@cgohlke wrote on 2011-04-17

The following code returns unexpected results with any of my 64 bit msvc9 builds of numpy 1.5.1 or 1.6.x, and also with EPD 7.0-1 64 bit:

>>> import numpy as np
>>> np.random.seed(1234567890)
>>> np.random.noncentral_f(dfnum=5, dfden=2, nonc=1.0, size=(3, 2))
array([[ 1.62003345,  1.7253997 ],
       [ 0.96735921,  0.42933718],
       [ 0.71714872,  6.24979552]])

Expected result:

array([[ 1.405981  ,  0.34207973],
       [ 3.57715069,  7.92632663],
       [ 0.43741599,  1.17742088]])

The code fails whether MKL is used/linked or not.

The 32 bit msvc9 and 64 bit "Windows Server 2003 R2 Platform SDK" builds are OK. So this seems to be a 64 bit msvc9 specific issue.

Other numpy.random tests pass with the same seed. Specifically, chisquare and noncentral_chisquare return expected results.

The noncentral_f function is implemented in numpy\random\mtrand\distributions.c as

double rk_noncentral_f(rk_state *state, double dfnum, double dfden, double nonc)
{
    return ((rk_noncentral_chisquare(state, dfnum, nonc)*dfden) /
            (rk_chisquare(state, dfden)*dfnum));
}

I changed the function to the following, which does return the correct results:

double rk_noncentral_f(rk_state *state, double dfnum, double dfden, double nonc)
{
    double t1, t2;
    t1 = rk_noncentral_chisquare(state, dfnum, nonc) * dfden;
    t2 = rk_chisquare(state, dfden) * dfnum;
    t1 /= t2;
    return t1;
}

So far I only tested this with numpy-1.5.1.win-amd64-py2.6.

@rgommers wrote on 2011-04-17

Strange problem. Your change looks OK to me and works with Python 3.1 on OS X.

@cgohlke wrote on 2011-04-17

Could be that msvc9 compiled code for amd64 executed rk_chisquare() before rk_noncentral_chisquare(), unlike code compiled by other compilers. In that case the following should work reliable (not tested):

double rk_noncentral_f(rk_state *state, double dfnum, double dfden, double nonc)
{
double t = rk_noncentral_chisquare(state, dfnum, nonc) * dfden;
return t / (rk_chisquare(state, dfden) * dfnum);
}

@cgohlke wrote on 2011-04-18

The attached patch fixes the test failure on win-amd64 and should be safe for all platforms.

Attachment added by @cgohlke on 2011-04-18: mtrand_distributions.diff

@rgommers wrote on 2011-04-22

Thanks, applied in 137ec8e and 7bb23a7e.