Column slices of sparse matrices should be sparse vectors

[alsam]

Dear developers,

For the matrix created with

julia> x = sprandn(7,7,0.3)
7x7 sparse matrix with 14 Float64 entries:
    [1, 1]  =  1.31604
    [3, 1]  =  -0.879252
    [1, 2]  =  -0.23831
    [4, 2]  =  -0.904064
    [3, 3]  =  0.721235
    [4, 3]  =  -0.317153
    [6, 3]  =  1.80547
    [4, 4]  =  0.560326
    [2, 5]  =  0.68588
    [4, 5]  =  0.0368284
    [5, 5]  =  -0.360748
    [2, 6]  =  -0.52084
    [2, 7]  =  0.891165
    [3, 7]  =  -1.13065

The output for the slice operations are:

julia> x[2,:]
1x7 sparse matrix with 3 Float64 entries:
    [1, 5]  =  0.68588
    [1, 6]  =  -0.52084
    [1, 7]  =  0.891165

and

julia> x[:,2]
7x1 sparse matrix with 2 Float64 entries:
    [1, 1]  =  -0.23831
    [4, 1]  =  -0.904064

I.e. it prints 1 instead of 2.

Thanks

[jiahao]

?

[alsam]

Hello Jiahao,

Accidentally my finger slipped when I was in the process of typing, sorry
for it.

I fixed the issue description.

Thanks,
Alexander

On Wed, Apr 9, 2014 at 5:03 PM, Jiahao Chen notifications@github.comwrote:

?

Reply to this email directly or view it on GitHubhttps://github.com//issues/6479#issuecomment-39960790
.

[mschauer]

What you see is correct, the slice of the sparse 7x7 matrix is a sparse 1x7 respective 7x1 matrix and all the elements in the slice are by definition in its first row respective column.

[mschauer]

There is a small inconsistency though that column slices of sparse matrices are column matrices and column slices of matrices are vectors.

[nalimilan]

Maybe that's another argument for not dropping dimensions by default...

[alsam]

There is a small inconsistency though that column slices of sparse matrices are column matrices
and column slices of matrices are vectors.

Exactly, besides that sum for the slice of sparse matrices returns 1-element Array, but for the slice of dense matrices returns a scalar value.

see the
#6480

[jiahao]

There is a small inconsistency though that column slices of sparse matrices are column matrices and column slices of matrices are vectors.

The problem is that Julia currently doesn't support sparse vectors. (sparsevec produces Nx1 SparseMatrixCSCs also.) Ref: #6213.

[ViralBShah]

We can support sparse vectors for consistency with the rest of the codebase. So far, I have just been lazy about it - especially since it would be a fair amount of work to make the slightly different data structure work seamlessly all operators and such.

Much less attention has been paid to the consistency of sparse operations. I am going to dive into this again. Perhaps we should have a sparse issue tag.

[ViralBShah]

That said, we certainly do not have N-d sparse objects, although we probably should, with sorted N-tuple indices and values. I am sure there is an old issue for this somewhere.

[jiahao]

we certainly do not have N-d sparse objects, although we probably should, with sorted N-tuple indices and values

The coordinate representation is probably the easiest thing to support beyond rank 2, but we'd have to also support transformations between that and the CSC representation. I do think it's high time we tracked all the sparse array stuff.

[tkelman]

closed by #13440