[Paper Review] Asymptotic analysis of a particle system with mean-field interaction

Anatoly Manita, Vadim Shcherbakov|arXiv (Cornell University)|Aug 26, 2004

Simulation Techniques and Applications9 references12 citations

TL;DR

This paper studies a system of N interacting particles on Z with mean-field interaction, modeling rollback synchronization in distributed simulations. Under appropriate hydrodynamic scaling—time scaled as tN and interaction intensity as μ/N—the empirical measure converges weakly to a deterministic solution of a PDE; in the symmetric case (α=β), this PDE is the KPP equation, and the limit exhibits traveling wave behavior with a speed determined by drift and interaction parameters.

ABSTRACT

We study a system of $N$ interacting particles on $\bf{Z}$. The stochastic dynamics consists of two components: a free motion of each particle (independent random walks) and a pair-wise interaction between particles. The interaction belongs to the class of mean-field interactions and models a rollback synchronization in asynchronous networks of processors for a distributed simulation. First of all we study an empirical measure generated by the particle configuration on $\bf{R}$. We prove that if space, time and a parameter of the interaction are appropriately scaled (hydrodynamical scale), then the empirical measure converges weakly to a deterministic limit as $N$ goes to infinity. The limit process is defined as a weak solution of some partial differential equation. We also study the long time evolution of the particle system with fixed number of particles. The Markov chain formed by individual positions of the particles is not ergodic. Nevertheless it is possible to introduce relative coordinates and to prove that the new Markov chain is ergodic while the system as a whole moves with an asymptotically constant mean speed which differs from the mean drift of the free particle motion.

Motivation & Objective

• To analyze the hydrodynamic limit of a mean-field interacting particle system on Z, motivated by rollback synchronization in distributed simulations.
• To determine the appropriate scaling of time and interaction intensity (μN) such that the empirical measure of particle positions converges to a non-trivial deterministic limit as N→∞.
• To study the long-time behavior of the system with fixed N, focusing on relative dynamics and center-of-mass motion despite non-ergodicity of the original process.
• To establish that the relative particle coordinates form an ergodic Markov chain, implying relative stability, while the system as a whole moves with a non-trivial mean speed.

Proposed method

• Define the empirical tail function ξx,N(t) = (1/N)∑ᵢ 1{xᵢ(t)≥x} to track the proportion of particles at or above position x.
• Apply hydrodynamic scaling: tN = tN and μN = μ/N for α≠β, and tN = tN², μN = μ/N² for α=β, to balance particle motion and interaction.
• Prove weak convergence of the empirical measure to a deterministic limit process, characterized as a weak solution of a first-order PDE (for α≠β) or a second-order KPP-type PDE (for α=β).
• Introduce relative coordinates yi(t) = xi(t) − minⱼxⱼ(t) to decouple global motion and analyze the ergodicity of the relative system.
• Use Mitoma’s tightness theorem and Aldous’ criterion to establish tightness of the empirical measure in the Skorokhod space of tempered distributions.
• Leverage the duality between the particle system and the PDE limit to analyze long-time behavior, including the emergence of traveling wave solutions in the KPP case.

Experimental results

Research questions

• RQ1 What scaling of time and interaction intensity μN ensures a non-trivial hydrodynamic limit for the empirical measure of the particle system as N→∞?
• RQ2 How does the limiting PDE differ between the asymmetric (α≠β) and symmetric (α=β) cases, and what is the physical interpretation of the resulting dynamics?
• RQ3 Why is the original particle system non-ergodic, and how does transforming to relative coordinates restore ergodicity and relative stability?
• RQ4 What determines the asymptotic mean speed of the center of mass, and how does it differ from the free particle drift?
• RQ5 In the symmetric case, does the limiting PDE exhibit traveling wave solutions, and if so, what is the speed in terms of the model parameters?

Key findings

• Under scaling tN = tN and μN = μ/N, the empirical measure converges weakly to a deterministic solution of a first-order PDE, representing the hydrodynamic limit for α≠β.
• Under scaling tN = tN² and μN = μ/N², the empirical measure converges weakly to a solution of the KPP equation, a second-order PDE, in the symmetric case α=β.
• The KPP equation solution exhibits a traveling wave with speed v = −(γκ + μ/κ), where κ is a parameter related to the wave profile, showing that interaction alters propagation speed.
• The relative particle system (in terms of yi(t) = xi(t) − minⱼxⱼ(t)) is ergodic and converges exponentially fast to its stationary distribution, despite the original system being non-ergodic.
• The center of mass of the particle system moves with an asymptotically constant speed that differs from the free particle drift due to the mean-field interaction.
• The convergence of the empirical measure to the PDE solution is established via tightness in the Skorokhod space of tempered distributions, using Mitoma’s theorem and Aldous’ criterion.