Non-invertible change of variables?

[Bob_Carpenter]

@dlakelan

You can re-express that if you like in your code, and then the jacobian warning will go away.

This is a dangerous benchmark. it’s easy enough to get rid of Jacobian warnings where you really should be calculating one. Just using target += won’t catch any potential Jacobian issues.

@dlakelan

Would it be hard for Stan to determine that the left hand side of a sampling statement included a transform of both parameters and data, and then alter the warning?

In a very real sense these problems are all undecideable because Stan gives you a Turing-equivalent imperative language so you can’t decide statically which statements will get executed in a call to the log density. So it’s all going to come down to heuristics.

@dlakelan

I do think the Jacobian warnings are very helpful, but they currently don’t detect the difference between say a pure invertible parameter transform requiring a Jacobian correction, and a relationship between data and a parameter that defines an implicit likelihood and so they sometimes lead people astray.

I completely agree. I’m not the one to talk to about this becuase I barely understand implicit likelihood functions. Ben Goodrich and Michael Betancourt are the ones to talk to.