Upload
dinhthuan
View
223
Download
6
Embed Size (px)
Citation preview
Space-Time Coding Schemes for Wireless
Communications over Flat Fading Channels
A PhD Thesis Submitted to The Hong Kong University of Science and Technology
in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
in Electrical and Electronic Engineering
by
Meixia TAO
B.S., Fudan University, 1999
Department of Electrical and Electronic Engineering The Hong Kong University of Science & Technology
Clear Water Bay, Kowloon, Hong Kong
June 2003, Hong Kong
Authorization
I hereby declare that I am the sole author of the thesis.
I authorize the Hong Kong University of Science & Technology to lend this thesis
to other institutions or individuals for the purpose of scholarly research.
I further authorize the Hong Kong University of Science & Technology to
reproduce the thesis by photocopying or by other means, in total or in part, at the request
of other institutions or individuals for the purpose of scholarship research.
Meixia TAO
ii
Space-Time Coding Schemes for Wireless
Communications over Flat Fading Channels
by
Meixia TAO
This is to certify that I have examined the above PhD thesis
and have found that it is complete and satisfactory in all respects, and that any and all revisions required by
the thesis examination committee have been made.
Prof. Roger S. CHENG (Thesis Supervisor)
Prof. Shihe YANG (Committee Chairman)
Prof. Ross D. MURCH (Committee Member)
Prof. Wai Ho MOW (Committee Member)
Prof. Gary S. H. CHAN (Committee Member)
Prof. Khaled BEN LETAIEF (Acting Head of Department)
Department of Electrical and Electronic Engineering Hong Kong University of Science & Technology
June 2003
iii
Acknowledgements
I would like to express my sincere gratitude to my advisor, Prof. Roger S. Cheng, who is
an endless source of enthusiasm, ideas, and patience. It was him who led me into this
exciting area of wireless communications, and has offered me constant encouragement
and advice throughout the last four years. I hope I have learned from him not just his
broad knowledge, but his insights, inspiration, and his way of conducting research.
I especially thank Prof. Khaled Ben Letaief for introducing me to the academic world
before I started my Ph.D. research, and for giving me useful suggestions in my thesis
proposal. I am grateful to Prof. Ross Murch, Prof. Wai Ho Mow, Prof. Gary Chan
(Computer Science Department) and Prof. Xiang-Gen Xia (Chinese University of Hong
Kong) for being on my committee and providing me some constructive comments. I also
would like to thank Prof. Min Yan (Mathematics Department) and Prof. Kunrui Yu
(Mathematics Department) for helping me to solve some mathematical problems.
I gratefully acknowledge all my former and present colleagues in the wireless
research group for creating such a pleasant work environment and for having useful
discussions with me. Particular thanks go to Jason Leung (ASTRI), Daniel So, Zhiyu Xu
(UTStarcom), Yinjun Zhang, Defeng Huang, Xiaoli Chu, Tao Li, Ruly Choi, Sana Sfar,
Nejib Boubaker, and Peter Chan. It is really wonderful to work with them and talk with
them on many subjects.
I owe deepest appreciation to Fan Zhang for his support. He has always been with me
in the whole journey. Without him, this thesis would never have been completed.
Finally, I would like to thank my parents who always encourage and support me in all
of my decisions.
v
Contents
Acknowledgements v
Contents vi
Notations ii
Abbreviations iii
Abstract xv
Chapter 1 Introduction 1
1.1 Promises of Multiple Antennas 1
1.2 Problem Statement and Research Contribution 2
1.2.1 Enhanced Design of Space-Time Codes 3
1.2.2 Intensive Study on Generalized Layered ST Architecture 4
1.2.3 New Schemes for Non-Coherent ST Coding and Modulation 5
1.3 Outline of Thesis 6
1.4 Publication 7
Chapter 2 Preliminaries 9
2.1 System Model 9
2.1.1 Fading Model 9
2.1.2 Signal Model 10
2.1.3 MIMO Channel Model 12
2.1.4 Digital Transmission over MIMO Channels 13
2.2 Performance Measure 14
2.2.1 Channel Capacity 14
2.2.2 Error Probability and Pair-Wise Error Probability 16
2.2.3 Diversity 17
2.2.4 Coding Gain 17
2.3 Relevant MIMO Transmission Schemes 18
2.3.1 Rate-Oriented Layered Space-time Architecture (VBLAST) 18
2.3.2 Diversity-Oriented Space-Time Codes 19
2.3.3 Rate-Diversity-Oriented Space-Time Techniques 23
vi
2.3.4 Non-Coherent Diversity-Oriented (Differential) Unitary Space-
Time Modulation 23
Chapter 3 Improved design of Space-Time Codes at different SNR 25
3.1 System Model 26
3.2 Improved Design Criteria 26
3.2.1 Case 1: α ≈ 1 (Moderate SNR) 27
3.2.2 Case 2: α << 1 (Low SNR) 27
3.2.3 Case 3: α >> 1 (High SNR) 28
3.3 Computer Searched Trellis Codes for Moderate SNR 28
3.3.1 Code Examples 29
3.3.2 Simulation Results 32
3.4 Summary 33
Chapter 4 Diagonal Block Space-Time Coding 35
4.1 System Model and Performance Criteria 36
4.2 Diagonal Block Space-Time Codes 37
4.2.1 Code Structure 37
4.2.2 Performance Measure 38
4.2.3 Discussions on Diagonal Structure 41
4.3 (M, 1) Nonbinary Block Code Construction 42
4.3.1 Optimal Construction for Given Constellations 42
4.3.2 Linear Construction for PSK Modulation 46
4.3.3 Discussions 48
4.4 Simulation Results 49
4.4.1 Comparison with Delay Diversity Codes 49
4.4.2 Comparison with Other Existing Codes 53
4.5 Summary 55
Chapter 5 Generalized Layered Space-Time Architecture 56
5.1 System Model 57
5.1.1 Encoding 58
5.1.2 Decoding 59
5.2 Optimal Power Allocation 60
vii
5.3 Optimal Decoding Order 64
5.4 Interleaved GLST with Hard-Decision Iterative Decoding 68
5.5 Summary 70
Chapter 6 Differential Space-Time Block Codes 73
6.1 System Model 74
6.2 Differential Encoding 75
6.2.1 Data Matrix 75
6.2.2 Transmitted Matrix 76
6.3 Non-Coherent Decoding 76
6.3.1 Optimal Differential Decoder 77
6.3.2 Near Optimal Differential Decoder 78
6.3.3 Optimal DD versus Near Optimal DD 78
6.4 Simulation Results 79
6.5 Summary 84
Chapter 7 Trellis-Coded Differential Unitary Space-Time Modulation 86
7.1 Background on DUSTM 87
7.2 System Model and Performance Measure 88
7.2.1 System Model 88
7.2.2 Ideal Interleaver 89
7.2.3 No Interleaver 91
7.3 Code Construction 94
7.3.1 Ideal Interleaver 97
7.3.2 No Interleaver 98
7.4 Simulation Results 101
7.5 Extensions to Trellis-Coded Differential Space-Time Block Codes 104
7.6 Summary 106
Chapter 8 Conclusion and Future Work 107
8.1 Conclusion 107
8.2 Future work 109
Bibliography 111
viii
List of Figures
2.1 BPSK and QPSK constellations
2.2 16QAM and 32QAM constellations
2.3 System diagram of MIMO wireless communications.
2.4 QPSK 4-state ST code with 2 transmit antennas
3.1 Performance of the QPSK, 8-state, 2 bit/s/Hz space-time codes with 2 transmit
and 3,4 receive antennas.
3.2 Performance of the 8-PSK, 8-state, 3 bit/s/Hz space-time codes with 2 transmit
and 3, 4 receive antennas.
4.1 Transmitter diagram of the diagonal block space-time codes, “D” denotes one
symbol delay.
4.2 Trellis diagram for the DBST code with QPSK and M = 2.
4.3 SER performance of 8PSK codes with M = 3 transmit antennas over a quasi-static
fading channel.
4.4 SER performance of 8PSK codes with M = 3 transmit antennas over a rapid
fading channel.
4.5 SER performance of 8PSK codes with M = 3 transmit antennas over a time-
varying fading channel with . 05.0=sdTf
4.6 Histogram of the gains shown in Table 4.4.
4.7 FER performance of 8PSK codes with M = 3, 4 transmit antennas over a quasi-
static fading channel.
5.1 Encoder of (interleaved) GLST (a) main layout, (b) HGLST, and (c) DGLST.
5.2 Performance comparison of different power allocation in the (a) (4,4) and (b)
(8,8) GLST systems.
5.3 Performance comparison of optimal ordered decoding in the (a) (4,4) and (b) (8,8)
GLST and BLAST systems.
5.4 Iterative decoding of interleaved GLST (a) main block diagram, (b) sub-block
diagram for the “ST Dec-Enc j’” component.
5.5 Performance of iterative decoding in the (a) (4,4) and (b) (8,8) HGLST systems.
ix
5.6 Performance comparison of interleaved HGLST and interleaved DGLST with
iterative decoding in the (4,4) and (8,8) systems.
6.1 Performance of differential decoding and coherent decoding for G2 with 16QAM
data symbols at M = 2, N = 1, and R = 4 bit/s/Hz.
6.2 Performance of differential decoding and coherent decoding for H4 with 16QAM
data symbols at M = 4, N = 1, and R = 3 bit/s/Hz.
6.3 Performance of differential decoder for G2 with 16PSK and 16QAM data symbols
and cyclic group code at M = 2, N = 1 and R = 4 bit/s/Hz.
6.4 Performance of differential decoder for H4 with 16PSK and 16QAM data symbols
at M = 4, N = 1, and R = 3 bit/s/Hz.
6.5 Performance of differential decoder for H4 with 6PSK data symbols and cyclic
group code at M = 4, N = 1 and R ≈ 2 bit/s/Hz.
7.1 Transmission system diagram of trellis-coded differential unitary space-time
modulation.
7.2 Set-partition of G = ([1 3]; 8).
7.3 Set-partition of G = ([1 7]; 32).
7.4. Trellis diagrams of the codes designed for ideal interleaver case.
7.5 Trellis diagrams of the codes designed for no interleaver case: (a) rate 2/3 4-state
with G = ([1 3]; 8); (b) rate 4/5 16-state with G = ([1 7]; 32).
7.6 BER of rate 2/3 4-state and 8-state G = ([1 3]; 8) TC-DUSTM compared with
uncoded G = ([1 1]; 4) DUSTM, ML differential decoder and ideal interleaver, R
= 1 bit/s/Hz.
7.7 FER of rate 2/3 4-state G = ([1 3]; 8) TC-USTM compared with uncoded G = ([1
1]; 4) USTM, ML coherent decoder and no interleaver, frame length = 100, R = 1
bit/s/Hz.
7.8 FER of rate 2/3 4-state G = ([1 3]; 8) TC-DUSTM compared with uncoded G =
([1 1]; 4) DUSTM and rate 1/2 4-state TC-DPSK, suboptimal differential decoder
and no interleaver, frame length = 102, R = 1 bit/s/Hz.
x
List of Tables
3.1 Space-time codes with QPSK, 4 states, 2 bit/s/Hz.
3.2 Space-time codes with QPSK, 8 states, 2 bit/s/Hz.
3.3 Space-time codes with QPSK, 16 states, 2 bit/s/Hz.
3.4 Space-time codes with 8PSK, 8 states, 3 bit/s/Hz.
4.1 Optimum block codes used in DBST coding for P = 4 and 8 with PSK modulation
4.2 Optimum block codes used in DBST coding for P = 16 and 32 with PSK/QAM
modulation
4.3 Linear block ring codes used in DBST coding for P = 16, 32, and 64 with PSK
4.4 Operating SNR [dB] at SER = 2 for codes with M = 2 transmit antennas
over a rapid fading channel
410−×
7.1 Determinant distance profile of G = ([1 3]; 8)
7.2 Determinant distance profile of G = ([1 7]; 32)
7.3 Parameters of the codes designed for ideal interleaver case
xi
Notations
Throughout this work scalars are given by lowercase letters (a), vectors by boldface
lowercase letters (a), and matrices by boldface uppercase letters (A). Certain constants or
parameters are given by standard uppercase letters (A).
• 1−=j .
• P(C) is the probability of the event C.
• P(C|D) is the conditional probability of the event C knowing that the event D has
occurred.
• p(a) is the probability density function of the random variable a.
• E[a] is the expectation of the random variable a
• a* is the conjugate of the complex scalar a.
• AT is the transpose of A.
• AH is the complex conjugate transpose of A.
• 0T×N is the T × N zero matrix, the dimension may be dropped if there is no
confusion.
• IM is the M × M identity matrix, the dimension may be dropped if there is no
confusion.
• For a complex number a, )()( aaa jIR +=
• 22 )()( aaa IR += is the absolute value of the complex scalar a.
• ∑∑= =
=N
n
M
mnma
1 1
2,
2A is the squared Euclidean norm of the M × N matrix A with the (m,
n)th entry [A]m,n =am,n.
• rank(A) is the rank of the matrix A.
• det(A) is the determinant of the matrix A.
• tr(A) is the trace of the matrix A.
• diag{a1, a2, …, aM} is an M × M diagonal matrix with diagonal elements a1, a2, …,
aM.
xii
Abbreviations
3G Third Generation Mobile Telephony System
AWGN Additive White Gaussian Noise
BER Bit Error Rate
BLAST Bell-lab LAyered Space-Time architecture
CSI Channel State Information
DBST Diagonal Block Space-Time
DGLST Diagonal Generalized Layered Space-Time architecture
DPSK Differential Phase-Shift-Keying
DSTBC Differential Space-Time Block Codes
DUSTM Differential Unitary Space-Time Modulation
FER Frame Error Rate
GLST Generalized Layered Space-Time architecture
HGLST Horizontal Generalized Layered Space-Time architecture
i.i.d Independent and Identically Distributed
MAP Maximum A Posteriori Probability
MIMO Multiple-Input Multiple-Output
ML Maximum Likelihood
MLSE Maximum Likelihood Sequence Estimator
MMSE Minimum Mean Square Error
PAM Pulse Amplitude Modulation
PWEP Pair-Wise Error Probability
PSK Phase Shift Keying
QAM Quadrature Amplitude Modulation
RF Radio Frequency
Rx Receiver
SER Symbol Error Rate
SISO Single-Input Single-Output
SNR Signal to Noise Ratio
xiii
ST Space-Time
STBC Space-Time Block Code
STC Space-Time Coding
STTC Space-Time Trellis Code
TCM Trellis-Coded Modulation
TC-DPSK Trellis-Coded Differential Phase-Shift-Keying
TC-DUSTM Trellis-Coded Differential Unitary Space-Time Modulation
Tx Transmitter
USTM Unitary Space-Time Modulation
VBLAST Vertical Bell-lab LAyered Space-Time architecture
xiv
Space-Time Coding Schemes for Wireless
Communications over Flat Fading Channels
by
Meixia TAO
Department of Electrical and Electronic Engineering
The Hong Kong University of Science & Technology
Abstract
The increasing demand for higher data rates and higher quality in wireless
communications has motivated the use of multiple antenna elements at both the
transmitter and the receiver sides in a wireless link. The problem discussed in our
research is the development of fundamental space-time (ST) coding and modulation
methods to achieve the gains provided by multiple antennas, in terms of both improved
robustness of the link and a higher spectral efficiency. We focus on a point-to-point
wireless environment, in which the channel is modeled as flat fading, and channel
knowledge is not available at the transmitter. Several new and improved schemes tailored
for different applications are proposed.
We first consider the design of ST trellis codes that reduce the probability of error
without loss of spectral efficiency. It is found that the typical assumption of high signal-
to-noise ratio (SNR) and, consequently, traditional design criteria are invalid in certain
situations. Analyzing pair-wise error probability, we derive two sets of tighter design
criteria for low and moderate SNR regions, respectively. New ST trellis codes optimized
for moderate SNR are provided via a computer search. To avoid the prohibitively high
complexity of searching for good codes with a larger number of transmit antennas and
higher-level modulation, we introduce a novel systematic code construction method,
diagonal block ST coding. This two-step approach demonstrates promising results at the
commonly assumed high SNR.
xv
We then conduct an intensive study into generalized layered ST architecture that
allows a tradeoff between error probability and spectral efficiency. Our goal is to enhance
the tradeoff by further reducing the error probability. Techniques with no or little increase
in receiver complexity, such as optimal power allocation and optimal decoding order, are
introduced. A hard-decision iterative decoding algorithm that significantly enhances the
system performance is also proposed.
Finally, we consider the design of ST techniques that avoid channel estimation at the
receiver, but with minimal loss in error performance. A new differential ST modulation
scheme based on orthogonal ST block codes with square codeword matrices is
introduced. The important difference from previous differential ST modulation schemes
is the use of multiple amplitudes. This makes our scheme more power efficient. We then
introduce a joint design of channel coding and general differential ST modulation, called
trellis-coded differential unitary space-time modulation. Several examples that offer a
good tradeoff between the coding advantage and trellis complexity are presented.
xvi
Chapter 1
INTRODUCTION
In this chapter we introduce the motivation, provide the problem statement, present the
contribution of this work, and outline the organization of this thesis.
1.1 Promises of Multiple Antennas
Wireless is the fastest growing segment of the communications market in the world. It
has a wide range of services from satellites that provide low bit rates but global coverage
and cellular systems with continental coverage to high bit rate local area networks and
personal area networks with a maximum range of a few to a hundred meters. Using a
cellular system is by far the most common wireless method to access data or perform
voice dialing. In the near future, we will expect seamless global roaming across different
wireless networks and ubiquitous access to personalized applications and rich content via
a universal and user-friendly interface. Yet, in this climate, researchers still struggle with
the fundamental questions about the physical limitations of communicating over wireless
channel. These include multipath fading, limited spectrum resources, multiple-access
interference, and limited battery life of mobile devices.
