I am wondering if this might actually be a decent use for GMO+Laplace approximation to get a better approximation to the posterior curvature at the mode. That could then be used in place of non-centering based on the prior curvature to reparameterise the model.
It wouldn’t be as efficient as RMHMC, but it might be better than just the standard non-centering.
sparse cholesky factorisations
If you have a precision parameterization, doesn’t that mean you know K^-1 for y ~ multi_normal(0, K)? Then do you still need cholesky_decompose(K)?
GMO
I apologize but this acronym is not familiar to me. What is it?
Gradient based marginal optimization. Basically to let us do lme4 tricks and Bayesian equivalents. It’s in the works, but not ready for prime time yet apparently.
Don’t forget the determinant! If K is data, then no you don’t, as we exploit in the ICAR example, but if K contains parameters, then you need the log-determinant.
I hope you know I’m spending a portion of this week camped out in your office asking you, Sean, and Mitzi increasingly arcane questions about things like this. And teasing Betancourt, but that’s a given.