BUG: "RuntimeWarning: divide by zero encountered in log10" that does not appear with a `numpy` argument

[bersbersbers]

Pandas version checks

• I have checked that this issue has not already been reported.

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

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

Reproducible Example

import numpy as np, pandas as pd
x = np.array([0.0, 1.0])
df = pd.DataFrame(x, columns=["x"])
np.log10(x, out=np.zeros_like(x), where=x>0)     # no warning, expected result ([0, 0])
np.log10(df.x, where=x>0)                        # no warning, usually incorrect result
np.log10(df.x, out=np.zeros_like(x), where=x>0)  # warning, correct result

Issue Description

Warning:

Python\Python310\lib\site-packages\pandas\core\arraylike.py:397: RuntimeWarning: divide by zero encountered in log10
result = getattr(ufunc, method)(*inputs, **kwargs)

Expected Behavior

No warning

Installed Versions

(show_versions() fails with assert '_distutils' in core.file, core.file)

Pandas 1.4.1 from pip and pandas-1.5.0.dev0+583.g5414a2d59 from GitHub

[adamjliang]

take

[rhshadrach]

The warning is being generated by NumPy, not pandas. That said, for the penultimate line you say # no warning, usually incorrect result, can you explain what you mean here?

[bersbersbers]

for the penultimate line you say # no warning, usually incorrect result, can you explain what you mean here?

Sure: the second operation writes into an uninitialized array. Running it immediately after the first operation gave correct results in many of my tests, but repeating it then gave incorrect ones (as expected).

[rhshadrach]

but repeating it then gave incorrect ones (as expected).

Are you able to share a reproducible example along with the output you are stating is incorrect?

[bersbersbers]

The warning is being generated by NumPy, not pandas.

Correct - but it is somewhat specific to pandas in that the same does not happen for lists or tuples, while it does happen for pandas series and data frames:

import numpy as np, pandas as pd
np.seterr(all="warn")

def fun(var):
    return np.log10(var, out=np.ones_like(var), where=np.array(var) > 0)

x = [0.0, 1.0]

fun(x)                               # correct result: [1., 0.]
fun(tuple(x))                        # correct result: [1., 0.]
fun(np.array(x))                     # correct result: [1., 0.]
fun(pd.Series(x))                    # correct result, unexpected warning
fun(pd.DataFrame(x, columns=["x"]))  # correct result, unexpected warning

---

Are you able to share a reproducible example along with the output you are stating is incorrect?

No. Due to the nature of uninitialized variables, it is impossible for me to ensure that the uninitialized memory is not zero for some reason or another, especially when it's reusing memory that was zeroed out before.

import numpy as np
x = np.array([0.0, 1.0])
np.log10(x, where=x>0)
np.log10(x, where=x>0)
np.log10(x, where=x>0)
np.log10(x, where=x>0)

gives

>>> import numpy as np
>>> x = np.array([0.0, 1.0])
>>> np.log10(x, where=x>0)
array([2.99804668e+122, 0.00000000e+000])
>>> np.log10(x, where=x>0)
array([2.99804668e+122, 0.00000000e+000])
>>> np.log10(x, where=x>0)
array([2.99804668e+122, 0.00000000e+000])
>>> np.log10(x, where=x>0)
array([2.99804668e+122, 0.00000000e+000])

now, but it can be any other value in later runs. BUT, this behavior

1. is not the issue I am reporting.
2. is completely independent of pandas.
3. is expected and documented in NumPy (https://numpy.org/doc/stable/reference/generated/numpy.log10.html):

   Note that if an uninitialized out array is created via the default out=None, locations within it where the condition is False will remain uninitialized.

4. was posted only to show that when I omit the out parameter, then the warning goes away, which made me think whether pandas/NumPy apply np.log10 to the initialized out variable for some reason (which they should not).
   This hypothesis is falsified by the warning also appearing in this example, so I updated first code block in this comment:

   import numpy as np, pandas as pd
   x = np.array([0.0, 1.0])
   np.log10(pd.Series(x), where=x>0, out=np.ones_like(x))  # same warning

I hope that makes my original post clearer. Note that warnings are shown only once per session unless you do np.seterr(all="warn").

[bersbersbers]

Workaround:

import numpy as np, pandas as pd
x = np.array([0.0, 1.0])
np.log10(pd.Series(x).values, where=x>0, out=np.ones_like(x))  # no warning