Andrew Johnson

> FYI. I've tested this successfully with the experimental version RStan. > > PS: It conflicts with #2641. Great news! Yeah both PRs partially overlap in the same changes, so...

I've narrowed down the `mdivide` failures to something related to `vector` vs `row_vector` differences, since the tests all pass if `vectors` are used instead

I've traced the `mdivide` failures to the `finite_diff_grad_hessian_auto` function in the autodiff testing framework. The gradients calculated by the autodiff framework are consistent between a `row_vector(2)` and a `matrix (1,...

@SteveBronder I've narrowed things down to the `hessian` function, but I can't figure out why the behaviour is different. I've put together a minimal reproduction of the part of the...

Ah good catch, thanks. I knew that merge was too painless to be true

@SteveBronder this is now passing all tests (including `mdivide_`), but the `2pl_irt` model from the performance tests is failing. When I run locally it looks like the values are consistent...

The difficulty with implementing hypergeometric series and their extensions comes from the need to also be able to implement their derivatives/gradients, which as I've experienced in [this PR](https://github.com/stan-dev/math/pull/2510) can get...

Ah thanks for catching this. The `nan` pops up because an in `Inf` `x` argument to `log_inv_logit_diff` causes an evaluation of `Inf - Inf`, which resolves to `nan` rather than...

Note that the second case: ``` #> log_inv_logit_diff(y, x) #> [nan,nan,nan] ``` Is correct, since the difference in the inverse logits is negative and so their log is not defined:...

Nice to know I'm on the right track! Any chance you could point me towards some of the `rstan` source where this has been implemented? That way I can make...