Greedy Kernel Methods for Center Manifold Approximation

Bernard Haasdonk, Boumediene Hamzi, Gabriele Santin, Dominik WittwarLecture Notes in Computational Science and Engineering 2020 · DOI · OpenAlex

Abstract

For certain dynamical systems, stability analysis can be significantly simplified using center manifold theory. This theory isolates complex asymptotic behavior near non-hyperbolic equilibrium points, allowing for a reduced-dimensional analysis. However, since the manifold is typically unknown, approximation methods are essential. This paper presents a data-driven greedy kernel method to construct accurate manifold approximations. By using trajectories generated from high-accuracy numerical solvers, we build a surrogate model that satisfies theoretical constraints. Tests on various dynamical systems demonstrate the method's efficiency and high precision.

Introduction

In the study of complex dynamical systems—ranging from atmospheric models to chemical kinetics—researchers often encounter "non-hyperbolic" equilibrium points. At these points, standard linear stability analysis fails to provide a complete picture. Center manifold theory offers a solution by proving that the system's long-term behavior is confined to a lower-dimensional invariant manifold. Identifying this manifold allows a high-dimensional problem to be reduced to a much simpler, lower-dimensional one.

The primary challenge lies in the fact that finding the center manifold requires solving a quasilinear partial differential equation (PDE) that is often analytically intractable. Traditional Taylor series approximations are local and limited to the immediate vicinity of the equilibrium. This work introduces a data-based approximation approach using kernel methods, which provides greater flexibility and avoids the "curse of dimensionality" inherent in grid-based numerical methods.

Methodology

The proposed approach utilizes Kernel-based Approximation to develop a surrogate model of the manifold. The methodology incorporates several key technical innovations:

  • Matrix-Valued Kernels: To account for the vector-valued nature of the manifold mapping, matrix-valued kernels are used to capture inter-dependencies between system components.
  • Physical Constraints: The surrogate model is constrained to satisfy the mathematical properties of a center manifold: it must pass through the origin ($s(0)=0$) and be tangent to the center subspace ($Ds(0)=0$).
  • The P-Greedy Algorithm: To ensure computational efficiency, the authors employ the P-greedy method. This algorithm iteratively selects the most informative data points from a large simulation set, resulting in a sparse model that represents tens of thousands of observations using only a few dozen points.
  • Residual Minimization: Training involves minimizing a functional that balances the smoothness of the function within a Reproducing Kernel Hilbert Space (RKHS) against its fit to the observed trajectory data.

Experimental Results

The method was evaluated on synthetic datasets from 2D and 3D dynamical systems. A significant finding is that the training data does not need to lie perfectly on the center manifold; the kernel method effectively learns the manifold's structure from the asymptotic convergence of the trajectories.

Approximation of the center manifold components s1 and s2 showing near-perfect overlap between the model and the target.

As shown in the figures, the approximation is highly accurate. In Example 1, the greedy algorithm reduced over 38,000 data pairs to just 14 points while maintaining a pointwise residual—a measure of how well the surrogate satisfies the governing manifold PDE—on the order of $10^{-5}$ to $10^{-7}$.

Pointwise residuals r1 and r2 demonstrating the high precision of the kernel-based approximation.

Conclusion and Future Work

Greedy kernel methods provide a robust, efficient framework for center manifold approximation. By leveraging sparsity, the resulting models are extremely fast to evaluate, making them ideal for real-time stability analysis. Future work will focus on "data-driven splitting," where the coordinate transformation between center and stable subspaces is determined directly from data, further automating the model reduction process for high-dimensional engineering systems.

Summary updated

Abstract

Abstract For certain dynamical systems it is possible to significantly simplify the study of stability by means of the center manifold theory. This theory allows to isolate the complicated asymptotic behavior of the system close to a non-hyperbolic equilibrium point, and to obtain meaningful predictions of its behavior by analyzing a reduced dimensional problem. Since the manifold is usually not known, approximation methods are of great interest to obtain qualitative estimates. In this work, we use a data-based greedy kernel method to construct a suitable approximation of the manifold close to the equilibrium. The data are collected by repeated numerical simulation of the full system by means of a high-accuracy solver, which generates sets of discrete trajectories that are then used to construct a surrogate model of the manifold. The method is tested on different examples which show promising performance and good accuracy.
BibTeX
@inproceedings{haasdonk2020greedy,
  author = {Bernard Haasdonk and Boumediene Hamzi and Gabriele Santin and Dominik Wittwar},
  title = {{Greedy Kernel Methods for Center Manifold Approximation}},
  booktitle = {Lecture Notes in Computational Science and Engineering},
  year = {2020},
  doi = {10.1007/978-3-030-39647-3\_6},
}
Figure