Adjusted MCLMC

[reubenharry]

Presentation of the new sampler

MCLMC is an unadjusted algorithm, in the sense that no MH step is performed, and it is unbiased. As it turns out, it is possible to calculate the MH term, so we can implement an adjusted version.

How does it compare to other algorithms in blackjax?

Similar to MCLMC, but with adjustment. In most cases, it will be slower to converge (especially in high dimensions), but our tests suggest that it is generally faster than NUTS.

Where does it fit in blackjax

Next to HMC and NUTS. It may be possible even to share some code.

Are you willing to open a PR?

Yes, forthcoming.