Generative model for rough volatility: log-signatures + a learned Besov-wavelet decoder reconstruct high-frequency texture via differentiable IDWT. Pluggable MLP/attention/transformer backbones, scale-weighted wavelet loss, and a 5-dataset multi-domain registry (fBM, rough Bergomi, Burgers turbulence, CHB-MIT EEG, ESC-50 audio).

This Lean 4 project to formalize the main results on Besov spaces for measure spaces with good grids, as defined in Smania D. Besov-ish spaces through atomic decomposition. Analysis & PDE Vol. 15 (2022), No. 1, 123-174. We already formalized 90% of the paper.