Data-Rate-Constrained State

Estimation and Control of

Complex Networked Systems

This book presents research developments and novel methodologies on data-

rate-constrained control and state estimation for complex networked systems
with different kinds of encoding-decoding mechanisms. It describes frame-
work of state estimator and controller design, stability and performance
analysis for data-rate constrained complex systems with various kinds of
encoding-decoding schemes and so forth. Simulations given in this book are
constructed by applying MATLAB® software package.
Features:

• Gives a systematic investigation of the control and state estimation

for complex networked systems subject to the data rate constraint.
• Develops control/filtering algorithms in a unified framework.
• Includes comparisons for different coding-decoding techniques
proposed.
• Discusses theoretical value and practical application for the resource-
constrained communication environment.
• Provides performance analysis as well as the parameterizations of
filters and FD units.

This book is aimed at researchers and graduate students in electrical

engineering, signal processing, control systems and complex networks.
Data-Rate-Constrained
State Estimation and
Control of Complex
Networked Systems

Licheng Wang, Zidong Wang,

and Guoliang Wei
Designed cover image: Shutterstock
MATLAB® and Simulink® are trademarks of The MathWorks, Inc. and are used with
permission. The MathWorks does not warrant the accuracy of the text or exercises in this book.
This book’s use or discussion of MATLAB® or Simulink® software or related products does not
constitute endorsement or sponsorship by The MathWorks of a particular pedagogical approach
or particular use of the MATLAB® and Simulink® software.
First edition published 2025
by CRC Press
2385 NW Executive Center Drive, Suite 320, Boca Raton FL 33431
and by CRC Press
4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
CRC Press is an imprint of Taylor & Francis Group, LLC
© 2025 Licheng Wang, Zidong Wang, and Guoliang Wei
Reasonable efforts have been made to publish reliable data and information, but the author
and publisher cannot assume responsibility for the validity of all materials or the consequences
of their use. The authors and publishers have attempted to trace the copyright holders of all
material reproduced in this publication and apologize to copyright holders if permission to
publish in this form has not been obtained. If any copyright material has not been acknowledged
please write and let us know so we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted,
reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means,
now known or hereafter invented, including photocopying, microfilming, and recording, or in
any information storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, access
www.copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood
Drive, Danvers, MA 01923, 978-750-8400. For works that are not available on CCC please contact
[email protected]
Trademark notice: Product or corporate names may be trademarks or registered trademarks and
are used only for identification and explanation without intent to infringe.
ISBN: 978-1-032-64542-1 (hbk)
ISBN: 978-1-032-87862-1 (pbk)
ISBN: 978-1-003-53485-3 (ebk)
DOI: 10.1201/9781003534853
Typeset in Palatino
by SPi Technologies India Pvt Ltd (Straive)
To my supervisor and my family.
Contents

List of Figures...............................................................................................................................x
List of Tables.............................................................................................................................xiii
Preface..........................................................................................................................................xiv
Foreword....................................................................................................................................xvi
Acknowledgment..................................................................................................................xvii
Contributors...........................................................................................................................xviii
Symbols.......................................................................................................................................xix

1. Introduction...........................................................................................................................1
1.1 Theoretical Outlines................................................................................................2
1.2 Main Encoding-Decoding Schemes in NSs ................................................5
1.3 Encoding-Decoding-Based Control and Filtering Problems........... 13
1.4 Distributed Control and State Estimation Problems
with Limited Communication Capacity.................................................... 21
1.5 Outline........................................................................................................................ 25

2. Gain-Scheduled State Estimation for Discrete-Time Complex

Networks under Bit-Rate Constraints .................................................................29
2.1 Problem Formulation.......................................................................................... 30
2.1.1 System Model.......................................................................................... 30
2.1.2 Measurement Model Subject to Bit-Rate Constraint........... 31
2.1.3 Bit-Rate-Constrained Encoding-Decoding Scheme............. 31
2.2 Main Results............................................................................................................ 35
2.3 An Illustrative Example..................................................................................... 41
2.4 Summary................................................................................................................... 45

3. Partial-Neurons-Based State Estimation for Artificial Neural

Networks under Constrained Bit Rate: The Finite-Time Case................46
3.1 Problem Formulation.......................................................................................... 47
3.1.1 Encoding-Decoding Mechanism under Bit
Rate Constraint....................................................................................... 49
3.1.2 Estimator Structure .............................................................................. 51
3.2 Main Results............................................................................................................ 52
3.3 An Illustrative Example..................................................................................... 59
3.4 Summary................................................................................................................... 64

vii
viii Contents

4. Synchronization Control for a Class of Discrete-Time

Dynamical Networks with Packet Dropouts:
A Coding-Decoding-Based Approach...................................................................65
4.1 Problem Formulation.......................................................................................... 66
4.2 Main Results............................................................................................................ 70
4.2.1 Preliminaries............................................................................................ 70
4.2.2 State Estimator Design ....................................................................... 72
4.2.3 Detectability Analysis......................................................................... 74
4.2.4 Synchronization Control ................................................................... 80
4.3 An Illustrative Example..................................................................................... 83
4.4 Summary................................................................................................................... 87

5. Observer-Based Consensus Control for Discrete-Time Multi-Agent

Systems with Coding-Decoding Communication Protocol......................88
5.1 Problem Formulation.......................................................................................... 89
5.1.1 Traditional Observer Structure....................................................... 90
5.1.2 CDCP-Based Observer Structure .................................................. 90
5.2 Design of CDCP and Performance Analysis of MASs....................... 96
5.2.1 Coding-Decoding Communication Protocol........................... 96
5.2.2 Consensus Analysis ...........................................................................101
5.3 An Illustrative Example...................................................................................107
5.4 Summary.................................................................................................................109

6. Recursive Filtering with Measurement Fading: A Multiple

Description Coding Scheme....................................................................................110
6.1 Problem Formulation.........................................................................................111
6.1.1 Channel Fading.................................................................................... 112
6.1.2 Multiple Description Coding Scheme ...................................... 113
6.1.3 The Filter Structure............................................................................. 115
6.2 Design of the MDC and Analysis of the Decoding Error............... 116
6.2.1 Design of the MDC............................................................................. 116
6.2.2 Analysis of the Decoding Error.................................................... 118
6.2.3 A Unified Measurement Model ...................................................121
6.3 The Recursive Filtering Strategy.................................................................122
6.4 An Illustrative Example...................................................................................131
6.5 Summary.................................................................................................................140

7. Stabilization of Linear Discrete-Time Systems over

Resource-Constrained Networks under Dynamical Multiple
Description Coding Scheme....................................................................................141
7.1 Problem Formulation........................................................................................142
7.2 Main Results..........................................................................................................145
7.2.1 Decoding Scheme................................................................................148
7.2.2 Design of the Encoder-Decoder Pair.........................................149
Contents ix

7.3 Illustrative Examples ........................................................................................161

7.4 Summary.................................................................................................................164

8. An Event-Triggered Encoding Approach to Control of Linear

Systems under Bit Rate Conditions.....................................................................166
8.1 Problem Formulation........................................................................................168
8.2 Circumventing the “Zeno” Phenomenon...............................................170
8.3 Convergence Analysis of x̃a (t)......................................................................174
8.4 A Sufficient Condition of the Bit Rate to Guarantee the
Convergence of the Decoding Error..........................................................174
8.5 A Necessary Condition of the Bit Rate.....................................................175
8.6 Analysis and Synthesis of the Addressed Event-Based
Control Issue .........................................................................................................176
8.7 Illustrative Examples ........................................................................................180
8.8 Summary.................................................................................................................184

9. Event-Based State Estimation under Constrained Bit Rate:

An Encoding-Decoding Approach.......................................................................185
9.1 Problem Formulation........................................................................................186
9.1.1 System Model........................................................................................186
9.1.2 Event-Based Encoder-Decoder Design ....................................187
9.1.3 State Estimator Structure.................................................................190
9.2 Main Results..........................................................................................................191
9.2.1 Elimination of the Zeno Phenomenon......................................191
9.2.2 The Boundedness Analysis of the Decoding Error............196
9.2.3 The Design of the State Estimator...............................................201
9.3 An Illustrative Example...................................................................................204
9.4 Summary.................................................................................................................210

10. Conclusions and Future Topics..............................................................................211

Bibliography...........................................................................................................................213

Index............................................................................................................................................226
List of Figures

Figure 1.1 Structure of the general communication system.............................3

Figure 1.2 Framework of data-rate-constrained control and state
estimation issues. .............................................................................................5
Figure 1.3 The summarization of several typical encoding-
decoding schemes in NCSs.........................................................................6
Figure 1.4 Multiple description encoding-based data
communication in NSs. ...............................................................................11
Figure 1.5 NCSs with the encoding-decoding-based
communication protocol. .......................................................................... 14
Figure 1.6 Filtering system model with encoding-decoding-
based data transmission scheme........................................................... 16
Figure 1.7 The architecture of distributed systems with encoding-
decoding communication mechanism. .............................................. 22
Figure 2.1 State trajectory ui1 (t) (i = 1, 2, 3) and its estimation.................... 42
Figure 2.2 State trajectory ui2 (t) ( = 1, 2, 3) and its estimation..................... 43
Figure 2.3 Measurement output si (t) (i = 1, 2, 3) and its
decoded value................................................................................................. 43
Figure 2.4 MSE of node i (i = 1, 2, 3). ........................................................................ 44
Figure 2.5 Decoding error s̃i (t) (i = 1, 2, 3). ............................................................ 44
Figure 2.6 The time-varying probability p(t). ....................................................... 45
Figure 3.1 State trajectory u1 (t) and its estimation............................................. 60
Figure 3.2 State trajectory u2 (t) and its estimation............................................. 61
Figure 3.3 State trajectory u3 (t) and its estimation............................................. 61
Figure 3.4 Estimation error of neuron i (i = 1, 2, 3)............................................ 62
Figure 3.5 The measurement output s1 (t) of the first neuron and
its decoded value. ......................................................................................... 62
Figure 3.6 The measurement output s2 (t) of the second neuron
and its decoded value................................................................................. 63
Figure 3.7 Decoding error s̃i (t) (i = 1, 2).................................................................. 63
Figure 4.1 State trajectories xi1 (k) and their decoded values x̂i1 (k)
(i = 1, 2, 3). ........................................................................................................ 84
Figure 4.2 State trajectories xi2 (k) and their decoded values x̂i2 (k)
(i = 1, 2, 3). ........................................................................................................ 85
Figure 4.3 Decoding errors wij (k) (i = 1, 2, 3; j = 1, 2)....................................... 85
Figure 4.4 State trajectories xij (k) of the uncontrolled nodes
i (i = 1, 2, 3; j = 1, 2). .................................................................................... 86
Figure 4.5 State trajectories xij (k) of the controlled nodes
i (i = 1, 2, 3; j = 1, 2). .................................................................................... 87

x
List of Figures xi

Figure 5.1 State trajectories xij (k) (i = 1, 2 . . . , 5; j = 1, 2) with

control input..................................................................................................108
Figure 5.2 State trajectories xij (k) (i = 1, 2 . . . , 5; j = 1, 2) without
control input..................................................................................................108
Figure 5.3 The decoding errors ξEij (k) (i = 1, 2 . . . , 5; j = 1, 2). .....................109
Figure 6.1 Structure of two-description coding-decoding for NSs
with unreliable channels. ........................................................................ 113
Figure 6.2 Nested index assignment for d = 8 and p = 1 in [156]............ 117
Figure 6.3 The diagram of three-tank system.....................................................131
Figure 6.4 The fading measurement components yik (i = 1, 2) and
their decoded values ȳik ...........................................................................135
Figure 6.5 The actual state components xik (i = 1, 2, 3) and their
estimates x̂ik . ..................................................................................................135
Figure 6.6 The FE of x1k and its upper bound. ...................................................136
Figure 6.7 The FE of x2k and its upper bound. ...................................................136
Figure 6.8 The FE of x3k and its upper bound. ...................................................137
Figure 6.9 The upper bounds of filtering errors of xik (i = 1, 2, 3)
with single description and two descriptions. .............................137
Figure 6.10 The effect of fading channel parameter M on filtering
performance...................................................................................................138
Figure 6.11 The filtering errors of xik (i = 1, 2, 3) and their upper
bounds with different noise intensities............................................138
Figure 6.12 The filtering errors of xik (i = 1, 2, 3) and their upper
bounds with different packet-arrival probabilities....................139
Figure 6.13 The effect of the model uncertainty on the filtering
performance...................................................................................................139
Figure 7.1 The structure of the closed-loop control system
with the MDC scheme..............................................................................145
Figure 7.2 Nested index assignment for l = 8, d = 1.......................................147
Figure 7.3 State trajectories of the open-loop system. ....................................162
Figure 7.4 State trajectories of the closed-loop system...................................163
Figure 7.5 State trajectories and their decoded values...................................163
Figure 7.6 The trajectories of decoding errors....................................................164
Figure 7.7 The Bernoulli sequences γi (τ h) (i = 1, 2)........................................165
Figure 8.1 The schematic of the event-based control with
encoding-decoding scheme. ..................................................................168
Figure 8.2 Dynamical evolutions of auxiliary system
and decoder....................................................................................................181
Figure 8.3 The decoding errors between the actual system states
and their decoded values........................................................................182
Figure 8.4 State response without control signal u(t).....................................182
Figure 8.5 State response with control signal u(t). ...........................................183
Figure 8.6 The threshold of the event generation function and the
norm of x̃a (t). .................................................................................................183
xii List of Figures

Figure 9.1 Schematic of state estimation problem under

constrained bit rate.....................................................................................191
Figure 9.2 Measurement y(t) and its decoded value yd (t)............................205
Figure 9.3 The norm of the decoding error ỹd (t)...............................................206
Figure 9.4 The state trajectories xi (t) and their estimates
x̂i (t) (i = 1, 2)..................................................................................................206
Figure 9.5 The estimation errors x̃i (t) (i = 1, 2)..................................................207
Figure 9.6 Triggering threshold σ (t) and the norm of the
decoding error kỹd (t)k. .............................................................................207
Figure 9.7 The state trajectories xi (t) (i = 1, 2, 3) and their
estimates..........................................................................................................209
Figure 9.8 The estimation errors x̃i (t) (i = 1, 2, 3). ............................................209
List of Tables

Table 6.1 System Parameters ........................................................................................133

Table 9.1 Bit Rate R0 Under Different Threshold Values ̺ ...........................210
Table 9.2 Bit Rate R0 Under Different Disturbance Intensities...................210

xiii
Preface

With ever-rapid developments in sensing, computing, communication

and integration technologies, the modern industrial systems possess high
scalability and integrated architecture. In particular, compared with
traditional industrial control systems where information interactions are
implemented by a wire-based point-to-point way, modern networked control
systems upgrade in such a data exchange manner to the shared-network-
based one (e.g. field bus, IP/Ethernet, Bluetooth), which greatly enhances
the reliability, efficiency and flexibility of the systems. On the other hand,
the introduction of the communication network in industrial systems also
brings in some issues, a representative one of which is the data rate-
constraint. In practical applications, most of the systems suffer from limited
date rate of communication network due either to physical constraint or
to cost consideration, which gives rise to various undesirable network-
induced phenomena. The data encoding-decoding scheme is one of the
efficient approaches to deal with the scarcity issue of the communication
resource and thus might avoid the performance deteriorations of data
communication. The existing control/state estimation methods are designed
without consideration of the date rate issue and this necessitates re-designing
or developing new methodologies to accommodate the imbedding of
different encoding-decoding mechanisms.
The objective of this book is to present the state-of-the-art research
developments and novel methodologies on data-rate-constrained control
and state estimation for complex networked systems with different kinds of
encoding-decoding mechanisms. The contents of this book can be divided
into two parts, where the first part (Chapters 2–5) addresses the data-
constrained control/state estimation problems under uniform-quantization-
based encoding-decoding scheme for complicated systems (e.g. neural
networks, complex networks and multi-agent systems) and the second part
(Chapters 6–9) introduces both the multiple description encoding mecha-
nism and the event-based encoding protocol into the networked systems to
cope with the data constraint issue. The work provides a framework of state
estimator and controller design, stability and performance analysis for data-
rate constrained complex systems with various kinds of encoding-decoding
schemes, including the static/dynamic uniform-quantization-based scheme,
the multiple description coding scheme, the sign-based coding scheme,
etc. Some related techniques and theories are applied, which include
the recursive Riccati-like equations, hybrid system theory, matrix theory,
and mathematical optimization methods, to fulfill the specific analysis
and synthesis issues. In addition, this book provides up-to-date reference

xiv
Preface xv

materials for researchers who wish to explore the control and state estimation
problems with constrained data rates.
The concise frame and description of the book are given as follows.
Chapter 1 introduces the recent advances on data-rate-constrained control
and state estimation problems for networked systems and the outline
of the book. Chapter 2 investigates the state estimation for discrete-time
complex networks subject to date-rate constraints. Chapter 3 is concerned
with the finite-time state estimation problem for artificial neural networks
under constrained bit rate, where only partial neurons measurement
outputs are available for the state estimation task. Chapter 4 considers the
synchronization control for a class of discrete-time dynamical networks
with a uniform-based encoding-decoding method. Chapter 5 studies the
observer-based consensus control for discrete-time multi-agent systems
with coding-decoding communication protocol. Chapter 6 deals with the
recursive filtering problem for time-varying systems where a static multiple
description coding scheme is introduced to facilitate the data transmission.
Chapter 7 addresses the stabilization issue of linear discrete-time systems
over resource-constrained networks under dynamical multiple description
coding scheme. In Chapter 8, an event-triggered encoding approach is
employed to the control issue of linear systems under bit rate conditions.
Chapter 9 discusses a sign-based coding approach to the state estimation
problem under constrained bit rate. Chapter 10 gives the conclusion and
some possible future research directions. Simulations given in this book are
constructed by applying The MathWorks MATLAB software package.
This book is a research monograph whose intended audience is graduate
and postgraduate students as well as researchers.
Licheng Wang
Shanghai, China
Foreword

I am delighted to introduce the book Data-Rate-Constrained State Estimation

and Control of Complex Networked Systems. When I came to know about this
book project undertaken by three of the most active researchers in the field,
I was pleased that this book is coming in the early stage of a field that will
need it more than most fields do. In most emerging research fields, a book can
play a significant role in bringing some maturity to the field. Research fields
advance through research papers. In research papers, however, only a limited
perspective could be provided about the field, its application potential, and
the techniques required and already developed in the field. A book gives such
a chance. I liked the idea that there will be a book that will try to unify the
field by bringing in disparate topics already available in several papers that
are not easy to find and understand. I was supportive of this book project
even before I had seen any material on it. The project was a brilliant and bold
idea by three active researchers. Now that I have it on my screen, it appears
to be even a better idea.
Data-rate-constrained state estimation and control started gaining recog-
nition in the last century as a field. Signal processing, tracking, output
regulation, communication and optimization technologies had advanced
enough that researchers and technologists started developing different data
encoding approaches such that the resulting control and state estimation
performance is less influenced by the distortion of the transmitted data.
By properly selecting encoding-decoding criterion to restrain the signal
distortion, the data encoding-decoding approaches aspire to create a holistic
picture for complex networked systems subject to limited data rate.
Since the data distortion caused by the finite bit rate of the communication
media has a major impact on the overall system performance, how to well
attenuate or even eliminate the negative effects is of significant importance.
In addition, taking into account various encoding-decoding strategies,
different encoding-decoding approaches result in different degrees of data
distortion. Thus, the necessity (of researching state estimation and control
theories/methodologies with various encoding-decoding strategies under
a data-rate-constrained communication manner) arises, paving a way for
people to hypothesize relationships among data sources and explore support
for system monitoring/tracking.
Once successfully developed, such theories/methodologies will find vast
applications in many diverse domains and keep getting more applications.
In fact, many new filtering/control approaches are direct outgrowth of state
estimation and control subject to constrained bit rate and it is likely to become
a powerful system estimation/control tool.

xvi
Acknowledgment

The authors would like to express their deep appreciation to those who
have been directly involved in various aspects of the research leading to
this book. Special thanks go to Professor Bo Shen from Donghua University,
Shanghai, China, Professor Fuad E. Alsaadi from King Abdulaziz University,
Jeddah, Saudi Arabia, Professor Xiaohui Liu from Brunel University London,
London, UK, Professor Derui Ding from University of Shanghai for Science
and Technology, Shanghai, China, Professor Tingwen Huang from Texas
A&M University at Qatar, Doha, Qatar, Professor Hongli Dong from
Northeast Petroleum University, Daqing, China, and Professor Yurong Liu
from Yangzhou University, Yangzhou, China.
The writing of this book was supported in part by the National
Natural Science Foundation of China under Grants 62473247, 62473261,
62003213, 61903253, 61873148, 61933007, 61873169, 61873059 and 61973219,
the China Postdoctoral Science Foundation under Grants 2019TQ0202 and
2020M671172, the National Postdoctoral Program for Innovative Talents
under Grant BX20180202, the Shanghai Pujiang Program under Grant
19PJ1408100, the Royal Society of the UK and the Alexander von Humboldt
Foundation of Germany.

xvii
Contributors

Bo Shen Derui Ding

Donghua University University of Shanghai for Science
Shanghai, China and Technology
Shanghai, China
Fuad E. Alsaadi
King Abdulaziz University Hongli Dong
Jeddah, Saudi Arabia Northeast Petroleum University
Daqing, China
Xiaohui Liu
Brunel University London Shuai Liu
London, U.K. University of Shanghai for Science
and Technology
Tingwen Huang Shanghai, China
Texas A&M University
at Qatar Yurong Liu
Doha, Qatar Yangzhou University
Yangzhou, China

xviii
Symbols

⊗ The Kronecker product

Rn The n-dimensional Euclidean space

Rn×m The set of all n × m real matrices

R+ The set of all positive real numbers

N The set of natural numbers

n
S+ The set of n × m positive definite matrices

AT or A′ The transpose of matrix A

A† The Moore-Penrose pseudo inverse of A

A > 0 The matrix A is positive definite

A ≥ 0 The matrix A is positive semidefinite

A < 0 The matrix A is negative definite

A ≤ 0 The matrix A is negative semidefinite

k · k The Euclidian norm of real vectors or the spectral norm of real

matrices

k · kmax The maximum singular value of a matrix

ρ(A) The spectral radius of matrix A

δ( · ) The Kronecker delta function

tr(A) The trace of matrix A

kxk2P Equals to xT Px when x is a vector

P{·} The occurrence probability of the event “·”

E{x} The expectation of stochastic variable x

Var{x} The variance of stochastic variable x

E{x|y} The conditional expectation of x given y

I The identity matrix of compatible dimension

xix
xx Symbols

0 The zero matrix of compatible dimension

0n The n × n zero matrix
1n The n × 1 column vector with all elements equal to 1
 T
vec{x1 , x2 } The column vector xT1 x2T

vecn {xi } The column vector vec{xT1 , x2T , · · · xTn }

diag{xi } The block diagonal matrix with ith block being xi and all other
entries being zero
diagn {Ai } The block diagonal matrix diag{A1 , A2 , · · · , An }
{Mij }n×n The partitioned matrix with Mij being (i, j)-th block submatrix
1
Introduction

It has been well recognized that the publication titled A Mathematical

Theory of Communication [142, 143] by C. E. Shannon in 1948 marked
the establishment of the information theory. From then on, constant
research effort from the researchers of different fields has been put
on its developments and considerable progress has been made in the
aspects of theory and application. Coincidentally, in the same year,
another celebrated mathematician N. Wiener published his famous treatise
Cybernetics: Or Control and Communication in the Animal and the Machine
[169], which laid down the theoretical foundation for the servomechanisms
(whether electrical, mechanical or hydraulic), automatic navigation, analog
computing, artificial intelligence, neuroscience and reliable communications,
and made the birth of control theory. Nevertheless, during the past few
decades, the information theory and control theory have almost been
regarded as two independent disciplines with few common features and
connections due to their different concerns.
Broadly speaking, the information theory mainly focuses on how to obtain,
store, process and transmit the data but seldom pays attention to the specific
usage of the data itself. On the contrary, the control theory concerns how to
make use of the obtained data to achieve the prescribed tasks through certain
strategies. For general systems, the most basic but important idea to realize
desirable system performances is the feedback, which means the obtained
signals (usually the system outputs) should be fed back to the systems’ inputs
in certain manners by designing proper feedback mechanisms. However, in
most cases, such an idea is usually implemented upon the assumption that
the signal transmission between the components is with infinite precision,
namely, the signals can be perfectly transmitted and utilized to contribute to
the feedback schemes.
In recent years, owing to the rapid development of the networked
communication technology, networked systems (NSs) have emerged as a
new yet popular system architecture with an extensive utilization in the
industry for its low power cost, simple installation, easy maintenance and
high reliability. Different from the non-networked systems where the data
flows through the components in a point-to-point way, the NSs mean
that all the data communications between the components are executed

DOI: 10.1201/9781003534853-1 1
2 Data-Rate-Constrained State Estimation and Control

via a commonly shared communication network (channel). Although the

NSs possess various merits mentioned above, the introduction of the
communication network also brings about some new and noteworthy
problems including 1) how the network-based data exchange affects the
system performance, and 2) how to maximize the utilization of limited
network bandwidth while ensuring the required system performance.
If the network environment is ideal, that is, the bandwidth of the network
is sufficiently large, the assumption of the perfect data transmission would
make sense and it is reasonable to ignore the influence from the network
communication on the system performance. However, in situations that the
network resources are subject to certain constraints (e.g. limited bandwidth
of the channel), it is no longer proper to entirely neglect the distortion of
the transmitted data caused by the limited communication capacity of the
network. Especially, when it comes to large-scale NSs such as wireless sensor
networks (WSNs), multi-agent systems (MASs) and smart grids, the limited
network resources impose severe restrictions on the data communication
quality and even deteriorate the overall system performances. On the other
hand, from the standpoint of the network security, the NSs are usually
fragile and vulnerable to be attacked, for example, the data flowing through
the system components are easily stolen by the adversaries for some
hostile purposes, see [20, 50, 96, 183]. Therefore, in order to make the data
transmission of the NSs more efficient and secure via a resource-constrained
communication network, the encoding-decoding technique stands out as an
ideal candidate for its advantage in data compression and encryption.

1.1 Theoretical Outlines

Normally, as shown in Figure 1.1, a classic communication system always
contains three basic parts: the information source, the information channel
and the information sink, based on which, the encoding-decoding technique
can be accordingly divided into two main categories: the source encoding-
decoding and the channel encoding-decoding. In this paper, our attention
will be specifically focused on the source encoding-decoding technique,
which aims to transform the original data into certain codewords. Generally
speaking, the source information encoding technique possesses the following
three prominent features: 1) realizing the data compression, 2) facilitating
the digital data transmission and 3) enhancing the data security. From the
aspect of mathematics, the process of encoding can be regarded as a mapping
that maps the original data to certain codewords. Correspondingly, the
decoding process can be viewed as the “inverse” operation which intends to
recover the original data as exactly as possible by excavating the information
Introduction 3

FIGURE 1.1
Structure of the general communication system.

contained in the codewords. To be more specific, in the networked

communication systems, let x(k), F( · ), H, G( · ) and xd (k) be the original
signal before being encoded, the encoding mapping, the encoding alphabet,
the decoding mapping, and the decoded signal, respectively, where k is the
discrete time step and the H , {s1 , s2 , . . . , sm } is a set whose elements si (i = 1,
2, . . . , m) are certain deterministic codewords. In most existing literature,
the authors’ attempts have always been made to design a pair of encoder-
decoder appropriately such that at each time instant k, the decoding error
ed (k) , xd (k) − x(k) is possibly small at the minimum cost of the network
bandwidth.
In view of the aforementioned discussion, it is obvious that there are two
important performance indices to evaluate the designed encoding-decoding
procedure, namely, the decoding error and the bandwidth condition.
Intuitively, the large network bandwidth would result in a small decoding
error. For the NSs, if the network resource is not the concern (e.g. the
bandwidth can be sufficiently large), the decoding error would probably
be very small and hence has little impact on the system performance. In
this case, the coding-decoding-based NSs reduce to the traditional NSs for
the analysis and synthesis issues. However, in most practical situations,
due to the limited network resources, the network bandwidth serving as
a major concern should be incorporated into the system design. In this
sense, for the bandwidth-constrained systems, the concept of “minimum bit
rate” is proposed to characterize the smallest needed channel capacity for
the codeword transmission, above which the performance of the controlled
(or filtering error) systems is explored including asymptotical/exponential
stability, ultimately bounded stability, input-to-state stability (ISS), etc.
In the past few years, a rich body of research results with respect to
the minimum bit-rate-based control/filtering problem for both linear and
4 Data-Rate-Constrained State Estimation and Control

nonlinear NSs have appeared. Among them, it has been shown in [155] and
[115] that for linear systems, the minimum bit rate R (bits/sample) should
satisfy R > log1(2) Re(λi (A))>0 (λi (A)), where log is the base-e logarithm,
P

Re( · ) denotes the real part of a complex-valued scalar, A is the system

matrix, and λi {A} stands for the ith eigenvalue of matrix A. Furthermore, the
literature [123] revealed that in order to globally (or semiglobally) stabilize
the n−dimensional nonlinear systems with a feedforward structure, the
necessary bit rate should be n + 1 (or n).
Compared with the traditional NSs, the introduction of the encoding-
decoding communication mechanism would make the dynamics of the
NSs more complicated and diverse, which accordingly imposes certain
challenges on the analysis and synthesis issues. Moreover, because the
inevitable data distortion resulted from the encoding-decoding process,
the traditional approaches cannot be directly generalized to cope with
the performance analysis issues of data-rate-constrained NSs. In this
case, the influence of the encoding-decoding mechanism on the system
performance should be well taken into consideration. In other words,
how to choose suitable theoretical tools to quantitatively evaluate the
relationship between the system dynamics and the encoding-decoding
communication protocol appears to be particularly important. To date, there
are generally three theoretical methodologies to deal with the encoding-
decoding-based analysis and synthesis problems for the NSs subject to
data-rate constraint, namely, the information-theory-based methodology, the
ISS-based methodology and the hybrid-system-based methodology.
The theoretical basis of the information theory methodology is the
celebrated Shannon information theory, based on which, all information
including the original data, codewords and decoded data are described as
random variables [28, 29, 117]. Within such a framework, the main objective
is to find general necessary conditions for the channel capacity measured by
the concept of “entropy”, under which the original data can be reconstructed
from the decoded data with the acceptable distortion. Then, by resorting
to the decoded data, the desired control or filtering performance can be
guaranteed. On the other hand, for most of the effective encoding schemes,
a basic requirement is that the decoding error ed (k) should be bounded or
even convergent. Thus, it is natural to view the decoded error as the bounded
disturbance of the dynamical system, and the ISS-based approach [87,124]
has been proposed as a theoretical tool to handle the analysis and synthesis
issues for these systems. For the continuous-time systems with a digital
transmission manner, the received codewords by the decoder at a certain
discrete instant can be seen as an impulse sequence that affects the decoder
dynamics. Consequently, the hybrid-system-based approach [86,89] is valid
for analyzing such a kind of systems’ dynamical behaviors.
In this paper, our purpose is to offer an up-to-date review of existing
research results on encoding-decoding-based control and filtering problems
Introduction 5

FIGURE 1.2
Framework of data-rate-constrained control and state estimation issues.

for the NSs. Various usually adopted design schemes of the encoding-
decoding procedure in the NSs and their individual characteristics are
reviewed. Subsequently, recent advances on the encoding-decoding-based
control and filtering problems for the NSs are, respectively, surveyed. Then,
the control and filtering issues for distributed systems with encoding-
decoding communication scheme are discussed. Finally, conclusions and
future research directions are presented. The main contents that are reviewed
in this paper and the framework are shown in Figure 1.2.

1.2 Main Encoding-Decoding Schemes in NSs

In the framework of the coding-decoding-based data communications for
NSs, it is of crucial importance to construct a suitable pair of encoder and
decoder which could greatly restrain the impacts from the resulted data
distortion on the system performance. To this end, we are going to give a
clear look on some commonly utilized encoding-decoding schemes with the
following system model

x(t + 1) = Ax(t) + Bu(t) + Dw(t) (1.1)

or
ẋ(t) = Ax(t) + Bu(t) + Dw(t) (1.2)
6 Data-Rate-Constrained State Estimation and Control

FIGURE 1.3
The summarization of several typical encoding-decoding schemes in NCSs.

with the measurement output

y(t) = Cx(t) + Ev(t) (1.3)

where x(t) ∈ Rnx , u(t) ∈ Rnu , and y(t) ∈ Rny are, respectively, the system state,
the system input, and the measurement output. w(t) ∈ Rnw and v(t) ∈ Rnv
are the process and the measurement noises. A, B, C, D, and E are the
known matrices with compatible dimensions. For analysis convenience, in
this paper, it is assumed that the system state x(t) is accessible for the
stabilization problem, while for the filtering issue, it is supposed that only
the measurement output y(t) is observable. Till now, in the existing literature,
two representative design solutions of the encoding-decoding procedure
have been proposed to deal with the control and filtering problems in
the NSs, namely, the quantization-based encoding-decoding technique and
the symbolic-based encoding-decoding technique. Several typical encoding-
decoding schemes in the NCSs are exhibited in Figure 1.3. In the following,
we are in a position to successively illustrate these encoding-decoding design
approaches mentioned above.

• Quantization-Based Encoding-Decoding Mechanism

The quantization-based encoding-decoding method is one of the
most common methods for the NSs with the limited communication
capacity, based on which, a number of coder-decoder structures have
been presented in the existing literature, see [59,69,79,140,190]. In
addition, according to the case whether the quantizer parameters can
be dynamically adjusted, the quantizer can be classified into the static
quantizer and the dynamic one. In the following, in terms of the type
of the used quantizer, three typical coder-decoder structures are listed
and their corresponding characteristics are summarized.
Introduction 7

Dynamic-Quantization-Based Encoding-Decoding Scheme I

Such a method was proposed in [140] and a similar idea appeared
in the earlier literature [85]. For the sake of making this approach
more understandable, first, let us briefly recall some basic concepts
of the uniform quantization, and then we will show how the uniform
quantization approach is applied to the encoder-decoder procedure.
Let the scaling parameter (the length of the quantization interval)
a > 0 and integer q be given, and the Ba denotes the hyperrectangle
in Rn centered at 0 with the edge length a, that is, Ba = {δ ∈ Rn :
|δ|∞ ≤ a}, where | · |∞ is the infinite vector norm. Then, we uniformly
partition the hyperrectangles Ba into qn sub-hyperrectangles Is11 (a) ×
Is22 (a) × · · · × Isnn (a), where s1 , s2 , . . . , sn ∈ {1, 2, . . . , q} and
 
2a
I1i (a) , δi | − a ≤ δi < −a +
q
 
i 2a 4a
I2 (a) , δi | − a + ≤ δi < −a + (1.4)
q q
..
.
 
2a
Iqi (a) , δi | a − ≤ δi ≤ a
q

where δi is the ith element of the vector δ. Here, each sub-

hyperrectangle Is11 (a) × Is22 (a) × · · · × Isnn (a) is labeled with a sequence
{s1 , s2 , . . . , sn }, which represents the specific region of the sub-
hyperrectangle and will be sent to the decoder for the reconstruction
of the original data. Moreover, the center of the sub-hyperrectangle
Is11 (a) × Is22 (a) × · · · × Isnn (a) can be defined as

ηa (s1 , s2 , . . . , sn ) , [b1 b2 · · · bn ]T (1.5)

where bi = −a + (2si −1)a

q , i = 1, 2, . . . , n.
In what follows, we employ the proposed encoding-decoding
protocol to the continuous-time system (1.2) for the stabilization
problem. To be more specific, our goal is to design a pair of encoder-
decoder and certain decoded-signal-based control strategy such that
the closed-loop system is asymptotically stable. For simplicity of
illustration, we assume that the disturbance w(t) in (1.2) satisfies
w(t) ≡ 0 and the main idea for the design of the encoding-decoding-
based control strategy can be outlined in the following steps. To
start with, at each encoding time jT (j = 1, 2, . . . and “T” is the
encoding period), by calculating the error e(jT) between the system
8 Data-Rate-Constrained State Estimation and Control

state x(jT) and the decoded state x̂(jT) which will be defined later,
where e(jT) = x(jT) − x̂(jT− ), we are able to determine the uniform
quantization parameter “a” and thus partition the hyperrectangle Ba
into qn equal sub-hyperrectangles. Here, x̂(jT− ) denotes the limit of
x̂(jT) at the point of jT from the left, that is, x̂(jT− ) , limǫ>0,ǫ→0 x̂(jT −
ǫ). Subsequently, generate the corresponding codeword sequence
{s1 , s2 , . . . , sn } according to the region where the error “e(jT)” belongs
to and then transmit the codeword sequence {s1 , s2 , . . . , sn } to the
decoder to obtain the decoded state x̂(jT). Finally, design the
controller “u(jT)” in terms of the decoded state x̂(jT) to stabilize the
original system.
Based on the above analysis, the detailed form of the encoder and
decoder is given as follows:
Coder: For e(jT) ∈ Is11 (a(jT)) × Is22 (a(jT)) × · · · × Isnn (a(jT)) ⊂ Ba(jT) , we
have
g(e(jT)) = {s1 , s2 , . . . , sn } (1.6)
where a((j + 1)T) = ϕ(T, q, x(0))a(jT) for j = 1, 2, . . . with ϕ(T, q, x(0))
being the function of certain form with respect to the encoding
period T, the quantization level q and the initial condition x(0).
Decoder:

 x̂(0) = 0



 ẋˆ(t) = Ax̂(t) + Bu(t), ∀t ∈ [jT, (j + 1)T)
(1.7)


 x̂(jT) = x̂(jT− ) + ηa (s1 , s2 , . . . , sn )


u(t) = Kx̂(t).

From structures of the coder (1.6) and the decoder (1.7), we have the
following observations. 1) The proposed encoding approach is a kind
of differential encoding technique. It is obvious that, in (1.6), only
the decode error e(jT) is encoded at the encoding time jT. In fact, by
using the differential quantization encoding technique to encode the
signal difference, the amount of the information to be sent is greatly
reduced, see [27]. 2) It can be seen from the iteration relationship
between a((j+1)T) and a(jT) that the quantization region dynamically
varies at each encoding instant, and such a quantizer is referred to
as the dynamic quantizer. 3) In order to enforce the decoding error
be convergent, it is required that ϕ(T, q, x(0)) < 1, from which, we
can see that the convergence of the encoding-decoding algorithm is
closely associated with the encoding period, the quantization level
and the initial condition of the system.
Apart from the continuous-time case, for the discrete-time
counterpart, the encoding-decoding communication strategy for
Introduction 9

the NSs has been considered in [190], and the corresponding

extended control/filtering problems subject to the network-induced
phenomena have been investigated in [161,173].
Dynamic-Quantization-Based Encoding-Decoding Scheme II
This kind of encoding-decoding scheme is based on the so-called
“zooming in/zooming out” quantization technique which was
proposed in [171]. For the discrete-time system (1.1) with w(t) ≡ 0,
the following encoder is given by


 ξ(0) = 0

 ξ(t) = g(t − 1)s(t) + ξ(t − 1)
  (1.8)
 1
(x(t) − ξ(t − 1))

 s(t) = Q

g(t − 1)

where ξ(t) is the encoder’s internal state, s(t) is the generated

codeword which needs to be sent to the decoder, Q( · ) is a finite-
level uniform quantizer, and g(t) is a scaling function which is used
to adjust the amplitude of the signal to be quantized. In most of the
literature, the quantizer Q( · ) is chosen as the symmetric one with the
following form


 0, 1/2 ≤ z < 1/2

(2i − 1)/2 ≤ z < (2i + 1)/2



 i,
Q(z) = i = 1, 2, . . . , L − 1 (1.9)

z ≥ (2L − 1)/2

 L,




− Q( − z), z ≤ −1/2

where L is the quantizer’s saturation value and the associated

quantization level is 2L + 1.
By using the generated codeword s(t), the decoder 9 is designed as
(
x̂(0) = 0
(1.10)
x̂(t) = g(t − 1)s(t) + x̂(t − 1)

where x̂(t) is the output of the decoder, namely, the decoded value.
It can be concluded from the structure of the above encoder-
decoder that 1) although the saturation value L is fixed a priori,
due to the introduction of the scaling function g(t), the signal to be
quantized is enforced to be pre-processed first by the “zooming out”
or “zooming in” technique, and hence this type of encoding approach
is still of the dynamic-quantization-based encoding, 2) as discussed
above, encoding the difference between x(t) and ξ(t − 1) with scaling
rather than the system state x(t) itself possibly reduces the bits which
10 Data-Rate-Constrained State Estimation and Control

are utilized to represent the codeword, 3) the saturation value L is

closely related to the network bandwidth, that is, a small L results
in the requirement of a small network bandwidth. Especially, when
L = 1, it means that only one bit is necessary to send the codeword,
and 4) the encoding period is consistent with the system sampling
period, which is different from the former encoding approach that the
encoding period can be the multiple of the system sampling period,
see [140,173, 190].
Such a “zooming in/zooming out” quantization-based encoding-
decoding scheme was first proposed in [7], under which, a consensus
algorithm has been designed for a class of discrete-time MASs
that the information exchange among the agents was implemented
through a digital communication network. Recently, some extensions
have emerged, see [66, 70, 78, 79, 179]; however, most of which are
concerned with the discrete-time systems and the results for the
continuous-time counterpart are relatively few.
Static-Quantization-Based Encoding-Decoding Scheme
Having presented the above two different kinds of encoding-
decoding schemes, now we would like to give a simple discussion
about another encoding-decoding scheme, namely, the multiple
description (MD) encoding-decoding scheme. The original idea of the
MD encoding-decoding scheme comes from the investigation of the
audio encoding issue in the telephone network by Bell Laboratories
in the 1970s. As the rapid development of the network-based (e.g.
internet) communication techniques, in recent decades, the MD
encoding-decoding technology has gained considerable research
attention from many researchers in different disciplines.
The idea of the MD encoding-decoding scheme is based on the
diversity principle which could effectively enhance the transmission
capacity and reliability of the channels, see [23,37,59,156,157]. From
the point of the information theory view, the diversity principle
can be understood as follows: encode the information source into
multiple descriptions with the identical importance and then sent
to the decoder via the parallel individual channel. The framework
of the MD encoding-decoding scheme is built on the following two
assumptions: 1) there are multi-channels between the encoder and the
decoder and all the channels failing to work simultaneously is a small
probability event, and 2) the original signal can be reconstructed with
an acceptable quality as long as at least one of the descriptions is
successfully delivered to the decoder, and when more descriptions
are received by the decoder, the signal is reconstructed with a
higher quality. Therefore, such an approach is able to improve the
transmission reliability by rejecting the channel fluctuations resulted
from the fading [47], shadowing, crosstalk, etc.
Introduction 11

So far, according to the adopted method in the encoding-decoding

process, the multiple description encoding-decoding can be classified
into various classes including multiple description quantization
coding (MDQC), the multiple description transformation coding
(MDTC), the forward-error-correction-based multiple description
coding (FECMDC), etc. In this paper, our attention is focused on the
MDQC where the static uniform quantization method is applied to
the addressed encoding-decoding issue.
Loosely speaking, the static uniform quantizer-based multiple
description coding procedure can be divided into two steps. The first
step is the index generation process which mainly employs the scalar
quantizer to transform the information source into certain indices.
Then, represent the indices by multiple descriptions in terms of a
proper index assignment strategy. For convenience, a brief view on
the encoding scheme with two descriptions is provided where the
measurement output (1.3) is treated as the information source to be
encoded. The diagram of the MDC for the NSs with two descriptions
is shown in Figure 1.4.
Encoder:
(
θl (t) = f1l (yl (t)), l = 1, · · · , ny
(1.11)
ϑl (t) = f2l (yl (t)), l = 1, · · · , ny

Decoder:

 g1l (θl (t)),
 when λ1 (t) = 1, λ2 (t) = 0
ȳl (t) = g2l (ϑl (t)), when λ1 (t) = 0, λ2 (t) = 1 (1.12)

when λ1 (t) = 1, λ2 (t) = 1

gcl (θl (t), ϑl (t)),

FIGURE 1.4
Multiple description encoding-based data communication in NSs.
12 Data-Rate-Constrained State Estimation and Control

where f1l ( · ) and f2l ( · ) are two coding functions whose outputs
θl (t) and ϑl (t) can be regarded as the individual descriptions of
the source yl (t) with yl (t) being the lth component of y(t). g1l ( · )
and g2l ( · ) are two side decoding functions and gcl (·, ·) is the
central decoding function. λi (t) (i = 1, 2) are two independent
random variables characterizing the packet dropout phenomena
during the transmissions of the individual description, which
satisfy the Bernoulli binary distribution taking values on either
1 or 0 with mathematical expectations λ̄i and variances σi2 ,
respectively.
The early research effort on the MDC was mainly devoted to
its theoretical research. Since the scalar quantization-based MDC
method was first proposed in [156], the research direction has
been moved from the pure theoretical investigation to its practical
applications. From then on, the MDC has experienced a huge
development in the theory and applications, and myriad coding
schemes have appeared. It should be pointed out that most results
of the MDC have been centered on the applications in distributed
storage systems [60], diversity communication systems (antenna
diversity) [133], and image/audio/video encoding [3,5,13,176].
However, to the best of the authors’ knowledge, the relevant results
of the MDC with respect to the control and filtering problems of NSs
are very few.
• Symbolic-Based Encoding-Decoding Mechanism
This kind of encoding schemes, also called as the “binary encoding”,
mainly takes the sign of the source into account, and thus only one
bit is utilized to encode the signal. For convenience, in the sequel, we
call it as the binary encoding instead of the symbolic-based encoding.
Usually, the binary encoding technique including two versions, that
is, the absolute (static) encoding technique [32,33] and the differential
(dynamic) encoding technique [63, 89]. Since the dynamic binary
encoding technique is more complicated than the static one, in the
following, we will concentrate our attention on the dynamic binary
encoding-decoding technique.
As we know, such a kind of dynamic encoding scheme was first
proposed in [63], which is highly dependent on the change of certain
system parameters. The core of the encoding rule can be expressed as
follows with the system model (1.1) and (1.3).

Coder : s(tk ) = sign(y(tk ) − y(tk−1 ))

(1.13)
Decoder : ȳ(tk ) = g(s(tk ))
Introduction 13

where sign( · ) is the sign function, g( · ) is the decoding function,

s(tk ) is the codeword generated at tk , and ȳ(tk ) is the decoded
value of the measurement output y(tk ). It follows from (1.13) that
at each encoding instant tk , only the sign of the change between
the two successive measurement outputs is transmitted to the
decoder, which implies that only one bit is needed for the codeword
delivery during each encoding period. For the decoder design,
the expected decoding function g( · ) generally consists of two
parts, namely, g(s(tk )) = ȳ(tk−1 ) + g0 (s(tk )). The first term is the
decoded value at the last encoding instant and the second one is a
compensation function which aims to predict the difference between
y(tk ) and y(tk−1 ). It is obvious that such an encoding scheme greatly
reduces the communication burden and hence becomes a good
candidate for the network-based control/filtering systems subjected
to communication constraints. On the other hand, since only one-
bit information could be obtained at each decoding instant, how to
design an effective prediction function g0 (s(tk )) so as to ensure the
convergence of the decoding error appears to be the main challenge
in the design of the encoder-decoder procedure.

1.3 Encoding-Decoding-Based Control and Filtering Problems

In this section, we will go over the recent theoretical developments of both
the encoding-decoding-based control and filtering problems from various
aspects including controller/filter design and the system performance
evaluation.

• Encoding-Decoding-Based Stabilization Problems

Linear Systems
Among the various system performance indices for the networked
control systems (NCSs), stability is the most basic yet crucial one, and
thus it needs to be first considered in the system design. However,
it should be noted that the introduction of the encoding-decoding-
based communication scheme inevitably leads to the occurrence
of the signal distortion, which would pose considerable additional
complexities on the stability analysis and controller synthesis issues.
A schematic structure of the NCSs with encoding-decoding-based
data transmission protocol can be described in Figure 1.5.
14 Data-Rate-Constrained State Estimation and Control

FIGURE 1.5
NCSs with the encoding-decoding-based communication protocol.

In [16], the pioneering work has been reported on the NCSs with
quantization-based state feedback control strategy, which plays an
important role in the subsequent fruitful results on the encoding-
decoding schemes. In [171], the stabilization issue for a class of linear
time-invariant system subject to the bandwidth constraint has been
considered, where the encoding-decoding-based communication
protocol has been introduced. For the quantization-based encoding
scheme, a sufficient condition for the existence of the bit rate which
enables to stabilize the system has been derived. Moreover, when
the system reduces to a scalar one, a tight sufficient and necessary
condition has been obtained. Nevertheless, the concrete forms of
both the encoder and decoder are not explicitly given in their paper.
The quadratic stabilization problem for sampled-data systems has
been discussed in [53] under a bandwidth-constrained situation,
where the memoryless uniform and logarithmic quantizers have
been designed to overcome the circumstances of the limited bit
rate of the communication channel and an upper bound on the
bit rate has been acquired to achieve the stabilization purpose.
Furthermore, both the state encoding and output encoding schemes
have been developed in [85] for the continuous-time linear systems,
and the encoding procedure has been well described in terms
of the uniform quantization method. Meanwhile, the quantitative
relationship between the number of the values taken from the
encoder and the open-loop system parameters has been derived,
under which the globally asymptotic stability of the closed-loop
system has been guaranteed.
Nonlinear Systems
Subsequently, a number of research results have been extended to
the encoding-decoding issues for the nonlinear NCSs, see [55,80,
85,87,119,123–125, 140, 190]. In [87], by utilizing the decoded system
state, a feedback control strategy has been proposed to stabilize
a class of continuous-time Lipschitz-type nonlinear systems. The
ISS theory has been employed to characterize the impact of the
Introduction 15

decoding error on the closed-loop system state, based on which,

it has been demonstrated that the closed-loop system can achieve
globally asymptotical stability as long as the encoding frequency, the
amount of the information provided by the encoder and the system
parameters satisfy certain inequality conditions. In [124], the integral
ISS condition, which is a weaker assumption than the ISS, has been
obtained to ensure the stability of the closed-loop system under the
same requirement on the channel capacity with [87]. Then, [122]
extended the result to any smooth nonlinear systems, in which, it has
been revealed that if the original nonlinear systems can be stabilized
by certain control strategy without encoding process, it can also be
stabilized by the state feedback scheme with the designed encoding
communication protocol provided the channel capacity is larger
than a given lower bound. Both the detectability and stabilizability
problems have been investigated in [140] for the continuous-time
systems with Lipschitz nonlinearities and bounded noises.
With the help of “zooming” technique, an observer-based
encoding-decoding procedure has been put forward to globally
stabilize the closed-loop system without the knowledge of the bound
of the initial condition. Motivated by the approach proposed in
[140], the subsequent work has been done in a large body of
literature. For example, [190] has applied the approach to the discrete-
time descriptor systems, and the nonlinearities existing in both the
system state and the measurement equations have been considered
in [10]. Recently, the encoding-scheme-based stabilization issue for
the nonlinear uncertain systems has been studied in [80], and the
traditional ISS condition has been relaxed to the differentiable ISS
one. Moreover, by proposing a new uncertain nonlinear system
model, the results obtained in [87] have been further generalized.
• Encoding-Decoding-Based State Estimation Problems
In order to eliminate the confusion, it should be stated that the state
estimation problem considered here is different from the channel
state estimation issue [104, 186] which belongs to the scope of
information theory and aims to recover the state of the channel to
facilitate the information transmission. In the category of control
and signal processing, the state estimation (also called filtering)
problem is to provide a possible accurate reconstruction of the
state of the target plant and has long been a focus of research due
to its engineering insights in many branches, see [46,49,52,67,76,
77,100,103,106, 107, 145, 147, 159, 167, 184,185,191]. For the data-rate-
constrained NCSs, a typical filter structure with the encoder-decoder
pair is shown in Figure 1.6. In this subsection, an overview will
be given on the progress made in the encoding-decoding-based
filtering/state estimation problems in the past few years.
16 Data-Rate-Constrained State Estimation and Control

FIGURE 1.6
Filtering system model with encoding-decoding-based data transmission scheme.

Some initial attempts devoted to the state estimation problems

with the compressed measurement data can be found in [4,15],
and then a rich body of related results have appeared due to the
wide utilization of the prevalent network communication techniques
in a large range of the practical applications, see [12,31,111,112,
144]. For instance, the state estimation problem has been studied
in [111] for the linear discrete-time unstable system with a noisy
memoryless channel of communication constraints. The estimator
has been explicitly designed and an equivalent relationship between
estimation performance and the communication capacity of the
channel has been revealed. For the same system model, its almost
sure observability/stabilizability has been discussed in [112], where
the communication channel was noisy and subject to the limited
transmission capacity. It has been proved that the capability of the
channel to guarantee the almost sure stabilizability/observability
of the plant corresponds to its capability to deliver information
with zero probability of error. The communication-constrained
remote state estimation problem has been addressed in [12], where
both the source coding and channel coding have been concerned.
Without the assumption of the channel feedback existing in the
communication loop, two encoding approaches have been proposed,
and then the analysis results on the performance trade-off between
the convergence rate of the decoding error and computational
complexity have been derived.
It is noted that the results on the encoding-decoding-based state
estimation problems recalled above are generally based on the
information theoretic aspects, while, in recent years, there also has
a surge of research attention on this issue from the point of non-
information-theoretic view, see [41, 110, 117,150,177,178]. In [110], the
robust set-valued state estimation problem has been addressed for the
continuous-time uncertain systems via a digital communication with
Introduction 17

limited bit rate, where two encoding-decoding-based state estimation

schemes have been presented from the aspects of the estimation
precision and the algorithm implementation, respectively. The
concept of “maximin information functional” has been established in
[117] under a nonstochastic information theory framework, by which,
it has been verified that the largest maximin information rate through
a memoryless and error-prone channel reflects the zero-error capacity
of the channel and the result has been applied to the state estimation
issue for a linear time-invariant system. By applying the linear
temporal coding technique, the optimal state estimation problem has
been investigated in [41] for a class of discrete-time system with
an unreliable channel. Both the orthogonal projection principle and
the innovation sequence method have been employed to derive the
minimum mean square estimation algorithm. The extended work has
been done in [150], where the state estimation problem for the linear
discrete-time stochastic systems subject to the packet dropout has
been considered. In order to compensate the negative effect of the
packet loss on the estimation performance, a linear coding temporal
scheme has been employed to the raw measurements, by which some
necessary and sufficient conditions for the estimation error system
have been established in the mean square sense.
• Encoding-Decoding-Based Control/State Estimation Subject to
Network-Enhanced Complexities
In the remote networked control/estimation systems, because of the
power restriction, network congestion or error, it is unavoidable
that the data transmission via the digital channel suffers from
the communication imperfections, for example, transmission delay,
message loss, channel noise, etc. In this sense, for the encoding-
decoding-based control/filtering problem in NCSs, the inherent
complexities of the analysis and design issues have been typically
enhanced and a number of relevant results have been reported in
[24,26,38,54, 56, 101, 126, 158, 165, 168].
Time-Delays
In [126], the minimal data-rate stabilization problem has been
solved for the nonlinear NSs with transmission delays, which can
be arbitrarily large. By applying the tool of the hybrid system
theory, the original system has been converted into a couple of
impulsive delayed nonlinear systems, and the semi-global and local
stability of the closed-system has been guaranteed. The stabilization
problem has been concerned in [26] for a class of distributed
linear systems by a remote centralized control scheme, in which
the random packet dropout and communication delay caused by
an erasure network have been simultaneously considered. Three
18 Data-Rate-Constrained State Estimation and Control

scenarios of the availability of the feedback channel have been,

respectively, discussed and different encoding strategies have been
correspondingly provided to ensure the closed-loop stability in the
mean absolute sense. In [38], the channel coding technique has
been adopted to reduce the occurrence of packet dropouts for
the NCSs. However, the redundancy of the channel coding would
lead to time delays. In order to balance such a trade-off, the ON-
OFF channel coding technique has been utilized where the coding
operation was adaptively implemented according to the expectation
value of the tracking error. The data-rate-constrained stabilization
problem has been investigated in [158] for a class of continuous-
time linear systems. Due to the limited communication capacity of
the channel between the sensor and controller, the sampled data has
been encoded to certain codewords by the spherical polar coordinates
encoding scheme. Dependent on whether the control inputs were
needed to be sent to the decoder, two coding methods have been
developed to achieve the stability of the controlled system.
Packet-Dropouts
For the randomly occurring message loss during the data trans-
mission, a subband coding technique has been employed in [54]
to deal with the controller and encoder-decoder pair design issues.
To facilitate the analysis of the encoding-decoding procedure, the
addressed control problem has been tackled in the frequency
domain, and the stochastic stability and the H∞ performance have
been ensured in terms of the solution to certain linear matrix
inequalities (LMIs). In [56], the attention has been focused on the
stabilization problem for the discrete-time linear NCSs, where the
data transmission has been implemented via a lossy channel with
finite data rate. For the aim of the data compression, the control
signal has been encoded to finite values by the uniform quantization
method and a buffer has been designed in order to compensate the
packet-dropout effect.
For the sake of improving the data transmission reliability
via the erasure channel and obtaining a relatively satisfactory
control/estimation performance, the multiple description encoding
technique has been adopted in the NCSs in recent years. The MD
encoding approach has been first introduced in the NCSs in [59],
in which the Kalman filtering problem has been investigated for
linear stochastic systems. Due to the packet dropout existing in the
data packet transmission via the communication network, the MD
encoding scheme has counteracted the effects of the data loss on
the filtering performance. It has been demonstrated that compared
with the single description encoding scheme, the MD one could
Introduction 19

significantly enhance the filtering performance. In [120], the state

estimation problem has been addressed in the sensor networks
together with the power control mechanism. The sensors were
required to transmit the measurements to the gateway via a lossy
channel subject to random packet dropouts, and then a novel MD
quantizer has been applied to improve the estimation robustness. For
more details with respect to the encoding-decoding-based filtering
issues with lossy/noisy channels, we refer the readers to [92,121,130,
151] and the references therein.
Event-Triggering Mechanism
It is recognized that the signal delivery through the communication
channel, especially for the wireless network, would account for a
major ratio in the total power consumption of the NCSs. Therefore,
for the purpose of keeping the use of the energy resource as small as
possible and alleviating the network collision, the event-triggering
data transmission mechanism has been successfully applied in
the NCSs of digital communications [51,71,73,89,95,102,153,164].
In [64], an early attempt has been made on the combination of
the quantization-based control approach and the event-triggering
approach to address the stabilization problem for linear continuous-
time systems. Based on the quantization method utilized in [85], an
encoder-decoder pair has been proposed to reduce the amount of the
information to be sent via a bit-data-constrained channel; meanwhile,
the event-triggered scheme has been employed to decrease the
communication frequency. The boundedness of the closed-loop
system state and a lower upper of the minimum time interval
between two successive events have been explicitly studied. Very
recently, for a linear scalar system with the communication delay,
an event-triggered control problem combined with the quantization-
based encoding-decoding scheme has been investigated in [61]. The
additional information of the event-triggering time has been further
excavated, and it has been shown that for the sufficiently small
time delay, the timing information carried by the triggered events
is large enough to stabilize the closed-loop system with any small
information transmission rate. Subsequently, the related work has
been extended to the vector systems in [62] and the binary differential
coding approach in [89].
• Optimal Encoding-Decoding Schemes
In the context of bit-rate-constrained control/estimation problems,
the optimal coding scheme can be described from two aspects: 1)
deriving the optimal control/filtering performance with a fixed bit
rate of the network, and 2) determining the minimum information
needed to transmit in order to achieve a given control/filtering task.
20 Data-Rate-Constrained State Estimation and Control

Optimal Control/Estimation Performance

In the first sense, there has been attracting considerable attention in
recent years, see for example [2, 17, 30, 32,57,139,166,188]. In [139],
the optimal control problem for discrete-time nonlinear uncertain
systems via digital communication channels with limited capacity
has been taken into account. Based on the introduced concept
“topological entropy”, sufficient and necessary conditions, which are
related to the bit rate of the channel and the topological entropy of
the open-loop system, have been established for the solvability of the
addressed optimal control problem. In[2], the control problem for a
class of nonlinear systems subject to the communication constraints
has been investigated. A performance index characterizing the
difference between the quantized and unquantized systems has
been defined. By optimizing such an index, the optimal dynamic
quantizer in a closed form has been obtained. Zherlitsyn and Matveev
[188] have considered the distributed estimation issue through a
sensor network with the limited transmission capacity, in which an
optimal-quantization-based encoding-decoding algorithm has been
put forward to minimize the estimation error. The tracking problem
has been addressed in [30] via a signal-noise-ratio-constrained
communication network. Two practical network architectures have
been illustrated to model the transmission channel. Moreover, the
optimal encoder-decoder pair has been found, by which the tracking
error variance has been minimized with minimum information
transmissions.
Minimum Data-Rate
During the past 15 years, the minimum data-rate problem has gained
a great deal of research attention in the network-based control and
signal processing areas, see for example [44,114,115,131,138,155].
Some initial effort has been made in [114], in which the stabilization
problem has been considered for linear discrete-time systems with
the limited communication capacity. By resorting to the asymptotic
quantization theory, a sufficient and necessary condition has been
obtained for the existence of the minimum lower bound of the
bit rate, above which the closed-loop system can be stabilized
in the mth moment sense. The exponential stabilization problem
with limited feedback data rates has been studied in [115], and
moreover, the minimum data rate has been given which can achieve
system stability. In [44], the minimum bit rate has been obtained to
stabilize the linear time-invariant systems for both the continuous-
time and the discrete-time cases. By applying differential pulse code
modulation techniques, comparison results have been provided with
respect to the derived minimum bit rate for the fix-step quantization
Introduction 21

and the variable-step quantization approaches. Very recently, a joint

optimization problem has been discussed in [131]. For a NCS with the
rate-limited communication channel, a general performance index
including both the control cost and the communication cost has been
constructed, and the corresponding optimal control and encoding
problem has been solved.

