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 Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-07-20 Page: https://delphi.science/r/arxiv/2607.17865 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…