Bug in tanh? A solution, maybe for the GCL people?



 

> -----Original Message-----
> From: Raymond Toy [mailto:raymond.toy at ericsson.com] 
> Sent: Monday, March 19, 2007 9:38 AM
7
> 
> I think I wrote that (1987 is only 20 years ago)....

Actually, a relief, since I don't recall doing it...!
RJF

> 
>     Richard> ....
> 
>     Richard> We can compute  d(x):= exp(x) - 1, but do it 
> carefully to preserve precision
>     Richard> when
>     Richard> |x| is small.  X is a FP number, and a FP number 
> is returned.  
> 
>     Richard> Then tanh is   d(2*x)/(2+d(2*x))
> 
> Yes, this would work, if you have an accurate d(x).
> 
> Here is an alternative, based on Kahan's complex tanh, except modified
> to work with just real numbers.  Simple tests near zero and for large
> x indicates this this works.  Works on GCL too.
> 
>
<snip>