E E Entropy and Information Theory
Robert M. Gray
Entropy and Information Theory Second Edition
Robert M. Gray Department of Electrical Engineering Stanford University Stanford, CA 94305-9510 USA [emailprotected]
ISBN 978-1-4419-7969-8 e-ISBN 978-1-4419-7970-4 DOI 10.1007/978-1-4419-7970-4 Springer New York Dordrecht Heidelberg London Library of Congress Control Number: 2011920808 © Springer Science+Business Media, LLC 2011 All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights.
Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)
to Tim, Lori, Julia, Peter, Gus, Amy, and Alice and in memory of Tino
Preface
This book is devoted to the theory of probabilistic information measures and their application to coding theorems for information sources and noisy channels, with a strong emphasis on source coding and stationary codes. The eventual goal is a general development of Shannon’s mathematical theory of communication for single user systems, but much of the space is devoted to the tools and methods required to prove the Shannon coding theorems, especially the notions of sources, channels, codes, entropy, information, and the entropy ergodic theorem. These tools form an area common to ergodic theory and information theory and comprise several quantitative notions of the information in random variables, random processes, and dynamical systems. Examples are entropy, mutual information, conditional entropy, conditional information, and relative entropy (discrimination, Kullback-Leibler information, informational divergence), along with the limiting normalized versions of these quantities such as entropy rate and information rate. In addition to information we will be concerned with the distance or distortion between the random objects, that is, the accuracy of the representation of one random object by another or the degree of mutual approximation. Much of the book is concerned with the properties of these quantities, especially the long term asymptotic behavior of average information and distortion, where both sample averages and probabilistic averages are of interest. The book has been strongly influenced by M. S. Pinsker’s classic Information and Information Stability of Random Variables and Processes and by the seminal work of A. N. Kolmogorov, I. M. Gelfand, A. M. Yaglom, and R. L. Dobrushin on information measures for abstract alphabets and their convergence properties. The book also has as a major influence the work of D.S. Ornstein on the isomorphism problem in ergodic theory, especially on his ideas of stationary codes mimicking block codes implied by the entropy ergodic theorem and of the d-bar distance between random processes. Many of the results herein are extensions of their
vii
viii
Preface
generalizations of Shannon’s original results. The mathematical models adopted here are more general than traditional treatments in that nonstationary and nonergodic information processes are treated. The models are somewhat less general than those of the Russian school of information theory in the sense that standard alphabets rather than completely abstract alphabets are considered. This restriction, however, permits many stronger results as well as the extension to nonergodic processes. In addition, the assumption of standard spaces simplifies many proofs and such spaces include as examples virtually all examples of engineering interest. The information convergence results are combined with ergodic theorems to prove general Shannon coding theorems for sources and channels. The results are not the most general known and the converses are not the strongest available in the literature, but they are sufficiently general to cover most sources and single-user communications systems encountered in applications and they are more general than those encountered in most modern texts. For example, most treatments confine interest to stationary and ergodic sources or even independent identically distributed (IID) sources and memoryless channels; here we consider asymptotic mean stationary sources, both one-sided and two-sided sources, and nonergodic sources. General channels with memory are considered, in particular the class of d-bar continuous channels. Perhaps more important than the generality of the sources and channels is the variety of code structures considered. Most of the literature and virtually all of the texts on information theory focus exclusively on block codes, while many codes are more naturally described as a stationary or sliding-block code — a time-invariant possibly nonlinear filter, generally with a discrete output. Here the basic results of information theory are described for stationary or sliding-block codes as well as for the traditional block codes and the relationships between the two coding structures are explored in detail. Stationary codes arose in ergodic theory in the context of Ornstein’s proof of the isomorphism theorem in the 1970s, and they arise naturally in the communications context of classical information theory, including common coding techniques such as time-invariant convolutional codes, predictive quantization, sigma-delta coding, and wavelet transform based techniques that operate as slidingwindow or online filters rather than as block operations. Mathematically, stationary codes preserve many of the statistical properties of the source being coded such as stationarity, ergodicity, and mixing. In practice, stationary codes avoid the introduction of blocking artifacts not present in the original source. This book can be considered as a sequel to my book Probability, Random Processes, and Ergodic Properties [58], as the first edition of this book was a sequel to the first edition [56]. There the prerequisite results on probability, standard spaces, and ordinary ergodic properties
Preface
ix
may be found along with a development of the general sources considered (asymptotically mean stationary, not necessarily ergodic) and of the process distortion measures used here. This book is self contained with the exception of common (and a few less common) results which may be found in the first book. Results quoted from the first book are cited for both first and second editions as the numbering system in the two editions differs. It is my hope that the book will interest engineers in some of the mathematical aspects and general models of the theory and mathematicians in some of the important engineering applications of performance bounds and code design for communication systems.
What’s New in the Second Edition As in the second edition of the companion volume [58], material has been corrected, rearranged, and rewritten in an effort to improve the flow of ideas and the presentation. This volume has been revised to reflect the changes in the companion volume, and citations to specific results are given for both the first and second editions [55, 58]. A significant amount of new material has been added both to expand some of the discussions to include more related topics and to include more recent results on old problems. More general distortion measures are considered when treating the process distance and distortion measures, consistent with extensions or results in [55] on metric distortion measures to powers of metrics (such as the ubiquitous squared-error distortion) in [58]. Three new chapters have been added: one on the interplay between distortion and entropy, one on the interplay between distortion and information, and one on properties of good source codes — codes that are either optimal or asymptotically optimal in the sense of converging to the Shannon limit. The chapter on distortion and entropy begins with a classic result treated in the first edition, the Fano inequality and its extensions, but it expands the discussion to consider the goodness of approximation of codes and their relation to entropy rate. Pinsker’s classic result relating variation distance between probability measures and the divergence (Kullback-Leibler) distance is now treated along with its recent extension by Marton comparing Ornstein’s d-bar process distance to divergence rate. The chapter contains a preliminary special case of the coding theorems to come — the application of the entropy ergodic theorem to the design of both block and sliding-block (stationary) almost lossless codes. The example introduces several basic ideas in a relatively simple context, including the construction of a sliding-block code from a block code in a
x
Preface
way that preserves the essential properties. The example also serves to illustrate the connections between information theory and ergodic theory by means of an interpretation of Ornstein’s isomorphism theorem — which is not proved here — in terms of almost lossless stationary coding — which is. The results also provide insight into the close relationships between source coding or data compression and rate-constrained simulation of a stationary and ergodic process, the finding of a simple model based on coin flips that resembles as closely as possible the given process. The chapter on distortion and information adds considerable material on rate-distortion theory to the treatment of the first edition, specifically on the evaluation of Shannon distortion-rate and rate-distortion functions along with their easy applications to lower bounds on performance in idealized communications systems. The fundamentals of Csiszár’s variational approach based on the divergence inequality is described and some of the rarely noted attributes are pointed out. The implied algorithm for the evaluation of rate-distortion functions (originally due to Blahut [18]) is interpreted as an early example of alternating optimization. An entirely new chapter on properties of good codes provides a development along the lines of Gersho and Gray [50] of the basic properties of optimal block codes originally due to Lloyd [110] and Steinhaus [175] along with the implied iterative design algorithm, another early example of alternating optimization. An incomplete extension of these block code optimality properties to sliding-block codes is described, and a simple example of trellis encoding is used to exemplify basic relations between block, sliding-block, and hybrid codes. The remainder of the chapter comprises recent developments in properties of asymptotically optimal sequences of sliding-block codes as developed by Mao, Gray, and Linder [117]. This material adds to the book’s emphasis on stationary and sliding-block codes and adds to the limited literature on the subject. Along with these major additions, I have added many minor results either because I was annoyed to discover they were not already in the first edition when I looked for them or because they eased the development of results. The addition of three new chapters was partially balanced by the merging of two old chapters to better relate information rates for finite alphabet and continuous alphabet random processes.
Errors Typographical and technical errors reported to or discovered by me during the two decades since the publication of the first edition have been
Preface
xi
corrected and efforts have been made to improve formatting and appearance of the book. Doubtless with the inclusion of new material new errors have occurred. As I age my frequency of typographical and other errors seems to grow along with my ability to see through them. I apologize for any that remain in the book. I will keep a list of all errors found by me or sent to me at [emailprotected] and I will post the list at my Web site, http://ee.stanford.edu/∼gray/.
Acknowledgments The research in information theory that yielded many of the results and some of the new proofs for old results in this book was supported by the National Science Foundation. Portions of the research and much of the early writing were supported by a fellowship from the John Simon Guggenheim Memorial Foundation. Recent research and writing on some of these topics has been aided by gifts from Hewlett Packard, Inc. The book benefited greatly from comments from numerous students and colleagues over many years; including Paul Shields, Paul Algoet, Ender Ayanoglu, Lee Davisson, John Kieffer, Dave Neuhoff, Don Ornstein, Bob Fontana, Jim Dunham, Farivar Saadat, Michael Sabin, Andrew Barron, Phil Chou, Tom Lookabaugh, Andrew Nobel, Bradley Dickinson, and Tamás Linder. I am grateful to Matt Shannon, Ricardo Blasco Serrano, Young-Han Kim, and Christopher Ellison for pointing out typographical errors.
Robert M. Gray Rockport, Massachusetts November 2010
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii 1
Information Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Probability Spaces and Random Variables . . . . . . . . . . . . . . . . . 1.2 Random Processes and Dynamical Systems . . . . . . . . . . . . . . . . 1.3 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Standard Alphabets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6 Asymptotic Mean Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.7 Ergodic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 1 5 7 12 13 16 17
2
Pair Processes: Channels, Codes, and Couplings . . . . . . . . . . . . . . 2.1 Pair Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Stationarity Properties of Channels . . . . . . . . . . . . . . . . . . . . . . . 2.4 Extremes: Noiseless and Completely Random Channels . . . . 2.5 Deterministic Channels and Sequence Coders . . . . . . . . . . . . . 2.6 Stationary and Sliding-Block Codes . . . . . . . . . . . . . . . . . . . . . . . . 2.7 Block Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.8 Random Punctuation Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . 2.9 Memoryless Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.10 Finite-Memory Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.11 Output Mixing Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.12 Block Independent Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.13 Conditionally Block Independent Channels . . . . . . . . . . . . . . . . 2.14 Stationarizing Block Independent Channels . . . . . . . . . . . . . . . 2.15 Primitive Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.16 Additive Noise Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.17 Markov Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21 21 22 25 29 30 31 37 38 42 42 43 45 46 46 48 49 49
xiii
xiv
Contents
2.18 Finite-State Channels and Codes . . . . . . . . . . . . . . . . . . . . . . . . . . 2.19 Cascade Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.20 Communication Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.21 Couplings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.22 Block to Sliding-Block: The Rohlin-Kakutani Theorem . . . . . .
50 51 52 52 53
3
Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Entropy and Entropy Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Divergence Inequality and Relative Entropy . . . . . . . . . . . . . . . 3.3 Basic Properties of Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Entropy Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Relative Entropy Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Conditional Entropy and Mutual Information . . . . . . . . . . . . . . 3.7 Entropy Rate Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.8 Markov Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.9 Relative Entropy Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61 61 65 69 78 81 82 90 91 93
4
The 4.1 4.2 4.3 4.4 4.5
Entropy Ergodic Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stationary Ergodic Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stationary Nonergodic Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . AMS Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Asymptotic Equipartition Property . . . . . . . . . . . . . . . . . . . .
97 97 100 106 110 114
5
Distortion and Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Distortion Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Fidelity Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Average Limiting Distortion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Communications Systems Performance . . . . . . . . . . . . . . . . . . . 5.5 Optimal Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6 Code Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.7 Approximating Random Vectors and Processes . . . . . . . . . . . 5.8 The Monge/Kantorovich/Vasershtein Distance . . . . . . . . . . . . 5.9 Variation and Distribution Distance . . . . . . . . . . . . . . . . . . . . . . . 5.10 Coupling Discrete Spaces with the Hamming Distance . . . . . 5.11 Process Distance and Approximation . . . . . . . . . . . . . . . . . . . . . . 5.12 Source Approximation and Codes . . . . . . . . . . . . . . . . . . . . . . . . . 5.13 d-bar Continuous Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
117 117 120 121 123 124 124 129 132 132 134 135 141 142
6
Distortion and Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 The Fano Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Code Approximation and Entropy Rate . . . . . . . . . . . . . . . . . . . . 6.3 Pinsker’s and Marton’s Inequalities . . . . . . . . . . . . . . . . . . . . . . . . 6.4 Entropy and Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5 Almost Lossless Source Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.6 Asymptotically Optimal Almost Lossless Codes . . . . . . . . . . .
147 147 150 152 156 160 168
Contents
xv
6.7 Modeling and Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 7
Relative Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1 Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Conditional Relative Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 Limiting Entropy Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4 Information for General Alphabets . . . . . . . . . . . . . . . . . . . . . . . . 7.5 Convergence Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
173 173 189 202 204 216
8
Information Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1 Information Rates for Finite Alphabets . . . . . . . . . . . . . . . . . . . . 8.2 Information Rates for General Alphabets . . . . . . . . . . . . . . . . . . 8.3 A Mean Ergodic Theorem for Densities . . . . . . . . . . . . . . . . . . . . 8.4 Information Rates of Stationary Processes . . . . . . . . . . . . . . . . . 8.5 The Data Processing Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6 Memoryless Channels and Sources . . . . . . . . . . . . . . . . . . . . . . . .
219 219 221 225 227 234 235
9
Distortion and Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.1 The Shannon Distortion-Rate Function . . . . . . . . . . . . . . . . . . . . 9.2 Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3 Process Definitions of the Distortion-Rate Function . . . . . . . . 9.4 The Distortion-Rate Function as a Lower Bound . . . . . . . . . . . 9.5 Evaluating the Rate-Distortion Function . . . . . . . . . . . . . . . . . . .
237 237 239 242 250 252
10 Relative Entropy Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.1 Relative Entropy Densities and Rates . . . . . . . . . . . . . . . . . . . . . . 10.2 Markov Dominating Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3 Stationary Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4 Mean Ergodic Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
265 265 268 272 275
11 Ergodic Theorems for Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Stationary Ergodic Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Stationary Nonergodic Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3 AMS Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.4 Ergodic Theorems for Information Densities. . . . . . . . . . . . . . .
281 281 286 290 293
12 Source Coding Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.1 Source Coding and Channel Coding . . . . . . . . . . . . . . . . . . . . . . . 12.2 Block Source Codes for AMS Sources . . . . . . . . . . . . . . . . . . . . . . 12.3 Block Source Code Mismatch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4 Block Coding Stationary Sources . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5 Block Coding AMS Ergodic Sources . . . . . . . . . . . . . . . . . . . . . . . . 12.6 Subadditive Fidelity Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.7 Asynchronous Block Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.8 Sliding-Block Source Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.9 A Geometric Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
295 295 296 307 310 312 319 321 323 333
xvi
Contents
13 Properties of Good Source Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.1 Optimal and Asymptotically Optimal Codes . . . . . . . . . . . . . . . 13.2 Block Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.3 Sliding-Block Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
335 335 337 343
14 Coding for Noisy Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.1 Noisy Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.2 Feinstein’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3 Feinstein’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.4 Channel Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.5 Robust Block Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.6 Block Coding Theorems for Noisy Channels . . . . . . . . . . . . . . . 14.7 Joint Source and Channel Block Codes . . . . . . . . . . . . . . . . . . . . 14.8 Synchronizing Block Channel Codes . . . . . . . . . . . . . . . . . . . . . . 14.9 Sliding-block Source and Channel Coding . . . . . . . . . . . . . . . . .
359 359 361 364 367 372 375 377 380 384
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Introduction
Abstract A brief history of the development of Shannon information theory is presented with an emphasis on its interactions with ergodic theory. The origins and goals of this book are sketched.
Information theory, the mathematical theory of communication, has two primary goals: The first is the development of the fundamental theoretical limits on the achievable performance when communicating a given information source over a given communications channel using coding schemes from within a prescribed class. The second goal is the development of coding schemes that provide performance that is reasonably good in comparison with the optimal performance given by the theory. Information theory was born in a remarkably rich state in the classic papers of Claude E. Shannon [162, 163] which contained the basic results for simple memoryless sources and channels and introduced more general communication systems models, including finitestate sources and channels. The key tools used to prove the original results and many of those that followed were special cases of the ergodic theorem and a new variation of the ergodic theorem which considered sample averages of a measure of the entropy or self information in a process. Information theory can be viewed as simply a branch of applied probability theory. Because of its dependence on ergodic theorems, however, it can also be viewed as a branch of ergodic theory, the theory of invariant transformations and transformations related to invariant transformations. In order to develop the ergodic theory example of principal interest to information theory, suppose that one has a random process, which for the moment we consider as a sample space or ensemble of possible output sequences together with a probability measure on events
xvii
xviii
Introduction
composed of collections of such sequences. The shift is the transformation on this space of sequences that takes a sequence and produces a new sequence by shifting the first sequence a single time unit to the left. In other words, the shift transformation is a mathematical model for the effect of time on a data sequence. If the probability of any sequence event is unchanged by shifting the event, that is, by shifting all of the sequences in the event, then the shift transformation is said to be invariant and the random process is said to be stationary. Thus the theory of stationary random processes can be considered as a subset of ergodic theory. Transformations that are not actually invariant (random processes which are not actually stationary) can be considered using similar techniques by studying transformations which are almost invariant, which are invariant in an asymptotic sense, or which are dominated or asymptotically dominated in some sense by an invariant transformation. This generality can be important as many real processes are not well modeled as being stationary. Examples are processes with transients, processes that have been parsed into blocks and coded, processes that have been encoded using variable-length codes or finite-state codes, and channels with arbitrary starting states. Ergodic theory was originally developed for the study of statistical mechanics as a means of quantifying the trajectories of physical or dynamical systems. Hence, in the language of random processes, the early focus was on ergodic theorems: theorems relating the time or sample average behavior of a random process to its ensemble or expected behavior. The work of Hoph [77], von Neumann [190] and others culminated in the pointwise or almost everywhere ergodic theorem of Birkhoff [17]. In the 1940’s and 1950’s Shannon made use of the ergodic theorem in the simple special case of memoryless processes to characterize the optimal performance possible when communicating an information source over a constrained random medium or channel using codes. The ergodic theorem was applied in a direct fashion to study the asymptotic behavior of error frequency and time average distortion in a communication system, but a new variation was introduced by defining a mathematical measure of the entropy or information in a random process and characterizing its asymptotic behavior. The results characterizing the optimal performance achievable using codes became known as coding theorems. Results describing performance that is actually achievable, at least in the limit of unbounded complexity and time, are known as positive coding theorems. Results providing unbeatable bounds on performance are known as converse coding theorems or negative coding theorems. When the same quantity is given by both positive and negative coding theorems, one has exactly the optimal performance achievable in theory using codes from a given class to communicate through the given communication systems model.
Introduction
xix
While mathematical notions of information had existed before, it was Shannon who coupled the notion with the ergodic theorem and an ingenious idea known as “random coding” in order to develop the coding theorems and to thereby give operational significance to such information measures. The name “random coding” is a bit misleading since it refers to the random selection of a deterministic code and not a coding system that operates in a random or stochastic manner. The basic approach to proving positive coding theorems was to analyze the average performance over a random selection of codes. If the average is good, then there must be at least one code in the ensemble of codes with performance as good as the average. The ergodic theorem is crucial to this argument for determining such average behavior. Unfortunately, such proofs promise the existence of good codes but give little insight into their construction. Shannon’s original work focused on memoryless sources whose probability distribution did not change with time and whose outputs were drawn from a finite alphabet or the real line. In this simple case the well-known ergodic theorem immediately provided the required result concerning the asymptotic behavior of information. He observed that the basic ideas extended in a relatively straightforward manner to more complicated Markov sources. Even this generalization, however, was a far cry from the general stationary sources considered in the ergodic theorem. To continue the story requires a few additional words about measures of information. Shannon really made use of two different but related measures. The first was entropy, an idea inherited from thermodynamics and previously proposed as a measure of the information in a random signal by Hartley [75]. Shannon defined the entropy of a discrete time discrete alphabet random process {Xn }, which we denote by H(X) while deferring its definition, and made rigorous the idea that the the entropy of a process is the amount of information in the process. He did this by proving a coding theorem showing that if one wishes to code the given process into a sequence of binary symbols so that a receiver viewing the binary sequence can reconstruct the original process perfectly (or nearly so), then one needs at least H(X) binary symbols or bits (converse theorem) and one can accomplish the task with very close to H(X) bits (positive theorem). This coding theorem is known as the noiseless source coding theorem. The second notion of information used by Shannon was mutual information. Entropy is really a notion of self information — the information provided by a random process about itself. Mutual information is a measure of the information contained in one process about another process. While entropy is sufficient to study the reproduction of a single process through a noiseless environment, more often one has two or more distinct random processes, e.g., one random process representing an infor-
xx
Introduction
mation source and another representing the output of a communication medium wherein the coded source has been corrupted by another random process called noise. In such cases observations are made on one process in order to make decisions on another. Suppose that {Xn , Yn } is a random process with a discrete alphabet, that is, taking on values in a discrete set. The coordinate random processes {Xn } and {Yn } might correspond, for example, to the input and output of a communication system. Shannon introduced the notion of the average mutual information between the two processes: I(X, Y ) = H(X) + H(Y ) − H(X, Y ),
(1)
the sum of the two self entropies minus the entropy of the pair. This proved to be the relevant quantity in coding theorems involving more than one distinct random process: the channel coding theorem describing reliable communication through a noisy channel, and the general source coding theorem describing the coding of a source for a user subject to a fidelity criterion. The first theorem focuses on error detection and correction and the second on analog-to-digital conversion and data compression. Special cases of both of these coding theorems were given in Shannon’s original work. Average mutual information can also be defined in terms of conditional entropy H(X|Y ) = H(X, Y ) − H(Y ) and hence I(X, Y ) = H(X) − H(X|Y ) = H(Y ) − H(X|Y ).
(2)
In this form the mutual information can be interpreted as the information contained in one process minus the information contained in the process when the other process is known. While elementary texts on information theory abound with such intuitive descriptions of information measures, we will minimize such discussion because of the potential pitfall of using the interpretations to apply such measures to problems where they are not appropriate. (See, e.g., P. Elias’ “Information theory, photosynthesis, and religion” in his “Two famous papers” [37].) Information measures are important because coding theorems exist imbuing them with operational significance and not because of intuitively pleasing aspects of their definitions. We focus on the definition (1) of mutual information since it does not require any explanation of what conditional entropy means and since it has a more symmetric form than the conditional definitions. It turns out that H(X, X) = H(X) (the entropy of a random variable is not changed by repeating it) and hence from (1) I(X, X) = H(X)
(3)
Introduction
xxi
so that entropy can be considered as a special case of average mutual information. To return to the story, Shannon’s work spawned the new field of information theory and also had a profound effect on the older field of ergodic theory. Information theorists, both mathematicians and engineers, extended Shannon’s basic approach to ever more general models of information sources, coding structures, and performance measures. The fundamental ergodic theorem for entropy was extended to the same generality as the ordinary ergodic theorems by McMillan [123] and Breiman [20] and the result is now known as the Shannon-McMillan-Breiman theorem. Other names are the asymptotic equipartition theorem or AEP, the ergodic theorem of information theory, and the entropy theorem. A variety of detailed proofs of the basic coding theorems and stronger versions of the theorems for memoryless, Markov, and other special cases of random processes were developed, notable examples being the work of Feinstein [39] [40] and Wolfowitz (see, e.g., Wolfowitz [196].) The ideas of measures of information, channels, codes, and communications systems were rigorously extended to more general random processes with abstract alphabets and discrete and continuous time by Khinchine [87], [88] and by Kolmogorov and his colleagues, especially Gelfand, Yaglom, Dobrushin, and Pinsker [49], [104], [101], [32], [150]. (See, for example, “Kolmogorov’s contributions to information theory and algorithmic complexity” [23].) In almost all of the early Soviet work, it was average mutual information that played the fundamental role. It was the more natural quantity when more than one process were being considered. In addition, the notion of entropy was not useful when dealing with processes with continuous alphabets since it is generally infinite in such cases. A generalization of the idea of entropy called discrimination was developed by Kullback (see, e.g., Kullback [106]) and was further studied by the Soviet school. This form of information measure is now more commonly referred to as relative entropy, cross entropy, or Kullback-Leibler number, or information divergence and it is better interpreted as a measure of similarity or dissimilarity between probability distributions than as a measure of information between random variables. Many results for mutual information and entropy can be viewed as special cases of results for relative entropy and the formula for relative entropy arises naturally in some proofs. It is the mathematical aspects of information theory and hence the descendants of the above results that are the focus of this book, but the developments in the engineering community have had as significant an impact on the foundations of information theory as they have had on applications. Simpler proofs of the basic coding theorems were developed for special cases and, as a natural offshoot, the rate of convergence to the optimal performance bounds characterized in a variety of important
xxii
Introduction
cases. See, e.g., the texts by Gallager [47], Berger [11], and Csiszàr and Körner [27]. Numerous practicable coding techniques were developed which provided performance reasonably close to the optimum in many cases: from the simple linear error correcting and detecting codes of Slepian [171] to the huge variety of algebraic codes that have been implemented (see, e.g., [12], [192],[109], [113], [19]), the various forms of convolutional, tree, and trellis codes for error correction and data compression (see, e.g., [189, 81]), and the recent codes approaching the Shannon limits based on iterative coding and message passage ideas [126, 156], codes which have their roots in Gallager’s PhD thesis on low density parity check codes [48]. Codes for source coding and data compression include a variety of traditional and recent techniques for lossless coding of data and lossy coding of realtime signals such as voice, audio, still images, and video. Techniques range from simple quantization to predictive quantization, adaptive methods, vector quantizers based on linear transforms followed by quantization and lossless codes, subband coders, and model coders such as the linear preditive codes for voice which fit linear models to observed signals for local synthesis. A sampling of the fundamentals through the standards can be found in [50, 160, 144, 178]. The engineering side of information theory through the middle 1970’s has been well chronicled by two IEEE collections: Key Papers in the Development of Information Theory, edited by D. Slepian [172], and Key Papers in the Development of Coding Theory, edited by E. Berlekamp [13] and many papers describing the first fifty years of the field were collected into Information Theory: 50 Years of Discovery in 2000 [184]. In addition there have been several survey papers describing the history of information theory during each decade of its existence published in the IEEE Transactions on Information Theory. The influence on ergodic theory of Shannon’s work was equally great but in a different direction. After the development of quite general ergodic theorems, one of the principal issues of ergodic theory was the isomorphism problem, the characterization of conditions under which two dynamical systems are really the same in the sense that each could be obtained from the other in an invertible way by coding. Here, however, the coding was not of the variety considered by Shannon — Shannon considered block codes, codes that parsed the data into nonoverlapping blocks or windows of finite length and separately mapped each input block into an output block. The more natural construct in ergodic theory can be called a sliding-block code or stationary code — here the encoder views a block of possibly infinite length and produces a single symbol of the output sequence using some mapping (or code or filter). The input sequence is then shifted one time unit to the left, and the same mapping applied to produce the next output symbol, and so on. This is a smoother operation than the block coding structure since the outputs
Introduction
xxiii
are produced based on overlapping windows of data instead of on a completely different set of data each time. Unlike the Shannon codes, these codes will produce stationary output processes if given stationary input processes. It should be mentioned that examples of such sliding-block codes often occurred in the information theory literature: time-invariant convolutional codes or, simply, time-invariant linear filters are slidingblock codes. It is perhaps odd that virtually all of the theory for such codes in the information theory literature was developed by effectively considering the sliding-block codes as very long block codes. Slidingblock codes have proved a useful structure for the design of noiseless codes for constrained alphabet channels such as magnetic recording devices, and techniques from symbolic dynamics have been applied to the design of such codes. See, for example [3, 118]. Shannon’s noiseless source coding theorem suggested a solution to the isomorphism problem: If we assume for the moment that one of the two processes is binary, then perfect coding of a process into a binary process and back into the original process requires that the original process and the binary process have the same entropy. Thus a natural conjecture is that two processes are isomorphic if and only if they have the same entropy. A major difficulty was the fact that two different kinds of coding were being considered: stationary sliding-block codes with zero error by the ergodic theorists and either fixed length block codes with small error or variable length (and hence nonstationary) block codes with zero error by the Shannon theorists. While it was plausible that the former codes might be developed as some sort of limit of the latter, this proved to be an extremely difficult problem. It was Kolmogorov [102], [103] who first reasoned along these lines and proved that in fact equal entropy (appropriately defined) was a necessary condition for isomorphism. Kolmogorov’s seminal work initiated a new branch of ergodic theory devoted to the study of entropy of dynamical systems and its application to the isomorphism problem. Most of the original work was done by Soviet mathematicians; notable papers are those by Sinai [168] [169] (in ergodic theory entropy is also known as the Kolmogorov-Sinai invariant), Pinsker [150], and Rohlin and Sinai [157]. An actual construction of a perfectly noiseless sliding-block code for a special case was provided by Meshalkin [124]. While much insight was gained into the behavior of entropy and progress was made on several simplified versions of the isomorphism problem, it was several years before Ornstein [138] proved a result that has since come to be known as the Ornstein isomorphism theorem or the Kolmogorov-Ornstein or Kolmogorov-Sinai-Ornstein isomorphism theorem. Ornstein showed that if one focused on a class of random processes which we shall call B-processes, then two processes are indeed isomorphic if and only if they have the same entropy. B-process are also called
xxiv
Introduction
Bernoulli processes in the ergodic theory literature, but this is potentially confusing because of the usage of “Bernoulli process” as a synonym of an independent identically distributed (IID) process in information theory and random process theory. B-processes have several equivalent definitions, perhaps the simplest is that they are processes which can be obtained by encoding a memoryless process using a sliding-block code. This class remains the most general class known for which the isomorphism conjecture holds. In the course of his proof, Ornstein developed intricate connections between block coding and sliding-block coding. He used Shannon-like techniques on the block codes, then imbedded the block codes into sliding-block codes, and then used the stationary structure of the sliding-block codes to advantage in limiting arguments to obtain the required zero error codes. Several other useful techniques and results were introduced in the proof: notions of the distance between processes and relations between the goodness of approximation and the difference of entropy. Ornstein expanded these results into a book [140] and gave a tutorial discussion in the premier issue of the Annals of Probability [139]. Several correspondence items by other ergodic theorists discussing the paper accompanied the article. The origins of this book lie in the tools developed by Ornstein for the proof of the isomorphism theorem rather than with the result itself. During the early 1970’s I first become interested in ergodic theory because of joint work with Lee D. Davisson on source coding theorems for stationary nonergodic processes. The ergodic decomposition theorem discussed in Ornstein [139] provided a needed missing link and led to an intense campaign on my part to learn the fundamentals of ergodic theory and perhaps find other useful tools. This effort was greatly eased by Paul Shields’ book The Theory of Bernoulli Shifts [164] and by discussions with Paul on topics in both ergodic theory and information theory. This in turn led to a variety of other applications of ergodic theoretic techniques and results to information theory, mostly in the area of source coding theory: proving source coding theorems for sliding-block codes and using process distance measures to prove universal source coding theorems and to provide new characterizations of Shannon distortionrate functions. The work was done with Dave Neuhoff, like me then an apprentice ergodic theorist, and Paul Shields. With the departure of Dave and Paul from Stanford, my increasing interest led me to discussions with Don Ornstein on possible applications of his techniques to channel coding problems. The interchange often consisted of my describing a problem, his generation of possible avenues of solution, and then my going off to work for a few weeks to understand his suggestions and work them through. One problem resisted our best efforts–how to synchronize block codes over channels with memory, a prerequisite for constructing sliding-block codes for such channels. In 1975 I had the good fortune to meet and talk
Introduction
xxv
with Roland Dobrushin at the 1975 IEEE/USSR Workshop on Information Theory in Moscow. He observed that some of his techniques for handling synchronization in memoryless channels should immediately generalize to our case and therefore should provide the missing link. The key elements were all there, but it took seven years for the paper by Ornstein, Dobrushin and me to evolve and appear [68]. Early in the course of the channel coding paper, I decided that having the solution to the sliding-block channel coding result in sight was sufficient excuse to write a book on the overlap of ergodic theory and information theory. The intent was to develop the tools of ergodic theory of potential use to information theory and to demonstrate their use by proving Shannon coding theorems for the most general known information sources, channels, and code structures. Progress on the book was disappointingly slow, however, for a number of reasons. As delays mounted, I saw many of the general coding theorems extended and improved by others (often by J. C. Kieffer) and new applications of ergodic theory to information theory developed, such as the channel modeling work of Neuhoff and Shields [133], [136], [135], [134] and design methods for sliding-block codes for input restricted noiseless channels by Adler, Coppersmith, and Hasner [3] and Marcus [118]. Although I continued to work in some aspects of the area, especially with nonstationary and nonergodic processes and processes with standard alphabets, the area remained for me a relatively minor one and I had little time to write. Work and writing came in bursts during sabbaticals and occasional advanced topic seminars. I abandoned the idea of providing the most general possible coding theorems and decided instead to settle for coding theorems that were sufficiently general to cover most applications and which possessed proofs I liked and could understand. Only one third of this book is actually devoted to Shannon source and channel coding theorems; the remainder can be viewed as a monograph on sources, channels, and codes and on information and distortion measures and their properties, especially their ergodic properties. The sources or random processes considered include asymptotically mean stationary processes with standard alphabets, a subject developed in detail in my earlier book Probability, Random Processes, and Ergodic Properties, which was published by Springer-Verlag in 1988 [55] with a second edition published by Springer in 2009. That books treats advanced probability and random processes with an emphasis on processes with standard alphabets, on nonergodic and nonstationary processes, and on necessary and sufficient conditions for the convergence of long term sample averages. Asymptotically mean stationary sources and the ergodic decomposition are there treated in depth and recent simplified proofs of the ergodic theorem due to Ornstein and Weiss [141] and others are incorporated. The next chapter of this book reviews some of the basic notation of the first one in information theoretic terms, but results are
xxvi
Introduction
often simply quoted as needed from the first book without any attempt to derive them. The two books together are self-contained in that all supporting results from probability theory and ergodic theory needed here may be found in the first book. This book is self-contained so far as its information theory content, but it should be considered as an advanced text on the subject and not as an introductory treatise to the reader only wishing an intuitive overview. The border between the two books is the beginning of the treatment of entropy. Here the Shannon-McMillan-Breiman theorem is proved using the coding approach of Ornstein and Weiss [141] (see also Shield’s tutorial paper [165]) and hence the treatments of ordinary ergodic theorems in the first book and the ergodic theorems for information measures in this book are consistent. The extension of the Shannon-McMillan-Breiman theorem to densities is proved using the “sandwich” approach of Algoet and Cover [7], which depends strongly on the usual pointwise or Birkhoff ergodic theorem: sample entropy is asymptotically sandwiched between two functions whose limits can be determined from the ergodic theorem. These results are the most general yet published in book form and differ from traditional developments in that martingale theory is not required in the proofs. A few words are in order regarding topics that are not contained in this book. I have not included the increasingly important and growing area of multiuser information theory because my experience in the area is slight and I believe this topic can be better handled by others. Traditional noiseless coding theorems and actual codes such as the Huffman codes are not considered in depth because quite good treatments exist in the literature, e.g., [47], [1], [122]. The corresponding ergodic theory result — the Ornstein isomorphism theorem — is also not proved, because its proof is difficult and the result is not needed for the Shannon coding theorems. It is, however, described and many techniques used in its proof are used here for similar and other purposes. The actual computation of channel capacity and distortion rate functions has not been included because existing treatments [47], [18], [11], [25] [57] are quite adequate. New to the second edition, however, is a partial development of Csiszár’s [25] rigorous development of the information-theoretic optimization underlying the evaluation of the ratedistortion function. This book does not treat code design techniques in any depth, but in this second edition properties of optimal and asymptotically optimal source codes are developed and these properties provide insight into the structure of good codes and can be used to guide code design. The traditional Lloyd optimality properties for vector quantizers are described along with recent results for sliding-block codes which resemble their block coding cousins.
Introduction
xxvii
J. C. Kieffer developed a powerful new ergodic theorem that can be used to prove both traditional ergodic theorems and the extended Shannon-McMillan-Brieman theorem [96]. He has used this theorem to prove strong (almost everywhere) versions of the source coding theorem and its converse, that is, results showing that sample average distortion is with probability one no smaller than the distortion-rate function and that there exist codes with sample average distortion arbitrarily close to the distortion-rate function [99, 100].
Chapter 1
Information Sources
Abstract An information source or source is a mathematical model for a physical entity that produces a succession of symbols called “outputs” in a random manner. The symbols produced may be real numbers such as voltage measurements from a transducer, binary numbers as in computer data, two dimensional intensity fields as in a sequence of images, continuous or discontinuous waveforms, and so on. The space containing all of the possible output symbols is called the alphabet of the source and a source is essentially an assignment of a probability measure to events consisting of sets of sequences of symbols from the alphabet. It is useful, however, to explicitly treat the notion of time as a transformation of sequences produced by the source. Thus in addition to the common random process model we shall also consider modeling sources by dynamical systems as considered in ergodic theory. The material in this chapter is a distillation of [55, 58] and is intended to establish notation.
1.1 Probability Spaces and Random Variables A measurable space (Ω, B) is a pair consisting of a sample space Ω together with a σ -field B of subsets of Ω (also called the event space). A σ -field or σ -algebra B is a nonempty collection of subsets of Ω with the following properties: Ω ∈ B. (1.1) If F ∈ B, then F c = {ω : ω 6∈ F } ∈ B. [ Fi ∈ B. If Fi ∈ B; i = 1, 2, . . . , then
(1.2) (1.3)
i
From de Morgan’s “laws” of elementary set theory it follows that also
R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_1, © Springer Science+Business Media, LLC 2011
1
2
1 Information Sources ∞ \
Fi = (
i=1
∞ [
Fic )c ∈ B.
i=1
An event space is a collection of subsets of a sample space (called events by virtue of belonging to the event space) such that any countable sequence of set theoretic operations (union, intersection, complementation) on events produces other events. Note that there are two extremes: the largest possible σ -field of Ω is the collection of all subsets of Ω (sometimes called the power set) and the smallest possible σ -field is {Ω, ∅}, the entire space together with the null set ∅ = Ωc (called the trivial space). If instead of the closure under countable unions required by (1.3), we only require that the collection of subsets be closed under finite unions, then we say that the collection of subsets is a field. While the concept of a field is simpler to work with, a σ -field possesses the additional important property that it contains all of the limits of sequences of sets in the collection. That is, if Fn , n = 1, 2, · · · is an increasing sequence of sets in a σ -field, that is, if Fn−1 ⊂ Fn and if S∞ F = n=1 Fn (in which case we write Fn ↑ F or limn→∞ Fn = F ), then also F is contained in the σ -field. In a similar fashion we can define decreasing sequences of sets: If Fn decreases to F in the sense that Fn+1 ⊂ Fn and T∞ F = n=1 Fn , then we write Fn ↓ F . If Fn ∈ B for all n, then F ∈ B. A probability space (Ω, B, P ) is a triple consisting of a sample space Ω , a σ -field B of subsets of Ω , and a probability measure P which assigns a real number P (F ) to every member F of the σ -field B so that the following conditions are satisfied: • Nonnegativity: P (F ) ≥ 0, all F ∈ B;
(1.4)
P (Ω) = 1;
(1.5)
• Normalization: • Countable Additivity: If Fi ∈ B, i = 1, 2, · · · are disjoint, then P(
∞ [
i=1
Fi ) =
∞ X
P (Fi ).
(1.6)
i=1
A set function P satisfying only (1.4) and (1.6) but not necessarily (1.5) is called a measure and the triple (Ω, B, P ) is called a measure space. Since the probability measure is defined on a σ -field, such countable unions of subsets of Ω in the σ -field are also events in the σ -field. A standard result of basic probability theory is that T∞ if Gn ↓ ∅ (the empty or null set), that is, if Gn+1 ⊂ Gn for all n and n=1 Gn = ∅ , then we have
1.1 Probability Spaces and Random Variables
3
• Continuity at ∅: lim P (Gn ) = 0.
n→∞
(1.7)
similarly it follows that we have • Continuity from Below: If Fn ↑ F , then lim P (Fn ) = P (F ), n→∞
(1.8)
and • Continuity from Above: If Fn ↓ F , then lim P (Fn ) = P (F ). n→∞
(1.9)
Given a measurable space (Ω, B), a collection G of members of B is said to generate B and we write σ (G) = B if B is the smallest σ -field that contains G; that is, if a σ -field contains all of the members of G, then it must also contain all of the members of B. The following is a fundamental approximation theorem of probability theory. A proof may be found in Corollary 1.5.3 of [55] or Corollary 1.5 of [58]. The result is most easily stated in terms of the symmetric difference ∆ defined by F ∆G ≡ (F ∩ Gc ) ∪ (F c ∩ G). Theorem 1.1. Given a probability space (Ω, B, P ) and a generating field F , that is, F is a field and B = σ (F ), then given F ∈ B and > 0, there exists an F0 ∈ F such that P (F ∆F0 ) ≤ . Let (A, BA ) denote another measurable space. We will also use B(A) as a synonym for BA . A random variable or measurable function defined on (Ω, B) and taking values in (A,BA ) is a mapping or function f : Ω → A with the property that if F ∈ BA , then f −1 (F ) = {ω : f (ω) ∈ F } ∈ B.
(1.10)
The name “random variable” is commonly associated with the special case where A is the real line and B the Borel field, the smallest σ -field containing all the intervals. Occasionally a more general sounding name such as “random object” is used for a measurable function to implicitly include random variables (A the real line), random vectors (A a Euclidean space), and random processes (A a sequence or waveform space). We will use the terms “random variable” in the more general sense. Usually A will either be a metric space or a product of metric spaces, in which case the σ -field will be a Borel field BA or B(A) of subsets of A. If A is a product of metric spaces, then BA will be taken as the corresponding product σ -field, that is, the σ -field generated by the rectangles.
4
1 Information Sources
A random variable is just a function or mapping with the property that inverse images of “output events” determined by the random variable are events in the original measurable space. This simple property ensures that the output of the random variable will inherit its own probability measure. For example, with the probability measure Pf defined by Pf (B) = P (f −1 (B)) = P (ω : f (ω) ∈ B); B ∈ BA , (A, BA , Pf ) becomes a probability space since measurability of f and elementary set theory ensure that Pf is indeed a probability measure. The induced probability measure Pf is called the distribution of the random variable f . The measurable space (A, BA ) or, simply, the sample space A, is called the alphabet of the random variable f . We shall occasionally also use the notation P f −1 which is a mnemonic for the relation P f −1 (F ) = P (f −1 (F )) and which is less awkward when f itself is a function with a complicated name, e.g., ΠI→M . It is often convenient to abbreviate an English description the of a probability of an event to the pseudo mathematical form Pr(f ∈ F ), which can be considered shorthand for Pf (F ) = P (f −1 (F )) and can be read as “the probability that f is in F .” If the alphabet A of a random variable f is not clear from context, then we shall refer to f as an A-valued random variable. . If f is a measurable function from (Ω, B) to (A, BA ), we will say that f is B/BA -measurable if the σ -fields might not be clear from context. Given a probability space (Ω, B, P ), a collection of subsets G is a subσ -field if it is a σ -field and all its members are in B. A random variable f : Ω → A is said to be measurable with respect to a sub-σ -field G if f −1 (H) ∈ G for all H ∈ BA . Given a probability space (Ω, B, P ) and a sub-σ -field G, for any event H ∈ B the conditional probability m(H|G) is defined as any function, say g, which satisfies the two properties g is measurable with respect to G Z \ ghdP = m(G H); all G ∈ G.
(1.11) (1.12)
G
An important special case of conditional probability occurs when studying the distributions of random variables defined on an underlying probability space. Suppose that X : Ω → AX and Y : Ω → AY are two random variables defined on (Ω, B, P ) with alphabets AX and AY and σ fields BAX and BAY , respectively. Let PXY denote the induced distribution on (AXT× AY , BAX × BAY ), that is, PXY (F × G) = P (X ∈ F , Y ∈ G) = P (X −1 (F ) Y −1 (G)). Let σ (Y ) denote the sub-σ -field of B generated by Y , that is, Y −1 (BAY ). Since the conditional probability P (F |σ (Y )) is real-valued and measurable with respect to σ (Y ), it can be written as
1.2 Random Processes and Dynamical Systems
5
g(Y (ω)), ω ∈ Ω, for some function g(y). (See, for example, Lemma 5.2.1 of [55] or Lemma 6.1 of [58].) Define P (F |y) = g(y). For a fixed F ∈ BAX define the conditional distribution of F given Y = y by PX|Y (F |y) = P (X −1 (F )|y); y ∈ BAY . From the properties of conditional probability, Z PXY (F × G) = PX|Y (F |y)dPY (y); F ∈ BAX , G ∈ BAY .
(1.13)
G
It is tempting to think that for a fixed y, the set function defined by PX|Y (F |y); F ∈ BAX is actually a probability measure. This is not the case in general. When it does hold for a conditional probability measure, the conditional probability measure is said to be regular. This text will focus on standard alphabets for which regular conditional probabilities always exist.
1.2 Random Processes and Dynamical Systems We now consider two mathematical models for a source: A random process and a dynamical system. The first is the familiar one in elementary courses, a source is just a random process or sequence of random variables. The second model is possibly less familiar — a random process can also be constructed from an abstract dynamical system consisting of a probability space together with a transformation on the space. The two models are connected by considering a time shift to be a transformation. A discrete time random process or, simply, a random process is a sequence of random variables {Xn }n∈T or {Xn ; n ∈ T}, where T is an index set, defined on a common probability space (Ω, B, P ). We define a source as a random process, although we could also use the alternative definition of a dynamical system to be introduced shortly. We usually assume that all of the random variables share a common alphabet, say A. The two most common index sets of interest are the set of all integers Z = {· · · , −2, −1, 0, 1, 2, · · · }, in which case the random process is referred to as a two-sided random process, and the set of all nonnegative integers Z+ = {0, 1, 2, · · · }, in which case the random process is said to be one-sided. One-sided random processes will often prove to be far more difficult in theory, but they provide better models for physical random processes that must be “turned on” at some time or which have transient behavior. Observe that since the alphabet A is general, we could also model continuous time random processes in the above fashion by letting A
6
1 Information Sources
consist of a family of waveforms defined on an interval, e.g., the random variable Xn could in fact be a continuous time waveform X(t) for t ∈ [nT , (n + 1)T ), where T is some fixed positive real number. The above definition does not specify any structural properties of the index set T . In particular, it does not exclude the possibility that T be a finite set, in which case “random vector” would be a better name than “random process.” In fact, the two cases of T = Z and T = Z+ will be the only important examples for our purposes. Nonetheless, the general notation of T will be retained in order to avoid having to state separate results for these two cases. An abstract dynamical system consists of a probability space (Ω, B, P ) together with a measurable transformation T : Ω → Ω of Ω into itself. Measurability means that if F ∈ B, then also T −1 F = {ω : T ω ∈ F }∈ B. The quadruple (Ω,B,P ,T ) is called a dynamical system in ergodic theory. The interested reader can find excellent introductions to classical ergodic theory and dynamical system theory in the books of Halmos [73] and Sinai [170]. More complete treatments may be found in [16], [164], [149], [30], [191], [140], [46]. The term “dynamical systems” comes from the focus of the theory on the long term “dynamics” or “dynamical behavior” of repeated applications of the transformation T on the underlying measure space. An alternative to modeling a random process as a sequence or family of random variables defined on a common probability space is to consider a single random variable together with a transformation defined on the underlying probability space. The outputs of the random process will then be values of the random variable taken on transformed points in the original space. The transformation will usually be related to shifting in time and hence this viewpoint will focus on the action of time itself. Suppose now that T is a measurable mapping of points of the sample space Ω into itself. It is easy to see that the cascade or composition of measurable functions is also measurable. Hence the transformation T n defined as T 2 ω = T (T ω) and so on (T n ω = T (T n−1 ω)) is a measurable function for all positive integers n. If f is an A-valued random variable defined on (Ω, B), then the functions f T n : Ω → A defined by f T n (ω) = f (T n ω) for ω ∈ Ω will also be random variables for all n in Z+ . Thus a dynamical system together with a random variable or measurable function f defines a one-sided random process {Xn }n∈Z+ by Xn (ω) = f (T n ω). If it should be true that T is invertible, that is, T is one-to-one and its inverse T −1 is measurable, then one can define a two-sided random process by Xn (ω) = f (T n ω), all n in Z. The most common dynamical system for modeling random processes is that consisting of a sequence space Ω containing all one- or two-sided A-valued sequences together with the shift transformation T , that is, the transformation that maps a sequence {xn } into the sequence {xn+1 } wherein each coordinate has been shifted to the left by one time unit.
1.3 Distributions
7
Thus, for example, let Ω = AZ+ = {all x = (x0 , x1 , · · · ) with xi ∈ A for all i} and define T : Ω → Ω by T (x0 , x1 , x2 , · · · ) = (x1 , x2 , x3 , · · · ). T is called the shift or left shift transformation on the one-sided sequence space. The shift for two-sided spaces is defined similarly. The sequencespace model of a random process is sometimes referred to as the Kolmogorov representation of a process. The different models provide equivalent models for a given process — one emphasizing the sequence of outputs and the other emphasising the action of a transformation on the underlying space in producing these outputs. In order to demonstrate in what sense the models are equivalent for given random processes, we next turn to the notion of the distribution of a random process.
1.3 Distributions While in principle all probabilistic quantities associated with a random process can be determined from the underlying probability space, it is often more convenient to deal with the induced probability measures or distributions on the space of possible outputs of the random process. In particular, this allows us to compare different random processes without regard to the underlying probability spaces and thereby permits us to reasonably equate two random processes if their outputs have the same probabilistic structure, even if the underlying probability spaces are quite different. We have already seen that each random variable Xn of the random process {Xn } inherits a distribution because it is measurable. To describe a process, however, we need more than just probability measures on output values of separate individual random variables; we require probability measures on collections of random variables, that is, on sequences of outputs. In order to place probability measures on sequences of outputs of a random process, we first must construct the appropriate measurable spaces. A convenient technique for accomplishing this is to consider product spaces, spaces for sequences formed by concatenating spaces for individual outputs. Let T denote any finite or infinite set of integers. In particular, T = Z(n) = {0, 1, 2, · · · , n − 1}, T = Z, or T = Z+ . Define x T = {xi }i∈T . For example, x Z = (· · · , x−1 , x0 , x1 , · · · ) is a two-sided infinite sequence. When T = Z(n) we abbreviate x Z(n) to simply x n . Given alphabets Ai , i ∈ T , define the cartesian product space × Ai = {all x T : xi , ∈ Ai all i in T}.
i∈T
8
1 Information Sources
In most cases all of the Ai will be replicas of a single alphabet A and the above product will be denoted simply by AT . Thus, for example, A{m,m+1,··· ,n} is the space of all possible outputs of the process from time m to time n; AZ is the sequence space of all possible outputs of a two-sided process. We shall abbreviate the notation for the space AZ(n) , the space of all n dimensional vectors with coordinates in A, by An . To obtain useful σ -fields of the above product spaces, we introduce the idea of a rectangle in a product space. A rectangle in AT taking values in the coordinate σ -fields Bi , i ∈ J, is defined as any set of the form B = {x T ∈ AT : xi ∈ Bi ; all i in J},
(1.14)
where J is a finite subset of the index set T and Bi ∈ Bi for all i ∈ J. (Hence rectangles are sometimes referred to as finite dimensional rectangles.) A rectangle as in (1.14) can be written as a finite intersection of one-dimensional rectangles as \ \ B= {x T ∈ AT : xi ∈ Bi } = Xi −1 (Bi ) (1.15) i∈J
i∈J
where here we consider Xi as the coordinate functions Xi : AT → A defined by Xi (x T ) = xi . As rectangles in AT are clearly fundamental events, they should be members of any useful σ -field of subsets of AT . Define the product σ field BA T as the smallest σ -field containing all of the rectangles, that is, the collection of sets that contains the clearly important class of rectangles and the minimum amount of other stuff required to make the collection a σ -field. To be more precise, given an index set T of integers, let RECT (Bi , i ∈ T) denote the set of all rectangles in AT taking coordinate values in sets in Bi , i ∈ T . We then define the product σ -field of AT by BA T = σ (RECT (Bi , i ∈ T)). (1.16) Consider an index set T and an A-valued random process {Xn }n∈T defined on an underlying probability space (Ω, B, P ). Given any index set J ⊂ T , measurability of the individual random variables Xn implies that of the random vectors X J = {Xn ; n ∈ J}. Thus the measurable space (AJ , BA J ) inherits a probability measure from the underlying space through the random variables X J . Thus in particular the measurable space (AT , BA T ) inherits a probability measure from the underlying probability space and thereby determines a new probability space (AT , BA T , PX T ), where the induced probability measure is defined by PX T (F ) = P ((X T )−1 (F )) = P (ω : X T (ω) ∈ F ); F ∈ BA T .
(1.17)
1.3 Distributions
9
Such probability measures induced on the outputs of random variables are referred to as distributions for the random variables, exactly as in the simpler case first treated. When T = {m, m + 1, · · · , m + n − 1}, e.g., n when we are treating Xm = (Xn , · · · , Xm+n−1 ) taking values in An , the distribution is referred to as an n-dimensional or nth order distribution and it describes the behavior of an n-dimensional random variable. If T is the entire process index set, e.g., if T = Z for a two-sided process or T = Z+ for a one-sided process, then the induced probability measure is defined to be the distribution of the process. Thus, for example, a probability space (Ω, B, P ) together with a doubly infinite sequence of random variables {Xn }n∈Z induces a new probability space (AZ , BA Z , PX Z ) and PX Z is the distribution of the process. For simplicity, let us now denote the process distribution simply by m. We shall call the probability space (AT , BA T , m) induced in this way by a random process {Xn }n∈Z the output space or sequence space of the random process. Since the sequence space (AT , BA T , m) of a random process {Xn }n∈Z is a probability space, we can define random variables and hence also random processes on this space. One simple and useful such definition is that of a sampling or coordinate or projection function defined as follows: Given a product space AT , define the sampling functions Πn : AT → A by Πn (x T ) = xn , x T ∈ AT ; n ∈ T. (1.18) The sampling function is named Π since it is also a projection. Observe that the distribution of the random process {Πn }n∈T defined on the probability space (AT , BA T , m) is exactly the same as the distribution of the random process {Xn }n∈T defined on the probability space (Ω, B, P ). In fact, so far they are the same process since the {Πn } simply read off the values of the {Xn }. What happens, however, if we no longer build the Πn on the Xn , that is, we no longer first select ω from Ω according to P , then form the sequence x T = X T (ω) = {Xn (ω)}n∈T , and then define Πn (x T ) = Xn (ω)? Instead we directly choose an x in AT using the probability measure m and then view the sequence of coordinate values. In other words, we are considering two completely separate experiments, one described by the probability space (Ω, B, P ) and the random variables {Xn } and the other described by the probability space (AT , BA T , m) and the random variables {Πn }. In these two separate experiments, the actual sequences selected may be completely different. Yet intuitively the processes should be the “same” in the sense that their statistical structures are identical, that is, they have the same distribution. We make this intuition formal by defining two processes to be equivalent if their process distributions are identical, that is, if the probability measures on the output sequence spaces are the same, regardless of the functional form of the random variables of the underlying probability spaces. In the same way, we con-
10
1 Information Sources
sider two random variables to be equivalent if their distributions are identical. We have described above two equivalent processes or two equivalent models for the same random process, one defined as a sequence of random variables on a perhaps very complicated underlying probability space, the other defined as a probability measure directly on the measurable space of possible output sequences. The second model will be referred to as a directly given random process or a the Kolmogorov model for the random process. Which model is “better” depends on the application. For example, a directly given model for a random process may focus on the random process itself and not its origin and hence may be simpler to deal with. If the random process is then coded or measurements are taken on the random process, then it may be better to model the encoded random process in terms of random variables defined on the original random process and not as a directly given random process. This model will then focus on the input process and the coding operation. We shall let convenience determine the most appropriate model. We can now describe yet another model for the above random process, that is, another means of describing a random process with the same distribution. This time the model is in terms of a dynamical system. Given the probability space (AT , BA T , m), define the (left) shift transformation T : AT → AT by T (x T ) = T ({xn }n∈T ) = y T = {yn }n∈T , where yn = xn+1 , n ∈ T. Thus the nth coordinate of y T is simply the (n + 1)st coordinate of x T . (We assume that T is closed under addition and hence if n and 1 are in T, then so is (n + 1).) If the alphabet of such a shift is not clear from context, we will occasionally denote the shift by TA or TAT . The shift can easily be shown to be measurable. Consider next the dynamical system (AT , BA T , P , T ) and the random process formed by combining the dynamical system with the zero time sampling function Π0 (we assume that 0 is a member of T ). If we define Yn (x) = Π0 (T n x) for x = x T ∈ AT , or, in abbreviated form, Yn = Π0 T n , then the random process {Yn }n∈T is equivalent to the processes developed above. Thus we have developed three different, but equivalent, means of producing the same random process. Each will be seen to have its uses. The above development shows that a dynamical system is a more fundamental entity than a random process since we can always construct an equivalent model for a random process in terms of a dynamical system — use the directly given representation, shift transformation, and zero
1.3 Distributions
11
time sampling function. Two important properties of dynamical systems or random processes can be defined at this point, the implications will be developed throughout the book. A dynamical system (AT , BA T , P , T ) is said to be stationary (with respect to T ) if the distribution P is invariant with respect to P , that is, P (T −1 F ) = P (F ), all F ∈ BA T .
(1.19)
In other words, probabilities of process events are unchanged by shifting. The dynamical system is said to be ergodic if If T −1 F = F , then P (F ) = 0 or 1,
(1.20)
that is, all invariant events are trivial. Note that neither definition implies or excludes the other. The shift transformation on a sequence space introduced above is the most important transformation that we shall encounter. It is not, however, the only important transformation. When dealing with transformations we will usually use the notation T to reflect the fact that it is often related to the action of a simple left shift of a sequence, yet it should be kept in mind that occasionally other operators will be considered and the theory to be developed will remain valid, even if T is not required to be a simple time shift. For example, we will also consider block shifts. Most texts on ergodic theory deal with the case of an invertible transformation, that is, where T is a one-to-one transformation and the inverse mapping T −1 is measurable. This is the case for the shift on AZ , the two-sided shift. It is not the case, however, for the one-sided shift defined on AZ+ and hence we will avoid use of this assumption. We will, however, often point out in the discussion what simplifications or special properties arise for invertible transformations. Since random processes are considered equivalent if their distributions are the same, we shall adopt the notation [A, m, X] for a random process {Xn ; n ∈ T} with alphabet A and process distribution m, the index set T usually being clear from context. We will occasionally abbreviate this to the more common notation [A, m], but it is often convenient to note the name of the output random variables as there may be several, e.g., a random process may have an input X and output Y . By “the associated probability space” of a random process [A, m, X] we shall mean the sequence probability space (AT , BA T , m). It will often be convenient to consider the random process as a directly given random process, that is, to view Xn as the coordinate functions Πn on the sequence space AT rather than as being defined on some other abstract space. This will not always be the case, however, as often processes will be formed by coding or communicating other random processes. Context should render such bookkeeping details clear.
12
1 Information Sources
1.4 Standard Alphabets A measurable space (A, BA ) is a standard space if there exists a sequence of finite fields Fn ; n = 1, 2, · · · with the following properties: (1)Fn ⊂ Fn+1 (the fields are increasing). (2)BA is the smallest S∞ σ -field containing all of the Fn (the Fn generate BA or BA = σ ( n=1 Fn )). (3)An event Gn ∈ Fn is called an atom of the field if it is nonempty and and its only subsets which are also field members are itself and the empty set. If Gn ∈ Fn ; n = 1, 2, · · · are atoms and Gn+1 ⊂ Gn for all n, then ∞ \ Gn 6= ∅. n=1
Standard spaces are important for several reasons: First, they are a general class of spaces for which two of the key results of probability hold: (1) the Kolmogorov extension theorem showing that a random process is completely described by its finite order distributions, and (2) the existence of regular conditional probability measures. Thus, in particular, the conditional probability measure PX|Y (F |y) of (1.13) is regular if the alphabets AX and AY are standard and hence for each fixed y ∈ AY the set function PX|Y (F |y); F ∈ BAX is a probability measure. In this case we can interpret PX|Y (F |y) as P (X ∈ F |Y = y). Second, the ergodic decomposition theorem of ergodic theory holds for such spaces. The ergodic decomposition implies that any stationary process is equivalent to a mixture of stationary and ergodic processes; that is, a stationary nonergodic source can be viewed as a random selection of one of a family of stationary and ergodic sources. Third, the class is sufficiently general to include virtually all examples arising in applications, e.g., discrete spaces, the real line, Euclidean vector spaces, Polish spaces (complete separable metric spaces), etc. The reader is referred to [55] or [58] and the references cited therein for a detailed development of these properties and examples of standard spaces. Standard spaces are not the most general space for which the Kolmogorov extension theorem, the existence of conditional probability, and the ergodic decomposition theorem all hold. These results also hold for perfect spaces which include standard spaces as a special case. (See, e.g., [161],[174],[155], [114].) We limit discussion to standard spaces, however, as they are easier to characterize and work with and they are sufficiently general to handle most cases encountered in applications. Although standard spaces are not the most general for which the required probability theory results hold, they are the most general for which all finitely additive normalized measures extend to countably additive prob-
1.5 Expectation
13
ability measures, a property which greatly eases the proof of many of the desired results. Throughout this book we shall assume that the alphabet A of the information source is a standard space.
1.5 Expectation Let (Ω, B, m) be a probability space, e.g., the probability space of a directly given random process with alphabet A, (AT , BA T , m). A real-valued random variable f : Ω → R will also be called a measurement since it is often formed by taking a mapping or function of some other set of more general random variables, e.g., the outputs of some random process which might not have real-valued outputs. Measurements made on such processes, however, will always be assumed to be real. Suppose next we have a measurement f whose range space or alphabet f (Ω) ⊂ R of possible values is finite. Then f is called a discrete random variable or discrete measurement or digital measurement or, in the common mathematical terminology, a simple function. Given a discrete measurement f , suppose that its range space is f (Ω) = {bi , i = 1, · · · , N}, where the bi are distinct. Define the sets Fi = f −1 (bi ) = {x : f (x) = bi }, i = 1, · · · , N. Since f is measurable, the Fi are all members of B. Since the bi are distinct, the Fi are disjoint. Since every input point in Ω must map into some bi , the union of the Fi equals Ω. Thus the collection {Fi ; i = 1, 2, · · · , N} forms a partition of Ω. We have therefore shown that any discrete measurement f can be expressed in the form M X f (x) = bi 1Fi (x), (1.21) i=1
where bi ∈ R, the Fi ∈ B form a partition of Ω, and 1Fi is the indicator function of Fi , i = 1, · · · , M. Every simple function has a unique representation in this form with distinct bi and {Fi } a partition. The expectation or ensemble average or probabilistic average or mean of a discrete measurement f : Ω → R as in (1.21) with respect to a probability measure m is defined by Em f =
M X
bi m(Fi ).
(1.22)
i=0
An immediate consequence of the definition of expectation is the simple but useful fact that for any event F in the original probability space, Em 1F = m(F ),
14
1 Information Sources
that is, probabilities can be found from expectations of indicator functions. Again let (Ω, B, m) be a probability space and f : Ω → R a measurement, that is, a real-valued random variable or measurable real-valued function. Define the sequence of quantizers qn : R → R, n = 1, 2, · · · , as follows: n n≤r (k − 1)2−n (k − 1)2−n ≤ r < k2−n , k = 1, 2, · · · , n2n qn (r ) = −n −(k − 1)2 −k2−n ≤ r < −(k − 1)2−n ; k = 1, 2, · · · , n2n −n r < −n. We now define expectation for general measurements in two steps. If f ≥ 0, then define Em f = lim Em (qn (f )). (1.23) n→∞
Since the qn are discrete measurements on f , the qn (f ) are discrete measurements on Ω (qn (f )(x) = qn (f (x)) is a simple function) and hence the individual expectations are well defined. Since the qn (f ) are nondecreasing, so are the Em (qn (f )) and this sequence must either converge to a finite limit or grow without bound, in which case we say it converges to ∞. In both cases the expectation Em f is well defined, although it may be infinite. If f is an arbitrary real random variable, define its positive and negative parts f + (x) = max(f (x), 0) and f − (x) = − min(f (x), 0) so that f (x) = f + (x) − f − (x) and set Em f = Em f + − Em f −
(1.24)
provided this does not have the form +∞ − ∞, in which case the expectation does not exist. It can be shown that the expectation can also be evaluated for nonnegative measurements by the formula Em f =
sup Em g. discrete g: g≤f
The expectation is also called an integral and is denoted by any of the following: Z Z Z Em f = f dm = f (x)dm(x) = f (x)m(dx). The subscript m denoting the measure with respect to which the expectation is taken will occasionally be omitted if it is clear from context. A measurement f is said to be integrable or m-integrable if Em f exists and is finite. A function is integrable if and only if its absolute value is
1.5 Expectation
15
integrable. Define L1 (m) to be the space of all m-integrable functions. Given any m-integrable f and an event B, define Z Z f dm = f (x)1B (x) dm(x). B
Two random variables f and g are said to be equal m-almost-everywhere or equal m-a.e. or equal with m-probability one if m(f = g) = m({x : f (x) = g(x)}) = 1. The m- is dropped if it is clear from context. Given a probability space (Ω, B, m), suppose that G is a sub-σ -field of B, that is, it is a σ -field of subsets of Ω and all those subsets are in B (G ⊂ B). Let f : Ω → R be an integrable measurement. Then the conditional expectation E(f |G) is described as any function, say h(ω), that satisfies the following two properties: h(ω) is measurable with respect to G Z
(1.25)
Z h dm = G
f dm; all G ∈ G.
(1.26)
G
If a regular conditional probability distribution given G exists, e.g., if the space is standard, then one has a constructive definition of conditional expectation: E(f |G)(ω) is simply the expectation of f with respect to the conditional probability measure m(.|G)(ω). Applying this to the example of two random variables X and Y with standard alphabets described in Section 1.2 we have from (1.26) that for integrable f : AX × A Y → R Z Z Z E(f ) = f (x, y)dPXY (x, y) = ( f (x, y)dPX|Y (x|y))dPY (y). (1.27) In particular, for fixed y, f (x, y) is an integrable (and measurable) function of x. Equation (1.27) provides a generalization of (1.13) from rectangles to arbitrary events. For an arbitrary F ∈ BAX ×AY we have that Z Z Z PXY (F ) = 1F (x, y)dPX|Y (x|y) dPY (y) = PX|Y (Fy |y)dPY (y), (1.28) where Fy = {x : (x, y) ∈ F } is called the section of F at y. If F is measurable, then so is Fy for all y. Alternatively, since 1F (x, y) is measurable with respect to x for each fixed y, Fy ∈ BAX and the inner integral is just Z x:(x,y)∈F
dPX|Y (x|y) = PX|Y (Fy |y).
16
1 Information Sources
1.6 Asymptotic Mean Stationarity Recall that a dynamical system (or the associated source) (Ω, B, P , T ) is said to be stationary if P (T −1 G) = P (G) for all G ∈ B. It is said to be asymptotically mean stationary or, simply, AMS if the limit n−1 1 X P (T −k G) n→∞ n k=0
P (G) = lim
(1.29)
exists for all G ∈ B. The following theorems summarize several important properties of AMS sources. Details may be found in Chapter 6 of [55] or Chapter 7 of [58]. Theorem 1.2. If a dynamical system (Ω, B, P , T ) is AMS, then P defined in (1.29) is a probability measure and (Ω, B, P , T ) is stationary. The distribution P is called the stationary mean of P . If an event G is invariant in the sense that T −1 G = G, then P (G) = P (G). If a random variable g is invariant in the sense that g(T x) = g(x) with P probability 1, then EP g = EP g. The stationary mean P asymptotically dominates P in the sense that if P (G) = 0, then lim sup P (T −n G) = 0. n→∞
Theorem 1.3. Given an AMS source {Xn } let σ (Xn , Xn+1 , · · · ) denote the σ -field generated by the random variables Xn , · · · , that is, the smallest σ -field with respect to which all these random variables are measurable. Define the tail σ -field F∞ by F∞ =
∞ \
σ (Xn , · · · ).
n=0
If G ∈ F∞ and P (G) = 0, then also P (G) = 0. The tail σ -field can be thought of as events that are determinable by looking only at samples of the sequence in the arbitrarily distant future. The theorem states that the stationary mean dominates the original measure on such tail events in the sense that zero probability under the stationary mean implies zero probability under the original source.
1.7 Ergodic Properties
17
1.7 Ergodic Properties Two of the basic results of ergodic theory that will be called upon extensively are the pointwise or almost-everywhere ergodic theorem and the ergodic decomposition theorem We quote these results along with some relevant notation for reference. Detailed developments may be found in Chapters 6–8 of [55] or Chapters 7–10 of [58]. The ergodic theorem states that AMS dynamical systems (and hence also sources) have convergent sample averages, and it characterizes the limits. Theorem 1.4. If a dynamical system (Ω, B, m, T ) is AMS with stationary mean m and if f ∈ L1 (m), then with probability one under m and m n−1 1 X f T i = Em (f |I), n→∞ n i=0
lim
where I is the sub-σ -field of invariant events, that is, events G for which T −1 G = G. The basic idea of the ergodic decomposition is that any stationary source which is not ergodic can be represented as a mixture of stationary ergodic components or subsources. Theorem 1.5. Ergodic Decomposition Given the standard sequence space (Ω, B) with shift T as previously, there exists a family of stationary ergodic measures {px ; x ∈ Ω}, called the ergodic decomposition, with the following properties: (a)pT x = px . (b)For any stationary measure m, Z m(G) = px (G)dm(x); all G ∈ B. (c) For any g ∈ L1 (m) Z
Z Z gdm =
gdpx dm(x).
It is important to note that the same collection of stationary ergodic components works for any stationary measure m. This is the strong form of the ergodic decomposition. The final result of this section is a variation on the ergodic decomposition. To describe the result, we need to digress briefly to introduce a metric on spaces of probability measures. A thorough development can be found in Chapter 8 of [55] or Chapter 9 of [58]. We have a standard sequence measurable space (Ω, B) and hence we can generate the σ -field B
18
1 Information Sources
by a countable field F = {Fn ; n = 1, 2, · · · }. Given such a countable generating field, a distributional distance between two probability measures p and m on (Ω, B) is defined by d(p, m) =
∞ X
2−n |p(Fn ) − m(Fn )|.
n=1
Any choice of a countable generating field yields a distributional distance. Such a distance or metric yields a measurable space of probability measures as follows: Let Λ denote the space of all probability measures on the original measurable space (Ω, B). Let B(Λ) denote the σ -field of subsets of Λ generated by all open spheres using the distributional distance, that is, all sets of the form {p : d(p, m) ≤ } for some m ∈ Λ and some > 0. We can now consider properties of functions that carry sequences in our original space into probability measures. The following is Theorem 8.5.1 of [55] and Theorem 10.1 of [58]. Theorem 1.6. A Variation on the Ergodic Decomposition Fix a standard measurable space (Ω, B) and a transformation T : Ω → Ω. Then there are a standard measurable space (Λ, L), a family of stationary ergodic measures {mλ ; λ ∈ Λ} on (Ω, B), and a measurable mapping ψ : Ω → Λ such that (a)ψ is invariant (ψ(T x) = ψ(x) all x); (b)if m is a stationary measure on (Ω, B) and Pψ is the induced distribution; that is, Pψ (G) = m(ψ−1 (G)) for G ∈ Λ (which is well defined from (a)), then Z Z m(F ) = dm(x)mψ(x) (F ) = dPψ (λ)mλ (F ), all F ∈ B, R and if f ∈ L1 (m), then so is f dmλ Pψ -a.e. and Z Z Em f = dm(x)Emψ(x) f = dPψ (λ)Emλ f . Finally, for any event F , mψ (F ) = m(F |ψ), that is, given the ergodic decomposition and a stationary measure m , the ergodic component λ is a version of the conditional probability under m given ψ = λ. The following corollary to the ergodic decomposition is Lemma 8.6.2 of [55] and Lemma 10.4 of [58]. It states that the conditional probability of a future event given the entire past is unchanged by knowing the ergodic component in effect. This is because the infinite past determines the ergodic component in effect. Corollary 1.1. Suppose that {Xn } is a two-sided stationary process with distribution m and that {mλ ; λ ∈ Λ} is the ergodic decomposition and ψ
1.7 Ergodic Properties
19
the ergodic component function. Then the mapping ψ is measurable with respect to σ (X−1 , X−2 , · · · ) and m((X0 , X1 , · · · ) ∈ F |X−1 , X−2 , · · · ) = mψ ((X0 , X1 , · · · ) ∈ F |X−1 , X−2 , · · · ); m − a.e.
Chapter 2
Pair Processes: Channels, Codes, and Couplings
Abstract We have considered a random process or source {Xn } as a sequence of random entities, where the object produced at each time could be quite general, e.g., a random variable, vector, or waveform. Hence sequences of pairs of random objects such as {Xn , Yn } are included in the general framework. We now focus on the possible interrelations between the two components of such a pair process. First consider the situation where we begin with one source, say {Xn }, called the input and use either a random or a deterministic mapping of the input sequence {Xn } to form an output sequence {Yn }. We generally refer to the mapping as a channel if it is random and a code if it is deterministic. Hence a code is a special case of a channel and results for channels will immediately imply corresponding results for codes. The initial point of interest will be conditions on the structure of the channel under which the resulting pair process {Xn , Yn } will inherit stationarity and ergodic properties from the original source {Xn }. We will also be interested in the behavior resulting when the output of one channel serves as the input to another, that is, when we form a new channel as a cascade of other channels. Such cascades yield models of a communication system which typically has a code mapping (called the encoder) followed by a channel followed by another code mapping (called the decoder). Lastly, pair processes arise naturally in other situations, including coupling two separate processes by constructing a joint distribution. This chapter develops the context for the development in future chapters of the properties of information and entropy arising in pair processes.
2.1 Pair Processes A common object throughout this book and the focus of this chapter is the idea of a pair process. The notation will vary somewhat dependR.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_2, © Springer Science+Business Media, LLC 2011
21
22
2 Pair Processes: Channels, Codes, and Couplings
ing on the specific application, but basically a pair process is a random process with two components, e.g., a sequence of random variables {(Xn , Yn ); n ∈ T} with alphabets A and B and process distribution p on (AT × B T , B(AT × B T ). When we wish to emphasize the names of the separate component random variables and processes, we will often write AX for A and AY for B and pXY for p. The process X or {Xn } will often have the interpretation of being the input of a code or channel or a cascade of such operations and Y or {Yn } the output. A pair process induces two “marginal” process {Xn } with process distribution, say µ, and {Yn }, with process distribution η. When we wish to emphasize the random variables we might write pX or µX instead of µ and pY or µY or ηY instead of η. All of these notations have their uses, and the added subscripts often help sort out which random process or variables are important. Often we will ˆn as the second component instead of Y when it is viewed as an use X approximation to the first component Xn .
2.2 Channels A channel converts one information source – typically called the input to the channel – into another – called the output. In general the operation is random and is specified by a conditional probability measure of output sequences given an input sequence. The combination of an input distribution with the channel yields a pair process, a process with an input component and an output component. If the channel is deterministic rather than random, the operation is called a code. In this section the basic definitions of channels and codes are introduced. A fundamental nuisance in the development of channels and codes is the notion of time. So far we have considered pair processes where at each unit of time, one random object is produced for each coordinate of the pair. In the channel or code example, this corresponds to one output for every input. Interesting communication systems do not always easily fit into this framework, and this can cause serious problems in notation and in the interpretation and development of results. For example, suppose that an input source consists of a sequence of real numbers and let T denote the time shift on the real sequence space. Suppose that the output source consists of a binary sequence and let S denote its shift. Suppose also that the channel is such that for each real number in, three binary symbols are produced. This fits our usual framework if we consider each output variable to consist of a binary three-tuple since then there is one output vector for each input symbol. One must be careful, however, when considering the stationarity of such a system. Do we consider the output process to be physically stationary if it is stationary with respect to S or with respect to S 3 ? The former might make more
2.2 Channels
23
sense if we are looking at the output alone, the latter if we are looking at the output in relation to the input. How do we define stationarity for the pair process? Given two sequence spaces, we might first construct a shift on the pair sequence space as simply the cartesian product of the shifts, e.g., given an input sequence x and an output sequence y define a shift T ∗ by T ∗ (x, y) = (T x, Sy). While this might seem natural given only the pair random process {Xn , Yn }, it is not natural in the physical context that one symbol of X yields three symbols of Y . In other words, the two shifts do not correspond to the same amount of time. Here the more physically meaningful shift on the pair space would be T 0 (x, y) = (T x, S 3 y) and the more physically meaningful questions on stationarity and ergodicity relate to T 0 and not to T ∗ . The problem becomes even more complicated when channels or codes produce a varying number of output symbols for each input symbol, where the number of symbols depends on the input sequence. Such variable rate codes arise often in practice, especially for noiseless coding applications such as Huffman, Lempel-Ziv, and arithmetic codes. While we will not treat such variable rate systems in any detail, they point out the difficulty that can arise associating the mathematical shift operation with physical time when we are considering cartesian products of spaces, each having their own shift. There is no easy way to solve this problem notationally. We adopt the following view as a compromise which is usually adequate for fixed-rate systems. We will be most interested in pair processes that are stationary in the physical sense, that is, whose statistics are not changed when both are shifted by an equal amount of physical time. This is the same as stationarity with respect to the product shift if the two shifts correspond to equal amounts of physical time. Hence for simplicity we will usually focus on this case. More general cases will be introduced when appropriate to point out their form and how they can be put into the matching shift structure by considering groups of symbols and different shifts. This will necessitate occasional discussions about what is meant by stationarity or ergodicity for a particular system. The mathematical generalization of Shannon’s original notions of sources, codes, and channels are due to Khinchine [87] [88]. Khinchine’s results characterizing stationarity and ergodicity of channels were corrected and developed by Adler [2]. Say we are given a source [A, X, µ], that is, a sequence of A-valued random variables {Xn ; n ∈ T} defined on a common probability space (Ω, F , P ) having a process distribution µ defined on the measurable sequence space (B T , BA T ). We shall let X = {Xn ; n ∈ T} denote the sequence-valued random variable, that is, the random variable taking values in AT according to the distribution µ. Let B be another alphabet with a corresponding measurable sequence space (AT , BB T ). We assume as usual that A and B are standard and hence so are their sequence
24
2 Pair Processes: Channels, Codes, and Couplings
spaces and cartesian products. A channel [A, ν, B] with input alphabet A and output alphabet B (we denote the channel simply by ν when these alphabets are clear from context) is a family of probability measures {νx ; x ∈ AT } on (B T , BB T ) (the output sequence space) such that for every output event F ∈ BB T νx (F ) is a measurable function of x. This measurability requirement ensures that the set function p specified on the joint input/output space (AT × B T ), BA T × BB T ) by its values on rectangles as Z p(G × F ) = dµ(x)νx (F ); F ∈ BB T , G ∈ BA T , G
is well defined. The set function p is nonnegative, normalized, and countably additive on the field generated by the rectangles G × F , G ∈ BA T , F ∈ BB T . Thus p extends to a probability measure on the joint input/output space, which is sometimes called the hookup of the source µ and channel ν. We will often denote this joint measure by µν. The corresponding sequences of random variables are called the input/output process. Thus a channel is a probability measure on the output sequence space for each input sequence such that a joint input/output probability measure is well-defined. The above equation shows that a channel is simply a regular conditional probability, in particular, νx (F ) = p((x, y) : y ∈ F |x); F ∈ BB T , x ∈ AT . We can relate a channel to the notation used previously for conditional distributions by using the sequence-valued random variables X = {Xn ; n ∈ T} and Y = {Yn ; n ∈ T}: νx (F ) = PY |X (F |x).
(2.1)
Eq. (1.28) then provides the probability of an arbitrary input/output event: Z p(F ) = dµ(x)νx (Fx ), where Fx = {y : (x, y) ∈ F } is the section of F at x. If we start with a hookup p, then we can obtain the input distribution µ as µ(F ) = p(F × B T ); F ∈ BA T . Similarly we can obtain the output distribution, say η, via η(F ) = p(AT × F ); F ∈ BB T . Suppose one now starts with a pair process distribution p and hence also with the induced source distribution µ. Does there exist a channel ν
2.3 Stationarity Properties of Channels
25
for which p = µν? The answer is yes since the spaces are standard. One can always define the conditional probability νx (F ) = P (F × AT |X = x) for all input sequences x, but this need not possess a regular version, that is, be a probability measure for all x, in the case of arbitrary alphabets. If the alphabets are standard, however, we have seen that a regular conditional probability measure always exists.
2.3 Stationarity Properties of Channels We now define a variety of stationarity properties for channels that are related to, but not the same as, those for sources. The motivation behind the various definitions is that stationarity properties of channels coupled with those of sources should imply stationarity properties for the resulting source-channel hookups. The classical definition of a stationary channel is the following: Suppose that we have a channel [A, ν, B] and suppose that TA and TB are the shifts on the input sequence space and output sequence space, respectively. The channel is stationary with respect to TA and TB or (TA , TB )stationary if νx (TB−1 F ) = νTA x (F ), x ∈ AT , F ∈ BB T . (2.2) If the transformations are clear from context then we simply say that the channel is stationary. Intuitively, a right shift of an output event yields the same probability as the left shift of an input event. The different shifts are required because in general only TA x and not TA−1 x exists since the shift may not be invertible and in general only TB−1 F and not TB F exists for the same reason. If the shifts are invertible, e.g., the processes are two-sided, then the definition is equivalent to νTA x (TB F ) = νTA−1 x (TB−1 F ) = νx (F ), all x ∈ AT , F ∈ BB T
(2.3)
that is, shifting the input sequence and output event in the same direction does not change the probability. The fundamental importance of the stationarity of a channel is contained in the following lemma. Lemma 2.1. If a source [A, µ], stationary with respect to TA , is connected to channel [A, ν, B], stationary with respect to TA and TB , then the resulting hookup µν is stationary with respect to the cartesian product shift T = TA×B = TA × TB defined by T (x, y) = (TA x, TB y). Proof: We have that µν(T −1 F ) =
Z
dµ(x)νx ((T −1 F )x ).
26
2 Pair Processes: Channels, Codes, and Couplings
Now (T −1 F )x = {y : T (x, y) ∈ F } = {y : (TA x, TB y) ∈ F } = {y : TB y ∈ FTA x } = TB−1 FTA x and hence µν(T −1 F ) =
Z
dµ(x)νx (TB−1 FTA x ).
Since the channel is stationary, however, this becomes Z Z µν(T −1 F ) = dµ(x)νTA x (FTA x ) = dµTA−1 (x)νx (Fx ), where we have used the change of variables formula. Since µ is stationary, however, the right hand side is Z dµ(x)νx (F ), which proves the lemma.
2
Suppose next that we are told that a hookup µν is stationary. Does it then follow that the source µ and channel ν are necessarily stationary? The source must be since µ(TA−1 F ) = µν((TA × TB )−1 (F × B T )) = µν(F × B T ) = µ(F ). The channel need not be stationary, however, since, for example, the stationarity could be violated on a set of µ measure 0 without affecting the proof of the above lemma. This suggests a somewhat weaker notion of stationarity which is more directly related to the stationarity of the hookup. We say that a channel [A, ν, B] is stationary with respect to a source [A, µ] if µν is stationary. We also state that a channel is stationary µ-a.e. if it satisfies (2.2) for all x in a set of µ-probability one. If a channel is stationary µ-a.e. and µ is stationary, then the channel is also stationary with respect to µ. Clearly a stationary channel is stationary with respect to all stationary sources. The reason for this more general view is that we wish to extend the definition of stationary channels to asymptotically mean stationary channels. The general definition extends; the classical definition of stationary channels does not. Observe that the various definitions of stationarity of channels immediately extend to block shifts since they hold for any shifts defined on the input and output sequence spaces, e.g., a channel stationary with respect to TAN and TBK could be a reasonable model for a channel or code that puts out K symbols from an alphabet B every time it takes in N symbols from an alphabet A. We shorten the name (TAN , TBK )-stationary
2.3 Stationarity Properties of Channels
27
to (N, K)-stationary channel in this case. A stationary channel (without modifiers) is simply a (1,1)-stationary channel in this sense. The most general notion of stationarity that we are interested in is that of asymptotic mean stationarity We define a channel [A, ν, B] to be asymptotically mean stationary or AMS for a source [A, µ] with respect to TA and TB if the hookup µν is AMS with respect to the product shift TA × TB . As in the stationary case, an immediate necessary condition is that the input source be AMS with respect to TA . A channel will be said to be (TA , TB )-AMS if the hookup is (TA , TB )-AMS for all TA -AMS sources. The following lemma shows that an AMS channel is indeed a generalization of the idea of a stationary channel and that the stationary mean of a hookup of an AMS source to a stationary channel is simply the hookup of the stationary mean of the source to the channel. Lemma 2.2. Suppose that ν is (TA , TB )-stationary and that µ is AMS with respect to TA . Let µ denote the stationary mean of µ and observe that µν is stationary. Then the hookup µν is AMS with stationary mean µν = µν. Thus, in particular, ν is an AMS channel. Proof: We have that (T −i F )x = {y : (x, y) ∈ T −i F } = {y : T i (x, y) ∈ F } = {y : (TAi x, TBi y) ∈ F } = {y : TBi y ∈ FT i x } A
=
TB−i FT i x A
and therefore since ν is stationary Z µν(T −i F ) = dµ(x)νx (TB−i FT i x ) A Z Z = dµ(x)νT i x (FT i x ) = dµTA−i (x)νx (F ). A
A
Therefore n−1 1 X µν(T −i F ) = n i=0
n−1 Z 1 X dµTA−i (x)νx (F ) n i=0 Z → dµ(x)νx (F ) = µν(F )
n→∞
from Lemma 6.5.1 of [55] or Lemma 7.9 if [58]. This proves that µν is AMS and that the stationary mean is µν. 2 A final property crucial to quantifying the behavior of random processes is that of ergodicity. Hence we define a (stationary, AMS) channel
28
2 Pair Processes: Channels, Codes, and Couplings
ν to be ergodic with respect to (TA , TB ) if it has the property that whenever a (stationary, AMS) ergodic source (with respect to TA ) is connected to the channel, the overall input/output process is (stationary, AMS) ergodic. The following modification of Lemma 6.7.4 of [55] or Lemma 7.15 of [58] is the principal tool for proving a channel to be ergodic. Lemma 2.3. An AMS (stationary) channel [A, ν, B] is ergodic if for all AMS (stationary) sources µ and all sets of the form F = FA × FB , G = GA × GB ∞ for rectangles FA , GA ∈ B∞ A and FB , GB ∈ BB we have that for p = µν n−1 \ 1 X −i p(TA×B F G) = p(F )p(G), n→∞ n i=0
lim
(2.4)
where p is the stationary mean of p (p if p is already stationary). Proof: The proof parallels that of Lemma 6.7.4 of [55] or Lemma 7.15 of [58]. The result does not follow immediately from that lemma since the collection of given sets does not itself form a field. Arbitrary events F , G ∈ B∞ A×B can be approximated arbitrarily closely by events in the field generated by the above rectangles and hence given > 0 we can find finite disjoint rectanglesS of the given form Fi , Gi , i = 1, · · · , L such SL L that if F0 = i=1 Fi and G0 = i=1 Gi , then p(F ∆F0 ), p(G∆G0 ), p(F ∆F0 ), and p(G∆G0 ) are all less than . Then
|
n−1 \ 1 X p(T −k F G) − p(F )p(G)| ≤ n k=0
|
|
n−1 n−1 \ \ 1 X 1 X p(T −k F G) − p(T −k F0 G0 )|+ n k=0 n k=0
n−1 \ 1 X p(T −k F0 G0 ) − p(F0 )p(G0 )| + |p(F0 )p(G0 ) − p(F )p(G)|. n k=0
Exactly as in Lemma 6.7.4 of [55], the rightmost term is bound above by 2 and the first term on the left goes to zero as n → ∞. The middle term is the absolute magnitude of n−1 [ [ [ \[ 1 X p(T −k Fi Gj ) − p( Fi )p( Gj ) = n k=0 i j i j X \ X 1 n−1 p(T −k Fi Gj ) − p(Fi )p(Gj ) . n k=0 i,j
Each term in the finite sum converges to 0 by assumption. Thus p is ergodic from Lemma 6.7.4 of [55] or Lemma 7.15 of [58]. 2
2.4 Extremes: Noiseless and Completely Random Channels
29
Because of the specific class of sets chosen, the above lemma considered separate sets for shifting and remaining fixed, unlike using the same set for both purposes as in Lemma 6.7.4 of [55] or Lemma 7.15 of [58]. This was required so that the cross products in the final sum considered would converge accordingly.
2.4 Extremes: Noiseless and Completely Random Channels The first two examples of channels are the simplest, the first doing nothing to the input but reproducing it perfectly and the second being useless (at least for communication purposes) since the output is random and independent of the input. Both extremes provide simple examples of the properties of channels, and the completely random example will reappear when applying channel structure ideas to sources.
Noiseless Channels A channel [A, ν, B] is said to be noiseless if A = B and ( 1 x∈F νx (F ) = 0 x 6∈ F that is, with probability one the channel puts out what goes in, it acts as an ideal wire. In engineering terms, it is discrete-time linear system with impulse response equal to an impulse. A noiseless channel is clearly stationary and ergodic.
Completely Random Channels Suppose that η is a probability measure on the output space (B T , BB T ) and define a channel νx (F ) = η(F ), F ∈ BB T , x ∈ AT . Then it is easy to see that the input/output measure satisfies p(G × F ) = η(F )µ(G); F ∈ BB T , G ∈ BA T , and hence the input/output measure is a product measure and the input and output sequences are therefore independent of each other. This
30
2 Pair Processes: Channels, Codes, and Couplings
channel is called a completely random channel or product channel because the output is independent of the input. This channel is quite useless because the output tells us nothing of the input. The completely random channel is stationary (AMS) if the measure η is stationary (AMS). Perhaps surprisingly, such a channel need not be ergodic even if η is ergodic since the product of two stationary and ergodic sources need not be ergodic. (See, e.g., [22].) We shall later see that if η is also assumed to be weakly mixing, then the resulting channel is ergodic. A generalization of the noiseless channel that is of much greater interest is the deterministic channel. Here the channel is not random, but the output is formed by a general mapping of the input rather than being the input itself.
2.5 Deterministic Channels and Sequence Coders A channel [A, ν, B] is said to be deterministic if each input string x is mapped into an output string f (x) by a measurable mapping f : AT → B T . The conditional probability defining the channel is ( νx (G) =
1
f (x) ∈ G
f (x) 6∈ G.
Note that such a channel can also be written as νx (G) = 1f −1 (G) (x). A sequence coder is a deterministic channel, that is, a measurable mapping from one sequence space into another. It is easy to see that for a deterministic code the hookup is specified by \ p(F × G) = µ(F f −1 (G)) and the output process has distribution η(G) = µ(f −1 (G)). A sequence coder is said to be (TA , TB )-stationary (or just stationary) or (TAN , TBK )-stationary (or just (N, K)-stationary) if the corresponding channel is. Thus a sequence coder f is stationary if and only if f (TA x) = TB f (x) and it is (N, K)- stationary if and only if f (TAN x) = TBK f (x). Lemma 2.4. A stationary deterministic channel is ergodic. Proof: From Lemma 2.3 it suffices to show that
2.6 Stationary and Sliding-Block Codes
31
n−1 \ 1 X −i p(TA×B F G) = p(F )P (G) n→∞ n i=0
lim
for all rectangles of the form F = FA × FB , FA ∈ BB T , FB ∈ BA T and G = GA × GB . Then \ \ \ −i F G) = p((TA−i FA p(TA×B GA ) × (TB−i FB GB )) \ \ \ = µ((TA−i FA GA ) f −1 (TB−i FB GB )). Since f is stationary and since inverse images preserve set theoretic operations, \ \ f −1 (TB−i FB GB ) = TA−i f −1 (FB ) f −1 (GB ) and hence n−1 \ 1 X −i p(TA×B F G) = n i=0
n−1 \ \ \ 1 X µ(TA−i (FA f −1 (FB )) GA f −1 (GB )) n i=0 \ \ → µ(FA f −1 (FB ))µ(GA f −1 (GB ))
n→∞
= p(FA × FB )p(GA × GB ) since µ is ergodic. This means that the rectangles meet the required condition. Some algebra then will show that finite unions of disjoint sets meeting the conditions also meet the conditions and that complements of sets meeting the conditions also meet them. This implies from the good sets principle (see, for example, p. 14 of [55] or p. 50 in [58]) that the field generated by the rectangles also meets the condition and hence the lemma is proved. 2
2.6 Stationary and Sliding-Block Codes A stationary deterministic channel is also called a stationary code, so it follows that the output of a stationary code with a stationary input process is also stationary. A stationary code has a simple and useful structure. Suppose one has a mapping f : AT → B, that is, a mapping that maps an input sequence into a single output symbol. We can define a complete output sequence y corresponding to an input sequence x by yn = f (TAn x); n ∈ T,
(2.5)
that is, we produce an output, then shift or slide the input sequence by one time unit, and then we produce another output using the same function, and so on. A mapping of this form is called a sliding-block code
32
2 Pair Processes: Channels, Codes, and Couplings
because it produces outputs by successively sliding an infinite-length input sequence and each time using a fixed mapping to produce the output. The sequence-to-symbol mapping implies a sequence coder, say f , defined by f (x) = {f (TAn x); n ∈ T}. Furthermore, f (TA x) = TB f (x), that is, a sliding-block code induces a stationary sequence coder. Conversely, any stationary sequence coder f induces a sliding-block code f for which (2.5) holds by the simple identification f (x) = (f (x))0 , the output at time 0 of the sequence coder. Thus the ideas of stationary sequence coders mapping sequences into sequences and sliding-block codes mapping sequences into letters by sliding the input sequence are equivalent. We can similarly define an (N, K)-sliding-block code which is a mapping f : AT → B K which forms an output sequence y from an input sequence x via the construction K ynK = f (TANn x).
By a similar argument, (N, K)-sliding-block coders are equivalent to (N, K)-stationary sequence coders. When dealing with sliding-block codes we will usually assume for simplicity that K is 1. This involves no loss in generality since it can be made true by redefining the output alphabet. The following stationarity property of sliding-block codes follows from the properties for stationary channels, but the proof is given for completeness. Lemma 2.5. If f is a stationary coding of an AMS process, then the process {fn = f T n } is also AMS. If the input process is ergodic, then so is {fn }. Proof: Suppose that the input process has alphabet AX and distribution P and that the measurement f has alphabet Af . Define the sequence ∞ mapping f : A∞ X → Af by f (x) = {fn (x); n ∈ T}, where fn (x) = f (T n x) and T is the shift on the input sequence space A∞ X . If T also denotes the shift on the output space, then by construction f (T x) = −1
−1
T f (x) and hence for any output event F , f (T −1 F ) = T −1 f (F ). Let m denote the process distribution for the encoded process. Since m(F ) = −1
P (f (F )) for any event F ∈ B(Af )∞ , we have using the stationarity of the mapping f that n−1 n−1 −1 1 X 1 X m(T −i F ) = lim P (f (T −i F )) n→∞ n n→∞ n i=0 i=0
lim
n−1 −1 −1 1 X P (T −i f (F )) = P (f (F )), n→∞ n i=0
= lim
where P is the stationary mean of P . Thus m is AMS. If G is an invariant output event, then f
−1
(G) is also invariant since T −1 f
−1
(G) =
2.6 Stationary and Sliding-Block Codes
33
−1
f (T −1 G). Hence if input invariant sets can only have probability 1 or 0, the same is true for output invariant sets. 2
Finite-length Sliding-Block Codes Stationary or sliding-block codes have a simple description when the sequence-to-symbol mapping characterizing the code depends on only a finite number of the sequence values; that is, the mapping is measurable with respect to a finite number of coordinates. As a particularly simple example, consider the code depicted in Figure 2.1, where an IID process {Zn } consisting of equiprobable coin flips is shifted into a length 3 shift register at the completion of the shift the table is used to produce one output value given the three binary numbers in the shift register. For
{Zn }
- Zn
Zn−1
Zn−2
shift register
@ R ? @ -Yn = φ(Zn , Zn−1 , Zn−2 ) φ function, table
Zn Zn−1 Zn−2 000 001 010 011 100 101 110 111
Yn 0.7683 -0.4233 -0.1362 1.3286 0.4233 0.1362 -1.3286 -0.7683
Fig. 2.1 A length 3 stationary code
the curious, this simple code tries to map coin flips into a sequence that looks approximately Gaussian. The output values correspond to eight possible values of an inverse cdf for a 0 mean Gaussian random variable with variance 3/4 evaluated at 8 equally spaced points in the unit interval. The values are “scrambled” to reduce correlation, but the marginal distribution is an approximation to Shannon optimal distribution when simulating or source coding an IID Gaussian sequence with mean 0 and variance 1. All of these ideas will be encountered later in the book. More generally, suppose that we consider two-sided processes and that we have a measurable mapping D
φ : × Ai → B i=−M
and we define a sliding-block code by f (x) = φ(x−M , · · · , x0 , · · · , xD ),
34
2 Pair Processes: Channels, Codes, and Couplings
so that the output process is Yn = φ(Xn−M , · · · , Xn , · · · , Xn+D ), a mapping of the contents of a shift register as depicted in Figure 2.2. Note that the time order is reversed in the shift-register representa-
···
Xn- Xn+D
Xn
···
Xn−M
XXX XXX ? z φ X )
XX
? Yn = f (Xn+D , · · · , Xn , · · · , Xn−M ) Fig. 2.2 Stationary or sliding-block code
tion since in the shift register new input symbols flow in from the left and exit from the right, but the standard way of writing a sequence is · · · , Xn−2 , Xn−1 , Xn , Xn+1 , Xn+2 , · · · with “past” symbols on the left and “future” symbols on the right. The standard shift is the left shift so that shifting the above sequence results in the new sequence · · · , Xn−1 , Xn , Xn+1 , Xn+2 , Xn+3 , · · · . Rather than adding to the notational clutter by formally mapping sequences or vectors into a reversedtime form, we shall suffer the minor abuse of notation and follow tradition by using the first format (time increases to the left) for shiftregisters, and the second notation (time increases to the right) when dealing with theory and stationary mappings. Context should make the usage clear and clarification will be added when necessary. The length of the code is the length of the shift register or dimension of the vector argument, L = D + M + 1. The mapping φ induces a sequence-to-symbol mapping f and a corresponding stationary sequence coder f . The mapping φ is also called a sliding-block code or a finite-length sliding-block code or a finite-window sliding-block code. M is called the memory of the code and D is called the delay of the code since M past source symbols and D future symbols are required to produce the current output symbol. The window length or constraint length of the code is M + D + 1, the number of input symbols viewed to produce an output symbol. If D = 0 the code is said to be causal. If M = 0 the code is said to be memoryless.
2.6 Stationary and Sliding-Block Codes
35
There is a problem with the above model if we wish to code a onesided source since if we start coding at time 0, there are no input symbols with negative indices. Hence we either must require the code be memoryless (M = 0) or we must redefine the code for the first M instances (e.g., by “stuffing” the code register with arbitrary symbols) or we must only define the output for times i ≥ M. For two-sided sources a finite-length sliding-block code is stationary. In the one-sided case it is not even defined precisely unless it is memoryless, in which case it is stationary. While codes that depend on infinite input sequences may not at first glance seem to be a reasonable physical model of a coding system, it is possible for such codes to depend on the infinite sequence only through a finite number of coordinates. In addition, some real codes may indeed depend on an unboundedly large number of past inputs because of feedback.
