{"id":"arxiv/2603.15091","url":"https://arxiv.org/abs/2603.15091","tags":["math.NA","cs.LG","math.DS","math.OC"],"score":0,"title":"Trustworthy Koopman Operator Learning: Invariance Diagnostics and Error Bounds","domain":"math.NA","source":"arxiv","authors":["Gustav Conradie","Nicolas Boullé","Jean-Christophe Loiseau","Steven L. Brunton","Matthew J. Colbrook"],"reviews":0,"subject":"math.NA","summary":"Koopman operator theory provides a global linear representation of nonlinear dynamics and underpins many data-driven methods. In practice, however, finite-dimensional feature spaces induced by a user-chosen dictionary are rarely invariant, so closure failures and projection errors lead to spurious eigenvalues, misleading Koopman modes, and overconfident forecasts. This paper addresses a central validation problem in data-driven Koopman methods: how to quantify invariance and projection errors for an arbitrary feature space using only snapshot data, and how to use these diagnostics to produce…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-03-16","reactions":{}}