Hi there,
I’d suggest reading through Albert’s response
So in line with that, I would try:
- add init messages for each separate node and init marginals (you can remove them later)
- take out the broadcast of sigma and create an auxiliary variable (as per Albert’s response), where you provide explicit factorization via
constraint - look for inspiration in the test suite
But it’s all a guesswork from me, as I haven’t tried to build MvNormal. I’m not sure how the factorization between location and covariance in McNormal works in ReactiveMP, because you should use Cholesky there…