Sliding-Block Codes and Partitions Codes mapping sequences (or vectors) into discrete alphabets have an alternative representation in terms of partitions and range spaces or codebooks. Given a sliding-block code f : A∞ → B where B is discrete, suppose that we index the members of the set B as B = {bi ; i ∈ I} where I is a finite or infinite collection of positive integers. Since codes are assumed to be measurable mappings, the sets Pi = {x : x ∈ A∞ : f (x) = bi } = f −1 (bi ), i ∈ I, collectively form a measurable partition P = {Pi , i ∈ I} of A∞ ; that is, they are disjoint and collectively exhaustive. The sets Pi are referred to as the atoms of the partition. The range space B = {bi ; i ∈ I} is called the codebook of the code f or output alphabet and it will be assumed without loss of generality that its members are distinct. The code f can be expressed in terms of its partition and codebook by X f (x) = bi 1Pi (x), (2.6) i
where 1P (x) is the indicator function for a set P . Conversely, given a partition and a codebook, (2.6) describes the corresponding code.
B-Processes One use of sliding-block codes is to provide an easy yet powerful generalization of the simplest class of random processes. IID random pro-
36
2 Pair Processes: Channels, Codes, and Couplings
cesses provide the simplest nontrivial example of a random process, typically the first example of a random process encountered in introductory courses is that of sequence of coin flips or rolls of a die. IID processes have no memory and are generally the easiest example to analyze. Stationary or sliding-block coding of an IID process preserves many of the most useful properties of the IID process, including stationarity, ergodicity, and mixing. In addition to providing a common mathematical model of many real processes, processes formed this way turn out to be one of the most important classes of processes in ergodic theory in a way that is relevant to this book — the class of stationary codings of IID processes is exactly the class of random processes for which equal entropy rate is both necessary and sufficient for two processes to be isomorphic in the sense that one can be coded by a stationary code into the other in an invertible way. This result is Ornstein’s isomorphism theorem, a result far beyond the scope of this book. But the importance of the class was first recognized in ergodic theory, and adds weight to its emphasis in this presentation of entropy and information theory. A process is said to be a B-process if it can be represented as a finitealphabet stationary coding of an independent identically distributed (IID) process, where the IID process need not have a finite alphabet. Such processes are also called or Bernoulli processes in ergodic theory, but in information theory that name usually implies IID processes (often binary) and not the more general case of any stationary coding of an IID process, so here the name B-process will be used exclusively. The definition also extends to continuous alphabet processes, for example a stationary Gaussian autoregressive processes is also a B-process since it can be represented as the result of passing an IID Gaussian process through a stable autoregressive filter, which is a stationary mapping [173]. The emphasis here, however, will be on ordinary finite-alphabet B-processes. There are many other characterizations of this class of random processes, but the class of stationary codings of IID processes is the simplest and most suitable for the purposes of this book. Let µ denote the original distribution of the IID process and let η denote the induced output distribution. Then for any output events F and G \ \ \ −1 −1 −1 η(F TB−n G) = µ(f (F TB−n G)) = µ(f (F ) TA−n f (G)), since f is stationary. But µ is stationary and mixing since it is IID (see Section 6.7 of [55] or Section 7.7 of [58]) and hence this probability converges to µ(f
−1
(F ))µ(f
−1
(G)) = η(F )η(G)
and hence η is also mixing. Thus a B-process is mixing of all orders and hence is ergodic with respect to TBn for all positive integers n.
2.7 Block Codes
37
B-processes can be thought of as the most random of random processes since they have at their heart an IID process such as coin flips or dice rolls.
2.7 Block Codes Another case of sequence coding arises when we have a measurable mapping α : AN → B K and we define a sequence coder f (x) = y by N K = (ynK , ynK+1 , · · · , y(n+1)K−1 ) = α(xnN ), ynK
that is, the input is parsed into nonoverlapping blocks of length N and each is successively coded into a block of length K outputs without regard to past or previous input or output blocks. Clearly N input time units must correspond to K output time units in physical time if the code is to make sense. A code of this form is called a block code and it is a special case of an (N, K) sliding block code so that such a code is (TAN , TAK )-stationary.
Block Independent Processes As sliding-block coding of an IID process leads to a more general class of random processes, one can also apply a block code to an IID process to obtain a more general class of random processes including IID processes as a special case (with blocklength = 1). The resulting process will be block independent in the sense that successive K-blocks will be independent since they depend on independent input N blocks. Unlike B-processes, however, the new processes are not in general stationary or ergodic even if the input was. The process can be modified by inserting a random uniformly distributed start time to convert the K-stationary process into a stationary process, but in general ergodicity is lost and sample functions will still exhibit blocking artifacts.
Sliding-Block vs. Block Codes We shall be interested in constructing sliding-block codes from block codes and vice versa. Each has its uses. The random process obtained in the next section by sliding-block coding a stationary and ergodic process
38
2 Pair Processes: Channels, Codes, and Couplings
will provide a key tool in constructing stationary codes and channels from block-stationary ones.
2.8 Random Punctuation Sequences This section develops an example of a sliding-block coding of a stationary and ergodic process to obtain a special random process called a random punctuation sequence which can be used to imbed a block structure into a stationary process in a way that preserves ergodicity and mixing properties in codes and channels. Any stationary and ergodic process can be used in the construction, but if the initial process is a B-process, then the resulting puncutation process will also be a B-process. The results are a a variant of a theorem of Shields and Neuhoff [167] as simplified by Neuhoff and Gilbert [131] for sliding-block codings of finite-alphabet processes. Lemma 2.6. Suppose that {Xn } is a stationary and ergodic process. Then given N and δ > 0 there exists a stationary (or sliding-block) coding f : AT → {0, 1, 2} yielding a ternary process {Zn } with the following properties: (a) {Zn } is stationary and ergodic. (b) {Zn } has a ternary alphabet {0, 1, 2} and it can output only N-cells of the form 011 · · · 1 (0 followed by N − 1 ones) or individual 2’s. In particular, each 0 is always followed by at exactly N − 1 1’s. (c) For all integers k 1−δ 1 ≤ Pr(ZkN = 011 · · · 1) ≤ N N and hence for any n Pr(Zn is in an N − cell) ≥ 1 − δ. Comment: A process {Zn } with these properties is called an (N, δ)random blocking process or punctuation sequence {Zn }. As a visual aid, a segment of a typical punctuation sequence might look like · · · 111][0 |11 ·{z· · 1}][0 11 | ·{z· · 1}][0 |11 ·{z· · 1}][0 11 | ·{z· · 1}]2[0 11 | ·{z· · 1}][0 |11 ·{z· · 1}]
N−11’s N−11’s N−11’s N−11’s N−11’s N−11’s [0 |11 ·{z· · 1}][0 11 · · · 1 ][0 11 · · · 1 ][0 11 · · · 1 ][0 11 · · · 1 ]222[0 11 · · · 1 ]0111 ··· | {z } | {z } | {z } | {z } | {z } N−11’s
N−11’s
N−11’s
N−11’s
N−11’s
N−11’s
with the most of the sequence taken up by N-cells with a few 2’s interspersed.
2.8 Random Punctuation Sequences
39
Proof. A sliding-block coding is stationary and hence coding a stationary and ergodic process will yield a stationary and ergodic process (Lemma 2.4), which proves the first part. Pick an > 0 such that N < δ. Given the stationary and ergodic process {Xn } (that is also assumed to be aperiodic in the sense that it does not place all of its probability on a finite set of sequences) we can find an S event G ∈ BA T having probability N−1 less than . Consider the event F = G − i=1 T −i G, that is, F is the collection of sequences x for which x ∈ G, but T i x 6∈ G for i = 1, · · · , N − 1. We next develop several properties of this set. First observe that obviously µ(F ) ≤ µ(G) and hence µ(F ) ≤ The sequence of sets T −i F are disjoint since if y ∈ T −i F , then T i y ∈ F ⊂ G and T i+l y 6∈ G for l = 1, · · · , N − 1, which means that T j y 6∈ G and hence T j y 6∈ F for N − 1 ≥ j > i. Lastly we need to show that although F may have small probability, S it is not 0. To see this suppose the contrary, N−1 that is, suppose that µ(G − i=1 T −i G) = 0. Then \ N−1 \ N−1 [ [ −i µ(G ( T G)) = µ(G) − µ(G ( T −i G)c ) = µ(G) i=1
i=1
SN−1 and hence µ( i=1 T −i G|G) = 1. In words, if G occurs, then it is certain to occur again within the next N shifts. This means that with probability 1 the relative frequency of G in a sequence x must be no less than 1/N since if it ever occurs (which it must with probability 1), it must thereafter occur at least once every N shifts. This is a contradiction, however, since this means from the ergodic theorem that µ(G) ≥ 1/N when it was assumed that µ(G) ≤ < 1/N. Thus it must hold that µ(F ) > 0. We now use the rare event F to define a sliding-block code. The general idea is simple, but a more complicated detail will be required to handle a special case. Given a sequence x, define n(x) to be the smallest i for which T i x ∈ F ; that is, we look into the future to find the next occurrence of F . Since F has nonzero probability, n(x) will be finite with probability 1. Intuitively, n(x) should usually be large since F has small probability. Once F is found, we code backwards from that point using blocks of a 0 prefix followed by N − 1 1’s. The appropriate symbol is then the output of the sliding block code. More precisely, if n(x) = kN + l, then the sliding-block code prints a 0 if l = 0 and prints a 1 otherwise. This idea suffices until the event F actually occurs at the present time, that is, when n(x) = 0. At this point the sliding-block code has just completed printing an N-cell of 0111 · · · 1. It should not automatically start a new N-cell, because at the next shift it will be looking for a new F in the future to code back from and the new cells may not align with the old cells. Thus the coder looks into the future for the next F ; that is, it again seeks n(x), the smallest i for which T i x ∈ F . This time n(x) must be greater than or equal to N since x is now in F and T −i F are disjoint for i = 1, · · · N − 1. After finding n(x) = kN + l, the coder again codes back
40
2 Pair Processes: Channels, Codes, and Couplings
to the origin of time. If l = 0, then the two codes are aligned and the coder prints a 0 and continues as before. If l 6= 0, then the two codes are not aligned, that is, the current time is in the middle of a new code word. By construction l ≤ N − 1. In this case the coder prints l 2’s (filler poop) and shifts the input sequence l times. At this point there is an n(x) = kN for such that T n(x) x ∈ F and the coding can proceed as before. Note that k is at least one, that is, there is at least one complete cell before encountering the new F . By construction, 2’s can occur only following the event F and then no more than N 2’s can be produced. Thus from the ergodic theorem the relative frequency of 2’s (and hence the probability that Zn is not in an N-block) is no greater than n−1 n−1 1 X 1 X δ = δ, 12 (Z0 (T i x)) ≤ lim 1F (T i x)N = Nµ(F ) ≤ N n→∞ n n→∞ n N i=0 i=0 (2.7) that is, Pr(Zn is in an N − cell) ≥ 1 − δ.
lim
Since Zn is stationary by construction, Pr(ZkN = 011 · · · 1) = Pr(Z0N = 011 · · · 1) for all k. Thus Pr(Z0N = 011 · · · 1) =
N−1 1 X Pr(ZkN = 011 · · · 1). N k=0
The events {ZkN = 011 · · · 1}, k = 0, 1, . . . , N − 1 are disjoint, however, since there can be at most one 0 in a single block of N symbols. Thus NPr(Z N = 011 · · · 1) =
N−1 X
Pr(ZkN = 011 · · · 1)
k=0 N−1 [
= Pr(
{ZkN = 011 · · · 1}).
(2.8)
k=0
Thus since the rightmost probability is between 1 − δ and 1, 1 1−δ ≥ Pr(Z0N = 011 · · · 1) ≥ N N which completes the proof.
2
The following corollary shows that a finite-length sliding-block code can be used in the lemma.
2.8 Random Punctuation Sequences
41
Corollary 2.1. Given the assumptions of the lemma, a finite-length slidingblock code exists with properties (a)-(c). Proof. The sets G and hence also F can be chosen in the proof of the lemma to be finite dimensional, that is, to be measurable with respect to σ (X−K , · · · , XK ) for some sufficiently large K. Choose these sets as before with δ/2 replacing δ. Define n(x) as in the proof of the lemma. Since n(x) is finite with probability one, there must be an L such that if BL = {x : n(x) > L}, then µ(BL )
L, then the sliding-block code prints a 2. Thus if there is no occurrence of the desired finite dimensional pattern in a huge bunch of future symbols, a 2 is produced. If n(x) < L, then f is chosen as in the proof of the lemma. The proof now proceeds as in the lemma until (2.7), which is replaced by n−1 n−1 n−1 1 X 1 X 1 X 12 (Z0 (T i x)) ≤ lim 1BL (T i x) + lim 1F (T i x)N ≤ δ. n→∞ n n→∞ n n→∞ n i=0 i=0 i=0
lim
The remainder of the proof is the same. 2 Application of the lemma to an IID source and merging the symbols 1 and 2 in the punctuation process immediately yield the following result since coding an IID process yields a B-process. Corollary 2.2. Given an integer N and a δ > 0 there exists an (N, δ)punctuation sequence {Zn } with the following properties: (a) {Zn } is B-process (and hence stationary, ergodic, and mixing). (b) {Zn } has a binary alphabet {0, 1} and it can output only N-cells of the form 011 · · · 1 (0 followed by N − 1 ones) or individual ones; that is, each zero is always followed by at least N − 1 ones. (c) For all integers k 1−δ 1 ≤ Pr(ZkN = 011 · · · 1) ≤ N N and hence for any n Pr(Zn is in an N − cell) ≥ 1 − δ. Random punctuation sequences are closely related to the RohlinKakutani theorem, a classic result of ergodic theory. The language and notation is somewhat different and we shall return to the topic at the end of this chapter.
42
2 Pair Processes: Channels, Codes, and Couplings
2.9 Memoryless Channels Suppose that qx0 (·) is a probability measure on BB for all x0 ∈ A and that for fixed F ,qx0 (F ) is a measurable function of x0 . Let ν be a channel specified by its values on output rectangles by Y qxi (Fi ), νx ( × Fi ) = i∈J
i∈J
for any finite index set J ⊂ T. Then ν is said to be a memoryless channel. Intuitively, Y Pr(Yi ∈ Fi ; i ∈ J|X) = Pr(Yi ∈ Fi |Xi ). i∈J
In fact two forms of memorylessness are evident in a memoryless channel. The channel is input memoryless in that the probability of an output event involving {Yi ; i ∈ {k, k + 1, · · · , m}} does not involve any inputs before time k, that is, the past inputs. The channel is also input nonanticipatory since this event does not depend on inputs after time m, that is, the future inputs. The channel is also output memoryless in the sense that for any given input x, output events involving nonoverlapping times are independent, i.e., \ νx (Y1 ∈ F1 Y2 ∈ F2 ) = νx (Y1 ∈ F1 )νx (Y2 ∈ F2 ).
2.10 Finite-Memory Channels A channel ν is said to have finite input memory of order M if for all one-sided events F and all n νx ((Yn , Yn+1 , · · · ) ∈ F ) = νx 0 ((Yn , Yn+1 , · · · ) ∈ F ) whenever xi = xi0 for i ≥ n − M. In other words, for an event involving Yi ’s after some time n, knowing only the inputs for the same times and M time units earlier completely determines the output probability. Similarly ν is said to have finite anticipation of order L if for all one-sided events F and all n νx ((· · · , Yn ) ∈ F ) = νx 0 ((· · · , Yn ) ∈ F ) provided xi0 = xi for i ≤ n + L. That is, at most L future inputs must be known to determine the probability of an event involving current and past outputs. Channels with finite input memory were introduced by Feinstein [41].
2.11 Output Mixing Channels
43
A channel ν is said to have finite output memory of order K if for all one-sided events F and G and all inputs x, if k > K then νx ((· · · , Yn ) ∈ F
\ (Yn+k , · · · ) ∈ G) = νx ((· · · , Yn ) ∈ F )νx ((Yn+k , · · · ) ∈ G);
that is, output events involving output samples separated by more than K time units are independent. Channels with finite output memory were introduced by Wolfowitz [195]. Channels with finite memory and anticipation are historically important as the first real generalizations of memoryless channels for which coding theorems could be proved. Furthermore, the assumption of finite anticipation is physically reasonable as a model for real-world communication channels. The finite memory assumptions, however, exclude many important examples, e.g., finite-state or Markov channels and channels with feedback filtering action. Hence we will emphasize more general notions which can be viewed as approximations or asymptotic versions of the finite memory assumption. The generalization of finite input memory channels requires some additional tools and is postponed to the next chapter. The notion of finite output memory can be generalized by using the notion of mixing.
2.11 Output Mixing Channels A channel is said to be output mixing (or asymptotically output independent or asymptotically output memoryless) if for all output rectangles F and G and all input sequences x \ lim |νx (T −n F G) − νx (T −n F )νx (G)| = 0. n→∞
More generally it is said to be output weakly mixing if n−1 \ 1 X |νx (T −i F G) − νx (T −i F )νx (G)| = 0. n→∞ n i=0
lim
Unlike mixing systems, the above definitions for channels place conditions only on output rectangles and not on all output events. Output mixing channels were introduced by Adler [2]. The principal property of output mixing channels is provided by the following lemma.
44
2 Pair Processes: Channels, Codes, and Couplings
Lemma 2.7. If a channel is stationary and output weakly mixing, then it is also ergodic. That is, if ν is stationary and output weakly mixing and if µ is stationary and ergodic, then also µν is stationary and ergodic. Proof: The process µν is stationary by Lemma 2.1. To prove that it is ergodic it suffices from Lemma 2.3 to prove that for all input/output rectangles of the form F = FB × FA , FB ∈ BA T , FA ∈ BB T , and G = GB × GA that n−1 \ 1 X µν(T −i F G) = µν(F )µν(G). lim n→∞ n i=0 We have that n−1 \ 1 X µν(T −i F G) − m(F )m(G) = n i=0 n−1 \ \ 1 X µν((TB−i FB GB ) × (TA−i FA GA )) − µν(FB × FA )µν(GB × GA ) = n i=0 n−1 Z \ 1 X dµ(x)νx (TB−i FB GB ) − µν(FB × FA )µ(GB × GA ) = T −i n i=0 TA FA GA n−1 Z \ 1 X dµ(x)νx (TB−i FB GB ) T n i=0 TA−i FA GA ! Z −i dµ(x)νx (TB FB )νx (GB ) + − T TA−i FA
1 n
n−1 X i=0
GA
!
Z TA−i FA
T
GA
dµ(x)νx (TB−i FB )νx (GB )
− µν(FB × FA )µν(GB × GA ) .
The first term is bound above by n−1 Z \ 1 X dµ(x)|νx (TB−i FB GB ) − νx (TB−i FB )νx (GB )| ≤ T −i n i=0 TA FA GA
Z dµ(x)
n−1 \ 1 X |νx (TB−i FB GB ) − νx (T −i FB )νx (GB )| n i=0
which goes to zero from the dominated convergence theorem since the integrand converges to zero from the output weakly mixing assumption. The second term can be expressed using the stationarity of the channel as Z n−1 1 X dµ(x)νx (GB ) 1F (T i x)νT i x (FB ) − µν(F )µν(G). A n i=0 A A FA
2.12 Block Independent Channels
45
The ergodic theorem implies that as n → ∞ the sample average goes to its expectation Z dµ(x)1FA (x)νx (FB ) = µν(F )
2 and hence the above formula converges to 0, proving the lemma. The lemma provides an example of a completely random channel that is also ergodic in the following corollary. Corollary 2.3. Suppose that ν is a stationary completely random channel described by an output measure η. If η is weakly mixing, then ν is ergodic. That is, if µ is stationary and ergodic and η is stationary and weakly mixing, then µν = µ × η is stationary and ergodic. Proof: If η is weakly mixing, then the channel ν defined by νx (F ) = η(F ), all x ∈ AT , F ∈ BB T is output weakly mixing. Thus ergodicity follows from the lemma. 2
2.12 Block Independent Channels The idea of a memoryless channel can be extended to a block memoryless or block independent channel. Given integers N and K (usually N N K = N) and a probability measure qx N (·) on BK B for each x ∈ A such K N N that qx (F ) is a measurable function of x for each F ∈ BB . Let ν be specified by its values on output rectangles by n bK c
νx (y : yi ∈ Gi ; i = m, · · · , m + n − 1) =
Y i=0
qx N (Gi ), iN
where Gi ∈ BB , all i, where bzc is the largest integer contained in z, and where Gi =
m+(i+1)K−1
×
j=m+iK
Fj
with Fj = B if j ≥ m + n. Such channels are called block memoryless channels or block independent channels.. A deterministic block independent and block stationary channel is a sequence coder formed by a block code. The primary use of block independent channels is in the construction of a channel given finite-dimensional conditional probabilities; that is, one has probabilities for output K-tuples given input N-tuples and one wishes to model a channel consistent with these finite-dimensional distributions. The finite-dimensional distributions themselves may be the result of an optimization problem or an estimate based on observed behavior. An immediate problem is that a channel constructed in this
46
2 Pair Processes: Channels, Codes, and Couplings
manner may not be stationary, although it is clearly (N, K)-stationary. In Section 2.14 it is seen how to modify a block independent channel so as to produce a stationary channel. The basic idea is to occasionally insert some random spacing between the blocks so as to “stationarize” the channel. Block independent channels are a special case of the class of conditionally block independent channels, which are considered next.
2.13 Conditionally Block Independent Channels A conditionally block independent (CBI) channel resembles the block independent channel in that for a given input sequence the outputs are block independent. It is more general, however, in that the conditional probabilities of the output block may depend on the entire input sequence (or at least on parts of the input sequence not in the same time block). A channel is CBI if its values on output rectangles satisfy n bK c
νx (y : yi ∈ Fi ; i = m, · · · , m + n − 1) =
Y
N νx (y : yiN ∈ Gi ).
i=0
where as before Gi =
m+(i+1)K−1
×
j=m+iK
Fj
with Fj = B if j ≥ m + n. Block memoryless channels are clearly a special case of CBI channels. These channels have only finite output memory, but unlike the block independent channels they need not have finite input memory or anticipation.
2.14 Stationarizing Block Independent Channels Block memoryless channels (and CBI channels) are both block stationary channels. Connecting a stationary input to a block stationary channel will yield a block stationary input/output pair process, but it is sometimes desirable to have a stationary model. In this section we consider a technique of “stationarizing” a block independent channel in order to produce a stationary channel. Intuitively, a stationarized block independent (SBI) channel is a block independent channel with random spacing inserted between the blocks according to a random punctuation process.
2.14 Stationarizing Block Independent Channels
47
That is, when the random blocking process produces N-cells (which is most of the time), the channel uses the N-dimensional conditional distribution. When it is not using an N cell, the channel produces some arbitrary symbol in its output alphabet. We now make this idea precise. Let N, K, and qx N (·) be as in block independent channel of Section 2.12. We now assume that K = N, that is, one output symbol is produced for every input symbol and hence output blocks have the same number of symbols as input blocks. This is done for simplicity as the more general case adds significant notational clutter for minimal conceptual gain. Given δ > 0 let γ denote the distribution of an (N, δ)-random punctuation sequence {Zn }. Let µ × γ denote the product distribution on (AT × {0, 1}T , BTA × BT{0,1} ); that is, µ × γ is the distribution of the pair process {Xn , Zn } consisting of the original source {Xn } and the random punctuation source {Zn } with the two sources being independent of one another. Define a regular conditional probability (and hence a channel) πx,z (F ), F ∈ {BB }T , x ∈ AT , z ∈ {0, 1}T by its values on rectangles as follows: Given z, let J2 (z) denote the collection of indices i for which zi = 2 and hence for which zi is not in an N-cell and let J0 (z) denote those indices i for which zi = 0, that is, those indices where N-cells begin. Let q∗ denote a trivial probability mass function on B placing all of its probability on a reference letter b∗ . Given an output rectangle F = {y : yj ∈ Fj ; j ∈ J} = × Fj , j∈J
define Y
πx,z (F ) = i∈J
T
J2 (z)
Y
q∗ (Fi ) i∈J
T
J0 (z)
i+N−1
qx N ( × i
j=i
Fi ),
where we assume that Fi = B if i 6∈ J. Connecting the product source µ × γ to the channel π yields a hookup process {Xn , Zn , Yn } with distribution, say, r , which in turn induces a distribution p on the pair process {Xn , Yn } having distribution µ on {Xn }. If the alphabets are standard, p also induces a regular conditional probability for Y given X and hence a channel ν for which p = µν. A channel of this form is said to be an (N, δ)-stationarized block independent or SBI channel. Lemma 2.8. An SBI channel is stationary and ergodic. Thus if a stationary (and ergodic) source µ is connected to an SBI channel ν, then the output is stationary (and ergodic). Proof: The product source µ × γ is stationary and the channel π is stationary, hence so is the hookup (µ × γ)π or {Xn , Zn , Yn }. Thus the pair process {Xn , Yn } must also be stationary as claimed. The product source µ × γ is ergodic from Corollary 2.3 since it can be considered as the input/output process of a completely random channel described by a mixing (hence also weakly mixing) output measure. The channel π is output
48
2 Pair Processes: Channels, Codes, and Couplings
strongly mixing by construction and hence is ergodic from Lemma 2.4. Thus the hookup (µ × γ)π must be ergodic. This implies that the coordinate process {Xn , Yn } must also be ergodic. This completes the proof.
2
The block independent and SBI channels are useful primarily for proving theorems relating finite-dimensional behavior to sequence behavior and for simulating channels with specified finite-dimensional behavior. The SBI channels will also play a key role in deriving sliding-block coding theorems from block coding theorems by replacing the block distributions by trivial distributions, i.e., by finite-dimensional deterministic mappings or block codes. The SMB channel was introduced by Pursley and Davisson [29] for finite-alphabet channels and further developed by Gray and Saadat [70], who called it a randomly blocked conditionally independent (RBCI) channel. We opt for the first name because these channels resemble block memoryless channels more than CBI channels.
2.15 Primitive Channels Primitive channels were introduced by Neuhoff and Shields [136, 133] as a physically motivated general channel model. The idea is that most physical channels combine the input process with a separate noise process that is independent of the signal and then filter the combination in a stationary fashion. The noise is assumed to be IID since the filtering can introduce dependence. The construction of such channels strongly resembles that of the SBI channels. Let γ be the distribution of an IID process {Zn } with alphabet W , let µ × γ denote the product source formed by an independent joining of the original source distribution µ and the noise process Zn , let π denote the deterministic channel induced by a stationary sequence coder f : AT × W T → B T mapping an input sequence and a noise sequence into an output sequence. Let r = (µ × γ)π denote the resulting hookup distribution and {Xn , Zn , Yn } denote the resulting process. Let p denote the induced distribution for the pair process {Xn , Yn }. If the alphabets are standard, then p and µ together induce a channel νx (F ), x ∈ AT , F ∈ BB T . A channel of this form is called a primitive channel. Lemma 2.9. A primitive channel is stationary with respect to any stationary source and it is ergodic. Thus if µ is stationary and ergodic and ν is primitive, then µν is stationary and ergodic. Proof: Since µ is stationary and ergodic and γ is IID and hence mixing, µ × ν is stationary and ergodic from Corollary 2.3. Since the deterministic channel is stationary, it is also ergodic from Lemma 2.4 and the
2.17 Markov Channels
49
resulting triple {Xn , Zn , Yn } is stationary and ergodic. This implies that the component process {Xn , Yn } must also be stationary and ergodic, completing the proof. 2
2.16 Additive Noise Channels Suppose that {Xn } is a source with distribution µ and that {Wn } is a “noise” process with distribution γ. Let {Xn , Wn } denote the induced product source, that is, the source with distribution µ × γ so that the two processes are independent. Suppose that the two processes take values in a common alphabet A and that A has an addition operation +, e.g., it is a semi-group. Define the sliding-block code f by f (x, w) = x0 + w0 and let f denote the corresponding sequence coder. Then as in the primitive channels we have an induced distribution r on triples {Xn , Wn , Yn } and hence a distribution on pairs {Xn , Yn } which with µ induces a channel ν if the alphabets are standard. Example 2.1. A channel of this form is called a additive noise channel or a signal-independent additive noise channel. If the noise process is a B-process, then this is easily seen to be a special case of a primitive channel and hence the channel is stationary with respect to any stationary source and ergodic. If the noise is only known to be stationary, the channel is still stationary with respect to any stationary source. Unless the noise is assumed to be at least weakly mixing, however, it is not known if the channel is ergodic in general.
2.17 Markov Channels We now consider a special case where A and B are finite sets with the same number of symbols. For a fixed positive integer K, let P denote the space of all K × K stochastic matrices P = {P (i, j); i, j = 1, 2, · · · , K}. Using the Euclidean metric on this space we can construct the Borel field P of subsets of P generated by the open sets to form a measurable space (P, P). This, in turn, gives a one-sided or two-sided sequence space (PT , P T ). A map φ : AT → PT is said to be stationary if φTA = TP φ. Given a sequence P ∈ PT , let M(P ) denote the set of all probability measures on (B T , BT ) with respect to which Ym , Ym+1 , Ym+2 , · · · forms a Markov chain with transition matrices Pm , Pm+1 , · · · for any integer m, that is, λ ∈ M(P ) if and only if for any m
50
2 Pair Processes: Channels, Codes, and Couplings
λ[Ym = ym , · · · , Yn = yn ] = λ[Ym = ym ]
n−1 Y
Pi (yi , yi+1 ),
i=m
n > m, ym , · · · , yn ∈ B. In the one-sided case only m = 1 need be verified. Observe that in general the Markov chain is nonhom*ogeneous. A channel [A, ν, B] is said to be Markov if there exists a stationary measurable map φ : AT → PT such that νx ∈ M(φ(x)), x ∈ AT . Markov channels were introduced by Kieffer and Rahe [98] who proved that one-sided and two-sided Markov channels are AMS. Their proof is not included as it is lengthy and involves techniques not otherwise used in this book. The channels are introduced for completeness and to show that several important channels and codes in the literature can be considered as special cases. A variety of conditions for ergodicity for Markov channels are considered in [69]. Most are equivalent to one already considered more generally here: A Markov channel is ergodic if it is output mixing.
2.18 Finite-State Channels and Codes The most important special cases of Markov channels are finite-state channels and codes. Given a Markov channel with stationary mapping φ, the channel is said to be a finite-state channel (FSC) if we have a collection of stochastic matrices Pa ∈ P; a ∈ A and that φ(x)n = Pxn , that is, the matrix produced by φ at time n depends only on the input at that time, xn . If the matrices Pa ; a ∈ A contain only 0’s and 1’s, the channel is called a finite-state code. There are several equivalent models of finitestate channels and we pause to consider an alternative form that is more common in information theory. (See Gallager [47], Ch. 4, for a discussion of equivalent models of FSC’s and numerous physical examples.) An FSC converts an input sequence x into an output sequence y and a state sequence s according to a conditional probability Pr(Yk = yk , Sk = sk ; k = m, · · · , n|Xi = xi , Si = si ; i < m) = n Y
P (yi , si |xi , si−1 ),
i=m
that is, conditioned on Xi , Si−1 , the pair Yi , Si is independent of all prior inputs, outputs, and states. This specifies a FSC defined as a special case of a Markov channel where the output sequence above is here the joint state-output sequence {yi , si }. Note that with this setup, saying the Markov channel is AMS implies that the triple process of source, states,
2.19 Cascade Channels
51
and outputs is AMS (and hence obviously so is the Gallager input-output process). We will adapt the Kieffer-Rahe viewpoint and call the outputs {Yn } of the Markov channel states even though they may correspond to state-output pairs for a specific physical model. In the two-sided case, the Markov channel is significantly more general than the FSC because the choice of matrices φ(x)i can depend on the past in a very complicated (but stationary) way. One might think that a Markov channel is not a significant generalization of an FSC in the onesided case, however, because there stationarity of φ does not permit a dependence on past channel inputs, only on future inputs, which might seem physically unrealistic. Many practical communications systems do effectively depend on the future, however, by incorporating delay in the coding. The prime example of such look-ahead coders are trellis and tree codes used in an incremental fashion. Such codes investigate many possible output strings several steps into the future to determine the possible effect on the receiver and select the best path, often by a Viterbi algorithm. (See, e.g., Viterbi and Omura [189].) The encoder then outputs only the first symbol of the selected path. While clearly a finite-state machine, this code does not fit the usual model of a finite-state channel or code because of the dependence of the transition matrix on future inputs (unless, of course, one greatly expands the state space). It is, however, a Markov channel.
2.19 Cascade Channels We will often wish to connect more than one channel in cascade in order to form a communication system, e.g., the original source is connected to a deterministic channel (encoder) which is connected to a communications channel which is in turn connected to another deterministic channel (decoder). We now make precise this idea. Suppose that we are given two channels [A, ν (1) , C] and [C, ν (2) , B]. The cascade of ν (1) and ν (2) is defined as the channel [A, ν, B] given by Z νx (F ) = νu(2) (F ) dνx(1) (u). CT
In other words, if the original source sequence is X, the output to the first channel and input to the second is U, and the output of the second (1) channel is Y , then νx (F ) = PU |X (F |x), νu (G) = PY |U (G|u), and νx (G) = PY |X (G|x). Observe that by construction X → U → Y is a Markov chain. Lemma 2.10. A cascade of two stationary channels is stationary. Proof: Let T denote the shift on all of the spaces. Then
52
2 Pair Processes: Channels, Codes, and Couplings
νx (T
−1
Z F) = CT
νu(2) (T −1 F )dνx(1) (u)
(1)
Z = CT
νu(2) (F )dνx(1) T −1 (u). (1)
(1)
But νx (T −1 F ) = νT x (1) (F ), that is, the measures νx T −1 and νT x are identical and hence the above integral is Z (1) νu(2) (F ) dνT x (u) = νT x (F ), CT
2
proving the lemma.
2.20 Communication Systems A communication system consists of a source [A, µ], a sequence encoder f : AT → B T (a deterministic channel), a channel [B, ν, B 0 ], and a seˆT . The overall distribution r is specified by quence decoder g : B 0 T → A its values on rectangles as Z \ r (F1 × F2 × F3 × F4 ) = dµ(x)νf (x) (F3 g −1 (F4 )). T F1
f −1 (F2 )
Denoting the source by {Xn }, the encoded source or channel input process by {Un }, the channel output process by {Yn }, and the decoded proˆn }, then r is the distribution of the process {Xn , Un , Yn , X ˆn }. If cess by {X ˆ denote the corresponding sequences, then observe we let X,U,Y , and X ˆ are Markov chains. We abbreviate a that X → U → Y and U → Y → X communication system to [µ, f , ν, g]. It is straightforward from Lemma 2.10 to show that if the source, channel, and coders are stationary, then so is the overall process. A key topic in information theory, which is a mathematical theory of communication systems, is the characterization of the optimal performance one can obtain for communicating a given source over a given channel using codes within some available class of codes. Precise definitions of optimal will be based on the notion of the quality of a system as determined by a measure of distortion between input and output to be introduced in Chapter 5.
2.21 Couplings So far in this chapter the focus has been on combining a source [A, µ] and a channel [A, ν, B] or a code which together produce a pair or input/output process [A × B, π ], where π = µν. The pair process in turn
2.22 Block to Sliding-Block: The Rohlin-Kakutani Theorem
53
induces an output process, [B, η]. Given a pair process [A × B, π ], the induced input process [A, µ] and output process [B, η] can be thought of as the marginal processes and µ and η the marginal distributions of the pair process [A × B, π ] and its distribution. From a different viewpoint, we could consider the two marginal processes [A, µ] and [B, η] as being given and define a coupling or joining of these two processes as any pair process [A × B, π ] having the given marginals. Here we can view π as a coupling of µ and η. In general, given any two processes [A, µ] and [B, η], let P(µ, η) denote the class of all pair process distributions corresponding to couplings of the two given distributions. This class is not empty because, for example, we can always construct a coupling using product measures. This corresponds to the pair process with the given marginals where the two processes are mutually independent or, in other words, the example of the completely random channel given earlier. When it is desired to place emphasis on the names of the random processes rather than the distributions, we will refer to a pair process distribution πX,Y with marginals πX and πY . If we begin with two separate processes with distributions µX and µY , say, then P(µX , µY ) will denote the collection of all pair processes with marginals πX = µX and πY = µY . Occasionally πX,Y ∈ P(µX , µY ) will be abbreviated to πX,Y ⇒ µX , µY . If one is given two sources and forms a coupling, then in the case of processes with standard alphabets the coupling implies a channel since the joint process distribution and the input process distribution together imply a conditional distribution of output sequences given input sequences, and this conditional distribution is a regular conditional distribution and hence describes a channel. Couplings can also be defined for pairs of random vectors rather than random processes in a similar manner.
2.22 Block to Sliding-Block: The Rohlin-Kakutani Theorem The punctuation sequences of Section 2.14 provide a means for converting a block code into a sliding-block code. Suppose, for example, that {Xn } is a source with alphabet A and γN is a block code, γN : AN → B N . (The dimensions of the input and output vector are assumed equal to simplify the discussion.) Typically B is binary. As has been argued, block codes are not stationary. One way to stationarize a block code is to use a procedure similar to that used to stationarize a block independent channel: send long sequences of blocks with occasional random spacing to make the overall encoded process stationary. Thus, for example, one could use a sliding-block code to produce a punctuation sequence {Zn } as in Corollary 2.1 which produces isolated 0’s followed by KN 1’s
54
2 Pair Processes: Channels, Codes, and Couplings
and occasionally produces 2’s. The sliding-block code uses γN to encode N N a sequence of K source blocks XnN , Xn+N , · · · , Xn+(K−1)N if and only if Zn = 0. For those rare times l when Zl = 2, the sliding-block code produces an arbitrary symbol b∗ ∈ B. The resulting sliding-block code inherits many of the properties of the original block code, as will be demonstrated when proving theorems for sliding-block codes constructed in this manner. This construction suffices for source coding theorems, but an additional property will be needed when treating the channel coding theorems and other applications. The shortcoming of the results of Lemma 2.6 and Corollary 2.1 is that important source events can depend on the punctuation sequence. In other words, probabilities can be changed by conditioning on the occurrence of Zn = 0 or the beginning of a block code word. In this section we modify the simple construction of Lemma 2.6 to obtain a new punctuation sequence that is approximately independent of certain prespecified events. The result is a variation of the Rohlin-Kakutani theorem of ergodic theory [157] [83]. The development here is patterned after that in Shields [164]. See also Shields and Neuhoff [167]. We begin by recasting the punctuation sequence result in different terms. Given a stationary and ergodic source {Xn } with a process distribution µ and a punctuation sequence {Zn } as in Section 2.14, define the set F = {x : ZN (x) = 0}, where x ∈ A∞ is a two-sided sequence x = (· · · , x−1 , x0 , x1 , · · · ). Let T denote the shift on this sequence space. Restating Corollary 2.1 yields the following. Lemma 2.11. Given δ > 0 and an integer N, an L sufficiently large and a set F of sequences that is measurable with respect to (X−L , · · · , XL ) with the following properties: (A) (B)
The sets T i F , i = 0, 1, · · · , N − 1 are disjoint. 1−δ 1 ≤ µ(F ) ≤ . N N
(C) 1 − δ ≤ µ(
N−1 [
T i F ).
i=0
So far all that has been done is to rephrase the punctuation result in more ergodic theory oriented terminology. One can think of the lemma as representing sequence space as a “base” F together with its disjoint shifts T i F ; i = 1, 2, · · · , N − 1, which make up most SN−1 of the space, together with whatever is left over, a set G = A∞ − i=0 T i F , a set which has probability less than δ which will be called the garbage set. This picture is called a tower or Rochlin-Kakutani tower. The basic construction is pictured in Figure 2.3.
2.22 Block to Sliding-Block: The Rohlin-Kakutani Theorem
55 G
TNF .. .
.. . T 3F
6 T 2F
6 TF
6 F Fig. 2.3 Rohlin-Kakutani Tower
We can relate a tower to a punctuation sequence by identifying the base of the tower, the set F , as the set of sequences of the underlying process which yield Z0 = 0, that is, the punctuation sequence at time 0 yields a 0, indicating the beginning of an N-cell.
Partitions We now add another wrinkle — consider a finite partition P = {Pi ; i = 0, 1, · · · , kPk − 1} of A∞ . One example is the partition of a finitealphabet sequence space into its possible outputs at time 0, that is, Pi = {x : x0 = ai } for i = 0, 1, · · · , kAk − 1. This is the zero-time partition for the underlying finite-alphabet process. Another possible partition would be according to the output of a sliding-block coding of x, the zero-time partition of the sliding-block coding (or the zero-time partition of the encoded process). In general there is a finite collection of important events that we wish to force to be approximately independent of the punctuation sequence and P is chosen so that the important events are unions of atoms of P. Given a partition P, we define the label function
56
2 Pair Processes: Channels, Codes, and Couplings
labelP (x) =
kPk−1 X
i1Pi (x),
i=0
where as usual 1P is the indicator function of a set P . Thus the label of a sequence is simply the index of the atom of the partition into which it falls. As P partitions the input space into which sequences belong to atoms of P, T −i P partitions the space according to which shifted sequences T i x belong to atoms of P, that is, x ∈ T −i Pl ∈ T −i P is equivalent to T i x ∈ Pl and hence labelP (T i x) = l. The join PN =
N−1 _
T −i P
i=0
partitions the space into sequences sharing N labels in the following sense: Each atom Q of P N has the form Q = {x : labelP (x) = k0 , labelP (T x) = k1 , · · · , labelP (T N−1x) = kN − 1} =
N−1 \
T −i Pki
i=0
for some N tuple of integers k = (k0 , · · · , kN − 1). In the ergodic theory literature k is clled the P-N-name of the atom Q. For this reason we index the atoms of P N = Q as Qk . Thus P N breaks up the sequence space into groups of sequences which have the same labels for N shifts.
Gadgets In ergodic theory a gadget is a quadruple (T , F , N, P) where T is a transformation (for us a shift), F is an event such that T i F ; i = 01, . . .S , N −1 are N−1 disjoint (as in a Rohlin-Kakutani tower), and P is a partition of i=0 T i F . For concreteness, suppose that P is the zero-time partition of an underlying process, say a binary IID process. Consider the partition induced in F , the base of the gadget, by P N = {Qk }, that is, the collection of sets of the form Qk ∩ F . By construction, this will be the collection of all infinite sequences for which the punctuation sequence at time zero is 0 (Z0 = 0) and the P-n label of the next N outputs of the process is k, in the binary example there are 2N such binary N-tuples since kPk = 2. The set Qk ∩ F SN−1 together with its N − 1 shifts (that is, the set i=0 T i (Qk ∩ F )) is called a column of the gadget. Gadgets provide an extremely useful structure for using a block code to construct a sliding-block code. Each atom Qk ∩ F in the base partition
2.22 Block to Sliding-Block: The Rohlin-Kakutani Theorem
57
contains all sequences corresponding the next N input values being a given binary N-tuple following a punctuation event Z0 = 0.
Strengthened Rohlin-Kakutani Theorem Lemma 2.12. Given the assumptions of Lemma 2.11 and a finite partition P, L and F can be chosen so that in addition to properties (A)-(C) it is also true that (D) µ(Pi |F ) = µ(Pi |T l F ); l = 1, 2, · · · , N − 1, µ(Pi |F ) = µ(Pi |
N−1 [
T kF )
(2.9) (2.10)
k=0
and µ(Pi
\
F) ≤
1 µ(Pi ). N
(2.11)
Comment: Eq. (2.11) can be interpreted as stating that Pi and F are approximately independent since 1/N is approximately the probability of F . Only the upper T bound is stated as it is all we need. Eq. (2.9) also implies that µ(Pi F ) is bounded below by (µ(Pi ) − δ)µ(F ). Proof: Eq. (2.10) follows from (2.9) since µ(Pi |
N−1 [
T lF ) =
l=0
T SN−1
µ(Pi
l=0
SN−1
µ( PN−1 l=0
=
l=0
T lF )
T lF )
PN−1
T lF ) l=0 µ(Pi PN−1 l l=0 µ(T F )
=
T
N−1 µ(Pi |T l F )µ(T l F ) 1 X µ(Pi |T l F ) = Nµ(F ) N l=0
= µ(Pi | F ) Eq. (2.11) follows from (2.10) since µ(Pi
\
F ) = µ(Pi |F )µ(F ) = µ(Pi |
N−1 [
T k F )µ(F )
k=0
= µ(Pi |
N−1 [ k=0
=
T kF )
N−1 [
1 T k F )) µ( N k=0
\ N−1 [ 1 1 µ(Pi T k F ) ≤ µ(Pi ) N N k=0
since the T k F are disjoint and have equal probability, The remainder of this section is devoted to proving (2.9).
58
2 Pair Processes: Channels, Codes, and Couplings
We first construct using Lemma 2.11 a huge tower of size KN N, the height of the tower to be produced for this lemma. Let S denote the base of this original tower and let by the probability of the garbage set. This height KN tower with base S will be used to construct a new tower of height N and a base F with the additional desired property.T First consider the restriction of the partition P N to F defined by T N P F = {Qk F ; all KN-tuples k with coordinates taking values in T {0, 1, · · · , kPk − 1}}. P N F divides up the original base according to T the labels of NK shifts of base T sequences. For each atom Qk F in this base partition, the sets {T l (Qk F ); k = 0, 1, · · · , KN − 1} are disjoint and together form a column of the tower {T l F ; k = 0, 1, · · · , KN − 1}. A T l set of the form T (Qk F ) is called the lth level of the column containing T T it. Observe that if y ∈ T l (Qk F ), then y = T l u for some u ∈ Qk F and T l u hasTlabel kl . Thus we consider kl to be the label of the column level T l (Qk F ). This complicated structure of columns and levels can be used to recover the original partition by [ \ \ \ T l (Qk Pj = F ) (Pj G), (2.12) l,k:kl =j
that is, Pj is the union of all column levels with label j together with that part of Pj in the garbage. We will focus on the pieces of Pj in the column levels as the garbage has very small probability. We wish to construct a new tower with base F so that the probability of Pi for any of N shifts of F is the same. To do this we form F dividing each column of the original tower into N equal parts. We collect a group of these parts to form F so that F will contain only one part at each level, the N shifts of F will be disjoint, and the union of the N shifts will almost contain all of the original tower. By using the equal probability parts the new base will have conditional probabilities for Pj given T l equal for all l, as will be shown. T T Consider the atom Q = Qk S in the partition P N S of the base of the original tower. If the source is aperiodic in the sense of placing zero probability on individual sequences, then the set Q can be divided into N disjoint sets of equal probability, say W0 , W1 , · · · , WN−1 . Define the set FQ by (K−2)N [
FQ = (
T iN W0 )
[
(K−2)N [
(
i=0
=
i=0
N−1 [ (K−2)N [ l=0
T 1+iN W1 )
[
(K−2)N [
···(
T N−1+iN WN−1 )
i=0
T l+iN Wl .
i=0
FQ contains (K − 2) N shifts of W0 , of T W1 , · · · of T l Wl , · · · and of T N−1 WN−1 . Because it only takes N-shifts of each small set and because
2.22 Block to Sliding-Block: The Rohlin-Kakutani Theorem
59
it does not include the top N levels of the original column, shifting FQ fewer than N times causes no overlap, that is, T l FQ are disjoint for j = 0, 1, · · · , N − 1. The union of these sets contains all of the original column of the tower except possibly portions of the top and bottom N − 1 levels (which the construction may not include). The new base F is now defined to be the union of all of the FQk T S . The sets T l F are then disjoint (since all the pieces are) and contain all of the levels of the original tower except possibly the top and bottom N − 1 levels. Thus µ(
N−1 [
(K−1)N−1 [
T l F ) ≥ µ(
l=0
T i S) =
i=N
≥ K−2
(K−1)N−1 X
µ(S)
i=N
1− 1− 2 = − . KN N KN
By choosing = δ/2 and K large this can be made larger than 1 − δ. Thus the new tower satisfies conditions (A)-(C) and we need only verify the new condition (D), that is, (2.9). We have that T µ(Pi T l F ) µ(Pi |T l F ) = . µ(F ) Since the denominator does not depend on l, we need only show the numerator does not depend on l. From (2.12) applied to the original tower we have that X \ \ \ µ(Pi µ(T j (Qk T lF ) = S) T l F ), j,k:kj =i
that is, the sum over all column levels (old tower) labeled i of the probability of the intersection of the column level and the lth shift of the new base F . The intersection of a column level in the jth level of the original tower with any shift of F must be an intersection of that column level with the jth shift of one of the sets W0 , · · · , WN−1 (which particular set depends on l). Whichever set is chosen, however, the probability within the sum has the form \ \ \ \ µ(T j (Qk S) T l F ) = µ(T j (Qk S) T j Wm ) \ \ = µ((Qk S) Wm ) = µ(Wm ), whereT the final step follows since Wm was originally chosen as a subset of Qk S. Since these subsets were all chosen to have equal probability, this last probability does not depend on m and hence on l and µ(T j (Qk
\
S)
\
T lF ) =
\ 1 µ(Qk S) N
60
2 Pair Processes: Channels, Codes, and Couplings
and hence µ(Pi
\
T lF ) =
X j,k:kj
\ 1 µ(Qk S), N =i
which proves (2.9) since there is no dependence on l. This completes the 2 proof of the lemma.
Chapter 3
Entropy
Abstract The development of the idea of entropy of random variables and processes by Claude Shannon provided the beginnings of information theory and of the modern age of ergodic theory. Entropy and related information measures will be shown to provide useful descriptions of the long term behavior of random processes and this behavior is a key factor in developing the coding theorems of information theory. Here the various notions of entropy for random variables, vectors, processes, and dynamical systems are introduced and and there fundamental properties derived. In this chapter the case of finite-alphabet random processes is emphasized for simplicity, reflecting the historical development of the subject. Occasionally we consider more general cases when it will ease later developments.
3.1 Entropy and Entropy Rate There are several ways to introduce the notion of entropy and entropy rate. The difference between the two concepts is that entropy is relevant to a single random variable or random vector or, equivalently, to a partition of the sample space, while entropy rate describes a limiting entropy per time unit is we look at sample vectors with increasing dimensions or iterates of a partition. We take some care at the beginning in order to avoid redefining things later. We also try to use definitions resembling the usual definitions of elementary information theory where possible. Let (Ω, B, P , T ) be a dynamical system. Let f : Ω → A be a finite-alphabet measurement (a simple function) defined on Ω and define the random process fn = f T n ; n ∈ T. For the moment we focus on one sided processes with T = {0, 1, 2, . . .}. If the transformation T is invertible, we can extend the definition to all integer n and obtain a two-sided process with T = Z. This process can be viewed as a coding of the original space, that R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_3, © Springer Science+Business Media, LLC 2011
61
62
3 Entropy
is, one produces successive coded values by transforming (e.g., shifting) the points of the space, each time producing an output symbol using the same rule or mapping. If Ω is itself a sequence space and T is a shift, then f is a sliding-block code as considered in Section 2.6 and it induces a stationary sequence code f = {f T n ; n ∈ T}. In the usual way we can construct an equivalent Kolmogorov model of this process. Let A = {a1 , a2 , . . . , akAk } denote the finite alphabet of f Z and let (AZ+ , BA+ ) be the resulting one-sided sequence space, where BA is the power set. We abbreviate the notation for this sequence space to (A∞ , B∞ A ). Let TA denote the shift on this space and let X denote the time zero sampling or coordinate function and define Xn (x) = X(TAn x) = xn . Let m denote the process distribution induced by the original space and −1
the f T n , i.e., m = Pf = P f where f (ω) = (f (ω), f (T ω), f (T 2 ω), . . .). Observe that by construction, shifting the input point yields an output sequence that is also shifted, that is, f (T ω) = TA f (ω). Sequence-valued measurements of this form are called stationary or invariant codings (or time-invariant or shift-invariant codings in the case of the shift) since the coding commutes with the transformations. Stationary codes will play an important role throughout this book and are discussed in some detail in Chapter 2. If the input space Ω is itself a sequence space and T is a shift, then the code is also called a sliding-block code to reflect the fact that the code operates by shifting the input sequence (sliding) and applying a common measurement or mapping to it. Both the sequence-to-symbol mapping f and the sequence-to-sequence mapping f are referred to as a sliding-block code, each implies the other. The entropy and entropy rates of a finite-alphabet measurement depend only on the process distributions and hence are usually more easily stated in terms of the induced directly given model and the process distribution. For the moment, however, we point out that the definition can be stated in terms of either system. Later we will see that the entropy of the underlying system is defined as a supremum of the entropy rates of all finite-alphabet codings of the system. The entropy of a discrete alphabet random variable f defined on the probability space (Ω, B, P ) is defined by X HP (f ) = − P (f = a) ln P (f = a). (3.1) a∈A
We define 0ln0 to be 0 in the above formula. We shall often use logarithms to the base 2 instead of natural logarithms. The units for entropy are “nats” when the natural logarithm is used and “bits” for base 2 logarithms. The natural logarithms are usually more convenient for mathe-
3.1 Entropy and Entropy Rate
63
matics while the base 2 logarithms provide more intuitive descriptions. The subscript P can be omitted if the measure is clear from context. Be forewarned that the measure will often not be clear from context since more than one measure may be under consideration and hence the subscripts will be required. A discrete alphabet random variable f has a probability mass function (PMF), say pf , defined by pf (a) = P (f = a) = P ({ω : f (ω) = a}) and hence we can also write X pf (a) ln pf (a). H(f ) = − a∈A
It is often convenient to consider the entropy not as a function of the particular outputs of f but as a function of the partition that f induces on Ω. In particular, suppose that the alphabet of f is A = {a1 , a2 , . . . , akAk } and define the partition Q = {Qi ; i = 1, 2, . . . , kAk} by Qi = {ω : f (ω) = ai } = f −1 ({ai }). In other words, Q consists of disjoint sets which group the points in Ω together according to what output the measurement f produces. We can consider the entropy as a function of the partition and write HP (Q) = −
kAk X
P (Qi ) ln P (Qi ).
(3.2)
i=1
Clearly different mappings with different alphabets can have the same entropy if they induce the same partition. Both notations will be used according to the desired emphasis. We have not yet defined entropy for random variables that do not have discrete alphabets; we shall do that later. Return to the notation emphasizing the mapping f rather than the partition. Defining the random variable P (f ) by P (f )(ω) = P (λ : f (λ) = f (ω)) we can also write the entropy as HP (f ) = EP (− ln P (f )). Using the equivalent directly given model we have immediately that HP (f ) = HP (Q) = Hm (X0 ) = Em (− ln m(X0 )).
(3.3)
At this point one might ask why we are carrying the baggage of notations for entropy in both the original space and in the sequence space. If we were dealing with only one measurement f (or Xn ), we could confine interest to the simpler directly-given form. More generally, however, we will be interested in different measurements or codings on a common system. In this case we will require the notation using the original system. Hence for the moment we keep both forms, but we shall often focus on the second where possible and the first only when necessary.
64
3 Entropy
The nth order entropy of a discrete alphabet measurement f with respect to T is defined as (n)
HP (f ) = n−1 HP (f n ) where f n = (f , f T , f T 2 , . . . , f T n−1 ) or, equivalently, we define the discrete alphabet random process Xn (ω) = f (T n ω), then f n = X n = X0 , X1 , . . . , Xn−1 . As previously, this is given by (n) (X) = n−1 Hm (X n ) = n−1 Em (− ln m(X n )). Hm
This is also called the entropy (per-coordinate or per-sample) of the random vector f n or X n . We can also use the partition notation here. The partition corresponding to f n has a particular form: Suppose thatWwe have two partitions, Q = {Qi } and P = {Pi }. Define their join Q P as T the partition containing all nonempty intersection sets of the form Qi Pj . Define also T −1 Q as the partition containing the atoms T −1 Qi . Then f n induces the partition n−1 _
T −i Q
i=0
and we can write (n)
(n)
HP (f ) = HP (Q) = n−1 HP (
n−1 _
T −i Q).
i=0
As before, which notation is preferable depends on whether we wish to emphasize the mapping f or the partition Q. The entropy rate or mean entropy of a discrete alphabet measurement f with respect to the transformation T is defined by (n)
H P (f ) = lim sup HP (f ) n→∞
(n)
= H P (Q) = lim sup HP (Q) n→∞
(n) (X). = H m (X) = lim sup Hm n→∞
Given a dynamical system (Ω, B, P , T ), the entropy H(P , T ) of the system (or of the measure with respect to the transformation) is defined by H(P , T ) = sup H P (f ) = sup H P (Q), (3.4) f
Q
3.2 Divergence Inequality and Relative Entropy
65
where the supremum is over all finite-alphabet measurements (or codings) or, equivalently, over all finite measurable partitions of Ω. (We emphasize that this means alphabets of size M for all finite values of M.) The entropy of a system is also called the Kolmogorov-Sinai invariant of the system because of the generalization by Kolmogorov [102] and Sinai [168] of Shannon’s entropy rate concept to dynamical systems and the demonstration that equal entropy was a necessary condition for two dynamical systems to be isomorphic. Note that the entropy rate is well-defined for a continuous-alphabet random process as the supremum over the entropy rates over all finitealphabet codings of the process. Such an entropy rate is usually infinite, but it is defined. Suppose that we have a dynamical system corresponding to a finitealphabet random process {Xn }, then one possible finite-alphabet measurement on the process is f (x) = x0 , that is, the time 0 output. In this case clearly H P (f ) = H P (X) and hence, since the system entropy is defined as the supremum over all simple measurements, H(P , T ) ≥ H P (X).
(3.5)
We shall later see in Theorem 6.1 that (3.5) holds with equality for finite alphabet random processes and provides a generalization of entropy rate for processes that do not have finite alphabets.
3.2 Divergence Inequality and Relative Entropy Many of the basic properties of entropy follow from a simple result known as the divergence inequality. A slight variation is well-known as the log-sum inequality). The divergence or relative entropy is a variation on the idea of entropy and it crops up often as a useful tool for proving and interpreting results and for comparing probability distributions. In this section several fundamental definitions and results are collected together for use in the next section in developing the properties of entropy and entropy rate. Lemma 3.1. Given two probability mass functions {pi } and {qi }, that is, two countable or finite sequences of nonnegative numbers that sum to one, then X pi pi ln ≥0 qi i with equality if and only if qi = pi , all i. Proof: The lemma follows easily from the elementary inequality for real numbers
66
3 Entropy
ln x ≤ x − 1
(3.6)
(with equality if and only if x = 1) since X i
pi ln
X X X qi qi ≤ pi ( − 1) = qi − pi = 0 pi pi i i i
with equality if and only if qi /pi = 1 all i. Alternatively,Tthe inequality follows from Jensen’s inequality [74] since ln is a convex function: X qi X qi pi ln ≤ ln pi = 0 p pi i i i with equality if and only if qi /pi = 1, all i. 2 The inequality has a simple corollary that we record now for later use. Corollary 3.1. The function x ln(x/y) of real positive x, y is convex in (x, y). Proof. Let (xi , yi ), i = 1, 2, be pairs of real positive numbers, 0 ≤ λ ≤ 1, and define x = λx1 +(1−λ)x2 and y = λy1 +(1−λ)y2 . Apply Lemma 3.1 to the probability mass functions p and q defined by λx1 λx1 + (1 − λ)x2 (1 − λ)x2 p2 = λx1 + (1 − λ)x2 λy1 q1 = λy1 + (1 − λ)y2 (1 − λ)y2 q2 = λy1 + (1 − λ)y2 p1 =
yields 0≤ λx1 (1−λ)x2 (1 − λ)x λx1 λx +(1−λ)x λx +(1−λ)x 2 1 2 1 2 + . ln ln (1−λ)y λy1 2 λx1 + (1 − λ)x2 λx1 + (1 − λ)x2 λy1 +(1−λ)y2
λy1 +(1−λ)y2
Cancelling the positive denominator and rearranging λx1 ln
x1 x2 + (1 − λ)λx2 ln ≥ y1 y2 (λx1 + (1 − λ)x2 ) ln
proving the claimed convexity.
λx1 + (1 − λ)x2 λy1 + (1 − λ)y2
! ,
2
3.2 Divergence Inequality and Relative Entropy
67
The quantity used in Lemma 3.1 is of such fundamental importance that we introduce another notion of information and recast the inequality in terms of it. As with entropy, the definition for the moment is only for finite-alphabet random variables. Also as with entropy, there are a variety of ways to define it. Suppose that we have an underlying measurable space (Ω, B) and two measures on this space, say P and M, and we have a random variable f with finite alphabet A defined on the space and that Q is the induced partition {f −1 (a); a ∈ A}. Let Pf and Mf be the induced distributions and let p and m be the corresponding probability mass functions, e.g., p(a) = Pf ({a}) = P (f = a). Define the relative entropy of a measurement f with measure P with respect to the measure M by kAk X p(a) P (Qi ) HP kM (f ) = HP kM (Q) = = . p(a) ln P (Qi ) ln m(a) M(Q i) a∈A i=1
X
Observe that this only makes sense if p(a) is 0 whenever m(a) is, that is, if Pf is absolutely continuous with respect to Mf or Mf Pf . Define HP kM (f ) = ∞ if Pf is not absolutely continuous with respect to Mf . The measure M is referred to as the reference measure. Relative entropies will play an increasingly important role as general alphabets are considered. In the early chapters the emphasis will be on ordinary entropy with similar properties for relative entropies following almost as an afterthought. When considering more abstract (nonfinite) alphabets later on, relative entropies will prove indispensible. Analogous to entropy, given a random process {Xn } described by two process distributions p and m, if it is true that mX n pX n ; n = 1, 2, . . . , then we can define for each n the nth order relative entropy n −1 Hpkm (X n) and the relative entropy rate H pkm (X) ≡ lim sup n→∞
1 Hpkm (X n ). n
When dealing with relative entropies it is often the measures that are important and not the random variable or partition. We introduce a special notation which emphasizes this fact. Given a probability space (Ω, B, P ), with Ω a finite space, and another measure M on the same space, we define the divergence of P with respect to M as the relative entropy of the identity mapping with respect to the two measures: D(P kM) =
X ω∈Ω
P (ω) ln
P (ω) . M(ω)
68
3 Entropy
Thus, for example, given a finite-alphabet measurement f on an arbitrary probability space (Ω, B, P ), if M is another measure on (Ω, B) then HP kM (f ) = D(Pf kMf ). Similarly, Hpkm (X n ) = D(PX n kMX n ), where PX n and MX n are the distributions for X n induced by process measures p and m, respectively. The theory and properties of relative entropy are therefore determined by those for divergence. There are many names and notations for relative entropy and divergence throughout the literature. The idea was introduced by Kullback for applications of information theory to statistics (see, e.g., Kullback [106] and the references therein) and was used to develop information theoretic results by Perez [145] [147] [146], Dobrushin [32], and Pinsker [150]. Various names in common use for this quantity are discrimination, discrimination information, Kullback-Leibler (KL) number , directed divergence, informational divergence, and cross entropy. The lemma can be summarized simply in terms of divergence as in the following theorem, which is commonly referred to as the divergence inequality. The assumption of finite alphabets will be dropped later. Theorem 3.1. Divergence Inequality Given any two probability measures P and M on a common finite-alphabet probability space, then D(P kM) ≥ 0
(3.7)
with equality if and only if P = M. In this form the result is known as the divergence inequality. The fact that the divergence of one probability measure with respect to another is nonnegative and zero only when the two measures are the same suggest the interpretation of divergence as a “distance” between the two probability measures, that is, a measure of how different the two measures are. It is not a true distance or metric in the usual sense since it is not a symmetric function of the two measures and it does not satisfy the triangle inequality. The interpretation is, however, quite useful for adding insight into results characterizing the behavior of divergence and it will later be seen to have implications for ordinary distance measures between probability measures. The divergence plays a basic role in the family of information measures all of the information measures that we will encounter — entropy, relative entropy, mutual information, and the conditional forms of these information measures — can be expressed as a divergence. There are three ways to view entropy as a special case of divergence. The first is to permit M to be a general measure instead of requiring it to
3.3 Basic Properties of Entropy
69
be a probability measure and have total mass 1. In this case entropy is minus the divergence if M is the counting measure, i.e., assigns measure 1 to every point in the discrete alphabet. If M is not a probability measure, then the divergence inequality (3.7) need not hold. Second, if the alphabet of f is Af and has kAf k elements, then letting M be a uniform PMF assigning probability 1/kAk to all symbols in A yields D(P kM) = ln kAf k − HP (f ) ≥ 0 and hence the entropy is the log of the alphabet size minus the divergence with respect to the uniform distribution. Third, we can also consider entropy a special case of divergence while still requiring that M be a probability measure by using product measures and a bit of a trick. Say we have two measures P and Q on a common probability space (Ω, B). Define two measures on the product space (Ω × Ω, B(Ω × Ω)) as follows: Let P × Q denote the usual product measure, that is, the measure specified by its values on rectangles as P × Q(F × G) = P (F )Q(G). Thus, for example, if P and Q are discrete distributions with PMF’s p and q, then the PMF for P × Q is just p(a)q(b). Let P 0 denote the “diagonal” meaT sure defined by its values on rectangles as P 0 (F × G) = P (F G). In the discrete case P 0 has PMF p0(a, b) = p(a) if a = b and 0 otherwise. Then HP (f ) = D(P 0 kP × P ). Note that if we let X and Y be the coordinate random variables on our product space, then both P 0 and P × P give the same marginal probabilities to X and Y , that is, PX = PY = P . P 0 is an extreme distribution on (X, Y ) in the sense that with probability one X = Y ; the two coordinates are deterministically dependent on one another. P × P , however, is the opposite extreme in that it makes the two random variables X and Y independent of one another. Thus the entropy of a distribution P can be viewed as the relative entropy between these two extreme joint distributions having marginals P .
3.3 Basic Properties of Entropy For the moment fix a probability measure m on a measurable space (Ω, B) and let X and Y be two finite-alphabet random variables defined on that space. Let AX and AY denote the corresponding alphabets. Let PXY , PX , and PY denote the distributions of (X, Y ), X, and Y , respectively. First observe that since PX (a) ≤ 1, all a, − ln PX (a) is positive and hence X H(X) = − PX (a) ln PX (a) ≥ 0. (3.8) a∈A
70
3 Entropy
From (3.7) with M uniform as in the second interpretation of entropy above, if X is a random variable with alphabet AX , then H(X) ≤ ln kAX k. Since for any a ∈ AX and b ∈ AY we have that PX (a) ≥ PXY (a, b), it follows that X H(X, Y ) = − PXY (a, b) ln PXY (a, b) a,b
≥−
X
PXY (a, b) ln PX (a) = H(X).
a,b
Using Lemma 3.1 we have that since PXY and PX PY are probability mass functions, H(X, Y ) − (H(X) + H(Y )) =
X
PXY (a, b) ln
a,b
PX (a)PY (b) ≤ 0. PXY (a, b)
This proves the following result. Lemma 3.2. Given two discrete alphabet random variables X and Y defined on a common probability space, we have 0 ≤ H(X)
(3.9)
and max(H(X), H(Y )) ≤ H(X, Y ) ≤ H(X) + H(Y )
(3.10)
where the right hand inequality holds with equality if and only if X and Y are independent. If the alphabet of X has kAX k symbols, then HX (X) ≤ ln kAX k.
(3.11)
There is another proof of the left hand inequality in (3.10) that uses an inequality for relative entropy that will be useful later when considering codes. The following lemma gives the inequality. First we introduce a definition. A partition R is said to refine a partion Q if every atom in Q is a union of atoms of R, in which case we write Q < R. Lemma 3.3. Suppose that P and M are two measures defined on a common measurable space (Ω, B) and that we are given a finite partitions Q < R. Then HP kM (Q) ≤ HP kM (R) and HP (Q) ≤ HP (R).
3.3 Basic Properties of Entropy
71
Comments: The lemma can also be stated in terms of random variables and mappings in an intuitive way: Suppose that U is a random variable with finite alphabet A and f : A → B is a mapping from A into another finite alphabet B. Then the composite random variable f (U) defined by f (U )(ω) = f (U (ω)) is also a finite alphabet random variable. If U induces a partition R and f (U) a partition Q, then Q < R (since knowing the value of U implies the value of f (U)). Thus the lemma immediately gives the following corollary. Corollary 3.2. If M P are two measures describing a random variable U with finite alphabet A and if f : A → B, then HP kM (f (U)) ≤ HP kM (U) and HP (f (U)) ≤ HP (U ). Since D(Pf kMf ) = HP kM (f ), we have also the following corollary which we state for future reference. Corollary 3.3. Suppose that P and M are two probability measures on a discrete space and that f is a random variable defined on that space, then D(Pf kMf ) ≤ D(P kM). The lemma, discussion, and corollaries can all be interpreted as saying that taking a measurement on a finite-alphabet random variable lowers the entropy and the relative entropy of that random variable. By choosing U as (X, Y ) and f (X, Y ) = X or Y , the lemma yields the promised inequality of the previous lemma. Proof of Lemma: If HP kM (R) = +∞, the result is immediate. If HP kM (Q) = +∞, that is, if there exists at least one Qj such that M(Qj ) = 0 but P (Qj ) 6= 0, then there exists an Ri ⊂ Qj such that M(Ri ) = 0 and P (Ri ) > 0 and hence HP kM (R) = +∞. Lastly assume that both HP kM (R) and HP kM (Q) are finite and consider the difference HP kM (R) − HP kM (Q) =
X
P (Ri ) ln
i
X X = j
i:Ri ⊂Qj
X P (Qj ) P (Ri ) P (Qj ) ln − M(Ri ) M(Q j) j
) P (Q P (Ri ) j P (Ri ) ln − P (Qj ) ln . M(Ri ) M(Qj )
We shall show that each of the bracketed terms is nonnegative, which will prove the first inequality. Fix j. If P (Qj ) is 0 we are done since then also P (Ri ) is 0 for all i in the inner sum since these Ri all belong to Qj . If P (Qj ) is not 0, we can divide by it to rewrite the bracketed term as
72
3 Entropy
X P (Qj ) i:Ri ⊂Qj
P (Ri )/P (Qj ) P (Ri ) ln , P (Qj ) M(Ri )/M(Qj )
where we also used the fact that M(Qj ) cannot be 0 sinceTthen P (Qj ) would also have to be zero. Since R i ⊂ Qj , P (Ri )/P (Qj ) = P (Ri Qj )/P (Qj ) = P (Ri |Qj ) is an elementary conditional probability. Applying a similar argument to M and dividing by P (Qj ), the above expression becomes X i:Ri ⊂Qj
P (Ri |Qj ) ln
P (Ri |Qj ) M(Ri |Qj )
which is nonnegative from Lemma 3.1, which proves the first inequality. The second inequality follows similarly. Consider the difference X X P (Qj ) P (Ri ) ln HP (R) − HP (Q) = P (Ri ) j i:Ri ⊂Qj X X = P (Qj ) − P (Ri |Qj ) ln P (Ri |Qj ) j
i:Ri ⊂Qj
and the result follows since the bracketed term is nonnegative since it is an entropy for each value of j (Lemma 3.2). 2
Concavity of Entropy The next result provides useful inequalities for entropy considered as a function of the underlying distribution. In particular, it shows that enT tropy is a concave (or convex ) function of the underlying distribution. The concavity follows from Corollary 3.5, but for later use we do a little extra work to obtain an additional property. Define the binary entropy function (the entropy of a binary random variable with probability mass function (λ, 1 − λ)) by h2 (λ) = −λ ln λ − (1 − λ) ln(1 − λ). Lemma 3.4. Let m and p denote two distributions for a discrete alphabet random variable X and let λ ∈ (0, 1). Then for any λ ∈ (0, 1) λHm (X) + (1 − λ)Hp (X) ≤ Hλm+(1−λ)p (X) ≤ λHm (X) + (1 − λ)Hp (X) + h2 (λ).
(3.12)
3.3 Basic Properties of Entropy
73
Proof: Define the quantities X I = − m(x) ln(λm(x) + (1 − λ)p(x)) x
and J = Hλm+(1−λ)p (X) = −λ
X
m(x) ln(λm(x) + (1 − λ)p(x))−
x
(1 − λ)
X
p(x) ln(λm(x) + (1 − λ)p(x)).
x
First observe that λm(x) + (1 − λ)p(x) ≥ λm(x) and therefore applying this bound to both m and p X m(x) ln m(x) = − ln λ + Hm (X) I ≤ −ln λ − x
J ≤ −λ
X
m(x) ln m(x) − (1 − λ)
x
X
p(x) ln p(x) + h2 (λ)
x
= λHm (X) + (1 − λ)Hp (X) + h2 (λ).
(3.13)
To obtain the lower bounds of the lemma observe that I=−
X
=−
X
m(x) ln m(x)(λ + (1 − λ)
x
m(x) ln m(x) −
x
X
p(x) ) m(x)
m(x) ln(λ + (1 − λ)
x
p(x) ). m(x)
Using (3.6) the rightmost term is bound below by −
X
m(x)((λ + (1 − λ)
x
X p(x) − 1) = −λ − 1 + λ p(X = a) + 1 = 0. m(x) a∈A
Thus for all n I≥−
X
m(x) ln m(x) = Hm (X).
(3.14)
x
and hence also J ≥ −λ
X
m(x) ln m(x) − (1 − λ)
x
X
p(x) ln p(x)
x
= λHm (X) + (1 − λ)Hp (X).
2
74
3 Entropy
Convexity of Divergence Relative entropy possesses a useful convexity property with respect to the two probability measures, as described in the following lemma. Lemma 3.5. D(P kM) is convex in (P , M) for probability measures P , M on a common finite-alphabet probability space, that is, if (Pi , Mi ), i = 1, 2 are pairs of probability measures, all of which are on a common finitealphabet probability space, and (P , M) = λ(P1 , M1 ) + (1 − λ)(P2 , M2 ), then D(P kM) ≤ λD(P1 kM1 ) + (1 − λ)D(P2 kM2 ). Proof. The result follows from the convexity of a ln(a/b) in (a, b) from Corollary 3.1. 2
Entropy and Binomial Sums The next result presents an interesting connection between combinatorics and binomial sums with a particular entropy. We require the familiar definition of the binomial coefficient: ! n n! = . k k!(n − k)! 1 Lemma 3.6. Given δ ∈ (0, 2 ] and a positive integer M, we have
X i≤δM
M i
! ≤ eMh2 (δ) .
(3.15)
If 0 < δ ≤ p ≤ 1, then X i≤δM
! M p i (1 − p)M−i ≤ e−Mh2 (δkp) , i
where h2 (δkp) = δ ln
(3.16)
1−δ δ + (1 − δ) ln . p 1−p
Proof: We have after some simple algebra that e−h2 (δ)M = δδM (1 − δ)(1−δ)M . If δ < 1/2, then δk (1 − δ)M−k increases as k decreases (since we are having more large terms and fewer small terms in the product) and hence
3.3 Basic Properties of Entropy
75
if i ≤ Mδ, δδM (1 − δ)(1−δ)M ≤ δi (1 − δ)M−i . Thus we have the inequalities ! ! M X M X M i M−i ≥ δ (1 − δ) δi (1 − δ)M−i 1= i i i=0 i≤δM ! X M ≥ e−h2 (δ)M i i≤δM which completes the proof of (3.15). In a similar fashion we have that eMh2 (δkp) = (
δ δM 1 − δ (1−δ)M ) ( ) . p 1−p
Since δ ≤ p, we have as in the first argument that for i ≤ Mδ (
δ δM 1 − δ (1−δ)M δ 1 − δ M−i ) ( ) ) ≤ ( )i ( p 1−p p 1−p
and therefore after some algebra we have that if i ≤ Mδ then p i (1 − p)M−i ≤ δi (1 − δ)M−i e−Mh2 (δkp) and hence X i≤δM
! ! X M M i M−i −Mh2 (δkp) ≤e p (1 − p) δi (1 − δ)M−i i i i≤δM ≤e
−nh2 (δkp)
! M X M i δ (1 − δ)M−i = e−Mh2 (δkp) , i i=0
which proves (3.16). 2 The following is a technical but useful property of sample entropies. The proof follows Billingsley [16]. Lemma 3.7. Given a finite-alphabet process {Xn } (not necessarily stationary) with distribution m, let Xkn = (Xk , Xk+1 , . . . , Xk+n−1 ) denote the random vectors giving a block of samples of dimension n starting at time k. Then the random variables n−1 ln m(Xkn ) are m-uniformly integrable (uniform in k and n). Proof: For each nonnegative integer r define the sets Er (k, n) = {x : − and hence if x ∈ Er (k, n) then
1 ln m(xkn ) ∈ [r , r + 1)} n
76
3 Entropy
r ≤−
1 ln m(xkn ) < r + 1 n
or e−nr ≥ m(xkn ) > e−n(r +1) . Thus for any r Z 1 − ln m(Xkn ) dm < (r + 1)m(Er (k, n)) n Er (k,n) X X m(xkn ) ≤ (r + 1) e−nr = (r + 1) xkn ∈Er (k,n)
xkn
= (r + 1)e−nr kAkn ≤ (r + 1)e−nr , where the final step follows since there are at most kAkn possible ntuples corresponding to thin cylinders in Er (k, n) and by construction each has probability less than e−nr . To prove uniform integrability we must show uniform convergence to 0 as r → ∞ of the integral Z γr (k, n) =
≤
1 x:− n ln m(xkn )≥r
∞ X
(−
∞ Z X 1 1 (− ln m(Xkn ))dm ln m(Xkn ))dm = n n Er +i (k,n) i=0
(r + i + 1)e−n(r +i) kAkn ≤
i=0
∞ X
(r + i + 1)e−n(r +i−ln kAk) .
i=0
Taking r large enough so that r > ln kAk, then the exponential term is bound above by the special case n = 1 and we have the bound γr (k, n) ≤
∞ X
(r + i + 1)e−(r +i−ln kAk)
i=0
a bound which is finite and independent of k and n. The sum can easily be shown to go to zero as r → ∞ using standard summation formulas. (The exponential terms shrink faster than the linear terms grow.) 2
Variational Description of Divergence Divergence has a variational characterization that is a fundamental property for its applications to large deviations theory [182] [31]. Although this theory will not be treated here, the basic result of this section provides an alternative description of divergence and hence of relative entropy that has intrinsic interest. The basic result is originally due to Donsker and Varadhan [35].
3.3 Basic Properties of Entropy
77
Suppose now that P and M are two probability measures on a common discrete probability space, say (Ω, B). Given any real-valued random variable Φ defined on the probability space, we will be interested in the quantity EM eΦ . (3.17) which is called the cumulant generating function of Φ with respect to M and is related to the characteristic function of the random variable Φ as well as to the moment generating function and the operational transform of the random variable. The following theorem provides a variational description of divergence in terms of the cumulant generating function. Theorem 3.2. D(P kM) = sup EP Φ − ln(EM (eΦ )) . Φ
(3.18)
Proof: First consider the random variable Φ defined by Φ(ω) = ln(P (ω)/M(ω)) and observe that EP Φ − ln(EM (eΦ )) =
X ω
P (ω) ln
X P (ω) P (ω) − ln( M(ω) ) M(ω) M(ω) ω
= D(P kM) − ln 1 = D(P kM). This proves that the supremum over all Φ is no smaller than the divergence. To prove the other half observe that for any bounded random variable Φ, ! ! X M Φ (ω) eΦ Φ EP Φ − ln EM (e ) = EP ln = , P (ω) ln EM (eΦ ) M(ω) ω where the probability measure M Φ is defined by M(ω)eΦ(ω) M Φ (ω) = P . Φ(x) x M(x)e We now have for any Φ that D(P kQ) − EP Φ − ln(EM (eΦ )) = ! X X P (ω) M Φ (ω) − = P (ω) ln P (ω) ln M(ω) M(ω) ω ω X P (ω) ≥0 P (ω) ln Φ M (ω) ω
78
3 Entropy
using the divergence inequality. Since this is true for any Φ, it is also true for the supremum over Φ and the theorem is proved. 2
3.4 Entropy Rate For simplicity we focus on the entropy rate of a directly given finitealphabet random process {Xn }. We also will emphasize stationary measures, but we will try to clarify those results that require stationarity and those that are more general. As a reminder, we recall the underlying structure. Let A be a finite set. Let Ω = AZ+ and let B be the sigma-field of subsets of Ω generated by the rectangles. Since A is finite, (A, BA ) is standard, where BA is the power set of A. Thus (Ω, B) is also standard by Lemma 2.4.1 of [55] or Lemma 3.7 of [58]. In fact, from the proof that cartesian products of standard spaces are standard, we can take as a basis for B the fields Fn generated by the finite dimensional rectangles having the form {x : X n (x) = x n = an } for all an ∈ An and all positive integers n. (Members of this class of rectangles are called thin cylinders.) The union of all such fields, say F , is then a generating field. Again let {Xn ; n = 0, 1, . . .} denote a finite-alphabet random process and apply Lemma 3.2 to vectors and obtain H(X0 , X1 , . . . , Xn−1 ) ≤ H(X0 , X1 , . . . , Xm−1 ) + H(Xm , Xm+1 , . . . , Xn−1 ); 0 < m < n.
(3.19)
Define as usual the random vectors Xkn = (Xk , Xk+1 , . . . , Xk+n−1 ), that is, Xkn is a vector of dimension n consisting of the samples of X from k to k + n − 1. If the underlying measure is stationary, then the distributions of the random vectors Xkn do not depend on k. Hence if we define the sequence h(n) = H(X n ) = H(X0 , . . . , Xn−1 ), then the above equation becomes h(k + n) ≤ h(k) + h(n); all k, n > 0. Thus h(n) is a subadditive sequence as treated in Section 7.5 of [55] or Section 8.5 of [58]. A basic property of subadditive sequences is that the limit h(n)/n as n → ∞ exists and equals the infimum of h(n)/n over n. (See, e.g., Lemma 7.5.1 of [55] or Lemma 8.5.3 of [58].) This immediately yields the following result. Lemma 3.8. If the distribution m of a finite-alphabet random process {Xn } is stationary, then 1 1 Hm (X n ) = inf Hm (X n ). n→∞ n n≥1 n
H m (X) = lim
3.4 Entropy Rate
79
Thus the limit exists and equals the infimum. The next two properties of entropy rate are primarily of interest because they imply a third property, the ergodic decomposition of entropy rate, which will be described in Theorem 3.3. They are also of some independent interest. The first result is a continuity result for entropy rate when considered as a function or functional on the underlying process distribution. The second property demonstrates that entropy rate is acS T tually an affine functional (both convex and convex ) of the underT lying distribution, even though finite order entropy was only convex and not affine. We apply the distributional distance described in Section 1.7 to the Z standard sequence measurable space (Ω, B) = (AZ+ , BA+ ) with a σ -field generated by the countable field F = {Fn ; n = 1, 2, . . .} generated by all thin rectangles. Corollary 3.4. The entropy rate H m (X) of a discrete alphabet random process considered as a functional of stationary measures is upper semicontinuous; that is, if probability measures m and mn , n = 1, 2, . . . have the property that d(m, mn ) → 0 as n → ∞, then H m (X) ≥ lim sup H mn (X). n→∞
Proof: For each fixed n Hm (X n ) = −
X
m(X n = an ) ln m(X n = an )
an ∈An
is a continuous function of m since for the distance to go to zero, the probabilities of all thin rectangles must go to zero and the entropy is the sum of continuous real-valued functions of the probabilities of thin rectangles. Thus we have from Lemma 3.8 that if d(mk , m) → 0, then 1 1 Hm (X n ) = inf lim Hmk (X n ) n n n n k→∞ 1 n ≥ lim sup inf Hmk (X ) = lim sup H mk (X). n n k→∞ k→∞
H m (X) = inf
2 The next lemma uses Lemma 3.4 to show that entropy rates are affine functions of the underlying probability measures. Lemma 3.9. Let m and p denote two distributions for a discrete alphabet random process {Xn }. Then for any λ ∈ (0, 1), λHm (X n ) + (1 − λ)Hp (X n ) ≤ Hλm+(1−λ)p (X n ) ≤ λHm (X n ) + (1 − λ)Hp (X n ) + h2 (λ),
(3.20)
80
3 Entropy
and Z lim sup(− n→∞
1 ln(λm(X n (x)) + (1 − λ)p(X n (x)))) n Z 1 = lim sup − dm(x) ln m(X n (x)) = H m (X). n n→∞
dm(x)
(3.21)
If m and p are stationary then H λm+(1−λ)p (X) = λH m (X) + (1 − λ)H p (X)
(3.22)
and hence the entropy rate of a stationary discrete alphabet random process is an affine function of the process distribution. Comment: Eq. (3.20) is simply Lemma 3.4 applied to the random vectors X n stated in terms of the process distributions. Eq. (3.21) states that if we look at the limit of the normalized log of a mixture of a pair of measures when one of the measures governs the process, then the limit of the expectation does not depend on the other measure at all and is simply the entropy rate of the driving source. Thus in a sense the sequences produced by a measure are able to select the true measure from a mixture. Proof: Eq. (3.20) is just Lemma 3.4. Dividing by n and taking the limit as n → ∞ proves that entropy rate is affine. Similarly, take the limit supremum in expressions (3.13) and (3.14) and the lemma is proved. 2 We are now prepared to prove one of the fundamental properties of entropy rate, the fact that it has an ergodic decomposition formula similar to property (c) of Theorem 1.5 when it is considered as a functional on the underlying distribution. In other words, the entropy rate of a stationary source is given by an integral of the entropy rates of the stationary ergodic components. This is a far more complicated result than property (c) of the ordinary ergodic decomposition because the entropy rate depends on the distribution; it is not a simple function of the underlying sequence. The result is due to Jacobs [80]. Theorem 3.3. The Ergodic Decomposition of Entropy Rate Let (AZ+ , B(A)Z+ , m, T ) be a stationary dynamical system corresponding to a stationary finite alphabet source {Xn }. Let {px } denote the ergodic decomposition of m. If H px (X) is m-integrable, then Z H m (X) = dm(x)H px (X). Proof: The theorem follows immediately from Corollary 3.4 and Lemma 3.9 and the ergodic decomposition of semi-continuous affine funtionals as in Theorem 8.9.1 of [55] or Theorem 8.5 of [58]. 2
3.5 Relative Entropy Rate
81
3.5 Relative Entropy Rate The properties of relative entropy rate are more difficult to demonstrate. In particular, the obvious analog to (3.19) does not hold for relative entropy rate without the requirement that the reference measure be memoryless, and hence one cannot immediately infer that the relative entropy rate is given by a limit for stationary sources. The following lemma provides a condition under which the relative entropy rate is given by a limit. The condition, that the dominating measure be a kth order (or k-step) Markov source will occur repeatedly when dealing with relative entropy rates. A discrete alphabet source is kth order Markov or k-step Markov (or simply Markov if k is clear from context) if for any n and any N≥k P (Xn = xn |Xn−1 = xn−1 , . . . , Xn−N = xn−N ) = P (Xn = xn |Xn−1 = xn−1 , . . . , Xn−k = xn−k ); that is, conditional probabilities given the infinite past depend only on the most recent k symbols. A 0-step Markov source is a memoryless source. A Markov source is said to have stationary transitions if the above conditional probabilities do not depend on n, that is, if for any n P (Xn = xn |Xn−1 = xn−1 , . . . , Xn−N = xn−N ) = P (Xk = xn |Xk−1 = xn−1 , . . . , X0 = xn−k ). Lemma 3.10. If p is a stationary process and m is a k-step Markov process with stationary transitions, then H pkm (X) = lim
n→∞
1 Hpkm (X n ) = −H p (X) − Ep [ln m(Xk |X k )], n
where Ep [ln m(Xk |X k )] is an abbreviation for Ep [ln m(Xk |X k )] =
X
pX k+1 (x k+1 ) ln mXk |X k (xk |x k ).
x k+1
Proof: If for any n it is not true that mX n pX n , then Hpkm (X n ) = ∞ for that and all larger n and both sides of the formula are infinite, hence we assume that all of the finite dimensional distributions satisfy the absolute continuity relation. Since m is Markov, mX n (x n ) =
n−1 Y l=k
Thus
mXl |X l (xl |x l )mX k (x k ).
82
3 Entropy
1 1 1X pX n (x n ) ln mX n (x n ) Hpkm (X n ) = − Hp (X n ) − n n n xn 1X 1 pX k (x k ) ln mX k (x k ) = − Hp (X n ) − n n k x
n−k X pX k+1 (x k+1 ) ln mXk |X k (xk |x k ). − n k+1 x
Taking limits then yields H pkm (X) = −H p −
X
pX k+1 (x k+1 ) ln mXk |X k (xk |x k ),
x k+1
where the sum is well defined because if mXk |X k (xk |x k ) = 0, then so 2 must pX k+1 (x k+1 ) = 0 from absolute continuity. Combining the previous lemma with the ergodic decomposition of entropy rate yields the following corollary. Corollary 3.5. The Ergodic Decomposition of Relative Entropy Rate Let (AZ+ , B(A)Z+ , p, T ) be a stationary dynamical system corresponding to a stationary finite alphabet source {Xn }. Let m be a kth order Markov process for which mX n pX n for all n. Let {px } denote the ergodic decomposition of p. If H px km (X) is p-integrable, then Z H pkm (X) = dp(x)H px km (X).
3.6 Conditional Entropy and Mutual Information We now turn to other notions of information. While we could do without these if we confined interest to finite-alphabet processes, they will be essential for later generalizations and provide additional intuition and results even in the finite alphabet case. We begin by adding a second finite-alphabet measurement to the setup of the previous sections. To conform more to information theory tradition, we consider the measurements as finite-alphabet random variables X and Y rather than f and g. This has the advantage of releasing f and g for use as functions defined on the random variables: f (X) and g(Y ). It should be kept in mind, however, that X and Y could themselves be distinct measurements on a common random variable. This interpretation will often be useful when comparing codes. Let (Ω, B, P , T ) be a dynamical system. Let X and Y be finite-alphabet measurements defined on Ω with alphabets AX and AY . Define the conditional entropy of X given Y by
3.6 Conditional Entropy and Mutual Information
83
H(X|Y ) ≡ H(X, Y ) − H(Y ). The name conditional entropy comes from the fact that X P (X = a, Y = b) ln P (X = a|Y = b) H(X|Y ) = − x,y
=−
X
pX,Y (x, y) ln pX|Y (x|y)
x,y
" =−
X
# X
pY (y)
y
pX|Y (x | y) ln pX|Y (x|y) ,
x
where pX,Y (x, y) is the joint PMF for (X, Y ) and pX|Y (x|y) = pX,Y (x, y)/pY (y) is the conditional PMF. Defining X H(X|Y = y) = − pX|Y (x|y) ln pX|Y (x|y) x
we can also write H(X|Y ) =
X
pY (y)H(X|Y = y).
y
Thus conditional entropy is an average of entropies with respect to conditional PMF’s. We have immediately from Lemma 3.2 and the definition of conditional entropy that 0 ≤ H(X|Y ) ≤ H(X).
(3.23)
The inequalities could also be written in terms of the partitions induced by X and Y . Recall that according to Lemma 3.2 the right hand inequality will be an equality if and only if X and Y are independent. Define the average mutual information between X and Y by I(X; Y ) = H(X) + H(Y ) − H(X, Y ) = H(X) − H(X|Y ) = H(Y ) − H(Y |X). In terms of distributions and PMF’s we have that
84
3 Entropy
I(X; Y ) =
X
P (X = x, Y = y) ln
x,y
=
X
pX,Y (x, y) ln
X pX,Y (x, y) pX|Y (x|y) = pX,Y (x, y) ln pX (x)pY (y) x,y pX (x)
pX,Y (x, y) ln
pY |X (y|x) . pY (y)
x,y
=
X x,y
P (X = x, Y = y) P (X = x)P (Y = y)
Note also that mutual information can be expressed as a divergence by I(X; Y ) = D(PX,Y kPX × PY ), where PX × PY is the product measure on X, Y , that is, a probability measure which gives X and Y the same marginal distributions as PXY , but under which X and Y are independent. Entropy is a special case of mutual information since H(X) = I(X; X). We can collect several of the properties of entropy and relative entropy and produce corresponding properties of mutual information. We state these in the form using measurements, but they can equally well be expressed in terms of partitions. Lemma 3.11. Suppose that X and Y are two finite-alphabet random variables defined on a common probability space. Then 0 ≤ I(X; Y ) ≤ min(H(X), H(Y )). Suppose that f : AX → A and g : AY → B are two measurements. Then I(f (X); g(Y )) ≤ I(X; Y ). Proof: The first result follows immediately from the properties of entropy. The second follows from Lemma 3.3 applied to the measurement (f , g) since mutual information is a special case of relative entropy. 2 The next lemma collects some additional, similar properties. Lemma 3.12. Given the assumptions of the previous lemma,
3.6 Conditional Entropy and Mutual Information
85
H(f (X)|X) = 0, H(X, f (X)) = H(X), H(X) = H(f (X)) + H(X|f (X), I(X; f (X)) = H(f (X)), H(X|g(Y )) ≥ H(X|Y ), I(f (X); g(Y )) ≤ I(X; Y ), H(X|Y ) = H(X, f (X, Y ))|Y ), and, if Z is a third finite-alphabet random variable defined on the same probability space, H(X|Y ) ≥ H(X|Y , Z). Comments: The first relation has the interpretation that given a random variable, there is no additional information in a measurement made on the random variable. The second and third relationships follow from the first and the definitions. The third relation is a form of chain rule and it implies that given a measurement on a random variable, the entropy of the random variable is given by that of the measurement plus the conditional entropy of the random variable given the measurement. This provides an alternative proof of the second result of Lemma 3.3. The fifth relation says that conditioning on a measurement of a random variable is less informative than conditioning on the random variable itself. The sixth relation states that coding reduces mutual information as well as entropy. The seventh relation is a conditional extension of the second. The eighth relation says that conditional entropy is nonincreasing when conditioning on more information. Proof: Since g(X) is a deterministic function of X, the conditional PMF is trivial (a Kronecker delta) and hence H(g(X)|X = x) is 0 for all x, hence the first relation holds. The second and third relations follow from the first and the definition of conditional entropy. The fourth relation follows from the first since I(X; Y ) = H(Y ) − H(Y |X). The fifth relation follows from the previous lemma since H(X) − H(X|g(Y )) = I(X; g(Y )) ≤ I(X; Y ) = H(X) − H(X|Y ). The sixth relation follows from Corollary 3.3 and the fact that I(X; Y ) = D(PX,Y kPX × PY ). The seventh relation follows since H(X, f (X, Y ))|Y ) = H(X, f (X, Y )), Y ) − H(Y ) = H(X, Y ) − H(Y ) = H(X|Y ).
86
3 Entropy
The final relation follows from the second by replacing Y by Y , Z and setting g(Y , Z) = Y . 2 In a similar fashion we can consider conditional relative entropies. Suppose now that M and P are two probability measures on a common space, that X and Y are two random variables defined on that space, and that MXY PXY (and hence also MX PY ). Analagous to the definition of the conditional entropy we can define HP kM (X|Y ) ≡ HP kM (X, Y ) − HP kM (Y ). Some algebra shows that this is equivalent to HP kM (X|Y ) =
X
=
X
pX,Y (x, y) ln
x,y
pY (y)
y
X x
pX|Y (x|y) mX|Y (x|y)
pX|Y (x|y) pX|Y (x|y) ln mX|Y (x|y)
! .
(3.24)
This can be written as HP kM (X|Y ) =
X
pY (y)D(pX|Y (·|y)kmX|Y (·|y)),
y
an average of divergences of conditional PMF’s, each of which is well defined because of the original absolute continuity of the joint measure. Manipulations similar to those for entropy can now be used to prove the following properties of conditional relative entropies. Lemma 3.13. Given two probability measures M and P on a common space, and two random variables X and Y defined on that space with the property that MXY PXY , then the following properties hold: HP kM (f (X)|X) = 0, HP kM (X, f (X)) = HP kM (X), HP kM (X) = HP kM (f (X)) + HP kM (X|f (X)),
(3.25)
If MXY = MX ×MY (that is, if the PMFs satisfy mX,Y (x, y) = mX (x)mY (y)), then HP kM (X, Y ) ≥ HP kM (X) + HP kM (Y ) and HP kM (X|Y ) ≥ HP kM (X). Eq. (3.25) is a chain rule for relative entropy which provides as a corollary an immediate proof of Lemma 3.3. The final two inequalities resemble inequalities for entropy (with a sign reversal), but they do not hold for all reference measures.
3.6 Conditional Entropy and Mutual Information
87
The above lemmas along with Lemma 3.3 show that all of the information measures thus far considered are reduced by taking measurements or by coding. This property is the key to generalizing these quantities to nondiscrete alphabets. T We saw in Lemma 3.4 that entropy was a convex function of the underlying distribution. The following lemma provides similar properties of mutual information considered as a function of either a marginal or a conditional distribution. Lemma 3.14. Let µ denote a PMF on a discrete space AX , µ(x) = Pr(X = x), and let ν be a conditional PMF, ν(y|x) = Pr(Y = y|X = x). Let µν denote the resulting joint PMF µν(x, y) = µ(x)ν(y|x). Let Iµν = S Iµν (X; Y ) be the average mutual information. Then Iµν is a convex function of ν; that is, given two conditional PMF’s ν1 and ν2 , a λ ∈ [0, 1], and ν = λν1 + (1 − λ)ν2 , then Iµν ≤ λIµν1 + (1 − λ)Iµν2 , T function of µ, that is, given two PMF’s µ1 and µ2 , and Iµν is a convex λ ∈ [0, 1], and µ = λµ1 + (1 − λ)µ2 , Iµν ≥ λIµ1 ν + (1 − λ)Iµ2 ν . Proof. First consider the case of a fixed source µ and consider conditional PMFs ν1 , ν2 , and for 0 ≤ λ ≤ 1 define ν = λν1 + (1 − λ)ν2 . Denote the corresponding input/output pair processes by pi = µνi , i = 1, 2, and p = λp1 + (1 − λ)p2 Let η (respectively, η1 , η2 , η) denote the PMF for Y P resulting from ν (respectively ν1 , ν2 , ν), that is, η(y) = Pr(Y = y) = x µ(x)ν(y|x). Note that p1 , p2 , and p all have a common input marginal PMF µ. We have that µ × η = λµ × η1 + (1 − λ)µ × η2 so that from Lemma 3.5 Iµν = D(µν||µ × η) = D(λp1 + (1 − λ)p2 kλµ × η1 + (1 − λ)µ × η2 ) ≤ λD(p1 kµ × η1 ) + (1 − λ)D(p2 kµ × η2 ) = λIµν1 + (1 − λ)Iµν2 , proving the convexity of mutual information with respect to the channel. The author is indebted to T. Linder for suggesting this proof, which is much simpler than the one in the first edition. Similarly, let µ = λµ1 + (1 − λ)µ2 and let η1 , η2 , and η denote the corresponding induced output PMF’s. Then
88
3 Entropy
Iµν = λ
X
µ1 (x)ν(y|x) log
x,y
+ (1 − λ)
X
ν(y|x) η1 (y) ν(y|x) η(y) ν(y|x) η1 (y)
ν(y|x) η2 (y) ν(y|x) η(y) ν(y|x) η2 (y)
µ2 (x)ν(y|x) log
x,y
= λIµ1 ν + (1 − λ)Iµ2 ν − λ
X
µ1 (x)ν(y|x) log
x,y
− (1 − λ)
X
!
µ2 (x)ν(y|x) log
x,y
!
