Wasserstein asymptotics for the empirical measure of fractional Brownian motion on a flat torus

Martin Huesmann, Francesco Mattesini, Dario Trevisan

TL;DR

This paper establishes the convergence rates of the empirical measure of Fractional Brownian Motion (fBm) toward a uniform distribution on a flat torus using Wasserstein distances. The authors discover a "phase transition" in these rates at a critical dimension $d = 2 + 1/H$, drawing a profound parallel between the memory-heavy paths of fBm and the classical matching problems of independent points.

What Problem Does This Paper Address?

In statistics and probability, the "occupation measure" describes how much time a stochastic process spends in different parts of its state space. As time $T \to \infty$, we expect this measure to converge to a steady-state distribution (like the uniform Lebesgue measure on a torus).

While this convergence is well-understood for independent variables or Markov processes (where the future depends only on the present), it is much harder to quantify for processes with "long-range dependence." Fractional Brownian Motion (fBm) is the quintessential example of such a process; its increments are correlated over time, governed by the Hurst index ($H$). The authors tackle the challenge of finding precise asymptotic bounds for the Wasserstein distance—often called the "Earth Mover's Distance"—between the fBm's path and the uniform measure.

Methods

The authors employ a sophisticated blend of partial differential equations (PDEs) and probabilistic analysis:

  • The PDE Approach: Instead of calculating optimal transport maps directly, they bound the Wasserstein distance using negative Sobolev norms. They solve a Poisson equation ($-\Delta u_\epsilon = \mu_T - T$) where the source term is the difference between the empirical measure and the target measure.
  • Smoothing and Heat Semigroups: Because the empirical measure is singular (supported on a 1D path), they use the heat semigroup to "smooth" the measure, allowing for more stable analysis before taking the limit.
  • Fourier Analysis: The core estimates are performed in the frequency domain. By decomposing the measure into Fourier coefficients, the authors can analyze the correlations between different parts of the fBm trajectory.
  • Local Non-Determinism: To handle the non-Markovian nature of fBm, they rely on a property called "local non-determinism." This provides a lower bound on the covariance of the process increments, essentially proving that while the process has memory, it still explores new space with enough "innovation" to be analyzed.

Datasets

This is a theoretical mathematics paper. No empirical datasets were used. The results are derived through rigorous mathematical proof and apply to the mathematical model of Fractional Brownian Motion.

Key Insights and Ideas

  • The Phase Transition: The most striking finding is that the convergence rate changes fundamentally at the critical dimension $d = 2 + 1/H$.
  • In low dimensions, the rate is dominated by global fluctuations (similar to the Central Limit Theorem).
  • In high dimensions, the rate is limited by the "curse of dimensionality," where the "roughness" of the path (determined by $H$) cannot fill the space fast enough.
  • Heuristic Interpretation: The path of an fBm is roughly $1/H$-dimensional. The results suggest that the optimal way to transport mass from this "fractal" path to a uniform $d$-dimensional volume depends on how these two dimensions interact.
  • Connection to AKT: The results mirror the famous Ajtai-Komlós-Tusnády (AKT) theorem for i.i.d. points. This suggests that even though fBm has a complex memory structure, its large-scale spatial distribution eventually behaves like a collection of independent samples as $H \to 0$.

Key Results

The paper provides explicit asymptotic rates for the expected $p$-th order Wasserstein distance $E[W_p^p]$ as $T \to \infty$:

  • Subcritical ($d < 2 + 1/H$): The rate is $T^{-p/2}$, independent of the dimension.
  • Critical ($d = 2 + 1/H$): The rate includes a logarithmic correction: $(\log T / T)^{p/2}$.
  • Supercritical ($d > 2 + 1/H$): The rate becomes dimension-dependent: $T^{-p/(d - 1/H)}$.

The authors also extend these results to discrete-time approximations, showing how the sampling frequency ($\tau \sim T^{-\alpha}$) interacts with the continuous-time rates.

Limitations and Future Work

  • Compactness: The upper bounds are currently restricted to a flat torus ($\mathbb{T}^d$). Extending these sharp upper bounds to the full Euclidean space $\mathbb{R}^d$ remains an open challenge due to the lack of compactness.
  • The $L^\infty$ Case: The methods used (Sobolev norms) do not easily translate to the $W_\infty$ distance, which is closely related to the "cover time" problem (how long it takes for a process to visit every corner of the space).
  • Almost Sure Convergence: The results are presented in expectation. While the authors suspect "strong" (almost sure) convergence holds, proving it would require more intense concentration of measure arguments.

Summary updated

BibTeX
@inproceedings{huesmann2022wasserstein,
  author = {Martin Huesmann and Francesco Mattesini and Dario Trevisan},
  title = {{Wasserstein asymptotics for the empirical measure of fractional Brownian motion on a flat torus}},
  booktitle = {Stochastic},
  year = {2022},
  doi = {10.1016/j.spa.2022.09.008},
}
Figure