Discrepancy Analysis of a New Randomized Diffusion Algorithm

TL;DR

This paper introduces a new randomized diffusion algorithm for load balancing in graphs, achieving significantly lower discrepancy than deterministic methods, with broad applicability to various graph structures.

Contribution

The paper proposes a novel randomized diffusion algorithm that reduces discrepancy to O(√(d log N)) in regular graphs and generalizes to arbitrary graphs, outperforming deterministic approaches.

Findings

Achieves O(√(d log N)) discrepancy in d-regular graphs with high probability.

Deterministic diffusion algorithms have an Ω(d) lower bound on discrepancy.

Generalizes to any symmetric round matrix, achieving O(√(d_max log N)) discrepancy.

Abstract

For an arbitrary initial configuration of discrete loads over vertices of a distributed graph, we consider the problem of minimizing the {\em discrepancy} between the maximum and minimum loads among all vertices. For this problem, this paper is concerned with the ability of natural diffusion-based iterative algorithms: at each discrete and synchronous time step on an algorithm, each vertex is allowed to distribute its loads to each neighbor (including itself) without occurring negative loads or using the information of previous time steps. In this setting, this paper presents a new {\em randomized} diffusion algorithm like multiple random walks. Our algorithm archives $O(\sqrt{d \log N})$ discrepancy for any $d$-regular graph with $N$ vertices with high probability, while {\em deterministic} diffusion algorithms have $\Omega(d)$ lower bound. Furthermore, we succeed in generalizing our algorithm to any symmetric round matrix. This yields that $O(\sqrt{ d_{\max} \log N})$ discrepancy for arbitrary graphs without using the information of maximum degree $d_{\max}$.