η(y) η1 (y)
η(y) η2 (y)
= λIµ1 ν + (1 − λ)Iµ2 ν + λD(η1 kη) + (1 − λ)D(η2 kη) ≥ λIµ1 ν + (1 − λ)Iµ2 ν
(3.26)
2 from the divergence inequality. We consider one other notion of information: Given three finitealphabet random variables X, Y , Z, define the conditional mutual information between X and Y given Z by I(X; Y |Z) = D(PXY Z kPX×Y |Z )
(3.27)
where PX×Y |Z is the distribution defined by its values on rectangles as PX×Y |Z (F × G × D) =
X
P (X ∈ F |Z = z)P (Y ∈ G|Z = z)P (Z = z).
z∈D
(3.28) PX×Y |Z has the same conditional distributions for X given Z and for Y given Z as does PXY Z , but now X and Y are conditionally independent given Z. Alternatively, the conditional distribution for X, Y given Z under the distribution PX×Y |Z is the product distribution PX |Z × PY |Z. Thus I(X; Y |Z) =
X
pXY Z (x, y, z) ln
pXY Z (x, y, z) pX|Z (x|z)pY |Z (y|z)pZ (z)
pXY Z (x, y, z) ln
pXY |Z (x, y|z) . pX|Z (x|z)pY |Z (y|z)
x,y,z
=
X x,y,z
Since
(3.29)
pXY Z pXY Z pX pXY Z pY = × = × pX|Z pY |Z pZ pX pY Z pX|Z pXZ pY pY |Z
we have the first statement in the following lemma. Lemma 3.15. I(X; Y |Z) + I(Y ; Z) = I(Y ; (X, Z)),
(3.30)
I(X; Y |Z) ≥ 0,
(3.31)
3.6 Conditional Entropy and Mutual Information
89
with equality if and only if X and Y are conditionally independent given Z, that is, pXY |Z = pX|Z pY |Z . Given finite valued measurements f and g, I(f (X); g(Y )|Z) ≤ I(X; Y |Z). Proof: The second inequality follows from the divergence inequality (3.7) with P = PXY Z and M = PX×Y |Z , i.e., the PMF’s pXY Z and pX|Z pY |Z pZ . The third inequality follows from Lemma 3.3 or its corollary applied to the same measures. 2 Comments: Eq. (3.30) is called Kolmogorov’s formula. If X and Y are conditionally independent given Z in the above sense, then we also have that pX|Y Z = pXY |Z /pY |Z = pX|Z , in which case Y → Z → X forms a Markov chain — given Z, X does not depend on Y . Note that if Y → Z → X is a Markov chain, then so is X → Z → Y . Thus the conditional mutual information is 0 if and only if the variables form a Markov chain with the conditioning variable in the middle. One might be tempted to infer from Lemma 3.3 that given finite valued measurements f , g, and r I(f (X); g(Y )|r (Z))
(?) I(X; Y |Z). ≤
This does not follow, however, since it is not true that if Q is the partition corresponding to the three quantizers, then D(Pf (X),g(Y ),r (Z)kPf (X)×g(Y )|r (Z)) is HPX,Y ,Z kPX×Y |Z (f (X), g(Y ), r (Z)) because of the way that PX×Y |Z is constructed; e.g., the fact that X and Y are conditionally independent given Z implies that f (X) and g(Y ) are conditionally independent given Z, but it does not imply that f (X) and g(Y ) are conditionally independent given r (Z). Alternatively, if M is PX×Z|Y , then it is not true that Pf (X)×g(Y )|r (Z) equals M(f gr )−1 . Note that if this inequality were true, choosing r (z) to be trivial (say 1 for all z) would result in I(X; Y |Z) ≥ I(X; Y |r (Z)) = I(X; Y ). This cannot be true in general since, for example, choosing Z as (X, Y ) would give I(X; Y |Z) = 0. Thus one must be careful when applying Lemma 3.3 if the measures and random variables are related as they are in the case of conditional mutual information. We close this section with an easy corollary of the previous lemma and of the definition of conditional entropy. Results of this type are referred to as chain rules for information and entropy. Corollary 3.6. Given finite-alphabet random variables Y , X1 , X2 , . . ., Xn , H(X1 , X2 , . . . , Xn ) =
n X
H(Xi |X1 , . . . , Xi−1 )
i=1
Hpkm (X1 , X2 , . . . , Xn ) =
n X i=1
Hpkm (Xi |X1 , . . . , Xi−1 )
90
3 Entropy
I(Y ; (X1 , X2 , . . . , Xn )) =
n X
I(Y ; Xi |X1 , . . . , Xi−1 ).
i=1
3.7 Entropy Rate Revisited The chain rule of Corollary 3.6 provides a means of computing entropy rates for stationary processes. We have that n−1 1 1 X H(X n ) = H(Xi |X i ). n n i=0
First suppose that the source is a stationary kth order Markov process, that is, for any m > k Pr(Xn = xn |Xi = xi ; i = 0, 1, . . . , n − 1) = Pr(Xn = xn |Xi = xi ; i = n − k, . . . , n − 1). For such a process we have for all n ≥ k that k ) = H(Xk |X k ), H(Xn |X n ) = H(Xn |Xn−k
where Xim = Xi , . . . , Xi+m−1 . Thus taking the limit as n → ∞ of the nth order entropy, all but a finite number of terms in the sum are identical and hence the Cesàro (or arithmetic) mean is given by the conditional expectation. We have therefore proved the following lemma. Lemma 3.16. If {Xn } is a stationary kth order Markov source, then H(X) = H(Xk |X k ). If we have a two-sided stationary process {Xn }, then all of the previous definitions for entropies of vectors extend in an obvious fashion and a generalization of the Markov result follows if we use stationarity and the chain rule to write n−1 1 1 X n H(X ) = H(X0 |X−1 , . . . , X−i ). n n i=0
Since conditional entropy is nonincreasing with more conditioning variables ((3.23) or Lemma 3.12), H(X0 |X−1 , . . . , X−i ) has a limit. Again using the fact that a Cesàro mean of terms all converging to a common limit also converges to the same limit we have the following result.
3.8 Markov Approximations
91
Lemma 3.17. If {Xn } is a two-sided stationary source, then H(X) = lim H(X0 |X−1 , . . . , X−n ). n→∞
It is tempting to identify the above limit as the conditional entropy given the infinite past, H(X0 |X−1 , . . .). Since the conditioning variable is a sequence and does not have a finite alphabet, such a conditional entropy is not included in any of the definitions yet introduced. We shall later demonstrate that this interpretation is indeed valid when the notion of conditional entropy has been suitably generalized. The natural generalization of Lemma 3.17 to relative entropy rates unfortunately does not work because conditional relative entropies are not in general monotonic with increased conditioning and hence the chain rule does not immediately yield a limiting argument analogous to that for entropy. The argument does work if the reference measure is a kth order Markov, as considered in the following lemma. Lemma 3.18. If {Xn } is a source described by process distributions p and m and if p is stationary and m is kth order Markov with stationary transitions, then for n ≥ k Hpkm (X0 |X−1 , . . . , X−n ) is nondecreasing in n and H pkm (X) = lim Hpkm (X0 |X−1 , . . . , X−n ) n→∞
= −H p (X) − Ep [ln m(Xk |X k )]. Proof: For n ≥ k we have that Hpkm (X0 |X−1 , . . . , X−n ) = − Hp (X0 |X−1 , . . . , X−n ) −
X
pX k+1 (x k+1 ) ln mXk |X k (xk |x k ).
x k+1
Since the conditional entropy is nonincreasing with n and the remaining term does not depend on n, the combination is nondecreasing with n. The remainder of the proof then parallels the entropy rate result. 2 It is important to note that the relative entropy analogs to entropy properties often require kth order Markov assumptions on the reference measure (but not on the original measure).
3.8 Markov Approximations Recall that the relative entropy rate H pkm (X) can be thought of as a distance between the process with distribution p and that with distribution m and that the rate is given by a limit if the reference measure m is Markov. A particular Markov measure relevant to p is the distribution
92
3 Entropy
p (k) which is the kth order Markov approximation to p in the sense that it is a kth order Markov source and it has the same kth order transition probabilities as p. To be more precise, the process distribution p (k) is specified by its finite dimensional distributions (k)
pX k (x k ) = pX k (x k ) (k)
pX n (x n ) = pX k (x k )
n−1 Y
k pXl |X k (xl |xl−k ); n = k, k + 1, . . . l−k
l=k
so that
(k) k k |X
pX
= pXk |X k .
It is natural to ask how good this approximation is, especially in the limit, that is, to study the behavior of the relative entropy rate H pkp(k) (X) as k → ∞. Theorem 3.4. Given a stationary process p, let p (k) denote the kth order Markov approximations to p. Then lim H pkp(k) (X) = inf H pkp(k) (X) = 0.
k→∞
k
Thus the Markov approximations are asymptotically accurate in the sense that the relative entropy rate between the source and approximation can be made arbitrarily small (zero if the original source itself happens to be Markov). Proof: As in the proof of Lemma 3.18 we can write for n ≥ k that Hpkp(k) (X0 |X−1 , . . . , X−n ) = −Hp (X0 |X−1 , . . . , X−n ) −
X
pX k+1 (x k+1 ) ln pXk |X k (xk |x k )
x k+1
= Hp (X0 |X−1 , . . . , X−k ) − Hp (X0 |X−1 , . . . , X−n ). (k)
Note that this implies that pX n pX n for all n since the entropies are finite. This automatic domination of the finite dimensional distributions of a measure by those of its Markov approximation will not hold in the general case to be encountered later, it is specific to the finite alphabet case. Taking the limit as n → ∞ gives H pkp(k) (X) = lim Hpkp(k) (X0 |X−1 , . . . , X−n ) n→∞
= Hp (X0 |X−1 , . . . , X−k ) − H p (X). The corollary then follows immediately from Lemma 3.17.
2
Markov approximations will play a fundamental role when considering relative entropies for general (nonfinite-alphabet) processes. The ba-
3.9 Relative Entropy Densities
93
sic result above will generalize to that case, but the proof will be much more involved.
3.9 Relative Entropy Densities Many of the convergence results to come will be given and stated in terms of relative entropy densities. In this section we present a simple but important result describing the asymptotic behavior of relative entropy densities. Although the result of this section is only for finite alphabet processes, it is stated and proved in a manner that will extend naturally to more general processes later on. The result will play a fundamental role in the basic ergodic theorems to come. Throughout this section we will assume that M and P are two process distributions describing a random process {Xn }. Denote as before the sample vector X n = (X0 , X1 , . . . , Xn−1 ), that is, the vector beginning at time 0 having length n. The distributions on X n induced by M and P will be denoted by Mn and Pn , respectively. The corresponding PMF’s are mX n and pX n . The key assumption in this section is that for all n if mX n (x n ) = 0, then also pX n (x n ) = 0, that is, Mn Pn for all n.
(3.32)
If this is the case, we can define the relative entropy density hn (x) ≡ ln where fn (x) ≡
pX n (x n ) = ln fn (x), mX n (x n )
pX n (x n )
if mX n (x n ) 6= 0
0
otherwise
mX n (x n )
(3.33)
(3.34)
Observe that the relative entropy is found by integrating the relative entropy density: HP kM (X n ) = D(Pn kMn ) =
X
pX n (x n ) ln
xn n
Z =
pX n (x n ) mX n (x n )
ln
pX n (X ) dP mX n (X n )
(3.35)
Thus, for example, if we assume that HP kM (X n ) < ∞, all n, then (3.32) holds.
(3.36)
94
3 Entropy
The following lemma will prove to be useful when comparing the asymptotic behavior of relative entropy densities for different probability measures. It is the first almost everywhere result for relative entropy densities that we consider. It is somewhat narrow in the sense that it only compares limiting densities to zero and not to expectations. We shall later see that essentially the same argument implies the same result for the general case (Theorem 7.4), only the interim steps involving PMF’s need be dropped. Note that the lemma requires neither stationarity nor asymptotic mean stationarity. Lemma 3.19. Given a finite-alphabet process {Xn } with process measures P , M satisfying (3.32), Then lim sup n→∞
and lim inf n→∞
1 hn ≤ 0, M − a.e. n
(3.37)
1 hn ≥ 0, P − a.e.. n
(3.38)
If in addition M P , then lim
n→∞
1 hn = 0, P − a.e.. n
(3.39)
Proof: First consider the probability M(
EM (fn ) 1 hn ≥ ) = M(fn ≥ en ) ≤ , n en
where the final inequality is Markov’s inequality. But Z X pX n (x n ) mX n (x n ) EM (fn ) = dMfn = mX n (x n ) x n : mX n (x n )6=0 X pX n (x n ) ≤ 1 = x n : mX n (x n )6=0
and therefore M( and hence
∞ X n=1
M(
1 hn ≥ ) ≤ 2−n n
∞ X 1 e−n < ∞. hn > ) ≤ n n=1
From the Borel-Cantelli Lemma (e.g., Lemma 4.6.3 of [55] or Lemma 5.17 of [58]) this implies that M(n−1 hn ≥ i.o.) = 0 which implies the first equation of the lemma. Next consider
3.9 Relative Entropy Densities
P (−
1 hn > ) = n
95
X 1 x n :− n
pX n (x n )
ln pX n (x n )/mX n (x n )>
X
=
1 x n :− n ln pX n (x n )/mX n (x n )>
pX n (x n ) and
mX n (x n )6=0
where the last statement follows since if mX n (x n ) = 0, then also pX n (x n ) = 0 and hence nothing would be contributed to the sum. In other words, terms violating this condition add zero to the sum and hence adding this condition to the sum does not change the sum’s value. Thus 1 hn > ) n
P (−
X
=
1 x n :− n ln pX n (x n )/mX n (x n )>
Z
Z = =e
fn <e−n −n
dMfn ≤
M(fn < e
and
dMe−n
fn <e−n −n −n
)≤e
mX n (x n )6=0
pX n (x n ) mX n (x n ) mX n (x n )
.
Thus as before we have that P (n−1 hn > ) ≤ e−n and hence that P (n−1 hn ≤ − i.o.) = 0 which proves the second claim. If also M P , then the first equation of the lemma is also true P -a.e., which when coupled with the second equation proves the third. 2
Chapter 4
The Entropy Ergodic Theorem
Abstract The goal of this chapter is to prove an ergodic theorem for sample entropy of finite-alphabet random processes. The result is sometimes called the ergodic theorem of information theory or the asymptotic equipartition (AEP) theorem, but it is best known as the ShannonMcMillan-Breiman theorem. It provides a common foundation to many of the results of both ergodic theory and information theory.
4.1 History Shannon [162] first demonstrated the convergence in probability of sample entropy to the entropy rate for stationary ergodic Markov sources. McMillan [123] proved L1 convergence for stationary ergodic sources and Breiman [20] [21] proved almost everywhere convergence for stationary and ergodic sources. Billingsley [16] extended the result to stationary nonergodic sources. Jacobs [79] [78] extended it to processes dominated by a stationary measure and hence to two-sided AMS processes. Gray and Kieffer [62] extended it to processes asymptotically dominated by a stationary measure and hence to all AMS processes. The generalizations to AMS processes build on the Billingsley theorem for the stationary mean. Breiman’s and Billingsley’s approach requires the martingale convergence theorem and embeds the possibly one-sided stationary process into a two-sided process. Ornstein and Weiss [141] developed a proof for the stationary and ergodic case that does not require any martingale theory and considers only positive time and hence does not require any embedding into two-sided processes. The technique was described for both the ordinary ergodic theorem and the entropy ergodic theorem by Shields [165]. In addition, it uses a form of coding argument that is both more direct and more information theoretic in flavor than the traditional martingale proofs. We here follow the Ornstein and Weiss approach for R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_4, © Springer Science+Business Media, LLC 2011
97
98
4 The Entropy Ergodic Theorem
the stationary ergodic result. We also use some modifications similar to those of Katznelson and Weiss for the proof of the ergodic theorem. We then generalize the result first to nonergodic processes using the “sandwich” technique of Algoet and Cover [7] and then to AMS processes using a variation on a result of [62]. We next state the theorem to serve as a guide through the various steps. We also prove the result for the simple special case of a Markov source, for which the result follows from the usual ergodic theorem. We consider a directly given finite-alphabet source {Xn } described by a distribution m on the sequence measurable space (Ω, B). Define as previously Xkn = (Xk , Xk+1 , · · · , Xk+n−1 ). The subscript is omitted when it is zero. For any random variable Y defined on the sequence space (such as Xkn ) we define the random variable m(Y ) by m(Y )(x) = m(Y = Y (x)). Theorem 4.1. The Entropy Ergodic Theorem Given a finite-alphabet AMS source {Xn } with process distribution m and stationary mean m, let {mx ; x ∈ Ω} be the ergodic decomposition of the stationary mean m. Then − ln m(X n ) = h; m − a.e. and in L1 (m), n→∞ n lim
(4.1)
where h(x) is the invariant function defined by h(x) = H mx (X).
(4.2)
Furthermore, Em h = lim
n→∞
1 Hm (X n ) = H m (X); n
(4.3)
that is, the entropy rate of an AMS process is given by the limit, and H m (X) = H m (X).
(4.4)
Comments: The theorem states that the sample entropy using the AMS measure m converges to the entropy rate of the underlying ergodic component of the stationary mean. Thus, for example, if m is itself stationary and ergodic, then the sample entropy converges to the entropy rate of the process m-a.e. and in L1 (m). The L1 (m) convergence follows immediately from the almost everywhere convergence and the fact that sample entropy is uniformly integrable (Lemma 3.7). L1 convergence in turn immediately implies the left-hand equality of (4.3). Since the limit exists, it is the entropy rate. The final equality states that the entropy rates of an AMS process and its stationary mean are the same. This result follows from (4.2)-(4.3) by the following argument: We have that H m (X) = Em h and H m (X) = Em h, but h is invariant and hence the two expectations are
4.1 History
99
equal (see, e.g., Lemma 6.3.1 of [55] or Lemma 7.5 of [58]). Thus we need only prove almost everywhere convergence in (4.1) to prove the theorem. In this section we limit ourselves to the following special case of the theorem that can be proved using the ordinary ergodic theorem without any new techniques. Lemma 4.1. Given a finite-alphabet stationary kth order Markov source {Xn }, then there is an invariant function h such that − ln m(X n ) = h; m − a.e. and in L1 (m), n→∞ n lim
where h is defined by h(x) = −Emx ln m(Xk |X k ),
(4.5)
where {mx } is the ergodic decomposition of the stationary mean m. Furthermore, h(x) = H mx (X) = Hmx (Xk |X k ). (4.6) Proof of Lemma: We have that −
n−1 1 X 1 ln m(Xi |X i ). ln m(X n ) = − n n i=0
Since the process is kth order Markov with stationary transition probabilites, for i > k we have that m(Xi |X i ) = m(Xi |Xi−k , · · · , Xi−1 ) = m(Xk |X k )T i−k . The terms − ln m(Xi |X i ), i = 0, 1, · · · , k − 1 have finite expectation and hence are finite m-a.e. so that the ergodic theorem can be applied to deduce − ln m(X n )(x) n
= −
k−1 n−1 1 X 1 X ln m(Xi |X i )(x) − ln m(Xk |X k )(T i−k x) n i=0 n i=k
k−1 n−k−1 1 X 1 X i = − ln m(Xi |X )(x) − ln m(Xk |X k )(T i x) n i=0 n i=0
→ Emx (− ln m(Xk |X k )),
n→∞
proving the first statement of the lemma. It follows from the ergodic decomposition of Markov sources (see Lemma 8.6.3 of [55] or Lemma 10.5 of [58]) that with probability 1, mx (Xk |X k ) = m(Xk |ψ(x), X k ) = m(Xk |X k ), where ψ is the ergodic component function. This completes the proof. 2
100
4 The Entropy Ergodic Theorem
We prove the theorem in three steps: The first step considers stationary and ergodic sources and uses the approach of Ornstein and Weiss [141] (see also Shields [165]). The second step removes the requirement for ergodicity. This result will later be seen to provide an information theoretic interpretation of the ergodic decomposition. The third step extends the result to AMS processes by showing that such processes inherit limiting sample entropies from their stationary mean. The later extension of these results to more general relative entropy and information densities will closely parallel the proofs of the second and third steps for the finite case. In subsequent chapters the definitions of entropy and information will be generalized and corresponding generalizations of the entropy ergodic theorem will be developed in Chapter 11.
4.2 Stationary Ergodic Sources This section is devoted to proving the entropy ergodic theorem for the special case of stationary ergodic sources. The result was originally proved by Breiman [20]. The original proof first used the martingale convergence theorem to infer the convergence of conditional probabilities of the form m(X0 |X−1 , X−2 , · · · , X−k ) to m(X0 |X−1 , X−2 , · · · ). This result was combined with an an extended form of the ergodic theorem stating that if gk → g as k → ∞ and if gk is L1 -dominated (supk |gk | is in Pn−1 Pn−1 L1 ), then 1/n k=0 gk T k has the same limit as 1/n k=0 gT k . Combining these facts yields that that n−1 n−1 1 1 X 1 X k )T k ln m(X n ) = ln m(Xk |X k ) = ln m(X0 |X−k n n k=0 n k=0
has the same limit as n−1 1 X ln m(X0 |X−1 , X−2 , · · · )T k n k=0
which, from the usual ergodic theorem, is the expectation E(ln m(X0 |X− ) ≡ E(ln m(X0 |X−1 , X−2 , · · · )). As suggested at the end of the preceeding chapter, this should be minus the conditional entropy H(X0 |X−1 , X−2 , · · · ) which in turn should be the entropy rate H X . This approach has three shortcomings: it requires a result from martingale theory which has not been proved here or in the companion volume [55] or [58], it requires an extended ergodic theo-
4.2 Stationary Ergodic Sources
101
rem which has similarly not been proved here, and it requires a more advanced definition of entropy which has not yet been introduced. Another approach is the sandwich proof of Algoet and Cover [7]. They show without theory or the extended ergodic theorem that Pn−1using martingale i 1/n i=0 ln m(X0 |X−i )T i is asymptotically sandwiched between the entropy rate of a kth order Markov approximation: n−1 1 X k k k ln m(X0 |X−k )T i → Em [ln m(X0 |X−k )] = −H(X0 |X−k ) n→∞ n i=k
and n−1 1 X ln m(X0 |X−1 , X−2 , · · · )T i → Em [ln m(X0 |X1 , · · · )] n→∞ n i=k
= −H(X0 |X−1 , X−2 , · · · ). By showing that these two limits are arbitrarily close as k → ∞, the result is proved. The drawback of this approach for present purposes is that again the more advanced notion of conditional entropy given the infinite past is required. Algoet and Cover’s proof that the above two entropies are asymptotically close involves martingale theory, but this can be avoided by using Corollary 7.4 as will be seen. The result can, however, be proved without martingale theory, the extended ergodic theorem, or advanced notions of entropy using the approach of Ornstein and Weiss [141], which is the approach we shall take in this chapter. In a later chapter when the entropy ergodic theorem is generalized to nonfinite alphabets and the convergence of entropy and information densities is proved, the sandwich approach will be used since the appropriate general definitions of entropy will have been developed and the necessary side results will have been proved. Lemma 4.2. Given a finite-alphabet source {Xn } with a stationary ergodic distribution m, we have that − ln m(X n ) = h; m − a.e., n→∞ n lim
where h(x) is the invariant function defined by h(x) = H m (X). Proof: Define hn (x) = − ln m(X n )(x) = − ln m(x n ) and h(x) = lim inf n→∞
1 − ln m(x n ) hn (x) = lim inf . n→∞ n n
102
4 The Entropy Ergodic Theorem
Since m((x0 , · · · , xn−1 )) ≤ m((x1 , · · · , xn−1 )), we have that hn (x) ≥ hn−1 (T x). Dividing by n and taking the limit infimum of both sides shows that h(x) ≥ h(T x). Since the n−1 hn are nonnegative and uniformly integrable (Lemma 3.7), we can use Fatou’s lemma to deduce that h and hence also hT are integrable with respect to m. Integrating with respect to the stationary measure m yields Z Z dm(x)h(x) = dm(x)h(T x) which can only be true if h(x) = h(T x); m − a.e., that is, if h is an invariant function with m-probability one. If h is invariant almost everywhere, however, it must be a constant with probability one since m is ergodic (Lemma 6.7.1 of [55] or Lemma 7.12 of [58]). Since it has a finite integral (bounded by H m (X)), h must also be finite. Henceforth we consider h to be a finite constant. We now proceed with steps that resemble those of the proof of the ergodic theorem in Section 7.2 of [55] or Section 8.1 or [58]. Fix > 0. We also choose for later use a δ > 0 small enough to have the following properties: If A is the alphabet of X0 and kAk is the finite cardinality of the alphabet, then δ ln kAk < , (4.7) and −δ ln δ − (1 − δ) ln(1 − δ) ≡ h2 (δ) < .
(4.8)
The latter property is possible since h2 (δ) → 0 as δ → 0. Define the random variable n(x) to be the smallest integer n for which n−1 hn (x) ≤ h + . As in the proof of the ergodic theorem, n(x) in general will be large in order to well approximate the limit infimum, but by definition of the limit infimum there must be infinitely many n for which the inequality is true and hence n(x) is everywhere finite, but it is not bounded. Still mimicking the proof of the ergodic theorem, define a set of “bad” sequences B = {x : n(x) > N} where N is chosen large enough to ensure that m(B) < δ/2. Define a bounded modification of n(x) by ( ˜ n(x) =
n(x)
x 6∈ B
1
x∈B
˜ so that n(x) ≤ N for all x ∈ B c . We now parse the sequence into variablelength blocks. Iteratively define nk (x) by
4.2 Stationary Ergodic Sources
103
n0 (x) = 0 ˜ n1 (x) = n(x) ˜ n1 (x) x) = n1 (x) + l1 (x) n2 (x) = n1 (x) + n(T .. . ˜ nk (x) x) = nk (x) + lk (x), nk+1 (x) = nk (x) + n(T where lk (x) is the length of the kth block: ˜ nk (x) x). lk (x) = n(T We have parsed a very long sequence x L = (x0 , · · · , xL−1 ), where l (x) L N, into long blocks xnk (x) , · · · , xnk+1 (x)−1 = xnkk (x) which begin at time nk (x) and have length lk (x) for k = 0, 1, · · · . We refer to this parsing as the block decomposition of a sequence. The kth block, which begins at time nk (x), must either have sample entropy satisfying l (x)
− ln m(xnkk (x) ) lk (x)
≤h+
(4.9)
or, equivalently, probability at least l (x)
m(xnkk (x) ) ≥ e−lk (x)(h+) ,
(4.10)
or it must consist of only a single symbol. Blocks having length 1 (lk = 1) could have the correct sample entropy, that is, 1 ) − ln m(xn k (x)
1
≤ h + ,
or they could be bad in the sense that they are the first symbol of a sequence with n > N; that is, n(T nk (x) x) > N, or, equivalently, T nk (x) x ∈ B. Except for these bad symbols, each of the blocks by construction will have a probability which satisfies the above bound. Define for nonnegative integers n and positive integers l the sets S(n, l) = {x : m(Xnl (x)) ≥ e−l(h+) }, that is, the collection of infinite sequences for which (4.9) and (4.10) hold for a block starting at n and having length l. Observe that for such blocks there cannot be more than el(h+) distinct l-tuples for which the
104
4 The Entropy Ergodic Theorem
bound holds (lest the probabilities sum to something greater than 1). In symbols this is ||S(n, l)|| ≤ el(h+) . (4.11) The ergodic theorem will imply that there cannot be too many single symbol blocks with n(T nk (x) x) > N because the event has small probability. These facts will be essential to the proof. ˜ Even though we write n(x) as a function of the entire infinite sequence, we can determine its value by observing only the prefix x N of x since either there is an n ≤ N for which n−1 ln m(x n ) ≤ h + or ˆ N ) such that n(x) ˜ ˆ N ). there is not. Hence there is a function n(x = n(x N N ˆ Define the finite length sequence event C = {x : n(x ) = 1 and − ln m(x 1 ) > h + }, that is, C is the collection of all N-tuples x N that are prefixes of bad infinite sequences, sequences x for which n(x) > N. Thus in particular, x ∈ B if and only if x N ∈ C.
(4.12)
Recall that we parse sequences of length L N and define the set GL of “good” L-tuples by GL = {x L :
1 L−N
L−N−1 X
1C (xiN ) ≤ δ},
i=0
that is, GL is the collection of all L-tuples which have fewer than δ(L − N) ≤ δL time slots i for which xiN is a prefix of a bad infinite sequence. From (4.12) and the ergodic theorem for stationary ergodic sources we know that m-a.e. we get an x for which n−1 n−1 1 X 1 X δ 1C (xiN ) = lim 1B (T i x) = m(B) ≤ . n→∞ n n→∞ n 2 i=0 i=0
lim
(4.13)
From the definition of a limit, this means that with probability 1 we get an x for which there is an L0 = L0 (x) such that 1 L−N
L−N−1 X
1C (xiN ) ≤ δ; for all L > L0 .
(4.14)
i=0
This follows because if the limit is less than δ/2, there must be an L0 so large that for larger L the time average is at least no greater than 2δ/2 = δ. We can restate (4.14) as follows: with probability 1 we get an x for which x L ∈ GL for all but a finite number of L. Stating this in negative fashion, we have one of the key properties required by the proof: If x L ∈ GL for all but a finite number of L, then x L cannot be in the complement GLc infinitely often, that is,
4.2 Stationary Ergodic Sources
105
m(x : x L ∈ GLc i.o.) = 0.
(4.15)
We now change tack to develop another key result for the proof. For each L we bounded above the cardinality ||GL || of the set of good Ltuples. By construction there are no more than δL bad symbols in an L-tuple in GL and these can occur in any of at most ! X L (4.16) ≤ eh2 (δ)L k k≤δL places, where we have used Lemma 3.6. Eq. (4.16) provides an upper bound on the number of ways that a sequence in GL can be parsed by the given rules. The bad symbols and the final N symbols in the L-tuple can take on any of the kAk different values in the alphabet. Eq. (4.11) bounds the number of finite length sequences that can occur in each of the remaining blocks and hence for any given block decomposition, the number of ways that the remaining blocks can be filled is bounded above by P Y elk (x)(h+) = e k lk (x)(h+) = eL(h+) , (4.17) k:T nk (x) x6∈B
regardless of the details of the parsing. Combining these bounds we have that ||GL || ≤ eh2 (δ)L × kAkδL × kAkN × eL(h+) = eh2 (δ)L+(δL+N) ln kAk+L(h+) or
N
||GL || ≤ eL(h++h2 (δ)+(δ+ L ) ln kAk) . Since δ satisfies (4.7)–(4.8), we can choose L1 large enough so that N ln kAk/L1 ≤ and thereby obtain ||GL || ≤ eL(h+4) ; L ≥ L1 .
(4.18)
This bound provides the second key result in the proof of the lemma. We now combine (4.18) and (4.15) to complete the proof. Let BL denote a collection of L-tuples that are bad in the sense of having too large a sample entropy or, equivalently, too small a probability; that is if x L ∈ BL , then m(x L ) ≤ e−L(h+5) or, equivalently, for any x with prefix x L hL (x) ≥ h + 5. The T upper bound on ||GL || provides a bound on the probability of BL GL :
106
4 The Entropy Ergodic Theorem
m(BL
\
X
GL ) =
x L ∈B
L
X
m(x L ) ≤
T
x L ∈G
GL
e−L(h+5) L
≤ ||GL ||e−L(h+5) ≤ e−L . Recall now that the above bound is true for a fixed > 0 and for all L ≥ L1 . Thus ∞ X L=1
m(BL
\
GL ) =
LX 1 −1
m(BL
\
m(BL
\
GL )
L=L1
L=1
≤ L1 +
∞ X
GL ) +
∞ X
e−L < ∞
L=L1
and hence from the Borel-Cantelli T lemma (Lemma 4.6.3 of [55] or Lemma 5.17 of [58]) m(x : x L ∈ BL GL i.o.) = 0. We also have from (4.15), however, that m(x : x L ∈ GLc i.o. ) = 0 and hence x L ∈ GL for T all but a finite number of L. Thus x L ∈ BL i.o. if and only if x L ∈ BL GL i.o. As this latter event has zero probability, we have shown that m(x : x L ∈ BL i.o.) = 0 and hence lim sup hL (x) ≤ h + 5. L→∞
Since is arbitrary we have proved that the limit supremum of the sample entropy −n−1 ln m(X n ) is less than or equal to the limit infimum and therefore that the limit exists and hence with m-probability 1 − ln m(X n ) = h. n→∞ n lim
(4.19)
Since the terms on the left in (4.19) are uniformly integrable from Lemma 3.7, we can integrate to the limit and apply Lemma 3.8 to find that Z − ln m(X n (x)) h = lim dm(x) = H m (X), n→∞ n which completes the proof of the lemma and hence also proves Theorem 4.1 for the special case of stationary ergodic measures. 2
4.3 Stationary Nonergodic Sources Next suppose that a source is stationary with ergodic decomposition {mλ ; λ ∈ Λ} and ergodic component function ψ as in Theorem 1.6. The source will produce with probability one under m an ergodic component mλ and Lemma 4.2 will hold for this ergodic component. In other words, we should have that
4.3 Stationary Nonergodic Sources
lim −
n→∞
107
1 ln mψ (X n ) = H mψ (X); m − a.e., n
(4.20)
that is, m({x : − lim ln mψ(x) (x n ) = H mψ(x) (X)}) = 1. n→∞
This argument is made rigorous in the following lemma. Lemma 4.3. Suppose that {Xn } is a stationary not necessarily ergodic source with ergodic component function ψ. Then m({x : − lim ln mψ(x) (x n ) = H mψ(x) (X)}) = 1; m − a.e.. n→∞
(4.21)
Proof: Let G = {x : − lim ln mψ(x) (x n ) = H mψ(x) (X)} n→∞
and let Gλ denote the section of G at λ, that is, Gλ = {x : − lim ln mλ (x n ) = H mλ (X)}. n→∞
From the ergodic decomposition (e.g., Theorem 1.6 or [55], Theorem 8.5.1, [58], Theorem 10.1) and (1.28) Z m(G) = dPψ (λ)mλ (G), where mλ (G) = m(G|ψ = λ) = m(G
\
{x : ψ(x) = λ}|ψ = λ)
= m(Gλ |ψ = λ) = mλ (Gλ ) which is 1 for all λ from the stationary ergodic result. Thus Z m(G) = dPψ (λ)mλ (Gλ ) = 1. It is straightforward to verify that all of the sets considered are in fact measurable. 2 Unfortunately it is not the sample entropy using the distribution of the ergodic component that is of interest, rather it is the original sample entropy for which we wish to prove convergence. The following lemma shows that the two sample entropies converge to the same limit and hence Lemma 4.3 will also provide the limit of the sample entropy with respect to the stationary measure. Lemma 4.4. Given a stationary source {Xn }, let {mλ ; λ ∈ Λ} denote the ergodic decomposition and ψ the ergodic component function of Theorem 1.6. Then
108
4 The Entropy Ergodic Theorem
mψ (X n ) 1 ln = 0; m − a.e. n→∞ n m(X n ) Proof: First observe that if m(an ) is 0, then from the ergodic decomposition with probability 1 mψ (an ) will also be 0. One part is easy. For any > 0 we have from the Markov inequality that lim
m(
m(X n ) −n m(X n ) m(X n ) 1 n ) ≤ E ( . ln > ) = m( > e )e m n mψ (X n ) mψ (X n ) mψ (X n ) (λ)
The expectation, however, can be evaluated as follows: Let An = {an : mλ (an ) > 0}. Then ! Z Z X m(an ) m(X n ) n Em = dPψ (λ) mλ (a ) = dPψ (λ)m(A(λ) n ) ≤ 1, mψ (X n ) mλ (an ) an ∈A n
where Pψ is the distribution of ψ. Thus m( and hence
∞ X
m(X n ) 1 ln > ) ≤ e−n . n mψ (X n )
m(
n=1
1 m(X n ) ln > ) < ∞ n mψ (X n )
and hence from the Borel-Cantelli lemma m(
1 m(X n ) ln > i.o.) = 0 n mψ (X n )
and hence with m probability 1 lim sup n→∞
1 m(X n ) ln ≤ . n mψ (X n )
Since is arbitrary, lim sup n→∞
1 m(X n ) ln ≤ 0; m − a.e. n mψ (X n )
(4.22)
For later use we restate this as lim inf n→∞
mψ (X n ) 1 ln ≥ 0; m − a.e. n m(X n )
(4.23)
Now turn to the converse inequality. For any positive integer k, we can construct a stationary k-step Markov approximation to m as in Section 3.7 that is, construct a process m(k) with the conditional probabilities
4.3 Stationary Nonergodic Sources
109
k k m(k) (Xn ∈ F |X n ) = m(k) (Xn ∈ F |Xn−k ) = m(Xn ∈ F |Xn−k )
and the same kth order distributions m(k) (X k ∈ F ) = m(X k ∈ F ). Consider the probability m(
1 m(k) (X n ) −n m(k) (X n ) m(k) (X n ) n ) ≤ E ( . ln ≥ ) = m( ≥ e )e m n m(X n ) m(X n ) m(X n )
The expectation is evaluated as X m(k) (x n ) xn
m(x n )
m(x n ) = 1
and hence we again have using Borel-Cantelli that lim sup n→∞
1 m(k) (X n ) ln ≤ 0. n m(X n )
Apply the usual ergodic theorem to conclude that with probability 1 under m lim sup n→∞
1 1 1 1 ln ≤ lim ln (k) n = Emψ [− ln m(Xk |X k )]. n n→∞ n m(X ) n m (X )
Combining this result with (4.20) and Lemma 3.10 yields lim sup n→∞
mψ (X n ) 1 ln ≤ −H mψ (X) − Emψ [ln m(Xk |X k )]. = H mψ ||m(k) (X). n m(X n )
This bound holds for any integer k and hence it must also be true that m-a.e. the following holds: lim sup n→∞
mψ (X n ) 1 ln ≤ inf H mψ ||m(k) (X) ≡ ζ. n m(X n ) k
(4.24)
In order to evaluate ζ we apply the ergodic decomposition of relative entropy rate (Corollary 3.5) and the ordinary ergodic decomposition to write Z Z dPψ ζ = dPψ inf H mψ ||m(k) (X) k Z ≤ inf dPψ H mψ ||m(k) (X) = inf H m||m(k) (X). k
k
From Theorem 3.4, the right hand term is 0. If the integral of a nonnegative function is 0, the integrand must itself be 0 with probability one. Thus (4.24) becomes
110
4 The Entropy Ergodic Theorem
lim sup n→∞
mψ (X n ) 1 ln ≤ 0, n m(X n )
which with (4.23) completes the proof of the lemma.
2
We shall later see that the quantity in (X n ; ψ) =
mψ (X n ) 1 ln n m(X n )
is the sample mutual information (in a generalized sense so that it applies to the usually non-discrete ψ) and hence the lemma states that the normalized sample mutual information between the process outputs and the ergodic component function goes to 0 as the number of samples goes to infinity. The two previous lemmas immediately yield the following result. Corollary 4.1. The conclusions of Theorem 4.1 hold for sources that are stationary.
4.4 AMS Sources The principal idea required to extend the entropy theorem from stationary sources to AMS sources is contained in Lemma 4.6. It shows that an AMS source inherits sample entropy properties from an asymptotically dominating stationary source (just as it inherits ordinary ergodic properties from such a source). The result is originally due to Gray and Kieffer [62], but the proof here is somewhat different. The tough part here is handling the fact that the sample average being considered depends on a specific measure. From Theorem 1.2, the stationary mean of an AMS source dominates the original source on tail events, that is, events in F∞ . We begin by showing that certain important events can be recast as tail events, that is, they can be determined by looking at only samples in the arbitrarily distant future. The following result is of this variety: It implies that sample entropy is unaffected by the starting time. Lemma 4.5. Let {Xn } be a finite-alphabet source with distribution m. Recall that Xkn = (Xk , Xk+1 , · · · , Xk+n−1 ) and define the information density i(X k ; Xkn−k ) = ln Then lim
n→∞
m(X n ) m(X k )m(Xkn−k )
1 i(X k ; Xkn−k ) = 0; m − a.e. n
.
4.4 AMS Sources
111
Comment: The lemma states that with probability 1 the per-sample mutual information density between the first k samples and future samples goes to zero in the limit. Equivalently, limits of n−1 ln m(X n ) will be the same as limits of n−1 ln m(Xkn−k ) for any finite k. Note that the result does not require even that the source be AMS. The lemma is a direct consequence of Lemma 3.19. Proof: Define the distribution p = mX k ×mXk ,Xk+1 ,··· , that is, a distribution for which all samples after the first k are independent of the first k samples. Thus, in particular, p(X n ) = m(X k )m(Xkn ). We will show that p m, in which case the lemma will follow from Lemma 3.19. Suppose that p(F ) = 0. If we denote Xk+ = Xk , Xk+1 , · · · , then 0 = p(F ) =
X
m(x k )mXk+ (Fx k ),
xk
where Fx k is the section {xk+ : (x k , xk+ ) = x ∈ F }. For the above relation to hold, we must have mXk+ (Fx k ) = 0 for all x k with m(x k ) 6= 0. We also have, however, that X m(F ) = m(X k = ak , Xk+ ∈ Fak ) ak
=
X
m(X k = ak |Xk+ ∈ Fak )m(Xk+ ∈ Fak ).
ak
But this sum must be 0 since the rightmost terms are 0 for all ak for which m(X k = ak ) is not 0. (Observe that we must have m(X k = ak |Xk+ ∈ Fak ) = 0 if m(Xk+ ∈ Fak ) 6= 0 since otherwise m(X k = ak ) ≥ m(X k = ak , Xk+ ∈ Fak ) > 0, yielding a contradiction.) Thus p m and the lemma is proved. 2 For later use we note that we have shown that a joint distribution is dominated by a product of its marginals if one of the marginal distributions is discrete. Lemma 4.6. Suppose that {Xn } is an AMS source with distribution m and suppose that m is a stationary source that asymptotically dominates m (e.g., m is the stationary mean). If there is an invariant function h such that 1 lim − ln m(X n ) = h; m − a.e., n→∞ n then also, lim −
n→∞
1 ln m(X n ) = h; m − a.e. n
Proof: For any k we can write using the chain rule for densities
112
4 The Entropy Ergodic Theorem
−
1 1 1 ln m(X n ) + ln m(Xkn−k ) = − ln m(X k |Xkn−k ) n n n 1 1 = − i(X k ; Xkn−k ) − ln m(X k ). n n
From the previous lemma and from the fact that Hm (X k ) = −Em ln m(X k ) is finite, the right hand terms converge to 0 as n → ∞ and hence for any k lim −
n→∞
1 ln m(X k |Xkn−k ) = n 1 1 lim (− ln m(X n ) + ln m(Xkn−k )) = 0; m − a.e. n→∞ n n
(4.25)
This implies that there is a subsequence k(n) → ∞ such that 1 1 1 n−k(n) n−k (n)) → 0; m−a.e. ln m(X k(n) |Xk(n) ) = − ln m(X n )− ln m(Xk(n) n n n (4.26) To see this, observe that (4.25) ensures that for each k there is an N(k) large enough so that N(k) > N(k − 1) and −
m(| −
1 N(k)−k ln m(X k |Xk )| > 2−k ) ≤ 2−k . N(k)
(4.27)
Applying the Borel-Cantelli lemma implies that for any , N(k)−k
m(| − 1/N(k) ln m(X k |Xk
)| > i.o.) = 0.
Now let k(n) = k for N(k) ≤ n < N(k + 1). Then n−k(n)
m(| − 1/n ln m(X k(n) |Xk(n)
)| > i.o.) = 0
and therefore lim
n→∞
−
1 1 n−k(n) ln m(X n ) + ln m(Xk(n) ) = 0; m − a.e. n n
as claimed in (4.26). In a similar manner we can also choose the sequence so that 1 1 n−k(n) lim − ln m(X n ) + ln m(Xk(n) ) = 0; m − a.e., n→∞ n n that is, we can choose N(k) so that (4.27) simultaneously holds for both m and m. Invoking the entropy ergodic theorem for the stationary m (Corollary 4.3) we have therefore that lim −
n→∞
1 n−k(n) ln m(Xk(n) ) = h; m − a.e.. n
(4.28)
4.4 AMS Sources
113
From Markov’s inequality (Lemma 4.4.3 of [55] or Lemma 5.8 of [58]) m(−
m(Xkn ) 1 1 ln m(Xkn ) ≤ − ln m(Xkn ) − ) = m( ≥ en ) n n m(Xkn ) ≤ e−n Em
m(Xkn−k ) m(Xkn−k ) X
= e−n
xkn−k :m(xkn−k )6=0
m(xkn−k ) m(xkn−k )
m(xkn−k )
≤ e−n . Hence taking k = k(n) and again invoking the Borel-Cantelli lemma we have that m(−
1 1 n−k(n) n−k(n) ln m(Xk(n) ) ≤ − ln m(Xk(n) ) − i.o.) = 0 n n
or, equivalently, that n−k(n)
lim inf − n→∞
m(Xk(n) ) 1 ln ≥ 0; m − a.e. n−k(n) n m(X )
(4.29)
k(n)
Therefore from (4.28) 1 n−k(n) (4.30) ln m(Xk(n) ) ≥ h; m − a.e.. n T The above event is in the tail σ -field F∞ = n σ (Xn , Xn+1 , · · · ) since it can be determined from Xk(n) , · · · for arbitrarily large n and since h is invariant. Since m dominates m on the tail σ -field (Theorem 1.3), we have also 1 n−k(n) lim inf − ln m(Xk(n) ) ≥ h; m − a.e. n→∞ n lim inf − n→∞
and hence by (4.26) lim inf − n→∞
1 ln m(X n ) ≥ h; m − a.e. n
which proves half of the lemma. Since m( lim − n→∞
1 ln m(X n ) 6= h) = 0 n
and since m asymptotically dominates m (Theorem 1.2), given > 0 there is a k such that m( lim − n→∞
1 ln m(Xkn ) = h) ≥ 1 − . n
114
4 The Entropy Ergodic Theorem
Again applying Markov’s inequality and the Borel-Cantelli lemma as in the development of (4.28) we have that lim inf − n→∞
m(Xkn ) 1 ln ≥ 0; m − a.e, n m(Xkn )
which implies that m(lim sup − n→∞
1 ln m(Xkn ) ≤ h) ≥ 1 − n
and hence also that m(lim sup − n→∞
1 ln m(X n ) ≤ h) ≥ 1 − . n
Since can be made arbitrarily small, this proves that m-a.e. lim sup −n−1 ln m(X n ) ≤ h, n→∞
which completes the proof of the lemma. 2 The lemma combined with Corollary 4.3 completes the proof of Theorem 4.1. 2 Theorem 4.1 and Lemma 2.5 immediately yield the following corollary stating that a stationary coding of an AMS process has a well defined entropy rate given by a limit, as in the case of a stationary process. Corollary 4.2. Theorem 4.1 If f is a stationary coding of an AMS process, then 1 H(f ) = lim H(f n ). n→∞ n
4.5 The Asymptotic Equipartition Property Since convergence almost everywhere implies convergence in probability, Theorem 4.1 has the following implication: Suppose that {Xn } is an AMS ergodic source with entropy rate H. Given > 0 there is an N such that for all n > N the set Gn = {x n : |n−1 hn (x) − H| ≥ } = {x n : e−n(H+) ≤ m(x n ) ≤ e−n(H−) } has probability greater then 1 − . Furthermore, as in the proof of the theorem, there can be no more than en(H+) n-tuples in Gn . Thus there are two sets of n-tuples: a “good” set of approximately enH n-tuples having approximately equal probability of e−nH and the complement of this set which has small total probability. The set of good sequences are
4.5 The Asymptotic Equipartition Property
115
often referred to as “typical sequences” or “entropy-typical sequences” in the information theory literature and in this form the theorem is called the asymptotic equipartition property or the AEP. As a first information theoretic application of an ergodic theorem, we consider a simple coding scheme called an “almost noiseless” or “almost lossless” source code. As we often do, we consider logarithms to the base 2 when considering specific coding applications. Suppose that a random process {Xn } has a finite alphabet A with cardinality kAk and entropy rate H. Suppose that H < log kAk, e.g., A might have 16 symbols, but the entropy rate is slightly less than 2 bits per symbol rather than log 16 = 4. Larger alphabets cost money in either storage or communication applications. For example, to communicate a source with a 16 letter alphabet sending one letter per second without using any coding and using a binary communication system we would need to send 4 binary symbols (or four bits) for each source letter and hence 4 bits per second would be required. If the alphabet only had 4 letters, we would need to send only 2 bits per second. The question is the following: Since our source has an alphabet of size 16 but an entropy rate of less than 2, can we code the original source into a new source with an alphabet of only 4 = 22 letters so as to communicate the source at the smaller rate and yet have the receiver be able to recover the original source? The AEP suggests a technique for accomplishing this provided we are willing to tolerate rare errors. We construct a block code of the original source by first picking a small and a δ small enough so that H + δ < 2. Choose a large enough n so that the AEP holds giving a set Gn of good sequences as above with probability greater than 1 − . Index this collection of fewer than 2n(H+δ) < 22n sequences using binary 2n-tuples. The source Xk is parsed into n blocks of length n as Xkn = (Xkn , Xkn+1 , · · · , X(k+1)n ) and each block is encoded into a binary 2n-tuple as follows: If the source n-tuple is in Gn , the codeword is its binary 2n-tuple index. Select one of the unused binary 2n-tuples as the error index and whenever an n-tuple is not in Gn , the error index is the codeword. The receiver or decoder than uses the received index and decodes it as the appropriate n-tuple in Gn . If the error index is received, the decoder can declare an arbitrary source sequence or just declare an error. With probability at least 1 − a source n-tuple at a particular time will be in Gn and hence it will be correctly decoded. We can make this probability as small as desired by taking n large enough, but we cannot in general make it 0. The above simple scheme is an example of a block coding scheme as considered in Section 2.7. If considered as a mapping from sequences into sequences, the map is not stationary, but it is block stationary in the sense that shifting an input block by n results in a corresponding block shift of the encoded sequence by 2n binary symbols.
Chapter 5
Distortion and Approximation
Abstract Various notions of the distortion between random variables, vectors, and processes as well as between different codings of a common source are quantified in this chapter. A distortion measure is not a “measure” in the sense used so far — it is an assignment of a nonnegative real number which indicates how bad an approximation one symbol or random object or coding is of another. The smaller the distortion, the better the approximation. If the two objects correspond to the input and output of a communication system, then the average distortion provides a measure of the performance or fidelity of the system. Small average distortion means high fidelity and good performance, while large average distortion means low fidelity and poor performance. Distortion measures generalize the idea of a distance or metric and they need not have metric properties such as the triangle inequality and symmetry, but such properties can be exploited when available and unsurprisingly the most important notions of distortion are either metrics or simple functions of metrics. We shall encounter several notions of distortion and a diversity of applications, with the most important application being the average distortion between input and output as a measure of the performance of a communications system. Other applications include extensions of finite memory channels to channels which approximate finite memory channels, geometric characterizations of the optimal performance of communications systems, approximations of complicated codes by simpler ones, and modeling random processes.
5.1 Distortion Measures Given two measurable spaces (A, BA ) and (B, BB ), a distortion measure on A × B is a nonnegative measurable mapping ρ : A × B → [0, ∞) which assigns a real number ρ(x, y) to each x ∈ A and y ∈ B which can be R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_5, © Springer Science+Business Media, LLC 2011
117
118
5 Distortion and Approximation
thought of as the cost of reproducing x and y. The principal practical goal is to have a number by which the goodness or badness of communication systems can be compared. For example, if the input to a communication system is a random variable X ∈ A and the output is Y ∈ B, then one possible measure of the performance or quality of the system is the average distortion Eρ(X, Y ). A distortion measure is essentially the same as a loss, risk, or cost function in statistics. Ideally one would like a distortion measure to have three properties: • It should be tractable so that one can do useful theory. • It should be computable so that it can be measured in real systems. • It should be subjectively meaningful in the sense that small (large) distortion corresponds to good (bad) perceived quality. Unfortunately these requirements are often incompatible and one is forced to compromise between tractability and subjective significance in the choice of distortion measures. Among the most popular choices for distortion measures are metrics or distances, but some practically important distortion measures are not metrics in that they are not symmetric or do not satisfy the triangle inequality. Two specific examples are by far the most important in information theory, signal processing, communications, and statistics: the per symbol Hamming distortion and squared-error distortion. While neither provides a panacea, the Hamming distortion is the most common distortion measure for discrete alphabets and the squared-error for continuous alphabets for several reasons. Both are tractable and simple and easy to compute. The Hamming distortion is arguably the most unforgiving distortion measure possible since the maximum distortion is assigned to every pair unless they two match exactly. Average closeness with respect to any distortion measure with a maximum value can be assured by ensuring average closeness in the Hamming sense. The Hamming distance is primarily used with discrete alphabet variables. Squared-error does not play a similar “worst case” role, but it is an intuitive measure since its average is the energy of the error between two variables. Variations allowing linear weightings of variables and signals before evaluating the average of a quadratic error yield a variety of distortion measures that have proved useful in speech, voice, audio, image, and video processing, especially in techniques incorporating “perceptual coding” where distortion is measured according to its estimated impact on human perception. Suppose that A, B ⊂ R. The Hamming distance is defined by ( ρ(x, y) = dH (x, y) =
0 x=y 1
x ≠ y.
The Hamming distance is a distortion measure which is also a metric. Given a probability measure p on (A × B, B(A × B)) let X and Y denote
5.1 Distortion Measures
119
the coordinate random variables X(x, y) = x, Y (x, y) = y, then Ep dH = Pr(X ≠ Y ). The squared-error distortion is defined by ρ(x, y) =| x − y |2 . Squared error is not a metric, but it is the square of the metric | x − y |, which means it inherits properties of a metric: it is symmetric and its square root satisfies the triangle inequality. The average squared error E[(X − Y )2 ] is often interpreted as the energy in the error signal. The Hamming distance is well-defined if we consider vector alphabets A, B ⊂ Rn : dH (x n , y n ) = 0 if x n = y n and 0 otherwise. A far more useful extension from scalars to vectors, however, is construction of the vector distortion from the scalar Hamming distance in an additive or linear fashion: n−1 X ρ(x n , y n ) = d(xi , yi ), (5.1) i=0
which is the number of coordinates in which the two vectors differ. This distortion measure is referred to as the average Hamming distance or mean Hamming distance. Extending the squared-error distortion in a similar additive fashion to vector spaces yields a squared-error distortion ρ(x n , y n ) =
n−1 X
| xi − yi |2 .
(5.2)
i=0
This is not a metric, but it is the square of the Euclidean or `2 distance: ρ(x n , y n ) = kx n − yn k22 . These examples typify distortion measures and are the most important special cases, but most of the results hold more generally and the development will focus on distortion measures formed as a positive power of a metric. Many of the results developed here will be for the case where A is a Polish space, a complete separable metric space under a metric d, and B is either A itself or a Borel subset of A. The distortion measure is assumed to be a positive power of d. In this case the distortion measure is fundamental to the structure of the alphabet and the alphabets are standard since the space is Polish.
120
5 Distortion and Approximation
5.2 Fidelity Criteria It is often of interest to consider distortion between sequences as well as between scalars and vectors. One approach to a distortion between sequences is to define a family of distortion measures between vectors of all dimensions and consider limits. This is the idea behind a fidelity criterion as defined by Shannon [162, 163]. Given “scalar” spaces A and B, a fidelity criterion ρn , n = 1, 2, · · · , is a sequence of distortion measures on An × B n . A candidate for the distortion between infinite sequences is then the limit supremum of the per symbol distortion ρ∞ (x, y) = lim sup n→∞
1 ρn (x n , y n ). n
If one has a pair random process, say {Xn , Yn } with process distribution p, then it is of interest to find conditions under which there is a limiting per symbol distortion in the sense that the limit exists with p probability 1: 1 ρ∞ (x, y) = lim ρn (x n , y n ). (5.3) n→∞ n As one might guess, the distortion measures in the sequence need to be interrelated in order to get useful behavior. The simplest and most common example is that of an additive or single-letter fidelity criterion which has the form ρn (x n , y n ) =
n−1 X
ρ1 (xi , yi ).
i=0
Here if the pair process is AMS and ρ1 satisfies suitable integrability assumptions, then the limiting per-symbol distortion n−1 1 X 1 ρn (x n , y n ) = ρ1 (xi , yi ). n n i=0
will exist and be invariant from the ergodic theorem. If the pair process is stationary rather than only AMS, then one can consider the more general case where ρn is subadditive in the sense that ρn (x n , y n ) ≤ ρk (x k , y k ) + ρn−k (xkn−k , ykn−k ). In this case then stationarity of the pair process and integrability of ρ1 will ensure that n−1 ρn converges from the subadditive ergodic theorem. For example, if d is a distortion measure (possibly a metric) on A × B, then
5.3 Average Limiting Distortion
121
1/p
n−1 X
ρn (x n , y n ) =
d(xi , yi )p
i=0
for p > 1 is subadditive from Minkowski’s inequality. By far the bulk of the information theory literature considers only additive fidelity criteria and we will share this emphasis. The most common examples of additive fidelity criteria involve defining the perletter distortion ρ1 in terms of an underlying metric d. For example, set ρ1 (x, y) = d(x, y), in which case ρn is also a metric for all n, or ρ1 (x, y) = dp (x, y) for some p > 0. If 1 > p > 0, then again ρn is a 1/p metric for all n. If p ≥ 1, then ρn is not a metric, but ρn is a metric. We do not wish to include the 1/p in the definition of the fidelity criterion, however, because what we gain by having a metric distortion we more than lose when we take the expectation. For example, a popular distortion measure is the expected squared error, not the expectation of the square root of the squared error. The fidelity criteria introduced here all are context-free in that the distortion between n successive input/output samples of a pair process does not depend on samples occurring before or after these n-samples. Some work has been done on context-dependent distortion measures (see, e.g., [107]), but we do not consider their importance sufficient to merit the increased notational and technical difficulties involved. Hence we shall consider only context-free distortion measures.
5.3 Average Limiting Distortion Suppose that {Xn , Yn } is an AMS pair process with alphabet A × B. Let p denote the corresponding distribution of the pair process. One measure of the quality (or rather the lack thereof) of approximation of X by Y is given by the average limiting distortion with respect to a fidelity criterion. Given two sequences x and y and a fidelity criterion ρn ; n = 1, 2, · · · , define the limiting sample average distortion or sequence distortion by ρ∞ (x, y) = lim sup n→∞
1 ρn (x n , y n ) n
and define the average sequence distortion ∆(p) = Ep ρ∞ = Ep
! 1 n n lim sup ρn (X , Y ) . n→∞ n
122
5 Distortion and Approximation
We focus on two important special cases. The first and most important is that of AMS pair processes and additive fidelity criteria. We also consider the case of subadditive distortion measures and systems that are either two-sided and AMS or are one-sided and stationary. Unhappily the overall AMS one-sided case cannot be handled as there is not yet a general subadditive ergodic theorem for that case (from Theorem 8.5 of [58] there is a theorem if in addition p is absolutely continuous with respect to its stationary mean p ). In all of these cases we have that if ρ1 is integrable with respect to the stationary mean process p, then ρ∞ (x, y) = lim
n→∞
1 ρn (x n , y n ); p − a.e., n
(5.4)
and ρ∞ is an invariant function of its two arguments, i.e., ρ∞ (TA x, TB y) = ρ∞ (x, y); p − a.e..
(5.5)
When a process distribution and fidelity criterion are such that (5.4) and (5.5) are satisfied (at least with probability 1) we say that we have a convergent fidelity criterion. This property holds, for example, by an underlying assumption that p is AMS and that the fidelity criterion is additive and the single-letter distortion is integrable with respect to the stationary mean. Since ρ∞ is invariant, we have from Lemma 6.3.1 of [55] or Corollary 7.10 of [58] that ∆(p) = Ep ρ∞ = Ep ρ∞ . (5.6) If the fidelity criterion is additive, then we have from the stationarity of p that the average limiting distortion is given by ∆(p) = Ep ρ1 (X0 , Y0 ).
(5.7)
If the fidelity criterion is subadditive and the processes stationary, then this is replaced by ∆(p) = inf N
1 Ep ρN (X N , Y N ). N
(5.8)
Assume for the remainder of this section that ρn is an additive fidelity criterion. Suppose that we know that p is N-stationary; that is, if T = TA × TB denotes the shift on the input/output space AT × B T , then the overall process is stationary with respect to T N . In this case ∆(p) =
1 Ep ρN (X N , Y N ). N
(5.9)
We can also consider the behavior of the N-shift more generally when the system is only AMS This will be useful when considering block codes.
5.4 Communications Systems Performance
123
Suppose now that p is AMS with stationary mean p. Then from Theorem 7.3.1 of [55] or Theorem 8.2 of [58], p is also T N -AMS with an N-stationary mean, say p N . Applying the ergodic theorem to the N shift then implies that if ρN is p N -integrable, then n−1 1 X N N (N) ˆiN ρN (xiN ,x ) = ρ∞ n→∞ n i=0
(5.10)
lim
(N)
exists p N (and hence also p) almost everywhere. In addition, ρ∞ is Ninvariant and (N) (N) ˆN ). Ep ρ∞ = EpN ρ∞ = EpN ρN (X N , X (5.11) (N)
Comparison of (5.4) and (5.11) shows that ρ∞ ∆(p) =
= Nρ∞ p-a.e. and hence
1 1 (N) ˆ N ) = E p ρ∞ ˆ0 ) = ∆(p). = Ep ρ1 (X0 , X EpN ρN (X N , X N N
(5.12)
The key point here is that the measure of the quality or fidelity in terms of one component of an AMS pair process approximating the other is the same as that of the induced stationary mean, which can be described in terms of the time 0 samples of the two random processes.
5.4 Communications Systems Performance The primary application of the idea of distortion is to the quantification of quality or fidelity in a communications system. Suppose that [µ, f , ν, g] is a communications system with overall input/output proˆ Let p denote the corresponding ˆn } and alphabet A × A. cess is {Xn , X distribution of the pair process comprised of the input and output. As in Section 5.3, a natural measure of the (lack of) quality of the output ˆ as an approximation to the original input sigor reproduction signal X nal X is given by the average limiting distortion ∆(p), which in the case of a communications system we call the performance of the system. In this case there is much going on between the original input and ˆn } that determines the final output, but it is still the pair process {Xn , X the performance of the system. Note that in the communications example, the input/output process can be N-stationary if the source source is N-stationary, the first sequence coder (N, K)-stationary, the channel K-stationary (e.g., stationary), and the second sequence coder (K, N)stationary. It is the overall properties that matter when looking at the performance. If the source and codes are such that the input/output process is AMS, then the results of Section 5.3 show that the performance satisfies
124
5 Distortion and Approximation
∆(p) =
1 1 (N) ˆN ) = Ep ρ∞ ˆ0 ). = Ep ρ1 (X0 , X Ep ρN (X N , X N N N
(5.13)
5.5 Optimal Performance Given a notion of the performance of a communication system, it makes sense to define the optimal performance achievable when communicating a source {Xn } with distribution µ over a channel ν. Suppose that E is some class of sequence coders f : AT → B T . For example, E might consist of all sequence coders generated by block codes with some constraint or by finite-length sliding-block codes. Similarly let D denote a ˆT . Define the operational distortionclass of sequence coders g : B 0 T → A rate function (DRF) for the source µ, channel ν, and code classes E and D by ∆(µ, ν, E, D) = inf ∆(µ, f , ν, g). (5.14) f ∈E,g∈D
When the code classes E, D are clear from context, the notation is simplified to ∆(µ, ν). When the channel is assumed to be noiseless with alphabet B with kBk letters, then all that will matter when considering block and sliding-block codes is the channel rate R = log kBk and the notation is simplified to δ(R, µ) = ∆(µ, ν) = ∆(µ, ν, E, D).
(5.15)
When the class of codes being considered is that of sliding-block codes, the operational DRF δ(R, µ) will be subscripted as δSBC (R, µ). The goal of the coding theorems of information theory is to relate the operational DRF of a source, channel, and code class to (hopefully) computable functions of the source and channel. We will do this in stages in later chapters: first we will focus on the source coding by assuming a noiseless channel, then we will focus on reliable communication over a noisy channel, and lastly we will combine the two.
5.6 Code Approximation Suppose that {Xn } is an information source with process distribution µ and suppose that f and g are two sliding-block codes which share a common reproduction alphabet B. For the moment assume for simplicity that the source is stationary. Let P and Q be the corresponding partitions of sequence space, e.g., P = {Pi ; i ∈ I} where Pi = f −1 (bi ). Average distortion can be used to measure how good an approximation one code is for another. This is of interest, for example, if one code is
5.6 Code Approximation
125
nearly optimal in some sense, but too complicated to be practical. If another, simpler, code has approximately the same behavior, then it may be a better choice for implementation. Given a distortion measure ρ, the distortion between two codes f , g applied to a common source with distribution µ can be defined as ∆(f , g) = Eµ ρ(f , g).
(5.16)
In the case where both f and g have discrete output alphabets, then a natural distortion is the Hamming distortion and this becomes ∆H (f , g) = Eµ dH (f , g) = Pr(f ≠ g) = Pe ,
(5.17)
where Pe is a common notation for error probability, the probability that the two discrete random variables f and g differ. This distance between codes can be related to the partition distance of ergodic theory between the two partitions P and Q which is defined by X | P − Q |= µ(Pi ∆Qi ). (5.18) i∈I
We have that ∆H (f , g) = 1 − Pr(f = g) X µ(Pi ∩ Qi ) = 1− i
1X = (µ(Pi ) + µ(Qi ) − µ(Pi ∩ Qi )) 2 i 1X 1 = µ(Pi ∆Qi ) = | P − Q | . 2 i 2
(5.19)
So far we have considered only a single output of the code, which suffices if the source is stationary. In general, however, we may wish to consider AMS sources and an additive fidelity criterion based on the Hamming distance, in which case the mean distortion is given by n−1 1 X Pr(fn ≠ gn ), n i=0
where as usual fn = f T n , and its limit as n → ∞ are of interest. Since stationary codings of an AMS source are jointly AMS, this average converges and we can define a code distance n−1 1 X Pr(fn ≠ gn ). n→∞ n i=0
∆H (f , g) = P e = lim
(5.20)
126
5 Distortion and Approximation
If the source is stationary, then from the stationarity of sliding-block codes this simplifies to Pe = Pr(f ≠ g). The next lemma and corollary provide tools for approximating complicated codes by simpler ones. Lemma 5.1. Given a probability space (Ω, B, P ) suppose that F is a generating field: B = σ (F ). Suppose that B-measurable Q is a partition of Ω and > 0. Then there is a partition Q0 with atoms in F such that |Q − Q0 | ≤ . Proof: Let kAk = K. From Theorem 1.1 given γ > 0 we can find sets Ri ∈ F such that P (Qi ∆Ri ) ≤ γ for i = 1, 2, · · · , K − 1. The remainder of the proof consists of set theoretic manipulations showing that we can construct the desired partition from the Ri by removing overlapping pieces. The algebra is given for completeness, but it can be skipped. Form a partition from the sets as Q0 i = Ri −
i−1 [
Rj , i = 1, 2, · · · , K − 1
j=1 K−1 [
Q0 K = (
Q0 i )c .
i=1
For i < K \ Q0 i ) − P (Qi Q0 i ) [ [ \ ≤ P (Qi Rj )). Ri ) − P (Qi (Ri −
P (Qi ∆Q0 i ) = P (Qi
[
(5.21)
j
The rightmost term can be written as [ [ \ \ \ \ Rj )) = P ((Qi P (Qi (Ri − Ri ) − ( Qi Ri Rj )) j
j
= P (Qi
\
Ri ) − P (
[
Qi
\
Ri
\
Rj ), (5.22)
j
where we have used the fact that a set difference is unchanged if the portion being removed is intersected with the set it is being removed from and we have used the fact that P (F − G) = P (F ) − P (G) if G ⊂ F . Combining (5.21) and (5.22) we have that
5.6 Code Approximation
127
P (Qi ∆Q0 i ) ≤ P (Qi
[
Ri ) − P (Qi
\
Ri ) + P (
[
Qi
\
Ri
\
Rj )
j
= P (Qi ∆Ri ) + P (
[
Qi
\
Ri
\
Rj )
j
≤γ+
X
P (Qi
\
Rj ).
j
For j 6= i, however, we have that \ \ \ \ P (Qi Rj ) = P (Qi Rj Qjc ) ≤ P (Rj Qjc ) ≤ P (Rj ∆Qj ) ≤ γ, which with the previous equation implies that P (Qi ∆Q0 i ) ≤ Kγ; i = 1, 2, · · · , K − 1. For the remaining atom: P (QK ∆Q0 K ) = P (QK
\
c
Q0 K
[
c QK
\
Q0 K ).
(5.23)
We have QK
\
c
Q0 K = QK
\ [ \ [ \ Q0 j ) = QK ( Q0 j ( Qjc ), j 0. We will refer to this as the dp distortion. We will also allow the case p = 0 which is defined by ρ1 (x, y) = d0 (x, y) = dH (x, y), the Hamming distance dH (x, y) = 1 − δx,y . • The random process distributions µ considered will possess a reference letter in the sense of Gallager [47]: given a stationary µ it is assumed that there exists an a ∈ A for which Z dp (x, a∗ )dµ 1 (x) < ∞. (5.37) If the process is not stationary, we assume that there exists an a ∈ A 1 and ρ ∗ < ∞ such that for all marginal distributions µn Z 1 dp (x, a∗ )dµn (x) < ρ ∗ (5.38) so that one reference letter works for all times n. This is trivially satisfied if the distortion is bounded. Define P p (A, d) to be the space of all stationary process distributions µ on (A, B(A))T satisfying (5.37), where both T = Z and Z+ will be considered. Note that previously Pp (A, d) referred to a space of distributions of random objects taking values in a metric space A. Now it refers to a space of process distributions with all individual component random variables taking values in a common metric space A, that is, the process distributions are on (A, B(A))T . In the process case we will consider Monge/Kantorovich distances on the spaces of random vectors with alphabets An for all n = 1, 2, . . ., but we shall define a process distortion and metric as the supremum over all possible vector dimensions. Given two process distributions µX and µY describing random processes {Xn } and {Yn }, let the induced n-dimensional distributions be µX n , µY n . For each positive integer n we can define the n-dimensional or n-th order optimal coupling distortion as the distortion between the induced n-dimensional distributions: ρ n (µX n , µY n ) =
inf
π ∈Pp (µX n ,µY n )
Eπ ρn (X n , Y n )
(5.39)
5.11 Process Distance and Approximation
137
The process optimal coupling distortion (or ρ distortion) between µX and µY is defined by ρ(µX , µY ) = sup n
1 ρ (µX n , µY n ). n n
(5.40)
The extension of the optimal coupling distortion or transportation cost to processes was developed in the early 1970s by D.S. Ornstein for the case of ρ1 = dH and the resulting process metric, called the d distance or d-bar distance or Ornstein distance, played a fundamental role in the development of the Ornstein isomorphism theorem of ergodic theory (see [138, 139, 140] and the references therein). The idea was extended to processes with Polish alphabets and metric distortion measures and the square of metric distortion measures in 1975 [66] and applied to problems of quantizer mismatch [60], Shannon information theory [63, 132, 64, 52, 67, 53], and robust statistics [142]. While there is a large literature on the finite-dimensional optimal coupling distance, the literature for the process optimal coupling distance seems limited to the information theory and ergodic theory literature. Here the focus is on the process case, but the relevant finite-dimensional results are also treated as needed. The key aspect of the process distortion is that if it is small, then necessarily the distortion between all sample vectors produced by the two processes is also small.
The dp -distance The process distortion is a metric for the important special case of an additive fidelity criterion with a metric per letter distortion. This subsection shows how a process metric can be obtained in the most important special case of an additive fidelity criterion with a per letter distortion given by a positive power of a metric. The result generalizes the special cases of process metrics in [139, 66, 179] and extends the well known finite-dimensional version of optimal transportation theory (e.g., Theorem 7.1 in [186]). Theorem 5.1. Given a Polish space (A, d) and p ≥ 0, define the additive fidelity criterion ρn : An × An → R+ ; n = 1, 2, . . . by ρn (x n , y n ) =
n−1 X
dp (xi , yi ),
n=0
where d0 is shorthand for the Hamming distance dH . Define
138
5 Distortion and Approximation min(1,1/p)
dp (µX , µY ) = sup n−1 ρ n n
(µX n , µY n ) = ρ min(1,1/p) (µX , µY ). (5.41)
Then dp is a metric on P p (A, d), the space of all stationary random processes with alphabet A. The theorem together with Lemma 5.3 says that if ρ1 = dp , then dp = ρ is a metric for both the vector and process case if p ≥ 1, and dp = ρ is a metric if 0 ≤ p ≤ 1. The two cases agree for p = 1 and the p = 0 case is simply shorthand for the p = 1 case with the Hamming metric. In the case of p = 0, the process distance d0 is Ornstein’s d-distance and the notation is usually abbreviated to simply d to match usage in the ergodic theory literature. It merits pointing out that d0 (µX n , µY n ) is not the transportation distance with a Hamming distance on An , it is the transportation distance with respect to the distance d(x n , y n ) = P n−1 i=0 dH (xi , yi ), the sum of the Hamming distances between symbols. This is also n times the average Hamming distance. The two metrics on An are related through the simple bounds 1/p
n−1 X
dH (xi , yi ) ≥ dH (x n , y n ) ≥
i=0
n−1 1 X dH (xi , yi ). n i=0
(5.42)
The next theorem collects several more properties of the dp distance between stationary processes, including the facts that the supremum defining the process distance is a limit, that the distance between IID processes reduces to the Monge/Kantorovich distance between the first order marginals, and a characterization of the process distance as an optimization over processes. Theorem 5.2. Suppose that µX and µY are stationary process distributions with a common standard alphabet A and that ρ1 = dp is a positive power of a metric on A and that ρn is defined on An in an additive fashion as before. Then (a)
limn→∞ n−1 ρ n (µX n , µY n ) exists and equals supn n−1 ρ n (µX n , µY n ). min(1,1/p)
(µX n , µY n ). Thus dp (µX , µY ) = limn→∞ n−1 ρ n (b) If µX and µY are both IID, then ρ(µX , µY ) = ρ 1 (µX0 , µY0 ) and hence min(1,1/p)
dp (µX , µY ) = ρ 1 (µX0 , µY0 ) (c) Let Ps = Ps (µX , µY ) denote the collection of all stationary distributions πXY having µX and µY as marginals, that is, distributions on {Xn , Yn } with coordinate processes {Xn } and {Yn } having the given distributions. Define the process distortion measure ρ 0 ρ 0 (µX , µY ) = Then
inf EπXY ρ(X0 , Y0 ).
πXY ∈Ps
5.11 Process Distance and Approximation
139
ρ(µX , µY ) = ρ 0 (µX , µY ); that is, the limit of the finite dimensional minimizations is given by a minimization over stationary processes. (d) Suppose that µX and µY are both stationary and ergodic. Define Pe = Pe (µX , µY ) as the subset of Ps containing only ergodic processes, then ρ(µX , µY ) = inf EπXY ρ(X0 , Y0 ). πXY ∈Pe
(e)
Suppose that µX and µY are both stationary and ergodic. Let GX denote a collection of frequency-typical or generic sequences for µX in the sense of Section 8.3 of [55] or Section 7.9 of [58]. Frequencytypical sequences are those along which the relative frequencies of a set of generating events all converge and hence by measuring relative frequencies on frequency-typical sequences one can deduce the underlying stationary and ergodic measure that produced the sequence. An AMS process produces frequency-typical sequences with probability 1. Similarly let GY denote a set of frequency-typical sequences for µY . Define the process distortion measure ρ 00 (µX , µY ) =
inf
x∈GX ,y∈GY
lim sup n→∞
n−1 1 X ρ1 (x0 , y0 ). n i=0
Then ρ(µX , µY ) = ρ 00 (µX , µY ). that is, the ρ distortion gives the minimum long term time average distance obtainable between frequency-typical sequences from the two processes. Proofs of these results can be found in Section 9.4 of [58], but the proof of part (c) is not correct. For completeness the proof is presented here. Proof. (c) (c) Given > 0 let π ∈ Ps (µX , µY ) be such that Eπ ρ1 (X0 , Y0 ) ≤ ρ 0 (µX , µY ) + . The induced distribution on {X n , Y n } is then contained in Pn (µX n , µY n ), and hence using the stationarity of the processes ρ n (µX n , µY n ) ≤ Eρn (X n , Y n ) = nEρ1 (X0 , Y0 ) ≤ n(ρ 0 (µX , µY ) + , and therefore ρ 0 ≥ ρ since is arbitrary. Let π n ∈ Pn , n = 1, 2, . . . be a sequence of measures such that Eπ n [ρn (X n , Y n )] ≤ ρ n + n where n > 0 and n → 0 as n → ∞. Let qn denote the product (block independent) measure (AT , B(A)T )2 induced by the π n as explained
140
5 Distortion and Approximation
next. Let G denote a countable generating field for the standard space (A, B(A)). For any N and N-dimensional rectangle or cylinder of the form F = ×i∈T Fi with all but a finite number N of the Fi being A2 and the remainder being in G 2 define Y π n (Fjn × Fjn+1 × · · · × Fjn+n−1 ). qn (F ) = j∈T
Thus qn assigns a probability to rectangles in a way that treats successive n-tuples as independent. Next “stationarize” qn to form a measure on rectangles by averaging over n-shifts to form πn (F ) =
n−1 n−1 1 X 1 X Y n qn (T −i F ) = π (Fjn+i ×Fjn+i+1 ×· · ·×Fjn+i+n−1 ). n i=0 n i=0 j∈T
This measure on the rectangles extends to a stationary pair process distribution. For any m = 1, 2, . . . , n we can relate the mth marginal restrictions of πn to the corresponding original marginals. For example, m consider the Y marginal and let G = ×m−1 k=0 Gi ∈ G . Then qn ({x, y : x m ∈ Am , y m ∈ G) = π n (An × (G × An−m )) = µY n (G × An−m ) = µY m (G) and similarly if G ∈ B(A)m then qn ({x, y : x m ∈ G, y m ∈ B m }) = µX m (G). Thus πnm (Am × G) = πn ({(x, y) : x =
(5.43) m
m
∈ A ,y
1 n−m+1 µY m (G) + n n
m
m−1 X
∈ G}) i−1 µY m−i (×m−i k=i A)µY i (×k=0 Gk ) (5.44)
i=1
with a similar expression for G × Am . Since there are a countable number of finite dimensional rectangles in BT with coordinates in G, we can use a diagonalization argument to extract a subsequence πnk of πn which converges on all of the rectangles. To do this enumerate all the rectangles, then pick a subsequence converging on the first, then a further subsequence converging on the second, and so on. The result is a limiting measure π on the finitedimensional rectangles, and this can be extended to a measure also denoted by π on (A, B(A))2 , that is, to a stationary pair process distribution. Eq. (5.44) implies that for each fixed m
5.12 Source Approximation and Codes
141
lim πn (Am × G) = π m (Am × G) = µY m (G)
n→∞
lim πn (G × Am ) = π m (G × Am ) = µX m (G)
n→∞
and hence for any cylinder F ∈ B(A) that π (AT × F ) = µY (F ) π (F × AT ) = µX (F ) Thus π induces the desired marginals and hence π ∈ Ps and ρ 0 (µX , µY ) ≤ Eπ ρ1 (X0 , Y0 ) = lim Eπnk ρ1 (X0 , Y0 ) k→∞
= lim n−1 k k→∞
nX k −1
Eqnk ρ1 (Xi , Yi ) = lim (ρ nk + nk ) = ρ(µX , µY ).
i=0
k→∞
2
Evaluating Process Distortion Evaluation of the rho-bar distortion or d-bar distance can in general be difficult analytically. Theorem 5.2 provides an important exception, if both of the processes are IID then the process distance is given by the distance between the zero-time samples, the first-order coupling distance. From Lemma 5.5, if the distance is with respect to the Hamming distance, this in turn is given by half the variation distance. In [66] the distance between Gaussian processes was shown to be the L2 distance between the square roots of their power spectral densities.
5.12 Source Approximation and Codes In Section 5.6 the approximation of the output of two codes applied to a common source was considered. A natural variation on this idea is to fix a code and look at the approximation of the two outputs resulting from different sources. We again focus on the d-bar distance. Lemma 5.6. Let µX and µY be distributions of two stationary random processes on a common discrete alphabet, let f be a sliding-block code of length N, and let µf (X) and µf (Y ) denote the corresponding output distributions of coding µX or µY with f . Then d(µf (X) , µf (Y ) ) ≤ Nd(µX , µY ).
142
5 Distortion and Approximation
Proof. If f is a sliding-block code of length N then it depends only on N Xm = (Xm , Xm+1 , . . . , Xm+N−1 ) for some fixed m. Choose a coupling of the two processes yielding Pr(X0 ≠ Y0 ) = d(µX , µY ). We have not shown that d-bar can be hit with equality like this, but it turns out to be the case (and the following argument works with the addition of a small > 0). This coupling of input processes implies a coupling of the output processes, so that with a slight abuse of notation we have from the union bound that N N d(µf (X) , µf (Y ) ) ≤ Pr(f (Xm ) ≠ f (Ym )) N N ≠ Ym )≤ ≤ Pr(Xm
m+N−1 X
Pr(Xi ≠ Yi )
i=m
= N Pr(X0 ≠ Y0 ) = Nd(µX , µY ).
2 Thus in particular the output of a finite-length sliding-block code is a continuous function of the input in d-bar distance (with respect to the Hamming distance).
5.13 d-bar Continuous Channels The d distance can be used to generalize some of the notions of discretealphabet channels by weakening the definitions. The first definition is the most important for channel coding applications. We confine interest to the d-bar or d0 distance, the ρ-distortion for the special case of the Hamming distance: ( ρ1 (x, y) = d1 (x, y) =
if x = y
1
if x 6= y.
Suppose that [A, ν, B] is a discrete alphabet channel and let νxn denote the restriction of the channel to B n , that is, the output distribution on Y n given an input sequence x. The channel is said to be d-continuous if for any > 0 there is an n0 such that for all n > n0 dn (νxn , νxn0 ) ≤ whenever xi = x 0 i for i = 0, 1, · · · , n. Alternatively, ν is d-continuous if lim sup sup n→∞
sup
an ∈An x,x 0 ∈c(an )
dn (νxn , νxn0 ) = 0,
where c(an ) is the rectangle defined as all x with xi = ai ; i = 0, 1, · · · , n−1. d-continuity implies the distributions on output n-tuples Y n given two input sequences are very close provided that the input sequences are identical over the same time period and that n is large.
5.13 d-bar Continuous Channels
143
This generalizes the notions of 0 or finite input memory and anticipation since the distributions need only approximate each other and do not have to be exactly the same. More generally we could consider ρ-continuous channels in a similar manner, but we will focus on the simpler discrete d-continuous channel. d-continuous channels possess continuity properties that will be useful for proving block and sliding-block coding theorems. They are “continuous” in the sense that knowing the input with sufficiently high probability for a sufficiently long time also specifies the output with high probability. The following two lemmas make these ideas precise. Lemma 5.7. Suppose that x, x ∈ c(an ) and d(νxn , νxn ) ≤ δ2 . This is the case, for example, if the channel is d continuous and n is chosen sufficiently large. Then νxn (Gδ ) ≥ νxn (G) − δ and hence inf
x∈c(an )
νxn (Gδ ) ≥
sup νxn (G) − δ.
x∈c(an )
Proof: Again we assume that the infima defining the d distance are actually minima and hence there is a pmf p on B n × B n such that X p(y n , bn ) = νxn (y n ) bn ∈B n
and X
p(bn , y n ) = νxn (y n );
bn ∈B n
that is, p has νxn and νxn as marginals, and 1 n Ep dn (Y n , Y ) = d(νxn , νxn ). n As previously done, this is true within > 0 and the proof follows in the same way with inequalities. Using the Markov inequality we can write νxn (Gδ ) = p(Y n ∈ Gδ ) ≥ p(Y = 1 − p(Y
n n
n
n
∈ G and dn (Y n , Y ) ≤ nδ) n
6∈ G or dn (Y n , Y ) > nδ) n
6 G) − p(dn (Y n , Y ) > nδ) ∈ 1 n ≥ νxn (G) − E(n−1 dn (Y n , Y )) ≥ νxn (G) − δ δ
≥ 1 − p(Y
144
5 Distortion and Approximation
proving the first statement. The second statement follows from the first.
2 Next suppose that [G, µ, U] is a stationary source, f is a stationary encoder which could correspond to a finite length sliding block encoder or to an infinite length one, ν is a stationary channel, and g is a length m sliding-block decoder. The probability of error for the resulting hookup is defined by Z ˆ0 ) = µν(E) = dµ(u)νf (u) (Eu ), Pe (µ, ν, f , g) = Pr(U0 6= U where E is the error event {u, y : u0 6= gm (Y− qm )} and Eu = {y : (u, y) ∈ E} is the section of E at u. Lemma 5.8. Given a stationary channel ν, a stationary source [G, µ, U], a length m sliding-block decoder, and two encoders f and φ, then for any positive integer r |Pe (µ, ν, f , g) − Pe (µ, ν, φ, g)| ≤ m + r Pr(f 6= φ) + m max sup dr (νxr , νxr 0 ). ar ∈Ar x,x 0 ∈c(ar ) r Proof: Define Λ = {u : f (u) = φ(u)} and Λr = {u : f (T i u) = φ(T i u); i = 0, 1 · · · , r − 1} =
r\ −1
T i Λ.
i=0
From the union bound µ(Λcr ) ≤ r µ(Λc ) = r Pr(f 6= φ).
(5.45)
m ) then From stationarity, if g = gm (Y−q
Z Pe (µ, ν, f , g) =
m dµ(u)νf (u) (y : gm (y−q ) 6= u0 )
r −1 Z 1 X m dµ(u)νf (u) (y : gm (yi−q ) 6= u0 ) r i=0 r −q Z m 1 X m ≤ + dµ(u)νfr (u) (y r : gm (yi−q ) 6= ui ) + µ(Λcr ). r r i=q Λr
=
Fix u ∈ Λr and let pu yield dr (νfr (u),φ(u) ); that is, P r νfr (u) (y r ), y r pu (y r , w r ) = νφ(u) (w r ), and
P
wr
(5.46)
pu (y r , w r ) =
5.13 d-bar Continuous Channels
145
r −1 1 X pu (y r , w r : yi 6= wi ) = dr (νfr (u),φ(u) ). r i=0
(5.47)
We have that r −q 1 X r m ν (y r : gm (yi−q ) 6= ui ) r i=q f (u) r −q 1 X m = pu (y r , w r : gm (yi−q ) 6= ui ) r i=q r −q r− q 1 X 1 X m m m r r ≤ pu (y , w : gm (yi−q ) 6= wi−q )+ pu (y r , w r : gm (wi−q ) 6= ui ) r i=q r i=q
≤
r −q 1 X r r pu (y r , w r : yi−q 6= wi−q ) + Pe (µ, ν, φ, g) r i=q
≤
r −q i−q+m 1 X X pu (y r , w r : yj 6= wj ) + Pe (µ, ν, φ, g) r i=q j=i−q
r ≤ mdr (νfr (u) , νφ(u) ) + Pe (µ, ν, φ, g),
which with (5.45)-(5.47) proves the lemma.
2
The following corollary states that the probability of error using sliding-block codes over a d-continuous channel is a continuous function of the encoder as measured by the metric on encoders given by the probability of disagreement of the outputs of two encoders. Corollary 5.2. Given a stationary d-continuous channel ν and a finite length decoder gm : B m → A, then given > 0 there is a δ > 0 so that if f and φ are two stationary encoders such that Pr(f 6= g) ≤ δ, then |Pe (µ, ν, f , g) − Pe (µ, ν, φ, g)| ≤ . Proof: Fix > 0 and choose r so large that max r a
sup x,x 0 ∈c(ar )
3m m ≤ , r 3
dr (νxr , νxr 0 ) ≤
and choose δ = /(3r ). Then Lemma 5.8 implies that |Pe (µ, ν, f , g) − Pe (µ, ν, φ, g)| ≤ .
2
146
5 Distortion and Approximation
Given an arbitrary channel [A, ν, B], we can define for any block length ˜, B] as the CBI channel with the same N a closely related CBI channel [A, ν probabilities on output N-blocks, that is, the same conditional probabilN ities for YkN given x, but having conditionally independent blocks. We ˜ the N-CBI approximation to ν. A channel ν is said to be conshall call ν ditionally almost block independent or CABI if given there is an N0 such that for any N ≥ N0 there is an M0 such that for any x and any N-CBI ˜ to ν approximation ν d(˜ νxM , νxM ) ≤ , all M ≥ M0 , where νxM denotes the restriction of νx to BN B , that is, the output distribution on Y N given x. A CABI channel is one such that the output distribution is close (in a d sense) to that of the N-CBI approximation provided that N is big enough. CABI channels were introduced by Neuhoff and Shields [133] who provided several examples alternative characterizations of the class. In particular they showed that finite memory channels are both d-continuous and CABI. Their principal result, however, requires the notion of the d distance between channels. Given two channels [A, ν, B] and [A, ν 0 , B], define the d distance between the channels to be N d(ν, ν 0 ) = lim sup sup d(νxn , ν 0 x ). n→∞
x
Neuhoff and Shields [133] showed that the class of CABI channels is exactly the class of primitive channels together with the d limits of such channels.
Chapter 6
Distortion and Entropy
Abstract Results are developed relating the goodness of approximation as measured by average Hamming distance between codes and and the d-bar distance between sources to the closeness of entropy rate. A few easy applications provide important properties of entropy rate. One might suspect that if two codes for a common source closely approximate each other, then the resulting output entropies should also be close. Similarly, it seems reasonable to expect that if two random processes well approximate each other, than their entropy rates should be close. Such notions of continuity of entropy with respect to the goodness of approximation between codes and processes are the focus of this chapter and are fundamental to to the development and extensions to follow. A few easy applications are collected in this chapter.
6.1 The Fano Inequality A classic result of this type, showing that closeness in the average Hamming distance between two codes forces the entropy to be close, was first proved by Fano and is called Fano’s inequality [38]. The result has a variety of extensions and applications. Lemma 6.1. Given two finite alphabet measurements f and g on a common probability space (Ω, B, P ) having a common alphabet A or, equivalently, given the corresponding partitions Q = {f −1 (a); a ∈ A} and R = {g −1 (a); a ∈ A}, define the error probability Pe = |Q − R| = Pr(f 6= g). Then H(f |g) ≤ h2 (Pe ) + Pe ln(kAk − 1) and |H(f ) − H(g)| ≤ h2 (Pe ) + Pe ln(kAk − 1) R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_6, © Springer Science+Business Media, LLC 2011
147
148
6 Distortion and Entropy
and hence entropy is continuous with respect to partition distance for a fixed measure. Proof: Let M = kAk and define a measurement r : A × A → {0, 1, · · · , M − 1} by r (a, b) = 0 if a = b and r (a, b) = i if a 6= b and a is the ith letter in the alphabet Ab = A − b. If we know g and we know r (f , g), then clearly we know f since either f = g (if r (f , g) is 0) or, if not, it is equal to the r (f , g)th letter in the alphabet A with g removed. Since f can be considered a function of g and r (f , g), H(f |g, r (f , g)) = 0 and hence H(f , g, r (f , g)) = H(f |g, r (f , g)) + H(g, r (f , g)) = H(g, r (f , g)). Similarly H(f , g, r (f , g)) = H(f , g). From Lemma 3.2 H(f , g) = H(g, r (f , g)) ≤ H(g) + H(r (f , g)) or H(f , g) − H(g) = H(f |g) ≤ H(r (f , g)) = −P (r = 0) ln P (r = 0) −
M−1 X
P (r = i) ln P (r = i).
i=1
Since P (r = 0) = 1 − Pe and since
P
i6=0
H(f |g) ≤ −(1 − Pe ) ln(1 − Pe ) − Pe
P (r = i) = Pe , this becomes
M−1 X i=1
P (r = i) P (r = i) ln − Pe ln Pe Pe Pe
≤ h2 (Pe ) + Pe ln(M − 1) since the entropy of a random variable with an alphabet of size M − 1 is no greater than ln(M − 1). This proves the first inequality. Since H(f ) ≤ H(f , g) = H(f |g) + H(g), this implies H(f ) − H(g) ≤ h2 (Pe ) + Pe ln(M − 1). Interchanging the roles of f and g completes the proof.
2
6.1 The Fano Inequality
149
The lemma can be used to show that related information measures such as mutual information and conditional mutual information are also continuous with respect to the partition metric. The following corollary extends the the lemma to repeated measurements. Similar extensions may be found in Csiszár and Körner [27]. Again let f and g denote finite-alphabet measurements on a common probability space, but now interpret them as sliding-block codes as in Section 2.6; that is, let T denote a transformation on the common space (e.g., the shift on a sequence space) and define fi = f T i , g = gT i so that {fn , gn } is a pair process with a common finite alphabet. Define the nth order per-symbol or mean probability of error Pe(n) =
n−1 1 X Pr(fi 6= gi ) n i=0
and observe that this is simply the normalized average of the additive fidelity criterion corresponding to the Hamming distance: n−1 X 1 Pe(n) = E dH (xi , yi ) . n i=0 (n)
If the transformation (or the pair process) is stationary, then Pe (1) Pe = Pe .
=
Corollary 6.1. Given two sequences of measurements {fn } and {gn } with finite alphabet A on a common probability space, 1 H(f n |g n ) ≤ Pe(n) ln(kAk − 1) + h2 (Pe(n) ) n and |
1 1 H(f n ) − H(g n )| ≤ Pe(n) ln(kAk − 1) + h2 (Pe(n) ). n n
If {fn , gn } are also AMS and hence the limit P e = lim Pe(n) n→∞
exists, then if we define H(f | g) = lim
n→∞
1 1 H(f n | g n ) = lim (H(f n , g n ) − H(g n )), n→∞ n n
where the limits exist since the processes are AMS, then H(f |g) ≤ P e ln(kAk − 1) + h2 (P e ) |H(f ) − H(g)| ≤ P e ln(kAk − 1) + h2 (P e ).
150
6 Distortion and Entropy
Proof: From the chain rule for entropy (Corollary 3.6), Lemma 3.12, and Lemma 6.1 H(f n |g n ) =
n−1 X i=0
≤
n−1 X
H(fi |f i , g n ) ≤
n−1 X
H(fi |g i ) ≤
i=0
n−1 X
H(fi |gi )
i=0
Pr(fi 6= gi ) ln(kAk − 1) + h2 (Pr(fi 6= gi ))
i=0
from the previous lemma. Dividing by n yields the first inequality which implies the second as in the proof of the previous lemma. If the processes are jointly AMS, then the limits exist and the entropy rate results 2 follows from the continuity of h2 by taking the limit.
6.2 Code Approximation and Entropy Rate Corollary 6.1 has two simple but extremely important implications summarized in the next corollary. The first part is immediate, that two sliding-block codes which closely approximate each other in the code distance must have approximately the same entropy rate. The second part applies this observation to draw the similar conclusion that entropy rate of a source is a continuous function with respect to the Ornstein d-bar distance. Corollary 6.2. For a fixed AMS source, entropy rate of a sliding-block encoding of the source is a continuous function of the (Hamming) code distance. Entropy rate is a continuous function of the source with respect to the Ornstein d-bar distance. Proof. The first part follows since P e in Corollary 6.1 is the code distance between sliding-block codes f and g. For the second part, if we can make dH between two processes {fn } and {gn } arbitrarily small, then there is a coupling which yields an average Hamming distortion P e as small as we would like, which in turn implies from Corollary 6.1 that the two entropy rates are small. 2 Combining Lemma 5.2 and Corollary 6.1 immediately yields the following corollary, which permits us to study the entropy rate of general stationary codes by considering codes which depend on only a finite number of inputs (and hence for which the ordinary entropy results for random vectors can be applied). Corollary 6.3. Let f be a stationary code of an AMS process X. As in (5.24–5.25) define for positive integers n define Fn = σ (X0 , X1 , · · · , Xn )
6.2 Code Approximation and Entropy Rate
151
in the one-sided case and σ (X−n , · · · , Xn ) in the two-sided case. Given > 0 there exists for sufficiently large n a code g measurable with respect to Fn such that | H(f ) − H(g) |≤ . Corollary 6.3 can be used to show that entropy rate, like entropy, is reduced by coding. The general stationary code is approximated by a code depending on only a finite number of inputs and then the result that entropy is reduced by mapping (Lemma 3.3) is applied. Corollary 6.4. Given an AMS process {Xn } and a stationary coding f of the process, then H(X) ≥ H(f ), that is, stationary coding reduces entropy rate. Proof: For integer n define Fn = σ (X0 , X1 , · · · , Xn ) in the one-sided case and σ (X−n , · · · , Xn ) in the two-sided case. Then Fn asymptotically generates B(AX )∞ . Hence given a code f and an > 0 we can choose using the finite alphabet special case of the previous lemma a large k and a Fk -measurable code g such that |H(f ) − H(g)| ≤ . We shall show that H(g) ≤ H(X), which will prove the lemma. To see this in the one-sided case observe that g is a function of X k and hence g n depends only on X n+k and hence H(g n ) ≤ H(X n+k ) and hence 1 1 n H(g n ) ≤ lim H(X n+k ) = H(X). n→∞ n n→∞ n n + k
H(g) = lim
In the two-sided case g depends on {X−k , · · · , Xk } and hence gn depends on {X−k , · · · , Xn+k } and hence H(g n ) ≤ H(X−k , · · · , X−1 , X0 , · · · , Xn+k ) ≤ H(X−k , · · · , X−1 )+H(X n+k ). Dividing by n and taking the limit completes the proof as before.
2
Dynamical Systems and Random Processes It is instructive to apply Corollary 6.4 to relate the idea of the entropy of a dynamical system with the entropy rate of a random process. The result is not required for later coding theorems, but it provides insight into the connections between entropy as considered in ergodic theory and entropy as used in information theory. In addition, the development involves some ideas of coding and approximation which are useful in
152
6 Distortion and Entropy
proving the ergodic theorems of information theory used to prove coding theorems. Let {Xn } be a random process with alphabet AX . Let A∞ X denote the one or two-sided sequence space. Consider the dynamical system ∞ (Ω, B, P , T ) defined by (A∞ X , B(AX ) , P , T ), where P is the process distribution and T the shift. Recall from Section 3.1 and Section 2.6 that a stationary coding or infinite length sliding-block coding of {Xn } is a measurable mapping f : A∞ X → Af into a finite alphabet which produces an encoded process {fn } defined by fn (x) = f (T n x); x ∈ A∞ X. The entropy H(P , T ) of the dynamical system was defined in (3.4) by H(P , T ) = sup H P (f ), f
the supremum of the entropy rates of finite alphabet stationary codings of the original process. We shall show that if the original alphabet is finite, then the entropy of the dynamical system is exactly the entropy rate of the process. Theorem 6.1. Let {Xn } be a random process with finite alphabet AX . Let A∞ X denote the one or two-sided sequence space. Consider a dynamical sys∞ tem (Ω, B, P , T ) defined by (A∞ X , B(AX ) , P , T ), where P is an AMS process distribution and T is the shift. Then H(P , T ) = H(X). Proof: From (3.5), H(P , T ) ≥ H(X). Conversely suppose that f is a code which yields H(f ) ≥ H(P , T ) − . Since f is a stationary coding of the AMS process {Xn }, Corollary 6.4 implies that H(f ) ≤ H(X). Thus H(P , T ) − ≤ H(X), which completes the proof since is arbitrary. 2
6.3 Pinsker’s and Marton’s Inequalities Fano’s inequality shows that small probability of error between two processes implies that the entropy rates of the processes must be close. A converse of sorts for the special case where one of the two processes is IID follows from an inequality of Marton, which in turn follows from an inequality of Pinsker. In this section these inequalities are derived and discussed. The following inequality provides an upper bound to the variation distance between to probability measures in terms of the divergence (see [151], p. 58 of [27], (13) of [166]).
