Power method does not converge

[dennlinger]

I have the problem that for a given transition matrix, I cannot reach convergence with the implemented _power_method. Instead, after only a few iterations, I am left with only NaN values in the eigenvector guess.

Given that there are existing functions in numpy etc. for computing eigenvectors, is there any particular reason to use the given power method implementation?

[dennlinger]

I've realized this is a problem of my transition matrix, which has non-normalized transition scores (with potentially negative values) in it. Since this violates the non-periodicity and irreducibility assumption about a Markov chain, it may happen that the iterative power method actually diverges.

The problem can be resolved by adding a custom normalization, for example, with scipy.special.softmax(), in create_markov_matrix().
Since this doesn't regularly happen with the traditional graph degree-based approach, I'll close this issue here :)