1.4 Distributed Control and State Estimation Problems with

Limited Communication Capacity
Compared with the simple NCS which consists of only a single plant and
controller/estimator, distributed networked systems are more widespread
in the real-world, such as the chemical process systems, the biological
systems, the unmanned aerial vehicles, etc. For those large-scale distributed
systems, each individual can be viewed as an agent which has the capacity
of sensing, communication, computation, and implementation. In practical
applications, the MASs and the WSNs are two typical examples of the
distributed systems, and the corresponding distributed control and state
estimation problems have attracted increasing research attention in recent
years. Since the frequent information exchange among agents would
certainly occupy considerable network resource and power energy, effective
data-compression and energy-saving transmission mechanisms are vitally
meaningful in real engineering systems. The typical framework of the
distributed networked system with encoding-decoding communication is
provided in Figure 1.7. In this section, we aim to provide some recent
developments with respect to the encoding-decoding-based distributed
control problem for the MASs and the state estimation problem for the WSNs,
respectively.

• Multi-Agent Systems with Encoding-Decoding-Based Information

Transmission
Without Network-Induced Phenomena
Consensus is one of the most important collective behaviors for the
MASs, and the related control problems have received remarkable
research interests in the past decade, see for example [39,94,109,
172,179,181]. It is known from [78] that the encoding-decoding-
based consensus control protocol for MASs was first proposed
in [7], in which, the average consensus performance has been
investigated for a class of linear MASs by the bit-rate-constrained
control protocol. According to different quantization strategies,
22 Data-Rate-Constrained State Estimation and Control

FIGURE 1.7
The architecture of distributed systems with encoding-decoding communication mechanism.

two encoding-decoding schemes have been provided which would

guarantee the desired average consensus if the communication
network is strongly connected. Under a connected communication
network with limited data transmission capacity, [69] has explored
the average consensus control problem for the first-order linear
discrete-time MASs. By designing a dynamic encoding-decoding-
based distributed control law, the exponential average consensus has
been ensured for the requirement that only one bit data exchanged
between two adjacent agents. Furthermore, an energy index has
been defined to characterize the communication energy consumption
for the consensus task, by optimizing which, a trade-off between
the consensus rate and the quantization levels has been presented.
Then, [78] has extended to the time-varying topologies case which
is of jointly connected and the quantized observer-based encoding-
decoding consensus control problem has been considered in [79] for
the second-order discrete-time MASs. The related results with respect
to high-order MASs can be found in [127].
Subject to Network-Induced Complexities
In the above results, a common assumption is that the communication
channel is noiseless, that is, no time delays, dropouts, external
disturbances, etc. during the data transmission process. However,
this is not always true in the engineering practical owing to the
limited network resources and the environmental complexity. In [93],
the authors have discussed the average consensus problem of the
MASs subject to bit-rate constraint, where the communication delays
Introduction 23

have been taken into consideration. By means of constructing a pair of

dynamic encoder and decoder, a consensus control protocol has been
designed to reach the average consensus performance with bounded
time-delays. Because of the merits in the reduction of the energy
and network resource, the event-triggering transmission mechanism
has recently been adopted in the distributed control problem of the
MASs with bandwidth restrictions, see for example [9,34,66,180]. In
[66], in order to handle the consensus control problem for a class
of discrete-time MASs via an imperfection communication network,
both the event-triggering transmission mechanism and the dynamic
encoding-decoding scheme have been employed. The proposed
encoding scheme could adaptively select the number of the channel
quantization level at each time step, and the relationship among the
convergence rate, the number of the agents, the quantization interval,
the network structure, and control gain has been well illustrated.
Based on [66], very recently, the distributed consensus optimization
problem for the MASs has been put forward in [70]. For each agent, a
convex cost function has been defined and the objective optimization
function was the sum of those individual cost functions. It also has
been proved that as long as the system parameters satisfy certain
conditions, the optimality of the consensus control scheme can be
guaranteed with only one bit information exchange for each pair of
adjacent agents. The similar problem with some extensions has been
investigated in [72].
• Wireless Sensor Networks with Encoding-Decoding-Based
Communication Scheme
A wireless sensor network (WSN) is usually composed of numerous
simple and low-cost sensor nodes geographically deployed in certain
areas. With the collaboration such as integrated sensing, computation
and communication, those sensor nodes can jointly perform
various tasks in wide application fields including environmental
monitoring, battlefield surveillance, civil infrastructure monitoring
and control, intelligent solar farms, etc., [18,19,21,22,40,74,108,160,
162,168]. However, it should be stressed that due to the frequent
communications among sensor nodes powered by irreplaceable
batteries, the scarce network bandwidth and energy resources would
give rise to a severe limitation on the practical applications of
the WSNs. Therefore, developing certain useful communication
techniques, which aims to improve the efficiency of the bandwidth
utilization and reduce the energy consumption, appears to be an
urgent task.
So far, an alternative way to facilitate the information exchange in
a bit-rate-constrained WSN is to design appropriate data encoding-
24 Data-Rate-Constrained State Estimation and Control

decoding schemes so as to reduce the number of transmitted

bits at each data transmission process, see [27,136,137]. In [137],
the distributed filtering problem has been considered for a class
of continuous-time stochastic systems with a low communication
capacity. For each sensor, the observation value is required to
be encoded into one bit information to be transmitted, based on
which, a recursive algorithm of the Kalman filtering type has been
designed. Also, it has been revealed that encoding-decoding-based
filtering algorithm can achieve a close performance with the standard
Kalman filtering algorithm. The distributed parameter estimation
issue via the bandwidth-restricted WNSs has been studied in [25],
in which the distributed adaptive quantization approach and a
sequential transmission strategy have been employed. Based on
the adaptive quantization approach, three quantization schemes
have been proposed and the corresponding maximum likelihood
estimators have been designed.
By using the observations from multiple sensors, [65] has
researched the distributed state estimation problem for a discrete-
time linear system. Because of the data-rate constraint, the individual
sensors’ innovations are required to be quantized before sending
to the fusion center. The stability of the estimation scheme can be
achieved if the data rate for each sensor is sufficiently large. In
addition, an optimal rate allocation issue for the individual sensors
has been discussed. [113] has studied the distributed parameter
estimation problem in an energy-constrained WSN, where only
one bit for each sensor has been allowed to be sent to the fusion
center. On the basis of the variance of the measurement noises
of individual sensors, the mean square estimation error can be
minimized by designing an optimal bit assignment scheme. In [132],
based on the Kalman filter structure, a distributed state estimation
problem has been dealt with for the microgrid, and the systematic
convolution encoding approach has been developed to facilitate
smart grid communications. By applying the energy measurements
from the bit-rate-constrained sensor networks, the target location
estimation problem has been addressed in [91]. A position-based
adaptive quantization scheme has been provided to improve the
estimation performance, and the maximum likelihood estimator and
the corresponding Cramér–Rao lower bound have been obtained. In
[129], the distributed state estimation issue has been handled via a
lossy WSN. By using the encoded measurements, the Kalman filter
has been constructed to estimate the system state at the gateway,
which also has controlled the transmission power of the radio power
amplifier and the bit rate needed for each sensor.
Introduction 25

1.5 Outline
Throughout this paper, we have summarized the existing results on the
encoding-decoding-based control and state estimation/filtering issues for
NSs. Several typical encoding-decoding schemes and the corresponding
encoder-decoder structures have been systematically reviewed. Subse-
quently, for different kinds of NSs, a variety of encoding-decoding-based
control and state estimation problems have been thoroughly surveyed.
Finally, based on the literature review, possible relative directions for the
further research work are listed as follows.
The framework of this book is shown as follows:

• In Chapter 1, the research background is first introduced, where

the state estimation and control problems with constrained data
rate are comprehensively reviewed. It is clearly pointed out that
most existing literature has focused on the theoretical analysis for
the data rate condition to certain control or state estimation issue.
The corresponding results for providing concrete encoding-decoding
procedures/algorithms have been much fewer, especially in the
context of networked systems suffering multiple network induced
phenomena. After presenting the motivation, the outline of this book
is then listed.
• In Chapter 2, the gain-scheduled state estimation problem is studied
for a class of complex networks subject to randomly occurring
nonlinearities under bit-rate constraints. A set of random variables
is employed to describe randomly occurring nonlinearities with a
time-varying but bounded occurrence probability taking value in a
known interval. Given a bit rate constraint, an encoding-decoding
scheme is developed by which the boundedness of data distortion
is guaranteed. The objective of this chapter is to propose a gain-
scheduled state estimation algorithm by means of the decoded output
measurements. With the help of the stochastic analysis technique as
well as the Lyapunov stability theory, a sufficient condition is derived
under which the prescribed estimation performance is ensured.
Moreover, the estimator parameter is explicitly determined in terms
of the solution to certain matrix inequalities.
• Chapter 3 studies the partial-neurons-based finite-time state esti-
mation problem for a class of artificial neural networks with
time-varying delays. Measurements information from only a small
fractional of the artificial neurons are applied to the state estimation
process. The data transmission from the sensor to the estimator
is implemented via a bit-rate-constrained communication channel,
26 Data-Rate-Constrained State Estimation and Control

and a data encoding-decoding scheme is developed to convert the

original analog sensor measurements into certain digital codewords
with fewer occupations of the network bandwidth. With the help of
the Lyapunov stability theory, sufficient conditions are presented to
guarantee the finite-time boundedness of the estimation error and
the estimator gain matrix is parameterized in terms of the solution
to certain matrix inequalities.
• Chapter 4 concerns the synchronization control problem for a class of
discrete-time dynamical networks with packet dropouts via a coding-
decoding-based approach. The data is transmitted through digital
communication channels and only a sequence of finite coded signals
is sent to the controller. A series of mutually independent Bernoulli
distributed random variables is utilized to model the packet dropout
phenomenon occurring in the transmissions of coded signals. The
purpose of the addressed synchronization control problem is to
design a suitable coding-decoding procedure for each node, based
on which an efficient decoder-based control protocol is developed
to guarantee that the closed-loop network achieves the desired
synchronization performance.
• In Chapter 5, we deal with the consensus control problem for
a class of discrete-time networked multi-agent systems with the
coding-decoding communication protocol (CDCP). Under a directed
communication topology, an observer-based control scheme is
proposed for each agent by utilizing the relative measurement
outputs between the agent itself and its neighboring ones. The signal
delivery is in a digital manner, which means that only the sequence
of finite coded signals is sent from the observer to the controller. To
be specific, the observed data is encoded to certain codewords by a
designed coder via the CDCP, and the received codewords are then
decoded by the corresponding decoder at the controller side. The
purpose of the addressed problem is to design an observer-based
controller such that the close-loop multi-agent system achieves the
expected consensus performance.
• In Chapter 6, the recursive filtering problem is investigated for a
class of discrete-time stochastic nonlinear systems subject to fading
channels where the multiple description coding scheme is employed
during the data exchange between the devices. In order to facilitate
the data transmission in a resource-constrained communication
network, the multiple description coding scheme is adopted
to encode the fading measurements into two descriptions with
the identical importance. Two independent Bernoulli distributed
random variables are introduced to govern the occurrences of the
packet dropouts in two channels from the encoders to the decoders.
Introduction 27

The channel fading phenomenon is characterized by the Mth-order

Rice fading model whose coefficients are mutually independent
random variables obeying certain probability distributions. The
purpose of the problem addressed is to design a recursive filter such
that, in the simultaneous presence of the stochastic noises, the channel
fading and the data coding-decoding mechanism, an upper bound of
the filtering error variance is obtained and then minimized at each
time step. In virtue of the Riccati difference equation technique and
the stochastic analysis approach, the explicit form of the desired filter
parameters is derived by solving a sequence of coupled algebraic
Riccati-like difference equations.
• Chapter 7 solves the stabilization problem is investigated for a class
of linear discrete-time systems in a resource-constrained network
environment where only the data with finite bits is permitted to
transmit. The data transmissions from the sensors to the controller
are subject to random packet dropouts. To improve the reliability
of the data transmission, a novel dynamical multiple description
coding scheme is developed with guaranteed convergence of the
decoding error. By recurring to the matrix inequality theory and the
stochastic analysis technique, sufficient conditions are established to
ensure the stochastic stability of the closed-loop system. Furthermore,
a parametric relationship is discovered that characterizes the
influences from the packet-dropout probabilities and the encoder
parameters on the stability of the closed-loop system.
• In Chapter 8, the control problem is taken into consideration
for a class of continuous-time linear systems under the event-
triggering encoding strategy. In order to accommodate the digital
communication and save the network resource, a symbol-based
difference coding scheme is proposed to convert the raw data into a
symbolic one with only one-bit size. An event-triggering mechanism
is put forward for the sake of reducing the frequency of both coder
and controller updates. By deriving a lower bound of the inter-
event execution interval, the Zeno behavior is successfully excluded.
Then, a sufficient condition with respect to the required bit rate of
the communication channel is derived to guarantee the convergence
of the proposed encoding-decoding scheme. Moreover, a necessary
condition of the bit rate is also given below which the encoding
error is divergent. By utilizing the input-delay approach, the closed-
loop system is transformed into a continuous time-delay one with
a bounded time-varying delay. Sufficient conditions are established
to guarantee the exponential stability of the closed-loop system. In
addition, the explicit expression of the desired controller parameter is
further obtained in terms of the solution to certain matrix inequality.
28 Data-Rate-Constrained State Estimation and Control

• In Chapter 9, we cope with the state estimation problem for a class of

linear continuous-time systems under bit rate constraints. The data
transmission between the sensors and the estimator is implemented
through a digital communication network with limited bandwidth.
By resorting to the singular value decomposition technique, an event-
based encoding strategy is developed to encode the measurement
signals into 1-bit codewords so as to reduce network resource
consumption. The Zeno phenomenon is first proven to be excluded
and then a sufficient condition is established under which the
decoding error is ultimately bounded. Subsequently, a necessary
condition is proposed to derive a lower bound of the bit rate below
which the decoding error diverges. Moreover, the explicit expression
of the desired state estimator is parameterized and the boundedness
of the estimation errors is also analyzed.
• In Chapter 10, the conclusions and some potential topics for future
work are given.
2
Gain-Scheduled State Estimation for
Discrete-Time Complex Networks under
Bit-Rate Constraints

Complex networks have been successfully applied in a broad scope of

research areas such as social networks, transport networks, neural networks
and power grids. By now, some popular models of complex networks (CNs)
include, but are not limited to, switching CNs, time delay CNs and stochastic
CNs. Built on such models, much effort has been focused on the analysis of
the dynamical behaviors of various CNs such as stability, synchronization,
control and state estimation.
It is noticeable that the state information is quite important for the
dynamical analysis of CNs, which leads to the prosperity of the research
on the filtering/state estimation problem for CNs [75,149]. On the other
hand, due to the complexities caused by the strong couplings between
nodes, the nonlinear function widely exists in the CN. However, due to
unexpected broken links and sudden changes in the external environment, it
is quite common that the nonlinear function appears in a random manner.
Moreover, it is worth mentioning that the occurring probability of the
randomly occurring phenomena may be time varying in engineering practice
due to the unpredictably environmental change. In this context, an important
question is how to reflect the effect of the time-varying probability on the
control/filtering performance.
With the sustained and fast development of network technology,
digitalized communication has displaced traditional analog communication
as a new communication trend due to its distinct advantages in low-cost,
easy wiring, and easy installation, operation and maintenance. This has
also brought up the flourish of the networked systems with plenty of
excellent results on the research of communication protocols and network-
induced phenomena. Therefore, it is an interesting yet potential research
topic to introduce digital communication mechanisms to the state estimation
problem of CNs, which motivates our current investigation.
It is well recognized that the bit rate is an important indicator to
evaluate the network communication capability, that is when the bit rate

DOI: 10.1201/9781003534853-2 29
30 Data-Rate-Constrained State Estimation and Control

is big, it means that the network has a strong communication capability.

However, due to the limited network resources in real systems, the bit-rate-
constrained problem has gained some attention and a variety of coding-
decoding schemes have been developed, see [8,78,79,88,89,116,140,141,
190]. In particular, due to the large-scale and strong couplings, there is a
great deal of data communication between nodes, which leads to a sizable
demand in the bit rate. Therefore, it is of practical significance and theoretical
importance to investigate the bit-rate-constrained state estimation problem
for CNs with the following two crucial difficulties/challenges: 1) how to
allocate the limited network resources to each network node and 2) how
to design the coding-decoding scheme to take full advantage of the limited
bit rate.
Concluding the above discussions, in this chapter, we strive to investigate
the gain-scheduled state estimation problem for a kind of discrete-time
complex networks (DTCNs) under the bit-rate constraints. An array of
random variables is introduced to govern the nonlinearities whose occurring
probability is a time-varying but bounded value with known upper and
lower bounds. A bit-rate constraint model is established and an encoding-
decoding mechanism is proposed, under which an upper bound of the
decoding error is acquired. The primary purpose of the issue considered
is to design a gain-scheduled state estimator to obtain an estimate of the
network state with an acceptable accuracy according to available output
measurements. By means of the stochastic analysis and Lyapunov stability
theory, a sufficient condition is provided such that the estimation error
dynamics achieve the exponentially mean square ultimate boundedness. The
required estimator gain matrix is parameterized by solving a series of matrix
inequalities. A numerical simulation is exploited to show the usefulness of
the obtained gain-scheduled state estimator.

2.1 Problem Formulation

2.1.1 System Model
The dynamics of the considered DTCN with M nodes is described as follows:

M
X
ui (t + 1) = Ai ui (t) + γ (t)fi (ui (t)) + σij 3uj (t) + Di wi (t) (2.1)
j=1

where ui (t) ∈ Rnu is the state vector of the i-th node. The process noise
wi (t) ∈ R is a zero-man Gaussian noise sequence with variance w̄i . σij > 0
is the element located in the i-th row and the j-th column of the coupling
Gain-Scheduled State Estimation 31

configuration matrix 6 = [σij ]M×M . 3 = diag{λ1 , λ2 , . . . , λnu } is the inner-

coupling matrix which determines the connection relationship between the
elements of different nodes. Ai and Di are constant matrices of appropriate
dimensions. γ (t) is a random variable satisfying the Bernoulli distribution
with the following statistical properties:

Prob{γ (t) = 1} = p(t), Prob{γ (t) = 0} = 1 − p(t) (2.2)

where p(t) is a time-varying value and satisfies p(t) ∈ [p1 p2 ] ⊆ [0 1] with

0 < p1 < p2 being two positive scalars.
It is assumed that the nonlinear function fi ( · ) : Rnu → Rnu satisfies the
following constraint:

[fi (x) − fi (y) − H1 (x − y)]T [fi (x) − fi (y) − H2 (x − y)] ≤ 0 (2.3)

for any x, y ∈ Rnu , where H1 and H2 are known constant matrices.

2.1.2 Measurement Model Subject to Bit-Rate Constraint

The measurement output of the i-th node is given as follows:

si (t) = Ci ui (t) + Ei vi (t) i ∈ {1, 2, . . . , M} (2.4)

where si (t) ∈ Rns is the measurement output of node i collected by sensor i,

vi (t) ∈ R is the zero-mean measurement noise with variance v̄i . Ci and Ei are
given constant matrices with appropriate dimensions.
In this chapter, the data transmission from the sensor to the estimator is
implemented by a wireless communication network of limited data rate,
which would result in possible data collision or even packet dropout. Thus,
it is necessary to allocate the available network resource to each node under
certain protocols to alleviate those network-induced phenomena. As such,
the model of the bit rate constraint is formulated as follows:
M
X
Ri ≤ R (2.5)
i=1

where R ∈ R+ is the total available bit rate in terms of the practical

engineering specification and Ri ∈ R+ is the assigned bit rate of the
communication channel for node i.

2.1.3 Bit-Rate-Constrained Encoding-Decoding Scheme

For the digitized data transmission pattern, the raw data collected by
the sensor should be converted into certain binary codes which are
then sent to the destination via the digital channel. In this subsection, a
32 Data-Rate-Constrained State Estimation and Control

uniform-quantization-based encoding-decoding scheme is presented for the

data transmission under limited bit rate. In what follows, we introduce the
quantization
 process for the measurement output si (t). First, a hyperrectangle
Smi = |sij (t)| ≤ li , j = 1, 2, . . . , ns is employed for node i in which the
corresponding measurement output si (t) is contained in, where li > 0 is a
scaling parameter and sij (t) is the j-th component of si (t). Then, for node i,
partitioning the interval [−li li ] into qi subintervals uniformly which are
described by
 
(1) 2li
Iij (li ) , sij − li ≤ sij < −li +
qi
 
(2) 2li 4li
Iij (li ) , sij − li + ≤ sij < −li +
qi qi (2.6)
..
.
 
(q ) 2li
Iij i (li ) , sij li − ≤ sij ≤ li .
qi

As such, the hyperrectangle Smi = {|sij (t)| ≤ li , j = 1, 2, . . . , ns } is uniformly

(α )
divided into qins sub-hyperrectangles Iij i1 (li ) × Iij i2 (li )×· · ·× Iij ins (li ) where
(α ) (α )

αi1 , αi2 , . . ., αins ∈ {1, 2, · · · , qi }.

Taking the bit rate constraint condition (2.5) into account, for the uniform
quantization scheme, the maximum quantization level for each sub-hyper-
(α ) (α ) (α )
rectangle Iij i1 (li ) × Iij i2 (li ) × · · · × Iij ins (li ) is calculated as
j np
s
k
q̄i = 2R i (2.7)

where ⌊·⌋ is the floor function. Thus, the center of the hyperrectangle Smi is
determined by
 T
ηlii (αi1 , αi2 , . . . , αins ) , bi1 bi2 · · · bins (2.8)

(2αij −1)li
where bij = −li + q¯ i , j = 1, 2, . . . , ns . Thus, for any measurement output
(α ) (α )
si (t), there exists a unique sub-hyperrectangle Iij i1 (li ) × Iij i2 (li ) × · · · ×
(α )
Iij ins (li ) such that si (t) is contained in. Moreover, the distance between any
(α ) (α ) (α )
si (t) in the sub-hyperrectangle Ii1 i1 (li ) × Ii2 i2 (li ) ×· · ·× Iinsins (li ) and its center
ηli (αi1 , αi2 , . . . , αins ) is bounded by
i
Gain-Scheduled State Estimation 33

√
n s li
si (t) − ηlii (αi1 , αi2 , . . . , αins ) ≤ np . (2.9)
s R
2 i

Based on the above analysis, the following encoding-decoding scheme is

proposed for node i subject to the constrained bit rate Ri .
Encoding Scheme for Node i:
(α ) (α )
For the measurement output si (t), if si (t) ∈ Ii1 i1 (li ) × Ii2 i2 (li ) × · · · ×
(α )
Iinsins (li ), then the codewords corresponds to si (t) is given as

R  T
Ci i = αi1 αi2 . . . αins ∀i ∈ {1, 2, . . . , M} (2.10)

Decoding Scheme for Node i:

R
For the decoder i, according to the received codewords Ci i for node i, the
corresponding decoded measurement output ŝi (t) is obtained as
 (2αqi1 −1)li 
−li + ns
 ⌊ 2Ri ⌋ 
 
 −l + (2αqi2 −1)li 
R  i ns 
sˆi (t) = ηlii (Ci i ) = 
 ⌊ 2Ri ⌋  . (2.11)

 · · · 

 (2α −1)li 
−li + nqins
⌊ 2Ri ⌋
s

Define the decoding error for the measurement output of node i as s˜i (t) ,
si (t) − ŝi (t) and the augmented form as s̃(t) , [s̃T1 (t) s˜T2 (t) · · · s̃TM (t)]T . It
follows from (2.9)–(2.11) that the decoding error satisfies
√
n s li
ks̃i (t)k ≤ p
ns
(2.12)
⌊ 2R i ⌋

and
v
u ns
uX ns l2i
ks̃(t)k ≤ u
t  p 2 , ̺. (2.13)
ns
i=1 ⌊ 2R i ⌋

By utilizing the decoded measurement output ŝi (t) (i = 1, 2, . . . , M), the

corresponding state estimator for ui (t) is designed as follows:
34 Data-Rate-Constrained State Estimation and Control

M
X
ûi (t + 1) = Ai ûi (t) + p(t)fi (ûi (t)) + σij 3ûj (t)
j=1

+ Ki (t)(ŝi (t) − Ci ui (t)) (2.14)

where p(t) is the time-varying probability given in (2.2), Ki (t) = Ki1 + p(t)Ki2
with Ki1 and Ki2 being the estimator gain matrices to be determined. Denoting
ei (t) , ui (t) − ûi (t) as the estimation error, it is inferred from (2.1)–(2.14) that
the dynamics of the estimation error is described as

ei (t + 1) = (Ai − Ki (t)Ci )ei (t) + γ̃ (t)fi (ui (t)) + p(t)f̃i (ui (t))
M
X
+ σij Ŵej (t) − Ki (t)Ei vi (t) + Ki (t)s̃i (t) (2.15)
i=1

where γ̃ (t) , γ (t) − p(t) and f̃i (ui (t)) , fi (ui (t)) − f̂i (ui (t)). Then, for
node i, compacting the state vector ui (t) and the estimation error ei (t) as
ηi (t) , [eiT (t) uTi (t)]T , we have the following error dynamics:

ηi (t + 1) = Ai (t)ηi (t) + p(t)Fi (ηi (t)) + γ̃ (t)WFi (ηi (t))

M
X
+ σij 3ηj (t) + ε1i (t)ζi (t) + ε2i (t)s̃i (t) (2.16)
i=1

where
 
Ai − Ki (t)Ci 0  T
Ai (t) , , ζi (t) , wTi (t) viT (t) ,
0 Ai
 
˜ 0 −Ki (t)Ei ˜  T
K1i (t) , , K2i (t) , KiT (t) 0
Di 0
 
0 I h iT
W, , Fi (ηi (t)) , f̃iT (ei (t)) fiT (ui (t)) .
0 I

Setting η(t) , [η1T (t) η2T (t) . . . ηM

T (t)]T and stacking the error dynamics

ηi (t) of M nodes together, one has the following augmented system:

¯ (η(t))
η(t + 1) = A(t)η(t) + p(t)F(η(t)) + γ̃ (t)WF
+ K̃1 (t)ζ (t) + K̃2 (t)s̃(t) (2.17)

where

A(t) , diagM {Ai (t)} + 6 ⊗ Ŵ, K̃1 (t) , diagM {K̃1i (t)},

Gain-Scheduled State Estimation 35

K̃2 (t) , diagM {K̃2i (t)}, s̃(t) , [s̃T1 (t) s˜T2 (t) · · · s̃TM (t)]T ,
W̄ , W ⊗ IM , ζ (t) , [ζ1T (t) ζ2T (t) . . . ζM (t)]T
F(η(t)) , [FT1 (η1 (t)) FT2 (η2 (t)) . . . FTM (ηM (t))]T .

The objective of our study is to develop a gain-scheduled state estimation

strategy for DTCNs (2.1) where the signal transmission from the sensor to
the estimator is executed by an encoding-decoding mechanism (2.10)–(2.11)
subject to the bit rate constraint (2.5) such that the dynamics of the augmented
system (2.17) is exponentially mean square ultimately bounded, that is,

E{kη(k)k2 } ≤ e−νt β0 E{kη(0)k2 } + δ (2.18)

where β0 > 0, ν > 0 and δ are certain positive values.

2.2 Main Results

In this section, first, we shall first establish certain sufficient conditions
to guarantee the exponentially mean square ultimate boundedness of the
estimation error dynamics and then acquire the desired gain-scheduled state
estimator gain by solving a series of matrix inequalities.

Theorem 2.1: For a known positive scalar 0 < ν < 1 and the time-varying
probability p(t) belonging to [p1 , p2 ] as well as known gain matrices Ki1 and Ki2 , the
augmented estimation error dynamics (2.17) is exponentially ultimately bounded in
mean square sense if there exist two positive definite matrices Q(p(t)), Q(p(t + 1))
and two positive scalars ǫ1 and ǫ2 such that the following matrix inequality

5(p(t)) , 40 + 4T1 Q(p(t + 1))41 + θ (p(t))4T2 Q(p(t + 1))42 < 0 (2.19)

holds where
 11 
40 412
0 0
40 ,  ∗ 422
0 0 ,
∗ ∗ −ǫ1 IM×ns
41 , [A(t) p(t)IM×nu K̃2 (t)], I , diagM {I2nu },
42 , [0 W̄ 0], θ (p(t)) , p(t)(1 − p(t)),
4011 , −(1 − ν)Q(p(t)) − ǫ2 φ2 , 412 T
0 , −ǫ2 φ1 ,
422
0 , −ǫ2 I, φ̂1 , diag{φ̃1 , φ̃1 }, φ̂2 , diag{φ̃2 , φ̃2 },
36 Data-Rate-Constrained State Estimation and Control

I , diagM {I2nu }, φ1 , diagM {φ̂1 }, φ2 , diagM {φ̂2 },

φ̂1 , diag{φ̃1 , φ̃1 }, φ̂2 , diag{φ̃2 , φ̃2 }, φ1 , diagM {φˆ 1 },
H1T + H2T HT H2 + H2T H1
φ2 , diagM {φˆ 2 }, φ˜ 1 , − , φ˜ 2 , 1 .
2 2

Proof: The Lyapunov function is chosen as

V(t) = ηT (t)Q(p(t))η(t) (2.20)

and its difference along with the trajectory of (2.17) is defined as

1V (t) = ηT (t + 1)Q(p(t + 1))η(t + 1) − ηT (t)Q(p(t))η(t). (2.21)

Next, the mathematical expectation of the term 1V(t) + νV(t) is calculated as

E{1V(t) + νV (t)} = E{ηT (t + 1)Q(p(t + 1))η(t + 1) − (1 − ν)ηT (t)Q(p(t))η(t)}

h iT
= E A(t)η(t) + p(t)F(η(t)) + K˜ 2 (t)s̃(t) Q(p(t + 1))

× [A(t)η(t) + p(t)F(η(t)) + K̃2 (t)s̃(t)]

¯ T Q(p(t + 1))W̄F(η(t))
+ θ (p(t))FT (η(t))W
ζ T (t)K̃1T (t)Q(p(t + 1))K̃1 (t)ζ (t)

T
− (1 − ν)η (t)Q(p(t))η(t) . (2.22)

It is easy to verify that

E{ζ T (t)K̃1T (t)Q(p(t + 1))K̃1 (t)ζ (t)}

n M
oX
≤ λmax {Q(p(t + 1))}E K̃1T (t)K̃1 (t) (w̄i + v̄i ) (2.23)
i=1
, ϕ(t).

In addition, it is seen from (2.3) that

 T   
fi (ui (t)) Inu φ̃1 fi (ui (t))
≤0 (2.24)
ui (t) ∗ φ̃2 ui (t)

and  T   
fi (ei (t)) Inu φ̃1 fi (ei (t))
≤ 0, (2.25)
ei (t) ∗ φ˜ 2 ei (t)
Gain-Scheduled State Estimation 37

which further implies that

 T   
Fi (ηi (t)) I2nu φˆ 1 Fi (ηi (t))
≤0 (2.26)
ηi (t) ∗ φ̂2 ηi (t)

and  T   
F(η(t)) I φ1 F(η(t)) (2.27)
≤ 0.
η(t) ∗ φ2 η(t)
Considering the decoding error (2.13), it follows from (2.22)–(2.27) that

E{1V(t) + νV(t)}
= E{ηT (t + 1)Q(p(t + 1))η(t + 1) − (1 − ν)ηT (t)Qη(t)}
h iT
≤ E A(t)η(t) + p(t)F(η(t)) + K̃2 (t)s̃(t) Q(p(t + 1))

×[A(t)η(t) + p(t)F(η(t)) + K̃2 (t)s̃(t)]

T (2.28)
+ θ (p(t))FT (η(t))W̄ Q(p(t + 1))W̄F(η(t))
− (1 − ν)ηT (t)Q(p(t))η(t) − ǫ1 (s̃T (t)s̃(t) − ρ 2 ) + ϕ(t)
 T   )
F(η(t)) I φ1 F(η (t))
− ǫ2
η(t) ∗ φ2 η(t)
= ξ T (t)5(t)ξ(t) + ϕ̄(t)

where ξ(t) , [ηT (t) FT (η(t)) s̃T (t)]T and ϕ̄(t) , ǫ1 ρ 2 + ϕ(t).
According to the condition (2.19) in Theorem 2.1, one further calculates that

λmin {Q(p(t + 1))}E{kη(t)k2 }

≤ E{V(t + 1)} ≤ (1 − ν)E{V(t)} + ϕ
≤ (1 − ν)2 E{V(t − 1)} + (1 − ν)ϕ̄(t) + ϕ̄(t)
≤ ···
(2.29)
≤ (1 − ν)t+1 E{V (0)} + (1 − ν)t ϕ̄(t)
+ (1 − ν)t−1 ϕ̄(t) + · · · + (1 − ν)ϕ̄(t) + ϕ̄(t)
1 − (1 − ν)t+1
≤ (1 − ν)t+1 E{V(0)} + ϕ¯ (t),
ν
which indicates
1
E{kη(t)k2 } ≤ (1 − ν)t β0 E{kη(0)k2 } + ϕ̄(t) (2.30)
ν
λmax {Q(p(0))}
with β0 , (1 − ν) λ . Hence, the proof of this theorem is ended.
min {Q(p(t+1))}
38 Data-Rate-Constrained State Estimation and Control

Up to now, the analysis issue of the addressed state estimation problem has
been investigated for DTCNs with bit-rate-constrained encoding-decoding
scheme. A sufficient condition has been provided in Theorem 2.1 to show
the ultimate boundedness of the estimation error dynamics (2.17). In what
follows, we shall focus our attention on the estimator design for the
parameters Ki1 and Ki2 .

Theorem 2.2: For a known positive scalar 0 < ν < 1 and the time-varying
probability p(t) belonging to [p1 , p2 ], if there exist positive definite matrices Q(p(t)),
Q(p(t + 1)), two positive scalars ǫ1 and ǫ2 , real-valued matrices Ki1 and Ki2 and a
nonsingular matrix S such that the following matrix inequality
 
40 41T θ (p(t))(S42 )T
4(p(t)) ,  ∗ 43 0 <0 (2.31)
∗ ∗ θ (p(t))4T4

holds where 43 = 44 , Q(p(t + 1)) − ST − S.

Proof: Constructing the congruence transformation diag{I, I, S−1 } to (2.31) leads to

 
40 4T1 θ (p(t))4T2
∗ 43 0 <0 (2.32)
∗ ∗ T
θ (p(t))4̄4

where 4̄4 , S−T Q(p(t + 1))S−1 − S−T − S−1 . Then, with the help of the
Schur Complement Lemma, the inequality (2.19) holds if and only if the following
inequality is true
 
40 4T1 θ (p(t))4T2
∗ −Q(p(t + 1))−1 0  < 0. (2.33)
∗ ∗ −θ (p(t))Q−1 (p(t + 1))

In addition, based on the fact (S−1 − Q−1 (p(t + 1)))T Q(p(t + 1))(S−1 − Q−1 (p(t +
1))) ≥ 0, one obtains the relationship that −Q(p(t + 1))−1 ≤ S−T Q(p(t + 1))S−1 −
S−T − S−1 . Then, it is not difficult to verify that (2.33) holds as long as (2.31) is
true. As such, we can give the conclusion that the exponentially mean square ultimate
boundedness of the augmented system (2.17) is attained for all p(t) ∈ [p1 , p2 ]. Now,
the proof of this theorem is ended.

It is worth mentioning that both the time-varying probabilities p(t + 1) and

p(t) are involved in the matrix inequality (2.31) simultaneously, which leads
to that the matrix inequality (2.31) varies along with the time instant. Thus,
the total number of the LMIs to be solved is accumulated and trends to be
Gain-Scheduled State Estimation 39

infinity when the time horizon reaches to infinite. Unfortunately, this is not
acceptable in engineering practice because the computation burden caused
by the implementation of the state estimation algorithm. In the next theorem,
we will make our effort to overcome such a technical difficulty and propose
an alternative method to transform the infinite number of LMIs to be solved
into finite ones. To facilitate the technical development, we assume that the
time-varying positive definite matrix is with the form Q(p(t)) = Qa + p(t)Qb ,
where Qa and Qb are two positive definite matrices.

Theorem 2.3: For a known positive scalar 0 < ν < 1 and the time-varying
probability p(t) belonging to [p1 , p2 ], if there exist two positive definite matrices
Qa and Qb , two positive scalars ǫ1 and ǫ2 , real-valued matrices Ki1 and Ki2 and a
nonsingular matrix S such that the following LMIs

θ¯ (pjk )(S82 )T
 
80j 8T1j
ϒ jkl , ∗ 83l 0  < 0, j, k, l = 1, 2 (2.34)
∗ ∗ θ̄ (pjk )84l

are satisfied, where 80j has the same form with 40 where only 411
0 is replaced by
−(1 − ν)(Qa + pj Qb ) − ǫ2 φ2 and

81j , [Āj pj I ˜ 2j , diag {K̃2ij }

K̃2j ], 82 , 42 , K M

83l , 84l , Qa + pl Qb − ST − S, θ̄ (pjk ) , pj (1 − pk )

Āj , diag{Ā1j , Ā2j , · · · , A¯ Mj },

   
Ai − K¯ ipj 0 ˜ K¯ ipj
Āij , , K2ij , ,
0 Ai 0

K̄ipj , xKi1 Ci + pj Ki2 Ci .

Moreover, we derive an ultimate upper bound of dynamics (2.17) in mean square

sense as γ , ν1 ρ̄, where ρ̄ , ǫ1 ρ 2 + 2λmax {Q0 + p2 Qp }λmax {5} M
P
 T i=1 (w̄i + v̄i )
T T 2 T
with 5 , diagM {5i } and 5i = diag Di Di , 2Ei (K1i K1i + p2 K2i K2i )Ei .

Proof: Introduce four variables as follows

p2 − p(t) p(t) − p1
χ1 (t) , , χ2 (t) , ,
p2 − p1 p2 − p1
p2 − p(t + 1) p(t + 1) − p1
ψ1 (t) , , ψ2 (t) , .
p2 − p1 p2 − p1
40 Data-Rate-Constrained State Estimation and Control

Then, we can rewrite p(t) and p(t + 1) as

p(t) = χ1 (t)p1 + α2 (t)p2

and
p(t + 1) = ψ1 (t)p1 + ψ2 (t)p2 ,
P2 P2
respectively, where χi (t) ≥ 0, ψi (t) ≥ 0, i=1 χi (t) = 1 and i=1 ψi (t) = 1.
Moreover, one can further calculate that

2
X
Q(p(t)) = Qa + χi (t)pi Qb
i=1

and
2
X
Q(p(t + 1)) = Qa + ψi (t)pi Qb .
i=1

Based on the discussions made above, it is straightforward to turn out that 4(p(t)) =
P2 jkl
j,k,l=1 χj (t)χk (t)ψl (t)ϒ , which means that (2.31) is satisfied if (2.34) is met.
On the other hand, by some n algebraic manipulations,
o one has λmax {Q(p(t + 1))} ≤
T ˜
λmax {Q0 + p2 Qp } and E K̃ (t)K1 (t) ≤ λmax {5}I. Thus, we can now conclude
1
1−µ 1−µ
that limk→∞ E{kξk k2 } ≤ µ limk→∞ ρk+1 ≤ µ ρ̄ = γ , which finishes the proof
of this theorem.

Remark 2.1: Till now, we have handled the state estimation issue for a kind of
DTCNs subject to RONs under bit-rate constraints. Compared with the existing
results on the state estimation of DTCNs, our main results have the following
distinct characteristics: 1) a model of the bit-rate constraint is established for the
DTCN according to the allowable bandwidth of the communication network under
which an encoding-decoding mechanism is proposed; and 2) a general theoretical
framework is provided for the estimation error dynamics, by which the exponentially
ultimate boundedness is explored in the mean square sense. To be more specific, a
sufficient condition is established on the performance analysis in Theorem 2.1 where
the estimation error dynamics is exponentially mean square ultimately bounded.
Then, the probability-dependent state estimator design problem is given in Theorem
2.2. To get rid of the difficulty resulted from the time-varying probability, Theorem
2.3 subsequently provides a finite number of matrix inequalities by replacing the
time-varying probability by its known upper and lower bounds and the desired
estimator gain matrix is further obtained.
Gain-Scheduled State Estimation 41

2.3 An Illustrative Example

A numerical example is given in this section to demonstrate the validness
of the developed bit-rate-constrained state estimation strategy for DTCNs
subject to RONs with a time-varying occurrence probability.
In this example, we consider that the network is composed of three nodes
and the corresponding parameters (e.g. the system matrices, the coupling
matrices, etc.) in the simulation are chosen as follows:
   
0.32 0.4 0.25 0.4
A1 = , A2 = ,
0.5 0.15 0.3 0.25
 
0.53 0.1  
A3 = , C1 = 1.0 0.9 ,
0.7 0.5
   
C2 = 1.0 0.6 , C3 = 1.2 0.7 ,
 T
Di = 0.5 0.5 , 3 = I2 , Ei = 0.1,
 
−0.2 0.1 0.1
6 =  0.1 −0.2 0.1  .
0.1 0.1 −0.2

Choose the nonlinear function fi ( · ) as

 
0.5xi1 (t) + tanh (0.2xi1 (t))
fi (xi (t)) =
− tanh (0.45xi2 (t)) + 0.65xi2 (t)

and thus the matrices H1 and H2 are, respectively, given as

   
−0.5 0 −0.3 0
H1 = , H2 = .
0 0.65 0 0.2

Set the time-varying probability p(t) of RONs as p(t) = p1 + (p2 − p1 )| cos (k)|
with p1 = 0.3 and p2 = 0.9. Let the variances of the process noise wi (t) and the
measurement noise vi (t) as w̄i = 0.1 and v̄i = 0.1. The total available bit rate
of the DTCN is given as 18bps. According to the average allocation protocol,
the individual bit rate for each node is R1 = R2 = R3 = 6bps and the
quantization level is consequently calculated as q̄i = 64. The initial condition
of the system state ui (t) and its estimation are set as u1 (t) = [−0.10 0.14]T ,
u2 (t) = [−0.28 −0.10]T , u3 (t) = [1.7 0.15]T and ûi (t) = [0 0]T . By solving
the LMIs (2.34) with the help of the Matlab LMI toolbox, we obtain the
following estimator parameters:
42 Data-Rate-Constrained State Estimation and Control

 T  T
K11 = 0.2250 0.3776 , K12 = −0.1465 0.1798
 T  T
K21 = 0.4399 0.4323 , K22 = −0.3928 0.0480
 T  T
K31 = 0.1941 0.4535 , K32 = −0.2699 0.0605 .

The simulation results are shown in Figures 2.1–2.6. Figures 2.1–2.2

describe the state strategies of node i and their estimates, and Figure 2.3
illustrates the measurement output of the DTCN and their decoded values.
Figure 2.4 gives the root mean square error of the state estimation for each
node. Figure 2.5 exhibits the decoding error for each component of the
measurement output s(t). Figure 2.6 shows the curve of the time-varying
probability p(t). From the provided simulation results, we can see that
the developed gain-scheduled state estimation scheme achieves a desired
estimation performance and the proposed encoding-decoding mechanism
maintains an acceptable decoding accuracy under a bandwidth resource
constraint. Consequently, the simulation results carried out above well
confirm the applicability of the derived theoretical results.

0.4
0.2
0

0 50 100 150

0.5

-0.5
0 50 100 150

0
0 50 100 150

FIGURE 2.1
State trajectory ui1 (t) (i = 1, 2, 3) and its estimation.
Gain-Scheduled State Estimation 43

0.4

0.2

0 50 100 150

0.2
0.1
0
-0.1

0 50 100 150

0.5

0 50 100 150

FIGURE 2.2
State trajectory ui2 (t) ( = 1, 2, 3) and its estimation.

0.5

0 50 100 150

-1
0 50 100 150

0 50 100 150

FIGURE 2.3
Measurement output si (t) (i = 1, 2, 3) and its decoded value.
44 Data-Rate-Constrained State Estimation and Control

0.4

0.2

0
0 50 100 150

0.4

0.2

0
0 50 100 150

0.5

0
0 50 100 150

FIGURE 2.4
MSE of node i (i = 1, 2, 3).

0.2

-0.2

0 50 100 150

0.2

-0.2

0 50 100 150

0.2

-0.2

0 50 100 150

FIGURE 2.5
Decoding error s̃i (t) (i = 1, 2, 3).
Gain-Scheduled State Estimation 45

0.9

0.8

0.7

0.6

0.5

0.4

0.3
0 50 100 150

FIGURE 2.6
The time-varying probability p(t).

2.4 Summary
This chapter has dealt with the gain-scheduled state estimation issue for
a kind of DTCNs with RONs under bit-rate constraints. In a resource-
constrained network circumstance, a bit-rate constraint model has been
established under which an encoding-decoding scheme has been given and
an upper bound of the decoding error has been acquired. The phenomenon
of RONs has been characterized by an array of Bernoulli random variables.
Sufficient conditions have been provided to ensure that the estimation error
dynamics is exponentially mean square ultimate bounded. The required
gain-scheduled state estimator gain matrix has been obtained in terms of
the solution to certain matrix inequalities. In the end, the usefulness of the
presented gain-scheduled state estimator design scheme has been validated
via a numerical simulation example.
3
Partial-Neurons-Based State Estimation for
Artificial Neural Networks under
Constrained Bit Rate: The Finite-Time Case

With the blooming development of computer science as well as data science,

artificial neural networks (ANNs) have been found extensive applications
ranging from biomedicine, manufacturing and transportation, to smart
home due to their excellent performance in simulating the human brain
for decision-making and ratiocination. In general, the ANN is a kind of
man-made complex network composed of a large number of interconnected
neurons, which is able to perform various high-level tasks such as nonlinear
approximation, associative memory, pattern recognition and optimization.
Due to the strong coupling, highly nonlinearities and large scale of the
network structure, the states of certain neurons are usually difficult to
access directly. In order to have a clear understanding with respect to
the relationship between the dynamical behaviors of ANNs (e.g. the
stability, periodic oscillation, bifurcation, or even chaotic dynamics) and
the performance in handling those complicated tasks, it is desired to acquire
the state information of the neurons. Accordingly, the state estimation issue
for ANNs has been an active research topic and received considerable
research interest in the past few decades., see [118,128,152].
An implicit assumption made in most of the literature considering the state
estimation problem for ANNs is that the measurement outputs of all the
neurons are accessible to the state estimator. However, such a prerequisite
requirement is unrealistic due either to the limited measurement technology
or to the cost of acquiring all the measurements of a large scale of neuron
nodes. As such, in engineering practice, it is neither possible nor necessary
to collect all the neurons’ measurement outputs, which gives rise to the
partial-neurons-based state estimation problem and some relevant results
have been reported in [14, 68]. For example, in [98] the partial-neurons-
based state estimation problem has been addressed for a class of neural
networks in the presence of both state-dependent noises and time-varying
delays. Moreover, for the complex networks with unbounded distributed

DOI: 10.1201/9781003534853-3 46
Partial-Neurons-Based State Estimation 47

delays and energy-bounded measurement noises, the partial-nodes-based

state estimation problem has been discussed in [98].
Serving as one of the essential steps of a networked control/state
estimation process, data exchange has a major effect on the overall system
performance. Generally speaking, the communication quality of a digital
channel is heavily dependent on the bandwidth of the communication
network, which is also known as the bit rate of the channel. In an ideal
case, a communication channel with an infinite bit rate provides a reliable
communication process without any distortion of the original data. However,
this is only an extreme situation and can hardly be satisfied in real
applications due mainly to the limitation of the network resource. In practice,
it is often the case that the bit rate of the communication media is limited,
which is particularly true for those complicated systems with a large scale
of unit systems and the tight communication coupling among them, such as
the ANNs, sensor networks and multi-agent system, etc. Therefore, how to
develop an efficient data transmission mechanism for the ANNs in a bit-rate-
constrained environment is of both theoretical significance and engineering
importance.
It is worth pointing out that most of the existing literature concerning the
state estimation problem for ANNs is based on the Lyapunov asymptotic
stability and the system performance is evaluated over an infinite time
horizon. However, such a performance metric is not suitable in certain
practical scenarios since most of the real systems (e.g. the missile system,
the robot control system, and the calculation of drug dose) require that
the dynamics achieve the prescribed performance requirement in a finite
time interval, which is referred to as the finite time stability [1]. Compared
with the traditional asymptotic stability, the performance index of the finite-
time stability is more appealing since the system trajectory is confined to a
specified state space in a finite time, which would achieve a better transient
performance [1]. So far, although some research efforts have been dedicated
to the finite-time control/estimation problem [1,175], the corresponding
results for the ANN with time-varying delays subject to partial neuron
measurements have been scattered, not to mention the communication
channel is of constrained bit rate, which constitutes the primary motivation
of the investigation in this chapter.

3.1 Problem Formulation

The artificial neural network considered here is composed of nu neurons and
each of them is described by the following discrete-time nonlinear dynamics:
48 Data-Rate-Constrained State Estimation and Control

 nu nu
 X X



 u i (t + 1) = ai u i (t) + b ij p j (u j (t)) + dij qj (uj (t − π(t)))

 j=1 j=1
(3.1)
i = 1, 2, · · · , n,



 + Ji (t) + li wi (t),



ui (r) = φi (r), r ∈ [−πM , 0]

where ui (t) ∈ R stands for the ith neuron’s state information and ai represents
the state feedback weight. pj (·) and qj (·) are the activation functions of neuron
j. wi (t) refers to a bounded exogenous disturbance satisfying kwi (t)k ≤ w̄i
with w̄i being a given scalar. For the ith neuron, the parameters bij and dij are
the connection weights for the non-delayed term pj (uj (t)) and the delayed
term qj (uj (t − πj (t))), respectively. Ji (t) is the external input of the ith neuron.
πj (t) is a positive integer denoting the time-varying delay of the jth neuron,
which satisfies πm ≤ πj (t) ≤ πM with πm and πm being known scalars.
Moreover, it is assumed that φi (r) is a given initial condition sequence.
In this chapter, among the nu neuron nodes, only a small fraction of
measurement outputs can be obtained. Without loss of generality, we employ
a set Na , {1, 2, . . . , ns } ⊂ {1, 2, . . . , nu } to denote the available measurements,
in which the element i ∈ Na corresponds to the ith neuron whose output is
measurable with the following form:

si (t) = ci ui (t) + ri vi (t), i = 1, 2, · · · , ns (3.2)

where si (t) is the measurement output collected by the ith sensor. vi (t) ∈ R
is an amplitude bounded disturbance sequence, that is kvi (t)k ≤ v̄i with v̄i
being a given scalar. ci and ri are given constant scalars.
For the activation functions pi ( · ) and qi ( · ) of the ith neuron, the following
assumption is given.

Assumption 3.1: For the neuron i (i ∈ {i = 1, 2, · · · , nu }), the neuron activation

functions pi ( · ) and qi ( · ) in (3.1) are constrained by the following conditions:

[pi (s1 ) − pi (s2 ) − hi1 (s1 − s2 )][pi (s1 ) − pi (s2 ) − hi2 (s1 − s2 )] ≤ 0,

[qi (s1 ) − qi (s2 ) − gi1 (s1 − s2 )][qi (s1 ) − qi (s2 ) − gi2 (s1 − s2 )] ≤ 0 (3.3)

for any s1 , s2 ∈ R, where hi1 , hi2 , gi1 and gi2 are known constants.
In order to facilitate the further development, we rewrite the component-
based artificial neural network model (3.1) to the following compact form:
(
u(t + 1) = Au(t) + Bp(u(t)) + Dq(u(t − π(t))) + J(t) + Lω(t),
(3.4)
u(r) = φ(r), r ∈ [−πM , 0]
Partial-Neurons-Based State Estimation 49

where
u(t) , vecnu {ui (t)}, p(u(t)) , vecnu {pi (ui (t))},
φ(r) , vecnu {φi (r)}, q(u(t)) , vecnu {qi (ui (t))},
A , diagnu {ai }, qπ (t) (u(t)) , vecnu {q(u(t − πi (t)))},
B , (bij )nu ×nu , D , (dij )nu ×nu ,
J(t) , vecnu {Ji (t)}, L , vecnu {li }.
Similarly, the augmented measurement output model of (3.2) is given as
follows:

s(t) = Cu(t) + Rv(t) (3.5)

where

s(t) , vecns {si (t)}, v(t) , vecns {vi (t)},

   
C , C1 0ns ×(nu −ns ) , R , R1 0ns ×(nu −ns ) ,
C1 , diag{c1 , c2 , · · ·, cns }, R1 , diag{r1 , r2 , · · ·, rns }.

In this chapter, due to the scarcity of the communication resource, the data
delivery from the sensor to the estimator is implemented through a channel
with limited bandwidth. To be more specific, it is assumed that the total
available bit rate of the communication media is R. According to the average
allocation protocol [68], the bit rate of each sensor-to-estimator transmission
channel is assigned as Ri , which means

ns
X
R= Ri . (3.6)
i=1

In order to comply with the digitalized data transmission under the

bandwidth constraint, a uniform-quantization-based encoding-decoding
mechanism is introduced.

3.1.1 Encoding-Decoding Mechanism under Bit Rate Constraint

One of the purposes of the encoding-decoding mechanism is to convert
the raw analogue signal to a digital one with less bits. To do this, in
this subsection, we shall develop an easy-to-implement encoding-decoding
mechanism for the data exchange between the sensor to the state estimator.
First, a brief introduction of the uniform quantization scheme is provided.
For the measurement output s(t) ∈ Rns , we introduce a hyperrectangle

Sb = |sj (t)| ≤ b, j = 1, 2, . . . , ns (3.7)
50 Data-Rate-Constrained State Estimation and Control

in which s(t) is contained, where b is the scaling parameter, which is

determined in terms of the practical requirement. Next, we divide the
interval [−b b] into N subsection uniformly and each of the subsection can
be represented by
 
2b
I1j (b) , sj − b ≤ sj < −b +
N
 
2b 4b
I2j (b) , sj − b + ≤ sj < −b +
N N
..
.
 
2b
INj (b) , sj b − ≤ sj ≤ b . (3.8)
N

Accordingly, the hyperrectangle Sb is divided into Nns sub-hyperrectangles

Iβ1 (b) × Iβ2 (b) × · · · × Iβns (b) where β1 , β2 , . . ., βns ∈ {1, 2, · · · , N}.
Because an individual artificial neuron is allocated Ri bit rate for the
data communication, according to the average allocation protocol, the
quantization level for each Iβi (b) (i = 1, 2, . . . , ns ) is determined as N̄i = 2Ri .
Thus, the center of the hyperrectangle Sb is given by
 T
θb (β1 , β2 , . . . , βns ) , g1 g2 · · · gn s (3.9)

i −1)b
where gi = −b+ (2βN̄ , i = 1, 2, . . . , ns . It is easy to find that the measurement
i
output s(t) uniquely corresponds to a sub-hyperrectangle Iβ1 (b) × Iβ2 (b) ×
· · · × Iβns (b). Moreover, the distance between s(t) and the center of the
hyperrectangle Sb satisfies
v
u ns
uX 1
s(t) − θb (β1 , β2 , . . . , βns ) ≤ t b. (3.10)
i=1
22Ri

After introducing the uniform quantization scheme and giving related

discussions (3.7)–(3.10), we are now in a position to employ the following
encoding-decoding mechanism which will be utilized for the digitalized data
transmission.
Encoding Scheme:
Given the measurement output s(t), if s(t) ∈ Iβ1 (b) × Iβ2 (b) × · · · × Iβns (b),
then s(t) is encoded with the following codewords

 
χR = β1 β2 . . . βn s . (3.11)
Partial-Neurons-Based State Estimation 51

Decoding Scheme:
Having received the generated codewords χR , the measurement output
s(t) is correspondingly decoded as

s̄(t) = ηb (χR )
(2βns −1)b T
h i
(2β −1)b (2β2 −1)b
= −b + 1R −b + · · · −b + . (3.12)
2 1 2R2 2Rns

Denote the decoding error as s̃(t) , s(t) − s̄(t) and it follows from (3.9)–(3.12)
that the decoding error s̃(t) is bounded by
v
u ns
uX 1
ks̃(t)k ≤ t b , b¯ . (3.13)
22Ri
i=1

3.1.2 Estimator Structure

After obtaining the decoded measurement output s̄(t), the following state
estimator is constructed for the dynamical system (3.4)

 û(t + 1) = Aû(t) + Bp(û(t)) + Dq(û(t − π(t))) + J(t)

+ K(s̄(t) − Cû(t)), (3.14)


û(r) = φ̂(r), r ∈ [−πM , 0]

where û(t) is the estimations of u(t) and K ∈ Rnu ×ns is the estimator gain
matrix to be determined later.
Setting ũ(t) , u(t) − û(t), one derives the following estimation error
dynamics:
(
ũ(t + 1) = Ãũ(t) + Bp̃(ũ(t)) + Dq̃(ũ(t − π(t))) + Eξ(t),
(3.15)
ũ(r) = φ̃(r), r ∈ [−πM , 0]

where

à , A − KC, φ̃(r) , φ(r) − φ̂(r),

E , [L −KR −K],
ξ(t) , [wT (t) vT (t) s̃T (t)]T ,
p̃(ũ(t)) , p(u(t)) − p̂(û(t)),
q̃(ũ(t − π(t))) , q(u(t − π(t))) − q̂(û(t − π(t))).

