Deprecate sparse.spmatrix.todense

Is your feature request related to a problem? Please describe.

np.matrix is deprecated, and is slowly wiped from the public code. New users, I hope, don’t learn what np.matrix is, and that’s to the better.

At the same time, sparse matrices have a public interface for creating numpy matrices, which is actually used quite a lot in scipy tests. A new user who is unaware about the existence of numpy matrix type does not have a clear way to anticipate that a matrix is returned, or that it’s deprecated. So a possible scenario (from my personal experience, even though I’m not a new user) is that:

• The user calls spmatrix.todense()
• The following code either emits a deprecation warning for doing matrix operations, or does the usual confusing stuff with matrix vector multiplication.
• The user realizes they should have used spmatrix.toarray(), switches to it and carries on.

Describe the solution you’d like

I propose to move towards making .todense an alias for .toarray through a deprecation cycle. This could be done by:

1. Eliminating .todense from scipy own codebase
2. Adding a keyword-only argument return_array=False to .todense and emitting a deprecation warning if it isn’t set to True.
3. Making return_array=True a default
4. Eventually eliminating it and making .todense an alias to .toarray.

Describe alternatives you’ve considered

Other options are:

• Keeping todense around until numpy removes it.
• Making it an alias right away and hoping nobody will notice 🙃 idk

[stefanv]

Will be partially addressed by #14822, but not clear how to address the problem of sparse.eye returning a sparse matrix by default.

[akhmerov]

Perhaps, I’d start from pinging @perimosocordiae as the code owner of scipy.sparse: is now the time to start getting rid of np.matrix legacy?

#14131 and https://github.com/scipy/scipy/pull/14488#discussion_r678451919 indicate that there is ongoing cost both to users and developers from scipy.sparse preferring a np.matrix interface.