On the reconstruction limits of complex networks

TL;DR

This paper establishes theoretical limits on network reconstruction from data, introduces a new index to assess reconstruction quality, and demonstrates its effectiveness on empirical networks.

Contribution

It provides a first-principles, information-theoretic framework for understanding the limits of network reconstruction and introduces a practical index for evaluating reconstructed networks.

Findings

Reconstruction index predicts error without knowing the true network.

Theoretical limits relate reconstructability to data-generating processes.

Method validated on empirical time series and networks.

Abstract

Network reconstruction consists in retrieving the hidden interaction structure of a system from observations. Many reconstruction algorithms have been proposed, although less research has been devoted to describe their theoretical limitations. In this work, we take a first-principles approach and build on our earlier definition of reconstructability-the fraction of structural information recoverable from data. We relate this quantity to the true data-generating (TDG) process and delineate an information-theoretic reconstruction limit, i.e., the upper bound of the mutual information between the true underlying graph and any graph reconstructed from observations. These concepts lead us to a principled numerical method to assess the validity of empirically reconstructed networks, based on model selection and a quantity we introduce: the reconstruction index. This index approximates the reconstructability from data, quantifies the variability of the reconstructed network ensemble, and is shown to predict reconstruction error without requiring knowledge of the true underlying network. We characterize this method and test it on empirical time series and networks.