Re: [Bug-gsl] overflow in gsl_ran_beta_pdf, gsl 1.14

[Brian Gough]

At Thu, 15 Jul 2010 15:17:00 -0700,
Michael Radford wrote:
> For example:
> (gdb) call gsl_ran_beta_pdf(1,100,100)
> $26 = 0
> (gdb) call gsl_ran_beta_pdf(0,100,100)
> $27 = 0
> (gdb) call gsl_ran_beta_pdf(1,1000,1000)
> $28 = -nan(0x8000000000000)
> (gdb) call gsl_ran_beta_pdf(0,1000,1000)
> $29 = -nan(0x8000000000000)
> 
> but the correct result is always 0 when x == 0 or 1 and a, b > 1.
> 
> The reason is that the call to exp() overflows when a and b are
> large enough.  However, it's unnecessary to call exp() when either
> of the two calls to pow() will return 0.
> 
> I'm attaching a minimal patch which simply checks for a, b > 1
> and explicitly returns 0 in that case.
> 

Thanks for the bug report and patch! I've applied it.  Sorry for the
delay in replying.

-- 
Brian Gough