🔢 mathematics

Data-Driven Discovery and Rigorous Certification of Periodic Solutions in Infinite-Dimensional Dynamical Systems

This paper proposes a novel diagnostic framework that combines low-dimensional data-driven regression with a-posteriori computer-assisted proofs to rigorously certify the existence of periodic solutions in infinite-dimensional dynamical systems, successfully validating the approach on the Mackey-Glass delay differential equation and the Kuramoto-Sivashinsky partial differential equation.

Original author: Aaron Loh, J. D. Mireles James

Published 2026-11-13✓ Author reviewed ⓘ

📖 6 min read🧠 Deep dive

CC BY 4.0

Original author: Aaron Loh, J. D. Mireles James

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 by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the study of complex systems, from the beating of a human heart to the swirling of weather patterns, scientists often face a problem of scale. The real world is messy and infinite, with countless variables interacting in ways that are impossible to track one by one. To make sense of this, researchers build simplified models, or proxies, that capture the essential behavior of a system using far fewer numbers. These models are like a sketch of a landscape; they might miss the texture of every leaf, but they show the shape of the mountains and the flow of the river. For decades, the standard way to check if a sketch is good has been to see how well it predicts the next few steps of a journey. If the model's path matches the real path for a short time, it is considered useful. However, this short-term success is deceptive. A model can mimic a path perfectly for a while and then drift wildly off course, or worse, it might get the short-term details right while completely missing the deep, repeating structures that define the system's long-term behavior. The critical question has remained: can a simplified model truly find the hidden, repeating patterns of the real world, or is it just a clever imitation that breaks down under scrutiny?

A team of researchers has developed a new way to answer this question, moving beyond simple prediction to a form of mathematical proof. Instead of just watching how long a model's path stays close to reality, they ask whether the model can find a specific, repeating cycle and then prove that a true, exact version of that cycle exists nearby in the complex, real system. They treat the simplified model as a scout that locates a potential treasure, and then they use rigorous computer-assisted mathematics to verify that the treasure is actually there. This method does not just say the model is close; it provides a mathematically guaranteed boundary, a small bubble of certainty, within which a true, perfect solution must exist. If the proof succeeds, it confirms that the simplified model has successfully identified a genuine, repeating structure of the complex system, even if the model itself is a rough approximation.

The researchers tested this approach on two famous mathematical systems known for their chaotic and infinite complexity. The first is the Mackey-Glass equation, a model originally designed to describe how the body regulates blood cell production, which behaves like a chaotic system with an infinite number of possible states. The second is the Kuramoto-Sivashinsky equation, a model for how flames flicker or fluids mix, which also lives in an infinite-dimensional space. In both cases, the researchers first used data from these complex systems to build a much simpler, low-dimensional model. They then asked this simple model to find a repeating cycle, a pattern that returns to its starting point again and again. Once the simple model found a candidate cycle, the researchers froze that candidate and used it as a starting point for a rigorous computer proof. This proof checked whether a true, exact cycle existed in the original, complex system very close to the candidate found by the simple model.

The results were striking. For the blood cell model, the simple model found a repeating cycle, and the rigorous proof confirmed that a true cycle exists within a tiny distance of the model's guess. The proof showed that the true cycle is no more than 0.0006 units away from the model's prediction, a margin of error so small it is practically invisible. Crucially, this confirmation was achieved without needing to tweak or correct the model's guess using the complex equations; the simple model's guess was good enough on its own to trigger the proof. For the fluid dynamics model, the researchers tested two different ways of gathering data. When they used data from a full mathematical description of the system, the simple model again found a cycle that was proven to exist within a distance of 0.00001 units. This success demonstrated that the simplified model had captured the true repeating structure of the complex system.

However, the study also revealed the limits of this approach when the data is imperfect. In a second test for the fluid model, the researchers tried to build the simple model using data from only a few fixed points, similar to how weather stations measure temperature at specific locations. As they added more sensors, the model's predictions improved, and the candidates it found got closer to the true cycles. But once they added more than about twenty-four sensors, the model began to fail. The extra data contained information about very fast, tiny movements that the simple model could not learn to dampen correctly. As a result, the model became unstable and could not produce a valid candidate for the rigorous proof. This failure was not because the proof method was broken, but because the simple model had reached the limit of what it could learn from that specific type of data. The study showed that while the method can rigorously certify success, it can also clearly identify when a model is insufficient, a distinction that short-term prediction tests often miss.

The significance of this work lies in its ability to provide a definitive "yes" or "no" regarding the structural fidelity of a model. Traditional methods can tell you that a model is wrong over the long run, but they cannot easily tell you if a specific feature the model found is real or an illusion. This new method bridges that gap. By combining data-driven modeling with computer-assisted proofs, the researchers created a tool that can validate the existence of complex, repeating patterns in infinite systems. The findings confirm that simple models can indeed uncover deep truths about complex systems, provided the data is rich enough to support them. At the same time, the method acts as a precise diagnostic, revealing exactly when a model has learned too little to be trusted. This approach offers a new standard for evaluating how well our mathematical shortcuts reflect the intricate, infinite reality they are meant to describe.

Technical Summary: Data-Driven Discovery and Rigorous Certification of Periodic Solutions in Infinite-Dimensional Dynamical Systems

Problem Statement

The paper addresses a critical gap in data-driven modeling of complex, infinite-dimensional dynamical systems (such as delay differential equations and partial differential equations). While standard regression techniques can fit low-dimensional proxy models to high-dimensional time series data, standard diagnostics (e.g., short-term trajectory tracking) fail to determine whether the long-term invariant structures (such as periodic orbits) of the proxy model correspond to genuine invariant structures in the original target system. The authors ask whether recurrent invariant structures identified in a learned surrogate can be used to rigorously establish the existence of corresponding structures in the original, potentially infinite-dimensional, system.