In this thesis we consider the use of multiple antenna elements at both the transmitter
and the receiver ends to improve a wireless connection. The use of multiple antennas has
been a recent significant breakthrough in wireless technologies. It creates a multi-input
multi-output (MIMO) channel in which each path from one transmit antenna to one
receive antenna can be viewed as one signaling branch. MIMO systems have two major
attractive advantages that conventional single-input single-output (SISO) systems do not
have. These are:
1
Multiplexing gain (or spectral efficiency gain): As supported by information-theoretic
studies [1, 2, 51, 52], the channel capacity of a multiple-antenna system is considerably
higher than that of a single-antenna system. In particular, it is widely understood that
channel capacity increases asymptotically linearly with the minimum number of transmit
and receive antennas when channel knowledge is available at the receiver. Therefore, the
degree of freedom for communications is increased. As a result, the transmission rate
increases linearly without an increase in the total transmission power or channel
bandwidth.
Diversity gain: If the antennas at both ends have no, or very low, correlation, the
signaling branches between different transmit-receive antenna pairs in a MIMO system
can be assumed to be statistically independent. These independent branches create
diversity gain. By transmitting the same data (in the same, or different, representations)
over multiple independent branches, fading can be effectively mitigated and, hence, link
reliability significantly improved.
MIMO systems also provide other types of gains such as array gain and interference
suppression gain. Consequently, multiple antennas are expected to play an important role
in advanced wireless systems, for example, 3G and beyond.
1.2 Problem Statement and Research Contribution
The problem discussed in our research is how to develop fundamental transmission
strategies adapted to a point-to-point wireless link with flat fading channels to utilize the
promises of multiple antennas jointly or individually. This topic has, in fact, received
much attention in the past few years. As the core idea is complementing the traditional
time dimension with the space dimension inherently brought by multiple antennas,
MIMO-related transmission strategies are often referred to as space-time (ST) techniques.
In this thesis, we focus on developing ST coding and modulation schemes that do not
require channel knowledge, i.e., channel state information (CSI), at the transmitter. Both
cases in which CSI is available (coherent) and unavailable (non-coherent) at the receiver,
respectively, are considered.
2
1.2.1 Enhanced Design of Space-Time Codes
Our first goal is the design of space-time codes that fully utilize the diversity advantage to
improve the error probability behavior. The concept of space-time coding was introduced
by Tarokh et al [8]. This family of code design performs coding across both time and
space (transmit antennas) dimensions. It works with multiple transmit antennas and does
not necessarily need multiple receive antennas. One of the fundamental difficulties of
space-time codes, a fact which has made its design challenging, is that the design criteria
apply to the complex domain of the baseband modulated signals rather than to the binary
or discrete domain in which the underlying codes are traditionally designed. Current
space-time codes include space-time trellis codes (STTC) [8] and space-time block codes
(STBC) [9, 11, 12, 14, 57]. Carefully designed STTC can achieve maximum antenna
diversity gain as well as a certain amount of coding gain, while, in STBC, only diversity
gain, not coding gain, can be achieved. The focus of previous work on the design of
STTC was either on finding good codes through a global search [14, 15, 16] or on
proposing new code constructions [17, 21, 22, 32, 33].
There are two problems associated with previous work. First, most of the codes were
designed using the traditional rank and determinant criteria [7, 8], and these criteria are
valid under the assumption of high signal-to-noise ratio (SNR). It is observed, however,
that, in space-time coded systems, the SNR needed to satisfy a particular system
specification depends heavily on the number of antennas, especially the number of
receive antennas. This renders the high-SNR assumption and, consequently, the
traditional criteria invalid in certain situations. In this thesis, analyzing the pair-wise error
probability, we propose two sets of new design criteria for low and moderate operating
SNR regions, respectively. One of our results is that the traditional full-rank criterion,
quantifying the order of transmit antenna diversity, is not always necessary for good
space-time codes. In addition, the minimum Euclidean distance is a good performance
measure at low SNR. Several new STTCs using two transmit antennas and optimized for
moderate SNR are found via a computer search. Simulation results demonstrate that they
outperform existing codes based on traditional criteria over a wide SNR range.
3
The second problem is that the computer-searched codes in [25, 26, 27, 28] are only
available for a small number of transmit antennas (two or three) with low-level
modulation (BPSK or QPSK) due to the time complexity of searching, and that the code
constructions in [34, 35, 46, 47] are not always efficient. Thus, we propose a novel two-
step code construction, i.e., a one-dimensional block code in conjunction with diagonal
transmission pattern in space-time grid. This scheme is referred to as diagonal block
space-time coding. It is highly systematic and suitable for an arbitrary number of transmit
antennas with any signal constellation. It is also more efficient than previous systematic
code constructions.
1.2.2 Intensive Study on Generalized Layered ST Architecture
Our second goal in this work is the study of generalized layered ST architecture (GLST)
that achieves both diversity gain and multiplexing mentioned in Section 1.1, and ensures
a balance between the two. The framework of this architecture is to partition all the
available transmit antennas into groups and apply an individual space-time encoder for
each group independently. Obviously, by varying the grouping methods, a flexible
tradeoff between diversity order and multiplexing order, or equivalently a tradeoff
between error probability and spectral efficiency, can be ensured. Previous examples of
this architecture can be found in [19] and [40], where the employed component space-
time encoders are STTC and STBC, respectively. Only the basic transmission and
detection methods, but no advances on this topic, were considered.
In this work we generalize both of [19] and [40], and study several key aspects
embedded in the layering architecture to enhance the tradeoff in terms of the further
reduction in the probability of error. We first construct different GLST systems according
to different mappings from ST coded symbols to antenna groups. We then propose two
approaches, namely, optimal power allocation (no CSI at the transmitter) and optimal
decoding order, to improve the system performance with no or little increase in receiver
complexity. Finally, a computationally efficient hard-decision iterative decoding
algorithm is proposed. This algorithm can efficiently utilize full receive antenna diversity
and, hence, dramatically enhance performance of the overall system.
4
1.2.3 New Schemes for Non-Coherent ST Coding and Modulation
The previous two subjects we considered require channel estimation at the receiver to
identify CSI before detection/decoding. Our goal in this part is the design of non-coherent
ST techniques that prevent channel estimation, but with minimal loss in error
performance. Non-coherent schemes are desirable in a mobile environment where
precisely tracking the channel variation is difficult, especially when there are a large
number of antenna elements used. Previous techniques include unitary space-time
modulation [53, 54], differential unitary space-time modulation (DUSTM) [55, 56, 57,
62], and differential schemes based on Alamouti’s space-time block codes (DSTBC) [58,
63]. All these schemes are designed to avoid estimation of the channel but still enjoy
maximum transmit and receive antenna diversity. These schemes are usually known as
non-coherent modulation methods in the space-time dimension, rather than coding
schemes.
We first propose a new differential ST modulation scheme based on orthogonal ST
block codes [9, 11] with square codeword matrices, namely, differential space-time block
codes (DSTBC). Compared with DUSTM, our proposed scheme does not necessarily
have unique amplitude. As a consequence, the spectral efficiency is increased by carrying
information not only on orientations but also on amplitudes. Compared with [58], our
scheme is different in two ways. First, the restriction of PSK constellation on information
symbols is relaxed so that more power efficient constellations, such as QAM, can be
applied with little increase of complexity. Second, this differential scheme is not only for
the Alamouti’s code with two transmit antennas [11], but also for orthogonal codes with
an arbitrary number of transmit antennas as long as the codeword matrices are square [9,
12, 15].
To further enhance the performance of differential space-time modulation, we
introduce a joint design of channel coding and general differential ST modulation, called
trellis-coded differential unitary space-time modulation (TC-DUSTM). This is a
combined trellis coding and space-time modulation scheme, similar to the conventional
trellis-coded modulation (TCM) in single-antenna systems. The advantage of this
5
combination is that carefully designed trellis codes can increase the minimum distance of
DUSTM. It results in coding gain and, possibly, time diversity gain if an interleaver is
applied. We thoroughly study the performance measures and trellis code design rules for
systems with either an ideal interleaver or no interleaver. Several examples that offer a
good tradeoff between the coding advantage and trellis complexity are presented.
Extensions to trellis-coded differential orthogonal space-time block codes are also
discussed. Therein, we show that the inherent orthogonality allows the simplification of
the trellis encoding and decoding, and that the conventional well-developed one-
dimensional TCM can be directly applied.
1.3 Outline of Thesis
Following the introduction in this chapter, we review in Chapter 2 some background
knowledge, including channel model and performance measure, and then present the state
of the art on space-time techniques.
In Chapter 3 and Chapter 4 we present the first part of our research. Chapter 3 starts
with the derivation of improved design criteria in Section 3.2, followed by the computer-
searched codes and their simulation results in Section 3.3. Chapter 4 presents the system
code construction, diagonal block space-time coding. The code structure and its
performance criteria are described in Section 4.2. The design of the 1-D block code is
presented in Section 4.3, along with some code examples. In Section 4.4 the performance
of the proposed codes is evaluated and compared with that of existing codes.
The study on GLST is provided in Chapter 5. In Section 5.1, we present the basic
framework. The optimal power allocation and optimal decoding order are derived in
Section 5.2 and 5.3, respectively. The iterative decoding is proposed in Section 5.4.
In Chapter 6 and Chapter 7 we present the third part of our research. The differential
orthogonal STBC is provided in Chapter 6. The differential encoding and non-coherent
decoding are described in Section 6.2 and 6.3, respectively. Some simulation results are
shown and compared with that of other schemes in Section 6.4. In Chapter 7 we present
the results on TC-DUSTM. In Section 7.1 the system model of TC-DUSTM is
introduced, along with the differentially non-coherent decision metrics and the trellis
6
code design criteria. Section 7.2 describes code construction, as well as some code
examples. Some simulation results are shown in Section 7.3. Extensions to the trellis-
coded differential STBC are discussed in Section 7.4.
Finally, we provide some concluding remarks and suggestions for future research in
Chapter 8.
1.4 Publications
Journal Papers
[Tao1] Meixia Tao and Roger S. Cheng, "Trellis-coded differential unitary space-time
modulation over flat fading channels", IEEE Trans. on Communications, vol.
51, no. 4, pp. 587-596, April 2003.
[Tao2] Meixia Tao and Roger S. Cheng, "Improved design criteria and new trellis
codes for space-time coded modulation in slow flat fading channels", IEEE
Communications Letters, vol. 5, no. 7, pp. 313-315, July 2001.
[Tao3] Meixia Tao and Roger S. Cheng, “Diagonal block space-time code design for
diversity and coding advantage over flat Rayleigh fading channels”, accepted in
IEEE Trans. on Signal Processing.
[Tao4] Meixia Tao and Roger S. Cheng, "Generalized layered space-time codes for
high data rate wireless communications", accepted in IEEE Trans. on Wireless
Communications.
Conference Papers
[Tao5] Meixia Tao and Roger S. Cheng, “Space code design in delay diversity
transmission for PSK modulation”, in Proc. IEEE Vehicular Technology
Conference (VTC) 2002-Fall, pp. 444-448, Vancouver, Canada, Sept. 2002.
[Tao6] Meixia Tao and Roger S. Cheng, “Trellis coded differential unitary space-time
modulation in slow flat fading channels with interleaver”, in Proc. IEEE
Wireless Communications and Networking Conference (WCNC) 2002, pp. 285-
290, Florida, USA, Mar. 2002.
7
[Tao7] Meixia Tao and Roger S. Cheng, “Differential space-time block codes”, in
Proc. IEEE Global Telecommunications Conference (GLOBECOM) 2001, pp.
1098-1102, Texas, USA, Nov. 2001.
[Tao8] Meixia Tao and Roger S. Cheng, “Low complexity post-ordered iterative
decoding for generalized layered space-time coding systems”, in Proc. IEEE
International Conference on Communications (ICC) 2001, pp. 1137-1141,
Helsinki, Finland, June 2001.
[Tao9] H. C. J. Leung, Meixia Tao, and Roger S. Cheng, “Optimal power allocation
scheme on generalized layered space-time coding systems”, in Proc. IEEE
International Conference on Communications (ICC) 2001, pp. 1706-1710,
Helsinki, Finland, June 2001.
8
Chapter 2
PRELIMINARIES In this chapter we present some background knowledge of wireless communications with
multiple antennas, including channel model, performance measures, and previous
transmission schemes. We introduce the signal and channel model adopted throughout
this thesis in Section 2.1. In Section 2.2 we present several performance measures over
MIMO channels. Then, in Section 2.3 we provide the state of the art on relevant space-
time techniques.
2.1 System Model
2.1.1 Fading Model
What wireless communication essentially means is the propagation of information-
bearing electromagnetic waves transmitted and received from some kind of antenna
without any physical wave-guides. Therefore, it is subject to thermal noise, interference
from other wireless systems, propagation loss (large-scale), and self-interference (small-
scale). The last effect, the most distinct difficulty of wireless system design, originates
from buildings, trees, cars and other objects surrounding the transmitter and receiver. The
result is multiple paths of the same signal arriving at different times. These arrivals
combine constructively or destructively and, hence, introduce randomness, called
multipath fading, or simply fading. This is the major and the most challenging problem
that needs to be combated in wireless communications.
Many physical factors in the radio propagation channel influence fading. These
include the presence of reflecting objects and scatters, the relative motion between the
transmitter and the receiver, the movement of surrounding objects, and the transmission
9
bandwidth of signals. Depending on these factors, different transmitted signals undergo
different types of fading. One type of such fading is flat fading. It applies by definition to
systems where the bandwidth of the transmitted signal is much less than the coherent
bandwidth of the channel. All the frequency component of the transmitted signal
undergoes the same attenuation and phase shift when propagating through the channel. In
the time domain, flat fading corresponds on a channel delay spread which is much less
than the symbol time, hence the channel induces no inter-symbol-interference (ISI). This
type of fading is, historically, the most common type of fading described in the technical
literature and is assumed throughout this thesis. When the fading is non-flat, i.e.,
frequency-selective, most of the general results still apply when the receiver compensates
with equalization techniques. Throughout this thesis, slow fading is also assumed. That is,
the channel impulse response changes at a rate much slower than the transmitted
baseband signal. In this case, we further assume that fading is quasi-static, i.e., constant
during a (long) burst or transmission frame and then changes in an independent manner.
Another slow fading model is block fading. It applies when several adjacent symbols (a
block) are subject to the same fading value. We also consider time-varying fading which
follows a certain autocorrelation function. A detailed description on the types of fading
can be found in [77].
The statistical nature of fading can be modeled with various distributions. In this
thesis, we employ the commonly used Rayleigh distribution. That is, the amplitude of the
channel gain (or channel coefficient) follows a Rayleigh distribution while its phase is
uniformly distributed from 0 to 2π. This is valid when there are a large number of scatters
and no direct line of sight between the transmitter and receiver, and it accurately models
many indoor or urban environments. We discuss the explicit form of the channel gain
with Rayleigh distribution in Section 2.1.3.
2.1.2 Signal Model
In the transmission of digital information over a communication channel, the modulator is
the interface device that maps the digital information into analog waveforms that match
the characteristics of the channel. The mapping is generally performed by taking blocks
10
of k = log2L binary digits at a time from the information sequence and selecting one of L
= 2k deterministic, finite, energy waveforms for transmission over the channel. When
analyzing communication systems, it is often unnecessary to model the up- and down-
conversion between the baseband and the carrier frequency. So, one can choose to work
with baseband models, or equivalent low pass signals and channels, which then become
complex valued. Throughout this thesis, we use complex baseband representation of
signals. The following modulations are considered.
Phase shift keying (PSK): the baseband representation of PSK signals is
PSK = { }1,,1,0,/2 −= Lke Lkj Kπ
The special case when L = 2 is the binary phase shift keying (BPSK). Signal space
diagrams, or signal constellations, for L = 2 and 4, are shown in Fig. 2.1.
0 (0) 1 (1)
0 (00)
1 (01)
2 (11)
3 (10)
Fig. 2.1: BPSK and QPSK constellations
Pulse amplitude modulation (PAM): the baseband representation of this
modulation is
PAM = { } )1(,,3,1 −±±± LK
Quadrature amplitude modulation (QAM): the baseband representation of QAM
signals is
QAM = {a + jb ; a, b∈ {±1, ±3, ± ( − 1)}} 2/1L
Examples of signal constellations are shown in Fig. 2.2 for L = 16 and 32.
Representation of other modulation schemes can be found in [78].
11
12
8
4
0
13
9
5
1
14
10
6
2
15
11
7
3
23
17
11
5
24
18
12
6
25
19
13
7
26
20
14
8
28 29 30 31
0 1 2 3
16
10
4
21
15
9
22 27
Fig. 2.2: 16QAM and 32QAM constellations
2.1.3 MIMO Channel Model
A multi-antenna system is simply an arbitrary wireless communication system in which
the transmitter side as well as the receiver side are equipped with multiple antenna
elements. Fig. 2.3 illustrates the simplified baseband system diagram of a point-to-point
wireless communication link with multiple antennas.
coding
modulation
weighting/mapping
010011
weighting/demapping
demodulation
decoding
010011
N RxM Tx Fig. 2.3: System diagram of MIMO wireless communications.
As can be seen, the underlying nature of using multiple antennas compared with
traditional single-antenna systems is to create a MIMO channel, in which each path from
one transmit (Tx) antenna to one receive (Rx) antenna can be viewed as one signaling
branch. By collecting all of the branches, the MIMO channel can be fully described using
an N × M matrix H, where N is the number of Rx antennas, M is the number of Tx
antennas, and the (n, m)th entry hn,m denotes the channel coefficient from transmit
antenna m to receive antenna n. In general, the correlation between all of the entries in the
12
channel matrix depends on the physical separation of the antenna elements at both sides,
the antenna polarization patterns, the wavelength and the location of surrounding scatters.
In a rich-scattering environment with devices capable of providing enough space to
allocate multiple antennas without coupling, it is usually valid to assume that they are
independent and identically distributed (i.i.d). Hence, based on the considered flat
Rayleigh fading model, the entries can be modeled as samples of independent complex
Gaussian random processes with zero mean and unit variance:
=mnh , Normal (0, 0.5) + ⋅j Normal (0, 0.5).
Consequently, is a chi-square random variable with degree of 2, , but
normalized to = 1.
2, || mnh
|[| 2,mnhE
22χ
]
Let ct denote the M × 1 baseband transmitted signal vector at discrete time slot t, and
rt be the N × 1 baseband received signal vector. As the signal at each Rx antenna is
simply a noisy superposition of the M transmitted signals corrupted by different fading,
the input-output relationship of the MIMO channel is modeled compactly as
ttt wHcr += (2.1)
where wt is a vector of additive white Gaussian noise (AWGN) terms.
