Gradient Mean-Field Dynamics with Measure-Valued States: Well-Posedness, Chaos, and Long-Time Stability
Anderson Melchor Hernandez · 2026
Read Gradient Mean-Field Dynamics with Measure-Valued States: Well-Posedness, Chaos, and Long-Time Stability by Anderson Melchor Hernandez (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-24385.txt.
Introduction
Hernandez lays the foundations for particle systems whose members carry entire probability distributions as part of their state. In this mean field system each agent has a place in space, evolving under Brownian diffusion. It also has a measure valued part: a distribution over a compact space, driven by projected cylindrical noise.
Strong solutions exist and are unique, first. They cover both the finite particle system and the limiting nonlinear McKean-Vlasov equation, under locally Lipschitz drift with linear growth. Propagation of chaos follows. As the particle count grows, the sample distribution converges on average, in the Wasserstein distance, to the McKean-Vlasov solution. Particles thus become independent in the limit. Hernandez then investigates fast convergence of the nonlinear flow toward a unique invariant measure, with stability conditions for the long time behaviour. The 2026 analysis gives measure valued mean field dynamics a complete treatment.
Catalog record
| CREATOR | Anderson Melchor Hernandez |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 99 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.24385
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Gradient Mean-Field Dynamics with Measure-Valued States: Well-Posedness, Chaos, and Long-Time Stability Anderson Melchor Hernandez Abstract . We study a stochastic mean-field interacting particle system whose state space is 𝒴 = ℝ d × 𝒫 ( U ) {\mathscr{Y}}={\mathbb{R}}^{d}\times{\mathscr{P}}(U) , where the first component represents a spatial variable and the second one is a probability measure over a compact metric space U U . The dynamics are driven by locally Lipschitz drift operators: the spatial component evolves according to a Brownian diffusion, while the measure-valued component is perturbed by a projected cylindrical noise acting in the Arens–Eells space. We first establish existence and uniqueness of strong solutions for both the N N -particle system and the associated nonlinear McKean–Vlasov equation under locally Lipschitz and linear growth assumptions on the drift coefficients. We then prove propagation of chaos: as N → ∞ N\to\infty , the empirical measure converges in expectation in Wasserstein–1 distance towards the unique McKean–Vlasov solution. Further, we investigate exponential convergence of the nonlinear McKean–Vlasov dynamics towards a unique invariant measure. 1. Introduction Mean-field interacting particle systems provide a powerful framework for describing the collective behavior of large populations of agents subject to deterministic interactions and stochastic fluctuations. Such models arise naturally in kinetic theory, statistical mechanics, interacting diffusions, and in the mathematical analysis of socio-economic systems, crowd dynamics, and multi-agent systems. In particular, kinetic and mean-field approaches have provided a systematic framework for the derivation of macroscopic models from microscopic interaction rules in crowd and swarming dynamics [ 11 ] . More recently, mean-field techniques have also played a central role in the analysis of learning dynamics and neural-type systems [ 13 , 28 , 31 ] . From a modeling perspective, our approach is conceptually related to the active particle framework proposed by [ 12 ] . In that setting, the dynamics of a large population is derived from interaction rules prescribed at the level of individual agents. A statistical description of the system is then introduced, leading to a kinetic equation governing the evolution of the population density. Similarly, in our framework, we start from a system of interacting stochastic differential equations describing the microscopic behavior of each agent. The collective dynamics emerges by considering the empirical measure of the system and passing to the limit as the number of agents tends to infinity, which yields a nonlinear McKean–Vlasov equation. However, unlike the classical kinetic PDE formulation, our model involves a measure-valued internal state for each particle and evolves in an infinite-dimensional functional space. In this work, we investigate a stochastic interacting particle system whose state space is {\mathscr{Y}}={\mathbb{R}}^{d}\times{\mathscr{P}}(U) , where the first component represents a spatial state variable evolving on {\mathbb{R}}^{d} , and the second component is a probability measure over a compact set U . The dynamics couple a spatial diffusion with a measure-valued