Accounting for the role of long walks on networks via a new matrix function

Abstract

We introduce a new matrix function for studying graphs and real-world networks based on a double-factorial penalization of walks between nodes in a graph. This new matrix function is based on the matrix error function. We find a very good approximation of this function using a matrix hyperbolic tangent function. We derive a communicability function, a subgraph centrality and a double-factorial Estrada index based on this new matrix function. We obtain upper and lower bounds for the double-factorial Estrada index of graphs, showing that they are similar to those of the single-factorial Estrada index. We then compare these indices with the single-factorial one for simple graphs and real-world networks. We conclude that for networks containing chordless cycles-holes-the two penalization schemes produce significantly different results. In particular, we study two series of real-world networks representing urban street networks, and protein residue networks. We observe that the subgraph centrality based on both indices produce significantly different ranking of the nodes. The use of the double factorial penalization of walks opens new possibilities for studying important structural properties of real-world networks where long-walks play a fundamental role, such as the cases of networks containing chordless cycles.

The Deformed Graph Laplacian and Its Applications to Network Centrality Analysis

“…The lower degree (triangle) node has 8 neighbors, none of which are leaves. Taking the top 10 nodes in each ranking, from first to tenth, and recording their degrees, we obtain 44,45,26,24,21,13,25,11,8,10 for Katz and 44,26,24,13,8,45,10,11,10,5 for NBTW. Overall, this example shows how the new NBTW measure (i) is able to move further from basic degree centrality through the use of a larger downweighting parameter, and (ii) gives less credence to low quality, peripheral neighbors.…”

“…We finish by using a network where some external information is available concerning importance. We use a yeast protein interaction network that has been studied in [17,18,21]. There are n = 2224 nodes connected by 6609 edges.…”

Topological Analysis of SARS CoV-2 Main Protease

“…That is, although G = exp (A) accounts for all walks connecting every pair of nodes, it penalizes very much those walks of relatively large length, then making more emphasis in shorter walks around a given node. In order to include longer walks in the analysis we study the following matrix function [36]:…”

High-order couplings in geometric complex networks of neurons

“…Therefore, following the same mathematical framework developed in Ref. [19][20][21], we allow that the injection current I i accounts for the contribution not only from neurons j ∈ N i whose topological distance is d ij = 1, but also from neighbors at higher topological distances d ij > 1., that is:…”