{"id":"arxiv/2607.17865","url":"https://arxiv.org/abs/2607.17865","tags":["math.ST","stat.ME","stat.ML"],"score":0,"title":"How Fast Do Signatures Learn? Statistical Theory and Applications for Path Regression","domain":"math.ST","source":"arxiv","authors":["Blanka Horvath","Wen Su","Wu Su","Binnan Wang","Ruixun Zhang"],"reviews":0,"subject":"math.ST","summary":"Many prediction and decision-making problems in operations research involve path-valued covariates -- data that evolve over time -- for which path signatures have become a canonical feature representation. Their use is justified by a universal approximation theorem, but this is an existence result: it guarantees that a finite-level signature can approximate any continuous path functional, without quantifying how fast the approximation error decreases as the truncation level grows. This paper develops approximation and statistical theory for signature-based path regression. We establish an…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-07-20","reactions":{}}