pandas, how to add columns to a multiindex column DataFrame

Question

This is my original DataFrame (with multiindex column):

In [72]:df
Out[72]: 
          a                   b          
          x         y         x         y
0  1.545293 -0.459270  0.899254 -1.010453
1  0.458760  0.275400 -0.190951  0.169195
2 -0.941817  1.109823  0.077953 -0.247074
3  1.790101 -1.643470  0.979625 -1.704657
4 -2.044814 -0.243726 -0.039724  0.600066

and I have another DataFrame:

In [77]:df2
Out[77]: 
          x         y
0 -1.085869 -0.952949
1  0.601585  0.570050
2  0.328601  0.802610
3 -0.415952 -0.090088
4  0.757545 -0.736933

how can I add df2's columns to df to get a new DataFrame like this:

In [83]:df3
Out[83]: 
          a                   b                   c          
          x         y         x         y         x         y
0  1.545293 -0.459270  0.899254 -1.010453 -1.085869 -0.952949
1  0.458760  0.275400 -0.190951  0.169195  0.601585  0.570050
2 -0.941817  1.109823  0.077953 -0.247074  0.328601  0.802610
3  1.790101 -1.643470  0.979625 -1.704657 -0.415952 -0.090088
4 -2.044814 -0.243726 -0.039724  0.600066  0.757545 -0.736933

My current approach is to use a for loop:

for col in df2.columns:
    df['c', col] = df2[col]

is there any method to avoid the loop?

Answer 1

Try pd.concat:

pieces = {'a' : df1['a'],
          'b' : df1['b'],
          'c' : df2}
df3 = pd.concat(pieces, axis=1)

Answer 2

I discovered another way to do this in the general case (running Python 3.6), without having to explicitly deconstruct the DataFrame. You can use pd.concat with the dictionary argument,

df3 = pd.concat({**df1, **{('c',nm):val for nm,val in df2.items()})

** expansion on DataFrame objects seems to return a dictionary of Series objects with "names" equal to the column name string/value, or if the columns are MultiIndexed, the tuple containing the hieararchy of column string/values. Then, when read back into pd.concat as a dictionary, Pandas re-constructs the MultiIndexed columns from the tuples.

Note this is much less efficient than the direct assignment you were doing! Since it has to deconstruct each column and MultiIndex of the dataframe, then re-combine.