Methodology

The proposed approach integrates data-driven regression with a-posteriori computer-assisted proofs (CAP). The workflow consists of three main stages:

Surrogate Construction (DPSR):
The authors employ Dense Polynomial Surrogate Reconstruction (DPSR). This involves fitting continuous-time polynomial vector fields to time series data using least squares or ridge regression. Unlike sparse identification methods (e.g., SINDy), this approach retains a complete polynomial library in prescribed observation coordinates.
- Mackey–Glass (MG): Uses delay embedding of a scalar time series to reconstruct a 6-dimensional phase space.
- Kuramoto–Sivashinsky (KS): Uses two setups: (a) a spectral semidiscretization with periodic boundary conditions using spatial node values, and (b) a Neumann boundary condition problem using sine series coefficients derived from fixed spatial sensors ("weather stations").
Candidate Generation:
Using the learned low-dimensional vector field, the authors employ harmonic balance methods to numerically compute Fourier expansions of periodic orbits. These solutions serve as "frozen" candidates. Crucially, the waveform, frequency, and coordinate conversions are fixed before the validation step to ensure the certificate applies specifically to the surrogate's output.
Rigorous Certification:
The fixed candidates are exported into the original high-dimensional system (the target PDE or DDE). The authors apply Newton–Kantorovich type contraction arguments in suitable Banach spaces (weighted sequence spaces for MG, trigonometric coefficient spaces for KS).
- The method constructs a bounded approximate inverse and bounds the residual and nonlinear terms.
- If specific inequalities (involving the residual bound $Y$ , linear defect $\alpha$ , and quadratic constant $Z_2$ ) are satisfied, a unique true periodic solution is guaranteed to exist within a specific radius of the surrogate candidate.
- This process is performed without applying iterative Newton corrections to the candidate within the target system for the initial certification, ensuring the error bound reflects the surrogate's raw accuracy.

Key Results

The paper validates this methodology on two infinite-dimensional systems:

1. Mackey–Glass Delay Differential Equation

Setup: A cubic polynomial surrogate was fitted to delay coordinates.
Outcome: A periodic candidate generated by the surrogate was certified to exist within a distance of $6.00 \times 10^{-4}$ of a true periodic solution of the original equation, with zero Newton corrections applied to the target system.
Refinement: Applying five Newton updates in the target system refined the solution to a radius of $1.2 \times 10^{-12}$ , confirming the initial candidate was a valid approximation.

2. Kuramoto–Sivashinsky (KS) Equation

Periodic Boundary Conditions: A quadratic surrogate based on 14 spatial nodes was assessed via forecasting and attractor projections. While these provided numerical evidence, the rigorous certification was performed on a separate Neumann problem.
Neumann Boundary Conditions (Exact Derivatives): A quadratic surrogate trained on exact Galerkin derivatives produced a candidate certified within $10^{-5}$ of a true periodic orbit without Newton corrections.
Neumann Boundary Conditions (Sensor Data): Surrogates were learned from time series at fixed spatial sensors (mimicking weather stations).
- Convergence: As the number of sensors ( $M$ ) increased from 12 to 24, the error of the candidates decreased geometrically, and all candidates converged to true orbits after 2–9 Newton corrections.
- Limitation: At $M > 24$ , the regression became unstable. The data lacked sufficient information to resolve the damping of fine modes, causing long-time rollouts to diverge even though short-term derivative errors remained low.
- Certification: For the sensor data, the uncorrected surrogate candidates were too far from the true orbits (e.g., a distance of 0.315 for the $M=24$ case) to be certified directly. Consequently, the certification followed a Newton-refined route: the candidate was refined using Newton iterations in the true equation, and then a solution was certified within a radius of $10^{-5}$ of this refined center. Unlike the exact-derivative case, this process did not quantify the accuracy of the original frozen surrogate export.

Significance and Claims

The paper makes the following modest but specific claims regarding its contribution:

Structural Fidelity Assessment: The primary contribution is a validation procedure that uses the success (or failure) of rigorous existence proofs to assess how well a learned surrogate reproduces specific invariant structures. This provides mathematically rigorous, local information about structural fidelity that short-term trajectory comparisons cannot provide.
Decoupling of Discovery and Verification: The method demonstrates that one can locate periodic solutions in a low-dimensional proxy and rigorously certify their existence in the high-dimensional target without needing to iteratively correct the candidate in the target system for the initial proof (as demonstrated in the MG and exact-derivative KS cases).
Diagnostic Power: The certification process acts as a diagnostic tool. It successfully identified a limitation in polynomial surrogates trained on sensor data: while they could approximate dynamics for low sensor counts, they failed to capture the damping of high-frequency modes necessary for stability at higher resolutions, a failure not revealed by standard derivative error metrics.
Scope Limitations: The authors explicitly state that these results do not establish global equivalence of dynamics, orbital stability, or agreement of long-time statistical properties. They do not claim superiority over other validation methods or a general population certification rate. The method is presented as a complementary tool to forecasting and numerical diagnostics, offering a specific, rigorous check on selected invariant sets.

In conclusion, the work demonstrates that combining data-driven regression with computer-assisted proofs allows for the rigorous validation of specific dynamical features in infinite-dimensional systems, providing a precise measure of where a learned model succeeds and where it fails structurally.