2.1.4 Digital Transmission over MIMO Channels
A basic transmission and detection procedure over MIMO channels is described as
follows. Consider the wireless communication link shown in Fig. 2.3. A compressed
digital source in the form of a binary data stream is fed to a transmitting block
encompassing the functions of error control coding and (possibly joined with) mapping to
complex modulation symbols (QAM, PSK, etc.). The latter produces several separate
symbol streams which range from independent, to partially redundant, to fully redundant.
Each is then mapped onto one of the multiple transmit antennas. Mapping may include
linear spatial weighting of the antenna elements or linear space-time precoding. After
upward frequency conversion, filtering and amplification, the signals are launched into
the wireless channel. At the receiver, the signals are captured by multiple antennas and
demodulation and demapping operations are performed to recover the message.
13
In general, the design of above channel coding, modulation and mapping is very
different from that of traditional schemes used in SISO channels. This is mainly due to
the presence of a new signaling dimension: space, inherently brought by multiple
antennas, especially multiple transmit antennas. Hence, the MIMO-related transmission
techniques can be referred to as space-time techniques. The selection of detailed
techniques varies a lot depending on whether one or both sides have knowledge of fading
coefficients, i.e. CSI. Typically, it is very difficult to obtain channel knowledge at the
transmitter side and, hence, transmitter knowledge is not discussed in this thesis. In fact,
receive knowledge is a fairly common assumption and is possible through a training
sequence and/or a separate pilot channel. In mobile environment where the channel
changes rapidly, however, precisely tracking the channel variation is difficult. Moreover,
while a large number of transmit antennas increase the training period, which
significantly reduces the system efficiency, a large number of receive antennas increase
the complexity of channel estimation. Therefore, in our research, both the coherent and
non-coherent cases in which CSI is available and unavailable, respectively, at the receiver
are considered. We review existing transmission schemes in Section 2.3
2.2 Performance Measure
2.2.1 Channel Capacity
Radio spectrum is a very scarce and expensive resource in wireless communications. The
goal has always been to try to transmit as much information as possible over a given
limited spectrum. Channel capacity is the information-theoretic measure of maximum
possible information transfer rate per unit bandwidth (in bit/s/Hz) with reliable
transmission over a channel, subject to specified constraints.
The very famous capacity formula of a MIMO channel given that the channel matrix
H is known to the receiver can be expressed as [1, 2]
+= H
N MC HHI ρ
2log (2.2)
14
where H is the channel matrix and ρ is the total transmitted SNR. This capacity is
achieved when the transmitted signal vector in (2.1) is composed of M statistically
independent equal power components each with a zero-mean Gaussian distribution. The
capacity in (2.2) is, in fact, a random variable because of the randomness of the channel
matrix H. One way to characterize it is to use the average (or ergodic) capacity which is
obtained by taking the average over all H. Let K = min{M, N} and K’ = max{M, N}, then
a lower bound of the average capacity at high SNR can be derived as [1]
[∑+−=
+≥'
1'
2222 loglog
K
KKiicoherent E
MKC χρ ] , (2.3)
where is a chi-square random variable with a degree of 2i. This lower bound is
asymptotically tight at high SNR. It is observed that at high SNR, a 3-dB increase in ρ
yields a K-bit/s/Hz increase on the capacity, in contrast to a 1-bit/s/Hz increase for
traditional single-input single-output (SISO) channels. This result suggests that the
MIMO channel can be viewed as K parallel spatial channels, and that K = min{M, N} is
the total number of degrees of freedom to communicate. Therefore, independent
information symbols can be transmitted in parallel through the spatial channels to
increase the spectral efficiency. This is also called spatial multiplexing in [72].
22iχ
The study of the channel capacity when H is unknown at the receiver is a little bit
more complicated. It was initiated by Marzetta and Hochwald in [51] for block fading
channels with a block length equal to T discrete time intervals. It is shown that for any
fixed N Rx antennas, the capacity obtained with M > T transmit antennas is the same as
the capacity obtained with M = T transmit antennas. Thus, in what follows M ≤ T is
always assumed. Second, the capacity-achieving random signal matrix may be
constructed as a product C = VΦ, where Φ is an M × T istropically distributed matrix
with orthonormal rows and V is an independent M × M real non-negative diagonal
matrix. Notice that this capacity-achieving input distribution is very different from the
Gaussian distribution in the coherent case. The non-coherent channel capacity was further
analyzed by Zheng and Tse in [52] using a geometric approach. It was shown that the
average channel capacity at high SNR is asymptotically equal to
15
aT
MMC coherentnon +
−=− ρ2
** log1 (2.4)
where M* = min{M, N, T/2} and a is some constant independent of ρ. In particular, if M
< N and T ≥ M + N, the degree of freedom for non-coherent communication without
imposing a training sequence is only . ( )TMM /1−
2.2.2 Error Probability and Pair-Wise Error Probability
Although channel capacity can somehow motivate the design of transmitted signals (e.g.
[53] as well as receiver structures (e.g. [1, 3]), the drawback is that it usually does not
reflect the performance achieved by actual transmission systems, since it only provides an
upper bound realized by codes with boundless complexity or latency. In practice, the
probability of error is used to measure the reliability or robustness of a communication
strategy.
Pair-wise error probability (PWEP) between any two possible codewords is defined
as the probability that a certain receiver makes an error in favor of one codeword when
the other is actually transmitted. In fading channels, PWEP reveals the effect of some
dominant factors (e.g., diversity gain and coding gain, see Section 2.2.4) in coded system
performance. And, it usually can be calculated with a closed-form tight upper bound.
Hence, in MIMO systems, the worse case PWEP over all possible distinct codeword pairs
is often used to specify the design criteria for constructing proper coding and/or
modulation.
With PWEP, the bit error probability (BER), symbol error probability (SER), or
frame error probability (FER) can be consequently derived. In general, however, it is
difficult to give an explicit expression for these error probabilities in a system.
Alternatively, computer simulation is usually used. The error probability is obtained by
counting the error numbers with respect to the total number of transmission realizations.
The accuracy of simulation results depends on the number of realizations. Usually, 100
error numbers are required to obtain ± 0.1 accurate results.
16
2.2.3 Diversity
Diversity is a powerful technique in wireless communications used to mitigate random
fading at relatively low cost. The basic idea of diversity is to exploit the random nature of
radio propagation by finding independent (or at least highly uncorrelated) signaling
branches for communication. If one channel branch undergoes severe attenuation, another
independent branch may have a strong signal. A key concept is the diversity order, which
is defined by the number of independent channel branches. The probability of losing the
signal vanishes exponentially with the diversity order. Hence, diversity order is one of the
parameters used to evaluate system error performance in fading channel. A simple
example of diversity technique is channel coding coupled with interleaving.
In a MIMO system with M Tx and N Rx antennas, there are MN independent spatial
channels, through which replicas of the same information data can be transmitted.
Intuitively, the maximum possible antenna diversity order is up to MN.
Of the antenna diversity techniques, receive antenna diversity is already well
developed. It can be achieved using selection diversity (selecting the antenna with the
highest signal power) or maximum-ratio-combing (weighting and combing the received
signals to maximize the SNR) when CSI is available at the receiver. It is, however, not
easy to achieve transmit diversity when the transmitter cannot access CSI. In general,
channel coding (either trellis-based or block-based) should be applied across both the
time domain and space (Tx antennas) domain so as to achieve transmit diversity. This is
the general concept of the so-called space-time coding [8].
2.2.4 Coding Gain
Beside diversity, coding gain is another useful parameter used to measure the error
performance of a system over fading channels. Originally it was used in AWGN
channels as the ratio of the minimum free Euclidean distance of a coded system to the
minimum Euclidean distance of an uncoded system. This value asymptotically reflects
the SNR reduction of a coded system over an uncoded system for achieving the same
17
amount of error probability. This term is now also applied in coded MIMO systems with
fading channels, but may not refer to Euclidean distance.
In [8] coding gain/advantage is defined from the pair-wise error probability as
follows. If the PWEP is upper bound as dSNRPWEP −⋅≤ δ
in the region of high SNR, then the system is said to have diversity advantage of d and
coding advantage of δ. As can be seen, the coding gain in MIMO systems shifts the upper
bound of the PWEP up or down and is the approximate measure of the gain over an
uncoded system operating with the same diversity order. Later on it will be shown that, in
a space-time coded system with diversity advantage of d = MN, the coding gain refers to
the minimum determinant of the Hermitian square of codeword error matrix [8].
2.3 Relevant MIMO Transmission Schemes
In this section we present an overview of relevant MIMO transmission schemes. As can
been seen from Section 2.2, MIMO systems can provide spatial multiplexing gain and
spatial diversity gain. These two types of gains reflect the spectral efficiency of a wireless
link and its reliability, respectively, and they are often mutually conflicting. Most of the
current ST techniques can be, therefore, classified into three categories: rate-oriented,
diversity-oriented, and rate-diversity-oriented.
2.3.1 Rate-Oriented Layered Space-Time Architecture (VBLAST)
The goal of rate-oriented schemes is to increase the spectral efficiency by sending
multiple independent data streams simultaneously on the same frequency band. The
number of independent data streams, or the order of the multiplexing gain, is equal to the
minimum number of transmit and receive antennas, which agrees with the capacity
behavior in (2.3). These schemes usually work when the transmitter and the receiver are
both equipped with multiple antennas.
Among these schemes, layered space-time architecture (LST) is innovative and
practical scheme. It was developed by Bell-Lab [3, 4, 5], and, thus, also known as
18
VBLAST (Vertical Bell-lab LAyered Space-Time architecture). In this scheme, each Tx
antenna simply sends an independent data stream (also called layer) simultaneously with
all of the other antennas on the same frequency band. The detection strategy at the
receiver is somehow motivated by the information-theoretic result in (2.3). Each layer is
regarded as one virtual user and detected successively with certain ordering. To be
specific, when detecting one layer, all of the other not-yet-detected layers are treated as
interference and nulled out based on zero-forcing (ZF) or minimum mean square error
(MMSE) criterion. After this layer is detected, its contribution is subtracted and the
detection jumps to the next layer. What makes the VBLAST scheme famous is, in fact,
the ordering method, the basic idea of which is to select the strongest layer from all the
not-yet-detected candidates to detect at each detection level. It has been proved that this
ordering method can maximize the minimum post-detection SNR over all layers and,
hence, minimize the overall error probability [6]. Obviously, the spectral efficiency in
this scheme increases linearly with the number of Tx antennas, M. However, it should be
noticed that the number of Rx antennas should not be less than that of Tx antennas, N ≥
M, in order to apply the layer-by-layer detection algorithm. In addition, the overall
system performance is dominated by the layer with the worst error probability and, hence,
is usually very poor when N is not large enough, relative to M.
The complexity of VBLAST detection algorithm can be further reduced by a square-
root algorithm [42]. And its performance can be enhanced through the use of outer
channel coding [41] or better decoding algorithms, such as maximum-likelihood (ML)
decoder [377], sphere decoder [67], and iterative algorithm [38, 39, 40]. Notice that the
use of outer channel coding also reduces the spectral efficiency.
2.3.2 Diversity-Oriented Space-Time Codes
Space-time coding (STC), also called space-time coded modulation, is a revolutionary
diversity-oriented development that aims at improving the error probability behavior by
performing coding across time and space (transmit antennas). It works with multiple
transmit antennas and does not necessarily need multiple receive antennas. Delay
diversity transmission, proposed in [43, 44], is probably the first STC scheme. It
19
transmits delayed copies of the same information signal sequence on multiple antennas,
and is seen at the receiver as a single-antenna transmission with increased channel delay
spread. The spatial diversity is, thus, artificially transformed to multipath diversity where
the gain can be realized at the receiver using the Viterbi-algorithm based maximum
likelihood sequence estimator (MLSE) [45]. This transmission scheme can be designed
for an arbitrary number of transmit antennas with arbitrary signal constellations. From a
coding perspective, it can be viewed as a systematic approach of designing ST codes and
is, hence, referred to as the delay diversity (DD) coding. The more general concept of
STC was later introduced by Tarokh et al [8].
Based on code structure, space-time codes can be divided into space-time trellis codes
and space-time block codes (STBC). STTC can be fully specified using a trellis diagram.
At each time t, depending on the state of the ST encoder and the input bits, a transition
branch is chosen. If the label of this branch is , then transmit antenna m is used
to send constellation symbols , m = 1, 2, …, M and all these transmissions are
simultaneous. Presented in Fig. 2.4 is an example of the trellis diagram for the 4-state
QPSK-modulated code using two transmit antennas [8, Fig. 4]. The mapping of the
QPSK constellation symbols is given in Fig. 2.1. At the receiver, a vector-based Viterbi
algorithm is applied to perform ML decoding.
Mttt xxx L21
mtx
30, 31, 32, 33
20, 21, 22, 23
10, 11, 12, 13
00, 01, 02, 03
Fig. 2.4: QPSK 4-state ST code with 2 transmit antennas
Under different channel environments, different criteria are needed for the design of
good space-time codes. In particular, the well-known and fundamental design criteria
over quasi-static flat Rayleigh fading channels are the rank criterion and the determinant
criterion [8, 7]. The rank criterion is used for achieving maximum diversity gain, while
the determinant criterion is for maximizing the coding gain. These two criteria have been
20
widely used to construct many classes of space-time trellis codes. In STTC, to achieve
full transmit diversity, the minimum required number of trellis states grows exponentially
with the number of transmit antennas and the transmission efficiency (in bit/s/Hz). The
result is an exponential increase in the decoding complexity. All the handcrafted space-
time trellis codes using two transmit antennas provided in [8] are full rank, thus attaining
maximum diversity gain, but may not have maximum coding gain. Subsequent computer
searches were carried out in [25, 26, 27, 28] to find codes with higher coding gains. New
code construction methods were also proposed, including the algebraic approach for
BPSK and QPSK modulation [34, 35], the systematic approach [47], generalized delay
diversity coding [46], and our proposed diagonal block space-time (DBST) coding
[Tao3].
Instead of following the traditional design criteria, several investigations were done
by [30, 29], and [Tao2] into different design criteria where, in particular, the role of the
Euclidean distance was studied at low SNR, or equivalently at a large number of receive
antennas. The corresponding codes in [31, 32, 33] and our codes in [Tao2] with a large
minimum Euclidean distance and non-full transmit diversity perform well with enough
receive antennas but poorly when the number of receive antennas is small.
A space-time block code is essentially a linear mapping from a group of modulated
data symbols onto codeword matrices in the space-time grid. There are several types of
STBCs based on the different theories employed in the code construction, such as
orthogonal designs [11, 9], amicable orthogonal designs [12, 13, 15], and algebraic
designs with constellation rotation [70]. The first orthogonal STBC for two antennas was
discovered by Alamouti [11] and, hence, is often referred to as Alamouti’s scheme. In
this scheme, the symbols input to the space-time block encoder are divided into groups of
two symbols each. At a given symbol periods, the two symbols in each group {c1, c2} are
transmitted simultaneously from the two antennas. That is, the signal transmitted from
antenna 1 is c1 and the signal transmitted from antenna 2 is c2. In the next symbol period,
the signal is transmitted from antenna 1 and the signal is transmitted from antenna
2. Hence, the codeword matrix (denoted by G
*2c− *
1c
2) can be written as
−= *12
*21
2 ccccG . (2.5)
21
Assume the channel coefficients keep constant over these two consecutive symbol
periods. Then the two symbols c1, c2 can be decoupled and decoded individually at the
receiver based on simple linear processing, and each of them achieves the transmit
diversity of order 2. This is due to the orthogonality attained in the codeword structure
(2.5). The very simple structure and linear processing of Alamouti’s scheme makes it a
very attractive scheme, and it is currently part of both the W-CDMA and CDMA-2000
standards. This scheme was later generalized by [9] to an arbitrary number of transmit
antennas using the theory of orthogonal designs. Define the code rate as the ratio of the
number of information symbols contained in each codeword to the number of symbol
periods the codeword occupies. It was shown that, for real signal modulation, e.g. PAM,
orthogonal STBC with rate 1 can be constructed while, for general complex modulation
like multi-level QAM and multi-level PSK, a rate-1 orthogonal STBC with simple linear
processing based ML detection and achieving full antenna diversity does not exist for M
> 2. There is a tradeoff among the orthogonality, the rate of the code, and the order of
transmit diversity. For example, the rate ¾ orthogonal code for M = 4 (denoted by H4)
achieving full transmit diversity is given by
−−+−−++−
−−
=
211233
122133
*3
*3
*12
*3
*3
*21
4
2222
2222
jyxjyxccjyxjyxcc
cccccccc
H (2.6)
in which xi and yi are the real part and imaginary part of the complex symbol ci
respectively. The design of quasi-orthogonal codes can be found in [14, 16, 17, 18], some
of which may sacrifice the diversity gain to have a higher code rate. The diagonal
algebraic STBC proposed in [70] solves this problem. It achieves full transmit diversity at
rate 1. However, it creates new issues as well, e.g. peak-to-average power ratio and
receiver complexity.
In general, STBC performs worse than STTC due to lower coding advantage. Some
efforts have been made in [20, 21, 22, 23, 24] on the concatenation of orthogonal STBC
and outer channel coding in order to enhance the coding advantage.
22
2.3.3 Rate-Diversity-Oriented Space-Time Techniques
There are two types of rate-diversity-oriented space-time techniques. One achieves both
diversity gain and multiplexing gain, and ensures a balance between the two, while the
other, more ambitious, endeavors to achieve full diversity gain and maximum
multiplexing gain simultaneously. Examples of the former technique, found in [19, 40],
can be viewed as a direct compromise between diversity-oriented space-time codes
(STTC or STBC) and rate-oriented VBLAST. The study on this rate-diversity tradeoff
technique was then intensively conducted in [Tao4]. Examples without a tradeoff include
the linear dispersion codes [68, 69], the threaded algebraic space-time codes [73], and the
algebraic code for two transmit antennas [71]. All these codes apply (linear) precoding at
the transmitter and use sphere decoder [66] at the receiver to do ML decoding. The
disadvantage of these codes compared with the tradeoff-type scheme is the expansion of
transmitted signal points due to the use of complex-field precoding. So far, there has been
only limited work on this non-tradeoff topic.
2.3.4 Non-Coherent Diversity-Oriented (Differential) Unitary Space-Time
Modulation
The information-theoretic study in [51] suggests a capacity-achieving signal structure
which comprises complex-valued signals that are orthonormal with respect to time among
the transmit antennas. Specifically, the M × T transmitted signal matrix on M antennas
over T time intervals has the partitioned form , where the M
rows, representing the temporal signals fed to the M Tx antennas, are mutually
orthogonal. It is further shown that setting v
[ ]TTMM
TT vvv φφφ ,,, 2211 K=C
1 = v2 = … = vM = T achieves capacity for
either T >> M or for high SNR and T > M. Unitary space-time modulation (USTM) [53]
is, hence, defined as the transmission of ΦT=C , where . I=HΦΦ
To design a USTM with good error performance, one should follow the criterion: for
any pair of codewords C1 and C2, none of the singular values of Φ should be 1. In
other words, C
H21Φ
1 and C2 should be made as orthogonal as possible. In this case, the
23
number of singular values that are not equal to 1 quantifies the order of achieved transmit
diversity. Several examples are provided in [53] and [54].
Most of current USTM schemes were designed with T = 2M, and are somehow not
very spectrally efficient. A differential USTM was then proposed in [56 57, ]. There can
be two different ways to look at DUSTM. First, it can be regarded as an overlapped
USTM with T = 2M, in which the first half of a USTM signal matrix is made the same as
the second half of a previously transmitted signal matrix by a certain information-lossless
unitary transformation. Therefore, when transmitting this signal matrix, it is only
necessary to send the second half. Also, DUSTM can be viewed as an extension of the
traditional differential PSK modulation used in single-antenna systems. The main
difference is that the signal constellation is no longer the set of scalar symbols with unit
amplitude, but the set of M × M complex-valued unitary matrices.
The encoding processing of DUSTM is briefly reviewed. As the signals are
transmitted in the unit of block, each containing M time intervals, it is convenient to use k
= 0, 1, … to denote the block index; within the kth block, the time index is denoted as t =
kM, kM + 1, …, kM + M − 1. We let Ck denote the M × M unitary signal matrix
transmitted over M transmit antennas during the kth block. The differential transmission is
initiated by sending the identity matrix, i.e. C0 = IM. Then, with differential encoding, we
have, at block k = 1, 2, …,
)(1 klkk VCC −= , (2.7)
where Vl(k), with l(k) ∈ {0, 1, …, L−1}, is the M × M data matrix at time block k and is
selected from a unitary matrix constellation V with size L, i.e. Vl(k) ∈ V ≡ {Vl | VlH
lV = I,
l = 0,1,…, L−1}. The transmitted signal matrix Ck generally does not belong to the
constellation unless the constellation itself forms a group under matrix multiplication.
The design of signals constellations follows a similar rank and determinant criteria.
Existing constellations for DUSTM include the diagonal cyclic group constellations
discussed in [57], group constellation of [56], Alamouti’s scheme in [44, 63], our
proposed orthogonal STBC [Tao7], amicable orthogonal codes [59], parametric codes
[62], and other group and non-group constellations [55]. For a large number of receive
antennas, diagonal cyclic group codes were presented in [61].
24
Chapter 3
IMPROVED DESIGN OF SPACE-TIME TRELLIS
CODES AT DIFFERENT SNR
In Chapter 2, we reviewed the concept and several key design examples of space-time
coding. It is known that the traditional rank and determinant criteria been widely used to
construct many classes of space-time trellis codes. Several handcrafted space-time trellis
codes using two transmit antennas were provided in [8]. These codes are full rank but
may not achieve the maximum coding gain. Subsequent computer searches were carried
out in [25, 26] to find codes with larger coding gain. However, it is worth to notice that
the derivation of traditional rank and determinant criteria was based on the assumption
that the SNR is sufficiently high. From existing simulation results [8, 25, 26], we observe
that to achieve a frame error rate (FER) of 10-2, the codes with QPSK modulation and 2
bit/s/Hz transmission efficiency require only around 10 dB transmit SNR when the
number of receive antennas, N, is equal to the number of transmit antennas, M (M = N = 2
in this example). Even smaller SNR is required when N > M. In this situation, the
assumption of high SNR is not valid and thus the two criteria are not tight.
In this chapter, we address the issue of non-high SNR and derive more precise criteria
for designing space-time trellis codes. Based on our new criteria, we also provide several
code examples using computer search. The remainder of this chapter is organized as
follows. We start by introducing the system model in Section 3.1. In Section 3.2, we
derive the improved design criteria. The trellis code examples and their simulation results
are provided in Section 3.3. Finally Section 3.4 summarizes this Chapter. The results in
this chapter are published in [Tao2].
25
3.1 System Model
We consider a point-to-point wireless link with M transmit antennas and N receive
antennas as shown in Fig. 2.3. It is assumed that the channel is quasi-static flat Rayleigh
fading and the channel coefficients are perfectly known to the receiver. The received
signal on antenna n at time slot t is given by
nt
mt
M
mmns
nt wchEr += ∑
=1, (3.1)
where hn,m is the normalized channel coefficient from transmit antenna m to receive
antenna n; c is transmitted symbol by antenna m at time t and chosen from a certain
constellation (e.g. PSK and QAM) with unit average energy; is the additive complex
white Gaussian noise with zero mean and variance ; and E
mt
ntw
0N s is the average energy per
symbol. The SNR is defined to be total transmitted signal energy to noise power spectral
density ratio, and given by 0NMEs=ρ .
3.2 Improved Design Criteria
A space-time codeword is defined as an M × T matrix, with T being the codeword length
=
MT
MM
T
T
ccc
cccccc
L
MOMM
L
L
21
222
21
112
11
C (3.2)
in which the t-th column represents the space-time symbol transmitted at time t and the
m-th row be the symbol sequence transmitted from antenna m. Suppose C, E ∈ are
two possible space-time codewords. Let B be the codeword error matrix defined by
. Further define A = BB
TM ×C
ECB −= H, in which the (p, q)-th entry can be written as
[ ] ( )( *
1,
pt
pt
T
t
pt
ptqp ecec −−=∑
=A ) . (3.3)
The Chernoff bound of the pair-wise error probability (PWEP) has been derived in [8] as
26
( ) ( )[ NM
NM
iiM
P −−
=+=
+≤→ ∏ AIEC αλρ det
41
1] (3.4)
where α = M4ρ , and λi, i = 1, 2, …, M, are the eigenvalues of A with λ1 ≥ λ2 ≥ …λn ≥
0. Evidently, to minimize the PWEP, the optimal design criterion is to maximize
( AI )α+det over all C . Unlike [8] where high SNR was assumed, we study the
design criteria for different ranges of SNR. In practical system design, the FER is
required to meet the system specification. Based on the number of transmit and receive
antennas, this FER requirement translates to a minimum required value for α. Given the
FER requirement, the number of antennas and the code complexity, α falls into a certain
range and the objective of code design is to minimize the FER for that SNR range. The
designed code need not be optimal for a higher SNR range as the FER at that higher SNR
range is much better than the requirement already. Similarly, code that works better at a
lower SNR range will not be useful because the FER cannot meet the minimum system
specification anyway. In this contribution, we consider the following three SNR ranges.
E≠
3.2.1 Case 1: α ≈ 1 (Moderate SNR)
As discussed above, the required SNR to achieve 10-2 FER is around 10 dB when M = N
= 2, which results in α=10/(4⋅2)≈1. In general, it is valid to assume moderate SNR for
achieving a FER of most interest when M ≈ N, which implies α ≈ 1. Hence from (3.3) we
reach the following criterion when α ≈ 1:
The minimum determinant of the matrix I + A over all possible distinct codword
pairs C and E must be maximized to minimize the PWEP for moderate SNR.
3.2.2 Case 2: α << 1 (Low SNR)
To see how this assumption can be valid, we observe that, to achieve a target FER, the
required SNR decreases as the number of receive antennas increases. As a result, when N
>> M, α will be much less than 1. Based on this assumption, we ignore the high order
terms of α and further upper bound the PWEP tightly by
27
. (3.5) ( )NM
iiP
−
=
+≤→ ∑
11 λαEC
Since the sum of the eigenvalues is equal to the trace, and the trace of A is exactly the
squared Euclidean distance between the codewords C and E by observing the definition
of A in (3.3), we reach the following conclusion:
The design criteion for low SNR is just to maximize the minimum squared Euclidean
distance of the space-time code.
The role of the squared Euclidean distance in the design of space-time codes was also
analyzed in [23]. The difference is that the problem considered in [23] still assumed high
SNR and a criterion of equating the eigenvalues of A for all pairs of C and E was
reached.
3.2.3 Case 3: α >> 1 (High SNR)
If N << M, we need a high SNR to achieve a desired FER and this is the case most
commonly assumed. The PWEP is now upper bounded by
( )Nr
ii
rNP−
=
−
≤→ ∏
1λαEC (3.6)
where r is the rank of B. Therefore, this reduces to the same criteria as in [7, 8] for high
SNR:
1) Rank Criterion: In order to achieve the maximum diversity MN, the matrix B has
to be full rank for any two distinct codewords C and E.
2) Determinant Criterion: If a diversity of MN is the target, then the minimum
determinant of A which corresponds to coding gain must be maximized.
Note that, except in case 3, the code optimized for case 1 or 2 need not be full rank,
contrary to the common belief that full rank is always needed for good code.
3.3 Computer Searched Trellis Codes for Moderate SNR
In the literatures, there have been several trellis codes designed for space-time coding.
The original one is the handcrafted code in [8] (TSC codes). All of them are full rank in
28
spatial domain. The others are the systematic global searched codes in [14] (BBH codes)
and [15](YB codes). They are also full rank and try to achieve the coding gain as much as
possible by maximizing the minimum determinant of A. In other words, they are all
designed for case 3 and may not be optimal in case 1 and case 2. In this work, we are
more interested in case 1 and consider designing space-time trellis codes for two transmit
antennas functioning optimally at moderate SNR. Hence our searching program will seek
the codes that have the largest minimum determinant of the matrix I + A over all distinct
codeword pairs C and E.
3.3.1 Code Examples
Following both [25] and [26], we use a generator matrix G to represent a space-time
code. We take the space-time code for 2 bit/s/Hz with QPSK modulation and 8 trellis
states in [8, Fig. 5] as an example. Let (at , bt) be the sequence of binary inputs at time t,
the output signal pair 1 2( , )t tx x is given by
( ) ( ) 4mod, 21121 G⋅= −−− ttttttt ababaxx ( )
where
=
2202012010
G .
Tables 3.1 to 3.4 list some of the search results. Three parameters, the minimum
squared Euclidean distance ( ), the minimum determinant of A (det(A)), and the
minimum determinant of I + A (det(I +A)), are given for numerical comparison with the
TSC, BBH and YB codes.
2mind
For the QPSK, 4-state, 2 bit/s/Hz code in Table 3.1, there are many different codes
found by global search that satisfy det(I+A) = 17 but have different and det(A).
Among them, we select the one with the largest d as our new code. We can see that
this new code has the largest out of all the codes listed in this table, the same
det(I+A) as the YB code and smaller det(A) than both the BBH and YB codes. For the Q-
2mind
2min
2mind
29
PSK, 8-state, 2 bit/s/Hz code in Table 3.2, there are also many codes satisfying the same
three parameters. We just randomly pick one as our new code. The new code has the
largest d and det(I+A), but smaller det(A) than the YB code. For the Q-PSK, 16-state,
2 bit/s/Hz code in Table 3.3, we also find multiple codes satisfying det(I+A) = 45. All of
them have the same but different det(A), which is either 28 or 32. We randomly
select one as our new code from those with det(A) = 32. It is seen that the YB code and
our new code have the same three parameters although the generator matrices are
different.
2min
2mind
2min
)A
)A+
The 8PSK, 8-state, 3 bit/s/Hz codes are listed in Table 3.4. No BBH and YB codes
are available in this case. For our new code, the last row of the generator matrix G can be
any permutation of 6 and 2 since they are symmetric in the 8-PSK constellation. We
claim that this is the best code [maximizing det(I+A)] we have found so far. It may not be
the optimal one over global search. As seen in the table, the new code is not full rank as
det(A) = 0, but has larger and det(I+A) than the TSC code. 2mind
Table 3.1: Space-time codes with QPSK, 4 states, 2 bit/s/Hz
G
New
02211220
TSC
01021020
BBH
13012022
YB
12212022
d 10 4 6 8
det( 4 4 8 8
det(I 17 9 15 17
30
Table 3.2: Space-time codes with QPSK, 8 states, 2 bit/s/Hz
G
New
2022112032
TSC
2202012010
BBH
2202012210
YB
2220011220
2mind 12 8 8 10
)det(A 12 12 12 16
)det( AI + 29 25 21 27
Table 3.3: Space-time codes with QPSK, 16 states, 2 bit/s/Hz
G
New
220201102022
TSC
022021021020
BBH
022021021220
YB
200221112220
2mind 12 8 12 12
)det(A 32 12 20 32
)det( AI + 45 21 37 45
Table 3.4: Space-time codes with 8PSK, 8 states, 3 bit/s/Hz
G
New
)2(6)2(64404040120
TSC
050204102040
2mind 6 4
)det(A 0 2
)det( AI + 9.929 7
31
3.3.2 Simulation Results
The simulation results for the QPSK, 8-state, 2 bit/s/Hz code and the 8PSK, 8-state, 3
bit/s/Hz code are of interest since our new codes have the largest det(I+A) in these two
cases. We illustrate the FER performance of these two cases in Fig. 3.1 and Fig. 3.2,
respectively, with two, three and four Rx antennas. Each frame contains 130 transmitted
symbols out of each transmit antenna. The channel coefficients are kept constant within
one frame and changed randomly from one frame to the other frame. From Fig. 3.1 it is
observed that the new code is 0.5~0.8 dB better than the TSC code and slightly better
than the YB code in systems with three and four receive antennas. From Fig. 3.2 it is seen
that the new code performs slightly worse that TSC code at high SNR with two Rx
antennas, but much better with three and four Rx antennas and the gains are 0.8 dB and
1.0 dB, respectively, at a FER of 10-2.
0 2 4 6 8 10 12 1410
−3
10−2
10−1
100
SNR [dB]
Fra
me
Err
or P
roba
bilit
y
2 Rx3 Rx4 Rx
TSCYBNew
Fig. 3.1: Performance of the QPSK, 8-state, 2 bit/s/Hz space-time codes with 2 transmit and 2, 3,4
receive antennas
32
4 6 8 10 12 14 16 1810
−3
10−2
10−1
100
SNR [dB]
Fra
me
Err
or P
roba
bilit
y
2 Rx3 Rx4 Rx
TSCNew
Fig. 3.2: Performance of the 8-PSK, 8-state, 3 bit/s/Hz space-time codes with 2 transmit and 2, 3,
4 receive antennas
3.4 Summary
In this chapter we considered the design of space-time codes in practical systems at
different SNR regions or, equivalently, at different number combinations between
transmit antennas and receive antennas. We derived tighter design criteria than the
traditional ones for low and moderate SNRs. In particular, we showed that the traditional
full-rank criterion is not always necessary. Moreover, the minimum Euclidean distance is
the dominating factor at low SNR in the design of space-time codes minimizing the error
probability. We also provided several new trellis code examples for moderate SNR
33
through computer search. These codes showed better performance than existing codes
under a wide range of low to moderately high SNR conditions.
34
Chapter 4
DIAGONAL BLOCK SPACE-TIME CODING
In the last chapter we discussed the improved design criteria of space-time trellis codes
and also provided several code examples for moderate SNR via exhaustive search. In this
chapter, we are interested in the design of space-time trellis codes when the number of
receive antennas is limited so that the rank and determinant criteria for high SNR are still
valid.
A large number of existing space-time trellis codes are found based on exhaustive
search. Hence they are limited to a small number of transmit antennas (2 or 3) and low-
level modulation (BPSK or QPSK) due to the extraordinary time complexity of
searching. In this chapter, we propose an efficient systematic coding approach that is
suitable for an arbitrary number of transmit antennas with arbitrary signal constellations.
The key of this approach is to separate the traditional space-time code design into two
parts. It first encodes the information symbols by a one-dimensional (M, 1) nonbinary
block code, with M being the number of transmit antennas, and then transmits the coded
symbols through the M antennas in a diagonal pattern. Hence, we refer to this scheme as
diagonal block space-time (DBST) coding. We show that, regardless of the channel time-
selectivity, this new class of space-time codes always achieves a transmit diversity of
order M with the minimum number of trellis states and a coding advantage equal to the
minimum product distance of the employed block codes. Traditional delay diversity
codes can be viewed as a special example in this category where a repetition block code
is employed. To maximize the coding advantage, we introduce an optimal construction of
the nonbinary block code for given modulation schemes. We also propose an efficient
suboptimal solution for multi-level PSK modulation.
Similar work on generalized delay diversity codes was done in [46], where only the
quasi-static fading was considered and it is not clear how to apply its design method to
high-level modulation. A recent work in [47] suggested another systematic design of
space-time trellis codes for quasi-static fading by manually assigning the channel output
35
symbols for each trellis state transition with certain rules. This scheme can achieve the
maximum possible antenna diversity order but its coding advantage is less efficient, as
will be shown in detail. An algebraic approach was presented in [34, 35], but it is only for
BPSK and QPSK modulation.
This chapter is organized as follows. In Section 4.1, we review the system model and
the fundamental performance criteria of space-time codes at both quasi-static and fast
fading channels. The proposed diagonal block space-time code structure and its pair-wise
error probability are described in Section 4.2. The nonbinary block code construction is
presented in Section 4.3, along with some code examples. In Section 4.4, the performance
of the proposed codes is evaluated and compared with that of existing codes. Finally,
Section 4.5 concludes this chapter. The results in this chapter are published in [Tao3,
Tao5]
4.1 System Model and Performance Criteria
The system model considered in this chapter is the same as Chapter 3, except that a fast
fading channel model is included here. While the channel coefficients keep unchanged
within a space-time codeword length in quasi-static fading channels, they vary
independently at different time slots in rapid fading channels.
The pair-wise error probability of mistaking codeword E for C with ML decoder at
both quasi-static fading and rapid fading channels for high SNR can be summarized as
( ) ( ) NP
NE
EM
PH
−⋅−
≤→
4ρEC , (4.1)
in which
(a) Quasi-static fading: EH is the rank of the codeword error matrix B = C−E, and EP
is the product of nonzero eigenvalues of matrix A = BBH
(b) Rapid fading: EH is the size of the time index set of 1 ≤ t ≤ T with c , denoted
by ζ, where c
tt e≠
t (et) is the t-th column in the M × T codeword matrix C (E); and EP
is the product of 2tt ec − over ζ∈t .
36
In both cases, EH can be called the effective Hamming distance and EP the effective
product distance. These two parameters quantify the transmit (either in the space domain
or time domain) diversity order and the coding advantage, respectively, of a space-time
code.
4.2 Diagonal Block Space-Time Codes
4.2.1 Code Structure
Fig. 4.1 depicts the simplified transmission diagram of the proposed diagonal block
space-time codes in a system with M transmit antennas. Assume the information bit
stream is divided into b-bit long blocks, forming P-ary (P = 2b) source symbols, denoted
by st, , at time t (t = 1, 2, ….). As can be seen from Fig. 4.1, the
encoding framework can be separated into two parts. In the first part, each information
symbol s
{ 1,,1,0 −∈ Pst K
MtsK
1ts
tc
mts
}
t is encoded by an (M, 1) nonbinary block code with output codeword
[ ]. In the second part, the M elements of each output codeword are
transmitted by the M antennas in a diagonal pattern across the space-time grid. That is,
while the first element is transmitted by antenna one at time t, the second element
is transmitted by antenna two at time t+1, the third one by antenna three at t+2, and so
forth. As a consequence, the baseband version of the transmitted signal at time t on
antenna m is given by , for m = 1, 2, …, M, where f is the modulator
mapping function and = s
tt ss 21
2ts
3ts
)( 1m
mtm sf +−=
t = 0 when t ≤ 0. Hence, the space-time codeword pattern is
formulated as
=
+−+−
−+−
−++
LLL
MMOMMM
LLL
LLL
)()()(
)()()()()()(
21
22
221
11
11
1
Mt
MMt
MMt
Mttt
Mttt
sfsfsf
sfsfsfsfsfsf
C . (4.2)
The design criterion of the (M, 1) nonbinary block code will be discussed in detail in
the next subsection. It is noted here that the Hamming distance between any two distinct
37
block code outputs is equal to M, i.e., there is a one-to-one mapping from each input
symbol to every element in the output codeword [32]. The original delay diversity code
falls in the special case when this block code is a repetition code, i.e., = smts t for all =
1, 2, …, M.
m
(M, 1) Nonbinary Block Code
D
Tx 1
Tx 2
Tx M
st
1ts
D D
Mapper
Mapper
Mapper
2ts
Mts
Fig. 4.1: Transmitter diagram of the diagonal block space-time codes, “D” denotes one symbol
delay
Note that the signal vector ct transmitted at a given time t is governed by the current
input st and the M−1 most recent inputs st−1, st−2, …, and st−Μ+1. The encoder thus forms a
finite-state-machine and we can define the trellis state at time t as
( )121 ,,, +−−−= Mtttt sssS K .
Given that the information symbols are P-ary, the total number of trellis states is equal to
PM−1, which is the minimum number of states for a space-time trellis code to achieve full
antenna diversity over a quasi-static fading channel [8]. During the transition from state St
to state St+1 produced by input st, the encoder outputs M channel symbol indices [
… ], one for each transmit antenna. This procedure can be illustrated as
1ts 2
1−ts
MMts 1+−
( )1
12
11
+ → +−−t
sssst SS
MMtttt K .
Therefore, the Viterbi algorithm can be applied at the receiver to do ML decoding.
4.2.2 Performance Measure
Proposition: In both quasi-static fading and rapid fading channels, the diagonal block
space-time code with M transmit antennas satisfies
MEH =min (4.3)
38
and
∏=−≤<≤
−=M
m
mm
PssP sfsfE1
2
1~0min )~()(min , (4.4)
where EHmin and EPmin are the minimum effective Hamming distance and the associated
minimum effective product distance in the PWEP formula (4.1), [s1 … sM] and
]~~[ 1 Mss K are the two block code outputs generated by inputs s and s~ , respectively.
Note from this Proposition that the proposed DBST codes can always achieve a
transmit diversity of order M, and its coding advantage is governed by the minimum
product distance (PDmin) of the employed (M, 1) nonbinary block code over a chosen
modulation scheme.
Proof: Consider an error event in the Viterbi-algorithm based ML decoder. The
correct and estimated trellis states are denoted by St and tS~ , respectively. Similarly, the
transmitted information symbol sequence and the estimated sequence are denoted by s =
{st} and }~{~ts=s , respectively. Suppose without loss of generality that, in this error
event, the estimated path through the trellis diverges from the correct path at time k and
remerges with the correct path at time k+T. Then, we have
=
−=≠
≠=
++
++
TkTk
tktk
kkkk
SS
TtSS
ssSS
~1,,1for,~
~and~
K . (4.5)
Because every single error in the information symbol sequence propagates M time
intervals by the nature of diagonal transmission, it follows that T ≥ M. The corresponding
space-time codeword difference matrix within this time period is of the form
TMM
MTMM
MT
MT
×−
−
−
∆∆∆
∆∆∆
∆∆∆
=−=
LL
MOOM
L
LL
10
222
20
111
10
00
0 0
00
ECB (4.6)
where )~()( mik
mik
mi sfsf ++ −=∆ , for m = 1, 2, …, M and i = 0, 1, …, T − M.
(a) In quasi-static fading: Partition the codeword difference matrix (4.6) into B = [U
R], where U is an M × M upper triangular matrix and R is an M × (T−M) matrix. Since
39
the inequality s kk s~≠ results in for all m = 1, 2, …, M, U is a full rank matrix.
Then B is also a full rank matrix. Hence the effective Hamming distance E
00 ≠∆m
H is equal to
M. It also yields
( ) ( )( ) ∏
=∆==≥
+==M
m
mH
HHHPE
1
20
2)det(det
detdet
UUU
RRUUBB , (4.7)
where “≥” holds directly from the Minkowski’s determinant inequality [74] as a
Hermitian matrix of the form RRH is nonnegative definite. The equality in (4.7), indeed,
holds if and only if R = 0, i.e., the sequence pair s and s~ are different at time k only. In
the sequel, the final expression in (4.7) corresponds to the product distance between the
two block code outputs [ … ] and [1ks M
ks 1~ks … M
ks~ ] generated by input sk and ks~ ,
respectively. Thus, the results for the quasi-static fading case are proved.
(b) In rapid fading: By definition, EH is now equal to the number of nonzero columns
in B shown in (4.6) associated with the sequence pair s and s~ . It is observed based on
(4.5) that EH = T. Therefore, we have EHmin = M as T ≥ M. Indeed, the minimum occurs if
and only if s and s~ are distinct at time k only. Under this circumstance, the effective
product distance EP is equal to
∏=
∆=M
m
mPE
1
20 . (4.8)
The results in rapid fading are also proved. ■
As in the proof, EH of a DBST code in quasi-static fading channels is always equal to
M for any distinct information sequences s and s~ , but EPmin is obtained if and only if
there is only one symbol error between s and s~ . While, in rapid fading channels, EH
depends on the error sequence and it can be greater than M. Nevertheless, EHmin is still
equal to M and it occurs if and only if there is only one symbol error between s and s~
too. This is because eventually there is no outer coding across the information symbols.
The quasi-static fading and the rapid fading are just the two extreme cases of a general
time-varying fading model. It is, therefore, reasonable to expect that, regardless of the
channel time-selectivity, a DBST code always achieves the diversity advantage and
coding advantage as shown in (4.3) and (4.4), respectively. Simulation results in Section
40
V will demonstrate this statement.
With this Proposition, the optimization of a DBST code simply amounts to finding
the optimal block code that maximizes the minimum product distance given in (4.4). In
particular, the minimum product distance of a repetition code is given by PDmin (rep) =
d2M, where d is the minimum Euclidean distance of the signal constellation. Later on, to
characterize the theoretical performance of our proposed coding scheme, we will treat
delay diversity codes as references and define the asymptotic improvement of a DBST
code as
]dB[rep)(
)new(log10
min
min10 PD
PDM
≡∆∞ . (4.9)
4.2.3 Discussions on Diagonal Structure
The diagonal transmission pattern in the proposed DBST coding has been frequently
utilized in the literature for MIMO systems. It first appeared in [3] as the diagonally
layered space-time architecture (D-BLAST). Recent work includes the trellis coded D-
BLAST [48] and wrapped space-time coding [49]. Most of existing work, designed for
the case when the number of receive antennas N ≥ M, relies on the diagonal structure to
perform a simple ZF or MMSE decision-feedback detection coupled with constituent
decoder at the receiver. This work, instead, applies the diagonal structure to achieve full
transmit diversity, which is essential for reliable transmissions when N < M. In the
sequel, our scheme achieves far lower error probability than the variants of D-BLAST
when N takes a small value (In particular, N = 1 in the downlink of most personal
wireless communication systems).
As another merit of the diagonal transmission pattern in this work, the proposed
DBST coding scheme can be easily extended to frequency-selective fading channels. As
done in [50] for delay diversity codes, we can change the delay step in Fig. 4.1 from one
symbol period to L symbol periods, with L being the number of channel taps. Therefore,
the maximum possible combined transmit diversity of order LM is achieved, as shown in
[50].
41
4.3 (M, 1) Nonbinary Block Code Construction
In the last section we introduced the DBST code structure and derived the minimum
product distance criterion (4.4) for designing the employed (M, 1) nonbinary block code
with M transmit antennas. In this section we discuss the construction of this 1-D code in
detail.
4.3.1 Optimal Construction for Given Constellations
We first consider the optimal code construction. Let c , which is the m-th
element in the block code output corresponding to input s∈ , with m = 1, 2,
…, M. Due to the one-to-one mapping between s and every , the P-long sequence
, , …, forms a permutation of the numbers 0, 1, …, P−1. The
ultimate code design is, therefore, to find these permutations for m = 1, 2, …, M that give
the largest minimum product distance over a given modulator mapping function. As the
numbers 0, 1, …, P−1 can be arranged in P! different ways, the size of the exhaustive
search space is . As each transmit antenna is statistically equivalent to every other
in the space domain, the permutation on the first antenna can be fixed. Without loss of
generality, we simply let it be the natural order {0, 1, …, P−1} and form a systematic
block code with . For constellations in symmetrical shape, the size can be further
reduced, as done for QPSK modulation in [46]. Yet, as the increase of P and M, the
complexity of exhaustive search still increases prohibitively.
mm ss =)(
,,1,0{ PK
)(scm
}1−
)0(mc )1(mc )1( −Pcm
M
s=
( )P!
s1
To solve this permutation optimization problem efficiently, the general branch and
bound algorithm [75] can be applied. The thrust of this algorithm is to form a tree
structure (branching operation) and establish a lower bound (bounding operation). We
take M = 2 for example to illustrate its application in our problem. As discussed above,
our problem is to find the permutation on the second antenna that can give the largest
minimum product distance. In the first level of the tree, the root has P children each
denoting an integer number between [0, P-1]. Each node in the first level further has P-1
children, each denoting an integer selected from [0, P-1]. During the construction of the
42
tree, a child node must be distinct from its ancestor nodes. The tree has P levels. Each
path from the root to a leaf corresponds to a possible permutation, while the whole tree
enumerates all P! permutations. The algorithm then traverses the tree in the depth-first
manner. When reaching a node of the tree, a local minimum product distance is
calculated. If it is greater than a given lower bound of the largest minimum product
distance, the search continues. Otherwise, the remaining tree associated with this node is
pruned. Once a permutation is found, it is used to form or improve the lower bound of the
largest minimum product distance. The extension to M > 2 is straightforward. The height
of the tree is still P. But the number of nodes in each level grows exponentially with M.
Table 4.1 and 4.2 list some of search results. Notice that the solution for P = 2 with
BPSK modulation at any M is just the repetition code and no more gain can be obtained
using other permutations. The modulator mapping function for P-ary PSK modulation is
given by f(s) = , while the mapping for QAM modulation with P = 16 and 32 is
shown in Fig. 2.2. For each code, all the codewords are arranged in an M × P matrix,
where each M × 1 column vector represents one codeword and P is the total number of
codewords. To illustrate the mapping of the nonbinary block codes onto the proposed
DBST codes, Fig. 4.3 gives an example of the trellis diagram for P = 4 (QPSK) and M =
2, in which the branch label xy denotes the symbols on antenna 1 and 2 respectively.
( Psie /2π )
Table 4.1 and 4.2 also shows the asymptotic improvement ∆∞, defined in (4.9), for
comparison. In particular, the ∆∞’s of the QAM codes in Table 4.2 are over the PSK
repetition codes. As we can see, although no improvement can be obtained in terms of
PDmin, the QPSK code with M = 2 has less multiplicity (multi = 2) than the repetition
code (multi = 4).
Notice that the optimal code for each pair of P and M listed in these two tables is not
unique due to the symmetric constellations. As a result, the 8PSK code with M = 2 is the
same as the 8PSK 8-state code designed by Tarokh et al for two transmit antennas in [8].
43
Table 4.1: Optimum block codes used in DBST coding for P = 4 and 8 with PSK modulation
P M Codeset PDmin(new) PDmin(rep) ∆∞ [dB]
2
23103210 4 (multi=2) 4(multi=4) 0
3 ref. [46]
312023103210
16 8 1.00
4 ref. [47]
3120231032103210
32 16 0.75
5
31202310231032103210
64 32 0.60
4
6 ref. [47]
312031202310231032103210
256 64 1.00
2
5274163076543210 ref. [8] 2 0.3431 3.83
3
526147304613752076543210
4 0.2010 4.33
4
36157240623751405267431076543210
4.6863 0.1177 4.00
5
5472163016472530475261305276341076543210
13.6569 0.0690 4.59
8
6
631572404516273051472630743165203745621076543210
32 0.0404 4.83
44
Table 4.2: Optimum block codes used in DBST coding for P =16 and 32 with PSK/QAM
modulation
P M Codeset PDmin(new) ∆∞ [dB] 2
PSK
81331015612194147211501514131211109876543210 0.4210 6.30
2 QAM
53814121017241591113601514131211109876543210 0.64 7.21
3 PSK
49142115158121610313707133106112815511214940
1514131211109876543210 0.8929 8.01
3 QAM
11071251492834156131014613101211072514915831
1514131211109876543210 1.28 8.53
16
4 PSK
12269133158410141511707101123613815294111450
12610151511841429137301514131211109876543210
2 8.93
2 PSK
162772121429924419311226617122102815320825133018523110
313029282726252423222120191817161514131211109876543210
L
L
0.1196 9.54
32
2 QAM
162118141229731282526152211962353027242019817131021430
313029282726252423222120191817161514131211109876543210
L
L
0.2 10.66
02, 12, 22, 32
03, 13, 23, 33
01, 11, 21, 31
00, 10, 20, 30
Fig. 4.2: Trellis diagram for the DBST code with QPSK and M = 2.
45
4.3.2 Linear construction for PSK modulation
The branch-and-bound algorithm can usually reduce the time of searching optimal codes.
However, as the search space grows exponentially, the searching time could still be
dramatically long for large P and M and renders this algorithm impractical. Hence, we
propose an efficient approach which may produce suboptimal solutions. This approach is
particularly applied for PSK modulation. By observing that the signal points in a P-ary
PSK constellation are evenly distributed on a unit circle, we can construct a linear block
code over a ring of integers, denoted by ZP = {0, 1, …, P−1}. The mapping from the ring
ZP onto the constellation is given by , for . Let the generator be
formed as a 1 × M row vector G = [g
( ) ( Psiesf /2π= )
)
PZs ∈
m Zg ∈1 g2 … gM], in which for m = 1, 2, …, M.
With input s∈ Z
P
P, the block code output c(s) = [s1 s2 … sM] is generated by
c(s) = s⋅G (mod P). (4.10)
The minimum product distance of this block code can thus be written as
( ) (∏=−≤<≤
−=M
m
PgsjPsgj
Pssmm eePD
1
2/~2/2
1~0min min ππ . (4.11)
After simple manipulation, we further write PDmin as
∏=−≤<
=M
m
mM
Ps Psg
PD1
2
10min sin4minπ
. (4.12)
Thus, designing the optimal linear code becomes finding the solutions of gm, m = 1, 2, …,
M, that maximize (4.12), and it can be done by performing a simple search in the set ZP
for every gm. The following four properties can be applied to further reduce the size of the
search space. First, it is seen that equation (4.12) does not change if gm is replaced by P
− gm. Hence, the search can be restricted in the new set {0, 1, …, P/2}. Second, to
guarantee nonzero PDmin, each gm must be relatively prime to P. Third, since each
transmit antenna is statistically equivalent to every other in the space domain, we can
impose the ordering g1 ≤ g2 ≤…≤ gM. Last, the codewords generated by [g1 g2 … gM] and
[αg1 αg2 … αgM] are identical for any α that is relatively prime to P. There exists an α in
the set ZP such that αg1 = 1 (mod P). By multiplying g2, … gM with this same α, we can
let g1 = 1.
46
Table 4.3 shows some of the search results for P =16 (16PSK), 32 (32PSK), and 64
(64PSK) with M = 2, 3, …, 6. Again, due to the symmetrical shapes of the constellations,
the optimum solution of the generator G is not unique. From this table, it is observed that
this linear construction, though suboptimal, provides reasonably good results besides
having significantly low complexity of searching. It is further noticed that the code with
P = 16 (16PSK) and M = 4 is indeed optimal in terms of PDmin as compared with the one
in Table 4.2.
Another suboptimal approach to combat the complexity of searching global optimal
block codes was recently reported in [47]. This approach at QPSK modulation, as a
matter of fact, provides the optimal solution that maximizes PDmin as indicated in Table I.
However, at higher-level modulation (P ≥ 8) it becomes much less efficient. For example,
the 8PSK code with M = 3 in [47] only achieves PDmin = 0.6863, while the optimal code
has PDmin(new) = 4 as shown in Table I. Similarly, the 16PSK code with M = 3 in [47]
only achieves PDmin = 0.1101, while the linear code we found in Table 4.3 using our
proposed suboptimal approach has PDmin(new) = 0.5198.
