`xp.vecdot` output arrayness depends on NumPy version

[mdhaber]

With NumPy >= 2.0:

import numpy as np
from scipy._lib._array_api import array_namespace
x1 = np.asarray([0.667, 0.667, 0.667])
x2 = np.asarray([0.123, 0.456, 0.789])
xp = array_namespace(x1, x2)
xp.vecdot(x1, x2)
# np.float64(0.912456)

With NumPy < 2.0, the result is a 0d array.

I'd suggest that array-api-compat should either:

• make modern NumPy produce an array (for compliance with the standard) or
• make old NumPy produce a NumPy float (for consistency with all NumPy functions).

The latter would be my preference for now. (Actually fixing NumPy would be the long term preference - mdhaber/numpy#2).

[ev-br]

This, I suppose, depends on whether vecdot is a reduction (then it should return np scalars, shudder), or a linear algebra function (then it should produce an array).
FWIW I think of it as a linalg function.

Would be great to have some clarity from numpy itself.

[rgommers]

FWIW I think of it as a linalg function.

Me too

[mdhaber]

Well, in case it helps.

import numpy as np
x = np.arange(10.)
A = np.eye((3))

x @ x  # scalar
np.vdot(x, x)  # scalar
np.inner(x, x)  # scalar
np.matmul(x, x)  # scalar
np.einsum('i,i', x, x)  # scalar
np.linalg.det(A)  # scalar
np.linalg.cond(A) # scalar
np.linalg.slogdet(A)  # scalar
np.linalg.matrix_rank(A)  # scalar
np.linalg.vector_norm(x)  # scalar
np.linalg.matrix_norm(A)  # scalar
np.linalg.trace(A)  # scalar

Which linalg functions did you have in mind? The only function I could find that returns a 0d array is tensordot.

[ev-br]

Ouch, my bad.
I did not realize that "linalg" functions return numpy scalars not 0D arrays, too. Falls out of them being (g)ufuncs, I guess. In that case I agree np.vecdot returning scalars is at least consistent with the rest of numpy and is not a numpy 2.x artefact.

In the compat layer, the numpy version dependence is then coming from

array-api-compat/array_api_compat/numpy/_aliases.py

Line 150 in 1609bf4

# (i.e., numpy 2.0), use the library version instead of our wrapper.

for numpy < 2 it drops through to the "generic" version,

array-api-compat/array_api_compat/common/_aliases.py

Line 554 in 1609bf4

def vecdot(x1: Array, x2: Array, /, xp: Namespace, *, axis: int = -1) -> Array:

which returns an array indeed.

While a fix is not too difficult (basically stick a [()] in the numpy/_aliases.py variant), how much of an annoyance it is? Given that scipy's dropping numpy 1.x for 1.18.x

[mdhaber]

With the new minimum version, it won't bother me anymore.