Assortativity and the Distribution of Spectral Gaps in Directed Networks

Abstract

This paper studies how the distribution of spectral gaps of a weighted directed network varies with its assortativity. We prove that the sum of the distances of each of the eigenvalues from the Perron root along the real axis, which we call the `spectral depth', decreases as the network is rewired to be less assortative. The result is conditional on the network's Perron vectors ordering the rewired vertices in the same way as their strengths do. We also develop a method to determine whether a particular spectral gap contracts or expands under a rewiring that changes the network's assortativity. It turns out that the direction of any one gap depends on far more than the network's assortativity. In computational experiments on heavy-tailed networks, disassortative rewiring lowers the spectral depth by about two fifths, contracts most strongly the gaps in the middle of the spectrum, and collapses a growing part of the spectrum to zero. The experiments also show that a sequence of disassortative rewirings tends to narrow the first spectral gap, while a sequence of assortative rewirings widens it.