📊 statistics

Quickest Detection with Rough Path Signatures

This paper introduces a framework for quickest distributional change detection in rough path signals by formulating the problem as an optimal stopping task where the optimal rule is a first hitting time of a linear functional of the path's signature, offering statistical guarantees and robust performance against adversarial perturbations, particularly in fractional Brownian settings.

Original authors: Mingrui Wang, Prakash Chakraborty

Published 2026-07-28

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Mingrui Wang, Prakash Chakraborty

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a security guard watching a live feed of a busy factory floor. Your job is to spot the exact moment a machine goes haywire. In the old days, you might have just watched the speedometer; if the needle jumped, you'd hit the alarm. But what if the machine doesn't just speed up? What if it starts vibrating in a weird, chaotic rhythm, or if the noise in the room is so messy and unpredictable that a simple speedometer is useless? This is the world of "rough paths"—signals that are too jagged, too irregular, and too dependent on their own history to be described by simple, smooth curves.

In the world of math and statistics, scientists have long tried to solve this "quickest detection" problem: how to raise an alarm as fast as possible without crying wolf too often. The classic tools work great for smooth, predictable signals (like a calm river), but they break down when the signal is a stormy, churning ocean. This is where a branch of math called "rough path theory" comes in. It treats these messy signals not as a single point, but as a complex shape that remembers every twist and turn it has ever taken. The key tool here is the "signature," which is like a unique fingerprint or a DNA sequence for the entire history of a signal's movement. It captures the essence of the path in a way that smooth tools can't.

Now, imagine you have a new, super-smart guard who doesn't just look at the speedometer. Instead, this guard looks at the entire "DNA fingerprint" of the machine's movement. A new paper by Mingrui Wang and Prakash Chakraborty proposes exactly this kind of guard. They built a framework that uses these "rough path signatures" to detect changes in signals that are messy, irregular, and full of long-term memory. They found that the best way to spot a change is to watch for when a specific mathematical "fingerprint" of the signal crosses a certain line. Their method works surprisingly well even when the signal is driven by chaotic, fractional noise (like the jagged movements of a stock market or a turbulent fluid) where old methods fail. Even better, they showed that this new guard can be trained to stay calm even if a sneaky hacker tries to nudge the signal just enough to trick the alarm, making the system robust against adversarial attacks.

The Story of the Signature Detective

So, how does this actually work? Let's break it down into a story.

The Problem: The Messy Signal
Imagine you are listening to a radio station. Before a certain moment (the "change-point"), the station is just playing static noise. After that moment, a song starts playing, but the song is mixed with a lot of static, and the static itself is weird—it's not just random white noise; it has a rhythm, a memory, and it's very "rough." In the past, mathematicians tried to solve this by assuming the noise was smooth and simple (like a gentle breeze). But in the real world, signals are often like a hurricane: they are jagged, they remember where they've been, and they don't follow simple rules. When the signal is this messy, the old "speedometer" methods (like CUSUM or Shiryaev rules) get confused. They can't tell if the signal is just being noisy or if it's actually changing.

The Solution: The Signature Fingerprint
The authors of this paper say, "Let's stop looking at the signal as a single point and start looking at its whole history." They use a mathematical tool called a signature. Think of a signature like a unique ID card for a path. If you walk through a park, your signature isn't just where you ended up; it's the story of every step you took, every turn you made, and how those turns interacted with each other. Even if two paths end up at the same spot, their signatures are different if they took different routes.

The paper proposes that instead of trying to guess the rules of the noise, we should just feed the entire history of the signal into a "signature machine." This machine converts the messy, jagged path into a list of numbers (the signature coefficients). The authors proved that the best way to detect a change is to watch for when a specific combination of these numbers crosses a threshold. It's like having a detector that listens to the shape of the signal's history rather than just its current speed.

