[Numpy-discussion] Bug in logaddexp2.reduce

Charles R Harris charlesr.harris@gmail....
Thu Apr 1 22:01:36 CDT 2010


On Thu, Apr 1, 2010 at 8:16 PM, Charles R Harris
<charlesr.harris@gmail.com>wrote:

>
>
> On Thu, Apr 1, 2010 at 7:59 PM, David Cournapeau <david@silveregg.co.jp>wrote:
>
>> Anne Archibald wrote:
>>
>> >
>> > First I guess we should check which systems don't have log1p
>>
>> This is already done - we do use the system log1p on linux (but note
>> that log2_1p is not standard AFAIK). I would guess few systems outside
>> windows don't have log1p, given that msun has an implementation,
>>
>>
> I see that msun uses the same series I came to. However, my rational (Pade)
> approximation is a lot better than their polynomial.
>
>
In fact I get better than 119 bits using the same range as sun and the ratio
of two 7'th degree polynomials. I suspect it's better than that, but I only
have mpmath set to that precision. Sun got 58 bits with a 14'th degree
polynomial. So we can definitely improve on sun.

Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/numpy-discussion/attachments/20100401/8cbd20aa/attachment.html 


More information about the NumPy-Discussion mailing list