Characteristic Length and Clustering (1410.3173v1)

Abstract: We explore relations between various variational problems for graphs like Euler characteristic chi(G), characteristic length mu(G), mean clustering nu(G), inductive dimension iota(G), edge density epsilon(G), scale measure sigma(G), Hilbert action eta(G) and spectral complexity xi(G). A new insight in this note is that the local cluster coefficient C(x) in a finite simple graph can be written as a relative characteristic length L(x) of the unit sphere S(x) within the unit ball B(x) of a vertex. This relation L(x) = 2-C(x) will allow to study clustering in more general metric spaces like Riemannian manifolds or fractals. If eta is the average of scalar curvature s(x), a formula mu ~ 1+log(epsilon)/log(eta) of Newman, Watts and Strogatz relates mu with the edge density epsilon and average scalar curvature eta telling that large curvature correlates with small characteristic length. Experiments show that the statistical relation mu ~ log(1/nu) holds for random or deterministic constructed networks, indicating that small clustering is often associated to large characteristic lengths and lambda=mu/log(nu) can converge in some graph limits of networks. Mean clustering nu, edge density epsilon and curvature average eta therefore can relate with characteristic length mu on a statistical level. We also discovered experimentally that inductive dimension iota and cluster-length ratio lambda correlate strongly on Erdos-Renyi probability spaces.