Linear solvers don't use the `init` argument to warm start solvers. about jaxopt HOT 3 CLOSED

Indeed, good catch. We need to set x0=init in the gmres call. Also I'm not sure if forwarding the kwargs is a good idea in the end. This could lead to us breaking user code if we make changes to argument names...

from jaxopt.

That's true... Should we explicitly call the arguments from the underlying solvers in jax.scipy.sparse.linalg ? The main point of these wrappers is to add the ridge argument... Should we write the with_ridge as a decorator over the underlying solvers?

from jaxopt.

I guess one of my early motivations was to have matvec support even for direct linear solvers (by materializing the matrix in memory).

from jaxopt.