Sparsity-Promoting Algorithms for Informative Koopman Invariant Subspaces
This presentation explores a novel framework for discovering informative Koopman-invariant subspaces in nonlinear dynamical systems. The authors address the challenge of spurious modes in high-dimensional systems by introducing sparsity-promoting algorithms that extend Dynamic Mode Decomposition techniques. Through applications in fluid dynamics, including two-dimensional cylinder flow and three-dimensional ship air-wake simulations, the work demonstrates how these methods reveal multi-scale dynamics and improve identification of stable Koopman modes, offering significant advances for modal analysis and control in complex systems.Script
When you watch fluid swirl around a cylinder or ship, you're seeing nonlinear chaos. Yet hidden within that turbulence are linear patterns called Koopman modes, and finding the right ones has been like searching for signal in an ocean of noise.
Traditional Dynamic Mode Decomposition methods face a critical challenge. When analyzing complex flows in high-dimensional spaces, they produce spurious modes that mask the actual physics, making it nearly impossible to extract the patterns that matter.
The authors propose a fundamentally different approach: if you can't eliminate noise, starve it.
The framework treats Koopman subspace discovery as multi-task feature learning. By actively penalizing modes that deviate from linear evolution and applying hard thresholding to insignificant features, the algorithm forces the system to reveal only its most informative structures.
The difference is dramatic. Where traditional methods drown you in modes of questionable relevance, the sparsity-promoting approach distills the dynamics down to essential features, exposing mode clusters that reveal how the system behaves across multiple scales.
Consider these error trends from discrete-time kernel DMD applied to cylinder wake flow at Reynolds numbers 70, 100, and 130. The linearly evolving error Q measures how well modes stay linear over time, while reconstruction error R captures how accurately the sparse set reproduces the full dynamics. The sharp drops reveal where the algorithm locks onto genuinely informative subspaces, and notice how the error behavior shifts with flow regime.
The authors validated their framework on two demanding fluid scenarios. In cylinder flow, the method successfully isolated vortex shedding dynamics. In ship air-wake simulations, it captured intricate turbulent structures. Most remarkably, the discovered mode clusters revealed how the dynamics fundamentally shift as Reynolds number increases, exposing regime transitions traditional methods miss.
This figure shows the trade-off between sparsity and accuracy. Each curve represents a different Reynolds number, plotting reconstruction error against the number of non-zero terms retained. The blue circles mark the optimal sparsity level selected by the algorithm. Notice that as flow complexity increases with Reynolds number, the optimal representation requires more modes, but still far fewer than the full space.
This work transcends standard modal decomposition. By revealing nonlinear transient behavior and connecting sparsity promotion to rigorous optimization theory, the authors provide both practical tools and theoretical foundation for the next generation of data-driven flow analysis and control.
From chaos emerges structure, but only if you know which patterns to keep and which to discard. Visit EmergentMind.com to learn more and create your own research videos.
