Original ticket http://projects.scipy.org/numpy/ticket/1250 on 2009-10-06 by @huard, assigned to @pv.

For complex arguments, np.sign returns the following:

> [10]: np.sign(1j)
 <[10]: (1+0j)

> [11]: np.sign(-1j)
 <[11]: (-1+0j)

> [12]: np.sign(1-1j)
 <[12]: (1+0j)

So the sign of the real part takes precedence on the sign of the complex part when there is a real part. I'm not sure if this qualifies as a bug, but it's not intuitive, nor documented.