numeric_limits<double>::min() is positive!

Contrary to what one might expect, it is the smallest representable
*positive* number.

Using -max() works because by IEEE 754, floats are symmetric!