The method assures good statistical properties for the marginal probabilities, and a uniform distribution across the joint probability space. All good things.
If you download Roger's code, I suggest you run this test to compare methods. (I'll post a figure with the result later, if I can.)
B1 = randfixedsum(10,10000,1,0,1);
A = rand(100000,10);
B2 = bsxfun(@rdivide,A,sum(A,2));
figure
subplot(2,1,1),hist(B1(1,:),25)
subplot(2,1,2),hist(B2(:,1),25)
Best Answer