Computations are done in complex float. Casting these values to
complex short format introduces an error. Of course, the main
advantage is a factor 2 reduction in the size of the **output file**.

Now an example follows for the error in amplitude and phase for a complex value of about (100,100). If the actual interpolated complex value equals (100.5,100.5), then the error in the magnitude approximately is

(U1) |

If the actual value did equal (100.5,100.0), then the error in the phase is approximately

(U2) |

These are worst case scenarios. If the complex value is larger, then the relative error decreases. Note that the maximum for a signed short integer is .

Leijen 2009-04-14