On Matching, and Even Rectifying, Dynamical Systems through Koopman Operator Eigenfunctions

Erik M. Bollt, Qianxiao Li, Felix Dietrich, Ioannis G. KevrekidisSIAM J. Appl. Dyn. Syst 2018 · OpenReview

Abstract

This paper presents a novel data-driven framework, EDMD-M, for identifying topological conjugacies between dynamical systems by matching their Koopman operator eigenfunctions. By leveraging the linear properties of the Koopman operator, the authors demonstrate how complex nonlinear systems can be "rectified" into simpler forms, such as linear systems or constant flows. The method is validated on several synthetic systems, including the Van der Pol oscillator, showing that neural-network-based dictionaries can effectively recover these transformations directly from data.

Introduction

A fundamental challenge in the study of dynamical systems is classification: determining if two seemingly distinct systems are topologically conjugate. Two systems are conjugate if there exists a mapping $h$ that translates the motion of one into the other. While the existence of such mappings is theoretically guaranteed for many smooth flows (e.g., the flow box theorem), constructing them explicitly for nonlinear systems is notoriously difficult. This work addresses this gap by utilizing the spectral properties of the Koopman operator to discover these transformations directly from data.

Methodology

The core of the proposed approach is the Koopman operator framework, which shifts the focus from state-space trajectories to the evolution of functions (observables) defined on the state space. Unlike the underlying nonlinear dynamics, the Koopman operator is linear, making it amenable to spectral analysis.

Spectral Matching and EDMD-M

The authors prove that if two systems are conjugate, their Koopman eigenfunctions are related via the conjugacy mapping $h$. Specifically, for eigenfunctions $g^{(1)}$ and $g^{(2)}$ associated with the same eigenvalue, the relationship $g^{(1)}(x) = g^{(2)}(h(x))$ holds. To find these mappings numerically, the authors introduce EDMD-M (Matching), an extension of Extended Dynamic Mode Decomposition. By using a shared dictionary of functions—often implemented via multi-layer neural networks—the algorithm approximates the Koopman operator and its eigenfunctions for both systems simultaneously.

Visualizing the mapping of state trajectories to Koopman eigenfunctions. The left panel shows the evolution on a manifold, while the right panel illustrates the simplified functional evolution.

Rectification

A key application of this method is "rectification," where a nonlinear system is matched to a target system with a constant vector field (e.g., $\dot{z} = [1, 0, \dots, 0]$). This process effectively "straightens" complex trajectories into linear flows in the transformed coordinate space, providing a systematic way to discover natural coordinates for complex physics.

Experimental Evaluation

The EDMD-M framework was evaluated on several benchmark systems:

  • 1D and 2D Quadratic Systems: Used to demonstrate the derivation of closed-form mappings and the handling of coordinate transformations in higher dimensions.
  • Van der Pol Oscillator: A classic nonlinear system with a limit cycle. This served as a testbed for the neural-network-based dictionary approach.

Results and Discussion

The method successfully recovered analytical mappings for simple quadratic systems, transforming them into linear diagonal forms. In the more complex Van der Pol oscillator case, the neural-network-based EDMD-M recovered the nonlinear transformation with a relative error of approximately 5% at initial conditions, with errors decaying as the system evolved.

One significant challenge identified is the multiplicity problem in higher dimensions, where a single eigenvalue may correspond to multiple eigenfunctions. The authors addressed this by forcing both systems into a shared functional subspace, which "pins" the eigenfunctions and makes the mapping $h$ recoverable.

Conclusion

This work provides a robust, data-driven methodology for discovering conjugacies between dynamical systems. By matching Koopman eigenfunctions, researchers can systematically "rectify" nonlinear dynamics, transforming them into simpler, more interpretable forms. Future work will explore the extension of this method to systems with continuous spectra and the handling of global singularities where local "flow box" mappings typically fail.

Summary updated

BibTeX
@article{bollt2018matching,
  author = {Erik M. Bollt and Qianxiao Li and Felix Dietrich and Ioannis G. Kevrekidis},
  title = {{On Matching, and Even Rectifying, Dynamical Systems through Koopman Operator Eigenfunctions}},
  journal = {SIAM J. Appl. Dyn. Syst.},
  year = {2018},
  url = {https://openreview.net/forum?id=ba55dAb4ZQ},
}
Figure