Spatiotemporal pattern extraction by spectral analysis of vector-valued observables (Q2327837)

This paper is a self-contained analysis on a data-driven framework for extracting complex spatio-temporal patterns generated by ergodic dynamical systems. More precisely, the authors present a Vector-valued Spectral Analysis (VSA) for spatio-temporal pattern extraction using operator-valued kernels. The main purpose of VSA is to construct a decomposition of the observation map \(\vec{F}\) via an expansion of the form \(\vec{F}\approx \sum_{j=0}^{l-1}\vec{F}_{j}\), \(\vec{F}_j=c_{j}\vec{\phi}_{j}\), where \(c_{j}\) and \(\vec{\phi}_{j}\) are real-valued coefficients and vector-valued observables. The authors describe operator-valued kernel, vector-valued eigenfunctions, operator-valued kernels with delay-coordinate maps and Markov normalization. The authors also give some general observations on the topological structure of spatio-temporal data in a delay-coordinate space. The application of VSA to the Kuramoto-Sivashinsky model in periodic and chaotic regimes demonstrates a significant performance.