Before proceeding further, we introduce the following definition and

assumption.
52 Data-Rate-Constrained State Estimation and Control

Definition 3.1: The estimation error dynamics (3.15) is said to be finite-time

bounded with respect to (ε1 , ε2 , S, ξ̄ , ks ), if, for ∀t = 1, 2, . . . , ks , the following
relationship is true
 T

 ũ (k)Sũ(k) ≤ ε1 , k ∈ [−πM , 0]

ks
X
T
=⇒ ũT (t)Sũ(t) ≤ ε2 (3.16)


 ξ (t)ξ(t) ≤ ξ̄
t=0

where ξ̄ , (ks + 1)(nu w̄2 + ns v̄2 + b̄), 0 ≤ ε1 < ε2 are known positive scalars
and S > 0 is a prescribed weighted matrix.

Assumption 3.2: The initial values of the error dynamics (3.15) meet the following
requirements ( T
ũ (i)Sũ(i) ≤ ε1
(3.17)
(ũ(i + 1) − ũ(i))T S(ũ(i + 1) − ũ(i)) ≤ ̟
for i = −πM , −πM + 1, · · · , −1, 0.

The objective of this chapter is to design a filter structure (3.14) based on

partial measurable neuron outputs subject to constrained bit-rates such that
the estimation error dynamics (3.15) is finite-time bounded with respect to
(ε1 , ε2 , S, ξ̄ , ks ).

3.2 Main Results

Theorem 3.1: For the ANN (3.4) with the bit-rate constraint (3.6), the estimator
gain matrix K is given. The corresponding state estimation error dynamics (3.15)
is finite-time bounded with respect to (ε1 , ε2 , S, ξ̄ , ks ) if there exist positive matrices
Qi (i = 1, 2, 3, 4), some positive scalars α ∈ (0, 1), τi > 0 (i = 1, 2) and σi > 0 (i =
0, 1) such that the following inequality constraints
 
Ŵ11 0 0 0 Ŵ15 Ŵ16 Ŵ17 Ŵ18
 ∗
 Ŵ22 Ŵ23 0 0 0 0 0 

 ∗
 ∗ Ŵ33 Ŵ34 0 Ŵ36 0 0 

 ∗ ∗ ∗ Ŵ44 0 0 0 0 
 < 0,
Ŵ, (3.18)
 ∗
 ∗ ∗ ∗ Ŵ55 0 0 Ŵ58 

 ∗
 ∗ ∗ ∗ ∗ Ŵ66 0 Ŵ68 

 ∗ ∗ ∗ ∗ ∗ ∗ Ŵ77 Ŵ78 
∗ ∗ ∗ ∗ ∗ ∗ ∗ Ŵ88
Partial-Neurons-Based State Estimation 53

σ0 S ≤ Qj ≤ σ1 S, j = 1, 2, 3, (3.19)
(1 − α)t ρ2 + (1 − α)λmax {Q4 }ξ̄
≤ ε2 (3.20)
σ0

hold, where
  α  T   α 
Ŵ11 , A − KC − 1 − I Q1 + Q1 A − KC − 1 − I
2 2
+ (π + 1)Q2 − τ1 H ˜ 1,
˜ 2 , Ŵ16 , Q1 D, Ŵ17 , Q1 E,
Ŵ15 , Q1 B − τ1 H
(1 − α)τM
Ŵ18 , (A − KC − I)T , Ŵ22 = Ŵ23 = Ŵ34 = Ŵ44 , − Q3 ,
π
2(1 − α )τM
Ŵ33 , − Q3 − τ2 G˜ 1 , Ŵ36 , −τ2 G˜ 2 , Ŵ55 , −τ1 I,
π
Ŵ58 , BT , Ŵ66 , −τ2 I, Ŵ68 , DT , Ŵ77 , −Q4 , Ŵ78 , ET ,
(1 − α)πM
Ŵ88 , −Q −1 , Q , Q1 + π Q3 , π , πM − πm , µ , ,
π
ρ2 , (1 + πM + πM (πM − πm ))σ1 ε1 + (πM − πm )πM ̟ ,
1 1
H̃1 , (H1T H2 + HT2 H1 ), H̃2 , − (HT1 + HT2 ),
2 2
1 T 1
G̃1 , (G1 G2 + GT2 G1 ), G̃2 , − (GT1 + GT2 ),
2 2
H1 , diag{h11 , h21 , h31 }, H2 , diag{h12 , h22 , h32 },
G1 , diag{g11 , g21 , g31 }, G2 , diag{g12 , g22 , g32 }.

Proof: Construct a Lyapunov functional as

V(t) , V1 (t) + V2 (t) + V3 (t) (3.21)

where

V1 (t) , ũT (t)Q1 ũ(t),

t−1
X
V2 (t) , (1 − α)t−i−1 u˜ T (i)Q2 ũ(i),
i=t−π(t)
t−π
Xm t−1
X
V3 (t) , (1 − α)t−i−1 ũT (i)Q2 ũ(i),
j=t−πM +1 i=j
−πm
tX t
X
V4 (t) , (1 − α)t−i (ũ(i) − ũ(i − 1))T Q3 (ũ(i) − ũ(i − 1)).
j=t−πM +1 i=j
54 Data-Rate-Constrained State Estimation and Control

Then, calculating the difference along the trajectory of (3.15), one obtains

4
X
1V(t) = 1Vi (t) (3.22)
i=1

where 1Vi (t) , Vi (t + 1) − Vi (t). Next, terms 1Vi (t) + αVi (t) (i = 1, 2, 3, 4) are,
respectively, calculated as follows:

1V1 (t) + αV1 (t) = V1 (t + 1) − (1 − α)V1 (t)

= ũT (t + 1)Q1 ũ(t + 1) − (1 − α)ũT (t)Q1 ũ(t)}
(3.23)
= (ũ(t + 1) − ũ(t))T Q1 (ũ(t + 1) − ũ(t))
  α 
+ 2ũT (t)Q1 ũ(t + 1) − 1 − ũ(t) ,
2
t
X
1V2 (t) + αV2 (t) = (1 − α)t−i ũT (i)Q2 ũ(i)
i=t−π(t+1)+1
t−1
X
− (1 − α)t−i−1 ũT (i)Q2 ũ(i)
i=t−π(t)
t−1
X
+α (1 − α)t−i−1 ũT (i)Q2 ũ(i)
i=t−π(t)
t
X
= (1 − α)t−i ũT (i)Q2 ũ(i)
i=t−π(t+1)+1
t−1
X
− (1 − α)t−i ũT (i)Q2 ũ(i) (3.24)
i=t−π(t)

= ũT (t)Q2 ũ(t) − (1 − α)π(t) ũT (t − π(t))Q2

t−1
X
× ũ(t − π(t)) + (1 − α )t−i ũT (i)Q2 ũ(i)
i=t−π(t+1)+1
t−1
X
− (1 − α)t−i ũT (i)Q2 ũ(i)
i=t−π(t)+1

≤ ũ (t)Q2 ũ(t) − (1 − α)πM ũT (t − π(t))Q2 ũ(t − π(t))

T

t−π
Xm
+ (1 − α)t−i ũT (i)Q2 ũ(i),
i=t−π(t+1)+1
Partial-Neurons-Based State Estimation 55

t−π
X m +1 t
X
1V3 (t) + αV3 (t) = (1 − α)t−i ũT (i)Q2 ũ(i)
j=t−πM +2 i=j
t−π
Xm t−1
X
− (1 − α)t−i−1 ũT (i)Q2 ũ(i)
j=t−πM +1 i=j
t−π
Xm t−1
X
+α (1 − α)t−i−1 ũT (i)Q2 ũ(i) (3.25)
j=t−πM +1 i=j
 
−πm
tX t
X t−1
X
=  −  (1 − α)t−i ũT (i)Q2 ũ(i)
j=t−πM +1 i=j+1 i=j
t−π
Xm
= πũT (t)Q2 ũ(t) − (1 − α)t−i ũT (i)Q2 ũ(i),
i=t−πM +1

1V4 (t) + αV4 (t)

t−π
X m +1 X
t+1
= (1 − α)t−i+1 (ũ(i) − ũ(i − 1))T Q3 (ũ(i) − ũ(i − 1)) (3.26)
j=t−πM +2 i=j

k−π
Xm t
X
− (1 − α)t−i (ũ(i) − ũ(i − 1))T Q3 (ũ(i) − ũ(i − 1))
j=t−πM +1 i=j

t−π
Xm t
X
+α (1 − α)t−i (ũ(i) − ũ(i − 1))T Q3 (ũ(i) − ũ(i − 1))
j=t−πM +1 i=j

t−π
Xm t+1
 X t 
X
= − (1 − α)t−i+1
j=t−πM +1 i=j+1 i=j

T
× (ũ(i) − ũ(i − 1)) Q3 (ũ(i) − ũ(i − 1))
= π(ũ(t + 1) − ũ(t))T Q3 (ũ(t + 1) − ũ(t))
−πm
tX
− (1 − α)t−i+1 (ũ(i) − ũ(i − 1))T Q3 (ũ(i) − ũ(i − 1))
i=t−πM +1

≤ π(ũ(t + 1) − ũ(t))T Q3 (ũ(t + 1) − ũ(t))

− µ(ũ(t − π(t)) − ũ(t − πM ))T Q3 (ũ(t − π(t)) − ũ(t − πM ))
− µ(ũ(t − πm ) − ũ(t − π(t)))T Q3 (ũ(t − πm ) − ũ(t − π(t))). (3.27)
56 Data-Rate-Constrained State Estimation and Control

Noting the nonlinear constraint for p(u(t)) and q(u(t)) in Assumption 3.1, one
further obtains
 T   
ũ(t) H̃1 H ˜2 ũ(t)
≤ 0, (3.28)
p̃(ũ(t − π(t))) ∗ I p̃(ũ(t))
 T   
ũ(t − π(t)) G̃1 G˜ 2 ũ(t − π(t))
≤ 0. (3.29)
q̃(ũ(t − π(t))) ∗ I q̃(ũ(t − π(t)))

Based on the discussions from (3.22)–(3.29), one derives

1V(t) + αV(t) ≤ (ũ(t + 1) − ũ(t))T Q (ũ(t + 1) − ũ(t))

  α 
+ 2ũT (t)Q1 ũ(t + 1) − 1 − ũ(t) + (π + 1)ũT (t)Q2 ũ(t)
2
− (1 − α)πM ũT (t − π(t))Q2 ũ(t − π(t))

T

− µ ũ(t − πm ) − ũ(t − π(t) Q3 ũ(t − πm ) − ũ(t − π(t))

T

− µ ũ(t − π(t)) − ũ(t − πM ) Q3 ũ(t − π(t)) − ũ(t − πM )
 T   
u˜ (t) H̃1 H̃2 ũ(t)
− τ1
p̃(ũ(t)) ∗ I p̃(ũ(t))
 T   
ũ(t − π(t)) G̃1 G˜ 2 ũ(t − π(t))
− τ2 . (3.30)
q̃(ũ(t − π(t))) ∗ I q̃(ũ(t − π(t)))

In light of (3.30), one has

1V (t) + α V (t) − ξ T (t)Q4 ξ(t) ≤ ηT (t)Ŵ̃η(t) (3.31)

where
h
η(k) , ũT (t) ũT (t − πm ) ũT (t − π(t)) ũT (t − πM )
iT
p̃T (ũ(t)) q̃T (ũ(t − π(t))) ξ T (t) ,
h i
 , Â 0 0 0 B D E ,
   
Aˆ , A − KC − 1 − α I , Ŵ̃ , Ŵ̄ + T Q ,
 2 
Ŵ11 0 0 0 Ŵ15 Ŵ16 Ŵ17
 ∗ Ŵ22 Ŵ23
 0 0 0 0 
 ∗
 ∗ Ŵ 33 Ŵ 34 0 Ŵ 36 0 
Ŵ̄ , 
 ∗ ∗ ∗ Ŵ44 0 0 0 .
 ∗
 ∗ ∗ ∗ Ŵ 55 0 0 

 ∗ ∗ ∗ ∗ ∗ Ŵ66 0 
∗ ∗ ∗ ∗ ∗ ∗ Ŵ77
Partial-Neurons-Based State Estimation 57

In terms of the Schur Complement Lemma, it follows from the condition (3.18) in
Theorem 3.1 that Ŵ̃ < 0 and thus one finds 1V(t) + αV(t) − ξ T (t)Q4 ξ(t) ≤ 0,
which further implies

V(t + 1) ≤ (1 − α)V(t) + λmax {Q4 }ξ T (t)ξ(t)

≤ (1 − α)2 V(t − 1) + (1 − α)λmax {Q4 }ξ T (t − 1)ξ(t − 1)

+ λmax {Q4 }ξ T (t)ξ(t)

.. (3.32)
.
t
X
≤ (1 − α)t+1 V(0) + (1 − α)t−i λmax {Q4 }ξ T (i)ξ(i)
i=0

≤ (1 − α )t+1 V(0) + (1 − α)λmax {Q4 }ξ̄ .

Based on the Assumption 3.2, the initial condition V(0) satisfies

V(0) ≤ (1 + πM + πM (πM − πm ))σ1 ε1 + (πM − πm )πM ̟

, ρ2 .

It is easy to see from (3.22) that V(t) ≥ σ0 ũT (t)Sũ(t). On the basis of (3.19) and the
discussions made above, we have

σ0 ũT (t)Sũ(t) ≤ V (t) ≤ (1 − α)t ρ2 + (1 − α)λmax {Q4 }ξ̄ , (3.33)

which further indicates

(1 − α)t ρ2 + (1 − α )λmax {Q4 }ξ¯

ũT (t)Su(t
˜ )≤ = ε2 . (3.34)
σ0

Consequently, it follows from the definition (3.1) that the estimation error dynamics
(3.15) is finite-time bounded and the proof of this theorem is complete.

Theorem 3.2: For the ANN (3.4) with the bit-rate constraint (3.6), the
corresponding state estimation error dynamics (3.15) is finite-time bounded with
respect to (ε1 , ε2 , S, ξ̄ , ks ) if there exist positive matrices Qi (i = 1, 2, 3, 4), a matrix
X, some positive scalars α ∈ (0, 1), τi > 0 (i = 1, 2) and σi > 0 (i = 0, 1) such that
the following inequality constraints
58 Data-Rate-Constrained State Estimation and Control

 
Ŵ̂11 0 0 0 Ŵ15 Ŵ16 Ŵ̂17 Ŵ̂18
 ∗
 Ŵ22 Ŵ23 0 0 0 0 0 

 ∗
 ∗ Ŵ33 Ŵ34 0 Ŵ36 0 0 

 ∗ ∗ ∗ Ŵ44 0 0 0 0 
 < 0,
Ŵ, (3.35)
 ∗
 ∗ ∗ ∗ Ŵ55 0 0 Ŵ̂58 

 ∗
 ∗ ∗ ∗ ∗ Ŵ66 0 Ŵ68 

 ∗ ∗ ∗ ∗ ∗ ∗ Ŵ77 Ŵ̂78 
∗ ∗ ∗ ∗ ∗ ∗ ∗ Ŵ̂88
σ0 S ≤ Qj ≤ σ1 S, j = 1, 2, 3 (3.36)
(1 − α)t ρ2 + (1 − α)λmax {Q4 }ξ̄
≤ ε2 (3.37)
σ0

hold, where
  α  T   α 
Ŵ̂11 , Q1 A − XC − 1 − Q1 + Q1 A − XC − 1 − Q1
2 2
˜1
+ (π + 1)Q2 − τ1 H
˜ 2 , Ŵ16 , Q1 D, Ŵ̂17 , [Q1 L −XR −X],
Ŵ15 , Q1 B − τ1 H
Ŵ̂18 , (Q1 A − X1 C − Q1 )T ,
(1 − α)τM
Ŵ22 = Ŵ23 = Ŵ34 = Ŵ44 , − Q3 ,
π
2(1 − α)τM
Ŵ33 , − Q3 − τ2 G̃1 , Ŵ36 , −τ2 G˜ 2 , Ŵ55 , −τ1 I,
π
Ŵ̂58 , (Q1 B)T , Ŵ66 , −τ2 I, Ŵ̂68 , (Q1 D)T , Ŵ77 , −Q4 ,
Ŵ̂78 , [Q1 L −XR −X]T , Ŵ88 , Q − QT1 − Q1 .

Moreover, the estimator gain matrix K is determined by

K = Q1−1 X. (3.38)
1
Proof: Implementing the congruence transformation diag{I, I, I,I, I, I, I, Q−
1 } to
(3.35) results in
 
Ŵ̂11 0 0 0 Ŵ15 Ŵ16 Ŵ̂17 Ŵ18
 ∗
 Ŵ22 Ŵ23 0 0 0 0 0 

 ∗
 ∗ Ŵ33 Ŵ34 0 Ŵ36 0 0 

 ∗ ∗ ∗ Ŵ44 0 0 0 0 
<0
 (3.39)
 ∗
 ∗ ∗ ∗ Ŵ55 0 0 Ŵ58 

 ∗
 ∗ ∗ ∗ ∗ Ŵ66 0 Ŵ68 

 ∗ ∗ ∗ ∗ ∗ ∗ Ŵ77 Ŵ78 
∗ ∗ ∗ ∗ ∗ ∗ ∗ Ŵˆ 1
88
Partial-Neurons-Based State Estimation 59

1 , Q−T QQ−1 − Q−T − Q−1 and the other parameters are as the same
where Ŵ̂88 1 1 1 1
−1
as given in Theorem 3.1. Applying the useful inequality −Q −1 ≤ Q− T
1 QQ1 −
−T −1
Q1 − Q1 and (3.38), it is not difficult to see that (3.18) is true provided that
(3.35) is met. Therefore, according to the proof of Theorem 3.1, the desired finite-time
bounded requirement is accordingly achieved and the state estimator gain matrix can
be designed as K = Q−1 1 X. The proof of this theorem is now ended.

3.3 An Illustrative Example

This section presents an illustrative numerical example to show the effec-
tiveness of the proposed partial-neurons-based finite-time state estimation
problem for a class of discrete-time delayed neural networks subject to
constrained bit rates.
In this example, the considered neural network is assumed to be composed
of three neurons and the relevant parameters (e.g. the system matrices, the
connection weights, etc.) are in the simulation are given as follows:
   
0.81 0 0 0.2 0.1 0.4
A = 0 0.85 0  , B = 0.3 0.2 0  ,
0 0 0.75 0 0.3 0.4
 
  0.5 0.4 0.13
2.2 0 0
C= , D = 0.5 0.1 0.1  ,
0 2.8 0
0.5 0.1 0.5
 
  0.1 0 0
0.05 0
R= , L =  0 0.3 0.3 .
0 0.04
0 0 0.2

It is seen from the expression of matrix C that only the first and the second
neurons’ state information can be measured and the status of the third neuron
is unavailable, which is consistent with the theoretical analysis for the partial-
neuron-based state estimation problem. The nonlinear activation function
p( · ) and q( · ) are chosen as
 
−0.5u1 (t) + tanh (0.2u1 (t))
p(u(t)) = − tanh (0.45u2 (t)) + 0.65u2 (t) ,
tanh (0.7u3 (t)) + 0.3u3 (t)
 
−0.5u1 (t − π(t)) + tanh (0.2u1 (t − π(t)))
q(u(t − π(t))) = − tanh (0.45u2 (t − π(t))) + 0.65u2 (t − π(t)) .
tanh (0.7u3 (t − π(t))) + 0.3u3 (t − π(t))
60 Data-Rate-Constrained State Estimation and Control

t
Set the varying time delay as π(t) = 2 + 1+(−1 2
)
and consequently the upper
and lower bounds of π(t) are, respectively, chosen as πm = 2 and πM = 3.
The input signal is set as J(t) = [0.1 exp ( − t) 0 0.08 sin (t/2)]T . Let the
upper bound of the process disturbance wi (t) and the measurement noise
vi (t) take values over the interval [−0.5 0.5] uniformly and thus their upper
bounds are, respectively, given as w̄i = 0.5 and v̄i = 0.5. The total available
bit rate of the complex network is given as 24bps. According to the average
allocation protocol in [68], the individual bit rate for each neuron node is
R1 = R2 = R3 = 8bps and the quantization level is consequently calculated
as q̄i = 256.
The initial condition of the system state u(t) and its estimation are given as
u(t) = [−1 1.4 −0.8]T and û(t) = [0 0 0]T . Let the matrix S in Definition
3.1 be S = 0.1I. With the help of the Matlab toolbox, the matrix inequalities
in (3.35) can be solved and the estimator gain matrix is determined as
 
0.2683 0.1564
K = 0.0788 0.2607 .
0.3369 0.2755

The corresponding simulation results are exhibited in Figures 3.1–3.7.

Figures 3.1–3.3 show the curves of the states of three neurons and their

0.5

-0.5

-1

-1.5

-2

-2.5

-3

-3.5
0 50 100 150 200 250

FIGURE 3.1
State trajectory u1 (t) and its estimation.
Partial-Neurons-Based State Estimation 61

1.5

0.5

-0.5

-1

-1.5

-2

-2.5
0 50 100 150 200 250

FIGURE 3.2
State trajectory u2 (t) and its estimation.

-1

-2

-3

-4

-5

-6

-7

-8

-9
0 50 100 150 200 250

FIGURE 3.3
State trajectory u3 (t) and its estimation.
62 Data-Rate-Constrained State Estimation and Control

-1

-2
0 50 100 150 200 250

0.5

0
0 50 100 150 200 250

-1

-2
50 100 150 200 250

FIGURE 3.4
Estimation error of neuron i (i = 1, 2, 3).

-2

-4

-6

-8

-10

-12

-14
0 50 100 150 200 250

FIGURE 3.5
The measurement output s1 (t) of the first neuron and its decoded value.
Partial-Neurons-Based State Estimation 63

-1

-2
0 50 100 150 200 250

FIGURE 3.6
The measurement output s2 (t) of the second neuron and its decoded value.

0.1

0.05

-0.05

-0.1
0 50 100 150 200 250

0.1

0.05

-0.05

-0.1
0 50 100 150 200 250

FIGURE 3.7
Decoding error s̃i (t) (i = 1, 2).
64 Data-Rate-Constrained State Estimation and Control

estimates. Figure 3.4 gives the estimation error of each neuron state.
Figures 3.5–3.6 describe the measurement output si (t) (i = 1, 2) of the first
and second neurons and their decoded values. Figure 3.7 plots the decoding
error for the measurement output s(t) of each neuron. It can be observed
from the above simulation results that the proposed partial-neurons-based
state estimation strategy achieves an acceptable estimation performance in
a bandwidth resource-constrained communication scenario with the aid of
the proposed encoding-decoding mechanism. Therefore, all the simulation
results well support the validness of the derived theoretical results.

3.4 Summary
In this chapter, the finite-time state estimation issue has been addressed
for a class of time-delayed ANNs subject to certain bit rate constraint. For
the purpose of meeting the network bandwidth requirement, a uniform-
quantization-based encoding-decoding mechanism has been applied to
encode the raw measurement outputs into specific digital codewords which
occupy fewer communication resource. During the design of the state
estimation scheme, only a small fraction of the neurons’ outputs have
been assumed to be measurable and utilized. By constructing a time-delay-
dependent Lyapunov functional, sufficient conditions have been established
to ensure the finite-time boundedness of the estimation error dynamics. The
desired estimator gain matrix has been obtained by solving certain matrix
inequalities. In addition, the effectiveness of the developed state estimator
has been further demonstrated by a numerical simulation.
4
Synchronization Control for a Class of
Discrete-Time Dynamical Networks with
Packet Dropouts: A Coding-Decoding-Based
Approach

The synchronization problem, which is concerned with a typical collective

behavior in nature, has long been serving as an important research topic
for dynamical networks because of its practical background and potential
applications. However, in reality, there may be a case that the synchronization
phenomenon cannot be attained for autonomous networks by their local
connections, that is, the dynamic evolutions exhibit the asynchronous
nature. For such kind of asynchronous networks, there is a vital need to
develop effective synchronization control approaches in order to enforce the
trajectories of all nodes to the desired synchronization manifold.
Concerning the synchronization control problems of dynamical networks,
the available control strategies include, but are not limited to, the sampled-
data control [146], the impulsive control [42], the pinning control [105] and
the intermittent control [48]. Note that most reported synchronization control
protocols have been assumed to work within the analog communication
framework where the information is translated into electric pulses of varying
amplitude. Compared with the traditional data transmission scheme in
an analog manner, up to now, very little attention has been devoted to
the synchronization control issue of dynamical networks with a digital
communication mechanism, despite the fact that the digital information
interaction has already become a dominant communication fashion for its
obvious advantages of strong disturbance rejection capacity, low power
consumption as well as high reliability. It is worth mentioning that, in
the digitalization era, the digital communication strategy has been widely
utilized in the areas of signal processing and control engineering.
In the real world, almost all the physical signals are virtually of analog
nature and it would be improper to directly transmit them through digital
channels. As such, before transmitting such analog signals in a digital
pattern, one should first transform them into the digital form by an analog-
to-digital (A/D) converter with the steps including sampling, reservation,

DOI: 10.1201/9781003534853-4 65
66 Data-Rate-Constrained State Estimation and Control

quantization and encoding. To some extent, the data coding stage is most
crucial in realizing the digitization transmission since it is closely associated
with the detectability issue, that is, whether and how the coded data could
be recovered from the codewords with a prescribed accuracy requirement.
Compared with the traditional analog communication schemes, the
distinct features of the digital communication protocols lie in that they a)
have a good noise immunity; b) possess a high communication reliability;
and c) are convenient for the data encryption. Here, the features a) and b)
facilitate the long distant data transmission, and the feature c) prevents the
transmitted data from being vulnerable to attacks launched by adversaries
in response to the increasingly important security requirement of the
information transmission. In particular, it is significantly important to design
an appropriate coding-decoding procedure so as to ensure a satisfactory
control performance via digital communication channels.
In the past few years, the coding-decoding communication scheme has
gained some initial research attention. Up to now, almost all results
concerning the coding-decoding issues have been on linear/nonlinear
systems without network topology specifications, and the corresponding
results on dynamical networks have not been addressed yet despite their
wide applications in engineering practice.
The coding-decoding-based synchronization control problem for discrete-
time dynamical networks appears to be an essentially difficult issue with
three fundamental challenges identified as follows. 1) The first challenge
stems from the mathematical difficulty in analyzing the addressed problem.
As a typical dynamical network usually comes with both high dimension and
tight coupling among the nodes, rather complicated/difficult mathematical
analysis is expected and this is particularly true when the random
packet dropout phenomenon is taken into account during the codeword
transmissions [154]. On the other hand, due to the existence of the decoding
error between the decoded system state and actual system state, it is fairly
challenging to research into the issue of how to utilize imperfect decoded
information to attain the desired control objective. In addition, since the
performance of the coding-decoding protocol is influenced by a number
of factors (e.g. the coding period, the size of the coding alphabet and the
probabilistic packet dropouts), the third challenge is how to examine the
effects of these combined factors on the prescribed system performance
indices. It is, therefore, the main aim of this chapter to deal with the listed
challenges by addressing the synchronization control problem for dynamical
networks using a coding-decoding-based approach.

4.1 Problem Formulation

Consider a discrete-time dynamical network consisting of N coupled nodes
described by:
Synchronization Control for a Class of Discrete-Time Networks 67

 N
 X
1)



 x i (k + = f (x i (k)) + wij Ŵxj (k) + ui (k),

 j=1
(4.1)



 yi (k) = Ci xi (k),



xi (0) = xi0 ∈ X0 , i = 1, 2, . . . , N

where, for the ith node, xi (k) ∈ Rn is the state vector which is inaccessible
directly, and yi (k) ∈ Rm and ui (k) ∈ Rn are the measurement output and
the control input, respectively. Ŵ = diag{γ1 , γ2 , . . . , γn } is the inner coupling
matrix linking the jth state variable if γj 6 = 0. W = [wij ]N×N is the coupled
configuration matrix of the network with wij ≥ 0 (i 6 = j) but not all zero. xi0 is
the initial value of the ith node belonging to a known set X0 . Ci is a constant
matrix of appropriate dimensions. The nonlinear vector-valued function
f ( · ) : Rn 7 → Rn is continuous which satisfies f (0) = 0 and the following
condition:

[f (x) − f (y) − U1 (x − y)]T [f (x) − f (y) − U2 (x − y)] ≤ 0 (4.2)

for all x, y ∈ Rn , where U1 and U2 are known real matrices of appropriate

dimensions.
Note that full state information of the dynamical network (4.1) may not be
obtained directly and only the network output is available. In this case, we
construct the following state estimator for node i:

 N
 X
1

 x̃
 i
 (k + ) = f (x̃ i (k)) + wij Ŵx̃j (k)

 j=1
(4.3)



 + Li (yi (k) − Ci x̃i (k)) + ui (k),



x̃i (0) = x̃i0 ∈ X0

where x̃i (k) ∈ Rn is the estimator state for the ith node, Li ∈ Rn×m is the
estimator gain matrix to be designed and x̃i0 is the ith estimator’s initial
condition.
Let s(k) ∈ Rn be the solution to the unforced isolate node
(
s(k + 1) = f (s(k)),
(4.4)
s(0) = s0 ∈ X0

where s0 is the initial condition.

Denote by esi (k) , xi (k) − s(k) the synchronization error vector. Then, for
node i, we have the following synchronization error dynamics:
68 Data-Rate-Constrained State Estimation and Control

 N
 X
 esi (k + 1) = f (esi (k)) +

 wij Ŵesj (k) + ui (k),
j=1 (4.5a)




esi (0) = esi0 ∈ X0

where f (esi (k)) , f (xi (k)) − f (s(k)) and esi0 , xi0 − s0 is the initial value.
To comply with the digital communication fashion, the purpose of this
chapter is to address the synchronization control problem for the dynamical
network (4.1) by using a coding-decoding communication protocol. To be
exact, by developing an efficient coding-decoding procedure based on the
estimated network states, we shall design a decoder-based control protocol
such that all the network nodes can be synchronized to the dynamics (4.4).
Before proceeding, a general form of the coding-decoding procedure for
dynamical networks (4.1) is given as follows:
Coder for Node i:

gi (lh) = Fil (x̃i (h), x˜ i (2h), . . . , x̃i (lh)), (4.6)

Decoder for Node i:

X̂i (lh) = Gil (αi (h)gi (h), αi (2h)gi (2h), . . . , αi (lh)gi (lh)) (4.7)

for l = 1, 2, . . ., where gi (lh) is the codeword generated from the coder i at

the coding instant lh, and X̂i (lh) is defined as

X̂i (lh) , {x̂i (lh), x̂i (lh + 1), . . . , x̂i ((l + 1)h − 1)}

with x̂i (k) being the decoded state of x̃i (k) for k ∈ [lh, (l+1)h). Fli (·) and Gil (·) are
coder and decoder functions to be designed, respectively. αi (lh) is a Bernoulli
distributed random variable which characterizes the packet dropouts and
satisfies the following probability distribution:

Prob{αi (lh) = 0} = ᾱi , Prob{αi (lh) = 1} = 1 − ᾱi

where ᾱi ∈ [0, 1] is a given scalar. In addition, αi (lh) and αj (lh) are mutually
independent random variables for i 6 = j with i, j = 1, 2, . . . , N.
For node i, in terms of the decoded state x̂i (k), the decoder-based
synchronization controller is given as

ui (k) = Kc (x̂i (k) − s(k)) (4.8)

where Kc is the controller parameter to be designed.

Synchronization Control for a Class of Discrete-Time Networks 69

Remark 4.1: In the coding-decoding procedure (4.6)–(4.7), by employing the

Bernoulli distributed white sequence αi (lh), the phenomenon of probabilistic packet
loss occurring in the transmission of codewords is well described, thereby better
reflecting the reality. Also, because the coding period is h times of the system sampling
period, it can be observed from (4.7) that, for a fixed coding period [lh, (l + 1)h),
the decoding function Gil ( · ) of node i applies the codeword gi (lh) to generate a
set of decoded state x̂i (k) (k ∈ [lh, (l + 1)h)) which indicates that the decoder can
generate the decoded states at both the coding instant lh and those non-coding time
instants lh + 1, · · · , (l + 1)h − 1. Here, the decoded state x̂i (k) k ∈ (lh, (l + 1)h)
generated at non-coding instants can be viewed as a “prediction” of the actual
system state.

For notation simplicity, we denote

x(k) = vecN {xi (k)}, x̃(k) = vecN {x̃i (k)},

x̂(k) = vecN {x̂i (k)}, es (k) = vecN {esi (k)},
x0 = vecN {xi0 }, x̃0 = vecN {x̃i0 },
(4.9)
es0 = vecN {esi0 }, F(x(k)) = vecN {f (xi (k)},
F(x̃(k)) = vecN {f (x̃i (k)}, C = diagN {Ci },
F(es (k)) = vecN {f (esi (k))}, L = diagN {Li }.

By recurring to the Kronecker product, the original dynamical network (4.1)

can be rewritten in the following compact form:

 x(k + 1) = F(x(k)) + (W ⊗ Ŵ)x(k) + u(k),

y(k) = Cx(k), (4.10)


x(0) = x0 .

Accordingly, we develop the following state estimator for the dynamical

network (4.10):

˜ + 1) = F(x̃(k)) + (W ⊗ Ŵ)x̃(k)
 x(k

+ LC(x(k) − x̃(k)) + u(k), (4.11)


x̃(0) = x̃0

and obtain the compact form of synchronization error dynamics (4.5a) as

follows:
(
es (k + 1) = F(es (k)) + (W ⊗ Ŵ)es (k) + u(k),
(4.12)
es (0) = es0 .
70 Data-Rate-Constrained State Estimation and Control

To facilitate the subsequent developments, we introduce the following

definitions which are needed for stating the problem to be investigated.

Definition 4.1: Consider the discrete-time dynamical network of the compact form
(4.10) subject to packet dropouts during the codeword transmissions. The network
(4.10) is said to be detectable if there exist families of coder-decoder pairs (4.6) and
(4.7) with a coding alphabet H of size χ such that

lim E{kx(k) − x̂(k)k2 } = 0 (4.13)

k→∞

holds for any solution of (4.10).

Definition 4.2: The discrete-time dynamical network (4.10) is said to be

synchronized to the isolated node (4.4) if there exist families of decoder-controller
pairs (4.8) such that

lim E{kes (k)k2 } = 0 (4.14)

k→∞

holds for any solution of closed-loop system (4.12).

In this chapter, we investigate the synchronization control problem for the

dynamical network (4.10) by using a decoder-based control protocol. For the sake
of realizing the synchronization criterion, one has to examine the detectability of the
whole dynamical networks first. In other words, we are interested in developing an
efficient coding-decoding procedure (4.6) and (4.7) for each node such that
a) The dynamical network (4.10) subject to the packet dropouts is detectable.
b) Based on the established detectability criterion, the dynamical network (4.10)
is synchronized to the prescribed dynamics s(k) by the decoder-based control
protocol (4.8).

4.2 Main Results

In this section, our attention is focused on the analysis and design problems
of the coding-decoding procedure for dynamical network (4.10). Then, based
on the established coding-decoding procedure, we plan to deal with the
decoder-based synchronization control problem for the addressed dynamical
network to achieve the desired performance requirements.

4.2.1 Preliminaries
To start with, we give the following useful lemma which will be needed for
the subsequent derivation of our main results.
Synchronization Control for a Class of Discrete-Time Networks 71

Lemma 4.1: Let the scalar µ1 > 0 be given. If there exist a positive definite matrix
P > 0 and a scalar ε1 > 0 satisfying the following linear matrix inequality (LMI)
 
−511 −ε1 8T23 (W ⊗ Ŵ)T P
51 =  ∗ −ε1 INn P  < 0, (4.15)
∗ ∗ −P

then we have

kv1 (k + 1) − v2 (k + 1)k2 ≤ c0 kv1 (k) − v2 (k)k2 (4.16)

√ (1+µ1 )λmax {P}
where 511 = (1 + µ1 )P − ε1 813 , c0 = β, β = λmin {P} , and v1 (k) and v2 (k)
are any two solutions of (4.10).

Proof: First, denoting z(k) , v1 (k) − v2 (k) and F(z(k)) , F(v1 (k)) − F(v2 (k)), it
can be inferred immediately from (4.10) that z(k+1) = F(z(k))+(W⊗Ŵ)z(k). Then,
consider the Lyapunov function V(k) = zT (k)Pz(k). The term 1V(k) − µ1 V(k),
where 1V(k) is the difference of V(k) along the dynamics of z(k + 1) = F(z(k)) +
(W ⊗ Ŵ)z(k), can be subsequently calculated as

1V(k) − µ1 V(k) = zT (k)[(W ⊗ Ŵ)T P(W ⊗ Ŵ) − (1 + µ1 )P]z(k)

(4.17)
+ 2FT (z(k))P(W ⊗ Ŵ)z(k) + FT (z(k))PF(z(k)).

It is easy to see from (4.2) that

 T   
z(k) 813 8T23 z(k)
(4.18)
≤0
F(z(k)) ∗ INn F(z(k))

U1T U2 +U2T U1
where 813 = IN ⊗ 81 , 823 = IN ⊗ 82 , 81 = 2 and 82 = − U1 +2 U2 .
Then, substituting (4.18) into (4.17) yields

1V(k) − µ1 V(k) ≤ zT (k)[(W ⊗ Ŵ)T P(W ⊗ Ŵ) − (1 + µ1 )P]z(k)

+ 2FT (z(k))P(W ⊗ Ŵ)z(k) + FT (z(k))PF(z(k))
 T " #  (4.19)
z(k) 813 8T23 z(k)
− ε1
F(z(k)) ∗ INn F(z(k))

¯ 1 ξ1 (k)
= ξ1T (k)5

where
 T
ξ1 (k) = zT (k) FT (z(k)) ,
72 Data-Rate-Constrained State Estimation and Control

 
5¯ 11 (W ⊗ Ŵ)T P − ε1 8T23
¯1 =
5
∗ P − ε1 INn

with 5̄11 = (W ⊗ Ŵ)T P(W ⊗ Ŵ) − (1 + µ1 )P − ε1 813 .

By using the Schur Complement Lemma, it can be obtained from (4.15) that
5̄1 < 0, and therefore we have that V(k + 1) ≤ (1 + µ1 )V(k). In addition, it can
be readily seen from the definition of V(k) that λmin {P}kz(k + 1)k22 ≤ V(k + 1) ≤
√
(1 + µ1 )λmax {P}kz(k)k22 , which implies that kz(k + 1)k2 ≤ β kz(k)k2 , namely,
kv1 (k + 1) − v2 (k + 1)k2 ≤ c0 kv1 (k) − v2 (k)k2 . The proof is complete.

4.2.2 State Estimator Design

The following lemma provides a sufficient condition for the design problem
of the state estimator (4.11) to ensure that the estimation error dynamics is
convergent. Furthermore, the estimator gain matrix for each node is explicitly
characterized.

Lemma 4.2: Let the positive scalar 0 < µ2 < 1 be given and consider the dynamical
network (4.10) with estimator (4.11). It is assumed that there exist a positive definite
matrix Q = diagN {Qi } > 0, a matrix X = diagN {Xi } and a positive scalar ε2 > 0
satisfying
 
−(1 − µ2 )Q − ε2 813 T
−ε2 823 Q̄
Q <0 (4.20)
 
52 =  ∗ −ε2 INn
∗ ∗ −Q

where Q̄ = (Q(W ⊗ Ŵ) − XC)T . Then, for the augmented dynamical network (4.10)
and estimator (4.11), there always exist an integer h ∈ Z+ and a scalar c1 (0 < c1 < 1)
such that
kx(k + h) − x̃(k + h)k2 < c1 kx(k) − x̃(k)k2 . (4.21)

Moreover, the desired estimator gain matrix for each node can be obtained in the
form of
1
Li = Q −
i Xi , i = 1, 2, . . . , N. (4.22)

Proof: Following the similar line in the proof of Lemma 4.1, we denote the estimation
error as e(k) , x(k) − x̃(k). Then, subtracting (4.11) from (4.10) results in the
estimation error dynamics

e(k + 1) = F(e(k)) + (W ⊗ Ŵ − LC)e(k),

where F(e(k)) , F(x(k)) − F(x̃(k)). Choosing the Lyapunov function as V(k) =

eT (k)Qe(k), we have
Synchronization Control for a Class of Discrete-Time Networks 73

1V(k) + µ2 V(k) = [F(e(k)) + (W ⊗ Ŵ − LC)e(k)]T Q[F(e(k))

+ (W ⊗ Ŵ − LC)e(k)] − (1 − µ2 )eT (k)Qe(k)
n
= eT (k) [(W ⊗ Ŵ) − LC]T Q[(W ⊗ Ŵ) − LC]
− (1 − µ2 )Q} e(k) + 2FT (e(k))Q[(W ⊗ Ŵ) − LC]e(k)
+ FT (e(k))QF(e(k)).
(4.23)
Similarly, by noting (4.2), one has
 T   
e(k) 813 8T23 e(k)
(4.24)
≤ 0.
F(e(k)) ∗ INn F(e(k))

Then, it can be easily obtained from (4.23) and (4.24) that

h
1V(k) + µ2 V(k) ≤ e(k)T (W ⊗ Ŵ − LC)T Q(W ⊗ Ŵ − LC)

− (1 − µ2 )Q] e(k) + 2FT (e(k))Q(W ⊗ Ŵ − LC)e(k)

+ FT (e(k))QF(e(k))
 T   
e(k) 813 8T23 e(k)
− ε2 (4.25)
F(e(k)) ∗ INn F(e(k))
= ξ2T (k)5̄2 ξ2 (k)

where
 T
ξ2 (k) = eT (k) FT (e(k)) ,
 
5¯ 21 (W ⊗ Ŵ − LC)T Q − ε2 8T
¯
52 = 2 3
∗ Q − ε2 INn

with 5̄21 = (W ⊗ Ŵ − LC)T Q(W ⊗ Ŵ − LC) − (1 − µ2 )Q − ε2 813 .

By noticing (4.22), it follows from (4.20) and the Schur Complement Lemma that
1V(k) + µ2 V(k) ≤ 0, which implies that V(k + h) ≤ (1 − µ2 )V(k + h − 1) ≤
(1 − µ2 )2 V(k + h − 2) ≤ · · · ≤ (1 − µ2 )h V(k). Based on such a fact, one can further
obtain that

λmin {Q}ke(k + h)k22 ≤ V(k + h) ≤ (1 − µ2 )h V(k)

(4.26)
≤ (1 − µ2 )h λmax {Q}ke(k)k22 ,

which means that

kx(k + h) − x̃(k + h)k2 < c1 kx(k) − x̃(k)k2 (4.27)

74 Data-Rate-Constrained State Estimation and Control

q
where c1 = (1−µλ2 ) λ{Q}
h
max {Q}
. Therefore, it is not difficult to verify that there always
min
exists a proper integer h to ensure 0 < c1 < 1. The proof of Lemma 4.2 is complete.

4.2.3 Detectability Analysis

Motivated by the uniform quantization approach proposed in [140,190],
in this chapter, a modified uniform quantization approach is put forward,
which would play an important role in the development of the coding-
decoding procedure. For such a procedure, at each coding instant lh, since
the current decoded state x̂i (lh) of the ith node is inaccessible at this stage,
we introduce an auxiliary state x̄i (lh) which is determined by the previously
decoded state x̂i (lh − 1) and will be defined later, see (4.32). Then, a modified
uniform quantization approach is applied to the error between the estimated
state x̃i (lh) and the auxiliary state x̄i (lh). In this regard, denote by δi (lh) ,
x̃i (lh) − x̄i (lh) the error vector for the ith node at the coding instant lh. In
the following, a brief introduction of the proposed quantization method is
given to facilitate the readers. For presentation convenience, sometimes, the
arguments of a function or a matrix will be omitted in the analysis when no
confusion can arise.
For the given scaling parameters ai > 0 (i = 1, 2, . . . , N) and integer q, we
(j)
can partition the hyperrectangles Bai = {δi ∈ Rn : |δi | ≤ ai , j = 1, . . . , n}
into qn hyperrectangles Ii1i (ai ) × Ii2i (ai ) × · · · × Iini (ai ), where si1 , si2 , . . . , sin ∈
s1 s2 sn
{1, 2, . . . , q} and
 
ij (j) (j) 2ai
I1 (ai ) , δi − ai ≤ δi < −ai + ,
q
 
ij (j) 2ai (j) 4ai
I2 (ai ) , δi − ai + ≤ δi < −ai + ,
q q (4.28)
..
.
 
ij (j) 2ai (j)
Iq (ai ) , δi ai − ≤ δi ≤ ai
q

(j)
where δi is the jth element of the vector δi . As a result, for each Bai , the center
of the hyperrectangle Ii1i (ai ) × Ii2i (ai ) × · · · × Iini (ai ) can be defined as
s1 s2 sn

 T
ηai i (si1 , s2i , . . . , sin ) , bi1 bi2 · · · bin (4.29)

(2sij −1)ai
where bij = −ai + q , j = 1, 2, . . . , n.
Synchronization Control for a Class of Discrete-Time Networks 75

Hence, for any δi ∈ Bai , there exist unique integers si1 , s2i , . . . , sin ∈ {1, 2, . . . , q}
such that δi ∈ Ii1i (ai ) × Ii2i (ai ) × · · · × Iini (ai ), which amounts to the following
s1 s2 sn

√
nai
kδi − ηai i (si1 , si2 , . . . , sni )k2 ≤ . (4.30)
q

ij
In the coding-decoding procedure to be proposed, each segment I i (ai )
sj
(j = 1, 2, . . . , n) of the hyperrectangle Bai determines an integer sji , which is
one component of the codeword. So, the transmitted codeword corresponds
to a certain integer sequence s11 , . . . , s1n , s21 , . . . , s2n , . . . , sN N
1 , . . . , sn , which is
vitally important to the decoding procedure.
Assume that Lemma 4.1 and Lemma 4.2 hold for some proper constants
c0 and c1 , respectively. The following specific form of coding-decoding
procedure is proposed to handle the detectability analysis problem for the
ith node.
Coder for Node i: For δi (lh) , x̃i (lh) − x̄i (lh) ∈ Ii1i (ai (lh)) × Ii2i (ai (lh)) × · · · ×
s1 s2
Iini (ai (lh)) ⊂ Bai (lh) , we have
sn

gi (lh) = [si1 , si2 , . . . , sin ] (4.31)

where x̄i (lh) is defined by

x̄i (0) = 0,
x̄i (k) = x̂i (k), k 6 = lh,
(4.32)
N
X
x̄i (lh) = f (x̂i (lh − 1)) + wij Ŵx̂j (lh − 1) + ui (lh − 1).
j=1

Decoder for Node i:

x̂i (0) = 0,
N
X
x̂i (k + 1) = f (x̂i (k)) + wij Ŵx̂j (k) + ui (k), k 6 = lh − 1, (4.33)
j=1

x̂i (lh) = x̄i (lh) + αi (lh)ηai i (lh) (si1 , s2i , . . . , sin ).

By utilizing the Kronecker product, we reformulate the compact coding-

decoding procedure as follows.
76 Data-Rate-Constrained State Estimation and Control

Coder: For δ(lh) , x̃(lh) − x̄(lh) ∈ I111 (a1 (lh)) × · · · × I11n (a1 (lh)) × I212 (a2 (lh)) ×
s1 sn s1
2 n N 1 Nn
· · · × I 2 (a2 (lh)) × · · · × I N (aN (lh)) × · · · × I N (aN (lh)) ⊂ Ba(lh) , we have
sn s1 sn

g(lh) = [s11 , . . . , s1n , s21 , . . . , s2n , . . . , sN N

1 , . . . , sn ] (4.34)

where Ba(lh) = {x(lh) ∈ RNn : |x(j) (lh)| < a(lh), j = 1, . . . , Nn} and x̄(lh) =
vecN {x̄i (lh)} is defined by

x̄(0) = 0,
x̄(k) = x̂(k), k 6= lh, (4.35)
x̄(lh) = F(x̂(lh − 1)) + (W ⊗ Ŵ)x̂(lh − 1) + u(lh − 1).

Decoder:

x̂(0) = 0,
x̂(k + 1) = F(x̂(k)) + (W ⊗ Ŵ)x̂(k) + u(k), k 6= lh − 1, (4.36)
x̂(lh) = x̄(lh) + α(lh)ηa(lh) (s11 , . . . , snN )

where a(lh) = max1≤i≤N {ai (lh)}, α(lh) = diagN {αi (lh)In } and ηa(lh) (s11 , . . . ,
sN i i i i
n ) = vecN {ηai (lh) (s1 , s2 , . . . , sn )}.
In the following, some definitions are given to facilitate the subsequent
developments.

s0 , sup kxi0 k22 ,

xi0 ∈X0
p (4.37)
a(h) , (2c1 + c0h ) Ns0 ,
p √
E{a((l + 1)h)} , 2cl1 Ns0 (c1 + ch0 ) + α˜ Nnch0 a(lh)
Q
where α̃ = ᾱq + (1 − ᾱ) and ᾱ = iN=1 (1 − ᾱi ).
The following lemma shows that the coding-decoding protocol developed
in this chapter is well posed. That is, the decoding condition E{x̃(lh)−x̄(lh)} ∈
Ba(lh) holds for all l = 1, 2, . . ..

Lemma 4.3: The coding-decoding procedure (4.34)–(4.36) satisfies the following

constraint
E{kx̃(lh) − x̄(lh)k∞ } ≤ a(lh), l = 1, 2, . . . . (4.38)

Proof: This lemma can be proved by using the mathematical induction. First, for
l = 1, by considering Lemma 4.1, Lemma 4.2 and the property of vector norm, we
have
Synchronization Control for a Class of Discrete-Time Networks 77

E{kx̃(h) − x̄(h)k2 } ≤ E{kx̃(h) − x(h)k2 } + E{kx(h) − x̄(h)k2 }

≤ c1 ke(0)k2 + c0 kx(h − 1) − x̂(h − 1)k2 (4.39)
p
≤ 2c1 Ns0 + ch0 kx(0) − x̂(0)k2 ,

which guarantees
E{kx̃(lh) − x̄(lh)k∞ } ≤ a(lh). (4.40)
√
Subsequently, assuming that E{kx̃(jh) − x̄(jh)k∞ } ≤ NnE{a(jh)} for all j = 2,
. . . , l, we arrive at

E{kx̃((l + 1)h) − x̄((l + 1)h)k2 }

≤ E{kx̃((l + 1)h) − x((l + 1)h)k2 }
+ E{kx((l + 1)h) − x̄((l + 1)h)k2 } (4.41)
≤ c1l+1 ke(0)k2 + c0 E{kx(lh + h − 1) − x̂(lh + h − 1)k2 }
p
≤ 2cl1+1 Ns0 + ch0 E{kx(lh) − x̂(lh)k2 }.

Furthermore, taking the packet dropout phenomenon into account, we obtain

kx(lh) − x̂(lh)k2 = kx(lh) − x̄(lh) − α(lh)ηa(lh) (s11 , . . . , sN

n )k2
≤ kx(lh) − x̃(lh)k2 + kx̃(lh) − x̄(lh) (4.42)
− α(lh)ηa(lh) (s11 , . . . , sN
n )k2
p
≤ 2c1 Ns0 + kx̃(lh) − x̄(lh) − α(lh)ηa(lh) (s11 , . . . , sN
l
n )k2 .

Since the codeword loss occurs in a random way for each node and is governed
by a set of mutually independent Bernoulli distributed sequence αi (lh), the term
kx̃(lh) − x̄(lh) − α(lh)ηa(lh) (s11 , . . . , sN
n )k2 in (4.42) can be further rewritten as

kx̃(lh) − x̄(lh) − α(lh)ηa(lh) (s11 , . . . , sN

n )k2
N
Y
= αi (lh) x̃(lh) − x̄(lh) − ηa(lh) (s11 , . . . , snN )
2
i=1
N
X N
Y
+ (1 − αi (lh))αj (lh) x̃(lh) − x̄(lh)
6
j=1 i=1,i=j
(4.43)

− diag{In , . . . , In , 0, In , . . . , In }ηa(lh) (s11 , . . . , sN

n)
| {z } | {z } 2
j−1 N−j
N
Y
+ ··· + (1 − αi (lh))kx̃(lh) − x̄(lh)k2
i=1
78 Data-Rate-Constrained State Estimation and Control

where the first term on the right-hand side of (4.43) implies the case that there is
no packet dropout occurring during the codeword transmissions, the second term
describes that only one node is subject to the codeword loss and the last term indicates
that all the nodes lose the codewords at the transmission stage.
For presentation convenience, only some cases of packet dropouts for the dynamical
network are described in (4.43). Specifically, without loss of generality, we assume
that the first m nodes suffer from the packet dropouts where m ∈ {1, 2, . . . , N}. Then,
we have

x̃(lh) − x̄(lh) − diag{0, . . . , 0, In , . . . , In }ηa(lh) (s11 , . . . , sN

n)
| {z } | {z } 2
m N−m

˜
= x(lh) ¯
− x(lh) − ηa(lh) (s11 , . . . , snN )

+ diag{In , . . . , In , 0, . . . , 0}ηa(lh) (s11 , . . . , sN

n)
| {z } | {z } 2
m N−m

≤ x̃(lh) − x̄(lh) − ηa(lh) (s11 , . . . , sN

n)
(4.44)
2

+ [ηa1T mT
1 (lh) , . . . , ηam (lh) , 0, . . . , 0]
T
| {z } | {z } 2
m N−m
 
√ a(lh) √ a(lh)
≤ Nn + Nn a(lh) −
q q
√
= Nna(lh).

It is not difficult to verify that, if we consider the case of packet dropouts for arbitrary
m nodes, the above result still holds. Accordingly, it can be derived from (4.42)–(4.44)
that
p √ α¯
E{kx(lh) − x̂(lh)k2 } ≤ 2cl1 Ns0 + Nn E {a(lh)}
q (4.45)
√
+ Nn (1 − ᾱ) E {a(lh)} .
Thus, it can be concluded from (4.37), (4.41) and (4.45) that

E kx̃((l + 1)h) − x̄((l + 1)h)k2 ≤ a((l + 1)h), (4.46)

which means

E kx̃((l + 1)h) − x̄((l + 1)h)k∞ ≤ a((l + 1)h). (4.47)

Consequently, it follows that the Lemma 4.3 holds for all l ≥ 1, and the proof of this
lemma is complete.
Synchronization Control for a Class of Discrete-Time Networks 79

The following theorem offers a sufficient condition to guarantee the

detectability of the dynamical network (4.10) by utilizing the proposed
coding-decoding protocol.

Theorem 4.1: The dynamical network (4.10) is detectable with the coding-decoding
procedure (4.34)-(4.36) if the inequality
 
ᾱ √
+ 1 − α¯ Nnch0 < 1 (4.48)
q

holds subject to (4.15) and (4.20) for some positive integer q, where c0 and h satisfy
Lemma 4.1 and Lemma 4.2, respectively.

Proof: It follows from (4.37), (4.48) and 0 < c1 < 1 that liml→+∞ E{a(lh)} = 0,
which indicates
lim E{kx̃(lh) − x̄(lh)k2 } = 0. (4.49)
l→∞

Then, it can be easily seen from (4.45) that, at each coding instant lh, we have

lim E{kx(lh) − x(
ˆ lh)k2 } = 0. (4.50)
l→∞

On the other hand, when lh < k < (l + 1)h, noticing that x(k) and xˆ(k) can be
viewed as two solutions of (4.10), one finds E{kx(k) − x̂(k)k2 } ≤ ck0−lh E{kx(lh) −
x̂(lh)k2 } by Lemma 4.1, which implies that E{kx(k) − x̂(k)k2 } is also bounded at
non-coding instants. Therefore, it follows immediately from the above analysis that

lim E{kx(k) − x̂(k)k2 } = 0, (4.51)

k→∞

which ends the proof.

Remark 4.2: In Theorem 4.1, a criterion has been established to ensure the
detectability of the dynamical network (4.10). It is not difficult to find that all the
information (including the system matrices, the coding period, the size of coding
alphabet and the statistics characteristics of packet dropouts) are reflected in the
condition (4.48). Hence, the detectability of the dynamical network (4.10) depends
heavily on the combined influences of those factors mentioned above. Let us now
evaluate the impact on the performance of the proposed coding-decoding
√
protocol
α¯ Nnc0h
from the following three aspects. 1) From (4.48), one has q > √ , which
¯ Nnch0
1−(1−α)
√
(α¯ Nnch0 )Nn
requires that the size of the coding alphabet H should satisfy χ> √ .
(1−(1−α)¯ Nnc0h )Nn

√ when ᾱ = 1, namely, there is no packet dropout occurring,

Especially, √ (4.48) reduces
to q > Nnc0h , then the size χ is required to satisfy χ > ( Nnc0 )Nnh . 2) The
80 Data-Rate-Constrained State Estimation and Control

√
q Nnc−h −1
inequality (4.48) also indicates 1 − ᾱ < 0
q−1 . Recalling the definition of ᾱ, it
is noticed that the term 1 − ᾱ can be regarded as the total packet dropout probability
of the whole network. Therefore, it is√easily seen that the upper bound of admissible
q Nnc−h −1
0
data packet dropout probability is q−1 . 3) For a given size χ of the coding
alphabet H, with increased packet dropout probabilities ᾱi , the ᾱ becomes smaller
and thus inevitably weakens the feasibility of the inequality (4.48) and even leads to
the possible divergence of the coding-decoding algorithm. Accordingly, this gives a
trade-off between the packet dropout probabilities and the convergence of the proposed
algorithm.

4.2.4 Synchronization Control

Based on the established results of the detectability analysis problem in the
last subsection, now let us tackle the decoder-based synchronization control
problem for the dynamical network (4.10). Due to the existence of control
input error induced by the decoding error, the ISS analysis approach is
employed.
For the convenience of the readers, let us recall some basic notions and
necessary foundations on the ISS theory before driving our main results.
Consider the following discrete-time nonlinear system

x(k + 1) = f (x(k), u(k)) (4.52)

where x(k) ∈ Rn is the system state vector, u(k) ∈ Rp is the control input and
f (·, ·) : Rn × Rp 7→ Rn is a continuous function satisfying f (0, 0) = 0.
First, let us introduce some important function classes which will be used
in the subsequent developments. K, K∞ and KL denote the different classes
of functions. A function γ (·) : R+ 7 → R+ belongs to class K if it is a continuous
strictly increasing function with γ (0) = 0 and belongs to class K∞ if γ ( · ) ∈ K
with γ (r) 7 → ∞ as r 7 → ∞. Also, a function σ (s, k) : R+ × R+ 7 → R+ belongs
to class KL if, for each fixed k, the function σ (·, k) ∈ K, and for each fixed s,
the function σ (s, ·) is decreasing and σ (·, k) 7 → 0 as k 7→ ∞. A function is said
to be a K (K∞ or KL) class function if it belongs to the class K (K∞ or KL).

