Delgadino et al. (2021) — ARXIV-2112.06304
Key takeaways
The paper establishes precise links between (i) the non-degeneracy of the log-Sobolev inequality (LSI) constant of the N-particle Gibbs measure as N→∞, (ii) absence of phase transitions for the McKean-Vlasov mean-field limit, and (iii) uniform-in-time propagation of chaos plus Gaussianity of equilibrium fluctuations. It provides gradient-flow / Wasserstein-calculus proofs of a generalized Talagrand inequality and quantitative propagation of chaos, and proves uniform LSI at high temperature for weakly-interacting, possibly non-convex potentials via the two-scale approach.
- Theorem 3.2 (generalised Talagrand). E^N[ρ^N] - E^N[M_N] ≥ (λ_LS^N/2) d̄_2^2(ρ^N, M_N); analogously E^MF[ρ] - inf E^MF ≥ (λ_LS^∞/2) d_2^2(ρ, K) Holds when: Holds for gradient-flow-regular energies; proof via Otto calculus on (P(Ω), d_2)
- Theorem 3.3. limsup_{N→∞} λ_LS^N ≤ λ_LS^∞ Holds when: Under Assumptions 2.1–2.2; mean-field LSI upper-bounds the large-N particle LSI
- Theorem 3.6 (Gibbs approximation). d̄_2^2(ρ_β^{⊗N}, M_N) ≤ (2/λ_LS^N) Ē(ρ_β^{⊗N}||M_N) ≤ C/N Holds when: Requires λ_LS^N bounded away from 0 uniformly in N
- Theorem 3.7 (uniform-in-time propagation of chaos). sup_{t≥0} d̄_2(ρ^N(t), ρ(t)^{⊗N}) ≤ C N^{-θ} for some θ>0 Holds when: Requires uniform LSI (λ_LS^N bounded below) and E^MF[ρ_in] < ∞; θ not sharp when K_V+K_W(1-1/N)<0
- Theorem 3.11 (equilibrium fluctuations). η^N = √N(μ^N - ρ_β) converges in law to Gaussian field η^∞ solving linear SPDE with covariance Q_G Holds when: Ω = T^d, V,W smooth, liminf λ_LS^N > 0; convergence in C([0,T]; H^{-m}) for m > d/2+3
- Theorem 3.14 (uniform LSI at high temperature). λ_LS^N ≥ c > 0 uniformly in N Holds when: Compact Ω, or bounded W with small ||W||∞ and ||D^2{xy} W||_∞ relative to temperature (β small); proved via two-scale LSI
- Classical propagation of chaos (Theorem 10.1). d̄_2(ρ^N(t), ρ(t)^{⊗N}) ≤ [(1-e^{-Kt/2})/K] · C/√N where K = K_V+K_W(1-1/N) Holds when: Constants blow up exponentially if K<0 (non-convex)
Standing in the anthology
Read — the reading is NOTE-146. 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.