Several reasons why RMHMC implementation is not exposed is explained in this post here with some summary as follows.
- has fragile numerical integration
- hard to detect the source (geometry or implementation) of problems
- can achieve same performance from model reparameterization instead of chaing metric
@betanalpha if you still hold this view, I wish to ask in what ways studies on Riemann manifold could contribute to Stan? Both for MCMC and optimization.
Optimization is included because I thought the second problem above could be relatively overlooked for optimizing function where performance has priority over diagnostics (of course not that they are unnecessary). Also its interface already provides several algorithms to choose from.