Dissertations
No. 623
Power Control in
Cellular Radio Systems:
Analysis, Design and Estimation
Fredrik Gunnarsson
Department of Electrical Engineering
Link opings universitet, SE581 83 Link oping, Sweden
Link oping 2000
Power Control in Cellular Radio Systems:
Analysis, Design and Estimation
c ( 2000 Fredrik Gunnarsson
fred@isy.liu.se
http://www.control.isy.liu.se
Division of Automatic Control
Department of Electrical Engineering,
Linkopings universitet,
SE581 83 Linkoping,
Sweden
ISBN 9172196890 ISSN 03457524
Printed by Linus & Linnea AB, Linkoping, Sweden 2000
To Soa
Abstract
The primary goal of cellular radio systems is to provide communications services
to a large number of mobile users. Due to the dramatic increase in number of
users and their demand for more advanced services, the available resources have
to be utilized eciently. Closedloop power control is considered as an important
component in this resource management.
For practical reasons, the powers have to be computed locally for each connec
tion, though performance and stability depend on how the dierent connections
interact. We consider the power control problem as a decentralized control system,
consisting of interconnected local control loops. Methods from control theory are
used to analyze existing algorithms locally and to design controllers with improved
performance. Thereby, performance degradation due to time delays and nonlin
earities, can be handled by careful controller design. On a global level, we provide
results on stability and convergence of the designed controllers. The results are il
lustrated by simulations using both small and largescale simulation environments.
i
ii
Acknowledgments
I am very grateful to all the people (yes, there are many) who have inspired and
supported me during the work on the thesis.
First of all I would like to thank my supervisor, Professor Fredrik Gustafsson, for
guidance through the evolution of the thesis, but even more so for all rewarding and
inspiring discussions. I especially appreciate his positive attitude and willingness
to share thoughts and ideas. Thanks goes also to Dr. Jonas Blom for times of
confusion and of clarication. During the time we shared room and project, we had
many fruitful discussions related to communications and control and to abridging
the two.
I would like to thank Professor Lennart Ljung for giving me the opportunity to
join the Automatic Control group, and for creating a stimulating and professional
atmosphere. I also value the freedom I have had in choosing research topic.
During my graduate studies I have had the opportunity to share ideas and dis
cussions with people at Ericsson Radio Systems. I am especially grateful to Magnus
Almgren for all energetic discussions related to practical and theoretical aspects
of power control and mobile communications. Thank you for sharing simulation
knowledge, but also for making me realize that simulations should be approached
with care. I also value discussions with Dr. Magnus Frodigh, Bengt Bergkvist and
Dr. Niclas Wiberg. In dierent phases of my project, I have had many questions
answered by Dr. Gunnar Bark, Tomas Rimhagen, Dr. Ke Wang Helmersson, Hakan
Olofsson, Magnus Persson and Maria Gustafsson, and several others.
The people in the Automatic Control group all have contributed to a great
atmosphere at work. Besides research activities and coursework, I appreciate the
social activities and spontaneous coee room discussions about the most shifting
things. Thanks also goes to our secretary Ulla for all administrative help and
cheerful attitude, and to Mattias Olofsson for maintaining the computers.
Fredrik Gustafsson and Jonas Blom proofread the entire manuscript and pro
vided me with many valuable comments and suggestions on improvements. Dr.
Niclas Bergman read parts and patiently answered many questions regarding L
A
T
E
X.
Mans
Ostring gave me insightful comments on the design. All of the above added
signicant improvements to the thesis.
This work was supported by the graduate school ECSEL and the Swedish Na
tional Board for Industrial and Technical Development (NUTEK), and they are
gratefully acknowledged.
To my parents and brothers: thanks for always being there and for being you.
Finally, I would like to thank Soa for abundant love and support. Your en
couragement and enormous patience, especially during the completion of this thesis,
made hardworking days lighter and brighter. I love you.
Linkoping, March 2000
Fredrik Gunnarsson
iii
iv
Contents
1 Introduction 1
1.1 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Reading Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Extended Summary 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Power Control Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Local Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Local Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Global Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Estimation and Outer Loop Control . . . . . . . . . . . . . . . . . . 17
2.7 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Cellular Radio Systems 21
3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Radio Communication . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Radio Wave Propagation . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.1 Path Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3.2 Shadow Fading . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3.3 Multipath Fading . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.4 Example: Spatially Correlated Propagation . . . . . . . . . . 27
3.4 MultiUser Communications . . . . . . . . . . . . . . . . . . . . . . . 27
3.4.1 Orthogonal Signals . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.2 NonOrthogonal Signals . . . . . . . . . . . . . . . . . . . . . 29
3.5 Cellular Radio Systems . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.6 Diversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 Radio Resource Management 37
4.1 Performance of Cellular Radio Systems . . . . . . . . . . . . . . . . . 38
4.2 Resource Allocation Algorithms . . . . . . . . . . . . . . . . . . . . . 39
4.3 Means to Control the Resources . . . . . . . . . . . . . . . . . . . . . 40
4.4 Aspects of Power Control . . . . . . . . . . . . . . . . . . . . . . . . 44
4.5 Survey over Proposed Power Control Algorithms . . . . . . . . . . . 47
v
vi Contents
5 Local Analysis 55
5.1 Motivating Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.2 Dynamical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.2.1 Power Control Algorithms . . . . . . . . . . . . . . . . . . . . 59
5.2.2 Time Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2.3 Nonlinearities . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.4 Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.3 LogLinear Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3.1 Stability of Linear Systems . . . . . . . . . . . . . . . . . . . 67
5.3.2 Local Analysis of Power Control Algorithms . . . . . . . . . . 68
5.3.3 Summary, Local Analysis of LogLinear Algorithms . . . . . . 71
5.4 Describing Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4.1 Describing Functions with Zero Phase Assumption . . . . . . 72
5.4.2 DiscreteTime Describing Functions . . . . . . . . . . . . . . 74
5.4.3 Simplied Nonzero Input Case . . . . . . . . . . . . . . . . . 76
5.4.4 Describing Function of a Relay . . . . . . . . . . . . . . . . . 76
5.5 LogLinear Algorithms with a Static Nonlinearity . . . . . . . . . . . 77
5.5.1 Summary, Describing Functions Analysis . . . . . . . . . . . 79
5.6 Eect of AutoInterference . . . . . . . . . . . . . . . . . . . . . . . . 79
5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.A Local Loop Analysis of the AAW Algorithm . . . . . . . . . . . . . . 83
6 Local Design 85
6.1 Fading Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.1.1 Shadow Fading . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.1.2 Multipath Fading . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.1.3 Fading in the Spatial Frequency Domain . . . . . . . . . . . . 91
6.1.4 Measurement Procedures . . . . . . . . . . . . . . . . . . . . 91
6.2 Time Delay Compensation . . . . . . . . . . . . . . . . . . . . . . . . 92
6.2.1 Fixed Step Power Control Algorithm . . . . . . . . . . . . . . 94
6.2.2 LogLinear Power Control Algorithms with TDC . . . . . . . 96
6.3 Pole Placement Design of LogLinear Algorithms . . . . . . . . . . . 98
6.3.1 Controller Structures . . . . . . . . . . . . . . . . . . . . . . . 98
6.3.2 Properties of Linear Systems . . . . . . . . . . . . . . . . . . 99
6.3.3 Pole Placement Design . . . . . . . . . . . . . . . . . . . . . . 100
6.4 Practical Design Aspects . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.4.1 LogLinear Power Control Algorithms . . . . . . . . . . . . . 104
6.4.2 Decision Feedback Algorithm . . . . . . . . . . . . . . . . . . 109
6.4.3 Summary, Design Aspects . . . . . . . . . . . . . . . . . . . . 110
6.5 Predictive Power Control . . . . . . . . . . . . . . . . . . . . . . . . 110
6.5.1 Disturbance Prediction . . . . . . . . . . . . . . . . . . . . . 111
6.5.2 MinimumVariance Control . . . . . . . . . . . . . . . . . . . 112
6.6 Nonlinear Components . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.6.1 AntiReset Windup . . . . . . . . . . . . . . . . . . . . . . . . 114
Contents vii
6.6.2 Selector or Switch . . . . . . . . . . . . . . . . . . . . . . . . 115
6.6.3 Relay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.7 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.7.1 Simulation of an Isolated Local Loop . . . . . . . . . . . . . . 116
6.7.2 MonteCarlo Simulations . . . . . . . . . . . . . . . . . . . . 119
6.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7 Global Analysis 123
7.1 Optimal Global Power Assignments . . . . . . . . . . . . . . . . . . . 124
7.1.1 Optimal Balancing Assignments . . . . . . . . . . . . . . . . 124
7.1.2 Optimal Assignments Relative a Target SIR . . . . . . . . . . 127
7.2 Convergence in Linear Scale . . . . . . . . . . . . . . . . . . . . . . . 130
7.2.1 Standard Interference Functions . . . . . . . . . . . . . . . . 131
7.2.2 Convergence of Some LogLinear Algorithms . . . . . . . . . 132
7.2.3 Convergence of Some Decision Feedback Algorithms . . . . . 134
7.3 LogLinear Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 135
7.3.1 The Global System as Interconnected Local Loops . . . . . . 136
7.3.2 Interference Linearization . . . . . . . . . . . . . . . . . . . . 140
7.3.3 Robust Stability of LogLinear Power Control Algorithms . . 142
7.3.4 Two Mobile Case . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.3.5 The Conservativeness of Theorem 7.20 . . . . . . . . . . . . . 150
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.A Proof of Theorem 7.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.B Proof of Lemma 7.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
7.C Proof of Theorem 7.14 . . . . . . . . . . . . . . . . . . . . . . . . . . 156
7.D Robust Stability of Power Control Algorithms . . . . . . . . . . . . . 158
7.D.1 Robust Stability . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.D.2 Proof of Power Control Related Theorems . . . . . . . . . . . 161
8 Nonlinear Estimation 165
8.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
8.1.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
8.1.2 Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.2 Maximum Likelihood Estimation . . . . . . . . . . . . . . . . . . . . 169
8.2.1 A Simple Example . . . . . . . . . . . . . . . . . . . . . . . . 169
8.2.2 Adaptive Estimation . . . . . . . . . . . . . . . . . . . . . . . 171
8.3 SemiLinear Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.4 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.4.1 Abrupt Interference Step . . . . . . . . . . . . . . . . . . . . 174
8.4.2 Runtime Estimation . . . . . . . . . . . . . . . . . . . . . . . 175
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.A The Measurement Reports in GSM . . . . . . . . . . . . . . . . . . . 179
viii Contents
9 Outer Loop Control 181
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.2 Quality Measures in Cellular Radio Systems . . . . . . . . . . . . . . 182
9.2.1 Speech Quality Measures . . . . . . . . . . . . . . . . . . . . 182
9.2.2 Practical Quality Measures . . . . . . . . . . . . . . . . . . . 183
9.2.3 Example: GSM . . . . . . . . . . . . . . . . . . . . . . . . . . 185
9.3 Quality Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.3.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.3.2 Parameter Ranges in PointMass Approximations . . . . . . . 191
9.3.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 192
9.4 Stating Priorities in the Outer Loop . . . . . . . . . . . . . . . . . . 193
9.4.1 Fading Margins . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.4.2 Rate Adaption . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.4.3 Soft Dropping Power Control . . . . . . . . . . . . . . . . . . 195
9.4.4 Mode Switching Controllers . . . . . . . . . . . . . . . . . . . 196
9.4.5 Low Interference Problem . . . . . . . . . . . . . . . . . . . . 196
9.5 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
9.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
10 Network Simulations 199
10.1 Simulation Studies in General . . . . . . . . . . . . . . . . . . . . . . 199
10.2 Simple Network Simulation Environment . . . . . . . . . . . . . . . . 200
10.2.1 MOSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
10.2.2 Stabilizing Controllers in the Motivating Example . . . . . . 201
10.2.3 Conservativeness of Global Stability Requirements . . . . . . 202
10.2.4 Abrupt Change of Target SIR . . . . . . . . . . . . . . . . . . 203
10.3 FHGSM Simulation Environment . . . . . . . . . . . . . . . . . . . 204
10.3.1 System Level Eects of Controlling the Power . . . . . . . . . 206
10.3.2 Quality Measures . . . . . . . . . . . . . . . . . . . . . . . . . 209
10.3.3 AntiReset Windup . . . . . . . . . . . . . . . . . . . . . . . . 209
10.3.4 System Level Eects of Unstable Controllers . . . . . . . . . 210
10.3.5 Inuence of Estimation . . . . . . . . . . . . . . . . . . . . . 210
10.3.6 Quality Mapping . . . . . . . . . . . . . . . . . . . . . . . . . 210
10.4 WCDMA Simulation Environment . . . . . . . . . . . . . . . . . . . 211
10.4.1 Eect of Command Bit Errors on TDC . . . . . . . . . . . . 212
10.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
10.A MOSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
10.B Plots Associated with Simulations in Sections 10.3 and 10.4. . . . . . 216
11 Conclusions 225
Bibliography 229
Index 243
Notation
Symbols
Both values in linear and logarithmic scale (for more on units, see below) are used
in the thesis. Therefore, the convention of indicating linear scaled values by a bar,
. Hence, p is a value in linear scale, and p denote the corresponding value in
logarithmic scale. Moreover, the index i denote a specic terminal
p
i
, p
i
transmitter power
p, p vectors with components as above
p
, p
, i)
p
t
i
, p
t
i
powers at steady state (corresponds to
i
=
t
i
, i)
p
t
, p
t
vectors with components as above
p
i
internally monitored power used in TDC
p equilibrium deviation p
i
p
t
i
g
ij
, g
ij
power gain between terminals i and j
G,
G gain matrix (or Gmatrix) with entries as above
C
i
,
C
i
desired signal power
1
C
t
i
,
C
t
i
target value when using Cbased power control
I
i
ltered interference power,
I
i
(t) = F
I
(q)C
I
(t)
I
i
() interference function with respect to terminal i
I() interference function (vector with cpomponents as above)
m
I
mean interference at a specic receiver
I
interference standard deviation at a specic receiver
i
,
i
signaltointerference ratio (SIR)
, vectors with components as above
t
i
,
t
i
individual target SIR:s
t
diagonal matrix with diagonal elements
t
i
i
ltered SIR,
i
(t) = F(q)
i
(t)
ix
x Notation
i
adjusted SIR using TDC
i
,
i
receiver eciency
i
.
cc
crosscoupling matrix
cc
degree of crosscoupling,
cc
= 
cc

i
crosscoupling parameters
m
feasibility margin
L
r
relative load
q timeshift operator (cf. z in the ztransform)
R(q) loglinear controller
F(q), F
g
(q), F
I
(q) loglinear lters
Units
W, mW power units in linear scale.
dB a ratio of powers in linear scale p/ p
ref
is given in dB by
10 log
10
p
p
ref
= 10 log
10
( p) 10 log
10
( p
ref
) = p p
ref
.
dBW a power level compared to 1 W.
dBm a power level compared to 1 mW. Note that
p dBW = (p + 30) dBm.
Abbreviations
The list below only include abbreviations most frequently used in the thesis.
AAW algorithm proposed by (Almgren, Andersson, and Wallstedt,
1994).
BER Bit Error Rate.
DPC Distributed Power Control algorithm.
FER Frame Erasure Rate.
FSPC Fixed Step Power Control algorithm.
SIR SignaltoInterference Ratio.
TDC Time Delay Compensation
1
Introduction
The global communications system today (the telephone system yesterday) is con
sidered as the largest manmade system all categories. While the demand for access
to services in such systems is exponentially growing, an increased interest in uti
lizing the available resources eciently can be observed. In this thesis, the subset
of wireless cellular communications systems will be in focus.
A consequence of the limited availability of radio resources is that the users
have to share these resources. Power control is seen as an important means to
reduce mutual interference between the users, while compensating for timevarying
propagation conditions. The powers are controlled using feedback, and feedback
result in a dynamical behavior that critically aect the performance. In this work
we use methods from control theory to analyze the dynamical eects and to design
appropriate control strategies.
1.1 Thesis Outline
The subtitle Analysis, Design and Estimation describes the core of the thesis, but
not in a consecutive order.
Chapter 2 illustrates the main ideas in the thesis by considering a simple case:
a small wireless network with two users. Yet simple, the case provides valuable
insight into some of the rather theoretical issues brought up in the thesis. Primarily,
this concerns emphasizing the relevance and importance of considering dynamical
1
2 Introduction
eects in power controlled cellular radio systems. For example, time delays are
identied as critical for the performance if not handled appropriately.
Some background material on cellular radio systems are provided in Chapter 3.
The objective is to provide a reader without prior knowledge with an understanding
of the problems and challenges in these systems as well as a familiarity with the
terminology used. After this foundational chapter, power control is put into the
context of radio resource management algorithms in Chapter 4. The focus of the
thesis is further claried, and various aspects of power control are characterized.
Typically, power control algorithms are either based on information feedback, where
essentially real valued measurements are available, or based on decision feedback,
where only decisions 1 are available. The chapter is concluded with survey over
previous contributions in the literature.
For practical reasons, it is necessary to control the powers in a distributed fash
ion. These distributed algorithms can be seen as interconnected local control loops.
In Chapter 5 we consider these local loops separately with respect to dynamical
behavior. Previous proposals are seen to fall into any of two categories: Algorithms
based on information feedback are associated with a linear local loop, which can
be analyzed with methods from linear systems theory. Conversely, decision feed
back algorithms are not solely linear, but rather linear with a static nonlinear
component. Such a control loop is approximatively analyzed using discretetime
describing functions.
Since the local loops are all or partly linear, we apply methods from control
theory in Chapter 6 to improve the performance of previous proposals and to
relate the performance to disturbance properties and physical parameters. When
relevant disturbance models are at hand, these are used for disturbance prediction
and controller design.
Chapter 7 addresses dierent global issues, such as capacity, load and whether
it is possible to meet the requirements from all users. Furthermore, it provides
results on stability and convergence of some central algorithms. Global stability
requirements are also posed as requirements on the local control loops.
In GSM, the measurements are subject to coarse quantization and made avail
able to the algorithm in measurement reports. The core problem in Chapter 8
is to locally extract as much relevant information from these reports. We discuss
the applicability of maximum likelihood estimation to the problem. Yet another
practical issue is to design algorithms to operate in an outer loop, providing the
inner loop with relevant set points. This is the main issue in Chapter 9.
The ambition with Chapter 10 is to illustrate dierent issues from the preceeding
chapters using network simulations. Both small and largescale simulations are
used, all with dierent purposes.
Finally, Chapter 11 provides some conclusive remarks, and point out possible
directions for future research.
1.2 Reading Directions 3
1.2 Reading Directions
The main ambition with the many introductory parts in the thesis is that every
body should skip something. Readers with a strong background in mobile commu
nications can skip Chapter 3 and the introductory sections in Chapter 4. When
necessary, references backwards help the reader to back track to cover up nota
tional discrepancies. Similarly, a reader with a background in automatic control or
systems engineering may skip sections on properties of linear systems. Hopefully,
every reader nds it useful to read the extended summary in Chapter 2. This
chapter also serves as a road map to the rest of the thesis. Chapter 8 on nonlinear
estimation can be read relatively independently from the other chapters. Chapter 9
is to some extent based on material in the preceeding chapter.
1.3 Contributions
The main contributions in this thesis can be summarized as:
The unifying approach to power control algorithms in Chapter 5, where they
are considered as distributed control loops. Furthermore the applicability
of linear systems theory methods to assess local stability, or rather local
instability, with respect to controller parameters and time delays. These
methods are also used to conclude that autointerference, which is due to
nonideal receivers, does not violate local loop stability.
Discretetime describing functions analysis of decision feedback power control
algorithms in Chapter 5.
Time delay compensation for general power control algorithms, as described
in Section 6.2.
Local loop design of power control algorithms in Chapter 6. This involves
linear design using pole placement and minimumvariance control, the ap
plicability of nonlinear components and the practical aspects discussed in
the same chapter. Moreover, the related issues on outer loop control in Sec
tion 9.4.
The results on capacity and relative load in Section 7.1, and generalization
of some results to include autointerference.
Stability and convergence of some central power control algorithms with time
delay compensation, when subject to delays in Section 7.2.
Results on global stability with respect to the approximate global system
in Section 7.3. Furthermore, the interpretation of the results as local loop
design requirements.
The concept of qualitybased control with respect to GSM in Chapters 8
and 9.
4 Introduction
1.4 Publications
Some of the material in this thesis has been, or will be, published elsewhere. Tech
nical reports and nonreviewed material are not included. The licentiate thesis
relates to several parts of this thesis:
J. Blom and F. Gunnarsson. Power Control in Cellular Radio Systems. Licen
tiate Thesis, Linkopings universitet, Sweden, June 1998.
The original formulation of power control algorithms as linear distributed control
appear in
J. Blom, F. Gunnarsson, and F. Gustafsson. Constrained power control subject
to time delays. In Proc. International Conference on Telecommunications,
Chalkidiki, Greece, June 1998.
The applicability of discretetime describing functions to relay control in Chapter 5
is addressed in
F. Gunnarsson, F. Gustafsson, and J. Blom. Dynamical eects of time delays
and time delay compensation in power controlled WCDMA. Submitted to
IEEE Journal on Selected Areas in Communications, WCDMA Special Issue.,
2000.
Preliminary results of the above and analysis of loglinear distributed control loops
in Chapter 5 are covered by
F. Gunnarsson, J. Blom, and F. Gustafsson. Power control in cellular systems
subject to constraints and time delays. In Proc. IEEE Global Telecommuni
cations Conference, Sydney, Australia, November 1998d.
F. Gunnarsson, J. Blom, and F. Gustafsson. Power control algorithms and
stability analysis for radio network control. In Proc. IEEE Conference on
Decision and Control, Tampa, FL, USA, December 1998c.
Some of the design strategies in Chapter 6 are also found in
F. Gunnarsson, F. Gustafsson, and N. Wiberg. Transmit power control time
delay compensation in a wireless communications system. US Patent Appli
cation No. 09/346316, 1999e.
F. Gunnarsson and F. Gustafsson. Time delay compensation in power controlled
cellular radio systems. Submitted to IEEE Communications Letters, 1999.
F. Gunnarsson, F. Gustafsson, and J. Blom. Pole placement design of power
control algorithms. In Proc. IEEE Vehicular Technology Conference, Houston,
TX, USA, May 1999d.
F. Gunnarsson, F. Gustafsson, and J. Blom. Improved performance using non
linear components in power control algorithms. In Proc. IEEE Vehicular
Technology Conference, Houston, TX, USA, May 1999c.
1.4 Publications 5
The nonlinear estimation techniques applied to the GSM measurement reports in
Chapter 8 also appeared in
J. Blom, F. Gunnarsson, and F. Gustafsson. Estimation in cellular radio systems.
In Proc. IEEE International Conference on Acoustics, Speech, and Signal
Processing., Phoenix, AZ, USA., March 1999.
F. Gunnarsson, J. Blom, and F. Gustafsson. Estimation of the carrierto
interference ratio in cellular radio systems. In Proc. IEEE Vehicular Technol
ogy Conference, Houston, TX, USA, May 1999b.
Some of the outer loop strategies in Chapter 9 are covered by
J. Blom, F. Gunnarsson, and F. Gustafsson. A new concept of power control
in cellular systems reecting challenges of todays systems. In Proc. IEEE
Global Telecommunications Conference, Sydney, Australia, November 1998b.
F. Gunnarsson, J. Blom, and F. Gustafsson. A concept of power control in
cellular radio systems. In Proc. IFAC World Congress, Beijing, P. R. China,
July 1999a.
F. Gunnarsson, J. Blom, and F. Gustafsson. Method and system for quality
based power control in cellular communications systems. US Patent Applica
tion No. 09/031,635, 1998a.
Results on convergence and stability from Chapter 7 have been submitted to IEEE
Vehicular Technology Conference, Fall 2000.
6 Introduction
2
Extended Summary
The objective of the thesis is to cover control theory related issues of power control
in cellular radio systems. This chapter is introduced to bridge the gap between the
short abstract and the whole thesis. The main ambition is to exemplify the main
results by considering simplistic situations. Primarily, this concerns emphasizing
the relevance and importance of considering dynamical eects in power controlled
cellular radio systems. Most of the material is more rigorously treated in the
following chapters. For clarity, the organization of this chapter is similar to the
entire thesis.
In order to tickle the intuition, the considered examples sometimes might seem
naive to a welleducated reader. However, recall the ambition to provide insight
into power control problems, which might not track the theoretically optimal path.
2.1 Introduction
The primary goal of cellular radio systems is to provide reliable communication
to users irrespective of location and situation. This involves a great number of
users, and complex timevarying phenomenons. For clarity, we rst consider a
foundational situation: digital communication between two points. The available
bandwidth between the transmitter and the receiver occupies W Hz. Assume that
the received signal consists of the transmitted signal and additive white Gaussian
noise with power
0
. Furthermore, the power used by the transmitter is denoted
7
8 Extended Summary
by p
0
. Such a communication link has limited capacity. A fundamental result
(Shannon, 1956) states that the maximum data rate in bits per second is given by
maximum number of bits/s = W log
2
1 +
p
0
W
0
. (2.1)
If the bandwidth is xed, the capacity thus depends on the signaltonoise ratio
(SNR) p
0
/
0
of the connection.
The channel is generally more complex. Typically, the received signal power
level depends on the distance between the receiver and transmitter and eects
from the environment. The latter comes about due to wave propagation eects,
such as diraction and reection phenomenons. Hence, a more relevant model is
to associate the channel with a timevarying multiplicative power gain
1
g(t). The
signaltonoise ratio can thus be expressed as
SNR =
p
0
g(t)
0
Signal power gains and power levels can be expressed using dierent scales. Log
arithmic (e.g., dB or dBW) or linear are often used. To avoid confusion we will
adopt the convention of indicating linearly scaled values with a bar. Thus g is a
value in linear scale and g the corresponding value in dB. Recall that the conversion
from linear scale to dB is g(t) = 10 log
10
( g(t)). Using values in dB, SNR is given
by
SNR = p
0
+g(t)
0
, [dB].
A time varying power gain g(t) is illustrated in Figure 2.1a. Gain variations as
depicted in the gure degrade the performance and reduce the worstcase capacity
of the connection. An intuitive approach is to compensate for the variations by
using a timevarying power p(t). If the channel variations are known to the trans
mitter, we could literally invert the channel to obtain a constant SNR. Ideally, we
would like to use the power p(t) = p
0
g(t), which is exemplied by the solid line
in Figure 2.1b. A dierent approach is to use excessive power to overcome even
the worstcase situation, see the dashed line in Figure 2.1b. At some level, we
could guarantee a minimum signaltonoise ratio, and thereby provide a required
data rate. Seemingly, the only drawback with using excessive power is high battery
consumption. However, when other connections are using the same channel, other
aspects apply. This is further discussed below.
Consider a simple multiuser scenario with two connections using the same chan
nel. The two mobile stations MS
1
and MS
2
are connected to the base stations BS
1
and BS
2
respectively, see Figure 2.2. The transmitted signal from BS
1
to MS
1
also
reaches MS
2
and vice versa. These unwanted signals are normally weaker than the
desired signals, but still create a mutual interference. In Figure 2.2, desired signals
are indicated by solid arrows, and undesired interfering signals by dashed arrows.
1
Sometimes, the term attenuation is used to emphasize that g(t) < 1.
2.1 Introduction 9
Channel gain [dB]
Transmission power [dBW]
Time [s]
a)
b)
Figure 2.1 An intuitive approach to compensate for a timevarying chan
nel as in a) is to invert the channel. This corresponds to using
the power as described by the solid line in b). An alternative
is to use excess power as given by the dashed line in b).
We will focus on the downlink (from BS to MS), to keep the notation clear. Each
communication link from base station BS
j
to mobile station MS
i
is characterized
by the signal power gain g
ij
(t). Thus, all possible power gains in this small network
are described by the gain matrix (or Gmatrix)
G =
g
11
g
12
g
21
g
22
.
Denote the transmitted power from base station BS
j
at time instant t by p
j
(t). Mo
bile station 1 will thus receive the desired signal power p
1
(t) g
11
(t) and interference
power
I
1
(t), given by
I
1
(t) = p
2
(t) g
12
(t) +
1
,
where
1
represents noise power as in the singleuser case discussed earlier in the
chapter. The signaltonoise ratio is now of less importance. Instead, the signalto
interference ratio (SIR) plays a similar role. For example, the SIR at mobile MS
1
,
is given by
1
(t)
=
p
1
(t) g
11
(t)
I
1
(t)
=
p
1
(t) g
11
(t)
p
2
(t) g
12
+
1
.
10 Extended Summary
g
11
(t)
g
12
(t)
g
21
(t)
g
22
(t)
MS
1
MS
2
BS
1
BS
2
Figure 2.2 In the considered communication situation, two connections are
using the same channel and create mutual interference (dashed
arrows). The desired signal paths are indicated by solid arrows.
The corresponding SIR in dB is
1
(t) = p
1
(t) +g
11
(t) I
1
(t).
As indicated by the capacity expression in Equation (2.1), the required data rate
is related to SIR. We therefore assume that the requirements of the two users
are provided as target SIR:s,
t
i
. Essentially, power control is about assigning
appropriate p
i
(t):s to meet these requirements. The time instants t will from now
on be considered as time instants when the power is updated in the base stations.
2.2 Power Control Algorithms
If all four power gains g
ij
(t) and both the noise powers
i
were known at every
time instant t, it would be possible to compute the optimal powers to use at each
base station. This is not realistic in a practical situation. The simple reason is
that it requires a centralized solution, which would involve extensive signaling of
information. Instead, we require that each base station updates its own power
based only on SIR measurements from the connected mobile (distributed control ).
Dierent protocols may be used to feed back these measurements. We will consider
two important feedback situations:
Information feedback. The mobile feeds back the exact SIR measurements
or the error
e
i
(t) =
t
i
i
(t). (2.2)
The important characterization is essentially that real numbers are fed back.
Decision feedback. In a power control setting, we typically associate this
with feedback of the sign of the error in (2.2)
s
i
(t) = sign(
t
i
i
(t)) = sign(e
i
(t)) (2.3)
2.3 Local Analysis 11
Thus, only one bit is needed for command signaling, which makes the scheme
bandwidth ecient.
In this introductory chapter, we will consider a simple integrator as control al
gorithm. Depending on feedback assumptions, the powers will thus be controlled
as
Information feedback: p
i
(t + 1) = p
i
(t) +e
i
(t) (2.4a)
Decision feedback: p
i
(t + 1) = p
i
(t) +s
i
(t) (2.4b)
Yet simple, these two algorithms include most of the proposed algorithms to
date. For a more detailed discussion, see Section 5.2. The actual time between
consecutive power updates, the sample interval T
s
, varies from systems to system.
For example T
s
= 0.48 s in GSM and T
s
= 1/1500 s in WCDMA.
As seen above, the controller itself contains a delay of one sample interval.
However, both measuring and control signaling take time, resulting in additional
time delays in the system. A relevant model is that the measured SIR available
to the algorithm at time instant t, depends on the power level at t 1. This is
typically the situation in GSM and WCDMA. The SIR at mobile MS
i
is thus given
by
i
(t) = p
i
(t 1) +g
ii
(t) I
i
(t). (2.5)
2.3 Local Analysis
The global system can be seen as two local distributed control loops interconnected
via the mutual interference. Valuable insight can be obtained by studying these
local loops separately. In fact, as will be motivated in Chapter 5, local loop stability
is a necessary, but not sucient condition for global system stability.
The terms instability and stability of a local loop has to be well dened in
order to avoid confusion. Rather simplied, instability comes about when relying
too much on outdated information such as in Equation (2.5). This results in over
compensation, which may be aggravated over time. A thorough study of local loops
with respect to various algorithms is provided in Chapter 5.
Considering information feedback and the situation at mobile MS
1
, the local
loop dynamics is described by Equations (2.2), (2.4a) and (2.5), which combine to
p
1
(t + 1) p
1
(t) +p
1
(t 1) = (
t
1
g
11
(t) +I
1
(t)). (2.6)
The power p
i
(t) thus satises a linear dierence equation. Its solution can therefore
be associated with the characteristic equation
z
2
z + = 0.
Considering results from linear systems theory (see Section 5.3) we conclude that
the solution to the dierence equation is stable if and only if the roots (possibly
12 Extended Summary
complex) to the characteristic equation are located within the unit circle. The
product of the roots is equal to . Therefore, the local control loop is stable if
0 < < 1. For example the choice = 0.9 provides a locally stable controller.
General power control algorithms in the information feedback case are analyzed
with respect to local stability, using similar root locus analysis in Chapter 5.
The sign function, which is a vital component in decision feedback, has both a
stabilizing and a destabilizing eect. It is stabilizing in the sense that we cannot
overcompensate an outdated measurement by more than as seen in (2.4b). On
the other hand it is destabilizing, since the error is never compensated for by less
than either.
The corresponding local loop using decision feedback can be seen as a linear
system with a static nonlinearity (the sign function). Figure 2.3 illustrates the
local dynamics of the algorithm by a block diagram. Simulations of this local
t
i
a.
1
ei(t)
f(e)
G
si(t) i(t)
0 10 20 30 40 50
34
33
32
31
30
29
b.
p
i
(
t
)
Time instants t
Figure 2.3 When using decision feedback, the local loop can be seen as a
linear system G in series with a static nonlinearity f(e) (the
sign function). These components combine to the local loop
depicted in a. Simulations of this local loop indicate oscillating
power levels, as seen in b.
loop indicate an oscillative behavior, as seen in Figure 2.3b. Chapter 5 introduces
the discretetime describing function of the sign function to describe its dynamical
eects together with time delays and the integrating controller in (2.4b). The anal
ysis predicts an oscillation in the power levels characterized by a triangular wave
with a period of N = 6 samples and an amplitude 1.5. This is in accordance with
the observed behavior in Figure 2.3b. Discretetime describing function analysis
is presented more rigorously in Chapter 5. The analysis predicts several modes of
oscillation, and external disturbances may stimulate switching between the modes.
2.4 Local Design
As disclosed in Section 2.3, instability of local loops is essentially a result of too
much trust in outdated information. The behavior can be improved by either
relying less on the measurements or by predicting future measurements.
2.4 Local Design 13
Relying less on the measurements corresponds to a smaller controller parameter
. This will ensure local stability in case of information feedback. Stability is of
course necessary, but dierent stabilizing controllers provide very dierent behav
ior. Therefore, some design methods from control theory are applied to obtain good
performance in Chapter 6. Using pole placement techniques, the local loop perfor
mance is optimized with respect to a criterion, resulting in the optimal parameter
value = 0.34. In case of decision feedback, the amplitude of the oscillations in
the local loop is directly proportional to . Hence, the amplitude can be reduced
by choosing a smaller .
The core problem with outdated information is that the most recently com
puted output powers are not reected in the SIR measurements in (2.5). However,
these power levels can be monitored in the mobile station, and used to adjust the
measurements. The strategy for a general control algorithm is referred to as time
delay compensation (TDC) in Section 6.2. It is essentially a simple prediction of
future SIR measurements.
If we utilize information feedback, the powers can be monitored by an internal
copy of the power control algorithm in (2.4a). Then, the SIR in (2.5) is naturally
adjusted. The adjustments in the mobile thus consists of the following steps:
1. Adjust the SIR measurement using internally monitored powers:
i
(t) =
i
(t) + p
i
(t) p
i
(t 1).
2. Feed back the control error based on the adjusted measurement:
e
i
(t) =
t
i
i
(t).
3. Monitor power to be used by the base station:
p
i
(t + 1) = p
i
(t) +e
i
(t).
This scheme can be simplied when using decision feedback, since the dierence
between two consecutive powers is always . Therefore, TDC reduces to a single
equation:
1. Feed back the decision:
s
i
(t) = sign(
t
i
i
(t) +s
i
(t 1)).
The simple prediction in the compensation above only involve the transmitted
powers. If dynamical models of the power gains are at hand, they can be utilized for
prediction of future power gains or to design appropriate power control algorithms.
Good methods for estimating such models online are available. When using this
information, the local control performance can be improved signicantly.
Chapter 6 contains more details on design of local control algorithms. Further
more, it addresses some practical design aspects, such as various controllers ability
to mitigate timevarying power gains. These properties are related to the update
rate of the controller.
14 Extended Summary
2.5 Global Analysis
Hitherto, the interconnections between the two loops have been neglected. If there
were no such interconnections, it would always be possible to meet the target SIR
requirement in each loop (at least if sucient power is available). Otherwise, this
ability is strongly dependent on the global system. Here, global issues like capacity,
load and stability are in focus.
It is natural to discuss the power control problem at two levels. The local
control loops were discussed in the preceding two sections. Capacity and load are
typically discussed at a global level. We assume that the local loops successfully
compensates for the fast variations, so that the power gains can be considered as
constant. Moreover, stability and convergence are also discussed at a global level.
Again, the power gains are assumed constant, but the motivation is that we are
considering the short term behavior of the global system, and the power gains are
assumed to uctuate more slowly. To exemplify the somewhat theoretical results
in this section, we consider the particular case:
G = [ g
ij
] =
1 0.08
0.01 0.8
(2.7)
The load and capacity of a system relate to whether specic user requirements
can be supported or not. Therefore, they are naturally discussed in terms of the
target SIR:s
t
i
. Conceptually, the capacity of a system describes the optimal
utilization of the resources. An intuitive optimization strategy is to maximize the
smallest SIR of the two users. This leads to the balanced situation, where both
mobiles experience the same SIR (Zander, 1992b). In the current situation with
two connections the expression for optimal balanced SIR simplies to
=
1
max eig
0
g12
g11
g21
g22
0
=
g
11
g
22
g
12
g
21
. (2.8)
The example gains in (2.7) yield the optimal balanced SIR
= 31.6 (
= 15 dB).
Unfortunately, it is not possible to achieve the optimal balanced SIR in practice,
since it requires innite powers. Instead it serves as an upper bound on the achiev
able capacity.
If the two mobiles are requesting dierent services, they will probably require
dierent target SIR:s. Consequently it is interesting to study the load of the system
given target SIR requirements. For that matter we introduce the relative load
L
r
in Chapter 7. In the case of two connections, it is given by
L
r
= max eig
0
g12
g11
t
1
g21
g22
t
2
0
g
12
g
21
g
11
g
22
t
1
t
2
=
1
t
1
t
2
An important result in Chapter 7 is that it is possible to meet the target SIR
requirements if and only if
L
r
< 1. Then, the power control problem is said to be
2.5 Global Analysis 15
feasible. In the twoconnection case, the feasibility requirement reduces to
L
r
< 1
t
1
t
2
<
1
2
(
t
1
+
t
2
) <
. (2.9)
Consider the particular case in (2.7). Both the requirements (
t
1
,
t
2
) = (10, 10) dB
and (
t
1
,
t
2
) = (10, 16) dB correspond to feasible power control problems, while
(
t
1
,
t
2
) = (18, 16) dB does not.
Other important global issues include stability and convergence. Assume that
the considered power control problem is feasible. Previous results on convergence
have almost exclusively omitted time delays. For example, the specic case of
information feedback and = 1 provides appealing convergence properties when
omitting time delays. This does not hold true when subject to time delays. As
seen in Section 2.3, the algorithm corresponds to an unstable system. However,
when using the time delay compensation in Section 2.4, the combination results
in a controller that is stable and converges for all possible time delays. This is
formally proven in Chapter 7.
The benets of the time delay compensation are also evident when considering
decision feedback. This algorithm is stable and converges (Herdtner and Chong,
2000) to the bounded region
[
t
i
i
(t)[ (2n + 2)
When using TDC we show in Chapter 7, that the resulting controller is stable and
converges to the region
[
t
i
i
(t)[ (n + 2)
Hence, the longer the time delay, the more emphasized improvements using TDC.
Note, however, that only the fact that the bound is tighter does not imply that the
error variance is smaller.
When considering other algorithms and lters, the foundation on which these
proofs rely, is insucient. Instead we will discuss stability in terms of an approxima
tive analysis, based on linearization around a specic equilibrium point. A general
stability result is provided in Chapter 7. To illustrate the main ideas, we will focus
on the particular example of information feedback power control in (2.4a). Even
though the local loops are linear, the interconnections are not. The interference at
mobile MS
1
in dB is given by
I
1
(t) = 10 log
10
( p
2
(t) g
12
+
1
), (2.10)
which clearly is nonlinear. One way of circumvent this is to linearize the interference
at a particular point to obtain a linear model valid in a neighborhood of that point.
Assume that the power control problem is feasible. Hence, there exist power
assignments p
t
1
, p
t
2
corresponding to the target SIR:s
t
1
,
t
2
. The interference
at mobile MS
1
when using these powers is denoted by I
t
1
. If we linearize the
16 Extended Summary
interference around the equilibrium power level of the other base station p
t
2
, we get
I
1
(t) I
t
1
+
I
1
(t)
p
2
p2(t)=p
t
2
(p
2
(t) p
t
2
).
Using Equation (2.10) yields
I
1
(t) I
t
1
+
g
12
p
t
2
I
t
1
(p
2
(t) p
t
2
) =
1
(p
2
(t) p
t
2
), (2.11)
where
1
describes the inuence from the other mobile and can be rewritten as
1
=
g
12
p
t
2
I
t
1
=
I
t
1
1
I
t
1
.
Since all quantities are positive, we conclude that 0
1
< 1. It is instructive to
establish the relation between the relative load
L
r
and the
i
:s. It is easy to show
that the following holds
L
r
=
1
2
.
Hence, there is a clear connection between the load of the system and the extent
to which the loops aect each other.
If we insert the linearized interference in (2.11) into Equations (2.2) and (2.4a),
we obtain
p
1
(t + 1) = p
1
(t) +
t
1
p
1
(t 1) g
11
+I
t
1
+
1
(p
2
(t 1) p
t
2
)
. (2.12)
Note that the power from the other base station p
2
(t 1) also is delayed by one
sample. Since the linearization is valid in a neighborhood of p
t
1
, we are interested
in the dynamics in terms of the deviation from that point. Therefore, we introduce
p
1
(t) = p
1
(t) p
t
1
.
The same analysis applies to the situation of the other mobile MS
2
. If we use
p
i
(t) above to rewrite Equation (2.12), we obtain the two equations (one for each
connection) describing the global system dynamics as
p
1
(t + 1) = p
1
(t) p
1
(t 1) +
1
p
2
(t 1)
p
2
(t + 1) = p
2
(t) p
2
(t 1) +
2
p
1
(t 1).
This approximation is linear and can be analyzed using linear methods. Stability
results regarding the twomobile case and the general case can be found in Chap
ter 7. In the twoconnection case, these requirements allow a compact formulation.
We conclude that the global system is stable in a neighborhood of target SIR if the
following inequality holds
<
1
1 +
1
2
=
1
1 +
L
r
. (2.13)
2.6 Estimation and Outer Loop Control 17
Hence, the higher relative load, the tighter requirements on and vice versa. The
conclusion is that system with high loads is harder than average to control and
requires more careful local design. Moreover, 0.5 always provide globally
stable twoconnection systems. For example, the choice = 0.34 in Section 2.4
corresponds to a stable global system independent of the load. Conversely, we
cannot guarantee stability of the choice = 0.9 in Section 2.3. When applying
TDC to the control algorithm, the stability requirements are dierent. As shown
in Chapter 7, TDC results in a stable global system for 0 < 2.
2.6 Estimation and Outer Loop Control
In some practical cases, the power update computations take place in the base
stations, both for the up and downlink connections. The mobile acts as a slave and
reports requested (highly quantized) measurements. The base station then needs
to recover as much relevant information from these reports as possible. Another
practical issue is related to issuing appropriate target SIR:s. As was seen in the
previous section, not all target SIR requirements can be met. Both these issues are
dealt with in Chapters 8 and 9 respectively.
In GSM, the measurement reports comprise RXQUAL, which is related to the
perceived quality and RXLEV, which describes the total received signal power.
Since the focus is on one of the connections, we drop the index i for clarity. Assume
that we want to recover the desired signal power C(t) = p(t)+g(t) and the SIR (t)
from the measurements. The corresponding relations are nonlinear, which make
estimation a bit challenging. Chapter 8 proposes a maximum likelihood estimator
to address the problem. In essence, it is based on a probabilistic description of
the problem. The core operation is to consider the likeliness that the observed
measurements were generated from a model with parameters C and . Estimates
are obtained as the parameters that maximizes this likeliness, or likelihood function.
