Quantity comparison to dimensionless zero has an amusing edge case

[ayshih]

Description

Comparing a quantity with units to dimensionless zero always works, presumably because for a given comparison it is reasonable to assume that dimensionless zero is the same as zero times a unit. However, this results in an amusing edge case when working with two units where the zero point does not coincide. For example, you can't normally compare degrees Celsius and Kelvin without the temperature() equivalency:

>>> -1*u.Celsius < 1*u.Kelvin
...
UnitConversionError: 'K' (temperature) and 'deg_C' (temperature) are not convertible
...
>>> -1*u.Celsius < (1*u.Kelvin).to(u.Celsius, equivalencies=u.temperature())
False

But, if you throw dimensionless zero into the mix, you seemingly can compare them without the equivalency (through transitivity) and then obtain the wrong answer.

>>> -1*u.Celsius < 0
True
>>> 1*u.Kelvin > 0
True
>>> -1*u.Celsius < 0 < 1*u.Kelvin
True

Expected behavior

Something not amusing, but this example is rather contrived

How to Reproduce

See description

Versions

Windows-10-10.0.22621-SP0
Python 3.10.10 | packaged by conda-forge | (main, Mar 24 2023, 20:00:38) [MSC v.1934 64 bit (AMD64)]
astropy 5.2.2
Numpy 1.24.3
pyerfa 2.0.0.3
Scipy 1.10.1
Matplotlib 3.7.1

[byrdie]

This seems particularly hard to solve since the chained comparison operators are just syntactic sugar for (a < b) and (b < c), which would also have the same amusing behavior.

Seems like the only way out of this is to disable comparisons with dimensionless zero, but that would probably break a lot of downstream code.

Maybe the rich comparison chaining in PEP 535 would help?

[ayshih]

Heh, the example in my original report was inspired by a live situation, but it's even more amusing since of course the units compared don't have to be related at all:

>>> 0*u.Celsius == 0*u.m
False
>>> 0*u.Celsius == 0 == 0*u.m
True

[pllim]

Maybe don't use == twice in the same statement? This is Python specific thing right? Not all languages even support such comparison.

[mhvk]

I think the comparison with degrees C should just fail, i.e., any unit with an offset should not be allowed to compare with 0 without units.
In principle, this is not super-hard - deg_C and deg_F should become instances of a new function unit class (say OffsetUnit), which could disable this. But it needs time to implement...

[ayshih]

Maybe don't use == twice in the same statement? This is Python specific thing right? Not all languages even support such comparison.

Well, as @byrdie noted, it's just syntactic sugar, until there is rich comparison chaining. As a stilted example, one could have a convenience function that implicitly leans on transitivity:

>>> def all_three_equal(a, b, c):
...     return (a == b) and (b == c)

And then argument order matters:

>>> all_three_equal(0*u.Celsius, 0, 0*u.m)
True
>>> all_three_equal(0*u.Celsius, 0*u.m, 0)
False

I think the comparison with degrees C should just fail, i.e., any unit with an offset should not be allowed to compare with 0 without units.

Sure, but that doesn't address the equality comparisons between unrelated units without an offset:

>>> 0*u.s == 0*u.m
False
>>> 0*u.s == 0 == 0*u.m
True

[pllim]

Do you think acknowledging this limitation in Known Issues would help?

[mhvk]

Surely doesn't hurt. In the end, the reason for allowing 0 to have any dimension was simply that many things just worked a little easier (including some numpy functions such as np.sinc -- though these days one can just override).

[pllim]

Okay, I proposed text for known issue at #15154 . Do we keep this open even after that, or close this as "won't fix" ?

[byrdie]

I personally think it would be best to deprecate comparison with dimensionless zero, but I'd be interested to hear what others think about it. I'm sure it would cause lots of problems downstream.

[mhvk]

See #1254 for why this was added. Personally, I'm not sure I would decide the same way now, but it would definitely seem to be a case that one breaks a lot of code just for being slightly more "correct" -- which I don't think is worth it.

[pllim]

comparison with dimensionless zero

But I like that feature!