MATLAB Answers

Mark Whirdy

Sparse Matrix - More Efficient Assignment Operation

Asked by Mark Whirdy
on 2 Apr 2013

I have a large sparse matrix for which I am attempting to assign certain segments (simplified example as per below snippet) to some identity matrices. The assignment method below is extremely inefficient, taking up a large amount of memory and too much time (10sec for example below on my server, but my real dataset is much bigger). Looking at the screenshot from taskmanager, we can see clearly a large spike in memory usage while the operation is being carried out.

For somebody used to dealing with large sparse matrices, there should be a more efficient way of carrying out such an assignment operation. Probably a pretty simple solution but I have little experience with sparse-matrix syntax.

A = sparse(20000,60000);
A(10001:end,20001:50000) = ([speye(10000,10000),-speye(10000,10000),speye(10000,10000)]);



No products are associated with this question.

3 Answers

Answer by the cyclist
on 2 Apr 2013
Edited by the cyclist
on 2 Apr 2013
 Accepted answer

That operation took less than 3 milliseconds on my machine.

The resulting array is 960,008 bytes.


MATLAB R21013a on a Macbook Pro running Mountain Lion

We ran on linux server & got 15sec on a 2009 matlab. Maybe there have been some improvements in sparse handling between versions at some point - any mathworks developers reading this have any insights?

Hi Mark,

I can run the timings for you on any and all releases when I return to Natick on Friday.

Answer by Cedric Wannaz
on 2 Apr 2013
Edited by Cedric Wannaz
on 2 Apr 2013

The structure of sparse matrices in memory makes this kind of indexing operations slower than building the sparse matrix directly using vectors of row/col IDs and values.

 I = [10001:20000, 10001:20000, 10001:20000] ;
 J = [20001:50000] ;
 V = [ones(1, 1e4), -ones(1, 1e4), ones(1, 1e4)] ;
 S = sparse(I, J, V, 2e4, 6e4) ;
 spy(S) ;    % Check structure.

You can easily make it more flexible/concise. I can't test it now though.


ok, thanks guys

Slight change to Cedric's method to get the size as desired (added a 0 element at the end):

 I = [10001:20000, 10001:20000, 10001:20000 20000] ;
 J = [20001:50000 60000] ;
 V = [ones(1, 1e4), -ones(1, 1e4), ones(1, 1e4) 0] ;

Also, there IS a version specific issue here. The original code OP posted generates an out-of-memory on my Win32 XP machine through R2010b, but is fine for R2011a onwards. Which leads me to believe that there is something in the indexing or concatenating that changed between these versions with regards to sparse matrices and how they use temporary memory in the background. The final result isn't that large.

Regarding OP's comment that he needs to insert different values (other than 1's and -1's I presume), we would need to see some specifics in order to offer any suggestions.

It would be interesting to test with a smaller size matrix that would fit in memory as a dense matrix. There might be some conversion to dense in particular cases of indexing that could be spotted this way. I've been using quite intensely sparse matrices in MATLAB since ~2006 I guess, and I already had troubles with unsuspected conversions to dense a few times (but not while CAT-ing sparse matrices, as it is not an operation that I am doing frequently).

Answer by Teja Muppirala
on 3 Apr 2013

The performance of sparse matrix indexing was enhanced in R2011a.

If you can't upgrade, as somewhat of a workaround, you should be able to get away with something like this:

A = sparse(20000,60000);
A = spreplace(A,10001:20000,20001:50000,[speye(10000,10000),-speye(10000,10000),speye(10000,10000)]);

Where SPREPLACE is the following general purpose function:

function A = spreplace(A,I,J,B)
% Equivalent to 
% >> A(I,J) = B
% But does not support "end" indexing in I and J
I = I(:);
J = J(:);
[iA,jA,sA] = find(A);
[iB,jB,sB] = find(B);
trimA = ~(ismember(iA,I) & ismember(jA,J));
A = sparse([iA(trimA); I(iB)],...
           [jA(trimA); J(jB)],...
           [sA(trimA); sB],...

When I try this in R2007a, Your original code takes 15 seconds, The workaround above gets it done in about 5 milliseconds.


Join the 15-year community celebration.

Play games and win prizes!

Learn more
Discover MakerZone

MATLAB and Simulink resources for Arduino, LEGO, and Raspberry Pi

Learn more

Discover what MATLAB® can do for your career.

Opportunities for recent engineering grads.

Apply Today

MATLAB Academy

New to MATLAB?

Learn MATLAB today!