Hi, Quick question… what kind of guide makes sense for an LKJ Prior over a correlation matrix (for non-diagonal noise of parameters)? An AutoDiagonalNormal/AutoNormal would be incorrect, wouldn’t it… (since wouldn’t it treat each off-diagonal element of the correlation matrix as having an independent normal posterior)? So would that mean a MultivariateNormal would be the classical guide to use for this prior? Or are neither of these recommended and I need to use some distribution over matrices?
All the examples I’ve seen of LKJ priors use MCMC that don’t require specifying a variational posterior guide, so I wasn’t sure what’s typical here?
Thanks for any insights!