{"id":"arxiv/2610.39580","url":"https://arxiv.org/abs/2610.39580","tags":["math.DS","cs.LG"],"score":0,"title":"Learning Kolmogorov Backward Operators for Spectral Analysis and Forecasting: A Data-Driven Variational Galerkin Framework","domain":"math.DS","source":"arxiv","authors":["Yuanchao Xu","Claire Valva","Ben Gao","Isao Ishikawa"],"reviews":0,"subject":"math.DS","summary":"For a time-homogeneous Markov process, the fixed-lag solution operator \\( (T_τf)(x)=\\mathbb E[f(X_{t+τ})\\mid X_t=x] \\) of the Kolmogorov backward equation evolves observables linearly and deterministically, even when the dynamics are stochastic. We approximate this operator using a two-stage method where we first approximate this operator directly from data, without knowledge of the governing equation (e.g., Fourier Neural Operator), and then compress the Galerkin projection onto a finite-dimensional observable space that is fixed in advance or selected variationally. The error analysis…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-29","reactions":{}}