[SciPy-user] Inconsistencies between numpy & scipy corrcoef

David Cournapeau david@ar.media.kyoto-u.ac...
Mon Oct 15 23:43:08 CDT 2007


Pierre GM wrote:
>> So why not have the scipy methods just call down back the numpy
>> methods (or do some import magic in the scipy.stats module to just
>> import the numpy funcs into the namespace)?
>>     
>
> AAMOF, corrcoeff is not the only one that behaves inconsistently from one 
> version to another:
> scipy.var and numpy.var both use biased estimates of the variance, when 
> scipy.stats.var uses the unbiased estimates... Which one should take 
> precedence ? Would it be worth to add a "biased" keyword to the numpy.method 
> (that would default to True, so that we'd keep the current behavior) ?
>   
I already complained about this issue, and let it slipped. Actually, 
instead of complaining, I should have submitted a patch since then.

http://scipy.org/scipy/scipy/ticket/425

Having inconsistencies for basic things like mean, std is bad. Maybe we 
should have a list of those inconsistencies, to keep track of them 
(maybe by creating a numpy/scipy inconsistencies roadmap on scipy or 
numpy trac ?).

For corrcoeff, the problem was that the numpy version cannot be changed 
for backward compatibility: a new, fixed should be implemented anyway, 
but should have a different name (at least, that's what I remember from 
the discussion which happened at that time).

cheers,

David


More information about the SciPy-user mailing list