Definition 4.3: The system (4.52) is said to be input-to-state stable if there exist a
KL class function β(·, ·) and a K class function γ ( · ) such that the evolution of the
system state x(k) satisfies

kx(k)k2 ≤ β(kx(0)k2 , k) + γ (ku(k)k∞ ) (4.53)

for ∀k ≥ 0 and ∀x(0) ∈ Rn , where ku(k)k∞ , supk {ku(k)k2 }.

Synchronization Control for a Class of Discrete-Time Networks 81

Remark 4.3: It is clear from (4.53) that the bound of the state x(k) is closely
associated with the initial condition x(0) and the system input u(k). To be specific, if
the system input u(k) is bounded, then the state evolution trajectory x(k) is naturally
bounded. Also, it follows from the properties of functions β(·, ·) and γ ( · ) that the
state x(k) is ultimately bounded as k increases. In addition, as pointed out in [58], the
ISS property further implies the case of “converging-input” to “converging-state”,
that is, the system state x(k) converges to 0 if the input u(k) achieves 0 as k 7→ ∞.

To proceed, we present the following lemma.

Lemma 4.4: [58] The nonlinear discrete-time system (4.52) is said to be input-to-
state stable if there exist a positive definite function V(k, x(k)) : [0, +∞) × Rn 7→ R
(called an ISS-Lyapunov function), three K∞ class functions α1 (·), α2 (·) and α3 (·),
and a K class function ̟ ( · ) such that the following two inequalities

α1 (kx(k)k2 ) ≤ V(k, x(k)) ≤ α2 (kx(k)k2 ), (4.54)

V(k + 1, x(k + 1)) − V(k, x(k)) ≤ −α3 (kx(k)k2 ) + ̟ (ku(k)k2 ) (4.55)

hold for all x(k) ∈ Rn and u(k) ∈ Rp . Furthermore, if the above two inequalities
are met simultaneously, then the functions β(·, ·) and γ ( · ) in Definition 4.3 can be
selected as β(·, k) = α1−1 (φ k α2 ( · )) (0 < φ < 1) and γ ( · ) = α1−1 (α2 (α3−1 (̟ ( · )))),
respectively, where α1−1 ( · ) expresses the inverse function of the monotone function
α1 ( · ) and so does α3−1 ( · ).

Now, we rewrite the decoder-based control protocol (4.8) with the

following compact form:

u(k) = K̄c (x̂(k) − s̄(k)) (4.56)

where K̄c , IN ⊗ Kc and s̄(k) , 1N ⊗ s(k).

Next, let wi (k) , x̂i (k) − xi (k) be the decoding error vector for the ith node
and set w(k) , vecN {wi (k)}. In light of (4.56) and noting x̂(k) = w(k) + x(k),
the closed-loop system (4.12) can be rewritten as

es (k + 1) = (K̄c + (W ⊗ Ŵ))es (k) + F(es (k)) + K̄c w(k). (4.57)

Definition 4.3 and Lemma 4.4 reveal the relationship between the system
state evolution and the bound of inputs. On the other hand, based on the
obtained detectability analysis results, we have that the decoding error w(k)
is bounded. In this sense, w(k) can be viewed as a bounded input of (4.57) and
therefore the introduced ISS theory can be utilized to investigate the coding-
decoding-based synchronization control problem.
82 Data-Rate-Constrained State Estimation and Control

Theorem 4.2: Under the condition in Theorem 4.1, the dynamical network (4.10)
can be synchronized by the decoder-based controller (4.56) if there exist a positive
definite matrix R = IN ⊗ R0 > 0, a matrix Y = IN ⊗ Y0 and a positive scalar ε3 > 0
satisfying
 
−R − ε3 813 −ε3 8T23 0 R̄T
 ∗ −ε3 INn 0 R <0
53 = 
 (4.58)
∗ ∗ −R YT 
∗ ∗ ∗ −R

where R̄ = (R(W ⊗ Ŵ) + Y). Moreover, the desired controller gain matrix is given as

Kc = R−1
0 Y0 . (4.59)

Proof: Choosing the ISS-Lyapunov function as V(k) = eTs (k)Res (k) and calculating
its difference along the trajectory of (4.57), one obtains
 T
V(k + 1) − V(k) ≤ (K̄c + (W ⊗ Ŵ))es (k) + F(es (k)) + K̄c w(k) R
 
× (K̄c + (W ⊗ Ŵ))es (k) + F(es (k)) + K̄c w(k)
 T  
es (k) 813 82T3
− esT (k)Res (k) − ε3 (4.60)
F(es (k)) ∗ INn
 
es (k)
×
F(es (k))
= ξ3T (k)5̄3 ξ3 (k) + wT (k)Rw(k)
where
 T
ξ3 (k) = eTs (k) FT (es (k)) wT (k) ,
 
5 ¯ 31 KT R − ε3 8T KT K̄c
33
5̄3 =  ∗ R − ε3 INn K¯ c 
∗ ∗ T
K̄c RK̄c − R

with 5̄31 = KT RK − ε3 833 and K = K̄c + W ⊗ Ŵ.

By the Schur Complement and (4.59), inequality (4.58) means 5 ¯ 3 < 0,
which yields that V(k + 1) − V(k) ≤ −λmin {−5̄3 }kes (k)k22 + λmax {R}kw(k)k22 .
Choosing α1 (kes (k)k2 ) = λmin {R}kes (k)k22 , α2 (kes (k)k2 ) = λmax {R}kes (k)k22 ,
α3 (kes (k)k2 ) = λmin {−5̄3 }kes (k)k22 and ̟ (kw(k)k2 ) = λmax {R}kw(k)k22 , it is
inferred from Lemma 4.4 that the synchronization
q error system (4.57) is input-to-
{ R} k
state stable. So, let us choose β(kes0 k2 , k) = λλmax
min {R}
φ kes0 k2 and γ (kw(k)k2 ) =
r
λ2max {R}
cλ {R}λ {−5̄ }
kw(k)k2 , where 0 < c < 1. Then we have from Definition 4.3 that
min min 3
Synchronization Control for a Class of Discrete-Time Networks 83

s
λmax {R} k
kes (k)k2 ≤ φ kes0 k2
λmin {R}
s (4.61)
λ2max {R}
+ kw(k)k2 .
cλmin {R}λmin {−5¯ 3}

Taking the mathematical expectation of both sides of (4.61), it can be

immediately found from the detectability analysis result (4.51) and the property of
“converging-input” to “converging state” discussed in Remark 4.3 that
limk→∞ E{kes (k)k2 } = 0, which completes the proof.

4.3 An Illustrative Example

In this section, two simulation examples are presented to demonstrate
the detectability and the effectiveness of the proposed decoder-based
synchronization control protocol for the dynamical network (4.1).
Consider a dynamical network (4.1) with three nodes which are connected
according to the following coupling configuration matrix

−0.2 0.1 0.1

 

W =  0.1 −0.2 0.1 

0.1 0.1 −0.2

and the inner-coupling matrix Ŵ = diag{0.1, 0.1}.

The nonlinear functions are assumed to be
 
−0.5xi1 (k) + tanh(0.65xi1 (k)) − 0.15xi2 (k)
f (xi (k)) = .
1.1xi2 (k) − tanh(0.95xi2 (k))

Then, it is easy to see that the nonlinear function f ( · ) satisfies the sector
bounded condition (4.2) with
   
−0.5 −0.15 0.15 −0.15
U1 = , U2 = .
0 1.1 0 0.25

The available measurements of the dynamical network described above are

modeled in (4.1) with the following parameters
     
C1 = 0.41 0.52 , C2 = 0.33 0.44 , C3 = 0.32 0.60 .
84 Data-Rate-Constrained State Estimation and Control

In the following, we shall deal with the detectability as well as

synchronization control problems for the dynamical network (4.1) with given
parameters.

Example 4.1: In this example, according to our established criterion, we shall test
the detectability problem for the dynamical network (4.1). In this case, without
loss of generality, it is assumed that u(k) ≡ 0. Choosing µ1 = 0.3 and
solving the matrix inequality (4.15) by the Matlab software, we have that (4.16)
holds with c0 = 1.2701. For the given scalar µ2 = 0.1 and the coding
period h = 3, c1 = 0.9435 is obtained such that (4.21) holds. Moreover,
 T
the estimator gain matrices Li are obtained as L1 = −0.3476 0.8462 ,
 T  T
L2 = −0.4161 1.0285 and L3 = −0.3223 0.8768 . According to Theorem
4.1, the packet probability ᾱi (i = 1, 2, 3) is taken as 0.1 for q = 20.
To validate our analysis results and make our simulation non-trivial, we
consider the above dynamical network which is set to be unstable. The simulation
results are shown in Figures 4.1–4.3, where Figures 4.1–4.2 plot the actual
states and their decoded values for all the nodes of the network (4.1) and
Figure 4.3 shows the corresponding decoding errors wij (k) (i = 1, 2, 3;
j = 1, 2). It is noticed from Figure 4.3 that the errors between the actual states of
the whole network and their decoded states asymptotically approach zero. Therefore,

50

−50
0 10 20 30 40 50 60 70 80 90

50

−50
0 10 20 30 40 50 60 70 80 90

200

−200
0 10 20 30 40 50 60 70 80 90

FIGURE 4.1
State trajectories xi1 (k) and their decoded values x̂i1 (k) (i = 1, 2, 3).
Synchronization Control for a Class of Discrete-Time Networks 85

500

−500
0 10 20 30 40 50 60 70 80 90

500

−500
0 10 20 30 40 50 60 70 80 90

−1000

−2000
0 10 20 30 40 50 60 70 80 90

FIGURE 4.2
State trajectories xi2 (k) and their decoded values x̂i2 (k) (i = 1, 2, 3).

15

10

5
Decoding errors

−5

−10

−15

−20
0 10 20 30 40 50 60 70 80 90
k/time step

FIGURE 4.3
Decoding errors wij (k) (i = 1, 2, 3; j = 1, 2).
86 Data-Rate-Constrained State Estimation and Control

the desired detectability performance of the addressed dynamical networks is well

attained.

Remark 4.4: For comparison, if we use the infinity-vector norm-based uniform

quantization approach that has been widely used in the existing literature, we can
calculate that c0 = 3.1110 and c1 = 2.3111 for this example, which violates the
condition 0 < c1 < 1 and the feasibility of the coding-decoding procedure cannot
be guaranteed. Such a comparison result embodies the superiority of our proposed
modified uniform quantization approach.

Example 4.2: In the second example, we consider the decoder-based synchroniza-

tion control problem for the dynamical network (4.1). The control input signal is
taken as the form in (4.8) for each node. By solving (4.58) in Theorem 4.2, we obtain
the desired controller parameter as
 
1 0.1935 0.1521
Kc = R−
0 Y0 = .
−0.0003 −0.6601

As stated in Theorem 4.2, the considered dynamical network can be synchronized

with the designed controller parameters given above. Simulation results are
presented in Figures 4.4–4.5. From Figure 4.4, it can be observed that the state
response curves of all nodes diverge from the isolate node drastically when there is no

20

−20

−40

−60
0 10 20 30 40 50 60

1000

500

−500
0 10 20 30 40 50 60

FIGURE 4.4
State trajectories xij (k) of the uncontrolled nodes i (i = 1, 2, 3; j = 1, 2).
Synchronization Control for a Class of Discrete-Time Networks 87

20

−20

−40

−60
0 10 20 30 40 50 60

1000

500

−500
0 10 20 30 40 50 60

FIGURE 4.5
State trajectories xij (k) of the controlled nodes i (i = 1, 2, 3; j = 1, 2).

control signal to the dynamical network. Figure 4.5 shows that all the state response
curves of the controlled network can synchronize to the isolate node according to the
proposed control protocol. Therefore, the simulation examples have confirmed our
theoretical analysis very well.

4.4 Summary
This chapter has dealt with the synchronization control problem for a class
of dynamical networks in a coding-decoding framework. By employing a
modified uniform quantization approach, a set of coder-decoder-controller
pairs has been designed to comply with the digital communication
fashion. With the help of ISS theory, sufficient conditions have been
derived for the closed-loop network to achieve the desired synchronization
performance and the explicit expression of the controller gain matrix has
been characterized in terms of the solution to certain LMIs. Finally, two
illustrative examples have highlighted the detectability of the addressed
dynamical network and the effectiveness of the synchronization control
protocol presented in this chapter.
5
Observer-Based Consensus Control for
Discrete-Time Multi-Agent Systems with
Coding-Decoding Communication Protocol

The past few decades have witnessed particular research interests in the
problems of collective behaviors (e.g. consensus, flocking and swarming)
of networked multiple autonomous agents for their broad applications on
various aspects of engineering systems including cooperative control for
multi-robot systems, coordination of distributed sensor networks, formation
control of unmanned air vehicles and satellite flying, to name just a few. In
order to reach the goal of the coordination control for the individual agent in
a networked environment, a pivotal issue is how to develop certain simple
yet efficient distributed control protocols (or algorithms) only by sharing
the local information between the adjacent agents. Because of its theoretical
importance in distributed computing and application potentials in practical
engineering, the consensus problem has long been an active research topic
with a great many results reported in the literature.
It should be mentioned that almost all existing literature on the consensus
control problems for MASs has been based on the assumption that the signals
can be directly transmitted in an analog manner with infinite precision. Such
an assumption is, however, sometimes too stringent because the signals are
usually digitalized before being sent in a networked environment. As opposed
to the traditional analog communication scheme, the digital communication
strategy enjoys the distinct merits of well interference rejection capability,
high security as well as reduced power consumption, and has therefore
become increasingly popular in nowadays prevalent networked control
systems (NCSs). Roughly speaking, there are four steps in an analog-to-
digital (A/D) conversion, namely, sampling, reservation, quantization and
encoding. It is worth pointing out that, compared with sampling and
quantization issues, the signal encoding problem has received much less
research attention in the context of NCSs.
In NCSs, the limited communication resource (e.g. channel bandwidth)
is well known to be the main cause for quality deterioration of signal
transmissions and performance degradation of control systems . Especially,

DOI: 10.1201/9781003534853-5 88
Observer-Based Consensus Control 89

in networked distributed systems (e.g. MASs, sensor networks and complex

networks), the subsystems usually share the communication network cables
or wireless media, and it is therefore a critical issue to make adequate use
of the limited communication capacity through exploiting energy-saving
communication protocols.
To this end, another important communication protocol, namely, coding-
decoding communication protocol (CDCP), which sends the symbolic data
through communication channels and therefore occupies less resource as
compared to the original data, has largely been overlooked despite its
popularity in the environment of digital communication.
Motivated by the above discussion, in this chapter, our objective is to
design an observer-based consensus controller for a class of discrete-time
networked MASs with CDCPs. In doing so, three challenges are identified
as follows: 1) how to establish a unified theoretical framework that takes
both the observer-based control scheme and the CDCP into consideration?
2) how to obtain sufficient conditions for the existence of the desired easy-to-
implement consensus controllers? and 3) how to examine the impact from the
directed communication topology and the proposed CDCP on the consensus
performance? To address these three identified challenges, we shall first focus
our attention on the analysis/synthesis issues of the CDCP for the discrete-
time MASs, which is of vital importance to the further investigation on
the consensus performance. Then, based on the designed communication
protocol, the observer-based consensus controller is obtained to guarantee
the expected consensus performance of the closed-loop MAS.

5.1 Problem Formulation

In this chapter, it is assumed that the MAS has M agents which communicate
with each other according to a fixed network topology represented by a
directed graph G = (V , E , A ) with the set of agents V = {1, 2, . . . , M}, the set
of edges E ⊆ V × V , and the weighted adjacency matrix A = [aij ]M×M with
nonnegative adjacency element aij . An edge of G is denoted by the ordered
pair (i, j). The adjacency elements associated with the edges of the graph
are positive, that is, aij > 0 ⇐⇒ (i, j) ∈ E , which means that agent i can
obtain information from agent j. The neighborhood of agent i is denoted by
Ni = {j ∈ V : (j, i) ∈ E }. An element of Ni is called a neighbor of agent i.
In this chapter, self-edges (i, i) are not allowed, that is, (i, i) ∈/ E for any
P
i ∈ V . The in-degree of agent i is defined as degiin = j∈Ni aij . Denote the
diagonal matrix D = diagM {degiin }. The Laplacian matrix of G is given by
LG = [lij ]M×M = D − A .
90 Data-Rate-Constrained State Estimation and Control

Consider a class of discrete-time MASs consisting of M agents where the

dynamics of the ith agent is described by

xi (k + 1) = Axi (k) + Bui (k)
y (k) = Cxi (k) (5.1)
 i
xi (0) = xi0

where xi (k) ∈ Rn , yi (k) ∈ Rm and ui (k) ∈ Rp are, respectively, the state vector,
the measurement output and the control input of the ith agent. xi0 ∈ Rn is a
given initial condition. A, B and C are known real matrices with appropriate
dimensions.
In most existing results regarding the MASs, a common assumption is that
the relative full state information for each agent is available from itself and
its adjacent agents, which is defined by
X
ξi (k) , aij (xj (k) − xi (k))
(5.2)
j∈Ni

P
with the initial condition ξi (0) = si0 , where si0 = j∈Ni aij (xj0 − xi0 ). However,
from an engineering point of view, the relative full state information is often
unavailable in reality. Hence, for agent i, it is assumed that only the relative
measurement output of neighboring agents is obtained as follows:
X
ỹi (k) , aij (yj (k) − yi (k)).
(5.3)
j∈Ni

5.1.1 Traditional Observer Structure

With the available measurement output ỹi (k), we would be able to estimate
the relative full state information ξi (k) by using the following Luenberger-like
type observer for agent i:
X
ξ̂i (k + 1) = Aξ̂i (k) + B aij (uj (k) − ui (k)) + L(ỹi (k) − Cξ̂i (k))
(5.4)
j∈Ni

with the initial condition ξ̂i (0) = ŝi0 , where ŝi0 is a known vector. Moreover,
the control input ui (k) in (5.4) is given as ui (k) = Kξ̂i (k), where K is the
controller gain to be determined.

5.1.2 CDCP-Based Observer Structure

In order to realize the digital transmission, we introduce a CDCP between
the observer and the controller for each agent. To be more specific, a general
form of coder-decoder scheme for the MAS (5.1) is given as follows:
Observer-Based Consensus Control 91

Coder for Agent i:

gi (dh) = Fid (ξ̂i (h), ξ̂i (2h), . . . , ξ̂i (dh)), (5.5)

Decoder for Agent i:

4i (dh) = Gid (gi (h), gi (2h), . . . , gi (dh)) (5.6)

for d = 1, 2 . . ., where gi (dh) is the encoded data (codeword) generated

at the coding instant dh by coder i and 4i (dh) is defined by 4i (dh) ,
[ξ̌iT (dh) ξˇiT (dh + 1) · · · ξˇiT ((d + 1)h − 1)]T with ξ̌i (k) being the decoded
state of ξ̂i (k) for k ∈ [dh, (d + 1)h). Fdi ( · ) and Gid ( · ) are, respectively, the coder
and decoder functions to be designed for agent i.
Instead of using the estimated state ξ̂i (k) directly, only the decoded state
ξ̌i (k) is available at the controller side due to the introduction of the CDCP,
and thus the following decoder-based consensus control scheme is employed
for agent i:

ui (k) = Kξ̌i (k) (5.7)

where ξ̌i (k) ∈ Rn is the decoded signal generated by decoder i. Based on such
a fact, the control input ui (k) = Kξ̂i (k) in the observer (5.4) should be replaced
by (5.7). On the other hand, it should be pointed out that, owing to the
constraint of the network bandwidth, it is unrealistic and impossible to feed
the decoded signal ξ̌i (k) back to the state estimator side, which means that
the decoded state ξ̌i (k) cannot be directly utilized to construct state observer
(5.4). Therefore, an auxiliary control input ŭi (k) defined in the next section
(see (5.31)) is introduced to establish the following observer:
X
ξ̂i (k + 1) = Aξ̂i (k) + B aij (ŭj (k) − ŭi (k)) + L(ỹi (k) − Cξ̂i (k))
(5.8)
j∈Ni

subject to the initial condition ξ̂i (0) = ŝi0 , where ŝi0 is a known vector. As
will be shown in the later developments, the auxiliary control input ŭi (k)
equals ui (k) ∀k ≥ 0. Therefore, by letting ξ̃i (k) , ξi (k) − ξ̂i (k), it follows
from (5.2) and (5.8) that the following estimation error dynamics for agent i
is derived:

ξ̃i (k + 1) = (A − LC)ξ̃i (k). (5.9)

The following lemma is needed in deriving our main results.

92 Data-Rate-Constrained State Estimation and Control

Lemma 5.1: ([135]). Given a matrix A = [aij ]n×n , where aii ≤ 0, aij > 0 (∀i 6= j)
P
and nj=1 aij = 0 for each i. Then, A has at least one zero eigenvalue and all of the
nonzero eigenvalues are in the open left half plane. Furthermore, A has exactly one
zero eigenvalue if and only if the directed graph associated with A has a spanning tree.

For brevity, we set

x(k) = colM {xi (k)}, w(k) = colM {wi (k)},

v(k) = colM {vi (k)}, ξ(k) = colM {ξi (k)},

ξ̂ (k) = colM {ξ̂i (k)}, ξ̃ (k) = colM {ξ̃i (k)}, (5.10)

ξ̌ (k) = colM {ξ̌i (k)}, u(k) = colM {ui (k)},

 
L¯G = LGT LG = l̄ij M×M .

Subsequently, from (5.2) and (5.7), the augmentation of the dynamics ξi (k)
can be obtained with the following form:

ξ(k + 1) = (IM ⊗ A)ξ(k) − (LG ⊗ BK)ξ̌ (k). (5.11)

Moreover, denote ξE (k) , ξ(k) − ξ̌ (k) and rewrite u(k) , (IM ⊗ K)(ξ(k) − ξE (k)).
By augmenting the system state xi (k), we obtain the closed-loop system as
follows:

x(k + 1) = (IM ⊗ A − LG ⊗ BK)x(k) − (IM ⊗ BK)ξE (k). (5.12)

Similarly, we reformulate the observation error dynamics (5.9) as

follows:

ξ̃ (k + 1) = (IM ⊗ (A − LC))ξ̃ (k). (5.13)

PM
Next, introduce the weighted average state for all agents as x̄(k) = i=1
η1i xi (k) = (η1T ⊗ In )x(k), where η1i is the ith element of η1 = colM {η1i } ∈ RM
with η1 being the unique nonnegative left eigenvector of LG associated with
P
eigenvalue 0 satisfying M i=1 η1i = 1. Therefore, it derives from (5.12) and
η1T LG = 0 that

x̄(k + 1) = (η1T ⊗ In )x(k + 1)

= A(η1T ⊗ In )x(k) − (η1T ⊗ BK)ξE (k) (5.14)
= Ax̄(k) − (η1T ⊗ BK)ξE (k).
Observer-Based Consensus Control 93

By setting x̃i (k) , xi (k) − x̄(k) and x̃(k) , colM {x̃i (k)}, it is inferred from (5.12)
and (5.14) that the deviation from each agent’s state xi (k) to the weighted
average state x̄(k) is calculated with a compact form as follows:

x̃(k + 1) = x(k + 1) − (1M ⊗ In )x̄(k + 1)

= (IM ⊗ A − LG ⊗ BK)x(k) − (IM ⊗ BK)ξE (k)
− (1M ⊗ In )[Ax̄(k) − (ηT ⊗ BK)ξE (k)]
1 (5.15)
= (IM ⊗ A − LG ⊗ BK)x̃(k) − (H ⊗ BK)ξE (k)
= (IM ⊗ A − LG ⊗ BK)x̃(k) + Hδ(k),

where H = [hij ]M×M with hij = 1 − η1j for i = j and hij = −η1j for i 6 = j.
Subsequently, along the same line as in [84], denote 3 by the Jordan form
associated with the Laplace matrix LG , for example P−1 LG P = 3, where
P = [p1 p2 . . . pM ] ∈ CM×M and P−1 = [q1 q2 . . . qM ]T ∈ CM×M are
nonsingular matrices with pi ∈ CM and qi ∈ CM being the M-
dimensional complex vectors for i = 1, 2, . . . , M. Different from the case of
undirected graph G where LG is symmetric and 3 is a diagonal matrix,
for a directed graph G , LG could be a non-symmetric matrix whose
eigenvalues are complex and 3 is no more diagonal but with the form
3 = diag{31 , 32 , . . . , 3r }, where 3i ∈ CMi ×Mi is a Jordan block
 
λi 1 0 0

0 .. .. 
. . 0
3i =   . (5.16)

0 .. 
0 . 1
0 0 0 λi M ×M
i i

In (5.16), the diagonal elementsPλi are the eigenvalues of LG with multiplicity

Mi ≤ M for i = 1, 2, . . . , r and ir=1 Mi = M. Specially, when r = M, we have
Mi = 1 and 3 reduces to a diagonal matrix. It is assumed that the directed
graph G contains a directed spanning tree and, therefore, according to
Lemma 5.1, we denote 31 by the simple eigenvalue 0 with the corresponding
eigenvectors p1 = 1M and q1 = η1 , which also implies M1 = 1.
Based on the above discussion, we define δ(k) , (P−1 ⊗ In )x(k) ˜ ∈ CMn and
W , [q2 · · · qM ]T ∈ C(M−1)×M , and then make a partition to δ(k) as δ(k) ,
[ζ1T (k) δ˜ T (k)]T , where ζ1 (k) ∈ Rn contains the first nth entries of δ(k) and δ̃(k) is
of the form δ( ˜ k) , [δ˜ T (k), . . . , δ̃rT (k)]T with δ̃i (k) ∈ CMi n for i = 2, . . . , r. Bearing
2
in mind the facts that η1T H = 0 and W H = W, we immediately have that
(
ζ1 (k + 1) = (η1T ⊗ In )x̃(k + 1) = 0 (5.17a)
δ̃(k + 1) = (IM−1 ⊗ A − 3̃ ⊗ BK)δ̃(k) − (W ⊗ BK)ξ (k) E (5.17b)

where 3̃ = diag{32 , . . . , 3r }.
94 Data-Rate-Constrained State Estimation and Control

To proceed, we introduce the following definitions and lemmas which are

useful for the later developments.

Definition 5.1: The discrete-time dynamical system (5.11) is said to be detectable

if there exist families of coder-decoder pairs (5.5) and (5.6) with a coding alphabet H
of size χ such that

lim |ξ(k) − ξ̌ (k)| = 0. (5.18)

k→∞

Definition 5.2: Suppose that the directed graph G of the communication network
contains a directed spanning tree. The discrete-time MASs (5.1) is said to reach the
asymptotic consensus if

lim |xi (k) − xj (k)| = 0 (5.19)

k→∞

holds for ∀i, j ∈ V .

Lemma 5.2: Suppose that the directed graph G of the communication network
contains a directed spanning tree. The MAS (5.1) reaches the asymptotic consensus
if the component δ̃i (k) in (5.17b) satisfies

lim |δ˜i (k)| = 0 (5.20)

k→∞

for i = 2, . . . , r.
Pi−1
Proof: First, define p̃1 = p1 and p̃i , [pαi +1 pαi +2 · · · pαi +Mi ] with αi = s=1 Ms
| {z }
Mi
for i = 2, . . . , r. Thus, the matrix P can be rewritten as P = [˜p1 p˜ 2 · · · p˜ r ], and
then it follows from δ(k) = (P−1 ⊗ In )x̃(k) and ζ1 (k) = 0 that

lim |x̃(k)| = lim |(P ⊗ In )δ(k)|

k→∞ k→∞
r
X
= lim (p1 ⊗ In )ζ1 (k) + (p̃i ⊗ In )δ̃i (k)
k→∞
i=2
r
X
≤ lim |p̃i ⊗ In |F |δ˜i (k)|.
k→∞
i=2

Consequently, it follows immediately from (5.20) that the conclusion

limk→∞ |x̃(k)| = 0 is true, and the proof is complete.

Lemma 5.3: Let the positive scalar ǫ be given. If there exists a positive definite
matrix R > 0 satisfying

50 , (IM ⊗ A)T R(IM ⊗ A) − ǫ 2 R < 0, (5.21)

Observer-Based Consensus Control 95

then the following inequality

|ϕ(k + 1) − φ(k + 1)| ≤ c0 |ϕ(k) − φ(k)| (5.22)

is satisfied, where c0 > 0 is a positive scalar, and ϕ(k) and φ(k) are two trajectories
of system (5.11).

Proof: First, denoting z(k) , ǫ −k (ϕ(k) − φ(k)), it is readily seen from (5.11) that
z(k + 1) = ǫ −1 (IM ⊗ A)z(k). Then, by choosing the quadratic function V1 (k) =
ǫ 2 zT (k)Rz(k), we calculate the difference along the trajectory of z(k) as follows:

1V1 (k) , V1 (k + 1) − V1 (k)

= ǫ 2 zT (k + 1)Rz(k + 1) − ǫ 2 zT (k)Rz(k) (5.23)
2
≤ −λmin {−50 }|z(k)| .

Defining λ0 , λmin {−50 }, λ̄1 , λmax {R} and λ1 , λmin {R}, one has 1V1 (k) ≤
−λ0 λ̄1−1 ǫ −2 V1 (k), which further indicates that |ϕ(k+1)−φ(k+1)| ≤ c0 |ϕ(k)−φ(k)|
1 −2 − 1 1 /2
with c0 = ǫ((1 − λ0 λ̄− 1 ǫ )λ̄1 λ1 ) . The proof is now complete.

Lemma 5.4: Let the positive scalar 0 < µ < 1 be given. Consider the augmented
estimation error system (5.13). If there exist a positive definite matrix S̄ > 0 and a
matrix X̄ satisfying
 
−(1 − µ)S (IM ⊗ (SA ¯ −X ¯ C))T
51 , <0 (5.24)
∗ −S

where S = IM ⊗ S̄ and X = IM ⊗ X̄, then there always exist an integer h ∈ Z+ and

a scalar c1 (0 < c1 < 1) such that

|ξ̃ (k + h)| < c1 |ξ̃ (k)|. (5.25)

Moreover, the desired observer gain matrix for (5.9) can be obtained with the form of

L = S¯ −1 X̄. (5.26)

Proof: First, by selecting the quadratic function V2 (k) = ξ̃ T (k)Sξ̃ (k) for (5.13)
and defining its difference as 1V2 (k) , V2 (k + 1) − V2 (k), we calculate the term
1V2 (k) + µV2 (k) along the trajectory of (5.13) as

1V2 (k) + µV2 (k) = ξ̃ T (k + 1)Sξ̃ (k + 1) − (1 − µ)ξ̃ T (k)Sξ̃ (k)

= ξ̃ T (k)(IM ⊗ (A − LC))T S(IM ⊗ (A − LC))ξ̃ (k)
− (1 − µ)ξ̃ T (k)Sξ̃ (k).
96 Data-Rate-Constrained State Estimation and Control

It is inferred from (5.24), (5.26) and the Schur Complement Lemma that

1V2 (k) + µV2 (k) ≤ 0. (5.27)

After some straightforward algebraic manipulations, it follows from inequality (5.27)

that

µ1 |ξ˜ (k + h)|2 ≤ V2 (k + h) ≤ (1 − µ)h V2 (k) ≤ (1 − µ)h µ

¯ 1 |ξ˜ (k)|2

where µ̄1 = λmax {S} and µ1 = λmin {S}. Therefore, one further has

|ξ̃ (k + h)| ≤ c1 |ξ̃ (k)|

with c1 = (µ1−1 µ̄1 (1 − µ)h )1/2 . Noting the fact that 0 < µ < 1, there must exist a
positive integer h ensuring 0 < c1 < 1, which completes the proof of this lemma.

The objective of this chapter is to design an observer-based controller (5.7)

for the discrete-time MAS (5.1). More specifically, for a given directed
communication network topology, we are interested in looking for the
controller parameter K such that the consensus performance (5.19) can be
achieved under the CDCP.

5.2 Design of CDCP and Performance Analysis of MASs

In this section, we shall investigate the design issue of the CDCP (5.5)–(5.6)
for discrete-time MAS (5.1). Then, based on the established communication
protocol, we would like to handle the observer-based controller design
problems for the addressed system.

5.2.1 Coding-Decoding Communication Protocol

In this subsection, we are ready to tackle the synthesis issue for the CDCP
(5.5)–(5.6) by using the uniform quantization approach. For the convenience
of the readers, to begin with, we provide a brief introduction regarding the
quantization method which will be adopted later.
Let the scaling parameters ai > 0 (i = 1, 2, . . . , M) and integer q be given.
(j)
Then, the hyperrectangles Bai = {ϑi ∈ Rn : |ϑi | ≤ ai , j = 1, . . . , n} can be
respectively partitioned into qn hyperrectangles Ii1i (ai ) × Ii2i (ai )×· · ·× Iini (ai ),
s1 s2 sn
where s1i , si2 , . . . , sni ∈ {1, 2, . . . , q} and
Observer-Based Consensus Control 97

 
ij (j) (j) 2ai
I1 (ai ) , ϑi : −ai ≤ ϑi < −ai + ;
q
 
ij (j) 2ai (j) 4ai
I2 (ai ) , ϑi : −ai + ≤ ϑi < −ai + ;
q q
..
.
 
ij (j) 2ai (j)
Iq (ai ) , ϑi : ai − ≤ ϑi ≤ ai .
q

(j)
Here, ϑi is the jth element of the vector ϑi .
For each Bai , the center of the hyperrectangle Ii1i (ai ) × Ii2i (ai ) × · · · × Iini (ai )
s1 s2 sn
can be determined by
 T
ηai i (si1 , si2 , . . . , sin ) , υ1i υ2i · · · υni

(2si1 −1)ai (2si2 −1)ai (2sni −1)ai

where υ1i = −ai + q , υ2i = −ai + q and υni = −ai + q .
Consequently, for any ϑi ∈ Bai , there exist unique integers i i i
s1 , s2 , . . . , sn ∈
{1, 2, . . . , q} such that ϑi ∈ Ii1i (ai ) × Ii2i (ai ) × · · · × Iini (ai ), which indicates
s1 s2 sn
that
√
nai
|ϑi − ηai i (si1 , si2 , . . . , sin )| ≤ . (5.28)
q

In order to accomplish the control task for the MAS (5.1) under the digital
transmission mechanism, we present the following design scheme of the
CDCP: Coder for Agent i : For ξˆi (dh) − ξ¯i (dh) ∈ Ii1i (ai (dh)) × Ii2i (ai (dh)) ×
s1 s2
· · · × Iini (ai (dh)) ⊂ Bai (dh) , one obtains
sn

gi (dh) = {si1 , si2 , . . . , sin } (5.29)

where ξ̄i (dh) is governed by


 ξ̄i (0) = 0




 ξ̄i (k) = ξ̆i (k), k 6 = dh

(5.30)
 ξ̄i (dh) = Aξ̆i (dh − 1) + Bŭi (dh − 1)






ŭi (dh − 1) = Kξ̆ (dh − 1)
98 Data-Rate-Constrained State Estimation and Control

and the dynamics of ξ̆i (k) is determined by the following system



 ξ̆i (0) = 0


 ξ̆i (k + 1) = Aξ̆i (k) + Bŭi (k), k 6 = dh − 1

(5.31)
 ξ˘i (dh) = ξ̄i (dh) + ηi i (si1 , si2 , . . . , sni )


 a (dh)


ŭi (k) = Kξ̆i (k).

Decoder for Agent i :



 ξ̌i (0) = 0


 ξ̌i (k + 1) = Aξ̌i (k) + Bui (k), k 6 = dh − 1

(5.32)
ˇ i i i i
 ξi (dh) = ξ̄i (dh) + ηai (dh) (s1 , s2 , . . . , sn )




ui (k) = Kξ̌i (k).

Remark 5.1: Two auxiliary systems (5.30) and (5.31) are constructed in the design
of the coding-decoding protocol. The auxiliary system (5.30) generates a “prediction”
ξ̄i (dh) of the decoded state ξ̌i (dh) as the decoded state ξ̌i (dh) cannot be obtained at the
coding instant dh. On the other hand, it follows from (5.31) and (5.32) that ξ̆i (k) ≡
ξ̌i (k) for ∀k ≥ 0, and hence the auxiliary system (5.31) can be seen as a “copy” of
the system (5.32). Different from the procedure in [190] where the decoded state is
required to feed back to the coder side, in this chapter, such a requirement is removed
due to the introduction of an additional auxiliary system (5.31), which is closer to
the engineering practice.

Denoting ξ̄ (k) , colM {ξ̄i (k)} and ξ̆ (k) , colM {ξ̆i (k)}, we reformulate the
coder-decoder procedure in a compact form as follows:
Coder : For ξ̂ (dh) − ξ̄ (dh) ∈ I111 (a(dh)) × · · · × I11n (a(dh)) × I212 (a(dh)) × · · · ×
s1 sn s1
I22n (a(dh)) × · · · × IM1 Mn
M (a(dh)) × · · · × I M (a(dh)) ⊂ Ba(dh) , we have
sn s1 sn

g(dh) = {s11 , . . . , s1n , s21 , . . . s2n , . . . , s1M , . . . , sM

n } (5.33)

where Ba(dh) = {ϑ(dh) ∈ RMn : |ϑ (j) (dh)| ≤ a(dh), j = 1, . . . , Mn} with a(dh) =
max1≤i≤M {ai (dh)} and ξ̄ (dh) = colM {ξ̄i (dh} is constructed as


 ξ̄ (0) = 0


 ξ̄ (k) = ξ̆ (k), k 6 = dh
(5.34)
 ξ¯ (dh) = (IM ⊗ A)ξ̆ (dh − 1) − (LG ⊗ B)ŭ(dh − 1)




ŭ(dh − 1) = (IM ⊗ K)ξ̆ (dh − 1)
Observer-Based Consensus Control 99

with the evolution of the dynamics ξ̆ (k) as follows



 ξ̆ (0) = 0


 ξ̆ (k + 1) = Aξ̆ (k) + Bŭ(k), k 6= dh − 1
(5.35)


 ξ̆ (dh) = ξ̄ (dh) + ηa(dh) (s11 , . . . , sM
n )


ŭ(k) = (IM ⊗ K)ξ̆ (k).

Decoder :

 ξˇ (0) = 0



 ξ̌ (k + 1) = (IM ⊗ A)ξ̌ (k) − (L ⊗ B)u(k), k 6= dh − 1
G
1
(5.36)
 ξ̌ (dh) = ξ̄ (dh) + ηa(dh) (s , . . . , sM )


 1 n

u(k) = (IM ⊗ K)ξ̌ (k).

Before proceeding, let us give some notations as follows:

s̄0 , |s0 |, s0 , colM {si0 }, ŝ0 , colM {ŝi0 },

s̃0 , |s0 − ŝ0 |, a(h) , c1 s̃0 + ch0 s̄0 ,
√ (5.37)
d h h Mn
a((d + 1)h) , c1 s˜0 (c1 + c0 ) + c0 a(dh).
q

The following lemma is to verify that the CDCP proposed in this chapter
is well defined. In other words, we would like to deal with the decoding
condition ξ̂ (dh) − ξ̄ (dh) ∈ Ba(dh) for all d = 1, 2, . . ..

Lemma 5.5: The coding-decoding communication protocol (5.33) and (5.36)

satisfies the following requirement:

|ξ̂ (dh) − ξ̄ (dh)|∞ ≤ a(dh), d = 1, 2, . . . . (5.38)

Proof: The mathematical induction method is utilized in this proof. First, for d = 1,
it follows from Lemma 5.3, Lemma 5.4 and the property of vector norm that

|ξ̂ (h) − ξ̄ (h)| ≤ |ξ̂ (h) − ξ(h)| + |ξ(h) − ξ̄ (h)|

≤ c1 |ξ̂ (0) − ξ(0)| + |ξ(h) − ξ̄ (h)|
≤ c1 s̃0 + c0 |ξ(h − 1) − ξ̌ (h − 1)| (5.39)
≤ c1 s̃0 + c20 |ξ(h − 2) − ξ̌ (h − 2)|
≤ · · · ≤ c1 s̃0 + ch0 s̄0 = a(h)
100 Data-Rate-Constrained State Estimation and Control

or, equivalently,
|ξ̂ (dh) − ξ̄ (dh)|∞ ≤ a(h). (5.40)

Second, with the assumption that |ξ̂ (jh) − ξ̄ (jh)|∞ ≤ a(jh) for j = 2, . . . , d, one has

|ξ̂ ((d + 1)h) − ξ̄ ((d + 1)h)|

≤ |ξ̂ ((d + 1)h) − ξ((d + 1)h)| + |ξ((d + 1)h) − ξ̄ ((d + 1)h)|

(5.41)
≤ c1 |ξ̂ (dh) − ξ(dh)| + c0 |ξ((d + 1)h − 1) − ξ̌ ((d + 1)h − 1)|

≤ · · · ≤ c1d+1 s˜0 + c0h |ξ(dh) − ξˇ (dh)|.

By considering the dynamics of ξ̌ (dh) in (5.36), it is further derived that

|ξ(dh) − ξ̌ (dh)|

= |ξ(dh) − ξ̄ (dh) − ηa(dh) (s11 , . . . , sM

n )|

≤ |ξ(dh) − ξ̂ (dh)| + |ξ̂ (dh) − ξ̄ (dh) − ηa(dh) (s11 , . . . , sM (5.42)

n )|
√
Mn
≤ cd1 s˜0 + a(dh).
q

Therefore, in combination with (5.41) and (5.42), we can draw the conclusion that
√
Mn
|ξ̂ ((d + 1)h) − ξ̄ ((d + 1)h)| ≤ c1d s˜0 (c1 + ch0 ) + ch0 a(dh). (5.43)
q

Moreover, by noting (5.37), (5.43) also implies

|ξ̂ ((d + 1)h) − ξ̄ ((d + 1)h)|∞ ≤ a((d + 1)h). (5.44)

Finally, it follows readily from (5.39)–(5.44) that the requirement in Lemma 5.5 is
satisfied for d ≥ 1, which completes the proof of this lemma.

Theorem 5.1: The discrete-time system (5.11) is detectable with a CDCP (5.33)–
(5.36) if the following inequality
√
Mn
ch0 <1 (5.45)
q

holds subject to (5.21) and (5.24) for a certain proper positive integer q, where
the parameters c0 and h have been defined in Lemma 5.3 and Lemma 5.4,
respectively.
Observer-Based Consensus Control 101

Proof: According to the definition of a(dh) in (5.37) and 0 < c1 < 1, it is easy to see
that

lim a(dh) = 0, (5.46)

d→+∞

and then one can easily infer from (5.42) and (5.46) that

lim |ξ(dh) − ξ̌ (dh)| = 0.

d→∞

Moreover, for those non-coding instants k ∈ (dh, (d+1)h), it is noticed that ξ(k) and
ξ̌ (k) are actually the trajectories of (5.11) and, therefore, it follows immediately from
Lemma 5.3 that |ξ(k) − ξ̌ (k)|2 ≤ c0k−dh |ξ(dh) − ξ̌ (dh)|2 . So, we have that |ξ(k) −
ξ̌ (k)| is bounded at the non-coding instants, which implies that the dynamical system
(5.11) is detectable, namely, limk→∞ |ξ(k) − ξ̌ (k)| = 0. The proof is now complete.

5.2.2 Consensus Analysis

Let us start with introducing some basic notions and background on input-
to-state stability theory, which will be useful for the derivation of our main
results later.
Consider the following discrete-time nonlinear system with the following
form

ς (k + 1) = f (ς (k), τ (k)) (5.47)

where ς(k) ∈ Rn , τ (k) ∈ Rp and f (·, ·) : Rn × Rp 7 → Rn are, respectively,

the system state vector, the exogenous input and the continuous nonlinear
function with f (0, 0) = 0.

Definition 5.3: The system (5.47) is said to be input-to-state stable, if there exist a
KL function β(·, ·) and a K class function γ ( · ) such that the dynamics of the system
state ς(k) satisfies

|ς (k)| ≤ β(|ς (0)|, k) + γ (|τ (k)|∞ ) (5.48)

for ∀k ≥ 0 and ∀x(0) ∈ Rn , where |u(k)|∞ , supk {|u(k)|}..

Lemma 5.6: ([58]). The nonlinear discrete-time system (5.47) is said to be input-to-
state stable if there exist a positive definite function V(k, ς(k)) : [0, +∞) × Rn 7 → R
(called an ISS-Lyapunov function), three K∞ class functions α1 (·), α2 (·) and α3 (·),
and a K class function ̟ ( · ) such that the following two inequalities

α1 (|ς (k)|) ≤ V(k, ς (k)) ≤ α2 (|ς(k)|) (5.49)

102 Data-Rate-Constrained State Estimation and Control

and

V(k + 1, ς (k + 1)) − V(k, ς (k)) ≤ −α3 (|ς(k)|) + ̟ (|u(k)|) (5.50)

hold for all ς(k) ∈ Rn and u(k) ∈ Rp . Furthermore, if the above two inequalities are
met simultaneously, the functions β(·, ·) and γ (·) in Definition 5.3 can be selected as
β(·, k) = α1−1 (φ k α2 (·)) (0 < φ < 1) and γ (·) = α1−1 (α2 (α3−1 (̟ (·)))), respectively,
where α1−1 ( · ) expresses the inverse function of the monotone function α1 ( · ) and so
does α3−1 ( · ).
Subsequently, recalling the special form of (5.17b), we decouple it into the
following r − 1 parts

δ̃i (k + 1) = (IMi ⊗ A − 3i ⊗ BK)δ̃i (k) − ([W]i−1 ⊗ BK)ξE (k), i = 2, . . . , r

(5.51)

where [W]j ∈ RMj ×M stands for a matrix whose elements come from the 1st
Pj
row to the M2 th row of matrix W for j = 1, and from the ( s=2 Ms + 1)th
Pj+1
row to the ( s=2 Ms )th row of matrix W for j ≥ 2.
As discussed in [182], from the properties of the Jordan form (5.16), the
asymptotic property of system (5.51) is dominated by the diagonal terms,
and therefore, the consensus issue of system (5.1) can be now converted into
the problem of asymptotic stability of the following closed-loop system

θi (k + 1) =(A − λi BK)θi (k) − (Wi−1 ⊗ BK)ξE (k), i = 2, . . . , r (5.52)

where θi stands for an n-dimensional vector and Wj ∈ R1×M is a row vector

whose elements come from the 1st row of matrix W for j = 1 and from the
Pj
( s=2 Ms + 1)th row of matrix W for j ≥ 2.
It is worth mentioning that, since the topology structure of the
communication network is directed, some eigenvalues of the Laplacian
matrix LG may be complex values. To this end, for i = 1, . . . , r, we denote λi =
ai + bi j, θi (k) = θiR (k) + θiI (k)j and Wi = Wi−1
R + W I j, where a , θ R (k), W R
i−1 i i i−1
and b , θ I (k), W I are, respectively, the real and imaginary parts of λ , θ (k) and
i i i i i
Wi−1 . In addition, we assume that the real parts of the r different eigenvalues
of LG in an ascending order are written as 0 = a1 < a2 ≤ · · · ≤ ar . As such,
separating the real and imaginary parts of (5.52) for i = 2, · · · , r, one finds



θiR (k + 1) = (A − ai BK)θiR (k) + bi BKθiI (k)

 R ⊗ BK)ξE (k)
−(Wi−1
I (5.53)
θi (k + 1)

 = (A − ai BK)θiI (k) − bi BKθiR (k)

 I ⊗ BK)ξE (k).
−(Wi− 1
Observer-Based Consensus Control 103

It is not difficult to see that the systems (5.51) are asymptotically stable as
long as the systems (5.53) are asymptotically stable, which indicates from
the Lemma 5.2 that the asymptotic consensus can be reached in MASs (5.1).
Before providing the main results, we define

θ R (k) , colr−1 {θiR (k)}, θ I (k) , colr−1 {θiI (k)},

θ(k) , [(θ R (k))T (θ I (k))T ]T , Q , I2r ⊗ Q,

W̃ , max {(WiR )T WiR + (WiI )T WiI }.

1≤i≤r

Theorem 5.2: Suppose that the network topology G contains a directed spanning
tree. Let the positive scalars ̺1 , ̺2 and ̺3 be given. The asymptotic consensus of
system (5.1) can be reached with the CDCP (5.29)-(5.32) if there exist a matrix K
and a positive matrix Q satisfying


 5̄1 , AT QBK + (BK)T QA


− (a2 + ar )(BK)T Q(BK) > 0

 (5.54a)
T
 5̄2 , (1 + σ1 )(A − a2 BK) Q(A − a2 BK)




+ (1 + σ2 )b2m (BK)T Q(BK) − Q < 0

(5.54b)

subject to (5.21), (5.24) and (5.45), where σ1 = ̺1 + ̺2 , σ2 = ̺1−1 + ̺3 and

bm = max1≤i≤r {bi }.
P  
Proof: Choose V(k) = ri=2 (θiR (k))T QθiR (k) + (θiI (k))T QθiI (k) and calcu-
late the difference along the trajectory of (5.53) as
r n
X
V (k + 1) − V(k) = ((A − ai BK)θiR (k))T Q(A − ai BK)θiR (k)
i=2

+ bi2 (θiI (k))T (BK)T QBKθiI (k)

+ ξE T (k)(WiR T R E
−1 ⊗ BK) Q(Wi−1 ⊗ BK)ξ (k)

+ 2bi (θiR (k))T (A − ai BK)T QBKθiI (k)

− 2(θiR (k))T (A − ai BK)T Q(Wi−

R ⊗ BK)ξE (k)
1

− 2bi (θiI (k))T (BK)T Q(Wi−1

R ⊗ BK)ξE (k)

+ (θiI (k))T (A − ai BK)T Q(A − ai BK)θiI (k)

+ b2i (θiR (k))T (BK)T QBKθiR (k)

104 Data-Rate-Constrained State Estimation and Control

+ ξE T (k)(WiI−1 ⊗ BK)T Q(Wi−1

I ⊗ BK)ξE (k)

− 2bi (θiI (k))T (A − ai BK)T QBKθiR (k)

− 2(θiI (k))T (A − ai BK)T Q(WiI−1 ⊗ BK)ξE (k)

+ 2bi (θiR (k))T (BK)T Q(Wi−1

I ⊗ BK)ξE (k)
o
− (θiR (k))T QθiR (k) − (θ I (k))T QθiI (k) .

Then, considering the elementary inequality 2aT b ≤ δaT a + δ −1 bT b, we obtain

r n
X 
V(k + 1) − V(k) ≤ (θiR (k))T (1 + σ1 )(A − ai BK)T Q(A − ai BK)
i=2

+ (1 + σ2 )bi2 (BK)T Q(BK) − Q θiR (k)

+ (θiI (k))T (1 + σ1 )(A − ai BK)T Q(A − ai BK)
 o
+ (1 + σ2 )b2i (BK)T Q(BK) − Q θiI (k) + ξE T (k)ŴξE (k)

where Ŵ , (r − 1)(1 + σ3 )(W̃ ⊗ ((BK)T Q(BK))) with σ3 = ̺2−1 + ̺3−1 .

For 2 < i ≤ r, denoting 5̄i , (1 + σ1 )(A − ai BK)T Q(A − ai BK) + (1 + σ2 )
2
bi (BK)T Q(BK) − Q and noticing 0 < a2 ≤ ai , it is easy to obtain from (5.54a) that

5̄2 − 5̄i = (1 + σ1 )(AT QA − a2 AT QBK − a2 (BK)T QA

+ a22 (BK)T QBK − AT QA + ai AT QBK

+ ai (BK)T QA − ai2 (BK)T QBK)

+ (1 + σ2 )(b2m − b2i )(BK)T QBK

= (1 + σ1 )(ai − a2 )(AT QBK + (BK)T QA (5.55)

− (a2 + ai )(BK)T QBK)

+ (1 + σ2 )(b2m − b2i )(BK)T QBK

≥ (1 + σ1 )(ai − a2 )(AT QBK + (BK)T QA

− (a2 + ar )(BK)T QBK)

+ (1 + σ2 )(b2m − bi2 )(BK)T QBK ≥ 0.

Observer-Based Consensus Control 105

Then, combining (5.54b) and (5.55), one has 5̄i ≤ 5̄2 < 0. Consequently, in
accordance with the definition of θ (k), we further derive that

r n
X o
V(k + 1) − V(k) ≤ (θiR (k))T 5̄2 θiR (k) + (θiI (k))T 5̄2 θiI (k)
i=2

+ ξE T (k)ŴξE (k)
r n
X o
≤ −λmin {−5̄2 } (θiR (k))T θiR (k) + (θiI (k))T θiI (k)
i=2

+ ξ (k)ŴξE (k)
ET

= −λmin {−5̄2 }|θ(k)|2 + λmax {Ŵ}|ξE (k)|2 . (5.56)

Noting that V(k) can also be rewritten as V(k) = θ T (k)Q θ(k), we choose
 
α1 |θ (k)| = λmin {Q }|θ (k)|2 ,
 
α2 |θ (k)| = λmax {Q }|θ (k)|2 ,
 
α3 |θ (k)| = λmin {−5̄2 }|θ(k)|2 ,
̟ (|ξE (k)|) = λmax {Ŵ}|ξE (k)|2 .

Therefore, it is inferred from Lemma (5.53) is input-to-state r stable and so does

λmax {Q} k
the system (5.52). Here, we can choose β(|θ (0)|, k) = λmin {Q}
φ |θ(0)| and
r
{Q}λmax {Ŵ } E
γ (|ξE (k)|) = cλλmax{Q }λ {−5̄ }
|ξ (k)|, where 0 < c < 1. Furthermore, by using the
min min 2
property of “converging-input” to “converging state” in [58], it can be immediately
found that limk→∞ |δ̃i (k)| = 0 for i = 2, . . . , r, which completes the proof.

Remark 5.2: Both the ISS theory and the algebraic graph method are utilized to
derive the consensus conditions in this chapter. First, according to analysis results
in Theorem 1, the decoding error ξE (k) can be regarded as the exogenous disturbance
inputs which are bounded and asymptotically converge to zero, and hence the ISS
theory can be directly used to analyze the coder-decoder-based consensus problem.
In addition, the eigenvalue information (including both the real part and imaginary
part) of the Laplacian matrix is reflected in the consensus conditions, which shows
that the communication topology serving as an important factor has a non-negligible
impact on the consensus of the MASs.

Theorem 5.3: Suppose that the network G contains a directed spanning tree. Let
the positive scalars ̺1 , ̺2 and ̺3 be given. The asymptotic consensus in system
106 Data-Rate-Constrained State Estimation and Control

(5.1) can be reached with the CDCP (5.29)–(5.32) if there exist a positive matrix Q,
a nonsingular matrix S and a matrix K satisfying
  
−˜a−2 AT QA (ã−1 SUA − ãK̃)T
<0 (5.57a)





 ∗ Q − (SU)T − SU

  
˜T

−Q σ˜ 1 (SUA − a2 K˜ )T σ˜ 2 bm K
 
  ∗

 Q − (SU)T − SU 0 



 ∗ ∗ Q − (SU)T − SU


<0 (5.57b)

subject to (5.21), (5.24) and (5.45), where

 
S11 S12
U= [B((BT B)−1 )T B⊥ ]T ,
S= ,
0 S22
√ p
K̃ = [KT 0]T , a˜ = a2 + ar , σ̃1 = 1 + σ1 ,
p
σ̃2 = 1 + σ2 .

Here, B⊥ is an orthogonal basis of the null space for matrix BT , that is BT B⊥ = 0.

S11 ∈ Rp×p , S12 ∈ Rp×(n−p) , and S22 ∈ R(n−p)×(n−p) are arbitrary matrices
ensuring the non-singularity of S. Other parameters are defined as in Theorem 5.1
and Theorem 5.2. Furthermore, if inequalities (5.57a) and (5.57b) are feasible, the
controller gain matrix can be determined by

−1
K = S11 K. (5.58)

Proof: First, we rewrite (5.54a) as

−ã−2 AT QA + (ã−1 A − aBK

˜ )T Q(ã−1 A − aBK
˜ ) < 0. (5.59)

By applying the Schur Complement Lemma, one can conclude that (5.54a) is true if
and only if the following inequality
 
−ã−2 AT QA (ã−1 A − aBK
˜ )T
<0 (5.60)
∗ −Q−1

holds.
On the other hand, it is worth mentioning that −Q−1 ≤ (SU)−1 Q(SU)−T −
(SU)−T − (SU)−1 , which means that (5.60) can be guaranteed by
 
−ã−2 AT QA (ã−1 A − aBK
˜ )T
<0 (5.61)
∗ −WT QW − W − WT
Observer-Based Consensus Control 107

where W = (SU)−T . Then, pre- and post-multiplying (5.61) by diag{I, W−T } and
its transpose leads to
 −2 T 
−ã A QA (ã−1 SUA − a˜ SUBK)T
< 0. (5.62)
∗ Q − (SU)T − SU

By noticing (5.58), we have SUBK = K̃, which implies that (5.54a) can be guaranteed
by (5.57a). Similarly, (5.54b) can also be guaranteed by (5.57b), and therefore the
proof of this theorem is complete.

5.3 An Illustrative Example

In this section, a simulation example is presented to verify the effectiveness of
the proposed observer-based consensus control scheme for the discrete-time
MASs.
Consider the system (5.1) with the following parameters
   
1 0.1 0.2
A= , B= , C = [0.5 0.5].
0.15 0.5 0.25

In this example, we assume that there are five agents communicating with
each other via a directed topology G with the adjacency matrix A = [aij ]5×5 ,
where a13 = a15 = a21 = a24 = a34 = a35 = a41 = a45 = a51 = a53 = 1 and
other elements are 0.
The eigenvalues of the associated Laplacian matrix LG are, respectively,
calculated as 0, 2, 2.5 + 0.866j, 2.5 − 0.866j and 3, where “j” stands for the
imaginary unit.
For each agent, the initial state is set as xi0 = [0.6 − 0.1i 0.8 + 0.3i]T ,
i = 1, 2, . . . , 5. Choose ǫ = 1.1, µ = 0.3, h = 3 and q = 15. Then, solving matrix
inequality (5.21), (5.24), (5.57a) and (5.57b) by the Matlab software, we have
c0 = 1.3247 and c1 = 0.5556. Moreover, the desired observer and controller
gain matrices are obtained as L = [1.3030 0.5608]T and K = [1.5918 0.1592],
respectively.
The simulation results are shown in Figures 5.1–5.3. Figure 5.1 depicts the
sate trajectories of each agent for the MAS (5.1) with the designed control
scheme and the sate trajectories of each agent without the control input
are plotted in Figure 5.2. In addition, in order to show the applicability
of the coding-decoding-based communication protocol, Figure 5.3 displays
the corresponding decoding error for each agent. It is inferred from
the three figures that the closed-loop MAS (5.1) can attain the desired
consensus performance while the open-loop one cannot. The simulation
results have demonstrated that the developed consensus control scheme
performs very well.
108 Data-Rate-Constrained State Estimation and Control

1.5

0.5

0
0 10 20 30 40 50

0.5

−0.5

−1
0 10 20 30 40 50

FIGURE 5.1
State trajectories xij (k) (i = 1, 2 . . . , 5; j = 1, 2) with control input.

1
x
0.5
x
0 x
x
−0.5
x
−1
0 10 20 30 40 50

2
x
1
x
0 x
x
−1
x
−2
0 10 20 30 40 50

FIGURE 5.2
State trajectories xij (k) (i = 1, 2 . . . , 5; j = 1, 2) without control input.
Observer-Based Consensus Control 109

1.5

0.5

0
0 10 20 30 40 50

0.5

−0.5

−1
0 10 20 30 40 50

FIGURE 5.3
The decoding errors ξEij (k) (i = 1, 2 . . . , 5; j = 1, 2).

