BUG: Subtracting two series with unordered index and all-nan index produces unexpected result

[ssche]

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

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

• (optional) I have confirmed this bug exists on the master branch of pandas.

---

Code Sample, a copy-pastable example

>>> import pandas as pd
>>> import numpy as np
>>> a_index = pd.MultiIndex.from_tuples([
...   (81.0,  np.nan, '2018-06-01'),
...   (81.0,  np.nan, '2018-07-01'),
...   (82.0,  np.nan, '2018-07-01'),
...   (82.0,  np.nan, '2018-08-01'),
...   (np.nan,np.nan, np.nan)],
...    names=['id', 'sub_ix', 'data']
... )
>>> a_values = [25, 22, 20, 21, np.nan]
>>> b_index = pd.MultiIndex.from_tuples([
...   (81.0, np.nan,  '2018-06-01'),
...   (np.nan, np.nan, np.nan),
...   (81.0, np.nan,  '2018-07-01'),
...   (82.0, np.nan,  '2018-07-01'),
...   (82.0, np.nan,  '2018-08-01')],
...   names=['id', 'sub_ix', 'data']
... )
>>> b_values = [28.28, np.nan, 28.28, 25.25, 25.25]
>>> a = pd.Series(a_values, index=a_index)
>>> b = pd.Series(b_values, index=b_index)


>>> a
id    sub_ix  data      
81.0  NaN     2018-06-01    25.0
              2018-07-01    22.0
82.0  NaN     2018-07-01    20.0
              2018-08-01    21.0
NaN   NaN     NaN            NaN
dtype: float64


>>> b
id    sub_ix  data      
81.0  NaN     2018-06-01    28.28
NaN   NaN     NaN             NaN
81.0  NaN     2018-07-01    28.28
82.0  NaN     2018-07-01    25.25
              2018-08-01    25.25
dtype: float64



>>> a - b
id    sub_ix  data      
81.0  NaN     2018-06-01   -3.28
              2018-07-01     NaN <-- this shouldn't be NaN, the index (81.0, NaN, 2018-07-01) exists in both `a` and `b` (it's just not ordered in `b`)
82.0  NaN     2018-07-01   -8.28 <-- also wrong, expected -5.25
              2018-08-01   -4.25
NaN   NaN     NaN            NaN
dtype: float64
>>> a - b.sort_index()
id    sub_ix  data      
81.0  NaN     2018-06-01   -3.28
              2018-07-01   -6.28 <-- expected value
82.0  NaN     2018-07-01   -5.25
              2018-08-01   -4.25
NaN   NaN     NaN            NaN
dtype: float64

Problem description

When combining two series with both the same index and with an all-nan index row at different positions, the result of the arithmetic operation (+, -, /) is not as expected. The issue can be worked around by sorting both indices (series.sort_index). I tried a different example with unordered indices, but without the all-nan index row and the result is as expected (so it's not an issue of the unsorted indices).

id    sub_ix  data      
81.0  NaN     2018-07-01    25
              2018-06-01    22
82.0  NaN     2018-07-01    20
              2018-08-01    21
dtype: int64
>>> b
id    sub_ix  data      
81.0  NaN     2018-06-01    1
              2018-07-01    2
82.0  NaN     2018-07-01    3
              2018-08-01    4
dtype: int64
>>> a - b
id    sub_ix  data      
81.0  NaN     2018-06-01    21
              2018-07-01    23
82.0  NaN     2018-07-01    17
              2018-08-01    17
dtype: int64

Expected Output

Operands should be aligned as per index (despite all nan-rows in the index).

Output of pd.show_versions()

INSTALLED VERSIONS

commit : 65f0463
python : 3.8.6.final.0
python-bits : 64
OS : Linux
OS-release : 5.9.11-200.fc33.x86_64
Version : #1 SMP Tue Nov 24 18:18:01 UTC 2020
machine : x86_64
processor : x86_64
byteorder : little
LC_ALL : None
LANG : en_AU.UTF-8
LOCALE : en_AU.UTF-8

pandas : 1.2.0.dev0+1441.g65f0463d3
numpy : 1.19.4
pytz : 2020.4
dateutil : 2.8.1
pip : 20.2.4
setuptools : 50.3.2
Cython : 0.29.21
pytest : 5.1.1
hypothesis : None
sphinx : None
blosc : None
feather : None
xlsxwriter : 0.9.6
lxml.etree : None
html5lib : None
pymysql : None
psycopg2 : 2.8.3 (dt dec pq3 ext lo64)
jinja2 : 2.11.2
IPython : None
pandas_datareader: None
bs4 : None
bottleneck : 1.3.1
fsspec : None
fastparquet : None
gcsfs : None
matplotlib : None
numexpr : 2.7.1
odfpy : None
openpyxl : 1.8.6
pandas_gbq : None
pyarrow : 1.0.1
pyxlsb : None
s3fs : None
scipy : 1.4.1
sqlalchemy : 1.3.12
tables : 3.6.1
tabulate : None
xarray : None
xlrd : 1.2.0
xlwt : None
numba : None

