Dithered Quantization via Orthogonal Transformations

Abstract

Dithered quantization is a technique used to reduce or eliminate the statistical dependence between the signal and quantization error. This is most often achieved via adding pseudo-random noise prior to quantization. The present work develops a different dithering method, where dithering is accomplished by applying an orthogonal transformation to a vector of samples prior to quantization, and applying its inverse to the output of the quantizer. Focusing on uniform scalar quantization, it is shown that for any quantization rate, the proposed architecture approaches second-order independence, i.e., asymptotically vanishing correlation, as the dimension of the vector of samples processed jointly grows.