A mathematical preprint reports that, in a class of mean-field systems built from many coupled maps, collective fluctuations approach a Gaussian process as the number of maps N grows. Gaussian here means the normal-distribution pattern used to describe the limiting fluctuations. The statement is mathematical and holds under the paper’s specified regularity, positivity, independent-initial-data and small-coupling conditions over a finite time horizon.
The model consists of N coupled maps—rules that update positions in discrete time—on the one-dimensional torus T = R/Z. Here, mean-field refers to an interaction described through the population-level empirical measure. There is no sampled population or dataset behind the result: N is a generic particle-system size, and the analysis compares the particles’ empirical measure with a limiting distribution. Its central question is whether empirical-measure closeness, propagation of chaos and fluctuations can be controlled as the model evolves.
One result concerns the 1-Wasserstein distance, a way to quantify how far apart two distributions are. The paper proves an exponential-in-time upper bound: the distance at a later time is controlled by an exponential multiple of its initial value. This is a stability statement for the model, and the bound’s constant is existential, depending on the stated map and coupling norms.
The proof uses Kantorovich-Rubinstein duality and direct calculations, the tools the paper uses to obtain that Wasserstein estimate.
When particles begin to look alike
A second line of analysis tracks rescaled relative entropy, a measure of discrepancy between the full particle law and the tensorized law built from the limiting distribution. Under A1.2 and A2, and when the coupling strength is below the threshold Δ0, the quantity H_N(t) is bounded by e^(Ct)[H_N(0) + 1/N] for t = 0 through T. The form of the result combines an exponential-in-time factor with an inverse-particle-count remainder.
The entropy proof combines data processing, Taylor expansion, Gibbs inequality and large-deviation estimates.
That estimate is tied to a statement known as propagation of chaos. In this setting, the phrase means that any fixed number of particle coordinates has a joint distribution that converges weakly to a product of copies of the limiting distribution. Under A1.2–A3 and the same condition Δ < Δ0, the particle law is ρ̄_t-chaotic for t = 0 through T.
The assumptions are substantial. A3 requires independent and identically distributed initial positions on T with common distribution ρ̄0, while A2 requires a strictly positive C2 limiting solution with bounds κ and K over the stated time range. These conditions are part of the theorem’s setting, not evidence from an observed system.
The fluctuation limit
The Gaussian result follows a different path through the proof. Under A1.3, A2, A3 and Δ < Δ0, the fluctuation process η^N converges in distribution to a Gaussian process η as N tends to infinity. The convergence is stated in the negative-Sobolev path space (H^{-α}(T))^{T+1} for every α > 1/2, so it tracks the fluctuation object across the paper’s full discrete-time range.
The convergence proof uses relative-entropy Sobolev bounds to establish tightness, then identifies tight limits as solutions of the fluctuation equation.
The paper also identifies the Gaussian field’s covariance. For each test function, treated here as a chosen mathematical probe, its limiting variance is determined by the time-evolution operator Q0,t and the initial limiting distribution ρ̄0 through the stated covariance expression. This is a model-level characterization, not an empirical estimate of uncertainty.
For finite systems, the paper gives a quantitative expectation bound. For fixed T and m, the difference between expectations involving m test-function projections of the fluctuations and a specified smooth function Ψ is bounded at a rate proportional to 1/√N for t = 0 through T. The proof invokes the classical multivariate Central Limit Theorem and a multivariate Berry-Esseen/Stein estimate.
That rate is deliberately narrower than a universal error bar: it applies to the stated smooth test functions and Ψ, with constants that depend on their norms. It does not give a rate for every possible metric or observable.
What the result does not cover
The overall scope is a theoretical class of N-particle systems on the one-dimensional torus, and the theorem statements use different regularity requirements. The entropy, chaos and Gaussian conclusions are finite-horizon results requiring Δ < Δ0; the supplied analysis does not establish them for coupling strengths at or above that threshold.
The supplied document is an analysis of the stated mathematical model, not an empirical study of a population. The author says, to the best of the author’s knowledge, that the results are established for globally coupled maps for the first time. The acknowledgements report support from the National Key R&D Program of China and the NSFC.
Paper data and sources
Original title: Propagation of Chaos and Gaussian Fluctuations for Mean-field Coupled Maps
Authors: Ruicheng Cheng
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text