Quickly applying a function to all possible combinations of vectors

I have a function f that takes two vectors x and y and

 f(x,y) = Transpose(sin(x-y))*Z*sin(x-y)

where Z is a diagonal positive definite matrix of appropriate size. Now, I have a matrix X where each column is a vector (so X1 = X(:,1), X2 = X(:,2), etc). I want to make a matrix K where K(1,1) = f(X1,X1), K(1,2) = f(X1,X2), etc.

I know i can do this with nested for loops but it runs very slow (X contains over 1000 points and K is calculated a few hundred times and it ends up taking about two hours). Now the code I'm basing it off of has similar functions and can do it in about 2 minutes. Anyone know how to quickly do this?

[Matt J]

What is size(X)? You said X has over 1000 columns, but how many rows?

[Eric]

I want it to for X with any number of rows

[Matt J]

Yes, but will the number of rows typically be much greater, much less, or comparable to the number of columns? It would affect the approach used.

You can't perform the computation for any number of rows that you want. At some point, you're going to face limits of computer memory.

[Eric]

The number of rows will be significantly lower than the number of columns (i plan on using it for about 10 or so rows)

Answer 1

[Matt J]

Using only 1 loop,

    K=0;

    for i=1:size(X,1)

     xy=X(i,:);
     K=K + sin(bsxfun(@minus,xy.',xy)).^2*Z(i,i);

    end

[Eric]

A good idea, but it ran at about the same speed

[Matt J]

With what dimensions of X?

When I do it on my machine with size(X)=[2000, 2000], it completes in about 90 sec. If your data sizes are much bigger than that, you may have to face the fact that the computation time you're seeing is mainly dictated by data size and nothing more.

[Matt J]

The number of rows will be significantly lower than the number of columns (i plan on using it for about 10 or so rows)

When I run with size(X)=[10,2000], my proposed implementation finishes in less than 0.5 seconds. Even if you have a much poorer machine than me, there's no way it could still be taking 2 hours.

That's assuming of course that 2000 is a reasonable upper bound on the number of columns you're using. If it isn't, you might consider responding to my (third!!!) request for more full information on the size of X that you're benchmarking with.

[Eric]

Had to make some adjustments for my code and i made a mistake in translation, works great now, thanks!