I am sampling 440 parameters successfully with the metric set to
dense_e, using cmdStan.
This is done with 2000 data-points and a multivariate prior.
Increasing the model complexity to the next natural step results in 483 parameters.
In this case the run fails spectacularly, even though the amount of data and warm-up iterations are increase significantly.
Looking at the step-size during the warm-up period showed that it had converged as shown below:
But the mass matrix never gets close to what it is supposed to, even though I give it a very very warm guess. Wrongfully estimated mass matrix below (Actually it is the correlation matrix derived from the mass matrix):
As a reference I included the “correlation” mass matrix from the successful run with 440 model parameters
metric='diag_e' there is no problem. I can even go to 624 model parameters.
Has anyone experienced this kind of problem with the mass matrix?
Thanks in advance