Jacobian/Gradient Function


I am trying to implement Generalized Fiducial Inference into STAN. I was wondering if there already exists a function to calculate the Jacobian.

There is such a function in the Stan Math Library, but it is not exposed to the Stan language. So, you would either need to define your own C++ function that uses it or write the derivatives out in the Stan language. Let us know how it goes. I have never been able to figure out why someone would want to do GFI, but also always thought you would need to use Stan in order to do it.

I could also use a Jacobian function. It should be much faster to call the c++ function, right? Any pointers (maybe an analogous example) would also be welcome :)

Not sure what you mean by the C++ function. Everything’s C++, right?

The fastest thing to do is implement your own analytic derivatives rather than use nested autodiff. You can see that in all of our internal functions. There are helpers in rev like precomputed_gradients, operands_and_partials, and adj_jac_apply depending on what you’re trying to implement.

You’ll also need ones that are fully templated to deal with all the double and int instantiations and to deal with fvar if you want to add this to Stan itself.