Mohamed Tarek

Changing this line https://github.com/gdalle/ImplicitDifferentiation.jl/blob/1581d2e3b1b1ddf083f0d370c9f7b323aa98f610/src/implicit_function.jl#L64 to ```julia y = implicit(x; kwargs...) ``` and using a linear solver that's compatible with ForwardDiff makes the FD case work: ```julia linear_solver(A, x) = (Matrix(A)...

>Seems easy enough to fix in the source code, should we? Let's

In theory we could. That would require defining the linear solver as an implicit function and using ForwardDiffChainRules on that. We would need to think of a good way to...

Do you know if the Python package has higher order implicit derivates? If not, this could be an interesting conference paper.

I believe what the derivations in this link are doing is going to be computationally identical to making our linear solver an implicit function and letting ImplicitDifferentiation automate the rest...

I recall you tried LinearSolve before but didn't end up using it because it was still not mature enough. Maybe we can revisit it or roll our own, shouldn't be...

I don't think we need that many solvers so direct and GMRES should be enough.

Now (in `main`), you can pass `DirectLinearSolver()` as the `linear_solver` in the implicit function to do second order differentiation. For the more general case of making the linear solver an...

I won't have a chance to look into this for the next 3 weeks unfortunately.

I have a deadline in a few days and a queue of people to get back to. I will take a look but probably won't be in the coming week.