Seth Axen
Seth Axen
Okay! I may open a PR in the coming weeks.
Alternatively, as noted in the code, the initial Hessian is a preconditioner. We could use the preconditioner machinery here, but the preconditioner is updated using only the latest position. If...
@pkofod do you have any feedback to my previous comment?
In general we really don't consistently enforce that cotangents of reals are real if they interact with a complex number, i.e. in most places we do the Zygote thing. There...
The rules for determinant `y = det(A)` are ultimately [due to Giles](https://people.maths.ox.ac.uk/gilesm/files/NA-08-01.pdf), who implicitly assumed `A` is nonsingular. The forward-mode rule is `ẏ = y * tr(A \ Ȧ)`. We...
> It is easy to make a stable Ddet using the SVD. Yes it is slower, but it is not less accurate, where did you hear that? As a matter...
I came across this paper "On the adjugate matrix" by GW Stewart https://doi.org/10.1016/S0024-3795(98)10098-8, which does an error analysis to show that when working with decompositions of the form `A=XDY` for...
> This seems fine. How many other functions will need this? `cbrt` is currently a scalar rule. Powers are their own messy thing. At first glance, `/`, `\`, `^`, `inv`,...
I opened a PR to ChainRulesCore that would supersede this one if merged: https://github.com/JuliaDiff/ChainRulesCore.jl/pull/551
> Functions like `inv`, `log` etc. are a slightly different class to `sqrt`, since the primal is infinite. Is this difference important though? There are plenty of cases where in...