Comparisons result in different dtypes for empty DataFrames

[jcrist]

The comparison methods (lt, gt, etc...) return incorrect dtypes for empty dataframes. Interestingly, using the operators instead results in correct dtypes. Correct dtypes are also returned for empty series.

In [1]: import pandas as pd

In [2]: df = pd.DataFrame({'x': [1, 2, 3], 'y': [1., 2., 3.]})

In [3]: empty = df.iloc[:0]

In [4]: df.lt(2).dtypes
Out[4]:
x    bool
y    bool
dtype: object

In [5]: empty.lt(2).dtypes   # Should be all bool, but isn't
Out[5]:
x      int64
y    float64
dtype: object

In [6]: (df < 2).dtypes
Out[6]:
x    bool
y    bool
dtype: object

In [7]: (empty < 2).dtypes   # Things do work if you use the operator though
Out[7]:
x    bool
y    bool
dtype: object

In [8]: df.x.lt(2).dtype
Out[8]: dtype('bool')

In [9]: empty.x.lt(2).dtype    # Correct dtype for empty series
Out[9]: dtype('bool')

In [10]: pd.__version__
Out[10]: '0.19.2'

[mralgos]

@jcrist If the dataframe is empty, the comparison methods against a constant do not perform any operation. They simply return the input dataframe.

To fix the problem it seems to be enough to remove the first if statement in _combine_const function in pandas/core/frame.py. I'm running the tests.