The "Half-Space" Rule
The paper shows that the perfect rule for this detective is surprisingly simple: it's a "half-space hitting time." Imagine the signature of the signal is a point moving through a giant, multi-dimensional room. Before the change, this point wanders around in a safe zone. After the change, the point starts drifting toward a wall. The rule is simply: "Raise the alarm the moment the point touches the wall." The authors showed that this "wall" is just a flat plane (a half-space) defined by a linear function of the signature. This is a huge deal because it means you don't need to know the exact math of the noise to build the detector; you just need to find the right "wall" to watch.

Beating the Hackers
Here is where it gets really cool. In the real world, bad actors might try to trick your detector. They might add a tiny, invisible nudge to the signal to make it look like a change when there isn't one, or to hide a real change. The authors tested their method against these "adversarial" attacks. They created a scenario where an "adversary" tries to mess with the signal within a certain budget of energy. They found that by training their detector to expect these tricks (using a "minimax" approach, which means preparing for the worst-case scenario), the signature-based detector became incredibly tough. It didn't just survive the attacks; it actually performed better than the old methods in these messy, fractional environments.

The Results: Simulations and Surprises
The authors didn't just do the math on paper; they ran simulations to see if it worked in practice. They tested their method on two types of signals:

Brownian Motion: The classic, smooth-ish random walk (like a drunk person walking). Here, their new method performed just as well as the famous, established methods (CUSUM and Shiryaev).
Fractional Brownian Motion: The messy, rough, long-memory signal (like a stock market crash or turbulent wind). Here, the old methods struggled, but the signature-based method outperformed them significantly.

They also looked at what happens if you have multiple cameras (replications) watching the same event. They found that if you combine the data from many independent streams, the error rates (false alarms and missed detections) drop exponentially. It's like having 25 security guards instead of one; the chance of all of them missing the signal at the same time becomes tiny.

Why This Matters
The beauty of this approach is that it is "model-free." You don't need to know the exact equation of the noise. You don't need to assume the signal is smooth. You just feed the data into the signature, and the math finds the pattern. This makes it a powerful tool for the modern world, where signals are often messy, non-linear, and full of surprises. Whether it's detecting a fault in a power grid, spotting a cyber-attack, or monitoring financial markets, this method offers a way to see the change in the chaos that older tools miss.

In short, Wang and Chakraborty have built a new kind of alarm system. Instead of listening for a specific sound, it listens for a specific shape in the noise. And in a world full of rough, jagged, unpredictable signals, that might be the quickest way to catch a change before it's too late.

Technical Summary: Quickest Detection with Rough Path Signatures

Problem Formulation
The paper addresses the problem of quickest detection for distributional changes in signals modeled as rough paths. Unlike classical settings where observed processes are often Markovian or semimartingales, modern signals frequently exhibit non-Markovian dynamics, long memory, or high irregularity (e.g., fractional Brownian motion). In such regimes, the current state $X_t$ fails to summarize all relevant historical information, rendering finite-dimensional sufficient statistics (like those used in CUSUM or Shiryaev rules) unavailable. The authors model the observed signal $X$ as the concatenation of two independent geometric $p$ -rough paths, $X^{(1)}$ (pre-change) and $X^{(2)}$ (post-change), at an unknown random change-point $\theta$ . The objective is to find a stopping time $\tau$ adapted to the observation filtration that minimizes a cost function balancing detection delay and false alarm probability, formulated as an optimal stopping problem.

Methodology
The proposed framework replaces classical likelihood-ratio statistics with stopping rules derived from the rough path signature.