Figure 2.4 exemplies the likelihood at a specic time instant, and indicates how
to obtain the estimates.
This far we have assumed that the user requirements can be posed as target
SIR:s, which is said to correspond to a specic data rate. Unfortunately, the
relationship between a data rate and SIR is not perfectly known. In fact, it might
be time varying. Thus, there is a need to adapt the target SIR:s as well. This
is typically taken care of in an outer control loop, which operates at an order of
magnitude slower update rate. Chapter 9 proposes a method where statistics about
the relation between quality of service and target SIR is utilized.
The relative load
L
r
cannot be monitored online, since the measure is based
on the knowledge of power gains g
ij
. Therefore, there is a risk that the requested
target SIR:s correspond to an infeasible power control problem. If using the simple
integrating algorithms in (2.4) when the power control problem is infeasible, the
two connections will start competing about the target SIR:s. This race will not
end until one or both connections are using maximal powers. Such cases can be
handled by forcing connections that require high powers to aim a lower target
18 Extended Summary
110
105
100
95
90
0
5
10
15
20
25
30
4
3.5
3
2.5
2
1.5
1
0.5
0
C
t
2
= 10 dB, and that the control algorithms are perfectly initialized, i.e., p
i
(0) = p
t
i
.
At time instant t = 10, the second mobile upgrades his target SIR to
t
2
= 16 dB,
while the rst remains at
t
2
= 10 dB. Both situations correspond to feasible power
control problems as concluded in Section 2.5. The two mobiles remain at the same
locations, implying constant power gains g
ij
throughout the simulation. Figure 2.5
illuminate the recovering ability of dierent controller congurations. Subgures
ad consider information feedback, while ef address decision feedback. From the
simulation, we note the following for the feasible power control problem
a. Information feedback with = 1 does not correspond to a stable local loop
Therefore it is intuitive that the global system is unstable.
b. A smaller yields stable local loops, but the global system is still unstable.
This illustrates the fact that local stability is necessary but not sucient for
global stability.
c. An even smaller = 0.34 will always result in a stable global system.
d. When using TDC, the controller with = 1 corresponds to a stable system.
2.7 Simulations 19
0 5 10 15 20
40
20
0
20
40
0 5 10 15 20
40
20
0
20
40
0 10 20 30 40
5
10
15
20
0 10 20 30 40
5
10
15
20
0 10 20 30 40
5
10
15
20
0 10 20 30 40
5
10
15
20
i
(
t
)
i
(
t
)
i
(
t
)
i
(
t
)
i
(
t
)
i
(
t
)
Sample instants Sample instants
Sample instants Sample instants
Sample instants Sample instants
a. b.
c. d.
e. f.
Figure 2.5 The recovering ability of various controllers when one of the
mobiles upgrades his target SIR from 10 to 16 dB. Information
feedback: a. = 1, b. = 0.9, c. = 0.34, d. = 1 and
TDC. Decision feedback, e. = 1, f. = 1 and TDC.
e. Decision feedback always result in an oscillative behavior, even when perfectly
initialized.
f. The oscillations in the decision feedback case can be reduced by employing
TDC.
Simulations using other controllers and strategies are presented in Chapter 10.
Dierent phenomenons studied in the thesis are illuminated using both small and
largescale simulation environments.
20 Extended Summary
3
Cellular Radio Systems
In this chapter, the focus is on mobile communications systems. The objective is
to cover the main issues so that a reader without prior knowledge will understand
the problems and challenges in such systems. Subjects related to the scope of this
thesis are, however, presented in more detail. In the presentation we also introduce
the terminology used in the eld.
3.1 Overview
There are several books covering the development of mobile communications sys
tems, see e.g., Ahlin and Zander (1998); Garrard (1998); Jakes (1974); Lee (1982);
Rappaport (1995); Steele (1992) and St uber (1996).
History
The development of wireless communications stems from work during the second
half of the nineteenth century. In 1864, Maxwell predicted the existence of electro
magnetic radiation and formulated the basic theory that is still in use today, over a
century later. Maxwells theory was veried experimentally by Hertz in 1887, and
therefore the electromagnetic waves are sometimes referred to as Hertzian waves.
The origin of mobile radio is considered to be 1897, when Marconi was credited
with a patent for wireless telegraph, the rst radio related patent ever. On De
21
22 Cellular Radio Systems
cember 12, 1901, Marconi received a radio signal in Newfoundland, which was
transmitted from Cornwall, England. Thereby he showed that radio communica
tion was possible despite the curvature of the earth.
During the twentieth century, the radio technique was further improved, not the
least due to important inventions such as the vacuum tubes, amplitude modulation
(AM), frequency modulation (FM), transistors, integrated circuits etc. The rst
true mobile radio system used in practice was a police car radio system, introduced
at Detroit Police Department 1921, but the systems still had a long way to go
before they would reach the general public.
First Generation Cellular Systems
The concept of cellular systems originates from Bell Labs 1947. Instead of transmit
ting the signal as far as possible, the range would be deliberately limited, allowing
frequencies to be used again at a much shorter distance. However, the idea had to
wait for decades before it could be implemented. The technology needed was sim
ply not available. With the introduction of the Nordic Mobile Telephone (NMT)
system 1981 and the Advanced Mobile Phone Service (AMPS) system 1983, cellu
lar phones were not solely for the rich and famous anymore. The NMT standard
was dened in a collaboration between the Nordic countries, and an important
feature was international roaming. In the beginning it used radio frequencies in
the 450 MHz band, but in order to meet the needs, operation in the 900 MHz band
was introduced later. AMPS, which is operating in the 800 MHz band, became
widely spread over the world, and has the largest penetration among the analog
systems. Modied versions of AMPS include the British Total Access Communi
cations System (TACS) and the Japanese JTACS.
Second Generation Cellular Systems
The Global System for Mobile Communications (GSM)
1
was deployed in 1992
as the worlds rst digital cellular standard. Originally it operated only in the
900 MHz band, but nowadays it is adapted for 1800 MHz (DCS 1800) and 1900 MHz
(PCS 1900).
The incentive to consider digital systems was mainly dierent in the USA, since
AMPS had capacity problems. The problem was that, unlike in Europe, additional
spectrum for digital cellular systems at frequencies nearby the AMPSfrequencies
were not available. AMPS was also the dominating analog technology, and therefore
a new digital system had to be reverse compatible, preferably using most of the
infrastructure in AMPS. Two systems were proposed and brought into operation,
IS95 and DAMPS (Digital AMPS a.k.a. IS54, evolved into IS136). They were
both reverse compatible, but very dierent.
The rst digital cellular system in Japan, Personal Digital Cellular (PDC),
was introduced in 1994 and is very similar to IS54. It operates in the 800 MHz
1
The original meaning of the acronym was Groupe Special Mobile, which was a group established
in 1982 with the mission and mandate to dene future European cellular radio standards.
3.2 Radio Communication 23
and 1500 MHz bands and is the most spectrumecient among the systems of the
second generation.
The systems described above have an extensive coverage, but they will never
reach complete coverage. In order to ll in the gaps, there is a need for Mobile
Satellite Communications Systems, and at the present there are ve major systems
under development: Globalstar, ICO (formerly Inmarsat), Iridium, Odyssey, and
Teledesic (Keller et al., 1997).
Third Generation Cellular Systems
The needs for voice communication are more or less met by the narrowband systems
of the second generation. However, there are a number of other applications that
require more data to be transmitted. These include conferencing, video, database
access, Internet access etc. The concept name for this type of system is Interna
tional Mobile Telecommunications (IMT2000, see e.g., Chaudhury et al., 1999).
In traditional telephony systems, the calls have been circuitswitched, meaning
that the call occupies one channel throughout the call. Due to the bursty nature of
data communication this is inecient, and instead packet switching can be used,
where the data is divided into packets and transmitted independently. Therefore,
the second generation systems have evolved to include methods for data services.
The ambition is to gradually provide higher data rates, rst by circuitswitched
technologies, but eventually based on packet switching. For example, we mention
Cellular Digital Packet Data (CDPD, see e.g., Puk, 1996) in IS136 networks, High
Speed Circuit Switched Data service (HSCSD, see e.g., Scholeeld, 1997) and Gen
eral Packet Radio Services (GPRS, see e.g., Brasche and Walke, 1997; Cai and
Goodman, 1997) in GSM networks and High Speed Data (HSD, see e.g., Knisely
et al., 1998) in IS95 networks.
For higher data rates and a more exible resource allocation, the technology has
to be adopted accordingly. After several twists and turns, the dierent standardiza
tion bodies agreed to try to dene a global standard. From a technology perspec
tive, these systems are commonly referred to as either wideband CDMA (WCDMA,
see e.g., Dahlman et al., 1998) or cdma2000 (Knisely et al., 1998), from the two
dominating air interface proposals.
3.2 Radio Communication
The purpose of a radio communication system is to transmit a message from a
source to a user. These systems can be classied in a number of ways, but here
it is natural to treat analog and digital systems separately. However, all radio
communication systems are using the same medium the radio channel. The
transmitted signal is aected by radio propagation eects, thermal noise, interfering
signals and other types of disturbances. For further details on radio communication,
we refer to Ahlin and Zander (1998); Kucar (1991); Proakis (1995); St uber (1996).
An analog communication system consists of a transmitter, a radio channel and
a receiver. In the transmitter, the message signal is ltered and modulated onto
24 Cellular Radio Systems
the carrier signal. It is then transmitted over the radio channel and demodulated
in the receiver. The cellular systems of the rst generation are analog and use
Frequency Modulation (FM).
The basic elements of a digital communication system are illustrated in Fig
ure 3.1. It still consists of a transmitter and a receiver communicating over a radio
channel, but the internal components are a bit more sophisticated. Initially the
s
sq fer ber sir
I
Message
Estimated
Message
Source
Encoder
Channel
Encoder
Channel
Decoder
Source
Decoder
Channel
Modulation
Demodulation
Receiver
Transmitter
Figure 3.1 Block diagram of a digital communication system.
message is converted into a sequence of bits by the source encoder. In order to
enable a more reliable transmission, redundancy is added in a controlled manner
by the channel encoder. This redundancy in the bit sequence aids the receiver in
decoding the desired sequence. It can be used both for error detection and error
correction and is important to obtain good performance in a digital system, due to
the disturbances introduced by the channel.
The radio channel is in practice a channel for continuous signals, and therefore
the bit sequence has to be modulated onto an analog carrier signal. The primary
objective is to utilize the available bandwidth eciently, while achieving a pre
scribed performance, given by the number of erroneous bits received divided by
the number of bits transmitted. This is referred to as the Bit Error Rate (BER).
Note that the BER is primarily aected by the radio channel and the modulation
scheme used. Among the systems of the second generation, GSM and PDC are
using Gaussian Minimum Shift Keying (GMSK) and DAMPS and IS95 are using
/4shift Quaternary Phase Shift Keying (/4QPSK).
At the receiver, the signal is processed in the reversed order compared to the
transmitter, with the objective to reconstruct the original message. The demod
ulator will convert the corrupted waveform to an estimate of the transmitted bit
sequence, from which the channel decoder reconstructs the data stream by using
the redundancy in the transmitted bit sequence. Finally the source decoder tries
to reconstruct the original message signal.
3.3 Radio Wave Propagation 25
3.3 Radio Wave Propagation
The Maxwell equations provide, at least in principle, solutions to all problems
regarding electromagnetic elds. If all facts were known, we could in theory deter
mine the eects on a radio wave when it propagates through a medium. However,
these solutions tend to be far too complex to provide understanding of problems in
practical situations. Instead, simple models may be used to capture the physical
phenomenons coarsely, in order to provide an intuitive understanding of reallife
cases.
It is very hard to come up with simple models that are valid for all frequencies,
but since the issue here is mobile communications systems, we can focus on a
limited frequency band. The presented mobile communications systems are using
the UHF band (300  3000 MHz), which has well suited properties for wireless
communication. In this band it is possible to use signals with relatively large
bandwidth, which gives higher data transmission capacity. The eects of rain and
moisture are slight, but the radio waves are shielded and reected by mountains
and buildings.
The meaning of a communications channel is often used in an imprecise way.
Steele (1992) views a particular channel as the link between two points along a
path of communications. The following channel types are emphasized
The propagation channel is the physical medium that supports wave
propagation between the antennas.
The radio channel is the propagation channel and the transmitter and
receiver antennas viewed collectively.
The digital channel also incorporates the modulation and demodulation
stages.
The focus here will be on the radio channel. We are interested in the power gain
2
g
of the received signals, that is if a transmitter emits a signal using the power p the
receiver will observe a signal power of g p. In order to model this gain we separate
the propagation eects in three groups and use the model
g = g
p
g
s
g
m
.
The distance dependent attenuation and the antenna attenuation are modeled by
the path loss ( g
p
). Terrain variations are resulting in shielding and diraction
and is modeled by the shadow fading ( g
s
), while the eects of the reections are
captured by the multipath fading ( g
m
). These models are further explored below.
For more extensive material regarding radio wave propagation, see (St uber, 1996).
It will be important to note the dierence between g, which represents the
power gain in linear scale, and g which is the power gain in dB.
2
Often the term attenuation is used to stress the fact that the gain is less than one.
26 Cellular Radio Systems
3.3.1 Path Loss
On a long time average, the observed power at the receiver depends mainly on the
distance to the transmitter. One of the simplest models is based on the assumption
of propagation in free space, but except for satellite communication applications,
this is too coarse. When we consider propagation over plane earth an approximation
of the path loss is derived in (Jakes, 1974), and given by
g
p
=
C
p
r
, (3.1)
where r is the distance between the transmitter and the receiver, the path loss
exponent is a constant equal to 4, and C
p
is a constant which depends on antenna
specic parameters and the transmitted wavelength.
Empirical studies by Okumura et al. (1968) and Hata (1980) resulted in path
loss models for urban, suburban and rural areas. Equation (3.1) still holds for the
long time average gain if we note that the constants C
p
and depend on the type of
terrain. Typical values of are about 2 5, where the lowest value corresponds to
free space propagation, and the higher to urban environments with high buildings.
Note that the plane earth approximation is covered by this range in .
3.3.2 Shadow Fading
As mentioned above, the long time average is given by the path loss. Terrain varia
tions will however result in diraction and shielding phenomenons, which manifest
themselves as a slow variation in this average gain over a distance corresponding to
several tens of wavelengths. This eect is referred to as shadow fading. Okumura
et al. (1968) and Hata (1980) were pioneers in studying these variations. They used
values in dB and argued that the shadow fading can be modeled using a zeromean
Gaussian random variable, i.e.
g
s
N(0,
s
). (3.2)
Hence the linear shadow fading gain ( g
s
) can be modeled using a lognormal
distribution, i.e. the corresponding value in dB is normal distributed with zero
mean and variance
2
s
. This model is reasonably accurate when discussing separate
samples of the gain, but it is not sucient if we consider a receiver that is moving
around in the terrain. The eects from the terrain are correlated and if the receiver
is shielded at one instant, it will most likely be shielded for some time thereafter.
To include this spatial correlation, Gudmundson (1991) proposed an auto
correlation function that incorporates the velocity of the mobile v and relates it to
a correlation distance d. The basic idea is that two samples separated by a distance
of at least d are more or less uncorrelated. Spatially, the model can be described
as a zeromean Gaussian random variable ltered through a rst order lter. This
model was rened by Srensen (1998).
3.4 MultiUser Communications 27
3.3.3 Multipath Fading
In the presence of several large objects, there will be a great number of reected
signals that reach the receiver. Depending on their phase they interfere either
constructively or destructively resulting in multipath fading. For digital systems it
is important to determine if this fading can be assumed to be constant for each
received symbol. If that is the case, the fading is referred to as at, but if not,
we are experiencing intersymbol interference due to this frequency selective fading.
For low to moderate data rates the intersymbol interference can be counteracted
using channel equalization (St uber, 1996) and therefore we will model the multipath
fading as at.
A simplied worst case scenario arises when the LineofSight (LoS) ray is as
sumed completely shielded, but the reected rays reach the receiver and enable
transmission of information. If the arriving rays are modeled as isotropic, the
amplitude gain due to multipath is described by a Rayleigh distribution (Clarke,
1968). Therefore, multipath fading is commonly referred to as Rayleigh fading.
Just as in the case of shadow fading, the multipath fading has a spatial corre
lation as exemplied in Figure 3.2. Thus the fading is depending on the velocity of
the mobile station. One may also draw the erroneous conclusion, that the fading
is constant if the location of the mobile station is xed. The truth is that the
environment is also moving, and it is the movement relative to the environment
that matters. Since the origin of the multipath fading is dierent path lengths of
the reected waves, it is also depending on the wavelength of the signal. Therefore
we will also observe a correlation in the frequency domain, see Figure 3.2.
3.3.4 Example: Spatially Correlated Propagation
In Figure 3.3 we see a sample of how the power gain may vary when a mobile station
is moving in the terrain. It is assumed that the distance to the transmitter is almost
constant and that no LineofSight component reaches the receiver. Therefore the
variations are due to shadow and Rayleigh fading with spatial correlation.
3.4 MultiUser Communications
Consider a case of one base station (BS) serving a certain service area. The BS
transmits and receives data to and from the M mobile stations (MS) in the service
area. The communication goes in two directions, and one distinguishes between
the downlink (forward) channels from BS to MS, and the uplink (reverse) channels
from MS to BS. For simplicity we focus on the downlink from BS to MS number
1. Assume that the BS is transmitting the signals s
1
(t), s
2
(t), . . . , s
M
(t) to the
mobiles. Figure 3.4 provides a simple model of the situation at the receiver of MS
number 1. The receiver observes the desired signal and several interfering signals
plus thermal noise.
One important choice when designing a multiuser communication system is the
choice of carrier signals, and we will divide them into two categories: orthogonal and
28 Cellular Radio Systems
35
30
25
20
15
10
5
0
5
10
g
m
[dB]
Frequency
Spatial deection
Figure 3.2 A realization of the multipath fading as a function of spatial
deection and frequency.
nonorthogonal signals. The reason is that there are some characteristic dierences,
which will be pointed out in the following two sections.
3.4.1 Orthogonal Signals
When the signals are orthogonal, they can be seen as orthogonal base vectors
spanning the signal space. Therefore the receiver only has to extract the energy
along the s
1
dimension in the signal space in order to recover the information.
(In real systems, the channel aects the orthogonality at the receiver. Therefore,
signals are separated by a guard band to improve the orthogonality.) We will not
analyze the receiver structure in more detail, and instead focus on the commonly
used orthogonal signals.
The rst approach is to separate the users in the frequency domain. This is
referred to as Frequency Division Multiple Access, FDMA (Figure 3.5a), and it has
been very popular, mainly because it is well suited for analog technology. The main
challenge is the frequency synchronization, which is needed to extract the band of
interest. FDMA is the technology used in systems of rst generation. The channel
bandwidth used in these systems are 2530 kHz.
Similarly, we can separate the users in the time domain using Time Division
Multiple Access, TDMA (Figure 3.5b). Then each user is allowed to use the entire
frequency band, but only at assigned time slots in a round robin fashion. TDMA
has the advantage over FDMA that the receiver of the mobile station is only used
3.4 MultiUser Communications 29
Power Gain [dB]
Traveled Distance
g
p
g
s
g
m
Figure 3.3 The model for radio wave propagation comprises path loss (g
p
),
shadow fading (g
s
) and multipath fading (g
m
).
1/N
t
of the time. During the rest of the time it can be used for scanning other
channels. Since the message is sent discontinuously, TDMA is better suited for
digital technology than for analog and the challenge is time synchronization.
Pure TDMA solutions are rare and instead hybrid FDMA/TDMA (Figure 3.5c)
is used, where the spectrum is divided up in frequency bands, which each are
divided into time slots. One example is DAMPS, where each 30 kHz subband in
AMPS is divided into three time slots, resulting in a threefold capacity increase.
In GSM, the subbands of 200 kHz are divided up into eight time slots. Compared
to the 25 kHz channels in NMT there is no capacity increase at all. Instead this is
used to enable more reliable communication.
In GSM there is an optional frequencyhopping pattern (see Section 3.6) mean
ing that the mobile is using a dierent subband each time slot. This hopping
sequence is determined by a code, and therefore this technique is referred to as
Frequency Hopping Code Division Multiple Access, FHCDMA. Altogether, the
scheme can be seen as the hybrid FDMA/TDMA/CDMA (Figure 3.5d).
3.4.2 NonOrthogonal Signals
To avoid the problems of frequency and time synchronization, we can use signals
that are almost orthogonal. Each user is assigned a code and is then transmitting
using the entire frequency band. The receiver extracts the desired signal by corre
30 Cellular Radio Systems
.
.
.
+
s
1
(t)
s
2
(t)
s
M
(t)
g
1
(t)
g
2
(t)
g
M
(t)
Receiver
Transmitter
Transmitter
Transmitter
1
(t)
Figure 3.4 Simple model of a multiuser system. The focus is on the sit
uation at a specic mobile station. Note that the radio chan
nels are characterized by their power gains g
i
described in sec
tion 3.3. The signal
1
(t) represents the thermal noise.
lating the received signal and the code. This spreadspectrum technique is referred
to as Direct Sequence Code Division Multiple Access, DSCDMA. By spreading and
despreading the signal, the link becomes more robust against fading and interfer
ence. The performance gain obtained by spreading is related to the processing gain
(PG). In a system with the data rate R and the channel bandwidth, W, PG is
given by the ratio W/R (Viterbi, 1995).
Consider the uplink, and assume that there are two mobile stations in the
service area, where one is close to the base station and one is far away. If both
mobile stations are transmitting using the same power level, the received powers
at the base station might dier by several orders of magnitude, due to dierent
power gains of the signals (recall Section 3.3). The signal extraction at the receiver
based on correlation works ne when the powers of received signals are roughly the
same. This is not the case in the above situation, where the signal from the distant
mobile station is drowned by the signal from the close one. Therefore we have
to use appropriate power levels in order to tamper this nearfar eect (Gilhousen
et al., 1991). In the downlink, this is not a problem, since all the signals originate
from the same transmitter.
This technique is used today in the narrowband system IS95, where all ter
minals communicate on a frequency band of 1.25 MHz. In the systems of third
generation, similar technology will be used, but with a frequency band of 5 MHz.
That will enable higher data rates and mitigate fading better.
3.5 Cellular Radio Systems
When the service area is large, one base station will not suce. Instead the service
area is divided into a large number of cells, which has given rise to the name cellular
3.5 Cellular Radio Systems 31
1
1 1 1 1 1
1 1 1
1
2
2 2 2
2 2 2
3
3
3 3
3 3
3
t t
t t
f f
f f
a. b.
c. d.
Figure 3.5 Dierent multiple access techniques using orthogonal signals.
The users 1, 2, 3 may be separated either in frequency or in
time or both. a. FDMA, b. TDMA, c. FDMA/TDMA, d.
Frequency hopping FDMA/TDMA.
radio systems. Each cell is served by a base station (BS) and within the cell the
situation is similar to what is described in Section 3.4.
The size of dierent cells varies very much between rural and urban areas and
this is due to the limited number of users one cell can serve. In rural areas few and
large cells are sucient to meet the needs, and the radius may be several tenths
of kilometers. In denser populated areas where cellular phones are common, there
is a need for smaller and more numerous cells, serving an area of radius down to
about 100 meters. To meet the needs in the central parts of cities, it is getting
more common to use cells consisting of a short part of a street (micro cells) or a
room or a oor of a building (pico cells).
Usually the cells are depicted as hexagonal, but in reality their size and form are
irregular and depending on the terrain and the propagation conditions. Designing
a cellular system involves extensive cell planning. Among other things this involves
determining clever cell sites to get appealing propagation conditions and cell sizes
to meet the needs for communication at the present and in a near future. In order
to use the hardware eciently it is more and more common to colocate three base
station antennas and use sectorized antennas covering a sector of 120
each. When
the antenna is radiating in all directions the term omnidirectional antenna is used.
The operator of the cellular system is assigned a frequency band to use in
32 Cellular Radio Systems
the service area. This spectrum is split up in a number of channels (waveforms)
based on the multiple access scheme chosen for the operation of the system. When
assigning these channels to the base stations, the path loss is actually favorable.
Within the cell, the transmitted signal from the BS can be received by the MS at a
reasonable signal strength, but at some distant point outside the cell, the power of
the signal is negligible compared to the noise. Therefore the channels can be reused
in the service area. The available channels are divided into K channel groups and
each group is assigned to some base stations as exemplied in Figure 3.6. If we
assume a hexagonal cell pattern, the possible reuses K are given by K = a
2
+ab+b
2
,
where a and b are natural numbers. This yields K = 1, 3, 4, 7, 9, 12, 13, . . . . Note
that K = 1 corresponds to the trivial case where all channels are used in every cell.
Figure 3.6 The same channel group is reused in dierent cells in the net
work. This example shows the cells, to which one channel group
is assigned, when we are applying a reuse K = 9.
In order to keep the presentation clear we will consider the case of orthogonal
signals. Furthermore, let us focus on a specic downlink channel and on the mobile
stations and base stations using that channel. The terminals are numbered, so
that mobile station i is connected to base station i. (Since we focus on one single
channel, only one MS is connected to each BS on this channel.) Furthermore, the
power gain from base station j to mobile station i is denoted g
ij
. The information
about all downlink power gains on the channel at a time instant t can be condensed
into the G matrix of the channel,
G
c
(t). Assume that there are m connections
established on the channel. Then this Gmatrix is given by
G
c
(t) =
g
11
(t) g
1m
(t)
.
.
.
.
.
.
.
.
.
g
m1
(t) g
mm
(t)
. (3.3)
Thus the rst column contains the gains from base station 1 to the dierent mobile
3.5 Cellular Radio Systems 33
stations. This matrix is time variant, since each of the power gains is time variant
and since the dimension of the matrix varies with time when mobiles place new
calls or hang up. There are actually two Gmatrices: one for the uplink and one
for the downlink. This is due to the use of dierent frequency bands, as well as
dierent propagation conditions and antennas for the base stations and the mobile
stations.
Assume that base station i is transmitting using the power p
i
(t). The corre
sponding connected mobile station will experience a desired signal power
C
i
(t) =
g
ii
(t) p
i
(t), and an interference
I
i
(t), which is the sum of the powers from all other
base stations and thermal noise
i
(t). See Figure 3.7 for a situation of three inter
ferers. Since this interference is emanating from users on the same channel, it is
referred to as cochannel interference. Sometimes, the term multiple access inter
ference (MAI) is used instead. When we are employing a large frequency reuse or
there are few users in the system, the interference is neglectable compared to the
noise, and the system is noiselimited. In the opposite situation when we employ a
small reuse and several mobile stations are active, the system is interferencelimited
due to the dominating interference.
In a DSCDMA system, all the users are using the same frequency channel.
Hence the frequency planning process of assigning channels to the base stations is
avoided. It can be seen as employing a reuse of K = 1. Therefore everybody is
interfering with each other in the network and contributes to the interference at
each receiver. A DSCDMA system is typically interferencelimited.
Figure 3.7 Cochannel interference in a network employing frequency
reuse. The received signal at the mobile station consists of
the desired signal (solid), interfering signals from other base
stations (dashed) and thermal noise. The reuse pattern is the
same as in Figure 3.6 where the reuse K = 9.
34 Cellular Radio Systems
A quantity naturally related to the perceived quality is the signaltointerference
ratio (SIR) denoted . The SIR at mobile station i is dened by
i
(t) =
C
i
(t)
I
i
(t)
=
g
ii
(t) p
i
(t)
j=i
g
ij
(t) p
j
(t) +
i
(t)
. (3.4)
One other common term for roughly the same quantity is carriertointerference
ratio (C/I or CIR). These quantities are both used interchangeably, but more pre
cisely, SIR refers to the ratio at the base band, while CIR refers to the corresponding
ratio at the carrier frequency. Moreover, one diers between SIR before and after
despreading, and the dierence is equal to the processing gain. The main focus in
this thesis is the dynamical behavior of algorithms. Since the dierences are just a
matter of scale, they are neglectable.
Depending on the receiver design, propagation conditions and the distance to
the transmitter, the receiver is dierently successful in utilizing the available desired
signal power p
i
g
ii
. Assume that receiver i can utilize the fraction
i
(t) of the desired
signal power. Then the remainder
i
(t)
p
i
g
ii
acts as interference, denoted
autointerference (Godlewski and Nuaymi, 1999). Hence, the SIR expression in
Equation (3.4) transforms to
i
(t) =
i
(t) g
ii
(t) p
i
(t)
j=i
g
ij
(t) p
j
(t) +
i
(t)
p
i
(t) g
ii
(t) +
i
(t)
. (3.5)
From now on, this quantity will be referred to as SIR. For ecient receivers,
i
(t) = 1, and the expressions (3.4) and (3.5) are equal.
3.6 Diversity
The radio channel will result in a varying transmission quality due to fading and
thermal noise. In order to counteract these, one or several diversity methods can
be used. The concept of diversity is based on the natural idea that one should
let a single information bit be exposed to dierent channel qualities. Consider for
example the eects of multipath fading, which is varying with both time, frequency
and space (the position of the receiver), see Figure 3.2. With this in mind, it
is natural to categorize diversity into time, frequency, and space diversity. In
addition, polarization diversity is obtained by using clever antenna congurations.
Since diversity methods require the reception of several signals, which have been
transmitted over dierent channels, it is important to have a good algorithm for
how to combine them. This is a major research area, which will not be covered
here. Instead we refer the reader to (Ahlin and Zander, 1998), which also contains
more material on diversity methods.
Next, we describe the dierent diversity methods, with an emphasis on fre
quency diversity, since this will be a main issue in the sequel.
3.6 Diversity 35
Time Diversity
Transmission errors over the radio channel are more likely to occur in groups,
rather than regularly distributed over time. Recall for instance Figure 3.3, where
the signal strength at some times fall far below the average value. During this time
interval we will experience an extra high bit error probability. This bursty nature
of bit errors is troublesome and time diversity can be used to mitigate its eects.
Time diversity can be achieved by using interleaving, which means that the
bits are rearranged before they are sent over the channel, and restored in the
original order at the receiver. Thus, the eect is that the bit errors are likely
to become evenly distributed after this rearrangement, and the error correcting
codes can probably correct the errors, since they usually have been constructed for
independent errors. A consequence of the interleaving is that there will be a time
delay due to the fact that the receiver has to wait until all bits have arrived before
rearranging them.
Frequency Diversity
Frequency diversity can be achieved by transmitting the signals over two or more
frequency carriers. Since the fading is frequency dependent (see Figure 3.2) the
received signals will be more or less correlated depending on how well separated
the carrier frequencies are. Furthermore, one can also say that wideband carriers
yield frequency diversity, since they cover a wide frequency range.
Space Diversity
Space diversity methods utilize the fact that the fading becomes less correlated
at two points the further apart they are. It is not hard to imagine that by using
several antennas, it is unlikely that all of them receive dips in the fading pattern
simultaneously. An extreme case is when the antennas are connected to dierent
base stations, which is called macro diversity. More often an antenna array is used
at each base station (micro diversity).
Polarization Diversity
Polarization diversity is achieved when the antennas are aligned in dierent direc
tions, usually horizontally and vertically. This utilizes the fact that the electro
magnetic waves have dierent reection properties depending on their orientation,
and thus signals from dierent paths become more and more uncorrelated for each
reection.
Frequency Hopping
One of the frequency diversity methods is frequency hopping, which is used in
some of the hybrid FDMA/TDMA systems. The distinguishing feature is that
instead of using the same frequency channel all the time, it is changed at each time
36 Cellular Radio Systems
slot, according to a hop sequence, see Figure 3.5d. Usually frequency hopping is
combined with interleaving, and it is then also a time diversity method.
Frequency hopping falls into two categories, slow frequency hopping and fast
frequency hopping. The former means that there is more than one bit per hop,
i.e. the time it take to transmit one bit is shorter than the time during which the
receiver remains on the same frequency channel. Fast frequency hopping means
the opposite. This is used in some military systems for security reasons, since they
become hard to interfere. The slow frequency hopping on the other hand is used
for its diversity properties, for instance in GSM where it is part of the standard.
The eects of the slow frequency hopping is twofold. Firstly, we get the fre
quency diversity discussed above, and secondly we get interferer diversity. This
is the gain from having dierent cochannel interferers on each time slot, which
implies that the interference is shared between the users, an eect called inter
ference averaging. Furthermore, since information sent during sequential hops is
received with dierent interference, interleaving and coding will improve the trans
mission link quality, if it covers several hops. The interference diversity gain in
GSM is discussed in (Olofsson et al., 1995) .
One distinguishes between cyclic frequency hopping and random frequency hop
ping. The dierence is that in the former, all transmitters hop by using the same
sequence of frequencies, and thus there will be no interferer diversity, since the
receiver will experience interference from the same sources all the time.
DirectSequence CDMA
The eect of DirectSequence CDMA (DSCDMA) is similar to when using fre
quency hopping, but more emphasized. In such a system, the narrowbanded user
signal is spread onto the entire system bandwidth, using individual spreading codes.
The same code is used at the receiver to despread the received signal down to the
original narrowbanded signal. Spreading is implemented as a multiplication by a
code sequence, with a chip rate which is much higher than the symbol rate of the
system. The signals are essentially spread over a bandwidth approximately equal
to the chip rate. Thus frequency diversity is provided. Since all interferers in the
system are spread analogously, interference diversity is obtained. The despreading
also reduce the eect of possible narrowbanded signals transmitted in the same
frequency band. Introducing interleaving and appropriate antenna constellations
we can also obtain time, space and polarization diversity.
4
Radio Resource Management
During the last years, the market of mobile communications has been subject to a
rapid expansion, and as the demand for multimedia service increases, the available
bandwidth resources have to be utilized eciently. Therefore, appropriate Radio
Resource Management (RRM) is of great importance.
When managing the radio resources it is necessary to dene relevant perfor
mance measures and some brief denitions are given in Section 4.1. The available
resource is primarily bandwidth, and algorithms for allocating this resource are
surveyed in Section 4.2. The means to control the resource have to be more for
mally dened. From a discussion in Section 4.3 regarding a simple but yet relevant
quality measure they are dened rather naturally. Among those are the transmis
sion powers, and some aspects of power control is discussed in Section 4.4. Finally
the focus of this thesis is claried in the last section.
Throughout the thesis we will express power levels and gains using dierent
scales. Logarithmic (e.g. dB, dBW, dBm) and linear values are used concurrently.
In order to avoid confusion, we will adopt the convention of indicating linearly
scaled values with a bar. Thus p
i
(t) is a value in linear scale, while p
i
(t) is the
corresponding value in logarithmic scale.
37
38 Radio Resource Management
4.1 Performance of Cellular Radio Systems
Performance is hard to dene in a precise way, but loosely speaking one could say
that it refers to the quality as experienced by the users. With this description in
mind, it is obvious that we need to clarify not only what is meant by quality, but
also who the users are.
One obvious category of users are the subscribers, and their denition of quality
can be summarized as follows:
1. Primarily, the subscriber wants to be admitted to the network. Related to this
is the Grade of Service (GoS), which is dened when planning the network.
It denes the probability of the system being congested, and is typically set
to 2 5%.
2. When admitted, the subscriber requires an acceptable transmission qual
ity. This is naturally related to the nature of the service. Acampora and
Naghshineh (1994) categorize services in three classes. The rst class is
highly delaysensitive realtime trac, which when admitted can be seen as
a circuitswitched connection. This include speech and video. By relaxing
the realtime requirement, the second class is still delaysensitive, but not as
much as for realtime trac. Such services include le transfer and client
server applications. Paging and voice mail are examples of the last and the
most delayinsensitive class. Yet another example is download of old Seinfeld
episodes for viewing at later occasions. Resource allocation for multimedia
services is further studied by Anderlind (1997). In this thesis, the primary
examples are services of the rst class. However, the methods apply to the
other services as well while in transmission, but higher level reasoning may
be dierent.
3. Finally it is important that the subscriber remains connected for as long as
he requires. The process of disconnecting or removing is more infringing to
the user than not being admitted at all.
The second category of users are the operators, who have built their businesses
around the network, and thus it is of primary importance for them that the network
performs well. For an operator quality could mean the following:
1. If the subscribers are content, the operator will presumably get more sub
scribers.
2. Given the network, the operator is interested in serving as many subscribers
as possible.
In reality, a fundamental tradeo is between these two objectives. The goal is to
maximize the capacity, while oering a sucient quality to the subscribers.
4.2 Resource Allocation Algorithms 39
4.2 Resource Allocation Algorithms
In this section we briey survey the area of Resource Allocation Algorithms (RAA).
For a more extensive overview of RAA we refer to (Zander, 1997).
Consider a mobile communication system covering a certain service area as
described in Section 3.5. In this network the M active mobile stations (MS) are
served by B base stations (BS), numbered from the sets
= 1, 2, . . . , M.
B = 1, 2, . . . , B.
The frequency spectrum assigned to the network operator for radio links is divided
into C channel pairs (waveforms), numbered from the set
( = 1, 2, . . . , C.
The process of dividing up the spectrum into channels is based on the multiple
access method chosen, see Section 3.4. We need a pair of channels for each com
munication link, since it consists of an uplink and a downlink. Furthermore, the
channel has to be able to meet the service requirements in terms of data rate and
availability. When a mobile station is admitted to the network a radio link has to
be established. The following has to be assigned to the MS:
A base station from the set B.
A channel pair from the set (.
Transmitter powers for the BS and the MS.
The objective of a resource allocation algorithm (RAA) is to update these assign
ments during the call in order to maintain an acceptable quality while serving as
many users as possible.
The choice of base station mainly depends on the power gain from the dierent
BS:s to the MS. For systems based on orthogonal signals, the MS is often connected
to the BS with the most favorable propagation conditions, usually the one serving
the cell where the MS is located. When the MS is moving, it may cross a cell
border and enter a dierent cell. This triggers a handover or hando algorithm,
and the call is handed over to the BS serving the cell where entered. The situation
is somewhat dierent when considering a system based on nonorthogonal signals.
Then the MS can be connected to several BS:s. When the MS is crossing a cell
border, it is connected to both BS:s while being located near the border. This is
referred to as soft handover. When using sectorized cells, a related scheme is softer
handover, where the MS is connected to several sector antennas at the same BS.
Handover algorithms are surveyed by e.g., St uber (1996).
In the systems of the rst generation, the calls were assigned a xed channel
pair during the call. These xed channel allocation (FCA) schemes are also used
to some extent in the systems of the second generation. The problem is that these
40 Radio Resource Management
channels are usually narrowbanded in frequency, and as pointed out in Section 3.3,
the power gain is frequency dependent. In order to avoid the case when a slowly
moving MS might experience a deep fade in the power gain for a long time (see Fig
ure 3.2), random channel allocation (RCA) schemes are proposed, and the typical
example is the slow frequency hopping in GSM described in Section 3.6. The third
alternative is to use realtime measurements of propagation and/or trac condi
tions to reallocate the spectrum resources dynamically. A comprehensive survey of
these dynamic channel allocation (DCA) schemes as well as the others is provided
by Katzela and Naghshineh (1996).
Finally, the transmitter power levels have to be determined, and this is the
issue in a main part of the thesis. Proposed algorithms to date are surveyed
in Section 4.5. We will separate the power control problem from the other two
assignments, and consider them as given or managed by appropriate algorithms.
However, in previous work, joint approaches are proposed. For instance, Hanly
(1995); Yates and Huang (1995) discuss algorithms that combine power control
and base station selection, and Foschini and Miljanic (1995) provide preliminary
results regarding a combination of power control and DCA.
In a network operating close to its capacity limit, an arriving MS may have
a fatal eect on the established connections. Therefore, the RRM system has
to decide if this terminal should be allowed into the system. Apart from these
admission control algorithms, there is also a need for removal control algorithms,
to determine which MS that has to be disconnected. Generally, the removal process
is considered more infringing to the user, but it may be unavoidable when the
quality of the connection is too low, and eorts to change this situation have
failed. Algorithms for admission and removal control have been studied by Andersin
(1996); Andersin et al. (1996b, 1997) and Bambos et al. (1995).
4.3 Means to Control the Resources
In order to dene the controllable resources more formally and to see how they
aect the transmission quality, it is instructive to start from a simple but relevant
quality measure. The signaltointerference ratio SIR (see Section 3.5) is commonly
used as such a quality measure. Based on its denition, we will try to nd a general
framework which can be used both for orthogonal and nonorthogonal signals. The
framework is to some extent based on (Zander, 1996).
Consider the case of orthogonal signals. Focus on a specic downlink channel
and on the mobile stations and base stations using that channel. The m active
transmitters are using the powers p
i
(t), i = 1, . . . , m. Let i identify each link, so
that e.g., transmitter i is located in the base station to which mobile station i is
connected. Recall from Section 3.5 that the signal from transmitter j to mobile
station i is attenuated by the power gain g
ij
(t). Hence, the SIR at mobile station i
4.3 Means to Control the Resources 41
is given by Equation (3.5)
i
(t) =
i
(t) g
ii
(t) p
i
(t)
j=i
g
ij
(t) p
j
(t) +
i
(t)
p
i
(t) g
ii
(t) +
i
(t)
, (4.1)
where
i
(t) corresponds to the receivers ability to utilize the received desired signal
power. An ecient receiver utilizes all desired signal power and corresponds to
i
(t) = 1. A simplifying assumption is that perceived quality is dependent on
SIR and hence that the quality is acceptable when this ratio is above a certain
threshold, i.e.,
i
(t)
0
, i.
When extending the denition in (4.1) to include all the channels in the system
and to cover both orthogonal and nonorthogonal signals, we obtain an intuitive
denition of the control signals. These relate to the assignments of base stations,
channel pairs and transmission powers, which were listed in Section 4.2.
In addition, we will disclose that with some obvious redenitions, the SIR ex
pression in (4.1) holds for both orthogonal and nonorthogonal signals when soft
handover is not considered. In this thesis we assume that the mobile stations are
connected to one base station only if nothing else is stated. Without loss of gen
erality we can therefore adopt the notation of a downlink situation on a specic
channel in the orthogonal case, when analyzing several aspects of these systems.
Systems Based on Orthogonal Signals
Consider the entire network, and assume that there are B transmitters in base
stations covering the service area, serving the M active mobile stations. Denote
the power gain from transmitter j to mobile station i at time instant t by g
ij
(t).
We gather this information in the Gmatrix of the network,
G(t), dened by
G(t) =
g
11
(t) g
1B
(t)
.
.
.
.
.
.
.
.
.
g
M1
(t) g
MB
(t)
(4.2)
Note that apart from the Gmatrix of a single channel (see Section 3.5), this matrix
is most likely not square, since several transmitters (channels/waveforms) at the
base stations are probably not allocated all the time. Let mobile station i be
connected to transmitter b
i
. Thereby, the mobile station index i identies an
established connection, because of the onetoone correspondence. Furthermore,
dene
ij
as
ij
=
i
(t) g
ibi
(t) p
bi
(t)
M
k=1
g
ib
k
(t) p
b
k
(t)
ib
k
+
i
(t)
p
i
g
ibi
+
i
(t)
. (4.3)
Hence an established connection, identied by the connected mobile k, contributes
to the interference at mobile station i with a signal sent from transmitter b
k
, if
ib
k
is nonzero. The interference at mobile station i can thus be written as a sum over
the M active mobiles.
As mentioned earlier, the uplink is more or less analogous. One notational
dierence is that we have to discuss the SIR at a receiver with respect to a connected
mobile station i. Thus the SIR at receiver b
i
is given by
bi
(t) =
bi
(t) g
ibi
(t) p
i
(t)
M
k=1
g
kbi
(t) p
k
(t)
kbi
+
bi
(t)
p
i
g
ibi
+
bi
(t)
. (4.4)
Note that we have used g
ib
k
to denote the power gain between mobile station i and
receiver/transmitter b
k
both in the uplink and the downlink. In general these are
dierent, and thus the network is not reciprocal.
Adjacent Channel Interference
Even if the signals are orthogonal when transmitted, the radio channel may aect
this orthogonality. Time delays and lack of synchronization in the entire network,
as well as fading and other eects violate the orthogonality. Hence there is an
imminent risk of experiencing interference from adjacent channels at the receiver
as well.
The Gmatrix of the network is still dened as in (4.2). However, the matrix
ij
is dened slightly dierent. In the previous section, the element corresponding to
base station j and mobile station i was equal to zeros when using dierent channels.
