Run bmat in solve

[kinnala]

I think solve could automatically apply scipy.sparse.bmat on lists to allow syntax such as

solve([[A, B.T], [B, None]], [f, None])

Custom function is needed for RHS to have similar syntax for numpy arrays but it should be easy enough?

[kinnala]

What about codense et al? Should that do it also?

[gdmcbain]

I suppose in condense, it's usually only the first row that needs condensing in these saddle-point problems; e.g. for Stokes that's the velocity and one doesn't apply Dirichlet conditions to the pressure.

Is the idea here syntactic sugar or is there some way to exploit the block structure; e.g. for a saddle-point like your example?

[kinnala]

I was thinking syntactic sugar, however, I'm not sure how condensing etc. would work. Should D and I also be a list?

[kinnala]

Such as

solve(*condense([[A, B.T], [B, None]], [f, None], I=[abasis.get_dofs(), None]))

[gdmcbain]

Hmm. My first thought was to accept either a list[ndarray] or just an ndarray and in the latter case to interpret the indices in the big matrix block-assembled by bmat, but then I realized that this might be awkward if the indices are passed as an ‘array-like’ list. Perhaps it would suffice to check that the len of I or D matched that of A or A[0], in which case it should be interpreted in the first way, i.e. block-wise, and otherwise try the second way?

But then maybe being a bit stricter and insisting on the form you've shown (I=[abasis.get_dofs(), None]) is reasonable. It's not too verbose and it is clearer and less ambiguous. On reflection, I think the latter might be better than accepting a Union type.

I suppose that this would be done in

scikit-fem/skfem/utils.py

Line 248 in bc57d96

def _init_bc(A: spmatrix,