Design choices related to jvp/linearize/...

[pharringtonp19]

I should start off by saying that I enjoy working with jax.jvp and jax.linearize, so the following is not a complaint but rather a question about the design choice of these transformations.

After reading this paper, I am a bit surprised about the following example which I took from the docs

y, f_jvp = jax.linearize(f, x)

I would have thought that the following would have been the design choice. Am I missing something obvious? Thanks

y, f_jvp = jax.linearize(f)(x)

[AntixK]

Actually, the current design choice makes more sense because the linearization happens at the primal point x. Since f can be any function, it is crucial to linearize only at the points of concern i.e. x,

In your case, it would seem that the function is first linearized and then evaluated at x, which is certainly not the case.