[SciPy-user] Re: nan puzzle

[Grant Edwards]

On 2005-07-14, Jim Vickroy <Jim.Vickroy at noaa.gov> wrote:

> So without knowing the implementation of nan (e.g., is it
> truly unique package-wide),

I don't understand how there can even be "an implimentation of
nan" in this context.  In IEEE floating point a NaN is any of a
large set of bit patterns.  Assuming a 64-bit IEEE float (which
is what all the Python implimentations I know of use), there
are 16 million unique bit patterns that are NaNs. Half of them
are signalling NaNs, half are quiet NaNs. Half of them are
positive, half of them are negative.Though the sign of a NaN
is, in practice, meaningless, the sign bit can be either
postive or negative.

There is no unique NaN FP value. Since there is no unique
_value_ it's rather pointless to talk about a module-wide
implimentation of a NaN value. It's like talking about an
object that contains "the negative number", and wondering
whether that implimentation of "negative" is module-wide or
not.

If you want to know if a name is bound to a floating point
object object that has one of the NaN bit-patterns, then you've
got to look at the bit pattern.

It's handy to have a name bound to one of those bit patterns
for when you want to use one, but having a name bound to one of
those bit patterns is uselss when you want to decide if some
other FP object contains a NaN.

> I would not expect x is nan to evalutate to True.

-- 
Grant Edwards                   grante             Yow!  Now KEN and BARBIE
                                  at               are PERMANENTLY ADDICTED to
                               visi.com            MIND-ALTERING DRUGS...