Allocating column of sparseArrays adds explicit zeros

General Usage Performance
1

I’m not sure if this is a bug or a feature but copying a sparse vector into a column of a sparse matrix keeps (or rather adds) some of the zeros explicitly:

 julia> using SparseArrays
 julia> m=sparse([[0,1,2,0,0] [0,3,0,4,0]])
5×2 SparseMatrixCSC{Int64,Int64} with 4 stored entries:
  [2, 1]  =  1
  [3, 1]  =  2
  [2, 2]  =  3
  [4, 2]  =  4

julia> v=sparse([1,2,0,0,0])
5-element SparseVector{Int64,Int64} with 2 stored entries:
  [1]  =  1
  [2]  =  2

julia> m[:,1]=v
5-element SparseVector{Int64,Int64} with 2 stored entries:
  [1]  =  1
  [2]  =  2

julia> m
5×2 SparseMatrixCSC{Int64,Int64} with 5 stored entries:
  [1, 1]  =  1
  [2, 1]  =  2
  [3, 1]  =  0 <-----
  [2, 2]  =  3
  [4, 2]  =  4

julia> findnz(m)
([1, 2, 3, 2, 4], [1, 1, 1, 2, 2], [1, 2, 0, 3, 4])

I understand explicitly allocating a particular entry of the matrix to 0 does just that, but here the right hand side is a sparse vector, so I don’t get why this is happening.

2

It didn’t allocate an explicit zero, it just chose not to drop the stored entry (that was there from the beginning). Use dropzeros! to drop eventual stored zeros.

1 Like
3

I know about dropzeros, but the point is that it was not explicitly stored in v, so I wouldn’t say it was there from the beginning ! Besides, there were three zeros in v, and only one shows up in m so it’s definitely a strange behaviour don’t you think ?

4

Oh, I see, that’s because the 3 that was already there is replaced by 0 and not dropped ?

5

Yes.

6

I find this behavior surprising. Since assignment has to handle allocating new elements, it seems odd that it wouldn’t also deallocate elements when appropriate.

7

I do find this surprising too, at least as a default behavior, whenever I write A=B I’d expect A and B to be then, well, pretty much identical.

8

Rearranging the sparsity structure is quite costly. If you are iteratively updating a matrix using a loop, doing this for each element would be very expensive. Better to have it performed explicitly (eg when you are done updating).

1 Like
9

Sure. My point is, it already seemed to be rearranging the sparsity structure by allocating new elements. So I guess it’s just doing the minimal necessary restructuring (?).