Gradient flow structure for McKean--Vlasov equations on discrete spaces

Matthias Erbar, Max Fathi, Vaios Laschos, André SchlichtingOpen 2016 · DOI · arXiv

Abstract

In this work, we show that a family of non-linear mean-field equations on discrete spaces can be viewed as a gradient flow of a natural free energy functional with respect to a certain metric structure we make explicit. We also prove that this gradient flow structure arises as the limit of the gradient flow structures of a natural sequence of $N$-particle dynamics, as $N$ goes to infinity.

Introduction

In physics and biology, modeling systems with large populations of interacting individuals—such as particles, cells, or agents—is a fundamental challenge. When the number of individuals becomes sufficiently large, researchers typically transition from tracking individual trajectories to using "mean-field" equations, like the McKean–Vlasov equation, to describe the evolution of the population's density.

In continuous spaces (e.g., $\mathbb{R}^d$), these equations are well-understood through the lens of optimal transport. They are recognized as "gradient flows," meaning the population density evolves by following the path of steepest descent on a free energy landscape relative to the Wasserstein metric. However, for discrete spaces—such as networks, graphs, or grids—this geometric interpretation was previously incomplete. This paper establishes the missing link, providing a unified geometric theory for nonlinear mean-field interactions on discrete sets and connecting microscopic particle behavior to macroscopic limits.

Establishing the Discrete Geometric Framework

To construct a gradient flow in a discrete setting, the authors introduce several sophisticated mathematical constructs that mirror continuous calculus:

  • The Transportation Metric ($\mathcal{W}$): The authors define a new distance metric on the space of probability measures. This metric serves as the discrete analogue to the $L^2$-Wasserstein distance, providing the necessary geometry to measure the "cost" of moving mass between discrete points.
  • Onsager Operators and Logarithmic Means: Defining a "gradient" on a discrete set requires a way to handle transition rates. The authors utilize an operator based on the logarithmic mean of these rates. This choice is essential; it allows the discrete system to mirror the properties of continuous diffusion equations, such as the Fokker-Planck equation, ensuring physical consistency.
  • Free Energy Functional: The evolution of the system is driven by the minimization of a free energy functional $\mathcal{F}$. The paper proves that the discrete McKean–Vlasov equation is literally the steepest descent path on this energy landscape.

The Mean-Field Limit and Evolutionary $\Gamma$-convergence

A central contribution of this research is proving that the geometric structure of the $N$-particle system converges to the macroscopic mean-field structure as $N \to \infty$.

Rather than simply proving that the solutions to the equations converge, the authors employ evolutionary $\Gamma$-convergence. This technique demonstrates that the entire mathematical architecture—the energy functional and the metric structure—converges. To manage the complexity of the $N$-particle limit, they "lift" the dynamics to a higher-dimensional space of random measures (the Liouville equation), making the analysis of the limit more tractable. This provides a rigorous geometric perspective on "propagation of chaos," the phenomenon where particles become statistically independent as their number grows to infinity.

Key Theoretical Results

The paper establishes several landmark results for discrete dynamics:

  1. Gradient Flow Identification (Theorem 2.13): The authors formally identify the discrete McKean–Vlasov equation as a "curve of maximal slope" for the free energy functional $\mathcal{F}$. This explains why certain energy functions act as Lyapunov functions: the system is geometrically designed to minimize them as efficiently as possible.
  2. Convergence of Structures (Theorem 3.10): They prove that as $N \to \infty$, the $N$-particle gradient flow structure converges to the McKean–Vlasov gradient flow structure, validating the mean-field approximation from a geometric standpoint.
  3. Metric Foundations: The authors prove that the new transportation distance $\mathcal{W}$ makes the space of probability measures a complete, separable, and geodesic metric space, offering a robust foundation for future research in discrete optimal transport.

Conclusion and Future Directions

By providing a rigorous geometric framework for discrete mean-field equations, this work opens several new avenues for study. A primary next step is the investigation of "Ricci curvature" bounds for these discrete systems. In continuous settings, curvature bounds provide insights into how quickly a system converges to equilibrium; finding a discrete equivalent would be a significant breakthrough. Additionally, applying this framework to specific models—such as chemical reaction networks or the mean-field Ising model—could yield new quantitative insights into the long-term behavior of complex discrete systems.

Summary updated

Abstract

In this work, we show that a family of non-linear mean-field equations on discrete spaces can be viewed as a gradient flow of a natural free energy functional with respect to a certain metric structure we make explicit. We also prove that this gradient flow structure arises as the limit of the gradient flow structures of a natural sequence of N-particle dynamics, as N goes to infinity.
BibTeX
@inproceedings{erbar2016gradient,
  author = {Matthias Erbar and Max Fathi and Vaios Laschos and André Schlichting},
  title = {{Gradient flow structure for McKean--Vlasov equations on discrete spaces}},
  booktitle = {Open},
  year = {2016},
  doi = {10.20347/wias.preprint.2219},
}

Loading related content...

Figure