Table 4.3: Linear block ring codes used in DBST coding for P = 16, 32, and 64 with PSK
modulation
P M G PDmin(new) PDmin(rep) ∆∞ [dB] 2 [1 7] 0.3431 0.02318 5.85 3 [1 3 5] 0.5198 3.529×10-3 7.23 4 [1 3 5 7] 2 5.372×10-4 8.93 5 [1 1 3 7 7] 0.4020 8.178×10-5 7.38
16
6 [1 1 3 5 7 7] 1.1716 1.245×10-5 8.29 2 [1 7] 0.06186 1.477×10-3 8.11 3 [1 7 9] 0.08918 5.675×10-5 10.65 4 [1 7 9 15] 0.1177 2.181×10-6 11.83 5 [1 3 5 11 15] 0.1359 8.381×10-8 12.42
32
6 [1 3 5 11 13 15] 0.2702 3.221×10-9 13.21 2 [1 19] 0.02485 9.275×10-5 12.14 3 [1 11 27] 0.02862 8.932×10-7 15.02 4 [1 11 17 19] 0.04561 8.602×10-9 16.81 5 [1 3 23 25 27] 0.02364 8.284×10-11 16.91
64
6 [1 7 9 15 17 23] 0.03624 7.978×10-13 17.76
47
4.3.3 Discussions
As can be seen from Table 4.1 to Table 4.3, the asymptotic improvement of the new
space-time codes over delay diversity codes increases significantly as the constellation
size P increases. This is because the minimum product distance of a repetition code is
only a function of the minimum Euclidean distance of the given constellation.
Nevertheless, the minimum product distance of a new code depends on the whole
Euclidean distance profile, and the wider the profile distributes the more degree of
freedom the new code can exploit. This also explains why no gain can be obtained at
BPSK (P = 2) modulation. From Table 4.2, it is also observed that the asymptotic
advantage of using 16QAM constellation over 16PSK constellation is not as much as
expected traditionally.
Generally, for a given constellation, the product distance profile of the block code
should be made as dense as possible to maximize the minimum product distance. In the
ideal case, the product distance between any two distinct codewords should be a constant.
As a consequence, an upper bound of the optimum minimum product distance PDmin(opt)
can be obtained. Suppose the signal constellation has the Euclidean distances {d1, d2, …,
dk} with multiplicity {n1, n2, …, nk}, respectively. Then the PDmin(opt) of the (M, 1)
block code is upper bounded by
( ) ∑=×××≤k
iik
nMnk
nn dddPD 1212
21min )opt( K . (4.13)
Up to now, we have only considered the block code design over conventional
constellations, e.g. PSK and QAM. Notice, however, that our ultimate goal is to
maximize the value in (4.4) over an arbitrary constellation with unit average energy.
Therefore, a more general problem is to design the constellation shape at a given size P
with unit average energy that achieves the maximum value of PDmin. Yet, this is beyond
the scope of this paper.
48
4.4 Simulation Results
The analysis in the previous section provides the asymptotic performance improvement
of the proposed DBST codes over DD codes. In this section, simulation is carried out to
evaluate the actual performance gain in practical SNR region. The channel is set to be flat
Rayleigh fading, and the channel state information is available at the receiver but not at
the transmitter. Unless specified otherwise, ML decoding is obtained by the Viterbi
algorithm. The performances are plotted versus the total average transmitted SNR which,
by definition, is given as 0NMEs=ρ .
4.4.1 Comparison with Delay Diversity Codes
We first take the 8PSK code with M = 3 transmit antennas as shown in Table 4.1 for
example. Simulation is performed with three different channel autocorrelations in the
time domain. Since the frame error rate (FER) depends on the transmission frame length
and the bit error rate (BER) is a function of the bit-to-symbol mapping1, the information
symbol error rate (SER) is selected as the performance measure.
Fig. 4.3 plots the SER performance comparison over a quasi-static fading channel
(frame length T = 130). It is observed that the actual gain of the DBST code over the DD
code at a SER of 10-4 is about 1.8dB with one receive antenna. With two receive antennas
the gain increases to 3.5dB, which is less than 1dB away from the theoretically
asymptotic improvement of 4.33 dB shown in Table 4.1.
Fig. 4.4 shows the SER performance comparison over a rapid fading channel. Now
the actual gains at the SER of 10-4 are about 3.3dB and 3.9dB with one and two receive
antennas respectively, which are closer to the asymptotic improvement.
1 Gray-mapping is not necessarily the optimal mapping in space-time codes.
49
2 4 6 8 10 12 14 16 18 20 22 24 2610
−5
10−4
10−3
10−2
10−1
100
SNR [dB]
SE
R
1 Rx2 Rx
DBSTDD
Fig. 4.3: SER performance of 8PSK codes with M = 3 transmit antennas over a quasi-static fading
channel.
The SER performance comparison over a time-varying fading channel is illustrated in
Fig. 4.5. The channel autocorrelation function is modeled as ( kTfJ sd )π20 , where fd is
the maximum Doppler frequency, Ts is the symbol period, k is the discrete time index,
and is the zeroth order Bessel function of the first kind. The parameter f)(0 ⋅J dTs is set to
0.05 in this simulation. First, it is observed that the gain of the DBST code over the DD
code at 10-4 SER is around 2.5dB with one receive antenna, which is greater than the gain
over a quasi-static fading channel but less than that over rapid fading. It is also observed
that, with the same number of receive antennas, the performance curve of the DBST code
over this time-varying fading channel always lies somewhere between the curve in quasi-
static fading and that in rapid fading. This observation demonstrates clearly the
robustness of DBST codes over the time-selectivity of a fading channel.
50
2 4 6 8 10 12 14 16 18 20 22 24 2610
−5
10−4
10−3
10−2
10−1
100
SNR [dB]
SE
R
1 Rx2 Rx
DBSTDD
Fig. 4.4: SER performance of 8PSK codes with M = 3 transmit antennas over a rapid fading
channel.
To sketch the performance enhancement of other DBST codes in Table 4.1 to 4.3
over DD codes, we select the ones with M = 2 transmit antennas and evaluate their
required operating SNR’s, respectively, at a specified SER of 2 over a rapid fading
channel. The results are reported in Table 4.4, from which it is seen that the asymptotic
improvement predicts the actual SNR reduction very well, especially at more than
one receive antenna. A histogram of the results is also plotted in Fig. 4.6.
410−×
∞∆
51
2 4 6 8 10 12 14 16 18 20 22 24 2610
−5
10−4
10−3
10−2
10−1
100
SNR [dB]
SE
R
1 Rx2 Rx
DBSTDD
Fig. 4.5: SER performance of 8PSK codes with M = 3 transmit antennas over a time-varying
fading channel with . 05.0=sdTf
Table 4.4: Operating SNR [dB] at SER = 2 for codes with M = 2 transmit antennas over a
rapid fading channel
410−×
1 Rx 2 Rx SNRDBST SNRDD gain SNRDBST SNRDD gain QPSK 22 22.24 0.24 12.86 13.13 0.27 8PSK 25.33 27.74 2.41 15.48 18.5 3.02
16PSK∗ 29.34 33.6 4.26 18.9 24.47 5.57 32PSK* 33.15 39.56 6.41 21.84 30.42 8.58 64PSK 35.9 45.5 9.6 14.42 24.76 10.34
∗ Suboptimal codes from Table III
52
QPSK 8PSK 16PSK 32PSK 64PSK0
1
2
3
4
5
6
7
8
9
10
11
Modulation Level
Gai
n [d
B]
1Rx 2Rx
Fig. 4.6: Histogram of the gains shown in Table 4.4.
4.4.2 Comparison with Other Existing Codes
Through exhaustive search, authors in [33] provided several QPSK and 8PSK codes with
three and four transmit antennas for quasi-static flat fading channels based on the
Euclidean distance criterion. In particular, they designed 8PSK codes with up to 32 trellis
states with the order of transmit antenna diversity equal to 2. Comparisons in this
subsection are done between the 32-state 8PSK codes in [33] and our 8PSK codes at the
same number of transmit antennas. The transmission rate is the same for all the codes,
that is, 3 bit/s/Hz. But the number of trellis states is different. Our 8PSK codes with M =
3 and 4 have 64 and 512 trellis states, respectively. To have rather fair comparisons, a
suboptimal tree decoding algorithm is applied in the simulation of our 512-state code: the
M-algorithm [65]. In this algorithm, only a certain number of most likely states, denoted
53
as K, are kept and the remaining states are deleted at each decoding stage. Thus the
decoding complexity is O(K). In our case, K = 64. Fig. 4.7 illustrates the FER
performance comparison over a quasi-static fading channel with frame length T = 130. As
can be seen in this figure, a higher diversity order is achieved by our DBST codes. As a
result, even though our codes perform worse with two receive antennas, superior
performance is achieved with one receive antenna. This is because full transmit diversity
is necessary at high SNR with limited number of receive antennas, while the minimum
Euclidean distance is the dominating factor at low SNR with high enough total diversity
order, as claimed in Chapter 3. The codes in [33] have much larger minimum Euclidean
distance but smaller transmit diversity order than our codes. Hence, the observation in
Fig. 4.7 is not surprising.
6 8 10 12 14 16 18 20 22 2410
−3
10−2
10−1
100
SNR [dB]
FE
R
1 Rx2 Rx
M=3, 64−state DBSTM=4, 512−state (K = 64) DBSTM=3, 32−state in [33]M=4 32−state in [33]
Fig. 4.7: FER performance of 8PSK codes with M = 3, 4 transmit antennas over a quasi-static fading channel.
54
4.5 Summary
In this chapter we proposed an efficient and systematic space-time coding scheme:
diagonal block space-time coding. It is basically a two-step approach: first, construct a 1-
D nonbinary block code; then, apply the diagonal transmission pattern to send the block
code outputs through multiple transmit antennas. It was shown that the diagonal
transmission pattern promises a transmit (spatial and temporal) diversity of order M in a
system with M transmit antennas under both quasi-static and rapid flat fading channels,
while a carefully designed nonbinary block code assures good coding advantage. The
conventional delay diversity code is a special case of this coding scheme when the block
code is a repetition code. To design the optimal block code that maximizes the coding
advantage, two general problems were formulated, namely, the permutation optimization
for a given constellation and the constellation optimization. In particular, we proposed an
efficient linear block code construction over rings for multi-level PSK modulation.
Through simple computer search, we obtained some optimal and suboptimal code
examples using PSK and QAM modulations with 2~6 bit/s/Hz transmission rate and 2~6
transmit antennas. Simulation results showed that they possess significant advantage over
the original delay diversity codes in not only quasi-static fading and rapid fading channels
but also general time-varying fading channels. They also demonstrate superior
performance over the existing codes optimally designed based on the Euclidean distance
criterion with one receive antenna.
The proposed coding scheme is suitable for an arbitrary number of transmit antennas
with arbitrary signal constellations. It can also be easily extended to frequency-selective
fading channels. As it is designed for achieving full transmit antenna diversity with good
coding advantage, this coding scheme has particular application for the downlink
transmission where the base station is capable of installing multiple antennas while the
hand-held mobile station has only one antenna.
55
Chapter 5
GENERALIZED LAYERED SPACE-TIME
ARCHITECTURE
In Chapter 3 and 4, we proposed improvements on the design of space-time codes aiming
at minimizing the error probability at different scenarios. The transmission efficiency of
this type of transmission scheme is, however, constrained by the signal modulation level.
In this chapter, we provide a comprehensive study on generalized layered space-time
architecture (GLST), which can provide a flexible tradeoff between error probability and
transmission efficiency with reasonable complexity.
The basic framework of this architecture is to partition all the available transmit
antennas into several groups and apply space-time coding (STC) as component codes for
each group. At the receiver, to avoid the huge complexity of joint decoding of all groups,
each component space-time code is decoded individually through a serial processing
similar to VBLAST detection algorithm [4, 5] and combines group interference
suppression and group interference cancellation. Hence GLST can be viewed as a
combination of VBLAST and STC. On one hand, the signals transmitted within each
antenna group are space-time coded. Hence higher transmit antenna diversity can be
achieved compared with pure VBLAST. On the other hand, the overall spectral efficiency
is higher than pure STC due to the independence of the signals transmitted by different
antenna groups. Similar ideas can be found in [19] and [40], where space-time trellis
codes and space-time block codes are combined with VBLAST, respectively. We
generalize both of them in this chapter. In addition, we study several features embedded
in the generalized layered architecture to enhance the overall system performance,
including various signal-to-antenna mappings, power allocation and decoding order. To
further improve the system performance of GLST, we propose a computationally
efficient hard-decision iterative decoding scheme. This decoding algorithm performs
almost as good as the optimal ML decoding with much lower complexity.
56
This chapter is organized as follows. In Section 5.1 we introduce the symbol model of
GLST, including the basic encoding and decoding process. The optimal power allocation
and optimal decoding order are derived in Section 5.2 and 5.3, respectively. The iterative
decoding is proposed in Section 5.4. Finally Section 5.5 provides a summary.
The results in this chapter are published in [Tao4, Tao8, Tao9].
5.1 System Model
We consider a point-to-point wireless link with M transmit antennas and N receive
antennas as shown in Fig. 2.3. It is assumed that the channel is quasi-static flat Rayleigh
fading and the channel coefficients are perfectly known to the receiver. The basic
encoding and decoding process are described as follows.
S/P
STC1
B
B1
STC2
B2
STCq
Bq
C1 INT1
C2 INT2
Cq INTq
INT
(a)
α1 Tx 1,2
β1
γ1 Tx 5,6
Tx 3,4
α2 α3
β2 β3
γ2 γ3
t
α4
β4
γ4
α1 Tx 1,2
β1
γ1 Tx 5,6
Tx 3,4
γ2β3
α2γ3
β2α3
t
α4
β4
γ4
(b) (c)
Fig. 5.1: Encoder of (interleaved) GLST (a) main layout, (b) HGLST, and (c) DGLST.
57
5.1.1 Encoding
The encoding process of GLST is illustrated in Fig. 5.1. A block of B input information
bits is transformed by a serial-to-parallel (S/P) converter into q groups of bit stream of
length B1, B2, …, Bq, with . Each group of bit stream, BBBBB q =+++ K21 j, for 1 ≤ j ≤
q, is then separately encoded by a component space-time encoder STCj associated with
Mj Tx antennas, with M1 + M2 + … + Mq = M. The output Mj × T codeword matrix Cj is
to be transmitted by Mj antennas over T time intervals. The t-th column of Cj, denoted by
cj,t, is referred to as the symbol vector of group j at time t. Before transmission, the
symbol vectors from each group are passed through an individual temporal interleaver
followed by a spatial interleaver. The temporal interleavers are labeled by the dashed
block in Fig. 5.1(a) and will be discussed in Section 5.4. The spatial interleaver is
interpreted as a mapping from the symbol vectors to the transmit antennas at each symbol
period. We consider two special mappings, namely, horizontal mapping and diagonal
mapping, shown in Fig. 5.1(b) and (c), respectively.
The horizontal mapping is to simply pass all the symbol vectors from one group to a
fixed group of antennas. A simple example is shown in Fig. 5.1(b), where αt, βt, and γt,
with t = 1, 2, …, denote the symbol vectors from three different groups and each group is
associated with two transmit antennas. Thus, the transmitted symbol vectors from every
group form a row, or equivalently horizontal pattern and we refer to such GLST as
HGLST. In the diagonal mapping, the symbol vectors from each group cyclically utilize
different group of antennas at different symbol intervals with period p. A simple example
is illustrated in Fig. 5.2(c) with p = q = 3. Hence the transmitted symbol vectors from
every group are allocated in a diagonal pattern across the antenna and time dimensions.
We call such GLST as DGLST. Since the channel seen by the symbol vectors from each
group changes periodically, additional antenna diversity can be achieved in DGLST
compared with HGLST. It is noticed that the diagonal mapping considered in this paper
is different from the diagonal transmission pattern in [3], where zero-padding was
required at the beginning and the end of each transmission frame and hence the spectral
efficiency was reduced.
58
5.1.2 Decoding
Let rt denote the N × 1 received signal vector at time instance t in GLST. It can be written
as
ttqtqttttr wcHcHcHr ++++= ,,,2,2,1,1 L (5.1)
where Hj,t denotes the N × Mj subchannel matrix of group j at time t and wt denotes the
additive complex Gaussian noise term.
The layered structure of GLST allows a similar serial decoding as in VBLAST
systems, except that all the interference nulling and interference cancellation are group
based. The serial decoding order will be discussed in Section 5.3. In this subsection, we
assume it is already determined and denoted by, without loss of generality, {1’, 2’, …,
q’}. The decoding algorithm is the following.
At the first decoding level (j = 1), we let = r1tr t, for all 1 ≤ t ≤ T.
At the j-th decoding level, given that the previous j-1 groups 1’, 2’, …, (j-1)’ have
already been decoded and cancelled out from the received signals, the resulting received
signal for t =1, …, T, which still contains interference from the not-yet-decoded
groups (j+1)’, (j+2)’, …, q’, can now be written as,
jtr
jttqtqtjtjtjtj
jt vcHcHcHr ++++= ++ ,',',)'1(,)'1(,',' K . (5.2)
Then at each time t, we can find a set of orthonormal row vectors (not necessary unique)
in the null space of [H(j+1)’,t, …, Hq’,t], and form them into an (
nulling matrix . Multiplying with r suppresses all the signals from groups
(j+1)’ to q’ and generates the interference free decision signal for group j’
mnnnm j ×+++− )''1 L
jtW j
tW jt
jttjtj
jt
jttjtj
jt
jt
jt
jt vcHvW0cHWrWr ~~~
,',','', +=++== . (5.3)
It is seen that tj ,'~H
jtv
is the equivalent channel matrix at time t for group j’, whose entries
are independent and identically complex Gaussian distributed with mean zero and
variance 0.5 per dimension. Assuming perfect interference cancellation, the equivalent
noise vector ~ also contains independent and identically complex Gaussian distributed
entries [19].
59
As for the nulling matrix , we have = WjtW j
tW j for all 1 ≤ t ≤ T in HGLST since the
channel seen by each group is time-invariant in each transmission frame. This implies
that the nulling matrix is only calculated once for each frame. While in DGLST, since the
channel seen by each group varies with period p, we have = . Hence the total
number of p nulling matrices are required for each frame.
jtW j
pt modW
Now the j’-th codeword can be decoded using ML space-time decoder based on jtr
~
in (5.3)
∑=
−=T
tttj
jtj
1
2,''
~~minargˆ cHrCC
. (5.4)
Before moving on to the next decoding level, we subtract the contribution of the just
decoded group j’ from and it results in the modified received signal jtr 1+j
tr
Tttjtjj
tj
t ≤≤−=+ 1,ˆ ,','1 cHrr . (5.5)
We restart the above procedure for j = j + 1, until all groups are decoded (j = q).
