Re[4]: non linear regressions and histograms.

[Alan G Isaac <aisaac american edu>]

On Tue, 10 Jan 2006, Morten Welinder apparently wrote: 
> The problem with picking a random open (or closed, 
> reallyl) software product for numerical work and using it 
> is that you need to figure out if it is any good. 

I was not suggesting a random choice.  Just that a choice be 
made that provides a good range of functionality along with 
ease of use.  Taking advantage of Gnumeric's existing work 
on Python scripting made numpy look pretty interesting.

> A quick look at numpy suggests that the numerical quality is questionable. 
> Take for example 
> static double atanh(double x) 
> { 
>       return 0.5*log((1.0+x)/(1.0-x)); 
> } 
> This clearly has 100% loss of precision for small x.  
> (Luckily this function is only used if libc fails to 
> supply atanh.) 

This is a classic "naive" implementation, which 
unfortunately can be found in open and commercial software.  
If your claim is true, it could affect Windows users.
But where are you finding this: are you sure the current 
numpy supplies atanh separately from an available math 
library (possibly not necessarily libc)?  I see the 
following::

    Python 2.4.1 (#65, Mar 30 2005, 09:13:57) [MSC v.1310 32 
    bit (Intel)] on win32
    Type "help", "copyright", "credits" or "license" for more information.
    >>> from numpy import *
    >>> help(atanh)
    Traceback (most recent call last):
      File "<stdin>", line 1, in ?
    NameError: name 'atanh' is not defined

Cheers,
Alan Isaac