Re: cygwin syscalls precision problems

[Reini U. <ru...@x-...>]

2008/6/1 Sam Steingold <sd...@gn...>:
>> * Reini Urban <eh...@k-...> [2008-06-01 13:17:49 +0200]:
>>
>
> what do you get here?
>
> (loop :for i :from -10 :to 10
>  :for e = (os:erf i) :and c = (os:erfc i)
>  :for s = (+ e c)
>  :unless (= 1 s) :collect (list i e c (- 1 s)))

((-1 -0.8427007929497149d0 1.8427007929497148d0 1.1102230246251565d-16))

It's a deviation in the last digit representation,
but unfortunately it's a significant error for the erf(x)+erfc(x)=1 assumption.

And I found another newlib problem, which I also reported upstream:
(erfc 27) produces a floating point underflow
But that's all in the range [-100, 100], everything else is just noise
at the end
caused by a different implementation.

http://sourceware.org/ml/newlib/2008/msg00306.html
-- 
Reini Urban
http://phpwiki.org/ http://murbreak.at/