Implement SOL-HMC algorithm

[ben18785]

See paper. This also contains information on the algorithm.

[ben18785]

Unclear whether this method requires Gaussian priors and have emailed author.

[ben18785]

Contact with Michela Ottobre:

Dear Ben,

very happy to be contacted about this!
At least in finite dimensions (which is in practice always the place where you implement the algorithm), and assuming your target is supported everywhere in R^N, you can always multiply and divide by a Gaussian - basically artificially change the log-likelyhood, so you are back to the case treated in the paper. So, after using this trick, yes, even if the original prior was not Gaussian you can still bring your problem to having a Gaussian prior. Let me say though that the resulting log-likelyhood may not exactly satisfy all the assumptions of our theorem - so the full theory would not necessarily apply, but you can still implement the algorithm

The above makes less sense if your target is not supported on the whole R^N, and in this case we have worked on modifications of the algorithm which work for measures supported on a box. This is the variant in that case https://arxiv.org/abs/1903.06960

I hope I understood your question and the above makes sense. If not please let me know!

All the Best,
Michela