Need for signbit in the array API spec

The array specification for atan2 requires implementations to output -0.0 for some special values, e.g.

If x1_i is -0 and x2_i is +0, the result is -0.

which brings up a question on how implementors can distinguish between -0.0 and +0.0 in their test suite.

Equality testing spec mandates that these values be indistinguishable, and for a good reason.

The sign function also clamps these two values into +0.0.

I think this calls for the standard to have the signbit function:

numpy.signbit ref
cupy.signbit ref
torch.signbit ref
tf.experimental.numpy.signbit ref
dask.array.signbit ref
JAX does not seem to have signbit function, although it is implementable (example in one of JAX's issues)

Yeah it is a bit clunky right now. In array-api-tests we cast 0d array floats to builtin floats, before using math.copysign(1, your_builtin_float) == 1 to see whether its positively signed or not. Maybe it's not too big of a deal to add a function for it (or possibly modifying sign), don't have a sense for that.

There is a proposal for copysign at #593.

which brings up a question on how implementors can distinguish between -0.0 and +0.0 in their test suite.

For an implementer to test something, it isn't required to use only functions in the standard. This is checking the value of a bit, which can be done with a private function even.

Not saying that it's not a good idea to have this function - but especially given that JAX doesn't have it, it'd be nice to have a stronger reason.

I checked the SciPy code base. There's a lot of use of the C/C++ versions of signbit in compiled code, but only a couple in Python code:

scipy/special/tests/test_wrightomega.py
55:        assert_(np.signbit(res.imag) == np.bool_(False))

scipy/special/tests/test_trig.py
58:    assert np.signbit(y)
62:    assert not np.signbit(y)
66:    assert not np.signbit(y)

scipy/linalg/_matfuncs_expm.pyx.in
561:            S[:, t:] = np.signbit(Y)

Scikit-learn doesn't have any usage of signbit, and Pandas has four instances:

pandas/tests/indexes/test_numpy_compat.py
105:    "func", [np.isfinite, np.isinf, np.isnan, np.signbit], ids=lambda x: x.__name__
128:        is_complex_dtype(index) and func is np.signbit

pandas/tests/indexes/multi/test_analytics.py
253:    [np.isfinite, np.isinf, np.isnan, np.signbit],

pandas/tests/arithmetic/test_numeric.py
45:    if np.signbit(np.array(zero)).any():
48:        assert np.signbit(np.array(zero)).all()

pandas/core/ops/missing.py
111:        zneg_mask = zmask & np.signbit(y)

JAX does have signbit: https://jax.readthedocs.io/en/latest/_autosummary/jax.numpy.signbit.html

Okay, that is two votes for adding signbit and all libraries have it, so there are no blockers. The signature is as simple as it gets:

signbit(x: array, /) -> array

Let's add it?

I can create a PR later this week.

PR is up for review: gh-705

Navigation Menu

Repository navigation

Need for signbit in the array API spec #670

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Need for signbit in the array API spec #670

Description

Activity

honno commented on Aug 1, 2023

asmeurer commented on Aug 1, 2023

rgommers commented on Sep 6, 2023

jakevdp commented on Nov 13, 2023

rgommers commented on Nov 13, 2023

kgryte commented on Nov 13, 2023

rgommers commented on Nov 16, 2023

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions