Skip to main content
QUICK REVIEW

[论文解读] Large sparse networks of interacting diffusions

Daniel Lacker, Kavita Ramanan|arXiv (Cornell University)|Apr 4, 2019
Stochastic processes and statistical mechanics被引用 11
一句话总结

本文研究了大而稀疏的相互作用扩散网络,表明当相互作用图局部收敛时,粒子系统在局部弱收敛于极限图上的扩散过程。关键成果包括通过非马氏过程刻画了在单模Galton-Watson树上的极限动力学,以及证明了即使典型粒子的分布是确定性的,经验测度的极限仍可能保持随机性。

ABSTRACT

We consider interacting particle systems on a large sparse, possibly random, interaction graph $G_n$, where each particle evolves infinitesimally like a d-dimensional diffusion whose drift coefficient depends on the histories of its own state and the states of neighboring particles, and the diffusion coefficient depends only on the history of its own state. We study limits of such particle systems in the case when the average degree of $G_n$ remains almost surely bounded. Specifically, under suitable assumptions on the coefficients and initial conditions, we show that if $G_n$ converges in distribution in the sense of local convergence to a locally finite graph $G$, then the corresponding particle dynamics converge weakly locally to a certain limit diffusion on $G$. We also show that for certain graph sequences, including sparse Erdos-Renyi graph sequences, the limit of the empirical measure is deterministic and coincides with the law of a typical particle; but show that in general, the empirical measure limit may fail to coincide with the law of a typical particle, and it could even remain stochastic. Furthermore, when $G$ is a unimodular Galton-Watson tree, under suitable assumptions we characterize the limiting dynamics of the neighborhood of a typical particle in terms of a certain finite-dimensional non-Markovian stochastic process whose infinitesimal evolution at any time depends not only on the current state of the neighborhood, but also on the conditional law of the current state given the past of the neighborhood process until that time. This resolves the open problem of characterizing the limiting dynamics on sequences of sparse Erdos-Renyi graphs. Important ingredients of the proofs include correlation decay estimates, a second-order Markov random field property for particle trajectories, and a stochastic analytic result on mimicking Ito processes.

研究动机与目标

  • 理解大而稀疏相互作用图上相互作用扩散粒子系统的渐近行为。
  • 确定粒子系统经验测度收敛到确定性极限的条件。
  • 解决关于稀疏Erdos-Renyi图序列上极限动力学表征的开放问题。
  • 建立经验测度极限不同于典型粒子分布的条件。
  • 通过非马氏随机过程表征单模Galton-Watson树上的极限动力学。

提出的方法

  • 将随机图 $G_n$ 局部收敛于一个局部有限图 $G$ 作为关键结构假设。
  • 应用相关性衰减估计以控制大系统极限下网络中的依赖关系。
  • 利用粒子轨迹的二阶马尔可夫随机场性质来建模时间和空间依赖性。
  • 使用关于模仿伊藤过程的随机分析结果来构造极限动力学。
  • 通过依赖于过去条件分布的有限维非马氏过程,表征单模Galton-Watson树上的极限。
  • 在适当的系数和初始条件假设下,建立粒子动力学的弱局部收敛性。

实验结果

研究问题

  • RQ1在何种条件下,大而稀疏的相互作用扩散网络的经验测度收敛到确定性极限?
  • RQ2能否完全表征稀疏Erdos-Renyi图上的极限动力学?若能,如何实现?
  • RQ3在何种情况下,经验测度极限不与典型粒子的分布一致?
  • RQ4在单模Galton-Watson树上,极限动力学的非马氏性质如何在典型粒子邻域中体现?
  • RQ5给定过去邻域历史的当前状态的条件分布,在极限动力学中起什么作用?

主要发现

  • 当 $G_n$ 局部分布收敛于 $G$ 时,粒子系统在局部弱收敛于极限图 $G$ 上的极限扩散过程。
  • 对于某些图序列(包括稀疏Erdos-Renyi图),经验测度极限是确定性的,且与典型粒子的分布一致。
  • 一般情况下,即使典型粒子的分布是确定性的,经验测度极限仍可能保持随机性。
  • 在单模Galton-Watson树上,典型粒子邻域的极限动力学由一个有限维非马氏过程表征。
  • 该非马氏过程的无穷小演化依赖于当前状态以及给定邻域过程过去历史的条件分布。
  • 通过一种新颖的非马氏表示,本研究解决了稀疏Erdos-Renyi图上极限表征的开放问题。

更好的研究,从现在开始

从阅读论文到最终审阅,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。