In classical statistics, a probability distribution can often be characterized by its moments—mean, variance, skewness, and so on. If you know all the moments of a distribution, and if those moments do not grow too quickly, you can uniquely reconstruct the underlying distribution.
When moving from random variables to stochastic processes (paths moving through space and time), we face a similar challenge: how do we uniquely characterize and compare the probability distributions (or laws) of entire paths? The answer lies in rough path theory through a construct known as the path signature, and its probabilistic extension, the expected signature.
Recent mathematical developments have focused on understanding when and why these expected signatures uniquely identify a stochastic process, uncovering a fascinating structural breakdown when processes are stopped at random times.
The Path Analogue of a Moment Generating Function
The signature of a deterministic path is an infinite sequence of iterated integrals that captures its geometric essence. For a stochastic process, taking the expectation of each iterated integral yields the expected signature.
Just as the classical moment generating function packages all scalar moments into an analytic function, the expected signature acts as a non-commutative moment generating function for paths. In Signature Moments to Characterize Laws of Stochastic Processes, Ilya Chevyrev leverages this framework to show that "robust signature moments" can successfully characterize the laws of broad classes of stochastic processes, such as those with absolutely continuous or $p$-rough paths. This insight allows researchers to develop non-parametric tests to distinguish between different complex stochastic dynamics by directly comparing their signature moments.
For processes observed over a fixed, deterministic time horizon, a well-known sufficient condition guarantees that the expected signature uniquely determines the process's distribution: the expected signature must possess an infinite radius of convergence. This condition ensures that the path moments do not grow too fast, allowing the distribution to be uniquely recovered.
The Problem with Random Stopping Times
While fixed-time horizons are mathematically well-behaved, real-world stochastic phenomena are rarely so cooperative. We frequently care about processes stopped at a random time—for instance, when a particle hits a physical boundary or an asset price crosses a threshold.
When a process is stopped at a random time, the clean guarantees of the expected signature begin to fracture. The core issue is that the interaction between the spatial boundary and the path's trajectory can cause higher-order iterated integrals to grow much faster than they would over a fixed time interval.
This tension culminated in a breakthrough by Horatio Boedihardjo et al. in The expected signature of Brownian motion stopped on the boundary of a circle has finite radius of convergence. The authors investigated a deceptively simple system: a two-dimensional Brownian motion stopped the moment it exits a unit disk. They proved that the expected signature of this stopped process has a finite radius of convergence.
By demonstrating a finite radius of convergence, Horatio Boedihardjo et al. provided the first counterexample to the long-standing assumption that the expected signature universally maintains an infinite radius of convergence. This finding proved that an infinite radius of convergence is a sufficient, but not a necessary, condition for uniqueness, leaving the question of distribution identification for stopped processes wide open.
Universality of the Finite Radius
Is this finite radius of convergence an anomaly unique to circles and two dimensions? Subsequent research proved that it is actually a universal feature of stopped diffusion processes.
In Expected signature of stopped Brownian motion on d-dimensional C2,-domains has finite radius of convergence everywhere: 2 ≤ d ≤ 8, Hao Ni et al. expanded this boundary analysis significantly. By developing a novel "domain-averaging hyperbolic development" technique, the authors proved that for Brownian motion exiting any bounded, smooth ($C^2$) domain in dimensions 2 through 8, the expected signature invariably possesses a finite radius of convergence.
This universality highlights a fundamental truth about rough path descriptions of boundary-value problems: stopping a process based on spatial constraints fundamentally alters the analytic behavior of its path moments. While the expected signature remains a well-behaved and structured object, it does not automatically satisfy the classical conditions required for trivial uniqueness proofs.
Scaling Signatures to Practical Applications
While theorists map out the boundaries of uniqueness and convergence, applied researchers face a parallel hurdle: the dimensionality of the path signature explodes exponentially with the path's dimension and the depth of the iterated integrals. To use these representations for learning complex dynamics, such as Stochastic Differential Equations (SDEs), approximations are mandatory.
To bridge this gap, Enea Monzio Compagnoni et al. developed a computationally efficient alternative in Empirics on the expressiveness of Randomized Signature. Instead of computing the full, high-dimensional tensor series, the authors utilize randomized projections to map the path integrals into a lower-dimensional state space.
This randomized signature retains the expressiveness and underlying mathematical guarantees of the full signature while drastically reducing the computational footprint. It demonstrates that even when the theoretical boundaries of expected signatures are still being mapped out—such as extending convergence guarantees to broader classes of stopped SDEs—the empirical expressiveness of signature-based features remains a powerful tool for characterizing stochastic systems in practice.
Key Foundational Work
The expected signature of Brownian motion stopped on the boundary of a circle has finite radius of convergence (2020) — contributed 8 concept elements to this direction.
Empirics on the expressiveness of Randomized Signature (2021) — contributed 1 concept elements to this direction.