[IEEE Trans. on Communications, November 1992, pp. 1670-1674]
Robust Quantization of Memoryless Sources
using Dispersive FIR Filters
Kris Popat & Kenneth Zeger
Abstract
A novel approach to quantizing discrete-time memoryless sources is presented.
An important feature is that its performance is largely insensitive to errors
in modeling the input PDF. The method involves changing the amplitude
distribution of the source to be approximately Gaussian by all-pass filtering,
then applying a Lloyd-Max quantizer designed for a Gaussian source. After
quantization, the samples are passed through another all-pass filter, which is
an approximate inverse of the first filter. The mean-square error (MSE) for
the overall process is roughly equal to the quantization MSE for the
intermediate Gaussian signal, independent of the source statistics. For some
sources, this is actually an improvement over direct, correct-model Lloyd-Max
quantization. The cost of this technique is some delay due to filtering.