Complete hierarchies of SIR models on arbitrary networks with exact and approximate moment closure

TL;DR

This paper develops a systematic framework for deriving exact and approximate moment closure models for SIR epidemic dynamics on arbitrary networks, extending previous work with new theorems, concepts, and hierarchies.

Contribution

It introduces a general theorem for exact closures, defines transmission blocks, and constructs hierarchies of approximate models for epidemic processes on networks.

Findings

Exact closure conditions depend on the largest transmission block size.

Hierarchies of approximate models include well-known models as special cases.

New explicit definitions for higher-order epidemic models on networks.

Abstract

We first generalise ideas discussed by Kiss et al. (2015) to prove a theorem for generating exact closures (here expressing joint probabilities in terms of their constituent marginal probabilities) for susceptible-infectious-removed (SIR) dynamics on arbitrary graphs (networks). For Poisson transmission and removal processes, this enables us to obtain a systematic reduction in the number of differential equations needed for an exact `moment closure' representation of the underlying stochastic model. We define `transmission blocks' as a possible extension of the block concept in graph theory and show that the order at which the exact moment closure representation is curtailed is the size of the largest transmission block. More generally, approximate closures of the hierarchy of moment equations for these dynamics are typically defined for the first and second order yielding mean-field and pairwise models respectively. It is frequently implied that, in principle, closed models can be written down at arbitrary order if only we had the time and patience to do this. However, for epidemic dynamics on networks, these higher-order models have not been defined explicitly. Here we unambiguously define hierarchies of approximate closed models that can utilise subsystem states of any order, and show how well-known models are special cases of these hierarchies.