5.4 Summary
In this chapter, the observer-based consensus control problem has been
dealt with for a class of discrete-time networked MASs with CDCP. The
signal transmission of each agent between the observer and the controller
has been implemented via the digital communication channel. By means
of the uniform quantization technique, the CDCP has been employed to
encode the relative measurements between adjacent agents to finite symbolic
values before transmission and then decode the received symbolic data at
the controller side. In virtue of the decoded signal, an observer-based control
scheme has been put forward. Then, consensus criteria have been derived in
terms of the solutions to certain matrix inequality constraints such that the
closed-loop system achieves the desired performance requirement.
6
Recursive Filtering with Measurement
Fading: A Multiple Description Coding
Scheme

In recent years, in response to the ever-growing complexity of dynamical

systems, the filtering or state estimation problem has been gaining increasing
popularity from both control and signal processing communities. According
to the categories of system noises and the specifications of estimation
performance, a variety of filtering strategies have been developed in
the literature with successful applications in real-world plants, see for
example [45,99,148]. Among others, the notable extended Kalman filtering
(EKF) algorithm has been recognized as an effective tool to deal with the state
estimation problems for nonlinear stochastic systems. So far, a rich body of
research results regarding the EKF algorithm has been reported either on the
variants of the EKF algorithm accommodating more general systems or on its
engineering applications in more general practice, see for example [97,170].
In practical engineering, the signal transmission via wireless channels
might suffer from certain physical interferences such as reflection, diffraction,
and scattering, and this gives rise to the signal fading issue. From the
mathematical viewpoint, the fading phenomenon is usually modeled by a
stochastic process to reflect the random fluctuations in both the amplitude
and the phase of the transmitted signals. Till now, a great deal of research
effort has been made toward the network-induced channel fading problems.
Nevertheless, when it comes to the nonlinear recursive filtering issues with
fading measurements, the corresponding results have been scattered in the
literature.
In reality, the bandwidth of a communication network is usually limited
and only finite bits of data are allowed to be transmitted in a certain time. As
such, it is of both theoretical importance and practical significance to make
full use of the limited network resource in order to achieve the desired system
performance. In this case, the data-coding scheme appears to be particularly
efficient in resource saving as this scheme maps the original data to certain
codewords with fewer bits.
Among the existing coding algorithms, the multiple description coding
(MDC) scheme stands out as an effective one for its distinguished advantages

DOI: 10.1201/9781003534853-6 110

Recursive Filtering with Measurement Fading 111

in error resilience and rate-adaptive streaming. The MDC is based on the

diversity principle [134] that possesses three features: 1) the information
source is encoded into multiple descriptions with identical importance and
then sent to the decoder via parallel independent channels; 2) the original
data is decoded with an acceptable quality as long as at least one of the
descriptions is successfully transmitted to the decoder; and 3) the more
descriptions received by the decoder the higher accuracy of the decoded data
will be attained. Up to date, the MDC scheme has found a wide range of
applications including distributed storage systems, diversity communication
system and image/audio/video coding. Unfortunately, despite their clear
engineering insights, the MDC-based control/filtering issues for NSs have
received very little research attention mainly because of the essential
difficulties in quantifying the influences from the implementation of the
MDC on the performance indices.
Motivated by the above discussions, in this chapter, we aim to investigate
the recursive filtering problem for a class of discrete-time stochastic nonlinear
systems subject to fading channels where the multiple description coding
scheme is employed during the data exchange between the devices. In order
to facilitate the data transmission in a resource-constrained communication
network, the multiple description coding scheme is adopted to encode the
fading measurements into two descriptions with the identical importance.
Two independent Bernoulli distributed random variables are introduced to
govern the occurrences of the packet dropouts in two channels from the
encoders to the decoders. The channel fading phenomenon is characterized
by the Mth-order Rice fading model whose coefficients are mutually
independent random variables obeying certain probability distributions. The
purpose of the problem addressed is to design a recursive filter such that, in
the simultaneous presence of the stochastic noises, the channel fading and
the data coding-decoding mechanism, an upper bound of the filtering error
variance is obtained and then minimized at each time step. In virtue of the
Riccati difference equation technique and the stochastic analysis approach,
the explicit form of the desired filter parameters is derived by solving a
sequence of coupled algebraic Riccati-like difference equations. Finally, a
simulation experiment is provided to show the applicability of the developed
filtering scheme.

6.1 Problem Formulation

Consider a discrete time-varying nonlinear stochastic system described by
(
xk+1 = h(xk , ζk ) + B1k ωk
(6.1)
ỹk = Ck xk + B2k νk
112 Data-Rate-Constrained State Estimation and Control

where xk ∈ Rnx represents the state vector, ζk ∈ Rnu is the system input, and
ỹk ∈ Rny denotes the measurement output. h(·, ·) : Rnx × Rnζ 7→ Rnx is
a deterministic and continuously differential nonlinear function. ωk ∈ Rnω
and νk ∈ Rnν are two families of mutually independent truncated Gaussian
white noise sequences taking values over the interval [−ω̄, ω̄] and [−ν̄, ν̄],
respectively, where ω̄ and ν̄ are given positive scalars. The probability
density functions of the truncated Gaussian white sequences ωk and νk are,
respectively, defined as
 
√1 φ ω√ k −µω
p R0k R0k
fω (ωk ; µω , R0k , −ω̄, ω̄) ,    
¯ −µω
8 ω√ − 8 −√ ω̄−µω
R0k R0k

and  
√1 φ νk −µν
√
p Q0k Q0k
fν (νk ; µν , Q0k , −ν̄, ν̄ ) ,    
ν̄−µν −
√ν¯−µν
8 √ −8
Q0k Q0k
Rx 2
where φ(x) , √1 exp ( − 21 x2 ), 8(x) , −∞ √1 exp ( − s2 )ds, µω and R0k
2π 2π
are, respectively, the mean and variance of ωk without truncation. Similarly,
µν and Q0k are, respectively, the mean and variance of νk without truncation.
Without loss of generality, letting µω = µν = 0, it is not difficult to calculate
that the truncated ωk and νk are zero-mean random sequences with the
variances
      
−ω̄ ω̄
ω̄ φ √ +φ √
 R0k R0k 
Rk = R0k  1 −      
 √ 
R0k 8 √−ω̄ − 8 √ω̄
R0k R0k

and       
ν̄ φ √−ν¯ + φ √ν¯
 Q0k Q0k 
1 −
Qk = Q0k  p
    ,
 
Q0k 8 √−ν̄ − 8 √ν¯
Q0k Q0k

respectively.

6.1.1 Channel Fading

In this chapter, the phenomenon of channel fading is considered between the
sensor and the encoder due to the unreliable wireless communication. The
fading signal yk is described by the following Mth-order Rice fading model:
Recursive Filtering with Measurement Fading 113

M
X
yk = αmk y˜ k−m + B3k ̟k (6.2)
m=0

where M stands for the given number of the paths where the signals
get through. αmk (m = 0, 1, . . . , M) are the channel coefficients which are
mutually independent random variables in m and k. In addition, αmk
have the probability density functions pm (αmk ) on the interval [0, 1]
with mathematical expectations ᾱm and variances ςm . ̟k ∈ Rn̟ is a zero-
mean Gaussian random vector that characterizes the channel noise with
variance Sk . All the noise signals and random variables αmk (m = 0, 1, . . . , M)
are mutually uncorrelated. B3k is a known time-varying matrix with
appropriate dimension.
To facilitate the description, we set Mk , min{M, k} and rewrite the model
(6.2) as follows:
Mk
X
yk = αmk ỹk−m + B3k ̟k . (6.3)
m=0

6.1.2 Multiple Description Coding Scheme

In a networked environment, the phenomenon of packet dropouts is
frequently encountered during data transmissions owing primarily to the
limited capacity of the communication network. The MDC scheme is
known for its error resilience and has been widely used to improve the
communication quality. For example, with the use of the MDC scheme, the
negative impacts from the packet dropouts are mitigated in the sense
that the data stream is not interrupted at the cost of temporary quality
degradation. As such, in this chapter, we aim to examine how the
MDC scheme influences the state estimation performance. Without loss of
generality, the two-description case is considered. As shown in Figure 6.1, the

FIGURE 6.1
Structure of two-description coding-decoding for NSs with unreliable channels.
114 Data-Rate-Constrained State Estimation and Control

fading measurement is first encoded into two descriptions, which are then
simultaneously transmitted to the decoders via two mutually independent
channels.
According to Figure 6.1, the general form of the MDC is mathematically
characterized as follows:
Encoder:
(
ulk = f1l (ylk )
(6.4)
vlk = f2l (ylk )

Decoder:

when λ1k = 1, λ2k = 0


 g1l (ulk ),


ȳlk = g2l (vlk ), when λ1k = 0, λ2k = 1 (6.5)


when λ1k = 1, λ2k = 1

gcl (ulk , vlk ),

for l = 1, . . . , ny , where f1l ( · ) and f2l ( · ) are two coding functions whose
outputs ulk and vlk can be regarded as the individual descriptions of the
source ylk with ylk being the lth component of yk . ȳlk is the decoding value
corresponding to ylk . g1l ( · ) and g2l ( · ) are two side decoding functions and
gcl (·, ·) is the central decoding function. λik (i = 1, 2) are two independent
random variables which govern the packet dropout phenomena during the
transmissions of the descriptions and satisfy the Bernoulli binary distribution
taking values on either 1 or 0 with mathematical expectations λ̄i and
variances σi , respectively. Here, λki = 1 indicates that the ith channel works
well, and λki = 0 means that the ith channel suffers from the packet dropouts
at time instant k.
For notational convenience, we denote

uk , [u1k u2k · · · uny k ]T

vk , [v1k v2k · · · vny k ]T

ȳk , [ȳ1k ȳ2k · · · ȳny k ]T

f1 (yk ) , [f11 (y1k ) f12 (y2k ) · · · f1ny (yny k )]T

(6.6)
f2 (yk ) , [f21 (y1k ) f22 (y2k ) · · · f2ny (yny k )]T

g1 (yk ) , [g11 (y1k ) g12 (y2k ) · · · g1ny (yny k )]T

g2 (yk ) , [g21 (y1k ) g22 (y2k ) · · · g2ny (yny k )]T
gc (yk ) , [gc1 (y1k ) gc2 (y2k ) · · · gcny (yny k )]T
Recursive Filtering with Measurement Fading 115

where uk and vk are two description vectors to be transmitted as two packets

via different channels.
It is not difficult to see that there are a total of three cases at the decoder
side.

1) Only one packet, either uk or vk , is available to the corresponding

decoder g1 ( · ) or g2 ( · ).
2) Both the packets uk and vk are received by the decoder gc ( · ).
3) Neither of the packets is received by the decoders. In this case, both
the packets are lost and thus the zero-order holder (ZOH) strategy
is implemented by using the latest decoded measurement ȳk−1 to
compensate the current decoded value ȳk , that is, ȳk = ȳk−1 when
λ1k = λ2k = 0.

6.1.3 The Filter Structure

After obtaining the decoded measurement ȳk according to the decoding
scheme (6.5) and the zero-order holder strategy, we are now in a position
to design the recursive filter with the following form:

x̂k+1|k = h(x̂k|k , ζk ) (6.7a)




x̂k+1|k+1 = x̂k+1|k + Kk+1 ϑ ȳk+1 (6.7b)

where xˆ k|k ∈ Rnx is the estimate of the system state xk at time instant k, x̂k+1|k ∈
Rnx is the one-step prediction of xk+1 at k, Kk+1 ∈ Rnx ×ny is the time-varying
filter parameter to be designed, and ϑ( · ) : Rny 7→ Rny is the innovation
function of yk+1 whose explicit expression is to be given later.
The objective of this chapter is twofold.

1) Provide a rigorous mathematical analysis on the MDC scheme, prove

that the decoding error is uniformly bounded, and then establish a
unified model for the decoded measurement ȳk .
2) Design a time-varying recursive filter of the form (6.7a)-(6.7b) for the
stochastic nonlinear system (6.1) subject to fading measurements and
the MDC such that the filtering error variance satisfies the following
requirement:

E{(xk+1 − x̂k+1|k+1 )(xk+1 − x̂k+1|k+1 )T } ≤ 5k+1|k+1 (6.8)

where 5k+1|k+1 is certain upper bound of the filtering error variance,

which is further minimized by solving an optimization problem with
respect to the filter parameter Kk+1 .
116 Data-Rate-Constrained State Estimation and Control

6.2 Design of the MDC and Analysis of the Decoding Error

In this section, we first endeavor to design the MDC via the scalar
quantization approach. Then, for all different scenarios of the packet
delivery, the corresponding decoding errors are analyzed. In addition, a
comprehensive mathematical model of the decoded measurement ȳk is
established.

6.2.1 Design of the MDC

From the technical point of view, the design problem of the MDC procedure
can be generally divided into two steps: 1) the index generation and 2) the
index assignment. More specifically, in this chapter, we shall use a set of
scalar uniform quantizers to convert the source into a set of specific indices.
Then, by employing the so-called nested index assignment principle [156],
the generated indices are mapped into the corresponding cells of an index
mapping matrix. In what follows, we elaborate the multiple description
coding procedure in accordance to the above steps.

Index Generation
First, we define the scalar uniform quantizer q( · ) with the following
structure:


 L, τk ≥ L

q(τ ) = −L, τk < −L (6.9)

−L + (2s−1)L , −L + υ ≤ τ < −L + υ

ρ 1 2

where τ ∈ R is the signal to be quantized and L is the saturation value;

υ1 , 2(s−1)L
ρ and υ2 , 2sL
ρ for s ∈ {1, 2, . . . , ρ}. It can be seen from (6.9) that
the interval
h [−L, L] is partitioned
 into ρ regions Ii (L)(i = 1, 2, . . . , ρ), where
Ii (L) = −L + 2(i−1)L
ρ , −L+ 2iL
ρ and Ii (L) ∩ Ij (L) = ∅ for any i, j ∈ {1, 2, . . . , ρ}
and i 6= j.
y
For the measurement component ylk , define ql (ylk ) , µlk q( µlk ), where
lk
µlk > 0 is an adjustable parameter. In order to ensure that the quantizer ql ( · )
is not saturated, at each time instant, we select a proper µlk to guarantee that
the value ylk /µlk is located within the interval [−L, L] when |ylk | > L. Hence,
it is easily calculated that the quantization error ̺lk , ylk − ql (ylk ) satisfies

µlk L
|̺lk | ≤ . (6.10)
ρ
Recursive Filtering with Measurement Fading 117

FIGURE 6.2
Nested index assignment for d = 8 and p = 1 in [156].

Remark 6.1: It is observed from (6.10) that a larger µlk leads to a larger
quantization error, which means that the scaling parameter µlk should be chosen as
small as possible for the sake of decreasing the quantization error. On the other hand,
µlk needs to be large enough to avoid the quantizer saturation. As such, µlk should
be selected appropriately to balance the trade-off between the quantization error and
the quantizer saturation.

Next, for the lth quantized measurement component ql (ylk ), we define the
index generator function ψl ( · ) : R → N+ satisfying

ψl (ql (ylk )) = ilk , when ql (ylk ) ∈ Iilk (L) (6.11)

with ilk ∈ {1, 2, . . . , ρ} and l = 1, 2, . . . , ny .

Index Assignment
After generating the index ilk , we need to assign it into the corresponding cell
of the mapping matrix J according to the nested assignment principle. Then,
in terms of the location of ilk in J, we obtain a pair of descriptions (ulk , vlk )
accordingly. Without loss of generality, we suppose that the mapping matrix
J is a d−dimensional matrix with d being an even number.
Before proceeding, let us illustrate the nested assignment principle. It is
inferred from [156] that the indices are allocated in those cells that lie on
the main diagonal and on the 2p diagonals closest to the main diagonal.
To simplify the exposition, in this chapter, we discuss the index assignment
problem only for the case p = 1. As shown in Figure 6.2, all the indices are
allocated in the cells that lie on 2p + 1 = 3 diagonals including the main
diagonal and other 2 diagonals which are closest to the main diagonal.

Remark 6.2: It is apparent from (6.10) that the quantization level ρ has a significant
impact on the quantization accuracy. That is, a large quantization level ρ results in
a high quantization accuracy (or small quantization error) while occupying a large
amount of bandwidth. It is noted that a real network channel is often subjected to
the bandwidth constraint. Here, it is assumed that only χ-bits of data are allowed
118 Data-Rate-Constrained State Estimation and Control

to be transmitted at each time instant. In this case, for the traditional uniform-
quantization-based single description coding approach, the quantization level is
determined by ρ = 2χ . However, it is seen from Figure 6.2 that the proposed
two-description-based coding approach enlarges the quantization level from 2χ to
3 × 2χ − 2, thereby improving the quantization accuracy.

Now, we are in a position to conduct a rather detailed mathematical

analysis on the implementation of the proposed index assignment scheme.
Denote by ϕl ( · ) : N+ → N+ × N+ the index assignment function, which is
defined as ϕl (ilk ) , (ϕ1l (ilk ), ϕ2l (ilk )) =


(r + 1, rlk + 1), if mlk =1
 lk


 (rlk + 1, rlk ), if mlk = 0 and rlk is even



(rlk , rlk + 1), if mlk = 0 and rlk is odd (6.12)

lk + 2, rlk + 1), if mlk = 2 and rlk is even



 (r

(rlk + 1, rlk + 2), if mlk = 2 and rlk is odd


where ϕ1l (ilk ) and ϕ2l (ilk ) are, respectively, the row assignment function and
column assignment function; rlk = ⌊ 2pilk+1 ⌋ = ⌊ i3lk ⌋ and mlk = h 2p+1
ilk
i = h i3lk i.
It is easy to see that the index assignment function ϕl (ilk ) maps the single
description ilk into the corresponding description pair (ulk , vlk ), where ulk =
ϕ1l (ilk ) and vlk = ϕ2l (ilk ).
Based on the above discussion, the coding functions f1l ( · ) and f2l ( · ) in (6.4)
can be finally expressed as
(
f1l (ylk ) = ϕ1l (ψl (ql (ylk )))
(6.13)
f2l (ylk ) = ϕ2l (ψl (ql (ylk ))).

Up to now, by employing the index generator function ψl ( · ) and the index

assignment function ϕl ( · ), the component-based encoder design problem
has been dealt with. In the next stage, we shall address the decoder design
problem based on the established coding procedure.

6.2.2 Analysis of the Decoding Error

In light of the established index generation and assignment schemes, in this
subsection, we are to investigate the decoding accuracy of the proposed
coding-decoding strategy. For this purpose, the first step is to construct
an index estimation function to estimate each index ilk according to the
received descriptions. Then, based on the estimated index îlk , we will design
a decoding function to obtain the decoded value ȳlk .
Recursive Filtering with Measurement Fading 119

In the following, we discuss the possible situations of the decoding process

on a one-by-one basis in terms of the individual case of the received
descriptions at the decoder side.

• Case 1: both the descriptions ulk and vlk are successfully received at
the decoder side. Then, the central decoder is applied to uniquely
determine the index ilk and obtain the decoded measurement
component ȳlk .
• Case 2: one of the descriptions (ulk or vlk ) is lost during the packet
transmissions. The explicit location of the index ilk is unknown and
only its row number (or column number) can be acquired. Hence, in
order to generate the decoded measurement ȳlk by the side decoder,
we let the diagonal element in the same row (or column) with ilk be
an estimation of ilk .
• Case 3: neither of the description packets is available to the decoders.
In this case, îlk is set as zero and the latest decoded measurement
ȳl(k−1) is utilized to compensate the value of ȳlk by using the ZOH
strategy.

Based on the above analysis, the key point in the design of the decoding
procedure is to determine the estimated index value îlk for each measurement
component ylk . Obviously, with more descriptions received, we would
expect a less error between îlk and ilk , and hence a more accurate decoded
value ȳlk . As such, corresponding to the above three cases, we need to
examine the relationships between the received description(s) and the
estimated index îlk .
Let us start with the analysis for Case 1. In this case, it follows directly from
(6.12) that îlk , φcl (ulk , vlk ) =

 3ulk − 2,

if ulk = vlk


 3ulk − 3, if ulk = vlk + 1 and ulk is odd




3ulk , if ulk = vlk − 1 and ulk is odd (6.14)



 3ulk − 4, if ulk = vlk + 1 and ulk is even





3ulk − 1, if ulk = vlk − 1 and ulk is even

where φcl (·, ·) is the estimation function of the central decoder. Furthermore,
(6.14) means that the estimated index îlk can be explicitly calculated and the
index estimation error is ĩlk = îlk − ilk = 0.
For Case 2, suppose that only the description ulk (the row number) is
received while the description vlk (the column number) is lost. In such a case,
120 Data-Rate-Constrained State Estimation and Control

we let the index, which is located in the diagonal cell of the ulk th row, be the
estimation of ilk . So, we have

îlk , φ1l (ulk ) = 3ulk − 2 (6.15)

where φ1l ( · ) is the index estimation function of the side decoder 1.

Similarly, if only the description vlk is available, the index estimation
function for ilk is determined by

îlk , φ2l (vlk ) = 3vlk − 2 (6.16)

where φ2l ( · ) is the index estimation function of the side decoder 2.

For Case 3, there is no description packet received at the decoder side.
Therefore, instead of estimating ilk , the latest decoded measurement ȳl(k−1)
is directly employed to compensate the value of ȳlk .
Based on the above analysis, we introduce the following theorem that gives
a quantitative evaluation on the decoding error for different cases of the
descriptions received at the decoder. This theorem plays a crucial role for
the subsequent filter design issue.

Theorem 6.1: Under the framework of the multiple description coding scheme
(6.4)–(6.5), for each measurement component ylk (l = 1, 2, . . . , ny ), the
corresponding decoding error 1lk , ȳlk − ylk satisfies
(
1slk , |1slk | ≤ 5̺lk for the side decoder
1lk = c c (6.17)
1lk , |1lk | ≤ ̺lk for the central decoder

where ̺lk is defined previously as ̺lk , ylk − ql (ylk ).

Proof: First, recalling that ilk can be calculated as ilk = 3rlk + mlk , we confirm from
(6.12), (6.15)–(6.16) that the index estimation error satisfies
(
0, both descriptions are received
|ĩlk | ≤ (6.18)
2, only one description is received.

Then, based on the obtained index estimation îlk , we define the following inverse
quantization function q̂l ( · ) as follows:

(2îlk − 1)L
qˆ l (iˆlk ) , −L + (6.19)
ρ

for îlk ∈ {1, 2, . . . , ρ}.

Recursive Filtering with Measurement Fading 121

Accordingly, the decoder functions are given by


 g1l (ulk ) = q̂l (φ1l (ulk )), when λ1k = 1, λ2k = 0

g (v ) = q̂l (φ2l (vlk )), when λ1k = 0, λ2k = 1 (6.20)
 2l lk
gcl (ulk , vlk ) = q̂l (φcl (ulk , vlk )), when λ1k = λ2k = 1.


Then, it follows directly from (6.20) that the decoding error 1lk of ylk satisfies (6.17),
and the proof of this theorem is complete.

6.2.3 A Unified Measurement Model

Having analyzed the decoding error for each measurement component ylk ,
we are now ready to establish a unified measurement model accounting for
the MDC as well as the channel fading. For notational simplicity, define

s s · · · 1ns y k ]T
1sk , [11k 12k
c c · · · 1cny k ]T
1ck , [11k 12k (6.21)
βik , δ(λ̃k , i), λ̃k , λ1k + λ2k .

Based on the results in Theorem 6.1, the decoded measurement output ȳk
is described with the following form:

ȳk = β0k ȳk−1 + β1k (yk + 1ks ) + β2k (yk + 1ck ) (6.22)
qP qP
ny 2 and κ c = ny 2
where |1sk | ≤ κks and |1kc | ≤ κkc with κks = l=1 (5̺ lk ) k l=1 ̺lk .
P2
In addition, it is derived from the definition of βik that i=0 βik = 1 and

β̄0 , E{β0k } = (1 − λ̄1 )(1 − λ̄2 )

β̄1 , E{β1k } = λ̄1 (1 − λ̄2 ) + λ̄2 (1 − λ̄1 ) (6.23)
β̄2 , E{β2k } = λ̄1 λ̄2 .

Consequently, the decoded measurement output (6.22) is further converted

into
ȳk = β0k ȳk−1 + (1 − β0k )yk + β1k 1sk + β2k 1ck . (6.24)
which serves as a unified measurement model reflecting both the MDC
scheme and the channel fading phenomenon.

Remark 6.3: It follows from (6.17) and (6.21) that the decoding error can be
expressed as 1k , δ(λ̃k , 0)|1kd | + δ(λ̃k , 1)|1ks | + δ(λ˜ k , 2)|1kc |, where 1ks and
122 Data-Rate-Constrained State Estimation and Control

qP
ny 2
1ck , which have been defined in (6.21), satisfy |1sk | ≤ κks , l=1 (5̺lk ) and
qP
ny 2
|1kc | ≤ κkc , d
l=1 ̺lk , respectively. |1k | denotes the decoding error when both of
the descriptions are missing, and therefore it is reasonable to assume that its upper
bound is greater than κks . Then, in terms of the condition κks = 5κkc , there exists a
positive scalar p > 5 such that |1dk | ≤ pκkc . Next, introducing an average decoding
error which is defined as 1̄k , E{1k }, one calculates 1̄k = E{δ(λ̃k , 0)}|1kd | +
E{δ(λ̃k , 1)}|1sk | + E{δ(λ̃k , 2)}|1ck | = β̄0 |1dk | + β̄1 |1sk |+β̄2 |1kc | ≤ β̄0 pκkc + β̄1 κks +
β̄2 κkc = (pβ̄0 + 5β¯ 1 + β̄2 )κkc , ℵ(λ¯ 1 , λ̄2 , κkc ), where β¯ 0 = (1 − λ̄1 )(1 − λ̄2 ),
β̄1 = λ̄1 (1 − λ̄2 ) + λ̄2 (1 − λ̄1 ) and β̄2 = λ̄1 λ̄2 . It is not difficult to derive that
ℵ(λ̄1 , λ̄2 , κkc ) is monotonically decreasing with respect to λ¯ i (i = 1, 2). Therefore, one
can conclude that a larger packet dropout rate (i.e. a smaller λ̄i (i = 1, 2)) would lead
to a larger upper bound of the average decoding error, which complies with practical
engineering.

6.3 The Recursive Filtering Strategy

So far, we have discussed the design of the MDC procedure for the fading
measurement yk . In this section, we are interested in the filter design issue.
For convenience of later analysis, for i ∈ [−M, −1], we let yi = 0. Based on the
obtained decoded measurement ȳk in (6.24), the innovation function ϑ(ȳk+1 )
in (6.7b) is given as follows:

M
X
ϑ(ȳk+1 ) , ȳk+1 − β̄0 ȳk − (1 − β̄0 ) ᾱm Ck−m+1 x̂k−m+1|k−m . (6.25)
m=0

According to the proposed filter structure (6.7a)-(6.7b), we define the one-

step prediction error and filtering error by x̃k+1|k , xk+1 − x̂k+1|k and
x̃k+1|k+1 , xk+1 − x̂k+1|k+1 , respectively. Then, we denote by Pk+1|k ,
E{x̃k+1|k x̃Tk+1|k } and Pk+1|k+1 , E{x̃k+1|k+1 x˜ Tk+1|k+1 } the prediction error
variance and the filtering error variance, respectively.
By considering (6.1) and (6.7a), we have

x̃k+1|k = h(xk , ζk ) − h(x̂k|k , ζk ) + B1k ωk . (6.26)

In order to linearize the nonlinear function h(xk , ζk ), the Taylor series

expansion method is employed around x̂k|k , which leads to

h(xk , ζk ) = h(x̂k|k , ζk ) + Ak x̃k|k + o(|x̃k|k |) (6.27)

Recursive Filtering with Measurement Fading 123

∂h(xk ,ζk )
where Ak = ∂xk and o(|˜xk|k |) means the linearization error caused
xk =x̂k|k
by discarding the high-order terms. As discussed in [174], such a linearization
error is further transformed into the following tractable form:

o(|x̃k|k |) = Mk Fk Nk x̃k|k (6.28)

where Mk is a problem-dependent scaling matrix, the matrix Nk is used to

provide extra design freedom, and the matrix Fk (Fk FTk ≤ I) describes the
errors of the linear approximation manipulation.
Substituting (6.27) into (6.26) yields

x̃k+1|k = (Ak + Mk Fk Nk )x̃k|k + B1k wk . (6.29)

Then, we derive the following dynamics of the prediction error variance:

Pk+1|k = (Ak + Mk Fk Nk )Pk|k (Ak + Mk Fk Nk )T + B1k Qk BT1k . (6.30)

Denoting γm,k , (1 − β0k )αm,k , γ̄m , E{γm,k } = (1 − β̄0 )α¯ m , β̃i,k , βi,k − β̄i and
γ̃m,k , γm,k − γ̄m , we calculate the innovation of the filter as follows:

M
X
ϑ(ȳk+1 ) = ȳk+1 − β̄0 ȳk − (1 − β̄0 ) ᾱm Ck−m+1 x̂k−m+1|k−m
m=0

M
X
= β0,k+1 ȳk + (1 − β0,k+1 ) αm,k+1 ỹk−m+1
m=0
s c
+ (1 − β0,k+1 )B3,k+1 ̟k+1 + β1,k+1 1k+1 + β2,k+1 1k+1
M
X
− β̄0 ȳk − (1 − β̄0 ) ᾱm Ck−m+1 x̂k−m+1|k−m (6.31)
m=0

M
X M
X
= β̃0,k+1 ȳk + γ̃m,k+1 Ck−m+1 xk−m+1 + γm,k+1
m=0 m=0

M
X
× B2,k−m+1 νk−m+1 + γ̄m Ck−m+1 x̃k−m+1|k−m
m=0

+ (1 − β0,k+1 )B3,k+1 ̟k+1 + β1,k+1 1sk+1 + β2,k+1 1k+1

c
.

Hence, it follows from (6.1), (6.7b) and (6.31) that the filtering error dynamics
is governed by
124 Data-Rate-Constrained State Estimation and Control

M
X
x̃k+1|k+1 = (I − γ̄0 Kk+1 Ck+1 )x̃k+1|k − Kk+1 γ̄m Ck−m+1 x̃k−m+1|k−m
m=1
M
X M
X
− Kk+1 γ̃m,k+1 Ck−m+1 xk−m+1 − γm,k+1 Kk+1 B2,k−m+1 νk−m+1
m=0 m=0
− β̃0,k+1 Kk+1 ȳk − (1 − β0,k+1 )Kk+1 B3,k+1 ̟k+1
s c
− β1,k+1 Kk+1 1k+1 − β2,k+1 Kk+1 1k+1 . (6.32)

In the following, three useful lemmas are presented to facilitate the

subsequent filter analysis and design issues.

Lemma 6.1: Assume that A, M, N and F are real matrices of compatible dimensions
with FFT ≤ I. For any positive scalar π > 0, if there exists a positive definite matrix
X > 0 such that π −1 I − NXNT > 0, then the following inequality is true:

(A + MFN)X(A + MFN)T ≤ A(X−1 − π NT N)−1 AT + π −1 MMT . (6.33)

Lemma 6.2: For the positive definite matrix X > 0, real matrix functions Uk (X) =
UkT (X) ∈ Rn×n and Vk (X) = VkT (X) ∈ Rn×n , if

Uk (X) ≤ Uk (Y), ∀X ≤ Y = YT (6.34)

and
Uk (Y) ≤ Vk (Y) (6.35)
over the time horizon [0 N], then the solutions to the following difference equations:

Sk+1 = Uk (Sk ), Tk+1 = Vk (Tk ), S0 = T0 (6.36)

satisfy
Sk ≤ Tk . (6.37)

The proof of the following lemma is easily accessible from (6.32) and is
therefore omitted.

Lemma 6.3: The filtering error variance Pk+1|k+1 is determined by

Pk+1|k+1 = (I − γ̄0 Kk+1 Ck+1 )Pk+1|k (I − γ̄0 Kk+1 Ck+1 )T

+ E{Fk F T } + E{Gk+1 G T } + β¯ 0 (1 − β¯ 0 )Kk+1 y¯ k y¯ T KT
k k+1 k k+1
M
X
T T
+ (1 − β̄0 )Kk+1 B3,k+1 Kk+1 + γ̄m Kk+1 B2,k−m+1 Kk+1
m=0
Recursive Filtering with Measurement Fading 125

+ β̄2 Kk+1 1ck+1 1cT T ¯ s sT T

k+1 Kk+1 + β1 Kk+1 1k+1 1k+1 Kk+1
T T T
− Lk+1 − Lk+1 − Mk+1 − Mk+1 − Nk+1 − Nk+1 (6.38)
+ Ok+1 + OkT+1 + Pk+1 + Pk+1
T T
+ Qk+1 + Qk+1
T T T
+ Rk+1 + Rk+1 + Sk+1 + Sk+1 + Tk+1 + Tk+1
T
+ Uk+1 + Uk+1

where
M
!
X
Fk = Kk+1 γ̄m Ck−m+1 x̃k−m+1|k−m
m=1
M
!
X
Gk+1 = Kk+1 γ̃m,k+1 Ck−m+1 xk−m+1
m=0
 !T 
 XM 
Lk+1 = (I − γ̄0 Kk+1 Ck+1 )E x̃k+1|k γ̄m Ck−m+1 x̃k−m+1|k−m KT
  k+1
m=1
 sT T
Mk+1 = β̄1 (I − γ¯0 Kk+1 Ck+1 )E x̃k+1|k 1k+1 Kk+1

Nk+1 = β̄2 (I − γ̄0 Kk+1 Ck+1 )E x̃k+1|k 1cT T
k+1 Kk+1
( M !)
X
Ok+1 = β̄1 Kk+1 E γ̄m Ck−m+1 x̃k−m+1|k−m DeltasT T
k+1 Kk+1
m=1
( M
!)
X
cT T
Pk+1 = β̄2 Kk+1 E γ̄m Ck−m+1 x̃k−m+1|k−m 1k+1 Kk+1
m=1
( M
)
X
Qk+1 = Kk+1 E β˜ 0,k+1 γ˜m,k+1 Ck−m+1 xk−m+1 ȳkT Kk+1
T

m=0
( M
)
X
sT T
Rk+1 = Kk+1 E β1,k+1 γ̃m,k+1 Ck−m+1 xk−m+1 1k+1 Kk+1
m=0
( M
)
X
Sk+1 = Kk+1 E β2,k+1 γ̃m,k+1 Ck−m+1 xk−m+1 1kcT+1 Kk+1
T

m=0
n o
Tk+1 = Kk+1 ȳk E β̃0,k+1 β1,k+1 1sT T
k+1 Kk+1
n o
Uk+1 = Kk+1 ȳk E β̃0,k+1 β2,k+1 1cT T
k+1 Kk+1

B2,k−m+1 = B2,k−m+1 Rk−m+1 BT2,k−m+1

B3,k+1 = B3,k+1 Sk+1 BT3,k+1 .
126 Data-Rate-Constrained State Estimation and Control

On the basis of Lemmas 6.1–6.3, we shall now handle the design issue of
the filter parameter Kk . Due to uncertainty terms (e.g. Fk , 1sk and 1ck ) in the
filtering error variance matrix Pk|k , an upper bound of Pk|k is first determined,
and then the filter parameter Kk is calculated by minimizing the derived
upper bound at each sampling instant k.

Theorem 6.2: Consider the prediction error variance (6.30) and the filtering error
variance (6.38). Let the positive scalars ǫi , εh̄i (h¯ = l, o, p, q, r, s and i = 1, 2, . . . , M)
and εj (j = m, n, t, u) be given. Assume that the following discrete-time coupled
algebraic Riccati-like difference equations:

5k+1|k = V1k (5k|k )

−1 (6.39)
, Ak (5k|k − π Nk NkT )−1 ATk + π −1 Mk MTk + B1k Qk B1k
T

and

5k+1|k+1 = V2k (5k+1|k )

, ξ1 (I − γ̄0 Kk+1 Ck+1 )5k+1|k (I − γ̄0 Kk+1 Ck+1 )T
(2)
+ ξ20 Kk+1 Ck+1 5k+1|k CTk+1 Kk+1
T

M
X
+ ξ2m Kk+1 Ck−m+1 5k−m+1|k−m CTk−m+1 Kk+1
T

m=1
M
!
X
+ Kk+1 ξ3 θk2 I + ξ4 ȳk ȳTk + T
γ̄m B2,k−m+1 + (1 − β̄0 )B3,k+1 Kk+1
m=0
M
X
+ ξ5m Kk+1 Ck−m+1 X̂k−m+1|k−m CkT−m+1 KkT+1 (6.40)
m=1

subject to
(
5i|i = Pi|i > 0; i = −M, −M + 1, . . . , 0 (6.41a)
−1
π I − Nk 5k|k NkT >0 (6.41b)

have the positive definite solutions 5k+1|k and 5k+1|k+1 . Then, the desired filter
parameter Kk+1 is determined by

Kk+1 = ξ1 γ¯0 5k+1|k CTk+1 ξ¯1 Ck+1 5k+1|k CTk+1

M
X
+ ξ2m Ck−m+1 5k−m+1|k−m CTk−m+1 + ξ3 θk2 I
m=1
Recursive Filtering with Measurement Fading 127

M
X
+ ξ4 ȳk ȳTk + B2,k−m+1 + (1 − β̄0 )B3,k+1
m=0
M
!−1
X
+ ξ5m Ck−m+1 X̂k−m+1|k−m CTk−m+1 (6.42)
m=0

where
(1) (2)
ξ1 = 1 + ε̄l + εm + εn , ξ2m = ξ2m + ξ2m
(1) (2)
ξ3 = 25ξ3 + ξ3 , σγmn = E{γ˜m,k γ˜n,k }
ξ4 = β̄0 (ε̄q β̄0 + εt−1 β¯ 0 + εu−1 β̄0 + 1 − β̄0 )
)ℑm , ξ¯1 = ξ1 γ¯02 + ξ20
−1 (2)
ξ5m = (1 + ǫm
(1) −1 2 −1 2 −1 2
ξ2m = εlm γ¯m + εom γ̄m + εpm γ̄m + γ̄ γ¯m

= β̄12 (εm
−1
+ ε¯ o + ε¯ r + εt + β¯ 1−1 )
(1)
ξ3
= β̄22 (εn−1 + ε̄p + ε̄s + εu + β̄2−1 )
(2)
ξ3
−1 2 −1 ¯ 2 2
ℑm = εqm γ̄m + ε̄rm β0 α¯ m + ε¯ s−m1 β¯ 02 α¯ m
2
+ σγm
M
X
(2)
ξ2m = (1 + ǫm )ℑm , γ̄ = γ¯m
m=1
X̂k−m+1|k−m = x̂k−m+1|k−m x̂Tk−m+1|k−m
M
X
ε̄h̄ = εh¯ m , h¯ = l, o, p, q, r, s
m=1
 1
ny 2 M
2 L
X X
θk =  µlk , σγm = σγmn
ρ
l=1 n=0

V1k ( · ) : Rnx ×nx → Rnx ×nx and V2k ( · ) : Rnx ×nx → Rnx ×nx stand for two matrix-
valued functions. In addition, the matrix 5k+1|k+1 is an upper bound of Pk+1|k+1 ,
namely, Pk+1|k+1 ≤ 5k+1|k+1 . Also, such an upper bound is minimized by the filter
parameter Kk+1 at each sampling instant.

Proof: The proof is carried out by applying the mathematical induction. First,
according to (6.30), the prediction error variance Pk+1|k is actually a function
of the filtering error variance Pk|k , namely, Pk+1|k , U1k (Pk|k ), and then one
easily obtains that U1k (Xk ) ≤ U1k (Yk ), ∀ 0 ≤ Xk ≤ Yk . Noting that Pi|i = 5i|i
(i = −M, −M + 1, . . . , 0) and assuming that Pi|i ≤ 5i|i (i = 1, 2, . . . , k), we have
128 Data-Rate-Constrained State Estimation and Control

U1k (Pk|k ) ≤ U1k (5k|k ). Subsequently, it is readily observed from Lemma 6.1, (6.30),
(6.39) and (6.41b) that Pk+1|k = U1k (Pk|k ) ≤ U1k (5k|k ) ≤ V1k (5k|k ) = 5k+1|k . In
the following, we only need to show that Pk+1|k+1 ≤ 5k+1|k+1 . For this purpose, let
us deal with some terms of the right-hand side of (6.38).
1 1 1 1
By resorting to the elementary inequality (ε 2 x − ε− 2 y)(ε 2 x − ε− 2 y)T ≥ 0 for
any x, y ∈ Rn and ε > 0, one has

−LkT+1 − Lk+1 ≤ ε¯ l (I − γ¯0 Kk+1 Ck+1 )Pk+1|k (I − γ¯0 Kk+1 Ck+1 )T

M
X
−1 2
+ εlm γ¯m Kk+1 Ck−m+1 Pk−m+1|k−m CTk−m+1 KkT+1 , (6.43)
m=1
T
Mk+1 − Mk+1 ≤ εm (I − γ¯0 Kk+1 Ck+1 )Pk+1|k (I − γ¯0 Kk+1 Ck+1 )T
−1 ¯ 2
+ εm β1 Kk+1 1sk+1 1sT T
k+1 Kk+1 (6.44)
T T
−Nk+1 − Nk+1 ≤ εn (I − γ¯0 Kk+1 Ck+1 )Pk+1|k (I − γ¯0 Kk+1 Ck+1 )

+ εn−1 β¯ 22 Kk+1 1ck+1 1cT T

k+1 Kk+1 (6.45)
M
X
−1 2
OkT+1 + Ok+1 ≤ εom T
γ̄m Kk+1 Ck−m+1 Pk−m+1|k−m Ck−m+1 T
Kk+1
m=1
+ ε̄o β̄12 Kk+1 1sk+1 1sT T
k+1 Kk+1 (6.46)
M
X
T −1 2
Pk+1 + Pk+1 ≤ εpm γ¯m Kk+1 Ck−m+1 Pk−m+1|k−m CTk−m+1 Kk+1
T

m=1
+ ε̄p β̄22 Kk+1 1ck+1 1cT T
k+1 Kk+1 (6.47)
M
X
−1 2
QkT+1 + Qk+1 ≤ εqm γ¯m Kk+1 Ck−m+1 E{xk−m+1 xTk−m+1 }CTk−m+1 Kk+1
T

m=0
+ ε̄q β̄02 Kk+1 ȳk ȳTk KkT+1 (6.48)
M
X
T −1 ¯ 2 2
Rk+1 + Rk+1 ≤ εrm β0 α¯ m Kk+1 Ck−m+1 E{xk−m+1 xTk−m+1 }CTk−m+1 Kk+1
T

m=0
(6.49)
+ ε̄r β̄12 Kk+1 1k+1s
1sT T
k+1 Kk+1 (6.50)
XM
T −1 ¯ 2 2 T
Sk+ 1 + Sk+1 ≤ εsm β0 α¯ m Kk+1 Ck−m+1 E{xk−m+1 xk−m+1 }CTk−m+1 Kk+1
T

m=0
+ ε̄s β̄22 Kk+1 1k+1c
1cT T
k+1 Kk+1 (6.51)
T
Tk+1 + Tk+1 ≤ εt−1 β¯ 02 Kk+1 ȳk ȳkT Kk+1
T
+ εt β̄12 Kk+1 1sk+1 1sT T
k+1 Kk+1 (6.52)
Recursive Filtering with Measurement Fading 129

and
T
Uk+1 + Uk+1 ≤ εu−1 β̄02 Kk+1 ȳk ȳTk Kk+1
T
+ εu β̄22 Kk+1 1k+1
c
1cT T
k+1 Kk+1 . (6.53)

Moreover, invoking the Jensen inequality yields

M
X
E{Fk FkT } ≤ γ̄ γ̄m Kk+1 Ck−m+1 Pk−m+1|k−m CTk−m+1 Kk+1
T
(6.54)
m=1
M
X
E{Gk GkT } ≤ σrm Kk+1 Ck−m+1 E{xk−m+1 xTk−m+1 }CTk−m+1 Kk+1
T
. (6.55)
m=0

By utilizing the elementary inequality again, it is shown that the term

E{xk−m+1 xTk−m+1 } on the right-hand sides of (6.48)–(6.51) and (6.55) further
satisfies

E{xk−m+1 xTk−m+1 } ≤ ǭ1m Pk−m+1|k−m + ǭ2m x̂k−m+1k−m x̂Tk−m+1|k−m (6.56)

where ǭ1m = 1 + ǫm and ǭ2m = 1 + ǫm −1 . On the other hand, based on (6.41a) and the

assumption Pi|i ≤ 5i|i (i = 1, 2, . . . , k), it is obvious that Pi+1|i ≤ 5i+1|i (i = −M,

−M + 1, . . . , k). Hence, according to the above discussion, it is not difficult to verify
that

Pk+1|k+1 ≤ V2 (Pk+1|k ) ≤ V2 (5k+1|k ) = 5k+1|k+1 . (6.57)

In the rest of the proof, it remains to show that the given filter parameter (6.42)
minimizes the variance upper bound 5k+1|k+1 at each sampling instant. Applying
the partial derivative operation to the trace of 5k+1|k+1 with respect to Kk+1 , we have

∂tr{5k+1|k+1 }
∂Kk+1
(2)
= −2ξ1 γ¯0 (I − γ¯0 Kk+1 Ck+1 )5k+1|k CTk+1 + 2ξ20 Kk+1 Ck+1 5k+1 CTk+1
M
X
T 2
+2 ξ2m Kk+1 Ck−m+1 5k−m+1|k−m Ck−m ¯ kT
+1 + 2Kk+1 ξ3 θk I + ξ4 ȳk y
m=1
(6.58)
M
!
X
+ γ̄m B2,k−m+1 + (1 − β̄0 )B3k
m=0
M
X
+2 ξ5m Kk+1 Ck−m+1 X̂k−m+1|k−m CTk−m+1 .
m=1
130 Data-Rate-Constrained State Estimation and Control

∂tr{5k+1|k+1 }
Setting ∂ Kk+1 = 0, the filter parameter Kk+1 is immediately obtained with
the form in (6.42), and the proof of this theorem is thus complete.

Remark 6.4: The computational burden is a crucial issue which should be taken
into account during the filtering algorithm design. For the considered system (1)
and (2), the dimensions of system parameters are given as follows: xk ∈ Rnx ,
yk ∈ Rny , B1k ∈ Rnx ×nω , B2k ∈ Rny ×nν , B3k ∈ Rny ×n̟ and Ck ∈ Rny ×nx . By
carrying out some matrix manipulations (e.g. addition, multiplication, transpose,
inversion, etc.), it is not difficult to calculate that the proposed filtering algorithm
mainly contains a nonlinear operation, (7M + 18) addition operations, (18M + 35)
multiplication operations, (7M + 12) transpose operations and (3M + 7) inversion
operations, respectively. Based on the well-known computation complexity of the
matrix manipulations, the total computational complexity of the proposed filtering
algorithm for each iteration is O = maxi=1,2...,6 {Oi }, where O1 = O(h(x̂k|k , ζk )),
O2 , O(9n2.376 x + 3n2x + nx nω (nx + nω )), O3 , Oa3 + Om3 + Ot3 + Oi3 , O4 ,
O((M+1)nx ny +(M+2)ny ), O5 , O((1+nx )ny ), and O6 = Oa6 + Om6 + Ot6 with
Oa3 = O((3M+5)ny2 ), Om3 = O(2(1+M)nx ny (nx +ny )+ny2 +(M+1)ny nν (ny +
nν )+ny n̟ (ny +n̟ )), Ot3 = O((3M+4)nx ny +ny ) and Oi3 = O((3M +5)ny2.376 )
and Oa6 = O((7 + 3M)nx2 ), Om6 = O((M + 3)nx2.376 + (6M + 5)nx2 ny + (M + 1)
(ny n2ν + n2y nν ) + ny n2̟ + n2y n̟ ), Ot6 = O(n2x + (4M + 3)ny nx ). It is noted that the
computational complexity increases polynomially with the growth of the dimensions
of system parameters. Fortunately, research on matrix computations is a quite active
field in applied mathematics and operations research community, and substantial
speedups can be expected in the near future.

Remark 6.5: The main results of this chapter are dependent on some parameters
such as ǫi , εhi
¯ (h¯ = l, o, p, q, r, s and i = 1, 2, . . . , M) and εj (j = m, n, t, u).
In fact, these parameters are involved in the upper bound of the filtering error
variance (i.e. 5k|k ). In general, a basic principle for selecting these parameters is
to minimize the obtained upper bound of the filtering error variance (i.e. 5k|k ) at
time instant k, that is, it is desirable to determine a set of parameters that would
(locally) minimize 5k|k in the sense of matrix trace. Unfortunately, the upper bound
of the filtering error variance (i.e. 5k|k ) is actually a non-convex function of these
parameters, and this makes it extremely difficult (if not impossible) to develop an
analytic algorithm for selecting the best parameters in order to achieve the local
minimum of the obtained upper bound. In this case, a rather practical way would
be to select these parameters for optimal performance by using the evolutionary
computation algorithms (e.g. particle swarm optimization algorithm and genetic
algorithm), and this would be one of the future research topics.

Remark 6.6: In this chapter, we make tremendous efforts to examine the impact
from the MDC scheme on the filtering performance for a class of stochastic nonlinear
systems subject to measurement fading. It can be seen from Theorem 6.2 that all
Recursive Filtering with Measurement Fading 131

the factors contributing to the system complexity are reflected in the filter design
procedure, which include: 1) the stochasticity and nonlinearity of the system; 2) the
channel coefficient of the measurement fading phenomenon; 3) the coding accuracy
of the proposed MDC scheme; and 4) the occurrence probabilities of the packet
dropout during the transmission of the individual descriptions. In comparison with
the existing literature, there are three distinct features with our main results: 1) this
is the first attempt to consider the recursive filtering issue for the nonlinear stochastic
system with the MDC scheme; 2) a concrete design procedure is, for the first time,
proposed in a mathematically rigorous way on the MDC scheme; and 3) the developed
filtering algorithm is of a recursive nature and is therefore suitable for the online
application.

6.4 An Illustrative Example

The tank systems are widely applied in engineering practice such as process
control for petrochemical and metallurgical industries. As a simplified tank
system, the three-tank system is commonly adopted in the laboratory to
simulate the dynamics of the industrial tank systems.
In this section, in order to validate the effectiveness and the applicability
of the proposed nonlinear filtering strategy in a networked environment,
an experimental simulation is conducted on the internet-based three-tank
system [189]. Figure 6.3 shows the diagram of the three-tank system. It
can be seen from Figure 6.3 that the three tanks constitute the main body
of the system. The three tanks are of the equivalent cross section Sa and
interconnected by two cylindrical pipes with the cross section Sn . For Tank
1 and Tank 2, two pumps are equipped to pump water to each one with the

FIGURE 6.3
The diagram of three-tank system.
132 Data-Rate-Constrained State Estimation and Control

flow rate Q1 (t) and Q2 (t). Also, two sensors are installed to measure their
level heights, respectively.
In the simulation, it is assumed that the levels of three tanks satisfy the
condition h1 (t) > h3 (t) > h2 (t), that is, water always flows from Tank 1 to
Tank 2 via Tank 3. In terms of the “mass balance principle”, the dynamical
behavior of the level heights of the tanks can be described with the following
differential equations:
 1
˙
h1 (t) = Sa (Q1 (t) − Q13 (t)) + b11 ω1 (t)

1
ḣ2 (t) = Sa (Q2 (t) + Q32 (t) − Q20 (t)) + b12 ω2 (t) (6.59)

 1
ḣ3 (t) = Sa (Q13 (t) − Q32 (t)) + b13 ω3 (t)

where Qij (t) stands for the water flow rate from the ith tank to the jth tank
q
and can be calculated as Qij (t) = azi Sn sign(hi (t) − hj (t)) 2g|hi (t) − hj (t)|, in
which azi is the real value of the outflow coefficient of pipe i, Sn is the cross
section area of the connection pipe, and g = 9.8m/s2 is the acceleration of
gravity. Q20 (t) is the rate of outflow
p of the liquid from tank 2 and can be
determined by Q20 (t) = az2 Sn 2gh2 (t). ωi (t) is the process noise with the
intensive coefficient b1i . Consequently, substituting the detailed forms of the
above parameters to (6.59) yields
  p 
 ˙ 1 (t) = 1 − az1 Sn 2g(h1 (t) − h3 (t)) + Q1 (t)
h

 S a




 + b11 ω1 (t)

ḣ (t) = 1 az S p2g(h (t) − h (t)) − az S p2gh (t)

 2 Sa 3 n 3 2 2 n 2
 (6.60)

 + Q2 (t) + b12 ω2 (t)

 
ḣ (t) = 1 az S p2g(h (t) − h (t))



 3 S a
1 n 1 3

 p 
 − az S 2g(h (t) − h (t)) + b ω (t).
3 n 3 2 13 3

By applying the Euler discretization approach [43] to the above continuous-

time state equation at each sampling instant tk , the corresponding
approximated discrete-time counterpart is obtained as follows:
 TSn
p
h1 (tk+1 ) = h1 (tk ) − Sa az1 2g(h1 (tk ) − h3 (tk ))





 + STa Q1 (tk ) + Tb11 ω1 (tk )
 p
h2 (tk+1 ) = h2 (tk ) + TSn az3 2g(h3 (tk ) − h2 (tk ))



 pS a
 − TS T
Sa az2 2gh2 (tk ) + Sa Q2 (tk )
n
(6.61)


 + Tb12 ω2 (tk )
p
h3 (tk+1 ) = h3 (tk ) + TS

Sa az1 2g(h1 (tk ) − h3 (tk ))
 n

 p
− TS

Sa az3 2g(h3 (tk ) − h2 (tk ))

 n



+ Tb13 ω3 (tk ).
Recursive Filtering with Measurement Fading 133

Here, T is the constant sampling period, that is, T = tk+1 − tk . Similarly, the
measurement equation is obtained by the following formula:
(
y1 (tk ) = (1 + sin(tk ))h1 (tk ) + b21 ν1 (tk )
(6.62)
y2 (tk ) = h2 (tk ) + b22 ν2 (tk )

where sin(tk ) reflects the measurement error and νi (tk ) is the measurement
noise with the intensive coefficient b2i .
For convenience, in the sequel, we denote k as the shorthand of tk . Also, the
process noises ωik (i = 1, 2, 3) are assumed to be the same one and obey the
truncated Gaussian distribution, that is, ωik (i = 1, 2, 3) obey the Gaussian
distribution with zero mean and variance Qk and lie within the interval
[r1 , r2 ]. Moreover, the similar assumption is made on the measurement noises
νik (i = 1, 2), that is, νik = νk ∼ N(0, Rk ) (i = 1, 2). From (6.60)–(6.62), system
parameters are given as follows:
 T  T
B1k = 0.05 0.05 0.05 , B2k = 0.01 0.01 ,
 T
B3k = 0.01 0.01 , r1 = −0.5, r2 = 0.5, (6.63)
 
1 + 0.1sin(k) 0 0
Ck = , Qk = 0.05, Rk = Sk = 0.01
0 1 0

and other system parameters are listed in Table 6.1.

The order of the fading measurement model (6.2) is set to be M = 2 and
the probability density functions of channel coefficients are chosen as

TABLE 6.1
System Parameters

Parameters Symbol Value

Cross section area of tanks Sa 154 cm2

Cross section area of pipes Sn 0.5 cm2
Sampling period T 8s
Max. height of tanks Hmax 620 cm
Max. flow rate of tank 1 Q1 max 93 cm3 /s
Max. flow rate of tank 2 Q2 max 91 cm3 /s
The control law of tank 1 Q1 20 cm3 /s
The control law of tank 2 Q2 21 cm3 /s
The outflow coefficient of pipe 1 az1 0.48
The outflow coefficient of pipe 2 az2 0.58
The outflow coefficient of pipe 3 az3 0.48
134 Data-Rate-Constrained State Estimation and Control



 p0 (α0k ) = 0.0005(e9.8933α0k − 1), 0 ≤ α0k ≤ 1



 (
 10α1k , 0 ≤ α1k ≤ 0.2
p1 (α1k ) = (6.64)


 − 2.5(α1k − 1), 0.2 < α1k ≤ 1


 p (α ) = 8.5017e−8.5α2k ,

0 ≤ α ≤ 1,
2 2k 2k

from which it can be verified that mathematical expectations ᾱm of αmk

(m = 0, 1, 2) are 0.8991, 0.4000 and 0.1174, and the variances ςm are 0.0133,
0.0467 and 0.01364, respectively. In addition, the successful arrival rates
of the two description packets from the encoder to the decoder are set as
Prob{λik = 1} = 0.9 (i = 1, 2). Thus, the corresponding mathematical expecta-
tions and variances are easily obtained as λ̄i = 0.9 and σi = 0.09 (i = 1, 2),
respectively. For the uniform quantizer (6.9), let the saturation value be L = 1,
the number of the quantization level be ρ = 100 and the adjustable parameter
be µlk = 1. The sampling period T is taken as T = 13.
Letting xk = [x1k x2k x3k ]T = [h1k h2k h3k ]T , in this simulation, the
initial heights of the tanks and their estimates are, respectively, set as x0 =
[0.45 0.14 0.18]T and x̂0 = [0.15 0.04 0.08]T . At each time step k, for
the ith component (i = 1, 2, 3), define the filtering error (FE) as FEi ,
p
(xik − x̂ik )2 .
In order to illustrate the superiority of the proposed filtering scheme, the
performances (in terms of the upper bound of the filtering error (UBFE))
of the MDC-based filter and the single-description-coding-based filter are
compared in the simulation. Moreover, for the sake of revealing the effects
of the system complexities (e.g. the channel fading, the process noises and
the probability of the packet dropout) on the filtering performance, some
comparison simulations are also constructed.
The simulation results are shown in Figures 6.4–6.11. Figure 6.4 describes
the measurements received by the encoder and the filter. Figure 6.5 plots
the actual level heights of the tanks and their estimates. Figures 6.6–6.8
illustrate the FEi (i = 1, 2, 3) and their upper bounds. In Figure 6.9, the upper
bounds of the filtering error for each component with single description
coding scheme and two description coding scheme are, respectively, plotted,
which shows that the filtering performance with the two description coding
is superior to the single description coding case. Figure 6.10 gives the effect
of different channel coefficients on the UBFE, which indicates that, as the
fading parameter M grows from 0 (without channel fading) to 2, the UBFE
increases accordingly. The influence of the process noise is reflected in
Figure 6.11, where the UBFE increases with the growth of the noise intensity.
In Figure 6.12, the impact of the packet-arrival probability is exhibited,
which verifies that a higher arrival probability leads to a better filtering
performance. It is shown from the simulation results that the proposed
Recursive Filtering with Measurement Fading 135

0.8

0.6

0.4

0.2

0
0 100 200 300 400 500 600 700 800 900 1000

0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600 700 800 900 1000

FIGURE 6.4
The fading measurement components yik (i = 1, 2) and their decoded values ȳik .

0.4

0.2

0
0 100 200 300 400 500 600 700 800 900 1000

0.2

0.1

0
0 100 200 300 400 500 600 700 800 900 1000

0.2

0.1

0
0 100 200 300 400 500 600 700 800 900 1000

FIGURE 6.5
The actual state components xik (i = 1, 2, 3) and their estimates x̂ik .
136 Data-Rate-Constrained State Estimation and Control

0.3

0.25

0.2

0.15

0.1

0.05

0
0 100 200 300 400 500 600 700 800 900 1000

FIGURE 6.6
The FE of x1k and its upper bound.

