On the epidemic threshold of a network

TL;DR

This paper investigates the relationship between the largest eigenvalue of a network's adjacency matrix and epidemic spread, proposing a new centrality measure and evaluating intervention strategies based on spectral properties.

Contribution

It introduces the spread centrality measure based on eigenvalue changes and demonstrates its correlation with eigenvector centrality, aiding epidemic control strategies.

Findings

Largest eigenvalue change effectively measures intervention impact.

Spread centrality correlates strongly with eigenvector centrality.

Eigenvector centrality is a practical proxy for spread centrality in large networks.

Abstract

The graph invariant examined in this paper is the largest eigenvalue of the adjacency matrix of a graph. Previous work demonstrates the tight relationship between this invariant, the birth and death rate of a contagion spreading on the graph, and the trajectory of the contagion over time. We begin by conducting a simulation confirming this and explore bounds on the birth and death rate in terms of well-known graph invariants. As a result, the change in the largest eigenvalue resulting from removal of a vertex in the network is the best measure of effectiveness of interventions that slow the spread of a contagion. We define the spread centrality of a vertex $v$ in a graph $G$ as the difference between the largest eigenvalues of $G$ and $G-v$. While the spread centrality is a distinct centrality measure and serves as another graph invariant for distinguishing graphs, we found experimental evidence that vertices ranked by the spread centrality and those ranked by eigenvector centrality are strongly correlated. Since eigenvector centrality is easier to compute than the spread centrality, this justifies using eigenvector centrality as a measure of spread, especially in large networks with unknown portions. We also examine two strategies for selecting members of a population to vaccinate.