BUG: AD computed derivative with respect to a `Function` is a `Function` about firedrake HOT 4 CLOSED

I believe we can close this issue. Executing ReducedFunctional.derivative(options={"riesz_representation": self.inner_product}) results in a Cofunction. Feel free to reopen the issue if anyone has valid arguments against its closure.

Don't you mean: ReducedFunctional.derivative(options={"riesz_representation": None})

from firedrake.

I think we're doing what the pyadjoint API tells us to do. You can plausibly claim that the pyadjoint API is not well throught through, because having ReducedFunctional.derivative return a gradient, which is what it requires, is dumb. I have a lot of sympathy with that. However that API has actual users, so any change requires serious work.

from firedrake.

It seems to be standard for the gradient to be defined in the primal space (although gradient covectors are used in differential geometry). However for the derivative it should definitely be in the dual space. The workaround is easy, by accessing the adj_vector attribute after running the adjoint.

Since this isn't exactly a bug, it can be closed?

from firedrake.

I believe we can close this issue. Executing ReducedFunctional.derivative(options={"riesz_representation": None}) results in a Cofunction. Feel free to reopen the issue if anyone has valid arguments against its closure.

from firedrake.