[simonjayhawkins]

Thanks @ssche for the report. Tested with 0.25.3 and also giving same output as master.

[phofl]

This is a bug in MultiIndex.equals, which may hide a Bug in Series.align.
The following behaves differently:

a_index = pd.MultiIndex.from_tuples([(81.0,  np.nan), (np.nan,np.nan)])
b_index = pd.MultiIndex.from_tuples([(np.nan, np.nan),(82.0, np.nan)])
ser = Series([1, 2], index=a_index)
ser1 = Series([1, 2], index=b_index)
ser.align(ser1)

returns

(81.0  NaN    1
NaN   NaN    2
      NaN    2
82.0  NaN    2
dtype: int64, 81.0  NaN    1
NaN   NaN    1
      NaN    1
82.0  NaN    2
dtype: int64)

while

a_index = pd.MultiIndex.from_tuples([(81.0,  np.nan), (np.nan,np.nan)])
b_index = pd.MultiIndex.from_tuples([(np.nan, np.nan),(81.0, np.nan)])
ser = Series([1, 2], index=a_index)
ser1 = Series([1, 2], index=b_index)
ser.align(ser1)

returns

(81.0  NaN    1
NaN   NaN    2
dtype: int64, NaN   NaN    1
81.0  NaN    2
dtype: int64)

e.g. not duplicating the NaN rows. I think the second one is the correct one?

[mroeschke]

Seems like the above example is consistent now? Could use a test possibly

In [6]: a_index = pd.MultiIndex.from_tuples([(81.0,  np.nan), (np.nan,np.nan)])
   ...: b_index = pd.MultiIndex.from_tuples([(np.nan, np.nan),(82.0, np.nan)])
   ...: ser = Series([1, 2], index=a_index)
   ...: ser1 = Series([1, 2], index=b_index)
   ...: ser.align(ser1)
Out[6]:
(81.0  NaN    1
 NaN   NaN    2
 dtype: int64,
 81.0  NaN    1
 NaN   NaN    1
 dtype: int64)

In [7]: a_index = pd.MultiIndex.from_tuples([(81.0,  np.nan), (np.nan,np.nan)])
   ...: b_index = pd.MultiIndex.from_tuples([(np.nan, np.nan),(81.0, np.nan)])
   ...: ser = Series([1, 2], index=a_index)
   ...: ser1 = Series([1, 2], index=b_index)
   ...: ser.align(ser1)
Out[7]:
(81.0  NaN    1
 NaN   NaN    2
 dtype: int64,
 81.0  NaN    2
 NaN   NaN    1
 dtype: int64)

[trevorkask]

take

[trevorkask]

I have tried to test the code and the assertion has no errors

In[1]
import pandas as pd
import numpy as np
from pandas import Series, DataFrame
from pandas.testing import assert_series_equal
a_index = pd.MultiIndex.from_tuples([(81.0,  np.nan), (np.nan,np.nan)])
b_index = pd.MultiIndex.from_tuples([(np.nan, np.nan),(82.0, np.nan)])
ser = Series([1, 2], index=a_index)
ser1 = Series([1, 2], index=b_index)
ser.align(ser1)

a_index = pd.MultiIndex.from_tuples([(81.0,  np.nan), (np.nan,np.nan)])
b_index = pd.MultiIndex.from_tuples([(np.nan, np.nan),(81.0, np.nan)])
ser = Series([1, 2], index=a_index)
ser1 = Series([1, 2], index=b_index)
ser.align(ser1)

a = pd.Series([1, 2, 3, 4])
b = pd.Series([1, 2, 3, 4])
diff_index=a_index
diff = Series([0,0], index=diff_index)

print(assert_series_equal(ser-ser1,diff))

Out[2]: None

[ssche]

@trevorkask Great news and thanks for sharing your results. I think the intent is now to create a test case (your test case above) to lock in behaviour in Pandas. Do you need assistance?

[trevorkask]

Yes, please