BUG: Complex types are coerced to float when summing along level of MultiIndex

[elinoreroebber]

When trying to sum complex numbers along one level of a MultiIndex, I get a ComplexWarning and the resulting datatypes are float64 with imaginary components discarded. Summing without specifying the level or by unstacking the MultiIndex works as expected.

Based on the error message, I assume that groupby is being called internally, and that the bug is actually there. I'm using pandas version 17.1.

Code Sample

In [1]: df = pd.DataFrame(data=np.arange(4)*(1+2j), index=pd.MultiIndex.from_product(([0,1], [0,1]), names=['a','b']))
In [2]: print df
          0
a b        
0 0      0j
  1  (1+2j)
1 0  (2+4j)
  1  (3+6j)

In [3]: df.sum(level='a')
/Users/elinore/Library/Python/2.7/lib/python/site-packages/pandas/core/groupby.py:1629: ComplexWarning: Casting complex values to real discards the imaginary part
  values = _algos.ensure_float64(values)
Out[3]: 
   0
a   
0  1
1  5

Expected Output

Out[3]:
       0
a   
0  (1+2j)
1  (5+10j)

output of pd.show_versions()

INSTALLED VERSIONS

commit: None
python: 2.7.11.final.0
python-bits: 64
OS: Darwin
OS-release: 14.5.0
machine: x86_64
processor: i386
byteorder: little
LC_ALL: None
LANG: en_CA.UTF-8

pandas: 0.17.1
nose: 1.3.7
pip: 7.1.2
setuptools: 19.2
Cython: 0.23.5
numpy: 1.10.4
scipy: 0.17.0
statsmodels: None
IPython: 4.1.2
sphinx: 1.4
patsy: None
dateutil: 2.5.1
pytz: 2016.3
blosc: None
bottleneck: 1.0.0
tables: 3.2.2
numexpr: 2.5.1
matplotlib: 1.5.1
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
lxml: None
bs4: None
html5lib: None
httplib2: None
apiclient: None
sqlalchemy: None
pymysql: None
psycopg2: None
Jinja2: None

[jreback]

This is about here

pull-requests are welcome. For now we need to coerce this to object (rather than treat this as a float)

This does it. just need some tests and such.

diff --git a/pandas/core/groupby.py b/pandas/core/groupby.py
index 6996254..e2a4482 100644
--- a/pandas/core/groupby.py
+++ b/pandas/core/groupby.py
@@ -1747,7 +1747,7 @@ class BaseGrouper(object):
             values = _algos.ensure_float64(values)
         elif com.is_integer_dtype(values):
             values = values.astype('int64', copy=False)
-        elif is_numeric:
+        elif is_numeric and not com.is_complex_dtype(values):
             values = _algos.ensure_float64(values)
         else:
             values = values.astype(object)

[a-p-man]

The bug also occurs when using a Series and invoking groupby with level=0 (so it is not specific to MultiIndex, only indirectly via groupby, as suggested).

In [1]: series = pd.Series(data=np.arange(4)*(1+2j),index=[0,0,1,1])
In [2]: print series
0        0j
0    (1+2j)
1    (2+4j)
1    (3+6j)
dtype: complex128

In [3]: series.groupby(level=0).sum()
Out[3]: 
0    1.0
1    5.0
dtype: float64

Expected:

Out[3]: 
0     (1+2j)
1    (5+10j)
dtype: complex128