Proposal density
Using the covariance matrix to give the proposal directions typically
significantly increases the acceptance rate and gives faster movement
around parameter space.
We generate a random basis in the eigenvectors, then cycle through them
proposing changes to each, then generate a new random basis
The distance proposal in the random direction is given by a two-D Gaussian
radial function mixed with an exponential, which is quite robust to wrong width estimates
See http://cosmologist.info/notes/cosmomc.pdf