This is not the NaN you are looking for

Alternate title: "NaN - More than Not a Number"

• Python oddities concerning NaN
• Live demo a dict that allows use of NaN as a key the way it would be intuitively, so that all NaN values are treated as one.
• Live demo a Counter that counts how many occurences of NaN exist

Questions

• What about copy.copy?

Description

Have you ever wondered about float('NaN')? What it is? What you can do with it? What you shouldn't do with it? Let's take some time and explore this Not a Number and what it does (and doesn't do) in Python.

Outline

• Introduction. 3 minutes
  • Who am I?
  • What am I going to talk about?
  • What led me to write this talk?
• Introducing NaN. 3 minutes
  • What is NaN?
  • What purpose does NaN serve?
  • The differences beteween float('nan'), numpy.nan, and math.nan.
  • How we can tell if a "numeric" value is NaN.
• What makes NaN so unique? 5 minutes
  • No 2 instances of NaN are the same.
  • All instances of NaN have the same repr: "nan"
  • NaN is not equal to anything, not even itself.
  • NaN can be hashed but each instance of NaN will have a different hash.
    • A set can contain multiple instances of NaN.
    • A dictionary can have multiple NaN keys, but good luck accessing the data.
    • If you have an Iterable with multiple instances of NaN and create a Counter from it, each instance will appear with a count of 1.
• Creating a Pythonic dict that can only have one NaN. 12 minutes
  • Do a code walkthrough and live demo of a Python subclass of dict that can only have one instance of NaN.
    • Show how I created the new version of dict.
    • Simple live demo showing that we can use float('nan') as a dict key. In a regular dict it will give a KeyError when attempting to retrieve it; with the new version of dict it will work as expected.
  • Do a code walkthrough and live demo of a Python Counter class that will count how many instances of NaN are in it.
    • Show how I created a Counter class that will count all instances of NaN.
    • Live demo showing the behaviour of collections.Counter and the new Counter class.
• Conclusion. 2 minutes
  • Quick recap of what was discussed.
  • Where to find the talk slides.
  • Where to find me.

Things I've found

>>> float('nan') ** 0
1.0
>>> float('nan') > 1
False
>>> float('nan') < 1
False
>>> float('nan') == 1
False
>>> float('nan') != 1
True
>>>> float('inf') - float('inf')
nan
>>>> f'ba{1e234567 - 1e765432}as'
'bananas'

import struct
float_bytes = b"\xaa\xbb\xcc\xdd\xff\xff\xff\xff"
val = struct.unpack("d", float_bytes)[0]
print(val)
print(struct.pack("d", val))

In a single precision floating point number (32 bit) there are 2^24 - 2 potential NaN values.

Potential improvements for NaNDict:

• Different types of nans are distinct.
• Perserve the original supplied nan (maybe you do and check_key() does the coalescing).
• A dunder hook for hypothetical future other weird ass values with rogue behaviour like nan.
• An extensive test and performance suite.
• Accept any kind of object for the membership test.
• Accepting unhashable values.

json parsing in Python doesn't handle NaN properly. pydis

JSON does not represent IEEE-754 floats. It does not permit NaN and infinity as part of the "number" type.

YAML does support NaN

Signed NaN and things with NaN that don't return NaN Jeff Epler post

From Jeff Epler - NaN Cipher

[math.nan] == [math.nan] is True. Identity check on the objects before equality check

Prior to Python 3.11 CPython did not require IEEE-754 support and included a Py_NO_NAN macro. Current versions of CPython require IEEE-754 support and the macro was removed.

Quotes

NaN is a lot like infinity in that people want to treat it as a number but It's Not and trying to do that will only lead to surprise and disappointment. - Luna Celeste

NaN is a wild land governed by laws both alien and arcane - Luna Celeste

References

https://xkcd.com/851/

https://brassnet.biz/blog/nan-is-weird.html

https://github.com/python/cpython/issues/87641

https://twit.social/@brass75/116168712645625711

https://discuss.python.org/t/question-about-float-nan/106378

https://en.wikipedia.org/wiki/NaN