Can I avoid double forcycle by bsxfun?

Hi, I am wondering, if it is possible to use double bsxfun, or something similar. I have this piece of code:

N = 5;
prob = [0.1 0.2 0 0.1; 0 0 0.05 0.1; 0.2 0.2 0 0.1];
r = rand(size(prob,1),N);

P = zeros(N, size(prob,1));

for i = 1:N
 P(i,:) = sum(bsxfun(@ge,r(:,i),cumsum(normP([zeros(size(prob,1),1),prob]),2)),2);
end

Prob matrix contains rows with elements between 0 and 1, each row is probability distribution (after normalisation done by normP). First row of prob matrix is used to generate first element of vector P (values1,2 or 4), second row for the second element (values 3 or 4), and so on.

e.g.: P =

       2     4     2
       4     4     1
       2     4     2
       2     4     4
       2     3     1

I have already vectorised generating elements for one vector P, but I need to generate several (N) vectors. There must be way to avoid the for cycle.

In the attachment, there is the normP function. I will be glad for help, thank you.

Michal Roubalik

[Adam]

I had a quick look without managing to get the right syntax, but don't really have more time right now - I'm sure someone else will be able to suggest an answer though.

One thing I would say though is

cumsum(normP([zeros(size(prob,1),1),prob]),2)

is a constant. It never changes within the loop so take it outside the loop to only calculate it once instead of N times. Though of course that advice will likely be rendered irrelevant when someone posts a replacement for the for loop anyway so it is more just general advice on improving a for loop I guess!

[Michal Roubalik]

Thanks a lot, I have not noticed this, because I used this line for slightly different purpose before and I did not notice this change in problem while adapting the code to this problem.

It is faster by aproximately 30% now, thank you! :)