The whole point about the laplacian is that you are talking about edges (rather than the nodes) and gradient averaging. That is, the edge is the difference between two nodes values (perhaps). Eventually we want any residual bias that ties one plaintext bit to its successor to diminish. And we want trigrams, and quintograms to similarly show uniformity.
An expander is that set of graph that assures that the maximal use of edges will occur (diffusing Bayesian factors) And the profusion of edges for almost all cliques will quickly move one out of a local cycle to wider cycles more globally afield. Moreover as the edges cross the boundary between the clique and all the other potential rods, one wants the transitioning action to replace codependency on plaintext bits with dependency on key bits.
In a crypto avalanche one wants an average result that a change of one unit of distance in key or plaintext chooses half the edges that flip the cipher text bit on the next round. That is key and plaintext become isomorphs with even one bit flip inducing acting as an initial condition that causes an increase in uniformity.
So while the rotor wirings may be inducing long sequence of quantum conditional operations that preserve the dependency of each plaintext bit on key bits And preserve the randonmness of the key bit throughout, it’s the function of the expander to be guiding the walk through the qbit space.
Sent from my iPhone