Bug: Fill values 0 and -0 seen as different

[aulemahal]

sparse version checks

• I checked that this issue has not been reported before list of issues.

• I have confirmed this bug exists on the latest version of sparse.

• I have confirmed this bug exists on the main branch of sparse.

Describe the bug

I can't concatenate two arrays that have fill values of 0.0 and -0.0 as the values are seen as different, inconsistent.

Doesn't happen on the last release (0.15.4) but it does when I install the main branch.

Steps or code to reproduce the bug

import numpy as np
import sparse as sp

a = sp.COO(np.array([[0, 3]]), [1.0, 4.0], (4,))
b = sp.COO(np.array([[0, 3]]), [6.0, 8.0], (4,))

print(a)
# <COO: shape=(4,), dtype=float64, nnz=2, fill_value=0.0>

print(-b)
# <COO: shape=(4,), dtype=float64, nnz=2, fill_value=-0.0>

sp.concat([a, b])
# Works

sp.concat([a, -b])
# Fails

Expected results

<COO: shape=(8,), dtype=float64, nnz=4, fill_value=0>

Actual results

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[4], line 6
      4 a = sp.COO(np.array([[0, 3]]), [1.0, 4.0], (4,))
      5 b = sp.COO(np.array([[0, 3]]), [6.0, 8.0], (4,))
----> 6 sp.concat([a, -b])

File ~/miniforge3/envs/xesmf/lib/python3.12/site-packages/sparse/numba_backend/_common.py:1543, in concatenate(arrays, axis, compressed_axes)
   1540 if not builtins.all(isinstance(arr, GCXS) for arr in arrays):
   1541     from ._coo import concatenate as coo_concat
-> 1543     return coo_concat(arrays, axis)
   1545 from ._compressed import concatenate as gcxs_concat
   1547 return gcxs_concat(arrays, axis, compressed_axes)

File ~/miniforge3/envs/xesmf/lib/python3.12/site-packages/sparse/numba_backend/_coo/common.py:160, in concatenate(arrays, axis)
    134 """
    135 Concatenate the input arrays along the given dimension.
    136 
   (...)
    156 [`numpy.concatenate`][] : NumPy equivalent function
    157 """
    158 from .core import COO
--> 160 check_consistent_fill_value(arrays)
    162 if axis is None:
    163     axis = 0

File ~/miniforge3/envs/xesmf/lib/python3.12/site-packages/sparse/numba_backend/_utils.py:636, in check_consistent_fill_value(arrays)
    634 for i, arg in enumerate(arrays):
    635     if not equivalent(fv, arg.fill_value):
--> 636         raise ValueError(
    637             "This operation requires consistent fill-values, "
    638             f"but argument {i:d} had a fill value of {arg.fill_value!s}, which "
    639             f"is different from a fill_value of {fv!s} in the first "
    640             "argument."
    641         )

ValueError: This operation requires consistent fill-values, but argument 1 had a fill value of -0.0, which is different from a fill_value of 0.0 in the first argument.

Please describe your system.

1. Linux, Fedora 40, 6.11.5-200.fc40.x86_64
2. sparse version : 0.16.0a10.dev83+g440bd35
3. NumPy version : 2.0.2
4. Numba version : 0.60.0

Relevant log output

No response

[willow-ahrens]

These two values are indeed different though, right? Otherwise we would have 1/cat(A, B) != cat(1/A, 1/B)?

[aulemahal]

Ah, I didn't see it that way... Indeed, for that operation, numpy yields inf and -inf which are different.

Then I guess it's a new feature I simply didn't expect, and not a bug! Thanks.