Creation of chaos for interacting Brownian particles

Armand Bernou, Mitia Duerinckx, Matthieu Ménard · 2025

Read Creation of chaos for interacting Brownian particles by Armand Bernou, Mitia Duerinckx, Matthieu Ménard (2025) 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-2504-09917.txt.

Introduction

Brownian particles that push and pull on each other are a basic model in statistical physics. Armand Bernou, Mitia Duerinckx, and Matthieu Ménard studied such a system in 2025, with many particles linked by a weak, smooth, long-range force. They looked at a phenomenon called creation of chaos, where particles that start out linked become nearly independent over time.

The team prove a sharp version of that effect: the mean-field picture of a typical particle is accurate to within an error that shrinks as the particle count grows. The only extra error comes from pairs that start out correlated, and that error fades fast. Their method mixes basic calculus estimates with deeper equation tools.

Catalog record

CREATORArmand Bernou, Mitia Duerinckx, Matthieu Ménard
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE92 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-09917

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Creation of chaos for interacting Brownian particles Armand Bernou, Mitia Duerinckx, Matthieu Ménard Abstract . We consider a system of N N Brownian particles, with or without inertia, interacting in the mean-field regime via a weak, smooth, long-range potential, and starting initially from an arbitrary exchangeable N N -particle distribution. In this model framework, we establish a fine version of the so-called creation-of-chaos phenomenon: in weak norms, the mean-field approximation for a typical particle is shown to hold with an accuracy O ⁡ ( N − 1 ) O(N^{-1}) up to an error due solely to initial pair correlations, which is damped exponentially over time. Corresponding higher-order results are also derived in the form of higher-order correlation estimates. The approach is new and easily adaptable: we start from suboptimal correlation estimates obtained from an elementary use of Itô’s calculus on moments of the empirical measure, together with ergodic properties of the mean-field dynamics, and these bounds are then made optimal after combination with PDE estimates on the BBKY hierarchy. MSC-class: Primary 60K35; Secondary 35Q70, 82C22, 82C31, 35Q84. 1. Introduction 1.1. General overview We consider the Langevin dynamics for a system of N Brownian particles with mean-field interactions, moving in a confining potential in \mathbb{R}^{d} , d\geq 1 , as described by the following system of coupled SDEs: for 1\leq i\leq N, \displaystyle\left\{\begin{array}[]{ll}\mathrm{d}X^{i,N}_{t}=V^{i,N}_{t}\mathrm{d}t,&\\[5.69054pt] \mathrm{d}V^{i,N}_{t}=-\tfrac{\kappa}{N}\sum_{j=1}^{N}\nabla W(X^{i,N}_{t}-X_{t}^{j,N})\,\mathrm{d}t-\tfrac{\beta}{2}V^{i,N}_{t}\mathrm{d}t-\nabla A(X^{i,N}_{t})\,\mathrm{d}t+\mathrm{d}B_{t}^{i},&\quad t\geq 0,\end{array}\right. where \{Z^{i,N}:=(X^{i,N},V^{i,N})\}_{1\leq i\leq N} is the set of positions and velocities of the particles in the phase space \mathbb{X}:=\mathbb{R}^{d}\times\mathbb{R}^{d} , where W:\mathbb{R}^{d}\to\mathbb{R} is a long-range interaction potential with the action-reaction condition W(x)=W(-x) , where A is a uniformly convex confining potential, where \{B^{i}\}_{i} are i.i.d. d -dimensional Brownian motions, and where \kappa,\beta>0 are given constants. For simplicity, we choose the confining potential A to be quadratic, (1.3) A(x)\,:=\,\tfrac{1}{2}a|x|^{2},\qquad\text{for some $a>0$}, although perturbations can be considered as well (see Section 1.3 ). Turning to a statistical description of the system, in terms of a probability density F^{N} on the N -particle phase space \mathbb{X}^{N} , the Langevin dynamics ( 1.1 ) is equivalent to the corresponding Liouville equation (1.4) \partial_{t}F^{N}+\sum_{i=1}^{N}v_{i}\cdot\nabla_{x_{i}}F^{N}\,=\,\tfrac{1}{2}\sum_{i=1}^{N}{\operatorname{div}}_{v_{i}}((\nabla_{v_{i}}+\beta v_{i})F^{N})\\[-8.53581pt] +\tfrac{\kappa}{N}\sum_{i,j=1}^{N}\nabla W(x_{i}-x_{j})\cdot\nabla_{v_{i}}F^{N}+\sum_{i=1}^{N}\nabla A(x_{i})\cdot\nabla_{v_{i}}F^{N},\qquad t\geq 0. In this contribution, we investigate the behavior of the system in the regime of a large number N\gg 1 of particles in the general setting of initial data \{Z^{i,N}|_{t=0}\}_{1\leq i\leq N} that are only assumed to be exchangeable (that is, their law F^{N}|_{t=0} is invariant under permutation of the particles), but that may be strongly correlated. Let us recall two standard types of results for the system: — Propagation of chaos: If particles have weak correlations initially, then propagation of chaos is expected to hold, which means that only little correlations can build over time. As a consequence, the distribution of a typical particle is approximated by the solution \mu of the mean-field Vlasov–Fokker–Planck equation (1.5) \left\{\begin{array}[]{l}\partial_{t}\mu+v\cdot\nabla_{x}\mu=\tfrac{1}{2}{\operatorname{div}}_{v}((\nabla_{v}+\beta