Bob Carpenter

Sums of empty containers evaluate to 0 and products of empty containers evaluate to 1. These are the usual boundary conditions for empty containers because they generalize inductively properly, like...

I just wrote to the authors of that paper to ask what they recommend for `log_softmax`. They recommend against ``` softmax(x) = exp(x - max(x) - log_sum_exp(x - max(x))) ```...

I was going to try to fix this issue, but after spending about 8 hours trying to implement `log_softmax(x)` as `log(softmax(x))` following the advice of Nick Higham et al., I'm...

Thanks, @wds15. I'm giving up on this issue. I can't keep up with the C++ in the math lib with the limited amount of time I have to code. To...

@syclik or @mrbrubake: Could one of you verify this issue should still be outstanding. And if it is, will whoever works on it please get rid of all those redundant...

What we do with the unit vector coding is to make the transform `f` (constrained to unconstrained) one to one and the inverse transform `f_inv` many to one, making sure...

Thanks, @betanalpha, that's all really helpful. > my point is more semantics rather than practical. I agree. We shouldn't call a non-bijective function a bijection! I think TensorFlow Probability is...

@andrjohns: please let me know when this is ready to review and merge. Thanks!

> Is there any proposal for how this feature would be exposed in the language? We'd just use another decorator, like `_vec`, as in `normal_lpdf_vec`. I don't think we want...

I agree that "vectorization" is ridiculously overloaded. It technically refers to how the CPU pipelines operations in SSE, AVX, etc. Our general use for functions like `exp(vector)` is already a...