[Paper Review] Hierarchies, entropy, and quantitative propagation of chaos for mean field diffusions

Daniel Lacker|arXiv (Cornell University)|May 6, 2021

Advanced Thermodynamics and Statistical Mechanics4 citations

TL;DR

This paper introduces a novel local, non-asymptotic approach to quantify propagation of chaos in mean field diffusions by deriving a differential inequality from a modified BBGKY hierarchy. It establishes that the relative entropy between the k-particle law and its limiting product measure decays as O((k/n)²), which is optimal and improves upon prior O(k/n) bounds, under a functional inequality assumption on the limiting measure that covers irregular and infinite-range interactions.

ABSTRACT

This paper develops a non-asymptotic, local approach to quantitative propagation of chaos for a wide class of mean field diffusive dynamics. For a system of $n$ interacting particles, the relative entropy between the marginal law of $k$ particles and its limiting product measure is shown to be $O((k/n)^2)$ at each time, as long as the same is true at time zero. A simple Gaussian example shows that this rate is optimal. The main assumption is that the limiting measure obeys a certain functional inequality, which is shown to encompass many potentially irregular but not too singular finite-range interactions, as well as some infinite-range interactions. This unifies the previously disparate cases of Lipschitz versus bounded measurable interactions, improving the best prior bounds of $O(k/n)$ which were deduced from global estimates involving all $n$ particles. We also cover a class of models for which qualitative propagation of chaos and even well-posedness of the McKean-Vlasov equation were previously unknown. At the center of a new approach is a differential inequality, derived from a form of the BBGKY hierarchy, which bounds the $k$-particle entropy in terms of the $(k+1)$-particle entropy.

Motivation & Objective

• To develop a non-asymptotic, local framework for quantifying propagation of chaos in mean field diffusive systems.
• To establish a sharp O((k/n)²) rate for the relative entropy between k-particle laws and their limiting product measures.
• To unify the analysis of both Lipschitz and bounded measurable interactions under a single functional inequality assumption.
• To cover models for which qualitative propagation of chaos and even well-posedness of the McKean-Vlasov equation were previously unknown.
• To demonstrate optimality of the (k/n)² rate via a solvable Gaussian example.

Proposed method

• Derives a differential inequality linking the k-particle relative entropy to the (k+1)-particle relative entropy using a form of the BBGKY hierarchy.
• Employs a functional inequality assumption on the limiting measure to control the entropy dynamics.
• Uses a local approach that avoids global estimates involving all n particles, focusing only on finite k.
• Applies the chain rule for relative entropy and Gronwall-type arguments to bound the Wasserstein distance in terms of entropy.
• Adapts techniques from stochastic calculus, including Girsanov's theorem and martingale representation, to handle the path-space measures.
• Uses optimal couplings and the martingale representation theorem to relate initial entropy to pathwise distance.

Experimental results

Research questions

• RQ1 Can a non-asymptotic, local method achieve a sharp O((k/n)²) rate for relative entropy decay in mean field diffusions?
• RQ2 Is the O((k/n)²) rate optimal, and can it be achieved even when the initial law is fully chaotic?
• RQ3 Can the framework unify the analysis of both Lipschitz and bounded measurable interaction functions?
• RQ4 Does the method apply to models where qualitative propagation of chaos and well-posedness of the McKean-Vlasov equation were previously unknown?
• RQ5 Can the functional inequality assumption be satisfied by irregular or infinite-range interactions?

Key findings

• The relative entropy between the k-particle law and its limiting product measure decays as O((k/n)²) at all times, provided it holds at time zero.
• The O((k/n)²) rate is optimal, as demonstrated by a solvable Gaussian example where the bound cannot be improved.
• The method achieves a significant improvement over prior global estimates that yielded only O(k/n) bounds.
• The functional inequality assumption on the limiting measure encompasses a broad class of interactions, including finite-range, irregular, and some infinite-range interactions.
• The approach is local and does not require non-trivial global estimates over all n particles, enabling analysis of previously intractable models.
• The result implies a total variation convergence rate of O(k/n), which contradicts the widely held belief that O(√(k/n)) is optimal.