Wasserstein asymptotics for empirical measures of diffusions on four dimensional closed manifolds

Dario Trevisan, Feng‐Yu Wang, Jie-Xiang Zhu

Abstract

We identify the leading term in the asymptotics of the quadratic Wasserstein distance between the invariant measure and empirical measures for diffusion processes on closed weighted four-dimensional Riemannian manifolds. Unlike results in lower dimensions, our analysis shows that this term depends solely on the Riemannian volume of the manifold, remaining unaffected by the potential and vector field in the diffusion generator.

Introduction

A fundamental problem in the study of stochastic processes is quantifying how quickly the time-average of a process—its empirical measure—converges to its long-term steady state, or invariant measure. For a diffusion process $X_t$ on a Riemannian manifold $M$, the empirical measure is defined as:

$$\mu_T := \frac{1}{T} \int_0^T \delta_{X_t} dt$$

As $T \to \infty$, $\mu_T$ converges almost surely to the invariant measure $\mu$. To quantify the distance between these measures, researchers employ the quadratic Wasserstein distance ($W_2$), a metric derived from optimal transport theory that represents the minimum "work" required to transform one distribution into another.

While convergence rates are well-established for dimensions $d \leq 3$, the four-dimensional case represents a "critical" threshold. In 4D, the convergence involves logarithmic factors, and until now, the exact constant of the leading-order term remained unknown. This research provides a precise identification of this constant for closed 4D manifolds.

The Criticality of Dimension Four

The behavior of the Wasserstein distance is deeply linked to the spectral properties of the diffusion generator $L$. In four dimensions, the Green's function—which describes the response of the system to a point source—exhibits a logarithmic divergence at the diagonal.

Mathematically, the problem involves the eigenvalues $\lambda_i$ and eigenfunctions $\phi_i$ of the operator. In 4D, the sum $\sum \lambda_i^{-2}$ diverges logarithmically. This research demonstrates that in this specific dimension, the divergent part of the sum is dominated entirely by the symmetric part of the operator (the Laplacian), while the non-symmetric components (the drift and vector fields) contribute only to lower-order, convergent terms.

Methodology

The authors utilize a combination of spectral geometry, stochastic analysis, and optimal transport to derive the asymptotic limit. Key components of the proof include:

  • Smoothing via Semigroups: Directly comparing the empirical measure (a collection of points) to the smooth invariant measure is analytically challenging. The authors apply a heat kernel smoothing technique to create a regularized measure $\mu_{T,\epsilon}$, allowing them to isolate the asymptotic behavior.
  • Dacorogna-Moser Interpolation: To bridge the gap between probability and geometry, the authors use Dacorogna-Moser interpolation. This technique allows the $W_2$ distance to be bounded by the $H^{-1}$ Sobolev norm of the density, turning an optimal transport problem into a more tractable partial differential equation (PDE) problem.
  • Coupling by Reflection: To ensure the results hold regardless of the starting position $x$ of the diffusion, the authors employ "coupling by reflection." This method proves that the influence of the initial distribution decays exponentially, allowing the "stationary" case (where the process starts in equilibrium) to generalize to any starting point.

Key Results: The "Volume-Only" Phenomenon

The central finding of the paper is a precise limit theorem. For a 4-dimensional closed manifold $M$, the expected squared $W_2$ distance behaves as:

$$\mathbb{E}^x [W_2^2(\mu_T, \mu)] \approx \frac{\text{vol}(M)}{8\pi^2} \frac{\log T}{T}$$

The authors provide an explicit error bound for this approximation:

$$\left| \frac{T}{\log T} \mathbb{E}^x [W_2^2(\mu_T, \mu)] - \frac{\text{vol}(M)}{8\pi^2} \right| \leq C \sqrt{\frac{\log \log T}{\log T}}$$

This result is striking because of its simplicity. The diffusion generator is typically defined as $L = \Delta + \nabla V \cdot \nabla + Z$, where $V$ is a potential and $Z$ is a vector field. In lower dimensions, $V$ and $Z$ significantly influence the convergence constant. However, in 4D, these terms become higher-order noise. The global convergence is dictated solely by the total Riemannian volume of the manifold.

Implications and Universality

This "blindness" to local drift suggests a form of universality in 4D diffusions. Regardless of the specific forces (the "pushes" and "pulls") acting on a particle locally, the rate at which its occupation measure fills the space is governed by the sheer geometric size of the manifold. This highlights the unique role of $d=4$ as the critical dimension where the geometry of the space fully eclipses the dynamics of the generator in the leading term.

Limitations and Future Directions

While this work resolves the 4D case for $p=2$, several questions remain:

  • Higher Dimensions: For $d \geq 5$, the existence of a similar limit is still an open conjecture. It is expected that these cases will involve more complex structures, such as Brownian interlacements.
  • Different Transport Costs: The current proof relies on the specific relationship between $W_2$ and the $H^{-1}$ norm. Extending this to $W_p$ for $p \neq 2$ remains a challenge.
  • Manifold Geometry: The study is restricted to closed (compact, no boundary) manifolds. Future work could investigate how boundaries or non-compactness (where mass can escape to infinity) alter these asymptotics.

Summary updated

Abstract

We identify the leading term in the asymptotics of the quadratic Wasserstein distance between the invariant measure and empirical measures for diffusion processes on closed weighted four-dimensional Riemannian manifolds. Unlike results in lower dimensions, our analysis shows that this term depends solely on the Riemannian volume of the manifold, remaining unaffected by the potential and vector field in the diffusion generator.
BibTeX
@inproceedings{trevisan2025wasserstein,
  author = {Dario Trevisan and Feng‐Yu Wang and Jie-Xiang Zhu},
  title = {{Wasserstein asymptotics for empirical measures of diffusions on four dimensional closed manifolds}},
  booktitle = {Electronic},
  year = {2025},
  doi = {10.1214/25-ecp719},
}
Figure