[FIXED] Convert complex NumPy array into (n, 2)-array of real and imaginary parts

Issue

I have a complex-valued NumPy array that I'd like to convert into a contiguous NumPy array with real and imaginary parts separate.

This

import numpy

u = numpy.array([
    1.0 + 2.0j,
    2.0 + 4.0j,
    3.0 + 6.0j,
    4.0 + 8.0j
    ])

u2 = numpy.ascontiguousarray(numpy.vstack((u.real, u.imag)).T)

does the trick, but transposing, vstacking, and converting to a contiguous array is probably a step or two too much.

Is there a native NumPy function that does this for me?

Solution

None of the alternatives are native or save on reshape, transposes etc.

For example internally column_stack converts its inputs to 2d 'column' arrays. Effectively it is doing

In [1171]: np.concatenate((np.array(u.real,ndmin=2).T,np.array(u.imag,ndmin=2).T),axis=1)
Out[1171]: 
array([[ 1.,  2.],
       [ 2.,  4.],
       [ 3.,  6.],
       [ 4.,  8.]])

vstack passes its inputs through atleast_2d(m), making sure each is a 1 row 2d array. np.dstack uses atleast_3d(m).

A new function is np.stack

In [1174]: np.stack((u.real,u.imag),-1)
Out[1174]: 
array([[ 1.,  2.],
       [ 2.,  4.],
       [ 3.,  6.],
       [ 4.,  8.]])

It uses None indexing to correct dimensions for concatenation; effectively:

np.concatenate((u.real[:,None],u.imag[:,None]),axis=1)

All end up using np.concatenate; it and np.array are the only compiled joining functions.

Another trick is to use view

In [1179]: u.view('(2,)float')
Out[1179]: 
array([[ 1.,  2.],
       [ 2.,  4.],
       [ 3.,  6.],
       [ 4.,  8.]])

The complex values are saved as 2 adjacent floats. So the same databuffer can be viewed as pure floats, or with this view as a 2d array of floats. In contrast to the concatenate functions, there's no copying here.

Another test on the alternatives is to ask what happens when u is 2d or higher?

Answered By - hpaulj