Oh sorry, I probably should have mentioned - this isn’t my actual code; I ran into this bug and then reduced it down to the simplest program that still shows the same problem. Certainly I’m not worried about the fact that this model doesn’t sample well, I just don’t think it should give a theano error, and wanted to check if this was a bug or something I was doing wrong.
The reason I ended up using a 1-dimensional multivariate normal is that in my actual code I loop through several sets of observations, each of which corresponds to a different dimensional multivariate normal. So I don’t want to put in a special case for sets which only contain a single observation, when the same code ought to work fine.