6.3 Pinsker’s and Marton’s Inequalities
153
Lemma 6.2. Pinsker’s Inequality Given two probability measures M and P , the variation distance and divergence satisfy the inequality q var(P , M) ≤ 2D(P kM). Proof. Assume that M P since the result is trivial otherwise because the right-hand side is infinite. The inequality will follow from the first statement of Lemma 5.4 and the following inequality: Given 1 ≥ p, m ≥ 0, p 1−p p ln + (1 − p) ln − 2(p − m)2 ≥ 0. (6.1) m 1−m To see this, suppose the truth of (6.1). Since F can be chosen so that 2(P (F ) − M(F )) is arbitrarily close to d(P , M), given > 0 choose a set F such that [2(P (F ) − M(F ))]2 ≥ d(P , M)2 − 2. Since {F , F c } is a partition, d(P , M)2 ≥ 2 P (F ) 1 − P (F ) P (F ) ln + (1 − P (F )) ln − 2(P (F ) − M(F ))2 − . M(F ) 1 − M(F )
D(P kM) −
If (6.1) holds, then the right-hand side is bounded below by −, which proves the lemma since is arbitrarily small. To prove (6.1) observe that the left-hand side equals zero for p = m, has a negative derivative with respect to m for m < p, and has a positive derivative with respect to m for m > p. (The derivative with respect to m is (m − p)[1 − 4m(1 − m)]/[m(1 − m).) Thus the left hand side of (6.1) decreases to its minimum value of 0 as m tends to p from above or below. 2 The lemma together with Lemma 5.5 yield the following corollary. Corollary 6.5. Given two probability measures M and P , the transportation with respect to the Hamming distance (the first order Ornstein’s d-bar distance) and divergence satisfy the inequality s D(P kM) dH (P , M) ≤ . 2 Marton extended Pinsker’s inequality to vectors produced by processes when one of the processes is memoryless and thereby obtained in the limit an inequality between Ornstein’s d-bar distance between processes and the relative entropy rate of an arbitrary process with respect to the IID process [120]. She subsequently extended this result from IID to a class of Markov processes [121], but we shall concentrate on the IID result, which is stated in the following lemma. The proof follows Shields [166].
154
6 Distortion and Entropy
Lemma 6.3. Marton’s Inequality Suppose that X n and Y n are random vectors with a common finite alphabet and probability mass functions µX n and µY n and that Y n is memoryless (µY n is a product pmf). Then the d-bar distance (mean Hamming) satisfies s 1 D(µX n kµY n ) . (6.2) dn (µX n , µY n ) ≤ n 2n and hence in the limit s d(µX , µY ) ≤
D(µX kµY ) 2
(6.3)
Proof. Suppose that the common finite alphabet for Xn and Yn is A. First note that for the case n = 1 Marton’s inequality and Pinker’s inequality are the same so that s D(µX0 kµY0 ) . (6.4) d1 (µX0 , µY0 ) ≤ 2 Denote by p (1) (x0 , y0 ) the coupling yielding (with the usual caveat) the Hamming transportation distance d1 (µX0 , µY0 ), that is, the pmf on A × A with marginals µX0 and µY0 yielding the smallest E (dH (X0 , Y0 )) = Pr(X0 ≠ Y0 ). For n ≥ 2 consider Pinsker’s inequality applied to the distributions P and M corresponding to the probability mass functions µXn−1 |X n−1 (xn−1 | x n−1 ) and µYn−1 (yn−1 ): dH (µXn−1 |X n−1 (· | x n−1 ), µYn−1 ) ≤ s D(µXn−1 |X n−1 (· | x n−1 )kµYn−1 ) n , x ∈ An . 2
(6.5)
Let p (n) (xn−1 , yn−1 | x n−1 ) denote the coupling (on A × A) between the pmfs µXn−1 |X n−1 (· | x n−1 ) and µYn−1 on A yielding dH (µXn−1 |X n−1 (· | x n−1 ), µYn−1 ) (these exist from Theorem 5.2). Taking expectations in (6.5) yields X
µX n−1 (x n−1 )dH (µXn−1 |X n−1 (· | x n−1 ), µYn−1 ) ≤
x n−1
s X x n−1
µX n−1 (x
n−1
)
D(µXn−1 |X n−1 (· | x n−1 )kµYn−1 ) . 2
(6.6)
Use these pmfs to construct a new pmf π (n) (x n , y n ) on An × An defined by
6.3 Pinsker’s and Marton’s Inequalities
π (n) (x n , y n ) = p (1) (x0 , y0 )
155 n Y
p (i) (xi−1 , yi−1 | x i−1 ).
i=2
This joint pmf has as its marginals µX n and µY n and hence is a coupling of these two distributions. This implies that Eπ (n) (dn (X n , Y n )) ≥ dn (µX n , µY n ).
(6.7)
We also have that Eπ (n) (dn (X n , Y n )) n−1 n−1 n−1 X X X dH (Xi , Yi ) = Eπ (n) (dH (Xi , Yi )) = Eπ (i) (dH (Xi , Yi )) = Eπ (n) i=0
=
XX
p
(1)
i=0
i=0
(x0 , y0 )dH (X0 , Y0 )
x0 y0
+
n X X i=2
X
µX i−1 (x i−1 )
= dH (µX0 , µY0 ) +
p (i) (xi−1 , yi−1 | x i−1 )dH (xi−1 , yi−1 )
xi−1 ,yi−1
x i−1
n X X
µX i−1 (x i−1 )dH (µXi−1 |X i−1 (· | x i−1 ), Yi−1 ).
i=2 x i−1
Apply Pinsker’s inequality from (6.4) and (6.6) and use (6.7) to write dn (µX n , µY n ) ≤ dH (µX0 , µY0 ) +
n X X
µX i−1 (x i−1 )dH (µXi−1 |X i−1 (· | x i−1 ), Yi−1 )
i=2 x i−1
s ≤
n D(µX0 kµY0 ) X X µX i−1 (x i−1 ) + 2 i=2 i−1
s
D(µXi−1 |X i−1 (· | x i−1 )kµYi−1 )
x
2
.
Use the concavity of the square root twice to obtain s dn (µX n , µY n ) 1 D(µX0 kµY0 ) ≤ + n n 2 sP n i−1 )D(µ i−1 )kµ 1 X Yi−1 ) x i−1 µX i−1 (x Xi−1 |X i−1 (· | x n i=2 2 v u n X u X u 1 µX i−1 (x i−1 )D(µXi−1 |X i−1 (· | x i−1 )kµYi−1 ). ≤t D(µX0 kµY0 ) + 2n i=2 i−1 x
The following string of equalities shows that the term in parentheses is D(µX n kµY n ), which completes the proof of (6.2):
156
6 Distortion and Entropy
D(µX n kµY n ) X µX n (x n ) = µX n (x n ) ln µY n (x n ) xn Qn−1 X µX0 (x0 ) i=1 µXi |X i (xi | x i ) n µX n (x ) ln = Qn−1 µY0 (x0 ) i=1 µYi (yi ) xn n−1 X Y µX |X i (xi | x i ) (x ) µ X 0 i µX n (x n ) ln 0 = µY0 (x0 ) i=1 µYi (yi ) xn =
X
µX n (x n ) ln
xn
= D(µX0 kµY0 ) +
n−1 XX µX |X i (xi | x i ) µX0 (x0 ) µX n (x n ) ln i + µY0 (x0 ) µYi (yi ) i=1 x n n−1 XX i=1 x i
= D(µX0 kµY0 ) +
n−1 XX i=1
µX i (x i )
X
µXi |X i (xi | x i ) ln
µXi |X i (xi | x i )
xi
µYi (yi )
µX i (x i )D(µXi |X i (· | x i )kµYi ).
xi
2 While Fano’s inequality deals with conditional entropy at its most basic level, Marton’s inequality deals with relative entropy. Just as Fano’s inequality results in a relationship between the d-bar distance and the difference of entropy rates, Marton’s inequality also has an implication for entropy rates. If Y is IID and equiprobable as in the case of fair coin flips, Marton’s inequality immediately yields the following corollary. Corollary 6.6. Suppose that µX and µY are distributions of two stationary processes with a common finite alphabet and that Y is both IID and has equiprobable marginals. Then s H(Y ) − H(X) d(X, Y ) ≤ . 2 Thus if an arbitrary stationary process has entropy rate close to that of an IID equiprobable source with the same finite alphabet, then it must be also close in Ornstein’s d-bar distance.
6.4 Entropy and Isomorphism The results derived thus far in this chapter have as an easy application one of the most important results of ergodic theory, the KolmogorovSinai theorem demonstrating that a necessary condition for two dynamical systems to be isomorphic is that they have the same entropy rate.
6.4 Entropy and Isomorphism
157
Roughly speaking, two random processes are isomorphic if each can be coded into the other in a stationary and invertible way. The primary difficulty in making this result precise is developing the necessary definitions, which will be related to the coding language used here. The focus is on dynamical systems rather than on random processes because the latter are more general and form the traditional context for treating isomorphic processes. The initial material follows [58]. There are several notions of isomorphism: isomorphic measurable spaces, isomorphic probability spaces, and isomorphic dynamical systems. Isomorphic random processes are a special case of the latter.
Isomorphic Measurable Spaces Two measurable spaces (Ω, B) and (Λ, S) are isomorphic if there exists a measurable function φ : Ω → Λ that is one-to-one and has a measurable inverse φ−1 . In other words, the inverse image φ−1 (λ) of a point λ ∈ Λ consists of exactly one point in Ω and the inverse mapping so defined, say γ : Λ → Ω, γ(λ) = φ−1 (λ), is itself a measurable mapping. The function φ (or its inverse γ) with these properties is called an isomorphism. An isomorphism between two measurable spaces is thus an invertible mapping between the two sample spaces that is measurable in both directions.
Isomorphic Probability Spaces Two probability spaces (Ω, B, P ) and (Λ, S, Q) are isomorphic if there is an isomorphism φ : Ω → Λ between the two measurable spaces (Ω, B) and (Λ, S) with the added property that Q = Pφ and P = Qφ−1 that is, Q(F ) = P (φ−1 (F )); F ∈ S; P (G) = Q(φ(G)); G ∈ B. Two probability spaces are isomorphic 1. if one can find for each space a random variable defined on that space that has the other as its output space, and 2. the random variables can be chosen to be inverses of each other; that is, if the two random variables are φ and γ, then φ(γ(λ)) = λ and γ(φ(ω)) = ω.
158
6 Distortion and Entropy
Note that if the two probability spaces (Ω, B, P ) and (Λ, S, Q) are isomorphic and φ : Ω → Λ is an isomorphism with inverse γ, then the random variable φγ defined by φγ(λ) = φ(γ(λ)) is equivalent to the identity random variable i : Λ → Λ defined by i(λ) = λ.
Isomorphism Mod 0 A weaker notion of isomorphism between two probability spaces is that of isomorphism mod 0 . Two probability spaces are isomorphic mod 0 or metrically isomorphic if the mappings have the desired properties on sets of probability 1, that is, the mappings can be defined except for null sets in the respective spaces. Thus two probability spaces (Ω, B, P ) and (Λ, S, Q) are isomorphic (mod 0) if there are null sets Ω0 ∈ B and λ0 ∈ S and a measurable one-to-one onto map φ : Ω − Ω0 → Λ − Λ0 with measurable inverse such that Q(F ) = P (φ−1 (F )) for all F ∈ S. This weaker notion is the standard one in ergodic theory.
Isomorphic Dynamical Systems Roughly speaking, two dynamical systems are isomorphic if one can be coded or filtered onto the other in an invertible way so that the coding carries one transformation into the other, that is, one can code from one system into the other and back again and coding and transformations commute. Two dynamical systems (Ω, B, P , S) and (Λ, S, m, T ) are isomorphic if there exists an isomorphism f : Ω → Λ such that T φ(ω) = φ(Sω); ω ∈ Ω. As with the isomorphism of probability spaces, isomorphic mod 0 means that the properties need hold only on sets of probability 1, but we also require that the null sets in the respective spaces on which the isomorphism is not defined to be invariant with respect to the appropriate transformation. Henceforth isomorphism of dynamical systems will be taken to mean isomorphism mod 0 with this constraint.
Isomorphic Random Processes Suppose that the probability space (Λ, S, m) is the sequence space of a directly given finite-alphabet random process, say (ATX , BTAX , µX ), T is the shift on this space, and Π0 the sampling function on this space, then the random process Xn = Π0 T n defined on (Λ, S, m) is equivalent to the ran-
6.4 Entropy and Isomorphism
159
dom process Π0 (φS n ) defined on the probability space (Ω, B, P ). More generally, any random process of the form gT n defined on (Λ, S, m) is equivalent to the random process g(φS n ) defined on the probability space (Ω, B, P ). A similar conclusion holds in the opposite direction. Thus, any random process that can be defined on one dynamical system as a function of transformed points possesses an equivalent model in terms of the other dynamical system and its transformation. In addition, not only can one code from one system into the other, one can recover the original sample point by inverting the code (at least with probability 1). Isomorphism provides a variety of equivalent models for random processes. The models can be quite different in appearance, yet each can be transformed into the other by coding (for discrete alphabet processes) or filtering (for continuous alphabet processes). If X = {Xn } and Y = {Yn } are two finite-alphabet random processes and they are isomorphic, then each is equivalent to a random process formed by a stationary or sliding-block coding of the other. Equivalent processes have the same process distributions and hence the same entropy rates. Suppose that Y is equivalent to a stationary coding f of X and that X is equivalent to a stationary coding g = f −1 of Y . From Corollary 6.3, H(X) ≥ H(f (X)) = H(Y ) and H(Y ) ≥ H(g(X)) = H(X) and hence H(X) = H(Y ). This yields the random process special case of one of the most famous results of ergodic theory, originally due to Kolmogorov and Sinai. It is summarized in the following theorem. Theorem 6.2. Kolmogorov-Sinai Theorem A necessary condition for two AMS random processes to be isomorphic is that they have the same entropy rates. In the 1970s, Donald Ornstein proved in a remarkable series of papers that were summarized in [139, 140] that equal entropy was also a sufficient condition for two processes to be isomorphic if the processes were B-processes, stationary codings or filterings of IID processes. This result is now known as Ornstein’s isomorphism theorem or as the KolmogorovSinai-Ornstein isomorphism theorem. The general result is beyond the scope of this book and no attempt at a proof will be made here. On the other hand, stating the result reinforces the importance of the concept of entropy and entropy rate beyond the domain of information and coding theory, and the result is useful for obtaining and interpreting results relating source coding and simulating random processes, results which have not yet yielded to simpler proofs. For reference we state without proof the Ornstein theorem and a related result due to Sinai which is used in the proof of Ornstein’s theorem. Excellent accessible (with respect to the original papers) treatments of the Ornstein theory and the Sinai theorem can be found in books by Shields [164] and by Kalikow and McCutcheon [84].
160
6 Distortion and Entropy
Theorem 6.3. Ornstein Isomorphism Theorem Two B-processes are isomorphic if and only if they have the same entropy rate. Theorem 6.4. Sinai’s Theorem Suppose that {Un } is a stationary and ergodic process with entropy rate H(U), that {Xn } is a B-process with distribution µX and entropy rate H(X), and that H(X) ≤ H(U). Then there is a sliding-block coding of U that has distribution µX . Sinai’s theorem implies half the Ornstein theorem by showing that a specified B-process can always be obtained by stationary coding of any stationary and ergodic process having equal or greater entropy rate. The hard part, however, is the other half — the demonstration that if the entropy rates are equal and the process being encoded is also required to be a B-process, then the stationary code can be made invertible. The usual statement of the Sinai theorem is less general and considers the case where {Xn } is IID rather than a B-process. The more general result quoted here can be found, e.g., as 636 Theorem in [84], p. 141.
6.5 Almost Lossless Source Coding The Ornstein and Sinai theorems at the heart of modern ergodic theory are difficult to prove and require intricate and long arguments. Source coding ideas using the properties of entropy and the Rohlin theorem, however, yield approximate versions of the Ornstein results with proofs that are germane to the present development and provide a relatively simple example of techniques to be used in proving coding theorems in later chapters. The results also provide useful interpretations of the Ornstein and Sinai results as idealized source coding and simulation, and of almost lossless source coding as an approximation to isomorphism and the Sinai theorem. To simplify the notation and the presentation, this section concentrates on an archetypal problem of almost lossless source coding, that is, of converting a stationary and ergodic discrete-alphabet source with an arbitrary alphabet and known entropy rate into bits in a way that is almost lossless in that the coding allows recovery of the original sequence with small probability of error. This is a special case coding for small average distortion with respect to a Hamming distance on symbols. For simplicity and clarity, for the time being only the special case of a source with entropy rate of 1 bit per symbol is considered. The goal is to characterize the behavior of sliding-block codes. More general results will be developed later. Suppose that {Xn } is a stationary and ergodic source with a discrete alphabet A, process distribution µX , and entropy rate H(X) = H(µX ) = 1, where base 2 logarithms will be used throughout this section.
6.5 Almost Lossless Source Coding
161
Almost-Lossless Block Codes The first step is a slight variation of the block code constructed using the asymptotic equipartition property (AEP) of Section 4.5. Given > 0, there is an n0 sufficiently large so that for all n ≥ n0 the set of entropy-typical sequences Gn = {x n : 2−n(H(X)+) ≤ µX n (x n ) ≤ 2−n(H(X)−) } satisfies c µX n (Gn ) ≤ .
As outlined in Section 4.5, an initial approach to coding source n tuples into binary n-tuples is to index the length n sequences in Gn by binary n-tuples, encode each input vector into the index if the vector is in Gn and some arbitrary binary n-tuple otherwise, and then decode the binary n-tuple into the corresponding x n . With high probability (greater than 1 − ) the decoded vector will be the input vector, which in turn implies the symbols will be correct with high probability. An immediate problem, however, is that in general there will be too many vectors in Gn . Observe that the number of vectors Gn , kGn k, can be bound above using the inequality X µX n (x n ) 1 ≥ Pr(X n ∈ Gn ) = x n ∈Gn
≥
X x n ∈G
2−n(H(X)+) = kGn k2−n(1+) n
so that kGn k ≤ 2n(1+) , but there are only 2n available binary n-tuples as indices. Since there can be more than 2n entropy-typical sequences of a source with entropy rate 1, there can be too few indices in {0, 1}n for Gn . We could avoid this problem by requiring that the source have entropy rate H(X) strictly less than 1, but it is desirable to consider the case where the source has the maximal entropy rate that can be squeezed through the binary encoded process, which is the 1 bit per sample of fair coin flips. So instead we modify the scheme. One possible modification is as follows. We will use a slightly larger blocklength N for the code than the vector dimension n of the entropytypical vectors. Toward this end choose δ0 > 0 and set k = k(N) = bδ0 Nc + 1 = dδ0 Ne,
(6.8)
where as usual br c is the greatest integer less than or equal to r . and n = N − k so that N = n + k. Note for later use that
162
6 Distortion and Entropy
δ0 N ≤ k ≤ δ0 N + 1 (1 − δ0 )N − 1 ≤ n ≤ N(1 − δ0 ).
(6.9) (6.10)
The encoder ignores the first k symbols of each input block. If the n-tuple following the first k symbols is in Gn , it will be coded into the binary N-tuple index. Map all n-tuples not in Gn into a fixed binary ntuple, say the all 0 n-tuple. We need to reserve 2n(1+) indices as above, and we have 2N binary N-tuples, so there will be enough indices for all of the entropy-typical N-tuples if (N − k)(1 + ) ≤ N or N ≤ k = dδ0 Ne. 1+ This will be true if δ0 ≥ /(1 + ), so choose δ0 = . The decoder will map the first k ≈ N symbols of each block into an arbitrary k-tuple such as an all 0 k-tuple. The remaining n ≈ N(1 − ) binary symbols will be viewed as an index into Gn and the vector indexed by the binary n-tuple will be the output. Note that if the corresponding input n-tuple was in fact in Gn , all of the n corresponding output symˆi denote the resulting reconstruction bols will be decoded correctly. Let X symbols. The average error probability over a block satisfies Pe(N) =
N−1 k+n−1 n1 X 1 X k ˆi ) ≤ ˆi ). + Pr(Xi ≠ X Pr(Xi ≠ X N i=0 N N n i=k
ˆi } for some i ∈ {k, k + 1, . . . , k + n − 1} is a Since the event {Xi ≠ X n ˆn } , Pr(Xi ≠ X ˆi ) ≤ Pr(X n 6∈ Gn ) so that with subset of the event {Xk ≠ X k k stationarity, Pe(N) ≤
n k 1 c c + µX n (Gn + µX n (Gn ) ≤+ ), N N N
(6.11)
where we have used (6.9). Invoking the AEP of Section 4.5 with N satisfying N(1 − δ0 ) ≥ n0 we have Pe(N) ≤ 2 +
1 . N
(6.12)
Thus there is an N0 such that Pe(N) ≤ 3, all N ≥ N0 . so the average probability of error can be made as small as we would like by choosing a sufficiently large blocklength for the block code. Note that
6.5 Almost Lossless Source Coding
Pe(N) =
163
N−1 1 X ˆi ) , E dH (Xi , X N i=0
the average Hamming distortion for the block. In the limit of large block length, block codes can be used to achieve small average distortion in the sense that a discrete alphabet source with entropy rate 1 bit per symbol can be coded into a binary sequence from which we can recover the original source with asymptotically vanishing mean per symbol error probability. As previously discussed, however, a ˆn in a block coding system problem with the reproduction sequence X is that it will be neither stationary nor ergodic in general, and hence the reconstructed sequence lacks important statistical properties of the original source. In practical terms, there might (and often will) be artifacts in the reproduction due to the blocking, and these artifacts can be objectionable perceptually even if the average distortion is small. The end goal of this section will be to construct a sliding-block code with similar average distortion, but having the property that the reproduction (and the binary encoded sequence) is both stationary and ergodic.
Asynchronous Block Code Essential to the operation of an ordinary block code is the synchronization between the decoder and encoder — the decoder knows a priori where the code blocks begin so it knows how to interpret binary N-tuples as indices. When the block code is stationarized to form a sliding-block code, this synchronization is lost. As the next step towards constructing a sliding-block code we again modify the block code so that the binary codewords can be located even if the decoder does not know a priori where the block boundaries are. Before synchronizing the code, assume that is fixed as before and that the encoder blocks are divided as before into an initial k ≈ N symbols which will be ignored, followed by n ≈ N(1 − ) source symbols to be block coded using Gn . Now focus on the binary index codebook containing a subset of {0, 1}N and on the decoder. A classic method for accomplishing the goal of self-synchronization of binary N-tuples is to initiate each binary code block of length N with a synchronization sequence (or sync sequence, for short), a binary sequence of length, say, m that identifies the beginning of a code block. Each binary N-tuple codeword will consist of a common sync sequence of length m followed by a binary K-tuple with K = N − m. To ensure that the sync sequence always identifies the first m symbols of a binary code block, we no longer allow the remaining K binary symbols to be unconstrained — we now prohibit the appearance of the sync sequence
164
6 Distortion and Entropy
anywhere within the binary K-tuple that follows a sync sequence. Furthermore, for any ` = 1, 2, . . . , m − 1 we cannot allow the final ` symbols of a sync m-tuple followed by the first m − ` binary symbols of any allowed binary K-tuple (that is, any binary K-tuple appearing in the last K positions of an allowed binary N-tuple index) to equal the sync sequence. This last problem is easy to avoid. If m is even, then choose the first m/2 symbols of the sync sequence to be 0s and the remainder 1s. If m is odd, then choose the first (m − 1)/2 symbols to be 0s and the remainder 1s. If ` is greater than m/2 for m even or (m − 1)/2 if m is odd, then it is not possible for an overlap of sync and codeword to be mistaken for a sync since the possible false alarm begins with a 1 while a real sync must begin with a 0. if ` falls in the first half of a true sync, there will be insufficient 0s in the k-tuple to be mistaken for a sync. Thus we need only be concerned about avoiding a sync sequence inside a binary K-tuple following a sync. A sync sequence can occur in any of K − m = N − 2m positions in a binary K-tuple, and all of the 2N−2m binary K tuples containing a sync in any of the N − 2m possible positions are disallowed from the index set. After removing all of these disallowed sequences there will be at least 2K − (N − 2m)2N−2m K-tuples remaining for indexing the codebook Gn . Note that we have overcounted the number of sequences removed so that we have at least M remaining since K-tuples with two or more sync sequences within them get removed multiple times. Thus the condition required for ensuring that there are enough indices for the words in Gn is 2K − (N − 2m)2N−2m ≥ 2n(1+) (6.13) To relate , n, k chosen previously to parse the encoder block, we now derive the necessary conditions for K and m for obtaining arbitrarily small average probability of error. Analogous to the synchronous case, fix δ1 > 0 to be chosen shortly and define the sync sequence length similarly to (6.8) by m = m(N) = dδ1 Ne (6.14) and set K = N − m. As earlier, δ1 N ≤ m ≤ δ1 N + 1 (1 − δ1 )N − 1 ≤ K ≤ N(1 − δ1 ). From (6.13) and the definitions we have the inequalities 2K − (N − 2m)2N−2m ≥ 2N(1−δ1 ) − (N − 2δ1 N − 2)2N−2δ1 N−2 2(1−)N(1+) ≥ 2n(1+) and hence (6.13) will be satisfied and there will be sufficient indices for all of the vectors in Gn if
6.5 Almost Lossless Source Coding
165
2N(1−δ1 ) − (N − 2δ1 N − 2)2N−2δ1 N−2 ≥ 2(1− or 2−Nδ1 ≥ 2− or
2N
2 )N
+ (N − 2δ1 N − 2)2−2δ1 N−2
2
1 ≥ 2(δ1 − )N + [N(1 − 2δ1 N) − 2]2−δ1 N−2 . √ If we choose δ1 < , then the term on the right goes to zero with N and hence for sufficiently large blocklength N there are sufficient binary N tuples in a self-synchronized code for all vectors in Gn . The analysis of the mean probability of error follows exactly as before. Note that δ0 yields the fraction of symbols k in the initial and ignored symbols in the input source word, while δ1 yielded the fraction of initial symbols constituting the sync sequence in the output or encoded binary word.
Sliding-Block Code Let N and > 0 remain as before, where now N is chosen large enough to ensure that µX n (Gn ) ≤ in the previous subsections and construct an asynchronous block code of blocklength N as there described. The symbols k and m retain their meaning as the length of the input and output prefixes. Fix δ2 > 0, and use Lemmas 2.11–2.12 to construct a Rohlin tower with base F , height N, and µX (G) ≤ δ2 having the properties of the lemmas. In particular assume that the finite partition considered is P=
N−1 _
T −i P0 ,
i=0
where P0 is the zero-time partition for the finite-alphabet source, that is, if the stationary random process {Xn } has alphabet A = {ai ; i = 0, 1 . . . , kAk − 1} P0 = {{x : X0 (x) = x0 = ai }; i = 0, 1 . . . , kAk − 1}. Thus the atoms of P correspond to all sequences having initial N coordinates x N = aN , for some aN ∈ AN . In other words, the atoms are N-dimensional thin cylinders. The sliding-block encoder operates as follows to map a sequence x into a binary symbol. If x ∈ F , then use the asynchronous block code to map x N into a binary N-tuple nN and put out the first symbol b0 . This will be the first symbol in the sync sequence. If x ∈ T F , then T −1 x ∈ F . N Apply the block code to x−1 = (x−1 , x0 , . . . , xN−2 ) to obtain bN and put out the second symbol b1 . Continue in this way: if x ∈ T i F for i = N 0, 1, . . . , N − 1, then T −i x ∈ F and apply the block code to x−i to produce bN and put out the ith symbol bi . Lastly, if x ∈ G, put out a 0 (that is, an arbitrary symbol). This defines a stationary encoder for all infinite input
166
6 Distortion and Entropy
sequences x. Since F is measurable with respect to a finite window and block codes are used, the stationary code is described by a finite-length sliding-block code. The sliding-block decoder operates essentially in the same manner as the asynchronous block code decoder. Suppose that the sync sequence is a binary m-tuple s m . Given a received sequence b, look for an i = m 0, 1, . . . , N − 1 for which b−i = s m (there can be at most one). If there is no such i, put out an arbitrary fixed reference symbol, say b∗ . If there N is a match with a sync, form the binary N-tuple b−i and use the block ˆ N . Put out x ˆi . Recall decoder to map this into a reproduction vector x ˆ N will be a concatenation of k 0s followed by an n-tuple in Gn , the that x collection of entropy-typical source vectors. In a nutshell, the encoder uses the base of the Rohlin tower to initiate a block coding, and usually the block code will be used repetitively until eventually some spacing is thrown in to make things stationary. The resulting binary codewords all have a unique prefix that can occur only at the beginning of a code block. Consider the error probability resulting from this sliding-block code. ˆn denote the process resulting from encoding and decoding as Let X ˆ0 } be short hand for the set of sequences {x : above and let {X0 ≠ X ˆ X0 (x) ≠ X0 (x)}, where X0 (x) = x0 is just the coordinate function, and Xˆ0 (x) is the output at time 0 of the cascade of the sliding-block encoder and decoder. Then using total probability ˆ0 ) = µX ({X0 ≠ X ˆ0 }) Pe = Pr(X0 ≠ X ˆ0 } ∩ G) + = µX ({X0 ≠ X
N−1 X
ˆ0 } ∩ T i F ). µX ({X0 ≠ X
i=0
Since ˆ0 } ∩ G) ≤ µX (Gn ) ≤ δ2 µX ({X0 ≠ X and ˆ0 } ∩ T i F ) = µX ({x : X0 (x) ≠ X ˆ0 (x)} ∩ T i F ) µX ({X0 ≠ X −i ˆ0 (x)} ∩ F ) = µX (T {x : X0 (x) ≠ X ˆ0 (T i x)} ∩ F ) = µX ({x : X0 (T i x) ≠ X ˆi } ∩ F ) = µX ({Xi ≠ X using the stationarity of the process distribution and codes, we have that
6.5 Almost Lossless Source Coding
Pe ≤ δ2 +
N−1 X
167
ˆi } ∩ F ) µX ({Xi ≠ X
i=0
≤ δ2 +
k−1 X
ˆi } ∩ F ) + µX ({Xi ≠ X
i=0
≤ δ2 +
k−1 X
µX (F ) +
N−1 X
ˆn } ∩ F ) µX ({Xkn ≠ X k
i=k
k + N
≤ δ2 + +
ˆi } ∩ F ) µX ({Xi ≠ X
i=k
i=0
≤ δ2 +
N−1 X
N−1 X
µX ({Xkn 6∈ Gn } ∩ F )
i=k
N−1 X 1 + µX ({Xkn 6∈ Gn } ∩ F ). N i=k
Lemma 2.12 implies that for any x N , µX ({X N = x n } ∩ F ) ≤
1 µX ({X N = x n }) N
and hence µX ({Xkn 6∈ Gn } ∩ F ) =
X
µX ({X N = x n } ∩ F )
x N :xkn 6∈Gn
≤
X x N :xkn 6∈Gn
=
1 µX ({X N = x n }) N
1 1 µX ({Xkn 6∈ Gn }) = µX ({X n 6∈ Gn }) ≤ . N N N
Thus Pe ≤ δ2 + +
N −k 1 + ≤ δ2 + 2, N N
which can be made as small as desired by suitable choices of , δ2 . Summarizing, given any finite-alphabet stationary and ergodic source X with entropy rate 1 bit per sample and > 0, a sliding-block encoder with binary outputs and a sliding-block decoder mapping binary sequences into the source alphabet (both of finite length) can be constructed so that Pe ≤ . This implies that asymptotically optimal encoders fn and decoder gn can be constructed with resulting probability of error Pe (fn , gn ) going to 0 as n → ∞. This solves the theoretical problem of showing that a stationary and ergodic process with entropy rate 1 can be coded into bits and decoded in a stationary way so that the original source can be reconstructed without error. This can be viewed as a stationary coding analog of Shannon’s noiseless or lossless source coding theorem for variable-length block
168
6 Distortion and Entropy
codes (which are not stationary). The stationary coding leads to some additional properties, which are explored next.
6.6 Asymptotically Optimal Almost Lossless Codes Continuing with the almost lossless codes of the previous section, suppose that X is a stationary and ergodic finite-alphabet source with entropy rate 1 bit per symbol. Suppose that we construct a sequence of finite-length sliding block encoders fn and decoders gn which result in ˆ(n) . Since encoded processes U (n) and decoded reproduction processes X coding can only reduce entropy rate, we have immediately that ˆ(n) ). H(X) = 1 ≥ H(U (n) ) ≥ H(X
(6.15)
Since the codes are asymptotically optimal, (n)
ˆ0 ) = 0. lim Pr(X0 ≠ X
n→∞
(6.16)
The codes form a coupling between the input and output, and the probability of error is simply the expected Hamming distortion between the pair. This distortion is bound below by Ornstein’s d-bar distance between the processes, which means that ˆ(n) ) = 0. lim d(X, X
n→∞
(6.17)
Thus the output process converges in d-bar to the original source. Eq. (6.16) implies from the process version of the Fano inequality that ˆ(n) ) = H(X) = 1. lim H(X
(6.18)
lim H(U (n) ) = 1.
(6.19)
n→∞
From (6.15), this forces n→∞
This suggests that the binary encoded process is looking more and more like coin flips as n grows. Marton’s inequality as in Corollary 6.6 provides a rigorous proof of this fact. Suppose that Z is the fair coin flip process, an IID binary equiprobable process. Then from Corollary 6.6 s 1 (n) (H(Z) − H(U (n) )) d(U , Z) ≤ 2 which goes to 0 as n → ∞. Thus the encoded process converges to fair coin flips in d-bar if the codes are asymptotically optimal, and the output of the decoder converges to the original source in d-bar.
6.7 Modeling and Simulation
169
Together these facts provide an approximate variation on the Ornstein isomorphism and Sinai theorems of ergodic theory. The isomorphism theorem states that a B-process with entropy rate 1 can be coded into coin flips in an invertible manner. We have just seen that we can map the original process into a process that is very close to coin flips in d-bar, and then we can “invert” the mapping by another mapping which produces a process very close to the original in d-bar. Thus almost lossless source coding can be interpreted as an approximate version of the isomorphism theorem, and the isomorphism theorem can be interpreted as a limiting version of the lossless coding result. The source coding result holds generally for stationary and ergodic processes, but turning it into the vastly stronger isomorphism theorem requires B-processes. The Sinai theorem states that we can model a process with prescribed distributions and entropy rate 1 by a stationary coding of coin flips. We have shown that a stationary coding of something d-bar close to coin flips can be used to obtain something d-bar close to a prescribed process. This can be interpreted as a simulation result, generating a desired process from coin flips.
6.7 Modeling and Simulation Random processes and dynamical systems can be used to model real world phenomena in the sense that one can use statistical methods to fit a probabilistic description to observed measurements. Such models are based on an assumption that observed relative frequencies of measurements predict future behavior, the fundamental idea of the ergodic theorem and the notion of AMS processes. Often models can be of a specific form, such as Gaussian or Poisson, based on the observed or assumed physics describing the production of the measured quantities. It also often occurs that the form of a model is assumed simply for convenience, and its suitability for the signals in question may be controversial. In some situations one might wish to place constraints on the model, but not assume particular distributions. For example one might wish to allow only models of a particularly simple form with useful properties such as B-processes, processes formed by stationary coding or filtering of IID processes, or autoregressive processes, processes formed by linear filtering an IID process using a poles-only or autoregressive filter. We shall later see (in Lemma 14.13) that a source can be communicated with arbitrarily small distortion through a noisy channel if the source has entropy rate less than a quantity determined by the channel (called the channel capacity) and if the source is totally ergodic. This fundamental result of information theory implies that a particularly useful class of models is the class with an entropy constraint. Furthermore, totally
170
6 Distortion and Entropy
ergodic sources such as B-processes are amenable to reliable communication and hence a natural choice of model. This raises the issue of how good a model is as a fit to a “real” process, and process distortion measures provide a means of quantifying just how good a model approximates a target process. This situation can be idealized by a process distortion between a “true” distribution and the model distribution, but this assumes the existence of the former. Theorem 5.2 provides a practical means of estimating a process distortion by finding the best match between a sequence produced by the target process (which will produce a frequency-typical sequence with probability 1) and the model’s collection of frequency-typical sequences. A minimum distortion selection between a fixed sequence and a member of a collection of sequences resembles the encoder we will encounter in source coding in Chapter 12, so that the considerations of this section will shed insight on the source coding problem. With this introduction we formalize two optimization problems describing the fitting of a model from an interesting class of processes to a target process so as to minimize a process distortion between the target and the class. It is assumed that the target process distribution exists, but it should be kept in mind that in the case of stationary and ergodic processes, the process distortion can be estimated by finding the best match between an example target sequence and the collection of frequency-typical model sequences. Suppose that µX is the distribution of a stationary source {Xn }. For some class of random processes P define ρ(µX , P) = inf ρ(µX , µY ). µY ∈P
(6.20)
The previous discussion suggests the classes P(R) = {µY : H(Y ) ≤ R} PB (R) = {B-processes µY : H(Y ) ≤ R}
(6.21) (6.22)
where R ≥ 0. The first optimization problem originated in [65] and will be seen in Chapter 12 to provide a geometric interpretation of Shannon source coding. The second originated in [53], where it was dubbed the “simulation problem" because of its goal of generating a good model of a target process as a sliding-block coding of a simple finite alphabet IID process such as fair coin flips. The sliding-block simulating code was shown to provide a good decoder in a source coding system. Both of these results will be considered later. They are introduced here as a natural combination of distortion and entropy considerations. Since PB (R) ⊂ P(R), ρ(µX , PB (R)) ≥ ρ(µX , P(R)).
6.7 Modeling and Simulation
171
Of more practical interest than simply modeling or simulating a random process as a B-process is the possibility of using a stationary encoding of a particular IID process such as fair coin flips or dice rolls. For example, suppose that one is given an IID process Z with distribution µZ and entropy rate H(Z) and one wishes to simulate a random process X with distribution µX by applying a stationary code f to Z to produce a ˜ = f (Z) with entropy rate H(X) ˜ ≤ H(Z) which is as close as process X possible to X in rho-bar: ∆X|Z = inf ρ(µX , µf (Z) ).
(6.23)
∆X|Z ≥ ρ(µX , PB (H(Z)))
(6.24)
f
From the definitions,
since the optimization over B-processes formed by stationary codings of Z is a more constrained optimization. Suppose, however, that X is ˜ is a B-process approximately solving the itself a B-process and that X ˜ is a B-process with H(X) ˜ ≤ minimization of ρ(µX , PB (H(Z))) so that X H(Z) and ρ(µX , µX˜ ) ≤ for some small > 0. Then Sinai’s theorem ˜ (or rather a process equivalent to X) ˜ can be obtained as a implies that X stationary coding of Z and hence ∆X|Z = ρ(µX , PB (H(Z))). If R = H(Z), this means that ∆X|Z = ρ(µX , PB (R)) ≡ ∆X (R),
(6.25)
a quantity that depends only on R and not on the structure of Z other than its entropy rate!
Chapter 7
Relative Entropy
Abstract A variety of information measures have been introduced for finite alphabet random variables, vectors, and processes: entropy, mutual information, relative entropy, conditional entropy, and conditional mutual information. All of these can be expressed in terms of divergence and hence the generalization of these definitions and their properties to infinite alphabets will follow from a general definition of divergence. In this chapter the definition and properties of divergence in this general setting are developed, including the formulas for evaluating divergence as an expectation of information density and as a limit of divergences of finite codings. We also develop several inequalities for and asymptotic properties of divergence. These results provide the groundwork needed for generalizing the ergodic theorems of information theory from finite to standard alphabets. The general definitions of entropy and information measures originated in the pioneering work of Kolmogorov and his colleagues Gelfand, Yaglom, Dobrushin, and Pinsker
7.1 Divergence Given a probability space (Ω, B, P ) (not necessarily with finite alphabet) and another probability measure M on the same space, define the relative entropy or divergence of P with respect to M by D(P kM) = sup HP kM (Q) = sup D(Pf kMf ), Q
(7.1)
f
where the first supremum is over all finite measurable partitions Q of Ω and the second is over all finite alphabet measurements on Ω. The two forms have the same interpretation: the divergence is the supremum of the relative entropies or divergences obtainable by finite alphabet codings of the sample space. The partition form is perhaps more common R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_7, © Springer Science+Business Media, LLC 2011
173
174
7 Relative Entropy
when considering divergence per se, but the measurement or code form is usually more intuitive when considering entropy and information. This section is devoted to developing the basic properties of divergence, all of which will yield immediate corollaries for the measures of information. The first result is a generalization of the divergence inequality that is a trivial consequence of the definition and the finite alphabet special case. Lemma 7.1. The Divergence Inequality: For any two probability measures P and M D(P kM) ≥ 0 with equality if and only if P = M. Proof. Given any partition Q, Theorem 3.1 implies that X Q∈Q
P (Q) ln
P (Q) ≥0 M(Q)
with equality if and only if P (Q) = M(Q) for all atoms Q of the partition. Since D(P kQ) is the supremum over all such partitions, it is also nonnegative. It can be 0 only if P and M assign the same probabilities to all atoms in all partitions (the supremum is 0 only if the above sum is 0 for all partitions) and hence the divergence is 0 only if the measures are identical. 2 As in the finite alphabet case, Lemma 7.1 justifies interpreting divergence as a form of distance or dissimilarity between two probability measures. It is not a true distance or metric in the mathematical sense since it is not symmetric and it does not satisfy the triangle inequality. Since it is nonnegative and equals zero only if two measures are identical, the divergence is a distortion measure on probability distributions as considered in Chapter 5. This view often provides interpretations of the basic properties of divergence. We shall develop several relations between the divergence and other distance measures. The reader is referred to Csiszár [26] for a development of the distance-like properties of divergence. As the supremum definition of divergence in the general case permits an easy generalization of the divergence inequality, it also permits an easy generalization of the basic convexity property of Corollary 3.5. Lemma 7.2. The divergence D(P kM) is a convex function of the pair of probability measures (P , M). The lemma can also be proved using the integral representation of divergence as in Csiszár [25].
7.1 Divergence
175
The following two lemmas provide means for computing divergences and studying their behavior. The first result shows that the supremum can be confined to partitions with atoms in a generating field. This will provide a means for computing divergences by approximation or limits. The result is due to Dobrushin and is referred to as Dobrushin’s theorem. The second result shows that the divergence can be evaluated as the expectation of an entropy density defined as the logarithm of the Radon-Nikodym derivative of one measure relative to the other. This result is due to Gelfand, Yaglom, and Perez. The proofs largely follow the translator’s remarks in Chapter 2 of Pinsker [150] (which in turn follows Dobrushin [32]). Lemma 7.3. Suppose that (Ω, B) is a measurable space where B is generated by a field F , B = σ (F ). Then if P and M are two probability measures on this space, D(P kM) = sup HP kM (Q). Q⊂F
Proof. From the definition of divergence, the right-hand term above is clearly less than or equal to the divergence. If P is not absolutely continuous with respect to M, then we can find a set F such that M(F ) = 0 but P (F ) 6= 0 and hence the divergence is infinite. Approximating this event by a field element F0 by applying Theorem 1.1 simultaneously to M and G will yield a partition {F0 , F0c } for which the right hand side of the previous equation is arbitrarily large. Hence the lemma holds for this case. Henceforth assume that M P . Fix > 0 and suppose that a partition Q = {Q1 , · · · , QK } yields a relative entropy close to the divergence, that is, HP kM (Q) =
K X i=1
P (Qi ) ln
P (Qi ) ≥ D(P kM) − /2. M(Qi )
We will show that there is a partition, say Q0 with atoms in F which has almost the same relative entropy, which will prove the lemma. First observe that P (Q) ln[P (Q)/M(Q)] is a continuous function of P (Q) and M(Q) in the sense that given /(2K) there is a sufficiently small δ > 0 such that if |P (Q) − P (Q0 )| ≤ δ and |M(Q) − M(Q0 )| ≤ δ, then provided M(Q) 6= 0 P (Q) P (Q0 ) |P (Q) ln − P (Q0 ) ln |≤ . M(Q) M(Q0 ) 2K If we can find a partition Q0 with atoms in F such that |P (Qi0 ) − P (Qi )| ≤ δ, |M(Qi0 ) − M(Qi )| ≤ δ, i = 1, · · · , K, then
(7.2)
176
7 Relative Entropy
|HP kM (Q0 ) − HP kM (Q)| ≤
X
|P (Qi ) ln
i
≤K
P (Qi0 ) P (Qi ) − P (Qi0 ) ln | M(Qi ) M(Qi0 )
= 2K 2
and hence HP kM (Q0 ) ≥ D(P kM) − which will prove the lemma. To find the partition Q0 satisfying (7.2), let m be the mixture measure P /2 + M/2. As in the proof of Lemma 5.1, we can find a partition Q0 ⊂ F such that m(Qi ∆Qi0 ) ≤ K 2 γ for i = 1, 2, · · · , K, which implies that P (Qi ∆Qi0 ) ≤ 2K 2 γ and M(Qi ∆Qi0 ) ≤ 2K 2 γ; i = 1, 2, · · · , K. If we now choose γ so small that 2K 2 γ ≤ δ, then (7.2) and hence the lemma follow from the above and the fact that |P (F ) − P (G)| ≤ P (F ∆G).
(7.3)
2 Lemma 7.4. Given two probability measures P and M on a common measurable space (Ω, B), if P is not absolutely continuous with respect to M, then D(P kM) = ∞. If P M (e.g., if D(P kM) < ∞), then the Radon-Nikodym derivative f = dP /dM exists and Z Z D(P kM) = ln f (ω)dP (ω) = f (ω) ln f (ω)dM(ω). The quantity ln f (if it exists) is called the entropy density or relative entropy density of P with respect to M. Proof. The first statement was shown in the proof of the previous lemma. If P is not absolutely continuous with respect to M, then there is a set Q such that M(Q) = 0 and P (Q) > 0. The relative entropy for the partition Q = {Q, Qc } is then infinite, and hence so is the divergence. Assume that P M and let f = dP /dM. Suppose that Q is an event for which M(Q) > 0 and consider the conditional cumulative distribution function for the real random variable f given that ω ∈ Q: T M({f < u} Q) FQ (u) = ; u ∈ (−∞, ∞). M(Q) Observe that the expectation with respect to this distribution is
7.1 Divergence
177
Z∞ EM (f |Q) =
u dFQ (u) =
We also have that Z∞ u ln u dFQ (u) = 0
1 M(Q)
1 M(Q)
Z f (ω) dM(ω) = Q
P (Q) . M(Q)
Z f (ω) ln f (ω) dM(ω), Q
where the existence of the integral is ensured by the fact that u ln u ≥ −e−1 . S Applying Jensen’s inequality to the convex function u ln u yields the inequality Z Z 1 1 ln f (ω) dP (ω) = f (ω) ln f (ω) dM(ω) M(Q) Q M(Q) Q Z∞ u ln u dFQ (u) = 0 Z∞ Z∞ ≥[ u dFQ (u)] ln[ u dFQ (u)] 0
P (Q) P (Q) ln . = M(Q) M(Q) We therefore have that for any event Q with M(Q) > 0 that Z P (Q) . ln f (ω) dP (ω) ≥ P (Q) ln M(Q) Q
(7.4)
Let Q = {Qi } be a finite partition and we have Z XZ ln f (ω)dP (ω) = ln f (ω) dP (ω) i
Qi
X
≥
i:P (Qi )6=0
=
X
Z ln f (ω) dP (ω) Qi
P (Qi ) ln
i
P (Qi ) , M(Qi )
where the inequality follows from (7.4) since P (Qi ) 6= 0 implies that M(Qi ) 6= 0 since M P . This proves that Z D(P kM) ≤ ln f (ω) dP (ω). To obtain the converse inequality, let qn denote the asymptotically accurate quantizers of Section 1.5. From (1.23) Z Z ln f (ω) dP (ω) = lim qn (ln f (ω)) dP (ω). n→∞
178
7 Relative Entropy
For fixed n the quantizer qn induces a partition of Ω into 2n2n + 1 atoms Q. In particular, there are 2n2n − 1 “good” atoms such that for ω, ω0 inside the atoms we have that | ln f (ω) − ln f (ω0 )| ≤ 2−(n−1) . The remaining two atoms group ω for which ln f (ω) ≥ n or ln f (ω) < −n. Defining the shorthand P (ln f < −n) = P ({ω : ln f (ω) < −n}), we have then that X Q∈Q
P (Q) ln
P (Q) = M(Q)
P (ln f ≥ n) ln
X
P (Q) ln
good
Q
P (Q) + M(Q)
P (ln f < −n) P (ln f ≥ n) + P (ln f < −n) ln . M(ln f ≥ n) M(ln f < −n)
The rightmost two terms above are bounded below as P (ln f ≥ n) ln
P (ln f < −n) P (ln f ≥ n) + P (ln f < −n) ln M(ln f ≥ n) M(ln f < −n)
≥ P (ln f ≥ n) ln P (ln f ≥ n) + P (ln f < −n) ln P (ln f < −n). Since P (ln f ≥ n) and P (ln f < −n) → 0 as n → ∞ and since x ln x → 0 as x → 0, given we can choose n large enough to ensure that the above term is greater than −. This yields the lower bound X
P (Q) ln
Q∈Q
P (Q) ≥ M(Q)
X good
P (Q) ln Q
P (Q) − . M(Q)
¯ = sup ω∈Q lnf (ω) and h = infω∈Q ln f (ω) Fix a good atom Q and define h and note that by definition of the good atoms
¯ − h ≤ 2−(n−1) . h We now have that ¯≥ P (Q)h
Z ln f (ω) dP (ω) Q
and M(Q)eh ≤
Z f (ω)dM(ω) = P (Q). Q
Combining these yields P (Q) ln
P (Q) P (Q) = P (Q)h ≥ P (Q) ln M(Q) P (Q)e−h ¯ − 2−(n−1) ) ≥ P (Q)(h Z ≥ ln f (ω)dP (ω) − P (Q)2−(n−1) . Q
7.1 Divergence
179
Therefore X Q∈Q
P (Q) ln
P (Q) ≥ M(Q) ≥
X good X good Z
P (Q) ln Q
Z Q
P (Q) − M(Q)
ln f (ω) dP − 2−(n−1) − Q
ln f (ω) dP (ω) − 2−(n−1) − .
= ω:| ln f (ω)|≤n
Since this is true for arbitrarily large n and arbitrarily small , Z D(P kQ) ≥ ln f (ω)dP (ω),
2
completing the proof of the lemma.
It is worthwhile to point out two examples for the previous lemma. If P and M are discrete measures with corresponding PMF’s p and q, then the Radon-Nikodym derivative is simply dP /dM(ω) = p(ω)/m(ω) and the lemma gives the known formula for the discrete case. If P and M are both probability measures on Euclidean space Rn and if both measures are absolutely continuous with respect to Lebesgue measure, then there exists a density f called a probability density function or pdf such that Z P (F ) = f (x)dx, F
where dx means dm(x) with m Lebesgue measure. (Lebesgue measure assigns each set its volume.) Similarly, there is a pdf g for M. In this case, Z f (x) D(P kM) = dx. (7.5) f (x) ln g(x) Rn The following immediate corollary to the previous lemma provides a formula that is occasionally useful for computing divergences. Corollary 7.1. Given three probability distributions M Q P , then D(P kM) = D(P kQ) + EP (ln
dQ ). dM
Proof. From the chain rule for Radon-Nikodym derivatives (e.g., Lemma 5.7.3 of [55] or Lemma 6.6 of [58]) dP dQ dP = dM dQ dM
180
7 Relative Entropy
and taking expectations using the previous lemma yields the corollary.
2 The next result is a technical result that shows that given a mapping on a space, the divergence between the induced distributions can be computed from the restrictions of the original measures to the subσ -field induced by the mapping. As part of the result, the relation between the induced Radon-Nikodym derivative and the original derivative is made explicit. Recall that if P is a probability measure on a measurable space (Ω, B) and if F is a sub-σ -field of B, then the restriction PF of P to F is the probability measure on the measurable space (Ω, F ) defined by PF (G) = P (G), for all G ∈ F . In other words, we can use either the probability measures on the new space or the restrictions of the probability measures on the old space to compute the divergence. This motivates considering the properties of divergences of restrictions of measures, a useful generality in that it simplifies proofs. The following lemma can be viewed as a bookkeeping result relating the divergence and the RadonNikodym derivatives in the two spaces. Lemma 7.5. (a) Suppose that M, P are two probability measures on a space (Ω, B) and that X is a measurement mapping this space into (A, A). Let PX and MX denote the induced distributions (measures on (A, A)) and let Pσ (X) and Mσ (X) denote the restrictions of P and M to σ (X), the subσ -field of B generated by X. Then D(PX kMX ) = D(Pσ (X) kMσ (X) ). If the Radon-Nikodym derivative f = dPX /dMX exists (e.g., the above divergence is finite), then define the function f (X) : Ω → [0, ∞) by f (X)(ω) = f (X(ω)) =
dPX (X(ω)); dMX
then with probability 1 under both M and P f (X) =
dPσ (X) . dMσ (X)
(b) Suppose that P M. Then for any sub-σ -field F of B, we have that dPF dP = EM ( |F ). dMF dM Thus the Radon-Nikodym derivative for the restrictions is just the conditional expectation of the original Radon-Nikodym derivative. Proof. The proof is mostly algebra: D(Pσ (X) kMσ (X) ) is the supremum over all finite partitions Q with elements in σ (X) of the relative entropy
7.1 Divergence
181
HPσ (X) kMσ (X) (Q). Each element Q ∈ Q ⊂ σ (X) corresponds to a unique set Q0 ∈ A via Q = X −1 (Q0 ) and hence to each Q ⊂ σ (X) there is a corresponding partition Q0 ⊂ A. The corresponding relative entropies are equal, however, since HPX kMX (Q0 ) =
X
Pf (Q0 ) ln
Q0 ∈Q0
=
X
PX (Q0 ) MX (Q0 )
P (X −1 (Q0 )) ln
Q0 ∈Q0
=
X
PX (Q) ln
Q∈Q
P (X −1 (Q0 )) M(X −1 (Q0 ))
PX (Q) MX (Q)
= HPσ (X) kMσ (X) (Q). Taking the supremum over the partitions proves that the divergences are equal. If the derivative is f = dPX /dMX , then f (X) is measurable since it is a measurable function of a measurable function. In addition, it is measurable with respect to σ (X) since it depends on ω only through X(ω). For any F ∈ σ (X) there is a G ∈ A such that F = X −1 (G) and Z Z Z f (X)dMσ (X) = f (X)dM = f dMX F
F
G
from the change of variables formula (see, e.g., Lemma 4.4.7 of [55] or Lemma 5.12 of [58]). Thus Z f (X)dMσ (X) = PX (G) = Pσ (X) (X −1 (G)) = Pσ (X) (F ), F
which proves that f (X) is indeed the claimed derivative with probability 1 under M and hence also under P . The variation quoted in part (b) is proved by direct verification using iterated expectation. If G ∈ F , then using iterated expectation we have that Z Z dP dP |F ) dMF = EM (1G |F ) dMF . EM ( dM dM G Since the argument of the integrand is F -measurable (see, e.g., Lemma 5.3.1 of [55] or Lemma 6.3 of [58]), invoking iterated expectation (e.g., Corollary 5.9.3 of [55] or Corollary 6.5 of [58]) yields Z Z dP dP |F ) dMF = EM (1G |F ) dM EM ( dM dM G dP ) = P (G) = PF (G), = E(1G dM proving that the conditional expectation is the claimed derivative.
2
182
7 Relative Entropy
Part (b) of the lemma was pointed out to the author by Paul Algoet. Having argued above that restrictions of measures are useful when finding divergences of random variables, we provide a key trick for treating such restrictions. Lemma 7.6. Let M P be two measures on a space (Ω, B). Suppose that F is a sub-σ -field and that PF and MF are the restrictions of P and M to F Then there is a measure S such that M S P and dP dP /dM = , dS dPF /dMF dPF dS = , dM dMF and D(P kS) + D(PF kMF ) = D(P kM).
(7.6)
Proof. If M P , then clearly MF PF and hence the appropriate Radon-Nikodym derivatives exist. Define the set function S by Z Z dPF dP S(F ) = |F ) dM, dM = EM ( dM F dMF F using part (b) of the previous lemma. Thus M S and dS/dM = dPF /dMF . Observe that for F ∈ F , iterated expectation implies that S(F ) = EM (EM (1F
dP dP |F )) = EM (1F ) dM dM = P (F ) = PF (F ); F ∈ F
and hence in particular that S(Ω) is 1 so that dPF /dMF is integrable and S is indeed a probability measure on (Ω, B). (In addition, the restriction of S to F is just PF .) Define g=
dP /dM . dPF /dMF
This is well defined since with M probability 1, if the denominator is 0, then so is the numerator. Given F ∈ B the Radon-Nikodym theorem (e.g., Theorem 5.6.1 of [55] or Theorem 6.2 of [58]) implies that Z Z Z dP /dM dS dM = 1F gdS = 1F g dPF /dMF dM = P (F ), dM dPF /dMF F that is, P S and dP /dM dP = , dS dPF /dMF
7.1 Divergence
183
proving the first part of the lemma. The second part follows by direct verification: Z Z Z dPF dP /dM dP dP = ln dP + ln dP D(P kM) = ln dM dMF dPF /dMF Z Z dPF dP = ln dPF + ln dP dMF dS = D(PF kMF ) + D(P kS).
2 The two previous lemmas and the divergence inequality immediately yield the following result for M P . If M does not dominate P , then the result is trivial. Corollary 7.2. Given two measures M, P on a space (Ω, B) and a sub-σ field F of B, then D(P kM) ≥ D(PF kMF ). If f is a measurement on the given space, then D(P kM) ≥ D(Pf kMf ). The result is obvious for finite fields F or finite alphabet measurements f from the definition of divergence. The general result for arbitrary measurable functions could also have been proved by combining the corresponding finite alphabet result of Corollary 3.2 and an approximation technique. As above, however, we will occasionally get results comparing the divergences of measures and their restrictions by combining the trick of Lemma 7.6 with a result for a single divergence. The following corollary follows immediately from Lemma 7.3 since the union of a sequence of asymptotically generating sub-σ -fields is a generating field. Corollary 7.3. Suppose that M, P are probability measures on a measurable space (Ω, B) and that Fn is an asymptotically generating sequence of sub-σ -fields and let Pn and Mn denote the restrictions of P and M to Fn (e.g., Pn = PFn ). Then D(Pn kMn ) ↑ D(P kM). There are two useful special cases of the above corollary which follow immediately by specifying a particular sequence of increasing sub-σ fields. The following two corollaries give these results. Corollary 7.4. Let M, P be two probability measures on a measurable space (Ω, B). Suppose that f is an A-valued measurement on the space.
184
7 Relative Entropy
Assume that qn : A → An is a sequence of measurable mappings into finite sets An with the property that the sequence of fields Fn = F (qn (f )) −1 generated by the sets {qn (a); a ∈ An } asymptotically generate σ (f ). (For example, if the original space is standard let Fn be a basis and let qn map the points in the ith atom of Fn into i.) Then D(Pf kMf ) = lim D(Pqn (f ) kMqn (f ) ). n→∞
The corollary states that the divergence between two distributions of a random variable can be found as a limit of quantized versions of the random variable. Note that the limit could also be written as lim HPf kMf (qn ).
n→∞
In the next corollary we consider increasing sequences of random variables instead of increasing sequences of quantizers, that is, more random variables (which need not be finite alphabet) instead of ever finer quantizers. The corollary follows immediately from Corollary 7.3 and Lemma 7.5. Corollary 7.5. Suppose that M and P are measures on the sequence space corresponding to outcomes of a sequence of random variables X0 , X1 , · · · with alphabet A. Let Fn = σ (X0 , · · · , Xn−1 ), which asymptotically generates the σ -field σ (X0 , X1 , · · · ). Then lim D(PX n kMX n ) = D(P kM).
n→∞
We now develop two fundamental inequalities involving entropy densities and divergence. The first inequality is from Pinsker [150]. The second is an improvement of an inequality of Pinsker [150] by Csiszár [24] and Kullback [105]. The second inequality is more useful when the divergence is small. Coupling these inequalities with the trick of Lemma 7.6 provides a simple generalization of an inequality of [54] and will provide easy proofs of L1 convergence results for entropy and information densities. Recall from Section 5.9 that given two probability measures M, P on a common measurable space (Ω, B), the variation distance between them is defined by X var(P , M) ≡ sup |P (Q) − M(Q)|, Q
Q∈Q
where the supremum is over all finite measurable partitions. We will proceed by stating first the end goal — the two inequalities involving divergence — as a lemma, and then state a lemma giving the basic required properties of the variational distance. The lemmas will be proved in a different order.
7.1 Divergence
185
Lemma 7.7. Let P and M be two measures on a common probability space (Ω, B) with P M. Let f = dP /dM be the Radon-Nikodym derivative and let h = ln f be the entropy density. Then Z 2 D(P kM) ≤ |h|dP ≤ D(P kM) + , (7.7) e Z
q |h|dP ≤ D(P kM) + 2D(P kM).
(7.8)
Lemma 7.8. Recall from Lemma 5.4 that given two probability measures M, P on a common measurable space (Ω, B), var(P , M) = 2 sup |P (F ) − M(F )| = 2 tvar(P , M).
(7.9)
F ∈B
If S is a measure for which P S and M S (S = (P + M)/2, for example), then also Z dP dM var(P , M) = | − | dS (7.10) dS dS and the supremum in (5.35) is achieved by the set F = {ω :
dM dP (ω) > (ω)}. dS dS
Proof of Lemma 7.8: Suppose that a measure S dominating both P and M exists and define the set F = {ω :
dP dM (ω) > (ω)} dS dS
and observe that Z Z Z dP dM dM dM dP dP | − | dS = ( − ) dS − − ) dS ( dS dS dS dS F dS F c dS = P (F ) − M(F ) − (P (F c ) − M(F c )) = 2(P (F ) − M(F )). From the definition of F , however, Z Z dP dM dS ≥ dS = M(F ) P (F ) = F dS F dS so that P (F ) − M(F ) = |P (F ) − M(F )|. Thus we have that Z dM dP − | dS = 2|P (F ) − M(F )| ≤ 2 sup |P (G) − M(G)| = var(P , M). | dS dS G∈B
186
7 Relative Entropy
To prove the reverse inequality, assume that Q approximately yields the variational distance, that is, for > 0 we have X |P (Q) − M(Q)| ≥ var(P , M) − . Q∈Q
Then X
|P (Q) − M(Q)| =
Q∈Q
X
Z
Q∈Q
≤
Q
X Z Q∈Q
Z =
(
|
|
| Q
dM dP − ) dS| dS dS
dP dM − | dS dS dS
dP dM − | dS dS dS
which, since is arbitrary, proves that Z dM dP − | dS, var(P , M) ≤ | dS dS Combining this with the earlier inequality proves (7.10). We have already seen that this upper bound is actually achieved with the given choice of F , which completes the proof of the lemma. 2 Proof of Lemma 7.7: The magnitude entropy density can be written as |h(ω)| = h(ω) + 2h(ω)−
(7.11)
where a− = −min(a, 0). This inequality immediately gives the trivial lefthand inequality of (7.7). The right-hand inequality follows from the fact that Z Z h− dP = f [ln f ]− dM and the elementary inequality a ln a ≥ −1/e. The second inequality will follow from (7.11) if we can show that Z q 2 h− dP ≤ 2D(P kM). Let F denote the set {h ≤ 0} and we have from (7.4) that Z Z P (F ) 2 h− dP = −2 hdP ≤ −2P (F ) ln M(F ) F and hence using the inequality ln x ≤ x − 1 and Lemma 5.4 Z M(F ) 2 h− dP ≤ 2P (F ) ln ≤ 2(M(F ) − P (F )) P (F )
7.1 Divergence
187
q ≤ d(P , M) ≤ 2D(P kM),
2 completing the proof. Combining Lemmas 7.7 and 7.6 yields the following corollary, which generalizes Lemma 2 of [62]. Corollary 7.6. Let P and M be two measures on a space (Ω, B). Suppose that F is a sub-σ -field and that PF and MF are the restrictions of P and M to F . Assume that M P . Define the entropy densities h = ln dP /dM and h0 = ln dPF /dMF . Then Z 2 |h − h0 | dP ≤ D(P kM) − D(PF kMF ) + , (7.12) e and Z
|h − h0 | dP ≤ q D(P kM) − D(PF kMF ) + 2D(P kM) − 2D(PF kMF ).
(7.13)
Proof. Choose the measure S as in Lemma 7.6 and then apply Lemma 7.7 2 with S replacing M.
Variational Description of Divergence As in the discrete case, divergence has a variational characterization that is a fundamental property for its applications to large deviations theory [182] [31]. We again take a detour to state and prove the property without delving into its applications. Suppose now that P and M are two probability measures on a common probability space, say (Ω, B), such that M P and hence the density f =
dP dM
is well defined. Suppose that Φ is a real-valued random variable defined on the same space, which we explicitly require to be finite-valued (it cannot assume ∞ as a value) and to have finite cumulant generating function: EM (eΦ ) < ∞. Then we can define a probability measure M φ by
188
7 Relative Entropy
Z M Φ (F ) = F
eΦ dM EM (eΦ )
(7.14)
and observe immediately that by construction M M Φ and dM Φ eΦ = . dM EM (eΦ ) The measure M Φ is called a “tilted” distribution. Furthermore, by construction dM Φ /dM 6= 0 and hence we can write Z Z Z f f dM Φ f dM = P (F ) dQ = dM = φ Φ φ Φ F e /EM (e ) F e /EM (e ) dM F and hence P M Φ and dP f . = φ Φ dM e /EM (eΦ ) We are now ready to state and prove the principal result of this section, a variational characterization of divergence. Theorem 7.1. Suppose that M P . Then D(P kM) = sup EP Φ − ln(EM (eΦ )) ,
(7.15)
Φ
where the supremum is over all random variables Φ for which Φ is finitevalued and eΦ is M-integrable. Proof. First consider the random variable Φ defined by Φ = ln f and observe that Z Z EP Φ − ln(EM (eΦ )) = dP ln f − ln( dMf ) Z = D(P kM) − ln dP = D(P kM). This proves that the supremum over all Φ is no smaller than the divergence. To prove the other half observe that for any Φ, dP /dM H(P kM) − EP Φ − ln EM (eΦ ) = EP ln , dP /dM Φ where M Φ is the tilted distribution constructed above. Since M M Φ P , we have from the chain rule for Radon-Nikodym derivatives that dP = D(P kM Φ ) ≥ 0 H(P kM) − EP Φ − ln EM (eΦ ) = EP ln dM Φ
7.2 Conditional Relative Entropy
189
from the divergence inequality, which completes the proof. Note that equality holds and the supremum is achieved if and only if M Φ = P . 2
7.2 Conditional Relative Entropy Lemmas 7.5 and 7.6 combine with basic properties of conditional probability in standard spaces to provide an alternative form of Lemma 7.6 in terms of random variables that gives an interesting connection between the densities for combinations of random variables and those for individual random variables. The results are collected in Theorem 7.2. First, however, several definitions are required. Let X and Y be random variables with standard alphabets AX and AY and σ -fields BAX and BAY , respectively. Let PXY and MXY be two distributions on (AX × AY , BAX ×AY ) and assume that MXY PXY . Let MY and PY denote the induced marginal distributions, e.g., MY (F ) = MXY (AX × F ). Define the (nonnegative) densities (Radon-Nikodym derivatives): fXY =
dPXY dPY , fY = dMXY dMY
so that Z PXY (F ) = PY (F ) =
ZF F
fXY dMXY ; F ∈ BAX ×AY fY dMY ; F ∈ BAY .
Note that MXY PXY implies that MY PY and hence fY is well defined if fXY is. Define also the conditional density fXY (x,y) ; if f (y) > 0 Y fY (y) fX|Y (x|y) = 1; otherwise. Suppose now that the entropy density hY = ln fY exists and define the conditional entropy density or conditional relative entropy density by hX|Y = ln fX|Y . Again suppose that these densities exist, we (tentatively) define the conditional relative entropy
190
7 Relative Entropy
Z HP kM (X|Y ) = E ln fX|Y = dPXY (x, y) ln fX|Y (x|y) Z = dMXY (x, y)fXY (x, y) ln fX|Y (x|y). if the expectation exists. Note that unlike unconditional relative entropies, the above definition of conditional relative entropy requires the existence of densities. Although this is sufficient in many of the applications and is convenient for the moment, it is not sufficiently general to handle all the cases we will encounter. In particular, there will be situations where we wish to define a conditional relative entropy HP kM (X|Y ) even though it is not true that MXY PXY . Hence at the end of this section we will return to this question and provide a general definition that agrees with the current one when the appropriate densities exist and that shares those properties not requiring the existence of densities, e.g., the chain rule for relative entropy. An alternative approach to a general definition for conditional relative entropy can be found in Algoet [6]. The previous construction immediately yields the following lemma providing chain rules for densities and relative entropies. Lemma 7.9. fXY = fX|Y fY hXY = hX|Y + hY , and hence D(PXY kMXY ) = HP kM (X|Y ) + D(PY kMY ),
(7.16)
or, equivalently, HP kM (X, Y ) = HP kM (Y ) + HP kM (X|Y ),
(7.17)
a chain rule for relative entropy analogous to that for ordinary entropy. Thus if HP kM (Y ) < ∞ so that the indeterminate form ∞ − ∞ is avoided, then HP kM (X|Y ) = HP kM (X, Y ) − HP kM (Y ). Since the alphabets are standard, there is a regular version of the conditional probabilities of X given Y under the distribution MXY ; that is, for each y ∈ B there is a probability measure MX|Y (F |y); F ∈ BA for fixed F ∈ BAX MX|Y (F |y) is a measurable function of y and such that for all G ∈ BAY Z MXY (F × G) = E(1G (Y )MX|Y (F |Y )) = MX|Y (F |y)dMY (y). G
7.2 Conditional Relative Entropy
191
¯ ∈ BB to be Lemma 7.10. Given the previous definitions, define the set B the set of y for which Z fX|Y (x|y)dMX|Y (x|y) = 1. A
¯ by Define PX|Y for y ∈ B Z PX|Y (F |y) =
F
fX|Y (x|y)dMX|Y (x|y); F ∈ BA
and let PX|Y (.|y) be an arbitrary fixed probability measure on (A, BA ) for ¯ Then MY (B) ¯ = 1, PX|Y is a regular conditional probability for X all y 6∈ B. given Y under the distribution PXY , and PX|Y MX|Y ; MY − a.e., that is, MY ({y : PX|Y (·|y) MX|Y (·|y)}) = 1. Thus if PXY MXY , we can choose regular conditional probabilities under both distributions so that with probability one under MY the conditional probabilities under P are dominated by those under M and dPX|Y dPX|Y (·|y) (x|y) ≡ (x) = fX|Y (x|y); x ∈ A. dMX|Y dMX|Y (·|y) Proof. Define for each y ∈ B the set function Z fX|Y (x|y)dMX|Y (x|y); F ∈ BA . Gy (F ) = F
We shall show that Gy (F ), y ∈ B, F ∈ BA is a version of a regular conditional probability of X given Y under PXY . First observe using iterated expectation and the fact that conditional expectations are expectations with respect to conditional probability measures ([55], Section 5.9) that for any F ∈ BB Z Z [ fX|Y (x|y)dMX|Y (x|y)] dMY (y) F
A
fXY 1fY >0 ) = E(1F (Y )E[1A (X)fX|Y |Y ]) = E(1F (Y )1A (X) fY Z Z 1 1 = 1A×F 1{fY >0} fXY dMXY = 1{fY >0} dPXY fY f A×F Y Z Z 1 1 = 1{fY >0} dPY dPY , F fY F fY where the last step follows since since the function being integrated depends only on Y and hence is measurable with respect to σ (Y ) and
192
7 Relative Entropy
therefore its expectation can be computed from the restriction of PXY to σ (Y ) (see, for example, Lemma 5.3.1 of [55] or Lemma 6.3 of [58]) and since PY (fY > 0) = 1. We can compute this last expectation, however, using MY as Z Z Z 1 1 dPY = fY dMY = dMY = MY (F ) F fY F fY F which yields finally that Z Z fX|Y (x|y) dMX|Y (x|y) dMY (y) = MY (F ); all F ∈ BB . F
If
A
Z
Z F
g(y)dMY (y) =
F
1dMY (y), all F ∈ BB ,
however, it must also be true that g = 1 MY -a.e. (See, for example, Corollary 5.3.1 of [55] or Corollary 6.1 of [58].) Thus we have MY -a.e. and hence also PY -a.e. that Z fX|Y (x|y)dMX|Y (x|y)]dMY (y) = 1; A
¯ = 1. For y ∈ B, ¯ it follows from the basic properties of that is, MY (B) integration that Gy is a probability measure on (A, BA ) (see Corollary 4.4.3 of [55] or Corollary 5.4 of [58]). ¯ and hence this By construction, PX|Y (·|y) MX|Y (·|y) for all y ∈ B is true with probability 1 under MY and PY . Furthermore, by construction dPX|Y (·|y) (x) = fX|Y (x|y). dMX|Y (·|y) To complete the proof we need only show that PX|Y is indeed a version of the conditional probability of X given Y under PXY . To do this, fix G ∈ BA and observe for any F ∈ BB that Z Z Z PX|Y (G|y) dPY (y) = [ fX|Y (x|y)dMX|Y (x|y)] dPY (y) F ZF ZG = [ fX|Y (x|y) dMX|Y (x|y)]fY (y) dMY (y) F
G
= E[1F (Y )fY E[1G (X)fX|Y |Y ] = EM [1G×F fXY ], again using iterated expectation. This immediately yields Z Z Z PX|Y (G|y) dPY (y) = fXY dMXY = dPXY = PXY (G × F ), F
G×F
G×F
7.2 Conditional Relative Entropy
193
which proves that PX|Y (G|y) is a version of the conditional probability of X given Y under PXY , thereby completing the proof. 2 Theorem 7.2. Given the previous definitions with MXY PXY , define the distribution SXY by Z MX|Y (F |y)dPY (y), (7.18) SXY (F × G) = G
that is, SXY has PY as marginal distribution for Y and MX|Y as the conditional distribution of X given Y . Then the following statements are true: 1. MXY SXY PXY . 2. dSXY /dMXY = fY and dPXY /dSXY = fX|Y . 3. D(PXY kMXY ) = D(PY kMY ) + D(PXY kSXY ), and hence D(PXY kMXY ) exceeds D(PY kMY ) by an amount D(PXY kSXY ) = HP kM (X|Y ). Proof. To apply Lemma 7.6 define P = PXY , M = MXY , F = σ (Y ), P 0 = Pσ (Y ) , and M 0 = Mσ (Y ) . Define S by Z S(F × G) = F ×G
dPσ (Y ) dMXY , dMσ (Y )
for F ∈ BA and G ∈ BB . We begin by showing that S = SXY . All of the properties will then follow from Lemma 7.6. For F ∈ BAX and G ∈ BAY ! Z dPσ (Y ) dPσ (Y ) S(F × G) = dMXY = E 1F ×G , dMσ (Y ) F ×G dMσ (Y ) where the expectation is with respect to MXY . Using Lemma 7.5 and iterated conditional expectation (c.f. Corollary 5.9.3 of [55] or Corollary 6.5 of [58]) yields ! dPσ (Y ) dPY E 1F ×G (Y ) = E 1F (X)1G (Y ) dMσ (Y ) dMY dPY (Y )E[1F (X)|Y ] = E 1G (Y ) dMY dPY = E 1G (Y ) (Y )MX|Y (F |Y ) dMY Z Z dPY MX|Y (F |y) (Y )dMY (y) = MX|Y (F |y) dPY (y), dMY G proving that S = SXY . Thus Lemma 7.15 implies that MXY SXY PXY , proving the first property. From Lemma 7.5, dP 0 /dM 0 = dPσ (Y ) /dMσ (Y ) = dPY /dMY = fY , proving the first equality of property 2. This fact and the first property imply
194
7 Relative Entropy
the second equality of property 2 from the chain rule of Radon-Nikodym derivatives. (See, e.g., Lemma 5.7.3 of [55] or Lemma 6.6 of [58].) Alternatively, the second equality of the second property follows from Lemma 7.6 since dPXY dPXY /dMXY fXY = = . dSXY dMXY /dSXY fY Corollary 7.1 therefore implies that D(PXY kMXY ) = D(PXY kSXY ) + D(SXY kMXY ), which with Property 2, Lemma 7.4, and the definition of relative entropy rate imply Property 3. 2 It should be observed that it is not necessarily true that D(PXY kSXY ) ≥ D(PX kMX ) and hence that D(PXY kMXY ) ≥ D(PX kMX )+D(PY kMY ) as one might expect since in general SX 6= MX . These formulas will, however, be true in the special case where MXY = MX × MY . We next turn to an extension and elaboration of the theorem when there are three random variables instead of two. This will be a crucial generalization for our later considerations of processes, when the three random variables will be replaced by the current output, a finite number of previous outputs, and the infinite past. Suppose that MXY Z PXY Z are two distributions for three standard alphabet random variables X, Y , and Z taking values in measurable spaces (AX , BAX ), (AY , BAY ), (AZ , BAZ ), respectively. Observe that the absolute continuity implies absolute continuity for the restrictions, e.g., MXY PXY and MY PY . Define the Radon-Nikodym derivatives fXY Z , fY Z , fY , etc. in the obvious way; for example, fXY Z =
dPXY Z . dMXY Z
Let hXY Z , hY Z , hY , etc., denote the corresponding relative entropy densities, e.g., hXY Z = ln fXY Z . Define as previously the conditional densities fX|Y Z =
fXY Z fXY ; fX|Y = , fY Z fY
the conditional entropy densities hX|Y Z = ln fX|Y Z ; hX|Y = ln fX|Y , and the conditional relative entropies HP kM (X|Y ) = E(ln fX|Y ) and HP kM (X|Y , Z) = E(ln fX|Y Z ).
7.2 Conditional Relative Entropy
195
By construction (or by double use of Lemma 7.9) we have the following chain rules for conditional relative entropy and its densities. Lemma 7.11. fXY Z = fX|Y Z fY |Z fZ , hXY Z = hX|Y Z + hY |Z + hZ , and hence HP kM (X, Y , Z) = HP kM (X|Y Z) + HP kM (Y |Z) + HP kM (Z). Corollary 7.7. Given a distribution PXY , suppose that there is a product distribution MXY = MX × MY PXY . Then MXY dPXY d(PX × PY ) d(PX × PY ) dMXY D(PXY kPX × PY ) + HP kM (X)
PX × PY PXY , fX|Y fXY = , = fX fY fX = f X fY , = HP kM (X|Y ), and
D(PX × PY kMXY ) = HP kM (X) + HP kM (Y ). Proof. First apply Theorem 7.2 with MXY = MX × MY . Since MXY is a product measure, MX|Y = MX and MXY SXY = MX × PY PXY from the theorem. Next we again apply Theorem 7.2, but this time the roles of X and Y in the theorem are reversed and we replace MXY in the theorem statement by the current SXY = MX × PY and we replace SXY in the theorem statement by Z S 0 XY (F × G) = SY |X (G|x) dPX (x) = PX (F )PY (G); F
0 0 that is, SXY = PX × PY . We then conclude from the theorem that SXY = PX × PY PXY , proving the first statement. We now have that
MXY = MX × MY PX × PY PXY and hence the chain rule for Radon-Nikodym derivatives (e.g., Lemma 5.7.3 of [55] or Lemma 6.6 of [58]) implies that fXY =
dPXY dPXY d(PX × PY ) = . dMXY d(PX × PY ) d(MX × MY )
It is straightforward to verify directly that
196
7 Relative Entropy
d(PX × PY ) dPX dPY = = fX fY d(MX × MY ) dMX dMY and hence fXY =
dPXY )fX fY , d(PX × PY )
as claimed. Taking expectations using Lemma 7.4 then completes the proof (as in the proof of Corollary 7.1.) 2 The lemma provides an interpretation of the product measure PX × PY . This measure yields independent random variables with the same marginal distributions as PXY , which motivates calling PX × PY the independent approximation or memoryless approximation to PXY . The next corollary further enhances this name by showing that PX × PY is the best such approximation in the sense of yielding the minimum divergence with respect to the original distribution. Corollary 7.8. Given a distribution PXY let M denote the class of all product distributions for XY ; that is, if MXY ∈ M, then MXY = MX × MY . Then inf
MXY ∈M
D(PXY kMXY ) = D(PXY kPX × PY ).
Proof. We need only consider those M yielding finite divergence (since if there are none, both sides of the formula are infinite and the corollary is trivially true). Then D(PXY kMXY ) = D(PXY kPX × PY ) + D(PX × PY kMXY ) ≥ D(PXY kPX × PY ) with equality if and only if D(PX × PY kMXY ) = 0, which it will be if MXY = PX × PY . 2 Recall that given random variables (X, Y , Z) with distribution MXY Z , then X → Y → Z is a Markov chain (with respect to MXY Z ) if for any event F ∈ BAZ with probability one MZ|Y X (F |y, x) = MZ|Y (F |y). If this holds, we also say that X and Z are conditionally independent given Y . Equivalently, if we define the distribution MX×Z|Y by Z MX×Z|Y (FX × FZ × FY ) =
Fy
MX|Y (FX |y)MZ|Y (FZ |y)dMY (y); FX ∈ BAX ; FZ ∈ BAZ ; FY ∈ BAY ;
then Z → Y → X is a Markov chain if MX×Z|Y = MXY Z . (See Section 5.10 of [55] or Section 6.10 of [58].) This construction shows that a Markov chain is symmetric in the sense that X → Y → Z if and only if Z → Y → X.
7.2 Conditional Relative Entropy
197
Note that for any measure MXY Z , X → Y → Z is a Markov chain under MX×Z|Y by construction. The following corollary highlights special properties of the various densities and relative entropies when the dominating measure is a Markov chain. It will lead to the idea of a Markov approximation to an arbitrary distribution on triples extending the independent approximation of the previous corollary. Corollary 7.9. Given a probability space, suppose that MXY Z PXY Z are two distributions for a random vector (X, Y , Z) with the property that Z → Y → X forms a Markov chain under M. Then MXY Z PX×Z|Y PXY Z and fX|Y Z dPXY Z = dPX×Z|Y fX|Y dPX×Z|Y = fY Z fX|Y . dMXY Z
(7.19) (7.20)
Thus ln
dPXY Z + hX|Y = hX|Y Z dPX×Z|Y dPX×Z|Y ln = hY Z + hX|Y dMXY Z
and taking expectations yields D(PXY Z kPX×Z|Y ) + HP kM (X|Y ) = HP kM (X|Y Z) D(PX×Z|Y kMXY Z ) = D(PY Z kMY Z ) + HP kM (X|Y ). Furthermore, PX×Z|Y = PX|Y PY Z ,
(7.21)
that is, Z PX×Z|Y (FX × FZ × FY ) =
FY ×FZ
PX|Y (FX |y)dPZY (z, y).
(7.22)
Lastly, if Z → Y → X is a Markov chain under M, then it is also a Markov chain under P if and only if hX|Y = hX|Y Z
(7.23)
HP kM (X|Y ) = HP kM (X|Y Z).
(7.24)
in which case
198
7 Relative Entropy
Proof. Define g(x, y, z) =
fX|Y Z (x|y, z) fXY Z (x, y, z) fY (y) = fX|Y (x|y) fY Z (y, z) fXY (x, y)
and simplify notation by defining the measure Q = PX×Z|Y . Note that Z → Y → X is a Markov chain with respect to Q. To prove the first statement of the corollary requires proving the following relation: Z PXY Z (FX × FY × FZ ) =
gdQ; FX ×FY ×FZ
all FX ∈ BAX , FZ ∈ BAZ , FY ∈ BAY . From iterated expectation with respect to Q (e.g., Section 5.9 of [55] or Section 6.9 of [58]) E(g1FX (X)1FZ (Z)1FY (Y )) = E(1FY (Y )1FZ (Z)E(g1FX (X)|Y Z)) Z Z = 1FY (y)1FZ (z)( g(x, y, z) dQX|Y Z (x|y, z)) dQY Z (y, z). FX
Since QY Z = PY Z and QX|Y Z = PX|Y Q-a.e. by construction, the previous formula implies that Z Z Z g dQ = dPY Z gdPX|Y . FX ×FY ×FZ
FX
FY ×FZ
This proves (7.21. Since MXY Z PXY Z , we also have that MXY PXY and hence application of Theorem 7.2 yields Z Z Z gdQ = dPY Z gfX|Y dMX|Y FX ×FY ×FZ F ×F F Z Y Z Z X = dPY Z fX|Y Z dMX|Y . FY ×FZ
FX
By assumption, however, MX|Y = MX|Y Z a.e. and therefore Z Z Z g dQ = dPY Z fX|Y Z dMX|Y Z FX ×FY ×FZ F ×F F Z Y Z Z X = dPY Z dPX|Y Z FY ×FZ
FX
= PXY Z (FX × FY × FZ ), where the final step follows from iterated expectation. This proves (7.19) and that Q PXY Z . To prove (7.20) we proceed in a similar manner and replace g by fX|Y fZY and replace Q by MXY Z = MX×Y |Z . Also abbreviate PX×Y |Z to
7.2 Conditional Relative Entropy
199
Pˆ. As in the proof of (7.19) we have since Z → Y → X is a Markov chain under M that Z Z Z g dQ = dMY Z g dMX|Y FX ×FY ×FZ FY ×FZ FX ! Z Z = FY ×FZ
fZY dMY Z
FX
fX|Y dMX|Y !
Z
Z = FY ×FZ
dPY Z
FX
fX|Y dMX|Y
.
From Theorem 7.2 this is Z FY ×FZ
PX|Y (FX |y) dPY Z .
But PY Z = PˆY Z and PX|Y (FX |y) = PˆX|Y (FX |y) = PˆX|Y Z (FX |yz) since Pˆ yields a Markov chain. Thus the previous formula is Pˆ(FX × FY × FZ ), proving (7.20) and the corresponding absolute continuity. If Z → Y → X is a Markov chain under both M and P , then PX×Z|Y = PXY Z and hence fX|Y Z dPXY Z =1= , dPX×Z|Y fX|Y which implies (7.23). Conversely, if (7.23) holds, then fX|Y Z = fX|Y which with (7.19) implies that PXY Z = PX×Z|Y , proving that Z → Y → X is a Markov chain under P . 2 The previous corollary and one of the constructions used will prove important later and hence it is emphasized now with a definition and another corollary giving an interesting interpretation. Given a distribution PXY Z , define the distribution PX×Z|Y as the Markov approximation to PXY Z . Abbreviate PX×Z|Y to Pˆ. The definition has two motivations. First, the distribution Pˆ makes Z → Y → X a Markov chain which has the same initial distribution PˆZY = PZY and the same conditional distribution PˆX|Y = PX|Y , the only difference is that Pˆ yields a Markov chain, that is, PˆX|ZY = PˆX|Y . The second motivation is the following corollary which shows that of all Markov distributions, Pˆ is the closest to P in the sense of minimizing the divergence. Corollary 7.10. Given a distribution P = PXY Z , let M denote the class of all distributions for XY Z for which Z → Y → X is a Markov chain under MXY Z (MXY Z = MX×Z|Y ). Then inf
MXY Z ∈M
D(PXY Z kMXY Z ) = D(PXY Z kPX×Z|Y );
200
7 Relative Entropy
that is, the infimum is a minimum and it is achieved by the Markov approximation. Proof. If no MXY Z in the constraint set satisfies MXY Z PXY Z , then both sides of the above equation are infinite. Hence confine interest to the case MXY Z PXY Z . Similarly, if all such MXY Z yield an infinite divergence, we are done. Hence we also consider only MXY Z yielding finite divergence. Then the previous corollary implies that MXY Z PX×Z|Y PXY Z and hence D(PXY Z kMXY Z ) = D(PXY Z kPX×Z|Y ) + D(PX×Z|Y kMXY Z ) ≥ D(PXY Z kPX×Z|Y ) with equality if and only if D(PX×Z|Y kMXY Z ) = D(PY Z kMY Z ) + HP kM (X|Y ) = 0. But this will be zero if M is the Markov approximation to P since then MY Z = PY Z and MX|Y = PX|Y by construction. 2
Generalized Conditional Relative Entropy We now return to the issue of providing a general definition of conditional relative entropy, that is, one which does not require the existence of the densities or, equivalently, the absolute continuity of the underlying measures. We require, however, that the general definition reduce to that considered thus far when the densities exist so that all of the results of this section will remain valid when applicable. The general definition takes advantage of the basic construction of the early part of this section. Once again let MXY and PXY be two measures, where we no longer assume that MXY PXY . Define as in Theorem 7.2 the modified measure SXY by Z SXY (F × G) =
G
MX|Y (F |y)dPY (y);
(7.25)
that is, SXY has the same Y marginal as PXY and the same conditional distribution of X given Y as MXY . We now replace the previous definition by the following: The conditional relative entropy is defined by HP kM (X|Y ) = D(PXY kSXY ).
(7.26)
If MXY PXY as before, then from Theorem 7.2 this is the same quantity as the original definition and there is no change. The divergence of (7.26), however, is well-defined even if it is not true that MXY PXY and hence the densities used in the original definition do not work. The key
7.2 Conditional Relative Entropy
201
question is whether or not the chain rule HP kM (Y ) + HP kM (X|Y ) = HP kM (XY )
(7.27)
remains valid in the more general setting. It has already been proven in the case that MXY PXY , hence suppose this does not hold. In this case, if it is also true that MY PY does not hold, then both the marginal and joint relative entropies will be infinite and (7.27) again must hold since the conditional relative entropy is nonnegative. Thus we need only show that the formula holds for the case where MY PY but it is not true that MXY PXY . By assumption there must be an event F for which Z MXY (F ) = MX|Y (Fy ) dMY (y) = 0 but
Z PXY (F ) =
PX|Y (Fy ) dPY (y) 6= 0,
where Fy = {(x, y) : (x, y) ∈ F } is the section of F at Fy . Thus MX|Y (Fy ) = 0 MY -a.e. and hence also PY -a.e. since MY PY . Thus Z SXY (F ) = MX|Y (Fy ) dPY (y) = 0 and hence it is not true that SXY PXY and therefore D(PXY kSXY ) = ∞, which proves that the chain rule holds in the general case. It can happen that PXY is not absolutely continuous with respect to MXY , and yet D(PXY kSXY ) < ∞ and hence PXY SXY and hence Z dPXY , HP kM (X|Y ) = dPXY ln dSXY in which case it makes sense to define the conditional density fX|Y ≡
dPXY dSXY
so that exactly as in the original tentative definition in terms of densities (7.16) we have that Z HP kM (X|Y ) = dPXY ln fX|Y . Note that this allows us to define a meaningful conditional density even though the joint density fXY does not exist! If the joint density does ex-
202
7 Relative Entropy
ist, then the conditional density reduces to the previous definition from Theorem 7.2. We summarize the generalization in the following theorem. Theorem 7.3. The conditional relative entropy defined by (7.26) and (7.25) agrees with the definition (7.16) in terms of densities and satisfies the chain rule (7.27). If the conditional relative entropy is finite, then Z HP kM (X|Y ) = dPXY ln fX|Y , where the conditional density is defined by fX|Y ≡
dPXY . dSXY
If MXY PXY , then this reduces to the usual definition fX|Y =
fXY . fY
The generalizations can be extended to three or more random variables in the obvious manner.
7.3 Limiting Entropy Densities We now combine several of the results of the previous section to obtain results characterizing the limits of certain relative entropy densities. Lemma 7.12. Given a probability space (Ω, B) and an asymptotically generating sequence of sub-σ -fields Fn and two measures M P , let Pn = PFn , Mn = MFn and let hn = ln dPn /dMn and h = ln dP /dM denote the entropy densities. If D(P kM) < ∞, then Z lim |hn − h| dP = 0, n→∞
that is, hn → h in L1 . Thus the entropy densities hn are uniformly integrable. Proof. Follows from the Corollaries 7.3 and 7.6. The following lemma is Lemma 1 of Algoet and Cover [7].
2
Lemma 7.13. Given a sequence of nonnegative random variables {fn } defined on a probability space (Ω, B, P ), suppose that E(fn ) ≤ 1; all n.
7.3 Limiting Entropy Densities
203
Then lim sup n→∞
1 ln fn ≤ 0. n
Proof. Given any > 0 the Markov inequality and the given assumption imply that E(fn ) P (fn > en ) ≤ ≤ e−n . en We therefore have that P(
1 ln fn ≥ ) ≤ e−n n
and therefore ∞ X n=1
P(
∞ X 1 1 e−n = −1 < ∞, ln fn ≥ ) ≤ n e n=1
Thus from the Borel-Cantelli lemma (Lemma 4.6.3 of [55] or Lemma 5.17 of [58]), P (n−1 hn ≥ i.o.) = 0. Since is arbitrary, the lemma is proved.
2
The lemma easily gives the first half of the following result, which is also due to Algoet and Cover [7], but the proof is different here and does not use martingale theory. The result is the generalization of Lemma 3.19. Theorem 7.4. Given a probability space (Ω, B) and an asymptotically generating sequence of sub-σ -fields Fn , let M and P be two probability measures with their restrictions Mn = MFn and Pn = PFn . Suppose that Mn Pn for all n and define fn = dPn /dMn and hn = ln fn . Then lim sup n→∞
and lim inf n→∞
1 hn ≤ 0, M − a.e. n 1 hn ≥ 0, P − a.e.. n
If it is also true that M P (e.g., D(P kM) < ∞), then 1 hn = 0, P − a.e.. n→∞ n lim
Proof. Since EM fn = EMn fn = 1, the first statement follows from the previous lemma. To prove the second statement consider the probability
204
P (−
7 Relative Entropy
dPn 1 1 ln > ) = Pn (− ln fn > ) = Pn (fn < e−n ) n Mn n Z Z dPn = fn dMn = fn <e−n fn <e−n Z < e−n dMn = e−n Mn (fn < e−n ) ≤ e−n . fn <e−n
Thus it has been shown that P(
1 hn < −) ≤ e−n n
and hence again applying the Borel-Cantelli lemma we have that P (n−1 hn ≤ − i.o.) = 0 which proves the second claim of the theorem. If M P , then the first result also holds P -a.e., which with the second result proves the final claim. 2 Barron [8] provides an additional property of the sequence hn /n. If M P , then the sequence hn /n is dominated by an integrable function.
7.4 Information for General Alphabets We can now use the divergence results of the previous sections to generalize the definitions of information and to develop their basic properties. We assume now that all random variables and processes are defined on a common underlying probability space (Ω, B, P ). As we have seen how all of the various information quantities–entropy, mutual information, conditional mutual information–can be expressed in terms of divergence in the finite case, we immediately have definitions for the general case. Given two random variables X and Y , define the average mutual information between them by I(X; Y ) = D(PXY kPX × PY ),
(7.28)
where PXY is the joint distribution of the random variables X and Y and PX × PY is the product distribution. Define the entropy of a single random variable X by H(X) = I(X; X). From the definition of divergence this implies that
(7.29)
7.4 Information for General Alphabets
205
I(X; Y ) = sup HPXY kPX ×PY (Q). Q
From Dobrushin’s theorem (Lemma 7.3), the supremum can be taken over partitions whose elements are contained in generating field. Letting the generating field be the field of all rectangles of the form F × G, F ∈ BAX and G ∈ BAY , we have the following lemma which is often used as a definition for mutual information. Lemma 7.14. I(X; Y ) = sup I(q(X); r (Y )), q,r
where the supremum is over all quantizers q and r of AX and AY . Hence there exist sequences of increasingly fine quantizers qn : AX → An and rn : AY → Bn such that I(X; Y ) = lim I(qn (X); rn (Y )). n→∞
Applying this result to entropy we have that H(X) = sup H(q(X)), q
where the supremum is over all quantizers. By “increasingly fine” quantizers is meant that the corresponding par−1 titions Qn = {qn (a); a ∈ An } are successive refinements, e.g., atoms in Qn are unions of atoms in Qn+1 . (If this were not so, a new quantizer could be defined for which it was true.) There is an important drawback to the lemma (which will shortly be removed in Lemma 7.18 for the special case where the alphabets are standard): the quantizers which approach the suprema may depend on the underlying measure PXY . In particular, a sequence of quantizers which work for one measure need not work for another. An immediate corollary of Lemma 7.14 extends an inequality known for the finite case to general alphabets. It is useful when one of the random variables has finite entropy. Corollary 7.11. I(X; Y ) ≤ H(Y ). Proof. Let quantizers q and r be quantizers of X and Y as previously. The finite alphabet result of Lemma 3.11 implies that I(q(X); r (Y )) ≤ H(r (Y )). Taking the supremum over the quantizers yields the corollary.
2 Given a third random variable Z, let AX , AY , and AZ denote the alphabets of X, Y , and Z and define the conditional average mutual information
206
7 Relative Entropy
I(X; Y |Z) = D(PXY Z kPX×Y |Z ).
(7.30)
This is the extension of the discrete alphabet definition of (3.27) and it makes sense only if the distribution PX×Y |Z exists, which is the case if the alphabets are standard but may not be the case otherwise. We shall later provide an alternative definition due to Wyner [197] that is valid more generally and equal to the above when the spaces are standard. Note that I(X; Y |Z) can be interpreted using Corollary 7.10 as the divergence between PXY Z and its Markov approximation. Combining these definitions with Lemma 7.1 yields the following generalizations of the discrete alphabet results. Lemma 7.15. Given two random variables X and Y , then I(X; Y ) ≥ 0 with equality if and only if X and Y are independent. Given three random variables X, Y , and Z, then I(X; Y |Z) ≥ 0 with equality if and only if Y → Z → X form a Markov chain. Proof. The first statement follow from Lemma 7.1 since X and Y are independent if and only if PXY = PX × PY . The second statement follows from (7.30) and the fact that Y → Z → X is a Markov chain if and only if PXY Z = PX×Y |Z (see, e.g., Corollary 5.10.1 of [55] or Corollary 6.7 of [58]).
2
The properties of divergence provide means of computing and approximating these information measures. From Lemma 7.4, if I(X; Y ) is finite, then Z dPXY I(X; Y ) = ln dPXY (7.31) d(PX × PY ) and if I(X; Y |Z) is finite, then Z I(X; Y |Z) =
ln
dPXY Z dPXY Z . dPX×Y |Z
(7.32)
For example, if X, Y are two random variables whose distribution is absolutely continuous with respect to Lebesgue measure dxdy and hence which have a pdf fXY (x, y) = dPXY (xy)/dxdy, then Z fXY (x, y) I(X; Y ) = dxdyfXY (xy) ln , fX (x)fY (y) where fX and fY are the marginal pdf’s, e.g.,
7.4 Information for General Alphabets
207
Z fX (x) =
fXY (x, y) dy =
dPX (x) . dx
In the cases where these densities exist, we define the information densities dPXY iX;Y = ln d(PX × PY ) (7.33) iX;Y |Z = ln
dPXY Z . dPX×Y |Z
The results of Section 7.2 can be used to provide conditions under which the various information densities exist and to relate them to each other. Corollaries 7.7 and 7.8 combined with the definition of mutual information immediately yield the following two results. Lemma 7.16. Let X and Y be standard alphabet random variables with distribution PXY . Suppose that there exists a product distribution MXY = MX × MY such that MXY PXY . Then MXY PX × PY PXY , iX;Y = ln(fXY /fX fY ) = ln(fX|Y /fX ) and I(X; Y ) + HP kM (X) = HP kM (X|Y ).