In this case it will be dened as a small value less than one describing this leakage
between channels. Therefore the SIR expressions in Equations (4.3) and (4.4) still
holds.
Systems Based on NonOrthogonal Signals
The Gmatrix of the network can be dened as in (4.2). In a typical DSCDMA
setting, each base station (and maybe even the system) utilizes one frequency band,
used by all connected mobile stations. Thus, transmitter b
k
is analogous to base
station b
k
. Furthermore, if we dene
ib
k
by
ib
k
=
ib
k
=
1/N
c
, i = k
0 , i = k
where N
c
is related to the processing gain (PG), which is the ratio between the total
bandwidth and the data rate (Viterbi, 1995). This yields the following expression
for the SIR at base station b
i
bi
(t) =
bi
(t) g
ibi
(t) p
i
(t)
1
Nc
k=i
g
kbi
(t) p
k
(t) +
bi
(t)
p
i
g
ibi
+
bi
(t)
. (4.5)
One important dierence between the uplink and the downlink is that the or
thogonality is dierently aected by the channel. In the downlink, the part of the
interference at the mobile station that originates from the connected base station is
more or less orthogonal, since the waveforms are transmitted over the same chan
nel. In the uplink, this is not the case. The interfering waveforms from mobile
stations in the same cell are transmitted over dierent channel. Therefore, the
uplink situation is more critical. This is, however, reected in the s, but they are
dierent in the uplink and the downlink.
Theoretical Interpretation of the Control Signals
In Equation (4.3) we disclosed the following general formulation of the SIR for the
downlink
i
(t) =
i
(t) g
ibi
(t) p
bi
(t)
M
k=1
g
ib
k
(t) p
b
k
(t)
ib
k
+
i
(t)
p
i
g
ibi
+
i
(t)
, (4.6)
and from Equation (4.4), we see that the formulation for the uplink is essentially
analogous. It is instructive to relate this expression to the assignments performed
by a radio resource algorithm (see Section 4.2). Clearly the base station assignment
corresponds to a choice of b
i
:s and the power control amounts to the choice of p
i
and p
bi
. Finally, the channel allocation or the choice of waveforms or codes is
reected by
ij
.
Furthermore assume that the base stations and channels are assigned. Then we
can redene g
ij
by
g
ij
:= g
ibj
ibj
.
The motivation for redening g
ij
instead of using a new notation is to correlate
with the common notation used in the eld. Based on this denition, the SIR
is given by a similar expression as in the orthogonal and single channel case in
Equation (4.1). Therefore, when studying power control eects to the entire cellular
radio system, we can without loss of generality assume that the system is based
on orthogonal signals, and consider only the connected transmitters/receivers on a
certain channel.
44 Radio Resource Management
4.4 Aspects of Power Control
Let us assume that the appropriate base stations are assigned and channels allo
cated. Then the remaining problem is to update the output powers of the transmit
ters. This power control problem can be concerned with the following standard
block diagram if the signals are interpreted as below (with respect to mobile i).
u
w
y
z
Plant
Power
Control
Figure 4.1 The power control problem formulation.
The external disturbances w comprise the power gain g
ii
and the interference
I
i
(including the thermal noise
i
).
The control signal u is the power level p
i
.
The measurements available to the controller are represented by y. Normally
they consist of a quality related and/or a received signal strength related
measure.
Finally z represents an adequate quality measure. Reliable measurements are
most likely not available on a short time frame.
The objective is to assign power levels so that z meets the quality specications
z = z
t
(or z z
t
), with respect to the adequate quality measure. Additionally, it
is desirable to provide sucient quality to as many users as possible, in order to
maximize the capacity of the network.
When characterizing the necessary power control algorithms, we nd the fol
lowing aspects interesting:
Capacity and System Load. Only a limited number of mobile stations can
be accommodated in a cellular radio system. It is thus relevant to describe
the system load in a particular situation and also to relate it to the maximal
capacity of the network. Moreover, the cost to accommodate a particular
mobile varies from mobile to mobile.
Centralized/Decentralized Controller. A centralized controller has all
information about the established connections and power gains at hand, and
4.4 Aspects of Power Control 45
controls all the power levels in the network. A decentralized controller is only
controlling the power of one single transmitter, and the algorithm is only
relying on local information. The latter case is the only practical one, since a
centralized controller requires extensive control signaling in the network, and
will suer from additional time delays.
Note that the distinction is primarily between the type of information used
when computing the output powers. Even if an algorithm is considered to
be decentralized, it is not necessarily physically located in the mobile sta
tion. Instead, the power levels may be computed in the base stations, and
distributed to the corresponding transmitters. This masterslave relationship
facilitates software updates and support.
Global Stability and Performance. Interconnecting several distributed
power control algorithms aect properties of the overall global system, such
as stability and performance. The relations between local and global perfor
mance has thus to be established.
Quality Measure. Speech quality is a very subjective quantity, and rele
vant quality measures are dierent from measures applicable to data trac.
Therefore, the appropriate quality measure is service dependent.
Feedback Bandwidth. The feedback can be categorized by the necessary
bandwidth to transmit the information to the control algorithm. Typically
one diers between on the one end information feedback, where basically real
values are assumed available and on the other end decision feedback, where
only whether a value is above or below a threshold is assumed available.
These are the extreme situations and feedback schemes in between are also
implemented. In TDMA systems of the second generation, the feedback is
provided in measurement reports comprising a Quality Indicator (QI), re
ecting the quality and a Received Signal Strength Indicator (RSSI), reect
ing the received signal strength at the receiver. These values are coarsely
quantized in order to use few bits. Thus an important question is to deter
mine which quantities that can and should be estimated given the available
measurements.
Constraints. The output power levels are limited to a given set of values
due to hardware constraints. This includes quantizing and the fact that the
output power has an upper and a lower limit. Additionally, the various stan
dards include dierent constraints. For instance in GSM, there are channels
in the network, requiring the use of maximum power, and in DAMPS all
three time slots on the same carrier have to use the same power level.
Time Delays. Measuring and control signaling takes time, which results
in time delays in the network. These are primarily of two kinds. Firstly it
takes some time to measure and report the measurements to the algorithm,
resulting in a time delay of n
m
samples. Secondly we have a time delay of
46 Radio Resource Management
n
p
samples due to the time it takes before the computed power is actually
used in the transmitter. These time delays should be considered as additional
time delays, since the discretetime controller is assumed to always provide a
delay of one sample. As an example consider WCDMA (UMTS 30.06, 1997),
where each slot contains a power control command. The duration of a slot
is 1/1500 s, and consider the power update instants t as time reference. The
power computed at time instant t 1 is actually reected in measurements
not earlier than at time instant t. This measurement is used to update the
power at time instant t + 1. In this typical case n
p
= 1 and n
m
= 0.
Controller Bandwidth. Clearly, not all possible timevarying disturbances
can be compensated for using power control. The controller bandwidth
describes which variations that can be mitigated and which that can not.
Clearly, this is related to channel characteristics, modulation scheme, link
orientation (up or downlink), mobile velocities and the update rate of the
power control algorithms.
The power control strategy is naturally implemented by the cascade control
structure depicted in Figure 4.2. Relevant information is extracted from the (pos
sibly delayed) measurement y by the estimating device F. An outer control loop
computes a target SIR which on the medium time scale corresponds to the data
rate of the users service. The inner loop assigns transmission powers to track the
provided target SIR while compensating for disturbances w. Typically, the update
rate of the inner loop is an order of magnitude faster than the outer loop.
u
w
t
Environment
F
Inner
Loop
Outer
Loop
Figure 4.2 Cascade structure of the controller, where a fast inner loop
issues transmission powers to track a target value provided
by a slower outer loop, and to compensate for disturbances.
Relevant information is extracted from the measurements by
the device F.
Algorithmic properties are naturally studied at two levels. Capacity, load and
whether it is possible to accommodate all users with their data rate requirements
(reected by the SIR targets) are typically discussed at a global level. We assume
that the inner control loops successfully compensates for the fast variations, so that
4.5 Survey over Proposed Power Control Algorithms 47
the power gains can be considered as constant. Moreover, stability and convergence
are also discussed at a global level. Again, the power gains are assumed constant,
but the motivation is that we are considering the short term behavior of the global
system, and the power gains are assumed to uctuate more slowly. Conversely on
a local level, target SIR:s are assumed constant and that it is possible to accommo
date all users. Then the tracking capabilities are in focus in terms of inner control
loop bandwidth and stability.
4.5 Survey over Proposed Power Control
Algorithms
The ambition is not to provide a complete coverage of the area, but rather to
outline some important and central contributions. As in the previous discussion,
index i refers to a specic connection, and that the same algorithm is employed for
every connection. Some algorithms include design parameters, and these will be
denoted by Greek letters , etc.
Power Control to Improve Link Performance
In early wireless pointtopoint links, fading was primarily seen as a time varying
power gain g(t) (see Figure 4.3a). Power control was essentially used to compen
sate for the varying channel and the thermal noise power v(t). In other terms,
the control objective was to issue transmission powers p(t) to ensure a sucient
signaltonoise ratio (SNR) at the receiver (Ekberg and Laurent, 1964). The ideal
transmission power would be to invert the channel and to use a bias to fulll the
SNR requirement as given by the solid line in Figure 4.3b. This strategy is trouble
some to implement in practice, primarily due to varying channels and to inaccurate
and/or infrequent channel estimates. A possible solution would be to use excess
power as given by the dashed line in Figure 4.3b. The drawbacks include increased
energy consumption and disturbance to others if present.
Fundamental limits for digital communication systems were established by
Shannon (1948a,b, 1956). He formulated the basic problem of reliable transmission
of information in statistical terms, and demonstrated how constraints and noise
can be associated with a channel capacity. If the information rate from the source
is less than the channel capacity, then it is theoretically possible to achieve reliable
(errorfree) transmission through the channel by appropriate coding.
It is a wellknown fact that feedback can be used to improve stability and
performance. Normally the feedback is categorized by its information content.
Schalkwijk (1969) showed that the achieved performance using information feedback
is better than when using decision feedback, which in turn yields better performance
than when not using feedback at all. The cost is increased feedback bandwidth.
As an alternative to power control, Cavers (1972) discusses variable data rate
in the transmission. With constant average power, the proposed scheme adjusts
48 Radio Resource Management
Channel gain [dB]
Transmission power [dBW]
Time [s]
a.
b.
Figure 4.3 A natural approach to compensate for a timevarying channel
as in a) is to invert the channel. This corresponds to using the
power as described by the solid line in b). An alternative is to
use excess power as given by the dashed line in b).
the data rate to minimize the average error probability. Moreover, Cavers discusses
the eect of feedback delay, which is seen to aect the performance signicantly.
Srinivasan (1981) relaxed the assumption of a known channel, and included a
pilot symbol of constant power to the transmitted data symbol. The pilot symbol
was then used to estimated the channel. A generalization to a frame of data
symbols including one pilot symbol is provided by Saarinen and M ammela (1999).
In addition, the eect of channel coding is studied. Essentially, the purpose of
these minimum meansquare error (MMSE) receivers, is to nd the optimal linear
receiver that maximizes SNR.
Goldsmith (1997) proposed an optimal adaptive transmission scheme, which
achieves the Shannon capacity of a fading channel. This scheme is based on known
channel parameters, and utilizes both variable power and rate. The power adaption
uses more power when the channel is favorable, and conversely less power when
the channel is not as good. If the quality of a particular channel is below some
threshold, no power is allocated to that channel. It was also noted that the capacity
dierence between using variable power and constant power was neglectable for
most types of fading, when rate adaption is applied.
4.5 Survey over Proposed Power Control Algorithms 49
Power Control in Cellular Systems Based on Received Signal Power
The eect of controlling the power is more emphasized, however, when consider
ing multipleaccess systems. Then, the capacity of the overall network is in focus.
Some proposals concentrate on controlling the powers based on the received signal
strength. These include (Fujii and Sakamote, 1988; Nagatsu et al., 1983; Tchirks,
1989). Whitehead (1993) proposed a static control law to minimize the received
signal power variance under certain conditions. The scheme was rened by Carne
heim et al. (1994) and a static quality adaption was added .
Channel inversion was discussed above, and one algorithm with this objective
is the Constant Received Power (CRP) algorithm (Anderlind, 1997)
p
i
(t + 1) = p
i
(t)
C
t
C
i
(t)
, (4.7)
where
C
t
is the target signal power which the desired signal power
C
i
= p
i
g
ii
strives towards. Anderlind (1997) also provided a natural extension to handle
output power constraints. The index i relates to the transmitterreceiver pair i,
and this will be the convention throughout the rest of the section.
As pointed out in Section 3.5, tampering the nearfar eect in DSCDMA sys
tems is crucial for good performance. One possible solution is to control the powers
of the mobile stations so that the powers received at the connected base station
are all the same. The algorithm implemented in the IS95A standard (Adachi,
1997; Ariyavisitakul and Chang, 1993; Chang, 1996; Chang and Wang, 1996) can
be described by
p
i
(t + 1) =
i
p
i
(t),
C
i
(t)
C
t
i
1
i
p
i
(t),
C
i
(t) >
C
t
i
,
i
1 (4.8)
Since only one information bit is needed for signaling, the scheme is bandwidth
ecient, which enables the use of a high update frequency (800 Hz in IS95A).
As discussed above, the SNR is related to perceived transmission quality. In
cellular systems, not only thermal noise but also interfering signal power from other
users limit the performance and the capacity. A related problem is therefore to
assign appropriate target values
C
t
i
for these algorithms to control the interference
between cells (Salmasi and Gilhousen, 1991; Viterbi, 1995).
Centralized Power Control in Cellular Systems Based on SIR
As concluded by Ariyavisitakul (1994), a more ecient way to conduct power con
trol in cellular radio systems is to utilize the signaltointerference ratio at the
receivers. Thereby, both the nearfar problems and control of multiple access in
terference are addressed.
If all the signal power gains in the Gmatrix (see Section 4.3) were known,
the transmission powers of every transmitter could be computed in a centralized
fashion. Aein (1973) focused on satellite communications systems, and introduced
50 Radio Resource Management
the term SIR balancing, for a strategy of power control aiming for all users to have
the same SIR. The proposed algorithm was based on solving an eigenvalue problem.
Alavi and Nettleton (1982); Nettleton (1980); Nettleton and Alavi (1983) ex
tended these results and applied them to spread spectrum cellular radio systems.
This approach was further rened by Grandhi, Vijayan, Goodman, and Zander
(1993).
Zander (1992b) interpreted these algorithms as optimal solutions in the sense
that there exist no other power vector yielding a higher SIR
.
Thermal noise was not included in the analysis. However, in the noisy case, we
note that the inuence of the noise can be made arbitrarily small by choosing k
suciently large. A consequence is that the optimal balanced SIR in the noiseless
case serve as a bound for balanced SIR:s even in this case.
These results hold under the assumption that the receiver can utilize all the de
sired received signal power
C
i
(t). If the receiver only can utilize the fraction
i
(t) of
that power, the remainder acts as interference, autointerference (see Section 3.5).
Godlewski and Nuaymi (1999) proved that the autointerference will result in a
lower optimal balanced SIR.
In a practical case, a specic balanced target SIR,
t
at each receiver may be
desirable. This may be relevant if the objective for example is to provide speech
services of the same quality to all users. Zander (1993) proposed a centralized
algorithm to determine the corresponding balancing power vector. Moreover, this
is proven feasible if and only if
t
<
.
In the analyses above, it has been assumed that the transmitters can use any
power. In practice there are constraints on the power levels. A centralized scheme
for power control with upper bounds on the maximum transmitter power was dis
cussed by Grandhi et al. (1995).
Decentralized Power Control in Cellular Systems Based on SIR
The foundations for the distributed algorithms to be described were laid by Mey
erho (1974). The motivation in this article was not to obtain a distributed al
gorithm, but rather to make Aeins centralized algorithm more computationally
ecient. The proposed algorithms were based on results for iterative computa
tions of eigenvectors (Fadeev and Fadeeva, 1963; Pipes and Hovanessian, 1969).
Similar ideas appeared in an article by Zander (1992a), but this time applied
to cellular radio systems. Here the Distributed Balancing (DB) algorithm was
suggested:
p
i
(t + 1) =
p
i
(t)
1 +
1
i
(t)
. (4.9)
This result was also inspired by numerical methods. The DB algorithm is essen
tially the power method for nding the dominant eigenvalue and its corresponding
4.5 Survey over Proposed Power Control Algorithms 51
eigenvector (Carnahan et al., 1969), which directly relates it to the centralized al
gorithms discussed above. If the thermal noise is neglected, it can be shown (see
Zander, 1992a) that the DB algorithm converges to the balancing power vector,
which corresponds to
.
At rst sight the DB algorithm appears to be distributed, since it is based only
on
i
(t), which is a measurements of the SIR at the receiver. However, it turns
out that the choice of is problematic, since it may make p drift towards zero or
innity, if not appropriately chosen. This choice of
can be seen as a normalization
procedure, and it must be based on global information. Thus this is not a fully
decentralized algorithm.
The Distributed Power Control (DPC) algorithm, which is a slight modication
of the DB algorithm, was suggested by Grandhi, Vijayan, and Goodman (1994)
p
i
(t + 1) =
p
i
(t)
i
(t)
. (4.10)
When neglecting the thermal noise, the DPC algorithm was also shown to converge
to p
, and simulations indicated that the convergence was faster than when using
the DB algorithm. This is also shown analytically in (Zander, 1993). Furthermore,
simulation results presented in (Lee et al., 1995) show that the DPC algorithm
results in better performance than the DB algorithm. Nevertheless, the problems
of normalizing the powers by cleverly adapting
still remained. Yet another algo
rithm, similar to those described above was proposed by Lee and Lin (1996); Lin
(1995)
p
i
(t + 1) =
p
i
(t)
1
i
(t)
. (4.11)
This algorithm was referred to as the CDPCII algorithm and it also suers from
the problem about how to choose
.
The drawback with these distributed algorithms is that they are not fully decen
tralized. The normalization procedure requires signaling between the algorithms,
which is undesirable in practice. A turning point was when Foschini and Miljanic
(1993) proved convergence of the DPC algorithm in (4.10) in the noisy case. They
pointed out that the normalization procedure was unnecessary, when taking noise
into consideration. Instead the
could be seen as the target SIR, which the al
gorithm strives towards, if this is feasible (i.e.,
<
). Hence it is instructive to
rewrite the algorithm as
p
i
(t + 1) = p
i
(t)
t
i
(t)
. (4.12)
This will be denoted the DPC algorithm in the sequel.
In (Foschini and Miljanic, 1993), it was assumed that all transmitter powers are
updated synchronously in the network. This is not very realistic, and Mitra (1994)
showed that the result still holds if the powers are asynchronously updated.
52 Radio Resource Management
With the DPC algorithm as the base, extensions have been proposed when
considering systems with output power constraints. Grandhi, Zander, and Yates
(1995) proposed the Distributed Constrained Power Control (DCPC) algorithm,
and proved its convergence. The algorithm is given by
p
i
(t + 1) = min
p
max
, p
i
(t)
t
i
(t)
. (4.13)
It addresses systems where the output power is bounded from above by p
max
. If
the powers are bounded from below by p
min
(0), the scheme should be updated
accordingly. Additional complexity arises when considering systems where the out
put powers are quantized. Andersin, Rosberg, and Zander (1996a, 1998b) propose
the Distributed Discrete Power Control (DDPC) algorithm as a possible solution.
Their results intimate that there is an imminent risk of an oscillative behavior in
the system. The proposed solution incorporates some signaling between the dis
tributed algorithms, and is therefore not fully decentralized. An algorithm aimed
at the same problem is proposed in (Wu and Bertsekas, 1999), but with the restrict
ing assumption that global information is available at each transmitter to update
the powers distributedly.
This far primarily values in linear scale are utilized. Yates used values in loga
rithmic scale in the following algorithm (Yates, 1995)
x
i
(t) =
t
+p
i
(t)
i
(t)
p
i
(t + 1) = p
i
(t) + (1 )x
i
(t) , 0 1. (4.14)
Yates proved its convergence to
t
if feasible (
t
<
i
p
i
(t),
i
(t)
t
i
1
i
p
i
(t),
i
(t) >
t
i
(4.15)
Again, the benets include the low signaling bandwidth, which enables a high up
date frequency. This algorithm will be used in the emerging WCDMA standard,
with an update frequency of 1500 Hz (Dahlman et al., 1998; UMTS 30.06, 1997).
This algorithm will be referred to as the Fixedstep power control (FSPC) algo
rithm. The step size
i
can be adapted as well, normally at a much lower updating
frequency than the power control loop. Such algorithms will be denoted Adaptive
step power control (ASPC) algorithms. When in soft handover the mobile station
receives two possibly dierent control commands to update its uplink power. Then
the mobile adjusts the power with the largest commanded downward step (dierent
step sizes from the dierent base stations are possible).
The convergence analyses so far are based on an assumption of a constant
Gmatrix or at least subject to only neglectable variations with respect to the
4.5 Survey over Proposed Power Control Algorithms 53
time constants of the inner control loops. The case of a shadow fading Gmatrix
is studied by Andersin and Rosberg (1996), and the recommendation is that the
quality requirements should be multiplied by
const
a constant greater than one.
If the minimum required SIR at the receiver is given by
0
, the target SIR should
be chosen as
t
=
0
const
. (4.16)
Guidelines for an appropriate choice of the constant are also provided.
As with the algorithms based on received signal power, a related problem is to
provide a relevant target SIR. This is addressed in (Kawai et al., 1999; Sampath
et al., 1997; Won et al., 1998). An intuitive crucial problem is to issue target
SIR that corresponds to a feasible power control problem. If the value is set too
high, it is not possible to support all the mobiles. Instead the algorithms will
start competing to obtain the target SIR. This completion will last until several
transmitters are using maximal powers, still without achieving the target SIR at
the corresponding receiver. A proposal to counteract this partyeect and to
employ graceful degradation in the systems is described by Almgren et al. (1994).
The main idea is that a user with favorable propagation conditions may use a
higher power, and vice versa. Using logarithmic values, the algorithm is given by
p
i
(t + 1) = (
i
(t) p
i
(t)). (4.17)
A convergence proof of the algorithm in (4.17) is disclosed in Yates et al. (1997).
The result also holds when considering bounded output powers. For clarity, the
algorithm will be referred to as the AAW algorithm.
In real systems, SIR is not readily available. Ulukus and Yates (1998) based
the transmission power on the matched lter outputs. A probabilistic framework
as used to address convergence in probability. Other approaches including (Man
dayam, 1999; Manzanedo et al., 1996) are using the signals as standardized by the
link protocol.
As briey discussed in Section 4.2, some proposals aim at jointly control the
power and allocate the other resources. Similarly, some proposals aim at joint
power control and beamforming of antenna arrays (RashidFarrokhi et al., 1998;
Yener et al., 1999). The latter proposal also incorporate multiuser detection.
54 Radio Resource Management
5
Local Analysis
As disclosed in Section 4.4, it is desirable to employ distributed algorithms in the
network, due to the extensive control signaling involved with centralized schemes.
The only way to fully analyze the behavior of these algorithms, is to study the global
system and consider all the interconnections between the distributed algorithms.
However, valuable information can be obtained by studying these distributed algo
rithms separately and locally.
In this chapter, values in logarithmic scale will be used almost exclusively. This
leads to a loglinear model of the most interesting power control algorithms. A
general distributed control algorithm for mobile station i can be described by the
block diagram in Figure 5.1. It describes a general SIRbased power control algo
rithm, but as discussed in Section 5.2, the dynamics when using algorithms based
on received signal power are covered as well. The available measurements (possi
bly delayed)
i
(t n
m
) are processed by the ltering device F
i
, which also can be
seen as a model of an estimator. Based on a target value
t
i
(t) and a processed
measurement
i
(t), the control algorithm R
i
issues control commands u
i
(t). These
commands may be a single bit (decision feedback) or an analog value (informa
tion feedback) or anything in between. The channel may delay the commands or
corrupt them by a disturbance x
i
(t), modeled as additive. In the decision feed
back case, it manifests itself as command bit errors and in a practical case close
to information feedback, it may represent quantization noise. On the transmitter
side, the (possibly delayed) control command x
i
(t n
p
) is decoded by the device
55
56 Local Analysis
p
i
(t)
t
i
(t)
i
(t)
R
i
D
i
g
ii
(t) I
i
(t) x
i
(t)
u
i
(t)
n
p
T
n
m
T F
i
Receiver Transmitter
Figure 5.1 Block diagram of the receivertransmitter pair i when employ
ing a general SIRbased power control algorithm. In operation,
the controller results in a closed local loop. Time delays are
expressed in number of samples with respect to the sampling
interval T.
D
i
to the updated output transmission power p
i
(t). Note that in the typical infor
mation feedback case, D
i
is essentially a represented by a time delay. Therefore,
the quantization noise x
i
can without loss of generality be included in the additive
disturbance g
ii
(t) I
i
(t). The relation between the transmitted power and SIR
(neglecting autointerference) in Equation (3.5) is given in logarithmic scale by
i
(t) = p
i
(t) +g
ii
(t) I
i
(t). (5.1)
The physical location of devices R
i
, D
i
and F
i
and their interpretation are irrele
vant from a dynamical point of view. To emphasize the locality of the analysis, a
control algorithm in operation as in Figure 5.1 will be referred to as a local con
trol loop or just local loop. The local loops are interconnected via the interference
I
i
(t). In local loop analysis, we make the simplifying assumption that I
i
(t) is a
disturbance independent of p
i
(), t. In the case of power controlled cellular
networks, the interference at one receiver is a strictly increasing function of the
powers used by others. Therefore a competition between the users can be ob
served, and we can argue that local instabilities are not compensated for by the
interconnections. Hence, local stability is a necessary, but not sucient condition
for global stability. The dynamical eects of the interconnections are studied more
thoroughly in Chapter 7. An important conclusion is that local stability and some
additional requirements yield global stability.
To further motivate the local analysis, some introductory examples are provided
in Section 5.1. Section 5.2 focus on the components of the local loops. These are
identied as power control algorithms, time delays, static nonlinearities and lters.
Several interesting power control algorithms are naturally categorized as either
5.1 Motivating Example 57
linear or linear with static nonlinear components. For example some algorithms
using decision feedback belong to this category. Linear algorithms are analyzed
with respect to stability using root locus analysis in Section 5.3. As seen in the
following section, linear systems with static nonlinear components, often result in
an oscillatory behavior. For that purpose, describing functions are introduced in
Section 5.4 and applied in Section 5.5. As pointed out above, autointerference
is neglected throughout almost the entire chapter. The main motivation is that
its eect on stability can be treated together with the interconnections, as will
be further discussed in Chapter 7. However, some brief results are provided in
Section 5.6, where we conclude that autointerference does not violate stability of
linear algorithms nor aggravate oscillations of the FSPC algorithm.
5.1 Motivating Example
To further motivate local loop analysis, it is interesting to study the eect of a delay
of a single sample. The following example illustrates the behavior when employing
the DPC algorithm in (4.12) and the FSPC algorithm in (4.15)
Example 5.1 (Single Delay in the Closed Local Loop)
Consider the situation in Figure 5.2, where four mobile stations are connected
Figure 5.2 Conguration in Example 5.1, where four mobile stations ini
tially are allocated to the same channel. At time instant t = 0
a fth mobile station (indicated by the arrow) is allocated to
the channel.
58 Local Analysis
10 5 0 5 10 15 20 25 30 35 40
8
10
12
14
16
10 5 0 5 10 15 20 25 30 35 40
8
10
12
14
16
i
(
t
)
[
d
B
]
i
(
t
)
[
d
B
]
time instants
a.
b.
Figure 5.3 The recovering ability of a. DPC and b. FSPC algorithms
when subject to the situation in Figure 5.2 and a time delay of
one sample.
using the same channel. The simulation is initiated at time instant t = 10,
and the four mobile stations are initially using the optimal transmission powers
corresponding to a target SIR of 12 dB. The signal power gains g
ij
are constant
throughout the simulation. At time instant t = 0, the same channel is allocated
to a fth mobile station, initially using a transmission power of 1 dBW. The
target SIR is still feasible in the ve mobile station case. Power control is
employed as either DPC described by (4.12) or FSPC (step size =1 dB) given
by (4.15). It is illustrative to see how these algorithms recover in the changing
environment when subject to delayed power control commands. In this case,
the power control command (FSPC) or the transmission power itself (DPC) are
delayed by one sample interval.
The performance using the two algorithms is depicted in Figure 5.3. We note
that DPC is unstable in the sense that the oscillations in SIR are aggravated
over time. Up to t = 0, the powers are not updated since the target SIR and
the measured SIR are identical. The system is thus at rest in an equilibrium
point. Then, the admitted mobile increases the interference at each receiver.
This disturbance reveal the instability of the equilibrium point, and the SIR
uctuations get worse over time.
Despite being perfectly initialized, FSPC results in oscillations in the case of
four mobiles. That is due to the decision feedback, where the powers are either
increased or decreased each time instant. The oscillations are not that much
aected by the entering mobile station. However, one might expect that this
updown device would result in small oscillations (approximately 1 dB peak
topeak) around the target value. As seen in the plot, the amplitude is larger
(3 dB peaktopeak) than expected.
5.2 Dynamical Models 59
Some central questions are risen by the example.
Why is DPC unstable, when subject to time delays?
In what way does the time delays aect stability?
Why are the oscillations using FSPC worse than expected, when subject to
time delays?
One of the objectives with this chapter is to address these questions, using meth
ods from control theory. This example is reconsidered in Section 10.2.2, using
controllers designed in Chapter 6.
5.2 Dynamical Models
The local loops mainly comprise four components: power control algorithms, time
delays, static nonlinearities and lters. These are modeled in the following subsec
tions. The models relate primarily to two important feedback situations
Information feedback. The mobile feeds back the exact SIR measurements
or the error
e
i
(t) =
t
i
i
(t). (5.2)
The important characterization is essentially that real numbers are fed back.
Decision feedback. In a power control setting, we typically associate this
with feedback of the sign of the error in (5.2)
s
i
(t) = sign(
t
i
i
(t)) = sign(e
i
(t)) (5.3)
Thus, only one bit is needed for command signaling, which makes the scheme
bandwidth ecient.
In practice, the assumption of real valued measurements of information feedback
is unrealistic. Instead, quantized measurements can be made available. This is
further discussed in Chapter 8. However, the two cases describes two important
extremes, and constitute the situations considered in this and subsequent chapters.
5.2.1 Power Control Algorithms
Numerous algorithms have been proposed, and as the number of proposed algo
rithms increases, the picture tends to get more and more blur. We will focus on
some important fully distributed algorithms from Section 4.5, and see how the
algorithms relate to the following algorithm, based on information feedback
p
i
(t + 1) = p
i
(t)
t
i
(t)
i
(t)
. (5.4)
60 Local Analysis
Note that individually time varying target SIR:s
t
i
(t) are enabled. If the update
frequency of
t
i
(t) is much slower than the time constants of the control algorithm,
it can be considered as constant in the analysis. It is instructive to express this
control law in logarithmic scale.
p
i
(t + 1) = p
i
(t) +
t
i
(t)
i
(t)
(5.5)
This controller is identied as an integrating (I) controller.
The CDPCII algorithm proposed by Lee and Lin (1996); Lin (1995) and given
by (4.11) can be rewritten as
p
i
(t + 1) = p
i
(t)
1/
i
(t)
.
After some redenitions of controller parameters, the similarity to (5.4) using a
constant
t
is obvious. Similarly, the DPC algorithm (Foschini and Miljanic, 1993)
in (4.12)
p
i
(t + 1) = p
i
(t)
t
i
(t)
is obtained as the special case = 1 and constant
t
.
The two equations in (4.14) adopted from Yates (1995) compile to
p
i
(t + 1) = p
i
(t) + (1 )(
t
i
(t)),
which is the same as in (5.5) if the controller parameter is redened and a constant
target value is utilized.
The AAW algorithm proposed by Almgren, Andersson, and Wallstedt (1994)
is also related. From (4.17), we get
p
i
(t + 1) = (
i
(t) p
i
(t)) = p
i
(t) +
/
1
p
i
(t)
i
(t)
(5.6)
Thus it can be seen as the algorithm in (5.5) with a power level dependent target
SIR. This algorithm is further discussed in Section 9.4.3. The local dynamics are
more emphasized in the following form
p
i
(t + 1) = p
i
(t) +
t
i
i
(t)
, (5.7)
where
t
i
= / is a design parameter, possibly provided by an outer control loop.
Using decision feedback, the control error e
i
(t) =
t
i
(t)
i
(t) is not available,
but rather the sign of the same. The decision feedback algorithm FSPC in (4.15)
can be rewritten in logarithmic scale as
p
i
(t + 1) = p
i
(t) +
i
sign
t
i
(t)
i
(t)
, (5.8)
which illuminates the similarity to (5.5) where information feedback is available.
5.2 Dynamical Models 61
The focus in this chapter is on SIRbased power control algorithms. However,
most interesting algorithms based on desired signal power C
i
(t) are covered as well.
This hold true since the feedback signal in that case can be written as
C
i
(t) =
i
(t) +I
i
(t)
i
(t).
The interference I
i
(t) is considered as an independent disturbance to the local loop.
Therefore, the closedloop dynamics of the local loops are identical for SIR based
and C based algorithms. Hence, the local loop dynamics using CRP in (4.7) and
DPC are identical. Same thing holds for the local loop dynamics using the decision
feedback algorithm in (4.8) and the FSPC in (5.8). Thus the algorithms in (5.5),
(5.7) and (5.8) represent the mainstream of algorithms and will receive specic
attention.
The Icontroller and FSPC with integral action are only plausible when the
power control problem is feasible, i.e., when it is possible to nd transmission
powers to meet the desired target values. Therefore, these algorithms need this
underlying assumption. The feasibility assumption can be relaxed in case of the
AAW algorithm, since it is more or less designed to handle infeasible cases. The
assumptions do not aect the local loop analysis, but relates to global stability.
5.2.2 Time Delays
Both measuring and signaling in cellular systems take time, which results in delayed
signals. As pointed out in Section 4.4, there are three main reasons for time delays.
Firstly, the power control algorithm R
i
is assumed to result in a time delay of one
sample, since measurements at time t (e.g.,
i
(t)) are used to update the power level
at time t +1 (i.e., p
i
(t +1)). Secondly it takes some time before a computed power
level actually will be used and thereby observed by others. Additional delays are
caused by the fact that power update commands are only allowed to be transmitted
at certain time instants. Together they result in a total delay of n
p
samples.
Finally, the measuring procedure takes time, and again these measurements are
only reported to the power control algorithm at certain time instants, resulting in
a delay of n
m
samples. In total there is an additional delay of n = n
p
+n
m
samples
in the local loops. A typical delay situation is described in the following example.
Example 5.2 (Typical Delay Situations in WCDMA)
Since the command signaling is standardized, the delays are known exactly in
number of samples (or slots). Typical situations in WCDMA are depicted in
Figure 5.4. In a) the receiver estimates the SIR
i
(t) (possibly only the desired
signal power, which is combined with a separate interference measurement to
compile SIR) over some pilot bits and maybe data symbols, on a slot sent using
the power p
i
(t 1). The power control command (TPC) is included in a slot
in the opposite direction at time instant t. The transmitter nally updates the
power p
i
(t + 1). By shifting the slot synchronization as in b), the transmitter
power can be controlled without excessive loop delay. The drawback is that
62 Local Analysis
fewer bits for the SIR estimation is used, resulting in a larger variance of the
estimation error (see also Section 6.1.4). This is only possible for mobile stations
relatively close to the base station, and n
p
= 1, n
m
= 0 will be considered as a
typical WCDMA situation.
Transmitter
Transmitter
Receiver
Receiver
Update
Update
Est.
Est. TPC
TPC
t + 1
t + 1
t
t
t 1
t 1
a)
b)
Figure 5.4 Typical time delay situation in WCDMA. The transmission
powers are updated at time instants t (once per slot). Esti
mation is performed over the rst symbols of the slot (pilot
symbols and possibly some data symbols) Adachi et al. (1998).
Command signaling and power updates take place as indicated.
a) n
p
= 1, n
m
= 0, b) By shifting the slot congurations, n
p
can be brought down to zero when the mobile station is close
to the base station.
The same delay situation as in the example also applies to GSM. Systems subject
to longer delays include IS95A with n
p
= 2, n
m
= 0 (TIA/EIA/IS95, 1995),
and dierent satellite communication systems, which may inherit even longer time
delays (Mehta et al., 1998).
The time interval over which measurements are collected is normally a fraction
of the power control update interval, as illustrated by the example above. It might
even be timevariant. Therefore, it is natural to dene time instants t equal to the
update instants of the transmission power, which are xed and can be associated
with the power control update interval T
s
.
Time delays are conveniently represented using the timeshift operator q dened
by
q
n
p(t) = p(t n), q
n
p(t) = p(t +n)
5.2 Dynamical Models 63
To stress the direction of the time shift, the terms forwardshift operator and
backwardshift operator (or delay operator) are used for q and q
1
respectively.
Arithmetic operations of polynomials in q will be used frequently in this and sub
sequent chapters. For a more rigid discussion on a qoperator algebra, the reader is
referred to (
+
p
i
(t)
i
(t)
I
i
(t)
C
i
(t)
g
ii
(t)
F
I
(q)
F
g
(q)
Receiver
Figure 5.5 When the signal and the interfering powers both are available,
separate ltering could be employed.
5.3 LogLinear Algorithms 65
5.3 LogLinear Algorithms
In several cases, the nonlinearities in the local loops can be neglected, and thus
the remaining system to be analyzed is linear. This is illustrated by Example 5.3.
Local loop stability can therefore be addressed using linear systems theory methods.
In this section we apply root locus analysis, which was applied to power control
problems in (Blom et al., 1998a; Gunnarsson et al., 1998c,d).
Example 5.3 (Local Loop of DPC and Delayed Output Powers)
Consider the DPC algorithm
p
i
(t + 1) = p
i
(t) +
t
i
i
(t).
Using the timeshift operator, this corresponds to
p
i
(t) =
1
q 1
(
t
i
i
(t))
The computed output powers are delayed by one sample before they are actually
used by the corresponding transmitter. The SIR is thus given by
i
(t) = p
i
(t 1) +g
ii
(t) I
i
(t).
The algorithm in closedloop can thus be depicted as in Figure 5.6.
+ +
+ p
i
(t)
i
(t)
t
i
(t)
g
ii
(t) I
i
(t)
q
1
1
q 1
Figure 5.6 The local loop using DPC, when the computed output power
is delayed by one sample.
The power control algorithm in Example 5.3 generalizes to
p
i
(t) = R(q)(
t
i
(t)
i
(t)),
where R(q) in the case of the Icontroller in (5.5) is given by
R(q) =
q 1
. (5.12)
66 Local Analysis
Analogously, the AWWalgorithm is described by
R(q) =
q
(5.13)
The output from the controller p
i
(t) cannot depend on the input instantaneously.
Therefore, R(q) has to contain at least one delay. This is analogous to state that
R(q) has to be strictly proper, i.e., the degree of the denominator polynomial in q
has to be greater than the degree of the numerator polynomial. The lter, however,
is by denition proper, i.e., the denominator and numerator degrees may be equal.
Such a general power control algorithm with a lter and subject to time delays is
depicted in Figure 5.7. Note that time delays only aect the dynamics in terms of
+ +
+ p
i
(t)
i
(t)
t
i
(t)
g
ii
(t) I
i
(t)
q
np
q
nm
R(q)
F(q)
Figure 5.7 The local loop when employing the general linear control algo
rithm R(q) and the lter F(q).
the total round trip time delay n = n
p
+ n
m
, since g
ii
(t) and I
i
(t) are considered
as independent disturbances.
The terms instability and stability of a local loop have to be well dened in
order to avoid confusion. Rather simplied, instability comes about when relying
too much on outdated information. This result in overcompensation, which is ag
gravated over time. Eventually, the transmission power oscillates between its upper
and lower output constraints. A system may also be marginally stable, where a
stable oscillation in the output power is observed. In a stable local loop, the eects
of stepwise disturbances and target SIR changes decay to zero over time. We will
discuss stability in terms of the property uniform asymptotic stability. Basically
it implies that transients decay to zero asymptotically. One can compare to the
solution of a linear ordinary dierential equation. It consists of a solution to the
homogenous equation obtained by assuming zero input and a particular solution
obtained when considering an input. If the system described by this dierential
equation is asymptotically stable, then the homogenous, systemdependent, solu
tion will go to zero in steadystate. More specically, it is interesting to see how
the choice of controller parameter values aects the stability.
5.3 LogLinear Algorithms 67
5.3.1 Stability of Linear Systems
Consider a discretetime linear system with one input u(t) and one output y(t).
Using the timeshift operator q, such a system can be described using the transfer
function G(q) as
y(t) = G(q)u(t) =
B(q)
A(q)
u(t) =
b
1
q
1
+b
2
q
2
+. . . +b
n
b
q
n
b
1 +a
1
q
1
+. . . +a
na
q
na
u(t) (5.14)
The poles of this transfer function are dened as the roots of the A(q) polynomial.
When using feedback, the closedloop system can be depicted as in Figure 5.8,
where G(q) denotes the system to be controlled, R(q) denotes the controller, and
S(q) denotes for instance the sensor dynamics or a lter.
r(t) u(t) +
S(q)
R(q) G(q)
e(t)
v(t)
d(t)
n(t)
y(t)
Figure 5.8 The basic feedback loop.
Write G(q), R(q), and S(q) as ratios of coprime polynomials (i.e. polynomials
with no common factors):
G(q) =
B
G
(q)
A
G
(q)
, R(q) =
B
R
(q)
A
R
(q)
, S(q) =
B
S
(q)
A
S
(q)
,
Using these, we can describe the output y(t) as
y(t) =
1
B
G
B
R
B
S
+A
G
A
R
A
S
(B
G
A
R
A
S
d(t) +B
G
B
R
A
S
r(t) B
G
B
R
B
S
n(t))
The characteristic polynomial P(q) of the feedback system is then formed by taking
the product of the three numerators plus the product of the three denominators.
Hence
P(q) = B
G
(q)B
R
(q)B
S
(q) +A
G
(q)A
R
(q)A
S
(q) (5.15)
The closedloop poles are dened as the roots of the characteristic polynomial. A
dierent, but analogous representation of the linear system in Figure 5.8 is the
following state space representation
x(t + 1) = Ax(t) +B(r(t), d(t), n(t))
T
y(t) = Cx(t) +D(r(t), d(t), n(t))
T
68 Local Analysis
where A, B, C and D are matrices and x(t) is the state vector. Let E denote the
identity matrix of the same size as A. The characteristic polynomial is then given
by
P(q) = det(qE A)
Uniformly asymptotic stability of the local loop is disclosed in the following fun
damental theorem. For a general treatment, we refer to e.g., Doyle, Francis, and
Tannenbaum (1992); Franklin, Powell, and Workman (1997); Phillips and Nagle
(1990); Rugh (1996) or
Astrom and Wittenmark (1997).
Theorem 5.1 (Asymptotic Stability of Discretetime Linear Systems)
The feedback system in Figure 5.8 is uniformly asymptotically stable, if and only
if the closedloop poles are strictly within the unit disc. Otherwise the system is
unstable.
If the controller is parameterized using a parameter , it is interesting to see
how stability is aected by dierent parameter values. The natural thing to do is
to plot the locations of the poles for the dierent :s, and see whether they are
located within the unit disc or not. This technique is also known as root locus
analysis.
It is also interesting to study the margin to unstability. There are dierent
ways to express the stability margin, and here the focus is on the gain margin K
m
,
which is dened by
Denition 5.2 (Gain Margin)
The gain margin of a closedloop system with controller R(q) as in Figure 5.8, is the
smallest multiplicative constant K
m
such that the system with controller K
m
R(q),
is unstable. More formally,
K
m
= supK : KR(q) yields a stable closed loop system
5.3.2 Local Analysis of Power Control Algorithms
The closedloop system in Figure 5.7 is all linear, and thus root locus analysis is
applicable.
Consider the closedloop system in Figure 5.7, which includes time delays and
the use of a lter and the general linear power control algorithm R(q). The analysis
applies to a general R(q), but the focus will be on the integrating controller in (5.12)
R(q) =
q 1
.
