How to add two sparse matrices or ndSparse matrices in MATLAB?

I am aware that sparse matrices are most effciently worked with when they are created at once from a set of indices and values.

However, due to the size of my data, I am unable to store all indices and values at once, and am forced to batch my execution, so that a sparse matrix is built from a set of indices and values once in a while (say every 100th input file).

This introduces the problem of how to add two sparse matrices, which are not guaranteed to have the same dimension, as I need to accumulate the values (add them) if indices are identical.

I am currently using the ndSparse class from File Exchange, but would also be interested in an answer for MATLAB's built-in sparse function.

The main problem lies in making sure both sparse matrices have the same dimension. Is there any smart/efficient way to do this? Using padarray is not an option.

[James Tursa]

"The main problem lies in making sure both sparse matrices have the same dimension."

Can you elaborate on that more? Maybe provide short code snippets showing what your are currently doing and what you are proposing to change? The sparse() function in MATLAB does have some options for setting the matrix dimensions during the build.

[Ahmad Gad]

Hi Jakob,

I am having the same problem as you exactly described in paragraphs one and two. However, I don't know how efficient is to discretize your calculation for all indices. Could you tell me how you did this and was it efficient or not?

Thanks in advance.

[Jakob Havtorn]

As far as I recall, I ended up doing a couple of if-statements to check whether the dimensions of the two ndSparse matrices were identical. If not, I would obtain the dimension of the larger of the matrices and index into the smaller of the matrices at this index and assign a zero.

This dynamically allocates the extra space for the single zero while increasing the dimension available for assignment in the smaller matrix to match that of the larger. The two matrices can then be added as usual.

I found that the if statements and the dynamic allocation of a single double did not reduce the performance significantly (only a few scalar operations compared to addition of large, but sparse, matrices).

Answer 1

[Johannes Korsawe]

help sparse

fourth and fifth argument of sparse prescribe the dimensions. If you know them from the beginning of ayour code, you can define all matrices of the same size and add them regularly.