Sum/product of empty object array is False/True

[shoyer]

This strange behavior came up in pandas-dev/pandas#19813:

In [23]: np.array([], dtype=object).sum()
Out[23]: False

In [26]: np.array([], dtype=object).prod()
Out[26]: True

It's almost as if NumPy picks a dtype at random (e.g., bool in this case) to use for computing the result. A more obviously and likely more consistent choice would be to use the identities for the appropriate ufuncs (i.e., np.add.identity and np.mul.identity), which are the integers 0 and 1, respectively.

[eric-wieser]

Duplicate of #8860

This is fixed by #8955, and isbecause:

>>> np.add.identity
False
>>> np.multiply.identity
True

Prior to #8952, these would report 0 and 1, but the results of sum and prod would still be booleans.

This is fixed in #8952 (master), which now causes your example to return 0 and 1.

However, np.logical_or.reduce(np.array([], object)) now returns 0 instead of False, which is less good.
This would be fixed by #8955

[seberg]

Should it be the long term goal to get rid of identies for object arrays assuming we have an identity/default kwarg?

[TomAugspurger]

What's the status here? Do we think np.logical_or.reduce(np.array([], object)) will change, and if so, will that "break" the "fix" for the original issue?

[seberg]

Well, we have an initial kwarg now (next release). So you will be able to override the default in the future. Note that there are some limitations to this, in the sense that the initial has to be a valid initial value for non-empty reductions and is not a default that is ignored for empty ones.

With a good reason, pretty much only useful for object dtype (plausibly NaN can sometimes make sense in this regard, too) we could add an additional default argument.

I do not think we currently plan on a deprecation to force initial to be given for object dtypes or changing the default itself to 0 and 1, which would probably give ugly FutureWarnings though.

[WarrenWeckesser]

The examples shown above are all working now:

In [1]: import numpy as np

In [2]: np.__version__
Out[2]: '1.22.0.dev0+1504.gcfb981013'

In [3]: np.array([], dtype=object).sum()
Out[3]: 0

In [4]: np.array([], dtype=object).prod()
Out[4]: 1

In [5]: np.logical_or.reduce(np.array([], object))
Out[5]: False

I'm closing the issue, but if there are still problems lurking here, we can reopen or start a new issue.