Based on Figure 5.7 and Equation (5.15), we can form the characteristic polynomial
of the closedloop system as
P(q) = B
F
(q) +q
n
(q 1)A
F
(q) (5.16)
5.3 LogLinear Algorithms 69
The lter F(q) is discussed in Section 5.2.4, and when using a local average, it is
given by
F(q) =
q
L1
+. . . + 1
Lq
L1
=
B
F
(q)
A
F
(q)
With this choice of lter, the locations of the closedloop poles as functions of
for some dierent time delays n and window lengths L are plotted in Figure 5.9.
Note that L = 1 corresponds to no lter, i.e., F(q) = 1. The particular case of
1 0 1
1.5
1
0.5
0
0.5
1
1.5
Real Axis
I
m
a
g
A
x
i
s
1 0 1
1.5
1
0.5
0
0.5
1
1.5
Real Axis
I
m
a
g
A
x
i
s
1 0 1
1.5
1
0.5
0
0.5
1
1.5
Real Axis
I
m
a
g
A
x
i
s
1 0 1
1.5
1
0.5
0
0.5
1
1.5
Real Axis
I
m
a
g
A
x
i
s
a. n = 1, L = 1 b. n = 0, L = 3
c. n = 1, L = 3 d. n = 0, L = 10
Figure 5.9 Root locus plots for some dierent window lengths L and time
delays n when using a local average lter. The x:s mark the
roots corresponding to = 0, while o corresponds to the roots
when .
DPC, subject to a single delay and no lter applied, is addressed in the following
example.
Example 5.4 (Local Loop Poles of DPC, Subject to a Single Delay)
Consider the situation in Example 5.3, where the local loop when using DPC
is in focus. The computed output powers are delayed by one sample. The
corresponding characteristic polynomial P(q) is obtained from Equation 5.16
as
P(q) = q
2
q + 1.
70 Local Analysis
The roots of P(q) are (1
lim
,
since the controller K
m
R(q) is just about unstable.
In this section, the focus has been on the stability properties of the integrating
controller. Analyzing other loglinear algorithms is analogous, and the correspond
ing results for the AAW algorithm are summarized in Appendix 5.A.
5.3.3 Summary, Local Analysis of LogLinear Algorithms
Several interesting algorithms utilizing information feedback can be described by
linear local control loops in logarithmic scale. Root locus analysis is thus applica
ble to include the dynamical eects of time delays and to address local stability.
Thereby, we relate local stability to the choice of controller parameter values. Sim
ulations indicate that the DPC algorithm is globally unstable when subject to time
delays. This is natural, since the corresponding local loops are not asymptotically
stable. Furthermore, the eects of ltering are studied. The conclusion is that an
exponential forgetting lter has more appealing properties with respect to stability
than the local average lter.
5.4 Describing Functions
When the nonlinearities in the local loop cannot be neglected, the methods in the
previous section do not apply. As an alternative, we utilize describing functions,
72 Local Analysis
based on assumptions of oscillations or limit cycles in the loop. The relevance of
this assumption was illustrated in Example 5.1, where the use of decision feedback
in FSPC resulted in an oscillatory behavior. In this section we develop the under
lying theory, rst in a simple case to stress the main ideas, and then with focus to
power control local loops with decision feedback. Therefore, the analysis primarily
aim at studying the eects of relays in discretetime linear control loops. However,
decision feedback is common in many other applications related to communications
and signal processing. For example, we mention decision feedback equalization and
sigmadelta modulators. Moreover, it is also common in software sometimes for
no reason.
The theory of describing functions in continuous time is thoroughly discussed by
Atherton (1982); Glad (1994); Phillips and Nagle (1990). De la Sen (1993) provide
modications for discretetime describing functions of some specic nonlinearities.
The applicability to power control in cellular radio systems have been addressed in
(Blom and Gunnarsson, 1998; Blom et al., 1998a; Gunnarsson et al., 1998d, 1999c)
5.4.1 Describing Functions with Zero Phase Assumption
Basically, we are focusing on loops that consist of a linear part with transfer func
tion G(q) and a static nonlinearity described by the function f() resulting in a loop
as in Figure 5.11. Note that we have assumed a zero input to the loop. Nonzero
inputs are studied in Section 5.4.3.
0
1
e(t)
f(e) G(q)
w(t) y(t)
Figure 5.11 Block diagram of a nonlinear system, separated into one linear
and one nonlinear component.
Nonlinearities in the loop normally result in an oscillatory behavior. This will
be studied by assuming that there is an oscillation in the error signal e(t), and then
try to verify this assumption. We proceed by making the Nperiodic hypothesis
e(t) = E sin(
e
t) = E sin
2
N
t
,
where E is the amplitude of the oscillation and
e
is the normalized angular fre
quency. The hypothesis is based on a zero phase assumption, with the meaning
that the sample at t = 0 is zero (slightly more general, we assume that the sampled
5.4 Describing Functions 73
sinus is zero for some t in every period). In order to simplify the calculations we
assume for a moment that f() is odd. The computations using a general static
nonlinearity are analogous, but a little bit more complicated, and we will return
to a general f() further on. Since the nonlinearity is static, w(t) is Nperiodic as
well. Using discrete time Fourier series expansion, the signal w(t) is decomposed
into its Fourier components as
w(t) =f(E sin(
e
t)) = A
1
(E, N) sin(
e
t +
1
(E, N)) +
+A
2
(E, N) sin(2
e
t +
2
(E, N)) +. . .
Recall that a sinusoid
X sin(
X
t) fed through a linear system H(q) results in the
output
H(e
iX
)
sin(
X
t + arg(H(e
iX
))),
after the transients have decayed. Now let us make the assumption that the lin
ear system G(q) will attenuate the harmonics much more than the fundamental
frequency. This is the only approximation we will make, and it yields
y(t) A
1
(E, N)
G(e
ie
)
sin(
e
t +
1
(E, N) + arg(G(e
ie
))). (5.17)
Figure 5.11 yields
y(t) = e(t) = Esin(
e
t) = E sin(
e
t +). (5.18)
Recall that
e
=
2
N
. The loop is closed by combining Equations (5.17) and (5.18).
Thus an oscillative behavior is predicted if there exists a solution to the following
equations
A
1
(E, N)
G(e
i
2
N
)
= E (5.19a)
1
(E, N) + arg(G(e
i
2
N
)) = + 2, Z. (5.19b)
The complex Fourier series (Cadzow, 1973) is dened by
w(t) =
N1
k=0
C
k
(E, N)e
jekt
(5.20a)
C
k
(E, N) =
1
N
N1
t=0
w(t)e
jekt
(5.20b)
Moreover, dene the complex number
Y
f
(E, N) =
A
1
(E, N)e
i1(E,N)
E
=
2i
E
C
1
(E, N). (5.21)
Then (5.19) can be expressed more compactly as
Y
f
(E, N)G(e
i
2
N
) = 1. (5.22)
Equation (5.22) or the equations in 5.19 provide two equations and two unknowns.
If a solution exists, it describes an approximation of the oscillation in the loop.
Conversely, the lack of a solution motivates the absence of oscillative modes.
74 Local Analysis
5.4.2 DiscreteTime Describing Functions
The derivation in the previous section is based on a zero phase assumption. This
is not so important when the static nonlinearity is continuous, since phase shifts in
e(t) result in corresponding phase shifts in w(t) = f (e(t)). The situation is dierent
when considering discontinuous static nonlinearities. For example, consider a relay
and a phase shift of half a sample. The output is given by
w(t) = sign
E sin(
2
N
(t + 0.5))
.
Such a small phase shift does not aect the sign of the samples, and thus the output
will be the same as with zero phase shift, i.e.,
w(t) = sign
E sin(
2
N
(t + 0.5))
= sign
E sin(
2
N
(t))
.
If N is even, this relation hold for phase shifts
e
[0, 1[. Note that a phase shift
of an entire sample is the same as a time delay and should therefore be a part of
of the linear transfer function G(q). The unknown phase shift is thus among the
parameters characterizing an oscillation. Its eect depends the nonlinearity and
the period N. The discussion motivates the following denition of the discretetime
describing function
Denition 5.3 (DiscreteTime Describing Functions)
The discretetime describing function of the static nonlinearity f() is dened
(cf. (5.21)) by
Y
f
(E, N,
e
) =
2i
E
C
1
(E, N,
e
),
where the complex Fourier coecient C
1
(E, N,
e
) is given by
C
1
(E, N,
e
) =
1
N
N1
t=0
f
E sin(
2
N
(t +
e
))
e
i(
2
N
(t+e))
,
e
[0, 1[.
The denition can be written more compactly, as
Y
f
(E, N,
e
) =
2i
NE
N1
t=0
f (E sin(
e
(t +
e
))) e
i(e(t+e))
.
Recall the following central assumption from the previous section
Assumption 5.4
The linear part G(q) in Figure 5.11 attenuates the harmonics much more than the
fundamental frequency
e
. More formally, we assume that
[G(e
ike
)[
[G(e
ie
)[
e=
2
N
0
<< 1, k = 2, . . . ,
N
0
2
5.4 Describing Functions 75
With a analogous reasoning as in the previous section, we are interested in the
solution to
Y
f
(E, N,
e
)G(e
i
2
N
) = 1,
e
[0, 1[. (5.23)
This is essentially two equations and a constraint, which is more clear by separating
the magnitude and the phase of each side.
[Y
f
(E, N,
e
)[[G(e
i
2
N
)[ = 1 (5.24a)
arg Y
f
(E, N,
e
) + arg G(e
i
2
N
) = + 2, Z,
e
[0, 1[ (5.24b)
We thus have two equations and four unknowns (E, N,
e
and ), which enables
several possible solutions.
This far we have assumed that Assumption 5.4 holds. If this is not the case, the
analysis may still not be in vain. In such a case, the higher frequencies contribute
to the waveform of the error signal e(t). Describing functions focus on the funda
mental frequency of the oscillations. Therefore, the estimated fundamental period
N obtained from describing functions analysis is still informative. Instead, dierent
waveforms of the error signal may be assumed, depending on the application, the
nonlinearity and the system.
The discretetime describing functions analysis is summarized in the following
algorithm
Algorithm 5.1 (DiscreteTime Describing Functions Analysis)
Consider the situation in Figure 5.11, where the loop is separated in a linear
(G(q)) and a nonlinear (f()) part. Then the oscillation in the error signal e(t)
is approximated by the procedure
1. Determine the discretetime describing function of the nonlinearity as
Y
f
(E, N,
e
) =
2i
NE
N1
t=0
f (E sin(
e
(t +
e
))) e
i(e(t+e))
,
where
e
=
2
N
and
e
[0, 1[.
2. Compute G(q)[
q=e
2i/N .
3. Solve the following equation for E, N and
e
.
Y
f
(E, N,
e
)G
e
i
2
N
= 1 (5.25)
If one solution (E, N,
e
) exists, then the oscillation is approximated by
e(t) = E sin
2
N
(t +
e
)
.
If several solutions exist, then several modes of oscillation are possible.
4. Investigate the correctness of Assumption 5.4. If it does not hold, the
estimated periods are still informative, but alternative waveforms may be
discussed.
76 Local Analysis
5.4.3 Simplied Nonzero Input Case
When considering nonzero inputs in general, the analysis becomes more complex.
In some cases, however, such as the power control case, some approximative simpli
cations are readily available. Consider the FSPC algorithm in (5.8) and include
the delays as in Figure 5.1. This yields
p
i
(t + 1) = p
i
(t) +
i
sign
t
i
(t) p
i
(t n
p
n
m
) g
ij
(t n
m
) +I
i
(t n
m
)
.
Introduce p
i
(t) = p
i
(t)
t
i
(t) + g
ij
(t + n
p
) I
i
(t + n
p
), and assume that the
power gain, the interference and the target SIR are constant over the delay horizon
(n
p
+n
m
samples). Hence
p
i
(t + 1) = p
i
(t) +
i
sign( p
i
(t n
p
n
m
)),
which is the zero input case. This will only be an approximation. A speculative
proposition is that small deviations essentially eect the unknown phase
e
. There
fore, the resulting oscillative behavior possibly comprises several modes, and the
mode switching is stimulated by these deviations.
5.4.4 Describing Function of a Relay
An ideal relay or sign function (see Section 5.2.3) is dened by
f(e) =
1, e 0
1, e < 0
(5.26)
The corresponding describing function is obtained by applying Denition 5.3. In
the relay case, the complex Fourier coecient C
1
(E, N,
e
) can be computed as
C
1
(E, N,
e
) =
1
N
N1
t=0
f (E sin(
e
(t +
e
))) e
i(e(t+e))
=
=
1
N
N/21
t=0
e
i(e(t+e))
1
N
N1
t=N/2
e
i(e(t+e))
=
=
1
N
e
iee
1 e
i
N/21
t=0
e
iet
=
=
2
N sin
N
e
i(
2
e
2
N
)
The discretetime describing function is thus
Y
f
(E, N,
e
) =
4
NE sin
N
e
i(
N
e
2
N
)
(5.27)
5.5 LogLinear Algorithms with a Static Nonlinearity 77
5.5 LogLinear Algorithms with a Static
Nonlinearity
The analysis in this section is focused on the dynamical behavior of the FSPC
algorithm in (5.8)
p
i
(t + 1) = p
i
(t) +
i
sign
t
i
(t)
i
(t)
, (5.28)
In operation, the corresponding local loop can be associated with the block diagram
in Figure 5.12. With the relay as the static nonlinearity, and the remaining parts
all linear, this is in full analogy with the block diagram in Figure 5.11. Hence,
describing functions analysis is applicable.
t
i
(t)
p
i
(t)
e(t) + +
+
q
np
q
nm
i
q 1
F(q)
g
ii
(t) I
i
(t)
i
(t)
Figure 5.12 The decision feedback of the FSPC algorithm in (5.28) is vi
sualized as a relay block, which is a static nonlinearity.
In this case with a relay and an integrator, is is easy to realize that primarily
modes with even periods N are dominating the oscillations. This is formulated in
the following proposition.
Proposition 5.5 (Evenperiod Oscillations of FSPC)
An ideal relay together with an integrator as in the FSPC case, cannot have oscil
lations of odd periods. While in operation, single cycles of odd periods might be
present, but they will be regarded as transitions between evenperiod cycles.
Consider the process outlined in Algorithm 5.1. The describing function of the
ideal relay is provided in (5.27)
Y
f
(E, N,
e
) =
4
NE sin
N
e
i(
N
e
2
N
)
,
e
[0, 1[ (5.29)
The linear part of the block diagram in Figure 5.12 is
G(q) =
i
B
F
(q)
q
np+nm
(q 1)A
F
(q)
78 Local Analysis
For simplicity, consider unltered measurements. In that case, G(q) on the unit
circle is equal to
G(q)[
q=e
2i/N =
i
2 sin(/N)
e
i(
2
+
N
+
2
N
(np+nm))
(5.30)
As in the root locus analysis, only the total additional roundtrip delay n = n
p
+n
m
matters. Equations (5.25), (5.27) and (5.30) yield
2
i
NEsin
2
N
e
i(
2
+
2
N
(e+n))
= 1 = e
i(+2)
,
e
[0, 1[. (5.31)
Separating the phase and magnitude equalities, results in the two equations
E =
2
i
N sin
2
N
(5.32a)
2
+
2
N
(
e
+n) = + 2, Z,
e
[0, 1[ (5.32b)
Thus, it is the phase equality (5.32b) that allow dierent solutions. It can be
rewritten as
N =
4(
e
+n)
1 + 4
, , ; Z,
e
[0, 1[. (5.33)
Since N is positive, 0. The denominator describes an integer, thus also the nu
merator has to be an integer. Therefore, only
e
values in the set 0, 0.25, 0.5, 0.75
are plausible. Moreover, N is even according to Proposition 5.5, which excludes
the second and forth possible
e
values. Hence, the possible periods are obtained
from
N =
4(
e
+n)
1 + 4
, = 0, 1, 2, . . . ,
e
0,
1
2
. (5.34)
For each N, the corresponding amplitude is computed using (5.32a).
Finally, we have to verify the correctness of Assumption 5.4. If it does not hold,
the estimated periods are still informative, as pointed out in Section 5.4.2. With
the stepwise updates of the relay, an intuitive alternate waveform is a triangular
wave. Such a waveform of period N has the amplitude
E
=
N
i
4
. (5.35)
We exemplify the analysis by a concrete example
Example 5.5 (Analysis of FSPC in a Typical WCDMA Setting)
Consider the typical delay situation in a WCDMA system (see Example 5.2)
5.6 Effect of AutoInterference 79
n
p
= 1, n
m
= 0. The possible oscillation modes (identied by the period N)
are obtained from (5.34). Only v = 0 provides integer periods. Thus
N = 4
e
+ 4,
e
0,
1
2
,
which yield the solutions N
0
= 4 and N
0
= 6. To verify Assumption 5.4,
compute the ratio
[G(e
i2e
)[
[G(e
ie
)[
e=
2
N
0
=
sin(/N)
sin(2/N)
=
1
2 cos(/N)
,
where the expression in (5.30) is used in the second last equality. This is
equal to 0.71 and 0.58 for N
0
= 4 and N
0
= 6 respectively. It is therefore more
relevant to use the expression in (5.35) than in (5.32a) to compute the amplitude
of the modes. Table 5.3 provides the corresponding amplitudes using both
expressions. Compare the result to the introductory simulation in Example 5.1.
Oscillation mode E
0
E
0
N
0
= 4
i
i
N
0
= 6 1.33
i
1.5
i
Table 5.3 Predicted oscillation modes in the typical WCDMA situation.
The estimated second mode with period N
0
= 6 and amplitude E
0
= 1.5 dB
describes the oscillations well.
Corresponding oscillation analyses with respect to other delay situations are
summarized in Table 5.4.
5.5.1 Summary, Describing Functions Analysis
In summary, we conclude that static nonlinearities, such as a relay or a sign func
tion, results in an oscillatory behavior of the system. These oscillations may com
prise several modes. Switching between the modes may be stimulated by external
disturbances. For analysis purposes, describing functions are introduced and ap
plied to local loops consisting of one static nonlinearity and a linear part. Thereby,
prevalent oscillations can be predicted. A natural power control related example is
the FSPC algorithm. Using the discretetime describing functions, the oscillations
observed in simulations, can be predicted.
5.6 Effect of AutoInterference
The eect of the autointerference on local loop stability has been neglected this
far. Naturally, the eciency of the receiver has an impact, but the motivation in
80 Local Analysis
n Oscillation modes E
0
E
0
e
v r
G
0 N
0
= 2
i
i
0 0 
1 N
0
= 4
i
i
0 0 0.71
N
0
= 6 1.33
i
1.5
i
0.5 0 0.58
2 N
0
= 2
i
i
0.5 1 
N
0
= 8 1.7
i
2
i
0 0 0.54
N
0
= 10 2.1
i
2.5
i
0.5 0 0.53
3 N
0
= 12 2.5
i
3
i
0 0 0.52
N
0
= 14 2.9
i
3.5
i
0.5 0 0.51
Table 5.4 Predicted oscillation modes for various delays. The introduced
quantity r
G
represents the ratio in Assumptions 5.4 for k = 2.
Since r
G
is not small enough in most cases, E
0
is probably a
better amplitude approximation than E
0
.
the introduction of this chapter was that it can be described together with the
eect of the interconnections. Nevertheless, a brief analysis is included here. In
essence, the eect of the autointerference is brought into our linear systems theory
framework, by linearizing the interference. Then issues like stability are naturally
addressed.
Assume that the signal power gains g
ij
are slowly varying, and can be regarded
as constant. If all inner loops provide perfect control, and the power control prob
lem is feasible, all target SIR:s
t
i
are met. Let p
t
i
denote the corresponding power
assignments, and I
t
i
the resulting interference at each receiver. As stressed by
Equation (3.5), the interference at receiver i is given by (in linear scale)
I
i
(t) =
j=i
g
ij
(t) p
j
(t) +
i
(t)
p
i
(t) g
ii
(t) +
i
(t).
Delayed output powers p
i
(t) have to be considered. Moreover, an important as
sumption in this chapter is that the local loop dynamics is independent of other
loops. Therefore, we rewrite the interference expression as
I
i
(t) =
I
ol
i
(t) +
i
(t)
p
i
(t n
p
) g
ii
(t),
where
I
ol
i
(t) represents independent disturbances. The corresponding interference
in logarithmic scale (dB) is
I
i
(t) = 10 log
10
I
ol
i
(t) +
i
(t)
g
ii
(t)10
pi(t)/10
,
with respect to the power p
i
(t) in logarithmic scale. Assume that the receiver
eciency and the disturbances from others are constant on a short time scale.
A linearization (see e.g., Franklin et al., 1997) of the interference with respect to
5.6 Effect of AutoInterference 81
the equilibrium (p
t
i
, I
t
i
and
t
i
), provides an approximate relation that holds in a
neighborhood of the equilibrium. Further details are provided in Section 7.3, and
the linearization is given by
I
i
(t) = I
t
i
+ r
i
(p
i
(t n
p
) p
t
i
), 0 r
i
< 1. (5.36)
where r
i
is depending on
i
and other quantities.
Consider the use of a general loglinear power control algorithm R(q)
p
i
(t) = R(q)(
t
i
i
(t n
m
)). (5.37)
Moreover, SIR is given by
i
(t) =
i
+p
i
(t n
p
) +g
ii
I
i
(t). (5.38)
Equations (5.36), (5.37) and (5.38) approximate the closedloop behavior of the
local loop to
p
i
(t) =R(q)
t
i
i
p
i
(t n
p
n
m
) g
ii
+I
i
(t n
m
)
R(q)
p
t
i
p
i
(t n
p
n
m
) + r
i
(p
i
(t n
p
) p
t
i
)
=
=R(q)(1 r
i
)
p
t
i
p
i
(t n
p
n
m
)
.
A similar result is obtained when considering lters F(q) with the property F(1) = 1
(see Section 5.2.4).
p
i
(t) R(q)(1 r
i
)
p
t
i
F(q)p
i
(t n
p
n
m
)
t
i
i
(t n
m
)
. (5.39)
The closedloop system is described by
p
i
(t + 1) =p
i
(t) +
i
sign
t
i
i
(t) p
i
(t n
p
n
m
) g
ii
+I
i
(t)
p
i
(t) +
i
(1 r
i
) sign
p
t
i
p
i
(t n
p
n
m
)
.
As previously, including a lter F(q), F(1) = 1, we get
p
i
(t + 1) p
i
(t) +
i
(1 r
i
) sign
p
t
i
F(q)p
i
(t n
p
n
m
)
.
The eect of the autointerference is thus a smaller step size in reality. According to
the describing function analysis in Section 5.3.2, the amplitude of the oscillations is
smaller. Hence, the autointerference does not violate stability of linear algorithms
nor aggravate oscillations of the FSPC algorithm.
82 Local Analysis
5.7 Summary
The global system can be seen as a number of local control loops, interconnected
via the interfering powers between the users. We motivate why local loop stability
is necessary for global stability. Therefore, local loop analysis provides valuable
insight in the dynamical behavior of power control algorithms. Primarily two
classes of algorithms are considered. The rst class include most of the interesting
proposed algorithms utilizing information feedback. They can be seen as linear
controllers in logarithmic scale, and thus properties of linear systems are relevant.
In order to avoid confusion, stability is carefully dened. Using root locus analysis
from control theory, stability is addressed with respect to the controller parameters.
Controllers, such as the FSPC algorithms, that utilize decision feedback, behave
slightly dierent. The decision component can be seen as a sign function or a relay
and has both stabilizing and unstabilizing eects. This static nonlinearity result
in an oscillatory behavior, which neither vanishes nor aggravates over time. For
that matter, discretetime describing functions are introduced and applied to the
local loop when using FSPC. The conclusion is that several oscillation modes are
possible, and mode switching is stimulated by external disturbances.
Finally, the dynamical eects of imperfect receivers are studied. The conclusion
is that this autointerference does not violate stability of linear algorithms nor
aggravate oscillations of the FSPC algorithm.
Appendix
5.A Local Loop Analysis of the AAW Algorithm
The AAW algorithm, which was proposed by Almgren et al. (1994) and reviewed
in Section 5.2.1, is given by
p
i
(t + 1) = p
i
(t) +(
t
i
i
(t)),
where
t
i
= /. With the timeshift operator, this is rewritten as
p
i
(t) =
q
(
t
i
i
(t)) = R
AAW
(q)(
t
i
i
(t)). (5.A.1)
Using the same approach as in Example 5.3 and Figure 5.7 we can draw the
block diagram describing the local loop of the power control algorithm as in Fig
ure 5.1. As in Section 5.3, the use of a lter F(q) is optional. The local loop
+ +
+ p
i
(t)
i
(t)
t
i
(t)
g
ii
(t) I
i
(t)
q
np
q
nm
q
F(q)
Figure 5.1 The local loop when employing the AAW algorithm and a lter
F(q).
when neglecting constraints is linear, and therefore root locus techniques as in Sec
tion 5.3.2 are applicable. Instability limits for are summarized in the case of a
local average lter in Table 5.1 and in the case of an exponential forgetting lter
in Table 5.2.
83
84 Local Analysis
L = 1 L = 3 L = 5 L = 10
n = 0 > 1 > 1 0.981 0.842
n = 1 1.000 0.812 0.770 0.759
n = 2 0.707 0.671 0.673 0.700
Table 5.1 Local loop stability region given by strict upper bounds on ,
when using the AAW algorithm and a local average lter.
= 0 = 0.33 = 0.60 = 0.80
n = 0 > 1 > 1 > 1 > 1
n = 1 1.000 1.000 1.000 1.000
n = 2 0.707 0.731 0.782 0.858
Table 5.2 Local loop stability region given by strict upper bounds on ,
when using the AAW algorithm and an exponential forgetting
lter.
The conclusion in Section 5.3.2 that the exponential forgetting lter has the
more appealing stability properties still holds. The tabled upper stability bounds
are somewhat surprising. For instance, using window length L = 3 when subject
to a time delay of n = 1 actually puts stronger stability restrictions on than both
using no lter and when using window lengths L = 5 or L = 10. The reason is that
the controller gain of R
AAW
(q) in (5.A.1) does not linearly depend on .
6
Local Design
In the previous chapter, methods from control theory were applied to address
stability of distributed local control loops. Stability is of course necessary, but
dierent stabilizing controllers provide very dierent behavior. Therefore, some
design methods from control theory are applied to obtain good performance. The
considered controller structures form a local loop as depicted in Figure 6.1. As
in the previous chapter, the focus is on SIRbased power control algorithms, with
controller R
i
, controller command decoder D
i
and ltering operation F
i
. Again, the
system is subject to time delays, and possibly also transmission power constraints.
Disturbance models and statistics provide valuable information to the design
process. Such models are discussed in Section 6.1, and will be used throughout
the chapter. As disclosed in Chapter 5, instability of local loops is essentially a
result of too much trust in outdated information. The behavior can be improved
by either rely less or more careful on the measurements or by predicting future
measurements. The core problem with outdated information is that the most
recently computed output powers are not reected in the measurements. However,
these power levels are known to the algorithm and can be used to adjust the
measurements accordingly. The strategy is referred to as time delay compensation
in Section 6.2.
Loglinear algorithms are stable if the resulting closedloop poles are within the
unit circle (see Section 5.3.1). This naturally motivates a design strategy, where
85
86 Local Design
p
i
(t)
t
i
(t)
i
(t)
R
i
D
i
g
ii
(t) I
i
(t) x
i
(t)
u
i
(t)
q
np
q
nm
F
i
Receiver Transmitter
Figure 6.1 Block diagram of the receivertransmitter pair i when employ
ing a general SIRbased power control algorithm. In operation,
the controller result in a closed local loop.
the controller parameter values are chosen to place the poles at relevant locations.
Pole placement design is in focus in Section 6.3. The resulting SIR depend on the
disturbances, the target SIR and the power controller. These relations are naturally
discussed in the frequency domain and are further explored in Section 6.4.
The disturbance models in Section 6.1 are not only useful to evaluate perfor
mance. Given a disturbance model, possibly recursively estimated, future dis
turbances can be predicted. Such predictions are utilized by controllers and in
controller design in Section 6.5.
The discussion above mainly focus on linear controllers, possibly based on es
timated parameters. Nonlinear controllers stand out as natural extensions, but
are more dicult to design. In some situations, however, nonlinear components
in otherwise linear controllers provide control actions, that cannot be obtained by
only linear control. Nonlinear components with application to the power control
problem are discussed in Section 6.6.
A number of dierent design and control strategies are covered in this chapter.
The use of some central algorithms is exemplied in simulations in Section 6.7.
6.1 Fading Characteristics
The radio propagation channel can, as discussed in Section 3.3, be modeled as
comprising three main components: path loss, shadow fading and multipath fad
ing. As in Figure 6.1 it can be considered as an additive disturbance. The ability
to compensate for the timevariant propagation channel is therefore related to its
correlation properties. Since the timevariations are due to terrain and environ
ment variations, this is most naturally addressed in the spatial frequency domain
6.1 Fading Characteristics 87
[ m
1
]. Note that the spatial frequency domain is related to the more traditional
temporal frequency domain f [ s
1
] by the velocity v [ m/s] of the mobile station
such that f = v.
The path loss will be considered constant in the sequel, and thus timevariations
are captured by shadow and multipath fading. As in Section 3.3, the focus is on
the propagation channel. It is important to note that the generalization of power
gain to also include code correlation etc. (see Section 4.3) only aects the long time
mean and thus have no eect on the frequency domain representation except from
a bias.
6.1.1 Shadow Fading
The long term statistics of the shadow fading in logarithmic scale is well captured
by a Gaussian distribution. However, such a model is not describing the slow
correlation of the fading. The eects from the terrain are correlated and if the
receiver is shielded at one instant, it will most likely be shielded for some time
thereafter.
0 10 20 30 40 50 60 70 80 90 100
10
0
10
0 1 2 3 4 5 6 7 8 9 10
30
20
10
0
10
a.
b.
x [m]
Figure 6.2 Realizations of a. shadow fading gain and b. multipath fading
gain in dB based on the models discussed in the text. Note the
dierent spatial scales.
Gudmundson (1991) included spatial correlation by tting a rst order lter
to measured data at the 900 MHz band. Srensen (1998) rened the model using
88 Local Design
data from a 900 MHz small urban macro cell. The multipath contribution to the
measurement is assumed additive and white. It is deemphasized using spectral esti
mation techniques. The focus in this chapter will be on the latter model. Therefore
this model receives extra attention below. Srensen used data sampled by a spatial
sample interval x
s
= 0.1 m. Contributions from the multipath fading is assumed
to aect the measurements as a constant noise level. An ARMA(2,1)model is seen
to satisfactory represent the fading correlation in the observations.
g
s
(k) = a
1
g
s
(k 1) a
2
g
s
(k 2) +e
s
(k) +b
1
e
s
(k 1), Vare
2
s
(k) =
2
e
,
[a
1
, a
2
, b
1
]
T
= [1.8384, 0.8395, 0.9634]
T
,
2
e
= 2.01. (6.1)
where the sample instants k are with respect to the spatial sample interval x
s
and
e
s
(k) is zero mean stationary white Gaussian noise. The normalized autocorrelation
function of this model can be written as
r
gs
(0) = 1, r
gs
(1) = 0.930,
r
gs
(k) = 0.542e
k/142
+ 0.458e
k/6
Two decorrelation distances are thus naturally identied: 6 and 142 samples, i.e.,
x
1
= 14.2 m and x
2
= 0.6 m. The overall decorrelation distance (when the nor
malized autocorrelation is less than e
1
) is 5.5 m. A corresponding analysis of
the model provided by Gudmundson result in the decorrelation distance 8 m. The
decorrelation distances can be considered as a rough measure of the spatial fre
quency content. The low frequency content has components approximately up
to 1/x
1
= 0.07 [ m
1
] and the remainder up to 1/x
2
= 1.7 [ m
1
]. Figure 6.2a
illustrates a simulated instance of the model in (6.1.1) and indicates the spatial
correlation.
6.1.2 Multipath Fading
In an urban environment, the transmitted signal is reected by large objects along
the path. Therefore, several rays reach the receiver. Depending on their phase,
they interfere either constructively or destructively resulting in multipath fading.
A simplied worst case scenario arises when the LineofSight (LoS) ray is as
sumed completely shielded, but the reected rays reach the receiver and enable
transmission of information. If the arriving rays are modeled as isotropic, the am
plitude gain a due to multipath is described by a Rayleigh distribution (Clarke,
1968)
p
a
(x) =
2x
p
e
x
2
p
, x 0,
where
p
= E[ a
2
] is the average power gain. We are interested in the distribution
of the power gain g
m
= a
2
, which is obtained using the following transformation of
probability density functions
p
gm
(x) =
1
2
x
p
a
(
x). (6.2)
6.1 Fading Characteristics 89
The power gain thus has an exponential distribution with the pdf
p
gm
(x) =
1
p
e
x
p
, x 0.
When the LoS component cannot be neglected, the computations become a
little more complex, and the amplitude gain can be shown to have a Ricean distri
bution (Rice, 1948) with the pdf
p
a
(x) =
2x(K + 1)
p
e
K
(K+1)x
2
p
I
0
2x
K(K + 1)
, x 0,
where I
0
is the zeroorder modied Bessel function, and K is the ratio of the LoS
component power to the power of the other components.
The third alternative is to model the multipath fading using a Nakagami dis
tribution (Nakagami, 1960). It was selected to t empirical data, and in essence it
describes the amplitude gain by a
2
distribution with m degrees of freedom, i.e.
p
a
(x) =
2m
m
x
2m1
(m)
m
p
e
mx
2
p
, x 0,
where (m) is the Gamma function. The Nakagami distribution is related to
the other two models. Using m = 1 we get the Rayleigh distribution, and using
m = (K+1)
2
/(2K+1) we obtain a good approximation of the Ricean distribution.
The latter is particularly valuable when performing analytical computations, since
the Bessel function in the Ricean distribution is troublesome in those cases.
The power gain distributions for the last two models are left out, but they
are easily obtained using the transformation in (6.2). Further details on fading
channels can be found in a survey by Sklar (1997) or in (Larsson, 1997).
Just as with the lognormal distribution of the shadow fading, these distribu
tions only describe the longterm statistics. To include spatial correlation, Clarke
(1968) proposed a multipath fading model based on random scatterers in neareld
of the mobile station. The incident rays at the receiver combine to an envelope
mainly dependent on the scatterer locations and the wavelength. A simulated
corresponding power gain at 900 MHz is provided in Figure 6.2b, using a spatial
sampling interval x
m
= 0.01 m. This model is rened by Aulin (1979) to include
threedimensional eects.
Apart from scattering in the neareld, the transmitted signal may be reected
by objects further away. In such a case, several main rays enter the neareld and
are aected by neareld scattering. Typically, this intermediateled scattering is
modeled by an impulse response, which describes the relative strength and delays
of the incident rays to the neareld. Specic propagation scenarios are specied by
the dierent standardization organizations. For example, consider the Vehicular A
channel model in a WCDMA system (UMTS 30.06, 1997). It describes a situation,
where a vehicle is moving at 120 km/h in a classic macro cell environment with site
tosite distance of 6 km. The receiver observes ten incident rays to the nearled,
90 Local Design
but is only able to resolve the four strongest. Figure 6.3a describes the specied
impulse response and indicate which rays the receiver is able to resolve. The
resolving ability of the receiver clearly is critical. Advanced receiver structures
constitute an active research area, but is beyond the scope here. An interested
reader is referred to (Johansson, 1998; Tidestav, 1999) . If we assume that the
receiver can utilize all power in the four strongest rays, the multipath eect on the
power gain is just the sum of th respective powers. Figure 6.3b illustrates this power
gain for a simulated instance of the specied impulse response. The corresponding
periodogram is found Figure 6.3c. Note that the graphs are provided in the time
and temporal frequency domain respectively, since the velocity is specied.
10
1
10
0
10
1
10
2
10
2
10
1
10
0
10
1
0 1 2 3 4 5 6 7 8 9 10
15
10
5
0
5
10
0 0.5 1 1.5 2 2.5 3
30
20
10
0
a.
b.
c.
t [s]
t [s]
f [Hz]
Figure 6.3 The channel model Vehicular A is described by the impulse
response in a. It is assumed that the receiver is able to resolve
the four strongest paths, marked by o. A realization of the
power gain with corresponding periodogram are found in b.
and c. respectively.
6.1 Fading Characteristics 91
6.1.3 Fading in the Spatial Frequency Domain
The multipath fading is superimposed on the shadow fading. Using smoothed
periodograms (Ljung, 1999) of a realization at 900 MHz, an estimate of its spatial
frequency contents can be obtained. Figure 6.4 depicts this periodogram estimate
with respect to the spatial frequency [ m
1
]. For comparison, the discretetime
10
3
10
2
10
1
10
0
10
1
10
2
10
5
10
4
10
3
10
2
10
1
10
0
10
1
10
2
10
3
[m
1
]
Figure 6.4 Fading (in dB) at 900 MHz. Multipath fading is superimposed
on shadow fading, modeled as in (6.1.1). The thicker lines
represent the Fourier transform of the shadow fading model
with (solid) and without (dashed) a noise oor. The shadow
fading is dominant up to about 0.51 m
1
, and the multipath
fading has a characteristic resonance frequency at 1/ = 3 m
1
(the Doppler frequency).
Fourier transform of the model in (6.1.1) is included in the gure (solid). The
relevance of the model is further emphasized by adding (in linear scale) the noise
level, which was considered to represent multipath fading. In addition, we observe
a resonance at the Doppler frequency (1/) which is in accordance with theoretical
properties (Jakes, 1974).
As disclosed above, the relations to temporal frequency are described by f = v.
Note that v reects velocity relative to the terrain and environment. In the never
stationary world of today, v is always greater than zero.
6.1.4 Measurement Procedures
Measuring is not an instantaneous procedure, even though the measurements often
are considered as samples of a continuous process. This is a relevant approximation
92 Local Design
in most power control cases. However, some related issues are brought up here.
The actual interface in a GSM system provides measurements via reports, which
are available to the power control algorithm every 0.48 s. These reports are based
on a local average (see Section 5.2.4) of measurements from 104 bursts (see also
Chapter 8). The averaging is essentially a low pass lter.
In a WCDMA case, the situation is slightly dierent. Measurements are ob
tained from the fraction
s
of the slot (see Section 5.2.2), which in turn corresponds
to T
s
= 1/1500 s. Typical values (Adachi et al., 1998) of
s
include
s
= 0.1 (con
sidering only the four pilot symbols out of 40 symbols) and
s
= 0.25 (considering
the ten rst symbols). These values depend on the data rate assigned to the user.
A comparison of the ltering eects when considering a full slot average compared
to a fractional slot average is found in Figure 6.5. Aliasing is avoided if the fre
quency components over the Nyquist frequency are ltered out. This is almost the
case when using the local average of the full slot (or measurement period) as in
GSM. Conversely, aliasing eects are most likely when adopting local average over
fractional slots.
10
1
10
0
10
1
0
0.2
0.4
0.6
0.8
1
fT
s
f
N
Figure 6.5 Filtering eect with respect to normalized frequency of a local
average lter applied to the full slot (solid) and fractional slots
s
= 0.25 (dashed) and
s
= 0.1 (dashdotted). The normal
ization is with respect to the frequency after downsampling,
which in this case is 40 times smaller.
6.2 Time Delay Compensation
Time delays aect the stability and performance of any controlled system. Essen
tially, the core problem is that the measurements do not reect the most recent
power updates. However, these are known to the algorithm, and can be compen
sated for. In this section we discuss such a compensation strategy with respect to
the general power control algorithm in Figure 6.1. The resulting scheme is applied
to the important special cases of the FSPC algorithm and loglinear algorithms in
6.2 Time Delay Compensation 93
the preceeding subsections.
Primarily, time delays result in imperfect power control in two dierent ways:
Delayed reactions to changes in external disturbances.
Internal dynamics of the power control loop.
The former typically manifests itself as a phase lag between the input and the out
put, and is not addressed here. Instead the focus is on the latter. The core problem
is that the measurements do not reect the most recent power updates. These are
known to the algorithm, however, and can be used to adjust the measurement.
This is illustrated by the following example.
Example 6.1 (Measurement Adjustments to Compensate for Time De
lays)
Consider the typical case n
p
= 1 and n
m
= 0 and the FSPC algorithm in (5.8).
Using the notation from Section 5.2, this decision feedback algorithm can be
rewritten as
s
i
(t) = sign(
t
i
(t)
i
(t n
m
))
p
i
(t + 1) = p
i
(t) +
i
s
i
(t).
In a delayless situation, the SIR is given by
i
(t) = p
i
(t) +g
ii
(t) I
i
(t).
When subject to time delays as in this case, we instead get
i
(t) = p
i
(t 1) +g
ii
(t) I
i
(t).
Since the output powers are known, the following compensation is straightfor
ward
i
(t) =
i
(t) +p
i
(t) p
i
(t 1) =
i
(t) +
i
s
i
(t 1).
To generalize to the general algorithm in Figure 6.1, we have to consider the
lter F
i
and the power control mechanisms R
i
and D
i
. Assume that the ltering
operation is additive and timeinvariant
1
, i.e.,
Fx(t) +y(t) = Fx(t) + Fy(t), Fq
n
x(t) = q
n
Fx(t).
Thereby, the measurements can be adjusted by generalizing the Smith predictor
(
t
i
(t),
i
(t).
iii) Monitor output powers to be used: p
i
(t + 1) = D
i
s
i
(t).
Remark 6.1
If separate lters are applied to the interference power and the desired received
signal power (see Section 5.2.4), the lter operation to be considered in the above
algorithm is F
g
.
TDC is implemented in the receiver and can be seen as internal feedback as illus
trated by Figure 6.6.
+
p
i
(t)
t
i
(t)
i
(t)
i
(t) R
i
D
i
g
ii
(t) I
i
(t) x
i
(t)
u
i
(t)
q
np
q
nm
H
i
F
i
Receiver Transmitter
Figure 6.6 Block diagram of a receivertransmitter pair when employing a
general power control algorithm with time delay compensation.
TDC is implemented according to Algorithm 6.1 in the device
H
i
and can be seen as internal feedback in the receiver.
The use and applicability of TDC is illustrated by considering the important
situations of the FSPC algorithmand a general loglinear control algorithmutilizing
information feedback (see Sections 4.5 and 5.3 respectively). For clarity, we will
exclude ltering, but it is straightforwardly included.
6.2.1 Fixed Step Power Control Algorithm
Consider the FSPC algorithm with limited dynamic range
p
i
(t + 1) = max
p
min
, min(p
max
, p
i
(t) +
i
sign
t
i
(t)
i
(t)
. (6.3)
6.2 Time Delay Compensation 95
Corresponding components in Figure 6.1 are easily identied. Time delay compen
sation as in Algorithm 6.1 can thus be added as
Algorithm 6.2 (FSPC with TDC I)
i) Adjust measurements:
i
(t) =
i
(t) + p
i
(t) p
i
(t n
p
n
m
).
ii) Issue power control command: s
i
(t) = sign(
t
i
(t),
i
(t)).
iii) Monitor output powers to be used:
p
i
(t + 1) = max (p
min
, min(p
max
, p
i
(t) +
i
s
i
(t))).
In this particular case, the algorithm can be simplied. Consider the monitoring
of power levels (neglecting the limited dynamic range for the time being). It can
be written as
p
i
(t + 1) = p
i
(t) +
i
s
i
(t) p
i
(t) =
i
q 1
s
i
(t)
The measurements are adjusted according to Algorithm 6.1
i
(t) =
i
(t) + p
i
(t) p
i
(t n
p
n
m
).
The last two terms can be rewritten as
p
i
(t) p
i
(t n
p
n
m
) = (1 q
(np+nm)
) p
i
(t) =
i
q
1
1 q
(npnm)
1 q
1
s
i
(t) =
=
i
np+nm
j=1
q
j
s
i
(t) =
i
np+nm
j=1
s
i
(t j). (6.4)
Note that if the step size
i
is adapted and thus timevarying, it cannot be brought
outside the sum. Algorithm 6.2 is thus simplied, since the monitoring of the pow
ers is not necessary. Therefore, we do not need to consider the limited dynamic
range, since only dierential commands are used, and not the absolute power lev
els.
Algorithm 6.3 (FSPC with TDC II)
i) Adjust measurements:
i
(t) =
i
(t) +
i
np+nm
j=1
s
i
(t j).
ii) Issue power control command: s
i
(t) = sign(
t
i
(t),
i
(t)).
