Enhancing spectral analysis in nonlinear dynamics with pseudoeigenfunctions from continuous spectra

Itsushi Sakata, Yoshinobu KawaharaScientific 2024 · DOI

TL;DR

This paper introduces a novel clustering-based method to analyze complex dynamical systems with continuous spectra, a challenge for traditional Dynamic Mode Decomposition (DMD). By leveraging Residual Dynamic Mode Decomposition (ResDMD) to extract pseudoeigenfunctions and then clustering them, the authors reveal hidden dynamic patterns in noisy and chaotic systems, offering a more nuanced understanding than standard DMD.

What Problem Does This Paper Address?

Analyzing the behavior of complex dynamical systems, especially those exhibiting chaos or influenced by noise, is a significant challenge. Dynamic Mode Decomposition (DMD) is a powerful tool for uncovering spectral features of these systems, but it struggles with continuous spectra, which are common in chaotic and noisy data. These continuous spectra can obscure important dynamic patterns, making it difficult to fully understand the system's behavior. The authors aim to bridge this gap by developing a method that can effectively analyze and characterize dynamics represented by these continuous spectra.

Methods

The core of the proposed method lies in two key components:

  • Residual Dynamic Mode Decomposition (ResDMD): This technique is used to approximate the Koopman operator and, crucially, to compute pseudospectra and pseudoeigenfunctions. Unlike standard DMD, ResDMD is designed to handle continuous spectra and avoid spectral pollution (spurious eigenvalues). It provides a more robust way to extract spectral information from data, especially in the presence of noise.
  • Clustering Pseudoeigenfunctions: Once pseudoeigenfunctions are obtained from ResDMD, the authors employ a clustering approach to group similar ones. This involves:
  • Kernel Principal Angle: Used to compute a similarity matrix between the pseudoeigenfunctions, effectively comparing the subspaces they span.
  • Spectral Clustering: A graph-based clustering technique that uses the eigenvectors of a normalized graph Laplacian derived from the similarity matrix. This maps the data into a lower-dimensional space where clustering is performed.
  • Fuzzy C-Means: Used for soft clustering, allowing data points to belong to multiple clusters with varying degrees of membership. This is particularly useful for handling the inherent ambiguity in continuous spectra.

The authors also compare their method against standard DMD and a measure-preserving variant called mpEDMD.

Datasets

The paper evaluates the proposed method using two types of synthetic datasets:

  1. 1D Oscillators with Thermal Noise: A synthetic signal generated from two coupled harmonic oscillators with added thermal noise. This dataset is used to test the method's ability to separate underlying signals and handle noise.
  2. Coupled Hénon Maps: Time series data from two unidirectionally coupled Hénon maps, used to investigate generalized synchronization (GS) in chaotic systems. This dataset allows for the analysis of complex, aperiodic dynamics.

Key Insights and Ideas

  • Continuous Spectra are Key: The paper strongly emphasizes that continuous spectra are not just noise but can represent genuine dynamic properties, especially in chaotic and noisy systems.
  • ResDMD as a Superior Tool: ResDMD proves more effective than standard DMD and mpEDMD in capturing and separating spectral features, particularly in the presence of noise and continuous spectra.
  • Clustering Reveals Hidden Dynamics: Grouping pseudoeigenfunctions allows for the identification of distinct dynamic behaviors that are obscured by traditional spectral analysis methods.
  • Generalized Synchronization Detection: The method shows promise in detecting and characterizing generalized synchronization in coupled chaotic systems by analyzing the clustering of pseudoeigenfunctions.
  • Beyond Eigenvalues: The work highlights the limitations of relying solely on eigenvalues and demonstrates the value of analyzing the broader spectral properties captured by pseudoeigenfunctions.

Key Results

  • 1D Oscillator Analysis: The proposed method successfully decomposed a noisy 1D signal into its original components, a task that proved challenging for standard DMD and mpEDMD. ResDMD's pseudoeigenfunctions allowed for clear separation of the two underlying signals and the noise component.
  • Coupled Hénon Map Analysis: The clustering of pseudoeigenfunctions from coupled Hénon maps revealed distinct states corresponding to the presence or absence of generalized synchronization. The analysis showed that coupling leads to a convergence of attractors and increased correlation between variables, consistent with GS.
  • Improved Reconstruction: The reconstruction of the original signals from clustered modes showed significantly lower residuals for the ResDMD-based method compared to DMD and mpEDMD, indicating a more accurate representation of the underlying dynamics.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Measure-Preserving Assumption: The current ResDMD implementation requires the dynamical system to be measure-preserving for computing smoothed approximate spectral measures.
  • Optimal Cluster Number: Determining the optimal number of clusters remains a challenge, with the current approach relying on heuristics.
  • Time Series Length: Sufficiently long time series are needed to ensure measure-preserving behavior and accurate analysis.
  • Extending to Higher Dimensions: The method's applicability to larger state-space dimensions needs further investigation.
  • Kernel Methods: Combining ResDMD with kernel methods is proposed as a promising avenue for handling high-dimensional systems and implicitly computing similarities in feature spaces.
  • Rigged DMD Integration: Applying the clustering methodology to generalized eigenfunctions computed by Rigged DMD is suggested for more robust analysis, especially for measure-preserving systems.

Summary updated

Abstract

The analysis of complex behavior in empirical data poses significant challenges in various scientific and engineering disciplines. Dynamic Mode Decomposition (DMD) is a widely used method to reveal the spectral features of nonlinear dynamical systems without prior knowledge. However, because of its infinite dimensions, analyzing the continuous spectrum resulting from chaos and noise is problematic. We propose a clustering-based method to analyze dynamics represented by pseudoeigenfunctions associated with continuous spectra. This paper describes data-driven algorithms for comparing pseudoeigenfunctions using subspaces. We used the recently proposed Residual Dynamic Mode Decomposition (ResDMD) to approximate spectral properties from the data. To validate the effectiveness of our method, we analyzed 1D signal data affected by thermal noise and 2D-time series of coupled chaotic systems exhibiting generalized synchronization. The results reveal dynamic patterns previously obscured by conventional DMD analyses and provide insights into coupled chaos's complexities.
BibTeX
@inproceedings{sakata2024enhancing,
  author = {Itsushi Sakata and Yoshinobu Kawahara},
  title = {{Enhancing spectral analysis in nonlinear dynamics with pseudoeigenfunctions from continuous spectra}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-69837-y},
}
Figure