Neil Girdhar

> I am trying to add the generalized inverse Gaussian distribution. Awesome!! Very cool. > Please let me know if you have any suggestions on this issue. Have you looked...

> But can I use directly in the log normalizer without breaking ExpToNat? I Yes, I think so. You can see how I imported other Bessel functions into `_src.tools`.

> the Newton method used in ExpToNat does not support constraints Instead of adding constraints, the trick is to ensure that the flattened parametrization is over the entire plane. It...

FYI, these are the errors I get running this PR ``` ====== FAILED tests/test_distributions.py::test_conversion[GeneralizedInverseGaussian] - Failed: Conversion failure FAILED tests/test_entropy_gradient.py::test_nat_entropy_gradient[GeneralizedInverseGaussian] - TypeError: Gradient only defined for scalar-output functions. Output had...

Cool, just checked and it looks like you forgot the constraint on `p_minus_one: JaxRealArray = distribution_parameter(ScalarSupport())`?

It would be really cool to have this in Optimistix! > jax gives an error and suggests to use `jvp` directly: That's funny, I think I actually added that to...

> We do not probably need the flag explicitly, as input structure has the required information of which part is complex. Totally agree! > What we need is then during...

> The answer for all Python objects typically was "you can keep it on GPU and duck type it". Why don't you provide an abstract base class or protocol and...

> Sadly there is no AND logic join in types. Looks like RedKnot will support intersections (from reading pull requests), but yes, it's not in the typing standard yet. Anyway,...

Is [this usage](https://github.com/data-apis/array-api/issues/815#issuecomment-2182728244) guaranteed? If so, should it be added somewhere to the specification? I looked for it [here](https://data-apis.org/array-api/latest/API_specification/type_promotion.html#mixing-arrays-with-python-scalars). FWIW I also like the `item` method since it's all I've...