Mark (>): > I think the most sensible thing is to be consistent. sgn() fails for > non-real input as long as sqrt() returns NaN for negative input. > Change the latter behavior (via a pragma or whatever) so that sqrt() > returns complex numbers, and then sgn() should start behaving on such > numbers.
I like that. Both sign() and sqrt() will then behave like they usually do, without complex surprises. But for those who want the generalized behvaiour, it's only a pragma away. // Carl