Quickest Detection with Rough Path Signatures

What the math gives to ML

The paper identifies a structurally simple detector for highly non-Markovian signals: compute a truncated rough-path signature online, apply one learned linear functional, and stop when the result hits a threshold. The transferable asset is Chen's identity, which makes signatures compositional under streaming concatenation, together with the theorem-level claim that the optimal stopping boundary is a half-space in signature space rather than an arbitrary nonlinear boundary. This suggests a compact recurrent front-end for neural change detection and event-triggered inference, with robustness obtained by optimizing the signature hyperplane against adversarial path perturbations rather than training a large sequence model.