[Numpy-discussion] Bug in logaddexp2.reduce

[Charles R Harris]

On Wed, Mar 31, 2010 at 11:17 PM, Anne Archibald
<peridot.faceted at gmail.com>wrote:

> On 1 April 2010 01:11, Charles R Harris <charlesr.harris at gmail.com> wrote:
> >
> >
> > On Wed, Mar 31, 2010 at 11:03 PM, Anne Archibald <
> peridot.faceted at gmail.com>
> > wrote:
> >>
> >> On 31 March 2010 16:21, Charles R Harris <charlesr.harris at gmail.com>
> >> wrote:
> >> >
> >> >
> >> > On Wed, Mar 31, 2010 at 11:38 AM, T J <tjhnson at gmail.com> wrote:
> >> >>
> >> >> On Wed, Mar 31, 2010 at 10:30 AM, T J <tjhnson at gmail.com> wrote:
> >> >> > Hi,
> >> >> >
> >> >> > I'm getting some strange behavior with logaddexp2.reduce:
> >> >> >
> >> >> > from itertools import permutations
> >> >> > import numpy as np
> >> >> > x = np.array([-53.584962500721154, -1.5849625007211563,
> >> >> > -0.5849625007211563])
> >> >> > for p in permutations([0,1,2]):
> >> >> >    print p, np.logaddexp2.reduce(x[list(p)])
> >> >> >
> >> >> > Essentially, the result depends on the order of the array...and we
> >> >> > get
> >> >> > nans in the "bad" orders.  Likely, this also affects logaddexp.
> >> >> >
> >> >>
> >> >> Sorry, forgot version information:
> >> >>
> >> >> $ python -c "import numpy;print numpy.__version__"
> >> >> 1.5.0.dev8106
> >> >> __
> >> >
> >> > Looks like roundoff error.
> >>
> >> No. The whole point of logaddexp and logaddexp2 is that they function
> >> correctly - in particular, avoid unnecessary underflow and overflow -
> >> in this sort of situation. This is a genuine problem.
> >>
> >
> > Well, it did function correctly for me ;) Just some roundoff error.
> >
> >>
> >> I see it using the stock numpy in Ubuntu karmic:
> >> In [3]: np.logaddexp2(-0.5849625007211563, -53.584962500721154)
> >> Out[3]: nan
> >>
> >> In [4]: np.logaddexp2(-53.584962500721154, -0.5849625007211563)
> >> Out[4]: nan
> >>
> >> In [5]: np.log2(2**(-53.584962500721154) + 2**(-0.5849625007211563))
> >> Out[5]: -0.58496250072115608
> >>
> >> In [6]: numpy.__version__
> >> Out[6]: '1.3.0'
> >>
> >> In [8]: !uname -a
> >> Linux octopus 2.6.31-20-generic #58-Ubuntu SMP Fri Mar 12 05:23:09 UTC
> >> 2010 i686 GNU/Linux
> >>
> >> (the machine is a 32-bit Pentium M laptop.)
> >>
> >
> > I think this is related to 32 bits somehow. The code for the logaddexp
> and
> > logaddexp2 functions is almost identical. Maybe the compiler?
>
> Possibly, or possibly I just didn't find values that trigger the bug
> for logaddexp. I wonder if the problem is the handling of denormalized
> floats?
>
>
I'm suspecting the log2/exp2 function in the 32 bit gcc library or possibly
sse2. The exponents aren't large enough to denormalize the numbers.

Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20100331/84a00268/attachment.html>