(7.34)
Comment: This generalizes the fact that I(X; Y ) = H(X) − H(X|Y ) for the finite alphabet case. The sign reversal results from the difference in definitions of relative entropy and entropy. Note that this implies that unlike ordinary entropy, relative entropy is increased by conditioning, at least when the reference measure is a product measure. The previous lemma provides an apparently more general test for the existence of a mutual information density than the requirement that PX × PY PXY , it states that if PXY is dominated by any product measure, then it is also dominated by the product of its own marginals and hence the densities exist. The generality is only apparent, however, as the given condition implies from Corollary 7.7 that the distribution is dominated by its independent approximation. Restating Corollary 7.7 in terms of mutual information yields the following. Corollary 7.12. Given a distribution PXY let M denote the collection of all product distributions MXY = MX × MY . Then I(X; Y ) =
inf
MXY ∈M
HP kM (X|Y ) =
inf
MXY ∈M
D(PXY kMXY ).
208
7 Relative Entropy
The next result is an extension of Lemma 7.16 to conditional information densities and relative entropy densities when three random variables are considered. It follows immediately from Corollary 7.9 and the definition of conditional information density. Lemma 7.17. (The chain rule for relative entropy densities) Suppose that MXY Z PXY Z are two distributions for three standard alphabet random variables and that Z → Y → X is a Markov chain under MXY Z . Let fX|Y Z , fX|Y , hX|Y Z , and hX|Y be as in Section 7.2. Then PX×Z|Y PXY Z , hX|Y Z = iX;Z|Y + hX|Y
(7.35)
HP kM (X|Y , Z) = I(X; Z|Y ) + HP kM (X|Y ).
(7.36)
and Thus, for example, HP kM (X|Y , Z) ≥ HP kM (X|Y ). As with Corollary 7.12, the lemma implies a variational description of conditional mutual information. The result is just a restatement of Corollary 7.10. Corollary 7.13. Given a distribution PXY Z let M denote the class of all distributions for XY Z under which Z → Y → X is a Markov chain, then I(X; Z|Y ) =
inf
MXY Z ∈M
HP kM (X|Y , Z) =
inf
MXY Z ∈M
D(PXY Z kMXY Z ),
and the minimum is achieved by MXY Z = PX×Z|Y . The following corollary relates the information densities of the various information measures and extends Kolmogorov’s equality to standard alphabets. Corollary 7.14. (The chain rule for information densities and Kolmogorov’s formula.) Suppose that X,Y , and Z are random variables with standard alphabets and distribution PXY Z . Suppose also that there exists a distribution MXY Z = MX × MY Z such that MXY Z PXY Z . (This is true, for example, if PX × PY Z PXY Z .) Then the information densities iX;Z|Y , iX;Y , and iX;(Y Z) exist and are related by iX;Z|Y + iX;Y = iX;(Y ,Z)
(7.37)
I(X; Z|Y ) + I(X; Y ) = I(X; (Y , Z)).
(7.38)
and
7.4 Information for General Alphabets
209
Proof. If MXY Z = MX × MY Z , then Z → Y → X is trivially a Markov chain since MX|Y Z = MX|Y = MX . Thus the previous lemma can be applied to this MXY Z to conclude that PX×Z|Y PXY Z and that (7.35) holds. We also have that MXY = MX × MY PXY . Thus all of the densities exist. Applying Lemma 7.16 to the product measures MXY = MX × MY and MX(Y Z) = MX × MY Z in (7.35) yields iX;Z|Y = hX|Y Z − hX|Y = ln fX|Y Z − ln fX|Y = ln
fX|Y Z fX|Y − ln = iX;Y Z − iX;Y . fX fX
Taking expectations completes the proof.
2
The previous corollary implies that if PX × PY Z PXY Z , then also PX×Z|Y PXY Z and PX × PY PXY and hence that the existence of iX;(Y ,Z) implies that of iX;Z|Y and iX;Y . The following result provides a converse to this fact: the existence of the latter two densities implies that of the first. The result is due to Dobrushin [32]. (See also Theorem 3.6.1 of Pinsker [150] and the translator’s comments.) Corollary 7.15. If PX×Z|Y PXY Z and PX × PY PXY , then also PX × PY Z PXY Z and dPXY Z dPXY = . d(PX × PY Z ) d(PX × PY ) Thus the conclusions of Corollary 7.14 hold. Proof. The key to the proof is the demonstration that dPX×Z|Y dPXY = , d(PX × PY ) d(PX × PY Z )
(7.39)
which implies that PX × PY Z PX×Z|Y . Since it is assumed that PX×Z|Y PXY Z , the result then follows from the chain rule for Radon-Nikodym derivatives. Eq. (7.39) will be proved if it is shown that for all FX ∈ BAX , FY ∈ BAY , and FZ ∈ BAZ , Z dPXY PX×Z|Y (FX × FZ × FY ) = (7.40) d(PX × PY Z ). FX ×FZ ×FY d(PX × PY ) The thrust of the proof is the demonstration that for any measurable nonnegative function f (x, z) Z Z f (x, y) d(PX ×PY Z )(x, y, z) = f (x, y)PZ|Y (FZ |y)d(PX ×PY )(x, y). z∈FZ
The lemma will then follow by substituting
(7.41)
210
7 Relative Entropy
f (x, y) =
dPXY (x, y)1FX (x)1FY (y) d(PX × PY )
into (7.41) to obtain (7.40). To prove (7.41) first consider indicator functions of rectangles: f (x,y)= 1FX (x)1FY (y). Then both sides of (7.41) equal PX (FX )PY Z (FY × FY ) from the definitions of conditional probability and product measures. In particular, from Lemma 5.10.1 of [55] or Corollary 6.7 of [58] the left-hand side is Z Z Z 1FX (x)1FY (y) d(PX × PY Z )(x, y, z) = ( 1FX dPX )( 1FY ×FZ dPY Z ) z∈FZ
= PX (F )PY Z (FY × FZ ) and the right-hand side is Z 1FX (x)1FY (y)PZ|Y (FZ |y) d(PX × PY )(x, y) = Z Z ( 1FX (x) dPX (x))( 1FY (y)PZ|Y (FZ |y) dPY (y)) = PX (F )PY Z (FY × FZ ), as claimed. This implies (7.41) holds also for simple functions and hence also for positive functions by the usual approximation arguments. 2 Note that Kolmogorov’s formula (7.36) gives a formula for computing conditional mutual information as I(X; Z|Y ) = I(X; (Y , Z)) − I(X; Y ). The formula is only useful if it is not indeterminate, that is, not of the form ∞−∞. This will be the case if I(Y ; Z) (the smaller of the two mutual informations) is finite. Corollary 7.5 provides a means of approximating mutual information by that of finite alphabet random variables. Assume now that the random variables X, Y have standard alphabets. For, say, random variable X with alphabet AX there must then be an asymptotically generating sequence of finite fields FX (n) with atoms AX (n), that is, all of the members of FX (n) can be written S as unions of disjoint sets in AX (n) and FX (n) ↑ BAX ; that is, BAX = σ ( n FX (n)). The atoms AX (n) form a partition of the alphabet of X. Consider the divergence result of Corollary 7.5. with P = PXY , M = (n) (n) PX × PY and quantizer q(n) (x, y) = (qX (x), qY (y)). Consider the limit n → ∞. Since FX (n) asymptotically generates BAX and FY (n) asymptotically generates BAY and since the pair σ -field BAX ×AY is generated by rectangles, the field generated by all sets of the form FX × FY with FX ∈ FX (n), some n, and FY ∈ FY (m), some m, generates BAX ×AY . Hence Corollary 7.5 yields the first result of the following lemma. The
7.4 Information for General Alphabets
211
second is a special case of the first. The result shows that the increasingly fine quantizers of Lemma 7.14 can be chosen in a manner not depending on the underlying measure if the alphabets are standard. Lemma 7.18. Suppose that X and Y are random variables with standard (n) alphabets defined on a common probability space. Suppose that qX , n = 1, 2, · · · is a sequence of quantizers for AX such that the corresponding partitions asymptotically generate BAX . Define quantizers for Y similarly. Then for any distribution PXY (n)
(n)
I(X; Y ) = lim I(qX (X); qY (Y )) n→∞
and
(n)
H(X) = lim H(qX (X)); n→∞
that is, the same quantizer sequence works for all distributions. The following lemma generalizes Lemma 3.14 to standard alphabets. The concavity with respect to the source follows in a manner similar to the entropy result of Lemma 7.18 by combining the finite-alphabet result of Lemma 3.14 with limiting quantization. The convexity with respect to the channel does not readily follow in the same way because a channel can not be quantized without using an input distribution to form a joint distribution. The proof instead mimics the corresponding proof of the finite case based on the convexity of divergence of Lemma 7.2, which in turn follows from the finite alphabet result. Lemma 7.19. Let X and Y be random variables with standard alphabets AX and AY . Let µ denote a distribution on (AX , BAX ), and let ν be a regular conditional distribution ν(F |x) = Pr(Y ∈ F |X = x), x ∈ AX , F ∈ BAY . Let p = µν denote the resulting joint distribution. LetSIµν = Iµν (X; Y ) be the average T mutual information. Then Iµν is a convex function of ν and a convex function of µ. Proof. The proof of convexity was suggested by T. Linder. Consider a fixed source µ and consider channels ν1 , ν2 , and ν = λν1 + (1 − λ)ν2 . Denote the corresponding input/output pair processes by pi = µνi , i = 1, 2, and p = λp1 + (1 − λ)p2 and the corresponding output processes by ηi and η = λη1 + (1 − λ)η2 , e.g., η(G) = p(A∞ X × G) for all output events G. Note that p1 , p2 , and p all have a common input distribution µ. We have that µ × η = λµ × η1 + (1 − λ)µ × η2 so that from Lemma 7.2 Iµν = D(µν||µ × η) = D(λp1 + (1 − λ)p2 kλµ × η1 + (1 − λ)µ × η2 ) ≤ λD(p1 kµ × η1 ) + (1 − λ)D(p2 kµ × η2 ) = λIµν1 + (1 − λ)Iµν2 ,
212
7 Relative Entropy
proving the convexity of mutual information with respect to the channel in direct imitation of the proof for the finite case. The concavity with respect to the source distribution follows from the proof of the corresponding finite alphabet result, specifically the representation of (3.26), coupled with a sequence of asymptotically accurate quantizers for the input and output. As the quantization becomes asymptotically accurate, all of the terms in (3.26) converge upward to their limiting values, proving that (3.26) holds for the general distributions.
2 Next consider the mutual information I(f (X), g(Y )) for arbitrary measurable mappings f and g of X and Y . From Lemma 7.15 applied to the random variables f (X) and g(Y ), this mutual information can be approximated arbitrarily closely by I(q1 (f (X)); q2 (g(Y ))) by an appropriate choice of quantizers q1 and q2 . Since the composition of q1 and f constitutes a finite quantization of X and similarly q2 g is a quantizer for Y , we must have that I(f (X); g(Y )) ≈ I(q1 (f (X)); q2 (g(Y )) ≤ I(X; Y ). Making this precise yields the following corollary. Corollary 7.16. If f is a measurable function of X and g is a measurable function of Y , then I(f (X), g(Y )) ≤ I(X; Y ). The corollary states that mutual information is reduced by any measurable mapping, whether finite or not. For practice we point out another proof of this basic result that directly applies a property of divergence. Let P = PXY , M = PX × PY , and define the mapping r (x, y) = (f (x), g(y)). Then from Corollary 7.2 we have I(X; Y ) = D(P kM) ≥ D(Pr kMr ) ≥ D(Pf (X),g(Y ) kMf (X),g(Y ) ). But Mf (X),g(Y ) = Pf (X) × Pg(Y ) since Mf (X),g(Y ) (FX × FZ ) = M(f −1 (FX )
\
g −1 (FY )
= PX (f −1 (FX )) × PY (g −1 (FY )) = Pf (X) (FX ) × Pg(Y ) (FY ). Thus the previous inequality yields the corollary.
2
For the remainder of this section we focus on conditional entropy and information. Although we cannot express mutual information as a difference of ordinary entropies in the general case (since the entropies of nondiscrete random variables are generally infinite), we can obtain such a
7.4 Information for General Alphabets
213
representation in the case where one of the two variables is discrete. Suppose we are given a joint distribution PXY and that X is discrete. We can choose a version of the conditional probability given Y so that pX|Y (x|y) = P (X = x|Y = y) is a valid PMF (considered as a function of x for fixed y) with PY probability 1. (This follows from Corollary 5.8.1 of [55] since the alphabet of X is discrete; the alphabet of Y need not be even standard.) Define H(X|Y = y) =
X
pX|Y (x|y) ln
x
and
1 pX|Y (x|y)
Z H(X|Y ) =
H(X|Y = y) dPY (y).
Note that this agrees with the formula of Section 3.6 in the case that both alphabets are finite. The following result is due to Wyner [197]. Lemma 7.20. If X, Y are random variables and X has a finite alphabet, then I(X; Y ) = H(X) − H(X|Y ). Proof. We first claim that pX|Y (x|y)/pX (x) is a version of dPXY /d(PX × PY ). To see this observe that for F ∈ B(AX × AY ), letting Fy denote the section {x : (x, y) ∈ F } we have that Z X Z pX|Y (x|y) pX|Y (x|y) d(PX × PY ) = pX (x)dPY (y) pX (x) pX (x) F x∈Fy Z X = dPY (y) pX|Y (x|y) x∈Fy
Z =
dPY (y)PX (Fy |y) = PXY (F ).
Thus pX|Y (x|y) ) dPXY pX (x) Z X = H(X) + dPY (y) pX|Y (x|y) ln pX|Y (x|y). Z
I(X; Y ) =
ln(
x
2 We now wish to study the effects of quantizing on conditional information. As discussed in Section 3.6, it is not true that I(X; Y |Z) is always greater than I(f (X); q(Y )|r (Z)) and hence that I(X; Y |Z) can be written as a supremum over all quantizers and hence the definition of (7.30) and the formula (7.32) do not have the intuitive counterpart of a limit of informations of quantized values. We now consider an alternative (and
214
7 Relative Entropy
more general) definition of conditional mutual information due to Wyner [197]. The definition has the form of a supremum over quantizers and does not require the existence of the probability measure PX×Y |Z and hence makes sense for alphabets that are not standard. Given PXY Z and any finite measurements f and g on X and Y , we can choose a version of the conditional probability given Z = z so that pz (a, b) = Pr(f (X) = a, g(Y ) = b|Z = z) is a valid PMF with probability 1 (since the alphabets of f and g are finite and hence standard a regular conditional probability exists from Corollary 5.8.1 of [55] or Corollary 6.2 of [58]). For such finite measurements we can define I(f (X); g(Y )|Z = z) =
X
X
pz (a, b) ln P
a∈Af b∈Ag
pz (a, b) P , 0 0 a0 pz (a , b) b0 pz (a, b )
that is, the ordinary discrete average mutual information with respect to the distribution pz . Lemma 7.21. Define Z
I (X; Y |Z) = sup
dPZ (z)I(f (X); g(Y )|Z = z),
f ,g
where the supremum is over all quantizers. Then there exist sequences of quantizers (as in Lemma 7.18) such that I 0 (X; Y |Z) = lim I 0 (qm (X); rm (Y )|Z). n→∞
I 0 satisfies Kolmogorov’s formula, that is, I 0 (X; Y |Z) = I((X, Z); Y ) − I(Y ; Z). If the alphabets are standard, then I(X; Y |Z) = I 0 (X; Y |Z). Comment: The main point here is that conditional mutual information can be expressed as a supremum or limit of quantizers. The other results simply point out that the two conditional mutual informations have the same relation to ordinary mutual information and are (therefore) equal when both are defined. The proof follows Wyner [197]. Proof. First observe that for any quantizers q and r of Af and Ag we have from the usual properties of mutual information that I(q(f (X)); r (g(Y ))|Z = z) ≤ I(f (X); g(Y )|Z = z)
7.4 Information for General Alphabets
215
and hence integrating we have that Z I 0 (q(f (X)); r (g(Y ))|Z) = I(q(f (X)); r (g(Y ))|Z = z) dPZ (z) Z ≤ I(f (X); g(Y )|Z = z) dPZ (z) (7.42) and hence taking the supremum over all q and r to get I 0 (f (X); g(Y )|Z) yields Z 0 I (f (X); g(Y )|Z) = I(f (X); g(Y )|Z = z) dPZ (z). (7.43) so that (7.42) becomes I 0 (q(f (X)); r (g(Y ))|Z) ≤ I 0 (f (X); g(Y )|Z)
(7.44)
for any quantizers q and r and the definition of I 0 can be expressed as I 0 (X; Y |Z) = sup I 0 (f (X); g(Y )|Z),
(7.45)
f ,g
where the supremum is over all quantizers f and g. This proves the first part of the lemma since the supremum can be approached by a sequence of quantizers. Next observe that Z 0 I (f (X); g(Y )|Z) = I(f (X); g(Y )|Z = z) dPZ (z) = H(g(Y )|Z) − H(g(Y )|f (X), Z). Since we have from Lemma 7.20 that I(g(Y ); Z) = H(g(Y )) − H(g(Y )|Z), we have by adding these equations and again using Lemma 7.20 that I(g(Y ); Z) + I 0 (f (X); g(Y )|Z) = H(g(Y )) − H(g(Y )|f (X), Z) = I((f (X), Z); g(Y )). Taking suprema over both sides over all quantizers f and g yields the relation I(X; Z) + I 0 (X; Y |Z) = I((X, Z); Y ), proving Kolmogorov’s formula. Lastly, if the spaces are standard, then from Kolmogorov’s inequality for the original definition (which is valid for the standard space alphabets) combined with the above formula implies that I 0 (X; Y |Z) = I((X, Z); Y ) − I(X; Z) = I(X; Y |Z).
216
7 Relative Entropy
2
7.5 Convergence Results We now combine the convergence results for divergence with the definitions and properties of information densities to obtain several convergence results for information densities. Unlike the results to come for relative entropy rate and information rate, these are results involving the information between a sequence of random variables and a fixed random variable. Lemma 7.22. Given random variables X and Y1 , Y2 , · · · defined on a common probability space, lim I(X; (Y1 , Y2 , · · · , Yn )) = I(X; (Y1 , Y2 , · · · )).
n→∞
If in addition I(X; (Y1 , Y2 , · · · )) < ∞ and hence PX ×PY1 ,Y2 ,··· PX,Y1 ,Y2 ,··· , then iX;Y1 ,Y2 ,··· ,Yn → iX;Y1 ,Y2 ,··· n→∞
1
in Ł . Proof. The first result follows from Corollary 7.5 with X, Y1 , Y2 , · · · , Yn−1 replacing X n , P being the distribution PX,Y1 ,··· , and M being the product distribution PX × PY1 ,Y2 ,··· . The density result follows from Lemma 7.12.
2
Corollary 7.17. Given random variables X, Y , and Z1 , Z2 , · · · defined on a common probability space, then lim I(X; Y |Z1 , Z2 , · · · , Zn ) = I(X; Y |Z1 , Z2 , · · · ).
n→∞
If I((X, Z1 , · · · ); Y ) < ∞, ( e.g., if Y has a finite alphabet and hence I((X, Z1 , · · · ); Y ) ≤ H(Y ) < ∞), then also iX;Y |Z1 ,··· ,Zn → iX;Y |Z1 ,··· (7.46) n→∞
1
in L . Proof. From Kolmogorov’s formula I(X; Y |Z1 , Z2 , · · · , Zn ) = I(X; (Y , Z1 , Z2 , · · · , Zn )) − I(X; Z1 , · · · , Zn ) ≥ 0.
(7.47)
7.5 Convergence Results
217
From the previous lemma, the first term on the left converges as n → ∞ to I(X; (Y , Z1 , · · · )) and the second term on the right is the negative of a term converging to I(X; (Z1 , · · · )). If the first of these limits is finite, then the difference in (7.5) converges to the difference of these terms, which gives (7.46). From the chain rule for information densities, the conditional information density is the difference of the information densities: iX;Y |Z1 ,··· ,Zn = iX;(Y ,Z1 ,··· ,Zn ) − iX;(Z1 ,··· ,Zn ) which is converging in L1 x to iX;Y |Z1 ,··· = iX;(Y ,Z1 ,··· ) − iX;(Z1 ,··· ) , again invoking the density chain rule. If I(X; Y |Z1 , · · · ) = ∞ then quantize Y as q(Y ) and note since q(Y ) has a finite alphabet that I(X; Y |Z1 , Z2 , · · · , Zn ) ≥ I(X; q(Y )|Z1 , Z2 , · · · , Zn ) → I(X; q(Y )|Z1 , · · · ) n→∞
and hence lim inf I(X; Y |Z1 , · · · ) ≥ I(X; q(Y )|Z1 , · · · ). N→∞
Since the right-hand term above can be made arbitrarily large, the remaining part of the lemma is proved. 2 Lemma 7.23. If PX × PY1 ,Y2 ,··· PX,Y1 ,Y2 ,··· (e.g., I(X; (Y1 , Y2 , · · · )) < ∞), then with probability 1. lim
n→∞
1 i(X; (Y1 , · · · , Yn )) = 0. n
Proof. This is a corollary of Theorem 7.4. Let P denote the distribution of {X, Y1 , Y2 , · · · } and let M denote the distribution PX × PY1 ,··· . By assumption M P . The information density is i(X; (Y1 , · · · , Yn )) = ln
dPn , dMn
where Pn and Mn are the restrictions of P and M to σ (X, Y1 , · · · Yn ). Theorem 7.4 can therefore be applied to conclude that P -a.e. lim
n→∞
1 dPn = 0, ln n dMn
which proves the lemma. The lemma has the following immediate corollary. Corollary 7.18. If {Xn } is a process with the property that
2
218
7 Relative Entropy
I(X0 ; X−1 , X−2 , · · · ) < ∞, that is, there is a finite amount of information between the zero time sample and the infinite past, then lim
n→∞
1 i(X0 ; X−1 , · · · , X−n ) = 0. n
If the process is stationary, then also lim
n→∞
1 i(Xn ; X n ) = 0. n
Chapter 8
Information Rates
Abstract Definitions of information rate for processes with standard alphabets are developed and a mean ergodic theorem for information densities is proved. The relations among several different measures of information rate are developed.
8.1 Information Rates for Finite Alphabets Let {(Xn , Yn )} be a one-sided random process with finite alphabet A × B and let ((A × B)Z+ , B(A × B)Z+ ) be the corresponding one-sided sequence space of outputs of the pair process. We consider Xn and Yn to be the sampling functions on the sequence spaces A∞ and B ∞ and (Xn , Yn ) to be the pair sampling function on the product space, that is, for (x, y) ∈ A∞ × B ∞ , (Xn , Yn )(x, y) = (Xn (x), Yn (y)) = (xn , yn ). Let p denote the process distribution induced by the original space on the process {Xn , Yn }. Analogous to entropy rate we can define the mutual information rate (or simply information rate) of a finite alphabet pair process by 1 I(X, Y ) = lim sup I(X n , Y n ). n n→∞ If the finite alphabet pair process is AMS, then I(X; Y ) = H(X) + H(Y ) − H(X, Y )
(8.1)
and from Theorem 4.1 the entropy rates with respect to the AMS distribution equal those with respect to the stationary mean. These facts together with the properties of entropy rates of Theorems 3.3 and 4.1 yield the following lemma, where analogous to Theorem 4.1 we define the random variables p(X n , Y n ) by p(X n , Y n )(x, y) = p(X n = x n , Y n = y n ), p(X n ) by p(X n )(x, y) = p(X n = x n ), and similarly for p(Y n ). R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_8, © Springer Science+Business Media, LLC 2011
219
220
8 Information Rates
Lemma 8.1. Suppose that {Xn , Yn } is an AMS finite alphabet random process with distribution p and stationary mean p. Then the limits supremum defining information rates are limits and I p (X, Y ) = I p (X, Y ). I p is an affine function of the distribution p. If p has ergodic decomposition p xy , then Z I p (X, Y ) = dp(x, y)I pxy (X, Y ). If we define the information density in (X n , Y n ) = ln then lim
n→∞
p(X n , Y n ) , p(X n )p(Y n )
1 in (X n , Y n ) = I pxy (X, Y ) n
almost everywhere with respect to p and p and in L1 (p). The L1 results are extensions of the results of Moy [127] and Perez [148] for stationary processes, which in turn extended the Shannon-McMillan theorem from entropies of discrete alphabet processes to information densities. See also Kieffer [97]. The following lemmas follow either directly from or similarly to the corresponding results for entropy rate of Section 6.1. Lemma 8.2. Suppose that {Xn , Yn , X 0 n , Y 0 n } is an AMS process and n−1 1 X Pr((Xi , Yi ) 6= (X 0 i , Y 0 i )) ≤ n→∞ n i=0
P = lim
(the limit exists since the process is AMS). Then |I(X; Y ) − I(X 0 ; Y 0 )| ≤ 3( ln(kAk − 1) + h2 ()). Proof: The inequality follows from Corollary 6.1 since |(X; Y ) − I(X 0 ; Y 0 )| ≤ |H(X) − H(X 0 )| + |H(Y ) − H(Y 0 )| + |H(X, Y ) − H(X 0 , Y 0 )| and since Pr((Xi , Yi ) 6= (Xi 0 , Yi 0 )) = Pr(Xi 6= Xi 0 or Yi 6= Yi 0 ) is no smaller than Pr(Xi 6= Xi 0 ) or Pr(Yi 6= Yi 0 ). 2 Corollary 8.1. Let {Xn , Yn } be an AMS process and let f and g be stationary measurements on X and Y , respectively. Given > 0 there is an N sufficiently large, scalar quantizers q and r , and mappings f 0 and g 0 which
8.2 Information Rates for General Alphabets
221
depend only on {q(X0 ), · · · , q(XN−1 )} and {r (Y0 ), · · · , r (YN−1 )} in the one-sided case and {q(X−N ), · · · , q(XN )} and {r (Y−N ), · · · , r (YN )} in the two-sided case such that |I(f ; g) − I(f 0 ; g 0 )| ≤ . Proof: Choose the codes f 0 and g 0 from Lemma 5.2 and apply the previous lemma. 2 Lemma 8.3. If {Xn , Yn } is an AMS process and f and g are stationary codings of X and Y , respectively, then I(X; Y ) ≥ I(f ; g). Proof: This is proved as Corollary 6.4 by first approximating f and g by finite-window stationary codes, applying the result for mutual information (Lemma 3.12), and then taking the limit. 2
8.2 Information Rates for General Alphabets Suppose that we are given a pair random process {Xn , Yn } with distribution p. The most natural definition of the information rate between the two processes is the extension of the definition for the finite alphabet case: 1 I(X; Y ) = lim sup I(X n ; Y n ). n→∞ n This was the first general definition of information rate and it is due to Dobrushin [32]. While this definition has its uses, it also has its problems. Another definition is more in the spirit of the definition of information itself: We formed the general definitions by taking a supremum of the finite alphabet definitions over all finite-alphabet codings or quantizers. The above definition takes the limit of such suprema. An alternative definition is to instead reverse the order and take the supremum of the limit and hence the supremum of the information rate over all finite-alphabet codings of the process. This reversal of supremum and limit provides a definition of information rate similar to the definition of the entropy of a dynamical system. There is a question as to what kind of codings we permit, that is, do the quantizers quantize individual outputs or long sequences of outputs. We shall shortly see that it makes no difference. Suppose that we have a pair random process {Xn , Yn } with standard al∞ phabets AX and AY and suppose that f : A∞ X → Af and g : AY → Ag are stationary codings of the X and Y sequence spaces into a finite alphabet. Let {fn , gn } be the induced output process, that is, if T denotes the shift (on any of the sequence spaces) then fn (x, y) = f (T n x) and
222
8 Information Rates
gn (x, y) = g(T n y). Recall that f (T n (x, y)) = fn (x, y), that is, shifting the input n times results in the output being shifted n times. Since the new process {fn , gn } has a finite alphabet, its mutual information rate is defined. We define the information rate for general alphabets by I ∗ (X; Y ) = =
sup sliding-block codes sup sliding-block codes
I(f ; g) f ,g
lim sup f ,g
n→∞
1 I(f n ; g n ). n
We now focus on AMS processes, in which case the information rates for finite-alphabet processes (e.g., quantized processes) is given by the limit, that is, I ∗ (X; Y ) = =
sup sliding-block codes sup sliding-block codes
I(f ; g) f ,g
lim
f ,g
n→∞
1 I(f n ; g n ). n
Note for later use the following simple inequality which follows from the the above facts and Corollary 7.11. As with that corollary, the result is useful if the entropy rate of one component of the process is finite. Lemma 8.4. Given an AMS pair process (X, Y ) with standard alphabets, I ∗ (X; Y ) ≤ H(Y ). The following lemma shows that for AMS sources I ∗ can also be evaluated by constraining the sliding-block codes to be scalar quantizers. Lemma 8.5. Given an AMS pair random process {Xn , Yn } with standard alphabet, I ∗ (X; Y ) = sup I(q(X); r (Y )) = sup lim sup q,r
q,r
n→∞
1 I(q(X)n ; r (Y )n ), n
where the supremum is over all quantizers q of AX and r of AY and where q(X)n = q(X0 ), · · · , q(Xn−1 ). Proof: Clearly the right hand side above is less than I ∗ since a scalar quantizer is a special case of a stationary code. Conversely, suppose that f and g are sliding-block codes such that I(f ; g) ≥ I ∗ (X; Y ) − . Then from Corollary 8.1 there are quantizers q and r and codes f 0 and g 0 depending only on the quantized processes q(Xn ) and r (Yn ) such that I(f 0 ; g 0 ) ≥ I(f ; g) − . From Lemma 8.3, however, I(q(X); r (Y )) ≥ I(f 0 ; g 0 ) since f 0 and g 0 are stationary codings of the quantized pro-
8.2 Information Rates for General Alphabets
223
cesses. Thus I(q(X); r (Y )) ≥ I ∗ (X; Y ) − 2, which proves the lemma.
2 Corollary 8.2. I ∗ (X; Y ) ≤ I(X; Y ). If the alphabets are finite, then the two rates are equal. Proof: The inequality follows from the lemma and the fact that I(X n ; Y n ) ≥ I(q(X)n ; r (Y )n ) for any scalar quantizers q and r (where q(X)n is q(X0 ), · · · , q(Xn−1 )). If the alphabets are finite, then the identity mappings are quantizers and yield I(X n ; Y n ) for all n. 2 Pinsker [150] introduced the definition of information rate as a supremum over all scalar quantizers and hence we refer to this information rate as the Pinsker rate. The Pinsker definition has the advantage that we can use the known properties of information rates for finite-alphabet processes to infer those for general processes, an attribute the first definition lacks. Corollary 8.3. Given a standard alphabet pair process alphabet AX × AY , there is a sequence of scalar quantizers qm and rm such that for any AMS pair process {Xn , Yn } having this alphabet (that is, for any process distribution on the corresponding sequence space) I(X n ; Y n ) = lim I(qm (X)n ; rm (Y )n ) m→∞
I ∗ (X; Y ) = lim I(qm (X); rm (Y )). m→∞
Furthermore, the above limits can be taken to be increasing by using finer and finer quantizers. Comment: It is important to note that the same sequence of quantizers gives both of the limiting results. Proof: The first result is Lemma 7.18. The second follows from the previous lemma. 2 Observe that I ∗ (X; Y ) = lim lim sup m→∞
n→∞
whereas I(X; Y ) = lim sup lim n→∞
m→∞
1 I(qm (X); rm (Y )) n
1 I(qm (X); rm (Y )). n
Thus the two notions of information rate are equal if the two limits can be interchanged. We shall later consider conditions under which this is
224
8 Information Rates
true and we shall see that equality of these two rates is important for proving ergodic theorems for information densities. Lemma 8.6. Suppose that {Xn , Yn } is an AMS standard alphabet random process with distribution p and stationary mean p. Then Ip∗ (X; Y ) = Ip∗ (X; Y ). Ip∗ is an affine function of the distribution p. If p has ergodic decomposition p xy , then Z Ip∗ (X; Y ) = dp(x, y)I ∗ pxy (X; Y ). If f and g are stationary codings of X and Y , then Z Ip∗ (f ; g) = dp(x, y)Ip∗xy (f ; g). Proof: For any scalar quantizers q and r of X and Y we have that I p (q(X); r (Y )) = I p (q(X); r (Y )). Taking a limit with ever finer quantizers yields the first equality. The fact that I ∗ is affine follows similarly. Suppose that p has ergodic decomposition p xy . Define the induced distributions of the quantized process by m and mxy , that is, m(F ) = p(x, y : {q(xi ), r (yi ); i ∈ T} ∈ F ) and similarly for mxy . The mxy are stationary and ergodic since they are stationary codings of stationary ergodic processes and together they form an ergodic decomposition of m, which must also be stationary. Let Xn0 , Yn0 denote the coordinate functions on the quantized output sequence space (that is, the processes {q(Xn ), r (Yn )} and {Xn0 , Yn0 } are equivalent), then using the ergodic decomposition of mutual information for finite-alphabet processes (Lemma 8.1) we have that Z I p (q(X); r (Y )) = I m (X 0 ; Y 0 ) = I mx0 y 0 (X 0 ; Y 0 ) dm(x 0 , y 0 ) Z = I pxy (q(X); r (Y )) dp(x, y). Replacing the quantizers by the sequence qm , rm the result then follows by taking the limit using the monotone convergence theorem. The result for stationary codings follows similarly by applying the previous result to the induced distributions and then relating the equation to the original distributions. 2 The above properties are not known to hold for I in the general case. Thus although I may appear to be a more natural definition of mutual information rate, I ∗ is better behaved since it inherits properties from the discrete alphabet case. It will be of interest to find conditions under which the two rates are the same, since then I will share the properties
8.3 A Mean Ergodic Theorem for Densities
225
possessed by I ∗ . The first result of the next section adds to the interest by demonstrating that when the two rates are equal, a mean ergodic theorem holds for the information densities.
8.3 A Mean Ergodic Theorem for Densities Theorem 8.1. Given an AMS pair process {Xn , Yn } with standard alphabets, assume that for all n PX n × PY n PX n Y n and hence that the information densities iX n ;Y n = ln
dPX n ,Y n d(PX n × PY n )
are well defined. For simplicity we abbreviate iX n ;Y n to in when there is no possibility of confusion. If the limit lim
n→∞
1 I(X n ; Y n ) = I(X; Y ) n
exists and I(X; Y ) = I ∗ (X; Y ) < ∞, then n−1 in (X n ; Y n ) converges in L1 to an invariant function i(X; Y ). If the stationary mean of the process has an ergodic decomposition p xy , then the limiting density is I ∗ pxy (X; Y ), the information rate of the ergodic component in effect. Proof: Let qm and rm be asymptotically accurate quantizers for AX and ˆn = qm (Xn ) and Y ˆn = rm (Yn ). AY . Define the discrete approximations X n n n n Observe that PX × PY PX Y implies that PXˆn × PYˆ n PXˆn Yˆ n and hence we can define the information densities of the quantized vectors by dPXˆn Yˆ n iˆn = ln . d(PXˆn × PYˆ n ) For any m we have that
226
8 Information Rates
Z |
1 in (x n ; y n ) − I ∗ pxy (X; Y )| dp(x, y) ≤ n Z 1 1 | in (x n ; y n ) − iˆn (qm (x)n ; rm (y)n )| dp(x, y)+ n n Z 1ˆ n | in (qm (x) ; rm (y)n ) − I pxy (qm (X); rm (Y ))| dp(x, y)+ n Z |I pxy (qm (X); rm (Y )) − I ∗ pxy (X; Y )| dp(x, y), (8.2)
where qm (x)n = (qm (x0 ), · · · , qm (xn−1 )), rm (y)n = (rm (y0 ), · · · , rm (yn−1 )), and I p (qm (X); rm (Y )) denotes the information rate of the process {qm (Xn ), rm (Yn ); n = 0, 1, · · · , } when p is the process measure describing {Xn , Yn }. Consider first the right-most term of (8.2). Since I ∗ is the supremum over all quantized versions, Z
|I pxy (qm (X); rm (Y )) − Ip∗xy (X; Y )| dp(x, y) = Z (Ip∗xy (X; Y ) − I pxy (qm (X); rm (Y ))) dp(x, y).
Using the ergodic decomposition of I ∗ (Lemma 8.6) and that of I for discrete alphabet processes (Lemma 8.1) this becomes Z
|I pxy (qm (X); rm (Y )) − Ip∗xy (X; Y )| dp(x, y) = Ip∗ (X; Y ) − I p (qm (X); rm (Y )).
(8.3)
For fixed m the middle term of (8.2) can be made arbitrarily small by taking n large enough from the finite alphabet result of Lemma 8.1. The first term on the right can be bounded above using Corollary 7.6 with F = σ (q(X)n ; r (Y )n ) by 1 2 ˆn ; Y ˆn) + I(X n ; Y n ) − I(X n e which as n → ∞ goes to I(X; Y ) −I(qm (X); rm (Y )). Thus we have for any m that
8.4 Information Rates of Stationary Processes
Z lim sup n→∞
|
227
1 in (x n ; y n ) − Ip∗xy (X; Y )| dp(x, y) ≤ n I(X; Y ) − I(qm (X); rm (Y )) + I ∗ (X; Y ) − I(qm (X); rm (Y ))
which as m → ∞ becomes I(X; Y ) − I ∗ (X; Y ), which is 0 by assumption.
2
8.4 Information Rates of Stationary Processes In this section we introduce two more definitions of information rates for the case of stationary two-sided processes. These rates are useful tools in relating the Dobrushin and Pinsker rates and they provide additional interpretations of mutual information rates in terms of ordinary mutual information. The definitions follow Pinsker [150]. Henceforth assume that {Xn , Yn } is a stationary two-sided pair process with standard alphabets. Define the sequences y = {yi ; i ∈ T} and Y = {Yi ; i ∈ T} First define 1 I˜(X; Y ) = lim sup I(X n ; Y ), n→∞ n that is, consider the per-letter limiting information between n-tuples of X and the entire sequence from Y . Next define I − (X; Y ) = I(X0 ; Y |X−1 , X−2 , · · · ), that is, the average conditional mutual information between one letter from X and the entire Y sequence given the infinite past of the X process. We could define the first rate for one-sided processes, but the second makes sense only when we can consider an infinite past. For brevity we write X − = X−1 , X−2 , · · · and hence I − (X; Y ) = I(X0 ; Y |X − ). Theorem 8.2. I˜(X; Y ) ≥ I(X; Y ) ≥ I ∗ (X; Y ) ≥ I − (X; Y ). If the alphabet of X is finite, then the above rates are all equal. Comment: We will later see more general sufficient conditions for the equality of the various rates, but the case where one alphabet is finite is simple and important and points out that the rates are all equal in the finite alphabet case.
228
8 Information Rates
Proof: We have already proved the middle inequality. The left inequality follows immediately from the fact that I(X n ; Y ) ≥ I(X n ; Y n ) for all n. The remaining inequality is more involved. We prove it in two steps. First we prove the second half of the theorem, that the rates are the same if X has finite alphabet. We then couple this with an approximation argument to prove the remaining inequality. Suppose now that the alphabet of X is finite. Using the chain rule and stationarity we have that n−1 1 1 X I(X n ; Y n ) = I(Xi ; Y n |X0 , · · · , Xi−1 ) n n i=0
=
n−1 1 X n I(X0 ; Y−i |X−1 , · · · , X−i ), n i=0
n is Y−i , · · · , Y−i+n−1 , that is, the n-vector starting at −i. Since X where Y−i has finite alphabet, each term in the sum is bounded. We can show as in Section 7.4 (or using Kolmogorov’s formula and Lemma 7.14) that each term converges as i → ∞, n → ∞, and n − i → ∞ to I(X0 ; Y |X−1 , X−2 , · · · ) or I − (X; Y ). These facts, however, imply that the above Cesàro average converges to the same limit and hence I = I − . We can similarly expand I˜ as n−1 n−1 1 X 1 X I(Xi ; Y |X0 , · · · , Xi−1 ) = I(X0 ; Y |X−1 , · · · , X−i ), n i=0 n i=0
which converges to the same limit for the same reasons. Thus I˜ = I = I − for stationary processes when the alphabet of X is finite. Now suppose that X has a standard alphabet and let qm be an asymptotically accurate sequences of quantizers. Recall that the corresponding partitions are increasing, that is, each refines the previous partition. Fix > 0 and choose m large enough so that the quantizer α(X0 ) = qm (X0 ) satisfies I(α(X0 ); Y |X − ) ≥ I(X0 ; Y |X − ) − . Observe that so far we have only quantized X0 and not the past. Since Fm = σ (α(X0 ), Y , qm (X−i ); i = 1, 2, · · · ) asymptotically generates σ (α(X0 ), Y , X−i ; i = 1, 2, · · · ), given we can choose for m large enough (larger than before) a quantizer β(x) = qm (x) such that if we define β(X − ) to be β(X−1 ), β(X−2 ), · · · , then |I(α(X0 ); (Y , β(X − ))) − I(α(X0 ); (Y , X − ))| ≤
8.4 Information Rates of Stationary Processes
229
and |I(α(X0 ); β(X − )) − I(α(X0 ); X − )| ≤ . Using Kolmogorov’s formula this implies that |I(α(X0 ); Y |X − ) − I(α(X0 ); Y |β(X − ))| ≤ 2 and hence that I(α(X0 ); Y |β(X − )) ≥ I(α(X0 ); Y |X − ) − 2 ≥ I(X0 ; Y |X − ) − 3. But the partition corresponding to β refines that of α and hence increases the information; hence I(β(X0 ); Y |β(X − )) ≥ I(α(X0 ); Y |β(X − )) ≥ I(X0 ; Y |X − ) − 3. Since β(Xn ) has a finite alphabet, however, from the finite alphabet result the left-most term above must be I(β(X); Y ), which can be made arbitrarily close to I ∗ (X; Y ). Since is arbitrary, this proves the final inequality. 2 The following two theorems provide sufficient conditions for equality of the various information rates. The first result is almost a special case of the second, but it is handled separately as it is simpler, much of the proof applies to the second case, and it is not an exact special case of the subsequent result since it does not require the second condition of that result. The result corresponds to condition (7.4.33) of Pinsker [150], who also provides more general conditions. The more general condition is also due to Pinsker and strongly resembles that considered by Barron [8]. Theorem 8.3. Given a stationary pair process {Xn , Yn } with standard alphabets, if I(X0 ; (X−1 , X−2 , · · · )) < ∞, then I˜(X; Y ) = I(X; Y ) = I ∗ (X; Y ) = I − (X; Y ).
(8.4)
Proof: We have that 1 1 1 1 I(X n ; Y ) ≤ I(X n ; (Y , X − )) = I(X n ; X − ) + I(X n ; Y |X − ), n n n n
(8.5)
where, as before, X − = {X−1 , X−2 , · · · }. Consider the first term on the right. Using the chain rule for mutual information
230
8 Information Rates n−1 1 1 X I(X n ; X − ) = I(Xi ; X − |X i ) n n i=0
=
n−1 1 X (I(Xi ; (X i , X − )) − I(Xi ; X i )). n i=0
(8.6)
(8.7)
Using stationarity we have that n−1 1 X 1 I(X n ; X − ) = (I(X0 ; X − ) − I(X0 ; (X−1 , · · · , X−i )). n n i=0
(8.8)
The terms I(X0 ; (X−1 , · · · , X−i )) are converging to I(X0 ; X − ), hence the terms in the sum are converging to 0, i.e., lim I(Xi ; X − |X i ) = 0.
i→∞
(8.9)
The Cesàro mean of (8.7) is converging to the same thing and hence 1 I(X n ; X − ) → 0. n
(8.10)
Next consider the term I(X n ; Y |X − ). For any positive integers n,m we have I(X n+m ; Y |X − ) = I(X n ; Y |X − ) + I(Xnm ; Y |X − , X n ), (8.11) where Xnm = Xn , · · · , Xn+m−1 . From stationarity, however, the rightmost term is just I(X m ; Y |X − ) and hence I(X m+n ; Y |X − ) = I(X n ; Y |X − ) + I(X m ; Y |X − ).
(8.12)
This is just a linear functional equation of the form f (n + m) = f (n) + f (m) and the unique solution to such an equation is f (n) = nf (1), that is, 1 (8.13) I(X n ; Y |X − ) = I(X0 ; Y |X − ) = I − (X; Y ). n Taking the limit supremum in (8.5) yields I˜(X; Y ) ≤ I − (X; Y ),
(8.14)
which with Theorem 8.2 completes the proof.
2
Intuitively, the theorem states that if one of the processes has finite average mutual information between one symbol and its infinite past, then the Dobrushin and Pinsker information rates yield the same value and hence there is an L1 ergodic theorem for the information density. To generalize the theorem we introduce a condition that will often be useful when studying asymptotic properties of entropy and informa-
8.4 Information Rates of Stationary Processes
231
tion. A stationary process {Xn } is said to have the finite-gap information property if there exists an integer K such that I(XK ; X − |X K ) < ∞,
(8.15)
where, as usual, X − = (X−1 , X−2 , · · · ). When a process has this property for a specific K, we shall say that it has the K-gap information property. Observe that if a process possesses this property, then it follows from Lemma 7.17 I(XK ; (X−1 , · · · , X−l )|X K ) < ∞; l = 1, 2, · · ·
(8.16)
Since these informations are finite, (K)
PX n PX n ; n = 1, 2, . . . ,
(8.17)
(K)
where PX n is the Kth order Markov approximation to PX n . Theorem 8.4. Given a stationary standard alphabet pair process {Xn , Yn }, if {Xn } satisfies the finite-gap information property (8.15) and if, in addition, I(X K ; Y ) < ∞, (8.18) then (8.4) holds. If K = 0 then there is no conditioning and (8.18) is trivial, that is, the previous theorem is the special case with K = 0. Comment: This theorem shows that if there is any finite dimensional future vector (XK , XK+1 , · · · , XK+N−1 ) which has finite mutual information with respect to the infinite past X − when conditioned on the intervening gap (X0 , · · · , XK−1 ), then the various definitions of mutual information are equivalent provided that the mutual information betwen the “gap” X K and the sequence Y are finite. Note that this latter condition will hold if, for example, I˜(X; Y ) is finite. Proof: For n > K 1 1 1 I(X n ; Y ) = I(X K ; Y ) + I(XKn−K ; Y |X K ). n n n By assumption the first term on the left will tend to 0 as n → ∞ and hence we focus on the second, which can be broken up analogous to the previous theorem with the addition of the conditioning: 1 1 I(XKn−K ; Y |X K ) ≤ I(XKn−K ; (Y , X − |X K )) n n 1 1 = I(XKn−K ; X − |X K ) + I(XKn−K ; Y |X − , X K ). n n
232
8 Information Rates
Consider first the term n−1 1 1 X I(Xi ; X − |X i ), I(XKn−K ; X − |X K ) = n n i=K
which is as (8.7) in the proof of Theorem 8.3 except that the first K terms are missing. The same argument then shows that the limit of the sum is 0. The remaining term is 1 1 I(XKn−K ; Y |X − , X K ) = I(X n ; Y |X − ) n n exactly as in the proof of Theorem 8.3 and the same argument then shows that the limit is I − (X; Y ), which completes the proof. 2 One result developed in the proofs of Theorems 8.3 and 8.4 will be important later in its own right and hence we isolate it as a corollary. The result is just (8.9), which remains valid under the more general conditions of Theorem 8.4, and the fact that the Cesàro mean of converging terms has the same limit. Corollary 8.4. If a process {Xn } has the finite-gap information property I(XK ; X − |X K ) < ∞ for some K, then lim I(Xn ; X − |X n ) = 0
n→∞
and lim
n→∞
1 I(X n ; X − ) = 0. n
The corollary can be interpreted as saying that if a process has the the finite gap information property, then the mutual information between a single sample and the infinite past conditioned on the intervening samples goes to zero as the number of intervening samples goes to infinity. This can be interpreted as a form of asymptotic independence property of the process. Corollary 8.5. If a one-sided stationary source {Xn } is such that for some K K, I(Xn ; X n−K |Xn−K ) is bounded uniformly in n, then it has the finite-gap property and hence I(X; Y ) = I ∗ (X; Y ). Proof: Simply imbed the one-sided source into a two-sided stationary source with the same probabilities on all finite-dimensional events. For that source K I(Xn ; X n−K |Xn−K ) = I(XK ; X−1 , · · · , X−n−K |X K ) → I(XK ; X − |X K ). n→∞
8.4 Information Rates of Stationary Processes
233
Thus if the terms are bounded, the conditions of Theorem 8.3 are met for the two-sided source. The one-sided equality then follows. 2 The above results have an information theoretic implication for the ergodic decomposition, which is described in the next theorem. Theorem 8.5. Suppose that {Xn } is a stationary process with the finitegap property (8.15). Let ψ be the ergodic component function of Theorem 1.6 and suppose that for some n I(X n ; ψ) < ∞.
(8.19)
(This will be the case, for example, if the finite-gap information property holds for 0 gap, that is, I(X0 ; X − ) < ∞ since ψ can be determined from X − and information is decreased by taking a function.) Then lim
n→∞
1 I(X n ; ψ) = 0. n
Comment: For discrete alphabet processes this theorem is just the ergodic decomposition of entropy rate in disguise (Theorem 3.3). It also follows for finite-alphabet processes from Lemma 4.3. We shall later prove a corresponding almost everywhere convergence result for the corresponding densities. All of these results have the interpretation that the per-symbol mutual information between the outputs of the process and the ergodic component decreases with time because the ergodic component in effect can be inferred from the process output in the limit of an infinite observation sequence. The finiteness condition on some I(X n ; ψ) is necessary for the nonzero finite-gap case to avoid cases such as where Xn = ψ for all n and hence I(X n ; ψ) = I(ψ; ψ) = H(ψ) = ∞, in which case the theorem does not hold. Proof: Define ψn = ψ for all n. Since ψ is invariant, {Xn , ψn } is a stationary process. Since Xn satisfies the given conditions, however, I(X; ψ) = I ∗ (X; ψ). But for any scalar quantizer q, I(q(X); ψ) is 0 from Lemma 4.3. I ∗ (X; ψ) is therefore 0 since it is the supremum of I(q(X); ψ) over all quantizers q. Thus 0 = I(X; ψ) = lim
n→∞
1 1 I(X n ; ψn ) = lim I(X n ; ψ). 2 n→∞ n n
234
8 Information Rates
8.5 The Data Processing Theorem The following is a basic property of a communication system. If a communication system is stationary, then the mutual information rate between the overall input and output cannot exceed that over the channel. The result is often called the data processing theorem. Lemma 8.7. Suppose that a communication system is stationary in the ˆn } is stationary. Then sense that the process {Xn , Un , Yn , X ˆ I˜(U ; Y ) ≥ I(X; Y ) ≥ I(X; X).
(8.20)
If {Un } has a finite alphabet or if it has has the K-gap information property (8.15) and I(U K , Y ) < ∞, then ˆ ≤ I(U; Y ). I(X; X) ˆn } is a stationary deterministic encoding of the {Yn } Proof: Since {X ∗ ˆ I(X; X) ≤ I (X; Y ). From Theorem 8.2 the right hand side is bounded above by I(X; Y ). For each n I(X n ; Y n ) ≤ I((X n , U); Y n ) = I(Y n ; U) + I(X n ; Y n |U) = I(Y n ; U), where U = {Un , n ∈ T} and we have used the fact that X → U → Y is a Markov chain and hence so is X N → U → Y K and hence the conditional mutual information is 0 (Lemma 7.15). Thus I(X; Y ) ≤ lim I(Y n ; U) = I˜(Y ; U). n→∞
Applying Theorem 8.2 then proves that ˆ ≤ I˜(Y ; U). I(X; X) If {Un } has finite alphabet or has the K-gap information property and I(U K , Y ) < ∞, then from Theorems 8.2 or 8.4, respectively, I˜(Y ; U) = I((Y ; U ), completing the proof. 2 The lemma can be easily extended to block stationary processes. Corollary 8.6. Suppose that the process of the previous lemma is not stationary, but is (N, K)-stationary in the sense that the vector process N N K K ˆnN {XnN , UnK , YnK ,X } is stationary. Then ˆ ≤ I(X; X)
K I(U; Y ). N
8.6 Memoryless Channels and Sources
235
Proof: Apply the previous lemma to the stationary vector sequence to conclude that ˆN ) ≤ I(U K ; Y K ). I(X N ; X But
i h 1 ˆnN ) = lim E n−1 iX nN ,XˆnN , I(X nN ; X n→∞ n n→∞
ˆN ) = lim I(X N ; X
which is N times the limiting expectation of a subsequence of the densities n−1 iX n ,Xˆn , whose expectation converges to I(X; Y ). Thus ˆ I(X N ; X N ) = NI(X; X). A similar manipulation for I(U K ; Y K ) completes the proof.
2
8.6 Memoryless Channels and Sources A useful inequality is developed in this section for the mutual information between the input and output of a memoryless channel. For contrast we also describe the corresponding result for a memoryless source and an arbitrary channel. Lemma 8.8. Let {Xn } be a source with distribution µ and let ν be a channel. Let {Xn , Yn } be the hookup with distribution p. If the channel is memoryless, then for any n n
n
I(X ; Y ) ≤
n−1 X
I(Xi ; Yi )
i=0
If instead the source is memoryless, then the inequality is reversed: I(X n ; Y n ) ≥
n−1 X
I(Xi ; Yi ).
i=0
Thus if both source and channel are memoryless, I(X n ; Y n ) =
n−1 X
I(Xi ; Yi ).
i=0
Proof: First suppose that the process is discrete. Then I(X n ; Y n ) = H(Y n ) − H(Y n |X n ). Since by construction
236
8 Information Rates
PY n |X n (y n |x n ) =
n−1 Y
PY0 |X0 (yi |xi )
i=0
an easy computation shows that H(Y n |X n ) =
n−1 X
H(Yi |Xi ).
i=0
This combined with the inequality H(Y n ) ≤
n−1 X
H(Yi )
i=0
(Lemma 3.2 used several times) completes the proof of the memoryless channel result for finite alphabets. If instead the source is memoryless, we have I(X n ; Y n ) = H(X n ) − H(X n |Y n ) =
n−1 X
H(Xi ) − H(X n |Y n ).
i=0
Extending Lemma 3.2 to conditional entropy yields H(X n |Y n ) ≤
n−1 X
H(Xi |Y n )
i=0
which can be further overbounded by using Lemma 3.12 (the fact that reducing conditioning increases conditional entropy) as H(X n |Y n ) ≤
n−1 X
H(Xi |Yi )
i=0
which implies that I(X n ; Y n ) ≥
n−1 X i=0
H(Xi ) − H(Xi |Yi ) =
n−1 X
I(Xi ; Yi ),
i=0
which completes the proof for finite alphabets. To extend the result to standard alphabets, first consider the case where the Y n are quantized to a finite alphabet. If the Yk are conditionally independent given X k , then the same is true for q(Yk ), k = 0, 1, · · · , n − 1. Lemma 7.20 then implies that as in the discrete case, I(X n ; Y n ) = H(Y n ) − H(Y n |X n ) and the remainder of the proof follows as in the discrete case. Letting the quantizers become asymptotically accurate then completes the proof. 2
Chapter 9
Distortion and Information
Abstract A pair random process (X, Y ) = {Xn , Yn } can be generated by a source and a channel, a source and a code, a combination of a source, channel, encoder, and decoder, or two sources and a coupling. We have developed in some detail the properties of two quantities characterizing relations between the two components of a pair process: the average distortion between the components and their mutual information rate. In this chapter relations are developed between distortion and rate, where rate is measured by mutual information. The primary results concern the Shannon distortion-rate function and its dual, the Shannon rate-distortion function, which will be seen to provide bounds on the relationships between distortion and entropy and to characterize the optimal performance in source coding and rate-constrained simulation systems.
9.1 The Shannon Distortion-Rate Function Given a source [A, µ] and a fidelity criterion ρn ; n = 1, 2, . . . defined on ˆ where A ˆ is the reproduction alphabet, the Shannon distortion-rate A × A, function (DRF) is defined in terms of a nonnegative parameter called rate by 1 D(R, µ) = lim sup DN (R, µ N ) N N→∞ where DN (R, µ N ) =
inf
p N ∈RN (R,µ N )
EpN ρN (X N , Y N )
where RN (R, µ N ) is the collection of all distributions p N for the coordiN ˆN , BN nate random vectors X N and Y N on the space (AN × A A × BA ˆ ) with the properties that
R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_9, © Springer Science+Business Media, LLC 2011
237
238
9 Distortion and Information
ˆN ) = µ N (F ) for (1) p N induces the given marginal µ N ; that is, p N (F × A N all F ∈ BA , and (2) the mutual information satisfies 1 ˆN ) ≤ R. IpN (X N ; X N DN is called the Nth order distortion-rate function. If RN (R, µ N ) is empty, then DN (R, µ N ) is ∞. Readers familiar with the rate-distortion theory literature may notice that the definition is not the one usually encountered, which is an infimum over N and not a limit supremum. This is simply because the infimum is most useful for stationary processes, while the limit supremum turns out to be the right form in the general AMS case eventually considered here. It will be seen shortly that the two definitions are equal if the source is stationary. Various other notations are used for the distortion-rate function in the literature and here when convenience suggests it. The distribution µ may be dropped if it is fixed, writing D(R) instead of D(R, µ). Sometimes the random variable is used instead of the distribution, e.g., writing DX (R) for D(R, µ). An alternative approach to minimizing over joint distributions with a constrained input marginal is to minimize over test channels or regular conditional probabilities ν which induce a joint distribution by the hookup µν. This is equivalent since we are considering only standard alphabets and hence any joint distribution p with input marginal distribution µ will induce a regular conditional probability distribution ν for which p = µν. One can also define the dual or inverse function to the distortion-rate function, Shannon’s rate-distortion function, by 1 RN (D, µ N ) N ˆN ), inf IpN (X N ; X
R(D, µ) = lim sup
(9.1)
N→∞
RN (D, µ N ) =
p N ∈DN (D,µ N )
(9.2)
where DN (D, µ N ) is the collection of all distributions p N for the coordiˆN with the propnate random vectors X N and Y N on the space (AN × A erties that ˆN ) = µ N (F ) for (1) p N induces the given marginal µ N ; that is, p N (F × A N all F ∈ BA , and (2) the average distortion satisfies 1 EpN ρN (X N , Y N ) ≤ D. N
9.2 Basic Properties
239
In his original development of source coding subject to a fidelity criterion or rate-distortion theory, Shannon considered the rate-distortion function rather than the distortion-rate function [162, 163]. We emphasize the distortion-rate function because it is a better match to the formulation of source coding considered here. In particular, in communications applications it is usually the rate that is constrained by a communications or storage medium such as a noisy channel or a limited memory. The theory and algorithms for the evaluation of rate-distortion tradeoffs are usually simpler if stated using the rate-distortion viewpoint rather than the distortion-rate viewpoint, and most of the evaluation literature uses the rate-distortion approach. Hence we shall focus on the distortion-rate function for most of the development, the corresponding properties for rate-distortion functions follow similarly. For the discussion of bounding and evaluating the functions, we shall use the ratedistortion viewpoint.
9.2 Basic Properties LemmaS9.1. DN (R, µ) and D(R, µ) are nonnegative, nonincreasing, and convex functions of R and hence are continuous in R for R > 0. Proof: Nonnegativity is obvious from the nonnegativity of distortion. Nonincreasing follows since if R2 > R1 , then RN (R1 , µ N ) ⊂ RN (R2 , µ N ) and hence a minimization over the larger set can not yield a worse (larger) result. Suppose that pi ∈ RN (Ri , µ N ); i = 1, 2 yields Epi ρN (X N , Y N ) ≤ DN (Ri , µ) + . S From Lemma 7.19 mutual information is a convex function of the conditional distribution and hence if p = λp1 + (1 − λ)p2 , then Ip ≤ λIp1 + (1 − λ)Ip2 ≤ λR1 + (1 − λ)R2 and hence p ∈ RN (λR1 + (1 − λ)R2 ) and therefore DN (λR1 + (1 − λ)R2 ) ≤ Ep ρN (X N , Y N ) = λEp1 ρN (X N , Y N ) + (1 − λ)Ep2 ρN (X N , Y N ) ≤ λDN (R1 , µ) + (1 − λ)DN (R2 , µ). Since D(R, µ) is the limit of DN (R, µ), it too is nonincreasing and convex. It is well known from real analysis that convex functions are continuous except possibly at their end points. 2
240
9 Distortion and Information
The following lemma shows that when the underlying source is stationary and the fidelity criterion is subadditive (e.g., additive), then the limit defining D(R, µ) is an infimum. Lemma 9.2. If the source µ is stationary and the fidelity criterion is subadditive, then D(R, µ) = lim DN (R, µ) = inf N→∞
N
1 DN (R, µ). N
Proof: Fix N and n < N and let p n ∈ Rn (R, µ n ) yield Epn ρn (X n , Y n ) ≤ Dn (R, µ n ) +
2
and let p N−n ∈ RN−n (R, µ N−n ) yield EpN−n ρN−n (X N−n , Y N−n ) ≤ DN−n (R, µ N−n ) +
. 2
pn together with µ n implies a regular conditional probability q(F |x n ), N−n F ∈ Bn imply a regular conditional probability ˆ . Similarly pN−n and µ A N−n ). Define now a regular conditional probability t(·|x N ) by its r (G|x values on rectangles as N−n N−n t(F × G|x N ) = q(F |x n )r (G|xn ); F ∈ Bn . ˆ , G ∈ BA ˆ A
Note that this is the finite dimensional analog of a block memoryless channel with two blocks. Let p N = µ N t be the distribution induced by µ and t. Then exactly as in Lemma 8.8 we have because of the conditional independence that IpN (X N ; Y N ) ≤ IpN (X n ; Y n ) + IpN (XnN−n ; YnN−n ) and hence from stationarity IpN (X N ; Y N ) ≤ Ipn (X n ; Y n ) + IpN−n (X N−n ; Y N−n ) ≤ nR + (N − n)R = NR so that p
N
N
∈ RN (R, µ ). Thus
DN (R, µ N ) ≤ EpN ρN (X N , Y N ) ≤ EpN ρn (X n , Y n ) + ρN−n (XnN−n , YnN−n ) = Epn ρn (X n , Y n ) + EpN−n ρN−n (X N−n , Y N−n ) ≤ Dn (R, µ n ) + DN−n (R, µ N−n ) + . Thus since is arbitrary we have shown that if dn = Dn (R, µ n ), then
9.2 Basic Properties
241
dN ≤ dn + dN−n ; n ≤ N; that is, the sequence dn is subadditive. The lemma then follows immediately from the convergence of subadditive functions (e.g., Lemma 7.5.1 of [55] or Lemma 8.5.3 of [58]). 2
IID Sources If the source is IID, then the evaluation of the distortion-rate function becomes particularly simple. Lemma 9.3. If a source µ is IID, then N −1 DN (R, µ N ) = D1 (R, µ 1 ) = D(R, µ) for all N. Proof. Suppose that the distribution p N for (X N , Y N ) approximately yields DN (R, µ N ), that is, p N has µ N as input marginal, I(p N ) ≤ NR, and h i EpN dN (X N , Y N ) ≤ DN (R, µ N ) + for small > 0. Let pi denote the induced distribution for (Xi , Yi ). Since X is IID, all the pi have µ as input marginal. From Lemma 8.8, since µ is IID N−1 N−1 X X I(Xi ; Yi ) = I(pi ) NR ≥ I(p N ) = I(X N ; Y N )) ≥ i=0
i=0
which implies with the convexity of D1 (R, µ) in R that h i 1 1 (DN (R, µ N ) + ) ≥ EpN dN (X N , Y N ) N N N−1 N−1 1 X 1 X Epi [d(Xi , Yi )] ≥ D1 (I(pi ), µ) = N i=0 N i=0 ≥ D1 (
N−1 1 X I(pi ), µ) ≥ D1 (R, µ), N i=0
where the final step used the fact that D1 (R, µ) is nonincreasing in its argument. This proves that DN (R, µ) ≥ ND1 (R, µ). If p 1 approximately achieves D1 (R, µ) in the sense that I(p 1 ) ≤ R and Ep1 [d(X0 , Y0 )] ≤ D1 (R, µ) + , then from Lemma 8.8I(p N ) = NI(pi ) ≤ NR and hence i h DN (R, µ) ≤ EpN ρN (X N , Y N ) = NEp1 [ρ(X0 , Y0 )] ≤ N[D1 (R, µ 1 ) + ] which implies that DN (R, µ) ≤ ND1 (R, µ 1 ) since is arbitrary. Thus DN (R, µ) = ND1 (R, µ 1 ) 2
242
9 Distortion and Information
The corresponding result is true for the rate-distortion function.
9.3 Process Definitions of the Distortion-Rate Function As with the ρ distance, there are alternative characterizations of the distortion-rate function when the process is stationary. The remainder of this section is devoted to developing these results. The idea of a stationarized block memoryless (SBM) channel will play an important role in relating nth order distortion-rate functions to the process definitions. We henceforth assume that the input source µ is stationary and we confine interest to additive fidelity criteria based on a per-letter distortion ρ = ρ1 . The basic process DRF is defined by D s (R, µ) =
inf p∈Rs (R,µ)
Ep ρ(X0 , Y0 ),
where Rs (R, µ) is the collection of all stationary processes p having µ as an input distribution and having mutual information rate I p = I p (X; Y ) ≤ R. The original idea of a process rate-distortion function was due to Kolmogorov and his colleagues [101] [49] (see also [23]). The idea was later elaborated by Marton [119], Gray, Neuhoff, and Omura [63], and Hashimoto [76]. Recalling that the L1 ergodic theorem for information density holds when I p = Ip∗ ; that is, the two principal definitions of mutual information rate yield the same value, we also define the process DRF Ds∗ (R, µ) =
inf
p∈R∗ s (R,µ)
Ep ρ(X0 , Y0 ),
where R∗ s (R, µ) is the collection of all stationary processes p having µ as an input distribution, having mutual information rate I p ≤ R, and having I p = Ip∗ . If µ is both stationary and ergodic, define the corresponding ergodic process DRF’s by D e (R, µ) = De∗ (R, µ) =
inf p∈Re (R,µ)
inf
p∈R∗ e (R,µ)
Ep ρ(X0 , Y0 ), Ep ρ(X0 , Y0 ),
where Re (R, µ) is the subset of Rs (R, µ) containing only ergodic mea∗ sures and R∗ e (R, µ) is the subset of Rs (R, µ) containing only ergodic measures. Theorem 9.1. Given a stationary source which possesses a reference letter ˆ such that in the sense that there exists a letter a∗ ∈ A
9.3 Process Definitions of the Distortion-Rate Function
243
Eµ ρ(X0 , a∗ ) ≤ ρ ∗ < ∞.
(9.3)
Fix R > 0. If D(R, µ) < ∞, then D(R, µ) = D s (R, µ) = Ds∗ (R, µ). If in addition µ is ergodic, then also D(R, µ) = D e (R, µ) = De∗ (R, µ). The proof of the theorem depends strongly on the relations among distortion and mutual information for vectors and for SBM channels. These are stated and proved in the following lemma, the proof of which is straightforward but somewhat tedious. The theorem is proved after the lemma. Lemma 9.4. Let µ be the process distribution of a stationary source {Xn }. Let ρn ; n = 1, 2, · · · be a subadditive (e.g., additive) fidelity criterion. ˆ for which (9.3) holds. Let Suppose that there is a reference letter a∗ ∈ A N ˆN , BN p N be a measure on (AN × A × B ) having µ N as input marginal; A ˆ A ˆN ) = µ N (F ) for F ∈ BN that is, p N (F × A A . Let q denote the induced conditional probability measure; that is, qx N (F ), x N ∈ AN , F ∈ BN ˆ , is a regular A conditional probability measure. (This exists because the spaces are standard.) We abbreviate this relationship as p N = µ N q. Let X N , Y N denote ˆN and suppose that the coordinate functions on AN × A EpN and
1 ρN (X N , Y N ) ≤ D N
1 IpN (X N ; Y N ) ≤ R. N
(9.4)
(9.5)
If ν is an (N, δ) SBM channel induced by q as in Section 2.14 and if p = µν is the resulting hookup and {Xn , Yn } the input/output pair process, then 1 Ep ρN (X N , Y N ) ≤ D + ρ ∗ δ N
(9.6)
I p (X; Y ) = Ip∗ (X; Y ) ≤ R;
(9.7)
and that is, the resulting mutual information rate of the induced stationary process satisfies the same inequality as the vector mutual information and the resulting distortion approximately satisfies the vector inequality provided δ is sufficiently small. Observe that if the fidelity criterion is additive, the (9.6) becomes Ep ρ1 (X0 , Y0 ) ≤ D + ρ ∗ δ.
244
9 Distortion and Information
Proof: We first consider the distortion as it is easier to handle. Since the SBM channel is stationary and the source is stationary, the hookup p is stationary and Z 1 1 Ep ρn (X n , Y n ) = dmZ (z)Epz ρn (X n , Y n ), n n where pz is the conditional distribution of {Xn , Yn } given {Zn }, the punctuation process of Section 2.14. Note that the above formula reduces to Ep ρ(X0 , Y0 ) if the fidelity criterion is additive because of the stationarity. Given z, define J0n (z) to be the collection of indices of zn for which zi is not in an N-cell. (See the discussion in Section 2.14.) Let J1n (z) be the collection of indices for which zi begins an N-cell. If we define the event G = {z : z0 begins an N − cell}, then i ∈ J1n (z) if T i z ∈ G. From Corollary 2.2 mZ (G) ≤ N −1 . Since µ is stationary and {Xn } and {Zn } are mutually independent, X X nEpz ρn (X n , Y n ) ≤ Epz ρ(Xi , a∗ ) + N Epz ρ(XiN , YiN ) i∈J0n (z)
=
n−1 X
i∈J1n (z)
1Gc (T i z)ρ ∗ +
i=0
n−1 X
EpN ρN 1G (T i z).
i=0
Since mZ is stationary, integrating the above we have that Ep ρ1 (X0 , Y0 ) = ρ ∗ mZ (Gc ) + NmZ (G)EpN ρN ≤ ρ ∗ δ + EpN ρN , proving (9.6). ˆ Let rm and tm denote asymptotically accurate quantizers on A and A; that is, as in Corollary 8.2 define ˆn = rm (X)n = (rm (X0 ), · · · , rm (Xn−1 )) X ˆ n = tm (Y )n . Then and similarly define Y I(rm (X)n ; tm (Y )n ) → I(X n ; Y n ) m→∞
and I(rm (X); tm (Y )) → I ∗ (X; Y ). m→∞
We wish to prove that 1 I(rm (X)n ; tm (Y )n ) n 1 = lim lim I(rm (X)n ; tm (Y )n ) m→∞ n→∞ n = I ∗ (X; Y )
I(X; Y ) = lim lim
n→∞ m→∞
9.3 Process Definitions of the Distortion-Rate Function
245
Since I ≥ I ∗ , we must show that lim lim
n→∞ m→∞
1 1 I(rm (X)n ; tm (Y )n ) ≤ lim lim I(rm (X)n ; tm (Y )n ). m→∞ n→∞ n n
We have that ˆ n ) = I((X ˆn , Z n ); Y ˆ n ) − I(Z n , Y ˆ n |X ˆn ) ˆn ; Y I(X and ˆ n ) = I(X ˆn ; Y ˆ n |Z n ) + I(Y ˆ n ; Z n ) = I(X ˆn ; Y ˆ n |Z n ) ˆn , Z n ); Y I((X ˆn and Z n are independent. Similarly, since X ˆ n |X ˆn ) = H(Z n |X ˆn ) − H(Z n |X ˆn , Y ˆn) I(Z n ; Y ˆn , Y ˆ n ) = I(Z n ; (X ˆn , Y ˆ n )). = H(Z n ) − H(Z n |X Thus we need to show that 1 1 I(rm (X)n ; tm (Y )n |Z n ) − I(Z n , (rm (X)n , tm (Y )n )) ≤ n→∞ m→∞ n n 1 1 n n n lim lim I(rm (X) ; tm (Y ) |Z ) − I(Z n , (rm (X)n , tm (Y )n )) . m→∞ n→∞ n n
lim lim
Since Zn has a finite alphabet, the limits of n−1 I(Z n , (rm (X)n , tm (Y )n )) are the same regardless of the order from Theorem 8.2. Thus I will equal I ∗ if we can show that 1 I(rm (X)n ; tm (Y )n |Z n ) n 1 ≤ lim lim I(rm (X)n ; tm (Y )n |Z n ) = I ∗ (X; Y |Z).. (9.8) m→∞ n→∞ n
I(X; Y |Z) = lim lim
n→∞ m→∞
This we now proceed to do. From Lemma 7.21 we can write Z I(rm (X)n ; tm (Y )n |Z n ) = I(rm (X)n ; tm (Y )n |Z n = zn ) dPZ n (zn ). ˆn ; Y ˆ n ). This is simply Abbreviate I(rm (X)n ; tm (Y )n |Z n = zn ) to Iz (X n n ˆ and Y ˆ under the distribution for the mutual information between X ˆn , Y ˆ n ) given a particular random blocking sequence z. We have that (X ˆn ; Y ˆ n ) = Hz (Y ˆ n ) − Hz (Y ˆ n |X ˆn ). Iz (X Given z, let J0n (z) be as before. Let J2n (z) denote the collection of all indices i of zi for which zi begins an N cell except for the final such index (which may begin an N-cell not completed within zn ). Thus J2n (z)
246
9 Distortion and Information
is the same as J1n (z) except that the largest index in the latter collection may have been removed if the resulting N-cell was not completed within the n-tuple. We have using standard entropy relations that
X
ˆn ; Y ˆn) ≥ Iz (X
ˆi |Y ˆ i ) − Hz (Y ˆi |Y ˆi, X ˆi+1 ) Hz (Y
i∈J0n (z)
X
+
ˆ N |Y ˆ i ) − Hz (Y ˆ N |Y ˆi, X ˆi+N ) . H z (Y i i
(9.9)
i∈J2n (z)
For i ∈ J0n (z), however, Yi is a∗ with probability one and hence ˆi |Y ˆ i ) ≤ Hz (Y ˆi ) ≤ Hz (Yi ) = 0 Hz (Y and ˆi |Y ˆi, X ˆi+1 ) ≤ Hz (Y ˆi ) ≤ Hz (Yi ) = 0. Hz (Y Thus we have the bound X ˆn ; Y ˆn) ≥ Iz (X
ˆ N |Y ˆ i ) − Hz (Y ˆ N |Y ˆi, X ˆi+N ) . H z (Y i i
i∈J2n (z)
=
X
ˆ N ; (Y ˆi, X ˆi + N)) − Iz (Y ˆN ; Y ˆi) I z (Y i i
i∈J2n (z)
≥
ˆN ; X ˆN ) − Iz (Y ˆN ; Y ˆi) , I z (Y i i i
X
(9.10)
i∈J2n (z)
where the last inequality follows from the fact that I(U; (V , W )) ≥ I(U ; V ). For i ∈ J2n (z) we have by construction and the stationarity of µ that ˆN ; Y ˆ N ) = IpN (X ˆN ; Y ˆ N ). Iz (X i i
(9.11)
As before let G = {z : z0 begins an N − cell}. Then i ∈ J2n (z) if T i z ∈ G and i < n − N and we can write 1 ˆn ; Y ˆn) ≥ Iz (X n 1 ˆN ; Y ˆN ) IpN (X n
n−N−1 X i=0
1G (T i z) −
1 n
n−N−1 X
ˆN ; Y ˆ i )1G (T i z). Iz (Y i
i=0
All of the above terms are measurable functions of z and are nonnegative. Hence they are integrable (although we do not yet know if the integral is finite) and we have that
9.3 Process Definitions of the Distortion-Rate Function
247
1 ˆn) ≥ ˆn ; Y I(X n ˆN ; Y ˆ N )mZ (G) Ipn (X
1 n−N − n n
n−N−1 X Z
ˆN ; Y ˆ i )1G (T i z). dmZ (z)Iz (Y i
i=0
To continue we use the fact that since the processes are stationary, we can consider it to be a two-sided process (if it is one-sided, we can imbed it in a two-sided process with the same probabilities on rectangles). By construction ˆN ; Y ˆ i ) = IT i z (Y ˆ0N ; (Y−i , · · · , Y−1 )) Iz (Y i and hence since mZ is stationary we can change variables to obtain n−N 1 ˆn ; Y ˆ n ) ≥ Ipn (X ˆN ; Y ˆ N )mZ (G) I(X n n Z n−N−1 X 1 ˆ0N ; (Y ˆ−i , · · · , Y ˆ−1 ))1G (z). dmZ (z)Iz (Y − n i=0 We obtain a further bound from the inequalities ˆ0N ; (Y ˆ−i , · · · , Y ˆ−1 )) ≤ Iz (Y0N ; (Y−i , · · · , Y−1 )) ≤ Iz (Y0N ; Y − ) Iz (Y where Y − = (· · · , Y−2 , Y−1 ). Since Iz (Y0N ; Y − ) is measurable and nonnegative, its integral is defined and hence Z 1 ˆ n |Z n ) ≥ Ipn (X ˆN ; Y ˆ N )mZ (G) − ˆn ; Y lim I(X dmZ (z)Iz (Y0N ; Y − ). n→∞ n G We can now take the limit as m → ∞ to obtain Z ∗ N N dmZ (z)Iz (Y0N ; Y − ). I (X; Y |Z) ≥ Ipn (X ; Y )mZ (G) −
(9.12)
This provides half of what we need. Analogous to (9.9) we have the upper bound X ˆn ; Y ˆn) ≤ ˆ N ; (Y ˆi, X ˆi+N )) − Iz (Y ˆN ; Y ˆi) . Iz (Y Iz (X i i
(9.13)
G
i∈J1n (z)
We note in passing that the use of J1 here assumes that we are dealing with a one-sided channel and hence there is no contribution to the information from any initial symbols not contained in the first N-cell. In the two-sided case time 0 could occur in the middle of an N-cell and one could fix the upper bound by adding the first index less than 0 for which zi begins an N-cell to the above sum. This term has no affect on the limits. Taking the limits as m → ∞ using Lemma 7.14 we have that
248
9 Distortion and Information
Iz (X n ; Y n ) ≤
X
Iz (YiN ; (Y i , X i+N )) − Iz (YiN ; Y i ) .
i∈J1n (z)
Given Z n = zn and i ∈ J1n (z), (X i , Y i ) → XiN → YiN forms a Markov chain because of the conditional independence and hence from Lemma 7.15 and Corollary 7.14 Iz (YiN , (Y i , X i+N )) = Iz (XiN ; YiN ) = IpN (X N ; Y N ). Thus we have the upper bound n−1 n−1 X 1 1 X 1 1G (T i z) − Iz (YiN ; Y i )1G (T i z). Iz (X n ; Y n ) ≤ IpN (X N ; Y N ) n n n i=0 i=0
Taking expectations and using stationarity as before we find that 1 I(X n ; Y n |Z n ) ≤ n n−1 Z 1 X dmZ (z)Iz (Y0N ; (Y−i , · · · , Y−1 )). IpN (X ; Y )mZ (G) − n i=0 G N
N
Taking the limit as n → ∞ using Lemma 7.22 yields Z I(X; Y |Z) ≤ IpN (X N ; Y N )mZ (G) − dmZ (z)Iz (Y0N ; Y − ).
(9.14)
G
Combining this with (9.12) proves that I(X; Y |Z) ≤ I ∗ (X; Y |Z) and hence that I(X; Y ) = I ∗ (X; Y ). It also proves that I(X; Y ) = I(X; Y |Z) − I(Z; (X, Y )) ≤ I(X; Y |Z) 1 ≤ IpN (X N ; Y N )mZ (G) ≤ IpN (X N ; Y N ) N using Corollary 2.2 to bound mX (G). This proves (9.7).
2
Proof of the theorem: We have immediately that ∗ R∗ e (R, µ) ⊂ Rs (R, µ) ⊂ Rs (R, µ)
and R∗ e (R, µ) ⊂ Re (R, µ) ⊂ Rs (R, µ), and hence we have for stationary sources that D s (R, µ) ≤ Ds∗ (R, µ) and for ergodic sources that
(9.15)
9.3 Process Definitions of the Distortion-Rate Function
249
D s (R, µ) ≤ Ds∗ (R, µ) ≤ De∗ (R, µ)
(9.16)
D s (R, µ) ≤ D e (R, µ) ≤ De∗ (R, µ).
(9.17)
D s (R, µ) ≥ D(R, µ).
(9.18)
and We next prove that If D s (R, µ) is infinite, the inequality is obvious. Otherwise fix > 0 and choose a p ∈ Rs (R, µ) for which Ep ρ1 (X0 , Y0 ) ≤ D s (R, µ) + and fix δ > 0 and choose m so large that for n ≥ m we have that n−1 Ip (X n ; Y n ) ≤ I p (X; Y ) + δ ≤ R + δ. For n ≥ m we therefore have that p n ∈ Rn (R + δ, µ n ) and hence D s (R, µ) + = Epn ρn ≥ Dn (R + δ, µ) ≥ D(R + δ, µ). From Lemma 9.1 D(R, µ) is continuous in R and hence (9.18) is proved. Lastly, fix > 0 and choose N so large and p N ∈ RN (R, µ N ) so that EpN ρN ≤ DN (R, µ N ) +
2 ≤ D(R, µ) + . 3 3
Construct the corresponding (N, δ)-SBM channel as in Section 2.14 with δ small enough to ensure that δρ ∗ ≤ /3. Then from Lemma 9.2 we have that the resulting hookup p is stationary and that I p = Ip∗ ≤ R and hence p ∈ R∗ s (R, µ) ⊂ Rs (R, µ). Furthermore, if µ is ergodic then so is p and hence p ∈ R∗ e (R, µ) ⊂ Re (R, µ). From Lemma 9.2 the resulting distortion is Ep ρ1 (X0 , Y0 ) ≤ EpN ρN + ρ ∗ δ ≤ D(R, µ) + . ∗ Since > 0 this implies the exisitence of a p ∈ R∗ s (R, µ) (p ∈ Re (R, µ) if µ is ergodic) yielding Ep ρ1 (X0 , Y0 ) arbitrarily close to D(R, µ. Thus for any stationary source Ds∗ (R, µ) ≤ D(R, µ) and for any ergodic source De∗ (R, µ) ≤ D(R, µ). With (9.15)–(9.18) this completes the proof. 2
The previous lemma is technical but important in proving source coding theorems. It permits the construction of a stationary and ergodic pair process having rate and distortion near that of that for a finite dimensional vector described by the original source and a finite-dimensional conditional probability.
250
9 Distortion and Information
9.4 The Distortion-Rate Function as a Lower Bound The Shannon distortion-rate function provides a simple lower bound to the performance in source coding systems and in constrained rate simulation systems. Both results will follow from from a simple inequality which we now develop. Suppose that X is a stationary source with process distribution µ. Suppose also that X is encoded and then decoded in order to obtain ˆ but we do not know the details of the code a reproduction process X, ˆ with distribution structures except that the resulting pair process (X, X) ˆ ˆ ≤ R. Let p is AMS and the X process has a finite entropy rate H(X) ˆ η denote the distribution of X. Let p denote the stationary mean with marginals µ = µ, since µ is stationary, and η. From Theorem 4.1, ˆ = H(η) = H(η). H(X)
(9.19)
For example, in a source coding system the codes might be stationary codes or block stationary codes, or block codes, or possibly even variable length codes, but the cascade of the operations must be AMS and yield a finite entropy rate reproduction. If, for example, there is a common finite alphabet for the output of the encoder and input to the decoder with 2R letters, then since the decoder can not increase entropy rate, it must have entropy rate no greater than R. In the constrained simulation problem, ˆ as a coding of an IID process Z such the goal is to produce a process X ˆ ≤ R and the process X ˆ is as close as possible to X with respect that H(X) to the ρ distance. The simulation problem was earlier formulated for stationary coding, but for the moment we allow other coding structures ˆ We assume an additive fidelity provided they yield an AMS process X. criterion for which the single letter distortion is integrable with respect to the stationary mean (so that the fidelity criterion is convergent in the sense of Chapter 5). In this case the limiting distortion is ρ∞ = lim sup n→∞
n−1 n−1 1 X 1 X ˆi ) = lim ˆi ) ρ1 (Xi , X ρ1 (Xi , X n→∞ n n i=0 i=0
where the limit exists p-a.e. and p-a.e. and ˆ0 ) = ∆(p) ∆(p) = Ep ρ∞ = Ep ρ1 (X0 , X I ∗ (p) = I ∗ (p). From Lemma 8.4, I ∗ (p) ≤ H(η). Putting all of this together, we have that
9.4 The Distortion-Rate Function as a Lower Bound
251
ˆ0 ) ∆(p) = ∆(p) = Ep ρ1 (X0 , X I ∗ (p) = I ∗ (p) ≤ R and hence that ∆(p) ≥ Ds (µ, R).
(9.20)
The equation boils down to simply this: given any AMS pair process with a given performance with respect to a fidelity criterion and an entropy rate constraint on one component, then the average distortion can be no smaller that the stationary process distortion-rate function for the given constraint because the entropy rate of one component process overbounds the mutual information rate between the two components. Corollary 9.1. As in Section 5.4, consider a stationary source µ, a channel ν, and code classes E, D for which if f ∈ E, g ∈ D, then the pair process pX,Xˆ consisting of the input and output of the cascade µf νg is AMS. Then the operational DRF of (5.14) is bound below by the stationary process Shannon DRF: ∆(µ, ν, E, D) =
inf
f ∈E,g∈D
∆(µ, f , ν, g) ≥ D s (µ, H(pXˆ )).
If there exists a reference letter and H(pXˆ ) < ∞, then also ∆(µ, ν, E, D) ≥ D(µ, R). For example, if the channel ν is noiseless with input alphabet A equal to the output alphabet, then H(pXˆ ) ≤ R = log kAk and the bound becomes δ(R, µ) = ∆(µ, ν, E, D) ≥ D s (R, µ),
(9.21)
which is one form of the classic converse source coding theorem since from Theorem 9.1 the stationary process definition and the Shannon definitions are equal under the assumed conditions, that is, D s (R, µ) = D(µ, R). Consider next the constrained rate simulation problem of Section 6.7 of the best coding of an IID process Z with entropy rate H(Z). Suppose ˆ is a process with distribution µXˆ for which that X ρ(µX , µXˆ ) ≤ ∆X|Z + for a small > 0, which implies that there is a coupling π with marginals µX and µXˆ with H(µXˆ ) ≤ H(Z) and distortion ∆(π ) ≤ ∆X|Z + which from the lemma implies that ∆(π ) ≤ Ds (H(Z), µ).
252
9 Distortion and Information
Since was arbitrary, we have the following corollary. Corollary 9.2. Given an IID process Z with entropy rate H(Z) and an additive fidelity criterion ρn and a stationary process X, then ∆X|Z ≡ inf ρ1 (µX , µf (Z) ) ≥ Ds (H(Z), µ). f
Thus the constrained rate problem also has the Shannon distortion-rate function as an unbeatable lower bound.
9.5 Evaluating the Rate-Distortion Function The goal of this section is to consider an alternative characterization of the optimization that defines the RDF given by R(D) = RX (D) = RN (D, µX ). The Shannon DRF and RDF are defined by informationtheoretic optimization problems. Since average distortion is linear in terms of the joint distribution describing the random vectors and mutual information is a convex ∪ function of the conditional probability distribution of output given input — called the test channel, the techniques of convex optimization provide an approach to evaluating the DRF or RDF for specific sources and distortion measures of interest. Shannon [163] provided the first examples of evaluation of the RDF for memoryless sources with Hamming and squared-error distortion. Kolmogorov [101] considered the case of Gaussian vectors and processes with respect to a squared error distortion. Gallager [47] provided general Kuhn-Tucker conditions providing a variational approach to finding the RDF (see also Berger [11]). A key aspect of the variational approach is that the optimization over the test channel or pair distribution with constrained input marginal leads to an optimization of a reproduction distribution, the distribution of the output marginal of the pair distribution or source/test-channel hookup. Blahut [18] found an alternative formulation of the optimization in terms of relative entropy or divergence and an iterative algorithm for numerical solution, and Csiszár [25] extended these results and provided an elegant and rigorous development for general alphabets. For the rest of this section, we consider the finite-order Shannon ratedistortion function (RDF) for vectors X = X N with distribution µX . We drop the superscripts for the dimension as it is assumed fixed. We will often drop the random variable subscript on a distribution if it is clear from context, so that µ = µX throughout. We pause to summarize the notational shortcuts for this section. We consider joint distributions πX,Y and pX,Y for a pair of random vectors (X, Y ) with alphabet (AX × AY , BAX × BAY )). Given a joint distribution πXY , denote the induced marginal distributions by πX and πY ,
9.5 Evaluating the Rate-Distortion Function
253
that is, πX (G) = πXY (G × AY ); G ∈ BAX πY (G) = πXY (AX × G); G ∈ BAY . The marginals pX and pY are similarly defined. If we are focusing only on joint distributions, the subscripts indicating the random variables will be dropped, that is, π = πX,Y and p = pX,Y . Throughout this section πX = µX = µ and pX = µX = µ, either by assumption or by demonstration. This constraint on the input distribution will be denoted by π ∈ P(µX ). We will use η to denote a distribution on the reproduction Y , but it is not fixed. It is used simply as shorthand for an output marginal, which might be induced or optimized over. In particular, given a reproduction distribution η, we will construct a special joint distribution p which will be denoted pη . This admittedly takes liberties with notation, but pη will mean a joint distribution constructed using a given reproduction distribution η. The construction will be such that pη need not have η as its marginal output distribution — η is simply used in the construction and the form of pη depends on η. We can express the RDF as R(D) =
inf
πXY :πXY ∈P(µX ),ρ(πXY )≤D
I(πXY )
(9.22)
where Z ρ(πXY ) = EπXY (ρ(X, Y )) = dπXY (x, y)ρ(x, y) Z πXY (x, y) I(πXY ) = I(X; Y ) = dπXY (x, y) log d(πX × πY )(x, y)
(9.23) (9.24)
For simplicity it is assumed that R(D) < ∞ for D > 0 and that R(D) → 0 as D → ∞. These assumptions reflect typical behavior and the details required for removing these simplifying assumptions may be found in Csiszár [25]. The most basic properties of the RDF parallel those for the DRF in Lemma 9.1 as summarized in the following lemma. The proof is omitted since it is a minor variation of the DRF case. Lemma 9.5. The Shannon RDF R(D) is a nonnegative convex nonincreasing function of D. Note that convexity implies that R(D) can not be a constant other than 0 over an interval of D given our assumption that R(D) must go to zero as D grows.
254
9 Distortion and Information
Lemma 9.6. An equivalent definition for the RDF is R(D) =
inf
πXY :πXY ∈P(µX ),ρ(πXY )=D
I(πXY );
(9.25)
that is, the inequality constraint can be replaced by an equality constraint. Proof. Suppose the contrary and hence there is a D0 < D such that we can do better than D, that is for any > 0 we can find a π with D(π ) < D0 and I(π ) ≤ R(D) + . But then I(π ) ≥ R(D0 ) and hence, since is arbitrary, R(D0 ) ≤ R(D). But R(D) is nonincreasing in D and hence R(D0 ) = R(D), which violates the convexity of R(D). 2 The constrained optimization over all distributions is traditionally handled as an unconstrained minimization over distributions by focusing on the function F (s) =
inf
πXY ∈P(µX )
(I(πXY ) + sρ(πXY )) ; s ≥ 0.
The function can be thought of as a variational or Lagrange multiplier formulation to remove the distortion constraint and incorporate it into the functional being minimized, but we will not use calculus to accomplish the minimization as was done in the original derivations. Instead we follow Csiszár’s [26] approach and use the divergence inequality repeatedly to find conditions for global optimality. We have easily that R(D) =
inf
πXY :πXY ∈P(µX ),ρ(πXY )≤D
inf
=
πXY :πXY ∈P(µX ),ρ(πXY )≤D
I(πXY )
I(πXY ) + sρ(πXY ) −sρ(πXY ) {z } | ≥−sD
inf
≥
πXY :πXY ∈P(µX ),ρ(πXY )≤D
≥
inf
πXY :πXY ∈P(µX )
(I(πXY ) + sρ(πXY )) − sD
(I(πXY ) + sρ(πXY )) − sD = F (s) − sD.
Thus for any fixed D, R(D) + sD ≥ F (s) for all s ≥ 0.
(9.26)
Consider a plot of R(d) with rate on the vertical axis (the y axis) and distortion on the horizontal axis (the x axis). If s is fixed and d allowed to vary over nonnegative numbers, F (s) − sd traces out a straight line y = a − sx of slope −s in the plot with vertical axis intercept a = F (s). If we fix a value of d = D, then for any value of s we have seen that it must be true that R(D) ≥ F (s) − sD. But R(D) is a convex function, and hence at any point (D, R)=(D, R(D)) lying on the R(D) curve there must exist a straight line passing through the point with no points above the R(D)
9.5 Evaluating the Rate-Distortion Function
255
curve. Hence given a fixed D > 0, there must exist a slope −sD (perhaps more than one) and a straight line y(x) = a − sD x with slope −sD which passes through the point (D, R) such that y(D) = a − sD D = R(D) so that the y axis intercept of the straight line is at a = R(D) + sD D. We say that such an s is associated with D. Since this is a tangent line and no points on the R(D) curve can lie below it, for any D 0 we must have that R(D 0 ) ≥ y(D 0 ) = a − sD D 0 = R(D) + sD D − sD D 0 or, for all D 0 R(D 0 ) + sD D 0 ≥ R(D) + sD D.
(9.27)
Suppose that π ∈ P(µX ) approximately yields F (sD ) so that for small > 0 I(π ) + sD ρ(π ) ≤ F (s) + . Then from (9.27) and (9.26) F (sD ) + ≥ R(ρ(π )) + sD ρ(π ) ≥ R(D) + sD D ≥ F (sD ),
(9.28)
which since can be made arbitrarily small implies F (sD ) = R(D) + sD D,
(9.29)
showing that the lower bound of (9.26) is achieved if sD is associated with D. Conversely, suppose that if instead of starting with D, we fix s, and π ∗ achieves a minimum in F (s), that is, F (s) = I(π ∗ ) + sρ(π ∗ ) =
inf
π ∈P(µX )
(I(π ) + sρ(π )) .
Define Ds = ρ(π ∗ ) and Rs = I(π ∗ ), then F (s) = Rs + sDs ≥
inf
I(π ) + sDs = R(Ds ) + sDs .
inf
(I(π ) + sρ(π ))
inf
(I(π ) + sDs )
inf
I(π ) + sDs
π ∈P(µX );ρ(π )≤Ds
It is also true that F (s) = ≤
π ∈P(µX );ρ(π )≤Ds π ∈P(µX );ρ(π )≤Ds
=
π ∈P(µX );ρ(π )≤Ds
= R(Ds ) + sDs and hence F (s) = R(Ds ) + sDs , which means that s is associated with Ds and R(Ds ) = F (s) = sDs . Summarizing the preceding development yields the following result. Lemma 9.7. R(D) = max(F (s) − sD). s≥0
256
9 Distortion and Information
The maximum is attained iff s is associated with D. If π achieves a minimum in F (s), then s is associated with D = ρ(π ) and R(D) = I(π ). The implication of the lemma is that the R(D) curve can be thought of as being parametrized by s, the slope of the tangent to points on the curve. For each value of s we try to minimize F (s) over all joint distributions π with marginal fixed by the source. If the minimizing distribution is found, its mutual information and average distortion yield a point on the rate-distortion curve. Thus the problem is to minimize F (s). This topic is tackled next. Recall from (7.28) the average mutual information can also be written as a divergence as I(πXY ) = D(πXY kπX × πY ), so the problem is to find F (s) =
inf
πXY ∈P(µX )
(D(πXY kπX × πY ) + sρ(πXY )) .
A useful approach both for the mathematics of the solution and for suggesting an algorithm for computing the solution was introduced by Blahut [18]. The math and the algorithm have interesting parallels with the optimality properties of actual codes to be considered later and they provide an early example of an alternating optimization (AO) algorithm, an optimization that alternates between two steps, each of which optimizes one component of the function being optimized for the other [14]. To set up the method we make a change in the target function by introducing another distribution. This apparent complication will lead to several useful results. Define the functional J(πXY , η, s) = D(πXY kπX × η) + sρ(πXY )
(9.30)
and note that it differs from the function being optimized in F (s) only by the replacement of the actual marginal πY in the divergence by a separate distribution η. Corollary 7.12 implies an immediate relation between the two: D(πXY kπX × πY ) = inf D(πXY kπX × η). (9.31) η
If D(πXY kπX × η) is finite, then πXY will be absolutely continuous with respect to the product measure πX × η. Given a distribution η, we construct a new joint distribution pXY as follows. If ρ(πXY ) is finite, then {x, y : ρ(x, y) < ∞} has positive probability and hence γη,s (x) ≡ R
1 < ∞, µX − a.e.. e−ρ(x,y) dη(y)
9.5 Evaluating the Rate-Distortion Function
257
Given a reproduction distribution η, define a joint distribution pη by its density or Radon-Nikodym derivative with respect to the product measure µX × η by dpη (x, y) = γη,s (x)e−sρ(x,y) . d(µX × η)(x, y) In other words, the distribution is specified by its values on rectangles as Z pη (F × G) = γη,s (x)e−sρ(x,y) d(µX × η)(x, y). F ×G
Intuitively, the density is constructed so that if the logarithm is taken, the result is the negative average distortion multiplied by s plus a normalization term. This will be shortly seen to be useful in expressing J(πXY , η, s) as a combination of simple terms. The X marginal of p = pη is easily found using Fubini’s theorem to be pX (F ) = pη (F × AY ) Z dpη (x, y) d(µX × η)(x, y) = F ×AY d(µX × η)(x, y) Z = γη,s (x)e−sρ(x,y) d(µX × η)(x, y) F ×AY Z Z = dµX (x) γη,s (x) e−sρ(x,y) dη(y) ZF dµX (x) = µX (F ), = F
the source distribution. The output distribution pY , however, is not easily found in general and, perhaps surprisingly, need not equal η. The introduction of the additional distribution η and the construction of the implied joint distribution pη allows the following representations of the functional J(πXY , η, s). Lemma 9.8. The functional J(πXY , η, s) = D(πXY kπX × η) + sρ(πXY ) can be expressed as J(πXY , η, s) = J(πXY , πY , s) + D(πY kη) Z = dµX (x) log γη,s (x) + D(πXY kpη ) = J(pη , η, s) + D(πXY kpη ).
(9.32) (9.33) (9.34)
Proof. Csiszár [26] observes in his Lemma 1.3 that the equalities follow from the chain rule for Radon-Nikodym derivatives. We provide more detail to add insight. As in (9.31), Corollary 7.12 with MX = PX = πX = µX yields (9.32) and (9.35) . The second equality is a result of rewriting J(πXY , η, s) by replacing the average distortion as an expectation involv-
258
9 Distortion and Information
ing the specially constructed joint distribution pη : J(πXY , η, s) = D(πXY kπX × η) + sρ(πXY ) Z Z dπXY = dπXY (x, y) log (x, y) − dπXY (x, y) log e−sρ(x,y) d(πX × η) Z dπXY (x, y) = dπXY (x, y) log d(πX × η) Z Z − dπXY (x, y) log γη,s (x)e−sρ(x,y) + dπXY (x, y) log γη,s (x) Z dπXY = dπXY (x, y) log (x, y) d(µX × η) Z Z dpη (x, y) + dµX (x) log γη,s (x) − dπXY (x, y) log d(µX × η) " ! !# Z dpη dπXY (x, y) / (x, y) = dπXY (x, y) log d(µX × η) d(µX × η) Z + dµX (x) log γη,s (x) Z Z dπXY = dπXY (x, y) log (x, y) + dµX (x) log γη,s (x) dpη Z = D(πXY kpη ) + dµX (x) log γη,s (x), which shows explicitly the Radon-Nikodym derivative chain rule application. We have that Z dµX (x) log γη,s (x) Z Z = dpη (x, y) log γη,s (x) = dpη (x, y) log γη,s (x)e−sρ(x,y) esρ(x,y) Z Z dpη = dpη (x, y) log + s dpη (x, y)ρ(x, y) d(µX × η) = D(pη kpX × η) + sρ(pη ) = J(pη , η, s),
2 which completes the proof. The representations of the lemma imply immediate lower bounds to J(πXY , η, s) with obvious conditions for equality, as summarized in the following corollary.