Different from BLAST, the requirement for the number of receive antennas in GLST
is N ≥ M - M1’ + 1. Particularly, if the decoding order depends on the channel realization
for each transmission frame at the receiver, 1’ could be any integer from 1 to q. Hence to
guarantee that the system works, we let N ≥ . 1}{max +− jjMM
5.2 Optimal Power Allocation
It is seen from the above decoding procedure that, if the decoding order is pre-determined
and independent of the channel realizations at the receiver, group j’ decoded at the j-th
decoding level has the antenna diversity of order Mj’ (N−M + M1’ + … + Mj’). Assume M1
= M2 = … = Mq, and all groups employ the same STC and are assigned the same
transmission power, then group 1’ would have the worst performance. As the overall
system performance is usually limited by the group with the worst performance,
improving the performance of group 1’ will be most effective in enhancing the overall
system performance. Unless specified otherwise, the STCs employed for all groups are
identical in this paper. Then the most straightforward method to enhance the system
60
performance is to assign more power to the group that is decoded earlier and less to the
group decoded later under the constraint that the total transmission power is kept
constant.
Such unequal power allocation strategy was first applied in [19] and we call it
Tarokh’s power allocation in this paper. Tarokh’s power allocation basically reduces 3
dB power for each successive group along the decoding order. This scheme is novel but
has not been optimized. In this section, we will derive the optimal power allocation that
minimizes the probability of frame error.
For simplicity the pre-determined decoding order is set to be {1, 2, …, q}. Let Ej
denote the power level (in linear scale) assigned to group j and satisfy the total power
constraint
EEq
jj =∑
=1. (5.6)
Define Pj(Ej) as the individual frame error probability of group j as a function of Ej, given
that the previous j-1 groups are all decoded successfully. The probability of total frame
error, P(E), is then given by
( )∏=
−−=q
jjj EPEP
1)(11)( . (5.7)
When all Pj(Ej) are small, (9) can be tightly approximated by the sum of all Pj(Ej), i.e.
∑=
≈q
jjj EPEP
1)()( . (5.8)
Now the objective is to minimize P(E) in (5.8) subject to the constraint in (5.6). To
solve this problem, we can directly apply the Lagrange multiplier method
0)(1
=
−−
∂∂ ∑
=EEEP
E
q
jj
j
λ
λ=⇒j
jj
dEEdP )(
(5.9)
for all 1 ≤ j ≤ q, with λ determined by (5.6). Hence the optimal power allocation can be
achieved by (5.9), i.e. equalizing the derivative of the individual frame error probability
of all groups. An intuitive explanation of this result is as follows. Let us increase a very
small amount of power ∆E on a weaker group i and decrease the same ∆E from a stronger
61
group j. If the reduced individual frame error probability ∆Pi of group i is larger than the
increased individual frame error probability ∆Pj of group j, then we can say that the total
frame error probability is decreased. We can keep increasing the power on group i and
decreasing the power from group j as long as the corresponding ∆Pi is larger than ∆Pj.
The power exchange between the two groups can be stopped once ∆Pi and ∆Pj are equal,
i.e. the total frame error probability cannot be decreased anymore. Overall, the power
should be allocated in such a way that any small enough amount of power change will
have the same effect on all individual frame error probabilities. Mathematically, the
derivative of the individual frame error probabilities is the same, as shown in (5.9).
As the power allocation scheme does not depend on the channel information, the
optimization can be done offline for a target operating SNR. Hence the system
complexity is unchanged.
Theoretically obtaining the precise function for Pj(Ej) in GLST system is
complicated. In this paper we produce an approximate function using computer
simulation and curve fitting technique. We perform real interference nulling and space-
time decoder steps but ideal interference cancellation step, meaning the contribution from
the just decoded group is perfectly known and cancelled out at each decoding level. The
individual frame error probability for each group is then simulated and plotted as a
function of the power. Afterwards, curve fitting is applied on each curve and thus an
approximate function of the individual frame error probability is obtained for each group.
Simulation has been performed to evaluate the superiority of the proposed optimal
power allocation. For brevity, we use the notation (M, N) from now on to denote a MIMO
system with M transmit antennas and N receive antennas. The simulation setup is as
follows and it will be used in all the simulations in this paper. We take two GLST
systems as examples: (4, 4) GLST and (8, 8) GLST. Each system includes both HGLST
and DGLST versions. For simplicity, each group contains two antennas. The 2 bit/s/Hz
QPSK modulated 16-state space-time trellis code given in [8, Fig. 5] is taken as the
component code for all the groups. Thus the transmission rate of the two systems is 4
bit/s/Hz and 8 bit/s/Hz, respectively. We consider quasi-static flat Rayleigh fading
channels. The channel coefficients are assumed perfectly known at the receiver. Each
62
frame contains 130 transmission from each transmit antenna, which corresponds to 520
and 1040 information bits per frame at the (4, 4) and (8, 8) systems, respectively.
Fig. 5.2 shows the frame error rate (FER) performance versus the average received
SNR per receive antenna in the (4,4) and (8,8) systems. The curve fitting function for the
optimal power allocation we used in this simulation is ( ))(exp)( jjj EfEP = , where
is a third order polynomial function of E)( jEf j. We can see that, the improvement of the
optimal power allocation over equal power allocation (i.e. no power allocation) is about
1.5 dB and 3.5 dB at a FER of 0.01 in the (4,4) and (8,8) systems respectively. It is also
seen that, in the (4,4) system, the gain of the optimal power allocation over Tarokh’s
power allocation increases as the SNR increases, while in the (8,8) system, the gain at
low SNR is larger than the gain at high SNR.
Another observation from Fig. 5.2 is that DGLST generally performs better than
HGLST. This is as expected due to the additional diversity introduced by diagonal
mapping.
7 8 9 10 11 12 13 14 15
10−3
10−2
10−1
100
SNR [dB] per receive
Fra
me
Err
or R
ate
HGLST equal PA HGLST Tarokh PA HGLST optimal PADGLST equal PA DGLST Tarokh PA DGLST optimal PA
(a)
63
8 9 10 11 12 13 14 15 16 1710
−3
10−2
10−1
100
SNR [dB] per receive
Fra
me
Err
or R
ate
HGLST equal PA HGLST Tarokh PA HGLST optimal PADGLST equal PA DGLST Tarokh PA DGLST optimal PA
(b)
Fig. 5.2: Performance comparison of different power allocation in the (a) (4,4) and (b) (8,8) GLST systems.
5.3 Optimal Decoding Order
In the previous section we discussed the power allocation to enhance the system
performance when the decoding order is independent of CSI. In this section we will
demonstrate how to determine the decoding order based on the channel realizations at the
receiver when no power allocation is applied. Notice that, the decoding order in DGLST
systems is not important and can be arbitrary as different groups see the same time-
varying channels. Hence we only consider HGLST in this section.
At the first decoding level (j = 1), let H1 = [H1, …, Hq], where Hl is the N × Ml
subchannel matrix corresponding to group l, and define the candidate group set as S1≡{l
|1 ≤ l ≤ q}.
At the j-th decoding level, for each l∈ Sj, we find a set of orthonormal row vectors
in the null space of jl
H , where jl
H denotes the matrix obtained by eliminating Hl from
64
Hj, and use them to form an ( matrix . Then we
left-multiple H
NMMMMN ll ×++++− − ))'1('1 L
lj
ljl HWH =
jlW
l by to obtain the equivalent channel matrix of group l at this
decoding level
jlW
~
2⋅
2~maxarg' jlSl j
j H∈
=
}'{\1 jSS jj =+
jj
j'
1 HH =+
( 1 jMMM −−− K
)'1 −
.
As the post-detection SNR defined on the whole codeword from each group is
proportional to the squared Euclidean norm [74] of the corresponding equivalent
channel matrix, we choose the candidate with the largest squared Forbenius norm as the
group to be decoded at this level
. (5.10)
Then we update the candidate group set as, by removing j’ from Sj,
.
The channel matrix is correspondingly changed to, by eliminating Hj’ from Hj,
.
This process continues for j = j+1 until the candidate set becomes empty. The decoding
order is now obtained as {1’, 2’, …, q’}, which is a certain permutation of {1, 2, …, q}.
As BLAST order was shown to be optimal in the sense of minimizing the probability
of frame error [6], the above decoding order (5.10) can also be similarly shown to be
optimal in the same sense and we omit the proof.
To this stage, let us compare the complexity among HGLST with arbitrary (or pre-
determined) decoding order, HGLST with optimal decoding order and DGLST with
arbitrary decoding order (no optimal decoding order for DGLST). First we consider the
complexity of performing interference nulling measured by the computation of finding
the null space basis of a two-dimensional complex matrix. At the j-th decoding level of
HGLST with arbitrary decoding order, say {1, 2, …, q}, it requires computation of
finding the null space basis of an N × ) matrix. At the j-th decoding
level of HGLST with optimal decoding order {1’, 2’, …, q’}, to determine which group
will be decoded, the computation of finding the null space basis of an N ×
matrix has to be done for every l∈ S)( ('1 lj MMMM −−− −K j. While at the j-th
65
decoding level of DGLST with arbitrary decoding order, say {1, 2, …, q} also, it requires
computation of finding null space basis of p different N × matrices.
Let M
)
( 1 jMMM −−− K
1 = M 2 = … = M q and treat the complexity of performing interference nulling in
HGLST with arbitrary decoding order as benchmark. Then at the j-th decoding level with
1 ≤ j ≤ q − 12, the complexity in DGLST with arbitrary decoding order and HGLST with
optimal decoding order is q and q − j − 1 times higher, respectively. As the complexity of
computing the squared Euclidean norm in HGLST with optimal decoding order can be
ignored compared to that of finding null space basis, we see that before doing the ML
decoder process for each component STC, DGLST has the highest complexity and
HGLST with arbitrary decoding order has the lowest complexity. Once the decoding
order and the nulling matrices are computed, the complexity of the following decoding
process is the same for all of them.
Fig. 5.3 provides the FER performance of the optimal ordered decoding in the (4,4)
and (8,8) HGLST systems. We first observe that with optimal ordered decoding the FER
performance of HGLST is enhanced by around 2dB and 3.5dB in the (4,4) and (8,8)
systems, respectively. Comparing Fig. 5.3 with Fig. 5.2, we then observe that HGLST
with optimal ordered decoding performs slightly better than HGLST with optimal power
allocation, but slightly worse than DGLST with optimal power allocation.
For reference, we also provide the performance of the (4,4) and (8,8) BLAST and
coded BLAST systems in Fig. 5.3. In BLAST systems, the data symbols are BPSK
modulated. In coded BLAST systems, each layer is encoded individually by the rate ½
convolutional code with constraint length K = 5. The coded bits are Gray-mapped to
QPSK modulation. And the soft-decision Viterbi decoder, instead of the symbol-by-
symbol detector, is used for interference cancellation. From the slopes of the performance
curves, it is seen that the diversity order of HGLST is higher than BLAST and it results in
significant performance improvement even under roughly the same complexity.
2 Interference nulling is not required at the last decoding level.
66
7 9 11 13 15 17 1910
−3
10−2
10−1
100
SNR [dB] per receive
Fra
me
Err
or R
ate
BLAST optimal order coded BLAST optimal oderHGLST arbitrary order HGLST optimal order
(a)
8 10 12 14 16 18 20
10−3
10−2
10−1
100
SNR [dB] per receive
Fra
me
Err
or R
ate
BLAST optimal order coded BLAST optimal oderHGLST arbitrary order HGLST optimal order
(b)
Fig. 5.3: Performance comparison of optimal ordered decoding in the (a) (4,4) and (b) (8,8)
GLST and BLAST systems.
67
5.4 Interleaved GLST with Hard Decision Iterative Decoding
The above serial decoding scheme has not taken full advantage of the receive antenna
diversity yet. Particularly, the receive antenna diversity order of group j’ decoded at the j-
th decoding level is only N – M + M1’ + … + Mj’. Therefore, we propose an iterative
decoding process that can efficiently enable all groups to achieve full receive antenna
diversity. To eliminate the effect of burst error propagation among groups during
iterations we insert different temporal interleavers on the codewords from different
groups. This is shown in Fig. 5.1(a) and called interleaved GLST. Each interleaver is
vector based with vector size equal to the spatial dimension of the relevant codeword.
This ensures that the rank property of STC in each group is not destroyed.
The iteration procedure is depicted in Fig. 5.4. The first iteration is the same as
described in Section 5.1.2 except that interleaving and deinterleaving must be done
accordingly. The serial decoding order shown in this figure is {1’, 2’, …, q’}, without
loss of generality, for both optimal and arbitrary (or pre-determined) decoding order. At
the subsequent iterations, since all groups have been decoded already, interference
nulling is no longer needed. Instead, interference cancellation is performed for all groups.
Each iteration is composed of q levels, and each level is processed successively at the
same order as in the first iteration. At level j, we subtract the interference made by group
1’, …, (j-1)’, (j+1)’, …, q’ from the original received signals and re-decode group j’
using ML decoder. Once decoded, the new decision of group j’ is fed back to the next
level for decoding group (j+1)’. One iteration is finished when all the q levels are
finished. We can stop the iteration when the decoded signals are reliable.
An important note for the proposed iterative decoding algorithm is that, all the
interference cancellation is hard decision, and the ML space-time decoder becomes
Viterbi decoder or simple linear processing for space-time trellis codes and space-time
block codes, respectively. This guarantees low complexity compared with the
conventional turbo processing [38, 39, 40], where soft interference cancellation and
maximum a posteriori probability (MAP) decoder were used.
68
IC
P/SOutput
IN
adder
ST Dec-Enc 1’
IC
ST Dec-Enc 2’
IC
ST Dec-Enc q’
ST Dec-Enc 1’
IC
adderST Dec-Enc 2’
adder
ST Dec-Enc q’
adder
ST Dec-Enc 1’
adder
ST Dec-Enc 2’
adder
ST Dec-Enc q’
IC
IC
IC
IC
IC
B
C
B
C
B
C
C
1st iteration 2nd iteration kth iteration
IN
(a)
deint j’ STC j’ Decoder STC j’ Re-coder int j’ r C
B
(b)
Fig. 5.4: Iterative decoding of interleaved GLST (a) main block diagram, (b) sub-block diagram
for the “ST Dec-Enc j’” component.
To evaluate the results of the interleaved GLST with iterative decoding we first take
the HGLST version with optimal decoding order as an example to do simulation. All the
interleavers in our simulation are pseudo-random with depth equal to the frame length.
The results are shown in Fig. 5.5. We consider the (4, 4) HGLST system shown in Fig.
5.5(a) first. It is observed that the iterative decoding with interleaver dramatically
improves the performance. With three iterations it achieves 2.5 dB gains at 0.01 FER
when compared with the required SNR using the optimal ordered decoding found in Fig.
5.3(a). To study the role of interleaver we also provide the simulation results of the
iterative decoding without interleaver. We see that the interleaver has no gain at the first
iteration but brings significant gain at the subsequent iterations. As discussed above, the
first iteration procedure is the same as the basic serial decoding. Hence the performance
is limited by the group decoded first with the lowest diversity gain. As interleaver is
useless for the first group, it thus has little effect on the first iteration. However, from the
69
second iteration on, interleaver helps significantly in spreading out burst errors among
groups. It is seen that, with three iterations, the loss due to imperfect interference
cancellation (IC) with interleaver is only 0.6 dB, while the loss due to severe error
propagation without interleaver is 1.6 dB at 0.01 FER. Little gain can be achieved by
more than three iterations. Similar performance comparison for the (8, 8) HGLST system
is shown in Fig. 5.5(b). The improvement achieved by the iterative decoding with
interleaver is even larger. With four iterations, it provides a gain of 5.6 dB over the
optimal ordered decoding at 0.01 FER when compared with Fig. 5.4(b). The loss due to
imperfect IC with interleaver is now only 0.25 dB, while the loss due to severe error
propagation without interleaver is more than 2 dB at 0.01 FER. The gain achieved after
four iterations is very small.
We also provide the performance of interleaved DGLST with iterative decoding in
Fig. 5.6. We can see that with the same iterations HGLST is slightly inferior to DGLST
at low SNR but performs better at high SNR.
5.5 Summary
In this chapter, generalized layered space-time coding, a combination of BLAST and
STC for MIMO systems was introduced. In particular HGLST and DGLST were
classified according to different mappings from signals to transmit antennas. The basic
decoding of GLST is a certain ordered serial processing that combines group interference
nulling and group interference cancellation techniques. For the decoding with order pre-
determined and independent of the channel, optimal power allocation was derived to
enhance the system performance without increasing the complexity. This approach is
suitable for all GLST systems. For HGLST systems without power allocation, we also
discussed the optimal decoding order. To fully utilize the advantage of receive antenna
diversity, interleaved GLST with iterative decoding was further proposed. This iterative
decoding applies hard interference cancellation and ML decoder for component codes,
thus has much lower complexity than conventional turbo processing which requires soft
interference cancellation and MAP decoder.
70
4 5 6 7 8 9 1010
−3
10−2
10−1
100
SNR [dB] per receive
Fra
me
Err
or R
ate
1st ite w/ int2nd ite w/ int3rd ite w/ int1st ite w/o int2nd ite w/o int3rd ite w/o intPerfect IC bound
(a)
4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9
10−3
10−2
10−1
100
SNR [dB] per receive
Fra
me
Err
or R
ate
1st ite w/ int2nd ite w/ int4th ite w/ int1st ite w/o int2nd ite w/o int4th ite w/o intPerfect IC bound
(b)
Fig. 5.5: Performance of iterative decoding in the (a) (4,4) and (b) (8,8) HGLST systems.
71
2 3 4 5 6 7 8 9 10 1110
−3
10−2
10−1
100
SNR [dB] per receive
Fra
me
Err
or R
ate
(4,4) HGLST,3rd(4,4) DGLST,3rd(8,8) HGLST,4th(8,8) DGLST,4th
Fig. 5.6: Performance comparison of interleaved HGLST and interleaved DGLST with iterative
decoding in the (4,4) and (8,8) systems.
