larsgeb / hmc-linear-systems Goto Github PK

If a mass matrix is diagonal, a lot of computation can be sped up. Making this explicit will improve performance.

Using a general PD symmetric mass matrix is as of yet not as performant as it was in the 2017 version of the code. The cause is probably the following:

Cholesky decomposition was skipped as part of the optimization. At this point it might be useful to reimplement.

Inefficient integrator, requires a lot of transposes per sample. Consider precomputing.

I would be nice to create documentation on every method and field, as well as a theoretical introduction to HMC. The goal is to use Doxygen, and this will probably be done very incremental.

The code is much faster if positive definiteness is guaranteed for matrix A. It would be helpful if the settings accept a command line option for PD as to perform numerical optimization.

Energy levels should be evaluated before and after the propagation. In this implementation the energy level is first evaluated before the resampling, which is not valid for the hmc sampler.