In bpo-39288, I added math.nextafter(x, y) function. I propose to now add math.ulp() companion function.

Examples from tests of my PR:

self.assertEqual(math.ulp(1.0), sys.float_info.epsilon)
self.assertEqual(math.ulp(2.0 ** 52), 1.0)
self.assertEqual(math.ulp(2.0 ** 53), 2.0)
self.assertEqual(math.ulp(2.0 ** 64), 4096.0)

Unit in the last place:

• https://en.wikipedia.org/wiki/Unit_in_the_last_place
• Java provides a java.lang.Math.ulp(x) function: https://docs.oracle.com/javase/8/docs/api/java/lang/Math.html#ulp-double-

In numpy, I found two references to ULP:

• numpy.testing.assert_array_almost_equal_nulp:
  https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.testing.assert_array_almost_equal_nulp.html

• numpy.testing.assert_array_max_ulp:
  https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.testing.assert_array_max_ulp.html

Attached PR implements math.ulp(x).

See also math.isclose() and its PEP-485 which mentions ULP (Unit in the Last Place)... in the "Inappropriate uses" section :-)

• https://docs.python.org/dev/library/math.html#math.isclose
• https://www.python.org/dev/peps/pep-0485/#inappropriate-uses

Extract of an old python-ideas discussion on adding nextafter():

"It's also a little weird to jump from nextafter to isclose, since AFAIK
they have pretty much non-overlapping use cases..."

https://mail.python.org/pipermail/python-ideas/2017-February/044832.html

Other links:

• https://en.wikipedia.org/wiki/Unit_in_the_last_place
• https://en.wikipedia.org/wiki/Machine_epsilon
• https://en.wikipedia.org/wiki/IEEE_754

+1

Thank you Victor!

Any chance you could look at fma too? bpo-29282

(The reward for a job well done is to be given more jobs :-)

[Steven]

Any chance you could look at fma too? bpo-29282

fma is hard, for reasons explained in the issue you linked to. If you have suggestions for resolving the difficulties, please do add them to that issue.

New changeset 0b2ab21 by Victor Stinner in branch 'master':
bpo-39310: Add math.ulp(x) (GH-17965)
0b2ab21

New changeset 9362f85 by Victor Stinner in branch '3.8':
bpo-39310: Update sys.float_info documentation (GH-17982)
9362f85

New changeset dfe159c by Miss Islington (bot) in branch '3.7':
bpo-39310: Update sys.float_info documentation (GH-17982)
dfe159c

Can I just say that "ulp" is totally non-obvious what that even means unless you have a specific math background?

And sorry if that last response from me came off as grumpy; it wasn't meant to. I just had no clue what you were adding based on the name.

Can I just say that "ulp" is totally non-obvious what that even means unless you have a specific math background?

The math.ulp() documentation explicitly says: https://docs.python.org/dev/library/math.html#math.ulp

ULP stands for “Unit in the Last Place”.

The term "ulp" is commonly used when talking about IEEE 754 floating point numbers. It is used in numpy and Java for example.

test_math.py already had an ulp() function which was a pure Python implementation.

The term is also commonly used in math articles.

If you don't know the term "ulp", it's likely a *good thing*. You didn't have to suffer with rounding issues :-D

I hear what you are saying, but "ulp" is the standard term of art. Sure
it is non-obvious until you learn it, just like other technical terms
like "mro", "abc" or "ast".

Mathematics and numeric programming are rife with short names that are
non-obvious and often ambiguous with "ordinary" words, e.g.

sin, tan, log, nan

to mention just a few. "ulp" is a technical, and subtle, concept to
grasp, and no easier to understand when spelled out as "unit in last
place".

At least ulp is a TLA from English, unlike (say) "sine" which ultimately
derives from the Sanscrit word "jya" (chord), via Arabic and Latin. If
you've ever wondered if the trigonometric sin() function is related to
the sinus cavities in your nose, yes it is :-)

[Brett]

Can I just say that "ulp" is totally non-obvious what that even means unless you have a specific math background?

It's a good point. I guess we have a choice between using the domain-specific standard-ish name (which should be immediately meaningful to experts, but doesn't give much of a hint to non-experts) or using something more descriptive (which then risks confusing experts until they figure out "oh, that's just ulp").

There's also the option of spelling it out as "unit_in_last_place", but I'm not sure that benefits anyone.

For meaningful descriptive names, "float_resolution" or "gap_to_next" are about the best I can come up with. "precision" is too ambiguous.

NumPy has "numpy.spacing". But this exhibits exactly the trap of not using the "ulp" name: on a first glance, I incorrectly decided that NumPy didn't implement a ulp function. Then, having found numpy.spacing, I had to read the description carefully in order to recognise that "oh, this is just ulp". (Actually, I had to do more, since the description doesn't make all the corner cases clear, and in fact is currently wrong for powers of 2.)

If we can get people can coalesce around a preferred alternative name, we could consider changing this.

ulp() is the right name: universally understood by those who know how to use it, and easy to find exhaustive web explanations for those who don't.

In a different context, spelling out some variant of Hypertext_Transfer_Protocol would be as wrong-headed to avoid the "cryptic" http.