Loop-erased partitioning of a network: monotonicities & analysis of cycle-free graphs

TL;DR

This paper studies the properties and structure of loop-erased random walk-based partitions of graphs, revealing monotonicity in the parameter q and analyzing cluster formation on various graph types, including sparse and modular graphs.

Contribution

It introduces new monotonicity results for loop-erased partition measures and analyzes their asymptotic behavior on different graph models, especially sparse and modular graphs.

Findings

Two key monotonicity theorems for partition measures in q.

Correlation functions increase with q on undirected graphs.

Detailed analysis of cluster structures on line segments, trees, and modular graphs.

Abstract

We consider random partitions of the vertex set of a given finite graph that can be sampled by means of loop-erased random walks stopped at a random exponential time of parameter $q>0$. The related random blocks tend to cluster nodes visited by the random walk on time scale $1/q$. This random partitioning is induced by a measure of rooted spanning forest of the graph, which generalizes the classical uniform spanning tree measure and which can be obtained as a zero-limit of FK-percolation with an external cemetery state. Some general properties of this rooted forest measure and related determinantal observables, along with a number of applications in data analysis have been recently explored. We are here mainly interested in the structure the emergent partitioning, referred to as loop-erased partitioning, as the scale parameter $q$ varies. We first present two general results shedding light on subtle monoticity properties in $q$ of these rooted forest and loop-erased partitioning measures. The first theorem characterizes monotone events in $q$ by deriving a Russo-like formula. Our second general result concerns two-point correlations defined by the probability that two vertices do not belong to the same block of the partitioning. It states that, on undirected graphs, these correlation functions are increasing in $q$. We then explore other types of results aiming at understanding the emerging asymptotic clusters on simple growing graph models, as $q$ scales with the graph size. Some first results in this direction have been investigated in the recent [arXiv:1906.03858] on dense geometries. Here we look at very sparse graphs. We offer a detailed analysis of the resulting partitioning on line segments and we look at trees and other tree-like geometries, without and with implanted modular structures. For the latter, we characterize the asymptotic detection of these implanted modules.