v)\mu)+(\nabla A+\kappa\nabla W\ast\mu)\cdot\nabla_{v}\mu,\qquad t\geq 0,\\[2.84526pt] \mu|_{t=0}=\mu_{\circ},\end{array}\right. where \mu_{\circ} stands for the initial single-particle distribution. Such a result has been established in various settings and we refer e.g. to the recent work [ 5 , 4 ] in case of singular interactions. More precisely, for all 1\leq m\leq N , denoting by F^{N,m} the distribution of m typical particles in the system, as given by the m -th marginal of F^{N} , F^{N,m}(z_{1},\ldots,z_{m})=\int_{\mathbb{X}^{N-m}}F^{N}(z_{1},\ldots,z_{N})\,\mathrm{d}z_{m+1}\ldots\mathrm{d}z_{N},\qquad 1\leq m\leq N, we find that F^{N,m} converges to \mu^{\otimes m} as N\uparrow\infty for any fixed m . Beyond this qualitative convergence, an optimal O(N^{-1}) error estimate for the convergence of marginals was first obtained in [ 10 ] for smooth interactions, and later in [ 15 ] for bounded non-smooth interactions in a better metric: if correlations vanish initially, then there holds (1.6) \displaystyle\|F^{N,m}_{t}-\mu^{\otimes m}_{t}\|_{TV}\,\leq\,Ce^{Ct}mN^{-1},\qquad\text{for all $1\leq m\leq N$ and $t\geq 0$}. If in addition the intensity \kappa of interaction forces is small enough, which ensures the uniqueness of the mean-field equilibrium, then this error estimate ( 1.6 ) holds uniformly in time without the exponential factor [ 16 , 1 ] (see also previous work using coupling methods but only producing a suboptimal convergence rate [ 3 , 13 ] ). — Relaxation to equilibrium: If interaction forces are weak enough, even if particles are strongly correlated initially, their distribution is known to relax to Gibbs equilibrium on long times. For Lipschitz forces, the following uniform-in- N exponential estimate was established in [ 3 ] , \mathcal{W}_{2}(F_{t}^{N,m},M^{N,m})\,\leq\,C_{m}e^{-\frac{1}{C}t},\qquad\text{for all $1\leq m\leq N$ and $t\geq 0$}, where M^{N,m} stands for the m -th marginal of the N -particle Gibbs equilibrium M^{N} . As the latter exhibits weak correlations, this entails in particular that initial correlations between the particles must be exponentially damped over time. Both uniform-in-time propagation of chaos and convergence to Gibbs equilibrium rely essentially on the same phenomenon: due to the diffusion, correlations cannot accumulate too much over time and must somehow be damped. This leads to the natural related question of creation of chaos as first raised by Del Moral and Tugaut [ 7 ] (see also “generation of chaos” in [ 17 , 19 ] ): for general exchangeable initial data, is it possible to show that the mean-field approximation holds to order O(N^{-1}) up to an error coming from initial correlations, which is damped exponentially over time? Such a result can in fact be deduced immediately from the coupling argument of [ 3 ] , but only with a suboptimal convergence rate O(N^{-1/2}) : more precisely, for weak Lipschitz forces, it follows from [ 3 , Theorem 2 and Remark 12] that (1.7) \displaystyle\mathcal{W}_{2}(F^{N,m}_{t},\mu_{t}^{\otimes m})\leq C_{m}\,N^{-\frac{1}{2}}+C_{m}\,e^{-\frac{1}{C}t}\,\mathcal{W}_{2}(F^{N,m}_{0},\mu_{\circ}^{\otimes m}),\quad\hbox{ for all }1\leq m\leq N\hbox{ and }t\geq 0. Our first objective is to achieve the optimal rate O(N^{-1}) in this creation-of-chaos result, which to our knowledge is not possible via the same coupling methods. More precisely, we look for bounds of the following form, in some suitable metric, (1.8) \|F_{t}^{N,m}-\mu_{t}^{\otimes m}\|\,\leq\,C_{m}N^{-1}+C_{m}e^{-\frac{1}{C}t}\|F_{\circ}^{N,m}-\mu_{\circ}^{\otimes m}\|,\quad\text{for all $1\leq m\leq N$ and $t\geq 0$}. In the sequel, we establish estimates of this kind for weak smooth interactions, in negative Sobolev norms instead of Wasserstein distance; see Theorem 1.1 below. While virtually any proof of uniform-in-time propagation of chaos comes with some associated creation-of-chaos result showing the damping of initial correlations, we emphasize that clean estimates of the form ( 1.8 ) are more difficult to establish. In particular, there has been a lot of recent progress based on hierarchical approaches to uniform-in-time propagation of chaos [ 15 , 5 , 16 , 1 , 4 ] , starting from the BBGKY hierarchy of equations for marginals, but such methods require to deal with information on the whole hierarchy \{F^{N,m}-\mu^{\otimes m}\}_{1\leq m\leq N} at once and can thus a priori not be used to obtain estimates of the form ( 1.8 ) for any fixed m . For overdamped systems (see Section 1.3.1 below), modulated energy methods can also be used to establish creation of chaos, see [ 19 ] , but they require much stronger initial information in the form of the relative entropy of the N -particle distribution with respect to a tensorized state, which again cannot lead to clean estimates of the form ( 1.8 ) for marginals. Next to this question on the creation and propagation of chaos, we further refine our analysis to corresponding correlation estimates. Many-particle correlation

Read the complete text