9.5 Evaluating the Rate-Distortion Function
259
Corollary 9.3. J(πXY , η, s) ≥ J(πXY , πY , s) with equality if η = πY Z J(πXY , η, s) ≥ dµX (x) log γη,s (x)
(9.35)
= J(pη , η, s) with equality if πXY = pη
(9.37)
F (s) =
inf J(π , η, s) Z = inf dµX (x) log γη,s (x) η Z 1 . = inf dµX (x) log R η dη(y)e−ρ(x,y) η,π :πX =µX
(9.36)
(9.38)
(9.39)
Proof. The inequalities and conditions for equality (9.35-9.37) follow directly from the lemma and the divergence inequality. Eq. (9.38) follows from the definition of F (s) since inf
π :πX =µX ,η
J(π , η, s) = =
inf
π :πX =µX ,η
inf
π :πX =µX
[D(π kµX × η) + sρ(π )]
[D(π kπX × πY ) + sρ(π )] = F (s)
and (9.39) follows since if we choose a reproduction distribution η within of the infimum, then using pη yields J(pη , η, s) within of the infimum, and hence F (s) can be no farther than from the infimum. Since is arbitrary, F (s) must equal the infimum.
2 The corollary suggests a numerical algorithm for evaluating the rate distortion function. Given the input distribution µX , pick some reproduction distribution η(0) . This η(0) together with µX implies a joint distribution p (0) = pη(0) with input marginal µX resulting in J(pη(0) , η(0) , s). (0)
Replace η(0) by η(1) = pY , which yields J(pη(0) , η(1) , s) ≤ J(pη(0) , η(0) , s), that is, J can not increase. Then use the new reproduction marginal η(1) to form a new joint distribution p(1) , which results in J(p(1) , η(1) , s) ≤ J(pη(0) , η(1) , s). Continue in this matter, alternatively picking the best joint distribution for the reproduction and vice versa. Since J is monotonically nonincreasing and nonnegative, this is a descent algorithm and hence it must converge. This is the idea behind Blahut’s algorithm [18] for computing the rate-distortion function. Blahut discretizes the problem by quantizing the input and output spaces to make the algorithm amenable to numerical solution. As discused by Rose [158], the algorithm can be sensitive to the nature of the discretization. In particular, a fixed quantization of source and reproduction can yield a suboptimal support for the reproduction distribution The corollary shows that F (s) can be stated as a optimization over the reproduction distribution as in (9.39). If an optimal reproduction
260
9 Distortion and Information
distribution η∗ exists, then from Corollary 9.3 it must be true that the ∗ optimal joint distribution is πX,Y with ∗ = pη∗ πX,Y
(9.40)
πY∗
(9.41)
∗
η =
since otherwise either D(πXY kpη∗ ) or D(πY kη∗ ) would be nonzero and hence J(πX,Y , η, s) could be further decreased towards its infimum by substituting the appropriate joint or reproduction distribution. If the optimal reproduction distribution exists, it is called the Shannon optimal reproduction distribution. If these optimal distributions exist, then together they induce a regular conditional probability measure P (X ∈ F | Y = y) given by Z P (X ∈ F | Y = y) = dµ(x)γη∗ ,s (x)e−sρ(x,y) , F
so that γη∗ ,s (x)e−sρ(x,y) has the interpretation of being the backward test channel of the input given the output. The following theorem summarizes the results developed in this section. It comprises a combination of Lemma 1.2, corollary to Lemma 1.3, and equations (1.11) and (1.15) in Csiszár [25]. Theorem 9.2. If R(D) < ∞, then R(D) = max (F (s) − sD) s≥0
F (s) =
inf (I(π ) + sd(π )) Z 1 = inf dµX (x) log R µY dµY (y)e−sd(x,y) π ∈P(µX )
(9.42) (9.43) (9.44)
where the final line defines F (s) as an infimum over all distributions ˆ There exists a value s such that the straight line of slope −s on A. is tangent to the rate-distortion curve at (R(D), D), in which case s is said to be associated with D. If π achieves a minimum in (9.43), then D = d(π ), R(D) = I(π ). Thus for a given D there is a value of s associated with D, and for this value the evaluation of the rate-distortion curve can be accomplished by an optimization over all distributions µY on the reproduction alphabet. If a minimizing π exists, then the resulting marginal distribution for µY is called a Shannon optimal reproduction distribution. In general this distribution need not be unique. Csiszár [25] goes on to develop necessary and sufficient conditions for solutions to the optimizations defining F (s) and R(D), but the above results suffice for our purpose of demonstrated the role of the divergence
9.5 Evaluating the Rate-Distortion Function
261
inequality in the optimization, and sketching the basic ideas underlying numerical algorithms for computing the rate-distortion function and the properties of optimal distributions in the Shannon sense. Conditions for the existence of solutions and for their uniqueness are also developed in [25]. We here state without proof one such result which will be useful in the discussion of optimality properties of source codes. The result shows that under the assumptions of a distortion measure that is a power of a metric derived from a norm, there exists a π achieving the minimum of (9.2) and hence also a Shannon optimal reproduction distribution. Both the lemma and the subsequent corollary are implied by the proof of Csiszár’s Theorem 2.2 and the extension of the reproduction space from compact metric to Euclidean spaces discussed at the bottom of p. 66 of [25]. In the corollary, the roles of distortion and mutual information are interchanged to obtain the distortion-rate version of the result. Lemma 9.9. Given a random vector X with an alphabet A which is a finiteˆ= dimensional Euclidean space with norm kxk, a reproduction alphabet A r A, and a distortion measure d(x, y) = kx − yk , r > 0, then there exists a distribution π on A × A achieving the the minimum of (9.2). Hence a Shannon N-dimensional optimal reproduction distribution exists for the Nth order rate-distortion function. Corollary 9.4. Given the assumptions of the lemma, suppose that π (n) , ˆ with marginals µX and n = 1, 2, . . . is sequence of distributions on A × A µY (n) for which for n = 1, 2, . . . I(π (n) ) = I(X, Y (n) ) ≤ R, lim E[d(X, Y
n→∞
(n)
)] = DX (R).
(9.45) (9.46)
Then µY (n) has a subsequence that converges weakly to a Shannon optimal reproduction distribution. If the Shannon distribution is unique, then µY (n) converges weakly to it. The result is proved by showing that the stated conditions imply that any sequence of distributions π (n) has a weakly converging subsequence and that the limiting distribution inherits the properties of the individual π (n) . If the Shannon optimal distribution is unique, then we can assume (n) that µY0 converges weakly to it. Note that if there is a unique Shannon optimal reproduction distribution, then any sequence of π (n) for which (9.45–9.46) hold must converge weakly to the optimal distribution.
262
9 Distortion and Information
Support of Shannon Optimal Distributions We close this chapter with a discussion of some of the interesting aspects of the Shannon optimal distribution. It is rare that an analytical formula is known for the distribution, one of the notable exceptions being a Gaussian IID source with variance σ 2 . In this case the N-dimensional Shannon optimal reproduction distributions are known to be the product of N Gaussian distributions with variance σ 2 − D. In particular, the reproduction distribution is continuous. This turns out to be an exception, a fact which has had an effect on the evaluation of rate-distortion functions in the past. Some of the history and issues are discussed here. The discussion follows that of [117]. The basic ideas behind Blahut’s algorithm were described in Section 9.5. The algorithm works quite well for discrete sources, but historically it has been applied to continuous sources in a way that often provided incorrect or misleading results. In particular, the standard approach in the literature was to first quantize the source and reproduction alphabet and then run the algorithm on the resulting discrete source. As pointed out by S. Fix [43], the reproduction alphabet chosen in this way was arbitrary and unchangeable by the algorithm itself. Fix proved that in the case of the squared error distortion measure, it is often the case that the optimal reproduction alphabet has finite support, that is, is concentrated on a specific finite set. If the initial quantization prior to the Blahut algorithm does not take this into account, the subsequent optimization can yield a poor solution to the original problem. This is not an uncommon problem since, as Fix showed, the optimal reproduction algorithm has finite support whenever a lower bound to the RDF due to Shannon [163], the Shannon lower bound, does not hold with equality. This occurs often for common sources and distortion measures. In fact the IID Gaussian source with a squared error distortion and IID discrete sources with a Hamming distortion are the only commonly encountered cases where the Shannon lower bound does hold with equality. A classic example of the problem is with the simple uniform IID source and a squared error distortion. Here the optimum reproduction alphabet is not only finite, but it can be small — only three letters for a rate of 1 bit per symbol. Inaccurate values for the RDF for this case based on Blahut’s algorithm have been reported in the literature. A similar problem arises with discrete sources if one is given the option finding an optimal reconstruction alphabet instead of assuming that it is the same as the input alphabet. Early work on this problem was considered by T. Benjamin [9]. In such cases the Blahut algorithm only adjusts the probabilities assigned to the assumed reproduction alphabet, it does not seek an optimum alphabet. As pointed out by Fix, the general optimization problem can be formulated, but it is a nonlinear optimization and no one approach is clearly best. K. Rose [158] extended Fix’s result showing finite support
9.5 Evaluating the Rate-Distortion Function
263
of the optimal reproduction distribution when the Shannon lower bound is not met. He developed a deterministic annealing algorithm with supporting arguments and impressive experimental evidence showing that the algorithm found the optimal reproduction alphabets and the best existing estimates of the RDFs for several examples of IID sources. Regrettably Fix’s fascinating work was never formally published outside of his dissertation, and the mathematical details are long and complicated. His arguments are based on Csiszár’s [25] careful development, which is partially developed in Section 9.5. Csiszár considered in depth the issues of the existence of solutions and the asymptotics and his paper is the definitive reference for the most general known results of this variety. The results of Fix and Rose, however, add an important aspect to the problem by pointing out that the choice of the support of the reproduction distribution must be considered if accurate results are to be obtained. This is important not only for the evaluation of the ratedistortion functions, but to the characterization of approximately optimal codes, as will be considered in Chapter 13.
Chapter 10
Relative Entropy Rates
Abstract Many of the basic properties of relative entropy are extended to sequences of random variables and to processes. Several limiting properties of entropy rates are proved and a mean ergodic theorem for relative entropy densities is given. The principal ergodic theorems for relative entropy and information densities in the general case are given in the next chapter.
10.1 Relative Entropy Densities and Rates Suppose that p and m are two AMS distributions for a random process {Xn } with a standard alphabet A. For convenience we assume that the random variables {Xn } are coordinate functions of an underlying measurable space (Ω, B) where Ω is a one-sided or two-sided sequence space and B is the corresponding σ -field. Thus x ∈ Ω has the form x = {xi }, where the index i runs from 0 to ∞ for a one-sided process and from −∞ to +∞ for a two-sided process. The random variables and vectors of principal interest are Xn (x) = xn , X n (x) = x n = (x0 , · · · , xn−1 ), and Xlk (x) = (xl , · · · , xl+k−1 ). The process distributions p and m are both probability measures on the measurable space (Ω, B). For n = 1, 2, . . . let MX n and PX n be the vector distributions induced by p and m. We assume throughout this section that MX n PX n and hence that the Radon-Nikodym derivatives fX n = dPX n /dMX n and the entropy densities hX n = ln fX n are well defined for all n = 1, 2, . . . Strictly speaking, for each n the random variable fX n is defined on the measurable space (An , BAn ) and hence fX n is defined on a different space for each n. When considering convergence of relative entropy densities, it is necessary to consider a sequence of random variables defined on a common measurable space, and hence two notational modifications are introduced: The random variables fX n (X n ) : Ω → [0, ∞) are defined by R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_10, © Springer Science+Business Media, LLC 2011
265
266
10 Relative Entropy Rates
fX n (X n )(x) ≡ fX n (X n (x)) = fX n (x n ) for n = 1, 2, . . .. Similarly the entropy densities can be defined on the common space (Ω, B) by hX n (X n ) = ln fX n (X n ). The reader is warned of the potentially confusing dual use of X n in this notation: the subscript is the name of the random variable X n and the argument is the random variable X n itself. To simplify notation somewhat, we will often abbreviate the previous (unconditional) densities to fn = fX n (X n ); hn = hX n (X n ). For n = 1, 2, . . . define the relative entropy by Hpkm (X n ) = D(PX n kMX n ) = EPX n hX n = Ep hX n (X n ). Define the relative entropy rate by H pkm (X) = lim sup n→∞
1 Hpkm (X n ). n
Analogous to Dobrushin’s definition of information rate, we also define ∗ Hpkm (X) = sup H pkm (q(X)), q
where the supremum is over all scalar quantizers q. Define as in Chapter 7 the conditional densities fXn |X n =
dPXn |X n fX n+1 dPX n+1 /dMX n+1 = = fX n dPX n /dMX n dMXn |X n
(10.1)
provided fX n 6= 0 and fXn |X n = 1 otherwise. As for unconditional densities we change the notation when we wish to emphasize that the densities can all be defined on a common underlying sequence space. For example, we follow the notation for ordinary conditional probability density functions and define the random variables fXn |X n (Xn |X n ) =
fX n+1 (X n+1 ) fX n (X n )
and hXn |X n (Xn |X n ) = ln fXn |X n (Xn |X n ) on (Ω, B). These densities will not have a simple abbreviation as do the unconditional densities. Define the conditional relative entropy
10.1 Relative Entropy Densities and Rates
267
Hpkm (Xn |X n ) = EPX n (ln fXn |X n ) =
Z
dp ln fXn |X n (Xn |X n ).
(10.2)
All of the above definitions are immediate applications of definitions of Chapter 7 to the random variables Xn and X n . The difference is that these are now defined for all samples of a random process, that is, for all n = 1, 2, . . .. The focus of this chapter is the interrelations of these entropy measures and on some of their limiting properties for large n. For convenience define Dn = Hpkm (Xn |X n ); n = 1, 2, . . . , and D0 = Hpkm (X0 ). From Theorem 7.2 this quantity is nonnegative and Dn + D(PX n kMX n ) = D(PX n+1 kMX n+1 ). If D(PX n kMX n ) < ∞, then also Dn = D(PX n+1 kMX n+1 ) − D(PX n kMX n ). We can write Dn as a single divergence if we define as in Theorem 7.2 the distribution SX n+1 by Z SX n+1 (F × G) = MXn |X n (F |x n ) dPX n (x n ); F ∈ BA ; G ∈ BAn . (10.3) F
Recall that SX n+1 combines the distribution PX n on X n with the conditional distribution MXn |X n giving the conditional probability under M for Xn given X n . We shall abbreviate this construction by SX n+1 = MXn |X n PX n .
(10.4)
Dn = D(PX n+1 kSX n+1 ).
(10.5)
Then Note that SX n+1 is not in general a consistent family of measures in the sense of the Kolmogorov extension theorem since its form changes with n, the first n samples being chosen according to p and the final sample being chosen using the conditional distribution induced by m given the first n samples. Thus, in particular, we cannot infer that there is a process distribution s which has SX n ; , n = 1, 2, . . . as its vector distributions. We immediately have a chain rule for densities fX n =
n−1 Y i=0
fXi |X i
(10.6)
268
10 Relative Entropy Rates
and a corresponding chain rule for conditional relative entropies similar to that for ordinary entropies: D(P
Xn
kM
Xn
n
) = Hpkm (X ) =
n−1 X
i
Hpkm (Xi |X ) =
i=0
n−1 X
Di .
(10.7)
i=0
10.2 Markov Dominating Measures The evaluation of relative entropy simplifies for certain special cases and reduces to a mutual information when the dominating measure is a Markov approximation of the dominated measure. The following lemma is an extension to sequences of the results of Corollary 7.13 and Lemma 7.17. Theorem 10.1. Suppose that p is a process distribution for a standard alphabet random process {Xn } with induced vector distributions PX n ; n = 1, 2, . . .. Suppose also that there exists a process distribution m with induced vector distributions MX n such that (a)under m {Xn } is a k-step Markov source, that is, for all n ≥ k, X n−k → k Xn−k → Xn is a Markov chain or, equivalently, MXn |X n = MXn |X k , n−k
and (b)MX n PX n , n = 1, 2, . . . so that the densities fX n =
dPX n dMX n
are well defined. Suppose also that p (k) is the k-step Markov approximation to p, that is, (k) the source with induced vector distributions PX n such that (k)
PX k = PX k and for all n ≥ k (k)
PXn |X n = PXn |X k ; n−k
(k)
that is, p is a k-step Markov process having the same initial distribution and the same kth order conditional probabilities as p. Then for all n ≥ k (k)
MX n PX n PX n
(10.8)
10.2 Markov Dominating Measures
and
269
n−1 (k) Y dPX n (k) = fX n ≡ fX k fXl |X k , l−k dMX n l=k
dPX n (k) dPX n
=
fX n (k)
fX n
(10.9)
.
(10.10)
Furthermore hXn |X n = hXn |X k
n−k
+ iXn ;X n−k |X k
(10.11)
n−k
and hence Dn = Hpkm (Xn |X n ) k k ) + Hpkm (Xn |Xn−k ). = Ip (Xn ; X n−k |Xn−k
Thus hX n = hX k +
n−1 X
hXl |X k + iXl ;X l−k |X k l−k
l=k
l−k
(10.12)
and hence D(PX n kMX n ) = Hpkm (X k ) +
n−1 X
k k (Ip (Xl ; X l−k |Xl−k ) + Hpkm (Xl |Xl−k )).
(10.13)
l=k
If m = p (k) , then for all n ≥ k we have that hXn |X k
n−k
= 0 and hence
k )=0 Hpkp(k) (Xn |Xn−k
(10.14)
k Dn = Ip (Xn ; X n−k |Xn−k ),
(10.15)
and and hence (k)
D(PX n kPX n ) =
n−1 X
k Ip (Xl ; X l−k |Xl−k ).
(10.16)
l=k
Proof: If n = k + 1, then the results follow from Corollary 7.9 and Lemma 7.17 with X = Xn , Z = X k , and Y = Xk . Now proceed by induction and assume that the results hold for n. Consider the distribution QX (n+1) specified by QX n = PX n and QXn |X n = PXn |X k . In other words, n−k
QX n+1 = PXn |X k PX n n−k
k Application of Corollary 7.7 withright-hand Z = X n−k , Y = Xn−k , and X = Xn implies that MX n+1 QX n+1 PX n+1 and that
270
10 Relative Entropy Rates
fXn |X n dPX n+1 = . dQX n+1 fXn |X k n−k
This means that we can write Z Z dPX n+1 dPX n+1 PX n+1 (F ) = dQX n+1 = dQXn |X n dQX n n+1 F dQX F dQX n+1 Z dPX n+1 = dPXn |X k dPX n . n−k F dQX n+1 From the induction hypothesis we can express this as Z dPX n+1 dPX n (k) dPXn |X k dPX n PX n+1 (F ) = (k) n−k F dQX n+1 dP n X Z dPX n+1 dPX n (k) = dPX n+1 , (k) F dQX n+1 dP n X (k)
proving that PX n+1 PX n+1 and that dPX n+1 (k) dPX n+1
=
fXn |X n dPX n dPX n+1 dPX n = . (k) dQX n+1 dPX(k) f k dPX n n Xn |Xn−k
This proves the right-hand part of (10.9) and (10.10). Next define the distribution PˆX n by Z (k) PˆX n (F ) = fX n dMX n , F
(k) (k) where fX n is defined in (10.9). Proving that PˆX n = PX n will prove both the left hand relation of (10.8) and (10.9). Clearly
ˆ Xn dP (k) = fX n dMX n and from the definition of f (k) and conditional densities (k)
(k) . k n |Xn−k
fXn |X n = fX
(10.17)
k → Xn is a Markov From Corollary 7.7 it follows that X n−k → Xn−k ˆ n chain. Since this is true for any n ≥ k, PX is the distribution of a k-step Markov process. By construction we also have that (k) k n |Xn−k
fX and hence from Theorem 7.2
= fXn |X k
n−k
(10.18)
10.2 Markov Dominating Measures (k) k n |Xn−k
PX
271
= PXn |X k . n−k
(k) (k) Since also fX k = fX k , PˆX n = PX n as claimed. This completes the proof of (10.8)–(10.10). Eq. (10.11) follows since
fXn |X n = fXn |X k
n−k
×
fXn |X n . fXn |X k n−k
Eq. (10.12) then follows by taking expectations. Eq. (10.12) follows from (10.11) and n−1 Y fX n = fX k fXl |X l , l=k
whence (10.13) follows by taking expectations. If m = p (k) , then the claims follow from (7.23)–(7.24). 2 Corollary 10.1. Given a stationary source p, suppose that for some K there exists a K-step Markov source m with distributions MX n PX n , n = 1, 2, . . .. Then for all k ≥ K (10.8)–(10.10) hold. Proof: If m is a K-step Markov source with the property MX n PX n , n = 1, 2, . . ., then it is also a k-step Markov source with this property for all k ≥ K. The corollary then follows from the theorem. 2 Comment: The corollary implies that if any K-step Markov source dominates p on its finite dimensional distributions, then for all k ≥ K the k-step Markov approximations p (k) also dominate p on its finite dimensional distributions. The following variational corollary follows from Theorem 10.1. Corollary 10.2. For a fixed k let Let Mk denote the set of all k-step Markov distributions. Then infM∈Mk D(PX n kM) is attained by P (k) , and (k)
inf D(PX n kM) = D(PX n kPX n ) =
M∈Mk
n−1 X
k Ip (Xl ; X l−k |Xl−k ).
l=k
Since the divergence can be thought of as a distance between probability distributions, the corollary justifies considering the k-step Markov process with the same kth order distributions as the k-step Markov approximation or model for the original process: It is the minimum divergence distribution meeting the k-step Markov requirement.
272
10 Relative Entropy Rates
10.3 Stationary Processes Several of the previous results simplify when the processes m and p are both stationary. We can consider the processes to be two-sided since given a stationary one-sided process, there is always a stationary twosided process with the same probabilities on all positive time events. n and f n satisfy When both processes are stationary, the densities fXm X n = fXm
n dPXm dPX n m = fX n T m = T , n dMXm dMX n
and have the same expectation for any integer m. Similarly the condin tional densities fXn |X n , fXk |Xk−n , and fX0 |X−1 ,X−2 ,··· ,X−n satisfy n fXn |X n = fXk |Xk−n T n−k = fX0 |X−1 ,X−2 ,··· ,X−n T n
(10.19)
for any k and have the same expectation. Thus n−1 1 X 1 Hpkm (X0 |X−1 , · · · , X−i ). Hpkm (X n ) = n n i=0
(10.20)
Using the construction of Theorem 7.2 we have also that Di = Hpkm (Xi |X i ) = Hpkm (X0 |X−1 , · · · , X− i) = D(PX0 ,X−1 ,··· ,X−i kSX0 ,X−1 ,··· ,X−i ), where now SX0 ,X−1 ,··· ,X−i = MX0 |X−1 ,··· ,X−i PX−1 ,··· ,X−i ;
(10.21)
that is, SX0 ,X−1 ,··· ,X−i (F × G) = Z MX0 |X−1 ,··· ,X−i (F |x i ) dPX−1 ,··· ,X−i (x i ); F ∈ BA ; G ∈ BAi . F
As before the SX n distributions are not in general consistent. For example, they can yield differing marginal distributions SX0 . As we saw in the finite case, general conclusions about the behavior of the limiting conditional relative entropies cannot be drawn for arbitrary reference measures. If, however, we assume as in the finite case that the reference measures are Markov, then we can proceed. Suppose now that under m the process is a k-step Markov process. k Then for any n ≥ k (X−n , · · · , X−k−2 , X−k−1 ) → X−k → X0 is a Markov chain under m and Lemma 7.17 implies that
10.3 Stationary Processes
273
Hpkm (X0 |X−1 , · · · , X−n ) = Hpkm (Xk |X k ) + Ip (Xk ; (X−1 , · · · , X−n )|X k )
(10.22)
and hence from (10.20) ¯pkm (X) = Hpkm (Xk |X k ) + Ip (Xk ; X − |X k ). H
(10.23)
We also have, however, that X − → X k → Xk is a Markov chain under m and hence a second application of Lemma 7.17 implies that Hpkm (X0 |X − ) = Hpkm (Xk |X k ) + Ip (Xk ; X − |X k ).
(10.24)
Putting these facts together and using (10.2) yields the following lemma. Lemma 10.1. Let {Xn } be a two-sided process with a standard alphabet and let p and m be stationary process distributions such that MX n PX n all n and m is kth order Markov. Then the relative entropy rate exists and 1 Hpkm (X n ) n = lim Hpkm (X0 |X−1 , · · · , X−n )
H pkm (X) = lim
n→∞ n→∞
= Hpkm (X0 |X − ) = Hpkm (Xk |X k ) + Ip (Xk ; X − |X k ) = Ep [ln fXk |X k (Xk |X k )] + Ip (Xk ; X − |X k ).
Corollary 10.3. Given the assumptions of Lemma 10.1, Hpkm (X N |X − ) = NHpkm (X0 |X − ). Proof: From the chain rule for conditional relative entropy (equation (10.7), n−1 X N − Hpkm (Xl |X l , X − ). Hpkm (X |X ) = l=0
Stationarity implies that each term in the sum equals Hpkm (X0 |X − ), proving the corollary. 2 The next corollary extends Corollary 10.1 to processes. Corollary 10.4. Given k and n ≥ k, let Mk denote the class of all k-step stationary Markov process distributions. Then inf H pkm (X) = H pkp(k) (X) = Ip (Xk ; X − |X k ).
m∈Mk
274
10 Relative Entropy Rates
Proof: Follows from (10.22) and Theorem 10.1.
2
This result gives an interpretation of the finite-gap information property (8.15): If a process has this property, then there exists a k-step Markov process which is only a finite “distance” from the given process in terms of limiting per-symbol divergence. If any such process has a finite distance, then the k-step Markov approximation also has a finite distance. Furthermore, we can apply Corollary 8.4 to obtain the generalization of the finite alphabet result of Theorem 3.4 . Corollary 10.5. Given a stationary process distribution p which satisfies the finite-gap information property, inf inf H pkm (X) = inf H pkp(k) (X) = lim H pkp(k) (X) = 0. k m∈Mk
k
k→∞
Lemma 10.1 also yields the following approximation lemma. Corollary 10.6. Given a process {Xn } with standard alphabet A let p and m be stationary measures such that PX n MX n for all n and m is kth order Markov. Let qk be an asymptotically accurate sequence of quantizers for A. Then H pkm (X) = lim H pkm (qk (X)), k→∞
that is, the divergence rate can be approximated arbitrarily closely by that of a quantized version of the process. Thus, in particular, ∗ H pkm (X) = Hpkm (X).
Proof: This follows from Corollary 7.3 by letting the generating σ -fields be Fn = σ (qn (Xi ); i = 0, −1, . . .) and the representation of conditional relative entropy as an ordinary divergence. 2 Another interesting property of relative entropy rates for stationary processes is that we can “reverse time” when computing the rate in the sense of the following lemma. Lemma 10.2. Let {Xn }, p, and m be as in Lemma 10.1. If either H pkm (X) < ∞ or HP kM (X0 |X − ) < ∞, then Hpkm (X0 |X−1 , · · · , X−n ) = Hpkm (X0 |X1 , · · · , Xn ) and hence Hpkm (X0 |X1 , X2 , · · · ) = Hpkm (X0 kX−1 , X−2 , · · · ) = H pkm (X) < ∞.
10.4 Mean Ergodic Theorems
275
Proof: If H pkm (X) is finite, then so must be the terms Hpkm (X n ) = D(PX n kMX n ) (since otherwise all such terms with larger n would also be infinite and hence H could not be finite). Thus from stationarity Hpkm (X0 |X−1 , · · · , X−n ) = Hpkm (Xn |X n ) = D(PX n+1 kMX n+1 ) − D(PX n kMX n ) D(PX n+1 kMX n+1 ) − D(P
X1n
kM
X1n
) = Hpkm (X0 |X1 , · · · , Xn )
from which the results follow. If on the other hand the conditional relative entropy is finite, the results then follow as in the proof of Lemma 10.1 using the fact that the joint relative entropies are arithmetic averages of the conditional relative entropies and that the conditional relative entropy is defined as the divergence between the P and S measures (Theorem 7.3). 2
10.4 Mean Ergodic Theorems In this section we state and prove some preliminary ergodic theorems for relative entropy densities analogous to those first developed for entropy densities in Chapter 4 and for information densities in Section 8.3. In particular, we show that an almost everywhere ergodic theorem for finite alphabet processes follows easily from the sample entropy ergodic theorem and that an approximation argument then yields an L1 ergodic theorem for stationary sources. The results involve little new and closely parallel those for mutual information densities and therefore the details are skimpy. The results are given for completeness and because the L1 results yield the byproduct that relative entropies are uniformly integrable, a fact which does not follow as easily for relative entropies as it did for entropies.
Finite Alphabets Suppose that we now have two process distributions p and m for a random process {Xn } with finite alphabet. Let PX n and MX n denote the induced nth order distributions and pX n and mX n the corresponding probability mass functions (pmf’s). For example, pX n (an ) = PX n ({x n : x n = an }) = p({x : X n (x) = an }). We assume that PX n MX n . In this case the relative entropy density is given simply by hn (x) = hX n (X n )(x) = ln
pX n (x n ) , mX n (x n )
276
10 Relative Entropy Rates
where x n = X n (x). The following lemma generalizes Theorem 4.1 from entropy densities to relative entropy densities for finite alphabet processes. Relative entropies are of more general interest than ordinary entropies because they generalize to continuous alphabets in a useful way while ordinary entropies do not. Lemma 10.3. Suppose that {Xn } is a finite alphabet process and that p and m are two process distributions with MX n PX n for all n, where p is ¯ m is a kth order Markov source with stationAMS with stationary mean p, ¯x } is the ergodic decomposition of the stationary ary transitions, and {p mean of p. Assume also that MX n P¯X n for all n. Then lim
n→∞
1 hn = h; p − a.e. and in L1 (p), n
where h(x) is the invariant function defined by h(x) = −H p¯x (X) − Ep¯x ln m(Xk |X k ) 1 = lim Hp¯x km (X n ) = H p¯x km (X), n→∞ n where m(Xk |X k )(x) ≡
mX k+1 (x k+1 ) = MXk |X k (xk |x k ). mX k (x k )
Furthermore, Ep h = H pkm (X) = lim
n→∞
1 Hpkm (X n ), n
(10.25)
that is, the relative entropy rate of an AMS process with respect to a Markov process with stationary transitions is given by the limit. Lastly, H pkm (X) = H pkm (X); ¯
(10.26)
that is, the relative entropy rate of the AMS process with respect to m is the same as that of its stationary mean with respect to m. Proof: We have that n−1 1 1 1 1 X k h(X n ) = ln p(X n ) − ln m(X k ) + ln m(Xi |Xi−k ) n n n n i=k
=
n−1 1 1 X 1 ln m(Xk |X k )T i−k , ln p(X n ) − ln m(X k ) − n n n i=k
where T is the shift transformation, p(X n ) is an abbreviation for PX n (X n ), and m(Xk |X k ) = MXk |X k (Xk |X k ). From Theorem 4.1 the first term converges to −H p¯x (X)p-a.e. and in L1 (p).
10.4 Mean Ergodic Theorems
277
Since MX k PX k , if MX k (F ) = 0, then also PX k (F ) = 0. Thus PX k and hence also p assign zero probability to the event that MX k (X k ) = 0. Thus with probability one under p, ln m(X k ) is finite and hence the second term in 10.27 converges to 0 p-a.e. as n → ∞. Define α as the minimum nonzero value of the conditional probability m(xk |x k ). Then with probability 1 under MX n and hence also under PX n we have that n−1 1 1 1 X ≤ ln ln k n i=k α m(Xi |Xi−k ) since otherwise the sequence X n would have 0 probability under MX n and hence also under PX n and 0 ln 0 is considered to be 0. Thus the rightmost term of (10.27) is uniformly integrable with respect to p and hence from Theorem 1.6 this term converges to Ep¯x (ln m(Xk |X k )). This proves the leftmost equality of (10.25). ¯X n |x denote the distribution Rof X n under the ergodic component Let p ¯x . Since MX n P¯X n and P¯X n = dp(x) ¯ ¯X n |x , if MX n (F ) = 0, then p p ¯X n |x (F ) = 0 p-a.e. Since the alphabet of Xn if finite, we therefore also p ¯ that MX n p ¯X n |x and hence have with probability one under p Hp¯x km (X n ) =
X
¯X n |x (an ) ln p
an
¯X n |x (an ) p MX n (an )
¯ is well defined for p-almost all x. This expectation can also be written as Hp¯x km (X n ) = −Hp¯x (X n ) − Ep¯x [ln m(X k ) +
n−1 X
ln m(Xk |X k )T i−k ]
i=k n
k
= −Hp¯x (X ) − Ep¯x [ln m(X )] − (n − k)Ep¯x [ln m(Xk |X k )], where we have used the stationarity of the ergodic components. Dividing by n and taking the limit as n → ∞, the middle term goes to zero as previously and the remaining limits prove the middle equality and hence the rightmost inequality in (10.25). Equation (10.25) follows from (10.25) and L1 (p) convergence, that is, since n−1 hn → h, we must also have that Ep (n−1hn(X n)) = n−1 Hpkm(X n ) converges to Ep h. Since the former limit is H pkm (X), (10.25) follows. ¯x is invariant (Theorem 1.5) and since expectations of invariant Since p functions are the same under an AMS measure and its stationary mean (Lemma 6.3.1 of [55] or Lemma 7.5 of [58]), application of the previous ¯ proves that results of the lemma to both p and p Z Z ¯ H pkm (X) = dp(x)H p¯x km (X) = dp(x)H (X), ¯x km (X) = H pkm ¯ p which proves (10.27) and completes the proof of the lemma.
2
278
10 Relative Entropy Rates
Corollary 10.7. Given p and m as in the Lemma, then the relative entropy rate of p with respect to m has an ergodic decomposition, that is, Z H pkm (X) = dp(x)H p¯x km (X). Proof: This follows immediately from (10.25) and (10.25).
2
Standard Alphabets We now drop the finite alphabet assumption and suppose that {Xn } is a standard alphabet process with process distributions p and m, where p is stationary, m is kth order Markov with stationary transitions, and MX n PX n are the induced vector distributions for n = 1, 2, . . . . Define the densities fn and entropy densities hn as previously. As an easy consequence of the development to this point, the ergodic decomposition for divergence rate of finite alphabet processes combined with the definition of H ∗ as a supremum over rates of quantized processes yields an extension of Corollary 8.2 to divergences. This yields other useful properties as summarized in the following corollary. Corollary 10.8. Given a standard alphabet process {Xn } suppose that p and m are two process distributions such that p is AMS and m is kth order Markov with stationary transitions and MX n PX n are the induced ¯ denote the stationary mean of p and let {p ¯x } vector distributions. Let p ¯ Then denote the ergodic decomposition of the stationary mean p. Z ∗ Hpkm (X) = dp(x)Hp∗ (10.27) ¯x km (X). In addition, ∗ ∗ (X) = Hpkm (X) = H pkm (X) = H pkm (X); Hpkm ¯ ¯
(10.28)
that is, the two definitions of relative entropy rate yield the same values for AMS p and stationary transition Markov m and both rates are the same as the corresponding rates for the stationary mean. Thus relative entropy rate has an ergodic decomposition in the sense that Z H pkm (X) = dp(x)H p¯x km (X). (10.29) Comment: Note that the extra technical conditions of Theorem 8.3 for equality of the analogous mutual information rates I¯ and I ∗ are not needed here. Note also that only the ergodic decomposition of the sta-
10.4 Mean Ergodic Theorems
279
¯ of the AMS measure p is considered and not that of the tionary mean p Markov source m. Proof: The first statement follows as previously described from the finite alphabet result and the definition of H ∗ . The left-most and right-most equalities of (10.28) both follow from the previous lemma. The middle equality of (10.28) follows from Corollary 10.4. Eq. (10.29) then follows from (10.27) and (10.28). 2 Theorem 10.2. Given a standard alphabet process {Xn } suppose that p and m are two process distributions such that p is AMS and m is kth order Markov with stationary transitions and MX n PX n are the induced vector ¯x } denote the ergodic decomposition of the stationary distributions. Let {p ¯ If mean p. 1 lim Hpkm (X n ) = H pkm (X) < ∞, n→∞ n then there is an invariant function h such that n−1 hn → h in L1 (p) as n → ∞. In fact, h(x) = H p¯x km (X), ¯x with respect to the relative entropy rate of the ergodic component p m. Thus, in particular, under the stated conditions the relative entropy densities hn are uniformly integrable with respect to p. Proof: The proof exactly parallels that of Theorem 8.1, the mean ergodic theorem for information densities, with the relative entropy densities replacing the mutual information densities. The density is approximated by that of a quantized version and the integral bounded above using the triangle inequality. One term goes to zero from the finite alphabet case. Since H = H ∗ (Corollary 10.8) the remaining terms go to zero because the relative entropy rate can be approximated arbitrarily closely by that of a quantized process. 2 It should be emphasized that although Theorem 10.2 and Theorem 8.1 are similar in appearance, neither result directly implies the other. It is true that mutual information can be considered as a special case of relative entropy, but given a pair process {Xn , Yn } we cannot in general find a kth order Markov distribution m for which the mutual information rate I¯(X; Y ) equals a relative entropy rate H pkm . We will later consider conditions under which convergence of relative entropy densities does imply convergence of information densities.
Chapter 11
Ergodic Theorems for Densities
Abstract This chapter is devoted to developing ergodic theorems first for relative entropy densities and then information densities for the general case of AMS processes with standard alphabets. The general results were first developed by Barron using the martingale convergence theorem and a new martingale inequality. The similar results of Algoet and Cover can be proved without direct recourse to martingale theory. They infer the result for the stationary Markov approximation and for the infinite order approximation from the ordinary ergodic theorem. They then demonstrate that the growth rate of the true density is asymptotically sandwiched between that for the kth order Markov approximation and the infinite order approximation and that no gap is left between these asymptotic upper and lower bounds in the limit as k → ∞. They use martingale theory to show that the values between which the limiting density is sandwiched are arbitrarily close to each other, but in this chapter it is shown that martingale theory is not needed and this property follows from the results of Chapter 8.
11.1 Stationary Ergodic Sources Theorem 11.1. Given a standard alphabet process {Xn }, suppose that p and m are two process distributions such that p is stationary ergodic and m is a K-step Markov source with stationary transition probabilities. Let MX n PX n be the vector distributions induced by p and m. As before let hn = ln fX n (X n ) = ln
dPX n (X n ). dMX n
Then with probability one under p
R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_11, © Springer Science+Business Media, LLC 2011
281
282
11 Ergodic Theorems for Densities
1 hn = H pkm (X). n
lim
n→∞
Proof: Let p (k) denote the k-step Markov approximation of p as defined in Theorem 10.1, that is, p (k) has the same kth order conditional probabilities and k-dimensional initial distribution. From Corollary 10.1, if k ≥ K, then (10.8)–(10.10) hold. Consider the expectation (k)
fX n (X n ) fX n (X n )
Ep
(k)
!
fX n fX n
= EPX n
!
Z =
(k)
fX n fX n
! dPX n .
Define the set An = {x n : fX n > 0}; then PX n (An ) = 1. Use the fact that fX n = dPX n /dMX n to write (k)
EP
fX n (X n ) fX n (X n )
!
(k)
fX n fX n
Z = An
!
Z f
Xn
dM
Xn
(k)
= An
fX n dMX n .
From Theorem 10.1, (k)
(k)
fX n =
dPX n dMX n
and therefore (k)
Ep
fX n (X n ) fX n (X n )
!
Z = An
(k)
dPX n (k) dMX n = PX n (An ) ≤ 1. dMX n (k)
Thus we can apply Lemma 7.13 to the sequence fX n (X n )/fX n (X n ) to conclude that with p-probability 1 (k)
f n (X n ) 1 ln X ≤0 n→∞ n fX n (X n ) lim
and hence lim
n→∞
1 1 (k) ln fX n (X n ) ≤ lim inf fX n (X n ). n→∞ n n
(11.1)
The left-hand limit is well-defined by the usual ergodic theorem: n−1 1 1 X 1 (k) k ln fXl |X k (Xl |Xl−k ) + lim ln fX n (X n ) = lim ln fX k (X k ). l−k n→∞ n n→∞ n n→∞ n l=k
lim
Since 0 < fX k < ∞ with probability 1 under MX k and hence also under PX k , then 0 < fX k (X k ) < ∞ under p and therefore n−1 ln fX k (X k ) → 0 as n → ∞ with probability one. Furthermore, from the pointwise ergodic theorem for stationary and ergodic processes (e.g., Theorem 7.2.1 of [55] or Theorem 8.1 of [58]), since p is stationary ergodic we have with
11.1 Stationary Ergodic Sources
283
probability one under p using (10.19) and Lemma 10.1 that n−1 1 X k ln fXl |X k (Xl |Xl−k ) l−k n→∞ n l=k
lim
n−1 1 X ln fX0 |X−1 ,··· ,X−k (X0 | X−1 , · · · , X−k )T l n→∞ n l=k
= lim
= Ep ln fX0 |X−1 ,··· ,X−k (X0 |X−1 , · · · , X−k ) = Hpkm (X0 |X−1 , · · · , X−k ) = H p(k) km (X). Thus with (11.1) it follows that lim inf n→∞
1 ln fX n (X n ) ≥ Hpkm (X0 |X−1 , · · · , X−k ) n
(11.2)
for any positive integer k. Since m is Kth order Markov, Lemma 10.1 and the above imply that lim inf n→∞
1 ln fX n (X n ) ≥ Hpkm (X0 |X − ) = H pkm (X), n
(11.3)
which completes half of the sandwich proof of the theorem. If H pkm (X) = ∞, the proof is completed with (11.3). Hence we can suppose that H pkm (X) < ∞. From Lemma 10.1 using the distribution SX0 ,X−1 ,X− 2,··· constructed there, we have that Z D(PX0 ,X−1 ,··· kSX0 ,X−1 ,··· ) = Hpkm (X0 |X − ) = dPX0 ,X − ln fX0 |X − where fX0 |X − =
dPX0 ,X−1 ,··· . dSX0 ,X−1 ,···
It should be pointed out that we have not (and will not) prove that fX0 |X−1 ,··· ,X− n → fX0 |X − ; the convergence of conditional probability densities which follows from the martingale convergence theorem and the result about which most generalized Shannon-McMillan-Breiman theorems are built. (See, e.g., Barron [8].) We have proved, however, that the expectations converge (Lemma 10.1), which is what is needed to make the sandwich argument work. For the second half of the sandwich proof we construct a measure Q which will be dominated by p on semi-infinite sequences using the above conditional densities given the infinite past. Define the semi-infinite sequence Xn− = {· · · , Xn−1 } n − − for all nonnegative integers n. Let Bn k = σ (Xk ) and Bk = σ (Xk ) = σ (· · · , Xk−1 ) be the σ -fields generated by the finite dimensional random
284
11 Ergodic Theorems for Densities
vector Xkn and the semi-infinite sequence Xk− , respectively. Let Q be the process distribution having the same restriction to σ (Xk− ) as does p and the same restriction to σ (X0 , X1 , · · · ) as does p, but which makes X − and Xkn conditionally independent given X k for any n; that is, QXk− = PXk− , QXk ,Xk+1 ,··· = PXk ,Xk+1 ,··· , −
k
and X → X →
Xkn
is a Markov chain for all positive integers n so that Q(Xkn ∈ F |Xk− ) = Q(Xkn ∈ F |X k ).
The measure Q is a (nonstationary) k-step Markov approximation to P in the sense of Section 7.2 and Q = PX − ×(Xk ,Xk+1 ,··· )|X k (in contrast to P = PX − X k Xk∞ ). Observe that X − → X k → Xkn is a Markov chain under both Q and m. By assumption, Hpkm (X0 |X − ) < ∞ and hence from Lemma 10.1 Hpkm (Xkn |Xk− ) = nHpkm (Xkn |Xk− ) < ∞ and hence from Theorem 7.3 the density fXkn |Xk− is well-defined as fXkn |Xk− =
− dSXn+k
(11.4)
− PXn+k
where − SXn+k = MXkn |X k PXk− ,
and Z − − − ln fXkn |Xk− = D(PXn+k kSXn+k ) dPXn+k
= nHpkm (Xkn |Xk− ) < ∞. Thus, in particular, − − SXn+k PXn+k .
Consider now the sequence of ratios of conditional densities ζn =
fXkn |X k (X n+k ) − fXkn |Xk− (Xn+k )
.
(11.5)
11.1 Stationary Ergodic Sources
We have that
285
Z
Z dpζn =
Gn
ζn
where Gn = {x : fXkn |Xk− (x− n+k ) > 0} since Gn has probability 1 under p (or else (11.6) would be violated). Thus Z Z fXkn |X k (X n+k ) − dpζn = dPXn+k 1{fX n |X − >0} k k fXkn |Xk− Z fXkn |X k (X n+k ) − = dSXn+k fXkn |Xk− 1{fX n |X − >0} k k fXkn |Xk− Z − = dSXn+k fXkn |X k (X n+k )1{fX n |X − >0} k k Z − ≤ dSXn+k fXkn |X k (X n+k ). Using the definition of the measure S and iterated expectation we have that Z Z dpζn ≤ dMXkn |Xk− dPXk− fXkn |X k (X n+k ) Z = dMXkn |X k dPXk− fXkn |X k (X n+k ). Since the integrand is now measurable with respect to σ (X n+k ), this reduces to Z Z dpζn ≤ dMXkn |X k dPX k fXkn |X k . Applying Lemma 7.10 we have Z
Z
dMXkn |X k dPX k
dpζn ≤
dPXkn |X k dMXkn |X k
Z = Thus
dPX k dPXkn |X k = 1. Z dpζn ≤ 1
and we can apply Lemma 7.12 to conclude that p-a.e. lim sup ζn = lim sup n→∞
n→∞
fX n |X k 1 ln k ≤ 0. n fXkn |Xk−
(11.6)
286
11 Ergodic Theorems for Densities
Using the chain rule for densities, fXkn |X k fXkn |Xk−
=
fX n 1 × Qn−1 . fX k l=k fXl |X − l
Thus from (11.6)
n−1 X 1 1 1 ln fXl |Xl− ≤ 0. lim sup ln fX n − ln fX k − n n n l=k n→∞ Invoking the ergodic theorem for the rightmost terms and the fact that the middle term converges to 0 almost everywhere since ln fX k is finite almost everywhere implies that lim sup n→∞
1 ln fX n ≤ Ep (ln fXk |Xk− ) = Ep (ln fX0 |X − ) = H pkm (X). n
(11.7)
Combining this with (11.3) completes the sandwich and proves the theorem. 2
11.2 Stationary Nonergodic Sources Next suppose that the source p is stationary with ergodic decomposition {pλ ; λ ∈ Λ} and ergodic component function ψ as in Theorem 1.6. We first require some technical details to ensure that the various RadonNikodym derivatives are well-defined and that the needed chain rules for densities hold. Lemma 11.1. Given a stationary source {Xn }, let {pλ ; λ ∈ Λ} denote the ergodic decomposition and ψ the ergodic component function of Theorem 1.6. Let Pψ denote the induced distribution of ψ. Let PX n and PXλn denote the induced marginal distributions of p and pλ . Assume that {Xn } has the finite-gap information property of (8.15); that is, there exists a K such that Ip (XK ; X − |X K ) < ∞, (11.8) where X − = (X−1 , X−2 , · · · ). We also assume that for some n I(X n ; ψ) < ∞.
(11.9)
This will be the case, for example, if (11.8) holds for K = 0. Let m be a K-step Markov process such that MX n PX n for all n. (Observe that such a process exists since from (11.8) the Kth order Markov approximation p (K) suffices.) Define MX n ,ψ = MX n × Pψ . Then
11.2 Stationary Nonergodic Sources
287
MX n ,ψ PX n × Pψ PX n ,ψ ,
(11.10)
and with probability 1 under p ψ
MX n PX n PX n . Lastly,
ψ
dPX n ,ψ dPX n = fX n |ψ = . dMX n d(MX n × Pψ ) and therefore
ψ
(11.11)
ψ
fX n |ψ dPX n /dMX n dPX n = = . n n n dPX dPX /dMX fX n
(11.12)
Proof: From Theorem 8.5 the given assumptions ensure that lim
n→∞
1 1 Ep i(X n ; ψ) = lim I(X n ; ψ) = 0 n→∞ n n
(11.13)
and hence PX n × Pψ PX n ,ψ (since otherwise I(X n ; ψ) would be infinite for some n and hence infinite for all larger n since it is increasing with n). This proves the right-most absolute continuity relation of (11.10). This in turn implies that MX n × Pψ PX n ,ψ . The lemma then follows from Theorem 7.2 with X = X n , Y = ψ and the chain rule for RadonNikodym derivatives. 2 We know that the source will produce with probability one an ergodic component pλ and hence Theorem 11.1 will hold for this ergodic component. In other words, we have for all λ that lim
n→∞
1 ln fX n |ψ (X n |λ) = H pλ (X); pλ − a.e. n
This implies that lim
n→∞
1 ln fX n |ψ (X n |ψ) = H pψ (X); p − a.e. n
(11.14)
Making this step precise generalizes Lemma 4.3. Lemma 11.2. Suppose that {Xn } is a stationary not necessarily ergodic source with ergodic component function ψ. Then (11.14) holds. Proof: The proof parallels that for Lemma 4.3. Observe that if we have two random variables U , V (U = X0 , X1 , · · · and Y = ψ above) and a sequence of functions gn (U, V ) (n−1 fX n |ψ (X n |ψ)) and a function g(V ) (H pψ (X)) with the property lim gn (U, v) = g(v), PU |V =v − a.e.,
n→∞
288
11 Ergodic Theorems for Densities
then also lim gn (U, V ) = g(V ); PUV − a.e.
n→∞
since defining the (measurable) set G = {u, v : limn→∞ gn (u, v) = g(v)} and its section Gv = {u : (u, v) ∈ G}, then from (1.28) Z PU V (G) = PU |V (Gv |v)dPV (v) = 1
2
if PU |V (Gv |v) = 1 with probability 1.
It is not, however, the relative entropy density using the distribution of the ergodic component that we wish to show converges. It is the original sample density fX n . The following lemma shows that the two sample entropies converge to the same thing. The lemma generalizes Lemma 4.3 and is proved by a sandwich argument analogous to Theorem 11.1. The result can be viewed as an almost everywhere version of (11.13). Theorem 11.2. Given a stationary source {Xn }, let {pλ ; λ ∈ Λ} denote the ergodic decomposition and ψ the ergodic component function of Theorem 1.6. Assume that the finite-gap information property (11.8) is satisfied and that (11.9) holds for some n. Then lim
n→∞
fX n |ψ 1 1 i(X n ; ψ) = lim ln = 0; p − a.e. n→∞ n n fX n
Proof: From Theorem 7.4 we have immediately that lim inf in (X n ; ψ) ≥ 0, n→∞
(11.15)
which provides half of the sandwich proof. To develop the other half of the sandwich, for each k ≥ K let p (k) denote the k-step Markov approximation of p. Exactly as in the proof of Theorem 11.1, it follows that (11.1) holds. Now, however, the Markov approximation relative entropy density converges instead as ∞ 1 1 X (k) fXk |X k (Xk |X k )T k = Epψ fXk |X k (Xk |X k ). ln fX n (X n ) = lim n→∞ n n→∞ n l=k
lim
Combining this with (11.14 we have that lim sup n→∞
fX n |ψ (X n |ψ) 1 ln ≤ H pψ km (X) − Epψ fXk |X k (Xk |X k ). n fX n (X n )
From Lemma 10.1, the right hand side is just Ipψ (Xk ; X − |X k ) which from Corollary 10.4 is just H pkp(k) (X). Since the bound holds for all k, we have that
11.2 Stationary Nonergodic Sources
lim sup n→∞
289
fX n |ψ (X n |ψ) 1 ln ≤ inf H pψ kp(k) (X) ≡ ζ. n fX n (X n ) k
Using the ergodic decompostion of relative entropy rate (Corollary 10.7) that and the fact that Markov approximations are asymptotically accurate (Corollary 10.5) we have further that Z Z dPψ ζ = dPψ inf H pψ kp(k) (X) k Z ≤ inf dPψ H pψ kp(k) (X) k
= inf H pk p(k) (X) = 0 k
and hence ζ = 0 with Pψ probability 1. Thus lim sup n→∞
fX n |ψ (X n |ψ) 1 ln ≤ 0, n fX n (X n )
which with (11.15) completes the sandwich proof.
(11.16)
2
Simply restating the theorem yields and using (11.14) the ergodic theorem for relative entropy densities in the general stationary case. Corollary 8.3.1: Given the assumptions of Theorem 11.2, lim
n→∞
1 ln fX n (X n ) = H pψ km (X), p − a.e. n
The corollary states that the sample relative entropy density of a process satisfying (11.8) converges to the conditional relative entropy rate with respect to the underlying ergodic component. This is a slight extension and elaboration of Barron’s result [8] which made the stronger assumption that Hpkm (X0 |X − ) = H pkm (X) < ∞. From Corollary 10.5 this condition is sufficient but not necessary for the finite-gap information property of (11.8). In particular, the finite gap information property implies that H pkp(k) (X) = Ip (Xk ; X − |X k ) < ∞, but it need not be true that H pkm (X) < ∞. In addition, Barron [8] and Algoet and Cover [7] do not characterize the limiting density as the entropy rate of the ergodic component, instead they effectively show that the limit is Epψ (ln fX0 |X − (X0 |X − )). This, however, is equivalent since it follows from the ergodic decomposition (see specifically Lemma 8.6.2 of [55] or Lemma 10.4 of [58]) that fX0 |X − = fX0 |X − ,ψ with probability one since the ergodic component ψ can be determined from the infinite past X −.
290
11 Ergodic Theorems for Densities
11.3 AMS Sources The following lemma is a generalization of Lemma 4.5. The result is due to Barron [8], who proved it using martingale inequalities and convergence results. Lemma 11.3. Let {Xn } be an AMS source with the property that for every integer k there exists an integer l = l(k) such that Ip (X k ; (Xk+l , Xk+l+1 , · · · )|Xkl ) < ∞. Then lim
n→∞
(11.17)
1 i(X k ; (Xk + l, · · · , Xn−1 )|Xkl ) = 0; p − a.e. n
Proof: By assumption Ip (X k ; (Xk+l , Xk+l+1 , · · · )|Xkl ) = Ep ln
fX k |Xk ,Xk+1 ,··· (X k |Xk , Xk+1 , · · · ) fX k |X l (X k |Xkl )
< ∞.
k
This implies that PX k ×(Xk +l,··· )|X l PX0 ,X1 ,··· k
with dPX0 ,X1 ,··· dPX k ×(Xk +l,··· )|X l
=
k
fX k |Xk ,Xk +1,··· (X k |Xk , Xk + 1, · · · ) fX k |X l (X k |Xkl ).
.
k
Restricting the measures to X n for n > k + l yields dPX n dPX k ×(Xk +l,··· ,X n )|X l
k
=
fX k |Xk ,Xk +1,··· ,Xn (X k |Xk , Xk + 1, · · · ) fX k |X l (X k |Xkl ) k
= i(X k ; (Xk + l, · · · , Xn )|Xkl ). With this setup the lemma follows immediately from Theorem 7.4.
2
The following lemma generalizes Lemma 4.6 and will yield the general theorem. The lemma was first proved by Barron [8] using martingale inequalities. Theorem 11.3. Suppose that p and m are distributions of a standard alphabet process {Xn } such that p is AMS and m is k-step Markov. Let p be a stationary measure that asymptotically dominates p (e.g., the stationary mean). Suppose that PX n , P X n , and MX n are the distributions induced by p, p, and m and that MX n dominates both PX n and P X n for all n and that fX n and f X n are the corresponding densities. If there is an invariant function h such that
11.3 AMS Sources
291
lim
1 ln f X n (X n ) = h; p − a.e. n
lim
1 ln fX n (X n ) = h; p − a.e. n
n→∞
then also n→∞
Proof: For any k and n ≥ k we can write using the chain rule for densities 1 1 1 ln fX n − ln fX n−k = ln fX k |X n−k . k k n n n Since for k ≤ l < n 1 1 1 ln fX k |X n−k = ln fX k |X l + i(X k ; (Xk+l , · · · , Xn−1 )|Xkl ), k k n n n Lemma 11.3 and the fact that densities are finite with probability one implies that 1 lim ln fX k |X n−k = 0; p − a.e. k n→∞ n This implies that there is a subsequence k(n) → ∞ such that 1 1 n−k(n) ln fX n (X n ) − ln fX n−k(n) ) (Xk(n) ); → 0, p − a.e. k(n) n n To prove this, for each k chose N(k) large enough so that p(|
1 N(k)−k ln fX k |X N(k)−k (X k |Xk )| > 2−k ) ≤ 2−k k N(k)
and then let k(n) = k for N(k) ≤ n < N(k + 1). Then from the BorelCantelli lemma we have for any that p(|
1 N(k)−k ln fX k |X N(k)−k (X k |Xk )| > i.o.) = 0 k N(k)
and hence 1 1 n−k(n) ln fX n (X n ) = lim ln fX n−k(n) (Xk(n) ); p − a.e. n→∞ n n→∞ n k(n) lim
In a similar manner we can also choose the sequence so that lim
n→∞
1 1 n−k(n) ln f X n (X n ) = lim ln f X n−k(n) (Xk(n) ); p − a.e. n→∞ k(n) n n
From Markov’s inequality
292
11 Ergodic Theorems for Densities
1 1 p ln fX n−k (Xkn−k ) ≥ ln f X n−k (Xkn−k ) + k k n n n−k fX n−k (Xk ) = p( k ≥ en ) f X n−k (Xkn−k )
k
≤e
−n
Z dp
fX n−k (Xkn−k ) k
f X n−k (Xkn−k )
=e
−n
Z
dmfX n−k (Xkn−k ) = e−n . k
k
Hence again invoking the Borel-Cantelli lemma we have that p(
1 1 ln fX n−k (Xkn−k ) ≥ ln f X n−k (Xkn−k ) + i.o.) = 0 k k n n
and therefore 1 ln fX n−k (Xkn−k ) ≤ h, p − a.e. (11.18) k n T The above event is in the tail σ -field n σ (Xn , Xn+1 , · · · ) since h is invariant and p dominates p on the tail σ -field. Thus lim sup n→∞
lim sup n→∞
1 n−k(n) ln fX n−k(n) (Xk(n) ) ≤ h; p − a.e. k(n) n
and hence lim sup n→∞
1 ln fX n (X n ) ≤ h; p − a.e. n
which proves half of the lemma. Since p asymptotically dominates p, given > 0 there is a k such that p( lim n−1 f (Xkn−k ) = h) ≥ 1 − . n→∞
Again applying Markov’s inequality and the Borel-Cantelli lemma as previously we have that n−k(n)
fX n−k(n) (Xk(n) ) 1 k(n) lim inf ln ≥ 0; p − a.e. n−k(n) n→∞ n f n−k(n) (X ) Xk(n)
k(n)
which implies that p(lim inf n→∞
1 f n−k(n) (Xkn−k ) ≥ h) ≥ n Xk(n)
and hence also that p(lim inf n→∞
1 fX n (X n ) ≥ h) ≥ . n
11.4 Ergodic Theorems for Information Densities.
293
Since can be made arbitrarily small, this proves that p-a.e. lim inf n −1 hn ≥ h, which completes the proof of the lemma. 2 We can now extend the ergodic theorem for relative entropy densities to the general AMS case. Corollary 8.4.1: Given the assumptions of Theorem 11.3, lim
n→∞
1 ln fX n (X n ) = H pψ (X), n
where p ψ is the ergodic component of the stationary mean p of p. Proof: The proof follows immediately from Theorem 11.3 and Lemma 11.1, the ergodic theorem for the relative entropy density for the stationary mean. 2
11.4 Ergodic Theorems for Information Densities. As an application of the general theorem we prove an ergodic theorem for mutual information densities for stationary and ergodic sources. The result can be extended to AMS sources in the same manner that the results of Section 11.2 were extended to those of Section 11.3. As the stationary and ergodic result suffices for the coding theorems and the AMS conditions are messy, only the stationary case is considered here. The result is due to Barron [8]. Theorem 11.4. Let {Xn , Yn } be a stationary ergodic pair random process with standard alphabet. Let PX n Y n , PX n , and PY n denote the induced distributions and assume that for all n PX n × PY n PX n Y n and hence the information densities in (X n ; Y n ) =
dPX n Y n d(PX n × PY n )
are well-defined. Assume in addition that both the {Xn } and {Yn } processes have the finite-gap information property of (11.8) and hence by the comment following Corollary 10.1 there is a K such that both processes satisfy the K-gap property I(XK ; X − |X K ) < ∞, I(YK ; Y − |Y K ) < ∞. Then lim
n→∞
1 in (X n ; Y n ) = I(X; Y ); p − a.e.. n (K)
(K)
Proof: Let Zn = (Xn , Yn ). Let MX n = PX n and MY n = PY n denote the Kth order Markov approximations of {Xn } and {Yn }, respectively. The finite-gap approximation implies as in Section 11.2 that the densities
294
11 Ergodic Theorems for Densities
fX n =
dPX n dPY n and fY n = dMX n dMY n
are well-defined. From Theorem 11.1 1 ln fX n (X n ) = Hp kp(K) (X0 |X − ) = I(Xk ; X − |X k ) < ∞, X X n 1 lim ln fY n (Y n ) = I(Yk ; Y − |Y k ) < ∞. n→∞ n lim
n→∞
Define the measures MZ n by MX n × MY n . Then this is a K-step Markov source and since MX n × MY n PX n × PY n PX n ,Y n = PZ n , the density fZ n =
dPZ n dMZ n
is well-defined and from Theorem 11.1 has a limit lim
n→∞
1 ln fZ n (Z n ) = Hpkm (Z0 |Z − ). n
If the density in (X n , Y n ) is infinite for any n, then it is infinite for all larger n and convergence is trivially to the infinite information rate. If it is finite, the chain rule for densities yields 1 1 1 1 in (X n ; Y n ) = ln fZ n (Z n ) − ln fX n (X n ) − ln fY n (Y n ) n n n n → Hpkp(k) (Z0 |Z − ) − Hpkp(k) (X0 |X − ) − Hpkp(k) (Y0 |Y − ) n→∞
= H pkp(k) (X, Y ) − H pkp(k) (X) − H pkp(k) (Y ). The limit is not indeterminate (of the form ∞ − ∞) because the two subtracted terms are finite. Since convergence is to a constant, the constant must also be the limit of the expected values of n−1 in (X n , Y n ), that is, I(X; Y ). 2
Chapter 12
Source Coding Theorems
Abstract The source coding theorems subject to a fidelity criterion are develped for AMS sources and additive and subadditive distortion measures. The results are first developed for the classic case of block coding and then to sliding-block codes. The operational distortion-rate function for both classes of codes is shown to equal the Shannon distortion-rate function.
12.1 Source Coding and Channel Coding In this chapter and in Chapter 14 we develop the basic coding theorems of information theory. As is traditional, we consider two important special cases first and then later form the overall result by combining these special cases. In the first case in this chapter we assume that the channel is noiseless, but it is constrained in the sense that it can only pass R bits per input symbol to the receiver. Since this is usually insufficient for the receiver to perfectly recover the source sequence, we attempt to code the source so that the receiver can recover it with as little distortion as possible. This leads to the theory of source coding or source coding subject to a fidelity criterion or data compression, where the latter name reflects the fact that sources with infinite or very large entropy are “compressed” to fit across the given communication link. In Chapter 14 we ignore the source and focus on a discrete alphabet channel and construct codes that can communicate any of a finite number of messages with small probability of error and we quantify how large the message set can be. This operation is called channel coding or error control coding. We then develop joint source and channel codes which combine source coding and channel coding so as to code a given source for communication over a given channel so as to minimize average distortion. The ad hoc division into two forms of coding is convenient and will permit performance near R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_12, © Springer Science+Business Media, LLC 2011
295
296
12 Source Coding Theorems
that of the operational distortion-rate function function for the singleuser or point-to-point communication systems and codes considered in this book. The Shannon coding theorems quantify the optimal performance that can be achieved when communicating a given source through a given channel, but they do not say how to actually achieve such optimal performance. The Shannon theorems are at heart existence theorems and not constructive. There is a huge literature on constructing channel codes for reliable communication and source codes for analog-to-digital conversion and data compression. Coding theory in the sense of a rigorous approach to designing good codes is not treated or even surveyed here, but there are a collection of results which use the information theoretic techniques developed in this book to provide necessary conditions that optimal or asymptotically optimal source codes must satisfy. Such conditions highlight implications of the underlying theory for the behavior of good codes and provide insight into the structure of good codes and thereby suggest design techniques that can improve existing codes. These results are developed in Chapter 13.
12.2 Block Source Codes for AMS Sources We first consider a particular class of codes: block codes. For the time being we also concentrate on additive distortion measures. Extensions to subadditive distortion measures will be considered later. Let {Xn } be a source with a standard alphabet A. Recall that an (N, K) block code of N a source {Xn } maps successive nonoverlapping input vectors {XnN } into N K successive channel vectors UnK = α(XnN ), where α : AN → B K is called the source encoder. We assume that the channel is noiseless, but that it is constrained in the sense that N source time units corresponds to the same amount of physical time as K channel time units and that K log ||B|| ≤ R, N where the inequality can be made arbitrarily close to equality by taking N and K large enough subject to the physical stationarity constraint. R is called the source coding rate or resolution in bits or nats per input symbol. We may wish to change the values of N and K, but the rate is fixed. A reproduction or approximation of the original source is obtained by a source decoder, which we also assume to be a block code. The decoder ˆN which forms the reproduction process {X ˆn } is a mapping α : B K → A N K ˆnN = α(UnK ); n = 1, 2, . . .. In general we could have a reproduction via X dimension different from that of the input vectors provided they corre-
12.2 Block Source Codes for AMS Sources
297
sponded to the same amount of physical time and a suitable distortion measure was defined. We will make the simplifying assumption that they are the same, however. Because N source symbols are mapped into N reproduction symbols, we will often refer to N alone as the block length of the source code. Observe that the resulting sequence coder is N-stationary. Our immediate goal is now the following: Let E and D denote the collection of all block codes with rate no greater than R and let ν be the given channel. What is the optimal achievable performance ∆(µ, E, ν, D) for this system? Our first step toward evaluating the operational DRF is to find a simpler and equivalent expression for the current special case. Given a source code consisting of encoder α and decoder β, define the codebook to be C = { all β(uK ); uK ∈ B K }, that is, the collection of all possible reproduction vectors available to the receiver. For convenience we can index these words as C = {yi ; i = 1, 2, . . . , M}, where N −1 log M ≤ R by construction. Observe that if we are given only a decoder β or, equivalently, a codebook, and if our goal is to minimize the average distortion for the current block, then no encoder can do better than the encoder α∗ which maps an input word x N into the minimum distortion available reproduction word, that is, define α∗ (x N ) to be the uK minimizing ρN (x N , β(uK )), an assignment we denote by α∗ (x N ) = argminρN (x N , β(uK )). uK
The fact that no encoder can yield smaller average distortion than a minimum distortion encoder is an example of an optimality property of block codes. Such properties are the subject of Chapter 13. Observe that by construction we therefore have that ρN (x N , β(α∗ (x N ))) = min ρN (x N , y) y∈C
and the overall mapping of x N into a reproduction is a minimum distortion or nearest neighbor mapping. Define ρN (x N , C) = min ρN (x N , y). y∈C
To prove that this is the best encoder, observe that if the source µ is AMS and p is the joint distribution of the source and reproduction, then p is also AMS. This follows since the channel induced by the block code is N-stationary and hence also AMS with respect to T N . This means that
298
12 Source Coding Theorems
p is AMS with respect to T N which in turn implies that it is AMS with respect to T (Theorem 7.3.1 of [55] or Theorem 8.2 of [58]). Letting p denote the stationary mean of p and p N denote the N-stationary mean, we then have from (5.12) that for any block codes with codebook C ∆=
1 1 Ep ρN (X N , Y N ) ≥ EpN ρN (X N , C), N N N
with equality if the minimum distortion encoder is used. For this reason we can confine interest for the moment to block codes specified by a codebook: the encoder produces the index of the minimum distortion codeword for the observed vector and the decoder is a table lookup producing the codeword being indexed. We will be interested later in looking at possibly nonoptimal encoders in order to decouple the encoder from the decoder and characterize the separate effects of encoder and decoder on performance. A block code of this type is also called a vector quantizer or block quantizer. Denote the performance of the block code with codebook C on the source µ by ρ(C, µ) = ∆ = Ep ρ∞ . Lemma 12.1. Given an AMS source µ and a block length N code book C, let µ N denote the N-stationary mean of µ (which exists from Corollary 7.3.1 of [55] or Corollary 8.5 of [58]), let p denote the induced input/output distribution, and let p and p N denote its stationary mean and N-stationary mean, respectively. Then ρ(C, µ) = Ep ρ1 (X0 , Y0 ) = =
1 Ep ρN (X N , Y N ) N N
1 Eµ ρN (X N , C) = ρ(C, µ N ). N N
Proof: The first two equalities follow from (5.12), the next from the use of the minimum distortion encoder, the last from the definition of the performance of a block code. 2 It need not be true in general that ρ(C, µ) equal ρ(C, µ). For example, if µ produces a single periodic waveform with period N and C consists of a single period, then ρ(C, µ) = 0 and ρ(C, µ) > 0. It is the N-stationary mean and not the stationary mean that is most useful for studying an N-stationary code. We now define the operational distortion-rate function (DRF) for block codes to be δ(R, µ) = ∆(µ, ν, E, D) = inf δN (R, µ), N
δN (R, µ) =
1
inf
C: N log ||C||≤R
ρ(C, µ),
(12.1) (12.2)
12.2 Block Source Codes for AMS Sources
299
where ν is the noiseless channel described earlier and E and D are classes of block codes for the channel. δ(R, µ) is called the operational block coding distortion-rate function (DRF). Corollary 12.1. Given an AMS source µ, then for any positive integer N δN (R, µT −i ) = δN (R, µ N T −i ); i = 0, 1, . . . , N − 1. Proof: For i = 0 the result is immediate from the lemma. For i 6= 0 it follows from the lemma and the fact that the N-stationary mean of µT −i is µ N T −i (as is easily verified from the definitions). 2
Reference Letters Many of the source coding results will require a technical condition that is a generalization of the reference letter condition of Theorem 9.1 for stationary sources. An AMS source µ is said to have a reference letter ˆ with respect to a distortion measure ρ = ρ1 on A × A ˆ if a∗ ∈ A sup EµT −n ρ(X0 , a∗ ) = sup Eµ ρ(Xn , a∗ ) = ρ ∗ < ∞, n
(12.3)
n
that is, there exists a letter for which Eµ ρ(X n , a∗ ) is uniformly bounded above. If we define for any k the vector a∗ k = (a∗ , a∗ , · · · , a∗ ) consisting of k a∗ ’s, then (12.3) implies that 1 k sup EµT −n ρk (X k , a∗ ) ≤ ρ ∗ < ∞. k n
(12.4)
We assume for convenience that any block code of length N contains the reference vector a∗ N . This ensures that ρN (x N , C) ≤ ρN (x N , a∗ N ) and hence that ρN (x N , C) is bounded above by a µ-integrable function and hence is itself µ-integrable. This implies that δ(R, µ) ≤ δN (R, µ) ≤ ρ ∗ .
(12.5)
The reference letter also works for the stationary mean source µ since n−1 1 X ρ(xi , a∗ ) = ρ∞ (x, a∗ ), n→∞ n i=0
lim
µ-a.e. and µ-a.e., where a∗ denotes an infinite sequence of a∗ . Since ρ∞ is invariant we have from Lemma 6.3.1 of [55] or Lemma 7.5 of [58] and Fatou’s lemma (Lemma 4.4.5 of [55] or Lemma 8.5 of [58]) that
300
12 Source Coding Theorems
n−1 X 1 Eµ ρ(X0 , a∗ ) = Eµ lim ρ(Xi , a∗ ) n→∞ n i=0 ≤ lim inf n→∞
n−1 1 X Eµ ρ(Xi , a∗ ) ≤ ρ ∗ . n i=0
Performance and Distortion-Rate Functions We next develop several basic properties of the performance and the operational DRFs for block coding AMS sources with additive fidelity criteria. Lemma 12.2. Given two sources µ1 and µ2 and λ ∈ (0, 1), then for any block code C ρ(C, λµ1 + (1 − λ)µ2 ) = λρ(C, µ1 ) + (1 − λ)ρ(C, µ2 ) and for any N δN (R, λµ1 + (1 − λ)µ2 ) ≥ λδN (R, µ1 ) + (1 − λ)δN (R, µ2 ) and δ(R, λµ1 + (1 − λ)µ2 ) ≥ λδ(R, µ1 ) + (1 − λ)δ(R, µ2 ). Thus performance is linear in the source and the operational DRFs are T convex in R. Lastly, δN (R +
1 , λµ1 + (1 − λ)µ2 ) ≤ λδN (R, µ1 ) + (1 − λ)δN (R, µ2 ). N
Proof: The equality follows from the linearity of expectation since ρ(C, µ) = Eµ ρ(X N , C). The first inequality follows from the equality and the fact that the infimum of a sum is bounded below by the sum of the infima. The next inequality follows similarly. To get the final inequality, let Ci approximately yield δN (R, µi ); that is, ρ(Ci , µi ) ≤ δN (R, µi ) + . S Form the union code C = C1 C2 containing all of the words in both of the codes. Then the rate of the code is
12.2 Block Source Codes for AMS Sources
301
1 1 log ||C|| = log(||C1 || + ||C2 ||) N N 1 ≤ log(2NR + 2NR ) N 1 =R+ . N This code yields performance ρ(C, λµ1 + (1 − λ)µ2 ) = λρ(C, µ1 ) + (1 − λ)ρ(C, µ2 ) ≤ λρ(C1 , µ1 ) + (1 − λ)ρ(C2 , µ2 ) ≤ λδN (R, µ1 ) + λ + (1 − λ)δN (R, µ2 ) + (1 − λ). Since the leftmost term in the above equation can be no smaller than δN (R + 1/N, λµ1 + (1 − λ)µ2 ), the lemma is proved. 2 The first and last inequalities in the lemma suggest that δN is very nearly an affine function of the source and hence perhaps δ is as well. We will later pursue this possibility, but we are not yet equipped to do so. Before developing the connection between the distortion-rate functions of AMS sources and those of their stationary mean, we pause to develop some additional properties for operational DRFs in the special case of stationary sources. These results follow Kieffer [91]. Lemma 12.3. Suppose that µ is a stationary source. Then δ(R, µ) = lim δN (R, µ). N→∞
Thus the infimum over block lengths is given by the limit so that longer codes can do better. ˆn and CN−n ⊂ Proof: Fix an N and an n < N and choose codes Cn ⊂ A ˆN−n for which A ρ(Cn , µ) ≤ δn (R, µ) +
2
ρ(CN−n , µ) ≤ δN−n (R, µ) +
. 2
Form the block length N code C = Cn × CN−n . This code has rate no greater than R and has distortion
302
12 Source Coding Theorems
Nρ(C, µ) = E min ρN (X N , y) y∈C
= Ey n ∈Cn ρn (X n , y n ) + Ev N−n ∈CN−n ρN−n (XnN−n , v N−n ) = Ey n ∈Cn ρn (X n , y n ) + Ev N−n ∈CN−n ρN−n (X N−n , v N−n ) = nρ(Cn , µ) + (N − n)ρ(CN−n , µ) ≤ nδn (R, µ) + (N − n)δN−n (R, µ) + ,
(12.6)
where we have made essential use of the stationarity of the source. Since is arbitrary and since the leftmost term in the above equation can be no smaller than NδN (R, µ), we have shown that NδN (R, µ) ≤ nδn (R, µ) + (N − n)δN−n (R, µ) and hence that the sequence NδN is subadditive. The result then follows immediately from Lemma 7.5.1 of [55] or Lemma 8.5.3 of [58]. 2 S Corollary 12.2. If µ is a stationary source, then δ(R, µ) is a convex function of R and hence is continuous for R > 0. Proof: Pick R1 > R2 and λ ∈ (0, 1). Define R = λR1 +(1−λ)R2 . For large n define n1 = bλnc be the largest integer less than λn and let n2 = n − n1 . ˆni with rate Ri with distortion Pick codebooks Ci ⊂ A ρ(Ci , µ) ≤ δni (Ri , µ) + . Analogous to (12.6), for the product code C = C1 × C2 we have nρ(C, µ) = n1 ρ(C1 , µ) + n2 ρ(C2 , µ) ≤ n1 δn1 (R1 , µ) + n2 δn2 (R2 , µ) + n. The rate of the product code is no greater than R and hence the leftmost term above is bounded below by nδn (R, µ). Dividing by n we have since is arbitrary that δn (R, µ) ≤
n1 n2 δn1 (R1 , µ) + δn2 (R2 , µ). n n
Taking n → ∞ we have using the lemma and the choice of ni that δ(R, µ) ≤ λδ(R1 , µ) + (1 − λ)δ(R2 , µ), proving the claimed convexity.
2
Corollary 12.3. If µ is stationary, then δ(R, µ) is an affine function of µ. Proof: From Lemma 12.2 we need only prove that δ(R, λµ1 + (1 − λ)µ2 ) ≤ λδ(R, µ1 ) + (1 − λ)δ(R, µ2 ).
12.2 Block Source Codes for AMS Sources
303
From the same lemma we have that for any N δN (R +
1 , λµ1 + (1 − λ)µ2 ) ≤ λδN (R, µ1 ) + (1 − λ)δN (R, µ2 ) N
For any K ≤ N we have since δN (R, µ) is nonincreasing in R that δN (R +
1 , λµ1 + (1 − λ)µ2 ) ≤ λδN (R, µ1 ) + (1 − λ)δN (R, µ2 ). K
Taking the limit as N → ∞ yields from Lemma 12.3 that δ(R +
1 , µ) ≤ λδ(R, µ1 ) + (1 − λ)δ(R, µ2 ). K
From Corollary 12.2, however, δ is continuous in R and the result follows by letting K → ∞. 2 The following lemma provides the principal tool necessary for relating the operational DRF of an AMS source with that of its stationary mean. It shows that the DRF of an AMS source is not changed by shifting or, equivalently, by redefining the time origin. Lemma 12.4. Let µ be an AMS source with a reference letter. Then for any integer i δ(R, µ) = δ(R, µT −i ). Proof: Fix > 0 and let CN be a rate R block length N codebook for which ρ(CN , µ) ≤ δ(R, µ) + /2. For 1 ≤ i ≤ N − 1 choose J large and define the block length K = JN code CK (i) by CK (i) = a∗
(N−i)
J−2
i
× × CN × a∗ , j=0
where a∗ l is an l-tuple containing all a∗ ’s. CK (i) can be considered to be a code consisting of the original code shifted by i time units and repeated many times, with some filler at the beginning and end. Except for the edges of the long product code, the effect on the source is to use the original code with a delay. The code has at most (2NR )J−1 = 2KR 2−NR words; the rate is no greater than R. (i) For any K-block x K the distortion resulting from using CK is given by KρK (x K , CK (i)) ≤ (N − i)ρN−i (x N−i , a∗
(N−i)
i
i ) + iρi (xK−i , a∗ ).
(12.7)
ˆn } denote the encoded process using the block code CK (i). If n Let {x is a multiple of K, then
304
12 Source Coding Theorems n bK c
ˆn) ≤ nρn (x n , x
X
N−i ((N − i)ρN−i (xkK , a∗
(N−i)
i
i ) + iρi (x(k+1)K−i , a∗ ))
k=0 n bK cJ−1
+
X
N NρN (xN−i+kN , CN ).
k=0
If n is not a multiple of K we can further overbound the distortion by including the distortion contributed by enough future symbols to complete a K-block, that is, ˆ n ) ≤ nγn (x, x) ˆ nρn (x n , x n
=
b K c+1
X
N−i , a∗ (N − i)ρN−i (xkK
(N−i)
i
i ) + iρi (x(k+1)K−i , a∗ )
k=0 n (b K c+1)J−1
+
X
N NρN (xN−i+kN , CN ).
k=0
Thus n
b K c+1 X N −i 1 (N−i) ˆ )≤ ρN−i (X N−i (T kK x), a∗ ) ρn (x , x K n/K k=0 n
n
n
b K c+1 X i 1 i + ρi (X i (T (k+1)K−i x, a∗ ) K n/K k=0
1 + n/N
n (b K c+1)J−1
X
ρN (X N (T (N−i)+kN x), CN ).
k=0
Since µ is AMS these quantities all converge to invariant functions: ˆn) ≤ lim ρn (x n , x
n→∞
+
m−1 1 X N −i (N−i) lim ρN−i (X N−i (T kK x), a∗ ) K m→∞ m k=0
m−1 1 X i i lim ρi (X i (T (k+1)K−i x, a∗ ) K m→∞ m k=0 m−1 1 X ρN (X N (T (N−i)+kN x), CN ). m→∞ m k=0
+ lim
We now apply Fatou’s lemma, a change of variables, and Lemma 12.1 to obtain
12.2 Block Source Codes for AMS Sources
305
δ(R, µT −i ) ≤ ρ(CK (i), µT −i ) ≤
m N −i 1 X (N−i) lim sup EµT −i ρN−i (X N−i T kK , a∗ ) K m→∞ m k=0
+
m−1 1 X i i EµT −i ρi (X i T (k+1)K−i , a∗ ) lim K m→∞ m k=0 m−1 1 X ρN (X N T (N−i)+kN ), CN ). m→∞ m k=0
+ EµT −i lim
≤
m−1 1 X i N −i ∗ ρ + ρ ∗ + Eµ lim ρN (X N T kN CN ) m→∞ m K K k=1
≤
N ∗ ρ + ρ(CN , µ). K
Thus if J and hence K are chosen large enough to ensure that N/K ≤ /2, then δ(R, µT −i ) ≤ δ(R, µ), which proves that δ(R, µT −i ) ≤ δ(R, µ). The reverse implication is found in a similar manner: Let CN be a codebook for µT −i and construct a codebook CK (N − i) for use on µ. By arguments nearly identical to those above the reverse inequality is found and the proof completed. 2 Corollary 12.4. Let µ be an AMS source with a reference letter. Fix N and let µ and µ N denote the stationary and N-stationary means. Then for R>0 δ(R, µ) = δ(R, µ N T −i ); i = 0, 1, . . . , N − 1. Proof: It follows from the previous lemma that the δ(R, µ N T −i ) are all equal and hence it follows from Lemma 12.2, Theorem 7.3.1 of [55] or Theorem 8.2 of [58], and Corollary 7.3.1 of [55] or Corollary 8.5 of [58] that N−1 1 X δ(R, µ) ≥ δ(R, µ N T −i ) = δ(R, µ N ). N i=0 To prove the reverse inequality, take µ = µ N in the previous lemma and construct SN−1the codes CK (i) as in the previous proof. Take the union code CK = i=0 CK (i) having block length K and rate at most R + K −1 log N. We have from Lemma 12.1 and (12.7) that ρ(CK , µ) =
≤
N−1 1 X ρ(CK , µ N T −i ) N i=0 N−1 N 1 X ρ(CK (i), µ N T −i ) ≤ ρ ∗ + ρ(CN , µ N ) N i=0 K
306
12 Source Coding Theorems
and hence as before δ(R +
1 log N, µ) ≤ δ(R, µ N ). JN
From Corollary 12.1 δ(R, µ) is continuous in R for R > 0 since µ is stationary. Hence taking J large enough yields δ(R, µ) ≤ δ(R, µ N ). This completes the proof since from the lemma δ(R, µ N T −i ) = δ(R, µ N ). 2 We are now prepared to demonstrate the fundamental fact that the block source coding operational distortion-rate function for an AMS source with an additive fidelity criterion is the same as that of the stationary mean process. This will allow us to assume stationarity when proving the actual coding theorems. Theorem 12.1. If µ is an AMS source and {ρn } an additive fidelity criterion with a reference letter, then for R > 0 δ(R, µ) = δ(R, µ). Proof: We have from Corollaries 11.2.1 and 11.2.4 that δ(R, µ) ≤ δ(R, µ N ) ≤ δN (R, µ N ) = δN (R, µ). Taking the infimum over N yields δ(R, µ) ≤ δ(R, µ). Conversely, fix > 0 let CN be a block length N codebook for which ρ(CN , µ) ≤ δ(R, µ)+. From Lemma 12.1, Corollary 12.1, and Lemma 12.4 δ(R, µ) + ≤ ρ(CN , µ) =
N−1 1 X ρ(CN , µ N T −i ) N i=0
≥
N−1 N−1 1 X 1 X δN (R, µ N T −i ) = δN (R, µT −i ) N i=0 N i=0
≥
N−1 1 X δ(R, µT −i ) = δ(R, µ), N i=0
which completes the proof since is arbitrary.
2
Since the DRFs are the same for an AMS process and its stationary mean, this immediately yields the following corollary from Corollary 12.2: Corollary 12.5. If µ is AMS, then δ(R, µ) is a convex function of R and hence a continuous function of R for R > 0.
12.3 Block Source Code Mismatch
307
12.3 Block Source Code Mismatch In this section the mismatch results of [66] for metric distortion measures are extended to additive distortion mesures that are a power of a metric, the class of fidelity criteria considered in Section 5.11. The formulation of operational distortion-rate functions for block codes in terms of the decoder or reproduction codebook alone can be combined with the process metric using the same distortion to quantify the difference in performance when a fixed code is applied on two distinct sources. The topic is of primary interest in the case where one designs a codebook to be optimal for one source, but then applies the codebook to actually code a different source. This can happen, for example, if the codebook is designed for a source by using an empirical distribution based on a training sequence, e.g., a clustering or learning algorithm is used on data to estimate the source statistics. The code so designed is then applied to the source itself, which results in a mismatch between the distribution used to design the code and the true, but unknown, underlying distribution. Intuitively, if the sources are close in some sense, then the performance of the code on the separate sources should also be close. Suppose that {ρn } is an additive fidelity criterion with per-symbol distortion that is a positive power of a metric, ρ1 (x, y) = d(x, y)p , p ≥ 0, as considered in Section 5.11. Fix a blocklength N, an N-dimensional reproduction codebook C, and two stationary sources with distributions µX and µY . Recall from (5.40) that the rho-bar distortion between the two sources is ρ N (µX n , µY n ) =
inf
π ∈P(µX N ,µY N )
ρ(µX , µY ) = sup N
Eπ ρN (X N , Y N )
1 ρ (µX N , µY N ). N N
We assume the existence of a reference letter so that all of the expectations considered are finite. Fix N and suppose that π approximately yields ρ N (µX N , µY N ) in the sense that it has the correct marginals and for small > 0 Eπ ρN (X N , Y N ) ≤ ρ N (µX N , µY N ) + . For any x N , y N , zN we have that ρN (x N , zN ) =
N−1 X i=0
ρ1 (xi , zi ) =
N−1 X i=0
d(xi , zi )p .
(12.8)
308
12 Source Coding Theorems
If d is a metric and 0 ≤ p ≤ 1, then ρN (x N , zN ) is also a metric, in particular it satisfies the triangle inequality. Thus in this case we have that ρN (C, µX ) = EµX min ρN (X N , zN ) = Eπ min ρN (X N , zN ) zN ∈C zN ∈C h i = Eπ min ρN (X N , Y N ) + ρN (Y N , zN ) zN ∈C = Eπ ρN (X N , Y N ) + Eπ min ρN (Y N , zN ) zN ∈C
≤ ρ N (µX N , µY N ) + + ρN (C, µY ). Since is arbitrary, ρN (C, µX ) ≤ ρN (C, µY ) + ρ N (µX N , µY N ).
(12.9)
Reversing the roles of X and Y in (12.9) and using the fact that the process distortion is an upper bound for the normalized vector distortion implies the following bound on the mismatch in performance resulting from applying the same block code to different sources: | N −1 ρN (C, µX ) − N −1 ρN (C, µY ) |≤ ρ(µX , µY )
(12.10)
Taking the infimum of both sides (12.10)over all codes C with N −1 log||C||≤ R, normalizing the distortion, and using the fact that the process rho-bar distortion is an upper bound for all vector distortions yields
N −1 δN (R, µX ) − N −1 δN (R, µY ) ≤ ρ(µX , µY ). Reversing the roles of X and Y yields | N −1 δN (R, µX ) − N −1 δN (R, µY ) |≤ ρ(µX , µY )
(12.11)
which yields the conclusion that the N-th order operational distortionrate functions are continuous functions of the source under the rho-bar distortion, which in this case of 0 ≤ p ≤ 1 is the rho-bar distance. Since the sources are assumed stationary, the limits as N → ∞ exist so that | δ(R, µX ) − δ(R, µY ) |≤ ρ(µX , µY )
(12.12)
so that the operational block coding DRF is continuous with respect to rho-bar. Now consider the case where p > 1 so that ρN does not satisfy a triangle inequality. Now, however, p
ρN (x N , zN ) = dN (x N , zN ) =
N−1 X i=0
d(xi , zi )p
12.3 Block Source Code Mismatch
309
is the pth power of a metric dN , the `p norm on the vectors of the individual distortions. Analogous to the previous case consider ρN (C, µX ) = EµX min ρN (X N , zN ) zN ∈C = Eπ min dN (X N , zN )p zN ∈C = Eπ min [dN (X N , Y N ) + dN (Y N , zN )]p zN ∈C N N N N p = Eπ [ min (dN (X , Y ) + dN (Y , z ))] zN ∈C
since f (x) = x p is monotonically increasing for positive p and hence the minimum of the pth power of a quantity is the p power of the minimum. Continuing, ρN (C, µX ) ≤ Eπ [dN (X N , Y N ) + min (dN (Y N , zN ))]p zN ∈C
and hence application of Minkowski’s inequality yields 1/p ρN (C, µX )1/p ≤ Eπ [dN (X N , Y N ) + min (dN (Y N , zN ))]p zN ∈C
h
≤ Eπ dN (X N , Y N )p
i1/p
1/p + Eπ min dN (Y N , zN )p zN ∈C
1/p ≤ ρ N (µX N , µY N ) + + ρN (C, µY )1/p . which since > 0 is arbitrary, 1/p ρN (C, µX )1/p ≤ ρ N (µX N , µY N ) + ρN (C, µY )1/p which in a similar fashion to the previous case results in | (N −1 ρN (C, µX ))1/p − (N −1 ρN (C, µY ))1/p | ≤ ρ(µX , µY )1/p | (N −1 δN (R, µX ))1/p − (N −1 δN (R, µY ))1/p | ≤ ρ(µX , µY )1/p | δ(R, µX )1/p − δ(R, µY )1/p | ≤ ρ(µX , µY )1/p so that again the rho-bar distortion provides a bound on the performance mismatch of a single codebook used for different sources and the block source coding operational distortion-rate functions are continuous with respect to the rho-bar distortion. This completes the proof of the following lemma. Lemma 12.5. Assume an additive fidelity criterion with per-letter distortion ρ1 = dp , a positive power of a metric, µX and µY two stationary process distributions, and C a reproduction codebook of length N.
310
12 Source Coding Theorems
If 0 ≤ p ≤ 1, then | N −1 ρN (C, µX ) − N −1 ρN (C, µY ) | ≤ ρ(µX , µY ) |N
−1
δN (R, µX ) − N
−1
(12.13)
δN (R, µY ) | ≤ ρ(µX , µY )
(12.14)
| δ(R, µX ) − δ(R, µY ) | ≤ ρ(µX , µY )
(12.15)
and if 1 ≤ p, then | (N −1 ρN (C, µX ))1/p − (N −1 ρN (C, µY ))1/p | ≤ ρ(µX , µY )1/p (12.16) | (N −1 δN (R, µX ))1/p − (N −1 δN (R, µY ))1/p | ≤ ρ(µX , µY )1/p (12.17) | δ(R, µX )1/p − δ(R, µY )1/p | ≤ ρ(µX , µY )1/p (12.18) Thus the block source coding operational distortion-rate functions δN (R, µ) and δ(R, µ) are continuous functions of µ in the rho-bar distortion (and dp distance). These mismatch results can be used to derive universal coding results for block source coding for certaom classes of sources. Universal codes are designed to provide nearly optimal coding for a collection of sources rather than for one specific source. The basic idea is to carve up the class using the rho-bar distortion (or the corresponding dp -distance) and to design a code for a specific representative of each subclass. If the members of a subclass are close in rho-bar, then the representative will work well for all sources in the subclass. The overall codebook is then formed as the union of the subclass codebooks. Provided the number of subclasses is small with respect to the block length, each subclass codebook can have nearly the full rate in bits per symbol and hence provide nearly optimal coding within the class. A minimum distortion rule encoder will find the best word within all of the classes. This approach to universal coding is detailed in [66, 132]. A variety of other approaches exist to this problem of source coding with uncertainty about the source, see for example [92, 198, 199].
12.4 Block Coding Stationary Sources We showed in the previous section that when proving block source coding theorems for AMS sources, we could confine interest to stationary sources. In this section we show that in an important special case we can further confine interest to only those stationary sources that are ergodic by applying the ergodic decomposition. This will permit us to assume that sources are stationary and ergodic in the next section when the basic Shannon source coding theorem is proved and then extend the result to AMS sources which may not be ergodic.
12.4 Block Coding Stationary Sources
311
As previously we assume that we have a stationary source {Xn } with distribution µ and we assume that {ρn } is an additive distortion measure and there exists a reference letter. For this section we now assume in addition that the alphabet A is itself a Polish space and that ρ1 (r , y) is ˆ If the underlying alphabet a continuous function of r for every y ∈ A. has a metric structure, then it is reasonable to assume that forcing input symbols to be very close in the underlying alphabet should force the distortion between either symbol and a fixed output to be close also. The following theorem is the ergodic decomposition of the block source coding operational distortion-rate function. Theorem 12.2. Suppose that µ is the distribution of a stationary source and that {ρn } is an additive fidelity criterion with a reference letter. Assume also that ρ1 (·, y) is a continuous function for all y. Let {µx } denote the ergodic decomposition of µ. Then Z δ(R, µ) = dµ(x)δ(R, µx ), that is, δ(R, µ) is the average of the operational DRFs of its ergodic components. Proof: Analogous to the ergodic decomposition of entropy rate of Theorem 3.3, we need to show that δ(R, µ) satisfies the conditions of Theorem 8.9.1 of [55] or Theorem 8.5 of [58]. We have already seen (Corollary 12.3) that it is an affine function. We next see that it is upper semicontinuous. Since the alphabet is Polish, choose a distance dG on the space of stationary processes having this alphabet with the property that G is constructed as in Section 8.2 of [55] or Section 9.8 of [58]. Pick an N large enough and a length N codebook C so that δ(R, µ) ≥ δN (R, µ) −
≥ ρN (C, µ) − . 2
ρN (x N , y) is by assumption a continuous function of x N and hence so is ρN (x N , C) = miny∈C ρ(x N , y). Since it is also nonnegative, we have from Lemma 8.2.4 of [55] or Lemma 9.3 of [58] that if µn → µ then lim sup Eµn ρN (X N , C) ≤ Eµ ρN (X N , C). n→∞
The left hand side above is bounded below by lim sup δN (R, µn ) ≥ lim sup δ(R, µn ). n→∞
n→∞
Thus since is arbitrary, lim sup δ(R, µn ) ≤ δ(R, µ) n→∞
312
12 Source Coding Theorems
and hence δ(R, µ) upper semicontinuous in µ and hence also measurable. Since the process has a reference letter, δ(R, µx ) is integrable since δ(R, µX ) ≤ δN (R, µx ) ≤ Eµx ρ1 (X0 , a∗ ) which is integrable if ρ1 (x0 , a∗ ) is from the ergodic decomposition theorem. Thus Theorem 8.9.1 of [55] or Theorem 8.5 of [58] yields the desired result. 2 The theorem was first proved by Kieffer [91] for bounded continuous additive distortion measures. The above extension removes the requirement that ρ1 be bounded.
12.5 Block Coding AMS Ergodic Sources We have seen that the block source coding operational DRF of an AMS source is given by that of its stationary mean. Hence we will be able to concentrate on stationary sources when proving the coding theorem. Theorem 12.3. Let µ be an AMS ergodic source with a standard alphabet and {ρn } an additive distortion measure with a reference letter. Then δ(R, µ) = D(R, µ), where µ is the stationary mean of µ. If µ is stationary, then δ(R, µ) = D(R, µ). Comment: Coupling the theorem with Lemma 12.5 shows that if the persymbol distortion is a positive power of a metric, dp , then the Shannon distortion rate function is a continuous function of the source distribution µ in terms of the corresponding rho-bar process distortion or the corresponding dp -distance. This has the same flavor of Corollary 6.2 showing that entropy was a continuous function of the d-bar distance, which assumed a mean Hamming distortion. The dual result to shows that the Shannon rate-distortion function is a continuous function of the source with respect to the rho-bar distortion. Proof: From Theorem 12.1 δ(R, µ) = δ(R, µ) and hence we will be done if we can prove that δ(R, µ) = D(R, µ). This will follow if we can show that δ(R, µ) = D(R, µ) for any stationary ergodic source with a reference letter. Henceforth we assume that µ is stationary and ergodic. The negative or converse half of the theorem follows from Corollary 9.1. As the specific case is simpler and short, a proof is included.
