Distribution sampling

I have 2 million samples with three parameters (a,b,c). These are correlated each other and each have different distribution (not gausian or logarithmic). Now I need to collect 60,000 samples of them with same correlation and same distribution. Is there any particular method any one can suggest? Can any one help me?

Answer 1

[Doug Eastman]

I'm not a statistics expert but I believe randomly sampling a set of data should come close to preserving the distribution and correlation of the original data, so here's a way to take a random subset of length n of an array A:

i = randperm(numel(A));
subset = A(i(1:n));

Here's an example showing the preserved distribution:

N = 100000;
n = 10000;
x = randn(N,1)*3+12;
y = randn(N,1)*2+2;
A = [x;y];
i = randperm(numel(A));
subset = A(i(1:n));
hist(A,100);
figure
hist(subset,100);

[Lakshman Dontha]

Thanks, but will it apply if A is of mxn matrix?

[Doug Eastman]

Sorry, fixed a typo above, but yes, this will work for any dimension A because it uses linear indexing (only one number for the index).

If you have something like 1000x3 where you want 100x3 (100 of the 1000 original samples), you would do:

i = randperm(size(A,1));

subset = A(i(1:n),:);