To further illuminate the eect of TDC, some block diagram algebra exercise
is instructive. The block diagram of the FSPC algorithm with TDC operating
in closedloop is found in Figure 6.7. Power control command errors are more
naturally described by a multiplicative disturbance than an additive. Then x
i
(t) =
1 corresponds to correct reception of the command bit, while x
i
(t) = 1 indicates
96 Local Design
+
+ ei(t)
pi(t)
t
i
(t)
i(t)
i
q1
i
q1
gii(t) Ii(t)
pi(t)
q
np
q
nm
H(q)
Receiver
Transmitter
Figure 6.7 Time delay compensation (TDC) can be implemented as an
internal feedback by monitoring the powers to be used by the
transmitter, p
i
(t). The transfer function H(q) = 1q
(np+nm)
is obtained from (6.4)
a command bit error. Most of the blocks in Figure 6.7 commute. Hence, the
block diagram is easily rewritten as in Figure 6.8. The merits of TDC are evident,
since the internal roundtrip delays are cancelled in the loop. However, external
signals and disturbances are still delayed, and for example it takes some time before
changes in
t
i
(t) is reected in the measurement
i
(t).
The applicability of TDC to a controller described by the block diagram in
Figure 6.6 is intuitive. However, in some situations, that local loop description is
invalid. A typical example is the uplink when in soft handover. While in operation,
a base station is unaware of whether an issued power control command is applied
or not, since the command from another base station may have been prioritized
(see Section 4.5). Therefore, TDC should be disabled in the uplink while in soft
handover. The applicability in the downlink when in soft or softer handover is
depending on the combining strategy in the receiver.
A similar situation is prevalent when the power command bit error is high. The
measurements are adjusted to reect power levels to be used, but this is incorrect
when subject to bit errors. A simulation study of these eects is provided in
Section 10.4.1
6.2.2 LogLinear Power Control Algorithms with TDC
The class of loglinear power control algorithms utilizing information feedback in
clude several of the proposed algorithms to date (see Section 5.3). They can be
represented by
p
i
(t) = R(q)(
t
i
(t) F(q)
i
(t)).
6.2 Time Delay Compensation 97
PSfrag replacemen
+ ei(t)
pi(t)
t
i
(t)
i(t) i(t)
i
q1
gii(t) Ii(t)
q
np
q
nm
q
nm
1 q
np+nm
H(q)
. .. .
1 + q
np+nm
(1 q
(np+nm)
) = q
np+nm
Figure 6.8 By rewriting the diagram in Figure 6.7, it is evident how TDC
cancels the roundtrip delays in the control loop. External sig
nals and disturbances are still delayed before they are reected
in (t).
The device D
i
may represent quantization of limited dynamic range. Thus Algo
rithm 6.1 is naturally implemented as follows
Algorithm 6.4 (Loglinear Algorithms with TDC)
Adjust measurement by
i
(t) =
i
(t) +F(q)D
i
p
i
(t) F(q)D
i
p
i
(t n
p
n
m
).
TDC for this class of algorithms can thus be seen as an internal feedback H(q) from
the computed power to the control error. If separate lters are used it is given by
H(q) = F
g
(q)(1 q
npnm
) (6.5)
The stabilizing eect of TDC is illustrated by an example.
Example 6.2 (DPC with TDC)
Reconsider Examples 5.1, 5.3 and 5.4, where DPC subject to a delay of one
sample is in focus. With TDC and the more general I controller, the update
algorithm becomes
p
i
(t + 1) = p
i
(t) +
i
(
t
i
(t)
i
(t)), 0 <
i
1.
Considering delayed powers yields
p
i
(t + 1) =p
i
(t) +
t
i
(t) ((p
i
(t 1) +g
ii
(t) I
i
(t)) + p
i
(t) p
i
(t 1))
=
=
t
i
(t)
i
(g
ii
(t) I
i
(t)) + (1
i
)p
i
(t)
98 Local Design
The resulting SIR is thus
i
(t) =
i
q(q 1 +
i
)
t
i
(t)
q
2
(1
i
)q
i
q(q 1 +
i
)
(g
ii
(t) I
i
(t)) ,
which clearly is stable for 0 <
i
< 2. In particular, DPC with TDC yield a
deadbeat controller (all poles at the origin)
i
(t) =
1
q
2
t
i
(t)
q
2
1
q
2
(g
ii
(t) I
i
(t))
Whether this is a good design or not is further discussed in the sequel. For
comparison, the control action when applying an integrating controller without
TDC is described by
i
(t) =
q
2
q +
t
i
(t)
q
2
q
q
2
q +
(g
ii
(t) I
i
(t)) ,
which is unstable for = 1 (cf. Example 5.4).
6.3 Pole Placement Design of LogLinear
Algorithms
Stability of loglinear power control algorithms with respect to controller parame
ters was in focus in Chapter 5. Essentially, a loglinear closed local loop is stable
if and only if all closed local loop poles are within the unit circle. The set of stable
controllers, however, provides rather dierent performance in terms of fading track
ing. Since the performance is related to the pole locations, it is relevant to choose
the controller parameters to carefully place the closed local loop poles. Pole place
ment techniques was applied to design of power control algorithms in (Gunnarsson
et al., 1999d).
Section 6.3.1 describes the considered parameterized controller structures. Prop
erties of linear systems is the main topic in the following section and pole placement
design is applied to power control algorithms in Section 6.3.3.
6.3.1 Controller Structures
The closed local loop with a general loglinear power control algorithm R(q) in
operation is described by the block diagram in Figure 6.9. As in Chapter 5, we
consider an integrating controller
R(q) =
q 1
. (6.6)
6.3 Pole Placement Design of LogLinear Algorithms 99
+ +
+ p
i
(t)
i
(t)
t
i
(t)
g
ii
(t) I
i
(t)
q
np
q
nm
R(q)
F(q)
Figure 6.9 The local loop when employing the general linear control algo
rithm R(q) and the lter F(q).
and the AWWalgorithm
R(q) =
q
. (6.7)
A natural extension of the Icontroller is the PIcontroller (
Astrom
and Wittenmark, 1997; Franklin et al., 1997). Consider a continuoustime system
y(t) = G
c
(p)u(t) =
B
c
(p)
A
c
(p)
u(t),
100 Local Design
where p is the dierentiation operator pu(t) = u(t). The response to changes in
u(t) are described by the locations of the poles s
j
of G
c
(p), i.e. the roots of the
polynomial A
c
(p). As a rule of thumb, the response of the system is considered good
if the poles are located within the shaded area of Figure 6.10a (see e.g., Franklin
et al., 1997, Ch. 2). Valuable information about a specic systems behavior can
be obtained by studying the response to a abrupt step change of the input The
speed of reaction is determined by the pole distance to the origin. Commonly, this
is described in terms of the rise time t
r
, which we will dene as the time it takes
the system to reach the vicinity of of its new set point after the step is applied.
The slightly (or worse) oscillatory behavior during settling, by the angle to the
negative real axis. The poles closest to the origin will dominate the behavior and
are therefore referred to as dominating poles. Let
0
denote the dominating pole
distance to the origin. Given a specic system, the rise time is given by
t
r
=
K
0
,
for some constant K.
When discretizing the continuous system, the corresponding discretetime sys
tem G
d
(q) with sample interval (or update rate) T
s
will have poles at (see
Astrom
and Wittenmark, 1997)
z
j
= e
sjTs
. (6.11)
In addition, the poles will result in a number of zeros, i.e. roots to the numerator
polynomial. However, a good approximation for systems of low order is that the
characteristics are determined by the poles to G
d
(q) and specications can be stated
in the continuoustime domain using (6.11). Mapping the area in Figure 6.10a
using (6.11) yields the area in Figure 6.10b. Discretetime poles can thus be related
to corresponding continuoustime poles in the continuoustime domain, where they
have an intuitive interpretation.
6.3.3 Pole Placement Design
The relations between closedloop poles and performance were discussed above. An
appropriate design procedure is thus to choose the controller parameters to obtain
optimal pole locations. By analyzing the block diagram in Figure 6.9, the resulting
SIR can be written as
i
(t) =
1
1 +q
n
R(q)F(q)
q
np
R(q)
t
i
(t) +g
ii
(t) I
i
(t)
, (6.12)
where n = n
p
+n
m
is the total additional roundtrip delay (in addition to the delay
in R(q)). In steady state (q 1) the stable system converges to
i
(t) =
R(1)
1 +R(1)
t
i
(t) +
1
1 +R(1)
(g
ii
(t) I
i
(t)),
6.3 Pole Placement Design of LogLinear Algorithms 101
since F(1) = 1 when using the lters in Section 5.2.4. Clearly, there will be no
steady state control errors if and only if R(1) = , i.e., if the controller contains
an integrator (the factor (q 1) in the denominator).
A natural optimization criterion is to nd the controller parameters correspond
ing to fastest response (shortest rise time) to changes in target SIR. This is essen
tially an optimal strategy when the power gains are slowly varying. In other words
we maximize the distance from origin to the dominating poles, while remaining
within the shaded area in Figure 6.10b. If we denote the denominator and numer
ator of the general control algorithm R(q) by A
R
(q) and B
R
(q) respectively, the
characteristic polynomial (cf. Section 5.3.2) is given by
P(q) = q
n
A
R
(q)A
F
(q) +B
R
(q)B
F
(q)
2 1 0
2
1
0
1
2
Continuous time
1 0.5 0 0.5 1
1
0.5
0
0.5
1
Discrete time, T
s
= 1 s
b. a.
Figure 6.10 Poles located in the shaded area correspond to acceptable
behavior of the system in continuous time (a) and discrete
time (b). The +:s, *:s and o:s correspond to optimized I
controller, PIcontroller and AAW algorithm respectively. In
these cases no lter is considered ( = 0). Furthermore, the
:s describe the poles when using the Icontroller and a lter
( = 0.5). The angle to the negative real axis is = 0.25
in all cases.
I controller
The SIR expression in (6.12) using the Icontroller in (6.6) reduces to
i
(t) =
(q )
q
n1
(q 1)(q ) +(1 )
q
nm1
t
i
(t) +q
n1
(q 1)(g
ii
(t) I
i
(t))
.
102 Local Design
Consider the typical case n
p
= 1 and n
m
= 0. The characteristic polynomial is
then given by
P
I
(q) = q
2
(1 +)q +(1 ). (6.13)
It is desirable to place the corresponding continuoustime poles within the shaded
area in Figure 6.10a, i.e.,
s
1
, s
2
= re
i
, re
i
, 0 0.25, r r
0
.
where r
0
constitute a lower limit on the rise time. In this case we have heuristically
chosen r
0
= 0.5. The desirable discretetime characteristic polynomial is thus given
by
P
d
(q) =(q e
s1Ts
)(q e
s2Ts
) =
=q
2
q2e
rTs cos
cos(rT
s
sin) +e
2rTs cos
(6.14)
By equating the coecients of Equations (6.13) and (6.14), the following require
ments are obtained
= e
2rTs cos
(6.15a)
1 + = 2e
rTs cos
cos(rT
s
sin ) (6.15b)
The right hand side of (6.15b) is monotonically decreasing with increasing r. Thus,
any ltering decreases the maximal pole distance to the origin. If rise time is the
major concern, ltering should be excluded. However, if the lter represents a
model of an estimating procedure, is known. Thus, the optimal controller pa
rameter is obtained by solving (6.15b) for rT
s
, which in turn is inserted in (6.15).
Note that the optimal controller parameter is independent of T
s
, which is natural.
If we omit ltering and assign the poles on the border of the area, i.e., = 0.25,
then the optimal controller parameter is
I
= 0.340.
For comparison we consider ltering (exponential forgetting, = 0.5). Analogous
optimization yields
I,F
= 0.202.
The corresponding poles are plotted in Figure 6.10b. In particular, we note that
ltering result in a signicant degradation in terms of rise time. It is even slower
than the heuristically chosen lower rise time limit.
6.3 Pole Placement Design of LogLinear Algorithms 103
PI controller
The optimization when using a PIcontroller is more or less analogous. Equa
tions (6.12) and (6.9) yield that SIR is given by
i
(t) =
q
nm1
(( +)q )(q )
q
n
(q 1)(q ) + (1 )(( +)q )
t
i
(t)+
+
q
n
(q 1)(q )
q
n
(q 1)(q ) + (1 )(( +)q )
(g
ii
(t) I
i
(t))
e
rTs
+ 2e
rTs cos
cos(rT
s
sin )
+
+q
2e
rTs(1+cos )
cos(rT
s
sin )e
2rTs cos
e
rTs(1+2 cos )
(6.17)
Equating the coecients of Equations (6.16) and (6.17) result in a set of require
ments. Just as in the previous case, ltering only decreases the pole distance to
the origin. Omit ltering and use = 0.25 yield the optimal parameters
PI
= 0.0720,
PI
= 0.365
with corresponding poles in Figure 6.10b. These poles are located further out from
the origin compared to when using an Icontroller. At a cost of an extra parameter,
we have obtained a faster closed local loop. In this typical case, the system is
essentially the time delay. For longer time delays and maybe with a more complex
model of the estimating device, the ideas in (Kristiansson and Lennartsson, 1999)
may apply for PI design.
AAW algorithm
This case is rather similar to when designing the Icontroller. The resulting SIR
using the AAW transfer function in (6.7) is given by
i
(t) =
q
nm1
(q )
q
n1
(q )(q ) +(1 )
t
i
(t) +q
np
(q )(g
ii
(t) I
i
(t))
.
104 Local Design
In the specic case of a single additional delay, the computations are analogous to
the Icontroller and the optimal parameter is
AAW
= 0.0591.
This choice of parameter value yields the pole locations as in Figure 6.10b. The
poles are located even further out from the origin compared to when using the
previous controllers, proving a shorter rise time.
Comments
A procedure for optimizing the parameters in three dierent controller structures
has been outlined. It was noted that with careful design, AAW provides faster
responses than PI, which in turn is faster than I. If fast responses would be the
only concern, AAW seems to be the most appealing structure. However, since the
AAW algorithm is operating without integral action, it is not as suitable as an inner
loop. As pointed out in Section 5.2.1, the desired power level adaption of the target
SIR can be considered in an outer loop. This is further discussed Section 9.4.3.
A more complex controller structure provides more degrees of freedom and
moreover the ability to assign any closedloop poles. However, such a scheme would
introduce several zeros (roots of the numerator polynomial), which also aect the
performance.
6.4 Practical Design Aspects
An underlying assumption in the previous section is that the power gains are vary
ing slowly. However, this might be too simplifying when considering a real system.
Typically, such disturbances are roughly known in the frequency domain as briey
described in Section 6.1. Therefore, disturbance rejection properties are naturally
studied in the frequency domain. Furthermore, as will be shown in Chapter 7,
there are relations between global stability and local properties, which should be
addressed in the design. The focus is on inner loops tracking target SIR:s provided
by an outer loop. Thus, the AAW algorithm is not considered.
6.4.1 LogLinear Power Control Algorithms
As disclosed in the previous section, the resulting SIR in Figure 6.9, can be written
as
i
(t) =
q
np
R(q)
1 +q
n
R(q)F(q)
t
i
(t) +
1
1 +q
n
R(q)F(q)
g
ii
(t) I
i
(t)
, (6.18)
Time delay compensation, described in Section 6.2, can be applied to cancel round
trip delays in the local loops. Then, the SIR expression reduces to
i
(t) =
q
np
R(q)
1 +R(q)F(q)
t
i
(t) +
1 +R(q)F(q)(1 q
n
)
1 +R(q)F(q)
(g
ii
(t) I
i
(t))
. (6.19)
6.4 Practical Design Aspects 105
The relations between SIR on the one hand and target SIR and external distur
bances on the other can thus be associated with two central transfer functions
i
(t) = G
ll
(q)
t
i
(t) +S(q)(g
ii
(t) I
i
(t)), (6.20)
where G
ll
(q) ans S(q) will be referred to as the closed loop system and the sensi
tivity function respectively. We only consider controllers with integral action, i.e.,
with the factor (q 1) in the denominator. As is evident from Equations (6.18)
and (6.19), a factor (q 1) in the denominator of R(q) result in the factor (q 1)
in the numerator of S(q) This can be seen as a dierentiation, whose eects will be
discussed further on. We consider the typical case n
p
= 1, n
m
= 0 when nothing
else is stated. The controllers in focus are
1. The DPC algorithm in a delayless case, which yields
i
(t) =
1
q
t
i
(t) +
q 1
q
(g
ii
(t) I
i
(t)).
2. The optimized Icontroller in the previous section
i
(t) =
I
P
I
(q)
q
1
t
i
(t) + (q
2
q)(g
ii
(t) I
i
(t))
,
where P
I
(q) is given by (6.13).
3. The optimized PIcontroller in the previous section
i
(t) =
1
P
PI
(q)
((
PI
+
PI
)q
PI
)
t
i
(t) + (q
3
q
2
)(g
ii
(t) I
i
(t))
,
where P
PI
(q) is given by (6.16).
4. The integrating controller with = 0.5 and TDC. The SIR expression is
provided in Example 6.2.
5. As above, but with = 1 (i.e., DPC) to obtain a deadbeat controller.
We summarize the closedloop systems and sensitivity functions of the considered
controllers in Table 6.1. The dierentiation in the sensitivity function yields that
the controlled system is insensitive to the long term mean of the fading and in
terference. However, the dierentiation amplies high frequency components and
should therefore be balanced by at least one welldamped pole if high frequency
components are likely. The transfer functions G
ll
(q) and S(q) are plotted in the
frequency domain in Figure 6.11. The discretetime and continuoustime poles are
related via Equation (6.11).
z
j
= e
sjTs
.
Note that discretetime poles at the origin, lack counterparts in the continuous
time domain (such poles would have been located at the innity). If all closedloop
poles are placed at the origin, the controller is referred to as a deadbeat controller.
The reason is that the system is at rest when the desired output is reached. A
possible drawback is that unnecessarily large may be used. From these plots, a
number of interesting conclusions can be drawn.
106 Local Design
Controller Closedloop system Sensitivity function
1 G
ll
(q) =
1
q
S(q) =
q1
q
2 G
ll
(q) =
I
q
2
q+I
S(q) =
q
2
q
q
2
q+I
3 G
ll
(q) =
(PI+PI)qPI
PPI(q)
S(q) =
q
3
q
2
PPI(q)
4 G
ll
(q) =
0.5
q(q0.5)
S(q) =
(q1)(q+0.5)
q(q0.5)
5 G
ll
(q) =
1
q
2
S(q) =
(q1)(q+1)
q
2
Table 6.1 Closedloop systems G
ll
(q) and sensitivity functions S(q) of con
trollers studied in this section. The parameters from the pole
placement in Section 6.3.3 are
I
= 0.340,
PI
= 0.0720 and
PI
= 0.365, and the polynomial P
PI
(q) is obtained from (6.16).
Disturbance Rejection
The disturbances are considered rejected at frequencies where the sensitivity func
tion is small in Figure 6.11b. DPC (with TDC if needed) is thus preferable if the
disturbances primarily are of low frequency. On the other hand, the unbalanced
dierentiation of DPC amplify highfrequency components of the disturbances.
Such high frequency components arise naturally, when considering quantized pow
ers, since the additive quantization noise can be treated together with the additive
disturbance (cf. the introduction to Chapter 5).
The fourth studied controller (Icontroller with TDC and = 0.5) is motivated
by similar disturbance rejection as the optimized PIcontroller. Higher value yield
better low frequency disturbance rejection, but worse performance at intermediate
frequencies and vice versa. If we focus on frequencies attenuated by at least 0.5,
we see that normalized frequencies up to 0.03 (controllers 2, 3 and 4) and 0.04
(controller 5) are rejected.
Of course, it would be desirable to design the controller to obtain a small
sensitivity function at all frequencies. This is impossible, however. According
to Sung and Hara (1988), the sensitivity function in this case
2
has to meet the
following requirement
0
ln
S(e
iw
dw = 0.
Hence any frequency region where the sensitivity function is small (negative in
logarithmic scale) has to be compensated by a region where it is large (positive in
logarithmic scale).
2
When the open loop system have unstable poles, the requirement is slightly dierent.
6.4 Practical Design Aspects 107
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0.2
0.4
0.6
0.8
1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0.5
1
1.5
2 a.
b.
fT
s
Figure 6.11 The magnitude of the closed loop system G
ll
(q) (a) and sen
sitivity function S(q) (b) of the considered controllers, with
respect to normalized frequency. The curves correspond to
controller 1 (solid), 2 (dashed), 3 (dashed +), 4 (dashdotted)
and 5 (dotted).
Controller Bandwidth
The closed loop system G
ll
(q) captures how fast the inner loop adapts to changes
in target SIR. Essentially, fastest possible reactions are provided by deadbeat con
trollers 1 and 4. Normally, the controller bandwidth is dened as the normalized
frequency where the gain is half the DC gain (at q = 1). As seen in Figure 6.11,
the bandwidth of controller 2 is with this denition approximately 0.17 and of con
trollers 3 and 4 is approximately 0.22. Controllers 1 and 5 are deadbeat controllers
and have in this case unit gain.
The target SIR is normally updated at a slower rate than the inner loop. It is
therefore subject to abrupt steps, which implies high frequency components in the
input signal. Despite a low gain of the closed loop system at high frequencies, the
inner loop may overreact to these sudden changes. An alternative is to prelter
the target SIR using a simple low pass lter, e.g., an exponential forgetting lter.
Global Stability
The global system consists of local loops interconnected by the interference powers.
This will be studied thoroughly in Chapter 7, but it is instructive to consider one
important result from Theorem 7.20. The global system is stable if the closed loop
108 Local Design
system G
ll
(q) of all local loops fullls
sup
w
[G
ll
(q)[
q=e
iw 1.
Thus all ve controllers meet the global stability requirement.
Sample Interval
The sample interval, together with the mobile velocity, primarily relate the distur
bance rejection properties to the fading characteristics in Section 6.1. The shorter
sample interval, the better rejection, but also the more feedback bandwidth re
quired. Consider the shadow fading model, with low frequency content up to
1/x
1
= 0.07 [ m
1
] and the remainder up to 1/x
2
= 1.7 [ m
1
]. For a low velocity
(1 m/s) and using controller 3 above (rejects disturbances up to 0.03/T
s
Hz), we
need at least T
s
= 0.5 s to track the slowest shadow fading, and T
s
= 0.015 s to
track the faster one. At other velocities, the required sample intervals are divided
by the mobile velocity v. These gures should be related to the current sample
interval in GSM, T
s
= 0.5 s (Steele, 1992). In such a system, power control can
essentially be used to track path loss and possibly the slowest shadow fading. The
reminder has to be mitigated by fading margins and appropriate coding (see Chap
ter 9). The situation is better when considering the WCDMA proposal (Chaudhury
et al., 1999; Swarts et al., 1999; UMTS 30.06, 1997), where a sample interval of
T
s
= 1/1500 s is suggested. Assume that information feedback and an I controller
with TDC is used. That corresponds to disturbance rejection up to 0.04 1500 Hz
= 60 Hz, which is by far sucient to track the fast multipath fading in the Vehicu
lar A model (see Section 6.1.2). Moreover, if a measurement procedure as discussed
in Section 6.1.4 is employed, aliasing has to be considered in this analysis as well.
In such a case, excess bandwidth provided by a high sample interval is valuable.
Measurement Errors
Filtering is generally excluded in the thesis examples. As disclosed in Section 6.3.3,
ltering the measurements result in slower rise times. If that is a major concern,
the lter should be excluded if possible. However, the lter can for example be
motivated in a situation where we cannot neglect measurement errors. We model
the available measurements as subject to additive measurement errors e
m
(t).
m
(t) = (t) +e
m
(t),
where
m
(t) denote the available measurements. Consider an Icontroller with an
exponential forgetting lter, in the typical case n
p
= 1, n
m
= 0. Straightforward
computations (cf. Equation (6.18) and Figure 6.9) yield that the resulting SIR is
described by
i
(t) =
(q )
qP
I,F
(q)
t
i
(t) +
(q 1)(q )
P
I,F
(q)
(g
ii
(t) I
i
(t))
(1 )
P
I,F
(q)
e
m
(t) =
= G
ll
(q)
t
i
(t) +S(q)(g
ii
(t) I
i
(t)) +T(q)e
m
(t),
6.4 Practical Design Aspects 109
where the P
I,F
(q) is given by
P
I,F
(q) = (q 1)(q ) +(1 ).
The transfer function is referred to as the complimentary sensitivity function T(q)
and describes the measurement noise contribution to the resulting SIR. Since mea
surement errors typically include highfrequency components, it is desirable that
T(q) is small for those frequencies.
In this comparative study, we will use the two Icontrollers that were designed
with and without a lter in Section 6.3.3. Figure 6.12a illustrates the corresponding
magnitudes of the closedloop system. function. As expected, the controller band
width is better without a loop lter. However, the benets of ltering are evident
in Figure 6.12b, where the complimentary sensitivity function is much smaller at
high frequencies when using a lter. Hence, ltering should be considered, when
measurement errors is a major concern. A similar result is provided in (Su and
Geraniotis, 1998), where the eects from quantized measurements are in focus. The
authors minimized the power control error in simulations by a linear search over
the lter parameters, and concluded that ltering improved the performance.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0.2
0.4
0.6
0.8
1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0.2
0.4
0.6
0.8
1
a.
b.
fT
s
Figure 6.12 The magnitude of the closed loop system G
ll
(q) (a) and com
plimentary sensitivity function S(q) (b) of the Icontrollers,
with (dashed) and without (solid) a loop lter ( = 0.5).
6.4.2 Decision Feedback Algorithm
Decision feedback algorithms are characterized by a relay. This static nonlinearity
makes a bandwidth discussion more awkward. One approach could be to state that
it result in similar behavior as the corresponding information feedback algorithm.
110 Local Design
Clearly, the discrepancies would be considerable in some cases. For example, = 1
result in an unstable information feedback algorithm. The problem of analyzing the
performance is considered by Song et al. (2000), and we only include a brief review.
In essence, they replace the relay by a constant K
r
and an additive disturbance w
r
that minimize
E
(f(e) K
r
e)
2
,
given that
Varf(e) = VarK
r
e +w
r
.
Using a correlated shadow fading by Gudmundson (1991), the authors thereby
relate the performance to the speed of the mobile and the sample interval. The
conclusion in case of IS95, with T
s
= 1/800 s, is that FSPC satisfacory compen
sates for the shadow fading, but fails to track the fast fading except for very low
velocities.
6.4.3 Summary, Design Aspects
Generally, we have more design options when information feedback is possible. In
both cases, the ability to mitigate timevarying disturbances naturally depend on
the sample interval. If the spatial correlation of the disturbances can be modeled,
the expected mobile velocities also come into play to describe the corresponding
temporal correlation.
In case of information feedback it is natural to address disturbance rejection
and controller bandwidth in the frequency domain and with respect to the closed
loop system and the sensitivity function respectively. The sensitivity function also
captures the eects of quantization noise. If fastest possible response to updated
target SIR:s is the major concern, then loop ltering should be excluded. Con
versely, if the measurement are subject to considerable errors, ltering should be
considered to better suppress associated highfrequency components.
The stepwise updates of target SIR contains highfrequency components. To
prevent overcompensation, the target SIR:s should be lowpass ltered by employ
ing a prelter. When taking all these components into consideration, the local loop
can be described by the block diagram in Figure 6.13.
6.5 Predictive Power Control
Currently, we have almost exclusively used only the measurements as they are
for power control. An exception is TDC, which was introduced in Section 6.2.
However, improved performance can be gained by considering models of the dis
turbances to either predict the disturbances, as in Section 6.5.1, or to design the
appropriate controller as in Section 6.5.2.
6.5 Predictive Power Control 111
+ +
+
e
m
(t)
p
i
(t)
i
(t)
m
i
(t)
t
i
(t)
g
ii
(t) I
i
(t)
q
np
q
nm
R
F(q)
F
p
(q)
Figure 6.13 The local loop dynamics when employing the general linear
control algorithm R, the loop lter F(q) and the prelter
F
p
(q). The measurements are subject to additive errors e
m
(t).
In case of information feedback, the controller is a linear trans
fer function R = R(q). The only dierence when using deci
sion feedback is that controller block contains a sign function
at the input R = R(q)sign (e(t)). This local loop description
is slightly dierent from Figure 6.1.
6.5.1 Disturbance Prediction
Assume that the additive disturbance (power gain) can be modeled by ltered
white noise.
g
s
(t) =
B
g
(q)
A
g
(q)
e
s
(t), Vare
2
s
(t) =
2
e
, (6.21)
where e
s
(t) is zeromean and white. If an appropriate sampling interval is used, the
shadow fading model in Section 6.1.1 is an applicable model. However, for other
sample intervalvelocity relations, and when the averaging in the measurement
procedure in Section 6.1.4 is considered, other models are relevant. Due to the
uncorrelated samples from the noise sequence, we can without loss of generality
assume that A
g
and B
g
are of the same degree.
The parameterized model is tted to identication data either oine or re
cursively online (Ljung, 1999). Popular and implementationally simple recursive
algorithms include LMS and RLS (Gustafsson, 2000; Ljung, 1999). In these ap
proaches, simple models of parameter variations are considered. The tracking per
formance can be improved, by considering more relevant models. This is disclosed
in Ahlen et al. (2000); Lindbom (1993), where WLMS algorithms are introduced
and evaluated.
Linear algorithms are applied to mobile communications channels to improve
power control by e.g., Ericsson and Millnert (1996); Leung (1999); Mehta et al.
(1996); Moh et al. (1999). However, some of them fails to fully utilize the capa
bilities of modelbased predictions. Various nonlinear algorithms are considered in
(Ekman and Kubin, 1999; Tanskanen et al., 1998; Zhang and Li, 1997).
112 Local Design
Assume that the model in (6.21) is provided by some algorithm. For clarity, we
consider the standard notation
g
s
(t) =
C(q)
A(q)
e
s
(t)
In the typical case n
p
= 1 and n
m
= 0 and we are interested in predicting the
disturbances two samples ahead. If the polynomial C(q) is stable, the minimum
variance predictor over m steps is obtained from e.g.,
Astrom and Wittenmark
(1997, Thm. 12.1) as
g
s
(t +m[t) =
qG(q)
C(q)
. (6.22)
The polynomial G(q) is obtained as the solution to the following Diophantine equa
tion
q
m1
C(q) = A(q)F(q) +G(q), (6.23)
where degF = m1 and degG = degA 1.
The predicted disturbance may be used by the controller. However, it is impor
tant that the recursive estimation procedure has converged or at least is known to
provide reasonably accurate parameter estimates. The benets of using disturbance
prediction in power control is illustrated by simulations in Section 6.7.
6.5.2 MinimumVariance Control
When a disturbance model is known, this information can be utilized in the power
control design to obtain a controller that adapts to the disturbance. One strategy is
to control the power to obtain minimumvariance of the control error
t
i
(t)
i
(t).
The outline here will be rather brief. Further details can be found in (
Astrom and
Wittenmark, 1997).
Consider systems of the form
y(t) =
B(q)
A(q)
u(t) +
C(q)
A(q)
e(t), (6.24)
and assume that C(q) is a stable polynomial (the controller in the general case is
described by
Astrom and Wittenmark). Let d denote the pole excess, i.e., d =
degA degB. Furthermore, assume that the reference signal is zero. Then
the minimum variance controller is obtained as
u(t) =
G(q)
B(q)F(q)
, (6.25)
where F(q) and G(q) are obtained from the Diophantine equation (6.23) with
m = d.
6.6 Nonlinear Components 113
Consider a power control problem where the received power is only aected by
an additive power gain, modeled as in (6.21). The output powers are subject to a
delay of only one additional sample. The system to be controlled can thus be seen
as
i
(t) =
B
r
(q)
A
r
(q)
p
i
(t) +
B
g
(q)
A
g
(q)
e
s
(t) =
B
r
(q)A
g
(q)
A
r
(q)A
g
(q)
p
i
(t) +
B
g
(q)A
r
(q)
A
r
(q)A
g
(q)
e
s
(t)
B(q)
A(q)
p
i
(t) +
C(q)
A(q)
e
s
(t),
which is of the same form as Equation (6.24). The minimumvariance controller
is not strictly proper, which is required. Therefore, we consider two delays in the
system model, i.e., A
r
(q) = q
2
and B
r
(q) = 1, and then associate one of the delays
with the controller.
Furthermore, we require integral action in the local loop. As pointed out by
, (6.27)
where T
t
is the tracking parameter. If T
t
= T
s
, the state will exactly correspond to
the actual output. Note that when K
p
= 0, K
i
= 1, T
t
= T
s
and f() is a saturation,
we obtain the Distributed Constrained Power Control (DCPC) algorithm (Grandhi
et al., 1995) as a special case.
A more complicated power control algorithm of higher order (see e.g., (Gun
narsson et al., 1999d) or Section 6.5.2) needs special treatment. A general linear
algorithm can be described in state space form as
x
i
(t + 1) = Ax
i
(t) +By
i
(t)
p
i
(t) = Cx
i
(t), (6.28)
where A(n n), B(n m), C (m n), are matrices of given orders, y
i
(t) is the
measurement vector, and n is the dimension of the state vector. Note that the
algorithm is strictly proper. The order of the input vector m is in a typical power
control case equal to one (e.g., y
i
(t) = e
i
(t)) or two (e.g., y
i
(t) =
t
i
(t)
i
(t)
T
).
An observerbased antiwindup scheme (see Kothare et al. (1994) for details) can
be implemented as
x
i
(t + 1) =A
i
x
i
(t) +By
i
(t) +L
f (p
i
(t)) Cx
i
(t)
p
i
(t) =Cx
i
(t), (6.29)
where the (n m)matrix L is a design parameter.
The static function f() should not only incorporate the inherent nonlinearities,
but also all the intentional nonlinearities, such as those discussed in the rest of the
section.
6.6 Nonlinear Components 115
6.6.2 Selector or Switch
Sometimes it is desirable to employ several algorithms in parallel, and then use a
device to state priorities, i.e., to switch between the algorithms. Such a nonlin
earity is commonly referred to as a selector or a switch, depending on scientic
background.
As an example, consider the problem of an integrating power control algorithm
at low interference levels. A SIR balancing algorithm would compute a very low
power to use when the interference is low. The system requirement that a user
should use a low power in order not to disturb others does not hold, and instead the
user requirement of using a high power to obtain good quality should be prioritized.
An implementation is proposed in (Blom and Gunnarsson, 1998; Gunnarsson et al.,
1999b), where the estimated mean interference is used to adapt the minimum
allowable power level. When few users are in the network resulting in a low mean
interference, the minimum allowable power level is increased and vice versa. This
is implemented by using a maximum selector as
p
i
(t + 1) = max (p
R,i
, h(m
I
(t))) ,
where p
R,i
(t + 1) is computed by a nominal power control algorithm and h() is a
function that maps the measured or estimated mean interference onto a minimum
power level.
To rectify the purpose of the inner loop, we pose that the focus is to assign power
levels to track target SIR:s, provided by the outer loop. Therefore, adaptions and
priorities as above, should be placed in the outer loop.
However, there are situations when switches in the inner loop can be moti
vated. For example, consider the control strategies in Section 6.5, which are based
on disturbance prediction. When the estimation algorithm is not tracking the dis
turbance model parameters satisfactory, such schemes may result in poor control.
A relevant strategy could be to only adopt the prediction based methods when the
estimated parameter variance is less than a threshold. Hence, a mode switching is
benecial, where we switch between utilizing estimated parameters (e.g., for distur
bance prediction or minimumvariance control) and a controller as in Sections 6.2
and 6.3.
When employing switching between dierent algorithms, the idle algorithms
have to be updated on the current situation. Essentially, the situation is the same
as above, where output constraints aected the output. Analogously, antireset
windup techniques can be used to update the internal states when the actual output
is computed by a dierent control algorithm. This is referred to as bumpless transfer
i
(
t
)
[
d
B
]
i
(
t
)
[
d
B
]
t [s] t [s]
Figure 6.14 Performance of dierent power control algorithms when sub
ject to shadow fading and the typical delay situation n
p
= 1,
n
m
= 0. a. The PIcontroller from pole placements design
with (solid) and without (dashed) prelter. b. The same con
troller at T
s
= 0.015 (solid) and T
s
= 0.05 (dashed). c. An
optimized PIcontroller (solid) and an Icontroller with TDC
(dashed). d. The FSPC algorithm (stepsize 1 dB) with (dark
gray) and without (light gray) TDC.
Sample Interval
The sample interval aects both the controller bandwidth and the disturbance re
jection, as discussed in Section 6.4. The recommendation is that T
s
= 0.015 s is
needed to track the shadow fading. For comparison, the optimized PIcontroller
from pole placement design in operation with T
s
= 0.015 s and T
s
= 0.05 s is
considered. As seen in Figure 6.14b, this clearly has an aect on the disturbance
rejection. The recommendation T
s
= 0.015 is considered in the rest of the simula
tions.
Time Delay Compensation
As seen in Section 6.2, time delay compensation helps to stabilize controllers. How
ever, when information feedback is used, the same performance can be obtained by
118 Local Design
careful design. This is illustrated in Figure 6.14c, where the optimized PIcontroller
and an Icontroller with TDC yield almost identical performance.
The situation is dierent when considering controllers utilizing decision feed
back, such as the FSPC algorithm. Then, there are not that many other options
available. The dynamical eect of TDC is reduced oscillative behavior, which is
discussed using describing functions in Section 5.5. Figure 6.14d further illustrates
these oscillations and the benets of TDC.
Predictive Power Control
If a model of the disturbance (power gain) is at hand, it can be used to predict
the gain at future time instants. Section 6.5.1 provides a discussion on methods
to acquire such models. Here, the model is obtained oline, but from a dierent
shadow fading realization (identication data). Note that the model not necessar
ily is identical to the shadow fading model, since the model has to describe the
disturbance with respect to the sample interval T
s
.
10 20 30 40 50 60 70 80 90 100
8
9
10
11
12
13
14
10 20 30 40 50 60 70 80 90 100
8
9
10
11
12
13
14
a.
b.
i
(
t
)
[
d
B
]
i
(
t
)
[
d
B
]
t [s]
Figure 6.15 Predictive control improve the performance by predicting the
disturbances. a. PIcontroller with disturbance prediction
(solid) and a minimumvariance controller (dashed). b. The
FSPC algorithm with TDC and disturbance prediction.
These predictions are used together with the PIcontroller discussed above, and
the FSPC algorithm. Figure 6.15 indicates signicant improvements using both
algorithms. The optimal predictor given a model is utilized in Section 6.5.2 to
design a minimumvariance controller.
As seen in Figure 6.15a, the minimumvariance controller outperforms the other
studied controllers. Note, however, that the estimation procedure needed to acquire
the model is not considered. In a realistic case, the channel is not stationary, and
6.8 Summary 119
has the be recursively estimated. Yet another problem of the designed minimum
variance controller arises when considering global performance. Figure 6.16 depicts
the magnitude of the corresponding closedloop transfer function G
ll
(q). Hence,
the global stability requirement G
ll
(q) 1 in Section 6.4 is not met. However, as
will be disclosed in Chapter 7, the controller can still be applied to slightly loaded
systems, where the requirements on G
ll
(q) are relaxed.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
5
10
15
20
25
fT
s
Figure 6.16 The magnitude with respect to normalized frequency of the
closed loop system G
ll
(q) when using a minimumvariance
controller.
6.7.2 MonteCarlo Simulations
The same simulations as above are rerun 50 times using dierent shadow fading
realizations. As performance measure, we use the root mean squared error (RMSE)
1
50
50
i=1
1
N
N1
t=0
(
t
i
(t)
i
(t)).
The results are summarized in Table 6.2. In essence, the comparative conclusions
from the single realizations above hold.
The eect of the prelter is marginal on average, but as seen in Figure 6.14a,
it makes a clear dierence. It is further emphasized that TDC is not as impor
tant when information feedback is available, since the corresponding eect can be
obtained by linear design. Considering FSPC, however, the benets are signicant.
Further improvements are observed when considering disturbance prediction
and minimumvariance control. However, the eect of model estimation is not
included.
6.8 Summary
Stability is a necessary property of closedloop systems, but dierent stabilizing
controllers provide very dierent behavior. Therefore, methods from control theory
120 Local Design
Controller RMSE
PIcontroller, no prelter 0.27
PIcontroller 0.26
PIcontroller, T
s
= 0.05 s 0.89
Icontroller with TDC 0.33
FSPC algorithm 0.62
FSPC algorithm with TDC 0.38
PI with dist. pred. 0.19
FSPC with TDC and dist. pred. 0.30
Minimumvariance controller 0.081
Table 6.2 RMSE when using dierent control strategies. A prelter is
employed and the sample interval T
s
= 0.015 s is used if nothing
else is stated.
are applied to discuss controller design with respect to performance. Performance
in a completely general case is hard to determine. Valuable information is provided
by disturbance statistics and models, which facilitates the performance discussion.
Such models need to capture the spatial correlation of the disturbances, i.e., the
inuence of the environment on the received signal powers.
Outdated information was seen as a limiting factor in Chapter 5. The core
problem is that the measurements do not reect the most recent power updates.
Since the computed powers are known in the receiver, these can be used to adjust
the measurements in order to compensate for time delays. The proposed scheme is
discussed both with respect to a general control algorithm and applied to existing
ones.
The relations between stability and pole locations of the system are discussed
in Chapter 5. Therefore, an alternative design strategy is to chose controller pa
rameters to place the poles. Performance is naturally discussed with respect to
controller bandwidth and disturbance rejection. The former relates to how fast
the controller responds to modied target SIR:s, while the latter focuses on the
controllers ability to mitigate disturbances. Disturbance rejection can thereby be
parameterized by the sample interval and the velocity of the mobile station. The
performance evaluation also relates to the global stability requirements in Chap
ter 7, where global stability is related to local loop requirements.
Not all situations are readily handled by linear controllers. In some situations,
nonlinear components may improve the performance or attain a desirable behavior.
For example, output power constraints may result in internal discrepancies between
the output and the internal state of the controller. Such situations can be handled
by antireset windup techniques, which are discussed. In other situations, dierent
linear controllers are relevant in dierent cases, which can be seen as a nonlinear
switch between the algorithms.
6.8 Summary 121
The dierent design and control strategies in this chapter provide very dierent
performance. Some illustrative simulations conclude the chapter, to emphasize
similarities and dierences. MonteCarlo simulations over a large number of cases
stress the benets of prediction and the use of disturbance models in the controller
design.
General guidelines are hard to disclose. When using information feedback,
time delays can be compensated for by either pole placement design or time de
lay compensation. On the contrary, there is no natural alternative to time delay
compensation in case of decision feedback. If it is possible to adapt models to dis
turbances with good precision, this information should be utilized for disturbance
prediction independent of feedback strategy. In addition, disturbance models are
useful in slightly loaded systems, to signicantly improve the performance using
minimumvariance control.
122 Local Design
7
Global Analysis
For practical reasons, power control algorithms in cellular radio systems are im
plemented in a distributed fashion. The analysis of these local control loops in
Chapter 5 provide valuable insight in stability and performance issues. However,
the local loops are interconnected via the interference between the loops, which
aect the global dynamics. Therefore, as disclosed in Chapter 5, the local analy
sis only provide necessary, but not sucient, conditions on stability. To capture
the dynamical eects of the interconnections, the overall global system has to be
addressed in the analysis.
An important global discussion is whether it is possible to assign output powers
to meet each users requirement. This is the main issue in Section 7.1, where
we also introduce metrics to describe available capacity and current load in the
network. Proposed approaches for convergence analysis of power control algorithms
primarily consider the algorithms in linear scale. Section 7.2 provide some results
on algorithms in linear scale. Additional time delays have generally been omitted
in convergence analysis. However, here time delays are considered for two classes
of algorithms. In Section 7.3 we employ a robust control framework to analyze the
global system in logarithmic scale. Thereby, sucient conditions on global stability
are derived, with respect to an approximative global system. These results apply to
general loglinear local control loops, including the eect of time delays and general
loglinear power control algorithms and lters. These conditions can be formulated
as requirements on the local loops. The interesting conclusion is thus that global
123
124 Global Analysis
stability can be granted by proper design of the local loops (cf. Chapter 6).
