Unverified 2026

Signature Half-Space Change Detector

Implementation & benchmark of arXiv:2607.22958 — Quickest Detection with Rough Path Signatures

Usefulness6/10

Difficulty5/10

Novelty7/10

Architecture Accuracy Memory Stability Rnn Streaming Loss Inference speedup Rough path Tensor decomposition Stochastic processes Geometry

Source paper: Quickest Detection with Rough Path Signatures arXiv:2607.22958 ⓘ · analyzed Aug 30, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Replace a recurrent sequence classifier's unconstrained hidden-state alarm head with an online truncated-signature state and a first-hitting-time linear detector. The module summarizes local order information and cross-channel interactions while preserving exact compositional updates, making it suitable for long streaming sequences and early-exit decisions.

Formulas

$$\mathbb{X}_{0,t}=\mathbbm{1}_{\{t<\theta\}}\mathbb{X}^{(1)}_{0,t}+\mathbbm{1}_{\{t\geq\theta\}}\left(\mathbb{X}^{(1)}_{0,\theta}\otimes\mathbb{X}^{(2)}_{\theta,t}\right).$$

$$\pi_n(\mathbb{X}_{0,t})=\sum_{k=0}^{n}\pi_k(\mathbb{X}^{(1)}_{0,\theta})\otimes\pi_{n-k}(\mathbb{X}^{(2)}_{\theta,t}),\qquad 0\leq n\leq N.$$

$$S^{\leq N}(x\star y)=S^{\leq N}(x)\otimes S^{\leq N}(y),\qquad S^n(x)_{a,b}=\int_{a<u_1<\cdots<u_n<b}dx_{u_1}\otimes\cdots\otimes dx_{u_n}.$$

$$\tau=\inf\{t\geq 0:\ell^\top z_t\geq h\},\qquad z_t=\operatorname{vec}(S^{\leq N}(x)_{0,t}),\quad \ell\in\mathbb{R}^{D_N},\ h\in\mathbb{R}.$$

Mathematical statement

The observed rough path is the concatenation of a pre-change path $\mathbb{X}^{(1)}$ and a post-change path $\mathbb{X}^{(2)}$ at an unknown change time $\theta$: $\mathbb{X}_{0,t}=\mathbf{1}_{\{t<\theta\}}\mathbb{X}^{(1)}_{0,t}+\mathbf{1}_{\{t\geq\theta\}}(\mathbb{X}^{(1)}_{0,\theta}\otimes\mathbb{X}^{(2)}_{\theta,t})$. For a path $x:[a,b]\to\mathbb{R}^d$, its level-$n$ signature is the tensor of iterated integrals $S^n(x)_{a,b}=\int_{a<u_1<\cdots<u_n<b}dx_{u_1}\otimes\cdots\otimes dx_{u_n}$, and the truncated signature is $S^{\leq N}(x)=(1,S^1(x),\ldots,S^N(x))$. Chen's identity gives $S^{\leq N}(x\star y)=S^{\leq N}(x)\otimes S^{\leq N}(y)$, with level $n$ component $\sum_{k=0}^n S^k(x)\otimes S^{n-k}(y)$. The paper's transferable stopping structure is a half-space hitting rule: for signature feature vector $z_t=\operatorname{vec}(S^{\leq N}(x)_{0,t})$, alarm at $\tau=\inf\{t:\ell^\top z_t\geq h\}$, where $\ell$ is a learned linear functional and $h$ is a threshold. The neural adaptation learns $\ell$ and $h$ from labeled pre-change and post-change windows while retaining the mathematically compositional signature state.

Implementation notes

(1) Integration point: add this as a streaming change or event detector before an expensive RNN, transformer, or world-model update. Inputs are sampled multichannel increments $\Delta x_t\in\mathbb{R}^d$; use piecewise-linear interpolation and a truncated signature depth $N=2$ or $3$. The detector maintains the signature of the complete prefix and emits an alarm or early-exit flag.

(2) Pseudocode: initialize $Z=[1,0,\ldots,0]$, where $Z_n$ stores the level-$n$ tensor. For every increment $u=\Delta x_t$, compute the increment signature $U=(1,u,u\otimes u/2!,\ldots,u^{\otimes N}/N!)$ for a linear segment. Update $Z\leftarrow Z\otimes U$ using truncated tensor multiplication: $Z_n\leftarrow\sum_{k=0}^n Z_k\otimes U_{n-k}$. Flatten $Z$ to $z_t$, compute $s_t=\ell^\top z_t$, and stop when $s_t\geq h$. Optionally reset $Z$ after each alarm or maintain a sliding-window signature using an approximate inverse signature.

(3) Computed from the paper's mathematics: the tensor-product update, iterated-integral feature layout, and first-hitting rule are fixed. Estimated empirically: normalize each channel, choose depth $N$, fit $\ell$ with logistic loss or a ranking loss on pre-change versus post-change prefixes, and select $h$ on validation pre-change streams to satisfy a target false-alarm rate. For robustness, augment training with bounded perturbations $\|\delta x_t\|_2\leq\varepsilon$ and minimize the worst-case post-change score or maximize the negative-versus-positive margin under perturbations.

(4) First experiment: use synthetic Brownian and fractional-Brownian streams with a known drift, volatility, or Hurst-exponent change. Compare against CUSUM, a GRU alarm head, and a transformer classifier at equal feature-update FLOPs. Measure detection delay at fixed false-alarm probability, memory per stream, and performance under additive bounded path perturbations. Success means lower delay than CUSUM or GRU on fractional-Brownian data, stable false-alarm calibration, and substantially lower recurrent-state memory than the transformer.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.