Unbiased Rounding in Digital Signal Processing
Original Publication Date: 1987-Aug-01
Included in the Prior Art Database: 2005-Feb-01
In all signal processing operations, there is signal growth. This signal growth means a binary value gets larger. At some point the binary value will exceed the size of the holding device (register, storage cell, file, etc.) allocated for the value. In the majority of applications, it is only necessary to keep the most significant bits that represent the signal strength. Therefore, a value that has exceeded an acceptable binary width, for a processor, can be shortened by discarding the low-order bits. This constant discarding of low-order bits will result in an error buildup that is unacceptable. To compensate for this, the traditional approach is to execute a round operation. The round operation adds a value to the shortened number. Proper selection of this value results in an average error of zero.