Re: [Bug-gsl] possible bug in F distribution computation function gsl_ran_fdist_pdf

[Brian Gough]

At Tue, 5 Jan 2010 14:55:19 -0500,
Gunjan Verma wrote:
> When I run gsl_ran_fdist_pdf(.012, 8, 249) I get infinity.
> But when I run
> 
> gsl_ran_fdist_pdf(.012, 8, 248) I get .000074, which is correct (checked
> against MATLAB and R)
> 
> one d.f. change makes this difference - but I don't see why it should from
> the formula for F pdf.
> 
> i am running a Dell Latitude E6500 machine with ubuntu 9 linux , 64 bit and
> gcc 4.4.1   . I built the gsl library from source for my system.

Thanks for the bug report.  I can see the problem - there is a product
of terms exp(A)*B*C where the exponential overflows.  It needs to be
rewritten as exp(A+log(B)+log(C)) when there is a danger of overflow. 
I've logged it as a bug to be fixed in the next release 
https://savannah.gnu.org/bugs/index.php?28500

-- 
Brian Gough

GNU Scientific Library -
http://www.gnu.org/software/gsl/