Concise $(\varepsilon,r)$-representations of a path

What the math gives to ML

The paper treats a sequential path as a truncated collection of local log-signatures, exposing a principled tradeoff between the number of intervals and the signature degree. Its transferable asset is an explicit memory-versus-approximation-error objective: local signatures reduce the factorial truncation error of iterated-integral expansions while avoiding the large storage cost of one high-degree global signature. This suggests replacing fixed-length temporal patching with an adaptive hybrid representation whose interval count and log-signature degree are selected from estimated path variation and a target downstream CDE error. The same representation can serve as compressed tokens for transformers, state-space models, or neural CDE encoders.