Methinks this is a bit off :-)  True should be False.


Python 2.5b1 (trunk:47059, Jun 29 2006, 14:26:46)
[GCC 4.1.0 (SUSE Linux)] on linux2
>>> import decimal
>>> d = decimal.Decimal
>>> inf = d("inf")
>>> nan = d("nan")
>>> nan > inf
True
>>> nan < inf
False
>>> inf > nan
True
>>> inf < nan
False
b

Logged In: YES 
user_id=42444

It's still there in Beta 2.

Logged In: YES 
user_id=1581732

More specifically, any comparison with a NaN should equal
False, even inf, per IEEE 754.  A good starting point to
convince oneself of this is http://en.wikipedia.org/wiki/NaN.

The Decimal Arithmetic Specification says that NaN comparisons should return NaN.  The decimal module correctly implements this through the compare() method:

    >>> nan.compare(nan)
    Decimal('NaN')

Since python's < and > operators return a boolean result, the standard is silent on what should be done.  The current implementation uses the __cmp__ method which can only return -1, 0, or 1, so there is not a direct way to make both < and > both return False.

If you want to go beyond the standard and have both < and > return False for all NaN comparisons, then the __cmp__ implementation would need to be replaced with rich comparisons.  I'm not sure that this is desirable.  IMO, that would be no better than the current arbitrary choice where all comparisons involving NaN report self > other.  If someone has an application that would be harmed by the current implementation, then it should almost certainly be use the standard compliant compare() method instead of the boolean < and > operators.

Tim, what say you?

It seems likely that Decimal is going to grow rich comparisons anyway, 
either as a result of PEP 3141 or as a result of removal of __cmp__.

Given this, would there be any objections to making comparisons do the 
right thing (in the sense of IEEE-754) for NaNs?

See issue #1979 for a possible fix.

Fixed for Python 2.6 in r60630.