BUG: sparse.linalg: preserve column vector kind in iterative solvers

[pv]

Ensure iterative solvers return solution x as a similar vector as the
input b is, retaining shape and ndarray vs matrix.

The issue in the original make_system code is obvious, but there are backward
compat issues which need to be considered (in particular, whether the
behavior can even be changed any more or should just be documented better).

cf. gh-8722

• backward compatibility impact?

[tylerjereddy]

Should we still aim to decide on this before the approaching 1.2.0. @perimosocordiae any thoughts?

@rgommers Concerned re back compat?

[rgommers]

@rgommers Concerned re back compat?

I only had time for a quick browse. It seems like the main issue is that 1-D array input now results in a different output shape, (1, N) vs (N,)?

[tylerjereddy]

There seem to be quite a few iterative solvers, so I suspect we need to be comfortable changing the outputs from all of them -- both in terms of retaining shape & fidelity to input object type. Doesn't look like anything would change from buggy to correct or the other way around, but strict user or library code could certainly require some modest maintenance.

If we didn't have back-compat concerns it would sound straightforward, but I've not touched the code on the sparse linear solvers. @perimosocordiae thoughts on disruption vs. benefit?

[tylerjereddy]

Bumping milestone -- would be good to have a few people chime in before this kind of decision

[tylerjereddy]

Similar comment as for the last release -- I'm bumping the milestone because there's a behavior change that we may not be able to make at this point, and there's no feedback on this.

The linked issue has also been quiet for the last year, so perhaps not much evidence of major disruption to users.

[perimosocordiae]

Sorry I missed the pings on this one! I'm leaning toward documenting this behavior better, rather than changing output types. As we (the scientific Python ecosystem) continue to move away from using np.matrix, the impact of this bug should be decreasing.

For now, I'll drop the milestone.