Inserting 0 in sparse matrices without changing sparsity pattern?

Hello,

In my code I have very large sparse diagonal matrices, where the values of the diagonal change over time.

The code is very long, so I instead have an example of the issue profiled as seen below. If the new diagonal b contains zeros (line 17, 30% zeros), it is much slower compared to if b contains no zeros. In my own code b usually has between 5-30% zeros. My guess is this is because MATLAB reformats the sparse matrix when a zero is inserted(i.e. the element is deleted, indices are changed). Is it possible to tell MATLAB to not change the sparse matrix, but instead treat inserted zeros as any other value? By this I mean that the value 0 is actually saved in the diagonal, despite it being a sparse matrix.

Answer 1

[Bruno Luong]

Is it possible to tell MATLAB to not change the sparse matrix, but instead treat inserted zeros as any other value?

Short answer: No.

You can insert something that replaces 0, for example NaN or realmin and do specific treament with this value.

[Anders Melander]

Yeah, I had already used realmin as a "hack", just wondered if I could have avoided it, but seems not.

Thanks for the answer!