Enclosed is slow and dumb code for performing Swendsen-Wang sampling on
generalised Ising models.

*** 
*** On no account use this code for performance testing of Swendsen-Wang, I just
*** wasn't clever enough to run off a decent Matlab version before switching to C.
*** Also none of this has been tested that thoroughly. It comes with *no* guarantees.
*** 

See an overview in Radford Neal's MCMC tech. report. Or the Sokal and Edwards
generalisation in physical review. The check-coalescense version is even slower,
but allows you to see if different initial conditions ever couple together,
which might be useful for convergence analysis. Search for work on "bounding
chains" for Swendsen Wang to understand this.

Note that the clusters from Swendsen-Wang can give very pretty pictures (see the
pdf files). The one at zero temperature indicates how it can be very bad with
many negative couplings. The whole system has locked together and can only flip,
but not change structure. Even with positive-only couplings Swendsen-Wang can be
slow to mix. However, on Ising models it is *amazing* and I imagine it is pretty
useful in many machine learning problems --- especially in vision.

Iain Murray 2004