Improving the performance of a for loop in Matlab

Hello,

I am trying to improve the speed of the following for loop in Matlab. As it is now is incredibly slow. Maybe vectorizing? But one vector slides on the other and constantly changes.

for m = 1:size(phi,1) - (constant)/2

    phi(m) = phi(m).*(mean(conj(phi(1+m:(constant)/2+m))));

end

Thanks!

[dpb]

What are phi and constant? Is phi actually complex; and even if so,

mean(conj(x)) --> conj(mean(x))

so can be outside the loop.

[Yodish]

Hi and thanks for the answer. phi would be a complex vector whereas constant is a constant number (of elements) < than the length of the aforementioned vector phi

Michele

Answer 1

[David Goodmanson]

Hello Yodish,

'one vector slides on the other'. This is basically convolution, and since you are finding a mean it is convolution with a vector of ones, of length constant/2. Since the code is not iterative, all the means can be found first.

N = 10000;
constant = 5000;
phi_orig = rand(1,N)+i*rand(1,N);

tic                                      % first way
phi = phi_orig;
for m = 1:length(phi) - (constant)/2
    phi(m) = phi(m).*(mean(conj(phi(1+m:(constant)/2+m))));
end
toc

tic                                      % second way
phi2 = phi_orig;
c2 = constant/2;
A = conv(phi2,ones(1,c2),'valid')/c2;   % means
A(1) = [];       
sA = length(A);
phi2(1:sA) = phi2(1:sA).*conj(A);
toc

d = max(abs(phi2-phi))    % should be zero

In your code I used length(phi) rather than size(phi,1) so that phi could be either a row vector or a column vector.

[Yodish]

Yes thanks, it is indeed a convolution. It would also work with the movmean function. They both improve speed by a factor of 100.

Thanks for your precious help guys