Rough Path Modeling: The pre- and post-change dynamics are defined as geometric $p$ -rough paths. The observed process is constructed by concatenating these paths at $\theta$ . The authors prove that this concatenated process retains the regularity of a geometric $p$ -rough path (Lemma 3.6) and that the homogeneous $p$ -variation distance between the observed path and the pre-change path exclusively captures the post-change divergence (Lemma 3.7).
Signature-Based Stopping Rules: Leveraging the density of signature-based functionals (Lemma 2.6), the authors show that any continuous stopping policy can be approximated by a linear functional of the truncated rough path signature. Consequently, the optimal stopping rule takes the form of the first hitting time of a half-space by a linear functional of the signature:
$\tau_l = \inf \left\{ t \ge 0 : |\langle l, \hat{X}^{\le N}_{0,t} \rangle| \ge 1 \right\} \wedge T$
where $\hat{X}$ is the time-augmented rough path and $l$ is a coefficient vector in the tensor algebra.
Geometric Detector: Independently of the payoff structure, the authors construct a detector based on the intrinsic geometry of the path (specifically the $p$ -variation distance). They demonstrate that this geometric detector shares the same structural form (a signature half-space hitting time) as the optimal stopping rule derived from the payoff minimization.
Robust Formulation: The framework is extended to a distributionally robust setting where the pre- and post-change laws belong to uncertainty classes (e.g., Wasserstein ambiguity sets or total-variation budgets for adversarial perturbations). Under the assumption of a least favorable model pair, the robust problem reduces to an ordinary optimal stopping problem under that specific pair, solvable via the same signature-based rule.
Numerical Implementation: The stopping coefficients $l$ are learned via zeroth-order stochastic approximation (specifically using the exponentially-shifted Gaussian smoothing estimator) over truncated signature coefficients, as the objective function is non-smooth due to the discrete threshold crossing.

Key Contributions

Optimal Stopping on Rough Path Space: The authors formalize quickest detection for non-Markovian, path-irregular dynamics by modeling the signal as concatenated rough paths. They establish that the optimal stopping rule is a signature half-space hitting time, extending classical theory to settings where Markovian reduction is impossible.
Geometric Detector Equivalence: They derive a detector directly from the path's intrinsic geometry (using $p$ -variation distance) and prove it possesses the same structural form as the optimal stopping rule, providing a geometric justification for the signature-based approach.
Statistical Guarantees: Explicit bounds on detection delay and false alarm probability are derived. Furthermore, the paper shows that when multiple independent replications share a common change-point, aggregating information reduces both delay and false alarm probability exponentially in the number of replications.
Adversarial Robustness: A minimax formulation is developed for uncertainty classes defined on rough path space, including adversarial perturbations with bounded total variation.
Numerical Validation: The rules are implemented and evaluated under Brownian and fractional Brownian dynamics.

Results

Brownian Setting: In the classical Brownian motion setting (where optimal CUSUM and Shiryaev rules are known), the signature-based rules achieve performance comparable to these optimal methods.
Fractional Brownian Setting: In the fractional Brownian motion setting (non-Markovian), where classical methods like CUSUM are not directly applicable or require parametric assumptions not present in the data, the signature-based rules outperform the Page-Hinkley baseline (a non-parametric CUSUM variant).
Robustness: Under adversarial path perturbations (modeled via total variation budgets), adversarial training produces stopping rules that significantly improve worst-case performance compared to nominally trained rules, maintaining robustness against path distortions while incurring minimal cost in nominal performance.
Parameter Sensitivity: The framework successfully navigates the trade-off between detection delay and false alarm probability by tuning loss function parameters, demonstrating flexibility across different change-point distributions (Exponential and Weibull).

Significance and Claims
The paper claims to provide a "probabilistically flexible and analytically rigorous framework" for quickest detection in complex, modern signal environments characterized by irregularity and non-Markovianity. Its primary significance lies in:

Universality: The use of the rough path signature as a "universal, model-free feature" that encodes the entire observed history, bypassing the need for explicit parametric models or Markovian assumptions required by classical sufficient statistics.
Theoretical Unification: It unifies classical sequential change-point detection with rough path theory, showing that the optimal stopping structure (hitting time of a half-space) persists even when the underlying state space is infinite-dimensional and non-Markovian.
Robustness: It offers a principled approach to distributionally robust detection in the presence of adversarial perturbations, a critical requirement for modern sequential decision-making where data-generating distributions are rarely known exactly.

The authors position their work as a necessary extension of classical theory to handle signals that "evade description by classical Markovian or semimartingale models," providing a tractable solution to intractable, path-dependent optimal stopping problems.

The Story of the Signature Detective

Technical Summary: Quickest Detection with Rough Path Signatures

More like this