Zygote AD Jacobian-vector product

[pavanakumar]

The forward mode Jac-vec is described here in the following post

Efficient Vector^T * Jacobian matrix?

This is exactly what reverse-mode automatic differentiation is (e.g. if f is R^n to R you’re asking for a O(1) gradient). There are a few packages doing it, but it’s more tricky than forward diff.

Many thanks for pointing me in the right direction.