ntchg

Cheeger bounds on a network’s bottleneck

ntchg(from, to, w = NULL, nodes = NULL, tol = 1e-10, maxit = 1e+05)

The conductance of a set of nodes S is the weight of the edges leaving it over the smaller of its volume and that of the rest (volume: summed strength); the graph’s conductance h is the least over all sets, a measure of how cleanly the network splits in two. Finding h is hard, but the second-smallest eigenvalue lambda2 of the normalised Laplacian I - D^-1/2 A D^-1/2 brackets it: lambda2 / 2 <= h <= sqrt(2 lambda2) (Cheeger’s inequality; the lower half is proof("ntchg")). The sweep over the eigenvector finds a set whose conductance (sweep) is at most sqrt(2 lambda2); h is at most sweep.

Parameters

from, to: Labels of the two ends of each edge: vectors of the same length (zero when nodes gives every node) with no missing values.
w: NULL, or one positive weight per edge (a tie’s strength, such as the number of co-offences).
nodes: NULL, or every node’s label once: the node order of the results and the way to include isolated nodes.
tol: The stopping tolerance: one number between 0 and 1.
maxit: The most steps: one whole number, at least 1.

Returns

A list: bound, a named vector lambda2, lower (lambda2 / 2), upper (sqrt(2 lambda2)) and sweep; set, the labels of the sweep set; iter; conv.

Notes

lambda2 is found by power iteration on (I + D^-1/2 A D^-1/2) / 2 kept orthogonal to D^1/2 1, from a fixed pseudo-random start, until the residual is below tol or after maxit steps (conv FALSE). Convergence is slow when lambda2 and the next eigenvalue are close (long paths, large sparse graphs). The graph is undirected and must be connected; weights are tie strengths. Graph rules as in ntdeg().

References

Chung, F. R. K. (1997). Spectral Graph Theory. CBMS Regional Conference Series in Mathematics 92. American Mathematical Society. doi:10.1090/cbms/092

Examples

# two triangles joined by one edge
ntchg(c(1, 1, 2, 4, 4, 5, 3), c(2, 3, 3, 5, 6, 6, 4))

Generated from the R help page: ntchg in the R reference.