Is it possible to access mean and the covariance of vb?

Is it now possible to somehow retrieve mean and the covariance matrix of a vb run? (all my parameters are natively unconstrained).


Which interface? The mode of the approximation is stored in the csv file as the first numeric row. I don’t remember if there is option for getting covariance, but if you tell which interface I can check the corresponding user guide. I guess there is no such option as we for RStan we had to also wait to get log density computations, which use covariance, needed by advi diagnostic to C++ part.


Basically, my intention was that if I wanted to have an idea about the extreme tail of the produced posterior from vb, instead of sampling 30000 draws of 20000 parameters (too much RAM) I could get the means and covariance and calculate quantiles with some external functions (

I know it would be far from the true value, but I just need order of magnitute precision.

Or maybe if I have enough data I just take enough samples to define the mu and overdisperion of my negative binomials and calculate quantiles in R from those estimates, instead of generating 30000 quantities of those distributions. In this case I would loose the information about the uncertainty of those parameters, but maybe it would be the same result of modeling the vb multivariate normal posterior process.