Weak quantitative propagation of chaos via differential calculus on the space of measures

Jean-François Chassagneux, Łukasz Szpruch, Alvin TseThe Annals of Applied Probability 2022 · DOI · arXiv

Abstract

Consider the metric space $(\mathcal{P}_2(\mathbb{R}^d), W_2)$ of square-integrable laws on $\mathbb{R}^d$ with the topology induced by the 2-Wasserstein distance $W_2$. This work investigates the weak error expansion for empirical measures $\mu_N$ in two contexts: (a) i.i.d. samples from a law $\mu$, and (b) marginal laws of a particle system approximating a McKean–Vlasov PDE. The central result demonstrates that under suitable regularity conditions, the weak error follows an expansion of the form $|\Phi(\mu) - \mathbb{E}\Phi(\mu_N)| = \sum_{j=1}^{k-1} C_j N^{-j} + O(N^{-k})$, where $k$ denotes the degree of smoothness. By utilizing functional derivatives on the space of measures and the "master equation" from mean-field game theory, we establish quantitative estimates for propagation of chaos that are independent of the system's dimension. These results enable the use of Romberg extrapolation to significantly enhance the efficiency of interacting particle system simulations.

Introduction

In fields ranging from statistical physics to financial engineering, large systems of interacting agents are often modeled using McKean-Vlasov Stochastic Differential Equations (SDEs). These equations are unique because their evolution depends on the probability distribution of the entire population. In practice, we approximate these infinite-population limits using a system of $N$ discrete particles.

The study of how these $N$-particle systems converge to the limit as $N \to \infty$ is known as "Propagation of Chaos" (PoC). Traditional "strong" PoC results typically provide an error rate of $O(N^{-1/2})$, which often suffers from the "curse of dimensionality"—the error grows significantly as the dimension $d$ of the state space increases. This paper shifts the focus toward weak error: the difference between the expected value of a functional (such as a price or physical property) calculated on the $N$-particle system versus the true limit.

Theoretical Framework: Calculus on Measure Spaces

To derive a more precise error expansion, the authors employ a sophisticated variational approach rooted in differential calculus on the space of probability measures. Two primary tools are utilized:

  1. Lions Derivatives and Linear Functional Derivatives: The authors generalize standard calculus to functions whose inputs are probability distributions. This allows for Taylor-like expansions directly on the Wasserstein space.
  2. The Master Equation: For dynamic interacting particle systems, the error expansion is linked to the "master equation." This is a high-dimensional PDE on the space of measures, a concept central to Mean-Field Game theory, which describes the evolution of the system's value function.
  3. Itô Formula for Measures: To handle the temporal evolution of particle systems, the authors utilize a version of the Itô formula that describes how a functional changes along a flow of probability measures.

The Expansion Theorem

The core contribution of the paper is the proof that the weak error is not merely small, but structured as a power series in $1/N$.

Theorem (Weak Error Expansion): Under suitable regularity (smoothness) conditions, the error satisfies: $$|\mathbb{E}[\Phi(\mu_N)] - \Phi(\mu)| = \sum_{j=1}^{k-1} \frac{C_j}{N^j} + O\left(\frac{1}{N^k}\right)$$ where the constants $C_j$ are independent of $N$.

This result is significant because it shows that the error behaves predictably. If the system is $k$-times differentiable in the measure sense, one can explicitly account for $k-1$ terms of the expansion. Crucially, these estimates are dimension-independent, making the framework highly scalable for complex, high-dimensional models.

Numerical Implications: Romberg Extrapolation

The $1/N$ structure of the error allows for the application of Romberg extrapolation. By running multiple simulations with different particle counts (e.g., $N$ and $2N$) and taking a specific weighted average of the results, researchers can cancel out the lower-order error terms ($1/N, 1/N^2$, etc.).

This leads to a massive computational gain. To reach a target mean-square error of $\epsilon^2$, a standard particle system might require $O(\epsilon^{-4})$ interactions. Using the Romberg-extrapolated ensemble approach, the same level of precision can be achieved with significantly fewer interactions, provided the underlying system is sufficiently smooth.

Conclusion and Future Directions

This work provides a rigorous mathematical foundation for accelerating the simulation of McKean-Vlasov systems. By proving that weak propagation of chaos follows a power series in $1/N$, the authors bridge the gap between abstract measure theory and practical numerical analysis.

However, the current framework requires high regularity—the coefficients and functionals must be many times differentiable. Future research may look toward Malliavin calculus to relax these smoothness requirements, potentially extending these high-order convergence results to systems with non-smooth payoffs or discontinuous interactions, such as those found in collision physics or complex financial derivatives.

Summary updated

Abstract

Consider the metric space (P2(Rd),W2) of square integrable laws on Rd with the topology induced by the 2-Wasserstein distance W2. Let Φ:P2(Rd)→R and μ∈P2(Rd). In this work, we consider (a) μN being the empirical measure of N-samples from μ, and the other case in which (b) μN is the empirical measure of marginal laws of the particle system of a McKean–Vlasov PDE (μt)t. The main result of this paper is to show that under suitable regularity conditions, we have |Φ(μ)−EΦ(μN)|=∑j=1k−1Cj Nj+O(1 Nk), for some positive constants C1,…,Ck−1 that do not depend on N, where k corresponds to the degree of smoothness. The case where the samples are i.i.d. is studied using functional derivatives on the space of measures. The case of particle systems relies on an Itô-type formula for the flow of probability measures and is intimately connected to PDEs on the space of measures, called the master equation in the literature of mean-field games. We state general regularity conditions required for each case and analyze the regularity in the case of functionals of the laws of McKean–Vlasov PDEs. Ultimately, this work reveals quantitative estimates of propagation of chaos for interacting particle systems. Furthermore, we are able to provide weak propagation of chaos estimates for ensembles of interacting particles and show that these may have some remarkable properties.
BibTeX
@article{chassagneux2022weak,
  author = {Jean-François Chassagneux and Łukasz Szpruch and Alvin Tse},
  title = {{Weak quantitative propagation of chaos via differential calculus on the space of measures}},
  journal = {The Annals of Applied Probability},
  year = {2022},
  doi = {10.1214/21-aap1725},
}
Figure