12.5 Block Coding AMS Ergodic Sources
313
First suppose that we have a codebook C such that ρN (C, µ) = Eµ min ρN (X N , y) = δN (R, µ) + . y∈C
ˆN denote the resulting reproduction random vector and let If we let X N p denote the resulting joint distribution of the input/output pair, then ˆN has a finite alphabet, Lemma 7.20 implies that since X ˆN ) ≤ H(X ˆN ) ≤ NR I(X N ; X and hence p N ∈ RN (R, µ N ) and hence ˆN ) ≥ DN (R, µ). δN (R, µ) + ≥ EpN ρN (X N ; X Taking the limits as N → ∞ proves the easy half of the theorem: δ(R, µ) ≥ D(R, µ). Recall that both operational DRF and the Shannon DRF are given by limits if the source is stationary. The fundamental idea of Shannon’s positive source coding theorem is this: for a fixed block size N, choose a code at random according to a distribution implied by the distortion-rate function. That is, perform 2NR independent random selections of blocks of length N to form a codebook. This codebook is then used to encode the source using a minimum distortion mapping as above. We compute the average distortion over this double-random experiment (random codebook selection followed by use of the chosen code to encode the random source). We will find that if the code generation distribution is properly chosen, then this average will be no greater than D(R, µ) + . If the average over all randomly selected codes is no greater than D(R, µ) + , then there must be at least one code such that the average distortion over the source distribution for that one code is no greater than D(R, µ) + . This means that there exists at least one code with performance not much larger than D(R, µ). Unfortunately the proof only demonstrates the existence of such codes, it does not show how to construct them. To find the distribution for generating the random codes we use the ergodic process definition of the Shannon distortion-rate function. From Theorem 9.1 (or Lemma 9.4) we can select a stationary and ergodic pair process with distribution p which has the source distribution µ as one coordinate and which has Ep ρ(X0 , Y0 ) = and which has
1 EpN ρN (X N , Y N ) ≤ D(R, µ) + N
(12.19)
314
12 Source Coding Theorems
I p (X; Y ) = I ∗ (X; Y ) ≤ R
(12.20)
1
(and hence information densities converge in L from Theorem 8.1). Denote the implied vector distributions for (X N , Y N ), X N , and Y N by p N , µ N , and ηN , respectively. For any N we can generate a codebook C at random according to ηN as described above. To be precise, consider the random codebook as a large random vector C = (W0 , W1 , · · · , WM ), where M = beN(R+) c (where natural logarithms are used in the definition of R), where W0 is the fixed reference vector a∗ N and where the remaining Wn are independent, and where the marginal distributions for the Wn are given by ηN . Thus the distribution for the randomly selected code can be expressed as M
P C = × ηN . i=1
This codebook is then used with the optimal encoder and we denote the resulting average distortion (over codebook generation and the source) by Z ∆N = Eρ(C, µ) =
dPC (W )ρ(W , µ)
(12.21)
where ρ(W , µ) =
1 1 EρN (X N , W ) = N N
Z
dµ N (x N )ρN (x N , W ),
and where ρN (x N , C) = min ρN (x N , y). y∈C
Choose δ > 0 and break up the integral over x into two pieces: one over a set GN = {x : N −1 ρN (x N , a∗ N ) ≤ ρ ∗ + δ} and the other over the complement of this set. Then Z ∆N ≤
c GN
1 N ρN (x N , a∗ ) dµ N (x N ) N Z Z 1 + dµ N (x N )ρN (x N , W ), dPC (W ) N GN
(12.22)
where we have used the fact that ρN (x N , mW ) ≤ ρN (x N , a∗ N ). Fubini’s theorem implies that because Z N dµ N (x N )ρN (x N , a∗ ) < ∞ and
N
ρN (x N , W ) ≤ ρN (x N , a∗ ),
12.5 Block Coding AMS Ergodic Sources
315
the limits of integration in the second integral of (12.22) can be interchanged to obtain the bound ∆N ≤
1 N
Z
N
c GN
ρN (x N , a∗ )dµ N (x N ) Z Z 1 + dµ N (x N ) dPC (W )ρN (x N , W ) N GN
(12.23)
The rightmost term in (12.23) can be bound above by observing that 1 N
Z
Z dµ N (x N )[ dPC (W )ρN (x N , W )] GN Z Z 1 N N dµ (x )[ dPC (W )ρN (x N , W ) = N GN C:ρN (x N ,C)≤N(D+δ) Z 1 + dPC (W )ρN (x N , W )] N W :ρN (x N ,W )>N(D+δ) Z Z 1 ≤ dµ N (x N )[D + δ + (ρ ∗ + δ) dpC (W )] N GN W :ρN (x N ,W )>N(D+δ)
where we have used the fact that for x ∈ G the maximum distortion is given by ρ ∗ + δ. Define the probability Z P (N −1 ρN (x N , C) > D + δ|x N ) = dpC (W ) W :ρN (x N ,W )>N(D+δ)
and summarize the above bounds by ∆N ≤ D + δ+ 1 (ρ + δ) N ∗
Z
dµ N (x N )P (N −1 ρN (x N , C) > D + δ|x N ) Z 1 N + dµ N (x N )ρN (x N , a∗ ). N GNc
(12.24)
The remainder of the proof is devoted to proving that the two integrals above go to 0 as N → ∞ and hence lim sup ∆N ≤ D + δ.
(12.25)
N→∞
Consider first the integral Z Z 1 1 N N dµ N (x N )ρN (x N , a∗ ) = dµ N (x N )1GNc (x N ) ρN (x N , a∗ ). aN = c N GN N We shall see that this integral goes to zero as an easy application of the ergodic theorem. The integrand is dominated by N −1 ρN (x N , a∗ N ) which
316
12 Source Coding Theorems
is uniformly integrable (Lemma 4.7.2 of [55] or Lemma 5.23 of [58]) and hence the integrand is itself uniformly integrable (Lemma 4.4.4 of [55] or Lemma 5.9 of [58]). Thus we can invoke the extended Fatou lemma (Lemma 4.4.5 of [55] or Lemma 5.10 of [58]) to conclude that Z 1 N lim sup aN ≤ dµ N (x N ) lim sup 1GNc (x N ) ρN (x N , a∗ ) N N→∞ N→∞ Z 1 N ≤ dµ N (x N )(lim sup 1GNc (x N ))(lim sup ρN (x N , a∗ )). N→∞ N→∞ N c We have, however, that lim supN→∞ 1GNc (x N ) is 0 unless x N ∈ GN i.o. But this set has measure 0 since with µN probability 1, an x is produced so that N−1 1 X lim ρ(xi , a∗ ) = ρ ∗ N→∞ N i=0
exists and hence with probability one one gets an x which can yield N
N −1 ρN (x N , a∗ ) > ρ ∗ + δ at most for a finite number of N. Thus the above integral of the product of a function that is 0 a.e. with a dominated function must itself be 0 and hence lim sup aN = 0. (12.26) N→∞
We now consider the second integral in (12.24): Z 1 ∗ dµ N (x N )P (N −1 ρN (x N , C) > D + δ|x N ). bN = (ρ + δ) N Recall that P (ρN (x N , C) > D + δ|x N ) is the probability that for a fixed input block x N , a randomly selected code will result in a minimum distortion codeword larger than D + δ. This is the probability that none of the M words (excluding the reference code word) selected independently at random according to to the distribution ηN lie within D+δ of the fixed input word x N . This probability is bounded above by P(
1 1 ρN (x N , C) > D + δ|x N ) ≤ [1 − ηN ( ρN (x N , Y N ) ≤ D + δ)]M N N
where ηN (
1 ρN (x N , Y N ) ≤ D + δ)) = N
Z 1
y N : N ρN (x N ,y N )≤D+δ
dηN (y N ).
Now mutual information comes into the picture. The above probability can be bounded below by adding a condition:
12.5 Block Coding AMS Ergodic Sources
ηN (
317
1 ρN (x N , Y N ) ≤ D + δ) N 1 1 ≥ ηN ( ρN (x N , Y N ) ≤ D + δ and iN (x N , Y N ) ≤ R + δ), N N
where
1 1 iN (x N , y N ) = ln fN (x N , y N ), N N
where fN (x N , y N ) =
dp N (x N , y N ) , d(µ N × ηN )(x N , y N )
the Radon-Nikodym derivative of p N with respect to the product measure µ N × ηN . Thus we require both the distortion and the sample information be less than slightly more than their limiting value. Thus we have in the region of integration that 1 1 iN (x N ; y N ) = ln fN (x N , y N ) ≤ R + δ N N and hence ηN (ρN (x N , Y N ) ≤ D + δ) Z ≥ dηN (y N ) y N :ρN (x N ,y N )≤D+δ,fN (x N ,y N )≤eN(R+δ) Z dηN (y N )fN (x N , y N ) ≥ e−N(R+δ) y N :ρN (x N ,y N )≤D+δ,fN (x N ,y N )≤eN(R+δ)
which yields the bound 1 ρN (x N , C) > D + δ|x N ) N 1 ≤ [1 − ηN ( ρN (x N , Y N ) ≤ D + δ)]M N Z
P(
≤ [1 − e−N(R+δ)
1
1
y N : N ρN (x N ,y N )≤D+δ, N iN (x N ,y N )≤R+δ
Applying the inequality (1 − αβ)M ≤ 1 − β + e−Mα for α, β ∈ [0, 1] yields
dηN (y N )fN (x N , y N )]M ,
318
P(
12 Source Coding Theorems
1 ρN (x N , C) > D + δ|x N ) ≤ N Z 1−
1
1
y N : N ρN (x N ,y N )≤D+δ, N iN (x N ,y N )≤R+δ
dηN (y N ) × fN (x N , y N ) + e[−Me
−N(R+δ) ]
.
Averaging with respect to the distribution µ N yields Z bN = dµ N (x N )P (ρN (x N , C) > D + δ|x N ) ρ∗ + δ Z Z ≤ dµ N (x N ) 1 − 1
y N :ρN (x N ,y N )≤N(D+δ), N iN (x N ,y N )≤R+δ
×fN (x N , y N ) + e−Me
−N(R+δ)
Z =1−
1 1 yN : N ρN (x N ,y N )≤D+δ, N iN (x N ,y N )≤R+δ
= 1 + e−Me
−N(R+δ)
× fN (x N , y N ) + e−Me Z −
− p N (y N :
d(µ N × ηN )(x N , y N )
−N(R+δ)
1 1 yN : N ρN (x N ,y N )≤D+δ, N iN (x N ,y N )≤R+δ
= 1 + e−Me
dηN (y N )
dp N (x N , y N )
−N(R+δ)
1 1 ρN (x N , y N ) ≤ D + δ, iN (x N , y N ) ≤ R + δ). N N
(12.27)
Since M is bounded below by eN(R+) −1, the exponential term is bounded above by (N(R+) e−N(R+δ) +e−N(R+δ) ] N(−δ) +e−N(R+δ) ] = e[−e . e[−e If > δ, this term goes to 0 as N → ∞. The probability term in (12.27) goes to 1 from the mean ergodic theorem applied to ρ1 and the mean ergodic theorem for information density since mean convergence (or the almost everywhere convergence proved elsewhere) implies convergence in probability. This implies that lim sup bN = 0 n→∞
which with (12.26) gives (12.25). Choosing an N so large that ∆N ≤ δ, we have proved that there exists a block code C with average distortion less than D(R, µ) + δ and rate less than R + and hence δ(R + , µ) ≤ D(R, µ) + δ.
(12.28)
Since and δ can be chosen as small as desired and since D(R, µ) is a continuous function of R (Lemma 9.1), the theorem is proved. 2
12.6 Subadditive Fidelity Criteria
319
The source coding theorem is originally due to Shannon [162] [163], who proved it for discrete IID sources. It was extended to stationary and ergodic discrete alphabet sources and Gaussian sources by Gallager [47] and to stationary and ergodic sources with abstract alphabets by Berger [10] [11], but an error in the information density convergence result of Perez [148] (see Kieffer [89]) left a gap in the proof, which was subsequently repaired by Dunham [36]. The result was extended to nonergodic stationary sources and metric distortion measures and Polish alphabets by Gray and Davisson [59] and to AMS ergodic processes by Gray and Saadat [70]. The method used here of using a stationary and ergodic measure to construct the block codes and thereby avoid the block ergodic decomposition of Nedoma [129] used by Gallager [47] and Berger [11] was suggested by Pursley and Davisson [29] and developed in detail by Gray and Saadat [70].
12.6 Subadditive Fidelity Criteria In this section we generalize the block source coding theorem for stationary sources to subadditive fidelity criteria. Several of the interim results derived previously are no longer appropriate, but we describe those that are still valid in the course of the proof of the main result. Most importantly, we now consider only stationary and not AMS sources. The result can be extended to AMS sources in the two-sided case, but it is not known for the one-sided case. Source coding theorems for subadditive fidelity criteria were first developed by Mackenthun and Pursley [111]. Theorem 12.4. Let µ denote a stationary and ergodic distribution of a source {Xn } and let {ρn } be a subadditive fidelity criterion with a referˆ such that ence letter, i.e., there is an a∗ ∈ A Eρ1 (X0 , a∗ ) = ρ ∗ < ∞. Then the operational DRF for the class of block codes of rate less than R is given by the Shannon distortion-rate function D(R, µ). Proof: Suppose that we have a block code of length N, e.g., a block enˆN . Since the source coder α : AN → B K and a block decoder β : B K → A is stationary, the induced input/output distribution is then N-stationary and the performance resulting from using this code on a source µ is ∆N = Ep ρ∞ =
1 ˆN ), Ep ρN (X N , X N
ˆN } is the resulting reproduction process. Let δN (R, µ) denote where {X the infimum over all codes of length N of the performance using such
320
12 Source Coding Theorems
codes and let δ(R, µ) denote the infimum of δN over all N, that is, the operational distortion rate function. We do not assume a codebook/minimum distortion structure because the distortion is now effectively context dependent and it is not obvious that the best codes will have this form. Assume that given an > 0 we have chosen for each N a length N code such that δN (R, µ) ≥ ∆N − . As previously we assume that K log ||B|| ≤ R, N where the constraint R is the rate of the code. As in the proof of the converse coding theorem for an additive distortion measure, we have ˆN ) ≤ RN and hence that for the resulting process I(X N ; X ∆N ≥ DN (R, µ). From Lemma 9.2 we can take the infimum over all N to find that δ(R, µ) = inf δN (R, µ) ≥ inf DN (R, µ) − = D(R, µ) − . N
N
Since is arbitrary, δ(R, µ) ≤ D(R, µ), proving the converse theorem. To prove the positive coding theorem we proceed in an analogous manner to the proof for the additive case, except that we use Lemma 9.4 instead of Theorem 9.1. First pick an N large enough so that DN (R, µ) ≤ D(R, µ) +
δ 2
and then select a p N ∈ RN (R, µ N ) such that EpN
1 δ ρN (X N , Y N ) ≤ DN (R, µ) + ≤ D(R, µ) + δ. N 2
Construct as in Lemma 9.4 a stationary and ergodic process p which will have (10.6.4) and (10.6.5) satisfied (the right Nth order distortion and information). This step taken, the proof proceeds exactly as in the additive case since the reference vector yields the bound N−1 1 1 X N ρN (x N , a∗ ) ≤ ρ1 (xi , a∗ ), N N i=0
which converges, and since N −1 ρN (x N , y N ) converges as N → ∞ with p probability one from the subadditive ergodic theorem. Thus the exis-
12.7 Asynchronous Block Codes
321
tence of a code satisfying (12.28) can be demonstrated (which uses the minimum distortion encoder) and this implies the result since D(R, µ) is a continuous function of R (Lemma 9.1). 2
12.7 Asynchronous Block Codes The block codes considered so far all assume block synchronous communication, that is, that the decoder knows where the blocks begin and hence can deduce the correct words in the codebook from the index represented by the channel block. In this section we show that we can construct asynchronous block codes with little loss in performance or rate; that is, we can construct a block code so that a decoder can uniquely determine how the channel data are parsed and hence deduce the correct decoding sequence. This result will play an important role in the development in the next section of sliding-block coding theorems. The basic approach is that taken in the development of asynchronous and sliding-block almost lossless codes in Section 6.5. Given a source µ let δasync (R, µ) denote the operational distortion rate function for block codes with the added constraint that the decoder be able to synchronize, that is, correctly parse the channel codewords. Obviously δasync (R, µ) ≥ δ(R, µ) since we have added a constraint. The goal of this section is to prove the following result: Theorem 12.5. Given an AMS source with an additive fidelity criterion and a reference letter, δasync (R, µ) = δ(R, µ), that is, the operational DRF for asynchronous codes is the same as that for ordinary codes. Proof: A simple way of constructing a synchronized block code is to use a prefix code: Every codeword begins with a short prefix or source synchronization word or, simply, sync word, that is not allowed to appear anywhere else within a word or as any part of an overlap of the prefix and a piece of the word. The decoder then need only locate the prefix in order to decode the block begun by the prefix. The insertion of the sync word causes a reduction in the available number of codewords and hence a loss in rate, but ideally this loss can be made negligible if properly done. We construct a code in this fashion by finding a good codebook of slightly smaller rate and then indexing it by channel K-tuples with this prefix property.
322
12 Source Coding Theorems
Suppose that our channel has a rate constraint R, that is, if source N-tuples are mapped into channel K-tuples then K log ||B|| ≤ R, N where B is the channel alphabet. We assume that the constraint is achievable on the channel in the sense that we can choose N and K so that the physical stationarity requirement is met (N source time units corresponds to K channel time units) and such that ||B||K ≈ eNR ,
(12.29)
at least for large N. If K is to be the block length of the channel code words, let δ be small and define k(K) = bδKc + 1 and consider channel codewords which have a prefix of k(K) occurrences of a single channel letter, say b, followed by a sequence of K − k(K) channel letters which have the following constraint: no k(K)-tuple beginning after the first symbol can be bk(K) . We permit b’s to occur at the end of a K-tuple so that a k(K)-tuple of b’s may occur in the overlap of the end of a codeword and the new prefix since this causes no confusion, e.g., if we see an elongated sequence of b’s, the actual code information starts at the right edge. Let M(K) denote the number of distinct channel K-tuples of this form. Since M(K) is the number of distinct reproduction codewords that can be indexed by channel codewords, the codebooks will be constrained to have rate RK =
ln M(K) . N
We now study the behavior of RK as K gets large. There are a total of ||B||K−k(K) K-tuples having the given prefix. Of these, no more than (K − k(K))||B||K−2k(K) have the sync sequence appearing somewhere within the word (there are fewer than K − k(K) possible locations for the sync word and for each location the remaining K − 2k(K) symbols can be anything). Lastly, we must also eliminate those words for which the first i symbols are b for i = 1, 2, . . . , k(K) − 1 since this will cause confusion about the right edge of the sync sequence. These terms contribute k(K)−1 X
||B||K−k(K)−i
i=1
bad words. Using the geometric progression formula to sum the above series we have that it is bounded above by ||B||K−k(K)−1 . 1 − 1/||B||
12.8 Sliding-Block Source Codes
323
Thus the total number of available channel vectors is at least M(K) ≥ ||B||K−k(K) − (K − k(K))||B||K−2k(K) −
||B||K−k(K)−1 . 1 − 1/||B||
Thus 1 1 1 ln ||B||K−k(K) + ln 1 − (K − k(K))||B||−k(K) − N N ||B|| − 1 K − k(K) 1 ||B|| − 2 = ln ||B|| + ln − (K − k(K))||B||−k(K) N N ||B|| − 1 ≥ (1 − δ)R + o(N),
RK =
where o(N) is a term that goes to 0 as N (and hence K) goes to infinity. Thus given a channel with rate constraint R and given > 0, we can construct for N sufficiently large a collection of approximately eN(R−) channel K-tuples (where K ≈ NR) which are synchronizable, that is, satisfy the prefix condition. We are now ready to construct the desired code. Fix δ > 0 and then choose > 0 small enough to ensure that δ(R(1 − ), µ) ≤ δ(R, µ) +
δ 3
(which we can do since δ(R, µ) is continuous in R). Then choose an N large enough to give a prefix channel code as above and to yield a rate R − codebook C so that δ 3 2δ ≤ δ(R − , µ) + ≤ δ(R, µ) + δ. 3
ρN (C, µ) ≤ δN (R − , µ) +
The resulting code proves the theorem.
(12.30)
2
12.8 Sliding-Block Source Codes We now turn to sliding-block codes. For simplicity we consider codes which map blocks into single symbols. For example, a sliding-block encoder will be a mapping f : AN → B and the decoder will be a mapping ˆ In the case of one-sided processes, for example, the channel g : B K → A. sequence would be given by Un = f (XnN )
324
12 Source Coding Theorems
and the reproduction sequence by ˆn = g(UnL ). X When the processes are two-sided, it is more common to use memory as well as delay. This is often done by having an encoder mapping f : ˆ and the channel and reproduction A2N+1 → B, a decoder g : B 2L+1 → A, sequences being defined by Un = f (X−N , · · · , X0 , · · · , XN ), ˆn = g(U−L , · · · , U0 , · · · , UN ). X We emphasize the two-sided case. The final output can be viewed as a sliding-block coding of the input: ˆn = g(f (Xn−L−N , · · · , Xn−L+N ), · · · , f (Xn+L−N , · · · , Xn+L+N )) X = gf (Xn−(N+L) , · · · , Xn+(N+L) ), where we use gf to denote the overall coding, that is, the cascade of g and f . Note that the delay and memory of the overall code are the sums of those for the encoder and decoder. The overall window length is 2(N + L) + 1 Since one channel symbol is sent for every source symbol, the rate of such a code is given simply by R = log ||B|| bits per source symbol. The obvious problem with this restriction is that we are limited to rates which are logarithms of integers, e.g., we cannot get fractional rates. As previously discussed, however, we could get fractional rates by appropriate redefinition of the alphabets (or, equivalently, of the shifts on the corresponding sequence spaces). For example, regardless of the code window lengths involved, if we shift l source symbols to produce a new group of k channel symbols (to yield an (l, k)-stationary encoder) and then shift a group of k channel symbols to produce a new group of k source symbols, then the rate is k R = log ||B|| l bits or nats per source symbol and the overall code f g is l-stationary. The added notation to make this explicit is significant and the generalization is straightforward; hence we will stick to the simpler case. We can define the sliding-block operational DRF for a source and channel in the natural way. Suppose that we have an encoder f and a decoder g. Define the resulting performance by ρ(f g, µ) = Eµf g ρ∞ , where µf g is the input/output hookup of the source µ connected to the deterministic channel f g and where ρ∞ is the sequence distortion.
12.8 Sliding-Block Source Codes
325
Define δSBC (R, µ) = inf ρ(f g, µ) = ∆(µ, E, ν, D), f ,g
(12.31)
where E is the class of all finite-length sliding-block encoders and D is the collection of all finite-length sliding-block decoders. The rate constraint R is determined by the channel. Assume as usual that µ is AMS with stationary mean µ. Since the cascade of stationary channels f g is itself stationary (Lemma 2.10), we have from Lemma 2.2 that µf g is AMS with stationary mean µf g. This implies from (5.12) that for any sliding-block codes f and g Eµf g ρ∞ = Eµf g ρ∞ and hence δSBC (R, µ) = δSBC (R, µ). A fact we now formalize as a lemma. Lemma 12.6. Suppose that µ is an AMS source with stationary mean µ and let {ρn } be an additive fidelity criterion. Let δSBC (R, µ) denote the sliding-block coding operational distortion-rate function for the source and a channel with rate constraint R. Then δSBC (R, µ) = δSBC (R, µ). The lemma permits us to concentrate on stationary sources when quantifying the optimal performance of sliding-block codes. The principal result of this section is the following: Theorem 12.6. Given an AMS and ergodic source µ and an additive fidelity criterion with a reference letter, δSBC (R, µ) = δ(R, µ), that is, the class of sliding-block codes is capable of exactly the same performance as the class of block codes. If the source is only AMS and not ergodic, then δSBC (R, µ) ≥ δ(R, µ), (12.32) Proof: The proof of (12.32) follows that of Shields and Neuhoff [167] for the finite alphabet case, except that their proof was for ergodic sources and coded only typical input sequences. Their goal was different because they measured the rate of a sliding-block code by the entropy rate of its output, effectively assuming that further almost-noiseless coding was to be used. Because we consider a fixed channel and measure the rate in the usual way as a coding rate, this problem does not arise here. From the previous lemma we need only prove the result for stationary sources and hence we henceforth assume that µ is stationary. We first prove
326
12 Source Coding Theorems
that sliding-block codes can perform no better than block codes, that is, (12.32) holds. Fix δ > 0 and suppose that f : A2N+1 → B and g : B 2L+1 → ˆ are finite-length sliding-block codes for which A ρ(f g, µ) ≤ δSBC (R, µ) + δ. ˆ which we This yields a cascade sliding-block code f g : A2(N+L)+1 → A use to construct a block codebook. Choose K large (to be specified later). Observe an input sequence x n of length n = 2(N + L) + 1 + K and map it ˆ n as follows: Set the first and last (N + L) into a reproduction sequence x N+L symbols to the reference letter a∗ , that is, x0N+L = xn−N−L = a∗ (N+L) . Complete the remaining reproduction symbols by sliding-block coding the source word using the given codes, that is, 2(N+L)+1
ˆi = f g(xi−(N+L) ); i = N + L + 1, · · · , K + N + L. x Thus the long block code is obtained by sliding-block coding, except at the edges where the sliding-block code is not permitted to look at previous or future source symbols and hence are filled with a reference symbol. Call the resulting codebook C. The rate of the block code is less than R = log ||B|| because n channel symbols are used to produce a reproduction word of length n and hence the codebook can have no more that ||B||n possible vectors. Thus the rate is log ||B|| since the codebook is used to encode a source n-tuple. Using this codebook with a minimum distortion rule can do no worse (except at the edges) than if the original ˆi is the reproducsliding-block code had been used and therefore if X tion process produced by the block code and Yi that produced by the sliding-block code, we have (invoking stationarity) that nρ(C, µ) N+L−1 X
≤ E(
ρ(Xi , a∗ )) + E(
i=0
K+N+L X
ρ(Xi , Yi )) + E(
i=N+L
K+2(L+N) X
ρ(Xi , a∗ ))
i=K+N+L+1
≤ 2(N + L)ρ ∗ + K(δSBC (R, µ) + δ) and hence δ(R, µ) ≤
K 2(N + L) ρ∗ + (δSBC (R, µ) + δ). 2(N + L) + K 2(N + L) + K
By choosing δ small enough and K large enough we can make make the right hand side arbitrarily close to δSBC (R, µ), which proves (12.32). We now proceed to prove the converse inequality, δ(R, µ) ≥ δSBC (R, µ), which involves a bit more work.
(12.33)
12.8 Sliding-Block Source Codes
327
Before carefully tackling the proof, we note the general idea and an “almost proof” that unfortunately does not quite work, but which may provide some insight. Suppose that we take a very good block code, e.g., a block code C of block length N such that ρ(C, µ) ≤ δ(R, µ) + δ for a fixed δ > 0. We now wish to form a sliding-block code for the same channel with approximately the same performance. Since a slidingblock code is just a stationary code (at least if we permit an infinite window length), the goal can be viewed as “stationarizing” the nonstationary block code. One approach would be the analogy of the SBM channel. Since a block code can be viewed as a deterministic block memoryless channel, we could make it stationary by inserting occasional random spacing between long sequences of blocks. Ideally this would then imply the existence of a sliding-block code from the properties of SBM channels. The problem is that the SBM channel so constructed would no longer be a deterministic coding of the input since it would require the additional input of a random punctuation sequence. Nor could one use a random coding argument to claim that there must be a specific (nonrandom) punctuation sequence which could be used to construct a code since the deterministic encoder thus constructed would not be a stationary function of the input sequence, that is, it is only stationary if both the source and punctuation sequences are shifted together. Thus we are forced to obtain the punctuation sequence from the source input itself in order to get a stationary mapping. The original proofs for this result [65, 67] used a strong form of the Rohlin-Kakutani theorem of Section 2.22 given by Shields [164]. The Rohlin-Kakutani theorem demonstrates the existence of a punctuation sequence with the property that the punctuation sequence is very nearly independent of the source. Lemma 2.12 is a slightly weaker result than the form considered by Shields. The code construction described above can therefore be approximated by using a coding of the source instead of an independent process. Shields and Neuhoff [167] provided a simpler proof of a result equivalent to the Rohlin-Kakutani theorem and provided such a construction for finite alphabet sources. Davisson and Gray [28] provided an alternative heuristic development of a similar construction. We here adopt a somewhat different tack in order to avoid some of the problems arising in extending these approaches to general alphabet sources and to nonergodic sources. The principal difference is that we do not try to prove or use any approximate independence between source and the punctuation process derived from the source (which is code dependent in the case of continuous alphabets). Instead we take a good block code and first produce a much longer block code that is insensitive to shifts or starting positions using the same construction used to relate block cod-
328
12 Source Coding Theorems
ing performance of AMS processes to that of their stationary mean. This modified block code is then made into a sliding-block code using a punctuation sequence derived from the source. Because the resulting block code is little affected by starting time, the only important property is that most of the time the block code is actually in use. Independence of the punctuation sequence and the source is no longer required. The approach is most similar to that of Davisson and Gray [28], but the actual construction differs in the details. An alternative construction may be found in Kieffer [93]. Given δ > 0 and > 0, choose for large enough N an asynchronous block code C of block length N such that 1 log ||C|| ≤ R − 2 N and ρ(C, µ) ≤ δ(R, µ) + δ.
(12.34)
The continuity of the block operational distortion-rate function and the theorem for asynchronous block source coding ensure that we can do this. Next we construct a longer block code that is more robust against shifts. For i = 0, 1, . . . , N − 1 construct the codes CK (i) having length K = JN as in the proof of Lemma 12.4. These codebooks look like J − 1 repetitions of the codebook C starting from time i with the leftover symbols at the beginning and end S being filled by the reference letter. We then form the union code CK = i CK (i) as in the proof of Corollary 12.4 which has all the shifted versions. This code has rate no greater than R − 2 + (JN)−1 log N. We assume that J is large enough to ensure that 1 log N ≤ JN
(12.35)
so that the rate is no greater than R − and that 3 ∗ ρ ≤ δ. J
(12.36)
We now construct a sliding-block encoder f and decoder g from the given block code. From Corollary 2.1 we can construct a finite length sliding-block code of {Xn } to produce a two-sided (NJ, γ)-random punctuation sequence {Zn }. From the lemma P (Z0 = 2) ≤ γ and hence by the continuity of integration (Corollary 4.4.2 of [55] or Corollary 5.3 in [58]) we can choose γ small enough to ensure that Z ρ(X0 , a∗ ) ≤ δ. (12.37) x:Z0 (x)=2
12.8 Sliding-Block Source Codes
329
Recall that the punctuation sequence usually produces 0’s followed by NJ − 1 1’s with occasional 2’s interspersed to make things stationary. The sliding-block encoder f begins with time 0 and scans backward NJ time units to find the first 0 in the punctuation sequence. If there is no such 0, then put out an arbitrary channel symbol b. If there is such a K 0, then the block codebook CK is applied to the input K-tuple x−n to produce the minimum distortion codeword K uK = min −1 ρK (x−n , y) y∈CK
and the appropriate channel symbol, un , produced by the channel. The sliding-block encoder thus has length at most 2NJ + 1. The decoder sliding-block code g scans left N symbols to see if it finds a codebook sync sequence (remember the codebook is asynchronous and begins with a unique prefix or sync sequence). If it does not find one, it produces a reference letter. (In this case it is not in the middle of a code word.) If it does find one starting in position −n, then it produces the corresponding length N codeword from C and then puts out the reproduction symbol in position n. Note that the decoder sliding-block code has a finite window length of at most 2N + 1. We now evaluate the average distortion resulting from use of this sliding-block code. As a first step we mimic the proof of Lemma 9.4 up to the assumption of mutual independence of the source and the punctuation process (which is not the case here) to infer that if a long source sequence of length n yields the punctuation sequence z, then X X NJ NJ ˆn) = ˆi ), ρn (x n , x ρ(xi , a∗ ) + ρNJ (xi , x i∈J2n (z)
i∈J0n (z)
where J2n (z) is the collection of all i for which zi = 2 and hence zi is not in an NJ-cell (so that filler is being sent) and J0n (z) is the collection of all i for which zi is 0 and hence begins an NJ-cell and hence an NJ length codeword. Each one of these length NJ codewords contains at most N reference letters at the beginning and N reference letters at the end the end and in the middle it contains all shifts of sequences of length N codewords from C. Thus for any i ∈ J0n (z), we can write that NJ NJ ˆi ) ρNJ (xi , x
≤
N N N ρN (xiN , a∗ ) + ρN (xi+NJ−N , a∗ ) +
b Ni c+JN−1
X i
j=b N c
This yields the bound
ρN (xjN , C).
330
12 Source Coding Theorems
1 1 ˆn) ≤ ρn (x n , x n n
X
ρ(xi , a∗ )+
i∈J2n (z) n
1 n
X
∗N
ρN (xiN , a
∗N
N ) + ρN (xi+NJ−N ,a
i∈J0n (z)
=
bNc 1 X N ) + ρN (xjN , C) n j=0
n−1 1 X 12 (zi )ρ(xi , a∗ )+ n i=0 n
n−1 Nc 1 bX 1 X N N N N 10 (zi ) ρN (xiN , a∗ ) + ρN (xi+NJ−N , a∗ ) + ρN (xjN , C), n i=0 n j=0
where as usual the indicator function 1a (zi ) is 1 if zi = a and 0 otherwise. Taking expectations above we have that E
n−1 1 1 X ˆn ) ≤ E 12 (Zi )ρ(Xi , a∗ ) ρn (X n , X n n i=0 +
n−1 i 1 X h N N N E 10 (Zi ) ρN (XiN , a∗ ) + ρN (Xi+NJ−N , a∗ ) ) n i=0 n
bNc i 1 X h N E ρN (XjN , C) . + n j=0
Invoke stationarity to write E(
1 ˆn )) ≤ E(12 (Z0 )ρ(X0 , a∗ ))+ ρn (X n , X n 1 1 (2N+1) )) + ρN (X N , C). E(10 (Z0 )ρ2N+1 (X 2N+1 , a∗ NJ N
The first term is bounded above by δ from (12.37). The middle term can be bounded above using (12.36) by 1 1 (2N+1) (2N+1) E(10 (Z0 )ρ2N+1 (X 2N+1 , a∗ Eρ2N+1 (X 2N+1 , a∗ )≤ ) JN JN 1 2 = (2N + 1)ρ ∗ ≤ ( + 1)ρ ∗ ≤ δ. JN J Thus we have from the above and (12.34) that Eρ(X0 , Y0 ) ≤ ρ(C, µ) + 3δ.
12.8 Sliding-Block Source Codes
331
This proves the existence of a finite window sliding-block encoder and a finite window length decoder with performance arbitrarily close to that achievable by block codes. 2 The only use of ergodicity in the proof of the theorem was in the selection of the source sync sequence used to imbed the block code in a sliding-block code. The result would extend immediately to nonergodic stationary sources (and hence to nonergodic AMS sources) if we could somehow find a single source sync sequence that would work for all ergodic components in the ergodic decomposition of the source. Note that the source synch sequence affects only the encoder and is irrelevant to the decoder which looks for asynchronous codewords prefixed by channel synch sequences (which consisted of a single channel letter repeated several times). Unfortunately, one cannot guarantee the existence of a single source sequence with small but nonzero probability under all of the ergodic components. Since the components are ergodic, however, an infinite length sliding-block encoder could select such a source sequence in a simple (if impractical) way: proceed as in the proof of the theorem up to the use of Corollary 2.1. Instead of using this result, we construct by brute force a punctuation sequence for the ergodic component in effect. Suppose that G = {Gi ; i = 1, 2, . . .} is a countable generating field for the input sequence space. Given δ, the infinite length sliding-block encoder first finds the smallest value of i for which n−1 1 X 1Gi (T k x), n→∞ n k=0
0 < lim and
n−1 1 X 1Gi (T k x)ρ(xk , a∗ ) ≤ δ, n→∞ n k=0
lim
that is, we find a set with strictly positive relative frequency (and hence strictly positive probability with respect to the ergodic component in effect) which occurs rarely enough to ensure that the sample average distortion between the symbols produced when Gi occurs and the reference letter is smaller than δ. Given N and δ there must exist an i for which these relations hold (apply the proof of Lemma 2.6 to the ergodic component in effect with γ chosen to satisfy (12.37) for that component and then replace the arbitrary set G by a set in the generating field having very close probability). Analogous to the proof of Lemma 2.6 we construct a punctuation sequence {Zn } using the event Gi in place of G. The proof then follows in a like manner except that now from the dominated convergence theorem we have that
332
12 Source Coding Theorems n−1 1 X E(12 (Zi )ρ(Xi , a∗ ) n→∞ n i=0
E(12 (Z0 )ρ(X0 , a∗ )) = lim
n−1 1 X 12 (Zi )ρ(Xi , a∗ )) ≤ δ n→∞ n i=0
= E( lim
by construction. The above argument is patterned after that of Davisson and Gray [28] and extends the theorem to stationary nonergodic sources if infinite window sliding-block encoders are allowed. We can then approximate this encoder by a finite-window encoder, but we must make additional assumptions to ensure that the resulting encoder yields a good approximation in the sense of overall distortion. Suppose that f is the infinite window length encoder and g is the finite window-length (say 2L + 1) encoder. Let G denote a countable generating field of rectangles for the input sequence space. Then from Corollary 5.1 applied to G given > 0 we can find for sufficiently large N a finite window sliding-block code r : A2N+1 → B such that Pr(r 6= f 0 ) ≤ /(2L+1), that is, the two encoders produce the same channel symbol with high probability. The issue is when does this imply that ρ(f g, µ) and ρ(r g, µ) are therefore also close, which would complete the proof. Let r : AT → B denote the infinite2N+1 window sliding block encoder induced by r , i.e., r (x) = r (x−N ). Then Z X ˆ0 )) = ρ(f g, µ) = E(ρ(X0 , X dµ(x)ρ(x0 , g(b)), b∈B 2L+1
x∈Vf (b)
where Vf (b) = {x : f (x)2L+1 = b} and f (x)2L+1 is shorthand for f (xi ), i = −L, . . . , L, that is, the channel (2L + 1)-tuple produced by the source using encoder x. We therefore have that X Z ρ(r g, µ) ≤ dµ(x)ρ(x0 , g(b)) b∈B 2L+1
+
x∈Vf (b)
X
Z
b∈B 2L+1
= ρ(f , µ) +
x∈Vr (b)−Vf (b)
X b∈B 2L+1
≤ ρ(f , µ) +
X b∈B 2L+1
dµ(x)ρ(x0 , g(b))
Z x∈Vr (b)−Vf (b)
dµ(x)ρ(x0 , g(b))
Z x∈Vr (b)∆Vf (b)
dµ(x)ρ(x0 , g(b)).
By making N large enough, however, we can make
12.9 A Geometric Interpretation
333
µ(Vr (f )∆Vf (b)) ˆ2L + 1 and hence force all arbitrarily small simultaneously for all b ∈ A of the integrals above to be arbitrarily small by the continuity of integration. With Lemma 12.6 and Theorem 12.6 this completes the proof of the following theorem. Theorem 12.7. Theorem 11.7.2: Given an AMS source µ and an additive fidelity criterion with a reference letter, δSBC (R, µ) = δ(R, µ), that is, the class of sliding-block codes is capable of exactly the same performance as the class of block codes. The sliding-block source coding theorem immediately yields an alternative coding theorem for a code structure known as trellis encoding source codes wherein the sliding-block decoder is kept but the encoder is replaced by a tree or trellis search algorithm such as the Viterbi algorithm [44]. Details can be found in [53] and an example is discussed in Section 13.3.
12.9 A Geometric Interpretation We close this chapter on source coding theorems with a geometric interpretation of the operational DRFs in terms of the ρ distortion between sources. Suppose that µ is a stationary and ergodic source and that {ρn } is an additive fidelity criterion. Suppose that we have a nearly optimal sliding-block encoder and decoder for µ and a channel with rate R, that ˆn } and is, if the overall process is {Xn , X ˆ0 ) ≤ δ(R, µ) + δ. Eρ(X0 , X If the overall hookup (source/encoder/channel/decoder) yields a distriˆn } and distribution η on the reproduction process bution p on {Xn , X ˆ {Xn }, then clearly ρ(µ, η) ≤ δ(R, µ) + δ. Furthermore, since the channel alphabet is B the channel process must have entropy rate less than R = log ||B|| and hence the reproduction process must also have entropy rate less than B from Corollary 6.4. Since δ is arbitrary, δ(R, µ) ≥ inf ρ(µ, η). η:H(η)≤R
334
12 Source Coding Theorems
Suppose next that p, µ and η are stationary and ergodic and that H(η) ≤ R. Choose a stationary p having µ and η as coordinate processes such that Ep ρ(X0 , Y0 ) ≤ ρ(µ, ν) + δ. We have easily that I(X; Y ) ≤ H(η) ≤ R and hence the left hand side is bounded below by the process distortion rate function D s (R, µ). From Theorem 9.1 and the block source coding theorem, however, this is just the operational distortion-rate function. We have therefore proved the following result [63]. Theorem 12.8. Let µ be a stationary and ergodic source and let {ρn } be an additive fidelity criterion with a reference letter. Then δ(R, µ) =
inf
ρ(µ, η),
(12.38)
η:H(η)≤R
that is, the operational DRF (and hence the distortion-rate function) of a stationary ergodic source is just the “distance” in the ρ sense to the nearest stationary and ergodic process with the specified reproduction alphabet and with entropy rate less than R.
Chapter 13
Properties of Good Source Codes
Abstract Necessary conditions for a source code to be optimal or a sequence of source codes to be asymptotically optimal for a stationary source are developed for block and sliding-block codes.
13.1 Optimal and Asymptotically Optimal Codes In Eq. 5.14 the operational distortion-rate function (DRF) for the source µ, channel ν, and code classes E and D was defined by ∆(µ, ν, E, D) =
inf
f ∈E,g∈D
∆(µ, f , ν, g).
(13.1)
This chapter considers only source codes and hence the channel ν is assumed to be noiseless. A source code (f , g) is said to be optimal if it achieves the infimum, that is, if f ∈ E, g ∈ D, and ∆(µ, f , ν, g) = ∆(µ, ν, E, D).
(13.2)
Optimal codes might not exist, but from the definition of infimum we can always get close. Hence we define a sequence (fn , gn ), n = 1, 2, . . . to be asymptotically optimal or a.o. if lim ∆(µ, fn , ν, gn ) = ∆(µ, ν, E, D).
n→∞
(13.3)
This chapter is concerned with developing the implications of a code being optimal or a sequence of codes being asymptotically optimal (which can be interpreted as looking at good or nearly optimal codes). Note that any property obtained for a.o. codes implies a result for optimal codes if they exist by setting fn = f and gn = g for all n. We usually consider optimal and asymptotically optimal separately since the former is simpler
R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_13, © Springer Science+Business Media, LLC 2011
335
336
13 Properties of Good Source Codes
when it is applicable. These implications are in terms of necessary conditions for optimality or asymptotic optimality. The conditions describe attributes of encoders, decoders, and the distributions of the encoded and reproduction sequences. In the special case of squared-error distortion, the behavior of second order moments of the reproduction and the error sequence are quantified. We confine interest to stationary sources in order to keep things relatively simple. In (12.1-12.2) the definition of operational rate-distortion function was specialized to the case of block codes of dimension N and code rate R and to block codes of arbitrary dimension and code rate R, definitions which we repeat here to have them handy: δ(R, µ) = ∆(µ, ν, E, D) = inf δN (R, µ), N
δN (R, µ) =
inf
C∈K(N,R)
ρ(C, µ),
where ν is a noiseless channel as described in Section 12.2, E and D are classes of block codes for the channel, and K(N, R) is the class of all block length N codebooks C with 1 log ||C|| ≤ R. N
(13.4)
It was there argued that given a decoder of block codes of length N, an optimal encoder in the sense that no encoder could do better is given by a minimum distortion search of the decoder codebook. This observation is the original example (in Shannon [163]) of an optimality property of a source code — a necessary condition for a block code to be optimal is that the encoder be a minimum distortion (or “nearest neighbor”) mapping, at least with probability 1. Shannon defined his source codes to have this property. Here we allow a more general definition of an encoder to show how a decoder can be optimized for a fixed (but not necessarily optimum) encoder, but observe a necessary condition for overall optimality is that the encoder have this property, that is, that the encoder be a minimum distortion search matched to the decoder. The introduction of this extra degree of freedom results in several useful properties, analogous to the introduction of the extra distribution η in the evaluation of rate-distortion functions which led to useful conditions for an optimization and an alternating optimization algorithm in Section 9.5. Block codes have other such optimality properties, many of which were first observed in the scalar (quantization) case by Lloyd [110] and in the vector case by Steinhaus [175].
13.2 Block Codes
337
13.2 Block Codes Block source codes are also called block quantizers, multidimensional quantizers, and vector quantizers since they discretize or quantize a continuous (or large discrete) space into a relatively small discrete index set, and the indices are then decoded into approximations of the original input vector. While the emphasis of this book is on sliding-block codes, we treat the well-known optimality properties for block codes for two reasons. First, one way to prove sliding-block coding theorems is to embed a block code into a Rohlin tower to construct a stationarized version of the block code with approximately the same per-symbol average distortion. Thus having a good block code can lead to a good sliding-block code. Some proofs of the sliding-block coding theorems avoid directly using block codes by taking advantage of results from ergodic theory, but the ergodic theory results usually use block constructions in their proofs as well. Second, the optimality properties of block codes are simpler to state and prove and they provide some interesting comparisons with the stationary code results of the next section. This section treats well-known properties and methods for developing the properties for vector quantizers with notational changes as needed to be consistent with the book. The reader is referred to [50, 71] for further discussion. The usual formulation for a block source coder for a source X with alphabet AX involves two mappings, an encoder α : AN X → I, where I is an index set of size M = kIk, and a decoder β : I → C, where it is usually assumed that I is either a set of integers {0, 1, . . . , M − 1} (common when the block code is considered outside the context of a communications channel) or is a sequence space of channel symbols I = AK U , where AU = {0, 1} for binary channel codes or AU is some other finite set of the form {0, 1, . . . , m − 1}, in which case M = kAU kK . The collection of reproduction words C is called the reproduction codebook (or simply codebook if the usage is clear from context) and it is usually assumed for convenience that the decoder is a one-to-one mapping of indices into distinct reproduction codewords so that C = {β(i); i ∈ C} so that kCk = M. The code rate or transmission rate of the code is defined by R = log M, where the units are bits per input vector if the logarithm is base 2 and nats per vector if it is base e. It is common to consider the normalized or per-symbol code rate of R=
1 log M. N
338
13 Properties of Good Source Codes
Context should make clear if it is the normalized or unnormalized rate being considered. In this section we emphasize a fixed N and hence usually do not normalize. If I = AK U , then the normalized rate is R=
K log kAU k. N
In the important simple special case where the code is binary and kAU k = 2, then R = K/N, the number of binary channel symbols produced by the encoder for each N channel symbols put into it. We simplify the vector notation in this section by dropping the subscripts N and assume that X is a random vector with sample values x chosen in a vector alphabet AX . Rates will not be normalized and the dimension will be implicit. An encoder α is equivalent to a partition of the vector input space AX defined by P = {Pi ; i ∈ I}, where Pi = {x ∈ AX : α(x) = i}. The partition notation provides a useful representation of the encoder as X α(x) = 1Pi (x). i∈I
Similarly a decoder β is described by its reproduction codebook C. Assuming that no errors are made in the transmission of the channel codeword, then for most properties the specific nature of the index set I is unimportant and all that matters is the number of elements M and the codewords in C. The combined operation of an encoder and decoder is often referred to simply as a quantizer or vector quantizer. We will use Q to denote both the pair Q = (α, β) and the overall operation defined by Q(x) = β(α(x)). Two quantizers will be said to be equivalent if they have the same index set I and yield the same overall mapping Q. The rate of the quantizer (unnormalized) is given by R = R(Q) = log(α(AX )), the log of the size of the index set (or reproduction codebook). The principal goal in the design of vector quantizers is to find a codebook (decoder) and a partition (encoder) that minimizes an average distortion with respect to a distortion measure d. The average distortion for a vector quantizer Q = (α, β) applied to a random vector X with distribution µ is Z (13.5) ∆(α, β) = E[ρ(X, Q(X))] = ρ(x, Q(x))dµ(x).
13.2 Block Codes
339
Define as earlier the operational distortion-rate function ∆(µ) by ∆(R) =
inf
α,β:R(α)≤R
∆(α, β).
(13.6)
A code is optimal if ∆(α, β) = ∆(R). Recall that at present everything is for a fixed dimension N and that the clutter of more notation will be necessary to make N explicit when it is allowed to vary, for example if we wish to quantify the long term performance when the block code is applied to an AMS source as in Lemma 12.1. The most fundamental of the conditions for optimality of a quantizer (α, β) follows from the obvious inequality Z ∆(α, β) = E[ρ(X, Q(X))] = ρ(x, Q(x))dµ(x) Z ≥ min ρ(x, y)dµ(x), (13.7) y∈C
where the minimum exists since the codebook is assumed finite. This unbeatable lower bound to the average distortion for a quantizer with reproduction codebook C and hence for a given decoder is achieved with equality if the encoder is defined to be the minimum distortion encoder: ˆi ), α(x) = argmin ρ(x, x
(13.8)
i∈I
ˆi ; i ∈ I}. The encoder is not yet well defined in a strict where C = {x sense because there can be ties in distortion, in which case the encoder has to chose among multiple indices yielding the same distortion. In this case any tie-breaking rule can be used without affecting the distortion, for example choose the index lowest in lexicographical order. It can be assumed that the optimal encoder is of this form since, if it were not, changing to a minimum distortion encoder can not increase the average distortion. In the classic paper on source coding with a fidelity criterion [163], Shannon assumed that encoders were of this form. We do not make that assumption since it is useful to consider performance of a quantizer as a function of decoder (codebook) and encoder (partition) separately, but when all is said and done the best choice for an encoder (in terms of minimizing average distortion) is the minimum distortion encoder. An alternative means of describing an optimal encoder is in terms of the encoder partition by ˆi ) ≤ ρ(x, x ˆj ); j ≠ i}, Pi ⊂ {x : ρ(x, x
340
13 Properties of Good Source Codes
where again some tie-breaking rule is needed to assign an index to points on the border. Next suppose that the encoder α or its partition P is specified. Given ˆ (if it exists) any subset S ⊂ AX of probability µ(S) > 0, the point z ∈ A for which E[d(X, z) | X ∈ S] = inf E[d(X, y) | X ∈ S] ˆ y∈A
is called the centroid of S and denoted by cent(S). If the set S has zero probability, then the centroid can be defined in an arbitrary fashion. The name reflects the origins of the word as the centroid or center of gravity of a set in Euclidean space with respect to the squared-error distortion and Lebesgue measure. If the centroid exists, then cent(S) = argmin E[d(X, y) | X ∈ S]. ˆ y∈A
Centroids exist for many distortion measures and sets of interest. For example, if the AX = RN and the distortion is additive squared error (the square of the `2 norm of the vector difference), then the centroid is given by the conditional expectation E[X | X ∈ S], the minimum meansquared estimate of the source vector given the event X ∈ S. Other interesting distortion measures with centroids are considered in [50]. If the appropriate centroids exist, then the properties of conditional expectation yield the inequality X ˆi )1Pi (X)] E[d(X, Q(X))] = E[d(X, x i∈I
=
X
ˆi ) | X ∈ Pi ]µ(Pi ) E[d(X, x
i∈I
≥
X
cent(Pi )µ(Pi ),
(13.9)
i∈I
which holds with equality if the decoder output or reproduction codeword assigned to i is the centroid of Pi . In the squared error case this has the intuitive interpretation of being the minimum mean-squared estimate of the input given the received index. These two conditions were developed for squared error for vectors by Steinhaus [175] and in the scalar case for more general distortion measures by Lloyd [110] and they have since become known as the Lloyd conditions and quantizers which satisfy the conditions are sometimes referred to as Lloyd-optimal quantizers, although satisfaction of the two conditions does not ensure global optimality. Each condition, however, ensures conditional optimality given the other component of the quantizer and hence, as both Lloyd and Steinhaus observed, they provide an iterative algorithm for improving the quantizer performance. Begin-
13.2 Block Codes
341
ning with a distribution, distortion measure, and initial reconstruction codebook, the optimal encoder is a minimum distortion mapping. The decoder can then be replaced by the centroids. Each step can only decrease or leave unchanged the average distortion, hence the algorithm is a descent algorithm with respect to average distortion. The distribution can be an empirical distribution based on a training set of data, which makes the algorithm an early example of clustering and statistical (or machine) learning. The idea has been rediscovered in many fields, perhaps most famously a decade later as the “k-means” clustering algorithm of MacQueen [112]. This was one of the first examples of what has since become known as an alternating optimization (AO) algorithm, where a complicated optimization (e.g., nonconvex) can be broken down into two separate simpler optimizations [14]. Taken together, the conditions yield the following lemma. Lemma 13.1. Lloyd Quantizer Optimality Properties An optimal quantizer must satisfy the following two conditions (or be equivalent to a quantizer which does): Optimum encoder for a given decoder Given a decoder with reproˆi ; i ∈ I}, the optimal encoder satisfies duction codebook C = {x ˆi ). α(x) = argmin ρ(x, x
(13.10)
i∈I
Optimum decoder for a given encoder mal decoder satisfies
Given an encoder α, the opti-
β(i) = cent({x : α(x) = i}).
(13.11)
The implications of these optimality properties for the design of vector quantizers are explored in depth in the literature. For example, see [50, 61, 115, 128, 152, 71]. Two observations merit making. The decoder for a block code is simple in principal, it is simply a table lookup. A channel codeword provides an index and the output is the reproduction codeword in the codebook with that index. The encoder, however, requires a minimum distortion search of the reproduction codebook to find the best fit to the observed input vector. Search can be costly, however, especially as the dimension grows. There is a large literature on techniques for avoiding a full search of a codebook, from forcing a structure on the codebook (such as a tree or trellis or as a linear combination of basis vectors) to performing only a partial or approximate search. See, e.g., [71].
342
13 Properties of Good Source Codes
Moment Properties In the special case of the squared-error distortion where d(x, y) = kx − yk2 =
N−1 X
(xi − yi )2 ,
i=0
where N is the dimension of the vectors x, y, optimal quantizers have several additional interesting properties in terms of their moments. In fact, the quantizers need not be optimal to have these properties, the only requirement is that they satisfy the centroid condition so that the codewords are given by ˆi = E[X | α(X) = i], i ∈ I. x The following lemma collects these conditions. The results are simple consequences of basic properties of vector linear prediction. The lemma follows [50] Lemma 11.2.2. See also the scalar special case in Lemma 6.2.2. Lemma 13.2. A vector quantizer which satisfies the centroid condition for the squared-error distortion measure has the following properties: 1. 2. 3. 4.
E(Q(X)) = E(X) E(X t Q(X)) = E(kQ(X)k2 ) E((X − Q(X))t Q(X)) = 0 E(kQ(X)k2 ) = E(kXk2 ) − E(kX − Q(X)k2 ).
Proof. Let P = {Pi , i ∈ I} denote the partition associated with the partiˆi ; i ∈ I} the reproduction codebook. By assumption, the tion and C = {x ˆi are the centroids of the the partition cells Pi , which for squared error x are the conditional expectations of X given X ∈ Pi . Using conditional expectation, X X X X 1Pi (X) = E(X1Pi (X)) E(X) = E
=
X
i∈I
i∈I
µX (Pi )E(X | X ∈ Pi ) =
i∈I
X
ˆi = E[Q(X))], µX (Pi )x
i∈I
proving the first property. Since Q(X) =
X i∈I
we have that
ˆi 1Pi (X), x
13.3 Sliding-Block Codes
343
E(X t Q(X)) = E(X t
X
ˆi 1Pi (X)) = x
i∈I
=
X
X
ˆi E(1Pi (X)X t )x
i∈I
X ˆi ˆi = ˆit x µX (Pi )E(X | X ∈ Pi )x µX (Pi )x t
i∈I
=
X
i∈I
ˆit x ˆi = kQ(X)k2 , µX (Pi )x
i∈I
proving the second property. The third property follows from the second. The final property follows from expanding the left hand side and using the second and third properties E(kX − Q(X)k2 ) = E (X − Q(X))t (X − Q(X)) = E X t (X − Q(X)) − E Q(X)t (X − Q(X)) = E(kXk2 ) − E(X t Q(X)) − 0 = E(kXk2 ) − E(kQ(X)k2 ). .
2
The lemma has the intuition that the centroid condition is sufficient to ensure that the quantizer is an unbiased estimator of the input given the index (or the quantized value itself). The second property shows that the correlation of the quantizer output and the input equals the energy in the quantizer output. In particular, the input and output are not uncorrelated. and hence can not be independent. This conflicts directly with the frequently assumed model in the communications and signal processing literature where quantizer error is treated as signal-independent white noise. The third property shows that the error and the estimate quantizer output (which is an estimate of the input given the quantizer index) have 0 correlation, which is simply an example of the orthogonality property since, for a fixed encoder, Q(X) is an optimal linear estimate for X given Q(X).
13.3 Sliding-Block Codes A sliding-block code (f , g) for source coding is said to be optimum if it yields an average distortion equal to the operational distortion-rate function, ∆(f , g) = ∆X (R). Unlike the simple scalar quantizer case (or the nonstationary vector quantizer case), however, there are no simple conditions for guaranteeing the existence of an optimal code. Hence usually it is of greater interest to consider codes that are asymptotically optimal in the sense that their performance approaches the optimal in the limit, but there might not be a code which actually achieves the limit. Before
344
13 Properties of Good Source Codes
considering asymptotic optimality, consider the natural extensions of the quantizer optimality properties to a sliding-block code. Sliding-block decoders do have a basic similarity to quantization decoders in that one can view the contents of the decoder shift register as an index i in an index set I and the decoder is a mapping of I into a single reproduction ˆ say. Usually in practical systems the decoder has fisymbol g(i) ∈ A, nite length, in which case the index set I is finite. If the encoder is given, then the mapping of inputs into indices at a particular time is fixed, and hence the centroid condition must still hold if the centroids exist for the distortion measure being used. This yields the following lemma. Lemma 13.3. If the encoder f in a sliding-block source coder is fixed and results in decoder shift register contents U ∈ I at time 0, then a necessary condition for the decoder g to be optimal with respect to the encoder is that g(i) = cent(X0 | U = i) ≡ argmin E ρ(X0 , y) | U = i (13.12) ˆ y∈A
The lemma follows in the same manner as the quantizer result, and as in that case it implies for the squared-error distortion case that g(i) be the conditional expectation of the input at the same time given the contents of the shift register and knowledge of the encoder mapping of inputs into indexes. The application of Lloyd’s centroid condition for quantizers to sliding-block decoders was first treated by Stewart [176, 177]. This property can be used to tune a decoder to a training sequence using an empirical conditional expectation. Unfortunately the corresponding result for the encoder — that the minimum distortion rule is the optimal encoder for a fixed decoder — does not have a simple extension to sliding-block codes. There is, however, a hybrid system that couples a sliding-block decoder with a block encoder which does resemble the Lloyd alternating optimization for vector quantization, and it has the added advantage that the minimum distortion search required of the encoder does not require computational complexity increasing exponentially with the block length of the code. The trick is that the sliding-block coder structure allows a low complexity minimum distortion search. The technique is known as trellis source encoding and it was introduced by by Viterbi and Omura [188] and the connections with sliding-block codes and the Lloyd iteration were developed by Stewart et al. [53, 176, 177]. A brief overview is presented here to illustrate the similarities among the Lloyd conditions, sliding-block, and block codes. Details can be found in the cited references. To keep things simple, we focus on one bit per symbol codes so that the noiseless channel alphabet is binary and one reproduction symbol is produced for each channel bit. Consider the simple sliding-block code of Figure 2.1 with a more general decoder function g as in Fig-
13.3 Sliding-Block Codes
345
ure 13.1. The temporal operation of the decoder can be depicted as in
{Un }
- Un
Un−1 Un−2
@ R ? @ g
? ˆn = g(Un , Un−1 , Un−2 ) X Fig. 13.1 A simple sliding-block decoder
Figure 13.2, a directed graph called a trellis. The decoder is viewed as a finite-state machine where at time n the state consists of received channel symbols in the shift register except for the most recent — in this case Sn = (Un−1 , Un−2 ) so that there are four possible states {00, 01, 10, 11}. The states at a particular time are represented by the darkened circles stacked vertically and labeled on the far left. If at time n the shift register is in state sn = (un−1 , un−2 ) and a channel symbol un is received, then the output will be g(un , un−1 , un−2 ) and the state will change to sn+1 = (un , un−1 ). Thus the next-state rule given the current state and the current received channel symbol can be described in a state transition table as in Table 13.1. In general there is a next state rule sn 11 10 01 00 11 10 00 01 un 1 1 1 1 0 0 0 0 sn+1 11 11 10 10 01 01 00 00 Table 13.1 State transition table
sn+1 = r (un , sn ). The state transistions are noted in the trellis by connecting the states between times by a branch which is labeled by the decoder output produced by the transition between the two states connected by the branch. In the figure the upper branch shows the transition if the channel symbol is a 1, the lower branch shows the transition if the channel symbol is a 2. The picture for time n is replicated at every time instant. The leftmost column of states can be considered as time 0 and the trellis continues to replicate to the right. Continuing with this simple example, suppose that the decoder g is fixed, and we want to design a good encoder. While the end goal may be another sliding-block code to match the theoretical emphasis, suppose for the time being that a block
346
13 Properties of Good Source Codes
g(111)
11
v A A g(011) A
g(111)
g(111)
-v A A g(011) A
-v A A
g(011) A
g(111)
-v A A
g(111)
-v A A
-v
g(011) A g(011) A A A A g(110) g(110) g(110) g(110) g(110) A A A A v v v v v A v 10 A A A A A @ @ @ @ @ @ A g(010) @ A g(010) @ A g(010) @ A g(010) @ A g(010) @ A @ A @ A @ A @ A A A A A g(101) g(101) g(101) g(101) g(101) A @ @ @ @ @ A A A A A @ @ @ @ @ RA @ RA @ RA @ RAU v @ RAU v @ Uv Uv Uv v 01 @ @ @ @ @ g(001) @ g(001) @ g(001) @ g(001) @ g(001) @ @ @ @ @ @ g(100) g(100) g(100) g(100) g(100) @ @ @ @ @ @ @ @ @ @ R @ R @ R @ R @ Rv @ v -v -v -v -v 00
A
A
g(000)
g(000)
g(000)
g(000)
g(000)
Fig. 13.2 Trellis diagram of sliding-block code
encoder is to be used. In that case, the trellis provides a picture of all of the reproduction sequences available to the decoder, so the general principal would be to observe an input sequence of length, say, L, and find the best possible sequence of trellis branch labels through the trellis for L steps, that is, find the minimum distortion path through the trellis. An immediate issue how to initiate the encoding algorithm. In particular, if we allow the search algorithm to consider all possible initial states, then at the beginning of the block the decoder must be told what state to begin in, that is, what paths through the trellis are allowed. The optimal choice might be to place no constraint on the start, but this means that many bits would need to be sent at the beginning of a block, while only one bit per input sample would be required thereafter. The usual practical solution is to pick an arbitrary initial state, say the all-zero state, to begin with. Then the path through the trellis can be sent with one bit per branch for the current and future blocks. Hopefully the effects of an arbitrary and possibly bad initial state will wash out with time. Given that the encoder and decoder initialize their states to a common state, say σ0 , then as in the block source code case, the optimal encoder for a source block x L of length L will choose the sequence of bits uL that drives the decoder through the trellis so as to yield the smallest possible total distortion; that is, the encoder will find argmin uL
L−1 X i=0
ρ(xi , g(ui , si )).
13.3 Sliding-Block Codes
347
where the si are computed from the previous states and the inputs using Table 13.1. What makes this structure so useful is that instead of having to compute the distortions for all 2L possible reproduction L tuples with respect to an observed input tuple, the minimum distortion path can be found sequentially in time by a simple algorithm, as will be seen. An immediate point is that if the decoder is a good sliding-block code, then a natural choice for an encoder is a minimum distortion search through all possible decoder sequences, that is, find the binary channel sequence that will drive the decoder through the sequence of outputs that provide a good match to an observed input sequence. The 2L possible reproduction L-tuples correspond to the 2L possible sequences of reproduction symbols or branch labels resulting from the 2L possible binary path maps through the trellis from the initial state to one of the possible final states. A brute force minimization would be to compute for each of these 2L path maps and their distortion with respect to the input x L and choose the binary sequence which results in the minimum distortion sequence of branch labels. The trellis structure decreases the work, however, by eliminating the need for computing the distortion for all possible paths through the trellis. Many of the paths are bad and can not be candidates for the best path. Suppose that we know the best paths into each of the 2K−1 possible states at time n = L − 1 along with the resulting total distortion resulting from the associated sequence of reproductions/branch labels. Suppose that υL ≡ argmin uL
L−1 X
ρ(xi , g(ui , si ))
i=0
is the best binary path map and that the resulting distortion is ∆L =
L−1 X
ρ(xi , g((υL )i , si )) = min uL
i=0
L−1 X
ρ(xi , g(ui , si )),
i=0
where as always the state sequence is determined from the path map. As the notation is cluttered enough already, this dependence is not shown explicitly. Consider a time n < L. An optimal overall path map υL must have resulted in the decoder being in some particular state, say s, at time n, and in a binary path map υn (s) that resulted in the decoder being in the state s at time n, and in a running distortion of ∆n (s) =
n−1 X
ρ(xi , g((υL )i , si )),
i=0
with sn = s = r (un−1 , sn−1 ). Furthermore, this path must have been the best path from the initial state σ0 through the state s at time n; that is,
348
13 Properties of Good Source Codes
(υL0 , υL1 , . . . , (υL )n−1 ) = υn (s) ≡ argmin un
∆n =
n−1 X
n−1 X
ρ(xi , g(ui , si ))
i=0
ρ(xi , g((υL )i , si )) = min n
i=0
u
n−1 X
ρ(xi , g(ui , si )).
i=0
This follows from the additive nature of distortion since if the length n prefix of υL , ((υL )0 , (υL )1 , . . . , (υL )n−1 ), were not the minimum distortion path from the intitial state to state s at time n, then there would be another path into this node with strictly smaller distortion. Were that the case, however, that other path would be a better prefix to the overall path and it would yield smaller total distortion by yielding smaller distortion in the first n-tuple and no worse distortion for the remainder of the path. The remainder of the path is not changed because the choices from time n on depend only on the state of the decoder at time n. This argument assumes that we know the state through which the optimal path passes. In general we can only say before the optimal path is known that at time n the decoder must be in some state. This implies, however, that at time n at each state all we need track is what the best path so far into that state is and what the corresponding distortion is. All inferior paths into the state can be discarded as no longer being candidates for prefix of an optimal path map. This yields a search algorithm for finding the minimum distortion path through the trellis, which can be described informally as follows. Step 0 Given: input sequence x N , length K sliding-block decoder g, initial state σ0 . State space S = all 2K−1 binary (K − 1)-tuples. Next state mapping: If the old state is s = (b0 b1 · · · , bk−2 , bK−1 ) and the received channel symbol is u, then the next state is r (u, s) = (u, b0 , b1 , · · · , bk−2 ). Define υ0 (s) to be the empty set for all s ∈ S. Define ∆n (s) = 0 for all s ∈ S. Set n = 1. Step 1 For each s ∈ S: There can be at most two previous states sn−1 = σ0 and σ1 , say, for which sn = s = r (0, σ0 ) = r (1, σ1 ) and for which sn−1 is reachable from the initial state. (For n ≥ K − 1 all states at time n are reachable from the initial state.) If there are two such states, compare δ0 = ∆n−1 (σ0 ) + ρ(xn , g(0, σ0 )) with δ1 = ∆n−1 (σ1 ) + ρ(xn , g(1, σ1 )). If δ0 ≤ δ1 , then set υn (s) = (υn−1 (σ0 ), 0) ∆n (s) = ∆n−1 (σ0 ) + ρ(xn , g(0, σ0 )), otherwise set
13.3 Sliding-Block Codes
349
υn (s) = (υn−1 (σ1 ), 1) ∆n (s) = ∆n−1 (σ1 ) + ρ(xn , g(1, σ1 )). That is, choose the minimum distortion path available from the choice of two paths entering state s at time n. This extends one of the two candidate paths of length n − 1 available at the two allowed previous states and extinguishes the other, which can not be the suffix of an optimal path. If there is only one allowed previous state σ , choose the path from that single state using the update formula above. If there is no allowed previous state, do nothing (there is no update to a best path or associated distortion to the given state at time n). Set n ← n + 1. Step 2 If n < L, go to Step 1. If n = L, we have the best path maps υL (s) and corresponding distortions ∆L (s) at time L − 1 for all states s ∈ S. Set s ∗ = argmin ∆L (s) s∈S
and finish with binary path map υL (s ∗ ) as the encoded bit sequence to be communicated to the receiver to drive the decoder to produce the reproduction. Instead of computing 2L separate distortions for L-tuples, the algorithm computes 2K−1 incremental distortions at each time (level in the trellis) and does the corresponding addition to maintain and store the cumulative distortion up to that time. This is done for each of the L levels of the trellis. Thus the algorithm complexity grows roughly linearly and not exponentially in the encoder block size L, but it does grow with the state space size and hence exponentially with the decoder shift register length. Thus a relatively small decoder shift register length with a large block length can yield an overall reasonable complexity. As of this writing (2010) shift register lengths of over 20 and block lengths of a million are reasonable. Decoding is of minimal complexity. The basic optimality principle used here as that of dynamic programming, and its applications to channel coding and source coding were introduced by Andrew Viterbi. The algorithm is widely known as the Viterbi algorithm in communications and signal processing. See, e.g., [44]. At this point we have a hybrid code with a sliding block decoder and a block encoder. The block encoder can be made stationary using occasional input-dependent spacing as in the theoretical constructions of sliding-block codes from block codes, but in practice the Viterbi algorithm is usually run on very long blocks. There are many variations on the basic approach.
350
13 Properties of Good Source Codes
Asymptotically Optimal Sliding-Block Codes This subsection largely follows Mao, Gray, and Linder [117, 72]. A sequence of rate-R sliding-block codes fn , gn , n = 1, 2, . . ., for source coding is asymptotically optimal (a.o.) if lim ∆(fn , gn ) = ∆X (R) = DX (R).
n→∞
(13.13)
An optimal code (when it exists) is trivially asymptotically optimal and hence any necessary condition for an asymptotically optimal sequence of codes also applies to a fixed code that is optimal by simply equating every code in the sequence to the fixed code. Similarly, a simulation code g is optimal if ρ(µX , µg(Z) ) = ∆(X|Z) and a sequence of codes gn is asymptotically optimal if lim ρ(µX , µg n (Z) ) = ∆X|Z .
n→∞
(13.14)
Process approximation The following lemma provides necessary conditions for asymptotically optimal codes. The results are a slight generalization and elaboration of Theorem 1 of Gray and Linder [72] as given in [117]. Lemma 13.4. Given a real-valued stationary ergodic process X, suppose that fn , gn n = 1, 2, . . . is an asymptotically optimal sequence of stationary source codes for X with encoder output/decoder input alphabet B of size kBk = 2R for integer rate R. Denote the resulting reproduction proˆ(n) and the B-ary encoder output/decoder input processes by cesses by X (n) U . Then lim ρ(µX , µXˆ(n) ) = DX (R)
n→∞
ˆ(n) ) = lim H(U (n) ) = R lim H(X
n→∞
lim d(U
n→∞
n→∞
(n)
, Z) = 0,
where Z is an IID equiprobable process with alphabet size 2R . These properties are quite intuitive: • The process distance between a source and an approximately optimal reproduction of entropy rate less than R is close to the Shannon distortion rate function. Thus frequency-typical sequences of the reproduction should be as close as possible to frequency-typical source sequences.
13.3 Sliding-Block Codes
351
• The entropy rate of an approximately optimal reproduction and of the resulting encoded B-ary process must be near the maximum possible value in order to take advantage of all possible available information. • The sequence of encoder output processes approaches an IID equiprobable source in the Ornstein process distance. If R = 1, the encoder output bits should look like fair coin flips. Proof. The encoded and decoded processes are both stationary and ergodic since the original source is. From (12.38) and the source coding theorem, (n)
∆(fn , gn ) = E[d(X0 , X0 )] ≥ ρ(µX , µXˆ(n) ) ≥
inf
ν:H(ν)≤R
ρ(µX , ν) = ∆(µ, R) = DX (R).
The first inequality follows since the ρ distance is the minimum average distortion over all couplings yielding the marginal distributions for X0 (n) and X0 . The second inequality follows since stationary coding reduces ˆ(n) ). Since the leftmost term entropy rate so that R ≥ H(U (n) ) ≥ H(X converges to the rightmost, the first equality of the lemma is proved. From Lemma 8.4 ˆ(n) ) ≥ I(X, X ˆ(n) ). R ≥ H(X From the process definition of the rate-distortion function, the dual to the process definition of the distortion-rate function (the rate-distortion ˆ(n) ) ≥ RX (∆(fn , gn )). formulation can be found in [119, 63, 76]), I(X, X Taking the limit as n → ∞, RX (∆(fn , gn )) converges to R since the code sequence is asymptotically optimal and the Shannon rate-distortion function is a continuous function of its argument except possibly at D = ˆ(n) ) = R. 0, the dual of Lemma 9.1. Thus limn→∞ H(U (n) ) = limn→∞ H(X proving the second equality of the lemma. From Marton’s inequality of Corollary 6.6, N −1 dN (µU N , µZ N ) ≤
1/2
ln 2 (NR − H(U N )) 2N
and taking the limit as N → ∞ using property (a) of Theorem 5.2 yields d(µU , µZ ) ≤
1/2
ln 2 (R − H(U)) 2
.
Applying this to U (n) and taking the limit using the previous part of the lemma completes the proof. 2 If X is a B-process, then a sequence of a.o. simulation codes gn yield˜(n) satisfies limn→∞ ρ(µX , µX˜(n) ) = ∆X|Z = ing a reproduction processes X DX (R) and a similar argument to the proof of the previous lemma imˆ(n) ) = H(Z) = R. plies that limn→∞ H(X
352
13 Properties of Good Source Codes
Moment conditions The next set of necessary conditions concerns the squared-error distortion and resembles the standard result for scalar and vector quantizers described in Lemma 13.2. The proof differs, however, in that in the quantization case the centroid property is used, while here simple ideas from linear prediction theory accomplish a similar goal. Define in the usual way the covariance COV(X, Y ) = E[(X − E(X))(Y − E(Y ))]. Lemma 13.5. Given a real-valued stationary ergodic process X, suppose that fn , gn is an asymptotically optimal sequence of codes (with respect ˆ(n) with entropy rate to squared error) yielding reproduction processes X ˆ(n) ) ≤ R, then H(X (n)
ˆ0 ) = E(X0 ) lim E(X
n→∞
(13.15)
(n)
lim
ˆ0 ) COV(X0 , X σ 2ˆ(n)
n→∞
=1
(13.16)
X0
lim σ 2ˆ(n) = σX20 − DX (R).
n→∞ (n)
Defining the error as 0 come
X0
(13.17)
ˆ0(n) − X0 , then the necessary conditions be=X (n)
lim E(0 ) = 0
(13.18)
(n) ˆ (n) lim E(0 X 0 )) = 0
(13.19)
n→∞ n→∞
lim σ 2(n) = DX (R).
n→∞
(13.20)
The results are stated for time k = 0, but stationarity ensures that they hold for all times k. Proof: For any encoder/decoder pair (fn , gn ) yielding a reproduction ˆ(n) process X ∆(fn , gn ) ≥ inf ∆(fn , agn + b) a,b∈R
≥ DX (R) = inf ∆(f , g) f ,g
where the second inequality follows since scaling a sliding-block decoder by a real constant and adding a real constant results in another slidingblock decoder with entropy rate no greater than that of the input. The minimization over a and b for each n is solved by standard linear prediction techniques as
13.3 Sliding-Block Codes
353
an =
ˆ0(n) ) COV(X0 , X σ 2ˆ(n)
(13.21)
X0
ˆ0(n) ), bn = E(X0 ) − an E(X
(13.22)
inf ∆(fn , agn + b) = ∆(fn , an gn + bn ) a,b
= σX20 − a2n σ 2ˆ(n) . X0
(13.23)
Combining the above facts we have that since (fn , gn ) is an asymptotically optimal sequence, DX (R) = lim ∆(fn , gn ) ≥ lim ∆(fn , an gn + bn ) n→∞
n→∞
≥ DX (R)
(13.24)
and hence that both inequalities are actually equalities. The final inequality (13.24) being an equality yields lim a2n σ 2ˆ(n) = σX20 − DX (R).
n→∞
X0
(13.25)
Application of asymptotic optimality and (13.21) to ˆ0(n) )2 ∆(fn , gn ) = E (X0 − X ˆ0(n) − E(X ˆ0(n) )] = E ([X0 − E(X0 )] − [X ˆ0(n) )])2 + [E(X0 ) − E(X ˆ0(n) ) = σX20 + σ 2ˆ(n) − 2COV(X0 , X X0
(n)
ˆ0 )]2 + [E(X0 ) − E(X results in DX (R) = lim
n→∞
ˆ0(n) )]2 . σX20 + (1 − 2an )σ 2ˆ(n) + [E(X0 ) − E(X X0
Subtracting (13.25) from (13.26) yields ˆ0(n) )]2 = 0. lim (1 − an )2 σ 2ˆ(n) + [E(X0 ) − E(X n→∞
X0
(13.26)
(13.27)
Since both terms in the limit are nonnegative, both must converge to zero since the sum does. Convergence of the rightmost term in the sum proves (13.15). Provided DX (R) < σX20 , which is true if R > 0, (13.25) and (13.27) together imply that (an − 1)2 /a2n converges to 0 and hence that
354
13 Properties of Good Source Codes
lim an = lim
n→∞
ˆ0(n) ) COV(X0 , X
n→∞
σ 2ˆ(n)
= 1.
(13.28)
X0
This proves (13.16) and with (13.26) proves (13.17) and also that ˆ0(n) ) = σX2 − DX (R). lim COV(X0 , X 0
n→∞
(13.29)
Finally consider the conditions in terms of the reproduction error. Eq. (13.18) follows from (13.15). Eq. (13.19) follows from (13.15)–(13.29) and some algebra. Eq. (13.20) follows from (13.18) and the asymptotic optimality of the codes. 2 If X is a B-process so that ∆X|Z = DX (R), then a similar proof yields corresponding results for the simulation problem. If gn is an asymptotically optimal (with respect to ρ 2 distortion) sequence of stationary codes of an IID equiprobable source Z with alphabet B of size R = log kBk ˜(n) , then which produce a simulated process X (n)
˜0 ) = E(X0 ) lim E(X
n→∞
lim σ 2˜(n) = σX20 − ∆X|Z .
n→∞
X0
It is perhaps surprising that when finding the best matching process with constrained rate, the second moments differ.
Finite-order distribution Shannon conditions for IID processes Several code design algorithms, including randomly populating a trellis to mimic the proof of the trellis source encoding theorem [188], are based on the intuition that the guiding principle of designing such a system for an IID source should be to produce a code with marginal reproduction distribution close to a Shannon optimal reproduction distribution [193, 42, 143]. The following result from [117] formalizes this intuition. Lemma 13.6. Given a real-valued IID process X with distribution µX , assume that {fn , gn } is an asymptotically optimal sequence of stationary source encoder/decoder pairs with common alphabet B of size R = ˆ(n) . Then a subsequence log kBk which produce a reproduction process X of the marginal distribution of the reproduction process, µXˆ(n) converges 0 weakly and in quadratic transportation distortion (ρ-bar distortion with respect to squared error distortion) to a Shannon optimal reproduction
13.3 Sliding-Block Codes
355
distribution. If the Shannon optimal reproduction distribution is unique, then µXˆ(n) converges to it. 0
Proof: Given the asymptotically optimal sequence of codes, let πn deˆ(n) ). The encoded note the induced process joint distributions on (X, X R process has alphabet size 2 and hence entropy rate less than or equal to R. Since coding cannot increase entropy rate, the entropy rate of the reproduction (decoded) process is also less than or equal to R. Since the input process is IID, Lemma 8.8 implies that for all N that N−1 1 1 1 X ˆN ) ≥ ˆ(n) ) I(Xi , X I(πnN ) = I(X N , X i N N N i=0 (n)
ˆ0 ) = I(πn1 ). = I(X0 , X
(13.30)
The leftmost term converges to the mutual information rate between the input and reproduction, which is bound above by the entropy rate of the output so that ˆ0(n) ) ≤ R, all n. I(X0 , X (13.31) Since the code sequence is asymptotically optimal, (13.13) holds. Thus ˆ0(n) ) meets the condithe sequence of joint distributions πn for (X0 , X tions of Corollary 9.4 and hence µXˆ(n) has a subsequence which con0 verges weakly to a Shannon optimal distribution. If the Shannon optimal distribution µY0 is unique, then every subsequence of of µXˆ(n) has 0 a further subsequence which converges to µY0 , which implies that µXˆ(n) 0 converges weakly to µY0 . The moment conditions (13.15) and (13.17)) ˆ0 )2 ]. The weak ˆ0(n) )2 ] converges to E[(X of Lemma 13.5 imply that E[(X convergence of a subsequence of µXˆ(n) (or the sequence itself) and the convergence of the second moments imply convergence in quadratic transportation distortion (ρ-bar distortion with respect to squared error distortion) [187]. 2 Since the source is IID, the N-fold product of a one-dimensional Shannon optimal distribution is an N-dimensional Shannon optimal distribution. If the Shannon optimal marginal distribution is unique, then so is the N-dimensional Shannon optimal distribution. Since Csiszár’s [25] results as summarized in Corollary 9.4 hold for the N-dimensional case, we immediately have the first part of the following corollary. Corollary 13.1. Given the assumptions of the lemma, for any positive integer N let µXˆ(n) denote the N-dimensional joint distribution of the reproducˆ(n) . Then a subsequence of the N-dimensional reproduction tion process X distribution µXˆ(n) converges weakly and in quadratic transportation distortion to the N-fold product of a Shannon optimal marginal distribution (and hence to an N-dimensional Shannon optimal distribution). If the one dimensional Shannon optimal distribution is unique, then µXˆ(n) converges
356
13 Properties of Good Source Codes
weakly and in quadratic transportation distortion to its N-fold product distribution. Proof: The moment conditions (13.15) and (13.17)) of Lemma 13.5 imply ˆ(n) )2 ] converges to E[(X ˆk )2 ] for k = 0, 1, . . . , N − 1. The weak that E[(X k convergence of the N-dimensional distribution of a subsequence of µXˆ(n) (or the sequence itself) and the convergence of the second moments imply convergence in quadratic transportation distortion [187]. 2 There is no counterpart of this result for optimal codes as opposed to asymptotically optimal codes. Consider the Gaussian case where the Shannon optimal distribution is a product Gaussian distribution with variance σX2 −DX (R). If a code were optimal, then for each N the resulting Nth order reproduction distribution would have to equal the Shannon product distribution. But if this were true for all N, the reproduction would have to be the IID process with the Shannon marginals, but that process has infinite entropy rate. If X is a B-process, then a small variation on the proof yields similar results for the simulation problem: given an IID target source X, the Nth order joint distributions µX˜(n) of an asymptotically optimal se˜(n) will have a subsequence that quence of constrained rate simulations X converges weakly and in quadratic transportation distortion to an Ndimensional Shannon optimal distribution.
Asymptotic Uncorrelation ˆ(n) Define as usual the covariance function of the stationary process X (n) ˆ (n) ˆ by KXˆ(n) (k) = COV(Xi , Xi−k ) for all integer k. The following theorem states and proves an intuitive property nearly optimal codes for IID sources must yield approximately uncorrelated reproduction processes. The result is implied by the convergence of joint distributions to the Shannon optimal distribution along with a technical moment condition proved in the subsequent lemma. Theorem 13.1. Given a real-valued IID process X with distribution µX , assume that fn , gn is an asymptotically optimal sequence of stationary source encoder/decoder pairs with common alphabet B of size R = ˆ(n) . For all k ≠ 0, log kBk which produce a reproduction process X lim KXˆ(n) (k) = 0
n→∞
(13.32)
and hence the reproduction processes are asymptotically uncorrelated. Proof. If the Shannon optimal distribution is unique, then µXˆ(n) converges in quadratic transportation distortion to the N-fold product of the
13.3 Sliding-Block Codes
357
Shannon optimal marginal distribution by Corollary 13.1. Lemma 13.7 which follows shows that this implies the convergence of KXˆ(n) (k) = ˆ0(n) ) to 0 for all k ≠ 0. ˆ(n) , X 2 COV(X k Taken together these necessary conditions provide straightforward tests for code construction algorithms. Lemma 13.7. Let µ N denote the N-fold product of a probability distribuR tion µ on the real line such that x 2 dµ(x) < ∞. Assume {νn } is a sequence of probability distribution on RN such that limn→∞ ρ N (µ N , νn ) = (n) (n) (n) 0. If Y1 , Y2 , . . . , YN are random variables with joint distribution νn , (n) (n) (n) (n) then for all i ≠ j, limn→∞ E Yi − E(Yi ) Yj − E(Yj ) = 0. Proof. The convergence of νn to µ N in quadratic transportation distortion implies that there exist IID random variables Y1 , . . . , YN with com(n) (n) mon distribution µ and a sequence or N random variables Y1 , Y2 , . . . , (n) YN with joint distribution νn , all defined on the same probability space, such that (n) lim E[(Yi − Yi )2 ] = 0, i = 1, . . . , N. (13.33) n→∞
First note that this implies for all i (n) 2
) ] = E[Yi2 ].
lim E[(Yi
n→∞ (n)
Since lim E|Yi n→∞
(13.34)
− Yi | = 0, the Cauchy-Scwartz inequality implies that
for all i
(n)
lim E(Yi
n→∞
) = E(Yi ).
(13.35)
The statement is a direct consequence of the fact that in any inner product space, the inner product is jointly continuous. Letting hX, Y i = E(XY ) and kXk = [E(X 2 )]1/2 for random variables X and Y with finite second moment defined on this probability space, we have the bound (n) (n) hY , Y i − hYi , Yj i ≤ hY (n) , Y (n) − Yj i + hY (n) − Yi , Yj i i j i j i (n)
≤ kYi (n)
Since kYi
(n)
k kYj
(n)
− Yj k + kYi (n)
k converges to kYi k by (13.34) and kYi
zero by (13.33), we obtain that (n)
lim E(Yi
n→∞
(n)
Yi
(n) (n) hYi , Yj i
− Yi k kYj k.
− Yi k converges to
converges to hYi , Yj i, i.e,
) = E(Yi Yj ) = E(Yi )E(Yj )
since Yi and Yj are independent if i ≠ j. This and (13.35) imply the lemma statement. 2
Chapter 14
Coding for Noisy Channels
Abstract Reliable communication over a noisy channel is the focus of this chapter. The chapter begins with a development of the classic fundamental results of Feinstein regarding reliable communication of block codes and the relation of operational channel capacity to Shannon capacity for discrete channels. A technique of Dobrushin is used to extend Feinstein’s results for channels with no input memory or anticipation by making codes robust to small changes in the conditional distributions describing channels. This leads in turn to the extension of block coding theorems to d-bar continuous channels, discrete noisy channels where the noise distribution within a block can be well approximated in a d-bar sense with only finite knowledge of past and future inputs. Traditional channel coding theorems for block codes assume knowledge of synchronization — when the blocks begin. Another technique of Doburshin is used to synchronize block codes through noisy channels. Combining synchronized block codes with the Rohlin-Kakutani theorem yields a coding theorem for sliding-block channel coding. Finally, combining the source coding theorems with channel coding theorems yields joint-source and channel coding theorems.
14.1 Noisy Channels In the treatment of source coding the communication channel was assumed to be noiseless. If the channel is noisy, then the coding strategy must be different, some form of error control is required to undo the damage caused by the channel. The overall point-to-point communication problem is usually broken into two pieces: A source coder is designed for a noiseless channel with a given resolution or rate and an error correction code is designed for the actual noisy channel in order to make it appear almost noiseless. The combination of the two codes R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4_14, © Springer Science+Business Media, LLC 2011
359
360
14 Coding for Noisy Channels
then provides the desired overall code or joint source and channel code. This division is natural in the sense that optimizing a code for a particular source may suggest quite a different structure than optimizing it for a channel. The structures must be compatible at some point, however, so that they can be used together. This division of source and channel coding is apparent in the subdivision of this chapter. We begin with a fundamental lemma due to Feinstein [39] which is the basis of traditional proofs of coding theorems for channels. It does not consider a source at all, but finds for a given conditional distribution the maximum number of inputs which lead to outputs which can be distinguished with high probability. Feinstein’s lemma can be thought of as a channel coding theorem for a channel which is used only once and which has no past or future. The lemma immediately provides a coding theorem for the special case of a channel which has no input memory or anticipation. The difficulties enter when the conditional distributions of output blocks given input blocks depend on previous or future inputs. This difficulty is handled by imposing some form of continuity on the channel with respect to its input, that is, by assuming that if the channel input is known for a big enough block, then the conditional probability of outputs during the same block is known nearly exactly regardless of previous or future inputs. The continuity condition which we shall consider is that of d-continuous channels. Joint source and channel codes have been obtained for more general channels called weakly continuous channels (see, e.g., Kieffer [94] [95]), but these results require a variety of techniques not yet considered here and do not follow as a direct descendent of Feinstein’s lemma. Block codes are extended to sliding-block codes in a manner similar to that for source codes: First it is shown that asynchronous block codes can be synchronized and then that the block codes can be “stationarized” by the insertion of random punctuation. The approach to synchronizing channel codes is based on a technique of Dobrushin [33]. We consider stationary channels almost exclusively, thereby not including interesting nonstationary channels such as finite state channels with an arbitrary starting state. We will discuss such generalizations and we point out that they are straightforward for two-sided processes, but the general theory of AMS channels for one-sided processes is not in a satisfactory state. Lastly, we emphasize ergodic channels. In fact, for the sliding-block codes the channels are also required to be totally ergodic, that is, ergodic with respect to all block shifts. As previously discussed, we emphasize digital, i.e., discrete, channels. A few of the results, however, are as easily proved under somewhat more general conditions and hence we shall do so. For example, given the background of this book it is actually easier to write things in terms of measures and integrals than in terms of sums over probability mass func-
14.2 Feinstein’s Lemma
361
tions. This additional generality will also permit at least a description of how the results extend to continuous alphabet channels.
14.2 Feinstein’s Lemma Let (A, BA ) and (B, BB ) be measurable spaces called the input space and the output space, respectively. Let PX denote a probability distribution on (A, BA ) and let ν(F |x), F ∈ BB , x ∈ B denote a regular conditional probability distribution on the output space. ν can be thought of as a “channel” with random variables as input and output instead of sequences. Define the hookup PX ν = PXY by Z PXY (F ) = dPX (x)ν(Fx |x). Let PY denote the induced output distribution and let PX × PY denote the resulting product distribution. Assume that PXY 0 there exist xi ∈ A; i = 1, . . . , M and a measurable partition F = {Γi ; i = 1, . . . , M} of B such that ν(Γic |xi ) ≤ Me−a + PXY (i ≤ a). Proof: Define G = {x, y : i(x, y) > a} Set = Me−a + PXY (i ≤ a) = Me−a + PXY (Gc ). The result is obvious if ≥ 1 and hence we assume that < 1 and hence also that PXY (Gc ) ≤ < 1 and therefore that Z PXY (i > a) = PXY (G) =
dPX (x)ν(Gx |x) > 1 − > 0.
˜ = {x : ν(Gx |x) > 1 − and (14.2) holds} must This implies that the set A have positive measure under PX We now construct a code consisting of ˜ and define Γx1 = input points xi and output sets Γxi . Choose an x1 ∈ A ˜ for which ν(Gx2 − Γx1 |x2 ) > Gx1 . Next choose if possible a point x2 ∈ A 1 − . Continue in this way until either M points have been selected or ˜ have been exhausted. In particular, given the pairs all the points in A {xj , Γj }; j = 1, 2, . . . , i − 1, satisfying the condition, find an xi for which ν(Gxi −
[
Γxj |xi ) > 1 − .
(14.3)
j
If the procedure terminates before M points have been collected, denote the final point’s index by n. Observe that ν(Γxi c |xi ) ≤ ν(Gxi c |xi ) ≤ ; i = 1, 2, . . . , n and hence the lemma will be proved if we can show that necessarily n cannot be strictly less than M. We do this by assuming the contrary and finding a contradiction. Suppose that the selection has terminated at n < M and define the set Sn F = i=1 Γxi ∈ BB . Consider the probability PXY (G) = PXY (G
\
(A × F )) + PXY (G
\ (A × F c )).
The first term can be bounded above as PXY (G
\
(A × F )) ≤ PXY (A × F ) = PY (F ) =
n X i=1
PY (Γxi ).
(14.4)
14.2 Feinstein’s Lemma
363
We also have from the definitions and from (14.2) that Z Z Z f (xi , y) PY (Γxi ) = dPY (y) ≤ dPY (y) ≤ dPY (y) ea Γxi Gxi Gxi Z ≤ e−a dPY (y)f (xi , y) ≤ e−a and hence PXY (G
\ (A × F )) ≤ ne−a .
(14.5)
Consider the second term of (14.3): Z \ \ c PXY (G (A × F )) = dPX (x)ν((G (A × F c ))x |x) Z \ = dPX (x)ν(Gx F c |x) Z dPX (x)ν(Gx −
=
n [
Γi |x).
(14.6)
i=1
We must have, however, that ν(Gx −
n [
Γi |x) ≤ 1 −
i=1
with PX probability 1 or there would be a point xn+1 for which ν(Gxn+1 −
n+1 [
Γi |xn+1 ) > 1 − ,
i=1
that is, (14.3) would hold for i = n + 1, contradicting the definition of n as the largest integer for which (14.3) holds. Applying this observation to (14.6) yields \ PXY (G (A × F c )) ≤ 1 − which with (14.4) and (14.5) implies that PXY (G) ≤ ne−a + 1 − .
(14.7)
From the definition of , however, we have also that PXY (G) = 1 − PXY (Gc ) = 1 − + Me−a which with (14.7) implies that M ≤ n, completing the proof.
2
364
14 Coding for Noisy Channels
14.3 Feinstein’s Theorem Given a channel [A, ν, B] an (M, n, ) block channel code for ν is a collection {wi , Γi }; i = 1, 2, . . . , M, where wi ∈ An , Γi ∈ Bn B , all i, with the property that sup max νxn (Γi ) ≤ , (14.8) x∈c(wi ) i=1,...,M
where c(an ) = {x : x n = an } and where νxn is the restriction of νx to Bn B. The rate of the code is defined as n−1 log M. Thus an (n, M, ) channel code is a collection of M input n-tuples and corresponding output cells such that regardless of the past or future inputs, if the input during time 1 to n is a channel codeword, then the output during time 1 to n is very likely to lie in the corresponding output cell. Channel codes will be useful in a communication system because they permit nearly error free communication of a select group of messages or codewords. A communication system can then be constructed for communicating a source over the channel reliably by mapping source blocks into channel codewords. If there are enough channel codewords to assign to all of the source blocks (at least the most probable ones), then that source can be reliably reproduced by the receiver. Hence a fundamental issue for such an application will be the number of messages M or, equivalently, the rate R of a channel code. Feinstein’s lemma can be applied fairly easily to obtain something that resembles a coding theorem for a noisy channel. Suppose that [A, ν, B] is a channel and [A, µ] is a source and that [A × B, p = µν] is the resulting hookup. Denote the resulting pair process by {Xn , Yn } For any integer K K let p K denote the restriction of p to (AK × B K , BK A × BB ), that is, the K K distribution on input/output K-tuples (X , Y ). The joint distribution p K together with the input distribution µ K induce a regular conditional ˆK defined by ν ˆK (F |x K ) = Pr(Y K ∈ F |X K = x K ). In particuprobability ν lar, ˆK (G|aK ) = Pr(Y K ∈ G|X K = aK ) ν Z 1 = K K ν K (G)dµ(x). µ (a ) c(aK ) x
(14.9)
where c(aK ) = {x : x K = aK } is the rectangle of all sequences with a ˆK the induced K-dimensional common K-dimensional output. We call ν channel of the channel ν and the source µ. It is important to note that the induced channel depends on the source as well as on the channel, a fact that will cause some difficulty in applying Feinstein’s lemma. An exception to this case which proves to be an easy application is that of a channel without input memory and anticipation, in which case we have from the definitions that
14.3 Feinstein’s Theorem
365
ˆK (F |aK ) = νx (Y K ∈ F ); x ∈ c(aK ). ν Application of Feinstein’s lemma to the induced channel yields the following result, which was proved by Feinstein for stationary finite alphabet channels and is known as Feinstein’s theorem: Lemma 14.2. Suppose that [A × B, µν] is an AMS and ergodic hookup of a source µ and channel ν. Let I µν = I µν (X; Y ) denote the average mutual ∗ information rate and assume that I µν = Iµν is finite (as is the case if the alphabets are finite (Theorem 8.2) or have the finite-gap information property (Theorem 8.4)). Then for any R < I µν and any > 0 there exists for sufficiently large n a code {win ; Γi ; i = 1, 2, . . . , M}, where M = benR c, win ∈ An , and Γi ∈ Bn B , with the property that ˆn (Γic |win ) ≤ , i = 1, 2, . . . , M. ν
(14.10)
Comment: We shall call a code {wi , Γi ; i = 1, 2, . . . , M} which satisfies (14.10) for a channel input process µ a (µ, M, n, )-Feinstein code. The quantity n−1 log M is called the rate of the Feinstein code. Proof: Let η denote the output distribution induced by µ and ν. Define the information density in =
dp n d(µ n × ηn )
δ=
I µν − R > 0. 2
and define
Apply Feinstein’s lemma to the n-dimensional hookup (µν)n with M = benR c and a = n(R + δ) to obtain a code {wi , Γi }; i = 1, 2, . . . , M with ˆn (Γic |win ) max ν i
≤ Me−n(R+δ) + p n (in ≤ n(R + δ)) 1 = benR ce−n(R+δ) + p( in (X n ; Y n ) ≤ R + δ) n
(14.11)
and hence ˆn (Γic |win ) ≤ e−nδ + p( max ν i
1 in (X n ; Y n ) ≤ I µν − δ). n
(14.12)
From Theorem 8.1 n−1 in converges in L1 to I µν and hence it also converges in probability. Thus given we can choose an n large enough to ensure that the right hand side of (14.11) is smaller than , which completes the proof of the theorem. 2
366
14 Coding for Noisy Channels
We said that the lemma “resembled” a coding theorem because a real coding theorem would prove the existence of an (M, n, ) channel code, that is, it would concern the channel ν itself and not the induced channel ˆ, which depends on a channel input process distribution µ. The differν ence between a Feinstein code and a channel code is that the Feinstein code has a similar property for an induced channel which in general depends on a source distribution, while the channel code has this property independent of any source distribution and for any past or future inputs. Feinstein codes will be used to construct block codes for noisy channels. The simplest such construction is presented next. Corollary 14.1. Suppose that a channel [A, ν, B] is input memoryless and input nonanticipatory as defined in Section 2.9. Then a (µ, M, n, )Feinstein code for some channel input process µ is also an (M, n, )-code. Proof: Immediate since for a channel without input memory and anticipation we have that νxn (F ) = νun (F ) if x n = un . 2 The principal idea of constructing channel codes from Feinstein codes for more general channels will be to place assumptions on the channel which ensure that for sufficiently large n the channel distribution νxn and ˆn (·|x n ) are close. This general the induced finite dimensional channel ν idea was proposed by McMillan [123] who suggested that coding theorems would follow for channels that were sufficiently continuous in a suitable sense. The previous results did not require stationarity of the channel, but in a sense stationarity is implicit if the channel codes are to be used repeatedly (as they will be in a communication system). Thus the immediate applications of the Feinstein results will be to stationary channels. The following is a rephrasing of Feinstein’s theorem that will be useful. Corollary 14.2. Suppose that [A × B, µν] is an AMS and ergodic hookup of a source µ and channel ν. Let I µν = I µν (X; Y ) denote the average ∗ mutual information rate and assume that I µν = Iµν is finite. Then for any R < I µν and any > 0 there exists an n0 such that for all n ≥ n0 there are (µ, benR c, n, )-Feinstein codes. As a final result of the Feinstein variety, we point out a variation that applies to nonergodic channels. Corollary 14.3. Suppose that [A × B, µν] is an AMS hookup of a source µ and channel ν. Suppose also that the information density converges a.e. to a limiting density 1 i∞ = lim in (X n ; Y n ). n→∞ n (Conditions for this to hold are given in Theorem 11.4.) Then given > 0 and δ > 0 there exists for sufficiently large n a [µ, M, n, + µν(i∞ ≤ R + δ)] Feinstein code with M = benR c.
14.4 Channel Capacity
367
Proof: Follows from the lemma and from Fatou’s lemma which implies that 1 lim sup p( in (X n ; Y n ) ≤ a) ≤ p(i∞ ≤ a). n n→∞
2
14.4 Channel Capacity The form of the Feinstein lemma and its corollaries invites the question of how large R (and hence M) can be made while still getting a code of the desired form. From Feinstein’s theorem it is seen that for an ergodic channel R can be any number less than I(µν) which suggests that if we define the quantity CAMS,e =
I µν ,
sup
(14.13)
AMS and ergodic µ ∗ then if I µν = Iµν (e.g., the channel has finite alphabet), then we can construct for some µ a Feinstein code for µ with rate R arbitrarily near CAMS,e . CAMS,e is an example of a quantity called an information rate capacity or, simply, capacity of a channel. We shall encounter a few variations on this definition just as there were various ways of defining distortion-rate functions for sources by considering either vectors or processes with different constraints. In this section a few of these definitions are introduced and compared. A few possible definitions of information rate capacity are
CAMS = sup I µν ,
(14.14)
AMS µ
Cs =
I µν ,
sup
(14.15)
stationary µ
Cs,e =
I µν ,
sup
(14.16)
stationary and ergodic µ
Cns =
I µν ,
sup
(14.17)
n−stationary µ
Cbs =
I µν = sup
sup block stationary µ
n
sup
I µν .
