A Bound on the MSE of Oversampled Dithered Quantization With Feedback
Milan S. Derpich · IEEE Signal Processing Letters · 2009
We analyze the behavior of the mean squared error (MSE) achievable by oversampled, uniform scalar quantization using feedback, pre- and post-filters of unrestricted order, when encoding wide-sense stationary discrete-time random sources having (possibly) unbounded support. Our results are based upon the use of subtractively dithered uniform scalar quantizers. We consider the number of quantization levels,N, to be given and fixed, which lends itself to fixed-rate encoding, and focus on the cases in whichNis insufficient to avoid overload. In order to guarantee the stability of the closed-loop, we consider the use of a clipper before the scalar quantizer. Our results are valid for zero-mean sources having independent innovations whose moments satisfy some mild requirements, which are met by infinite-support distributions such as Gaussian and Laplacian. We show that, for fixedN, the MSE can be made to decay with the oversampling ratio lambda asO(e-c0lambda1/3) when lambda tends to infinity, wherec0[0.5(N-1)]2/3. We note that the latter bound is asymptotic in lambda but not inN, and that it includes clipping errors.