Koopman Regularization

Restoration, generalization, and dimensionality reduction of a vector field from samples are the most common and crucial tasks in dynamical system analysis. An optimization-based algorithm to fulfill these tasks is suggested. Given noisy, sparse, or redundant sampled vector fields, the optimization process encapsulates the inherent geometry of the dynamical system derived from the Koopman eigenfunction space. The dynamic geometry is revealed via the exact penalty method, compromising accuracy and smoothness. This algorithm is backed up by promising results of denoising and generalization with a concise dynamics representation leading to dimensionality reduction.