Nested for loop in parfor, indexing

Hello!

I'm running Monte Carlo simulation in Matlab 2012a, the following code returns an error

 if matlabpool('size') == 0 

    matlabpool open 

 end

 parfor i=1:M

        for j=2:N
             x(i,j)=x1(i,j-1)+sqrt(x(i,j-1))

             y(i,j)=y(i,j-1)+sqrt(y(i,j-1))

             z(i,j)=z(i,j-1)+x(i,j-1)+y(i,j-1)
       end
 end

 matlabpool close

(Definitions of x ,y, z are simplified but the main idea remains) As I understand it is impossible to use indexing this way in a nested 'for' 'cos Matlab cannot check if array elements are independent. But in my situation it's clear that index i (different MC paths are independent) is remaining unchanged in for loop, only j (time steps) is being changed. Could you please give me a piece of advice with this issue?

Thanks in advance, Alex.

P.S. the error is obvious for such cases: The variable x in a parfor cannot be classified.

Answer 1

[ChristianW]

You could use a seperated variable for the inner loop. On the example of x:

X = zeros(M,N); %preallocate
parfor i=1:M
    x1 = zeros(1,N); %preallocate
    x1(1) = x(i,1);
    for j=2:N
        x1(j)=x1(j-1)+sqrt(x1(j-1));
    end
    X(i,:) = x1;
end

But usually this should be much faster:

X1 = [x zeros(M,N-1)]; % first column and preallocate
for k=2:N
    X1(:,k)=X1(:,k-1)+sqrt(X1(:,k-1));
end

[alex]

Big thanks for you answer, Christian, the first part of your code works great! But still I cannot implement the second part, maybe it's really impossible to do this way in my case. The issues is that in real loop is a bit more complicated: I need to have a random vector and to apply each element of it to my X1 row in the way

X = zeros(M,N); %preallocate
parfor i=1:M
    eps1=randn(1,N)
    x1 = zeros(1,N); %preallocate
    x1(1) = x(i,1);

      for j=2:N
          x1(j)=x1(j-1)+eps1(j-1)*sqrt(x1(j-1))
      end

X(i,:) = x1;
end

The upper code works fine, but in order to follow your advice I tried the following code and clearly it didn't work

     eps1=randn(M,N);
      for j=2:N

          x1(:,j)=x1(:,j-1)+eps1(:,j-1)*sqrt(x1(:,j-1))

       end

The point is that the first index in eps isn't changin - so I have one and the same eps1 for all i indexes (1 to M). I also tried to put eps into for

    for j=2:N

eps1=randn(1,N);
        x1(:,j)=x1(:,j-1)+eps1(:,j-1)*sqrt(x1(:,j-1))

       end

This construction neither worked. Am I using this loop wrong or it is impossible to make such a short construction?

Thanks in advance!

[ChristianW]

Both work. The first one is faster. You have some errors in it. You need to give him the first column.

x2 = [x zeros(M,N-1)]; % first column and preallocate
eps1 = randn(M,N);
for j = 2:N
    x2(:,j) = x2(:,j-1)+eps1(:,j-1).*sqrt(x2(:,j-1));
end

x3 = [x zeros(M,N-1)]; % first column and preallocate
for j=2:N
    eps1 = randn(M,1);
    x3(:,j) = x3(:,j-1)+eps1.*sqrt(x3(:,j-1));
end