component evolving in a tangent-like space of {\mathscr{P}}(U) and perturbed by a generalized noise in the sense of Mamporia [ 23 , 24 ] . The interaction between particles is of mean-field type and depends on the empirical distribution of the full system [ 33 , 17 , 16 , 34 ] . More precisely, we consider a system of N particles (X^{i}(t),\lambda^{i}(t))_{i=1}^{N} solving a coupled stochastic differential system with locally Lipschitz drift fields satisfying linear growth conditions. The associated empirical measure is defined by (1.1) \displaystyle\mu^{N}(t)\coloneqq\frac{1}{N}\sum_{i=1}^{N}\delta_{(X^{i}(t),\lambda^{i}(t))}. The corresponding nonlinear McKean–Vlasov limit equation reads (1.2) \displaystyle\begin{cases}{\rm d}Y(t)=H(Y(t),\Lambda(t),\mu(t))\,{\rm d}t+\sqrt{2\sigma_{x}}\,{\rm d}B(t),\\ {\rm d}\Lambda(t)={\mathcal{R}}(Y(t),\Lambda(t),\mu(t))\,{\rm d}t+\sqrt{2\sigma_{\lambda}}\,\Pi_{\Lambda(t)}\,{\rm d}W^{Q}(t),\end{cases} where \mu(t)={\rm Law}(Y(t),\Lambda(t)) . The first equation describes the spatial evolution on {\mathbb{R}}^{d} , driven by the velocity field H and a d -dimensional Brownian motion. The second equation governs the evolution of the measure component \Lambda(t) and involves the velocity field {\mathcal{R}} together with a generalized cylindrical noise W^{Q}(t) acting in the Arens–Eells space {\mathscr{F}}(U) which is defined as the span of {\mathscr{P}}(U) with respect to the bounded Lipschitz norm [ 5 , 2 , 19 ] . Several applications have been proposed for systems of the form ( 1.2 ). One example is the follower–leader model, aimed at describing the collective behavior of a controlled population Y(t) influenced by a leader population distributed according to \Lambda(t) [ 1 , 6 ] . In [ 18 ] , this framework was further developed to analyze a neuronal system in which the random variable Y(t) represents the evolution of the membrane potential of a neuron receiving inputs from other neurons in the network. We emphasize that the model ( 1.2 ) differs from those considered in [ 18 , 9 , 6 ] due to the presence of the generalized noise W^{Q}(t) . Roughly speaking, this object mimics the properties of a Wiener process when Q is a nuclear operator; however, its definition is formulated via duality, allowing us to consider cylindrical operators [ 29 , 22 ] . Unlike the previously studied models, the introduction of this term may destroy the positivity of \Lambda(t) . To overcome this issue, we project the stochastic perturbation onto the tangent space at \Lambda(t) through a suitable operator, denoted by \Pi_{\Lambda(t)} . This projection mechanism, combined with structural growth conditions on Q and a martingale argument, ensures that \Lambda(t) remains in {\mathscr{P}}(U) almost surely [ 30 ] . We stress that the functional framework underlying the measure component relies on the structure of {\mathscr{F}}(U) , which is the canonical predual of the space of Lipschitz functions vanishing at a reference point. This dual characterization clarifies the compatibility between the bounded-Lipschitz norm on {\mathscr{F}}(U) and the Wasserstein- 1 distance (see ( 2.5 ) below). In particular, duality with Lipschitz functions naturally leads to estimates in {\mathcal{W}}_{1} , which can in turn be controlled by the {\rm BL} norm [ 26 , 2 , 9 , 18 ] . Contributions. Our analysis contributes to the mathematical theory of mean-field limits in a infinite-dimensional setting [ 6 , 7 , 8 , 18 ] . On the one hand, the measure-valued component of each particle evolves in the Arens–Eells space, and is driven by a projected cylindrical noise which is not assumed to be trace-class as usually done for the construction of a stochastic integral [ 22 , 29 ] . This goes beyond classical McKean–Vlasov dynamics with measure-dependent drifts, where the noise acts only on a finite-dimensional state and positivity is preserved by construction [ 9 , 18 ] . In our setting, a key technical difficulty stems from the fact that the noise does not preserve positivity a priori. In contrast to classical McKean–Vlasov models, where the law of each particle is automatically a probability measure and the stochastic perturbation does not act directly on the measure-valued coordinate, here the noise may drive \Lambda(t) out of {\mathscr{P}}(U) [ 35 ] . A central contribution of this work is to show that, despite this, the combination of the projection onto a suitable tangent space, structural conditions on the covariance operator Q , and a martingale argument yields almost-sure preservation of positivity for \Lambda(t) [ 15 , 30 ] . This allows us to treat a broader class of covariance operators than in previous works, while keeping the dynamics measure-valued. On the other hand, we obtain both well-posedness and propagation of chaos under