7.1 Optimal Global Power Assignments
Optimal power assignments were briey discussed in Section 4.5 as means to cen
trally control the transmission powers. Mainly, the interest is motivated by the
relations between these algorithms and capacity and performance bounds of dis
tributed algorithms. Here, we generalize the results to include autointerference
and individual target SIR:s. Essentially, at this global level, we assume that the
inner loops perfectly tracks the fast variations in the signal power gains, so that the
power gains can be considered constant. Then, the capacity and load discussion
mainly relates outer loops and whether it is possible to accommodate all users with
their service requirements (in terms of target SIR). Two metrics are introduced to
describe the available capacity and current load in the network. In the following
discussion only values at the same time instant t are considered, and therefore this
time index will be suppressed for clarity. Also adopt the convention that vector
inequalities, such as x y, hold if they hold elementwise.
7.1.1 Optimal Balancing Assignments
As disclosed in Section 4.3, the SIR at each receiver is a simplied measure of the
perceived quality. Assume that the quality is acceptable when this ratio is above
a certain threshold, i.e.,
i
=
i
g
ii
p
i
j=i
g
ij
p
j
+
p
i
g
ii
+
i
0
, i. (7.1)
An intuitive optimization strategy given a xed set of users, is to maximize the
smallest SIR in the network. This question is raised in (Grandhi et al., 1993;
Zander, 1992b) and some central results are covered here. The autointerference in
Equation (7.1) was omitted (i.e.,
i
= 1, i) in the analysis. Introduce the matrix
Z = [ z
ij
]
=
g
ij
g
ii
, p
= [ p
i
] , = [
i
]
=
i
g
ii
.
Note that the matrix
Z is a nonnegative matrix (i.e., all entries are nonnegative).
Furthermore,
Z has a unit diagonal, and if the links have been established between
the transmitters and receivers with the best power gain, the odiagonal elements
are less than one. Using the introduced quantities, Equation (7.1) can be rewritten
as
p
i
j
z
ij
p
j
p
i
+
i
0
, i. (7.2)
1 +
0
0
p
i
j
z
ij
p
j
i
(7.3)
7.1 Optimal Global Power Assignments 125
The corresponding equations for all users can be represented in vector notation.
1 +
0
0
p
Z p + . (7.4)
It is thus interesting to study solutions to (7.4) and their existence, with respect
to
0
. First, focus on the noiseless case, i.e.,
i
= 0, i. Since the elements of
Z
can be regarded as stochastic variables,
Z will have full rank with probability one.
Denote the eigenvalues of
Z by
1
, . . . ,
m
. The following results from the work
of Perron, Frobenius and Wielandt on positive matrices (Gantmacher, 1974):
Z has a positive real eigenvalue
= max[
i
[
m
i=1
and the corresponding
eigenvector is positive.
The minimum real
such that the inequality
Z p
p
has a solution for p 0 is
=
.
Bounds on the eigenvalues of
Z derived from row sums of
Z imply that
> 1.
Zander (1992b) applied these results to Equation (7.4) to prove the following the
orem,
Theorem 7.1 (Zander, 1992b)
With probability one, there exists a unique maximum achievable SIR in the noise
less case
= max
0
[ p 0 :
i
0
, i.
Furthermore, the maximum is given by
=
1
1
,
where
is the eigenvector of
(i.e., k p
for any k R
+
constitute an optimal power vector.).
An interpretation of this result is that it is possible to compute the maximum
achievable balanced SIR given a Gmatrix. This is an optimal solution in the sense
that there exists no other power vector yielding a higher SIR for all receivers. If
the network is reciprocal, the same Gmatrix is valid for both the uplink and the
downlink. To analyze the uplink situation, replace
Z by
Z
T
. However, since
Z
and
Z
T
have the same eigenvalues, Theorem 7.1 will result in the same optimal
SIR in both the uplink and the downlink. (The optimal power vectors are probably
dierent.)
126 Global Analysis
Example 7.1 (Optimal Balanced SIR)
Consider a situation where only two mobile stations are established on the same
channel:
G =
1 0.08
0.01 0.8
= 1.032. Hence,
the optimal SIR in this case (in both up and downlink) is
=
1
1
= 31.6 (
= 15 dB),
achieved by the optimal power assignment
p
= k(0.93, 0.37), k R
+
.
As pointed out in Section 4.5, the optimal power assignment is relative. The
inuence of the noise in Equation (7.4) can thus be made arbitrarily small by
choosing k suciently large. A consequence is that we for practical purposes may
achieve the same SIR as in the noiseless case, but only if sucient transmitter
power is available.
Godlewski and Nuaymi (1999) studied the eect of the autointerference to these
optimality results. It is found that when autointerference cannot be neglected, the
optimal SIR will be smaller. The detailed results are omitted here, but a simple
example provide some intuition. we will adopt a notation that relates to the results
by Zander (1992b). Introduce the diagonal matrix of receiver eciencies
= diag
1
, . . . ,
Z p +
1
. (7.5)
The degradation in optimal SIR due to inecient receivers can be considerable,
which is illustrated by the following example.
Example 7.2 (Optimal Balanced SIR Considering Autointerference)
Consider the case in Example 7.1, but include autointerference, which is as
sumed to be equal at each receiver,
i
=
0
, i. For example a receiver eciency
0
= 0.99 corresponds to the optimal SIR
0
is illustrated by Figure 7.1.
7.1 Optimal Global Power Assignments 127
0.7 0.75 0.8 0.85 0.9 0.95 1
0
5
10
15
20
25
30
35
1 +
t
t
E
Z
p
t
= , (7.6)
where E denotes the identity matrix. Furthermore, the leftmost matrix is non
singular with probability one.
In a realistic situation, power control is typically implemented as cascade con
trol (see Section 4.4), where the target SIR:s are provided by an outer control
loop. Below, we generalize the theory to include individual target SIR:s
t
i
and
autointerference. Thus, the objective is to assign transmitter powers to meet the
requirement
i
g
ii
p
i
j=i
g
ij
p
j
+
p
i
g
ii
+
i
=
t
i
, i (7.7)
Introduce the diagonal matrix
= diag(
t
1
, . . . ,
t
m
).
128 Global Analysis
The requirements in Equation (7.7) can thus be vectorized to
p =
Z E) p +
. (7.8)
Solvability of the equation above is related to feasibility of the related power
control problem, dened as
Denition 7.3 (Feasibility of Power Control Problems)
A set of target SIR:s
t
is said to be feasible with respect to a network described
by
Z and , and the receiver eciency
, if it is possible to assign transmitter
powers p so that the requirements in Equation (7.8) are met. Analogously, the
power control problem
Z, ,
,
/
t
(or dierence
t
in logarithmic scale).
However, in a situation with dierent individual target SIR:s as in Equation (7.7),
it is not as natural. One approach is to state that feasibility requires
t
i
<
, i.
This is a sucient but not necessary condition, and thus too conservative. Instead,
the feasibility margin as dened below, can be used to discuss this matter.
Denition 7.4
Given a power control problem (
Z, ,
,
t
), the feasibility margin
m
R
+
is
dened by
m
= sup
x R : x
t
is feasible
The concept has been adopted from Herdtner and Chong (1999, 2000), where
similar proofs of similar and additional theorems covering related situations also
are provided. Herdtner and Chong used the term feasibility index R
I
and omitted
autointerference. The motivation for introducing the name feasibility margin is to
stress the similarity to the gain margin discussed for local loops in Section 5.3.1.
As disclosed in that section, the gain margin is a relevant measure of the stability
margin of the local control loop. The feasibility margin can in a sense be seen
as a stability margin of the overall system, if we employ integrating inner control
loops. In such a case, feasibility is a necessary condition for global stability. In the
following theorem we capture the essentials regarding feasibility margins.
Theorem 7.5
Given a power control problem (
Z, ,
,
t
), the feasibility margin (of both up
and downlink) is obtained as
m
= 1/
is
= max eig
t
(
Z E)
.
The power control problem is feasible if and only if
m
> 1. If the power control
problem is feasible, there exists an optimal downlink power assignment, given by
p =
t
(
Z E)
1
.
The corresponding uplink power assignment is obtained by replacing
Z by
Z
T
.
Proof See Appendix 7.A 2
The feasibility margin is exemplied in the example below
Example 7.3 (Feasibility of Power Control Problems)
Consider the twomobile downlink case in Examples 7.1 and 7.2. Assume that
the receiver eciency at each receiver is
i
= 0.99, i = 1, 2, and that the target
SIR:s
t
i
= 10, i = 1, 2 are desirable. The corresponding feasibility margin is
obtained by applying Theorem 7.5
m
= 2.38.
The feasibility of the power control problem is not surprising, since the target
SIR:s are all less than the optimal SIR computed in Example 7.2. Instead,
consider the target SIR:s
t
1
= 10 and
t
2
= 40, with the feasibility margin
m
= 1.10.
This case thus also corresponds to a feasible problem, despite that not all target
SIR:s are below the optimal SIR.
The feasibility margin can also be related to the load of the system. When the
feasibility margin is one, the system clearly is fully loaded (only possible when un
limited transmission powers are available). Conversely, when the feasibility margin
is large, the load is low compared to a fully loaded system. Thus the following load
denition is logical.
Denition 7.6 (Relative Load)
The relative load
L
r
of a system is dened by
L
r
=
1
m
(=
in Theorem 7.5).
Feasibility of the power control problem is thus equivalent to a relative load less
than unity. This load denition is illustrated by Example 7.4.
130 Global Analysis
A
B
Figure 7.2 Locations of mobile stations in Example 7.4.
Example 7.4 (Relation Between Target SIR and Relative Load)
Consider a typical CDMA situation in Figure 7.2 where in this case three
mobile stations are located close to the base station. Their corresponding target
SIR:s=12 dB reect the data rate that they are currently using. A fourth mobile
(A), located far from the base station, is requiring a low data rate corresponding
to target SIR 10 dB. For simplicity, autointerference is not considered. The
power control problem is feasible, with feasibility margin and relative load
m
= 6.06,
L
r
= 0.165.
Assume that mobile A increases his data rate (aiming at a higher target SIR).
The relative load will increase accordingly. This is illustrated by the solid line in
Figure 7.3. We note that it is quite costly for the system to provide a high data
rate to the distant mobile. The load is for instance doubled if mobile A is using
the same data rate as the other mobiles (12 [dB]). For comparison, consider
the situation when mobile B, close to the base station, increases his data rate
(dashed line in Figure 7.3). Clearly, this is not as costly for the system.
7.2 Convergence in Linear Scale
Most convergence results presented to date utilize values in linear scale. Several
important algorithms can be written as matrix recursions, which facilitates the
analysis. A common approach is to rewrite the problem as a standard iterative
algorithm, see Section 7.2.1, and then relate to established results on convergence.
This approach is used in Section 7.2.2 to prove convergence of a general integrating
controller in a delayless situation. Section 5.3 disclosed that the DPC algorithm
is unstable, when subject to time delays. Here, we show that DPC with TDC
converges for any time delay.
These information feedback algorithms converges to an equilibrium. The sit
uation is dierent when considering decision feedback algorithms. Their behavior
7.2 Convergence in Linear Scale 131
10 5 0 5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
L
r
t
[dB]
Figure 7.3 The relative load with respect to dierent target SIR:s of mo
bile A (solid) and B (dashed).
can be characterized by persistent oscillations. It has been show that the power
control error using FSPC algorithm, converges to a bounded region. As will be
shown in Section7.2.3, the power control error, using the FSPC algorithm with
TDC, converges to a more smaller bounded region.
7.2.1 Standard Interference Functions
The framework for addressing pointwise convergence of algorithms is based on the
work of Bertsekas and Tsitsiklis (1989), which has been rened and put into the
context of power control algorithms by Yates (1995). An underlying assumption is
that the Gmatrix is constant. Let p denote the controllable resources (in a power
control case, it is the power vector). Assume that the requirements to be met can
be described by the inequality
p
I( p), (7.9)
where
I( p) is referred to as the interference function. A vector p 0 is a feasible
solution if it satises (7.9). Moreover, an interference function
I( p) is feasible
if (7.9) has a feasible solution. For an arbitrary interference function, the following
denition applies:
Denition 7.7 (Standard Interference Function)
The interference function is standard if the following properties are satised for all
p 0.
Positivity:
I( p) 0.
Monotonicity: if p p
then I( p) I( p
).
Scalability: I( p) > I( p) for all > 1.
Given the requirements in (7.9), we want to study the convergence properties of
the class of iterative algorithms described by
p(t + 1) = I( p(t)). (7.10)
132 Global Analysis
When I( p) is a standard interference function, the iteration (7.10) is denoted a
standard iterative algorithm. The corresponding equilibrium is thus
p(t + 1) = I( p(t)) = p(t).
The uniqueness of this equilibrium, is disclosed in the following theorem
Theorem 7.8 (Yates, 1995)
Let I( p) be standard and feasible. If the corresponding standard iterative algo
rithm (7.10) has an equilibrium, then it is unique.
Proof See Yates (1995) and Blom and Gunnarsson (1998). 2
Furthermore, the convergence of standard iterative algorithms is provided in
Theorem 7.9
Theorem 7.9 (Yates, 1995)
If I( p) is standard and feasible, then the iterative algorithm (7.10) converges to a
unique equilibrium p
t
i
i
(t)
i
, (7.11)
subject to the requirements
i
(t)
t
i
, i. (7.12)
The relations between (7.11) ans (7.12) and standard iterative algorithms are pro
vided in the lemma below.
Lemma 7.10
Assume that
t
i
are feasible target SIR:s according to Denition 7.3 and that
0 < 1. Then the distributed power control algorithm in (7.11) with require
ments as in (7.12) corresponds to a standard iterative algorithm.
Proof See Appendix 7.B. 2
The convergence of the integrating controller can thus be established.
7.2 Convergence in Linear Scale 133
Theorem 7.11 (Convergence of the Icontroller in a Delayless Case)
The integrating controller (7.10) with requirements (7.12) converges to a unique
equilibrium p
that meets the requirements with equality for any initial power
vector p
0
> 0.
Proof A direct consequence from Theorem 7.9 and Lemma 7.10. 2
The theorem does not hold when subject to time delays. Consider time delay
compensation as described in Section 6.2, where the measurements are adjusted as
i
(t) =
(
t n
m
) +p
i
(t) p
i
(t n
p
n
m
). (7.13)
This scheme compensates for the eect of outdated powers reected by the mea
surements. By employing TDC to the DPC algorithm (
i
= 1), stability and
convergence are easily proven. Stability and convergence in more general cases are
established in Section 7.3
Theorem 7.12 (Convergence of the DPC algorithm with TDC)
The DPC algorithm (7.10) and
i
= 1 employing TDC converges to a unique
equilibrium p
that meets the requirements in (7.12) with equality for any initial
power vector p
0
. The convergence rate is n+1 times slower compared to the DPC
algorithm applied to a delayless power control problem.
Proof The time delay compensation (7.13) in linear scale is
i
(t) = i(t nm)
pi(t)
pi(t nm np)
.
The updating procedure of the DPC algorithm in (7.11) and i = 1 is thus
pi(t + 1) = pi(t)
t
i
i
(t)
= pi(t nm np)
t
i
i(t nm)
=
t
i
Ii( p(t nm np))
gii
.
(7.14)
Introduce n = np + nm. Clearly, the update at time instant t + 1 only depend on the
powers at time instant t n. This is essentially the same as having n + 1 algorithms
operating independently in parallel. Thus if the sample interval is Ts, each of these
algorithms update every n + 1 sample interval. In essence, the situation is the same as
in Theorem 7.11, but the sample interval is longer, resulting in a factor n + 1 slower
convergence. 2
Convergence of the AAW algorithm (see Section 5.2.1) with respect to its pa
rameters is disclosed in (Yates et al., 1997). Time delays and autointerference are
omitted in the analysis, but the latter can be readily included.
134 Global Analysis
7.2.3 Convergence of Some Decision Feedback Algorithms
The FSPC algorithm results in an oscillative local loop, as discussed in Section 5.5.
Therefore, it will never converge to an equilibrium. The local loop oscillations is
reduced when employing TDC (see Section 6.2.1). Moreover, as will be shown
in this section, FSPC with TDC converges to a smaller region than without. A
slightly more general version of the FSPC algorithm is given by
p
i
(t + 1) =
i
p
i
(t),
i
(t)
t
i
i
p
i
(t),
i
(t) >
t
i
(7.15)
where
> 1 and < 1. Note that the dierence to (4.15) is that the algorithm
takes dierent steps upward and downward. Hence the FSPC algorithm is obtained
if = 1/
m
satises
n+1
/
n
<
m
. Then, when using the algorithm in (7.15), there exist
a t
FSPC
0 such that
n+1
t
i
i
(t)
n+1
, t t
FSPC
.
The requirement
n+1
/
n
<
m
essentially restricts the proof to cases which are
slightly below maximal relative load. This is essentially no loss of generality, since
the behavior close to the maximal load is unpredictable when the oscillations are
aected by the feasibility constraint. The authors omitted the autointerference,
but the proof holds after minor adjustments. Using TDC as described in (7.16)
together with the algorithm in (7.15) we form an algorithm that converges to a
region that is smaller than the region in Theorem 7.13.
Theorem 7.14 (Convergence of FSPC with TDC)
Assume that the power control problem is feasible and that the feasibility margin
m
satises
/
n
<
m
. Then, when using the algorithm in (7.16), there exist a
t
TDC
0 such that
n+1
t
i
i
(t)
n+1
, t t
TDC
.
7.3 LogLinear Algorithms 135
Proof See Appendix 7.C. 2
A comparative comment is provided in the following remark.
Remark 7.15
Consider the case of equally sized steps upward and downward ( = 1/
). The
bounds (in dB) in Theorems 7.13 and 7.14 are given by
FSPC, no TDC: [
t
i
i
(t)[ (2n + 2)
FSPC, with TDC: [
t
i
i
(t)[ (n + 2)
Hence, the longer the time delay, the more emphasized improvements using TDC.
Note, however, that the fact that the bounds are tighter does not imply that the
error variance is smaller.
7.3 LogLinear Algorithms
The techniques in the previous section apply primarily to structurally simple al
gorithms. When considering delays, lters and general loglinear controllers, it
is desirable to relate to the local loop analysis and design in Chapters 5 and 6.
The challenge is the the nonlinear interconnections between the loglinear control
loops. In this section we apply a robust stability approach, based on a linearized
interference. Essentially, we aim at rewriting the global system, so that it can
be associated with the multivariable system in Figure 7.4. This is the standard
setting to which the Small Gain Theorem (Zhou et al., 1995, and Appendix 7.D)
applies. Therefore, we proceed in three steps:
w
1
w
2
(q)
G(q)
Figure 7.4 Setting to which the Small Gain Theorem applies.
1. Rewrite the local loops as a diagonal multiinput, multioutput linear feed
back system G(q). This is the main issue in Section 7.3.1.
2. Represent the interconnections by a structured uncertainty (q). Since the
interference is nonlinear, this involves approximations. Section 7.3.2 ad
dresses the problem by linearizing the interference.
136 Global Analysis
3. Address stability using Small Gain Theorem. These results are found in
Section 7.3.3.
The established stability results put restrictions on the local loop design, and
thus stability can be (approximately) granted by appropriate local loop design. To
exemplify the results, the specic case of two mobiles receives extra attention in
Section 7.3.4. A wellknown drawback with Small Gain Theorem is that it is rather
conservative in some cases. Therefore, the two mobile case is used as a benchmark
example in Section 7.3.5.
7.3.1 The Global System as Interconnected Local Loops
The global system consists of interconnected local loops. Starting from the log
linear models of local loops derived in Chapter 5, the aim here is to obtain a model
of the global system. This model will contain nonlinearities from the nonlinear
interconnections via the interference, but the remainder will be linear. The focus
is on algorithms with the objective to meet individual target SIR:s provided by an
outer loop. Therefore, algorithms like AAW, which essentially adopts the provided
target, are not addressed. However, by rephrasing the algorithm as discussed in
Sections 5.2.1 and 9.4.3, this adaption is placed in the outer loop. Initially, time
delay compensation (TDC) described in Section 6.2 is omitted, but is included
further on. This section includes rather tedious formula manipulations. For con
venience, the resulting diagonal transfer functions are provided in a subsection in
the end of this section.
Operation without TDC
Recall that the local distributed control loops (considering autointerference) can
be depicted as in Figure 7.5 (cf. Section 5.3). The interconnections between the
+ +
+ p
i
(t)
i
(t)
t
i
(t)
i
(t) +g
ii
(t) I
i
(t)
q
np
q
nm
R(q)
F(q)
Figure 7.5 The local loop when employing the general linear control algo
rithm R(q) and the lter F(q).
loops are described by the interference I
i
(t). More specically, this is a function
of the powers from all interferers. In an ideal situation, the interference may be
regarded as an independent disturbance to the local loop, which was the underlying
7.3 LogLinear Algorithms 137
assumption in Chapter 5. However, in order to model the dynamical eects of
the interconnection, we adopt the more realistic model of the interference (see
Section 3.5) as
I
i
(t) =
j=i
p
j
(t) g
ij
(t) + (1
i
(t)) p
i
(t)g
ii
(t) +
i
(t) (7.17)
Denote the corresponding interference expression in logarithmic scale by
I(p, t) = [I
1
(p, t), . . . , I
m
(p, t)]
T
= 10 log
10
[
I
1
(t), . . . ,
I
m
(t)]
T
. (7.18)
In order to capture the global system behavior, the system model needs to be
vectorized. Therefore, introduce the following vectors
g(t)
= [
1
(t) +g
11
(t), . . . ,
m
(t) +g
mm
(t)]
T
(7.19)
p(t)
= [p
1
(t), . . . , p
m
(t)]
T
(7.20)
t
= [
t
1
, . . . ,
t
m
]
T
(7.21)
(t)
= [
1
(t), . . . ,
m
(t)]
T
. (7.22)
Note the correspondence between these vector quantities and the quantities in the
local loop in Figure 7.5. For a more compact notation we introduce
p(t) = p(t n
p
).
The dynamical eects of introduced lters also need to be described. Adopt the
general approach of applying separate lters F
I
(q) and F
g
(q) to the interference
and signal power respectively (see Section 5.2.4). In the specic case, when only
SIR is measured and ltered, F
I
(q) = F
g
(q) = F(q). To describe the ltering
eects in vector representation, introduce the mm matrices
F
I
(q)
= diag(F
I
(q), . . . , F
I
(q)) = F
I
(q)E (7.23)
F
g
(q)
= diag(F
g
(q), . . . , F
g
(q)) = F
g
(q)E (7.24)
Then the overall system (global system) can schematically be depicted as in Fig
ure 7.6. Since the target SIR is considered constant, and F
g
(q) and q
nm
commute,
it is possible to rearrange the block containing the measurement delay q
nm
as seen
in the gure. Moreover, denote the output of the interference lter F
I
(q) by
I( p, t),
and the diagonal transfer matrix of control algorithms and delays by
R
d
(q) = q
np
R(q)q
nm
E.
Using the introduced notation, Figure 7.6 yield
p(t) = R
d
(q)
I( p, t) +
t
F
g
(q)(g(t) + p(t))
.
138 Global Analysis
+ +
t
(t)
q
nm
R(q)q
np
0
.
.
.
0 q
nm
R(q)q
np
p(t)
I(, t) F
I
(q)
F
g
(q)
g(t)
Figure 7.6 The global system consists of interconnected local loops. The
interconnections consist of a diagonal lter F
I
(q) = F
I
(q)E
and a nonlinear, nondiagonal crosscoupling I(, t).
Solve for p(t) (except for the p(t) in
I( p, t)) results in
p(t) = (E +R
d
(q)F
g
(q))
1
R
d
(q)
I( p, t) +
t
F
g
(q)g(t)
.
Hence, it is natural to dene the diagonal linear feedback system G(q) as
G(q)
=G
ll
(q)E
= (E +R
d
(q)F
g
(q))
1
R
d
(q) =
=
R(q)
q
+np+nm
+F
g
(q)R(q)
E. (7.25)
Note that G
ll
(q) is the closedloop system of the local loops studied in Chapter 5.
The corresponding local poles are thus the roots of the denominator (the charac
teristic polynomial, cf. Section 5.16). Figure 7.7 provides a relevant block diagram
of the global system dynamics. It is intuitive that local loop analysis is sucient
if the interconnections via the interference are neglectable. Conversely, if the in
terference cannot be neglected, it does aect the global system performance and
stability.
Operation with TDC
Time delay compensation is implemented by the linear transfer function H(q) from
the computed transmission powers p
i
(t) as described in Section 6.2. This can be
described as the internal diagonal feedback H(q) = q
np+nm
H(q)E from p(t) (see
Section 6.2). Apart from the internal feedback, the resulting global system in
Figure 7.8, is analogous to the system in Figure 7.6. The global system behavior
7.3 LogLinear Algorithms 139
+
+
t
(t)
p(t)
I(, t) F
I
(q)
G(q)
F
g
(q)
g(t)
Figure 7.7 The global system can be seen as a diagonal system of closed
local loops G(q) and interconnections via the interference.
is thus captured by
p(t) = R
d
(q)
I( p, t) +
t
H(q) p(t) F
g
(q)(g(t) + p(t))
.
As in the case without TDC, solve for p(t), which yields
p(t) = (E +R
d
(q)F
g
(q) R
d
(q)H(q))
1
R
d
(q)
I( p, t) +
t
F
g
(q)g(t)
.
As in the case of operating without TDC, we dene a diagonal linear feedback
system G(q) by
G(q)
=G
ll
(q)E
= (E +R
d
(q)F
g
(q) R
d
(q)H(q))
1
R
d
(q) =
=
R(q)
q
+np+nm
(1 +q
npnm
F
g
(q)R(q) R(q)H(q))
E. (7.26)
The TDC implementation in Algorithm 6.4 reduces to the expression in Equa
tion (6.5), which is given by
H(q) = F
g
(q)(1 q
npnm
)
This choice of H(q) and Equation (7.26) results in the closed local loop
G
ll
(q) =
R(q)
q
+np+nm
(1 +F
g
(q)R(q))
. (7.27)
Clearly, this is the closed local loop system as obtained in Section 6.2. Thus
operation with TDC results in dierent local loops G
ll
(q) than operating without.
However, the global system interconnections are described by Figure 7.7 in both
cases. Since there is an imminent risk that the global and the local view get mixed
up in the sequel, we dene
140 Global Analysis
+ +
t
(t)
R
d
(q)
p(t)
I(, t) F
I
(q)
F
g
(q)
H(q)
g(t)
Figure 7.8 Time delay compensation can be seen as internal diagonal feed
back H(q) = q
np+nm
H(q)E from the computed power p(t).
Otherwise, the global system is identical to the case without
TDC in Figure 7.6.
Denition 7.16 (Local Loop and Global System Poles)
The term local loop poles refer to the closedloop poles of the local loop, while the
global system poles refer to the closedloop poles of the overall system.
Summary, Interconnected Local Loops
In summary, the interconnected local loops can be associated with a diagonal linear
feedback systemG(q) = G(q)
ll
E, where G
ll
(q) is the closedloop system of the local
control loops (cf.. Section 6.4.1). Assume that the power is controlled by R(q),
and the measurements are ltered by F
g
(q) (to emphasize which lter to consider
when using separate ltering, see Section 5.2.4). Then G
ll
(q) is given by
No TDC: G
ll
(q) =
R(q)
q
n
+F
g
(q)R(q)
(7.28)
TDC: G
ll
(q) =
R(q)
q
n
(1 +F
g
(q)R(q))
(7.29)
7.3.2 Interference Linearization
The global system in Figure 7.7 is nonlinear due to the nonlinear crosscoupling
I( p(t), t) between the local loops. One way of addressing stability issues, is to
consider a linearized interference. The resulting dynamics will be linear, and can
thus be analyzed using methods for linear systems. This linearization is relevant
in a neighborhood of an equilibrium point, where the iterative algorithms have
7.3 LogLinear Algorithms 141
settled.The existence of such a equilibrium point is provided in the following lemma,
based on the assumption that the power gains g
ij
and the receiver eciencies
i
(t)
are constant over the time frame of the analysis.
Lemma 7.17 (Momentary Equilibrium Point)
For all feasible power control problems, there exists a momentary equilibrium point
described by p
t
,
t
and
I
t
=
I( p
t
) (and corresponding values in logarithmic scale).
This is a valid equilibrium point over the time frame where g
ij
and
i
can be
considered constant.
Proof Follows directly from Theorem 7.5. 2
The interference function was dened in (7.18). A linearization of the same
about the equilibrium yields
I( p(t)) I
t
+
I( p(t))
p(t)
p(t)=p
t
p(t) p
t
=I
t
+
cc
p(t) p
t
, (7.30)
where the Jacobian
cc
will be referred to as the crosscoupling matrix. A closed
form expression for this matrix is derived in the following lemma.
Lemma 7.18 (Crosscoupling Matrix of the Power Control Problem)
Consider the interference function at receiver i as dened in (7.18). Then, based
on Lemma 7.17, the crosscoupling matrix in Equation 7.30 is given by
[
cc
]
ij
=
(1
i) gii p
t
i
Ii
i = j
gij p
t
j
Ii
i = j
(7.31)
Proof In logarithmic scale (dB), the interference function is given by
Ii(p) =10 log
10
k=i
g
ik
p
k
+ (1
i) pi gii + vi
=
=
10
ln10
ln
k=i
g
ik
e
ln10
10
p
k
+ (1
i) giie
ln10
10
p
i
+ vi
(7.32)
The partial derivative of (7.32) with respect to pi is
Ii(p)
pi
=
10
ln10
1
k=i
g
ik
p
k
+ (1
i) pi gii + vi
(1
i) gii pi
ln10
10
=
(1
i) gii pi
Ii
(7.33)
Taking the partial derivative with respect to pj, where j = i, yields
Ii(p)
pj
=
10
ln 10
1
k=i
g
ik
p
k
+ (1
i) pi gii + vi
gij pj
ln 10
10
=
gij pj
Ii
(7.34)
The proof is completed by inserting the equilibrium point from Lemma 7.17. 2
142 Global Analysis
The global system in Figure (7.7) can now be approximated by a linear system
using the interference linearization in Equation (7.30). Hence
p(t) =G(q)
I( p, t) +
t
F
g
(q)g(t)
=G(q)
F
I
(q)I
t
+F
I
(q)
cc
( p(t) p
t
) +
t
F
g
(q)g(t)
p(t)
cc
F
I
(q)
G(q) F
g
(q)
g(t)
Figure 7.9 Block diagram of the global system, when approximating the
interference by the corresponding linearization with respect to
the equilibrium deection p(t).
7.3.3 Robust Stability of LogLinear Power Control
Algorithms
Basically, the global system in Figure 7.9 can be seen as a diagonal system of local
loops with a parallel structured uncertainty. One way of addressing stability of
G(q) is to investigate robust stability (see e.g., Zhou et al., 1995) of G(q) with
respect to the uncertainty. Essentially, stability is granted if the roundtrip loop
gain of Figure 7.9 is less than one. Since the transfer function G(q) is frequency
dependent, an appropriate metric has to consider the worst case in the frequency
7.3 LogLinear Algorithms 143
domain. Relevant metrics or norms are discussed together with some robust sta
bility results in Appendix 7.D.1. A matrix norm has the natural interpretation
of input/output amplication gain. The corresponding gain of the crosscoupling
matrix is addressed in the following lemma.
Lemma 7.19
The following relations hold for the matrix norm of the cross coupling matrix
cc
in Lemma 7.18

cc

= max
1im
1
v
i
I
t
i
=
cc
< 1
The value of the norm
cc
will be referred to as the degree of crosscoupling.
Proof See Appendix 7.D.2 2
The degree of crosscoupling
cc
allows a natural interpretation. Essentially, it
reects the inuence of the interference to the global system stability. Two cases
are easily distinguished
cc
= 0. Corresponds to the case when the interference at every receiver
is only noise, i.e., the local loops are operating independently. In such a
situation, the local loop analysis in Chapter 5 is sucient.
cc
= 1. Impossible in practice, since it corresponds to the case when the
thermal noise is zero, or the interference at one receiver is innite. However,
in highly loaded interference limited systems,
cc
is close to 1.
The maximal gain of the local loop G
ll
(q), which is described in Equations (7.25)
or (7.27), is
sup
w
[G
ll
(q)[
q=e
iw
Provided that the local loop is stable, stability of the global system is granted if
this maximal gain meets a given restriction, related to the degree of crosscoupling.
Hence, it is possible to establish global system stability by careful local loop design.
This is formally proven in Theorem 7.20.
Theorem 7.20 (Global Stability of Power Control Algorithms)
Let G
ll
(q) be the stable closedloop transfer function of the local loop in Figure 7.5.
Then the global system in Figure 7.9 is stable if and only if any of the following
properties is satised
(i) sup
w
[G
ll
(e
iw
)[ 1/
FI
.
(ii) sup
w
[G
ll
(e
iw
)[ < 1/(
cc
FI
),
144 Global Analysis
where
FI
= sup
w
[F
I
(e
iw
)[
cc
= max
1im
1
v
i
I
t
i
(1)
2
(1
1
4
)
1
3
< 1
1 0 < <
1
3
7.3 LogLinear Algorithms 145
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
c
c
Z =
1
g12
g11
g21
g22
1
, =
1
g11
,
2
g22
T
,
t
=
t
1
0
0
t
2
(7.38)
From Denition 7.6, we obtain the relative load.
L
r
= max eig
t
(
Z E)
g
12
g
21
g
11
g
22
t
1
t
2
. (7.39)
The linearization described in Section 7.3.2 is naturally simplied. Thus the inter
ference at receiver i is approximated by
I
i
( p) I
t
i
+
I
i
( p(t))
p
j
p(t)=p
t
( p(t) p
t
) =
I
t
i
+
I
i
( p(t))
p
j
pj(t)=p
t
j
( p
j
p
t
j
) =
I
t
i
+
g
ij
p
t
j
I
t
i
( p
j
p
t
j
) = I
t
i
+
i
( p
j
p
t
j
), (7.40)
where i, j 1, 2, i = j. The crosscoupling is characterized by
1
and
2
, which
can be rewritten as
i
=
g
ij
p
t
j
I
t
i
=
I
t
i
i
I
t
i
=
t
i
g
ij
p
t
j
g
ii
p
t
i
(7.41)
It is instructive to establish the relations between the relative load and the
i
:s.
Equations (7.39) and (7.41) yield
L
r
=
1
2
. (7.42)
No Delays in the System
This case has been analyzed in linear scale in Section 7.2.1 and in (Blom and
Gunnarsson, 1998; Yates, 1995). Equations (7.37), (7.40) and (7.41) yield
p
i
(t + 1) = p
i
(t) +(
t
i
i
(t)) = p
i
(t) +(
t
i
p
i
(t) g
ii
+I
i
(t))
p
i
(t) +(
t
i
p
i
(t) g
ii
+I
t
i
+
i
(p
j
(t) p
t
j
) =
= p
i
(t) +(p
t
i
p
i
(t)) +
i
(p
j
(t) p
t
j
) (7.43)
Recall Equation (7.35), where
p
i
(t) = p
i
(t) p
t
i
.
This yields in (7.43)
p
i
(t + 1) = p
i
(t) p
i
(t) +
i
p
j
(t). (7.44)
7.3 LogLinear Algorithms 147
For the overall system of both mobile stations, we get
p(t + 1) =
1
1
2
1
p(t) = A
1
p(t). (7.45)
The global system poles are given by the roots to the following equation (cf. Sec
tion 5.3.1)
det(qE A
1
) = 0 (7.46)
Hence
0 = (q 1 +)
2
2
1
2
= (q 1 + +
L
r
)(q 1 +
L
r
),
where Equation 7.42 is used in the last equality. The global system poles are thus
given by
q = 1 (1
L
r
).
Since 0 < 1 and 0
L
r
< 1, the global system poles are always within the unit
circle, and the global system is thus always stable. This is the same conclusions as
in Section 7.2.1 and in (Blom and Gunnarsson, 1998; Yates, 1995).
Delay of One Sample
Since the gains g
ij
are assumed constant, there is no dynamical dierence between
a measurement delay n
m
and an output power delay n
p
. Both will delay the output
power contribution p
i
of the measurement
i
. This is more formally seen below
i
(t n
m
) = p
i
(t n
m
n
p
) +g
ii
(t n
m
) I
i
(p
j
(t n
m
n
p
), t).
Therefore, we notationwise study n
p
= 1, n
m
= 0. Analogous to (7.43), we get
p
i
(t + 1) = p
i
(t) +(
t
i
i
(t)) =
= p
i
(t) +(
t
i
p
i
(t 1) g
ii
+I
i
(p
j
(t 1), t))
p
i
(t) +(
t
i
p
i
(t 1) g
ii
+I
t
i
+
i
(p
j
(t 1) p
t
j
) =
= p
i
(t) +(p
t
i
p
i
(t 1)) +
i
(p
j
(t 1) p
t
j
).
Thus
p
i
(t + 1) = p
i
(t) p
i
(t 1) +
i
p
j
(t 1). (7.47)
Introduce the following state variables: w
1
(t) = p
1
(t), w
2
(t) = p
2
(t), w
3
(t) =
p
1
(t 1) and w
4
(t) = p
2
(t 1). Note that
w
3
(t + 1) = w
1
(t)
w
4
(t + 1) = w
2
(t)
148 Global Analysis
These combine to the state space representation of the global system dynamics
w(t + 1) =
1 0
1
0 1
2
1 0 0 0
0 1 0 0
w(t) = A
2
w(t). (7.48)
By solving
det(qE A
2
) = 0
we get the global system poles as below
q =
1/2 + 1/2
1 4 + 4
L
r
1/2 1/2
1 4 + 4
L
r
1/2 + 1/2
1 4 4
L
r
1/2 1/2
1 4 4
L
r
(7.49)
Not all combinations of
L
r
and will result in a stable global system. The global
system poles are within the unit circle if and only if the following condition is
satised
<
1
1 +
L
r
, 0
L
r
< 1 (7.50)
This stability region with respect to the parameter ranges is illustrated in Fig
ure 7.11. The conclusion is that a < 0.5 is needed in order to preserve stability
for all feasible crosscoupling and relative load situations. This should be compared
to the result from the localloop analysis, which is identical to the situation
L
r
= 0.
As in Example 5.4, localloop stability is achieved for < 1.
Delay of One Sample with Time Delay Compensation
As in the previous section, consider the case n
p
= 1, n
m
= 0. TDC is implemented
as described in Section 6.2.
i
(t) =
i
(t) +p
i
(t) p
i
(t 1)
This will aect the update equations in the following way
p
i
(t + 1) = p
i
(t) +(
t
i
i
(t)) = p
i
(t) +(
t
i
p
i
(t) +g
ii
I
i
(p
j
(t 1), t))
p
i
(t) +(
t
i
p
i
(t) g
ii
+I
t
i
+
i
(p
j
(t 1) p
t
j
) =
= p
i
(t) +(p
t
i
p
i
(t)) +
i
(p
j
(t 1) p
t
j
).
Thus
p
i
(t + 1) = (1 ) p
i
(t) +
i
p
j
(t 1). (7.51)
7.3 LogLinear Algorithms 149
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
L
r
1 0 0
1
0 1
2
0
1 0 0 0
0 1 0 0
w(t) = A
3
w(t). (7.52)
As in the previous, we solve
det(qE A
3
) = 0
to get the global system poles:
q =
1 2 +
2
+ 4
L
r
1/2 1/2 1/2
1 2 +
2
+ 4
L
r
1/2 1/2 + 1/2
1 2 +
2
4
L
r
1/2 1/2 1/2
1 2 +
2
4
L
r
(7.53)
Some computations yield that all combinations of
L
r
and result in global system
poles within the unit circle and thus stable global systems.
150 Global Analysis
The Effect of Cancelling the Power in the Measurement
An intuitive approach to delay compensation might be to cancel all inuence of
the power in the measurements. This corresponds to the following compensation
i
(t) =
i
(t) p
i
(t 1)
The update equations will change accordingly
p
i
(t + 1) = p
i
(t) +(
t
i
i
(t)) = p
i
(t) +(
t
i
g
ii
+I
i
(p
j
(t 1), t))
p
i
(t) +(
t
i
g
ii
+I
t
i
+
i
(p
j
(t 1) p
t
j
) =
= p
i
(t) +p
t
i
+
i
(p
j
(t 1) p
t
j
).
Thus
p
i
(t + 1) = p
i
(t) +
i
p
j
(t) +p
t
i
. (7.54)
Introduce state variables as previously, to form the state space representation
w(t + 1) =
1 0 0
1
0 1
2
0
1 0 0 0
0 1 0 0
w(t) = A
4
w(t). (7.55)
Solve
det(qE A
4
) = 0
to get the global system poles:
q =
1/2 + 1/2
1 + 4
L
r
1/2 1/2
1 + 4
L
r
1/2 + 1/2
1 4
L
r
1/2 1/2
1 4
L
r
(7.56)
It is easy to see that since the rst pole is always greater than or equal to one,
the global system is always unstable. This is not surprising, since that compensa
tion cancels the feedback. The result is thus an integrator in open loop, which is
unstable.
7.3.5 The Conservativeness of Theorem 7.20
As discussed in Section 7.3.3, Theorem 7.20 might yield rather conservative stability
requirements in some cases. Therefore, it is instructive to compare the stability
regions obtained directly in the two mobile case in Section 7.3.4 to the region
obtained via Theorem 7.20.
As in Section 7.3.4, the focus is on the use of the Icontroller, when the output
powers are delayed by one sample. Operation both with and without time delay
compensation will be studied.
7.3 LogLinear Algorithms 151
Icontroller without TDC
The corresponding closedloop system is discussed in Example 7.5, where the degree
of crosscoupling
cc
is related to the choice of controller parameter value . Con
versely, it is interesting to relate a given crosscoupling situation to corresponding
requirements on to ensure global stability. Theorem 7.20 yields the requirement
sup
w
[G
ll
(e
iw
)[ <
1
cc
=
1
max
1i2
i
(7.57)
The result in Example 7.5 simplies the left hand side of (7.57) to
( 1)
2
(1
1
4
)
<
1
max
1i2
i
,
1
3
< 1 (7.58)
This criterion is compared to the result from the root locus analysis in Section 7.3.4
(cf. Figure 7.11). Both criteria are illustrated in Figure 7.12. As we can see, the
0
0.2
0.4
0.6
0.8
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
L
r
1
2
L
r
L
r
=
m
. (7.59)
Note that this only should be considered as a rule of thumb. However, it further
motivates the name feasibility margin (cf. Section 7.1.2).
7.4 Summary 153
Icontroller with TDC
Example 6.2 exemplied the eect of TDC in this particular case, and the closed
loop system is given by
G
ll
(q) =
q(q 1 +)
.
Since
sup
w
[G
ll
(e
iw
)[ 1
Theorem 7.20 yields that stability is always granted. This is the same conclusion
as in Section 7.3.4.
7.4 Summary
Capacity, load and whether it is possible to accommodate all users with their
data rate requirements (reected by the SIR targets) are typically discussed at a
global level. We assume that the inner control loops successfully compensates for
the fast variations, so that the power gains can be considered as constant. With
the ambition to generalize previous results to include autointerference, we dene
feasibility, to include these eects. Furthermore, capacity and load are related to
the introduced metrics feasibility margin and relative load.
Moreover, stability and convergence are also discussed at a global level. Again,
the power gains are assumed constant, but the motivation is that we are consider
ing the short term behavior of the global system, and the power gains are assumed
to uctuate more slowly. Convergence results of some structurally simple algo
rithms are proven using values in linear scale. The DPC algorithm with TDC (see
Section 6.2) is proven stable for any time delays. This emphasizes the benets
using TDC, since DPC is unstable when subject to any additional time delays.
The FSPC, which is based on decision feedback, show a persistent oscillative be
havior. The resulting power control errors is bounded as proven in the literature.
Employing TDC with the FSPC algorithm, we prove that the power control error
converges to a bounded region with smaller bounds than above.
When considering general loglinear control algorithms, time delays and lters,
it is hard to rewrite the algorithm to t a linear framework as used above. Instead
it is desirable to relate stability and performance to the local loop analysis and
design in Chapters 5 and 6. Using, a linearized interference, we apply a robust
stability framework, to disclose results on stability. The conclusion is that global
stability can be provided by careful design of the local control loops.
Appendix
7.A Proof of Theorem 7.5
Study the feasibility of the same power control problem, but with the SIR target
x
t
. It is trivial that x 0 is corresponding to an infeasible power control problem.
