Repository navigation
X_logit_{rng, lpdf, lupdf}(theta) should allow -infinite values in theta #3201
Description
Activity
- marked categorical_logit_rng() doesn't accept negative_infinity() input stan#3331 as a duplicate of this issue
on Sep 1, 2026 Hi, was wondering if this is still open, if so I'd like to take a shot at it.
Yes, this is still open. I added one more case after reading it.
I'm not sure what we should do for autodiff if the value is infinite.
We also have to check if all the values are negative infinity and do something, either make it uniform, or throw an exception.
Hi @bob-carpenter, thanks for the update! I think throwing an error for the all
-infcase makes most sense rather than defaulting to uniform. For+infin_rng, I'll make one+infdeterministic and multiple+infs uniform among those spots, but I'll keep a bound check for the autodiff case for now. Let me know if that sounds good and I'll get started.Thanks. That sounds good, @godking123
Reacted by Rajit SareenHi @bob-carpenter I implemented all of the changes as discussed, this is my first PR in this repo so any tips on commenting/code quality would be really helpful and please do let me know if anything is wrong. Thanks you!
Hi @bob-carpenter I added a new commit to the PR with more comments properly explaining the new bounds checks and more doxygen notes
Thanks. I'm really hoping @SteveBronder can look at this. I'm going to go down the hall and ping him in person after this.
Reacted by Rajit SareenSounds good, thanks for the help!
The safety checks in the current implementations of
categorical_logit_rngandmultinomial_logit_rngare too strict. The doc sayscategorical(softmax(x))is the same ascategorical_logit(x), but that's not the case becausesoftmaxallows negative infinite inputs.There are two things to fix:
X_logit_rngfunctions to allow -infinity.X_logit_lpdfandX_logit_lupdfwhen the argument is a data variable.For 3, we still need to flag cases where the argument is an autodiff variable because the infinite values will wreak havoc with derivatives.