a specific Koopman mode describing the relaxation from the saddle to the attractor offers a new perspective on the "edge state" as a mediator of transitions, a feature largely unaccounted for in previous large deviation arguments.

The authors position this work as an initial step toward analyzing high-dimensional complex systems. They acknowledge that while the current study focuses on a symmetric 2D model, the principles (such as using the zero-isoline of subdominant modes to approximate basin boundaries) are expected to generalize to asymmetric systems and higher dimensions, potentially overcoming the "curse of dimensionality" through kernelized EDMD or residual DMD techniques. The work does not claim to solve all challenges of high-dimensional fractal basin boundaries but suggests that Koopman modes provide a "noise-regularized" estimate of such boundaries.

The Great Escape: Tracking Jumps with a Magic Lens

The Slow and the Fast

What They Found

Why This Matters

More like this