0.3

0.25

0.2

0.15

0.1

0.05

0
0 100 200 300 400 500 600 700 800 900 1000

FIGURE 6.7
The FE of x2k and its upper bound.
Recursive Filtering with Measurement Fading 137

0.3

0.25

0.2

0.15

0.1

0.05

0
0 100 200 300 400 500 600 700 800 900 1000

FIGURE 6.8
The FE of x3k and its upper bound.

0.3

0.2

0.1
0 20 40 60 80 100 120 140 160 180 200

0.3
0.07
0.2 0.06
0.05
0.1 50 100 150

20 40 60 80 100 120 140 160 180 200

0.2

0.1
0 20 40 60 80 100 120 140 160 180 200

FIGURE 6.9
The upper bounds of filtering errors of xik (i = 1, 2, 3) with single description and two
descriptions.
138 Data-Rate-Constrained State Estimation and Control

0.25
0.2
0.15
0.1
0 100 200 300 400 500 600 700 800 900 1000

0.2
0.07
0.15
0.06
0.1 0.05
300 350 400

0 100 200 300 400 500 600 700 800 900 1000

0.2

0.15

0.1

0 100 200 300 400 500 600 700 800 900 1000

FIGURE 6.10
The effect of fading channel parameter M on filtering performance.

0.3
0.13
0.125
0.12
0.2 0.115
300 400 500
0.1
0 100 200 300 400 500 600 700 800 900 1000

0.2
0.065
0.15
0.06
0.1
300 400 500
0 100 200 300 400 500 600 700 800 900 1000

0.2
0.1
0.15 0.09
0.08
0.1 300 400 500

0 100 200 300 400 500 600 700 800 900 1000

FIGURE 6.11
The filtering errors of xik (i = 1, 2, 3) and their upper bounds with different noise intensities.
Recursive Filtering with Measurement Fading 139

0.3

0.2

0.1
0 100 200 300 400 500 600 700 800 900 1000
Time (k)
0.2

0.15

0.1

0 100 200 300 400 500 600 700 800 900 1000
Time (k)

0.2

0.15

0.1
0 100 200 300 400 500 600 700 800 900 1000
Time (k)

FIGURE 6.12
The filtering errors of xik (i = 1, 2, 3) and their upper bounds with different packet-arrival
probabilities.

0.3

0.2

0.1
0 100 200 300 400 500 600 700 800 900 1000

0.2

0.15

0.1

0 100 200 300 400 500 600 700 800 900 1000

0.2

0.15

0.1
0 100 200 300 400 500 600 700 800 900 1000

FIGURE 6.13
The effect of the model uncertainty on the filtering performance.
140 Data-Rate-Constrained State Estimation and Control

MDC-based filtering scheme performs very well and the considered system
complexities do have a major effect on the filtering performance.
In order to examine the sensitivity of the developed filter to the model
uncertainty, a simulation test has been conducted. In this case, the terms
hi (tk ) (i = 1, 2, 3) in (6.61) have been replaced by (1 + ℜi (tk ))hi (tk ) where
ℜi (tk ) (i = 1, 2, 3) represent the parameter uncertainties and are taken as
ℜi (tk ) = ℜ̄i exp ( − 0.1tk ) with ℜ̄i being positive scalars. It is easy to see that ℜ̄i
satisfies |ℜi (tk )| ≤ ℜ̄i . It is observed from Figure 6.13 that as the degree of the
model uncertainties increases (i.e. the increase of ℜ̄i ), the upper bound of the
filtering error variance grows accordingly, which indicates that the proposed
filter is sensitive to the model uncertainties. One of our future investigations
would be focused on the improvement of the robustness for the developed
filtering scheme.

6.5 Summary
In this chapter, a multiple description coding scheme has been employed to
deal with the recursive filtering issue for a class of discrete-time nonlinear
stochastic networked systems. The Mth-order Rice fading model has been
adopted to describe the measurement fading from the sensor to the encoder.
By utilizing the distorted data resulting from the channel fading, coding
process and measurement noises, a recursive sub-optimal filter has been
designed to estimate the actual system state. In addition, a minimized upper
bound of filtering error variance has been derived and the corresponding
filter parameter has been explicitly determined by solving the coupled
algebraic Riccati-like difference equations. Finally, a simulation example has
been given to demonstrate the effectiveness of the proposed filtering strategy.
7
Stabilization of Linear Discrete-Time
Systems over Resource-Constrained Networks
under Dynamical Multiple Description
Coding Scheme

In networked control systems (NCSs), the communication among compo-

nents is realized via a shared network, thereby merits decreased system
wiring, easy installation and low maintenance cost. In NCS-related research,
the scarcity of communication resources is one of the main concerns that lead
to imperfect network communication and subsequently the degradation of
the overall system performance. In fact, NCSs have been acting as a research
frontier for almost two decades and a large amount of literature has been
available on the control/filtering problems in networked environments, see
for example [6,11, 187].
Network-induced phenomena (NIP), as a consequence of limited com-
munication capacity of NCSs, are likely to occur during data transmission
and pose a major impact on the system stability. An effective way of
mitigating the occurrence frequency of the undesirable NIPs is to improve the
communication quality of the shared channel by deploying the transmission
scheduling protocols/mechanisms. Up to now, much research effort has
been devoted to the analysis and/or design problems for NCSs subject
to scheduling protocols with some pioneering work appearing in the
literature. Roughly speaking, the aforementioned protocols/mechanisms
adopt the following two ways to improve the efficiency of utilizing network
resources: 1) reducing communication frequency (e.g. static, dynamic or
adaptive event-triggering mechanisms) and 2) reducing the communication
traffic (e.g. Round-Robin, random access and Try-Once-Discard protocols).
These protocols/mechanisms are, however, based on an implicit assumption
that the signals scheduled by the communication protocols are successfully
transmitted without any distortions, and this assumption is often unrealistic
in some engineering practices. To ensure distortion-free transmissions, a
huge communication resource would be occupied (which is especially
true for high precision/amplitude signals), and such a resource overuse

DOI: 10.1201/9781003534853-7 141

142 Data-Rate-Constrained State Estimation and Control

is simply unaffordable in a typically bit-rate-constrained communication

environment.
A recently popular strategy for data transmission is the data coding
scheme that aims at converting the raw data into specific codewords with
less bit occupancy according to a specific rule. It should be noted that
most of the existing coding schemes are only valid on the premise of
perfect codeword transmissions, that is, no packet loss occurs during the
codeword transmission. This premise is, unfortunately, sometimes incorrect
in high-throughput signal transmissions due primarily to the unreliable
network circumstance. As such, it appears both theoretically significant and
practically important to further improve existing coding schemes with the
aim to cope with the unreliability of the codeword transmission.
With the aspiration to handle the unreliability induced by the packet
dropout phenomenon during data transmission, the so-called multiple
description coding (MDC) scheme has been developed which comprises
the following three steps: 1) the encoder maps the raw data into several
descriptions with fewer bit occupancies and identical importance; 2) the
descriptions are transmitted to the decoder side through individual
channels; and 3) the decoder restores the decoded data in terms of the
received descriptions. Therefore, the primary motivation of current chapter
is to generalize the MDC strategy to examine the trade-off between
communication capacity and system stability.
Following the discussions made thus far, we conclude that there appears
to be a lack of systematic investigation on the stabilization problem for a
class of linear discrete-time NCSs subject to the randomly occurring packet
dropouts and bit-rate constraints. As such, the main objective of this paper is
to bridge such a gap by developing an MDC strategy capable of alleviating
the negative impact from randomly occurring packet dropouts and limited
network resources on system stability. This task is fairly demanding as we
are confronted with the following technical challenges: 1) how to construct
an appropriate coding-decoding framework to ensure that the decoding error
converges to 0 asymptotically? 2) how to develop a control scheme capable of
effectively handling the distorted signal while achieving the desired stability
condition? and 3) how to quantitatively analyze the influences from the
packet-dropout probabilities (of the channels and the encoder parameters)
on the system stability?

7.1 Problem Formulation

Consider the following linear discrete-time system:

x(k + 1) = Ax(k) + Bu(k) (7.1)

Stabilization of Linear Discrete-Time Systems 143

where x(k) ∈ Rn is the system state, u(k) ∈ Rr is the control input signal, and
A and B are known matrices with appropriate dimensions.
In an NCS framework, the system states are transmitted to the controller
via a resource-constrained communication media. Accordingly, the MDC
scheme is introduced with a view to enhancing the reliability of signal
transmission. Without loss of generality, we consider the case of two
descriptions in this paper. To be more specific, the data is first encoded
into two descriptions in terms of certain coding criterion, and the obtained
descriptions are subsequently transmitted to the decoding devices via the
two independent channels (labeled as “C1” and “C2”), respectively. In
addition, the component-based MDC is adopted in this paper in order to
meet the requirement of encoding each component xi (k) (i ∈ N , {1, 2, . . . , n})
of the system state x(k).
For the ith state component, a scalar-valued encoder is constructed as
follows:
(

εi (τ h) = f1i xi (τ h)

(7.2)
σi (τ h) = f2i xi (τ h)

for i ∈ N, where f1i ( · ) : R 7 → R and f2i ( · ) : R 7 → R are two coding functions,

εi (τ h) ∈ R and σi (τ h) ∈ R are two scalar-valued descriptions of xi (τ h)
(τ = 0, 1, 2, . . . ), and h ∈ N+ is the coding period.
The acquired descriptions εi (τ h) and σi (τ h) are, respectively, packed as two
data packets ε(τ h) and σ (τ h) with
 T
ε(τ h) , ε1 (τ h) ε2 (τ h) · · · εn (τ h)
 T
σ (τ h) , σ1 (τ h) σ2 (τ h) · · · σn (τ h) ,

which are correspondingly transmitted to the decoder via channels

“C1” and “C2”.
Due to the inevitable unreliability of the communication media, the
packet dropout phenomenon is frequently encountered during the packet
transmission. In this situation, for i ∈ N, the corresponding scalar-valued
decoder is characterized as follows:


 g1i εi (τ h) , if γ1 (τ h) = 1, γ2 (τ h) = 0


x̂i (τ h) = g2i σi (τ h) , if γ1 (τ h) = 0, γ2 (τ h) = 1 (7.3)



gci εi (τ h), σi (τ h) , if γ1 (τ h) = γ2 (τ h) = 1

where g1i ( · ) and g2i ( · ) are two side decoding functions, and gci (·, ·) is the
central decoding function. x̂i (τ h) is the decoded value of xi (τ h). The random
variables γi (τ h) (i = 1, 2) are two Bernoulli-distributed sequences with the
following probability distributions:
144 Data-Rate-Constrained State Estimation and Control

Prob{γi (τ h) = 1} = γ̄i , Prob{γi (τ h) = 0} = 1 − γ̄i (7.4)

where γ̄i ∈ [0, 1] (i = 1, 2) are given constants.

For presentation clarity, we set
 T
x̂( · ) , x̂1 ( · ) x̂2 ( · ) · · · x̂n ( · )
 T
f1 (x( · )) , f11 (x1 ( · ) f12 (x2 ( · )) · · · f1n (xn ( · ))
 T
f2 (x( · )) , f21 (x1 ( · )) f22 (x2 ( · )) · · · f2n (xn ( · ))
 T (7.5)
g1 (ε( · )) , g11 (ε1 ( · )) g12 (ε2 ( · )) · · · g1n (εn ( · ))
 T
g2 (σ ( · )) , g21 (σ1 ( · )) g22 (σ2 ( · )) · · · g2n (σn ( · ))
 T
gc (ε( · ), σ ( · )) , gc1 (ε1 ( · ), σ1 ( · )) · · · gcn (εn ( · ), σn ( · )) .

The scalar-valued encoder-decoder pair (7.2)–(7.3) can be compacted into the

following form: (
ε(τ h) =f1 (x(τ h))
(7.6)
σ (τ h) =f2 (x(τ h))
and 
 g1 (ε(τ h)), if γ1 (τ h) = 1, γ2 (τ h) = 0


 g2 (σ (τ h)), if γ1 (τ h) = 0, γ2 (τ h) = 1

x̂(τ h) = (7.7)
 gc (ε(τ h), σ (τ h)), if γ1 (τ h) = γ2 (τ h) = 1




g(x̂(τ h − 1)), if γ1 (τ h) = γ2 (τ h) = 0
where xˆ(τ h) is the vector-valued decoded data and g( · ) is a function of
x̂(τ h − 1) which will be given later.

Remark 7.1: It is observed from (7.4) that the random variables γi (τ h) (i = 1, 2)

are employed to regulate the randomly occurring packet-dropout phenomenon in
the communication channels “C1” and “C2”. In accordance with the MDC scheme
(7.6)–(7.7), at each coding time step τ h, the raw signal x(τ h) is encoded into two
descriptions ε(τ h) and σ (τ h) by the coding functions f1 ( · ) and f2 ( · ), respectively.
Then, ε(τ h) and σ (τ h) are individually transmitted to the decoder side via com-
munication channels “C1” and “C2”. As shown in Figure 7.1, if no packet-dropout
occurs, that is, γ1 (τ h) = γ2 (τ h) = 1, which means that the description packets ε(τ h)
and σ (τ h) are successfully transmitted and the central decoder (labeled as “CD”)
is triggered to generate the decoded value. If the packet-dropout occurs in channel
“C1” only (i.e. γ1 (τ h) = 0, γ2 (τ h) = 1), the description packet σ (τ h) is available
and the side decoder 2 (labeled as “SD2”) is activated to execute the decoding
procedure. If the packet-dropout occurs in channel “C2” only (i.e. γ1 (τ h) = 1,
Stabilization of Linear Discrete-Time Systems 145

FIGURE 7.1
The structure of the closed-loop control system with the MDC scheme.

γ2 (τ h) = 0), the description packet ε(τ h) is available and the side decoder 1
(labeled as “SD1”) is enabled to perform the decoding operation. Furthermore, if the
packet-dropout occurs in both channels “C1” and “C2” (i.e. γ1 (τ h) = γ2 (τ h) = 0),
none of the description packets is available and all the decoders fail to work. In this
case, without loss of generality, we assume that the decoded value x̂(τ h − 1) is used
to generate the control signal.

In this paper, the remote control scenario is considered that the encoder and
controller are deployed at different places. The schematic structure is shown
in Figure 7.1, where the network-based communication scheme is employed
and descriptions generated by two encoders are transmitted to the remote
controller via bandwidth-constrained network channels. The main purpose
of this paper is to investigate the remote control issue under the dynamical
MDC scheme. In particular, we aim to

• design the coding and decoding procedures based on the MDC

scheme;
• guarantee the convergence of the decoding error in the statistical
sense;
• provide sufficient conditions on the stochastic stability of the closed-
loop system; and
• reveal the influences of the major factors (e.g. system parameters,
encoding-decoding scheme and the probabilities of the channel
packet dropouts) on the stochastic stability of the closed-loop system.

7.2 Main Results

Coding Scheme
In this subsection, we are devoted to formalizing the coding procedure based
on the MDC scheme. The coding procedure developed in this paper includes
146 Data-Rate-Constrained State Estimation and Control

two steps, namely, the index generation step and the index assignment
step. For the step of index generation, the uniform quantization method is
employed to generate the indices and, for the step of index assignment, the
nested index assignment principle [156] is adopted to assign the generated
indices to a certain mapping matrix.
In the sequel, we slightly abuse the notation by using x, ε, σ , εi , σi and xi to
denote x(τ h), ε(τ h), σ (τ h), εi (τ h), σi (τ h) and xi (τ h), respectively.
Index Generation Scheme
For a scalar quantizer Q(·) : R 7→ R, the scaling parameter s and the positive
integer N correspond to the quantization range and the quantization level,
respectively.
Let us divide the hyperrectangles Hs = {x ∈ Rn : |x|∞ ≤ s} into Nn parts
H1,p1 (s) × H2,p2 (s) × · · · × Hn,pn (s) with equal volume, where p1 , p2 , . . . , pn ∈
{1, 2, . . . , N} and
 
2s
Hi,1 (s) , xi | − s ≤ xi < −s +
N
 
2s 4s
Hi,2 (s) , xi | − s + ≤ xi < −s +
N N
..
. (7.8)
 
2s
Hi,N (s) , xi |s − ≤ xi ≤ s
N

with i = 1, 2, . . . , n. Based on the above partition, we can see that, for an

arbitrary x ∈ Hs , there exists a sequence of positive scalars p1 , p2 , . . . , pn
such that x belongs to the sub-hyperrectangle H1,p1 (s) × H2,p2 (s) × · · · ×
Hn,pn (s). Thus, a vector-valued index p , [p1 p2 · · · pn ]T is exclusively
to determine the sub-hyperrectangle H1,p1 (s) × H2,p2 (s) × · · · × Hn,pn (s).
Having established the relationship between the sub-hyperrectangle
H1,p1 (s) × H2,p2 (s) × · · · × Hn,pn (s) and the vector p, we now construct the
index generator functions φi ( · ) (i = 1, 2, . . . , n) : R 7→ {1, 2, . . . , N} as
follows:

φi (xi ) = pi (7.9)

for x ∈ H1,p1 (s) × H2,p2 (s) × · · · × Hn,pn (s).

In light of (7.8), it is easy to verify that the center of the sub-hyperrectangle
H1,p1 (s) × H2,p2 (s) × · · · × Hn,pn (s) is
h iT
(2p1 −1)s (2p2 −1)s (2pn −1)s
ξ(p, s) , −s + N −s + N . . . −s + N
. (7.10)
Stabilization of Linear Discrete-Time Systems 147

Noticing the fact that x ∈ H1,p1 (s) × H2,p2 (s) × · · · × Hn,pn (s), one directly
obtains the following relationship:

s
|x − ξ(p, s)|∞ ≤ . (7.11)
N

Index Assignment Principle

Inspired by the nested assignment principle, we shall assign the generated
index pi associated with the component xi into the certain cell of the matrix
Mi . Without loss of generality, the mapping matrix Mi is assumed to be an
l-dimensional matrix with l being an even integer. Subsequently, in terms of
the locations of the row and the column of the cell containing pi , a pair of
descriptions εi and σi is thus obtained, where εi denotes the row location of
pi and σi stands for its column location.
The index assignment function is given as follows:

8i (pi ) = (ϕ1i (pi ), ϕ2i (pi )) , (εi , σi ) (7.12)

where ϕ1i ( · ) and ϕ2i ( · ) are the row assignment function and the column
assignment function, respectively, and εi and σi are two descriptions of xi .

Remark 7.2: As stated in [156], the nested assignment principle stipulates that the
indices are placed in the cells that lie on the main diagonal and its nearest 2d diagonals
of the mapping matrix. For the convenience of the exposition, the scenario d = 1 is
discussed in this paper, which means that the indices are placed in the cells lying
on the main diagonal and its nearest 2 diagonals of the mapping matrix. Figure 7.2
gives an example of the nested assignment principle with l = 8 and d = 1.

FIGURE 7.2
Nested index assignment for l = 8, d = 1.
148 Data-Rate-Constrained State Estimation and Control

The specific form of the index assignment function 8i ( · ) : N+ 7→ N+ × N+

is proposed as follows:

8i (pi ) ,(εi , σi )

(r + 1, ri + 1), if qi =1
 i


 (ri + 1, ri ), if qi = 0 and ri is even


(7.13)

= (ri , ri + 1), if qi = 0 and ri is odd

(ri + 2, ri + 1), if qi = 2 and ri is even





(ri + 1, ri + 2), if qi = 2 and ri is odd


p p
where ri = ⌊ 2d+1
i
⌋ and qi = h 2d+1
i
i. As such, it is not difficult to verify that all
the indices pi are mapped to the corresponding description pairs (εi , σi ).

7.2.1 Decoding Scheme

Due to the possible dropout of the description packets (i.e. ε and σ ) during
the data transmission, it is difficult to obtain an accurate value of the
vector-valued index p. In view of this, an index estimation strategy, which
is dependent on the reception of the descriptions, is first put forward to
estimate the index of the encoded signal.
Index Estimation Strategy
The index pi , which carries essential information of xi , plays an important
role in the design of the decoding procedure. Moreover, it is inferred from
the index assignment function (7.13) that an index pi exclusively determines
by two descriptions εi and σi , and vice versa. Unfortunately, the exact value
of index pi cannot be always accessible at the decoder side due to the
probabilistic packet dropout during the transmission of the descriptions. In
this sense, it is necessary to develop an index estimation strategy with the
hope to obtain an estimation of the index pi in terms of the reception of the
descriptions.
Case I: both the description packets ε and σ are successfully transmitted to
the central decoder. In this case, the index estimation function ψ(·, ·) : Z+ ×
Z+ 7→ Z+ for each index pi (i = 1, 2, . . . , n) is given as

 3εi − 2, if εi = σi


 3εi − 3, if εi = σi + 1 and εi is odd



ψ(εi , σi ) = 3εi , if εi = σi − 1 and εi is odd (7.14)

3εi − 4, if εi = σi + 1 and εi is even





3εi − 1, if εi = σi − 1 and εi is even


and thus, the value of ψ (εi , σi ) (denoted as p̂ci ) is the estimation of the index
pi . It is obvious that the estimation accuracy is 100%, namely, p̂ci = pi .
Stabilization of Linear Discrete-Time Systems 149

Case II: only one of the description packet (either ε or σ ) is received by the
side decoder. In this situation, the index estimation function ψ( · ) : Z+ 7→ Z+
for the index component pi is defined by

ψ(z) = 3z − 2, ∀ z ∈ {εi , σi }, (7.15)

and the value of ψ(z) is denoted as p̂si .

Case III: neither description packet is transmitted to the decoder. In this
scenario, the estimation for the index pi is set as 0.
Bearing pi = 3ri + qi in mind, it follows from (7.13)–(7.15) that the index
estimation errors p̃ci , pi − p̂ci and p̃si , pi − p̂si satisfy p̃ci = 0 for Case I and
|p̃si | ≤ 2 for Case II, respectively.

7.2.2 Design of the Encoder-Decoder Pair

After discussing the generation, assignment and estimation of the index
pi (i = 1, 2, . . . , n), we are in a position to specifically design the coding-
decoding procedure via a dynamical quantization approach.
Before presenting the encoder-decoder structure, let us introduce a random
sequence θ(τ h) , γ1 (τ h) + γ2 (τ h) satisfying the following probability
distribution:

Prob{θ (τ h) = 0} = θ̄0
Prob{θ (τ h) = 1} = θ̄1 (7.16)
Prob{θ (τ h) = 2} = θ̄2

where

θ̄0 , (1 − γ̄1 )(1 − γ̄2 )

θ̄1 , γ̄1 (1 − γ̄2 ) + (1 − γ̄1 )γ̄2
θ̄2 , 1 − θ̄0 − θ̄1 .

Encoder: at each coding instant τ h (τ = 0, 1, 2, . . . ), for the system state x(τ h),
the coding function is designed as follows:
(
εi (τ h) , ϕ1i (φi (x̃i (τ h)))
(7.17)
σi (τ h) , ϕ2i (φi (x̃i (τ h))), i = 1, 2, . . . n

where x̃i (τ h) is the ith component of x̃(τ h) and x̃(τ h) , x(τ h) − x̄(τ h) ∈
H1,p1 (s(τ h)) × H2,p2 (s(τ h)) × · · · × Hn,pn (s(τ h)). Moreover, the dynamics of
x̄(k) satisfies
150 Data-Rate-Constrained State Estimation and Control

(
x̄(k) = Ax̆(k − 1) + Bŭ(k − 1), ∀ k > 0
(7.18)
x̄(0) = 0

and the dynamics of x̆(k) is governed by the following auxiliary system:


 x̆(k + 1) = Ax̆(k) + Bŭ(k), k 6 = τ h − 1


 x̆(τ h) = Ax̆(τ h − 1) + Bŭ(τ h − 1)



+ ξ(p(τ h), s(τ h)) (7.19)




 ŭ(k) = Kx̆(k)

x̆(0) = 0


where ξ(·, ·) is defined in (7.10) and K is the controller gain matrix to be

designed later.
It follows from (7.2), (7.5) and (7.17) that

f1i (xi (τ h)) = ϕ1i (φi (x̃i (τ h))), f1 (x(τ h)) = ϕ1 (φ(x̃(τ h)))
f2i (x(τ h)) = ϕ2i (φi (x̃i (τ h))), f2 (x(τ h)) = ϕ2 (φ(x̃(τ h)))

where

ϕ1 (φ(x̃(τ h))) , [ϕ11 (φ1 (x̃1 (τ h))) · · · ϕ1n (φn (x̃n (τ h)))]T
ϕ2 (φ(x̃(τ h))) , [ϕ21 (φ1 (x̃1 (τ h))) · · · ϕ2n (φn (x̃n (τ h)))]T .

According to the reception of the description packets ε(τ h) and σ (τ h) at the

decoder side, the estimation model of the index vector p(τ h) is described by

p̂(τ h) = δ{θ (τ h), 2}p̂c (τ h) + δ{θ(τ h), 1}p̂s (τ h) (7.20)

where

p̂c (τ h) , [p̂c1 (τ h) p̂c2 (τ h) · · · p̂cn (τ h)]T

p̂s (τ h) , [p̂s1 (τ h) p̂s2 (τ h) · · · p̂sn (τ h)]T

and δ{·, ·} is the Kronecker delta function defined as

(
1, s = t
δ{s, t} =
0, s 6 = t,

Remark 7.3: It is observed from (7.16) and (7.20) that: 1) θ(τ h) = 2 indicates the
successful transmissions of both ε(τ h) and σ (τ h), under which the estimated index
p̂c (τ h) is obtained; 2) θ (τ h) = 1 implies that one of the packet (either ε(τ h) or
Stabilization of Linear Discrete-Time Systems 151

σ (τ h)) is lost during the transmission, and hence the index p̂s (τ h) is estimated; and
3) θ(τ h) = 0 means the simultaneous transmission failures of both ε(τ h) and σ (τ h),
under which the estimation of the index vector p(τ h) is set as 0, namely, p̂(τ h) = 0.

In terms of the estimation index p̂(τ h) obtained at the decoder

side, one derives the center of the sub-hyperrectangle H1,p1 (τ h) (s(τ h)) ×
H2,p2 (τ h) (s(τ h)) × · · · × Hn,pn (τ h) (s(τ h)) as ξ(p̂(τ h), s(τ h)). Moreover, recalling
the definitions of ξ(·, ·) in (7.10) and p̂(τ h) in (7.20), ξ(p̂(τ h), s(τ h)) can be
further rewritten as

ξ(p̂(τ h), s(τ h)) = δ{θ (τ h), 2}ξ(p̂c (τ h), s(τ h))
(7.21)
+ δ{θ (τ h), 1}ξ(p̂s (τ h), s(τ h)),

which infers that ξ(p̂(τ h), s(τ h)) is set to be 0 when both of the vector-valued
descriptions (i.e. ε(τ h) and σ (τ h)) are lost (i.e. θ (τ h) = 0).
It should be emphasized that the model (7.21) reflects all the cases of the
reception of ε(τ h) and σ (τ h) at the decoder side, based on which we intend
to tackle with the decoder design issue.
Decoder: at the decoder side, the decoding function is designed with the
following form:


 x̂(k + 1) = Ax̂(k) + Bu(k), k 6 = τ h − 1

 x̂(τ h) = Ax̂(τ h − 1) + Bu(τ h − 1)



+ ξ(p̂(τ h), s(τ h)) (7.22)




 u(k) = Kx̂(k)

x̂(0) = 0


where u(k) is the decoded-value-based control input of the plant (7.1).

Denoting the decoding error as e(k) , x(k) − x̂(k), the closed-loop system
is obtained as follows:

x(k + 1) = Ax(k) − BKe(k) (7.23)

where A , A + BK.

Definition 7.1: The discrete-time closed-loop system (7.23) is said to be stochasti-

cally stable if, for any initial condition x(0) ∈ Rn , the system dynamics satisfies

lim E{kx(k)k} = 0, k = 1, 2, . . . . (7.24)

k→∞

Next, we shall pay our attention to developing a new dynamical multiple

description coding scheme such that the closed-loop system (7.23) is
stochastically stable under the decoded-value-based control input u(k).
152 Data-Rate-Constrained State Estimation and Control

For convenience, some notations are defined as follows:

xE1 (k) , x(k) − x̆(k), xE2 (k) , x̂(k) − x̆(k)

s0 , Nkx(0)k∞ , µh , kAh k∞ , µ̄h , kAh k∞
h
X
ζh , Ah−i BKAi−1 .
i=1 ∞

Before proceeding, we introduce the following lemmas which will be

utilized in later discussions.

Lemma 7.1: Let the coding procedure (7.17)–(7.19) be given. The following
relationship

E{kx(τ h) − x̄(τ h)k∞ } ≤ s(τ h)

is guaranteed for all τ = 1, 2, . . ., where s(τ h) is governed by the iteration

¯h
µ ¯
s(τ h) = s((τ − 1)h) + ζh b((τ − 1)h) (7.25)
N

with s(0) , s0 and b̄(τ h) , E{kxE2 (τ h)k∞ }.

Proof: The proof is conducted by mathematical induction.

For τ = 1, one has

x(h) − x̄(h) = AxE1 (h − 1) + BKxE2 (h − 1)

= A2 xE1 (h − 2) + ABKxE2 (h − 2) + BKxE2 (h − 1)

= Ah xE1 (0) + Ah−1 BKxE2 (0) + Ah−2 BKxE2 (1)

+ Ah−3 BKxE2 (2) + · · · + ABKxE2 (h − 2)

+ BKxE2 (h − 1).

Noting the fact that

xE2 (i) = Ai xE2 (0), (i = 1, 2, . . . , h − 1),

we obtain
h
X
¯
x(h) − x(h) = Ah xE1 (0) + Ah−i BKAi−1 xE2 (0),
i=1
Stabilization of Linear Discrete-Time Systems 153

which gives

¯h
µ
E{kx(h) − x̄(h)k∞ } ≤ s(0) + ζh b̄(0).
N

Assuming that the inequality holds for τ = 2, 3, · · · , j, we explore the case of

τ = j + 1. Similar to the situation of τ = 1, one finds

x((j + 1)h) − x̄((j + 1)h) = AxE1 ((j + 1)h − 1) + BKxE2 ((j + 1)h − 1)
= A2 xE1 ((j + 1)h − 2) + ABKxE2 ((j + 1)h − 2)
+ BKxE2 ((j + 1)h − 1)
= Ah xE1 (jh) + Ah−1 BKxE2 (jh) + Ah−2 BKxE2 (jh + 1)
+ Ah−3 BKxE2 (jh + 2) + · · · + ABKxE2 ((j + 1)h − 2)
+ BKxE2 ((j + 1)h − 1).

By using the fact xE2 (i) = Ai−jh xE2 (jh) for i = jh + 1, jh + 2, · · · , (j + 1)h − 1 again,
one has
h
X
x((j + 1)h) − x̄((j + 1)h) = Ah xE1 (jh) + Ah−i BKAi−1 xE2 (jh),
i=1

which implies

E{kx((j + 1)h) − x̄((j + 1)h)k∞ } ≤ µ̄h E{kxE1 (jh)k∞ } + ζh b̄(jh).

In addition, it is inferred from (7.11) that

E{kxE1 (jh)k∞ } = E{kx(jh) − x̄(jh) − ξ(p(jh), s(jh))k∞ }

1
≤ s(jh).
N

Consequently, we have

µ¯n ¯ jh)
E{kx((j + 1)h) − x̄((j + 1)h)k∞ } ≤ s(jh) + ζh b(
N
= s((j + 1)h),

which completes the proof.

It should be pointed out that since the iteration (7.25) with the given initial
condition and system parameters A, B, K, h and N are known a prior
to both the encoder and the decoder, s(τ h) can be calculated and kept
154 Data-Rate-Constrained State Estimation and Control

synchronized at both the encoder side and the decoder side. This means
that the synchronization of the quantization interval s(τ h) can be always
guaranteed at the encoder side and the decoder side no matter whether the
packet dropout occurs or not.

Lemma 7.2: Let the coding procedure (7.17)–(7.19) and the decoding procedure
(7.22) be given. The error between the auxiliary system state x̆(k) and the decoded
value x̂(k) satisfies

E{kx̆(τ h) − x̂(τ h)k∞ } ≤ b(τ h),

where b(τ h) is determined by the recursive rule

b(τ h) = µh b((τ − 1)h) + νs(τ h)

4 ¯
with b(0) , 0, ν , N θ1 + θ̄0 and τ = 1, 2, . . ..

Proof: The mathematical induction is applied again to the proof of this lemma.
For τ = 1, it follows from (7.19) and (7.22) that

E{kx̆(h) − x̂(h)k∞ } ≤ µh kx̆(0) − x̂(0)k∞ + E{kξ(p(h), s(h))−ξ(p̂(h), s(h))k∞ }

= kξ(p(h), s(h)) − ξ(p̂c (h), s(h))k∞ × Prob{θ(h) = 2}
+ kξ(p(h), s(h)) − ξ(p̂s (h), s(h))k∞ × Prob{θ(h) = 1}
+ kξ(p(h), s(h))k∞ × Prob{θ(h) = 0} + µh kx̆(0) − x̂(0)k∞
≤ νs(h) + µh b(0)

and

E{kx̆(h) − x̂(h)k∞ } ≤ µh kx̆(0) − x̂(0)k∞ + E{kξ(p(h), s(h))−ξ(p̂(h), s(h))k∞ }

= kξ(p(h), s(h)) − ξ(p̂c (h), s(h))k∞ × Prob{θ(h) = 2}
+ kξ(p(h), s(h)) − ξ(p̂s (h), s(h))k∞ × Prob{θ(h) = 1}
+ kξ(p(h), s(h))k∞ × Prob{θ(h) = 0}
+ µh kx̆(0) − x̂(0)k∞
≤ µh b(0) + νs(h).

Assume that the inequality holds when τ = 2, 3, . . . , j. Then, for τ = j + 1, we have

E{kx̆((j + 1)h) − x̂((j + 1)h)k∞ }

= E{kAx̆((j + 1)h − 1) + ξ(p((j + 1)h), s((j + 1)h))
− Ax̂((j + 1)h − 1)
− δ{θ ((j + 1)h), 1} × ξ(p̂c ((j + 1)h), s((j + 1)h))
Stabilization of Linear Discrete-Time Systems 155

− δ{θ ((j + 1)h), 2} × ξ(p̂s ((j + 1)h), s((j + 1)h))k∞ }

≤ E{µh kx̆((jh) − x̂(jh)k∞ + kξ(p((j + 1)h), s((j + 1)h))
− δ{θ ((j + 1)h), 1} × ξ(p̂c ((j + 1)h), s((j + 1)h))
− δ{θ ((j + 1)h), 2} × ξ(p̂s ((j + 1)h), s((j + 1)h))k∞ }. (7.26)

By tedious calculations, one arrives at

E{kξ(p((j + 1)h), s((j + 1)h)) − δ{θ ((j + 1)h}, 1)

× ξ(p̂c ((j + 1)h), s((j + 1)h)) − δ{θ((j + 1)h), 2}
× ξ(p̂s ((j + 1)h), s((j + 1)h))k∞ }
= kξ(p((j + 1)h), s((j + 1)h))
− ξ(p̂c ((j + 1)h), s((j + 1)h))k∞ × Prob{θ(τ h) = 2}
+ kξ(p((j + 1)h), s((j + 1)h))
− ξ(p̂s ((j + 1)h), s((j + 1)h))k∞ × Prob{θ(τ h) = 1} (7.27)
+ kξ(p((j + 1)h), s((j + 1)h))k∞ × Prob{θ(τ h) = 0}
= θ̄1 kξ(p((j + 1)h), s((j + 1)h))
− ξ(p̂s ((j + 1)h), s((j + 1)h))k∞
+ θ̄0 kξ(p((j + 1)h), s((j + 1)h))k∞
 
4
= θ¯1 + θ̄0 s((j + 1)h).
N
Combining (7.26) and (7.27), it is easy to obtain that b(τ h) ≤ µh b((τ − 1)h) +
νs(τ h) for τ = j + 1, and the proof of this lemma is thus complete.

Lemma 7.3: The sequences s(τ h) and b(τ h) (τ = 1, 2, . . . ) in Lemma 7.1 and
Lemma 7.2 satisfy

lim s(τ h) = 0 and lim b(τ h) = 0

τ →∞ τ →∞

if ρ(h ) < 1, where

 µ̄ 
h
ζh ¯h
µ
h , N , µ̃ , µh + νζh , ν̃ , ν
ν̃ µ̃ N

and other parameters are defined in Lemmas 7.1–7.2.

Proof: On the basis of Lemmas 7.1–7.2, we obtain the recursive expressions

¯h
µ
s(τ h) = s((τ − 1)h) + ζh b̄((τ − 1)h)
N
b(τ h) = µh b((τ − 1)h) + νs(τ h).
156 Data-Rate-Constrained State Estimation and Control

Let us introduce two auxiliary sequences S(τ h) and B(τ h) satisfying


¯h
 S(τ h) = µ S((τ − 1)h) + ζh B((τ − 1)h)
N

S(0) = s(0)

and (
B(τ h) = µh B((τ − 1)h) + ν S(τ h)
B(0) = b(0).
It is obvious that

s(τ h) ≤ S(τ h) and b(τ h) ≤ B(τ h), (τ = 1, 2, . . . ).

Substituting the dynamics S(τ h) into B(τ h), we further obtain that

B(τ h) = µ̃B((τ − 1)h) + ν̃ S((τ − 1)h).

Denoting η(τ h) , [S(τ h) B(τ h)]T , the dynamics of the augmented system is
acquired as


η (τ + 1)h = h η(τ h)

and it follows immediately from ρ(h ) < 1 that

lim η(τ h) = 0,
τ →∞

which indicates that

lim s(τ h) = 0 and lim b(τ h) = 0.

τ →∞ τ →∞

The proof of this lemma is complete.

Lemma 7.4: The decoding error e(k) (k = 1, 2, . . . ) satisfies

lim E{ke(k)k} = 0
k→∞

if ρ(h ) < 1, where h is defined in Lemma 7.3.

Proof: Rewriting e(k) as e(k) = xE1 (k) − xE2 (k), we have

E{ke(k)k∞ } ≤ E{kxE1 (k)k∞ } + E{kxE2 (k)k∞ }.

Stabilization of Linear Discrete-Time Systems 157

In order to prove that the decoding error e(k) achieves the stochastic stability, in
what follows, we will show that

lim E{ke(k)k} = 0
k→∞

for k = τ h and

τ h ≤ k < (τ + 1)h (τ = 1, 2, . . . ).

It follows from the results obtained in Lemmas 7.1–7.3 that

E{ke(τ h)k∞ } ≤ E{kxE1 (τ h)k∞ } + E{kxE2 (τ h)k∞ }

1
≤ s(τ h) + b(τ h) (τ = 1, 2, . . . ),
N

which gives rise to

lim {ke(τ h)k∞ } = 0.

τ →∞

For τ h ≤ j < (τ + 1)h, it is easily seen from the proofs of Lemmas 7.1–7.2 that

E{kxE1 (j)k∞ } = E{kx(j) − x̄(j)k∞ }

≤ µ̄j kxE1 (τ h)k∞ + ζj−τ h kxE2 (τ h)k∞

and
j−τ h
E{kxE2 (j)k∞ } ≤ kAk∞ b(τ h)

where
j−τ h
X
j
kA k∞ , µ̄j , ζj−τ h , k Aj−τ h−i BKAi−1 k∞ .
i=1

Letting

j−τ h
ℓ , max{µ
¯ j , ζj−τ h kAk∞ } (τ h ≤ j < (τ + 1)h),

we have

E{kxE1 (j)k∞ } + E{kxE2 (j)k∞ } ≤ ℓ(s(τ h) + b(τ h)),

158 Data-Rate-Constrained State Estimation and Control

which shows that

lim E{ke(j)k∞ } = 0
j→∞

and the proof of this lemma is ended.

Having obtained the stochastic convergence condition for the decoding
error, it is time for us to deal with the stochastic stabilization problem for
the system (7.1). The following theorem is provided to establish a sufficient
condition for the stochastic stability of the closed-loop system (7.23).

Theorem 7.1: The closed-loop system (7.23) is stochastically stable if the conditions
ρ(A) < 1 and ρ(h ) < 1 are simultaneously satisfied, where h is defined in
Lemma 7.3.

Proof: It follows from the dynamics of the closed-loop system (7.23) that

k
X
x(k) = Ak x(0) − Ak−i BKe(i − 1),
i=1

which leads to
k
X
E{kx(k)k} ≤ kAk kkx(0)k + kAkk−i kBKkE{ke(i − 1)k}.
i=1

Since the modulus of the eigenvalues of h lies in the unit circle on the complex plane,
it is known from Lemma 7.4 that E{ke(i − 1)k} is bounded, whose upper bound is
¯ Accordingly, we further derive that
denoted by e.

k
X
E{kx(k)k} ≤ kAk kkx(0)k + kAkk−i kBKkē
i=1
kBKke¯(1 − kAkk )
= kAk kkx(0)k + .
1 − kAk

Moreover, by applying the property that

lim Ak = 0 ⇐⇒ ρ(A) < 1,

k→∞

we confirm that

kBKkē
lim E{kx(k)k} ≤ .
k→∞ 1 − kAk
Stabilization of Linear Discrete-Time Systems 159

Pk k−i kBKkE{ke(i − 1)k} h¯ k

In fact, the term i=1 kAk can be rewritten as ℓk , where

k
X
h̄k , kAk−i kBKkE{ke(i − 1)k}, ℓk , kAk−k .
i=1

Noting that ℓk is a strictly monotonously increasing positive sequence with

lim ℓk = ∞,
k→∞

it follows immediately from the Stolz-Cesàro Theorem that

h¯ k h¯ k+1 − h̄k
lim = lim = 0,
k→∞ ℓk k→∞ ℓk+1 − ℓk

which yields

lim E{kx(k)k} = 0.
k→∞

The proof is complete.

Remark 7.4: As a matter of fact, the L’Hôpital’s rule is recognized as a powerful

means in obtaining the convergence of the closed-loop system in the continuous-time
setting with additive convergence disturbances. Nevertheless, when it comes to the
discrete-time system, where the state trajectory is a sequence and the derivative of
system state is unavailable, the L’Hôpital’s rule is no longer valid. In view of this,
the Stolz-Cesàro Theorem (recognized as a discrete-time version of L’Hôpital’s rule)
is employed to derive the stochastic stability of closed-loop system (7.23) in the proof
of Theorem 7.1.

The following corollary is put forward to analyze the influences from the
encoder parameters and the probability of packet dropouts on the feasibility
of the addressed MDC-based control problem.

Corollary 7.1: Suppose that the spectral radius of matrix A satisfies ρ(A) < 1. The
closed-loop system (7.23) is stochastically stable if the probability of the simultaneous
packet dropout of both communication channels satisfies the following relationship:

µ̄h
(1 − N )(1 − µh ) 4
θ¯0 ≤ − θ¯1 . (7.28)
ζh N

Proof: Denote λi (i = 1, 2) as the two eigenvalues of matrix h , which are,

respectively, calculated as
160 Data-Rate-Constrained State Estimation and Control

r !
1 µ̄h µ̄h 2
λ1 = + νζh + µh + ( + νζh − µh ) + 4νζh µh
2 N N
r !
1 µ̄h µ̄h
λ2 = + νζh + µh − ( + νζh − µh )2 + 4νζh µh .
2 N N

Note that λ1 (ν) (λ2 (ν)) is a monotonically increasing (decreasing) function of ν on

the interval [0, + ∞) with
µ̄h
N + µh + | µ̄Nh − µh |
λ1 (0) =
2
µ̄h ¯h
N + µh − | µN − µh |
λ2 (0) = .
2

In addition, it is easy to verify that

 
¯h
µ
λi (0) ∈ , µh ⊂ (0, 1) (i = 1, 2),
N

which implies that the equation λ2 (ν) = −1 is infeasible. In this sense, letting
λ1 (ν) = 1 yields
¯h
µ
(1 − N )(1 − µh )
ν=
ζh
and the proof of this corollary is therefore complete.

Corollary 7.2: Suppose that the spectral radius of matrix A satisfies ρ(A) < 1.
The closed-loop system (7.23) is stochastically stable if the encoder parameter N
(i.e. quantization level) satisfies the following relationship:

4ζh θ¯1 + (1 − µh )µ̄h

N> (7.29)
1 − µh − θ̄0 ζh

where all the parameters on the right-hand side of the above inequality have been
defined previously.

Proof: It follows from the calculation of the eigenvalues of matrix h in

Corollary 7.1 that

λi > 0 (i = 1, 2).
1
Consider the fact that λ2 is a monotonically increasing function with respect to N
on [0, +∞). Then, solving the equation λ2 ( N1 ) = 1 gives
Stabilization of Linear Discrete-Time Systems 161

1 1 νζh
= − .
N µ̄h µ̄h (1 − µh )

It is inferred from (7.29) that the requirement h < 1 is met and, therefore, the
stochastic stability of the closed-loop system (7.29) is guaranteed. The proof of this
corollary is complete.

Remark 7.5: Corollary 7.1 (Corollary 7.2) reveals the relationship between the
occurrence probability packet dropouts (the quantization level of the coding function)
and the system stability. To be more specific, in Corollary 7.1, an upper bound of the
probability of the simultaneous packet dropouts of both channels (i.e. θ̄0 ) is obtained,
which poses a basic constraint on the communication quality of the channels. Then,
¯h
the relationship between system parameters (e.g. µN , µh , N and ζh ) and the bound
upper is explicitly characterized. In Corollary 7.2, a lower bound of the quantization
level N is determined, from which we can conclude that such a lower bound is
dependent on the major system parameters (e.g. µ̄h , θ̄0 , θ̄1 , ζh and µh ). In particular,
it is readily seen from the right-hand sides of (7.28) and (7.29) that 1) the derived
upper bound of the probability of the simultaneous packet dropouts increases with
the growth of the quantization level N, and 2) the acquired lower bound of the
quantization level N decreases with the descent of the packet-dropout probabilities
θ̄0 and θ̄1 , and such observations are in agreement with the practice.

Remark 7.6: In comparison with the existing studies, the main results of this paper
exhibit the following distinctive contributions: 1) the MDC strategy developed in this
paper is new in the sense that the dynamical uniform quantization approach is, for
the first time, adopted to cope with the unreliability of the communication channel;
2) the constructed control scheme is new that ensures the stochastic stability of the
controlled system subject to distorted control signals; and 3) the obtained stability
criterion is new that reveals the influence of both the system parameters and the
encoder parameters on the stability.

7.3 Illustrative Examples

In this section, a numerical example is provided to evaluate the effectiveness
and applicability of the proposed MDC-based control scheme.
Let the discrete-time system (7.1) have following parameters:
   
0.4 0.63 1.5 0.5
A= , B= .
0.18 0.83 0.8 −0.45

It is easy to calculate that the eigenvalues of A are λ1 = 0.2155 and λ2 =

1.0145, respectively, which indicates that the open-loop system is unstable.
162 Data-Rate-Constrained State Estimation and Control

The controller gain matrix K is chosen as

 
−0.4 −0.21
K=
−0.36 −0.45

under which the spectral radius of A satisfies ρ(A) = 0.8661 < 1.

The initial parameters are set as x(0) = [0.85 −1.14]T and x̆(0) = x̂(0) =
[0 0]T . The coding period is given by h = 3 and the quantization level is
selected as N = 20. For the two communication channels, the successful-
arrival probabilities of the packets are set as γ1 = 0.6 and γ2 = 0.7,
respectively.
The matrix h is calculated as
 
0.0532 1.0680
h = ,
0.0113 0.8876

whose eigenvalues are, respectively, 0.0390 and 0.9018, which satisfies

ρ(A) < 1.
The simulation results are illustrated in Figures 7.3–7.7. Figure 7.3
describes the state trajectories of the open-loop system. Figure 7.4 shows

1
The amplitude of the state strategies without

0.5

0
control scheme

-0.5

-1

-1.5

-2

-2.5

-3
0 10 20 30 40 50 60 70 80 90 100
k/time step

FIGURE 7.3
State trajectories of the open-loop system.
Stabilization of Linear Discrete-Time Systems 163

1
The amplitude of the state strategies with the

0.5
proposed control scheme

-0.5

-1

-1.5
0 10 20 30 40 50 60 70 80 90 100
k/time step

FIGURE 7.4
State trajectories of the closed-loop system.

1
The amplitude of the state
trajectories and their
decoded values

0.5

-0.5

-1
0 10 20 30 40 50 60 70 80 90 100
k/time step

0.5
The amplitude of the state
trajectories and their

0
decoded values

-0.5

-1

-1.5
0 10 20 30 40 50 60 70 80 90 100
k/time step

FIGURE 7.5
State trajectories and their decoded values.
164 Data-Rate-Constrained State Estimation and Control

1
Decoding error

0.5

-0.5

-1
0 10 20 30 40 50 60 70 80 90 100
k/time step

0
Decoding error

-0.5

-1

0 10 20 30 40 50 60 70 80 90 100
k/time step

FIGURE 7.6
The trajectories of decoding errors.

the state trajectories of the corresponding closed-loop system with the

proposed MDC-based control scheme. Figure 7.5 exhibits the controlled state
trajectories and their decoded values. Figure 7.6 displays the decoding errors
for different state components. Figure 7.7 plots the values of the Bernoulli
sequences γi (τ h) (i = 1, 2) that characterize the packet-dropout phenomenon
of the communication channels. All the figures verify the applicability of the
developed control strategy, which firmly confirms the obtained results.

7.4 Summary
This paper has dealt with the stabilization problem for a class of
linear discrete-time systems under a resource-constrained channel. A new
dynamical multiple description coding approach has been developed
to cope with the unreliability of the communication channel. Two
sequences of Bernoulli-distributed random variables with known probability
Stabilization of Linear Discrete-Time Systems 165

-1
0 10 20 30 40 50 60 70 80 90 100

-1
0 10 20 30 40 50 60 70 80 90 100

FIGURE 7.7
The Bernoulli sequences γi (τ h) (i = 1, 2).

distributions have been employed to regulate the randomly occurring

packet-dropout phenomenon in the sensor to controller channels. Based
on the designed dynamical MDC strategy, the convergence criterion of the
decoding error has been first obtained, and then the stochastic stability of
the closed-loop system has been guaranteed with the help of the matrix
theory and stochastic analysis technique. Moreover, a sufficient condition
has been established to guarantee the stability of the closed-loop system,
in which the packet-dropout probabilities and the encoder parameters have
been involved. Finally, a simulation example has been provided to manifest
the applicability of the derived main results.
8
An Event-Triggered Encoding Approach
to Control of Linear Systems under
Bit Rate Conditions

The past few decades have witnessed a surge of research interest on the
dynamics analysis of networked control systems (NCSs) due primarily to
their pervasive applications in real-world industrial systems. Because of the
digital nature of the communication network, the original analog signal is
required to be converted into the digitized one with finite number of bits,
which inevitably results in the conversion error. Generally speaking, a high
bandwidth (i.e. bit rate) of the communication channel gives rise to a small
conversion error and little control performance degradation. However, in
most of practical applications (such as micro-electro mechanical systems,
sensor networks and industrial control networks), the network resource is
scarce and the available bit rate of the network is often limited. Thus, it is of
both theoretical significance and practical importance to look for a bit-rate
condition as tight as possible under which the prescribed control objective is
guaranteed.
In general, the network bandwidth occupancy of a data transmission task
is determined by two factors: the bits of the data itself and its communication
frequency. From this point of view, in order to comply with the bit-rate
requirement of a digital communication channel, the signal usually needs to
be “compressed” before being transmitted. One of the effective compression
approaches is the utilization of certain coding algorithms to encode the raw
data into specific codewords which occupy fewer bits. More specifically, the
data encoding/decoding operation in closed-loop systems can be generally
described as follows: 1) the sender encodes the state/measurement signals
into the corresponding codewords by certain appropriate coding schemes;
and 2) the generated codewords are transmitted to the decoder and suited
decoding algorithms are applied to reconstruct the raw data by means of
the received codewords. It should be pointed out that since the codewords
carry only partial information of the raw data, how to develop an ingenious
encoding-decoding algorithm to recover the raw data with an acceptable
accuracy level appears to be quite crucial.

DOI: 10.1201/9781003534853-8 166

An Event-Triggered Encoding Approach 167

In recent years, considerable research effort has been directed toward the
control problems for networked systems under certain bit-rate conditions, in
most of which, the implementation of the encoding algorithms is governed
by a time-scheduled manner due to its simplicity in design, that is, the
encoder generates and sends the codewords periodically at those fixed
time instants. However, under such a circumstance, the encoder could
potentially cause unnecessary coding actions as well as a high codeword
transmission frequency. This would further induce a severe occupancy of the
network resource, which is actually a challenge for the practical networked
systems subject to the limited network bandwidth. In order to mitigate the
occupancy of the network bandwidth, apart from the adoption of certain
coding strategies mentioned above, another effective way is to reduce the
triggering ratio of the encoding operation.
Benefiting from the swift development of the digital control technology
in the past decades, the data interaction in most engineering applications
has been upgraded to the digital manner, and thus the event-triggered
control strategy has been attracting substantial research attention. It is
worth mentioning that, in engineering practice, the event-based control
strategy is usually executed by a digital control device, which constitutes
the so-called sampled-data control system. Among others, a prevalent
approach is to utilize the time-based periodic sampling strategy to update
the control signal by selecting a feasible sampling period. Instead of the time-
triggered sampled-data control problems, the corresponding topics related
to the event-triggered sampling mechanism have recently gained particular
research attention. The most outstanding advantage of the event-triggered
sampled-data control strategy is that the controller is updated only if some
conditions are met, and therefore it might be more efficient in reducing the
controller update frequency and prolonging the controller operating lifespan.
However, to the best of the authors’ knowledge, in most of the published
results, only the event-based sampled-data control problem has been taken
into account, when it comes to the simultaneous consideration of the bit-rate
condition, the related results on this topic are quite few.
In response to the discussions aforesaid, in this paper, we aim to deal
with the control problem for continuous-time linear systems via an event-
triggered difference coding strategy, under which the corresponding bit rate
conditions are derived. In fact, such an investigation is non-trivial with
the following three major challenges: 1) how to design an event-triggered
communication mechanism to schedule the updates of both the coder and
the controller in a unified way? 2) how to reveal the relationship between the
required bit rate and the system parameters? 3) how to develop appropriate
methodologies to analyze the dynamics of the closed-loop system and
synthesize the controller design? This motives our investigation in this
chapter.
168 Data-Rate-Constrained State Estimation and Control

8.1 Problem Formulation

Consider a continuous-time linear system described by

d
x(t) = Ax(t) + Bu(t) (8.1)
dt

where x(t) ∈ Rn represents the state vector and u(t) ∈ Rp is the control input.
The initial state x(0) is bounded belonging to a given set X0 . A and B are
known matrices with appropriate dimensions. In this paper, it is assumed
that at least one of the eigenvalues of A has positive real part, that is, the
open-loop system is unstable.
From the resource-saving point of view, in a networked circumstance, it is
preferable that the data transmission is executed in an event-driven manner.
To this end, inspired by the idea in [73, 89, 90], an event-based coding scheme
is developed in this paper, which is shown in Figure 8.1. It should be pointed
out that the coder design is based on the idea of the difference coding scheme,
which means that the input signal of the coder is the difference between the
system state and the previous decoder (to be designed later) state. In order
to access the decoder state at the encoder side and avoid the signal postback
(i.e. signal flows from the decoder to the encoder), the following auxiliary
system, which can be viewed as a “copy” of the decoder, is constructed:

 d x (t) = Ax (t) + Bu (t) (8.2a)
a a a
dt
xa (t+ −

k ) = xa (tk ) + Lθ (sk )η(tk ) (8.2b)

FIGURE 8.1
The schematic of the event-based control with encoding-decoding scheme.
An Event-Triggered Encoding Approach 169

where xa (t) ∈ Rn and ua (t) ∈ Rp are, respectively, the state and the control
input of the auxiliary system. xa (t+ −
k ) and xa (tk ) are, respectively, defined as
xa (t− +
k ) , limt→tk xa (t) and xa (tk ) , limt→tk xa (t). sk is the encoding instant
− +

and tk is the time instant when both the decoder and the auxiliary system
receive the data packet θ (sk ) , vecn {θi (sk )} with θi (sk ) (i = 1, 2, . . . , n) being
the codewords encoded by the i-th encoder at time instant sk . τ , tk − sk ≥ 0
is the time delay caused by the encoding process. L ∈ Rn×n and η(tk ) ∈ R are
the parameters to be designed later.
Denoting by 0 = s0 < s1 < · · · sk < sk+1 < · · · the sequence of event
occurring instants (encoding instants) and setting x̃a (sk ) , x(sk ) − xa (sk ) as
the encoder input. The encoding procedure is triggered when the following
event generation condition is satisfied:

f (x̃a (sk ), σ (sk )) = 0 (8.3)

where f (x̃a (sk ), σ (sk )) , kx̃a (sk )k − σ (sk ) is called the triggering function and
σ ( · ) is an exponentially decaying function defined as σ (t) , σ0 e−λt with σ0
being a given positive scalar and λ being a parameter to be determined.
The triggering instants (or encoding instants) can be successively
calculated by the following formula:

sk+1 = inf{t | t > sk , kx̃a (t)k = σ (t)}. (8.4)

Next, denote by x̃ai (t) the ith component of x̃a (t). As such, at the triggering
time sk , for each component x̃ai (t) (i = 1, 2, . . . , n), the individual codeword
θi (sk ) is determined by the following rule:

θi (sk ) = sign(x̃ai (sk )) (8.5)

where “sign( · )” stands for the signum function.

As stated before, due to the encoding delay, the codeword θi (sk ) can only be
available at time tk . In view of this, for t ∈ [tk , tk+1 ), we construct the decoder
with the following dynamics:

 d x (t) = Ax (t) + Bu(t) (8.6a)
d d
dt
xd (t+ −

k ) = xd (tk ) + Lθ (sk )η(tk ) (8.6b)

where xd (t) ∈ Rn is the decoder state, u(t) is the control input to be designed
and other parameters are the same as the auxiliary system (8.2a)-(8.2b).
It is seen from (8.2b) and (8.6b) that, with the obtained data packet θ(sk ),
both the auxiliary system state xa (t) and the decoding system state xd (t) are
simultaneously updated at time tk . Moreover, by employing the zero-order
170 Data-Rate-Constrained State Estimation and Control

holder (ZOH) strategy, the control input u(t) in (8.1) and (8.6a) is updated at
tk and will maintain constant until the next event time tk+1 , that is,

u(t) = Kxd (tk ), ∀t ∈ [tk , tk+1 ) (8.7)

and, along the same line, the controller ua (t) in (8.2a) is designed as

ua (t) = Kxa (tk ), ∀t ∈ [tk , tk+1 ) (8.8)

where K is the controller parameter to be designed. Combining (8.2a)–(8.2b)

and (8.6a)–(8.6b), it is not difficult to verify that xa (t) ≡ xd (t) ∀t > 0 provided
that xa (0) = xd (0).

Remark 8.1: In this paper, an event-based update mechanism is employed for the
purpose of reducing the frequency of both the data transmission and the controller
update. To be more specific, when the event generation condition (8.3) is satisfied
at the time instant sk , the encoder starts to implement the encoding procedure. In
addition, owing to the existence of the time delay during the encoding process, the
codeword θi (sk ) is only available to both the auxiliary system and the decoder system
at tk . Then, based on the renewed system states xa (tk ) and xd (tk ), the control input
is updated at tk . Therefore, not only the encoder executions but also the controller
updates are dependent on the event function (8.3), and such an event-triggered
mechanism achieves a great reduction of the update frequency of both the coder and
the controller.