(14.18)
n−stationary µ
Several inequalities are obvious from the definitions: CAMS ≥ Cbs ≥ Cns ≥ Cs ≥ Cs,e
(14.19)
CAMS ≥ CAMS,e ≥ Cs,e .
(14.20)
368
14 Coding for Noisy Channels
In order to relate these definitions we need a variation on Lemma 12.3.1 described in the following lemma. Lemma 14.3. Given a stationary finite-alphabet channel [A, ν, B], let µ be the distribution of a stationary channel input process and let {µx } be its ergodic decomposition. Then Z I µν = dµ(x)I µx ν . (14.21) Proof: We can write I µν = h1 (µ) − h2 (µ) where h1 (µ) = H η (Y ) = inf n
1 Hη (Y n ) n
is the entropy rate of the output, where η is the output measure induced by µ and ν, and where h2 (µ) = H µν (Y |X) = lim
n→∞
1 Hµν (Y n |X n ) n
is the conditional entropy rate of the output given the input. If µk → µ on any finite dimensional rectangle, then also ηk → η and hence Hηk (Y n ) → Hη (Y n ) so that it follows as in the proof of Corollary 3.4 that h1 (µ) is an upper semicontinuous function of µ. It is also affine because H η (Y ) is an affine function of η (Lemma 3.9) which is in turn a linear function of µ. Thus from Theorem 8.9.1 of [55] or Theorem 8.5 of [58] Z h1 (µ) = dµ(x)h1 (µx ). h2 (µ) is also affine in µ since h1 (µ) is affine in µ and I µν is affine in µ (since it is affine in µν from Lemma 8.6). Hence we will be done if we can show that h2 (µ) is upper semicontinuous in µ since then Theorem 8.9.1 of [55] will imply that Z h2 (µ) = dµ(x)h2 (µx ) which with the corresponding result for h1 proves the lemma. To see this observe that if µk → µ on finite dimensional rectangles, then Hµk ν (Y n |X n ) → Hµν (Y n |X n ). Next observe that for stationary processes
(14.22)
14.4 Channel Capacity
369
n−m H(Y n |X n ) ≤ H(Y m |X n ) + H(Ym |X n ) n−m n−m ≤ H(Y m |X m ) + H(Ym |Xm )
= H(Y m |X m ) + H(Y n−m |X n−m ) which as in Section 2.4 implies that H(Y n |X n ) is a subadditive sequence and hence 1 1 lim H(Y n |X n ) = inf H(Y n |X n ). n→∞ n n n Coupling this with (14.22) proves upper semicontinuity exactly as in the proof of Corollary 3.4, which completes the proof of the lemma. 2 Lemma 14.4. If a channel ν has a finite alphabet and is stationary, then all of the above information rate capacities are equal. Proof: From Theorem 8.2 I = I ∗ for finite alphabet processes and hence from Lemma 8.6 and Lemma 2.2 we have that if µ is AMS with stationary mean µ, then I µν = I µν = I µν and thus the supremum over AMS sources must be the same as that over stationary sources. The fact that Cs ≤ Cs,e follows immediately from the previous lemma since the best stationary source can do no better than to put all of its measure on the ergodic component yielding the maximum information rate. Combining these facts with (14.19)–(14.20) proves the lemma. 2 Because of the equivalence of the various forms of information rate capacity for stationary channels, we shall use the symbol C to represent the information rate capacity of a stationary channel and observe that it can be considered as the solution to any of the above maximization problems. Shannon’s original definition of channel capacity applied to channels without input memory or anticipation. We pause to relate this definition to the process definitions. Suppose that a channel [A, ν, B] has no input memory or anticipation and hence for each n there are regular condiˆn (G|x n ); x ∈ An , G ∈ Bn tional probability measures ν B , such that ˆn (G|x n ). νxn (G) = ν ˆn by Define the finite-dimensional capacity of the ν ν n ) = sup Iµn νˆn (X n ; Y n ), Cn (ˆ µn
where the supremum is over all vector distributions µ n on An . Define the Shannon capacity of the channel µ by 1 n n C (ˆ ν ) n→∞ n
CShannon = lim
370
14 Coding for Noisy Channels
if the limit exists. Suppose that the Shannon capacity exists for a channel ν without memory or anticipation. Choose N large enough so that CN is very close to CShannon and let µ N approximately yield CN . Then construct a block memoryless source using µ N . A block memoryless source is AMS and hence if the channel is AMS we must have an information rate I µν (X; Y ) = lim
n→∞
1 1 Iµν (X n ; Y n ) = lim Iµν (X kN ; Y kN ). n k→∞ kN
Since the input process is block memoryless, we have from Lemma 8.8 that k X N N I(X kN ; Y kN ) ≥ I(XiN ; YiN ). i=0
If the channel is stationary then {Xn , Yn } is N-stationary and hence if 1 IµN νˆN (X N ; Y N ) ≥ CShannon − , N then
1 I(X kN ; Y kN ) ≥ CShannon − . kN
Taking the limit as k → ∞ we have that CAMS = C ≥ I(X; Y ) = lim
k→∞
1 I(X kN ; Y kN ) ≥ CShannon − kN
and hence C ≥ CShannon . Conversely, pick a stationary source µ which nearly yields C = Cs , that is, I µν ≥ Cs − . Choose n0 sufficiently large to ensure that 1 Iµν (X n ; Y n ) ≥ I µν − ≥ Cs − 2. n This implies, however, that for n ≥ n0 Cn ≥ Cs − 2, and hence application of the previous lemma proves the following lemma. Lemma 14.5. Given a finite alphabet stationary channel ν with no input memory or anticipation, C = CAMS = Cs = Cs,e = CShannon .
14.4 Channel Capacity
371
The Shannon capacity is of interest because it can be numerically computed while the process definitions are not always amenable to such computation. With Corollary 14.2 and the definition of channel capacity we have the following result. Lemma 14.6. If ν is an AMS and ergodic channel and R < C, then there is an n0 sufficiently large to ensure that for all n ≥ n0 there exist (µ, benR c, n, ) Feinstein codes for some channel input process µ. Corollary 14.4. Suppose that [A, ν, B] is an AMS and ergodic channel with no input memory or anticipation. Then if R < C, the information rate capacity or Shannon capacity, then for > 0 there exists for sufficiently large n a (benR c, n, ) channel code. Proof: Follows immediately from Corollary 14.3 by choosing a stationary 2 and ergodic source µ with I µν ∈ (R, C). There is another, quite different, notion of channel capacity that we introduce for comparison and to aid the discussion of nonergodic stationary channels. Define for an AMS channel ν and any λ ∈ (0, 1) the quantile C ∗ (λ) = sup sup{r : µν(i∞ ≤ r ) < λ)}, AMS µ
where the supremum is over all AMS channel input processes and i∞ is the limiting information density (which exists because µν is AMS and has finite alphabet). Define the information quantile capacity C ∗ by C ∗ = lim C ∗ (λ). λ→0
The limit is well-defined since the C ∗ (λ) are bounded and nonincreasing. The information quantile capacity was introduced by Winkelbauer [194] and its properties were developed by him and by Kieffer [90]. Fix an R < C ∗ and define δ = (C ∗ −R)/2. Given > 0 we can find from the definition of C ∗ an AMS channel input process µ for which µν(i∞ ≤ R + δ) ≤ . Applying Corollary 14.3 with this δ and /2 then yields the following result for nonergodic channels. Lemma 14.7. If ν is an AMS channel and R < C ∗ , then there is an n0 sufficiently large to ensure that for all n ≥ n0 there exist (µ, f enR f , n, ) Feinstein codes for some channel input process µ. We close this section by relating C and C ∗ for AMS channels. Lemma 14.8. Given an AMS channel ν, C ≥ C ∗.
372
14 Coding for Noisy Channels
Proof: Fix λ > 0. If r < C ∗ (λ) there is a µ such that λ > µν(i∞ ≤ r ) = 1 − µν(i∞ > r ) ≥ 1 − I µν /r , where we have used the Markov inequality. Thus for all r < C ∗ we have that I µν ≥ r (1 − µν(i∞ ≤ r )) and hence C ≥ I µν ≥ C ∗ (λ)(1 − λ) → C ∗ . λ→0
2 It can be shown that if a stationary channel is also ergodic, then C = C ∗ by using the ergodic decomposition to show that the supremum defining C(λ) can be taken over ergodic sources and then using the fact that for ergodic µ and ν, i∞ equals I µν with probability one. (See Kieffer [90].)
14.5 Robust Block Codes Feinstein codes immediately yield channel codes when the channel has no input memory or anticipation because the induced vector channel is the same with respect to vectors as the original channel. When extending this technique to channels with memory and anticipation we will try to ensure that the induced channels are still reasonable approximations to the original channel, but the approximations will not be exact and hence the conditional distributions considered in the Feinstein construction will not be the same as the channel conditional distributions. In other words, the Feinstein construction guarantees a code that works well for a conditional distribution formed by averaging the channel over its past and future using a channel input distribution that approximately yields channel capacity. This does not in general imply that the code will also work well when used on the unaveraged channel with a particular past and future input sequence. We solve this problem by considering channels for which the two distributions are close if the block length is long enough. In order to use the Feinstein construction for one distribution on an actual channel, we will modify the block codes slightly so as to make them robust in the sense that if they are used on channels with slightly different conditional distributions, their performance as measured by probability of error does not change much. In this section we prove that this can be done. The basic technique is due to Dobrushin [33] and a similar technique was studied by Ahlswede and Gács [4]. (See also Ahlswede and Wolfowitz [5].) The results of this section are due to Gray, Ornstein, and Dobrushin [68]. A channel block length n code {wi , Γi ; i = 1, 2, . . . , M will be called δrobust (in the Hamming distance sense) if the decoding sets Γi are such that the expanded sets
14.5 Robust Block Codes
373
(Γi )δ ≡ {y n :
1 dn (y n , Γi ) ≤ δ} n
are disjoint, where dn (y n , un ) dn (y n , Γi ) = min n u ∈Γi
and dn (y n , un ) =
n−1 X
dH (yi , ui )
i=0
and dH (a, b) is the Hamming distance (1 if a 6= b and 0 if a = b). Thus the code is δ-robust if received n-tuples in a decoding set can be changed by an average Hamming distance of up to δ without falling in a different decoding set. We show that by reducing the rate of a code slightly we can always make a Feinstein code robust. 0
Lemma 14.9. Let {wi 0 , Γi0 ; i = 1, 2, . . . , M 0 } be a (µ, enR , n, )-Feinstein code for a channel ν. Given δ ∈ (0, 1/4) and R < R 0 − h2 (2δ) − 2δ log(kBk − 1), where as before h2 (a) is the binary entropy function −a log a − (1 − a) log(1 − a), there exists a δ-robust (µ, benR c, n, n )-Feinstein code for ν with 0 n ≤ + e−n(R −R−h2 (2δ)−2δ log(kBk−1)−3/n) . Proof: For i = 1, 2, . . . , M 0 let ri (y n ) denote the indicator function for (Γi )2δ . For a fixed y n there can be at most 2δn X i=0
! ! 2δn X n n 1 i (1 n−i i n ) ) (kBk − 1) = kBk (1 − kBk kBk i i i=0
n-tuples bn ∈ B n such that n−1 dn (y n , bn ) ≤ 2δ. Set p = 1 − 1/kBk and apply Lemma 3.6 to the sum to obtain the bound kBk
n
2δn X i=0
! n 1 i 1 n−i ≤ kBkn e−nh2 (2δkp) )( ) (1 − kBk kBk k = e−nh2 (2δkp)+n log kBk ,
where
374
14 Coding for Noisy Channels
1 − 2δ 2δ + (1 − 2δ) ln p 1−p kBk = −h2 (δ) + 2δ ln + (1 − 2δ) ln kBk kBk − 1 = −h2 (δ) + ln kBk − 2δ ln(kBk − 1).
h2 (2δkp) = 2δ ln
Combining this bound with the fact that the Γi are disjoint we have that 0
M X
n
ri (y ) ≤
i=1
2δn X
! n (kBk − 1)i ≤ e−n(h2 (2δ)+2δ ln(kBk−1) . i
i=0
Set M = benR c and select 2M subscripts k1 , · · · , k2M from {1, · · · , M 0 } by random equally likely independent selection without replacement so that each index pair (kj , km ); j, m = 1, . . . , 2M; j 6= m, assumes any unequal pair with probability (M 0 (M 0 − 1))−1 . We then have that 2M 2M X X \ 1 ˆ(Γk0j (Γk0m )2δ |wk0 j ) ν E 2M j=1 m=1,m6=j 0
2M M 2M M X X X X 1 1 X ˆ(y n |wk0 )ri (y n ) ν = 2M j=1 m=1,m6=j k=1 i=1,i6=k M 0 (M 0 − 1) n 0 y ∈Γk
≤
1 2M
M0 X
M X X 1 0 n ˆ |w ) ri (y n ) ν (y k 0 (M 0 − 1) M 0 n j=1 m=1,m6=j k=1 i=1,i6=k 2M X
2M X
y ∈Γk
2M n(h2 (2δ)+2δ log(kBk−1) e M −1 0 ≤ 4e−n(R −R−h2 (2δ)−2δ log(kBk−1) ≡ λn , ≤
where we have assumed that M 0 ≥ 2 so that M 0 − 1 ≥ M 0 /2. Analogous to a random coding argument, since the above expectation is less than λn , there must exist a fixed collection of subscripts i1 , · · · , i2M 0 such that 2M 2M X \ 1 X ˆ(Γi0j (Γi0m )2δ |wi 0j ) ≤ λn . ν 2M j=1 m=1,m6=j
Since no more than half of the above indices can exceed twice the expected value, there must exist indices k1 , · · · , kM ∈ {j1 , · · · , j2M } for which M X \ ˆ(Γk0j (Γk0m )2δ |wk0 j ) ≤ 2λn ; i = 1, 2, . . . , M. ν m=1,m6=j
Define the code {wi , Γi ; i = 1, . . . , M} by wi = wk0 i and
14.6 Block Coding Theorems for Noisy Channels
375
Γi =
Γk0i
M [
−
(Γk0m )2δ .
m=1,m6=i
The (Γi )δ are obviously disjoint since we have removed from Γk0i all words within 2δ of a word in any other decoding set. Furthermore, we have for all i = 1, 2, . . . , M that ˆ(Γk0i |wk0 i ) 1− ≤ ν c \ [ \ [ 0 0 0 0 0 ˆ(Γki ˆ(Γki =ν (Γkm )2δ |wki ) + ν (Γkm )2δ |wk0 i ) m6=i
≤
X
ˆ(Γk0i ν
\
(Γk0m )2δ |wk0 i )
m6=i
ˆ(Γi |wi ) +ν
m6=i
ˆ(Γi |wi ) < 2λn + ν and hence 0
ˆ(Γi |wi ) ≥ 1 − − 8e−n(R −R−h2 (2δ)−2δ log(kBk−1) , ν
2
which proves the lemma.
Corollary 14.5. Let ν be a stationary channel and let Cn be a sequence of 0 (µn , benR c, n, /2) Feinstein codes for n ≥ n0 . Given an R > 0 and δ > 0 such that R < R 0 − h2 (2δ) − 2δ log(kBk − 1), there exists for n1 sufficiently 0 large a sequence Cn ; n ≥ n1 , of δ-robust (µn , benR c, n, ) Feinstein codes. Proof: The corollary follows from the lemma by choosing n1 so that 0
e−n1 (R −R−h2 (2δ)−2δ ln(kBk−1)−3/n1 ) ≤
. 2
2 Note that the sources may be different for each n and that n1 does not depend on the channel input measure.
14.6 Block Coding Theorems for Noisy Channels Suppose now that ν is a stationary finite alphabet d-continuous channel. Suppose also that for n ≥ n1 we have a sequence of δ-robust (µn , benR c, n, ) Feinstein codes {wi , Γi } as in the previous section. We now quantify the performance of these codes when used as channel block codes, that is, used on the actual channel ν instead of on an inˆn be the n-dimensional channel induced channel. As previously let ν n duced by µn and the channel ν, that is, for µn (an ) > 0
376
14 Coding for Noisy Channels n
n
ˆ (G|a ) = Pr(Y ν
n
1 ∈ G|X = a ) = n n µn (a ) n
n
Z c(an )
νxn (G) dµ(x),
(14.23) where c(an ) is the rectangle {x : x ∈ AT ; x n = an }, an ∈ An , and where G ∈ Bn B . We have for the Feinstein codes that ˆn (Γic |wi ) ≤ . max ν i
We use the same codewords wi for the channel code, but we now use the expanded regions (Γi )δ for the decoding regions. Since the Feinstein codes were δ-robust, these sets are disjoint and the code well-defined. Since the channel is d-continuous we can choose an n large enough to ensure that if x n = x n , then dn (νxn , νxn ) ≤ δ2 . Suppose that we have a Feinstein code such that for the induced channel ˆ(Γi |wi ) ≥ 1 − . ν Then if the conditions of Lemma 5.7 are met and µn is the channel input source of the Feinstein code, then Z 1 ˆn (Γi |wi ) = n ν ν n (Γi ) dµ(x) ≤ sup νxn (Γi ) µn (wi ) c(wi ) x x∈c(wi ) ≤
inf
x∈c(wi )
νxn ((Γi )δ ) + δ
and hence inf
x∈c(wi )
ˆn (Γi |wi ) − δ ≥ 1 − − δ. νxn ((Γi )δ ) ≥ ν
Thus if the channel block code is constructed using the expanded decoding sets, we have that max sup νx ((Γi )cδ ) ≤ + δ; i
x∈c(wi )
that is, the code {wi , (Γi )δ } is a (benR c, n, + δ) channel code. We have now proved the following result. Lemma 14.10. Let ν be a stationary d-continuous channel and Cn ; n ≥ n0 , a sequence of δ-robust (µn , benR c, n, ) Feinstein codes. Then for n1 sufficiently large and each n ≥ n1 there exists a (benR c, n, + δ) block channel code. Combining the lemma with Lemma 14.6 and Lemma 14.7 yields the following theorem.
14.7 Joint Source and Channel Block Codes
377
Theorem 14.1. Let ν be an AMS ergodic d-continuous channel. If R < C then given > 0 there is an n0 such that for all n ≥ n0 there exist (benR c, n, ) channel codes. If the channel is not ergodic, then the same holds true if C is replaced by C ∗ . Up to this point the channel coding theorems have been “one shot” theorems in that they consider only a single use of the channel. In a communication system, however, a channel will be used repeatedly in order to communicate a sequence of outputs from a source.
14.7 Joint Source and Channel Block Codes We can now combine a source block code and a channel block code of comparable rates to obtain a block code for communicating a source over a noisy channel. Suppose that we wish to communicate a source {Xn } with a distribution µ over a stationary and ergodic d-continuous ˆ The channel coding theorem states that if K is chochannel [B, ν, B]. sen to be sufficiently large, then we can reliably communicate length K messages from a collection of beKR c messages if R < C. Suppose that R = C − /2. If we wish to send the given source across this channel, then instead of having a source coding rate of (K/N) log kBk bits or nats per source symbol for a source (N, K) block code, we reduce the source coding rate to slightly less than the channel coding rate R, say Rsource = (K/N)(R − /2) = (K/N)(C − ). We then construct a block source codebook C of this rate with performance near the operational DRF δ(Rsource , µ) defined in (12.1). Every codeword in the source codebook is assigned a channel codeword as index. The source is encoded by selecting the minimum distortion word in the codebook and then inserting the resulting channel codeword into the channel. The decoder then uses its decoding sets to decide which channel codeword was sent and then puts out the corresponding reproduction vector. Since the indices of the source code words are accurately decoded by the receiver with high probability, the reproduction vector should yield performance near that of δ((K/N)(C − ), µ). Since is arbitrary and δ(R, µ) is a continuous function of R, this implies that the optimal achievable performance for block coding µ for ν is given by δ((K/N)C, µ), that is, by the operational distortion-rate function for block coding a source evaluated at the channel capacity normalized to bits or nats per source symbol. Making this argument precise yields the block joint source and channel coding theorem. A joint source and channel (N, K) block code consists of an encoder ˆN . It is assumed that N source ˆK → A α : AN → B K and decoder β : B time units correspond to K channel time units. The block code yields
378
14 Coding for Noisy Channels
ˆT defined by ˆT → A sequence coders α : AT → B T and β : B N α(x) = {α(xiN ); all i} N β(x) = {β(xiN ); all i}.
Let E denote the class of all such codes (all N and K consistent with the physical stationarity requirement). Let ∆∗ (µ, ν, E) denote the block coding operational distortion-rate function function and D(R, µ) the Shannon distortion-rate function of the source with respect to an additive fidelity criterion {ρn }. We assume also that ρn is bounded, that is, there is a finite value ρmax such that 1 ˆ n ) ≤ ρmax ρn (x n , x n for all n. This assumption is an unfortunate restriction, but it yields a simple proof of the basic result. Theorem 14.2. Let {Xn } be a stationary source with distribution µ and let ν be a stationary and ergodic d-continuous channel with channel capacity C. Let {ρn } be a bounded additive fidelity criterion. Given > 0 there exists for sufficiently large N and K (where K channel time units correspond to N source time units) an encoder α : AN → B K and decoder ˆN such that if α : AT → B T and β : B ˆT are the induced ˆK → A ˆT → A β:B sequence coders, then the resulting performance is bounded above as ˆN ) ≤ δ( ∆(µ, α, ν, β) = EρN (X N , X
K C, µ) + . N
Proof: Given , choose γ > 0 so that δ(
K K (C − γ), µ) ≤ δ( C, µ) + N N 3
and choose N large enough to ensure the existence of a source codebook C of length N and rate Rsource = (K/N)(C − γ) with performance ρ(C, µ) ≤ δ(Rsource , µ) +
. 3
We also assume that N (and hence also K) is chosen large enough so that for a suitably small δ (to be specified later) there exists a channel (beKR c, K, δ) code, with R = C − γ/2. Index the beNRsource c words in the source codebook by the beK(C−γ/2 c channel codewords. By construction there are more indices than source codewords so that this is possible. We now evaluate the performance of this code. Suppose that there are M words in the source codebook and hence ˆi and wi denote corresponding M of the channel words are used. Let x
14.7 Joint Source and Channel Block Codes
379
ˆi is the minimum distortion source and channel codewords, that is, if x word in the source codebook for an observed vector, then wi is transmitted over the channel. Let Γi denote the corresponding decoding region. Then ˆN ) = EρN (X N , X
M Z M X X i=1 j=1
=
x:α(x N )=wi
M Z X x:α(x N )=wi
i=1
+
M X
M X
i=1 j=1,j6=i
≤
+
x:α(x N )=wi
x:α(x N )=wi
M X
M X
i=1 j=1,J6=i
ˆi ) dµ(x)νxK (Γi )ρN (x N , x
Z
M Z X i=1
ˆj ) dµ(x)νxK (Γj )ρN (x N , x
ˆj ) dµ(x)νxK (Γj )ρN (x N , x
ˆi ) dµ(x)ρN (x N , x
Z x:α(x N )=wi
ˆj ) dµ(x)νxK (Γj )ρN (x N , x
The first term is bounded above by δ(Rsource , µ) + /3 by construction. The second is bounded above by ρmax times the channel error probability, which is less than δ by assumption. If δ is chosen so that ρmax δ is less than /2, the theorem is proved. 2 Theorem 14.3. Let {Xn } be a stationary source source with distribution µ and let ν be a stationary channel with channel capacity C. Let {ρn } be a bounded additive fidelity criterion. For any block stationary communication system (µ, f , ν, g), the average performance satisfies Z ∆(µ, f , ν, g) ≤ dµ(x)D(C, µ x ), x
where µ is the stationary mean of µ and {µ x } is the ergodic decomposition of µ, C is the capacity of the channel, and D(R, µ) the Shannon distortionrate function. N N K K ˆnN Proof: Suppose that the process {XnN , UnK , YnK ,X } is stationary and ˆ consider the overall mutual information rate I(X; X). From the data processing theorem (Lemma 8.7)
ˆ ≤ I(X; X)
K K I(U; Y ) ≤ C. N N
Choose L sufficiently large so that 1 K ˆn ) ≤ C + I(X n ; X n N
380
14 Coding for Noisy Channels
and Dn (
K K C + , µ) ≥ D( C + , µ) − δ N N
for n ≥ L. Then if the ergodic component µx is in effect, the performance can be no better than ˆN ) ≥ Eµx ρN (X n , X
inf
K p N ∈RN ( N
C+,µxN )
ˆN ) ≥ DN ( ρN (X N , X
K C + , µx ) N
which when integrated yields a lower bound of Z K dµ(x)D( C + , µx ) − δ. N Since δ and are arbitrary, the lemma follows from the continuity of the distortion rate function. 2 Combining the previous results yields the block coding optimal achievable performance for stationary sources and stationary and ergodic dcontinuous channels. Corollary 14.6. Let {Xn } be a stationary source with distribution µ and let ν be a stationary and ergodic d-continuous channel with channel capacity C. Let {ρn } be a bounded additive fidelity criterion. The block coding operational DRF of (5.14) is given by Z ∆(µ, ν, E, D) = dµ(x)D(C, µ x ).
14.8 Synchronizing Block Channel Codes As in the source coding case, the first step towards proving a sliding block coding theorem is to show that a block code can be synchronized, that is, that the decoder can determine (at least with high probability) where the block code words begin and end. Unlike the source coding case, this cannot be accomplished by the use of a simple synchronization sequence which is prohibited from appearing within a block code word since channel errors can cause an unintended appearance of the sync word at the receiver. The basic idea still holds, however, if the codes are designed so that it is very unlikely that a non-sync word can be converted into a valid sync word. If the channel is d-continuous, then good robust Feinstein codes as in Corollary 14.5 can be used to obtain good codebooks . The basic result of this section is Lemma 14.11 which states that given a sequence of good robust Feinstein codes, the code length can be chosen large enough to ensure that there is a sync word for a slightly modified codebook; that is, the synch word has length a speci-
14.8 Synchronizing Block Channel Codes
381
fied fraction of the codeword length and the sync decoding words never appear as a segment of codeword decoding words. The technique is due to Dobrushin [33] and is an application of Shannon’s random coding technique. The lemma originated in [68]. The basic idea of the lemma is this: In addition to a good long code, one selects a short good robust Feinstein code (from which the sync word will be chosen) and then performs the following experiment. A word from the short code and a word from the long code are selected independently and at random. The probability that the short decoding word appears in the long decoding word is shown to be small. Since this average is small, there must be at least one short word such that the probability of its decoding word appearing in the decoding word of a randomly selected long code word is small. This in turn implies that if all long decoding words containing the short decoding word are removed from the long code decoding sets, the decoding sets of most of the original long code words will not be changed by much. In fact, one must remove a bit more from the long word decoding sets in order to ensure the desired properties are preserved when passing from a Feinstein code to a channel codebook. Lemma 14.11. Assume that ≤ 1/4 and {Cn ; n ≥ n0 } is a sequence of -robust {τ, M(n), n, /2} Feinstein codes for a d-continuous channel ν having capacity C > 0. Assume also that h2 (2) + 2 log(kBk − 1) < C, where B is the channel output alphabet. Let δ ∈ (0, 1/4). Then there exists an n1 such that for all n ≥ n1 the following statements are true. (A) If Cn = {vi , Γi ; i = 1, . . . , M(n)}, then there is a modified codebook Wn = {wi ; Wi ; i = 1, . . . , K(n)} and a set of K(n) indices Kn = {k1 , · · · , kK(n) ⊂ {1, · · · , M(n)} such that wi = vki , Wi ⊂ (Γi )2 ; i = 1, . . . , K(n), and max
sup νxn (Wjc ) ≤ .
1≤j≤K(n) x∈c(wj )
(14.24)
(B) There is a sync word σ ∈ Ar , r = r (n) = dδne = smallest integer larger than δn, and a sync decoding set S ∈ BrB such that sup νxr (S c ) ≤ .
(14.25)
x∈c(σ )
and such that no r -tuple in S appears in any n-tuple in S Wi ; that is, if G(br ) = {y n : yir = br some i = 0, . . . , n−r } and G(S) = br ∈S G(br ), then \ G(S) Wi = ∅, i = 1, . . . , K(n). (14.26) (C)
We have that k{k : k 6∈ Kn }k ≤ δM(n).
(14.27)
382
14 Coding for Noisy Channels
The modified code Wn has fewer words than the original code Cn , but (14.27) ensures that Wn cannot be much smaller since K(n) ≥ (1 − δ)M(n).
(14.28)
Given a codebook Wn = {wi , Wi ; i = 1, . . . , K(n)}, a sync word σ ∈ Ar , and a sync decoding set S, we call the length n + r codebook {σ × wi , S × Wi ; i = 1, . . . , K(n)} a prefixed or punctuated codebook. Proof: Since ν is d-continuous, n2 can be chosen so large that for n ≥ n2 max
sup
an ∈An x,x 0 ∈c(an )
n dn (νxn , νx0 )≤(
δ 2 ) . 2
(14.29)
From Corollary 14.5 there is an n3 so large that for each r ≥ n3 there exists an /2-robust (τ, J, r , /2)-Feinstein code Cs = {sj , Sj : j = 1, . . . , J}; J ≥ 2r Rs , where Rs ∈ (0, C − h2 (2) − 2 log(kBk − 1)). Assume that n1 is large enough to ensure that δn1 ≥ n2 ; δn1 ≥ n3 , and n1 ≥ n0 . Let 1F denote the indicator function of the set F and define λn by λn = J −1
M(n) \ X 1 ˆn (G((Sj ) ) Γi |vi ) ν M(n) i=1 j=1
= J −1
M(n) X X 1 M(n) i=1 b0∈(S j=1
J X
J X
j )
= J −1
1 M(n)
M(n) X
X
ˆn (y n |vi )1G(b0 ) (y n ) ν
y n ∈Γi
X
i=1 y n ∈Γi
J X ˆn (y n |vi ) ν
X
1G(b0 ) (y n ) . (14.30)
j=1 b0 ∈(Sj )
Since the (Sj ) are disjoint and a fixed y n can belong to at most n−r ≤ n sets G(br ), the bracket term above is bound above by n and hence λn ≤
M(n) X n 1 n ˆn (y n |vi ) ≤ ν ≤ n2−r Rs ≤ n2−δnRs → 0 n→∞ J M(n) i=1 J
so that choosing n1 also so that n1 2−δnRs ≤ (δ)2 we have that λn ≤ (δ)2 if n ≥ n1 . From (14.30) this implies that for n ≥ n1 there must exist at least one j for which M(n) X
ˆn (G((Sj ) ) ν
\
Γi |vi ) ≤ (δ)2
i=1
which in turn implies that for n ≥ n1 there must exist a set of indices Kn ⊂ {1, · · · , M(n)} such that
14.8 Synchronizing Block Channel Codes
ˆn (G((Sj ) ) ν
\
383
Γi |vi ) ≤ δ, i ∈ Kn , k{i : i 6∈ Kn }k ≤ δ.
T Define σ = sj ; S = (Sj )/2 , wi = vki , and Wi = (Γki G((Sj ) )c )δ ; i = 1, . . . , K(n). We then have from Lemma 14.10 and (14.29) that if x ∈ c(σ ), then since δ ≤ /2 ˆr (Sj |σ ) − νxr (S) = νxr ((Sj )/2 ) ≥ ν
≥ 1 − , 2
proving (14.25). Next observe that if y n ∈ (G((Sj ) )c )δ , then there is a bn ∈ G((Sj ) )c such that dn (y n , bn ) ≤ δ and thus for i = 0, 1, . . . , n − r we have that n δ dr (yir , bir ) ≤ ≤ . r 2 2 Since bn ∈ G((Sj ) )c , it has no r -tuple within of an r -tuple in Sj and hence the r -tuples yir are at least /2 distant from Sj and hence y n ∈ H((S)/2 )c ). We have therefore that (G((Sj ) )c )δ ⊂ G((Sj ) )c and hence \ \ \ G(S) Wi = G((Sj ) ) (Γki G((Sj ) )c )δ \ ⊂ G((Sj )/2 ) (G((Sj ) )c )δ = ∅,
2
completing the proof.
Combining the preceding lemma with the existence of robust Feinstein codes at rates less than capacity (Lemma 14.10) we have proved the following synchronized block coding theorem. Corollary 14.7. Le ν be a stationary ergodic d-continuous channel and fix > 0 and R ∈ (0, C). Then there exists for sufficiently large blocklength N, a length N codebook {σ × wi , S × Wi ; i = 1, . . . , M}, M ≥ 2NR , σ ∈ Ar , wi ∈ An , r + n = N, such that sup νxr (S c ) ≤ ,
x∈c(σ )
max νxn (Wjc ) ≤ ,
i≤j≤M
Wj
\
G(S) = ∅.
Proof: Choose δ ∈ (0, /2) so small that C − h(2δ) − 2δ log(kBk − 1) > (1 + δ)R(1 − log(1 − δ2 )) and choose R 0 ∈ ((1 + δ)R(1 − log(1 − δ2 )), C − h(2δ) − 2δ log(kBk − 1). From Lemma 14.10 there exists an n0 such that for n ≥ n0 there exist δ-robust (τ, µ, n, δ) Feinstein codes with 0 M(n) ≥ 2nR . From Lemma 14.11 there exists a codebook {wi , Wi ; i = 1, . . . , K(n)}, a sync word σ ∈ Ar , and a sync decoding set S ∈ BrB , r = dδne such that
384
14 Coding for Noisy Channels
max sup νxn (Wjc ) ≤ 2δ ≤ , j
x∈c(wj )
sup νxr (S) ≤ 2δ ≤ ,
x∈c(σ )
T G(S) Wj = ∅; j = 1, . . . , K(n), and from (14.28) M = K(n) ≥ (1 − δ2 )M(n). Therefore for N = n + r 0
N −1 log M ≥ (ndnδe)−1 log((1 − δ2 )2nR ) nR 0 + log(1 − δ2 ) R 0 + n−1 log(1 − δ2 ) = n + nδ 1+δ R 0 + log(1 − δ2 ) ≥ ≥ R, 1+δ
=
completing the proof.
2
14.9 Sliding-block Source and Channel Coding Analogous to the conversion of block source codes into sliding-block source codes, the basic idea of constructing a sliding-block channel code is to use a punctuation sequence to stationarize a block code and to use sync words to locate the blocks in the decoded sequence. The sync word can be used to mark the beginning of a codeword and it will rarely be falsely detected during a codeword. Unfortunately, however, an r -tuple consisting of a segment of a sync and a segment of a codeword may be erroneously detected as a sync with nonnegligible probability. To resolve this confusion we look at the relative frequency of sync-detects over a sequence of blocks instead of simply trying to find a single sync. The idea is that if we look at enough blocks, the relative frequency of the syncdetects in each position should be nearly the probability of occurrence in that position and these quantities taken together give a pattern that can be used to determine the true sync location. For the ergodic theorem to apply, however, we require that blocks be ergodic and hence we first consider totally ergodic sources and channels and then generalize where possible.
Totally Ergodic Sources Lemma 14.12. Let ν be a totally ergodic stationary d-continuous channel. Fix , δ > 0 and assume that CN = {σ × wi ; S × Wi ; i = 1, . . . , K} is a prefixed codebook satisfying (14.24)–(14.26). Let γn : GN → CN assign an N-tuple in the prefixed codebook to each N-tuple in GN and let [G, µ, U]
14.9 Sliding-block Source and Channel Coding
385
be an N-stationary, N-ergodic source. Let c(an ) denote the cylinder set or rectangle of all sequences u = (· · · , u−1 , u0 , u1 , · · · ) for which un = an . There exists for sufficiently large L (which depends on the source) a sync locating function s : B LN → {0, 1, . . . , N − 1} and a set Φ ∈ Bm G, m = N (L + 1)N, such that if um ∈ Φ and γN (ULN ) = σ × wi , then inf
x∈c(γm (um ))
νx (y : s(y LN ) = θ, θ = 0, . . . , N − 1; yLN ∈ S × Wi ) ≥ 1 − 3. (14.31)
Comments: The lemma can be interpreted as follows. The source is block encoded using γN . The decoder observes a possible sync word and then looks “back” in time at previous channel outputs and calculates s(y LN ) to obtain the exact sync location, which is correct with high probability. The sync locator function is constructed roughly as follows: Since µ and ν are N-stationary and N-ergodic, if γ : A∞ → B ∞ is the sequence encoder induced by the length N block code γN , then the encoded source µγ −1 and the induced channel output process η are all N-stationary and N-ergodic. The sequence zj = η(T j c(S))); j = . . . , −1, 0, 1, . . . is therefore periodic with period N. Furthermore, zj can have no smaller period than N since from (14.24)–(14.26) η(T j c(S)) ≤ , j = r + 1, . . . , n − r and η(c(S)) ≥ 1 − . Thus defining the sync pattern {zj ; j = 0, 1, . . . , N − 1}, the pattern is distinct from any cyclic shift of itself of the form {zk , · · · , zN−1 , z0 , · · · , xk−1 }, where k ≤ N − 1. The sync locator computes the relative frequencies of the occurrence of S at intervals of length N for each of N possible starting points to obˆN = (ˆ ˆ1 , · · · , z ˆN−1 ). The ergodic theorem implies tain, say, a vector z z0 , z ˆi will be near their expectation and hence with high probathat the z ˆN−1 ) = (zθ , zθ+1 , · · · , zN−1 , z0 , · · · , zθ−1 ), determining bility (ˆ z0 , · · · , z θ. Another way of looking at the result is to observe that the sources ηT j ; j = 0, . . . , N − 1 are each N-ergodic and N-stationary and hence any two are either identical or orthogonal in the sense that they place all of their measure on disjoint N-invariant sets. (See, e.g., Exercise 1, Section 6.7 of [55] or Section 8.2 of [58].) No two can be identical, however, since if ηT i = ηT j for i 6= j; 0 ≤ i, j ≤ N − 1, then η would be periodic with period |i − j| strictly less than N, yielding a contradiction. Since membership in any set can be determined with high probability by observing the sequence for a long enough time, the sync locator attempts to determine which of the N distinct sources ηT j is being observed. In fact, synchronizing the output is exactly equivalent to forcing the N sources ηT j ; j = 0, 1, . . . , N − 1 to be distinct N-ergodic sources. After this is accomplished, the remainder of the proof is devoted to using the properties of d-continuous channels to show that synchronization of the output source when driven by µ implies that with high probability the channel output can be synchronized for all fixed input sequences in a set of high µ probability.
386
14 Coding for Noisy Channels
The lemma is stronger (and more general) than the similar results of Nedoma [130] and Vajda [180], but the extra structure is required for application to sliding-block decoding. Proof: Choose ζ > 0 so that ζ < /2 and ζ
0 and θ = 0, 1, . . . , N − 1 define the sets ψ(θ, α) ∈ BLN B ˜ ψ(θ, α) ∈ Bm , m = (L + 1)N by B ψ(θ, α) = {y LN : |
L−2 1 X r 1S (yj+iN ) − zθ+j | ≤ α; j = 0, 1, . . . , N − 1} L − 1 i=0
˜ ψ(θ, α) = B θ × ψ(θ, α) × B N−θ . From the ergodic theorem L can be chosen large enough so that N−1 \
η(
N−1 \
T −θ c(ψ(θ, ζ))) = ηm (
θ=0
˜ ψ(θ, ζ)) ≥ 1 − ζ 2 .
(14.33)
θ=0
Assume also that L is large enough so that if xi = xi0 , i = 0, . . . , m − 1 then ζ dm (νxm , νxm0 ) ≤ ( )2 . (14.34) N From (14.33) N−1 \
ζ 2 ≥ ηm ((
˜ ψ(θ, ζ))c ) =
am ∈Gm
θ=0
=
N−1 \
Z
X
X
c(am )
m dµ(u)νγ(u) ((
µ m (am )ˆ ν ((
am ∈Gm
˜ ψ(θ, ζ)c ))
θ=0 N−1 \
˜ ψ(θ, ζ))c |γm (am ))
θ=0
and hence there must be a set Φ ∈ Bm B such that N−1 \
ˆm (( ν
˜ ψ(θ, ζ))c |γm (am )) ≤ ζ, am ∈ Φ,
(14.35)
θ=0
µ m (Φ) ≤ ζ.
(14.36)
Define the sync locating function s : B LN → {0, 1, · · · , N − 1} as follows: Define the set ψ(θ) = {y LN ∈ (ψ(θ, ζ))2ζ/N } and then define θ
y LN ∈ ψ(θ)
1
otherwise
( s(y
LN
)=
14.9 Sliding-block Source and Channel Coding
387
We show that s is well-defined by showing that ψ(θ) ⊂ ψ(θ, 4ζ), which sets are disjoint for θ = 0, 1, . . . , N − 1 from (14.32). If y LN ∈ ψ(θ), there is a bLN ∈ ψ(θ, ζ) for which dLN (y LN , bLN ) ≤ 2ζ/N and hence for any j ∈ {0, 1, · · · , N − 1} at most LN(2ζ/N) = 2ζL of the consecutive N nonoverlapping N-tuples yj+iN , i = 0, 1, . . . , L − 2, can differ from the N and therefore corresponding bj+iN
|
L−2 L−2 1 X 1 X r r 1S (yj+iN ) − zθ+j | ≤ | 1S (bj+iN ) − zθ+j | + 2ζ ≤ 3ζ L − 1 i=0 L − 1 i=0
˜ is defined to be B θ × ψ(θ) × B N−θ ∈ and hence y LN ∈ ψ(θ, 4ζ). If ψ(θ) m BB , then we also have that (
N−1 \
˜ ψ(θ, ζ))ζ/N ⊂
θ=0
N−1 \
˜ ψ(θ)
θ=0
TN−1 ˜ since if y n ∈ ( θ=0 ψ(θ, ζ))ζ/N , then there is a bm such that bθLN ∈ ψ(θ, ζ); θ = 0, 1, . . . , N −1 and dm (y m , bm ) ≤ ζ/N for θ = 0, 1, . . . , N−1. This implies from Lemma 14.10 and (14.34)–(14.36) that if x ∈ γ m (am ) and am ∈ Φ, then νxm (
N−1 \
N−1 \
˜ ψ(θ)) ≥ νxm ((
θ=0
˜ ψ(θ, ζ))ζ/N )
θ=0 N−1 \
ˆ( ≥ν
˜ ψ(θ, ζ)|γ m (am )) −
θ=0
≥ 1−ζ −
ζ ≥ 1 − . N
ζ N (14.37)
To complete the proof, we use (14.24)–(14.26) and (14.37) to obtain for am ∈ Φ and γm (aN LN ) = σ × wi that N ∈ S × Wi ) νx (y : s(yθLN ) = θ, θ = 0, 1, . . . , N − 1; yLN
≥ νxm (
N−1 \
ψ(θ)) − νTN−NL x (S × Wic ) ≥ 1 − − 2.
θ=0
2 Next the prefixed block code and the sync locator function are combined with a random punctuation sequence of Lemma 2.12 to construct a good sliding-block code for a totally ergodic source with entropy less than capacity. Lemma 14.13. Given a d-continuous totally ergodic stationary channel ν with Shannon capacity C, a stationary totally ergodic source [G, µ, U]
388
14 Coding for Noisy Channels
with entropy rate H(µ) < C, and δ > 0, there exists for sufficiently large n, m a sliding-block encoder f : Gn → A and decoder g : B m → G such that Pe (µ, ν, f , g) ≤ δ. Proof: Choose R, H < R < C, and fix > 0 so that ≤ δ/5 and ≤ (R −H)/2. Choose N large enough so that the conditions and conclusions of Corollary 14.7 hold. Construct first a joint source and channel block encoder γN as follows: From the asymptotic equipartition property (Lemma 4.2 or Section 4.5) there is an n0 large enough to ensure that for N ≥ n0 the set GN = {uN : |N −1 hN (u) − H| ≥ } = {uN : e−N(H+) ≤ µ(uN ) ≤ e−N(H−) }
(14.38)
has probability µU N (GN ) ≥ 1 − .
(14.39)
Observe that if M = kGN k, then 2N(H−) ≤ M 0 ≤ 2N(H+) ≤ 2N(R−) .
(14.40)
Index the members of GN as βi ; i = 1, . . . , M 0 . If uN = βi , set γN (uN ) = σ ×wi . Otherwise set γN (uN ) = σ ×wM 0 +1 . Since for large N, 2N(R−) +1 ≤ 2NR , γN is well-defined. γN can be viewed as a synchronized extension of the almost noiseless code of Section 3.5. Define also the block decoder ψN (y N ) = βi if y N ∈ S × Wi ; i = 1, . . . , M 0 . Otherwise set ψN (y N ) = β∗ , an arbitrary reference vector. Choose L so large that the conditions and conclusions of Lemma 14.12 hold for C and γN . The sliding-block ˆk = decoder gm : B m → G, m = (L + 1)N, yielding decoded process U m gm (Yk−NL ) is defined as follows: If s(yk−NL , · · · , yk − 1) = θ, form bN = ˆk (y) = gm (yk−NL , · · · , yk+N ) = bθ , ψN (yk−θ , · · · , yk−θ−N ) and set U the appropriate symbol of the appropriate block. The sliding-block encoder f will send very long sequences of block words with random spacing to make the code stationary. Let K be a large number satisfying K ≥ L + 1 so that m ≤ KN and recall that N ≥ 3 and L ≥ 1. We then have that 1 1 ≤ ≤ . KN 3K 6
(14.41)
Use Corollary 2.1 to produce a (KN, ) punctuation sequence Zn using a finite length sliding-block code of the input sequence. The punctuation process is stationary and ergodic, has a ternary output and can produce only isolated 0’s followed by KN 1’s or individual 2’s. The punctuation sequence is then used to convert the block encoder γN into a slidingblock coder: Suppose that the encoder views an input sequence u = · · · , u−1 , u0 , u1 , · · · and is to produce a single encoded symbol x0 . If u0
14.9 Sliding-block Source and Channel Coding
389
is a 2, then the encoder produces an arbitrary channel symbol, say a∗ . If x0 is not a 2, then the encoder inspects u0 , u−1 , u−2 and so on into the past until it locates the first 0. This must happen within KN input symbols by construction of the punctuation sequence. Given that the first 1 occurs at, say, Zl = 1, the encoder then uses the block code γN to encode successive blocks of input N-tuples until the block including the symbol at time 0 is encoded. The sliding-block encoder than produces the corresponding channel symbol x0 . Thus if Zl = 1, then for some J < Kx0 = (γN (ul+JN ))l mod N where the subscript denotes that the (l mod N)th coordinate of the block codeword is put out. The final sliding-block code has a finite length given by the maximum of the lengths of the code producing the punctuation sequence and the code imbedding the block code γN into the sliding-block code. We now proceed to compute the probability of the error event {u, y : ˆ0 (y) 6= U0 (u)} = E. Let Eu denote the section {y : U ˆ0 (y) 6= U0 (u)}, f U be the sequence coder induced by f , and F = {u : Z0 (u) = 0}. Note that if u ∈ T −1 F , then T u ∈ F and hence Z0 (T u) = Z1 (u) since the coding is stationary. More generally, if uT −i F , then Zi = 0. By construction any 1 must be followed by KN 1’s and hence the sets T −i F are disjoint for i = 0, 1, . . . , KN − 1 and hence we can write ˆ0 ) = µν(E) = Pe = Pr(U0 6= U ≤
LN−1 X Z i=0
T −i F
Z dµ(u)νf (u) (Eu )
dµ(u)νf (u) (Eu ) +
KN−1 X Z i=LN
T −i F
dµ(u)νf (u) (Eu )
Z +
SKN−1
(
= LNµ(F ) +
KN−1 X Z i=LN
+
KN−1 X
X
i=LN akN ∈GkN
i=0
T −i F
Z u0 ∈T −i (F
T
c(aK N))
T −i F )c
dµ(u)
dµ(u)νf (u) (Eu ) + a ≤ 2 ˆ0 (u0 )), dµ(u0 )νf (u0 ) (y 0 : U0 (u0 ) 6= U (14.42)
where we have used the fact that µ(F ) ≤ (KN)−1 (from Corollary 2.1) and hence LNµ(F ) ≤ L/K ≤ . Fix i = kN + j; 0 ≤ j ≤ N − 1 and define u = T j+LN u0 and y = T j+LN y 0 , and the integrals become
390
14 Coding for Noisy Channels
Z u0 ∈T −i (F
T
m dµ(u0 )νf (u0 ) (y 0 : U0 (u0 ) 6= gm (Y−NL (y 0 )) Z dµ(u0 )× = T
c(aKN ))
u∈T −(k−L)N (F
νf (T −(j+LN) u) (y : U0 (T
j+LN
Z = u∈T −(k−L)N (F
T
c(aKN ))
c(aKN ))
u) 6= gm (Y− NLm (T j+NL y)))
dµ(u0 )νf (T −(j+LN) u) (y : uj+LN 6= gm (yjm ))
Z
dµ(u0 ) T u∈T −(k−L)N (F c(aKN )) N LN νf (T −(j+LN) u) (y : uN LN = ψN (yLN ) or s(yj =
×
6= j)).
(14.43)
N N N If uN LN = βj ∈ GN , then uLN = ψN (yLN ) if yLN ∈ S × Wi . If u ∈ m T −(k−L)N c(aKN ), then um = a(k−L)N and hence from Lemma 14.12 and stationarity we have for i = kN + j that
Z
X aKN ∈GKN
T −i (c(aKN )
F)
dµ(u)νf (u) (Eu ) X
≤ 3 ×
+
T
µ(T −(k−L)N (c(aKN )
\
F ))
KN aKN ∈ T G LN m a(k−L)N ∈ Φ (G × GN ) \ X µ(T −(k−L)N (c(aKN ) F )) KN aKN ∈ T G LN 6∈ Φ (G × GN ) \ X µ(c(aKN ) F )) ≤ 3 ×
am (k−L)N
aKN ∈GKN
+
X c am (k−L)N ∈Φ
µ(c(aKN )
\
F ))
(GLN ×GN )c
S
≤ 3µ(F ) + µ(c(Φc )
\
F ) + µ(c(GN )
\
F ).
(14.44)
Choose the partition in Lemmas 2.11–2.12 to be that generated by the sets c(Φc ) and c(GN ) (the partition with all four possible intersections of these sets or their complements). Then the above expression is bounded above by 3 + + ≤5 NK NK NK NK and hence from (14.42) Pe ≤ 5 ≤ δ, which completes the proof. The lemma immediately yields the following corollary.
2
Corollary 14.8. If ν is a stationary d-continuous totally ergodic channel with Shannon capacity C, then any totally ergodic source [G, µ, U] with H(µ) < C is admissible.
14.9 Sliding-block Source and Channel Coding
391
Ergodic Sources If a prefixed blocklength N block code of Corollary 14.8 is used to block encode a general ergodic source [G, µ, U], then successive N-tuples from µ may not be ergodic, and hence the previous analysis does not apply. From the Nedoma ergodic decomposition [129] (see, e.g., [55], p. 232, or [58], p. 253), any ergodic source µ can be represented as a mixture of Nergodic sources, all of which are shifted versions of each other. Given an ergodic measure µ and an integer N, then there exists a decomposition of µ into M N-ergodic, N-stationary components where M divides N, that is, there is a set Π ∈ B∞ G such that TMΠ = Π µ(T i Π
\
M−1 [
µ(
T j Π) = 0; i, j ≤ M, i 6= j T i Π) = 1
i=0
µ(Π) =
1 , M
such that the sources [G, µi , U], where πi (W ) = µ(W |T i Π) = Mµ(W
\
T i Π)
are N-ergodic and N-stationary and µ(W ) =
M−1 M−1 \ 1 X 1 X πi (W ) = µ(W T i Π). M i=0 M i=0
(14.45)
This decomposition provides a method of generalizing the results for totally ergodic sources to ergodic sources. Since µ(·|Π) is N-ergodic, Lemma 14.13 is valid if µ is replaced by µ(·|Π). If an infinite length sliding-block encoder f is used, it can determine the ergodic component in effect by testing for T −i Π in the base of the tower and insert i dummy symbols and then encode using the length N prefixed block code. In other words, the encoder can line up the block code with a prespecified one of the N-possible N-ergodic modes. A finite-length encoder can then be obtained by approximating the infinite-length encoder by a finite length encoder. Making these ideas precise yields the following result. Theorem 14.4. If ν is a stationary d-continuous totally ergodic channel with Shannon capacity C, then any ergodic source [G, µ, U] with H(µ) < C is admissible. Proof: Assume that N is large enough for Corollary 14.7 and (14.38)– (14.40) to hold. From the Nedoma decomposition
392
14 Coding for Noisy Channels M−1 1 X N µ (GN |T i Π) = µ N (GN ) ≥ 1 − M i=0
and hence there exists at least one i for which µ N (GN |T i Π) ≥ 1 − ; that is, at least one N-ergodic mode must put high probability on the set GN of typical N-tuples for µ. For convenience relabel the indices so that this good mode is µ(·|Π) and call it the design mode. Since µ(·|Π) is N-ergodic and N-stationary, Lemma 14.12 holds with µ replaced by µ(·|Π); that is, there is a source/channel block code (γN , ψN ) and a sync locating function s : B LN → {0, 1, · · · , M − 1} such that there is a set Φ ∈ Gm ; m = (L + 1)N, for which (14.31) holds and µ m (Φ|Π) ≥ 1 − . The sliding-block decoder is exactly as in Lemma 14.12. The slidingblock encoder, however, is somewhat different. Consider a punctuation sequence or tower as in Lemma 2.12, but now consider the partition generated by Φ, GN , and T i Π, i = 0, 1, . . . , M S − 1. The infinite length NK−1 sliding-block code is defined as follows: If u 6∈T k=0 T k F , then f (u) = ∗ i a , an arbitrary channel symbol. If u ∈ T (F T −j Π) and if i < j, set f (u) = a∗ (these are spacing symbols to force alignment with the proper N-ergodic mode). If j ≤ i ≤ KN − (M − j), then i = j + kN + r for some N 0 ≤ k ≤ (K − 1)N, r ≤ N − 1. Form GN (uN j+kN ) = a and set f (u) = ar . This is the same encoder as before, except that if u ∈ T j Π, then block encoding is postponed for j symbols (at which time u ∈ Π). Lastly, if KN − (M − j) ≤ i ≤ KN − 1, then f (u) = a∗ . As in the proof of Lemma 14.13 Z Pe (µ, ν, f , gm ) = ≤ 2 +
m dµ(u)νf (u) (y : U0 (u) 6= gm (Y−LN (y)))
KN−1 X Z
ˆ0 (y)) u ∈ T i F dµ(u)νf (u) (y : U0 (u) 6= U
i=LN
= 2 +
KN−1 X M−1 X
X
i=LN j=0 aKN ∈GKN
Z u∈T i (c(aKN )
T
F
T
T −j Π)
≤ 2 +
ˆ0 (y)) dµ(u)νf (u) (y : U0 (u) 6= U
M−1 X KN−(M−j) X j=0
Z u∈T i (c(aKN )
T
F
T
T −j Π)
i=LN+j
X aKN ∈GKN
ˆ0 (y)) dµ(u)νf (u) (y : U0 (u) 6= U +
M−1 X j=0
Mµ(F
\
T −j Π),
(14.46)
14.9 Sliding-block Source and Channel Coding
393
where the rightmost term is M
M−1 X
µ(F
\
T −j Π) ≤
j=0
1 M ≤ ≤ . KN K
Thus
Pe (µ, ν, f , gm ) ≤ 3 +
M−1 X KN−(M−j) X j=0
X aKN ∈GKN
i=LN+j
Z u∈T i (c(aKN )
T
F
T
T −j Π)
ˆ0 (y)). dµ(u)νf (u) (y : U0 (u) 6= U
Analogous to (14.43) (except that here i = j + kN + r , u = T −(LN+r ) u0 ) Z u0 ∈T i (c(aKN )
T
F
T
T −j Π)
m dµ(u0 )νf (u0 ) (y 0 : U0 (u0 ) = gm (Y−LN (y 0 )))
Z dµ(u)×
≤ T j+(k−L)N (c(aKN )
T
νf (T i +LNu) (y :
T −j Π) N N uLN 6= ψN (yLN )
F
T
or s(yrLN ) 6= r ).
T T Since u ∈ T j+(k−L)N (c(aKN ) F T −j Π implies um = am j+(k−L)N , analogous to (14.44) we have that for i = j + kN + r X aKN ∈GKN
Z T i (c(aKN )
T
F
T
T −j Π)
X
=
dµ(u)νf (u) (y : U0 (u) 6= gm (Y− LN m (y))) \
µ(T j+(k−L)N (c(aKN )
F
\
aKN :am j+(k−L)N ∈Φ
+
X
\
µ(T j+(k−L)N (c(aKN )
F
aKN :am j+(k−L)N 6∈Φ
=
X
µ(c(aKN )
\
F
\
aKN :am j+(k−L)N ∈Φ
+
X
µ(c(aKN )
\
F
\
aKN :am j+(k−L)N 6∈Φ
= µ(T −(j+(k−L)N) c(Φ)
\
F
\
\
T −j Π))
T −j Π))
T −j Π)
T −j Π)
T −j Π)
+ µ(T −(j+(k−L)N) c(Φ)c
\
F
\
T −j Π).
From Lemma 2.12 (the Rohlin-Kakutani theorem), this is bounded above by
394
14 Coding for Noisy Channels
T T µ(T −(j+(k−L)N) c(Φ) T −j Π) µ(T −(j+(k−L)N) c(Φ)c T −j Π) + KN KN µ(T −(j+(k−L)N) c(Φ)|T −j Π)µ(Π) µ(T −(j+(k−L)N) c(Φ)c |T −j Π)µ(Π) + = KN KN µ(Π) 2 µ(Π) = µ(c(Φ)|Π) µ(c(Φ)c |Π) +≤ . KN KN MKN
With (14.45)–(14.46) this yields Pe (µ, ν, f , gm ) ≤ 3 +
MKN2 ≤ 5, MKN
(14.47)
which completes the result for an infinite sliding-block code. The proof is completed by applying Corollary 5.2, which shows that by choosing a finite length sliding-block code f0 from Lemma 5.2 so that Pr(f 6= f0 ) is sufficiently small, then the resulting Pe is close to that for the infinite length sliding-block code. 2 The theorem can be combined with the sliding block source coding theorem to prove a joint source and channel coding theorem similar to Theorem 14.2, that is, one can show that given a source with distortion rate function D(R) and a channel with capacity C, then sliding-block codes exist with average distortion approximately D(C). We have considered only discrete channels, which is less general than the continuous additive Gaussian noise channels considered in many classic information theory texts. On the other hand, we have considered more general memory structures than are usually encountered, and we have followed the common thread of the book to develop coding theorems for sliding-block codes as well as block codes.
References
1. N. M. Abramson. Information Theory and Coding. McGraw-Hill, New York, 1963. 2. R. Adler. Ergodic and mixing properties of infinite memory channels. Proc. Amer. Math. Soc., 12:924–930, 1961. 3. R. L. Adler, D. Coppersmith, and M. Hassner. Algorithms for sliding-block codes–an application of symbolic dynamics to information theory. IEEE Trans. Inform. Theory, IT-29:5–22, 1983. 4. R. Ahlswede and P. Gács. Two contributions to information theory. In Topics in Information Theory, pages 17–40, Keszthely,Hungary, 1975. 5. R. Ahlswede and J. Wolfowitz. Channels without synchronization. Adv. in Appl. Probab., 3:383–403, 1971. 6. P. Algoet. Log-Optimal Investment. PhD thesis, Stanford University, 1985. 7. P. Algoet and T. Cover. A sandwich proof of the Shannon-McMillan-Breiman theorem. Ann. Probab., 16:899–909, 1988. 8. A. R. Barron. The strong ergodic theorem for densities: generalized ShannonMcMillan-Breiman theorem. Ann. Probab., 13:1292–1303, 1985. 9. T.W. Benjamin. Rate Distortion Functions for Discrete Sources with Continuous Reproductions. PhD thesis, Cornell University, Ithaca, New York, 1973. 10. T. Berger. Rate distortion theory for sources with abstract alphabets and memory. Inform. and Control, 13:254–273, 1968. 11. T. Berger. Rate Distortion Theory. Prentice-Hall Inc., Englewood Cliffs,New Jersey, 1971. 12. E. Berlekamp. Algebraic Coding Theory. McGraw-Hill, New York, 1968. 13. E. Berlekamp, editor. Key Papers in the Development of Coding Theory. IEEE Press, New York, 1974. 14. James Bezdek and Richard Hathaway. Some notes on alternating optimiza˘T ˇ tion. In Nikhil Pal and Michio Sugeno, editors, Advances in Soft Computing âA AFSS 2002, volume 2275 of Lecture Notes in Computer Science, pages 187–195. Springer Berlin / Heidelberg, 2002. 10.1007/3-540-45631-7_39. 15. P. J. Bickel and D. A. Freedman. Some asymptotic theory for the bootstrap. Annals of Statistics, 9(1):196–1217, 1981. 16. P. Billingsley. Ergodic Theory and Information. Wiley, New York, 1965. 17. G. D. Birkhoff. Proof of the ergodic theorem. Proc. Nat. Acad. Sci., 17:656–660, 1931. 18. R. E. Blahut. Computation of channel capacity and rate-distortion functions. IEEE Trans. Inform. Theory, IT-18:460–473, 1972. 19. R. E. Blahut. Theory and Practice of Error Control Codes. Addison Wesley, Reading, Mass., 1987.
R.M. Gray, Entropy and Information Theory, DOI 10.1007/978-1-4419-7970-4, © Springer Science+Business Media, LLC 2011
395
396
References
20. L. Breiman. The individual ergodic theorem of information theory. Ann. of Math. Statist., 28:809–811, 1957. 21. L. Breiman. A correction to ‘The individual ergodic theorem of information theory’. Ann. of Math. Statist., 31:809–810, 1960. 22. J. R. Brown. Ergodic Theory and Topological Dynamics. Academic Press, New York, 1976. 23. T. M. Cover, P. Gacs, and R. M. Gray. Kolmogorov’s contributions to information theory and algorithmic complexity. Ann. Probab., 17:840–865, 1989. 24. I. Csiszár. Information-type measures of difference of probability distributions and indirect observations. Studia Scientiarum Mathematicarum Hungarica, 2:299–318, 1967. 25. I. Csiszár. On an extremum problem of information theory. Studia Scientiarum Mathematicarum Hungarica, pages 57–70, 1974. 26. I. Csiszár. I-divergence geometry of probability distributions and minimization problems. Ann. Probab., 3(1):146–158, 1975. 27. I. Csiszár and J. Körner. Coding Theorems of Information Theory. Academic Press/Hungarian Academy of Sciences, Budapest, 1981. 28. L. D. Davisson and R.M. Gray. A simplified proof of the sliding-block source coding theorem and its universal extension. In Conf. Record 1978 Int’l. Conf. on Comm. 2, pages 34.4.1–34.4.5, Toronto, 1978. 29. L. D. Davisson and M. B. Pursley. An alternate proof of the coding theorem for stationary ergodic sources. In Proceedings of the Eighth Annual Princeton Conference on Information Sciences and Systems, 1974. 30. M. Denker, C. Grillenberger, and K. Sigmund. Ergodic Theory on Compact Spaces, volume 57 of Lecture Notes in Mathematics. Springer-Verlag, New York, 1970. 31. J.-D. Deushcel and D. W. Stroock. Large Deviations, volume 137 of Pure and Applied Mathematics. Academic Press, Boston, 1989. 32. R. L. Dobrushin. A general formulation of the fundamental Shannon theorem in information theory. Uspehi Mat. Akad. Nauk. SSSR, 14:3–104, 1959. Translation in Transactions Amer. Math. Soc, series 2,vol. 33,323–438. 33. R. L. Dobrushin. Shannon’s theorems for channels with synchronization errors. Problemy Peredaci Informatsii, 3:18–36, 1967. Translated in Problems of Information Transmission, vol.,3,11–36 (1967),Plenum Publishing Corporation. 34. R.L. Dobrushin. Prescribing a system of random variables by conditional distributions. Theor. Prob. Appl., 15:458–486, 1970. 35. M. D. Donsker and S. R. S. Varadhan. Asymptotic evaluation of certain Markov process expectations for large time. J. Comm. Pure Appl. Math., 28:1–47, 1975. 36. J. G. Dunham. A note on the abstract alphabet block source coding with a fidelity criterion theorem. IEEE Trans. Inform. Theory, IT-24:760, November 1978. 37. P. Elias. Two famous papers. IRE Transactions on Information Theory, page 99, 1958. 38. R. M. Fano. Transmission of Information. Wiley, New York, 1961. 39. A. Feinstein. A new basic theorem of information theory. IRE Transactions on Information Theory, pages 2–20, 1954. 40. A. Feinstein. Foundations of Information Theory. McGraw-Hill, New York, 1958. 41. A. Feinstein. On the coding theorem and its converse for finite-memory channels. Inform. and Control, 2:25–44, 1959. 42. W. A. Finamore and W. A. Pearlman. Optimal encoding of discrete-time, continuous-amplitude, memoryless sources with finite output alphabets. IEEE Trans. Inform. Theory, 26:144–155, Mar. 1980. 43. S.L. Fix. Rate distortion functions for continuous alphabet memoryless sources. PhD thesis, University of Michigan, Ann Arbor, Michigan, 1977.
References
397
44. G. D. Forney, Jr. The Viterbi algorithm. Proc. IEEE, 61:268–278, March 1973. 45. M. Fréchet. Sur la distance de deux lois de probabilité. C. R. Acad. Sci. Paris, 244:689–692, 1957. 46. N. A. Friedman. Introduction to Ergodic Theory. Van Nostrand Reinhold Company, New York, 1970. 47. R. G. Gallager. Information Theory and Reliable Communication. John Wiley & Sons, New York, 1968. 48. R.G. Gallager. Low Density Parity Check Codes. MIT Press, 1963. 49. I. M. Gelfand, A. N. Kolmogorov, and A. M. Yaglom. On the general definitions of the quantity of information. Dokl. Akad. Nauk, 111:745–748, 1956. (In Russian.). 50. A. Gersho and R. M. Gray. Vector Quantization and Signal Compression. Kluwer Academic Publishers, Boston, 1992. 51. C. Gini. Sulla misura della concentrazione e della variabilit‘a del caratteri. Atti del R. Instituto Veneto, pages 1913–1914, 1914. 52. R. M. Gray. Sliding-block source coding. IEEE Trans. Inform. Theory, IT21(4):357–368, July 1975. 53. R. M. Gray. Time-invariant trellis encoding of ergodic discrete-time sources with a fidelity criterion. IEEE Trans. Inform. Theory, IT-23:71–83, 1977. 54. R. M. Gray. Tree-searched block source codes. In Proceedings of the 1980 Allerton Conference, Allerton IL, Oct. 1980. 55. R. M. Gray. Probability, Random Processes, and Ergodic Properties. SpringerVerlag, New York, 1988. 56. R. M. Gray. Spectral analysis of quantization noise in a single-loop sigma-delta modulator with dc input. IEEE Trans. Comm., COM-37:588–599, 1989. 57. R. M. Gray. Source Coding Theory. Kluwer Academic Press, Boston, 1990. 58. R. M. Gray. Probability, Random Processes, and Ergodic Properties, Second Ed. Springer, New York, 2009. 59. R. M. Gray and L. D. Davisson. Source coding without the ergodic assumption. IEEE Trans. Inform. Theory, IT-20:502–516, 1974. 60. R. M. Gray and L. D. Davisson. Quantizer mismatch. IEEE Trans. on Comm., 23(4):439–442, April 1975. 61. R. M. Gray and E. Karnin. Multiple local optima in vector quantizers. IEEE Trans. Inform. Theory, IT-28:256–261, March 1982. 62. R. M. Gray and J. C. Kieffer. Asymptotically mean stationary measures. Ann. Probab., 8:962–973, 1980. 63. R. M. Gray, D. L. Neuhoff, and J. K. Omura. Process definitions of distortion rate functions and source coding theorems. IEEE Trans. Inform. Theory, IT-21:524– 532, 1975. 64. R. M. Gray, D. L. Neuhoff, and J. K. Omura. Process definitions of distortion rate functions and source coding theorems. IEEE Trans. on Info. Theory, 21(5):524– 532, Sept. 1975. 65. R. M. Gray, D. L. Neuhoff, and D. Ornstein. Nonblock source coding with a fidelity criterion. Ann. Probab., 3:478–491, 1975. 66. R. M. Gray, D. L. Neuhoff, and P. C. Shields. A generalization of ornstein’s dbar distance with applications to information theory. Ann. Probab., 3:315–328, April 1975. 67. R. M. Gray and D. S. Ornstein. Sliding-block joint source/noisy-channel coding theorems. IEEE Trans. Inform. Theory, IT-22:682–690, 1976. 68. R. M. Gray, D. S. Ornstein, and R. L. Dobrushin. Block synchronization,slidingblock coding, invulnerable sources and zero error codes for discrete noisy channels. Ann. Probab., 8:639–674, 1980. 69. R. M. Gray, M. Ostendorf, and R. Gobbi. Ergodicity of Markov channels. IEEE Trans. Inform. Theory, 33:656–664, September 1987.
398
References
70. R. M. Gray and F. Saadat. Block source coding theory for asymptotically mean stationary sources. IEEE Trans. Inform. Theory, 30:64–67, 1984. 71. R.M. Gray and D. L. Neuhoff. Quantization. IEEE Transactions on Information Theory, 44:2325–2384, October 1998. 72. Robert M. Gray and Tamás Linder. Bits in asymptotically optimal lossy source codes are asymptotically bernoulli. In J. A. Storer and M. Cohn, editors, Proceedings 2009 Data Compression Conference (DCC), pages 53–62, March 2009. 73. P. R. Halmos. Lectures on Ergodic Theory. Chelsea, New York, 1956. 74. G. H. Hardy, J. E. Littlewood, and G. Polya. Inequalities. Cambridge Univ. Press, London, 1952. Second Edition,1959. 75. R. V. L. Hartley. Transmission of information. Bell System Tech. J., 7:535–563, 1928. 76. T. Hashimoto. A direct proof of the equality between the block definition and the process definition of distortion-rate functions for stationary ergodic sources. nform. Contr., 51:45–57, Oct. 1981. 77. E. Hoph. Ergodentheorie. Springer-Verlag, Berlin, 1937. 78. K. Jacobs. Die Übertragung diskreter Informationen durch periodishce und fastperiodische Kanale. Math. Annalen, 137:125–135, 1959. 79. K. Jacobs. Über die Struktur der mittleren Entropie. Math. Z., 78:33–43, 1962. 80. K. Jacobs. The ergodic decomposition of the Kolmogorov-Sinai invariant. In F. B. Wright and F. B. Wright, editors, Ergodic Theory. Academic Press, New York, 1963. 81. N. S. Jayant and P. Noll. Digital Coding of Waveforms. Prentice-Hall, Englewood Cliffs,New Jersey, 1984. 82. T. Kadota. Generalization of feinstein’s fundamental lemma. IEEE Trans. Inform. Theory, IT-16:791–792, 1970. 83. S. Kakutani. Induced measure preserving transformations. In Proceedings of the Imperial Academy of Tokyo, volume 19, pages 635–641, 1943. 84. S. Kalikow and R. McCutcheon. An Outline of Ergodic Theory, volume 122 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2010. 85. L.V. Kantorovich. On the transfer of masses. Uspekhi Mat. Nauk, 37:7–8, 1942. Villani has this as “On the translocation of masses” C.R. (Dikl.) Acad. Sci. URSS 37 (1942), 199–201. 86. L.V. Kantorovich. On a problem of monge. Dokl. Akad. Nauk, 3:225–226, 1948. 87. A. J. Khinchine. The entropy concept in probability theory. Uspekhi Matematicheskikh Nauk., 8:3–20, 1953. Translated in Mathematical Foundations of Information Theory,Dover New York (1957). 88. A. J. Khinchine. On the fundamental theorems of information theory. Uspekhi Matematicheskikh Nauk., 11:17–75, 1957. Translated in Mathematical Foundations of Information Theory,Dover New York (1957). 89. J. C. Kieffer. A counterexample to Perez’s generalization of the ShannonMcMillan theorem. Ann. Probab., 1:362–364, 1973. 90. J. C. Kieffer. A general formula for the capacity of stationary nonanticipatory channels. Inform. and Control, 26:381–391, 1974. 91. J. C. Kieffer. On the optimum average distortion attainable by fixed-rate coding of a nonergodic source. IEEE Trans. Inform. Theory, IT-21:190–193, March 1975. 92. J. C. Kieffer. A unified approach to weak universal source coding. IEEE Trans. Inform. Theory, IT-24:674–682, 1978. 93. J. C. Kieffer. Extension of source coding theorems for block codes to sliding block codes. IEEE Trans. Inform. Theory, IT-26:679–692, 1980. 94. J. C. Kieffer. Block coding for weakly continuous channels. IEEE Trans. Inform. Theory, IT-27:721–727, 1981. 95. J. C. Kieffer. Sliding-block coding for weakly continuous channels. IEEE Trans. Inform. Theory, IT-28:2–10, 1982.
References
399
96. J. C. Kieffer. An ergodic theorem for constrained sequences of functions. Bulletin American Math Society, 1989. 97. J. C. Kieffer. Elementary information theory. Unpublished manuscript, 1990. 98. J. C. Kieffer and M. Rahe. Markov channels are asymptotically mean stationary. Siam Journal of Mathematical Analysis, 12:293–305, 1980. 99. J.C. Kieffer. Sample converses in source coding theory. Information Theory, IEEE Transactions on, 37(2):263 –268, March 1991. 100. J.C. Kieffer. Strong converses in source coding relative to a fidelity criterion. Information Theory, IEEE Transactions on, 37(2):257 –262, March 1991. 101. A. N. Kolmogorov. On the Shannon theory of information in the case of continuous signals. IRE Transactions Inform. Theory, IT-2:102–108, 1956. 102. A. N. Kolmogorov. A new metric invariant of transitive dynamic systems and automorphisms in lebesgue spaces. Dokl. Akad. Nauk SSR, 119:861–864, 1958. (In Russian.). 103. A. N. Kolmogorov. On the entropy per unit time as a metric invariant of automorphisms. Dokl. Akad. Naud SSSR, 124:768–771, 1959. (In Russian.). 104. A. N. Kolmogorov, A. M. Yaglom, and I. M. Gelfand. Quantity of information and entropy for continuous distributions. In Proceedings 3rd All-Union Mat. Conf., volume 3, pages 300–320. Izd. Akad. Nauk. SSSR, 1956. 105. S. Kullback. A lower bound for discrimination in terms of variation. IEEE Trans. Inform. Theory, IT-13:126–127, 1967. 106. S. Kullback. Information Theory and Statistics. Dover, New York, 1968. Reprint of 1959 edition published by Wiley. 107. B. M. Leiner and R. M. Gray. Bounds on rate-distortion functions for stationary sources and context-dependent fidelity criteria. IEEE Trans. Inform. Theory, IT-19:706–708, Sept. 1973. 108. E. Levina and P. Bickel. The earth mover’s distance is the mallows distance: some insights from statistics. In Proceedings Eighth IEEE International Conference on Computer Vision (ICCV 2001), volume 2, pages 251–256, 2001. 109. S. Lin. Introduction to Error Correcting Codes. Prentice-Hall, Englewood Cliffs,NJ, 1970. 110. S. P. Lloyd. Least squares quantization in PCM. Unpublished Bell Laboratories Technical Note. Portions presented at the Institute of Mathematical Statistics Meeting Atlantic City New Jersey September 1957. Published in the March 1982 special issue on quantization of the IEEE Transactions on Information Theory, 1957. 111. K. M. Mackenthun and M. B. Pursley. Strongly and weakly universal source coding. In Proceedings of the 1977 Conference on Information Science and Systems, pages 286–291, Johns Hopkins University, 1977. 112. J. MacQueen. Some methods for classification and analysis of multivariate observations. In Proc. of the Fifth Berkeley Symposium on Math. Stat. and Prob., volume 1, pages 281–296, 1967. 113. F. J. MacWilliams and N. J. A. Sloane. The Theory of Error-Correcting Codes. North-Holland, New York, 1977. 114. A. Maitra. Integral representations of invariant measures. Transactions of the American Mathematical Society, 228:209–235, 1977. 115. J. Makhoul, S. Roucos, and H. Gish. Vector quantization in speech coding. Proc. IEEE, 73. No. 11:1551–1587, November 1985. 116. C.L. Mallows. A note on asymptotic joint normality. Ann. Math.l Stat., 43:508– 515, 1972. 117. Mark Z. Mao, Robert M. Gray, , and Tamás Linder. Rate-constrained simulation and source coding iid sources. submitted for possible publication. 118. B. Marcus. Sophic systems and encoding data. IEEE Trans. Inform. Theory, IT-31:366–377, 1985.
400
References
119. K. Marton. On the rate distortion function of stationary sources. Problems of Control and Information Theory, 4:289–297, 1975. 120. K. Marton. A simple proof of the blowing-up lemma. TIT, 32:445–446, 1986. 121. K. Marton. Bounding d-distance by informational divergence: a method to prove measure concentration. Annals of Probability, 24(2):857–866, 1996. 122. R. McEliece. The Theory of Information and Coding. Cambridge University Press, New York, NY, 1984. 123. B. McMillan. The basic theorems of information theory. Ann. of Math. Statist., 24:196–219, 1953. 124. L. D. Meshalkin. A case of isomorphisms of bernoulli scheme. Dokl. Akad. Nauk SSSR, 128:41–44, 1959. (In Russian.). 125. G. Monge. Mémoire sur la théorie des déblais et des remblais. 1781. 126. A. Montanari and R. Urbanke. Modern coding theory: the statistical mechanics and computer science point of view. In SUMMER SCHOOL ON COMPLEX SYSTEMS, LES HOUCHES, pages 704–2857, 2007. 127. Shu-Teh C. Moy. Generalizations of Shannon-McMillan theorem. Pacific Journal Math., 11:705–714, 1961. 128. N. M. Nasrabadi and R. A. King. Image coding using vector quantization: A review. IEEE Transactions on Communications, COM-36:957–971, August 1988. 129. J. Nedoma. On the ergodicity and r-ergodicity of stationary probability measures. Z. Wahrsch. Verw. Gebiete, 2:90–97, 1963. 130. J. Nedoma. The synchronization for ergodic channels. Transactions Third Prague Conf. Information Theory, Stat. Decision Functions, and Random Processes, pages 529–539, 1964. 131. D. L. Neuhoff and R. K. Gilbert. Causal source codes. IEEE Trans. Inform. Theory, IT-28:701–713, 1982. 132. D. L. Neuhoff, R. M. Gray, and L. D. Davisson. Fixed rate universal block source coding with a fidelity criterion. IEEE Trans. Inform. Theory, 21:511–523, 1975. 133. D. L. Neuhoff and P. C. Shields. Channels with almost finite memory. IEEE Trans. Inform. Theory, pages 440–447, 1979. 134. D. L. Neuhoff and P. C. Shields. Channel distances and exact representation. Inform. and Control, 55(1), 1982. 135. D. L. Neuhoff and P. C. Shields. Channel entropy and primitive approximation. Ann. Probab., 10(1):188–198, 1982. 136. D. L. Neuhoff and P. C. Shields. Indecomposable finite state channels and primitive approximation. IEEE Trans. Inform. Theory, IT-28:11–19, 1982. 137. David Lee Neuhoff. Source coding and distance measures on random processes. PhD thesis, Stanford University, Stanford, California, August 1974. 138. D. Ornstein. Bernoulli shifts with the same entropy are isomorphic. Advances in Math., 4:337–352, 1970. 139. D. Ornstein. An application of ergodic theory to probability theory. Ann. Probab., 1:43–58, 1973. 140. D. Ornstein. Ergodic Theory,Randomness,and Dynamical Systems. Yale University Press, New Haven, 1975. 141. D. Ornstein and B. Weiss. The Shannon-McMillan-Breiman theorem for a class of amenable groups. Israel J. of Math, 44:53–60, 1983. 142. P. Papantoni-Kazakos and R. M. Gray. Robustness of estimators on stationary observations. Ann. Probab., 7:989–1002, Dec. 1979. 143. W. A. Pearlman. Sliding-block and random source coding with constrained size reproduction alphabets. IEEE Trans. Commun., 30(8):1859–1867, Aug. 1982. 144. W. B. Pennebaker and J. Mitchell. JPEG Still Image Data Compression Standard. Kluwer, Norwell, Massachusetts, 2004. 145. A. Perez. Notions généralisees d’incertitude,d’entropie et d’information du point de vue de la théorie des martingales. In Transactions First Prague Conf. on Information Theory, Stat. Decision Functions,and Random Processes, pages 183–208. Czech. Acad. Sci. Publishing House, 1957.
References
401
146. A. Perez. Sur la convergence des incertitudes,entropies et informations échantillon vers leurs valeurs vraies. In Transactions First Prague Conf. on Information Theory, Stat. Decision Functions,and Random Processes, pages 245–252. Czech. Acad. Sci. Publishing House, 1957. 147. A. Perez. Sur la théorie de l’information dans le cas d’un alphabet abstrait. In Transactions First Prague Conf. on Information Theory, Stat. Decision Functions,Random Processes, pages 209–244. Czech. Acad. Sci. Publishing House, 1957. 148. A. Perez. Extensions of Shannon-McMillan’s limit theorem to more general stochastic processes processes. In Third Prague Conf. on Inform. Theory,Decision Functions,and Random Processes, pages 545–574, Prague and New York, 1964. Publishing House Czech. Akad. Sci. and Academic Press. 149. K. Petersen. Ergodic Theory. Cambridge University Press, Cambridge, 1983. 150. M. S. Pinsker. Dynamical systems with completely positive or zero entropy. Soviet Math. Dokl., 1:937–938, 1960. 151. M. S. Pinsker. Information and information stability of random variables and processes. Holden Day, San Francisco, 1964. Translated by A. Feinstein from the Russian edition published in 1960 by Izd. Akad. Nauk. SSSR. 152. M. Rabbani and P. W. Jones. Digital Image Compression Techniques, volume TT7 of Tutorial Texts in Optical Engineering. SPIE Optical Engineering Press, Bellingham, Washington, 1991. 153. S.T. Rachev. Probability Metrics and the Stability of Stochastic Models. John Wiley & Sons Ltd, Chichester, 1991. 154. S.T. Rachev and L. Rüschendorf. Mass Transportation Problems Vol. I: Theory, Vol. II: Applications. Probability and its applications. Springer- Verlag, New York, 1998. 155. D. Ramachandran. Perfect Measures. ISI Lecture Notes,No. 6 and 7. Indian Statistical Institute, Calcutta,India, 1979. 156. T. Richardson and R. Urbanke. Modern Coding Theory. Cambridge University Press, March 2008. 157. V. A. Rohlin and Ya. G. Sinai. Construction and properties of invariant measurable partitions. Soviet Math. Dokl., 2:1611–1614, 1962. 158. K. Rose. A mapping approach to rate-distortion computation and analysis. EEE Trans. Inform. Theory, 40(6):1939–1952, Nov. 1994. 159. Y. Rubner, C. Tomasi, and L.J. Guibas. A metric for distributions with applications to image databases. In Proceedings of the IEEE International Conference on Computer Vision, pages 59–66, Bombay, India, Jan 1998. 160. K. Sayood. Introduction to Data Compression. Morgan Kaufmann Publishers, San Francisco, second edition, 2000. 161. V. V. Sazanov. On perfect measures. Izv. Akad. Nauk SSSR, 26:391–414, 1962. American Math. Soc. Translations,Series 2, No. 48,pp. 229-254,1965. 162. C. E. Shannon. A mathematical theory of communication. Bell Syst. Tech. J., 27:379–423,623–656, 1948. 163. C. E. Shannon. Coding theorems for a discrete source with a fidelity criterion. In IRE National Convention Record,Part 4, pages 142–163, 1959. 164. P. C. Shields. The Theory of Bernoulli Shifts. The University of Chicago Press, Chicago,Ill., 1973. 165. P. C. Shields. The ergodic and entropy theorems revisited. IEEE Trans. Inform. Theory, IT-33:263–266, 1987. 166. P. C. Shields. The interactions between ergodic theory and information theory. IEEE Trans. Inform. Theory, 40:2079–2093, 1998. 167. P. C. Shields and D. L. Neuhoff. Block and sliding-block source coding. IEEE Trans. Inform. Theory, IT-23:211–215, 1977. 168. Ya. G. Sinai. On the concept of entropy of a dynamical system. Dokl. Akad. Nauk. SSSR, 124:768–771, 1959. (In Russian.).
402
References
169. Ya. G. Sinai. Weak isomorphism of transformations with an invariant measure. Soviet Math. Dokl., 3:1725–1729, 1962. 170. Ya. G. Sinai. Introduction to Ergodic Theory. Mathematical Notes,Princeton University Press, Princeton, 1976. 171. D. Slepian. A class of binary signaling alphabets. Bell Syst. Tech. J., 35:203–234, 1956. 172. D. Slepian, editor. Key Papers in the Development of Information Theory. IEEE Press, New York, 1973. 173. M. Smorodinsky. A partition on a bernoulli shift which is not weakly bernoulli. Theory of Computing Systems, 5(3):201–203, 1971. 174. A. D. Sokai. Existence of compatible families of proper regular conditional probabilities. Z. Wahrsch. Verw. Gebiete, 56:537–548, 1981. 175. H. Steinhaus. Sur la division des corp materiels en parties. Bull. Acad. Polon. Sci., IV(C1. III):801–804, 1956. 176. L. C. Stewart. Trellis data compression. Stanford university information systems lab technical report l905–1, Stanford University, July 1981. Stanford University Ph.D. thesis. 177. L. C. Stewart, R. M. Gray, and Y. Linde. The design of trellis waveform coders. IEEE Trans. Comm., COM-30:702–710, April 1982. 178. D.S. Taubman and M.W. Marcellin. JPEG2000: Image Compression Fundamentals, Standards and Practice. Kluwer, Norwell, Massachusetts, 2004. 179. A. H. Tchen. Inequalities for distributions with given marginals. The Annals of Probability, 8(4):814–827, Aug 1980. 180. I. Vajda. A synchronization method for totally ergodic channels. In Transactions of the Fourth Prague Conf. on Information Theory,Decision Functions,and Random Processes, pages 611–625, Prague, 1965. 181. S. S. Vallender. Computing the wasserstein distance between probability distributions on the line. Theory Probab. Appl., 18:824–827, 1973. 182. S. R. S. Varadhan. Large Deviations and Applications. Society for Industrial and Applied Mathematics, Philadelphia, 1984. 183. L. N. Vasershtein. Markov processes on countable product space describing large systems of automata. Problemy Peredachi Informatsii, 5:64–73, 1969. 184. S. Verdu and S. McLaughlin. Information Theory: 50 Years of Discovery. IEEE Press, Piscataway, New Jerseyu, 2000. 185. A.M. Vershik. The kantorovich metric: the initial history and little-known applications. Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov, (POMI) 312:69–85, 2004. 186. C. Villani. Topics in Optimal Transportation, volume 58 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2003. 187. C. Villani. Optimal Transport, Old and New, volume 338 of Grundlehren der Mathemtischen Wissenschaften. Springer, 2009. 188. A. J. Viterbi and J. K. Omura. Trellis encoding of memoryless discrete-time sources with a fidelity criterion. IEEE Trans. Inform. Theory, IT-20:325–332, May 1974. 189. A. J. Viterbi and J. K. Omura. Principles of Digital Communication and Coding. McGraw-Hill, New York, 1979. 190. J. von Neumann. Zur operatorenmethode in der klassischen mechanik. Ann. of Math., 33:587–642, 1932. 191. P. Walters. Ergodic Theory-Introductory Lectures. Lecture Notes in Mathematics No. 458. Springer-Verlag, New York, 1975. 192. E. J. Weldon, Jr. and W. W. Peterson. Error Correcting Codes. MIT Press, Cambridge, Mass., 1971. Second Ed. 193. S. G. Wilson and D. W. Lytle. Trellis coding of continuous-amplitude memoryless sources. IEEE Trans. Inform. Theory, 23:404–409, May 1977.
References
403
194. K. Winkelbauer. Communication channels with finite past history. Transactions of the Second Prague Conf. on Information Theory,Decision Functions,and Random Processes, pages 685–831, 1960. 195. J. Wolfowitz. Strong converse of the coding theorem for the general discrete finite-memory channel. Inform. and Control, 3:89–93, 1960. 196. J. Wolfowitz. Coding Theorems of Information Theory. Springer-Verlag, New York, 1978. Third edition. 197. A. Wyner. A definition of conditional mutual information for arbitrary ensembles. Inform. and Control, pages 51–59, 1978. 198. J. Ziv. Distortion-rate theory for individual sequences. IEEE Trans. Inform. Theory, IT-26:137–143, Mar. 1980. 199. J. Ziv. Universal quantization. IEEE Trans. Inform. Theory, IT-31:344–347, 1985.
Index
A-valued random variable, 4 ∆(µ, ν), 124 ∆(f , g), 125 ρ-distortion, 129 d-continuous, 142 k-step Markov source, 81 kth order Markov source, 81 AEP, xxi, 115 Algoet, P, xxvi almost lossless, 115 almost noiseless, 115 alphabet, 1 alphabets, standard, viii alternating optimization, 258, 343 AMS, 27 AMS, see asymptotically mean stationary, 16 asymptotic equipartion property (AEP), 115 asymptotic equipartition property (AEP), 161 asymptotically dominate, 16 asymptotically mean stationary, 16 asymptotically optimal (a.o.)source code, 337 atom, 12, 35 average Hamming distance, 119 B-process, 36 B-processes, xxiii backward test channel, 262 Berger, T., xxii Berlekamp, E., xxii Bernoulli processes, 36 binary entropy function, 72 Birkhoff, G., xviii
block code, synchronized, 323 block codes, xxii block codes, asynchronous, 323 block codes, synchronized, 382 block decoder, 339 block encoder, 339 block independent, 37 block length, 299 block quantizer, 300 Borel field, 3 Borel-Cantelli lemma, 106, 108, 203 branch, 347 Breiman, L., xxi capacity, 369 capacity, finite-dimensional, 371 capacity, information quantile, 373 capacity, information rate, 369 causal, 34 Cesàro mean, 90 chain rule, 89, 190, 208, 269 channel, 21, 24 channel code, 363 channel coding, 297 channel noiseless, 297 channel, d-continuous, 362 channel, d-continuous, 142 channel, additive noise, 49 channel, AMS, 27 channel, asymptotically mean stationary, 27 channel, block independent, 45 channel, block memoryless, 45 channel, CABI, 146 channel, CBI, 46, 146 channel, CBI approximation, 146 channel, completely random, 29, 30
405
406 channel, conditionally almost block independent, 146 channel, conditionally block independent, 46 channel, deterministic, 30 channel, ergodic, 28 channel, finite output memory, 43 channel, finite-state, 50 channel, induced, 366 channel, Markov, 50 channel, memoryless, 42 channel, memoryless nonanticipatory, 368 channel, noiseless, 29, 298 channel, noisy, 361 channel, output mixing, 43 channel, output weakly mixing, 43 channel, primitive, 48 channel, product, 30 channel, RBCI, 48 channel, SBI, 47 channel, weakly continuous, 362 ¯ channels, d-continuous, viii channels, cascade of, 21 code, 21 code, finite-state , 50 code, joint source and channel, 379 code, sliding-block, 31 codebook, 35, 299, 340 codebook, prefixed, 384 codes block, xxii sliding-block, xxii codes, block, 298 codes, joint source and channel, 297 coding, 61 coding theorem, noiseless, xix coding theorems, viii, xviii, 297 coding theorems, converse, xviii coding theorems, negative, xviii coding, channel, 297 coding, invariant, 62 coding, shift-invariant, 62 coding, source, 297 coding, stationary, 62 coding, time-invariant, 62 column, 58 communication system, 21, 52, 366 conditional density, 189 conditional entropy, 83 conditional entropy density, 189 conditional relative entropy, 189, 200 constraint length, 34 coupling, 53
Index Cover, T., xxvi cross entropy, xxi, 68 Csiszár, I., xxii cumulative generating function, 77 data compression, 297 data processing theorem, 234 Davisson, L. D., xxiv de Morgan’s laws, 1 decoder, 21 decoder, source, 298 delay, 34 density, conditional, 189 directed divergence, 68 discrimination, xxi, 68 discrimination information, 68 distance distribution, 133 variation, 132 distance, Hamming, 375 distortion, 117 distortion measure, 117, 174 distortion measure, additive, 298 distortion measure, subadditive, 321 distortion, ρ, 129 distortion-rate function, 239 distortion-rate function, Nth order, 240 distribution, 4 distribution distance, 133 distribution, conditional, 5 distributional distance, 18, 79 distributions, 7 divergence, 67 divergence inequality, 65, 68 Dobrushin’s theorem, 175, 205 Dobrushin, R. L., vii, 173 Dobrushin, R.L., xxi Dorbrushin, R. L., xxv DRF, 124, 239, 337 operational, 315 Shannon, 315 DRF, operational, 300, 301 DRF, process, 244 DRF, sliding-block codes, 326 dynamical system, 6 dynamical systems, 1 Elias, P., xx encoder, 21 encoder, source, 298 entropy, 61, 62, 64 cross, xxi entropy density, 176, 185, 267 entropy rate, 64, 78
Index entropy rate, ergodic decomposition of, 81 entropy, nth order, 64 entropy, relative, xxi entropy, uniform integrability, 75 entropy-typical sequences, 115 ergodic, 11 ergodic theorem, subadditive, 120 ergodic theory, xvii, xviii, 61 ergodicity, 27 error control coding, 297 event space, 1 event, invariant, 16 expectation, 13 expectation, conditional, 15 Fatou’s lemma, 301, 306, 318 Feinstein code, 367 Feinstein’s lemma, 362 Feinstein’s theorem, 367 Feinstein, A., xxi fidelity criterion, 120 fidelity criterion, additive, 120 fidelity criterion, context free, 121 fidelity criterion, convergent, 122 fidelity criterion, single-letter, 120 fidelity criterion, subadditive, 120, 321 finite anticipation, 42 finite input memory, 42 finite-gap information property, 276 Fubini’s theorem, 316 gadget, 56 Gallager, R. G., xxii Gelfand, I. M., vii, 173 Gelfand, K.Y., xxi good sets principle, 31 Halmos, P. R., 6 Hamming distance, 118 Hartley, R. V. L, xix hookup, 24 Hoph, E., xviii information densities, 219 information density, 110, 363 information divergence, xxi information property, K-gap, 231 information rate, 219, 221 information rate, Pinsker’s, 223 information source, 1 information theory, 61 informational divergence, 68 input, 21
407 input alphabet, 24 input memoryless, 42 input nonanticipatory, 42 input/output process, 24 integrable, 14 integral, 14 isomorphic, 65 dynamical systems, 158 measurable spaces, 157 probability spaces, 157 isomorphic, metrically, 158 isomorphism, 157 isomorphism mod 0, 158 isomorphism theorem, xxiii Jensen’s inequality, 66 join, 64 joining, 53 Körner, J., xxii Khinchine, A. J., xxi Kieffer, J. C, xxv KL number, 68 Kolmogorov extension theorem, 12 Kolmogorov model, 10 Kolmogorov’s formula, 89, 208, 210 Kolmogorov, A. N., vii, xxiii, 173 Kolmogorov, Y.G., xxi Kolmogorov-Sinai invariant, xxiii, 65 Kolmogorov-Sinai-Ornstein isomorphism theorem, xxiii Kullback, S., xxi Kullback-Leibler number, xxi, 68 label function, 55 Linder, T., 87 log-sum inequality, 65 marginal processes, 53 Markov approximation, 91, 199, 276 Markov chain, 89, 196, 206 Markov inequality, 108, 143 Markov source, 81 McMillan, B., xxi, 368 mean entropy, 64 measurable function, 3 measurable space, 1 memory, 34 memoryless, 34, 42 Meshalkin, L. D., xxiii metrically isomorphic, 158 Minkowski’s inequality, 121 mismatch, 309 mutual information, 84
408 mutual information density, 207 mutual information, conditional, 88 Neuhoff, D. L., xxiv nonanticipatory, 42 one-sided, 5 operational distortion-rate function, 298 block coding, 301 operational distortion-rate function (DRF), 124, 337 operational distortion-rate function, block codes, 300 optimal code, 341 optimal source code, 337 Ornstein isomorphism theorem, xxiii Ornstein, D. S., xxiii output alphabet, 24, 35 output memoryless, 42 pair process, 21 partition, 35 partition distance, 125 pdf, 179 Pinsker, M. S., vii, xxiii, 173 Pinsker, M.S., vii Polish space, 119, 313 prefixed codebook, 384 probability density function, 179 probability space, 2 probability, conditional, 4 probability, regular conditional, 5 process, directly given, 10 process, IID, 36 processes, Bernoulli, 36 processes, equivalent, 9 punctuation sequence, 38 quantile, 373 quantizer, 14, 300, 340 Radon-Nikodym derivative, 175, 176 random blocking process, 38 random coding, xix random process, 5 random variable, 3 random variable, invariant, 16 rate, 239 rate, channel code, 366 rate, source coding, 298 rectangle, 8 reference letter, 132, 301 reference measure, 67
Index refine, 70 relative entropy, xxi, 67, 268 relative entropy density, 93, 176 relative entropy rate, 67, 268 relative entropy rate, ergodic decomposition of, 82 relative entropy, nth order, 67 relative entropy, conditional, 86, 189 reproduction alphabet, 239 reproduction codebook, 339 resolution, 298 robust codes, 374 Rochlin, V. A., xxiii Rohlin theorem, 160 Rohlin-Kakutani theorem, 54 Rohlin-Kakutani tower, 54 sample space, 1 sandwich, xxvi section, 15, 24 sequence coder, 30 Shannon optimal reproduction distribution, 262 Shannon, C. E., xvii, xxii Shannon, Claude, 61 Shannon-McMillan theorem, 220 Shannon-McMillan-Breiman theorem, xxvi Shields, P. C., xxiv shift, xviii shift transformation, 7, 10 sigma field, 1 Sinai’s Theorem, 160 Sinai, J. G., xxiii Sinai, Ya. G., 6 Slepian, D., xxii sliding-block code, xxii, 31, 62 sliding-block code, finite window, 34 sliding-block code, finite-length, 34 source, 1, 5, 21 source coding, 297 source, AMS, 314 sources, AMS, 298 standard alphabet, 53 standard space, 12 stationarity, asymptotic mean, 27 stationary, xviii, 11, 16, 25 stationary code, xxii stationary coding, 221 stationary mean, 16 stationary transitions, 81 sub-sigma-field, 4 subadditive sequence, 79 sync locator, 387, 389
Index
409
sync word, 323 synchronization word, 323
two-sided, 5 typical sequences, 115
tail σ -field, 16 test channel, 240, 254 theorem:processmetric, 137 time, 22, 23 time shift, 22 tower, 54 transformation, invariant, xvii transformation, invertible, 11 transport, 131 transportation, 131 trellis, 347 trellis encoding, 335
uppersemicontinuous, 313 variation, 132 variation distance, 132, 184 vector quantizer, 300, 340 von Neumann, J., xviii window length, 34 Wolfowitz, J., xxi Yaglom, A. M., vii, 173