Abstract
The signature of a $d$-dimensional Brownian motion is a sequence of iterated Stratonovich integrals along the Brownian paths, an object taking values in the tensor algebra over $\mathbb{R}^{d}$. In this note, we derive the exact rate of convergence for the expected signatures of piecewise linear approximations to Brownian motion. The computation is based on the identification of the set of words whose coefficients are of the leading order, and the convergence is concentrated on this subset of words. Moreover, under the choice of projective tensor norm, we give the explicit value of the leading term constant.
Introduction
In stochastic analysis, the signature of a path serves as a powerful "fingerprint," capturing essential geometric data through an infinite series of iterated integrals. For a $d$-dimensional path, the signature is a sequence of tensors that provides a basis for the space of continuous functions on paths, making it a vital tool for solving stochastic differential equations and as a feature set in machine learning.
However, Brownian motion is inherently fractal-like and infinitely irregular. In practical applications, we cannot compute its signature directly; instead, we rely on approximations, such as piecewise linear paths constructed from discrete samples. While it is well-known that these approximations eventually converge to the true Brownian motion, a precise "speed limit" for the convergence of the expected signature was previously missing. This paper fills that gap, providing a rigorous convergence rate and an explicit constant that quantifies the error in numerical simulations.
Methodology
The authors employ techniques from Rough Path Theory and algebraic combinatorics to analyze the convergence behavior. The study focuses on three key methodological pillars:
1. Stratonovich Integrals
The signature is defined using Stratonovich iterated integrals. This choice is significant because Stratonovich calculus obeys the standard chain rule, making it more compatible with the algebraic structures of the tensor algebra than Itô calculus in this specific context.
2. The Projective Tensor Norm
A defining feature of this work is the use of the projective tensor norm. While many functional analysis papers utilize the Hilbert-Schmidt (Frobenius) norm, the authors demonstrate that the projective norm—which, in this context, is the sum of the absolute values of the tensor's coefficients—is uniquely suited for this problem. This choice allows for the calculation of an exact constant rather than a mere upper bound, revealing the intrinsic nature of the Brownian signature's convergence.
3. Word Analysis and Concentration
The signature is composed of "words" (sequences of indices). The authors categorize these words based on how their coefficients behave as the approximation mesh size $1/M$ (where $M$ is the number of segments) shrinks. They identify a specific subset of words that dominate the error term, showing that the error is not spread evenly but is concentrated on a small subset of words of length $2n$. Most other terms are negligible, vanishing at a faster rate of $O(1/M^2)$.
Key Insights
- Symmetry and Vanishing Terms: Due to the inherent symmetry of Brownian motion, the authors confirm that all odd-order expected signatures are zero for both the true process and its piecewise linear approximation.
- Independence of Increments: The proof leverages the fact that Brownian motion has independent increments. This allows the expected signature of the entire path to be decomposed into a product of the expected signatures of individual linear segments, simplifying the global analysis into a local one.
- Error Concentration: The discovery that the error concentrates on specific words allows for a much more precise asymptotic analysis than previous general bounds provided.
Main Results: The Convergence Rate
The central contribution of the paper is a precise limit theorem. Let $B^{(M)}$ be the piecewise linear approximation of Brownian motion with $M$ segments over time $T$, and let $\phi(T)$ be the true expected signature. The authors establish:
- Linear Convergence: The difference between the true expected signature and the approximation at the $2n$-th level vanishes at a rate of $1/M$.
- The Explicit Constant: Under the projective tensor norm, the authors prove the following limit: $$\lim_{M \to +\infty} \frac{M}{T} \|\pi_{2n}(\phi(T)) - \pi_{2n}(\phi^M(T))\| = \frac{d - 1}{3 \cdot (n - 2)!} \left( \frac{dT}{2} \right)^{n-1}$$
This formula provides a concrete way to calculate the expected error in approximations based on the dimension $d$, the time $T$, and the level of the signature $n$.
Limitations and Future Directions
While this paper provides a definitive rate for piecewise linear approximations, several avenues for future research remain:
- Approximation Schemes: The results are specific to piecewise linear interpolation. Other schemes, such as geodesic paths on manifolds, might yield different constants or rates.
- Norm Sensitivity: The explicit constant depends on the projective tensor norm. While the $1/M$ rate is likely universal, the constant would change if a different norm (such as the $l_2$ norm) were employed.
- Generalizing the Process: The current work focuses on standard Brownian motion. Extending these exact convergence rates to processes with memory, such as Fractional Brownian Motion, or processes with jumps, like Lévy processes, remains an open challenge.
Summary updated