[Paper Review] Mean-field limit of non-exchangeable systems

Pierre‐Emmanuel Jabin, David Poyato+1 more|arXiv (Cornell University)|Dec 31, 2021

Advanced Thermodynamics and Statistical Mechanics8 citations

TL;DR

This paper establishes the mean-field limit for non-exchangeable multi-agent systems on general sparse graphs, introducing extended graphons to model asymmetric, non-identical agent interactions. By combining PDE analysis, stochastic processes, and graph theory, it removes classical assumptions of dense graphs and symmetric interactions, proving convergence to a Vlasov-type PDE under minimal connectivity scaling conditions.

ABSTRACT

This paper deals with the derivation of the mean-field limit for multi-agent systems on a large class of sparse graphs. More specifically, the case of non-exchangeable multi-agent systems consisting of non-identical agents is addressed. The analysis does not only involve PDEs and stochastic analysis but also graph theory through a new concept of limits of sparse graphs (extended graphons) that reflect the structure of the connectivities in the network and has critical effects on the collective dynamics. In this article some of the main restrictive hypothesis in the previous literature on the connectivities between the agents (dense graphs) and the cooperation between them (symmetric interactions) are removed.

Motivation & Objective

• To derive the mean-field limit for multi-agent systems with non-identical agents and non-symmetric, sparse connectivity structures.
• To generalize classical mean-field theory beyond exchangeable systems by removing assumptions of dense graphs and symmetric interactions.
• To develop a new mathematical framework—extended graphons—that captures the structural impact of sparse, asymmetric connectivities on collective dynamics.
• To prove the convergence of finite-agent systems to a limiting Vlasov-type PDE under minimal scaling assumptions on interaction weights.
• To extend the applicability of mean-field theory to complex systems in biology, social sciences, and networks with heterogeneous, non-uniform connectivity.

Proposed method

• Introduces a generalized McKean SDE formulation to model the dynamics of N non-identical agents on a weighted, non-symmetric graph.
• Employs a hierarchy of observables indexed by trees to track the evolution of empirical measures and their moments.
• Develops a graphon-like representation of the system's connectivity structure, extending classical graphon theory to sparse, non-symmetric graphs.
• Applies artificial diffusion to stabilize the hierarchy and prove compactness, enabling convergence to a limiting measure.
• Uses a novel compactness lemma to handle unbounded interaction kernels and weight functions in the limit process.
• Applies the Banach fixed-point theorem in a suitable function space to establish local existence and uniqueness of the limiting solution, which is then extended globally in time.

Experimental results

Research questions

• RQ1 How can the mean-field limit be rigorously derived for non-exchangeable multi-agent systems with non-identical agents?
• RQ2 What mathematical framework is needed to represent and analyze the dynamics of sparse, non-symmetric networks in the mean-field regime?
• RQ3 To what extent can classical mean-field assumptions—such as symmetric interactions and dense graphs—be relaxed without losing convergence?
• RQ4 How do the structural properties of sparse graphs (e.g., degree distribution, asymmetry) influence the emergent collective behavior in large systems?
• RQ5 Can the limiting behavior of such systems be described by a Vlasov-type PDE even when the interaction graph is sparse and non-symmetric?

Key findings

• The paper establishes the mean-field limit for non-exchangeable systems on fully sparse graphs, removing the need for dense connectivity assumptions.
• The authors introduce extended graphons as a new tool to represent the limiting connectivity structure of sparse, non-symmetric networks.
• Convergence to a Vlasov-type PDE is proven under minimal scaling conditions: total interaction per agent is O(1), and no single agent dominates (max |w_ij| = o(1)).
• The limiting dynamics are governed by a nonlinear PDE that incorporates the graphon structure, reflecting how network topology shapes collective behavior.
• The proof relies on a hierarchy of observables indexed by trees, with stability established via artificial diffusion and a new compactness lemma.
• The limiting solution exists globally in time and is unique, with convergence established in the weak topology of probability measures.