Patrick Kidger
Patrick Kidger
Closing now that #158 is in, we'll merge #159 instead.
Hey there! So you're actually touching on a design point from way back in the early days of Lineax. We in fact originally had a `maybe_singular_tag` on the operator rather...
> It is true that AutoLinearSolver would be SVD for undecorated operators, but this is the only source of silent inefficiencies (and this is also analagous to the psd tag:...
> so the the only information it is missing is whether the operator is full rank So let's consider a solver that consumes a `2n`x`n` matrix and a `2n`-length vector...
Sorry for taking so long to get back to you; as this is a very technical + mostly internal change, it's been pushed down my priority list. Anyway, I think...
Okay, I am very belatedly getting back around to this PR now. (Personal life issues caught up for a while.) I've just merged this into our `dev` branch. I really...
As, good spot. This is what we document... but it's also very much a footgun. 😄 In the same spirit as #156 I'd be happy to take a PR on...
Ah, agreed. I like the sound of that approach. Very happy to take a PR on that!
Yup, this is totally possible. Take a look at the `InBoundsSolver` of https://github.com/patrick-kidger/diffrax/issues/200. (Which we should probably add to Diffrax's API directly!)
I think preconditioning in general should have a place here, so yes!