KARMAN

Low-rank approximation of Koopman operators for manifold-constrained dynamics

Funder: FIL
Role: Co-PI
Period: 2026–2029
Amount: €9k

Co-leaders

Anis Fradi

Partners

Stéphane Chrétien

The project aims to bridge the spectral theory of Koopman operators, kernel approximation on manifolds, and intrinsic dimensionality reduction in machine learning. The objective is to show that the complexity of certain Koopman approximations can be governed by the intrinsic dimension of the manifold supporting the dynamics, rather than by the ambient dimension of the observations. The project comprises a theoretical component devoted to deriving bounds on the approximate rank of dynamical kernels and projected operators, alongside an algorithmic component aimed at developing practical criteria for compression, dictionary selection, and rank estimation for EDMD and Kernel EDMD.