Idea: use df.index/df.columns names to automatically choose axis along which to broadcast

[supern8ent]

In writing some math code in pandas, I find it necessary to do things like

df2 = df.sub(ser, axis='columns')

instead of the shorter and more intuitive

df2 = df - ser

in order to control the axis along which the series is broadcast.

I think it would be a big improvement syntactically if pandas would automatically broadcast down the axis that didn't have a matching name.

Example:

df = pd.DataFrame(np.random.rand(3, 2), columns=['a', 'b'])
df.index.name = 'dim0'
df.columns.name = 'dim1'
df
dim1         a         b
dim0                    
0     0.755744  0.321682
1     0.915464  0.413154
2     0.647672  0.457927

subtract:

df - df['a']    # does not give the desired result
       a   b   0   1   2
dim0                    
0    NaN NaN NaN NaN NaN
1    NaN NaN NaN NaN NaN
2    NaN NaN NaN NaN NaN

subtract, specifying which axis to match on (broadcasting happens on the other axis):

df.sub(df['a'], axis='index')   # gives the desired result
dim1    a         b
dim0               
0     0.0 -0.434062
1     0.0 -0.502310
2     0.0 -0.189745

I am suggesting that the "-" operator would look at the names of the indices in the operands and match on the axis that has the same name in the two operands.

By way of motivation, I'm doing mass spectral matching of compounds, so I could name my indices 'chemical' and 'mass'.

[max-sixty]

The concept resonates. It's exactly how xarray works - check that out if you want labelled dimensions fully supported.

For this to work in pandas, my POV is named indexes / dimensions need to be supported throughout operations. If we only implemented it for this case, it could feel a bit too magical.

ref :#11373 & #4036

[jreback]

xref to #8162 .

so going to rerpose that as a master issue instead of having all of the issues we will never find open.