Sznitman (1991) — DOI:10.1007/BFb0085169
Key takeaways
Establishes the rigorous mathematical theory of propagation of chaos for systems of N interacting diffusion processes in the McKean-Vlasov (mean- field) regime. Proves that the empirical measure of N interacting particles converges to the solution of the McKean-Vlasov PDE as N→∞, and that any fixed k-tuple of particles becomes asymptotically independent in that limit. These lecture notes remain the canonical reference for mean-field limits of stochastic interacting particle systems.
- Propagation of chaos: if the initial N-particle law is μ_0^⊗N (product), then for any fixed k, the k-marginal of the N-particle law at time t converges to μ_t^⊗k as N→∞
- McKean-Vlasov limit: the one-particle marginal μ^N_t converges weakly to μ_t, the unique solution of the McKean-Vlasov nonlinear Fokker-Planck equation
- Empirical measure convergence: E[W₂(ρ^N_t, μ_t)²] = O(1/N) in d dimensions for d ≥ 3; logarithmic corrections in d = 2
- Uniqueness of the McKean-Vlasov equation under Lipschitz assumptions on b(x, μ)
- The nonlinear semigroup associated with the McKean-Vlasov SDE is well- posed in the space of probability measures
What the evidence does not cover
- Results are asymptotic in N; finite-N corrections are O(1/√N) which may be slow for moderate cluster sizes
- Requires Lipschitz continuity of the interaction kernel in Wasserstein distance — must be verified for the specific gossip weight matrices used in GELS
- All-to-all (mean-field) interaction assumption; gossip on sparse topologies (e.g., ring, fat-tree) is not directly covered and requires extension
- Ellipticity assumption on σ; degenerate noise (e.g., gradient noise only in certain directions) requires separate treatment
- Results are for continuous-time diffusions; SGLD is a discrete-time approximation and discretization error must be handled separately
Standing in the anthology
Read — the reading is NOTE-090. Arrived in the imported batch, which brought in the mean-field account of wide networks, where the object that moves is the distribution of neurons.