Re: Overhaul of statistical distribution functions

[Rik]

On 09/21/2011 11:17 AM, Michael D Godfrey wrote:
> On 09/21/2011 10:35 AM, Rik wrote:
>> Michael,
>>
>> I cleaned up a lot, but wasn't always trying to re-implement algorithms
>> which looked reasonably solid.  The Loader paper looks good though, and
>> particularly for large N as you said.  Are there any copyright issues
>> with filing a bug report on the Octave bug tracker and attaching the
>> Loader paper?
> I am sure not.  The paper is freely available and the algorithm from it
> is widely used, including, for
> example, in R which is also free software.  The algorithm is pretty
> simple, but n as a vector may
> need an extra line or two.
Great.  I'll make a bug report out of this so that anyone with a bit of
time can tackle this.
>
>
> About the binopdf problem: do you know for what values of n the Mac OS
> version fails?  If it is
> above n=10, then the solution in any case is to implement the Loader
> algorithm.  The current code
> should never be used for n > 10.
No, the tests I wrote were not stress cases for the algorithm.  The failing
test is

N = 2
P = 0.5
X = [0, 1, 2]

Expected results are [1/4, 1/2, 1/4] or basically p^2, p^1, p^2.  This
really does seem to be something about the Mac OS implementation of lgamma.

--Rik