[Numpy-discussion] ANN: NumPy 1.2.0

[Pauli Virtanen]

Mon, 29 Sep 2008 17:19:10 +0900, David Cournapeau wrote:

> On Mon, 2008-09-29 at 10:00 +0200, Matthieu Brucher wrote:
>> I've tested the latest Numpy on a RHEL 4, and I got this result :
>> 
>> 
> Hi Matthieu,
> 
> 	Are you on 32 or 64 bits ? As you can see, those errors are test
> errors, not errors in the function themselves (tolerance too low; maybe
> influenced by compilers - maybe gcc 3 ? - , etc...). Any way, those seem
> harmless,

I think the errors may imply that with your compiler, numpy gives

>>> np.log(-1 + 0j)
-3.1415926535897931j

or something similar "wrong" behavior at the branch cut of the logarithm, 
instead of

>>> np.log(-1 + 0j)
3.1415926535897931j
>>> np.log(-1 - 1e-99j)
-3.1415926535897931j
>>> np.log(-1 + 1e-99j)
3.1415926535897931j

I'd guess this is typically harmless, but may lead to problems if your 
code relies on the choice of branch at the branch cut.

Could you check if this is really the case? If log seems to work OK, then 
it's a problem with the test and not the functions.

The current implementation of complex log seemed to work on other 
platforms, but maybe icc has some features that break it in this respect.

-- 
Pauli Virtanen