Causal coding of stationary sources and individual sequences with high resolution

Tamás Linder*, Ram Zamir

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

Abstract

In a causal source coding system, the reconstruction of the present source sample is restricted to be a function of the present and past source samples, while the code stream itself may be noncausal and have variable rate. Neuhoff and Gilbert showed that for memoryless sources, optimum performance among all causal source codes is achieved by time-sharing at most two memoryless codes (quantizers) followed by entropy coding. In this work, we extend Neuhoff and Gilbert's result in the limit of small distortion (high resolution) to two new settings. First, we show that at high resolution, an optimal causal code for a stationary source with finite differential entropy rate consists of a uniform quantizer followed by a (sequence) entropy coder. This implies that the price of causality at high resolution is approximately 0.254 bit, i.e., the space-filling loss of the uniform quantizer. Then, we consider individual sequences and introduce a deterministic analogue of differential entropy, which we call "Lempel-Ziv differential entropy." We show that for any bounded individual sequence with finite Lempel-Ziv differential entropy, optimum high-resolution performance among all finite-memory variable-rate causal codes is achieved by dithered scalar uniform quantization followed by Lempel-Ziv coding. As a by-product, we also prove an individual-sequence version of the Shannon lower bound.

Original languageEnglish
Pages (from-to)662-680
Number of pages19
JournalIEEE Transactions on Information Theory
Volume52
Issue number2
DOIs
StatePublished - Feb 2006

Keywords

  • Causal source codes
  • Differential entropy
  • Finite memory codes
  • Individual sequences
  • Lempel-Ziv complexity
  • Stationary sources
  • Uniform quantizer

Fingerprint

Dive into the research topics of 'Causal coding of stationary sources and individual sequences with high resolution'. Together they form a unique fingerprint.

Cite this