Abstract
We study the Fokker-Planck equation as the hydrodynamic limit of a stochastic particle system on one hand and as a Wasserstein gradient flow on the other. We write the rate functional, that characterizes the large deviations from the hydrodynamic limit, in a way that the free energy appears explicitly. Next we use this formulation via the contraction principle to prove that the discreet time rate functional is asymptotically equivalent in the Gamma-convergence sense to the functional derived from the Wasserstein gradient discretization scheme.
Introduction
In 1998, a landmark paper by Jordan, Kinderlehrer, and Otto (JKO) established that the Fokker-Planck equation—which describes the evolution of a probability density under diffusion and a potential—can be viewed as a "gradient flow." Specifically, it represents the steepest descent of the free energy functional in the space of probability measures equipped with the Wasserstein metric ($W_2$).
While the JKO scheme is a powerful tool in analysis and optimal transport, a fundamental question remained: Why the Wasserstein metric? While it works mathematically, its physical origin in this context was not fully transparent. This paper addresses that gap by linking the macroscopic gradient flow structure to the microscopic behavior of many independent random particles. It demonstrates that the $W_2$ metric is not just a clever choice of geometry, but a direct consequence of how large systems of particles deviate from their average behavior, known as the thermodynamic limit.
Theoretical Framework
Large Deviation Principle (LDP)
The authors begin with a system of $N$ independent random particles. As $N \to \infty$, the empirical measure of these particles converges to the solution of the Fokker-Planck equation. The Large Deviation Principle (LDP) is used to quantify the probability of "rare events" where the system deviates from this limit. The cost of these deviations over a small time step $\tau$ is measured by the "rate functional" $J_\tau$.
Gamma-convergence
To bridge the gap between discrete-time stochastic dynamics and the continuous gradient flow, the authors employ Gamma-convergence. This functional analysis tool allows them to show that as the time step $\tau$ approaches zero, the sequence of rate functionals $J_\tau$ converges to the functional used in the JKO scheme. Crucially, Gamma-convergence ensures that the minimizers of the approximate functionals (the most likely paths of the particles) converge to the minimizer of the limit functional (the solution to the Fokker-Planck equation).
Key Insights: The Probabilistic Origin of JKO
The central contribution of this work is the demonstration that the rate functional $J_\tau$ from probability theory and the JKO functional from geometry are essentially two sides of the same coin. For small $\tau$, the authors establish the following approximation:
$$J_\tau(\rho|\rho_0) \approx \frac{1}{2}\mathcal{F}(\rho) - \frac{1}{2}\mathcal{F}(\rho_0) + \frac{1}{4\tau}W_2^2(\rho_0, \rho)$$
This formulation contains the exact terms found in the JKO variational scheme. By using the "contraction principle" from large deviation theory, the authors successfully rewrite the rate functional in a way that the Helmholtz free energy $\mathcal{F}$ and Fisher information appear explicitly. This provides a rigorous link between the stochastic dynamics of individual particles and classical thermodynamics.
Main Results
- Theorem 1.1: The authors rigorously prove the Gamma-convergence of the modified rate functional to the free energy difference as $\tau \to 0$.
- Asymptotic Development: They establish that the rate of convergence of the empirical measure is characterized by the Wasserstein distance. Specifically, they show that $\tau J_\tau$ Gamma-converges to $\frac{1}{4}W_2^2$, providing a probabilistic justification for the use of the $W_2$ metric in gradient flows.
- Generality of Potentials: The results are proven for a wide class of potentials $\Psi$, categorized into "subquadratic" (growing slower than $x^2$) and "superquadratic" (growing faster than $x^2$) cases.
Limitations and Future Directions
Despite the rigorous foundation provided, some constraints remain:
- Dimensionality: While the lower bound of the Gamma-convergence holds in any dimension, the proof for the recovery sequence (the upper bound) is currently restricted to one dimension ($d=1$). This is due to the reliance on specific properties of optimal transport maps (equal tails) that are unique to the real line.
- Potential Assumptions: The proof requires specific technical assumptions regarding the convexity and smoothness of the potential $\Psi$. Extending this to more singular or non-convex potentials remains an open challenge.
- Interacting Particles: This study focuses on independent particles. A natural progression would be extending this framework to systems with mean-field interactions, where particles influence one another's trajectories.
Summary updated