This post on Christian Robert’s blog points to this arXived paper, which quite explicitly proposes an algorithm to sample discrete variables in an HMC context.
My first reaction (I have not yet perused the 34 pages of the paper…) is that such an algorithm would be a nice addition to Stan…
What do you think ?
Please give us the benefit of the doubt that if there are any proposals about extending HMC that we will carefully read them immediately and if there were any validity to the proposal then we would already be considering them for inclusion into Stan. Unfortunately most proposals to extend HMC are ultimately ill-posed and will not yield any practical benefits. The paper you mentioned is one of them.
Please give us the benefit of the doubt that if there are any proposals about extending HMC that we will carefully read them immediately and if there were any validity to the proposal then we would already be considering them for inclusion into Stan.
May Your Highness forgive my unability to grasp that your omniscience made unhelpful my misguided attempt to signal a newly-arXived paper…
The paper you mentioned is one of them.
Did Your Highness somehow made available His critiscism of this paper ? Where can it be read by some common plebe ?
@betanalpha: Please try to be nice to people posting to the list and give them the benefit of the doubt. They have no idea just how much our inboxes light up whenever one of these papers comes out, nor do they understand why they won’t work.
@Emmanuel_Charpentier: We’re happy to get pointers to papers people think we might have missed. We’re reluctant to get involved in explicit criticism because it takes a lot of time to do carefully and the result just tends to piss off the original authors and make us look like bullies. Also, please keep the sarcasm out of the mailing lists.