[Paper Review] Mixing Time of Glauber Dynamics With Parallel Updates and Heterogeneous Fugacities

Mathieu Leconte, Jian Ni+1 more|arXiv (Cornell University)|Apr 2, 2010

Markov Chains and Monte Carlo Methods8 references3 citations

TL;DR

This paper analyzes the mixing time of a generalized Glauber dynamics with parallel updates and heterogeneous fugacities in the hard-core gas model. It establishes conditions under which the mixing time remains polynomial in the number of vertices, leveraging a novel drift analysis to bound the expected change in a potential function, with key results depending on vertex degrees, individual fugacities, and update probabilities.

ABSTRACT

Glauber dynamics is a powerful tool to generate randomized, approximate solutions to combinatorially difficult problems. Applications include Markov Chain Monte Carlo (MCMC) simulation and distributed scheduling for wireless networks. In this paper, we derive bounds on the mixing time of a generalization of Glauber dynamics where multiple vertices are allowed to update their states in parallel and the fugacity of each vertex can be different. The results can be used to obtain various conditions on the system parameters such as fugacities, vertex degrees and update probabilities, under which the mixing time grows polynomially in the number of vertices.

Motivation & Objective

• To study the mixing time of Glauber dynamics when multiple vertices update in parallel and each has a distinct fugacity.
• To determine under what conditions on fugacities, vertex degrees, and update probabilities the mixing time remains polynomial in the number of vertices.
• To extend the classical single-site Glauber dynamics to a parallel update model while preserving the product-form stationary distribution.
• To provide theoretical bounds on the convergence rate of the Markov chain under this generalized dynamics.
• To support the design of distributed, throughput-optimal scheduling algorithms in wireless networks with heterogeneous traffic demands.

Proposed method

• Proposes a parallel Glauber dynamics where an independent set of vertices is selected uniformly at random to update their states in each time slot.
• Maintains the product-form stationary distribution by ensuring that only independent sets are chosen as update sets.
• Introduces a potential function $\Phi$ that measures the Hamming distance between two configurations, used to analyze convergence.
• Applies a drift analysis to bound the expected change in the potential function over one step of the Markov chain.
• Derives upper bounds on the expected change in potential by analyzing contributions from adding, removing, or updating vertices in the configuration.
• Uses indicator functions and neighborhood dependencies to compute expected changes in potential due to state transitions, particularly focusing on vertices at distance 1 and 2 from a central vertex $v$.

Experimental results

Research questions

• RQ1 Under what conditions on the fugacities, vertex degrees, and update probabilities does the parallel Glauber dynamics with heterogeneous fugacities exhibit polynomial mixing time?
• RQ2 How can the product-form stationary distribution be preserved when allowing multiple vertices to update in parallel?
• RQ3 What is the rate of convergence (mixing time) of the generalized Glauber dynamics with parallel updates and heterogeneous fugacities?
• RQ4 How does the drift of the potential function relate to the mixing time in the presence of heterogeneous fugacities?
• RQ5 Can the mixing time be bounded independently of the graph structure under certain constraints on the fugacity values and update probabilities?

Key findings

• The mixing time of the parallel Glauber dynamics with heterogeneous fugacities is bounded above by a polynomial in the number of vertices $n$ under appropriate conditions on the system parameters.
• The expected change in the potential function $\Phi$ is shown to be at most $-2 + \sum_{w \in \mathcal{N}_v} \lambda_w$, which enables a drift-based bound on the mixing time.
• The analysis establishes that if $\lambda_v \leq \frac{1}{\Delta_v - 1}$ for all $v$, where $\Delta_v$ is the degree of vertex $v$, then the mixing time is polynomial in $n$.
• The bound on the expected drift depends on the sum of fugacities over neighbors of a vertex, suggesting that local fugacity control ensures fast convergence.
• The results generalize the classical single-site Glauber dynamics result (e.g., $\lambda < \frac{2}{\Delta - 2}$) to the parallel, heterogeneous case.
• The derived bounds are tight enough to support the design of distributed, throughput-optimal scheduling algorithms in wireless networks with heterogeneous queue lengths and interference constraints.