Efficient scaling for complex division

Source abstract

We develop a simple method for scaling to avoid overflow and harmful underflow in complex division. The method guarantees that no overflow will occur unless at least one component of the quotient must overflow, otherwise the normwise error in the computed result is at most a few units in the last place. Moreover, the scaling requires only four floating point multiplications and a small amount of integer arithmetic to compute the scale factor. Thus, on many modern CPUs, our method is both safer and faster than Smith's widely used algorithm.