Moreover, feasibility is analogous to solvability of Equation (7.8)
p = x
Z E) p +
.
It can be rewritten as
1
x
E
t
(
Z E)
p =
1
.
If the matrix on the left hand side is singular, solvability depends on whether the
vector on the right hand side is in the linear hull spanned by the basis of the left
hand side singular matrix. Since these vectors and matrices are stochastic, this
will be the case with probability zero. Thus solvability can be considered equal to
whether the left hand side matrix is singular or not. Hence, for insolvability, we
require
0 = det
1
x
E
t
(
Z E)
. .. .
. (7.A.1)
The matrix
B is positive with eigenvalues
1
, . . . ,
n
b
, and PerronFroebenius
theory (Gantmacher, 1974) yields that the matrix has a positive real eigenvalue
= max[
i
[
n
b
i=1
. Hence
= max
: detE
B = 0
m
= 1/
.
154
7.B Proof of Lemma 7.10 155
Naturally,
m
> 1 corresponds to a feasible target SIR and power control problem.
Since (E
B) is nonsingular, the optimal power assignment is given by
p = (E
B)
1
1
=
t
(
Z E)
1
,
which concludes the implication if. The implication only if when autointerfe
rence is omitted is provided in (Herdtner and Chong, 2000). The proof generalizes
to include autointerference with minor modications.
7.B Proof of Lemma 7.10
For simplicity, we will drop the time index t. Note that the power gains g
ij
, the
noise power
i
, the powers p
i
and the receiver eciency
i
all are positive and
assumed constant. Furthermore, (1
i
) 0, i. Consider the downlink (the
uplink is analogous). Dene the vector
( p) = [
i
( p)]
=
i
g
ii
j=i
g
ij
p
j
+ (1
i
) p
i
g
ii
+
i
, (7.B.2)
which is elementwise positive and satises
( p) ( p
) , ( p p
) (7.B.3a)
( p) >
1
t
i
i
t
i
p
i
i
( p)
. (7.B.4)
Multiply both sides of (7.B.4) by p
i
. An appropriate interference function
I( p) is
thus dened by
I( p) =
I
i
( p)
p
i
t
i
p
i
i
( p)
p
1
i
t
i
i
( p)
i
It is easy to verify that this is a standard interference function (see Denition 7.7).
This is important, since we need the scalability and monotonicity properties in the
sequel. The following lemma is a direct consequence from the feasibility margin
denition and Theorem 7.5 in Section 7.1.2. A more rigorous proof is found in
(Herdtner and Chong, 2000).
Lemma 7.C.1
If x [0,
m
[, then there exist an unique componentwise minimal power vector
that satises p xI(p). This vector is given by p = xI(p).
Denote this vector by p(x), which only is dened for x [0,
m
[. Note that each
component is a strictly increasing function of x.
To describe the update steps of the power control algorithm, we introduce
z
i
(t) =
p
i
(t)
I
i
(p(t n))
=
p
i
(t)
p
i
(t 1)
i
(t n)
t
i
(7.C.5)
y
i
(t n) =
p
i
(t n)
I
i
(p(t n))
=
i
(t n)
t
i
(7.C.6)
The updates using original algorithm without TDC is aected by the total delay
of n samples, and can thus be written as
p
i
(t + 1) =
p
i
(t), y
i
(t n) 1
p
i
(t), y
i
(t n) > 1
(7.C.7)
Using TDC, the delayed powers related to the same user is compensated for. The
updates are therefore given by
p
i
(t + 1) =
p
i
(t), z
i
(t) 1
p
i
(t), z
i
(t) > 1
(7.C.8)
Dene the following set of powers
C( p)
= p R
N
: p p.
This set of powers is invariant under power updates using (7.C.8), as is shown in
the following lemma.
7.C Proof of Theorem 7.14 157
Lemma 7.C.2
Assume that
n
<
m
and x [
n
,
m
[. Then C( p(x)) is invariant under power
updates using (7.C.8).
Proof We will componentwise show that p(t) p(x) implies p(t +1) p(x). Focus on
mobile i.
If zi(t) > 1 then pi(t + 1) = pi(t) pi(x) < pi(x).
If zi(t) 1 then
pi(t + 1) =pi(t)
(i)
= Ii(p(t n))zi(t) Ii(p(t n))
(ii)
Ii(p(t)/
n
)
(iii)
<
n
Ii( p(x))
(iv)
=
n
x
pi(x) pi(x).
Remarks: (i) follows directly from (7.C.5), (ii) the ratio is less than or equal to
the case when all other users have decreased their powers n times, (iii) scalability
extracts the factor, the remaining power vector is less than p(t) as stated, and yield
the inequality together with monotonicity, (iv) follows from Lemma 7.C.1.
This concludes the proof. 2
In order to formulate the stability theorem compactly, we dene
(p)
= min0 x <
m
: p p (7.C.9)
The stability using the power update in (7.C.8) is formally proven below.
Theorem 7.C.3
If /
n
<
m
then the algorithm in (7.C.8) is stable.
Proof Denote the initial power level by p(0) > 0. Moreover, let 0 = (p(0)), x =
max{
n
, 0}, p = p(x). Suppose that t 0, such thatp(t) C( p). Since
n
x <
m, Lemma 7.C.1 yields that p(t + 1) C( p). From corresponding denitions, we get
p(0) p(0) p(x) = p.
The proof is completed by induction. 2
Finally, all results are at hand, and Theorem 7.14 can be proven.
Proof (Theorem 7.14) We will only prove the lower bound, since the proof of
the upper bound is analogous. Recall the denitions of zi(t) and yi(t n) in (7.C.5)
and (7.C.6) respectively. The power updates can be written in terms of zi(t) as seen
in (7.C.8). Assume that yi(t) /
n
. Then it is desirable to prove that the same
inequality holds for yi(t + 1).
158 Global Analysis
zi(t) < 1:
yi(t + 1) =
pi(t)
Ii(p(t + 1))
pi(t)
Ii(p(t))
>
pi(t)
Ii(p(t))
= yi(t),
where scalability is used in the last inequality.
zi(t) < 1:
yi(t + 1) =
pi(t)
Ii(p(t + 1))
pi(t)
Ii(
n+1
p(t n))
>
pi(t)
n+1
Ii(p(t n))
=
=
n+1
zi(t)
n+1
.
Hence if there exist a ti, such that yi(ti) /
n
, then this hold true for t ti. We prove
this last step by contradiction. Assume that yi(t) < /
n
, t. Hence, the following hold
zi(t) =
pi(t)
Ii(p(t n))
n
pi(t n)
Ii(p(t n))
=
n
yi(t n) <
< 1.
Then, for t n, we compute the output power pi(t) =
tn
pi(0) , t . Since this
contradicts the stability in Theorem 7.C.3, we conclude that there exist a ti 0 for each
user such that yi(ti) /
n
holds. Finally, choosing tTDC = max{t1, . . . , tN} completes
the proof. 2
7.D Robust Stability of Power Control Algorithms
As pointed out in Section 7.3.3, the global system in Figure 7.9 can be seen as a
diagonal system of local loops with a parallel structured uncertainty. One way of
addressing stability of G(q) is to investigate robust stability (see e.g., Zhou et al.,
1995) of G(q) with respect to the uncertainty. The concept of robust stability is
discussed briey in Section 7.D.1, and applied to the power control problem in
Section 7.D.2
7.D.1 Robust Stability
In this section, we introduce vector and matrix norms. Furthermore, some central
theorems are discussed and in some cases proved.
Let X be a vector space. A realvalued function   dened on X is said to be
a norm on X if it satises the following properties:
(i) x 0;
(ii) x = 0 if and only if x = 0;
(iii) kx = [k[x for any scalar k;
(iv) x +y x +y.
7.D Robust Stability of Power Control Algorithms 159
There are numerous denitions of norms that satises these properties. For our
purpose, the focus will be on norms, which satises the properties above and
are dened as below.
Denition 7.D.4 (Vector norm)
Let x C
n
with components x
j
, 1 j n. The vector norm of x is dened by
x
= max
1jn
[x
j
[
In a sense, the vector norm can be thought of as an extension of our usual concept
of length. Similarly, we can introduce some kind of measure of a matrix.
Denition 7.D.5 (Matrix norm)
Let A = [a
ij
] C
mn
. The matrix norm induced by a vector norm is dened
by
A
= sup
x=0
Ax
x
Given a matrix, it is not easy to compute the norm from the denition. However,
it is easy to verify that the following lemma holds.
Lemma 7.D.6
The matrix norm in Denition 7.D.5 can be computed by
A
= max
1im
n
j=1
[a
ij
[ (row sum)
From a system perspective, the induced matrix norms have the natural interpre
tation of input/output amplication gains. This far only constant matrices have
been considered. With the system interpretation of input/output gains in mind, it
is interesting to extend the denition to transfer matrices G(q).
Denition 7.D.7 (Zhou et al., 1995, p. 114)
Let G(q) be a transfer matrix. Then the induced matrix norm is dened by
G(q)
= sup
w
G(e
jw
)
Remark 7.D.8
In a single input, single output (SISO) case, an interpretation is that the norm
of a stable transfer function is equal to the maximum magnitude of the steady
state response to all possible unit amplitude sinusoidal input signals. This is easily
concluded from the fact that
G
SISO
(q)
= sup
w
G
SISO
(e
jw
)
= sup
w
[G
SISO
(e
jw
)[.
160 Global Analysis
In a perfectly decoupled situation of identical SISO control loops operating in
parallel, the overall transfer matrix is given by
G(q) =
G
ll
(q) 0 . . . 0
0
.
.
.
.
.
.
.
.
.
.
.
. 0
0 . . . 0 G
ll
(q)
(7.D.10)
For this conceptually interesting situation, the following lemma can be formulated:
Lemma 7.D.9
Let G(q) be a diagonal transfer matrix as in (7.D.10). Then, the induced matrix
norm is given by
G(q)
= G
ll
(q)
= sup
w
[G
ll
(e
iw
)[
Proof For every w0, the following holds
G(e
jw
0
)
2
= max eig
{G(e
jw
0
)
G(e
jw
0
)
=
= max eig
{G
ll
(e
jw
0
)
G
ll
(e
jw
0
)
=
=
G
ll
(e
jw
0
)
2
,
where the second last equality follows from (7.D.10). Hence
sup
w
G(e
jw
)
= sup
w
G
ll
(e
jw
)
.
Since G
ll
(q) is a SISO system, the proof is concluded by considering Remark 7.D.8. 2
w
1
w
2
(q)
G(q)
Figure 7.14 Setting to which the Small Gain Theorem applies.
Theorem 7.D.10 (Small Gain Theorem, Zhou et al.,1995, p. 217)
Let G(q) be a stable real rational transfer matrix. Then the interconnected system
shown in Figure 7.14 is internally stable for all stable real rational (q) with
7.D Robust Stability of Power Control Algorithms 161
(i) (q)
1/.
(ii) (q)
< 1/.
A more restricted situation is considered in the following corollary.
Corollary 7.D.11
Let G(q) be a diagonal transfer matrix with identical stable real rational SISO
blocks G
ll
(q) as in (7.D.10). Then the interconnected system shown in Figure 7.14
is internally stable for all constant matrices (q) with
(i) (q)
cc
in Lemma 7.18

cc

= max
1im
1
v
i
I
t
i
=
cc
< 1
The value of the norm
cc
will be referred to as the degree of crosscoupling.
Proof According to Lemma 7.D.6 and Lemma 7.17, the matrix norm can be computed
as
cc = max
1im
m
j=1
Ii(p)
pj
p=p
t
.
Lemma 7.18 yields
cc = max
1im
j=i
t
i
gij p
t
j
gii p
t
i
,
since the gains and the powers all are positive. Furthermore,
cc = max
1im
t
i
I
t
i
vi
gii p
t
i
= max
1im
1
vi
I
t
i
= cc < 1.
2
162 Global Analysis
Theorem 7.15 (Global Stability of Power Control Algorithms)
Let G
ll
(q) be the stable closedloop transfer function of the local loop in Figure 7.5.
Then the global system in Figure 7.9 is stable if and only if any of the following
properties is satised
(i) sup
w
[G
ll
(e
iw
)[ 1/
FI
.
(ii) sup
w
[G
ll
(e
iw
)[ < 1/(
cc
FI
),
where
FI
= sup
w
[F
I
(e
iw
)[
cc
= max
1im
1
v
i
I
t
i
1
vi
I
t
i
= cc < 1 (7.11)
Transfer matrices like FI (q) are addressed in Lemma 7.D.9. Hence
FI(q) = sup
w
FI (q)(e
iw
) = F
I
. (7.12)
CauchySchwarz inequality (Zhou et al., 1995) yield
FI(q)cc FI(q)cc = ccF
I
(7.13)
Moreover, we note that the situation is analogous to the one in Corollary 7.D.11 with
FI(q)cc as the uncertainty matrix. The corollary together with Equation (7.11) leads
directly to the properties and concludes the proof. 2
Corollary 7.16
Assume that the interference lter F
I
(q) is either a local average lter or an expo
nential forgetting lter. Then the global system in Figure 7.9 is stable if and only
if any of the following properties is satised
(i) sup
w
[G
ll
(e
iw
)[ 1.
(ii) sup
w
[G
ll
(e
iw
)[ < 1/
cc
,
Proof Consider the local average lter FLA(q).
FLA(q) =
L1
k=0
q
k
.
Note that
FLA(e
iw
) =
1
L
L1
k=0
e
iwk
1
L
L1
j=0
e
iwk
= 1,
7.D Robust Stability of Power Control Algorithms 163
with equality for w = 0. The exponential forgetting lter FEF(q) is given by
FEF(q) =
(1 )q
q
.
The following hold
FEF(e
iw
) =
1
 cos w + i sin w
=
1
1 +
2
2cos w
1,
with equality for w = 0. Hence sup
w
FI(e
iw
) = 1 for both choices of lter. 2
164 Global Analysis
8
Nonlinear Estimation
Hitherto, the focus has been on algorithms that either utilize analog measurements
(information feedback) or has to rely on whether a measurement is either above or
below a threshold (decision feedback). In practice, the situation might be some
where in between. Consider e.g., GSM, where the measurements are subject to
coarse quantization, and made available in measurement reports. One core prob
lem is then to locally extract as much relevant information as possible from these
reports.
Proposed methods for SIR estimation include schemes that utilize analog sig
nal strength measurements (Andersin et al., 1998a; Ramakrishna et al., 1997;
T urkboylari and St uber, 1998). These algorithms are rather similar, and show
good performance. Such schemes are useful when the SIR:s are used in the receiver,
e.g., by employing decision feedback as in the FSPC algorithm (see Section 5.2.1).
However, in systems like GSM and DAMPS, the measurements are used for power
control by the connected base station. The receiver measurements in a downlink
situation, thus have to be transmitted over the radio interface facing a low sig
naling bandwidth requirement. These measurement reports commonly comprise
quantities reecting the perceived quality and signal strength. Hence, information
is lost due to quantization and sampling at a rate that might be as low as 2 Hz
(GSM case). The same signaling interface is used in an uplink situation between
the receiver and the unit computing output powers in the base station.
In this chapter, we discuss the applicability of Maximum Likelihood Estimation
165
166 Nonlinear Estimation
to the specic GSM example, and comment on readily generalizations. In essence,
it is based on a probabilistic description of the problem. The needed probability
functions of the measurements cannot be described analytically in general. Instead
pointmass approximations can be obtained from MonteCarlo simulations for each
point in a grid covering the interesting parameter space. This modeling is cov
ered in Section 8.1. Maximum Likelihood Estimation is reviewed and applied in
Section 8.2, and the chapter is concluded by some illuminating simulations. The
proposed estimation implementation is applied and further evaluated in network
simulations in Chapter 10 and in (Blom and Gunnarsson, 1998; Blom, 1999).
8.1 Modeling
The perceived quality and related quantities at a receiver can be seen as stochas
tic variables, parameterized by some parameters. In order to employ maximum
likelihood estimation, we need the parameterized probability functions of the mea
surements. First, the choice of parameters is motivated. In general, it is hard
to derive the measurements parameterized probability functions analytically. In
stead, we utilize pointmass approximations of the same, over a grid covering the
interesting parameter space.
The estimation approach outlined in this chapter is focused on the specic case
of GSM, and some generalizations are outlined where appropriate. Furthermore,
the approach is local and considers each user individually. Therefore, the mobile
station index i will be dropped for clarity. Simulations are central in this chapter.
Further details on the simulation environment are found in Section 10.3.
8.1.1 Parameters
Essentially, we are interested in the distribution of the interference power I and the
desired signal power C. The latter is subject to variations of the same order as the
sample interval, and it will therefore be considered as a deterministic parameter.
Another motivation is that the RXLEV measurement describes C well, so the
estimator could be able to track relatively fast variations in C. The interference I
is considered as a stochastic variable with distribution
I F(m
I
,
I
, x),
where m
I
is the mean,
I
the standard deviation, and x represents additional
parameters. The characteristics of the interference distribution are approximated
by using a simulation model. The gains of the transmitted powers in a random
frequency hopping network were modeled by the path loss, shadow fading and
multipath fading. Thermal noise was also included in the model. The results
are found in Figure 8.1, from which we conclude that it is reasonable to model
the interference as Gaussian. This result proved to be relatively independent not
only of network specic parameters such as cell radius and reuse, but also of the
distribution of the transmitted powers. Most probably, there are more accurate
8.1 Modeling 167
models, but the issue is essentially to nd a model that is good enough to serve
the purpose of modeling the interference for estimation. The conclusion is that the
120 115 110 105 100 95 90 85
0
50
100
150
200
250
300
Interference (dB)
N
o
.
o
f
s
a
m
p
le
s
in
in
t
e
r
v
a
l
Figure 8.1 Interference distribution in a random frequency hopping net
work.
interference distribution can approximately be characterized by its mean value m
I
and its standard deviation
I
(dierent for each user). This is a result in the same
direction as in (Schwartz and Yeh, 1982). Thus the parameters to be estimated in
the GSM case are
= [C, m
I
,
I
]
T
.
8.1.2 Measurements
As argued in the previous section, the interference is characterized by its mean,
m
I
, and standard deviation,
I
. These will aect the outcome of the measurements
together with the desired signal power, C = p + g. The measurements can be
modeled as realizations of stochastic variables with probability functions depending
on these parameters.
h
k
= h
k
(C, m
I
,
I
), k = 1, 2.
In GSM the measurement reports consist of RXLEV and RXQUAL (see Ap
pendix 8.A). RXLEV is a signal strength measure, which has been quantized in
64 levels, and RXQUAL is a logarithmic measure of the Bit Error Rate (BER),
quantized in 8 levels. The available measurements at the receiver in focus x =
[x
(1)
, x
(2)
]
T
can thus be expressed by
RXLEV : x
(1)
= h
1
(C, m
I
,
I
)
RXQUAL : x
(2)
= h
2
(C, m
I
,
I
) ,
where the functions h
k
() return realizations of stochastic variables.
Furthermore, a natural assumption is that RXQUAL is primarily related to SIR
and its distribution. By considering the fast variations in the desired signal power
C as noise, it is relevant to model the RXQUAL probability function as
RXQUAL : x
(2)
= h
2
(C m
I
,
I
)
168 Nonlinear Estimation
In general it is not possible to get simple analytical expressions for these functions.
However, pointmass approximations can be obtained from simulations for each
point in a grid covering the interesting parameter space. Consider GSM and dene
this grid as in Figure 8.2, where the grid resolution is a tradeo between computer
capacity and accuracy. The range of the grid has to cover all possible values in
all directions. The estimator is assumed to operate in a system employing power
control. Therefore it is more ecient to parameterize the grid in C m
I
instead
of m
I
, since the power control will result in less variations in the former. The
grid utilized in this chapter is characterized by the parameters in Table 8.1 More
Range, Cdirection 68 dB
Range, C m
I
direction 30 dB
Range,
I
direction 10 dB
Resolution 0.5 dB
Table 8.1 Grid parameters used in the implementation in this chapter.
on pointmass approximations of probability functions can be found in (Bergman,
1999) and the references therein.
I
C m
I C
Figure 8.2 Using pointmass approximations, the probability functions are
approximated over a grid covering the interesting parameter
space. The grid is parameterized in Cm
I
instead of m
I
, since
the power control will result in less variations in the former.
8.2 Maximum Likelihood Estimation 169
Given a point (i.e., a set of parameter values), C and Isequences can be gen
erated, from which the measurement report can be formed using models of the
modulation and coding. MonteCarlo simulations yield pointmass approximations
of the probability functions as described more detailed in the algorithm below
for the case of RXQUAL. The resulting probability function is depicted in Fig
ure 8.3. The corresponding procedure can be applied when forming the probability
function of RXLEV. Note that these tedious computations are made once and for
all.
Algorithm 8.1 (Pointmass Approximation of the RXQUAL Proba
bility Function)
1. Dene a grid covering the interesting parameter space as in Figure 8.2.
2. For each point k in the grid corresponding to [C
(k)
, m
I
(k)
,
I
(k)
], generate
an Isequence I
(k)
N(m
I
(k)
,
I
(k)
) over 104 bursts (GSM, 1994).
3. Compile a SIRsequence
(k)
= C
(k)
I
(k)
.
4. Use link models (Olofsson, 1997) of the modulation and coding to compute
the Bit Error Probability (BEP) for each burst.
5. Generate Bit Errors Bin(x, BEP), where x represents the number of con
sidered bits per burst, i.e., the 26 bits in the training sequence. Moreover,
compute the Bit Error Rate (BER).
6. Compute RXQUAL using the GSM specication (GSM, 1994).
7. Repeat the process N times (MonteCarlo) for each point k.
8. Return to 2. and repeat the process for the next point k + 1.
8.2 Maximum Likelihood Estimation
The estimator can be constructed in numerous ways, but we have chosen a Max
imum Likelihood (ML) estimator (Lehmann, 1991; Ljung, 1999), since it success
fully enables data fusion and is an implementationally simple algorithm. However,
it involves some demanding computations, which might be an important issue.
The method of ML estimation is based on a simple idea. Dierent probability
density functions generate dierent data samples and any given data sample is
more likely to have come from a particular distribution than from others. The
basic ideas will be illustrated by the following example.
8.2.1 A Simple Example
Consider an exponentially distributed stochastic variable X, parameterized by
X Exp() , f
X
(x; ) =
1
e
x/
170 Nonlinear Estimation
0
5
10 0
10
20
30
0.0
0.5
1.0
0
5
10 0
10
20
30
0.0
0.5
1.0
0
5
10 0
10
20
30
0.0
0.5
1.0
0
5
10 0
10
20
30
0.0
0.5
1.0
0
5
10 0
10
20
30
0.0
0.5
1.0
0
5
10 0
10
20
30
0.0
0.5
1.0
0
5
10 0
10
20
30
0.0
0.5
1.0
0
5
10 0
10
20
30
0.0
0.5
1.0
P
(
R
X
Q
U
A
L
=
0
)
P
(
R
X
Q
U
A
L
=
1
)
P
(
R
X
Q
U
A
L
=
2
)
P
(
R
X
Q
U
A
L
=
3
)
P
(
R
X
Q
U
A
L
=
4
)
P
(
R
X
Q
U
A
L
=
5
)
P
(
R
X
Q
U
A
L
=
6
)
P
(
R
X
Q
U
A
L
=
7
)
I I
I
I
I I
I I
C mI C mI
C mI C mI
C mI
C mI
C mI C mI
Figure 8.3 The probabilities that dierent RXQUAL values are measured,
for dierent values of C, m
I
and
I
.
Let the true value be = 5. Assume that a measurement x
1
= 4.3 has been
observed. In order to estimate , we form the likelihood function
l
1
() = f
X
(x
1
; ) =
1
e
x1/
computed for a number of grid points covering the interesting values of , see
Figure 8.4a. The natural estimate is the value maximizing this function. When
observing a second measurement (x
2
= 12.1), the procedure could be repeated
analogously as in Figure 8.4b. Even better is to utilize both measurements using
the joint likelihood function, dened as the product of the two likelihood functions.
The improvements of the estimate are illustrated in Figures 8.4cd. Numerically,
it may be more appealing to use logarithmic values. Then the joint likelihood can
8.2 Maximum Likelihood Estimation 171
0 2 4 6 8 10 0 2 4 6 8 10
0 2 4 6 8 10 0 2 4 6 8 10
a. f(x1; ) b. f(x2; )
c. f(x1; )f(x2; ) d. f(x1; ) . . . f(x5; )
Figure 8.4 The dierent likelihood functions in the simple example in sec
tion 8.2.1.
be dened as
log l
t
() =
t
i=1
log f
X
(x
i
; )
8.2.2 Adaptive Estimation
This far a xed parameter has been assumed. In a real situation, however, the
parameter may be timevarying, and therefore adaptivity is important. This is
introduced by employing exponential forgetting of the joint likelihood function,
which can be implemented by the following recursion
log l
t
() = (1 ) log f
X
(x
t
; ) +log l
t1
().
The estimate is then obtained as
ML
(t) = arg max
l
t
() = arg max
log l
t
().
When there are several parameters to be estimated and several measurements
available, all of the above apply. The measurement at time t, x
t
, and the parameter
are now vectors. A fundamental result from statistics states that when the
measurements are independent, the joint probability function is obtained as the
172 Nonlinear Estimation
product of the probability functions of each of the M measurements.
f
X
(x
t
; ) =
M
i=1
f
Xi
(x
(i)
t
; ).
If the probability function is (close to) zero at some grid points for a certain
measurement, the value of the likelihood function will remain (for a while close to)
zero at those points. Consequently, the likelihood function is blocked from growing
at those points, inhibiting the adaptivity to varying parameters. A solution is to
use a threshold value, f
min
, and when updating the likelihood use the probability
function
g
X
(x
t
; ) = maxf
X
(x
t
; ), f
min
.
Finally, the parameters may be changing dierently fast. The carrier and the
mean interference are typically fastvarying, while, the standard deviation of the
interference is varying more slowly. This is solved by postltering the estimates
using exponential ltering (Blom and Gunnarsson, 1998; Gunnarsson et al., 1998d).
with dierent forgetting factors
j
j
(t + 1) = (1
j
)
ML
j
(t + 1) +
j
j
(t).
This procedure is shown to be asymptotically ecient (Gustafsson, 2000; Kushner
and Yang, 1995; Polyak and Juditsky, 1992)
Thus, we propose the following estimation algorithm:
Algorithm 8.2 (Maximum Likelihood Estimation)
Let the measurements at time t, be given by the vector x
t
, with the correspond
ing probability function
f
X
(x
t
; ) =
M
i=1
f
Xi
(x
(i)
t
; ).
Dene g
X
(x
t
; ) by
g
X
(x
t
; ) = maxf
X
(x
t
; ), f
min
.
Update the likelihood function according to
log l
t
() = (1 ) log g
X
(x
t
; ) +log l
t1
(),
with the initial likelihood l
0
equal to a Gaussian probability density function.
The estimate
ML
(t) is given by
ML
(t) = arg max
l
t
().
Postlter the estimates separately as
j
(t + 1) = (1
j
)
ML
j
(t + 1) +
j
j
(t)
8.3 SemiLinear Estimation 173
8.3 SemiLinear Estimation
This far, complexity has not been an issue, and computations are not seen as costs.
For comparison, it is therefore interesting to design a simple estimation procedure,
where simplicity and computational complexity are main issues. In this section,
we discuss simple algorithms that can be posed as linear regressions. As in the
previous, the focus is on GSM with frequency hopping.
The probability functions of the measurements are modeled in Section 8.1. A
measurement report could be seen as outputs from two functions returning real
izations of stochastic variables.
RXLEV : x
(1)
= h
1
(C, m
I
,
I
)
RXQUAL : x
(2)
= h
2
(C, m
I
,
I
) .
Furthermore, a natural assumption is that RXQUAL is primarily related to SIR
and that RXLEV is related the total received power. Hence (with some abuse of
notation)
RXLEV : x
(1)
= h
1
(10 log(
C + m
I
)) = h
1
(C + 10 log(1 + m
I
/
C)) (8.1a)
RXQUAL : x
(2)
= h
2
(C m
I
) = h
2
(). (8.1b)
The relation between RXQUAL and SIR is static However, the SIR values are
correlated via the power gain correlations. It is thus natural to model this relation
using an ARmodel
A(q)(t) = RXQUAL(t) +
(t)
A(q) = 1 +a
1
q
1
+. . . +a
na
q
na
Even simpler is to approximate this model with an FIRmodel, given by the fol
lowing linear recursion
(t) =b
0
+b
1
RXQUAL(t) + . . . +b
n
b
RXQUAL(t n
b
) +
(t) =
=[1, RXQUAL(t), RXQUAL(t n
b
)] [b
0
, b
1
, . . . , b
n
b
]
T
+
(t) =
=
T
(t)
b
+
(t). (8.2)
The model error is thus represented by
t=0
(t)
T
(t)
b
2
. (8.3)
Thus, the optimal vector is given by the leastsquares estimate (Ljung, 1999), which
is the solution to the linear equation
1
N
N1
t=0
(t)
T
(t)
=
1
N
N1
t=0
(t)(t).
174 Nonlinear Estimation
The remaining design parameter is the model order n
b
. Consider a new data
set (validation data), also obtained from simulations. Then the model order is
chosen as the parameter value that minimizes the criterion in (8.3) considering the
validation data. The results from the validation are summarized in Table 8.2.
According to the GSM standard (see Appendix 8.A), RXLEV is a linear function
of the receiver signal power level (RSL)
RSL = 133.5 + RXLEV [dB].
Using (8.1a) and given a SIR estimate,
(linear scale), a reasonable Cestimate is
given by
C = RSL 10 log(1 + 1/
) [dB]. (8.4)
The corresponding t with respect to the validation data is found in Table 8.2. As
a rst comparison, the t when using the nonlinear estimation procedure, is also
included. A more thorough comparison is provided in Section 8.4. From the table
we conclude that model order n
b
= 4 results in the best t.
Linear Recursion Non
n
b
= 1 n
b
= 2 n
b
= 3 n
b
= 4 n
b
= 5 linear
Std 2.99 2.31 2.18 2.12 2.24 1.73
StdC
C 0.825 0.810 0.809 0.808 0.810 0.304
Table 8.2 The root mean squared error when employing the linear recur
sion of dierent model order to the validation data. For com
parison, the nonlinear estimation procedure is also included.
8.4 Simulations
Performance evaluation is naturally carried out via simulations. The simulation
environment is GSM specic and the measurements are available as measurement
reports as in the real system. For comparison, the true analog values can be
extracted as well. Further details on the simulation environment is provided in
Section 10.3.
To illustrate the operation of the proposed estimation procedure, an artical
situation of an abrupt interference step is studied in Section 8.4.1. A more realistic
situation is considered in Section 8.4.2, with data from network simulations.
8.4.1 Abrupt Interference Step
Step responses of linear systems provide valuable insight in how fast the system
adapts to abruptly changing inputs. Similarly, it is interesting to study the adap
tivity of the proposed estimation procedure, when subject to an abrupt change of
8.4 Simulations 175
110
105
100
95
90
0
5
10
15
20
25
30
4
3.5
3
2.5
2
1.5
1
0.5
0
C
C m
I
Figure 8.5 In ML estimation, the estimates are obtained as the values
corresponding to the likelihood summit. For clarity, the
I
direction has been eliminated by selecting the true value.
the mean interference. Such a situation may arise, when new users are admitted to
the network, or when a user is allowed to use a higher data rate (i.e., a higher target
SIR and a higher transmission power). As concluded in Section 8.1, we consider
the following measurements and parameters
x = (RXLEV, RXQUAL)
T
= (C, m
I
,
I
)
T
.
All parameters are xed except m
I
which changes abruptly. In this case, the
simulator is only used to generate a sequence of measurement reports, which are
fed to the estimator. The estimation procedure is illustrated by the snapshot of
the likelihood at time t in Figure 8.5. As seen in the gure, the estimates are
given by the values corresponding the the peak value of the likelihood. When
subject to the abrupt interference step, the estimator adapts to the new situation.
Figure 8.6 illustrates the adaption of the likelihood function at time instants before
and after the step. Resulting parameter tracking is visualized in Figure 8.7. Note
the relations between the performance and the slope of the likelihood. The steeper
the slope, the better the accuracy.
8.4.2 Runtime Estimation
The normal situation is that the desired signal power C and mean interference
m
I
are subject to fast variations due to shadow and multipath fading, see St uber
(1996). The measurement reports in Figure 8.8a,b are obtained from network simu
lations. These are fed to the estimator and the estimated parameters are compared
to the true values, see Figure 8.8c,d. When evaluating the overall performance, sev
176 Nonlinear Estimation
C C
C C
C m
I
C m
I
C m
I
C m
I
Figure 8.6 The shape and summit position of the likelihood function
change, when the value of m
I
change abruptly. The maximum
of this function is seen to move in the (C m
I
)direction. The
I
direction has been eliminated by selecting the true value.
eral connections have to be considered. In the sample case above there are in total
190 calls of dierent durations (from 15 to 170 seconds) established during the
simulation. Then it is natural to use the root mean squared (RMS) error over all
times and calls for performance evaluation of the estimator. To allow for a burnin
phase of the estimator, the rst 5 seconds (N
b
= 10 samples) of each call were not
considered in the RMS errors. Furthermore, only calls longer than 20 seconds were
considered, leaving M=188 calls for RMS error computations. The RMS error of
the desired signal power estimate
C(t) is
RMS
C
C
=
1
M
1
N
i
N
b
Ni1
t=N
b
C
i
(t)
C
i
(t)
2
,
where N
i
is the duration of the call i. The absolute value of the RMS errors are
not informative unless related to parameter variations. This is addressed by the
RMS deection of the parameters from their average value. Desired signal power
RMS deection is obtained as
RMS
CC
=
1
M
1
N
i
Ni1
t=0
C(t)
Ci(t)
2
,
where
Ci(t)
is the average desired signal power during call i. For good performance
it is desirable to obtain small RMS errors compared to the expected variations, de
scribed by the RMS deection from the average value. The results are summarized
in Table 8.3. We note that the estimation of the desired signal power C is very
accurate, which is in accordance with the slope discussion related to Figure 8.5.
8.5 Summary 177
0 10 20 30
0
1
2
3
4
5
6
7
0 10 20 30
32
33
34
35
36
0 10 20 30
102
101
100
99
98
0 10 20 30
120
115
110
105
0 10 20 30
5
10
15
20
0 10 20 30
0
5
10
a.
b.
c. d.
e. f.
Figure 8.7 Estimation when there is an abrupt change in the true m
I
.
The plots above represents estimated values based on a sin
gle realization (dashed) and MonteCarlo over ve realizations
(solid). The true values are dotted. Measurement reports: a)
RXQUAL and b) RXLEV. Estimated parameters: c) Carrier,
C, d) Mean interference, m
I
, e) SIR, (C m
I
), f) Interference
standard deviation,
I
.
Moreover, we get acceptable accuracy for the SIR and interference mean estima
tions. Unfortunately, the
I
estimates are not as good.
8.5 Summary
In this chapter the focus is on estimation in cellular radio systems based only
on the signals that are readily available. Previous work has demonstrated very
good performance relying on analog measurement. In a real system most of the
information is lost due to quantization and sampling at a rate that might be as
low as 2 Hz (GSM case). Therefore a dierent approach is required and for that
matter a Maximum Likelihood Estimator has been designed and exemplied in the
case of GSM. In addition, a (mainly) linear estimator is designed for comparison.
The ML estimator need probability functions of the measurements, which can
not be described analytically. Instead pointmass approximations can be obtained
from MonteCarlo simulations for each point in a grid covering the interesting
parameter space. The adaption rate of the estimator is parameterized by a forget
ting factor, and represents the tradeo between adaptivity and noise reduction.
Similarly, parameters varying at dierent rates can be tracked individually using
postltering with individual forgetting factors.
Simulations illustrate that it is possible to extract the parameters despite coar
178 Nonlinear Estimation
0 5 10 15 20 25 30 35 40 45 50
0
1
2
3
4
5
6
7
0 5 10 15 20 25 30 35 40 45 50
12
14
16
18
20
22
t [s]
a.
b.
0 5 10 15 20 25 30 35 40 45 50
140
135
130
125
120
115
0 5 10 15 20 25 30 35 40 45 50
10
0
10
20
30
t [s]
c.
d.
Figure 8.8 Measurement reports, consisting of a) RXQUAL and
b) RXLEV, are describing the perceived quality and signal
strength respectively. These values are used as input to the
proposed algorithm, which extracts e.g. c) SIR and d) mean
interference, m
I
. The estimated values (solid) are compared to
the true ones (dashed).
Parameter
C m
I
=
C m
I
I
Average RMS error 0.30 [dB] 1.7 [dB] 1.7 [dB] 2.0 [dB]
Average RMS error (LR) 1.0 [dB] 2.0 [dB] 2.1 [dB]
Average RMS deection 5.1 [dB] 3.9 [dB] 5.0 [dB] 1.5 [dB]
Table 8.3 Performance evaluation of the estimator. The rst and second
rows describes the average estimation errors when using the pro
posed estimator and linear estimator respectively. For compari
son, the corresponding average deections from mean values are
included.
sely quantized measurements. The performance is good compared to typical pa
rameter variations, both when the parameters are varying slowly, and when subject
to fast variations as in realistic cases.
The algorithm is ready for implementation in a second generation wireless sys
tem. The only component that needs to be updated is the software in the base
stations, where the output powers are computed. However, this estimation method
is general, and could be applied in a third generation wireless system as well.
Appendix
8.A The Measurement Reports in GSM
In this appendix, we provide some facts about the measurements in GSM which
are relevant for this chapter. For further information, see (Steele, 1992, chap. 8)
and (GSM, 1994).
Every 480 ms, a measurement report is sent from the receiver to the control
mechanism in the connected base station. This report comprises:
Received signal strength, RXLEV.
Perceived transmission quality, RXQUAL.
Other information is also included, but of no value for estimation. We consider the
other information to be important only for the purpose of handover.
RXQUAL
The received signal quality, RXQUAL, is formed by estimating the bit error rate
before channel decoding, using metrics from the channel equalizer and/or the de
coder. Eight values of RXQUAL span the range according to Table 8.4.
BER interval RXQUAL
< 0.2% 0
0.2% to 0.4% 1
0.4% to 0.8% 2
0.8% to 1.6% 3
1.6% to 3.2% 4
3.2% to 6.4% 5
6.4% to 12.8% 6
> 12.8% 7
Table 8.4 The relation between bit error rates and RXQUAL.
179
180 Nonlinear Estimation
RXLEV
RXLEV is a measure of the Received Signal Level (RSL). It takes values in the
range 0 to 63, where RXLEV = 0 if the received signal level is less than 103 dBm,
RXLEV = 1 if 103 dBm RSL < 102 dBm,..., RXLEV = 63 if RSL > 41 dBm.
9
Outer Loop Control
In the preceeding chapters, power control has been considered as a means to meet
provided target values. A popular strategy is to issue transmission powers to meet a
target signaltointerference ratio (SIR) at each receiver. The problem is, however,
to assign a relevant target SIR. In systems based on CDMA, it is well established
that the power control implementation is based on a cascade structure, where an
inner loop adjust the power in order to track a target value provided regularly by an
outer loop (see e.g., Viterbi, 1995, and Chapter 4). This organization is benecial
in systems based on TDMA/FDMA/CDMA and combinations as well.
9.1 Introduction
The motivation for the outer loop is that SIR is not necessarily well correlated to
quality. Instead the percentage of lost bits and frames is more relevant, since the
eect of modulation, coding and interleaving is included. Therefore, the outer loop
provides a SIR target so that Bit Error Rate (BER) and/or Frame Erasure Rate
(FER) (these concepts are dened in Section 9.2.2) remain below some threshold.
The work on outer loops to date are mainly focused towards CDMA systems. In
(Sampath et al., 1997) the SIR target is decreased stepwise while no frames are
corrupted, and increased by a larger step when a corrupted frame is received. Ac
cording to Won et al. (1998), this type of scheme results in capacity improvements
of up to 25% compared to using a xed SIR target that meets the worst case. One
181
182 Outer Loop Control
problem is that since a low FER is desirable, the time constants of the controller
should be large in order to get reasonable statistics. Kawai et al. (1999) propose a
technique, where the relations between FER and BER after decoding are utilized in
order to speed up the control. Another way of speeding up the control is to model
the frame erasure distribution with respect to relevant parameters, and instead use
the estimated parameters for control. This is the approach used in this work. The
subjective matter of perceived quality in cellular radio systems is the subject of
Section 9.2. The discussion ranges from theoretical models to practical measures.
In GSM, the Frame Erasure Ratio is one of the more relevant quality measures,
and it is utilized in the quality mappingtechnique introduced in Section 9.3. This
work is reported in more detail in (Blom and Gunnarsson, 1998).
For completeness, the approach in (Almgren et al., 1994; Yates et al., 1997)
should also be discussed. The authors propose a dierent inner loop, where users
that can do with a low power are allowed to aim at a higher SIR target and vice
versa. Here, the inner loop is assumed to adjust transmission powers to track a
provided target SIR. However, the same eect can instead be obtained by adapting
the target SIR in the outer control loop as described in Section 9.4. Furthermore,
there might be situations that need specic actions. Together with a target value
tracking inner loop, this type of reasoning is naturally incorporated in the outer
loop.
The impact of using an outer loop to adapt target values is briey studied in a
specic simulation situation in Section 9.5. Related network simulations are found
in Chapter 10. Throughout the chapter, the focus is on a specic user. Therefore,
the index i is dropped for clarity.
9.2 Quality Measures in Cellular Radio Systems
The concept of Quality of Service naturally depend on the type of service and
its corresponding quality requirements. Some services are reliable in the sense
that they never loose data, while others in the sense that they provide very low
transmission delay. Often one diers between connectionoriented services and
connectionless services (Tanenbaum, 1996). In the former, a low delay is the rst
priority, and that category include speech and video services. The latter use data
acknowledgment so that the sender is aware of successful transmission. After a
specic time without acknowledgment, the data is resent.
The focus in this and the next section are on speech services. There is a wide
range of quality measures in cellular radio systems, ranging from speech quality
measures, which try to describe how a listener experience the quality, to more easily
measured quantities, which are dened in terms of bit errors, etc.
9.2.1 Speech Quality Measures
The performance of speech transmission could of course be investigated by extensive
subjective listener test, but when implementing algorithms based on the speech
9.2 Quality Measures in Cellular Radio Systems 183
quality it is necessary to have an automatic procedure for this evaluation. Thus,
one distinguishes between subjective and objective speech quality measures, with
the former being based on the opinions of the listeners, while the latter tries to
construct an adequate quality measure from signal measurements.
Since the concept of speech quality is of subjective nature, the objective speech
quality measures should be designed to be good estimates of a subjective measure.
Most of the proposed speech quality measures can be divided into four categories
(see Quackenbush et al., 1988).
SignaltoNoise (SNR) Ratio Based Measures. This measure is ob
tained by computing the mean square error between the original and the
distorted speech waveform. This is usually a poor estimator and therefore
various improvements exist. Since the human perception of sound is frequency
variant, some of these improvements are based on perceptually weighted mean
square errors.
Linear Predictive Coding (LPC) Based Measures. These measures
are based on a parameterization of the linear prediction vocal tract model
of the original and the distorted speech. The measure is computed from the
change of these parameters between the original and the distorted speech.
Spectral Distance Based Measures. These measures are based on com
parisons of smoothed spectra from the original and distorted speech signals.
There are several variants since they can be weighted with respect to en
ergy, frequency etc., and also since they sometimes are based on logarithmic
spectra.
Psychoacoustically Based Measures. These are measures based on a
model of the human auditory system, see (see e.g., Beerends and Stemerdink,
1994).
Lam et al. (1996) show that SNRbased and LPCbased measures are not very
successful in estimating subjective measures well. Some of the spectral distance
based measures perform well, but it is the psychoacoustically motivated measures
that show the best performance.
9.2.2 Practical Quality Measures
Speech quality measures are the most sophisticated quality measures for voice
communication, since they describe the properties of the actual speech waveform
reaching the listener. However, in practice it is unusual that these are used as
quality measures from a power control point of view, and therefore we will discuss
other more common measures.
