A Scrapbook of Digital Signal Processing

Rounding

Rounding (especially binary rounding) is often done by adding 1 to the number. Be careful -- if the number is positive, add +1; if the number is negative, add -1. Otherwise, the rounding is done asymmetrically, with what can be rather deceptive results. Asymmetrical rounding at the end of each stage of a 16-bit fixed-point FFT, for instance, will cause an unexpectedly low noise floor with periodic spikes, instead of the expected smooth noise floor level at its normally higher value.

Home Page or Table of Contents