Documentation issue for sum on sparse matrices

EDIT: The original issue posted below is in fact working as intended. See first comment for remaining documentation issue.

There is a strange behavior when summing sparse arrays along a specific axis. The matrix dimensions and slicing of the result do not work as expected. The behavior contradicts the documentation (which mentions the axis will be dropped) and also does not correspond to behavior as seen in numpy.

Reproducing code example:

from scipy.sparse import csr_matrix
csr_matrix((2,3)).sum(axis=0)
# matrix([[0., 0., 0.]])
csr_matrix((2,3)).sum(axis=0)[0]
# matrix([[0., 0., 0.]])
csr_matrix((2,3)).sum(axis=0)[0][0][0][0][0][0]
# matrix([[0., 0., 0.]])
csr_matrix((2,3)).sum(axis=1)
# matrix([[0.],[0.]])
csr_matrix((2,3)).sum(axis=1)[0]
# matrix([[0.]])
csr_matrix((2,3)).sum(axis=1)[0][0][0][0]
# matrix([[0.]])

import numpy as np
np.zeros((2,3)).sum(axis=0)
# array([0., 0., 0.])
np.zeros((2,3)).sum(axis=1)
# array([0., 0.])

Scipy/Numpy/Python version information:

import sys, scipy, numpy
print(scipy.__version__, numpy.__version__, sys.version_info)
1.5.2 1.19.1 sys.version_info(major=3, minor=7, micro=8, releaselevel='final', serial=0)

Also tested on scipy 1.5.0 and 1.4.1

[kno10]

I find it irritating to get a matrix back, and usually want an array. Apparently numpy also by now suggests to prefer using ndarrays instead of matrix: https://numpy.org/doc/stable/reference/generated/numpy.matrix.html any plans on eventually migrating the sparse matrix classes to produce ndarrays instead of matrixes?

[stefanv]

Desired behavior implemented in #14822. Docstring for sum can still be improved.