The simplest and most usual quality measure is the signaltointerference ratio,
SIR. This is dened in terms of the properties of the radio wave, before entering
the receiver, see Figure 9.1. The main reason for using SIR as quality measure is
that it is directly related to the carrier power, which we can control. The result is a
184 Outer Loop Control
framework for power control where theoretical calculations can be done. Usually a
SIR which is above 10 dB is considered to correspond to good quality, and a smaller
SIR to poor quality. Due to the modulation and coding techniques, it is argued
that this is a threshold in the sense that the quality quickly becomes poor when
below this value, and increasing it above this level will not improve the quality
signicantly.
Message
Estimated
Message
Source
Encoder
Channel
Encoder
Channel
Decoder
Source
Decoder
Modulation
Demodulation
Receiver
Transmitter
Speech Quality FER BER
Figure 9.1 Dierent quality measures, and where they can be measured in
a digital communication system.
The next quality measure we address is the Bit Error Rate (BER). It can be
measured after the demodulator, and is dened as the average number of bits which
are erroneous as compared with the original bits entering the modulator. The actual
measurement is possible since some of transmitted bits are xed beforehand (a.k.a.
training sequence), and therefore already known at the receiver. If the carrier and
interference signal powers are constant, the BEP will be a function of the SIR, and
in this case they contain the equivalent information about the quality. However,
in reality the SIR is time variant, and thus the average SIR will not correspond to
the average BER, and in this sense BER is a better quality measure.
The purpose of the channel coding is to try to remove some (or all) of the bit
errors which have occured during the transmission. Depending on the design of the
channel encoder, the speech quality degradation will be more or less severe. After
the channel decoder, the next quality measure is dened. Here we could dene a
quality measure similar to the bit error rate, by considering individual bits after the
decoding. However, we dene a frame as the bits over which the coding is applied,
and assume that a frame will either be fully restored or completely useless. Then
the Frame Erasure Rate (FER) can be dened as the average number of frames
which are useless. As discussed before, the speech quality is the most appropriate
quality measure of a speech service, and this is obtained by comparing the original
speech waveform with the nal. It can be noted that the source coder will try to
represent the speech with as few bits as possible without loosing too much quality.
9.2 Quality Measures in Cellular Radio Systems 185
Therefore the estimated message will not be the same as the original even if no
interference would be present over the channel. Furthermore, if some of the bits in
the output from the source encoder are erroneous when entering the source decoder,
the speech quality degradation is strongly depending on how the encoder/decoder
has been designed. In order to use speech quality for control, we must have a model
describing how the source encoder/decoder is working.
To summarize, we have discussed dierent quality measures in a digital com
munication system, and how they are related to the functionality of the trans
mitter and receiver. By building models of the source encoder/decoder, channel
encoder/decoder and the modulation/demodulation procedures, more and more
adequate measures are produced, see Figure 9.2. The price for this is that we have
to put in more knowledge about the system, and the complexity of the algorithms
will increase.
SIR BER FER
Speech Quality
Model 1 Model 2 Model 3
Figure 9.2 The modeling necessary to obtain dierent quality measures.
Model 1 refers to the source encoder/decoder, model 2 to
the channel encoder/decoder and model 3 to the speech en
coder/decoder.
9.2.3 Example: GSM
For the mobile station user, the speech quality will in general get worse the more
data bits that are lost during the transmission. The speech quality is aected not
only by the fraction of bits that are erroneous, but also by the distribution of lost
bits over time.
We need to understand how the bits are arranged when transmitted over the
physical channel, see Figure 9.3. After the speech coding has been performed, the
resulting bits are divided into frames, with each frame corresponding to 20 ms of
speech. In the rst step coding is performed, resulting in 456 bits for each frame.
Furthermore, the bits are permuted, which is referred to as rst order interleaving.
Next, the second order interleaving splits the coded frame into 8 sets, each
containing 57 bits. Each of these sets is then mapped onto a burst, which is the
data from a certain user sent during one time slot. Thus the data from one speech
frame is separated on the channel. This implies that the information from one bit
in the original frame has been spread over dierent time instants on the channel.
Note that information from two dierent frames are mapped onto the same burst,
and therefore each burst contains 114 information bits. In practice the number of
bits in a burst is 156.25, since extra bits have been added to provide for proper
equalization and synchronization.
186 Outer Loop Control
Frame A Frame B Frame C Frame D
Coding and 1st order interleaving
8 57 = 456 bits
57 57 57 57 57 57 57 57
2nd order interleaving
Mapping on bursts
A
A A A B B B B
B B B B C C C C
TDMA frames
Figure 9.3 The mapping of speech frames onto the physical channel.
It is also interesting to reect over the dierent time scales. A typical distance
between multipath fades is about 0.15 m (half a wavelength) in GSM (Mehrotra,
1994). This can be compared to the distance a vehicle can travel during a burst.
Each burst is transmitted during 0.577 ms. Assuming that the vehicle travels
with 30 m/s, this would correspond to a distance of 0.577 10
3
30 0.017 m.
Thus, as a good approximation, the carrier can be regarded as constant during
the burst. During the 20 ms, which the speech frame is sent over, there will be a
considerable variation, since the same vehicle will travel 0.60 m during this time,
which corresponds to several fades. Since the information from one frame is spread
over this time interval on the channel, time diversity will be obtained. If frequency
hopping is used interferer diversity will also be obtained.
From SIR to Bit Error Probability
It is the SIR value during the burst, which gives a certain bit error probability
(BEP). It is important to notice the dierence between BEP and BER. The for
mer means the probability to get a bit error for a certain bit, while the latter is a
stochastic variable. In a received sequence of N bits with corresponding bit error
probability, BEP, the number of erroneous bits is a realization of the stochastic
variable X Bi(N, BEP). Furthermore, BER is equivalent to the stochastic vari
able
1
N
X. As argued in the previous section the SIR during a burst can be regarded
as constant. As a consequence, the probability for bit error during a burst, P
b
, is
9.2 Quality Measures in Cellular Radio Systems 187
a function of the SIR during that burst. For a typical result see (Olofsson, 1997)
and Figure 9.4
10 8 6 4 2 0 2 4 6 8 10
10
3
10
2
10
1
10
0
SIR [dB]
P
b
Figure 9.4 Uncoded performance for a GSM channel. A certain SIR for a
burst results in a certain bit error probability for that burst.
This plot has been generated by simulations for the case of two interferers. If
the number of interferers was not equal to two, we would obtain a dierent func
tion, since the demodulation performance is not only depending on the interference
power, but also on the interference characteristics. However, more interferers would
not signically change the bit error probability function. A more accurate BEP
vs. SIR function could be produced based on a similar simulation by modeling the
number of interferers as a stochastic variable. Usually the number of interferers
responsible for the main contribution of the interference is quite small. Therefore,
the twointerferer situation constitute a relevant good approximation. This is the
starting point for the rest of the discussion.
In GSM, RXQUAL is the quality related measure that is readily available. It
is based on estimated BER as discussed in Section 7A.
From Bit Error Probability to Bit Error Parameters
The frames are considered to either be fully restored or completely useless after
applying the error correcting codes. The probability that the frame is useless is
called the frame erasure probability (FEP). Note that we distinguish between FER
and FEP, just as we did for BER and BEP.
188 Outer Loop Control
This section follows Olofsson (1997); Olofsson et al. (1997) closely. Denote the
bit error probability at burst i by P
bi
, i = 1, . . . n, where n is the number of bursts
the frame is spread over, i.e. n = 8 in GSM. The number of bits in a burst is
denoted by k, which is 114 in GSM. Since the number of bit errors during a burst
is binomially distributed, the stochastic variable X
i
denoting the BER for burst i
will fulll
kX
i
Bin(k, P
bi
)
E[X
i
] = P
bi
V [X
i
] =
1
k
P
bi
(1 P
bi
)
Next, dene
M =
1
n
n
i=1
X
i
S
2
=
1
n 1
n
i=1
(X
i
M)
2
and
BER
and s
BER
as the corresponding expectation values, which can be shown
(Olofsson, 1997) to be related to P
bi
as
BER
= E(M) =
1
n
n
i=1
P
bi
(9.1a)
s
2
BER
= E(S
2
) =
1
nk
n
i=1
P
bi
(1 P
bi
) +
1
n 1
n
i=1
(P
bi
BER
)
2
(9.1b)
These parameters will now be used for estimating the frame erasure probability.
From Bit Error Parameters to Frame Erasure Rate
For GSM, FEP will in general be a function of P
bi
, i = 1, . . . , 8. However, we will
make the simplifying assumption (Olofsson, 1997) that FEP is only a function of
BER
and s
BER
.
FEP = f(P
b1
, P
b2
, . . . , P
b8
) g(
BER
, s
BER
)
By using a model of the channel encoding/decoding process in Figure 9.1, it is
possible to estimate the frame erasure probability from these parameters, i.e. the
function g, discussed above. In Figure 9.5 the probability for a frame erasure is
shown as a function of
BER
and s
BER
. Note that this plot contains some points
which can never occur in practice, for instance when
BER
= 0, s
2
BER
must also
be zero. These have been replaced by a zero value. The plot describes the ability
of the interleaving and error correcting codes to restore the frame in the presence
of bit errors.
9.2 Quality Measures in Cellular Radio Systems 189
0
0.1
0.2
0.3
0.4
0
0.05
0.1
0.15
0.2
0.25
0.3
0
0.2
0.4
0.6
0.8
1
s
BER
BER
F
E
P
Figure 9.5 FER performance for the GSM full rate speech channel as a
function of bit error characteristics.
From Frame Erasure Rate to Speech Quality
In this example we never reached the nal step, a quality measure. Instead the
frame erasure rate is used as a quality measure since it can be shown to describe
speech quality quite well and is well dened. Moreover, it is easily calculated from
the carrier and the interference characteristics according to the method described
above. The drawback with this approach is that the time distribution of frame
erasures is not taken into account.
Summary of the Example
The method for achieving the FEP can be summarized as follows:
1. A speech frame spreads its bits over eight bursts. Thus, given a sequence of
eight burst SIR values, P
bi
, i = 1, . . . , 8 are determined from Figure 9.4.
2.
BER
and s
BER
are determined from the P
bi
s using Equations (9.1a,b).
3. Map each parameter pair
BER
, s
BER
on a frame erasure probability using
the results in Figure 9.5.
This method for constructing the FEP from the SIR:s during subsequent bursts
will be important in the sequel, when we use the FEP as quality measure to be
used in the outer loop control.
190 Outer Loop Control
9.3 Quality Mapping
The need for a model in control highly depends on the time constants of the system
to be controlled. If the eect of the control signal is rapidly reected in the system
output, the output itself may be informative enough to control the system. The
perceived quality in cellular systems is indeed slowly varying, and therefore the
quality models derived in Section 9.3.1 are adequate. The parameter ranges in the
pointmass approximations are discussed in 9.3.2, and the use of the models for
outerloop control of target SIR is outlined in 9.3.3.
9.3.1 Modeling
Similar to the discussion in Section 8.1, the quality function is modeled as a prob
ability function, H, depending on the desired signal power and the interference
characteristics. In the GSM example, it is relevant to parameterize this function
using = C m
I
and
I
.
H = H(,
I
)
The procedure described in Section 8.1.2 together with the procedure in Sec
tion 9.2.3 can be used to form a pointmass approximation relating FER to the
estimated parameters. The result is provided in Figure 9.6, from which we conclude
that the mean SIR does not fully describe FER. It is dependent of the interfer
ence standard deviation as well. Moreover, it is depending on the coding scheme
and modulation employed. Thus when dierent coding modes are applied dur
ing the lifetime of the connection (rate adaption, see Section 9.4.2), corresponding
pointmass approximations can be formed for each mode.
10
5
0
5
10
15
20 0
2
4
6
8
10
0
0.2
0.4
0.6
0.8
1
[dB]
I
[dB]
F
E
R
Figure 9.6 The relations of FER and the estimated parameters = Cm
I
and
I
.
9.3 Quality Mapping 191
9.3.2 Parameter Ranges in PointMass Approximations
The objectives of the outer loop is to issue target values for the inner loop to track,
in order to meet a specied quality level, z
t
with respect to the quality measure z.
In GSM, it is relevant to use FER as the quality measure z. The requirement is
met if
H(,
I
) z
t
. (9.2)
The probability function H(, ) was dened and modeled in Section 9.3.1. It has
been assumed (without loss of generality) that a lower value of z corresponds to
better quality.
In the grid of the corresponding pointmass approximation, SIR is restricted to
values in the specied set
1
, . . . ,
r
, which are assumed sorted in ascending order.
The grid is similar in the
I
direction. The quality function computed in Sec
tion 9.3.1 can be used to determine if the chosen grid provides suciently coverage
of the parameter space in the direction. The maximum considered SIR (
r
) has
to correspond to a quality level that is less than the specied quality, otherwise we
can not expect to nd an appropriate SIR target, meeting this specication. More
formally. let =
r
be xed and search over the rest of the grid for the maximum,
z
0
, of the quality function
z
0
= max
H(
r
,
I
)
This process is illustrated for GSM in Figure 9.7. If the specied quality, z
t
, is less
5
0
5
10
15 0
2
4
6
8
10
0
0.2
0.4
0.6
0.8
1
C mI
I
F
E
R
z0
r
Figure 9.7 The FER plot can be used to nd the least quality value that
can be specied, z
0
, by a search along the maximum considered
SIR ( =
r
) in the grid.
than z
0
the grid has to be expanded to include higher values of .
192 Outer Loop Control
9.3.3 Implementation
Given measurements or estimates of the interference standard deviation
I
and a
quality specication z
t
, the aim is to nd the corresponding target SIR,
t
. If the
function H is invertible and monotonically decreasing in , the implementation is
straightforward. The required target SIR given the general grid point
(k)
I
is then
obtained by inverting Equation (9.2)
t
H
1
(
(k)
I
, z
t
).
The lookup table which spans over all grid points
(k)
I
is thus given by
t
= H
1
(
(k)
I
, z
t
). (9.3)
If H is not strictly decreasing in the direction, an alternate procedure is
needed when computing the lookup table. In such a situation, the table entry for
each point
(k)
I
in the grid is then computed as
j
k
= min
i
i[H(
i
,
(k)
I
, H(
i+1
,
(k)
I
, . . . , H(
r
,
(k)
I
) z
t
(9.4a)
t
(
(k)
I
) =
j
k
(9.4b)
For clarity, consider the following example. Assume that some H is approximated
in a specic grid point
(k)
by the values in Figure 9.8. Then the corresponding
target SIR for this grid point is given by the encircled value. By repeating the
procedure for each grid point, the lookup table is formed.
2 4 6 8 10 12 14 16 18
i
H(i,
(k)
I
z
t
Figure 9.8 The quality function might vary around the specied value in
the general case. Therefore we have to make sure to pick the
rst value after the last intersection as indicated in the gure.
The result of this procedure is a lookup table, that relates the measured or
estimated interference standard deviation
I
to a
t
. Interpolation may be used
9.4 Stating Priorities in the Outer Loop 193
when the parameters do not t the grid perfectly, which most likely will be the
case.
In the frequency hopping GSM example, it is reasonable to specify (McGregor ,
editor)
z
t
= FER
t
= 0.02.
As seen in Figure 9.6, the quality function H is monotonically decreasing in the
SIR direction. Thus, the procedure in (9.3) is applicable. (The procedure in (9.4)
would yield the same result.) The corresponding lookup table is thus given by
t
= H
1
(
I
, 0.02).
This is essentially the level curve of the plot in Figure 9.6 for z
t
= FER
t
= 0.02.
Target SIR as a function of the estimated interference standard deviation is illus
trated in Figure 9.9.
0 1 2 3 4 5 6 7 8 9 10
6
7
8
9
10
11
12
13
14
15
t
I
Figure 9.9 Quality mapping is implemented in the outer loop as a lookup
table, which relates estimated interference standard deviation
to appropriate target SIR:s. This lookup table is generated to
meet the quality specication FER 0.02.
Once again most of the computations can be completed before the runtime
operation of the system. The time constant of the adaption can be adjusted by
lowpass ltering the standard deviation of the interference. The resulting lookup
table requires more or less no computer capacity except memory.
If dierent coding modes are enabled, one lookup table per mode has to be
computed. In runtime, the computations using the lookup tables are again es
sentially of no concern.
9.4 Stating Priorities in the Outer Loop
Some situations may require special control actions. In order to use a target track
ing device in the inner loop, these actions may be realized by modifying the target
194 Outer Loop Control
SIR:s provided to the inner loop. In this section, we describe dierent algorithms
realized in the outer loop. The common framework is that the actual target SIR
provided by the outer loop is a function of a nominal target SIR,
t
, and other
relevant parameters. The nominal target SIR is possibly constant and predened
or adapted by a separate mechanism, such as the Quality Mapper described in the
previous section. The actual target SIR is thus given by
t
= f(
t
, ),
where f is a procedure as described in the sections below.
9.4.1 Fading Margins
When the bandwidth of the power control algorithm (see Section 6.4) is insucient
to track the varying environment, the perceived quality is hampered. Typically,
the algorithm can mitigate shadowing eects, but not fast fading or multipath. To
improve the performance, a fading margin
bias
may be employed (Andersin and
Rosberg, 1996; Blom and Gunnarsson, 1998, and Section 6.4) as
t
=
t
+
bias
. (9.5)
The design of the fading margin is essentially a tradeo between performance and
capacity. A high fading margin improves the performance, but increase the overall
interference in the system. Thereby, the capacity is reduced, and vice versa.
9.4.2 Rate Adaption
When a user is subject to less favorable propagation conditions, a number of actions
are possible. The transmission power could be increased to improve SIR. This, how
ever, increases the interference experienced by others and also signicantly increase
the load of the system (see Section 7.1 and Example 7.3 in particular). A dierent
approach is to lower the data rate. That might enable reliable communication due
to increased code protection, while not increasing the interference to other termi
nals. Further details on rate adaption are found in (Balachandran et al., 1999;
Bruhn et al., 1999; Eriksson et al., 1999; Furusk ar et al., 1999a,b; Lau and Maric,
1999; Uvliden and Bruhn, 1998) and the references therein. Typically, each base
station is adapting the rate of its connected mobiles. Such a distributed approach
might result in stability problems. Rezaiifar and Holtzman (1999) provides a proof
of convergence of one such strategy.
The achievable data rate is related to target SIR, and rate adaption can thus be
seen as target SIR adaption. A possible implementation is to compute a lookup
table for each of the coding modes as described in Section 9.3. Then, each rate is
associated with a lookup table in the outer loop control. In such a scheme, power
control is used in the inner loop to compensate for fast variations due to fading,
while rate adaption is used to map target SIR:s via lookup tables in the outer
loop.
9.4 Stating Priorities in the Outer Loop 195
9.4.3 Soft Dropping Power Control
A related approach is discussed in (Almgren et al., 1994; Yates et al., 1997), where
the quality requirements are adapted based on the current power level. Thereby,
a user with bad quality is prevented from disturbing others, since lower quality
requirements are enforced instead of enabling increased transmission power. Equa
tion (5.6) yield the following formulation of the algorithm
p
i
(t + 1) = p
i
(t) + ((
c
p
i
(t))
i
(t)). (9.6)
where and
c
are design parameters. Below, we discuss three dierent design
approaches:
1. A natural choice of
c
emanates from the statement (Almgren, 2000) that a
connection is allowed to use maximal power p
max
, if perfectly matching the
quality requirement SIR=
0
. Consider a connection using maximal power
while exactly achieving the quality requirement. Equation (9.6)
p
max
= p
max
+(
c
0
)
c
=
1
p
max
+
1
0
2. A dierent approach is to adapt
c
using the qualitymapping techniques in
Section 9.3, possibly via a direct adaption of
0
.
3. A third approach is essentially to implement the strategy in the outer loop.
To exemplify the ideas, we discuss the approach with respect to a specic
conguration. Consider a downlink situation, and assume that the FSPC
algorithm is used. Mobile station i observes the SIR
i
(t). In a traditional
setting (cf. Section 5.2.1), the observed SIR is compared to a threshold
t
i
(t)
to compute the power command bit s
i
(t). As advocated in Chapter 6, TDC
is benecial to utilize. One implementation is described in Algorithm 6.2,
which involves monitoring the powers to be used by the transmitter.
p
i
(t + 1) = max (p
min
, min(p
max
, p
i
(t) +
i
s
i
(t)))
Store the lowpassltered power p
i
(t) as
p
i
(t) = F(q) p
i
(t),
where F(q) might be one of the lters discussed in Section 5.2.4. This ltered
power can be used to realize Equation (9.6) as an integrating inner control
loop together with an outer loop with almost the same eect as the AAW
algorithm. Assume that the outer loop operates at a rate k times slower than
the inner loop, which in turn is updating the power every T
s
seconds. Hence,
196 Outer Loop Control
at times t = nkT
s
, n Z
+
, we update the target SIR
t
i
used in the mobile,
according to
t
i
(nk) =
c
p
i
(nk),
where is the step size of the inner control loop. The design parameter
0
may be chosen as in the two previous approaches. Furthermore, we possibly
use information from the base station, which can regularly be transmitted
using available outer loop protocols.
One important conclusion is that with this slower AAWadaption, the stability
results in Chapter 7 apply.
9.4.4 Mode Switching Controllers
In some systems it may not be plausible to design one inner loop that is relevant in
all situations. For instance, it may be desirable to utilize more aggressive control
actions when the system load is low (see Section 6.7), and soft dropping power
control when the load is high. Depending on the situation, the outer loop switches
between the dierent controllers. Then it is important to update the internal
states of the idle controllers, to obtain bumpless transfer when switching between
controllers. See also Section 6.6. Some of these switching philosophies may be
implemented in the outer loop, as illustrated in the following section.
9.4.5 Low Interference Problem
When the system load is low, the interference experienced by each user consists
basically of thermal noise. The target tracking of the inner loop will result in very
low powers, which might seem naive, since few others are disturbed by a higher
power (Almgren, 2000). However, when new users are admitted, the interference is
increased abruptly, resulting in bad quality until the controllers have adapted. In
this situation when literally now users are disturbed by others, it may be justied
to aim at a higher target SIR and better quality.
Using the estimated mean interference, dierent target SIR:s can be assigned,
for instance by adapting the fading margin in Equation (9.5). A possible imple
mentation is illustrated by the lookup table in Figure 9.10
9.5 Simulations
Outer loop control is most naturally studied in network simulations in Chapter 10.
Further simulations are also provided in (Blom and Gunnarsson, 1998). Therefore,
we only consider a single example of stating priorities in the outer loop. Further
details on the simulation environment is provided in Section 10.3.
In a specic situation, the load of the system is only 10 %. Then, the system
is essentially disturbance limited and the interference consists basically of thermal
9.6 Summary 197
Average Interference, mI
B
i
a
s
,
b
i
a
s
l
I
u
I
l
Figure 9.10 A possible implementation of fading margin adaption to im
prove the quality under favorable conditions.
noise. Low powers are sucient to provide acceptable connections, and low powers
are used if employing an integrating inner loop with xed target SIR. After about
14 s new users are admitted, resulting in increased interference. By using the mean
interference to adapt the fading margin as in Figure 9.10, the users are better
prepared to meet the increase in interference. The situation of a specic user is
illustrated in Figure 9.11. The main incentive for using a lower power is to decrease
system interference (and less important to reduce battery consumption). In this
case, interference to others is slight, and therefore, the benets of using a higher
power are more emphasized than the drawbacks.
9.6 Summary
The motivation for the outer loop in power control implementations is that SIR
is not necessarily well correlated to quality. The ratio of erasured frames is bet
ter correlated to quality, but hard to measure or estimate. Instead, the relation
between the estimated parameters and FER are described using pointmass ap
proximations. This model is used to relate a FER specication and parameter
estimates to a target SIR to the inner control loop. Moreover, it is emphasized
that the inner loop should focus on issuing power commands or power levels to
track a target value provided by the outer loop. Priorities and specic behavior in
certain situations should be reected in the provided SIR target. This also include
appropriate actions when rate adaption is employed.
198 Outer Loop Control
0 2 4 6 8 10 12 14 16 18 20
0
5
10
15
20
25
30
p
0 2 4 6 8 10 12 14 16 18 20
10
5
10
4
10
3
10
2
10
1
10
0
f
t [s]
t [s]
a.
b.
Figure 9.11 The perceived quality over time at a specic receiver given by
a. SIR and b. FER. The horizontal dashed lines corresponds
to = 10 dB and FER = 0.02. After about 14 s, new users
are admitted resulting in increased interference. The dashed
curves corresponds to using the xed target, and the solid to
prioritizing higher targets when the interference is low.
10
Network Simulations
The true impact of radio resource management algorithms is hard to describe
analytically. Instead, simulations provide valuable intuition into their behavior.
Numerous simulation studies have thoroughly considered the eects using such al
gorithms. The ambition here is not to thoroughly evaluate the performance and
capacity eects, but rather to illustrate the dynamical behavior of power control
algorithms. This is done in a small scale simulation environment, where most
other resource allocation algorithms are marginalized, as well as in more extensive
environments
1
where they are coexistent. Further details on the simulation envi
ronments are brought up in an appendix together with most plots from simulations.
10.1 Simulation Studies in General
Simulations may provide valuable insight into problems and behavior of proposed
solutions and algorithms. However, it is important not to jump the conclusions
from such simulation studies. As most scientists are aware of, simulations only
provide limited and uncertain information and knowledge. Conclusions never go
beyond underlying models. Verication by simulation is always hard to justify. Sir
Karl Popper is not yet outdated in his conclusion: all we know is but a woven web
1
Valuable simulation technology has been provided by dierent groups at Ericsson, which there
fore are acknowledged. Extensions, simulations and interpretations thereof are solely the
responsibility of the author.
199
200 Network Simulations
of guesses, that while empirical generalizations may not be veriable, they are, at
least, falsiable (Magee, 1973).
A typical simulation study includes problem description, modeling, algorithm
development and derivation, implementation and simulation. Of course, there are
numerous tractable paths to employ a study. Therefore, it is hard to compare and
evaluate results from dierent studies. Rosberg and Zander (1998) addressed the
problematic situation and emphasized the need for reference systems and bench
mark examples.
Absolute results depend on models of components in the simulator. It is always
important to separate the modeling of such components from implementation and
simulation. Even if the models are the same, the absolute results depend on im
plementation, parameter settings etc. Therefore, comparative studies of dierent
algorithms, which provide relative results, are more valuable.
The early phase of this work, reported in (Blom and Gunnarsson, 1998), was
to some extent a victim of the curse of simulationality
2
. The objective with the
simulations was essentially to verify or justify ideas and propositions. In this thesis
work, the ideas are well founded in the preceeding chapters, and the objective with
simulations is to illustrate dynamical eects.
10.2 Simple Network Simulation Environment
Several questions av been risen in several examples in the thesis, demonstrating
problems with and dierent aspect on power control. The objective here is to
illustrate the benets of careful power control design and to focus on dierent
properties of various controllers. The simulated cases are chosen with care, to
relate to examples and discussions from previous chapters.
10.2.1 MOSE
General purpose or technology specic simulation environments comprise numerous
control algorithms and degrees of freedom. It is therefore interesting to marginalize
the eects of related algorithms. The ambition with the MObile communications
Simulation Environment (MOSE) is to emphasize the dynamical eects of power
control algorithms. All mobiles are using the same frequency spectrum for commu
nication, a situation that might represent a specic channel in a system utilizing
orthogonal channels, or a simplied DSCDMA system. Channel allocation is not
considered, and base stations are allocated depending on either minimal distance
or maximal gain. The environment includes mobility models, path loss, shadow
fading and is described in more detail in Appendix 10.A and in (Gunnarsson et al.,
1997, 1998b). Since no models of modulation or coding are included, the natural
quality measure is SIR. Some central simulation parameters are summarized in
Table 10.1, and these are used if nothing else is stated.
2
A reader who fails to see the parallel to the curse of dimensionality has probably underestimated
the dimension of the simulation parameter space.
10.2 Simple Network Simulation Environment 201
Frequency reuse K = 1
Antennas Omnidirectional
Cell radius 1000 m
Cell layout 8 8 cells,
employing wrap around
Pathloss exponent = 3.5
Shadow fading std.dev.
s
= 6 dB
Shadow fading corr. dist. d = 100 m
Rayl. fading, avg. gain 0 dB
Time delays:
Power output delay n
p
= 1
Measurement delay n
m
= 0
Table 10.1 Central simulation parameters in the simple simulation envi
ronment.
10.2.2 Stabilizing Controllers in the Motivating Example
Example 5.1 illustrates the problematic eects of time delays in dynamical systems.
It rose a number of questions and thereby served as a motivation for the local loop
analysis in Chapter 5. Chapter 6 provides local design methods to deal with the
local dynamics. It is interesting to see how these algorithms improve the behavior
reported in Example 5.1
As in the example, we consider a situation where four mobiles are connected to
base stations using the same channel (or frequency spectrum). Initially (at time
instant t = 10), the mobiles are assigned optimal transmission powers correspond
ing to a target SIR of 12 dB. At time instant t = 0, the same channel is allocated to
a fth mobile station. The power control problem is feasible both before and after
the fth mobile entered the frequency spectrum. Furthermore, autointerference is
omitted, the mobiles are at xed locations, the Gmatrix is xed and the system
is subject to an additional time delay of one sample (n
p
= 1, n
m
= 0).
First, we focus on the same algorithms as in Example 5.1, with the dierence
that time delay compensation (TDC, see Section 6.2) is employed. The stabilizing
property of TDC is indicated by Figure 10.1a,b. The DPC algorithm now recovers
after the disturbance. It adapts quickly, which is in accordance with its deadbeat
behavior (see Section 6.4). As in the example, the oscillative behavior of FSPC
is only slightly aected by the entering mobile. It is also noteworthy, that the
amplitude of oscillations is signicantly less than when not using TDC. This is in
accordance with the describing functions analysis in Section 5.5. For comparison,
the performance, when using the poleplaced PIcontroller from Section 10.1, is
depicted in Figure 10.1c.
The simulations also exemplify the stability and convergence results in Chap
ter 7. Convergence of the DPC algorithm with TDC and the PI controller is proven
202 Network Simulations
10 5 0 5 10 15 20 25 30 35 40
8
10
12
14
16
g
10 5 0 5 10 15 20 25 30 35 40
8
10
12
14
16
g
10 5 0 5 10 15 20 25 30 35 40
8
10
12
14
16
g
a.
b.
c.
t [s]
Figure 10.1 The recovering ability of some central controllers when a new
mobile is allocated to the same channel. This is the same
situation as in Example 5.1. Considered controllers: a. DPC
with TDC, b. FSPC with TDC, c. PI.
in Theorems 7.12 and 7.20 respectively. Furthermore, FSPC with TDC converges
to [
t
i
i
(t)[ (n + 2) = 3, as stated by Theorem 7.14.
10.2.3 Conservativeness of Global Stability Requirements
To investigate Theorem 7.20 more carefully, we consider the integrating controller
of which the DPC is a special case (see Section 5.2.1).
p
i
(t + 1) = p
i
(t) +
i
(
t
i
(t)
i
(t)).
Chapter 5 discussed local stability requirements of this controller. When subject
to an additional time delay of one sample, it is considered locally stable if < 1.
Chapter 7 addressed global stability aspects, and the central stability result is
found in Theorem 7.20. Here we will compare three dierent parameter settings:
two corresponding to locally stable controllers and one based on a global stability
criterion. The objective is to illuminate the conservativeness of Theorem 7.20 (cf.
Section 7.3.5).
Reconsider the simulation conguration above, and make the power control
problem more challenging by aiming at SIR target 20 dB. This is a feasible target
SIR, since the corresponding feasibility margin is
m
= 2.53 > 1. The global
10.2 Simple Network Simulation Environment 203
stability result in Theorem 7.20 relates stability to the degree of crosscoupling
cc
.
In this case it is given by
cc
= max
1im
I
t
i
v
i
I
t
i
= 0.775.
Example 7.5 relates the degree of crosscoupling and to global stability, and the
stability requirements can be expressed as in Figure 7.10. From the example, we
conclude that = 0.46 provide a globally stable controller in this case. In addition,
we consider the rule of thumb provided in Section 7.3.5.
[G
ll
(e
iw
)[ <
m
.
which yield = 0.66. For comparison, we also consider = 0.9, which is locally
stable according to Section 5.3.2. As in the previous section, the entering mobile
starts by using 1 dBW. The recovering ability of each of the three controllers is
depicted in Figure 10.2. First, we again note that local stability does not imply
global stability since the controller with = 0.9 results in an unstable global
system. The controller designed to be globally stable indeed is stable with appealing
performance. However, even though global stability cannot be granted for = 0.66,
it does provide a stable global system in this case. On the other hand, this might
not be the case for every possible situation with the same degree of crosscoupling.
10.2.4 Abrupt Change of Target SIR
Abrupt changes appear naturally in the cellular radio systems. For example, the
entering mobile in the previous section, is observed as an abrupt increase of the
interference. A similar situation arises when a mobile station starts using a higher
data rate (i.e., aiming at a higher target SIR). Consider the situation in Exam
ple 7.4, where a more distant mobile is allocated a signicantly lower data rate due
to his less favorable propagation conditions. It is interesting to study the eect on
the system, when the user is upgraded from
t
i
= 10 dB to
t
i
= 12 dB, which is
the same as the others.
Mobility is modeled as a constant velocity (5 m/s) Manhattan random walk,
i.e., the direction of movement at each sample instant is any of the cardinal points
equally probable. Shadow fading is modeled as in Section 6.1.1, and the sample
interval is T
s
= 0.1 s. This is chosen to average out the eect of the fastest
shadow fading frequency components, to emphasize the eect of the upgraded SIR
target. All four mobiles are connected to the same base station in a cell of radius
1000 m, and the processing gain is PG=128. Power control commands are subject
to the additional delay n
p
= 1, n
m
= 0. Initially, the situation is analogous to
Example 7.4. Then the mobiles start to move, and the fth mobile is upgraded
after 10 s. Four algorithms are considered:
PI from Section 6.3
204 Network Simulations
5 0 5 10 15 20 25 30
20
0
20
40
5 0 5 10 15 20 25 30
10
20
30
5 0 5 10 15 20 25 30
10
20
30
a.
b.
c.
t [s]
i
(
t
)
[
d
B
]
i
(
t
)
[
d
B
]
i
(
t
)
[
d
B
]
Figure 10.2 The recovering ability, when a new mobile enters the system.
The plots correspond to the use of integrating controllers with
a. = 0.9, b. = 0.66 and c. = 0.46.
DPC with TDC
FSPC with stepsize 1 dB
FSPC with stepsize 1 dB and TDC
The resulting SIR:s are plotted in Figure 10.3. We note that DPC with TDC and
the PIcontroller adapt almost equally well. The FSPC algorithm adapts slower,
which is due to the rate limiting eect of the relay. This is also evident when
studying how the relative load
L
r
(see Denition 7.6) changes. As disclosed in
Example 7.4, it is relatively costly to upgrade the distant mobile. Figure 10.4
illustrates the relative load over time, when using DPC and FSPC with TDC.
The solid line describes the true relative load (based on the actual
i
:s), while the
dashed describes the target relative load (based on the
t
i
:s). Again, the limiting
eect of decision feedback on adaption rate is evident.
10.3 FHGSM Simulation Environment
In systems, such as GSM and DAMPS, there are numerous channels available in
the system. Therefore, dynamic channel allocation plays a more important role.
However, when adding frequency hopping, the critical eect of being allocated a
10.3 FHGSM Simulation Environment 205
0 5 10 15 20
30
20
10
0
10
20
30
0 5 10 15 20
30
20
10
0
10
20
30
0 5 10 15 20
30
20
10
0
10
20
30
0 5 10 15 20
30
20
10
0
10
20
30
a. b.
c. d.
t [s] t [s]
t [s] t [s]
i
(
t
)
[
d
B
]
i
(
t
)
[
d
B
]
i
(
t
)
[
d
B
]
i
(
t
)
[
d
B
]
Figure 10.3 The behavior of the system in Example 7.4, when the distant
mobile is upgraded from
t
i
= 10 dB to
t
i
= 12 dB. In
addition, the terminals are moving, resulting in slow shadow
fading signal powers. Controllers: a. PI, b. DPC with TDC,
c. FSPC, d. FSPC with TDC.
bad channel is averaged out, due to the hopping over several frequencies. In such
a system, coding and frequency hopping aim at mitigating the multipath fading,
while power control aims at compensating for the shadow fading to maintain a
satisfactory SIR level on a slower time scale.
The simulation environment considered here is FHGSM specic, but general
ize to describe the situation of a narrowbanded TDMA system. The simulation
environment is described in more detail in (Olofsson et al., 1997). Essentially, the
objective is to illuminate some characteristic eects in such a system. In FHGSM,
the measurements are available as measurement reports, and the output powers
are quantized and constrained from above and below. This is relaxed in some of
the simulations, where instead information feedback is assumed, with analog SIR
measurements and any power outputs.
Signal gains are modeled as discussed in Sections 3.3 and 6.1. Shadow fading is
described using the Gudmundson model, and the Clarkes model is used to repre
sent the fast fading correlation in both carrier frequency and spatial deection (cf.
Figure 3.2). Some of the more important simulation parameters are summarized
in Table 10.2. This yields a cell layout as in Figure 10.7.
The objective with these simulations is primarily to illustrate dynamical ef
206 Network Simulations
0 2 4 6 8 10 12 14 16 18 20
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16 18 20
0
0.2
0.4
0.6
0.8
1
t [s]
L
r
L
r
a.
b.
Figure 10.4 True (solid) and target (dashed) relative load over time, when
using a. DPC with TDC and b. FSPC with TDC, correspond
ing to the SIR plots in Figure 10.3b,d.
fects on a system level, rather than evaluating absolute performance or capacity
with respect to realistic scenarios. A system level perspective of the eects using
various controllers is provided by Section 10.3.1. Performance measures play an
important role when evaluating simulations. We will discuss it in terms of SIR and
FER as outlined in Section 10.3.2. The following section discusses the negative
eects of not handling output power constraints properly. Such constraints are al
ways present in real systems. Therefore, unstable control algorithms will not result
in powers increasing without bounds. Section 10.3.4 illustrates the performance
degradation when using a locally unstable controller such as DPC subject to time
delays. Finally, the behavior of the ML estimator from Chapter 8 and the quality
mapper from Section 9.3 are illustrated in the remaining sections. Models of dier
ent phenomenons are described in more detail in Appendix 10.B. All simulations
are based on the frequency reuse K = 3, except the comparative simulation of
using estimated or true values, in which the reuse K = 9 is employed. Numerous
plots are associated with the discussion in this section. In order to maintain the
continuity in the text, all plots are presented in Appendix 10.B.
10.3.1 System Level Effects of Controlling the Power
To study the system level eects of dierent power control algorithms, it is very in
structive to study scatter plots of the desired signal power (C) and the interference
(I) concurrently. Basically, we plot every CI pair (in dBm) for every user at every
sample instant. For comparison purposes, we consider four central control strate
gies, described in more detail below. The resulting plots are found in Figures 10.8
10.3 FHGSM Simulation Environment 207
Frequency band 900 MHz
No. of carrier freq. 27
Frequency reuse K = 3 or K = 9
Antennas Sectorized
Cell radius 1000 m
Cell layout 5 5 clusters of 9 cells,
employing wrap around
Adj. channel atten. 20 dB
Frequency hopping Pseudorandom
Pathloss exponent = 3.5
Shadow fading std.dev.
s
= 6 dB
Shadow fading corr. dist. d = 100 m
Rayl. fading, avg. gain 0 dB
Control sample interval T
s
= 0.48 s
Burst time 0.577 ms
Mean MS speed 50 km/h
Downlink power 13 dBW (constant)
Uplink MS: GSM class 4:
Uplink maximum power p
max
= 3 dBW
Uplink minimum power p
min
= 17 dBW
Uplink quantization 2 dB
Time delays:
Power output delay n
p
= 1
Measurement delay n
m
= 0
Table 10.2 FHGSM simulation parameters.
and 10.9.
1. Constant powers. When not controlling the powers at all, the twodimen
sional CIdistribution in Figure 10.8a, is approximately normal distributed.
This is natural, the desired signal is essentially described by the shadow fading
power gain, with normal distributed long term statistics (see Section 3.3.2).
Moreover, the normal distribution of the interference is further motivated in
Section 8.1.
2. Cbased integrating controller. Controllers aiming at a specic desired
signal power level are exemplied in Section 4.5. One natural example is
the CRP algorithm, which can be seen as an integrator in logarithmic scale.
Figure 10.8b illustrates how the controller compresses the distribution in the
Cdirection.
3. SIRbased integrating controller. A SIRbased integrating controller has
a similar eect on the systems, but it compresses the distribution in the CI
direction, see Figure 10.9a. By balancing the SIR:s, less power is needed to
208 Network Simulations
meet the requirements. This is also seen in the gure, since the thermal noise
(at 148 dBW) clearly is a major contribution of the interference. In this
case, all connections strive towards the same
t
.
4. AAW algorithm. The main idea is to favor good connections that need
relatively small powers to maintain an acceptable SIR level. As disclosed
above, the SIRbased controller strives towards the line C = I +
t
. AAW,
on the other hand strives towards the line C = I + A, which is indicated
by Figure 10.9b for = 0.7. This is easily seen by considering the converged
situation of the AAW algorithm (cf. Section 5.2.1)
p = p +(
c
)
Using C = p +g and = C I and simple manipulations yield
(C g)(1 ) = (
c
(C I))
C = I +g(1 ) +
c
. .. .
A
All the four strategies probably provide reasonable quality of service to the
users. When power economy is of no concern and the the network is sparsely
planned, using maximal powers is a plausible solution. The SIRbased Icontroller
reduces the power consumption, while maintaining an acceptable quality. However,
it is evident that this could be a low interference situation (see Section 9.4.5) to
some users. When the interference level is low, is could be desirable to use a higher
power. Conversely, the algorithm forces connections subject to high interference to
use high power to aim at higher carrier. This behavior is costly for the system.
The Cbased Icontroller is essentially neglecting high power requirements from
highinterference users, which is desirable from a systems perspective. The obvious
drawback is the algorithm essentially needs interference information, in order to
aim at relevant desired signal power levels.
AAW combine the nice features of Cbased and SIRbased integration. Thereby,
users with less favorable propagation conditions receive less priority than others.
This selfregulatory eect of target SIR adaption is rather unpredictable, which is
a drawback. The users that are the most costly to support are found in the upper
right area of the gures. They are subject to considerable interference and thus
requires a high desired signal power. Thus, from a system perspective it is desirable
to force the costly users to aim at lower target SIR:s, and shape the distribution
somewhat like a banana (Almgren, 2000).
In summary, we advocate a SIRbased Icontroller with the approach in Sec
tion 9.4.3, where the target SIR adaption is considered in an outer loop. As such,
the rate adaption approaches (see Section 9.4.2) incorporate nice features. Since
the rate is adapted to the individual propagation and interference situations, it
may enable all satised users with respect to the assigned data rate.
10.3 FHGSM Simulation Environment 209
0 2 4 6 8 10 12 14 16 18 20
10
0
10
20
30
0 2 4 6 8 10 12 14 16 18 20
10
5
10
4
10
3
10
2
10
1
10
0
a.
b.
t [s]
Figure 10.5 Two dierent quality measures. a) SIR is considered accept
able when above 10 dB. b) FER is acceptable when below
0.02.
10.3.2 Quality Measures
As pointed out in Section 9.2.1, speech quality is very subjective, and it is hard to
nd relevant objective measures. One common assumption is that the perceived
quality is dependant only on the SIR, and that is it acceptable if above a certain
threshold. Olofsson (1997) argues that this will not describe the quality satisfac
tory, and that FER is a more adequate measure.
These measures are depicted for a sample case in Figure 10.5, where the SIR
is considered acceptable when above 10 dB, and FER when below 0.02. The ques
tion is whether the user in Figure 10.5 is satised or not. A simple measure of
user dissatisfaction is the outage probability P(
i
) (Ahlin and Zander, 1998).
Consequently, 1P(
i
0
) relates to user satisfaction. A similar outage measure
can be dened in terms of FER. Therefore, cumulative distributions of SIR and
FER for all users collectively provide a relevant measure of the user satisfaction in
simulations. These will be used throughout the preceeding sections.
10.3.3 AntiReset Windup
In practical systems, the output power is constrained from above and below, which
aects the controller behavior. As pointed out in Section 6.6.1, the internal state of
the controller no longer matches the output. When a constraint becomes active, the
internal state winds up (