The objective of this paper is summarized as the following threefold.

1. Develop an event-triggered encoding-decoding mechanism (8.3)–

(8.6b), by which a sufficient condition for the bit-rate of the
communication network is established such that the decoding error
(i.e. x̃d (t) , x(t) − xd (t)) is convergent, that is limt→∞ kx̃d (t)k = 0.
2. Derive a necessary condition to obtain a lower bound of the required
bit rate, below which the decoding error diverges.
3. Design the event-based control strategy (8.7) for the system (8.1) to
guarantee that the closed-loop system is exponentially stable.

8.2 Circumventing the “Zeno” Phenomenon

It is worth pointing out that, for the event-triggered continuous-time system
(8.1), if the inter-event execution interval (namely, the length of two arbitrary
successive event-triggering instants) trends to be infinitesimal, the so-called
An Event-Triggered Encoding Approach 171

“Zeno” phenomenon happens. In this case, the triggering frequency of

the encoder would approach infinity and so does the update frequency
of the controller. Obviously, the “Zeno” behavior will adversely affect
the encoding/sampling process and even damage the relevant hardware
devices, which should be strictly avoided. To this end, in the following, a
useful lemma will be provided to eliminate the “Zeno” behavior. Also, an
upper bound of the required bit rate to stabilize the addressed system (8.1)
is derived.

Lemma 8.1: If the encoding time-delay τ (i.e. tk − sk ) satisfies 0 ≤ τ < κ, then

there exists a positive scalar λ such that the inter-event execution interval satisfies
ln δ1
q
3
sk+1 − sk ≥ κ, where κ = λ11 with δ1 = 1 − 4n and λ1 = λ + kAk

Proof: In order to make our statement clearer, the exclusion of the “Zeno” behavior
is conducted by the following two steps.
Step I. In this step, our intention is to ensure that there is no any other triggering
instant(s) over the interval (sk , tk ), which is equivalent to show that sk+1 > tk . To
begin with, it follows from (8.1) and (8.2a)–(8.2b) that the dynamics of the x̃a (t) is
governed by

 d x˜ (t) = Ax̃ (t) (8.9a)
a a
dt
x̃a (tk ) = x̃a (tk− ) − Lθ (sk )η(tk ).

(8.9b)

For t ∈ (sk , tk ), the evolution of the state trajectorypx̃a (t) can be determined by
x̃a (t) = eA(t−sk ) x̃a (sk ), which indicates that kx̃a (t)k = x̃Ta (sk )A(t − sk )x̃a (sk ) with
T
A(t − sk ) , eA (t−sk ) eA(t−sk ) . Due to the inverse property of the matrix exponential,
it is easy to verify that A(t − sk ) is a positive definite matrix and there exists
a positive scalar function ϕ(t − sk ) > 0 such that A(t − sk ) ≥ ϕ 2 (t − sk )I
for ∀t ∈ (sk , tk ). Based on such a fact, one can further confirm that kx̃a (t)k ≥
ϕ(t − sk )kx̃a (sk )k = ϕ(t − sk )σ (sk ) for ∀t ∈ (sk , tk ). Recalling that σ (t) can be
rewritten as σ (t) = e−λ(t−sk ) σ (sk ) for ∀t ∈ (sk , tk ), it is obvious that there must
exist a suitable λ meeting the requirement ϕ(t − sk ) > e−λ(t−sk ) , which leads to
kx̃a (t)k > σ (t). Consequently, the triggering event is excluded over the interval
(sk , tk ) by properly choosing the parameter λ.
Step II. In this step, we are interested in investigating the length of sk+1 − tk
and subsequently determining a lower bound of the inter-event execution interval
sk+1 − sk . To this end, our attention is focused on the dynamics evolution of kx̃a (t)k
over the time interval [tk , sk+1 ). For t ∈ [tk , tk+1 ), the upper right-hand Dini
derivative of kx̃a (t)k satisfies

D+ kx̃a (t)k ≤ kẋ˜a (t)k ≤ kAkkx̃a (t)k, (8.10)

172 Data-Rate-Constrained State Estimation and Control

which further indicates

kx̃a (t)k ≤ ekAk(t−tk ) kx̃a (tk )k. (8.11)

1 Aτ λτ
Letting L = 2n e e and η(t) = σ (t), it follows from (8.9b) that

1 Aτ λτ
x̃a (tk ) = x̃a (t−
k )− e e θ (sk )σ (tk )
2n
(8.12)
A(tk −sk ) 1 Aτ λτ
=e x̃a (sk ) − e e θ(sk )σ (tk ).
2n

Also, it is inferred
n from o (8.3) and (8.5) that x̃a (sk ) = 6(sk )θ(sk )σ (sk ), where
σi (sk )
6(sk ) = diagn σ (s ) with σi (sk ) = |x̃ai (sk )|. Substituting x̃a (sk ) = 6(sk )θ(sk )
k
σ (sk ) into (8.12) yields

x̃a (tk ) = eA(tk −sk ) 6(sk )θ (sk )σ (sk )

1 Aτ λτ
− e e θ (sk )σ (tk ) (8.13)
2n
 
Aτ σ (tk ) λτ
=e 6(sk )σ (sk ) − e In θ(sk ).
2n

In addition, it is observed from (8.9a) that, over the interval [tk , sk+1 ), the state
trajectory of x̃a (t) can be determined by

x̃a (t) = eA(t−tk ) x̃a (tk )

 
A(t−tk ) Aτ σ (tk ) λτ
=e e 6(sk )σ (sk ) − e In θ(sk ) (8.14)
2n
 
A(t−tk ) Aτ 1
=e e 6(sk ) − In θ(sk )eλτ eλ(t−tk ) σ (t)
2n

and, by means of the matrix norm inequalities, one can further derive that
 
A(t−tk ) Aτ 1
kx̃a (t)k = e e 6(sk ) − In θ(sk )eλτ eλ(t−tk ) σ (t)
2n
 
1
≤ eA(t−tk ) eAτ 6(sk ) − In θ(sk ) eλτ eλ(t−tk ) σ (t)
2n

1
≤ ekAk(t−tk ) ekAkτ (6(sk ) − In )θ(sk ) eλτ eλ(t−tk ) σ (t)
2n
An Event-Triggered Encoding Approach 173

 
1
= 6(sk ) − In θ (sk ) e(kAk+λ)(t+τ −tk ) σ (t)
2n
 
1
= 6(sk ) − In θ (sk ) e(kAk+λ)(t−sk ) σ (t). (8.15)
2n

Recalling the definition of 6(sk ), it is easy to verify that

1
(6(sk ) − In )θ (sk ) = 6̄(sk ) (8.16)
2n

where
v !
2
σi (sk )σ (sk ) σ 2 (sk )θi (sk )
u n
uX 1 2
¯ k) = t
6(s σi ( s k ) − + .
σ 2 (sk ) n 4n2
i=1

Noticing that σi (sk ) ≥ 0 (i = 1, 2, . . . , n) and

Pn 2 = σ 2 (sk ), we have
i=1 σi (sk )

σ1 (sk ) + · · · + σn (sk ) √
1≤ ≤ n. (8.17)
σ (sk )

Moreover, it is readily seen that 1 ≤

Pn 2 ≤ n, which indicates
i=1 θi (sk )
 
1
δ0 ≤ 6(sk ) − In θ (sk ) ≤ δ1 (8.18)
2n
q √ q
where δ0 = 1 − nn + 4n1 2 and δ1 = 1 − 4n 3
.
Therefore, it follows from (8.15) that, for t ∈ [tk , tk+1 ), one has

kx̃a (t)k ≤ δ1 eλ1 (t−sk ) σ (t) (8.19)

where λ1 = λ + kAk.
Letting δ1 eλ1 (t−sk ) = 1, we have

ln δ11
t − sk , κ = , (8.20)
λ1

namely,

t − tk = κ − τ = 1Ta , (8.21)

which further indicates that, for t ∈ [tk , tk + 1Ta ), kx̃a (t)k < σ (t). Consequently, it
is seen from the encoding event generation condition (8.3) that sk+1 − sk ≥ κ, and
the “Zeno” behavior is thus eliminated.
174 Data-Rate-Constrained State Estimation and Control

8.3 Convergence Analysis of x̃a (t)

Lemma 8.2: For the proposed event-based encoding-decoding scheme (8.3)–(8.6b),
the dynamics x̃a (t) is exponentially convergent, that is, kx̃a (t)k ≤ σ¯ 0 e−λt for ∀t ≥ 0,
where σ̄0 = σ0 eλ1 τ .

Proof: In the proof of Lemma 8.1, one has kx̃a (t)k < σ (t) for ∀t ∈ [tk , sk+1 ).
Subsequently, we intend to study the dynamics of kx̃a (t)k over the time interval
[sk+1 , tk+1 ). With the combination of (8.3) and (8.9a), it is readily seen that, for
∀t ∈ [sk+1 , tk+1 ), one has kx̃a (t)k ≤ ekAk(t−sk+1 ) kx̃a (sk+1 )k = eλ1 (t−sk+1 ) σ (t) ≤
eλ1 τ σ (t). Therefore, it is obvious that kx̃a (t)k ≤ σ̄0 e−λt for ∀t ≥ 0, which shows that
x̃a (t) exponentially converges to zero.

It is obvious from xa (t) ≡ xd (t) ∀t ≥ 0 that x̃a (t) ≡ x̃d (t) ∀t ≥ 0. Thus,
it follows immediately from Lemma 8.2 that the decoding error x̃d (t) is
exponentially convergent under the encoding-decoding scheme (8.3)–(8.6b).

Remark 8.2: It can be observed from Lemma 8.1 and the convergence analysis of
decoding error that the value of κ is associated with two parameters, namely, the
convergence rate of the triggering threshold λ (also the convergence rate of the
decoding error) and the system dimension n. On one hand, a larger λ leads to a
smaller κ, which means that as the convergence rate of the decoding error becomes
larger, the execution of the encoding operation would be more frequent. On the other
hand, the increase of the system dimension n would cause a decline of κ, which
signifies that a higher dimensional system may cost more resources in order to ensure
the feasibility of the coding-decoding algorithm.

8.4 A Sufficient Condition of the Bit Rate to Guarantee the

Convergence of the Decoding Error
The following theorem provides a sufficient condition of the required bit rate
of the communication channel to guarantee the convergence of the decoding
error.

Theorem 8.1: Consider the system (8.1) with the event-based encoding-decoding
scheme (8.3)–(8.6b). The convergence of the decoding error x̃d (t) is guaranteed
(i.e. limt→∞ kx̃d (t)k = 0) if the bit rate of the communication channel R satisfies
R ≥ R1 , n⌈ κ1 ⌉.
An Event-Triggered Encoding Approach 175

Proof: It follows from Lemmas 8.1-8.2 that the convergence of the decoding error is
guaranteed. In addition, it is easy to see from sk+1 − sk ≥ κ in Lemma 8.1 that there
are at most ⌈ κ1 ⌉ times’ packet transmissions per time unit. Since each transmission
of the packet θ(sk ) occupies n bits, it suffices to use the bit rate of n⌈ κ1 ⌉ to guarantee
that limt→∞ kx̃d (t)k = 0, which completes the proof of this theorem.

8.5 A Necessary Condition of the Bit Rate

Having obtained a sufficient condition of the bit rate under which the
convergence of the decoding error is guaranteed, we are also interested in
seeking a necessary condition of the bit rate below which the decoding error
is divergent. It is noted that the sufficient condition in Theorem 8.1 shows
that if the bit rate of the communication channel is greater than R1 , the
convergence of the decoding error x̃d (t) is guaranteed. While the necessary
condition of the bit rate aims to determine a bit rate R0 below which the
decoding error x̃d (t) diverges. In order to achieve this goal, the following
analysis is first presented.
As mentioned before, since at least one of the eigenvalues of A has positive
real part, it is easy to observe from (8.9a) that, for t ∈ [tk , tk+1 ), the evolution
of x̃a (t) is divergent. Also, noting that the triggering threshold (i.e. σ (t) =
σ0 e−λt ) is an exponential decay function, it is not difficult to verify that there
must exist an triggering instant sk+1 ∈ [tk , tk+1 ). Consequently, in the next
stage, we plan to determine an upper bound ι of the inter-event execution
interval, that is sk+1 − sk ≤ ι.

Lemma 8.3: For the proposed event-based encoding-decoding scheme (8.3)–(8.6b),

if the positive parameter λ satisfies λ > kAk, then an upper bound ι of the inter-
ln δ1
0
event execution interval can be obtained, that is, sk+1 − sk ≤ ι, where ι = λ0 with
λ0 = λ − kAk.

Proof: It follows from (8.14) that, for ∀t ∈ [tk , tk+1 ), the dynamics of x̃a (t) can be
rewritten as
 
1
x̃a (t) = eA(t−tk ) eAτ 6(sk ) − In θ(sk )eλτ eλ(t−tk ) σ (t)
2n
  (8.22)
1
= eA(t−sk ) 6(sk ) − In θ (sk )eλ(t−sk ) σ (t).
2n

In addition, noting that eA(t−sk ) is invertible with the inverse matrix e−A(t−sk ) , one
obtains
176 Data-Rate-Constrained State Estimation and Control

 
1
e−A(t−sk ) x̃a (t) = 6 (sk ) − In θ (sk )eλ(t−sk ) σ (t), (8.23)
2n
which further derives

ek−Ak(t−sk ) kx̃a (t)k2 ≥ ke−A(t−sk ) x̃a (t)k

(8.24)
 
1
= 6(sk ) − In θ(sk ) eλ(t−sk ) σ (t).
2n
 
1
Moreover, bearing in mind that 6(sk ) − 2n In θ(sk ) ≥ δ0 , one derives that, for
t ∈ [tk , tk+1 ), the following relation
 
1
kx̃a (t)k ≥ 6(sk ) − In θ (sk ) e(λ−kAk)(t−sk ) σ (t)
n (8.25)
≥ δ0 eλ0 (t−sk ) σ (t)

ln δ1
holds. Letting δ0 eλ0 (t−sk ) = 1 yields t−sk = λ00 , ι. Considering both the encoding
event generation condition (8.3) and the fact that the dynamics x̃a (t) satisfying
ln δ1
kx̃a (t)k ≥ σ (t) for t ∈ [tk + λ0
0
− τ , tk+1 ), it is easy to see that sk+1 − tk ≤ 1Tb ,
ln δ1
0
where 1Tb = λ0 − τ.

Theorem 8.2: Consider the system (8.1) with the event-based encoding-decoding
scheme (8.3)–(8.6b). The decode error x̃d (t) diverges if the bit rate of the
communication channel is less than R0 , n⌊ 1ι ⌋.

Proof: It is inferred from (8.25) that there are at least ⌊ 1ι ⌋ times’ data transmissions
each time unit. Along the similar line in the proof of Theorem 8.1 that each
transmission of the packet θ (sk ) occupies n bits, we can conclude that it is
impossible to transmit all the generated codewords to the decoder if the bit rate of
the communication channel is less than n⌊ 1ι ⌋, which consequently results in the
divergence of the decoding error x̃d (t). The proof of this theorem is ended.

8.6 Analysis and Synthesis of the Addressed Event-Based

Control Issue
In what follows, our effort is devoted to the analysis and synthesis of the
addressed event-based control problem. Moreover, the input-delay approach
An Event-Triggered Encoding Approach 177

is applied to derive the main results. To do this, first, defining dk (t) , t − tk

for ∀t ∈ [tk , tk+1 ), it is readily seen that 0 ≤ dk (t) < ι and ḋk (t) = 1. Then,
substituting the control input (8.7) into (8.1), the resulted closed-loop system
is obtained as follows:

ẋ(t) = Ax(t) + BKx(t − dk (t))

(8.26)
− BKx̃d (t − dk (t)), ∀ tk ≤ t < tk+1

where x̃d (t − dk (t)) , x(t − dk (t)) − xd (t − dk (t)).

By bringing in the function d(t) which is defined as d(t) , dk (t) for tk ≤ t <
tk+1 , k = 0, 1, . . . , ∞, the closed-loop system can be further rewritten as

ẋ(t) = Ax(t) + BKx(t − d(t)) − BKx̃d (t − d(t)). (8.27)

Now, the event-based control problem is converted into the stability

analysis issue of the time-delayed system (8.27). In the following, a sufficient
condition is provided to guarantee the exponential stability of the addressed
closed-loop system.

Theorem 8.3: Let the positive scalars µ, σ0 and α be given. Under the conditions
in Lemmas 8.1 and 8.3, the closed-loop system (8.27) is exponentially stable by the
event-based control strategy (8.8) subject to the bit-rate condition R ≥ R1 , if there
exist three positive definite matrices P̄ > 0, Q̄ > 0, R̄ > 0 and a real-valued matrix
X such that the condition

0
 
4̄11 4̄12 BX 4̄15
2 1¯ ¯ 25
 ∗
 − ι R̄ ι R 0 4 

0 0 <0 (8.28)
 
5= ∗ ∗ 4̄33
 
 ∗ ∗ ∗ −µ2 R̄ 4̄45 
∗ ∗ ∗ ∗ 4̄55

holds, where

1¯ 1¯
4̄11 = αP̄ + Q̄ − R + P̄AT + AP̄, 4̄12 = BX + R ,
ι ι
p p
4̄15 = β P̄AT , 4¯ 25 = β XT BT ,
1 1
1 p
¯ 33 = − R̄ − β0 Q̄, 4̄45 = − β XT BT ,
4 1
ι
eαι − 1
¯ 55 = R̄ − 2P̄, β0 = e−αι , β1 =
4 .
α

In addition, the expected controller parameter is designed as K = XP̄−1 .

178 Data-Rate-Constrained State Estimation and Control

Proof: First, define positive definite matrices P , P̄−1 , Q , P¯ −1 Q̄P̄−1 and R ,

P̄−1 R̄P̄−1 . Then, choose the Lyapunov-Krasovskii functional candidate as

V(t) = V1 (t) + V2 (t) + V3 (t) (8.29)

where

V1 (t) = xT (t)Px(t),
Z t
V2 (t) = eα(s−t) xT (s)Qx(s)ds,
t−ι
Z 0 Z t
V3 (t) = e−αθ eα(s−t) x˙ T (s)Rx˙(s)dsdθ.
−ι t+θ

Along the trajectory of (8.27), taking the derivative of V(t) with respect to the time
t results in
V̇(t) = V̇1 (t) + V̇2 (t) + V̇3 (t) (8.30)
where

V̇1 (t) = 2xT (t)Pẋ(t),

V̇2 (t) = −αV2 (t) + xT (t)Qx(t) − β0 xT (t − ι)Qx(t − ι),

(8.31)
Z t
T
V̇3 (t) = −αV3 (t) + β1 ẋ (t)Rẋ(t) − x˙ T (s)R˙(s)ds.
x
t−ι

By resorting to the well-known Jenson integral inequality, we have

Z t
− x˙ T (s)Rẋ(s)ds
t−ι
Z t Z t−d(t)
=− ẋT (s)Rẋ(s)ds − x˙ T (s)Rx(s)ds
˙
t−d(t) t−ι

1
≤− (x(t) − x(t − d(t))T R(x(t) − x(t − d(t))
d(t)
(8.32)
1
− (x(t − d(t) − x(t − ι))T R
ι − d(t)
× (x(t − d(t) − x(t − ι))
1
≤ − (x(t) − x(t − d(t)))T R(x(t) − x(t − d(t))
ι
1
− (x(t − d(t)) − x(t − ι))T R(x(t − d(t)) − x(t − ι)).
ι
An Event-Triggered Encoding Approach 179

Based on (8.29)–(8.32), it can be found that

V̇(t) + αV(t) ≤ αxT (t)Px(t) + 2xT (t)Pẋ(t) + xT (t)Qx(t)

− β0 xT (t − ι)Qx(t − ι) + β1 ẋT (t)Rẋ(t)
1
− (x(t) − x(t − d(t)))T R(x(t) − x(t − d(t)) (8.33)
d
1
− (x(t − d(t)) − x(t − ι))T R(x(t − d(t)) − x(t − ι))
d
= ξ (t)4ξ(t) + µ2 kx̃d (t − d(t))k2R
T

where
p
4 = 41 + 4T2 R42 , 42 = β 1 [A BK 0 −BK],
411 PBK + 1ι R
 
0 −PBK
 ∗ − 2ι R 1
ιR 0  ,
41 =  1
 ∗ ∗ − ι R − β0 Q 0 
∗ ∗ ∗ −µ2 R
ξ(t) = [xT (t) xT (t − d(t)) xT (t − ι) x̃Td (t − d(t))]T ,
1
411 = αP + Q − R + AT P + PA,
ι
1
kx̃d (t − d(t))kR = (x̃dT (t − d(t))Rx̃d (t − d(t))) 2 .

In order to draw the main results, we are in a position to show that 4 < 0 under the
condition (8.28). First, noting that P−1 RP−T −2P−1 ≥ −R−1 , it follows from (8.28)
that 5̄ < 0, where 5̄ has the same structure with 5 and only R̄ − 2P̄ is replaced by
−R−1 . Then, pre-multiplying and post-multiplying 5̄ by diag{P, P, P, P, I} and its
transpose, one has  
41 4T2
< 0. (8.34)
∗ −R−1
Using the Schur Complement Lemma to the inequality (8.34), it immediately finds
that 4 < 0. Moreover, considering that both the observation error and decoding error
2 ≤
satisfy kx̃a (t)k = kx̃d (t)k ≤ σ0 eλ1 τ e−λt , it is obvious that µ2 kx̃d (t − d(t))kR
2 2 2
h̄(t), where h̄(t) = µ λmax (R)σ0 e 1 e λ τ − 2λ(t−ι) . Consequently, it is inferred from
(8.33) that

V̇(t) ≤ −αV(t) + h̄(t). (8.35)

Furthermore, by applying the Comparison Lemma, (8.35) implies that

Z t
V (t) ≤ e−αt V(0) + e−α (t−θ) h¯ (θ)dθ. (8.36)
0
180 Data-Rate-Constrained State Estimation and Control

Then, according to the fact that V(t) ≥ λmin (P) and h̄(t) ≤ h̄(t0 ), one further has
s Rt
e−αt V(0) 0 e−α(t−θ ) h¯ (θ)dθ
kx(t)k ≤ +
λmin (P) λmin (P)
s
e−αt V(0) 1 − e−αt
≤ + h(t
¯ 0) (8.37)
λmin (P) αλmin (P)
s
e−αt V(0) 1
≤ + h(t
¯ 0 ).
λmin (P) αλmin (P)

In fact, since h̄(t) is a monotonically non-increasing function with limt→∞

h̄(t) = 0, it follows from the L’Hôpital’s rule that
Z t
lim e−α(t−θ) h¯ (θ )dθ = 0. (8.38)
t−→∞ 0

Therefore, we can draw the conclusion that the closed-loop

q control system (8.27)
e−αt V(0)
is exponentially stable, namely, limt→∞ kx(t)k ≤ λmin (P) = 0, which completes
this proof.

Remark 8.3: In this paper, we endeavor to offer a satisfactory solution to the

resource-saving-oriented control problem for a class of continuous-time systems.
By utilizing the event-triggered symbolic-based difference coding scheme, some bit
rate conditions are derived. In addition, in terms of the input-delay approach, a
unified framework has been established to deal with the proposed control issue.
Compared with the existing literature, the research conducted in this paper possesses
the following distinct features: 1) a sufficient condition of the bit rate is obtained to
guarantee the convergence of the decoding error and the explicit relationship among
the required bit rate of the control issue and the system dimension as well as the
convergence rate of the decoding error is revealed; 2) a necessary condition of the bit
rate is given below which the decoding error is divergent; and 3) the input-delay
approach is employed to facilitate the controller analysis and design issue of the
addressed systems.

8.7 Illustrative Examples

In this section, a simulation example is presented to demonstrate the
usefulness of both the proposed coding algorithm and the controller design
scheme for the addressed systems.
An Event-Triggered Encoding Approach 181

For system (8.1), the parameter matrices are given as follows:

   
0.01 0.02 1.12
A= , B= .
−0.01 0.01 −0.6

It is easily seen that the eigenvalues of matrix A are χ1 = 0.01 + 0.0141i

and χ2 = 0.01 − 0.0141i, respectively, which indicates that the open-loop
system (8.1) without control signal u(t) is unstable. For the event generation
condition (8.3), the threshold parameters are chosen as σ0 = 50 and λ = 0.23.
The encoding delay is assumed as τ = 0.004, the convergence rate of the
system state is set as α = 0.12 and the parameter µ is given by µ = 2.
Moreover, it can be confirmed from Lemma 8.1 that the inter-event execution
interval satisfies sk+1 − sk > κ = 0.9288 and an upper bound of the required
bit rate R guaranteeing the feasibility of the proposed coding strategy is
obtained as R ≤ R1 = κn = 2.1534. Let the initial conditions be x(0) =
[0.7 −0.5]T and xa (0) = xd (0) = [−23.7 13.5]T , respectively. By solving
the matrix inequality (8.28) in Theorem 8.3, the controller gain is calculated
as K = [0.4360 1.2921].
After setting the initial condition and deriving the controller parameter,
the simulation results are achieved and shown in Figures 8.2–8.6. Figure 8.2

20
Amplitude

0
x a1 (t)
−20 x d1 (t)

−40
0 5 10 15 20 25 30
Time(s)

20
x a2 (t)
10 x d2 (t)
Amplitude

−10

−20
0 5 10 15 20 25 30
Time(s)

FIGURE 8.2
Dynamical evolutions of auxiliary system and decoder.
182 Data-Rate-Constrained State Estimation and Control

25

x̃ a1 (t)
20
x̃ a2 (t)

15

10
Amplitude

−5

−10

−15
0 5 10 15 20 25 30
Time(s)

FIGURE 8.3
The decoding errors between the actual system states and their decoded values.

0.8
x 1 (t)
0.7
Amplitude

0.6

0.5

0.4
0 5 10 15 20 25 30
Time(s)

−0.5

−0.6 x 2 (t)
Amplitude

−0.7

−0.8

−0.9
0 5 10 15 20 25 30
Time(s)

FIGURE 8.4
State response without control signal u(t).
An Event-Triggered Encoding Approach 183

40

x 1 (t)
Amplitude

20

−20
0 5 10 15 20 25 30
Time(s)

10
Amplitude

0
x 2 (t)
−10

−20
0 5 10 15 20 25 30
Time(s)

FIGURE 8.5
State response with control signal u(t).

50

45

40

35

30
Amplitude

25

20

15

10

0
0 5 10 15 20 25 30
Time(s)

FIGURE 8.6
The threshold of the event generation function and the norm of x̃a (t).
184 Data-Rate-Constrained State Estimation and Control

depicts the dynamics of the auxiliary system state xa (t) and the decoder
state xd (t). Figure 8.3 describes the decoding error between the system state
x(t) and its decoded value xd (t). Figure 8.4 plots the state trajectory of the
open-loop system, while Figure 8.5 gives the state response of the closed-
loop system with control input u(t). Figure 8.6 illustrates the evolution of
the event generation function σ (t) and the triggering instants. It is inferred
from Figures 8.2–8.3 that the decoding error converges to zero as time goes to
infinity and hence the developed coding strategy works well. Figures 8.4–8.5
show that the proposed control scheme is effective for the unstable systems.
From Figure 8.6, it is not difficult to see that the intersections of the two curves
are the triggering instants. Consequently, all the simulation results confirm
the superiority of the proposed control policy.

8.8 Summary
In this paper, the event-triggered control issue has been considered for a class
of continuous-time linear systems via the digital communication channel.
A novel digital data transmission mechanism has been introduced, where the
communication media has been subjected to limited transmission capacity.
A difference coding scheme has been developed, which has transferred the
original information to one-bit symbolic data. In addition, a unified event-
based encoding and sampling mechanism has been proposed in order to
alleviate the communication burden as well as the equipment losses. The
“Zeno” behavior has been excluded by determining a lower bound of the
inter-event execution times. A sufficient condition of the required bit rate has
been obtained to guarantee the convergence of the decoding error. Moreover,
a necessary condition has also been derived with the bit rate below which
the decoding error diverges. By obtaining the decoded data, the input-delay
method has been adopted to deal with both the analysis and the design of the
addressed control problem. Finally, the effectiveness of our coding/control
strategies has been well verified by a simulation example.
9
Event-Based State Estimation under
Constrained Bit Rate: An
Encoding-Decoding Approach

In typical NSs, signal transmissions in the sensor-to-estimator and controller-

to-actuator channels are realized through a digital communication network.
Compared with traditional analogue communication, digital communication
has been preferred by industry because of its distinguishing features of
high reliability, strong anti-interference ability and low power consumption
[6,187]. In particular, data encoding-decoding serves as a crucially important
procedure in digital communication whose aim is to convert numerical
data to binary codeword, thereby catering to the needs for reliable,
secure and real-time communication. In recent years, encoding-decoding-
based analysis/synthesis problems have drawn much attention from both
academia and industry, and some representative research results can be
found in and the references therein.
In the encoding-decoding procedure, the numerical data is mapped to
the binary codeword with a finite number of bits through the coding
operation, and the codeword is then restored to the original numerical
data as accurately as possible through the decoding operation. It is worth
noting that, in the design process of the encoding-decoding mechanism, the
decoding accuracy is largely dependent on the bit rate of the network (i.e. the
number of transmitted bits per second), which is recognized as an important
indicator of the network bandwidth [36]. Intuitively, a larger bit rate leads
to less distortion of decoded data, thereby facilitating a higher decoding
accuracy. In most engineering practice, however, only limited bits are
allowed to be delivered for each transmission due mainly to limited network
bandwidth, and this brings great challenges to the analysis/synthesis of
communication-based systems/networks. As a consequence, a great number
of researchers have devoted themselves to developing efficient encoding-
decoding schemes with high accuracy for NSs under constrained bit rates.
In the context of NSs with bit rate constraints, most existing results have
been based on the periodic encoding/transmission principle under which
the data are encoded/exchanged at fixed time instants, see [161,163]. Such

DOI: 10.1201/9781003534853-9 185

186 Data-Rate-Constrained State Estimation and Control

a periodic mechanism, though easy to be implemented, might result in

unnecessarily frequent data transmissions (especially in the steady-state
situation) which, in turn, give rise to the overconsumption of the scarce bit
rates. For resource-saving purposes, a natural idea is to encode/transmit the
data only when certain prescribed “event” occurs, and such an even-based
mechanism has recently stirred much attention [35,192]. Compared with
its periodic counterpart, the event-based encoding-decoding scheme would
require much lower data bit rate (for achieving a favorable performance and
is therefore particularly suitable in analyzing NSs with constrained bit rate,
which constitutes the main focus of this paper).
Recently, there has been a trend in utilizing sign-based encoding approach
(SEA) to stabilizing control systems in networked environments, where the
main idea of SEA is to encode the original signal to certain sign information
occupying one bit. For example, pioneering yet excellent results have
been pursued in [89, 90] on the SEA-based state-feedback control problem.
In [89], a bit-rate condition has been obtained to stabilize a continuous-
time scalar linear systems and then the results have been extended to
multiple dimensional systems subject to processing delay and network
delay in [90]. In this chapter, we take the initiative to deal with another
fundamental research problem, namely, state estimation problem, in which
the measurement signal needs to be encoded before being transmitted to
the estimator. Note that the encoding approach adopted in [89,90] is highly
dependent on the system dynamics, and is therefore inapplicable to the state
estimation problem addressed here as the dynamics evolution of the system
measurement is not readily available.
Based on the above discussion, we are motivated to develop an appropriate
event-based encoding-decoding approach for solving the state estimation
problem under bit rate constraints. The difficulties we are going to face in
pursuit of our goal are foreseen as follows: 1) it is unclear as how to design
an effective encoding-decoding scheme for the measurement outputs whose
dynamical evolution is often unavailable; 2) it is nontrivial to derive a lower
bound of the required bit rate that ensures the convergence of the decoding
error; and 3) it is difficult to establish quantitative relationship between the
system parameters and the encoding/estimation performance.

9.1 Problem Formulation

9.1.1 System Model
Consider the following linear continuous-time system:

ẋ(t) = Ax(t) + Bw(t) (9.1)

Event-Based State Estimation under Constrained Bit Rate 187

where x(t) ∈ Rn is the state vector and w(t) ∈ Rp is the bounded disturbance
input satisfying kw(t)k ≤ w̄ with w̄ being a known positive scalar. A and B
are known matrices with appropriate dimensions.
The measurement model is given as follows:

y(t) = Cx(t) + Dv(t) (9.2)

where y(t) ∈ Rm is the measurement output and v(t) ∈ Rq is the measurement

noise satisfying kv(t)k ≤ v̄ with v̄ being a given positive scalar. Moreover, v(t)
is differentiable and its derivative satisfies kv̇(t)k ≤ ν̄ with ν̄ > 0. C and D are
known matrices with compatible dimensions and the pair (A, C) is assumed
to be observable.

9.1.2 Event-Based Encoder-Decoder Design

In practical applications of NSs, it is often the case that those measurement
signals of large bits fail to be directly transmitted to the receiver side due
to the bit-rate restriction of the communication network. To this end, an
event-based data encoding scheme is introduced to improve the efficiency
of the data transmission by i) compressing the data through encoding and ii)
reducing transmission frequency through event-based scheme.
The encoding scheme is characterized by the vector-valued function θ( · ) :
Rm → Rm with the following form:

θ (ℏ) , vec{θi (ℏi )} (9.3)

where θi (ℏi ) , sign(ℏi ) ∈ R denotes the scalar-valued codeword with fewer

bits, sign( · ) is the signum function, and ℏi is the ith component of the data ℏ
to be encoded.
Denoting the time sequence {sk }k∈Z as the transmission instants of the
codeword sequence, sk can be determined by the following rule:

sk+1 = inf{t|t > sk , ky(t) − ya (t)k = σ (t)} (9.4)

where ya (t) is the estimate of y(t) whose dynamics shall be given later and
σ (t) is the exponential decay function defined as

σ (t) , σ0 e−λt + ̺.

Here, σ0 and ̺ are given positive constants, and λ is a parameter to be

determined later. The encoding instant and the transmission instant of the
codeword are assumed to be the same one, that is, there is no time delay
caused by the data encoding process.
188 Data-Rate-Constrained State Estimation and Control

In this paper, the differential encoding scheme is adopted, which means

that the encoding operation is applied to the difference between the system
measurement y(t) and its estimate ya (t). Letting ỹa (t) , y(t) − ya (t), the
encoded signal is expressed as follows:



θi ỹai (sk ) = sign ỹai (sk ) (9.5)

where ỹai (t) (i = 1, 2, . . . , m) is the i-th component of ỹa (t). In what follows,


we slightly abuse the notation θi (sk ) to denote the encoded signal θi ỹai (sk ) .
Furthermore, we assume that θi (sk ) will not be sent to the decoder side if
y˜ ai (sk ) = 0.
In the sequel, we aim to characterize the dynamics of the measurement
output y(t) by using the following assumption and lemmas.

Assumption 9.1: The measurement matrix C is of full-row rank, namely,

rank{C} = m.

Lemma 9.1 (Singular value decomposition): For a given matrix C ∈ Rm×n of

full-row rank, there exist two orthogonal matrices U ∈ Rm×m and V ∈ Rn×n such
that

C = U CV T (9.6)

where σi > 0 (i = 1, 2, . . . , m) are the singular values of C and C , [C0 0m×(n−m) ]

with C0 , diag{σ1 , σ2 , . . . , σm }.

Lemma 9.2: For a given matrix A ∈ Rn×n and a given matrix C ∈ Rm×n of full-
row rank, there exists a matrix

Z , UC0 V1T AV1 (UC0 )−1 ∈ Rm×m

such that CA = ZC if A satisfies V1T AV2 = 0 where V1 ∈ Rn×m and V2 ∈ Rn×(n−m)

are defined through V , [V1 V2 ], with orthogonal matrices U ∈ Rm×m and
V ∈ Rn×n defined in Lemma 9.1.

Proof: The proof of this lemma is easily accessible from Lemma 9.1 and is therefore
omitted.

Lemma 9.3: The dynamic evolution of the measurement output y(t) is governed by
the differential equation

ẏ(t) = Zy(t) + Gξ(t) (9.7)

Event-Based State Estimation under Constrained Bit Rate 189

where
 
G , CB −ZD D
 T
ξ(t) , wT (t) vT (t) v̇T (t) .

Proof: It is seen from (9.1)–(9.2) that

ẏ(t) = CAx(t) + CBw(t) + Dv̇(t).

Then, in terms of Lemma 9.2, one further calculates that

ẏ(t) = Z(y(t) − Dv(t)) + CBw(t) + Dv̇(t)

(9.8)
= Zy(t) + Gξ(t),

which ends the proof.

Remark 9.1: Inspired by [89, 90], in this paper, we make the first attempt to
deal with the state estimation problem under the sign-information-based encoding-
decoding scheme. Note that, since the measurement matrix C is generally non-
invertible, there is a substantial difficulty in characterizing the dynamic evolution
of y(t) by means of a differential equation. To overcome such an intrinsic difficulty,
the singular value decomposition technique is applied to the measurement matrix
C, which plays a vital role in deriving the dynamic evolution of y(t) under certain
assumptions. In addition, the establishment of equation (9.8) facilitates, to a great
extent, the subsequent encoder/decoder analysis and synthesis.

In order to capture the dynamics of the measurement output y(t) at the

encoder side, we construct the following auxiliary system which is regarded
as a “copy” of the decoder function


 ẏa (t) = Zya (t), t ∈ [sk , sk+1 )


1

ya (s+k ) = ya (sk ) + m θ(sk )σ (sk )
(9.9)




ya (0) = 0, t = 0.


 T
Here, θ(sk ) , θ1 (sk ) θ2 (sk ) . . . θm (sk ) , σ (sk ) , σ0 e−λsk +̺ and s+
k stands
+
for the right limit of the triggering instant sk , namely, sk , limǫ>0,ǫ→0 {sk + ǫ}.
Similar to the results in [81–83], the auxiliary system (9.9) is a kind of hybrid
systems where sk can be viewed as the impulsive sequence determined by
the event function (9.4).
190 Data-Rate-Constrained State Estimation and Control

According to Lemma 9.2 and (9.9), the dynamics of ỹa (t) is derived as

ỹ˙ (t) = Zỹa (t) + Gξ(t), t ∈ [sk , sk+1 )
 a



1

ỹa (sk+ ) = ỹa (sk ) − θ(sk )σ (sk ) (9.10)


 m

ỹa (0) = y0 , t = 0


where y0 is the known initial value of the measurement output.

Following the dynamics ya (t) in (9.9), the decoder dynamics of the
measurement output y(t) is governed by

y˙ (t) = Zyd (t), t ∈ [sk , sk+1 )
 d



1

yd (s+k ) = yd (sk ) + m θ (sk )σ (sk )
(9.11)




 y (0) = 0, t = 0
d

where yd (t) is the decoded value of y(t).

Denoting ỹd (t) , y(t) − yd (t), the decoding error dynamics is derived as
follows:

ỹ˙ (t) = Zỹd (t) + Gξ(t), t ∈ [sk , sk+1 )
 d



1

ỹd (sk+ ) = ỹd (sk ) − θ (sk )σ (sk ) (9.12)


 m

ỹd (0) = y0 , t = 0.


It is not difficult to verify from (9.9)–(9.12) that ya (t) ≡ yd (t) and

ỹa (t) ≡ ỹd (t).

9.1.3 State Estimator Structure

After obtaining the decoded measurement ỹd (t) generated by (9.11), we
construct the following Luenberger-type state estimator to observe the states
of system (9.1):

ẋˆ(t) = Ax̂(t) + K(yd (t) − Cx̂(t)) (9.13)

where x̂(t) ∈ Rn is an estimate of x(t).

Denoting the estimation error as x̃(t) , x(t) − x̂(t), together with (9.1)–(9.2)
and the definition of ỹd (t), we have

ẋ˜(t) = Āx̃(t) + B̄ϑ(t) (9.14)

Event-Based State Estimation under Constrained Bit Rate 191

Encoder

FIGURE 9.1
Schematic of state estimation problem under constrained bit rate.

where

Ā , A − KC
B̄ , [B −KD K]
ϑ(t) , [wT (t) vT (t) ỹTd (t)]T .

In this paper, we investigate the state estimation problem under constrained

bit rate, where the schematic structure is shown in Figure 9.1. Our main
objective is to develop a state estimator under an event-based encoding-
decoding scheme for the continuous-time system (9.1)–(9.2) such that the
estimation error is exponentially bounded subject to the system noise w(t),
the measurement noise v(t) and the decoding error ỹd (t).

9.2 Main Results

In this section, a positive lower bound is first derived for the event-
triggered condition to exclude the Zeno phenomenon. Then, a sufficient
condition of the bit rate is established to ensure that the decoding error
ỹd (t) is ultimately bounded. Subsequently, a necessary condition of the
bit rate is derived to provide a lower bound of the bit rate required to
ensure the convergence of the decoding error. Furthermore, by utilizing the
decoded measurement yd (t), a state estimator is constructed with guaranteed
exponential boundedness of the estimation error.

9.2.1 Elimination of the Zeno Phenomenon

For event-triggered mechanism implemented in continuous time setting,
the notorious Zeno phenomenon could damage the data sampling/sensing
devices and induce the network congestion, and such a phenomenon should
192 Data-Rate-Constrained State Estimation and Control

be strictly avoided. In what follows, we shall exclude the Zeno phenomenon

from the proposed event-triggered mechanism by analyzing the dynamics
of ỹa (t).

Lemma 9.4: Under the event-triggered mechanism (9.4), if

ξ¯ kGk
ζ1 + < 1,
̺

then the triggering sequence {sk }k∈Z satisfies

sk+1 − sk ≥ δ

where
 
ξ̄ kGk
− ln ζ1 + ̺
δ, , ξ¯ , w̄ + v¯ + ν̄
kZk + λ¯
r
1
λ̄ , max{λ, 1}, ζ1 , 1− .
m

Proof: Let us first focus our attention on the dynamics evolution of ỹa (t) over the
time interval [sk , sk+1 ). Based on (9.10), ỹa (t) is calculated as
Z t
Z(t−sk )
ỹa (t) = e ỹa (s+
k )+eZ(t−s) Gξ(s)ds
sk
 
Z(t−sk ) 1
=e ỹa (sk ) − θ(sk )σ (sk ) (9.15)
m
Z t
+ eZ(t−s) Gξ(s)ds.
sk
n o
σi (sk )
Setting σi (sk ) , |yi (sk ) − yai (sk )| (i = 1, 2, . . . , m) and 6(sk ) , diagm σ (sk ) ,
together with the definition of θ (sk ), we rewrite ỹa (sk ) as

ỹa (sk ) = 6(sk )θ (sk )σ (sk ),

which results in
 
1
ỹa (t) = eZ(t−sk ) 6(sk ) − Im θ(sk )σ (sk )
m
Z t (9.16)
+ eZ(t−s) Gξ(s)ds.
sk
Event-Based State Estimation under Constrained Bit Rate 193

Subsequently, it is inferred from (9.16) that

 
Z(t−sk ) 1
kỹa (t)k ≤ e 6(sk ) − Im θ(sk )σ (sk )
m
Z t−sk
+ eZs Gξ(t − s)ds
0
  (9.17)
kZk(t−sk ) 1
≤e 6(sk ) − Im θ(sk ) σ (sk )
m
Z t−sk
+ eZs Gξ(t − s)ds .
0

By applying some algebraic manipulations, the second term in the right-hand side of
(9.16) can be further reduced to

Z t−sk
"Z
t−sk TZ t−sk # 12
eZs Gξ(t − s)ds = Zs
e Gξ(t − s)ds eZs Gξ(t − s) ds
0 0 0

 Z t−sk  T   12

Zs Zs
≤ (t − sk ) e Gξ(t − s) e Gξ(t − s) ds
0

 Z t−sk  12
2
Zs
= (t − sk ) e Gξ(t − s) ds
0

1 2
≤ (t − sk ) 2 (t − sk ) eZ(t−sk ) kGk2

n o 12
2
× max kξ(t − s)k
s∈[0, t−sk ]

≤ (t − sk )ξ¯ kGk eZ(t−sk )

≤ (t − sk )ξ¯ kGkekZk(t−sk ) (9.18)

where the first inequality is based on the well-known Jensen inequality.

Recalling that kw(t)k ≤ w, ¯ kv(t)k ≤ v, ¯ kv̇(t)k ≤ ν, ¯ ξ(t) =
 T T p
T T 2 2 2
w (t) v (t) v̇ (t) and ξ̄ , w̄ + v̄ + ν̄ , the third inequality follows from
  1 p
the fact maxs∈[0, t−sk ] kξ(t − s)k2 2 ≤ w̄2 + v̄2 + ν̄ 2 = ξ̄ . Then, it is readily
calculated that
194 Data-Rate-Constrained State Estimation and Control

 
1
6(sk ) − Im θ (sk )
m
v
u m 2  
uX θi (sk )
2 2σi (sk )σ (sk ) σ 2 (sk )
= t σ (sk ) − +
σ 2 (sk ) i m m2 (9.19)
i=1
v
u m 2
uX θi (sk )σi2 (sk ) m
X θi2 (sk )σi (sk ) Xm
θi2 (sk )
=t − 2 + .
σ 2 (sk ) mσ (sk ) m2
i=1 i=1 i=1

Furthermore, based on the facts σi (sk ) > 0 (i = 1, 2, . . . , m),

Pm 2 = σ 2 (sk )
i=1 σi (sk )
and θi2 (sk ) = 1, it is easy to verify that

m
θ 2 (sk )σ 2 (sk ) 2
Pm
X
i i i=1 σi (sk ) σ 2 (sk )
= = =1
σ 2 (s k) σ 2 (sk ) σ 2 (sk )
i=1

and
σ1 (sk ) + σ2 (sk ) + · · · + σm (sk ) √
1≤ ≤ m. (9.20)
σ (sk )

Pm θi2 (sk )σi (sk )

Thus, the term i=1 mσ (sk ) is bounded by

√
1
m
X θi2 (sk )σi (sk ) m
≤ ≤ ,
m mσ (sk ) m
i=1

which yields
√
2 m
m
X θi2 (sk )σi (sk ) 2
− ≤ −2 ≤− .
m mσ (sk ) m
i=1

Consequently, in accordance with the above analysis, we draw the conclusion that
 
1
ζ0 ≤ 6(sk ) − Im θ (sk ) ≤ ζ1 (9.21)
m

where
s
√ r
2 m 1 1
ζ0 , 1− + 2, ζ1 , 1− .
m m m
Event-Based State Estimation under Constrained Bit Rate 195

In addition, one derives that

σ (sk ) = σ0 e−λsk + ̺
= σ0 eλ(t−sk ) e−λt + ̺
≤ eλ(t−sk ) (σ0 e−λt + ̺)
= eλ(t−sk ) σ (t).

Also, it follows from (9.15)–(9.21) that

kỹa (t)k ≤ ζ1 e(kZk+λ)(t−sk ) σ (t) + (t − sk )ξ̄ kGkekZk(t−sk )

, ϕ(t). (9.22)

It is obvious that ϕ(t) is a monotonically increasing continuous function with regard

to t over the interval [sk , sk+1 ) with ϕ(sk ) < σ (sk ). As such, there must exist a time
instant t∗ ∈ [sk , sk+1 ) such that ϕ(t∗ ) = σ (t∗ ), that is,
∗ −s
σ (t∗ ) = ζ1 e(kZk+λ)(t k) σ (t∗ )
∗ −s
+ (t∗ − sk )ξ¯ kGkekZk(t k) . (9.23)

Noticing the form of σ (t), it is straightforward to see that

 ∗
 ∗
1 − ζ1 e(kZk+λ)(t −sk ) ̺ ≤ (t∗ − sk )ξ̄ kGkekZk(t −sk ) . (9.24)

Moreover, it can be concluded from the fact

t − sk < et−sk (∀t ≥ sk )

and the form of λ̄ that

¯ ∗ −s ξ¯ kGk (kZk+λ)(
¯ t∗ −sk )
1 − ζ1 e(kZk+λ)(t k) < e , (9.25)
̺

which means that

¯ ∗ −s 1
e(kZk+λ)(t k) > . (9.26)
ξ¯ kGk
ζ1 + ̺

Taking the logarithmic operation on both sides of (9.26) yields

1
ln  ¯

ζ1 + ξ kGk
̺
t∗ − sk > , δ, (9.27)
kZk + λ̄
196 Data-Rate-Constrained State Estimation and Control

which indicates that

sk+1 − sk > δ > 0.

Therefore, the Zeno behavior is eliminated and the proof of this lemma is ended.

9.2.2 The Boundedness Analysis of the Decoding Error

As we know, the decoding error has played a vital role in the estimation
performance. In other words, with the increase of the decoding error, the
estimation performance would accordingly get worse. Therefore, in this
subsection, we are interested in establishing some criteria to analyze the
boundedness of the decoding error under the bit-rate-constraint. First, a
sufficient condition on the network bit rate is given under which the
boundedness of the measurement decoding error ỹd (t) is ensured.
Bearing in mind that

ỹd (t) ≡ ỹa (t), ∀t ≥ 0,

we know that the boundedness of ỹd (t) is equivalent to that of ỹa (t). Then,
we can infer from (9.22)–(9.23) that

kỹa (t)k < σ (t), t ∈ [sk , t∗ ].

In addition, it is observed from the event-triggered mechanism (9.4) that

kỹa (t)k ≤ σ (t), ∀t ∈ [t∗ , sk+1 ].

Hence, we have

kỹd (t)k = kỹa (t)k ≤ σ (t), ∀t ≥ 0,

which implies that the decoding error ỹd (t) is ultimately bounded.

Theorem 9.1: Under the event-based encoding-decoding scheme (9.4), (9.5) and
(9.11), the decoding error ỹd (t) is ultimately bounded if the bit rate R of the
communication channel satisfies

m
R ≥ R0 , .
δ

Proof: It follows from the boundedness analysis of ỹd (t) that the decoding error is
ultimately bounded under the event-based encoding-decoding scheme (9.4), (9.5) and
(9.11).
Event-Based State Estimation under Constrained Bit Rate 197

Next, the encoding rule (9.5) indicates that, at each triggering instant sk ,
only one bit data is transmitted for each measurement component yi (t) (i = 1,
2, . . . , m), and thus m-bits data are required for y(t) during each data
transmission.
Now, based on Lemma 9.4, the required bit rate is mδ for the sake of guaranteeing
the boundedness of the decoding error. The proof is complete.

Remark 9.2: A sufficient condition is proposed in Theorem 9.1 to guarantee that

the decoding error ỹd (t) is ultimately bounded under the event-based encoding-
decoding scheme (9.4). It is observed from (9.27) that the derived bit-rate condition
is dependent on i) the system parameters (i.e. A which is reflected in Z), ii) the
dimension of the system measurement m, iii) the bound of the disturbances (i.e.
ξ̄ = w̄ + v̄ + ν̄), and iv) the threshold parameter ̺. To be specific, a large bound
of system disturbances leads to a large bit rate requirement. The increase of the
dimension of the system measurement leads to the increase of bit rate. Furthermore, a
larger threshold contributes to less bit rate while giving rise to a larger upper bound
of the decoding error. These observations reveal the relationship between the bit rate
and certain system performance, which is consistent with the practice.

Having established a sufficient condition to ensure the boundedness of the

decoding error, it is time for us to derive a corresponding necessary condition
regarding the bit rate of the communication network. To this end, we shall
obtain an upper bound of two arbitrarily adjacent triggering instants. Before
proceeding further, the following useful lemma is provided.

Lemma 9.5: Under the event-based encoding-decoding scheme (9.4), (9.5) and
(9.11), the decoding error is divergent if the bit rate of the communication channel
dξ̄ kGk
m ln (ζ0 χ)−ln
is less than R1 , namely, R < R1 , where R1 , δ1 , δ1 , 1
̺
, a is the
d +kZk−a
smallest real part of the eigenvalues of Z, d is a positive scalar and other parameters
are defined in Lemma 9.4.

Proof: The main idea of deriving the bit rate R1 is to seek an upper bound of the
time interval between two successive triggering instants. To proceed further, we first
give the following assertion.
Assertion: For the matrix Z ∈ Rm×m , there exists a positive real constant
ς > 0 and a positive-definite matrix P ∈ Rm×m such that the following inequality
holds

(eZ(t−sk ) )T eZ(t−sk ) ≥ χe2a(t−sk ) I, ∀t ≥ sk (9.28)

where χ , λmin {PT P}λmin {P−T P−1 }ς and a is the smallest real part of the
eigenvalues of Z.
198 Data-Rate-Constrained State Estimation and Control

In what follows, we aim to prove the matrix inequality (9.28). To begin with, let
us recall some properties of the matrix transformation.
Letting the matrix Z has r (r ≤ m) different eigenvalues λ1 , λ2 , . . . , λr , the
characteristic polynomial of Z is written as

det |λI − Z| = (λ − λ1 )m1 (λ − λ2 )m2 · · · (λ − λr )mr

where mi (i = 1, 2, . . . , r) is the algebraic multiplicity of λi and satisfies

P r
i=1 mi = m. P
Let the geometric multiplicity of λi be li and denote l , ri=1 li . It is known that

li ≤ mi , ∀i ∈ {1, 2, . . . , r},

which indicates l ≤ m.
In terms of the theory of linear algebra, there exists a non-singular matrix P ∈
Rm×m such that

P−1 ZP = J

where J , diag{J1 , J2 , . . . , Jl } with Ji (i = 1, 2, . . . , l) being the Jordan blocks.

By properly arranging the order of the Jordan blocks Ji , we know that the first l1
Jordan blocks J1 , J2 , . . . , Jl1 are with the eigenvalue λ1 , the eigenvalue of the Jordan
blocks from Jl1 +1 to Jl1 +l2 is λ2 and, by analogy, the eigenvalue of the Jordan blocks
from JPr−1 l +1 to Jl is λr .
i=1 i
Since λi (i = 1, 2, . . . , r) is probably a complex number, we denote λi , ai + bi i,
where ai and bi are, respectively, the real and imaginary parts of λi .
Without loss of generality, suppose that the real parts of λi (i = 1, . . . , r) satisfy
a1 ≤ a2 ≤ · · · ≤ ar . Then, for any Jordan block Ji (λs ) (i = 1, 2, . . . , l), it can be
rewritten as
 
λs 1 0 · · · 0
 0 λs 1 · · · 0 
Ji (λs ) ,  .  , s ∈ {1, 2, . . . , r}.
 
.. .. .
 .. . . .
. 1
0 0 0 · · · λs

In the sequel, for notation simplicity, Ji is slightly abused to denote Ji (λs ).

It follows from the property of the matrix exponent that
−1 ZP
P−1 eZ(t−sk ) P = e(t−sk )P
= e(t−sk )J
= diag{e(t−sk )J1 , e(t−sk )J2 , . . . , e(t−sk )Jl },
Event-Based State Estimation under Constrained Bit Rate 199

which indicates
−1 ZP
eZ(t−sk ) = Pe(t−sk )P P−1
= P diag{e(t−sk )J1 , e(t−sk )J2 , . . . , e(t−sk )Jl }P−1 .

As such, one further obtains

T (t−s T T T
eZ k) eZ(t−sk ) = P−T diag{e(t−sk )J1 , e(t−sk )J2 , . . . , e(t−sk )Jl }PT P
× diag{e(t−sk )J1 , e(t−sk )J2 , . . . , e(t−sk )Jl }P−1
T T T
≥ λmin {PT P}P−T diag{e(t−sk )J1 , e(t−sk )J2 , . . . , e(t−sk )Jl }
× diag{e(t−sk )J1 , e(t−sk )J2 , . . . , e(t−sk )Jl }P−1 . (9.29)

For Ji (i ∈ {1, 2, . . . , l}), noticing that

Ji = λs I + Ji (0)

and

(λs I)Ji (0) = Ji (0)λs I,

eJi can be rewritten as

eJi = eλs I+Ji (0) = eλs I eJi (0) .

Hence, one obtains

T T
e(t−sk )Ji e(t−sk )Ji = e(t−sk )λ̄s I e(t−sk )Ji (0) e(t−sk )Ji (0) e(t−sk )λs I

with λ̄s being the conjugate of λs .

T
Since e(t−sk )Ji (0) e(t−sk )Ji (0) is semi-positive definite and e(t−sk )Ji (0) is invertible, one
T
has that e(t−sk )Ji (0) e(t−sk )Ji (0) is a positive definite matrix. Thus, there exists a positive
scalar ςi such that
T
e(t−sk )Ji (0) e(t−sk )Ji (0) ≥ ςi I.

In this sense, one derives that

T
e(t−sk )Ji e(t−sk )Ji ≥ ςi e2(t−sk )as I

where as is the real part of λs and λ̄s .

200 Data-Rate-Constrained State Estimation and Control

Finally, it is inferred from (9.29) and the above discussion that

T (t−s
eZ k) eZ(t−sk ) ≥ λmin {PT P}λmin {P−T P−1 }ς e2a(t−sk ) I (9.30)

where ς , min{ςi } and a , min{ai } (i = 1, 2, . . . , r). The proof of the Assertion is

now complete.

Then, we focus our attention on analysis of the dynamics of the term

kỹa (t)k. It is inferred from (9.15) that
 
1
kỹa (t)k = eZ(t−sk ) 6(sk ) − Im θ(sk )σ (sk )
m
Z t
+ eZ(t−s) Gξ(s)ds
sk
  (9.31)
Z(t−sk ) 1
≥ e 6(sk ) − Im θ(sk )σ (sk )
m
Z t
− eZ(t−s) Gξ(s)ds .
sk

With the help of (9.21), (9.28) and the Jensen inequality, one further
calculates
Z t
a(t−sk )
ky(t) − ya (t)k ≥ ζ0 χ e σ (sk ) − eZ(t−s) Gξ(s)ds
sk

≥ ζ0 χea(t−sk ) σ (sk ) − (t − sk )ξ¯ kGkekZk(t−sk ) (9.32)

 
a(t−sk ) dξ¯ kGk ( 1 +kZk)(t−sk )
≥ ζ0 χe − e d σ (t)
̺

where the first inequality comes from (9.18) and the second inequality is
based on the facts 

 σ (sk ) ≥ σ (t)

t−sk
(t − sk ) ≤ de d (9.33)


 ̺ ≤ σ (t)
with d > 0.
Letting
 
a(t−sk ) dξ¯ kGk ( 1 +kZk)(t−sk )
ζ0 χ e − e d = 1,
̺
Event-Based State Estimation under Constrained Bit Rate 201

we can obtain

ln (ζ0 χ) − ln dξ̄ kGk

̺
t − sk ≤ 1
, δ1
d + kZk − a

as long as

dξ¯ kGk
ζ0 χ − > 0.
̺

Thus, by recalling the event-triggered mechanism (9.4), we draw a conclusion

that the time interval of two successive triggering instants is bounded by δ1 ,
namely,

sk+1 − sk ≤ δ1 .

As such, we conclude that the decoding error is divergent if

m
R < R1 , .
δ1

The proof of this Lemma is complete.

