Row sums by vector indices without for loop?
2 views (last 30 days)
Show older comments
Best illustrated by example
M=rand(5,10);
s=[1 3; 5 8; 9 10];
for i=1:size(s,1)
sm(:,i)=sum(M(:,s(i,1):s(i,2)),2);
end
Is there a one-liner or short-cut for the for loop? I do not know how large M or s is ahead of time, or their values. Above code shown only for illustration of problem trying to solve.
Thanks!
edit: Corrected typo of 'm' in for-loop to 'M'
0 Comments
Accepted Answer
Walter Roberson
on 7 Feb 2011
If your accuracy requirements are not high and your values are not extreme, you can use cumsum(), after which it becomes the difference of the cumsum indexed at two positions -- a transformation that gets around the problem that your indices do not always designate the same number of elements to be summed per row.
If I recall collectly, Bruno, and Matt Fig, both have posted non-cumsum() versions of this task in the past (it has come up before in cssm discussions.)
2 Comments
Oleg Komarov
on 7 Feb 2011
There's a submission by Bruno, mcolon, which will allow you to create the indexes and use accumarray avoiding the multiple call to sum.
More Answers (1)
Bruno Luong
on 7 Feb 2011
Yes, to be more specific my mcolon on FEX
can be used like this:
% Data
M = ceil(10*rand(5,10))
s = [1 3; 5 8; 9 10];
% Engine
[a i] = mcolon(s(:,1),s(:,2));
[r c] = ndgrid(1:size(M,1),i);
Ma = M(:,a);
sm = accumarray([r(:) c(:)], Ma(:))
Bruno
2 Comments
Jan
on 7 Feb 2011
Did you compare the speed with the OP's FOR loop method with adding a pre-allocation?
Bruno Luong
on 8 Feb 2011
Following Jan's question: I make the following test and I get the 25% acceleration by using MCOLON
Time relative for-loop/mcol = 1.26013/1
% Code
ntest = 10;
t = zeros(2,ntest);
for n=1:ntest
s = ceil(10*rand(1,2*10000));
s = cumsum(s);
s = reshape(s, 2, [])';
M=rand(5,s(end));
% for loop Engine
tic
sm = zeros(size(M,1),size(s,1));
for i=1:size(s,1)
sm(:,i)=sum(M(:,s(i,1):s(i,2)),2);
end
t(1,n) = toc;
% mcolon Engine
tic
[a c] = mcolon(s(:,1),s(:,2));
[r c] = ndgrid(1:size(M,1),c);
Ma = M(:,a);
sm = accumarray([r(:) c(:)], Ma(:));
t(2,n) = toc;
end
t = min(t,[],2);
t = t/min(t);
fprintf('Time relative for-loop/mcol = %g/%g\n', t);
See Also
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!