It is often useful to have preconditioning in MCLMC, so that the adaptation algorithm identifies the (approximate, diagonal) covariance of the distribution, and the integrator implicitly works in the mapped space with unit variance in all directions.
This PR adds that. The main change is a parameter std_mat to the esh momentum update.
It is often useful to have preconditioning in MCLMC, so that the adaptation algorithm identifies the (approximate, diagonal) covariance of the distribution, and the integrator implicitly works in the mapped space with unit variance in all directions.
This PR adds that. The main change is a parameter
std_mat
to the esh momentum update.