Remark 9.3: Except for the sufficient condition on the required bit rate R0 obtained
from Theorem 9.1, we further derive a bit rate condition R1 in Lemma 9.5,
whose theoretical and practical significance lies in that 1) a necessary condition
is established for the bit-rate-constrained state estimation problem, namely, below
which the estimation error would be divergent; 2) a lower bound of the bandwidth
requirement on the allocation of the network resource is given; and 3) a new
performance index 1R , R0 − R1 is provided to evaluate the conservativeness of
the sufficient condition in Theorem 9.1. To be more specific, the smaller the 1R, the
less the conservativeness of the bit rate condition in Theorem 9.1. For the extreme case
1R = 0, the sufficient condition in Theorem 1 becomes the necessary and sufficient
condition. Therefore, in terms of the bit rates R0 and R1 , a more accurate relationship
between the estimation performance and the communication resource is established,
which is of both theoretical significance and practical importance.

9.2.3 The Design of the State Estimator

For now, we have dealt with the boundedness analysis of the decoding error
dynamics ỹd (t) and the corresponding data rate condition. Next, we are going
to tackle the state estimation issue by utilizing the decoded measurement
yd (t).
202 Data-Rate-Constrained State Estimation and Control

In the following theorem, a sufficient condition is established to ensure the

ultimate boundedness of the estimation error.

Theorem 9.2: Let the bit rate of the communication network satisfy R ≥ R0 . The
error dynamics (9.14) of the state estimation is ultimately bounded under the event-
based encoding-decoding scheme (9.4), (9.5) and (9.11) if there exist positive scalars
α, γ , a positive definite matrix Q and a matrix K̄ satisfying
 
ĀT Q + QA¯ + αQ Q̄
4, <0 (9.34)
∗ −γ I

where

Q̄ , [QB −K̄D K̄].

Moreover, the estimator gain matrix is designed as

K = Q−1 K̄

and the ultimate upper bound of the estimation error kx̃(t)k is calculated as
s
γ ϑ̄ T ϑ̄
αλmin {Q}

with

ϑ̄ , [w̄ v̄ σ0 + ̺]T .

Proof: It is known from subsection 9.2.2 that the boundedness of the decoding error
ỹd (t) can be guaranteed under the bit rate condition R ≥ R0 , which implies that the
vector ϑ(t) is bounded.
Construct the Lyapunov function as follows:

V (t) = x̃T (t)Qx̃(t).

Then, the derivative of V(t) along the trajectory of the estimation error dynamics
(9.14) can be calculated as

V̇(t) = 2x̃T (t)Qx̃˙(t)

= 2x̃T (t)Q(Āx̃(t) + B̄ϑ(t)) (9.35)

= x̃T (t)(ĀT Q + QĀ)x̃(t) + 2x̃T (t)QB̄ϑ(t),

Event-Based State Estimation under Constrained Bit Rate 203

which, together with 4 < 0 in (9.34), results in

V̇(t) + αV(t) < γ ϑ T (t)ϑ(t).

As such, we have
Z t
−αt
V (t) ≤ e V (0) + γ e−α(t−s) ϑ T (s)ϑ(s)ds
0
γ
≤ e−αt V(0) + (1 − e−αt )ϑ̄ T ϑ̄,
α

which indicates that the estimation error satisfies

s s
λmax {Q} − αt γ (1 − e−α t )ϑ¯ T ϑ¯
kx̃(t)k ≤ e 2 kx˜(0)k + ,
λmin {Q} αλmin {Q}

and is therefore exponentially bounded. Accordingly, the ultimate upper bound can
be expressed by s
γ ϑ̄ T ϑ̄
αλmin {Q}
and the proof of this theorem is complete.

Remark 9.4: It follows from subsection 9.2.2 that the decoding error ỹd (t)
asymptotically tends to ̺. Based on this fact, we have

¯ 2 + v̄2 + ̺)
Z t ϑ T (t)ϑ(t) (w
lim e−α(t−s) ϑ T (s)ϑ(s)ds ≤ lim ≤ ,
t→∞ 0 t→∞ α α

which means that, when the time t tends to infinity, a tighter upper bound of the
estimation error is obtained as
s
γ (w̄2 + v̄2 + ̺)
.
αλmin {Q}

Remark 9.5: Till now, a systematic study has been conducted on the encoding-
decoding-based state estimation problem for a class of continuous-time systems
under a bit-rate constrained network. The emphasis is placed on the research with
respect to the joint design issue of encoder/decoder/estimator. To be more specific,
in Lemma 9.4, the unexpected Zeno phenomenon has been excluded in order to
ensure the normal operation of the designed event-based encoder. In Theorem 9.1,
a sufficient condition has been first established to guarantee the boundedness of
the decoding error. Then, in Lemma 9.5, a necessary condition has been given to
204 Data-Rate-Constrained State Estimation and Control

illustrate that the decoding error would be divergent when the bit rate is below a
given condition. Furthermore, Theorem 9.2 has provided a sufficient condition to
ensure the exponential boundedness of the estimation errors on the premise of the
convergence of the decoding errors.

Remark 9.6: Comparing with existing literature, the main novelties of our current
results lie in that: 1) the research problem addressed is new that represents the first
attempt to investigate the event-based encoding-decoding state estimation problem
under bit rate constraints; 2) the dynamical evolution of the measurement output
is proposed with the help of the singular value decomposition technique that offers
a better performance for the developed sign-information-based encoding-decoding
scheme; and 3) the derived bit rate condition is viewed as a bridge to establish the
tight connection among the network resources and the decoding errors as well as the
estimation performance.

9.3 An Illustrative Example

Two illustrative simulation examples are presented in this section to show
the effectiveness of the proposed state estimation method.

Example 9.1: Consider a linear continuous-time system with the following

parameters:
   
0.5 0.25 0.1
A= , B= ,
0.8 0.1333 0.1
C = [0.8 0.3], D = 0.1.

The eigenvalues of A are 0.8 and −0.1667, which indicate that system (9.1) is
unstable and it is therefore non-trivial to verify the developed estimation scheme on
this system.
By applying the singular value decomposition technique to matrix C, U, C0 and
V, one can obtain

U = 1, C0 = 0.8544
 
0.9363 −0.3511
V=
0.3511 0.9363

V1 = [0.9363 0.3511]T

V2 = [−0.3511 −0.9363]T ,
Event-Based State Estimation under Constrained Bit Rate 205

which ensures that V1T AV2 = 0. Moreover, the matrix Z is computed as

Z = 0.8. The disturbances w(t) and v(t) are chosen as w(t) = v(t) = 0.1 sin (t),
which imply that w̄ = v̄ = ν̄ = 0.1.
By solving the matrix inequality (9.34) using Matlab toolbox, the estimator gain
matrix K is obtained as

K = [2.3726 2.1875].

According to Theorem 9.1, the required bit rate to achieve the ultimate boundedness
of the estimation error is R = 0.6034 bits/second.
The simulation results are shown in Figures 9.2-9.6. The measurement output
y(t) and its decoded value are illustrated in Figure 9.2, and the corresponding
decoding error is depicted in Figure 9.3. The system states xi (t) (i = 1, 2) and
their estimations are described in Figure 9.4. The estimation errors are shown in
Figure 9.5. In Figure 9.6, the dynamics of the triggering function ky(t) − ya (t)k and
the corresponding threshold are given, from which it is observed that the x-axis of
the intersection point of the two curves is the triggering instants. All the simulation
results have confirmed the validity of the proposed encoding/estimation method.

Next, in order to make our theoretical results more convincing, a practical

example (i.e. a maneuvering target tracking system) is introduced.

250

200

150

100

50

0
0 1 2 3 4 5 6 7 8

FIGURE 9.2
Measurement y(t) and its decoded value yd (t).
206 Data-Rate-Constrained State Estimation and Control

6
The norm of the decoding error

0
0 5 10 15 20 25 30 35 40
time

FIGURE 9.3
The norm of the decoding error ỹd (t).
State strategies and their

1000

800
estimates

600

400

200

0
0 1 2 3 4 5 6 7 8 9 10
time

1500
State strategies and their
estimates

1000

500

0
0 1 2 3 4 5 6 7 8 9 10
time

FIGURE 9.4
The state trajectories xi (t) and their estimates x̂i (t) (i = 1, 2).
Event-Based State Estimation under Constrained Bit Rate 207

6
The estimation errors

0
0 5 10 15 20 25 30 35 40
time

8
The estimation errors

0
0 5 10 15 20 25 30 35 40
time

FIGURE 9.5
The estimation errors x̃i (t) (i = 1, 2).

12

10
Triggering threshold and the
norm of the decoding error

0
0 5 10 15 20 25 30 35 40
time

FIGURE 9.6
Triggering threshold σ (t) and the norm of the decoding error kỹd (t)k.
208 Data-Rate-Constrained State Estimation and Control

Example 9.2: The target plant is described as follows:


ẋ1 (t) = x2 (t) + 0.1w(t)

ẋ2 (t) = x3 (t) + 0.1w(t)


ẋ3 (t) = u(t)

where x1 (t), x2 (t) and x3 (t) stand for, respectively, the position, the velocity and the
acceleration of the target. In this case, the matrices A and B are
 
0 1 0  
A = 0 0 1 , B = 0.1 0.1 0 .
0 0 0

The measurement matrix is given as

 
0.01 0.1 0
C= .
0 0 0.9

Let u(t) = 0 and the other parameters are set the same as the ones in Example 9.1.
By solving the linear matrix inequality in Theorem 9.2, the estimator gain matrix is
given by
 
11.0827 −0.7379
K =  3.7083 1.3152  .
0.1407 0.1326

It is calculated that V1T AV2 = [0 0.0099] and thus the equality constraint
in Lemma 9.2 is considered to be met. The initial conditions are set as
x(0) = [0.2 0 −0.15]T , x̂(0) = [0 0 0]T , y(0) = Cx(0) and ya (0) =
yd (0) = [0.5 0]T . As expected, it is observed from Figures 9.7–9.8 that simulation
results well confirm the validity of the proposed estimation scheme. Moreover, in
order to examine the impact of the intensity of the disturbances w(t) and v(t) and
the triggering threshold ̺ on the bit-rate condition, some comparisons are made as
shown in Tables 9.1 and 9.2. It is can be concluded from Tables 9.1 and 9.2 that: 1)
the severer the intensities of the disturbances w(t) and v(t) are, the larger bit rate of
the communication channel is required and 2) the larger the triggering threshold ̺
is, the smaller the required bit rate would be. These comparison results are consistent
with the theoretical analysis.
Event-Based State Estimation under Constrained Bit Rate 209

The state trajectories

and their estimates
0

-1000

-2000
0 50 100 150
Time
The state trajectories
and their estimates

-10

-20
0 50 100 150
Time
The state trajectories
and their estimates

0
-0.2
-0.4
-0.6
0 50 100 150
Time

FIGURE 9.7
The state trajectories xi (t) (i = 1, 2, 3) and their estimates.
The estimation errors

50

0
0 50 100 150
Time
The estimation errors

10

0
0 50 100 150
Time
The estimation errors

0.5

0
0 50 100 150
Time

FIGURE 9.8
The estimation errors x̃i (t) (i = 1, 2, 3).
210 Data-Rate-Constrained State Estimation and Control

TABLE 9.1
Bit Rate R0 Under Different Threshold Values ̺

̺ 0.2 0.4 0.6 0.8 1.0

R0 27.83 11.16 9.17 8.39 7.98

TABLE 9.2
Bit Rate R0 Under Different Disturbance Intensities

w̄ = v̄ 0.1 0.15 0.2 0.25 0.3

R0 7.98 8.84 9.88 11.16 12.77

9.4 Summary
In this paper, the state estimation problem has been investigated for a class of
NSs under bit rate constraints. The measurement outputs have been encoded
into the binary codeword with 1 bit and then transmitted to the estimator via
a digital communication channel. For the resource-saving purpose, an event-
triggered coding-decoding mechanism has been employed in the process
of data encoding/transmission. A bit rate condition has been established
to guarantee the ultimate boundedness of both the decoding error and the
estimation error. Furthermore, the desired estimator gain has been acquired
in terms of the solution to a matrix inequality. Finally, a numerical simulation
has been proposed to demonstrate the validity of the main results.
10
Conclusions and Future Topics

In this book, the control and state estimation problems have been thoroughly
investigated for networked systems under a data-rate-constrained network
environment. Latest results on data-rate-constrained control and state
estimation issues have been first surveyed under different types of encoding-
decoding schemes and bandwidth constraints originating from real-world
engineering practice. Then, in each chapter, regarding the addressed
control and state estimation/filtering with data rate constraint, various
encoding-decoding strategies have been developed. Meanwhile, a set of
sufficient conditions have been established for designed encoding-decoding
strategies to achieve certain pre-specified system performances including
minimum variance specification, uniform boundedness index, consensus
and synchronization requirements. Subsequently, both numerical and
practical illustrative demonstrations have been presented that are capable
of verifying the feasibility and efficacy of all developed data-constrained
control/state estimation approaches.
This book has established a unified theoretical framework for the
synthesis/analysis problems for networked systems with constrained
network bandwidth. A great variety of engineering-oriented data rate
constraints and corresponding encoding-decoding schemes have been taken
into account in a systematical yet effective procedure. It is worth mentioning
that the acquired results are still quite limited, and some concluding remarks
and prospective topics are highlighted as follows:

(1) Channel Encoding. So far, most of the existing results focus their
attention on the aspect of source encoding and decoding. In fact, the
channel encoding-decoding is also very important for the reliability
of the data transmission in the communication network. Therefore,
it is quite meaningful to study the co-design issue of the source and
channel encoding-decoding in NSs.
(2) Combining with Other Communication Protocols. Recently, some effec-
tive data scheduling strategies such as the Round-Robin protocol, the
Try-Once-Discard protocol, the stochastic protocol, etc., have been
successfully applied to the NSs in order to cope with the limited
communication resources. For the purpose of making full use of the

DOI: 10.1201/9781003534853-10 211

212 Data-Rate-Constrained State Estimation and Control

network resources, the development of several encoding-decoding-

based data scheduling protocols would be a trend for future research.
(3) Data Encryption Issues. Since NSs share the common network links
for data exchange, they are prone to be attacked by adversaries,
and the secure communication in NSs becomes more and more
significant. Thereby, how to combine the encoding-decoding schemes
with the data encryption techniques in a unified framework is
another interesting topic.
(4) Combining with Self-Triggering Mechanism. Although there have been
some results with respect to the combination of the event-triggering
mechanism and the decoding-decoding communication schemes
recently, the additional hardware is still needed to detect the specific
“event”. In this sense, by replacing the detection task according to
certain admissible “prediction” [193], the self-triggering mechanism
is one preferable candidate to facilitate the encoding-decoding-based
control/filtering problems.
Bibliography

[1] F. Amato and M. Ariola, Finite-time control of discrete-time linear systems,

IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 724–729, 2005.
[2] S. Azuma and T. Sugie, Dynamic quantization of nonlinear control systems,
IEEE Transactions on Automatic Control, vol. 57, no. 4, pp. 875–888, 2012.
[3] H. Bai, M. Zhang, A. Wang, M. Liu, and Y. Zhao, Multiple description video
coding using inter- and intra-description correlation at macro block level, IEICE
Transactions on Information and Systems, vol. E97D, no. 2, pp. 384–387, 2014.
[4] Y. Barshalom, Redundancy and data compression in recursive estimation, IEEE
Transactions on Automatic Control, vol. AC17, no. 5, pp. 684–689, 1972.
[5] Y. Berrouche and R. E. Bekka, Improved multiple description wavelet
based image coding using Hadamard transform, AEU-International Journal of
Electronics and Communications, vol. 68, no. 10, pp. 976–982, 2014.
[6] R. Caballero-Águila, A. Hermoso-Carazo, J. Linares-Per ´ ez, and Z. Wang, A new
approach to distributed fusion filtering for networked systems with random
parameter matrices and correlated noises, Information Fusion, vol. 45, pp. 324–
332, 2019.
[7] R. Carli, F. Fagnani, and P. Frasca, Efficient quantized techniques for consensus
algorithms, NeCST Workshop, pp. 1–8, Nancy, France, 2007.
[8] J. Chen and Q. Ling, Bit-rate conditions for the consensus of quantized
multiagent systems with network-induced delays based on event triggering,
IEEE Transactions on Cybernetics, vol. 51, no. 2, pp. 984–993, 2021.
[9] X. Chen, X. Liao, L. Gao, S. Yang, H. Wang, and H. Li, Event-triggered
consensus for multi-agent networks with switching topology under quantized
communication, Neurocomputing, vol. 230, pp. 294–301, 2017.
[10] T. M. Cheng, Robust output feedback stabilization of nonlinear networked
systems via a finite data-rate communication channel, Journal of Systems Science
& Complexity, vol. 24, no. 1, pp. 1–13, 2011.
[11] D. Ciuonzo, A. Aubry, and V. Carotenuto, Rician MIMO channel- and jamming-
aware decision fusion, IEEE Transactions on Signal Processing, vol. 65, no. 15,
pp. 3866–3880, 2017.
[12] G. Como, F. Fagnani, and S. Zampieri, Anytime reliable transmission of real-
valued information through digital noisy channels, SIAM Journal on Control and
Optimization, vol. 48, no. 6, pp. 3903–3924, 2010.
[13] P. Correia, P. A. Assuncao, and V. Silva, Multiple description of coded video for
path diversity streaming adaptation, IEEE Transactions on Multimedia, vol. 14,
no. 3, pp. 923–935, 2012.
[14] Y. Cui, L. Yu, Y. Liu, W. Zhang, and F.-E. Alsaadi, Dynamic event-based non-
fragile state estimation for complex networks via partial nodes information,
Journal of the Franklin Institute-Engineering and Applied Mathematics, vol. 358,
no. 18, pp. 10193–10212, 2021.
[15] C. E. Curry, P. Mirchandani, and C. F. Price, State estimation with coarsely
quantized, high data-rate measurements, IEEE Transactions on Aerospace and
Electronic Systems, vol. 11, no. 4, pp. 613–621, 1975.

213
214 Bibliography

[16] D. F. Delchamps, Stabilizing a linear-system with quantized state feedback,

IEEE Transactions on Automatic Control, vol. 35, no. 8, pp. 916–924, 1990.
[17] S. Dey, A. Chiuso, and L. Schenato, Event-triggered stabilization of linear
systems under bounded bit rates, IEEE Transactions on Automatic Control,
vol. 62, no. 6, pp. 3054–3061, 2017.
[18] D. Ding, Z. Wang, D. W. C. Ho, and G. Wei, Distributed recursive filtering for
stochastic systems under uniform quantizations and deception attacks through
sensor networks, Automatica, vol. 78, pp. 231–240, 2017.
[19] D. Ding, Z. Wang, B. Shen, and H. Dong, Event-triggered distributed H∞ state
estimation with packet dropouts through sensor networks, IET Control Theory
& Applications, vol. 9, no. 13, pp. 1948–1955, 2015.
[20] D. Ding, G. Wei, S. Zhang, Y. Liu, and F. E. Alsaadi, On scheduling of deception
attacks for discrete-time networked systems equipped with attack detectors,
Neurocomputing, vol. 219, pp. 99–106, 2017.
[21] H. Dong, X. Bu, N. Hou, Y. Liu, F. E. Alsaadi, and T. Hayat, Event-triggered
distributed state estimation for a class of time-varying systems over sensor
networks with redundant channels, Information Fusion, vol. 36, pp. 243–250,
2017.
[22] H. Dong, Z. Wang, J. Lam, and H. Gao, Distributed filtering in sensor networks
with randomly occurring saturations and successive packet dropouts,
International Journal of Robust and Nonlinear Control, vol. 24, pp. 1743–1759, 2014.
[23] A. A. Elgamal and T. M. Cover, Achievable rates for multiple descriptions, IEEE
Transactions on Information Theory, vol. 28, no. 6, pp. 851–857, 1982.
[24] J. Fang and Q.-Q. Liu, Quantized feedback control over packet dropout
communication channels, Information Technology and Control, vol. 41, no. 3,
pp. 229–238, 2012.
[25] J. Fang and H. Li, Distributed adaptive quantization for wireless sensor
networks: From delta modulation to maximum likelihood, IEEE Transactions
on Signal Processing, vol. 56, no. 10, pp. 5246–5257, 2008.
[26] A. Farhadi, Feedback channel in linear noiseless dynamic systems controlled
over the packet erasure network, International Journal of Control, vol. 88, no. 8,
pp. 1490–1503, 2015.
[27] A. Farhadi and N. U. Ahmed, Tracking nonlinear noisy dynamic systems over
noisy communication channels, IEEE Transactions on Communications, vol. 59,
no. 4, pp. 955–961, 2011.
[28] A. Farhadi and C. D. Charalambous, Robust coding for a class of sources:
Applications in control and reliable communication over limited capacity
channels, Systems & Control Letters, vol. 57, no. 12, pp. 1005–1012, 2008.
[29] A. Farhadi and C. D. Charalambous, Stability and reliable data reconstruction
of uncertain dynamic systems over finite capacity channels, Automatica, vol. 46,
no. 5, pp. 889–896, 2010.
[30] Y. W. Feng and G. Ge, Optimal coding-decoding for systems controlled via a
communication channel, International Journal of Systems Science, vol. 44, no. 12,
pp. 2190–2198, 2013.
[31] A. L. Fradkov, B. Andrievsky, and M. Ananyevskiy, State estimation and
synchronization of pendula systems over digital communication channels,
European Physical Journal-Special Topics, vol. 223, no. 4, pp. 773–793, 2014.
Bibliography 215

[32] A. L. Fradkov, B. Andrievsky, and R. J. Evans, Chaotic observer-based

synchronization under information constraints, Physical Review E, vol. 73, no. 6,
Art. No. 066209, 2006.
[33] A. L. Fradkov, B. Andrievsky, and R. J. Evans, Synchronization of passifiable
Lurie systems via limited-capacity communication channel, IEEE Transactions
on Circuits And Systems I-Regular Papers, vol. 56, no. 2, pp. 430–439, 2009.
[34] L. Gao, X. Liao, H. Li, and G. Chen, Event-triggered control for multi-agent
network with limited digital communication, Nonlinear Dynamics, vol. 82, no. 4,
pp. 1659–1669, 2015.
[35] X. Ge, Q.-L. Han, X.-M. Zhang, L. Ding, and F. Yang, Distributed event-
triggered estimation over sensor networks: A survey, IEEE Transactions on
Cybernetics, vol. 50, no. 3, pp. 1306–1320, 2020.
[36] H. Geng, H. Liu, L. Ma, and X. Yi, Multi-sensor filtering fusion meets
censored measurements under a constrained network environment: Advances,
challenges and prospects, International Journal of Systems Science, vol. 52, no. 16,
pp. 3410–3436, 2021.
[37] V. K. Goyal, Multiple description coding: Compression meets the network,
IEEE Transactions on Signal Processing Magazine, vol. 18, no. 5, pp. 74–93, 2001.
[38] S. Hattori, K. Kobayashi, H. Okada, and M. Katayama, On-off error control
coding scheme for minimizing tracking error in wireless feedback, IEEE
Transactions on Industrial Informatics, vol. 11, no. 6, pp. 1411–1421, 2015.
[39] F. Han, G. Wei, D. Ding, and Y. Song, Finite-horizon H∞ -consensus control for
multi-agent systems with random parameters: The local condition case, Journal
of the Franklin Institute-Engineering and Applied Mathematics, vol. 354, no. 14,
pp. 6078–6097, 2017.
[40] F. Han, G. Wei, D. Ding, and Y. Song, Local condition based consensus
filtering with stochastic nonlinearities and multiple missing measurements,
IEEE Transactions on Automatic Control, vol. 62, no. 9, pp. 4784–4790, 2017.
[41] L. He, D. Han, X. Wang, and L. Shi, State estimation over a lossy network using
linear temporal coding, in Proceedings of the 51st IEEE Conference on Decision and
Control, pp. 2002–2007, Maui, HI, USA, 2012.
[42] W. He, G. Chen, Q.-L. Han, and F. Qian, Network-based leader-following
consensus of nonlinear multi-agent systems via distributed impulsive control,
Information Sciences, vol. 380, pp. 145–158, 2017.
[43] X. He, Z. Wang, X. Wang, and D. H. Zhou, Networked strong tracking filtering
with multiple packet dropouts: Algorithms and applications, IEEE Transactions
on Industrial Electronics, vol. 61, no. 3, pp. 1454–1463, 2014.
[44] J. Hespanha, A. Ortega, and L. Vasudevan, Towards the control of linear
systems with minimum bit-rate, in Proceedings of the 15th International
Symposium on Mathematical Theory of Networks and Systems, pp. 1–15, Notre
Dame, IN, USA, 2002.
[45] M. Hernandez-Gonzalez, M. Basin, and O. Stepanov, Discrete-time state
estimation for stochastic polynomial systems over polynomial observations,
International Journal of General Systems, vol. 47, no. 5, pp. 512–528, 2018.
[46] N. Hou, H. Dong, Z. Wang, W. Ren, and F. E. Alsaadi, Non-fragile state
estimation for discrete Markovian jumping neural networks, Neurocomputing,
vol. 179, pp. 238–245, 2016.
216 Bibliography

[47] N. Hou, H. Dong, Z. Wang, W. Ren, and F. E. Alsaadi, H∞ state estimation for
discrete-time neural networks with distributed delays and randomly occurring
uncertainties through fading channels, Neural Networks, vol. 89, pp. 61–73, 2017.
[48] C. Hu and H. Jiang, Pinning synchronization for directed networks with node
balance via adaptive intermittent control, Nonlinear Dynamics, vol. 80, nos. 1–2,
pp. 295–307, 2015.
[49] J. Hu, Z. Wang, and H. Gao, Recursive filtering with random parameter
matrices, multiple fading measurements and correlated noises, Automatica,
vol. 49, no. 11, pp. 3440–3448, 2013.
[50] L. Hu, Z. Wang, Q.-L. Han, and X. Liu, State estimation under false data
injection attacks: Security analysis and system protection, Automatica, vol. 219,
pp. 176–183, 2018.
[51] L. Hu, Z. Wang, Q.-L. Han, and X. Liu, Event-based input and state estimation
for linear discrete time-varying systems, International Journal of Control, vol. 91,
no. 1, pp. 101–113, 2018.
[52] J. Hu, Z. Wang, S. Liu, and H. Gao, A variance-constrained approach to
recursive state estimation for time-varying complex networks with missing
measurements, Automatica, vol. 64, pp. 155–162, 2016.
[53] H. Ishii and B. A. Francis, Quadratic stabilization of sampled-data systems with
quantization, Automatica, vol. 39, no. 10, pp. 1793–1800, 2013.
[54] H. Ishii and S. Hara, A subband coding approach to control under limited data
rates and message losses, Automatica, vol. 44, no. 4, pp. 1141–1148, 2008.
[55] A. Isidori, L. Marconi, and C. de Persis, Remote tracking via encoded
information for nonlinear systems, Systems & Control Letters, vol. 55, no. 10,
pp. 809–818, 2006.
[56] Y. Isidori, K. Takaba, and D. E. Quevedo, Stability analysis of networked control
systems subject to packet-dropouts and finite-level quantization, Systems &
Control Letters, vol. 60, no. 5, pp. 325–332, 2011.
[57] X.-W. Jiang, B. Hu, Z.-H. Guan, X.-H. Zhang, and L. Yu, Best achievable tracking
performance for networked control systems with encoder-decoder, Information
Sciences, vol. 305, pp. 184–195, 2015.
[58] Z. Jiang and Y. Wang, Input-to-state stability for discrete-time nonlinear
systems, Automatica, vol. 37, no. 6, pp. 857–869, 2001.
[59] Z. Jin, V. Gupta, and R. M. Murray, State estimation over packet dropping
networks using multiple description coding, Automatica, vol. 42, no. 9,
pp. 1141–1152, 2006.
[60] D. Kapetanovic, S. Chatzinotas, and B. Ottersten, Index assignment for
multiple description repair in distributed storage systems, in Proceedings of the
2014 IEEE International Conference on Communications, pp. 3896–3901, Sydney,
NSW, Australia, Jun. 10–14, 2014.
[61] M. J. Khojasteh, P. Tallapragada, J. Cortes, and M. Franceschetti, The value of
timing information in event-triggered control: The scalar case, in Proceedings of
the 54th Annual Allerton Conference on Communication, Control, and Computing,
pp. 1165–1172, Monticello, IL, USA, Sept. 27–30, 2016.
[62] M. J. Khojasteh, P. Tallapragada, J. Cortes, and M. Franceschetti, Time-
triggering versus event-triggering control over communication channels, in
Proceedings of the 2017 IEEE 56th Annual Conference on Decision and Control,
Melbourne, VIC, Australia, 2017.
Bibliography 217

[63] E. Kofman and J. H. Braslavsky, Level crossing sampling in feedback

stabilization under data-rate constraints, in Proceedings of the 45th IEEE
Conference on Decision and Control, pp. 4423–4428, San Diego, CA, USA, 2006.
[64] D. Lehmann and J. Lunze, Event-based control using quantized state
information, in Proceedings of the 2nd IFAC Workshop on Distributed Estimation
and Control in Networked Systems, pp. 1989–1994, Annecy, France, 2010.
[65] A. S. Leong, D. Subhrakanti, and G. N. Nair, Quantized filtering schemes for
multi-sensor linear state estimation: Stability and performance under high rate
quantization, IEEE Transactions on Signal Processing, vol. 61, no. 15, pp. 3852–
3865, 2013.
[66] H. Li, G. Chen, T. Huang, Z. Dong, and L. Gao, Event-triggered distributed
average consensus over directed digital networks with limited communication
bandwidth, IEEE Transactions on Cybernetics, vol. 46, no. 12, pp. 3098–3110, 2016.
[67] J. Li, H. Dong, F. Han, N. Hou, and X. Li, Filter design, fault estimation and
reliable control for networked time-varying systems: A survey, Systems Science
& Control Engineering: An Open Access Journal, vol. 5, pp. 331–341, 2017.
[68] J.-Y. Li, Z. Wang, R. Lu, and Y. Xu, Partial-nodes-based state estimation for
complex networks with constrained bit rate, IEEE Transactions on Network
Science and Engineering, vol. 8, pp. 2, pp. 1887–1899, 2021.
[69] T. Li, M. Fu, L. Xie, and J. Zhang, Distributed consensus with limited
communication data rate, IEEE Transactions on Automatic Control, vol. 56, no. 2,
pp. 279–292, 2011.
[70] H. Li, C. Huang, G. Chen, X. Liao, and T. Huang, Distributed consensus
optimization in multiagent networks with time-varying directed topologies
and quantized communication, IEEE Transactions on Cybernetics, vol. 47, no. 8,
pp. 2044–2057, 2017.
[71] L. Li, M. Lemmon, and X. Wang, Stabilizing bit-rates in quantized event
triggered control systems, in Proceedings of The 15th ACM International
Conference on Hybrid Systems: Computation and Control, pp. 245–254, Beijing,
P. R. China, 2012.
[72] H. Li, S. Liu, Y. C. Soh, and L. Xie, Event-triggered communication and
data rate constraint for distributed optimization of multiagent systems, IEEE
Transactions on Systems, Man, and Cybernetics: Systems, vol. 48, no. 11, pp. 1908–
1919, 2018.
[73] L. Li, X. Wang, and M. D. Lemmon, Efficiently attentive event-triggered
systems with limited bandwidth, IEEE Transactions on Automatic Control,
vol. 62, no. 3, pp. 1491–1497, 2017.
[74] Q. Li, B. Shen, Z. Wang, and F. E. Alsaadi, A sampled-data approach to
distributed H∞ resilient state estimation for a class of nonlinear time-delay
systems over sensor networks, Journal of the Franklin Institute, vol. 354, pp. 7139–
7157, 2017.
[75] W. Li, Y. Jia, and J. Du, Recursive state estimation for complex networks with
random coupling strength, Neurocomputing, vol. 219, pp. 1–8, 2017.
[76] W. Li, G. Wei, D. Ding, Y. Liu, and F. E. Alsaadi, A new look at boundedness
of error covariance of Kalman filtering, IEEE Transactions on Systems, Man, and
Cybernetics: Systems, vol. 48, no. 2, pp. 309–314, 2018.
[77] W. Li, G. Wei, F. Han, and Y. Liu, Weighted average consensus-based unscented
Kalman filtering, IEEE Transactions on Cybernetics, vol. 46, no. 2, pp. 558–567,
2016.
218 Bibliography

[78] T. Li and L. Xie, Distributed consensus over digital networks with limited
bandwidth and time-varying topologies, Automatica, vol. 47, no. 9, pp. 2006–
2015, 2011.
[79] T. Li and L. Xie, Distributed coordination of multi-agent systems with
quantized-observer based encoding-decoding, IEEE Transactions on Automatic
Control, vol. 57, no. 12, pp. 3023–3037, 2012.
[80] X. Li, C. Li, and M. Z. Q. Chen, Stabilisation of non-linear DISS systems with
uncertainty via encoded feedback, IET Control Theory & Applications, vol. 11,
no. 5, pp. 732–739, 2017.
[81] X. Li, Y. Yu, and L. Guo, Stability of time-delay systems with impulsive control
involving stabilizing delays, Automatica, vol. 124, Art. No. 109336, 2021.
[82] X. Li, D. Peng, and J. Cao, Lyapunov stability for impulsive systems via event-
triggered impulsive control, IEEE Transactions on Automatic Control, vol. 65,
no. 11, pp. 4908–4913, 2020.
[83] X. Li, T. Zhang, and J. Wang, Input-to-state stability of impulsive systems via
event-triggered impulsive control, IEEE Transactions on Cybernetics, vol. 52,
no. 7, pp. 7187–7195, 2022.
[84] Z. Li, Z. Duan, and G. Chen, Consensus of discrete-time linear multi-
agent systems with observer-type protocols, Discrete and Continuous Dynamical
Systems-Series B, vol. 16, no. 2, pp. 489–505, 2011.
[85] D. Liberzon, On stabilization of linear systems with limited information, IEEE
Transactions on Automatic Control, vol. 48, no. 2, pp. 304–307, 2003.
[86] D. Liberzon, Hybrid feedback stabilization of systems with quantized signals,
Automatica, vol. 39, no. 3, pp. 1543–1554, 2003.
[87] D. Liberzon and J. P. Hespanha, Stabilization of nonlinear systems with limited
information feedback, IEEE Transactions on Automatic Control, vol. 50, no. 6,
pp. 910–915, 2005.
[88] N. Lin and Q. Ling, Bit-rate conditions for the consensus of quantized
multiagent systems based on event triggering, IEEE Transactions on Cybernetics,
vol. 52, no. 1, pp. 116–127, 2022.
[89] Q. Ling, Bit rate conditions to stabilize a continuous-time scalar linear system
based on event triggering, IEEE Transactions on Automatic Control, vol. 62, no. 8,
pp. 4093–4100, 2017.
[90] Q. Ling, Bit-rate conditions to stabilize a continuous-time linear system with
feedback dropouts, IEEE Transactions on Automatic Control, vol. 63, no. 7,
pp. 2176–2183, 2018.
[91] G. Liu, H. Liu, H. Chen, C. Zhou, and L. Shu, Position-based adaptive
quantization for target location estimation in wireless sensor networks using
one-bit data, Wireless Communications & Mobile Computing, vol. 16, no. 8,
pp. 929–941, 2016.
[92] Q. Liu and J. Fan, State estimation for networked control systems using fixed
data rates, International Journal of Systems Science, vol. 48, no. 9, pp. 1818–1828,
2017.
[93] S. Liu, T. Li, and L. Xie, Distributed consensus for multiagent systems with
communication delays and limited data rate, SIAM Journal on Control and
Optimization, vol. 49, no. 6, pp. 2239–2262, 2011.
[94] S. Liu, T. Li, L. Xie, M. Fu, and J. Zhang, Continuous-time and sampled-
data-based average consensus with logarithmic quantizers, Automatica, vol. 49,
no. 11, pp. 3329–3336, Nov. 2013.
Bibliography 219

[95] S. Liu, Y. Song, D. Ding, and Y. Liu, Event-triggered dynamic output feedback
RMPC for polytopic systems with redundant channels: Input-to-state stability,
Journal of the Franklin Institute-Engineering and Applied Mathematics, vol. 354,
no. 7, pp. 2871–2892, 2017.
[96] S. Liu, Y. Song, G. Wei, and X. Huang, RMPC-based security problem
for polytopic uncertain system subject to deception attacks and persistent
disturbances, IET Control Theory & Applications, vol. 11, no. 10, pp. 1611–1618,
2017.
[97] S. Liu, Z. Wang, Y. Chen, and G. Wei, Protocol-based unscented Kalman
filtering in the presence of stochastic uncertainties, IEEE Transactions on
Automatic Control, vol. 65, no. 3, pp. 1303–1309, 2020.
[98] S. Liu, Z. Wang, B. Shen, and G. Wei, Partial-neurons-based state estimation
for delayed neural networks with state-dependent noises under redundant
channels, Information Sciences, vol. 547, pp. 931–944, 2021.
[99] S. Liu, Z. Wang, G. Wei, and M. Li, Distributed set-membership filtering for
multirate systems under the Round-Robin scheduling over sensor networks,
IEEE Transactions on Cybernetics, vol. 50, no. 5, pp. 1910–1920, 2020.
[100] S. Liu, G. Wei, Y. Song, and Y. Liu, Extended Kalman filtering for stochastic
nonlinear systems with randomly occurring cyber attacks, Neurocomputing,
vol. 207, pp. 708–716, 2016.
[101] W. Liu, Z. Wang, and N. Mingkang, Controlled synchronization for chaotic
systems via limited information with data packet dropout, Automatica, vol. 49,
no. 11, pp. 3329–3336, 2013.
[102] H. Liu, Z. Wang, B. Shen, and X. Liu, Event-triggered H∞ state estimation for
delayed stochastic memristive neural networks with missing measurements:
The discrete time case, IEEE Transactions on Neural Networks and Learning
Systems, vol. 29, no. 8, pp. 3726–3737, 2018.
[103] H. Liu, Z. Wang, B. Shen, and F. E. Alsaadi, H∞ state estimation for
discrete-time memristive recurrent neural networks with stochastic time-
delays, International Journal of General Systems, vol. 45, no. 5, pp. 633–647, 2016.
[104] M. B. Loiola, R. R. Lopes, and J. M. T. Romano, Modified Kalman filters for
channel estimation in orthogonal space-time coded systems, IEEE Transactions
on Automatic Control, vol. 60, no. 1, pp. 533–538, 2012.
[105] J. Lu, D. W. C. Ho, and Z. Wang, Pinning stabilization of linearly coupled
stochastic neural networks via minimum number of controllers, IEEE
Transactions on Neural Networks, vol. 20, no. 10, pp. 1617–1629, 2009.
[106] Y. Luo, Z. Wang, G. Wei, and F. E. Alsaadi, Robust H∞ filtering for a class
of two-dimensional uncertain fuzzy systems with randomly occurring mixed
delays, IEEE Transactions on Fuzzy Systems, vol. 25, no. 1, pp. 70–83, 2017.
[107] Y. Luo, Z. Wang, G. Wei, F. E. Alsaadi, and T. Hayat, State estimation for a class
of artificial neural networks with stochastically corrupted measurements under
Round-Robin protocol, Neural Networks, vol. 77, pp. 70–79, 2016.
[108] L. Ma, Z. Wang, Q.-L. Han, and H. K. Lam, Variance-constrained distributed
filtering for time-varying systems with multiplicative noises and deception
attacks over sensor networks, IEEE Sensors Journal, vol. 17, no. 7, pp. 2279–2288,
2017.
[109] L. Ma, Z. Wang, and H. K. Lam, Event-triggered mean-square consensus
control for time-varying stochastic multi-agent system with sensor saturations,
IEEE Transactions on Automatic Control, vol. 62, no. 7, pp. 3524–3531, 2017.
220 Bibliography

[110] V. Malyavej and A. V. Savkin, Set-valued robust Kalman state estimation via
digital communication channels with bit-rate constraints, in Proceedings of the
American Control Conference, Denver, CO, USA, pp. 1896–1901, 2003.
[111] A. S. Matveev, State estimation via limited capacity noisy communication
channels, Mathematics of Control Signals and Systems, vol. 20, no. 1, pp. 1–35,
2008.
[112] A. S. Matveev and A. V. Savkin, Shannon zero error capacity in the problems
of state estimation and stabilization via noisy communication channels,
International Journal of Control, vol. 80, no. 2, pp. 241–255, 2007.
[113] S. Movaghati and M. Ardakani, Optimum bit-sensor assignment for dis-
tributed estimation in inhomogeneous sensor networks, IEEE Communications
Letters, vol. 18, no. 4, pp. 668–671, 2014.
[114] G. N. Nair and R. J. Evans, Stabilization with data-rate-limited feedback:
Tightest attainable bounds, Systems & Control Letters, vol. 41, no. 1, pp. 49–56,
2000.
[115] G. N. Nair and R. J. Evans, Exponential stabilisability of finite-dimensional
linear systems with limited data rates, Automatica, vol. 39, no. 4, pp. 585–593,
2003.
[116] G. N. Nair and R. J. Evans, Stabilizability of stochastic linear systems with finite
feedback data rates, SIAM Journal on Control and Optimization vol. 43, no. 2,
pp. 413–436, 2004.
[117] G. N. Nair, A nonstochastic information theory for communication and state
estimation, IEEE Transactions on Automatic Control, vol. 58, no. 6, pp. 1479–1510,
2013.
[118] G. Nagamani, B. Adhira, and G. Soundararajan, Non-fragile extended
dissipative state estimation for delayed discrete-time neural networks:
Application to quadruple tank process model, Nonlinear Dynamics, vol. 104,
no. 1, pp. 451–466, 2023.
[119] D. Nešić and D. S. Laila, A note on input-to-state stabilization for nonlinear
sampled-data systems, IEEE Transactions on Automatic Control, vol. 47, no. 7,
pp. 1153–1158, 2002.
[120] J. Ostergaard, D. E. Quevedo, and A. Ahlen, Predictive power control and
multiple-description coding for wireless sensor networks, in Proceedings of the
IEEE International Conference on Acoustics, Speech, and Signal Processing, Taipei,
Taiwan, Apr. 19–24, 2009.
[121] J. Ostergaard and D. E. Quevedo, Multiple descriptions for packetized
predictive control, Eurasip Journal on Advances in Signal Processing, vol. 2016,
Art. No. 45, 2016.
[122] C. de Persis and A. Isidori, Stabilizability by state feedback implies
stabilizability by encoded state feedback, Systems & Control Letters, vol. 53,
nos. 3–4, pp. 249–258, 2004.
[123] C. de Persis, n-Bit stabilization of n-dimensional nonlinear systems in
feedforward form, IEEE Transactions on Automatic Control, vol. 50, no. 3, pp. 299–
311, 2005.
[124] C. de Persis, Nonlinear stabilizability via encoded feedback: The case of
integral ISS systems, Automatica, vol. 42, no. 10, pp. 1813–1816, 2006.
[125] C. de Persis and D. Nešic, ´ Practical encoders for controlling nonlinear systems
under communication constraints, Systems & Control Letters, vol. 57, no. 8,
pp. 654–662, 2008.
Bibliography 221

[126] C. de Persis, Minimal data rate stabilization of nonlinear systems over

networks with large delays, International Journal of Robust and Nonlinear Control,
vol. 20, no. 10, pp. 1097–1111, 2010.
[127] Z. Qiu, L. Xie, and Y. Hong, Quantized leaderless and leader-following
consensus of high-order multi-agent systems with limited data rate, IEEE
Transactions on Automatic Control, vol. 61, no. 9, pp. 2432–2447, 2016.
[128] F. Qu, E. Tian, and X. Zhao, Chance-constrained H∞ state estimation for
recursive neural networks under deception attacks and energy constraints: The
finite-horizon case, IEEE Transactions on Neural Networks and Learning Systems,
vol. 34, no. 9, pp. 6492–6503. DOI: 10.1109/TNNLS.2021.3137426.
[129] D. E. Quevedo, A. Ahlen, and J. Ostergaard, Energy efficient state estimation
with wireless sensors through the use of predictive power control and coding,
IEEE Transactions on Signal Processing, vol. 58, no. 9, pp. 4811–4823, 2010.
[130] D. E. Quevedo, J. Ostergaard, and A. Ahlen, Power control and coding
formulation for state estimation with wireless sensors, IEEE Transactions on
Control Systems Technology, vol. 22, no. 2, pp. 413–427, 2014.
[131] M. Rabi, C. Ramesh, and K. H. Johansson, Separated design of encoder and
controller for networked linear quadratic optimal control, SIAM Journal on
Control and Optimization, vol. 54, no. 2, pp. 662–689, 2016.
[132] M. M. Rana, L. Li, and S. Su, Distributed state estimation using RSC coded
smart grid communications, IEEE Access, vol. 3, pp. 1340–1349, 2015.
[133] M. A. Rana, F. Ndiaye, S. A. Chowdhury, and N. Mansoor, Multiple description
image transmission for diversity systems over unreliable communication
networks, in Proceedings of 10th International Conference on Computer and
Information Technology, pp. 240–244, Dhanmondi, Bangladesh, Dec. 27–29, 2007.
[134] T. S. Rappaport, Wireless Communications: Principles and Practice, Prentice Hall,
2001.
[135] W. Ren and R. W. Beard, Consensus seeking in multiagent systems under
dynamically changing interaction topologies, IEEE Transactions on Automatic
Control, vol. 50, no. 5, pp. 655–661, 2005.
[136] A. Ribeiro and B. Giannakis, Bandwidth-constrained distributed estimation
for wireless sensor networks-Part I: Gaussian case, IEEE Transactions on Signal
Processing, vol. 54, no. 3, pp. 1131–1143, 2006.
[137] A. Ribeiro, B. Giannakis, and S. I. Roumeliotis, SOI-KF: Distributed Kalman
filtering with low-cost communications using the sign of innovations, IEEE
Transactions on Signal Processing, vol. 54, no. 12, pp. 4782–4795, 2006.
[138] A. J. Rojas, J. S. Freudenberg, J. H. Braslavsky, and R. H. Middleton, Optimal
signal to noise ratio in feedback over communication channels with memory,
in Proceedings of the 45th IEEE Conference on Decision & Control, pp. 240–244,
San Diego, CA, USA, Dec. 13–15, 2006
[139] A. V. Savkin, Analysis and synthesis of networked control systems: Topological
entropy, observability, robustness and optimal control, Automatica, vol. 42,
no. 1, pp. 51–62, 2006.
[140] A. V. Savkin and T. M. Cheng, Detectability and output feedback stabilizability
of nonlinear networked control systems, IEEE Transactions on Automatic Control,
vol. 52, no. 4, pp. 730–735, 2007.
[141] A. V. Savkin and I. R. Petersen, Set-valued state estimation via a limited
capacity communication channel, IEEE Transactions on Automatic Control,
vol. 48, no. 4, pp. 676–680, 2003.
222 Bibliography

[142] C. E. Shannon, A mathematical theory of communication, Bell System Technical

Journal, vol. 27, no. 3, pp. 379–423, 1948.
[143] C. E. Shannon, A mathematical theory of communication, Bell System Technical
Journal, vol. 27, no. 4, pp. 623–666, 1948.
[144] A. Shirazinia, A. A. Zaidi, L. Bao, and M. Skoglund, Dynamic source-channel
coding for estimation and control over binary symmetric channels, IET Control
Theory & Applications, vol. 9, no. 9, pp. 1444–1454, 2015.
[145] B. Shen, Z. Wang, D. Ding, and H. Shu, H∞ state estimation for complex
networks with uncertain inner coupling and incomplete measurements, IEEE
Transactions on Neural Networks and Learning Systems, vol. 24, no. 12, pp. 2027–
2037, 2013.
[146] B. Shen, Z. Wang, and X. Liu, Sampled-data synchronization control of
dynamical networks with stochastic sampling, IEEE Transactions on Automatic
Control, vol. 57, no. 10, pp. 2644–2650, 2012.
[147] B. Shen, Z. Wang, and H. Qiao, Event-triggered state estimation for discrete-
time multidelayed neural networks with stochastic parameters and incomplete
measurements, IEEE Transactions on Neural Networks and Learning Systems,
vol. 28, no. 5, pp. 1152–1163, 2017.
[148] B. Shen, Z. Wang, D. Wang, and Q. Li, State-saturated recursive filter design for
stochastic time-varying nonlinear complex networks under deception attacks,
IEEE Transactions on Neural Networks and Learning Systems, vol. 31, no. 10,
pp. 3788–3800, 2020.
[149] B. Shen, Z. Wang, D. Wang, and H. Liu, Distributed state-saturated recursive
filtering over sensor networks under Round-Robin protocol, IEEE Transactions
on Cybernetics, vol. 50, no. 8, pp. 3605–3615, 2020.
[150] T. Sui, K. You, and M. Fu, Kalman filtering with intermittent observations
using measurements coding, in 10th IEEE International Conference on Control and
Automation (ICCA’ 2013), pp. 1127–1132, 2013.
[151] T. Sui, K. You, M. Fu, and D. Marelli, Stability of MMSE state estimators over
lossy networks using linear coding, Automatica, vol. 51, pp. 167–174, 2015.
[152] J. Suo, N. Li, and Q. Li, Event-triggered H∞ state estimation for discrete-
time delayed switched stochastic neural networks with persistent dwell-
time switching regularities and sensor saturations, Neurocomputing, vol. 455,
pp. 297–307, 2021.
[153] P. Tallapragada and J. Cortes, Event-triggered stabilization of linear systems
under bounded bit rates, IEEE Transactions on Automatic Control, vol. 61, no. 6,
pp. 1575–1589, 2016.
[154] Y. Tang, H. Gao, and J. Kurths, Robust H∞ self-triggered control of networked
systems under packet dropouts, IEEE Transactions on Cybernetics, vol. 46, no. 12,
pp. 3294–3305, 2016.
[155] S. Tatikonda and S. Mitter, Control-under communication constraints, IEEE
Transactions on Automatic Control, vol. 49, no. 7, pp. 1056–1068, 2004.
[156] V. A. Vaishampayan, Design of multiple description scalar quantizers, IEEE
Transactions on Information Theory, vol. 39, no. 3, pp. 8214–834, 1993.
[157] V. A. Vaishampayan and J. C. Batllo, Asymptotic analysis of multiple
description quantizers, IEEE Transactions on Information Theory, vol. 44, no. 1,
pp. 278–284, 1998.
Bibliography 223

[158] J. Wang and H. Li, Stabilization of a continuous linear system over channel
with network-induced delay and communication constraints, European Journal
of Control, vol. 31, pp. 72–78, 2016.
[159] F. Wang, J. Liang, Z. Wang, and X. Liu, A variance-constrained approach to
recursive filtering for nonlinear two-dimensional systems with measurement
degradations, IEEE Transactions on Cybernetics, vol. 48, no. 6, pp. 1877–1887,
2018.
[160] L. Wang, Z. Wang, Q.-L. Han, and G. Wei, Event-based variance-constrained
H∞ filtering for stochastic parameter systems over sensor networks with
successive missing measurements, IEEE Transactions on Cybernetics, vol. 48,
no. 3, pp. 1007–1017, 2018.
[161] L. Wang, Z. Wang, Q.-L. Han, and G. Wei, Synchronization control for a class of
discrete-time dynamical networks with packet dropouts: A coding-decoding-
based approach, IEEE Transactions on Cybernetics, vol. 48, no. 8, pp. 2437–2448,
2018.
[162] L. Wang, Z. Wang, T. Huang, and G. Wei, An event-triggered approach to
state estimation for a class of complex networks with mixed time delays and
nonlinearities, IEEE Transactions on Cybernetics, vol. 46, no. 11, pp. 2497–2508,
2016.
[163] L. Wang, Z. Wang, B. Shen, and G. Wei, Recursive filtering with measurement
fading: A multiple description coding scheme, IEEE Transactions on Automatic
Control, vol. 66, no. 11, pp. 5144–5159, 2021.
[164] L. Wang, Z. Wang, G. Wei, and F. E. Alsaadi, Finite-time state estimation for
recurrent delayed neural networks with component-based event-triggering
protocol, IEEE Transactions on Neural Networks and Learning Systems, vol. 29,
no. 4, pp. 1046–1057, 2018.
[165] J. Wang and Z. Yan, Coding scheme based on spherical polar coordinate for
control over packet erasure channel, International Journal of Robust and Nonlinear
Control, vol. 51, pp. 1159–1176, 2014.
[166] X. Wang, E. E. Yaz, and J. Long, Robust and resilient state-dependent control
of discrete-time nonlinear systems with general performance criteria, Systems
Science & Control Engineering: An Open Access Journal, vol. 2, pp. 48–54, 2014.
[167] G. Wei, S. Liu, Y. Song, and Y. Liu, Probability-guaranteed set-membership
filtering for systems with incomplete measurements, Automatica, vol. 60,
pp. 12–16, 2015.
[168] G. Wei, S. Liu, L. Wang, and Y. Wang, Event-based distributed set-membership
filtering for a class of time-varying non-linear systems over sensor networks
with saturation effects, International Journal of General Systems, vol. 45, no. 5,
pp. 532–547, 2016.
[169] N. Wiener, Cybernetics: Or Control and Communication in the Animal and the
Machine, Hermann & Cie and MIT Press, 1948.
[170] T. Wen, L. Zou, J. Liang, and C. Roberts, Recursive filtering for communication-
based train control systems with packet dropouts, Neurocomputing, vol. 275,
pp. 948–957, 2018.
[171] W. S. Wong and R. W. Brockett, Systems with finite communication
bandwidth constraints-II: Stabilization with limited information feedback,
IEEE Transactions on Automatic Control, vol. 44, no. 5, pp. 1049–1053, 1999.
224 Bibliography

[172] J. Wu, H. Li, and X. Chen, Leader-following consensus of nonlinear discrete-

time multi-agent systems with limited communication channel capacity,
Journal of the Franklin Institute-Engineering and Applied Mathematics, vol. 354,
no. 10, pp. 4179–4195, 2017.
[173] X. Xiao, L. Zhou, and G. Lu, Detection of singular systems via a limited
communication channel with missing measurements capacity, Information
Sciences, vol. 228, pp. 192–202, 2013.
[174] K. Xiong, C. Wei, and L. Liu, Robust extended Kalman filtering for nonlinear
systems with stochastic uncertainties, IEEE Transactions on Systems Man and
Cybernetics Part A-Systems and Humans, vol. 40, no. 2, pp. 399–405, 2010.
[175] T. Xu, S. Liu, and G. Wei, Event-based finite-time H∞ synchronization control
for switched complex networks with average dwell time, International Journal
of Robust and Nonlinear Control, vol. 36, no. 4, pp. 3923–3943, 2022.
[176] Y. Xu, C. Zhu, and L. Yu, Multipath routing of multiple description coded
images in wireless networks, Journal of Computer Science and Technology, vol. 29,
no. 4, pp. 576–588, 2014.
[177] K. You, Control over communication networks: A personal perspective, in
Proceedings of the 11th World Congress on Intelligent Control and Automation
(WCICA’2014), pp. 2002–2007, Shenyang, P. R. China, 2012.
[178] K. You, Recursive algorithms for parameter estimation with adaptive
quantizer, Automatica, vol. 52, pp. 192–201, 2015.
[179] K. You and L. Xie, Network topology and communication data rate for
consensusability of discrete-time multi-agent systems, IEEE Transactions on
Automatic Control, vol. 56, no. 10, pp. 2262–2275, 2011.
[180] M. Yu, C. Yan, and D. Xie, Event-triggered control for couple-group multi-agent
systems with logarithmic quantizers and communication delays, Asian Journal
of Control, vol. 19, no. 2, pp. 681–691, 2017.
[181] W. Yu, G. Chen, and M. Cao, Some necessary and sufficient conditions for
second-order consensus in multi-agent dynamical systems, Automatica, vol. 46,
no. 6, pp. 1089–1095, 2010.
[182] W. Yu, W. X. Zheng, G. Chen, W. Ren, and J. Cao, Second-order consensus
in multi-agent dynamical systems with sampled position data, Automatica,
vol. 47, no. 7, pp. 1496–1503, 2011.
[183] Y. Yuan, P. Zhang, L. Guo, and H. Yang, Towards quantifying the impact of
randomly occurred attacks on a class of networked control systems, Journal
of the Franklin Institute-Engineering and Applied Mathematics, vol. 354, no. 12,
pp. 4966–4988, 2017.
[184] S. Zhang, D. Ding, G. Wei, J. Mao, Y. Liu, and F. E. Alsaadi, Design and analysis
of H∞ filter for a class of T-S fuzzy system with redundant channels and
multiplicative noises, Neurocomputing, vol. 260, pp. 257–264, 2017.
[185] S. Zhang, D. Ding, G. Wei, Y. Liu, and F. E. Alsaadi, H∞ state estimation for
artificial neural networks over redundant channels, Neurocomputing, vol. 226,
pp. 117–125, 2017.
[186] W. Zhang, S. Vedantam, and U. Mitra, Joint transmission and state estimation:
A constrained channel coding approach, IEEE Transactions on Information
Theory, vol. 57, no. 10, pp. 7084–7095, 2011.
[187] X.-M. Zhang, Q.-L. Han, X. Ge, D. Ding, L. Ding, D. Yue, and C. Peng,
Networked control systems: A survey of trends and techniques, IEEE/CAA
Journal of Automatica Sinica, vol. 7, no. 1, pp. 1–17, 2020.
Bibliography 225

[188] G. Zherlitsyn and A. S. Matveev, Min-max optimal data encoding and fusion
in sensor networks, Automatica, vol. 46, no. 9, pp. 1546–1552, 2010.
[189] D. Zhou, X. He, Z. Wang, G. P. Liu, and Y. Ji, Leakage fault diagnosis for an
internet-based three-tank system: An experimental study, IEEE Transactions on
Control Systems Technology, vol. 20, no. 4, pp. 857–870, 2012.
[190] L. Zhou and G. Lu, Detection and stabilization for discrete-time descriptor
systems via a limited capacity communication channel, Automatica, vol. 45,
no. 10, pp. 2272–2277, 2009.
[191] L. Zou, Z. Wang, and H. Gao, Set-membership filtering for time-varying
systems with mixed time-delays under Round-Robin and Weighted Try-Once-
Discard protocols, Automatica, vol. 74, pp. 341–348, 2016.
[192] L. Zou, Z. Wang, J. Hu, Y. Liu, and X. Liu, Communication-protocol-
based analysis and synthesis of networked systems: Progress, prospects and
challenges, International Journal of Systems Science, vol. 52, no. 14, pp. 3013–3034,
2021.
[193] L. Zou, Z. Wang, and D. H. Zhou, Event-based control and filtering of
networked systems: A survey, International Journal of Automation and Computing,
vol. 14, no. 3, pp. 239–253, 2017.
Index

C N
channel fading, 27, 110–112, 121–122, network-induced phenomena, 9, 21,
134–135 31, 141
consensus control, 21–23, 26, 88, 91,
107, 109
O
D one-bit symbolic data, 184
dynamical multiple description coding,
27, 141, 151, 164 P
dynamical networks, 67–68, 70, 72, 85, 88
partial-neurons-based state estimation,
46, 64
E
encoding-decoding mechanism, 4, 6, 12,
R
35, 40, 42, 49, 64, 185
event-based encoding-decoding scheme, recursive filtering, 110, 122
174, 176, 186, 196–197, 204

S
G
synchronization control, 26, 28, 65–66,
gain-scheduled state estimation, 25, 68, 70, 80–81, 83–84, 86–87
29–30, 35, 42, 45

M Z

multi-agent systems, 2, 21, 26, 88 Zeno behavior, 27, 171, 173, 184,
multiple description coding, 11, 26–27, 196
110–111, 113, 116, 120, 140–142,
151, 164

226