Based on the simulation results we conclude that GLST provides a promising solution
achieving high data rate and high quality communications in multiple-antenna systems,
especially when the number of antennas is large.
72
Chapter 6
DIFFERENTIAL SPACE-TIME BLOCK CODES
The schemes we proposed in Chapter 3 to Chapter 5 are all based on the assumption that
the channel knowledge is available at the receiver. This is reasonable when the channel
varies slowly enough that pilot signals can be used for channel estimation. In mobile
environments, however, precisely tracking the channel variation becomes difficult,
especially in a system with a large number of antenna elements. Hence, in this chapter
and the following one, we propose enhancements on the non-coherent space-time
techniques that can avoid channel estimation at the receiver, but with minimal loss in
error performance.
In this chapter we propose a differential space-time modulation scheme based on the
orthogonal space-time block codes (STBC) having square codeword matrices in [9] when
neither the transmitter nor the receiver has access to channel state information. The
decoupling property retained by coherent STBC can still be exploited, which results in
simple linear decoding complexity. This work is different from the previous work by
Tarokh and Jafarkhani in [58] in two aspects. First, the restriction of constant-amplitude
modulation (PSK) on information data symbols is relaxed such that more efficient multi-
amplitude constellations, such as QAM, can be applied. Simulation results show that a
gain of 3 ~ 4 dB can be achieved by using 16QAM rather than 16PSK modulation with
almost no increase on the detection complexity. Second, this differential scheme is not
only suitable for Alamouti’s block code with two transmit antennas, but also for any
orthogonal block codes with an arbitrary number of transmit antennas as long as the
codeword matrices are square.
Moreover, we compare the proposed differential orthogonal space-time block codes
(DSTBC) with existing diagonal cyclic group codes [57, 55]. The advantages are not only
in lower decoding complexity, but also in better performance. It is shown that a
significant improvement up to 6~8 dB can be achieved in a system with two transmit
antennas and one receive antenna over a slowly time-varying flat fading channel.
73
This chapter is organized as follows. In Section 6.1 we briefly introduce the system
model. In Section 6.2 we describe the proposed DSTBC in detail. The differentially non-
coherent decoding methods are derived in Section 6.3. Section 6.4 provides some
simulation results and Section 6.5 summarizes this chapter.
The following terminology is used throughout this chapter. For an M × M orthogonal
matrix A with AAH = AHA = a2I, we say the nonnegative scalar, a, is the amplitude of A.
The results in this chapter are published in [Tao7].
6.1 System Model
We consider a point-to-point wireless link with M transmit antennas and N receive
antennas as shown in Fig. 2.3. The channel is assumed to be flat Rayleigh fading and
remains constant within T symbol periods. The M × T transmitted signal matrix is
denoted by C and satisfies the total energy constraint
E[tr(CCH)] = T. (6.1)
Let R denote the N × T received signal matrix. It can be written as
WHCR += sρ (6.2)
where H is the N × T channel matrix with i.i.d. and zero-mean unit-variance complex
Gaussian distributed entries, W contains i.i.d complex Gaussian noise terms with zero
mean and variance σ2, and ρs is the average transmitted signal energy per symbol period.
The total transmitted SNR can be computed as 2σρρ s= .
Without knowing H and conditioned on C, R contains i.i.d. rows and each row is
complex Gaussian distributed with mean vector 0 and covariance matrix .
Then the probability density function (pdf) of R is given by
CCI Hsρσ +2
( )( )
( )CCI
RCCIRCR H
sNTN
HHs
pρσπ
ρσ
+
+−
=
−
2
12
det
trexp| . (6.3)
One of the interesting observations from (6.3) is that P(R|C) will not change if we left
multiply C by any arbitrary M × M unitary matrix Φ. In other words, C and ΦC are
indistinguishable for non-coherent detection.
74
6.2 Differential Encoding
6.2.1 Data Matrix
In general, an orthogonal space-time block code is described by a set of M × P codeword
matrices, and can be viewed as a mapping from a sequence of modulated data symbols ci,
i = 1, …, K, to a codeword C which is to be transmitted by M antennas over P symbol
periods. Each element of C is a linear combination of c1, …, cK and their conjugates. The
rate for the code is defined as PK . The primary property of a STBC is its orthogonality
M
K
ii
H c ICC ∑=
=1
2 .
In this thesis, we concentrate on square STBCs, i.e., P = M. As we mentioned in Section
2.5.2, for rate-1 code with K = M the real-valued square codes exist for 2, 4, and 8 Tx
antennas, while the complex-valued square codes only exist for 2 Tx antennas (Alamout’s
scheme). For other code rate with K < M, complex-valued square codes can be
constructed.
We assume that each data symbol ci, i=1,…,K is taken from a constellation (e.g. PSK
or QAM, etc) with size q and unit average energy. Then the data matrix in our proposed
DSTBC is defined as V
CVK1= (6.4)
where the factor K1 is used to satisfy the energy constraint in (6.1). Now, or each V
we have
IVV 2aH = (6.5)
with ∑ == K
i icKa1
2/1 being the amplitude of V. If we consider the set of all possible
data matrices to form a constellation V, then V has size of L = qK and satisfies
MH
KH
K qqIVVVV
VV∑∑∈∈
==VV
11 .
75
Particularly, if ci is PSK modulated, a = 1 for all V and such V is called a
constellation with single amplitude. Otherwise, a will take some discrete values and V is
now called a constellation with multiple amplitudes. Note that this is the major difference
between our approach and traditional DUSTM which always has single amplitude. It
potentially allows us to increase the spectral efficiency by carrying information not only
on the orientation but also on the amplitude of a data matrix.
6.2.2 Transmitted Matrix
Similar to DPSK, an identity matrix which does not carry any information is sent by the
transmitter to initialize the transmission, i.e., C0 = IM. Thereafter, the data matrices are
differentially encoded and sent. The transmitted signal matrix at the kth time block is
given by
kkk VCC 1~
−= (6.6)
where Vk is the data matrix at time block k and 1~
−kC is the normalized version of Ck-1
defined as
11
11~
−−
− = kk
k aCC
with ak-1 being the amplitude of Ck-1. Note that the differential encoding function (6.6) is
a little bit different from the traditional equation shown in (2.7), where the amplitude of
Ck-1 is always equal to 1. In addition, the differential transmission scheme proposed in
[58] appears as a special case of our scheme for 2 transmit antennas and PSK modulated
data symbols.
6.3 Non-Coherent Decoding
Two consecutive received signal matrices are used to recover each data matrix at each
time block. We assume that the channel keeps unchanged within T = 2M symbol periods,
then the received matrices can be written as
111 −−− += kksk WHCR ρ (6.7)
76
kksk WHCR += ρ . (6.8)
We derive two non-coherent decoding methods, namely optimal differential decoder (in
the sense of maximum likelihood over two consecutive received signal matrices) and
near-optimal differential decoder.
6.3.1 Optimal Differential Decoder
We stack two consecutive received matrices to form a matrix with T = 2M rows
[ ] [ ] [ ] kkskkkkskkk WCHWWCCHRRR +=+== −−− ρρ 111 .
Then it is easy to show that
( ) MkkHkk aa ICC 22
1 += −
with ak being the amplitude of Ck, or equivalently Vk by (6.6). Substituting the above into
(6.3) yields
( ) ( )
( ) ( )( )
++−+++
=
=
−−−−
∈
∈
221
221122
12 lnmaxarg
|lnmaxargˆ
kksH
kkkkks
s
kkok
aaMNaaa
p
ρσρσ
ρVRR
CRV
V
V
V
V (6.9)
where the matrix formulae
( ) ( )BAIABI +=+ detdet
( ) ( ) 111111 DABDACBAABCDA −−−−−−− +−=+ 1
are employed. Hence the sequence of data symbols ci, i=1,…,K can be jointly decoded by
the inverse mapping from ( )okV̂ .
It is seen from (6.9) that the detection of Vk requires the knowledge of ak-1. This is
contrary to DUSTM where the detection of each data matrix is always independent of the
previously transmitted matrix. At first glance, one may expect that error propagation
would be introduced. However, by simulation in Section 6.3, we show that in the systems
using 16QAM modulated data symbols, there is virtually no error propagation.
77
6.3.2 Near-Optimal Differential Decoder
Substituting (6.6) into (6.8) and then applying (6.7) we can obtain
kkkkkkkk aa WVWVYR +−= −−−−
−− 1
111
11 .
Since the noise matrices at different time blocks are independent and have i.i.d entries,
we can rewrite the above equation as '22
1111 1 kkkkkkk aaa WVRR −
−−−− ++= (6.10)
where is an M × N equivalent noise matrix with entries i.i.d and complex Gaussian
distributed with zero mean and variance σ
'tW
21 ka
2. The insight from (6.10) is that a can
be treated as the known channel matrix for the system transmitting V
111 −
−− kk R
k with noise variance
. Ignoring the dependence of the noise variance on the transmitted signals
acquires the near-optimal differential decoder
21 ka −−+
( ) 21
11minargˆ VRRV
V −−−∈
−= kkknok aV
. (6.11)
Taking advantage of the structure of STBC, it has been shown in [9] that the data
symbols ci, i=1,…,K, can be decoupled and decoded individually from (6.11). As a result,
the near-optimal DD has linear complexity
6.3.3 Optimal DD versus Near-Optimal DD
In this subsection we will consider the constellations with single amplitude and multiple
amplitudes respectively. With single amplitude both ak-1 and ak in (6.9) and (6.11) are
equal to 1. In this case it can be easily verified that the two differential decoders become
equivalent. Moreover, the decoder of Vk is now independent of the previously decoded
matrices. With multiple amplitudes, the near-optimal DD is inferior to the optimal DD
since the variations in the noise variance are ignored. However, the former is much less
complex than the latter. Precisely the complexity of the optimal DD is equal to the size of
the constellation formed by data matrices, i.e. qK, while the complexity of the near-
optimal DD is only a linear function in K, i.e. K × q.
78
6.4 Simulation Results
In our simulation two channel models are assumed. One is quasi-static flat Rayleigh
fading, i.e. the channel is kept constant within one frame but changed randomly from
frame to frame. Each frame contains 1000 symbol periods besides an initial time block.
The other is time-varying flat Rayleigh fading with autocorrelation ( nTfJ sd )π20
( )⋅0
, where fd
is the maximum Doppler frequency, Ts is the sample period and is the zeroth order
Bessel function of the first kind. The parameter f
J
dTs is set to 0.002.
The two complex-valued square codes G2 with rate 1 and H4 with rate ¾ are taken for
example to demonstrate the performance of our proposed scheme. Their codeword
matrices are given in (2.5) and (2.6), respectively.
Fig. 6.1 and Fig. 6.2 compare the performance of the two non-coherent decoders in
terms of symbol error probability under the quasi-static channel model. The data symbols
are 16QAM, thus the transmission rates are 4 bit/s/Hz and 3 bit/s/Hz for G2 and H4
respectively. It can be seen that the loss due to the near-optimal DD is small compared
with the optimal DD. Hence, the near-optimal DD is selected as the default decoder for
later simulations. Compared with the results of the coherent decoding, the non-coherent
decoding achieves the same order of antenna diversity but with a SNR degradation of 3~4
dB as expected. In these two figures, we also show the performance of the “genied”
optimal DD to indicate the error propagation. “Genied” means that when we use the
proposed optimal DD to recover each data matrix, the amplitude of the past signal matrix
is assumed perfectly known. It is observed that the result of the “genied” optimal DD is
almost the same as the optimal DD, which implies no error propagation.
From Fig. 6.3 to Fig. 6.5 we illustrate the performance comparison in terms of block
error probability between the constellations based on DSTBC with single amplitude
(PSK) and multiple amplitudes (QAM) and the cyclic group constellations suggested for
DUSTM under the time-varying channel model. The block error is defined as the
codeword error. The cyclic group code shown in Fig. 6.3 from M = 2, R = 4 bit/s/Hz is
given in [57, Table I]. It is observed that our proposed DSTBC has significant advantage
over the cyclic group code. The improvement achieved by the constellation with multiple
amplitudes (16QAM) is up to 6~8 dB. Moreover for the same code G2, multiple-
79
amplitude (16QAM) outperforms single-amplitude (16PSK) by 2~3 dB. The cyclic group
code for M = 4, R = 3 bit/s/Hz is not available yet. We only compare the code H4 with
multiple amplitudes (16QAM) and with single amplitude (16PSK) as shown in Fig. 6.4. It
is seen that the former outperforms the latter by 3~4 dB. In Fig. 6.5, the cyclic group code
for M=4, N=1, R=1.98 bit/s/Hz is given in [55, Table I]. The code H4 with 6PSK
modulated data symbols has almost the same transmission rate which is R=1.94 bit/s/Hz
and achieves 1~2 dB improvement.
10 15 20 25 30 35
10−4
10−3
10−2
10−1
100
SNR [dB]
Sym
bol E
rror
Rat
e
Near Optimal DDOptimal DD"Genied" Optimal DDCoherent Decoder
Fig. 6.1: Performance of differential decoding and coherent decoding for G2 with 16QAM data
symbols at M=2, N=1, and R = 4 bit/s/Hz
80
Notice that for the constellations shown in Fig. 6.3 the decoding complexity of the
cyclic group code with differential decoder is equal to the constellation size, i.e.
L=2RM=256, while the complexity of the code G2 with 16QAM modulated data symbols
and with the proposed near-optimal DD is only 2×16=32. For the constellations shown in
Fig. 6.5, the complexities of the cyclic group code with differential decoder and the code
H4 with 6PSK modulated data symbols and with the near-optimal DD (here the near-
optimal DD is equivalent to the optimal DD) are 240 and 18 respectively.
12 14 16 18 20 22 24 26 28 3010
−5
10−4
10−3
10−2
10−1
100
SNR [dB]
Sym
bol E
rror
Rat
e
Near Optimal DDOptimal DD"Genied" Optimal DDCoherent Decoder
Fig. 6.2: Performance of differential decoding and coherent decoding for H4 with 16QAM data
symbols at M = 4, N = 1, and R = 3 bit/s/Hz.
81
10 15 20 25 30 35 4010
−4
10−3
10−2
10−1
100
SNR [dB]
Blo
ck E
rror
Rat
e
Cyclic Group CodeG2 16PSKG2 16QAM
Fig. 6.3: Performance of differential decoder for G2 with 16PSK and 16QAM data symbols and
cyclic group code at M = 2, N = 1 and R = 4 bit/s/Hz.
82
14 16 18 20 22 24 26 28 30 32 3410
−4
10−3
10−2
10−1
100
SNR [dB]
Blo
ck E
rror
Rat
e
H4 16PSKH4 16QAM
Fig. 6.4: Performance of differential decoder for H4 with 16PSK and 16QAM data symbols at M
= 4, N = 1, and R = 3 bit/s/Hz.
83
10 12 14 16 18 20 22 24 26 2810
−4
10−3
10−2
10−1
100
SNR [dB]
Blo
ck E
rror
Rat
e
Cyclic Group CodeH4 6PSK
Fig. 6.5. Performance of differential decoder for H4 with 6PSK data symbols and cyclic group
code at M = 4, N = 1 and R ≈ 2 bit/s/Hz.
6.5 Summary
We presented a differential modulation scheme based on the square STBC over flat
fading channels when neither the transmitter nor the receiver has knowledge of CSI. The
reason why such techniques works and achieves full antenna diversity is that the square
STBCs offer orthogonal property in both spatial and temporal dimensions. Compared
with previous work in [58], our main contributions are as follows. First, the restriction of
PSK modulation for data symbols is relaxed. Second, we extended the DSTBC design
from two transmit antennas to any number of transmitted antennas where square STBC
exists. The advantages of our scheme over cyclic group codes lie in two aspects too,
significant improvement in error performance and considerably low complexity in
84
differential decoding. However, there are also two limitations for our scheme. Since the
constellation formed by data matrices does not form groups, the transmitter needs to
perform matrix multiplication at each time block and the constellation formed by the
transmitted matrices is largely expanded, thus the complexity at the transmitter side is
increased. The second limitation is that, there are only a few square STBCs that have
been designed.
85
Chapter 7
TRELLIS-CODED DIFFERENTIAL UNITARY
SPACE-TIME MODULATION
In Chapter 6, we proposed differential space-time block codes which can be classified
into the general category of differential space-time modulation scheme. However, this
scheme cannot be designed for an arbitrary number of transmitted antennas due to
limitation of STBC itself. Similarly, the design of differential unitary space-time
modulation with good performance and low complexity is still an issue. Notice that all
the differential space-time modulation schemes are actually only non-coherent
modulation schemes over the space-time grid. From the perspective of information
theory, it is necessary to apply channel coding in front of DUSTM to further approach the
channel capacity and hence improve the system performance. In this chapter, we propose
a new trellis coding scheme based on DUSTM, namely trellis-coded DUSTM). This is a
combined trellis coding and space-time modulation scheme, similar to the conventional
trellis-coded modulation (TCM) in single-antenna systems. The advantage of this
combination is that carefully designed trellis codes can effectively help to increase the
minimum distance metric of DUSTM. It results in coding gain and possibly time
diversity gain if an interleaver (matrix-wise) is applied. We thoroughly study the
performance measures and trellis code design rules for systems with either an ideal
interleaver or no interleaver. Several code examples that are based on diagonal cyclic
group constellations [57] and offer a good tradeoff between the coding advantage and
trellis complexity are provided. Extensions to trellis-coded differential space-time block
codes are also discussed. Therein, it is shown that the inherent orthogonality allows to
simply the trellis encoding and decoding and that the conventional well-developed TCM
can be directly applied.
The rest of this chapter is organized as follows. Before we introduce the proposed
TC-DUSTM, some necessary background knowledge on DUSTM is reviewed in Section
86
7.1. In Section 7.2, we present the system model of TC-DUSTM, and derive the
differentially non-coherent decision metrics. Then we propose the trellis code design
criteria. Section 7.3 describes code construction, as well as some code examples. Some
simulation results are shown in Section 7.4. Extensions to the trellis-coded differential
STBC are discussed in Section 7.5. Finally, Section 7.6 summarizes this chapter.
The results in this chapter are published in [Tao1, Tao6].
7.1 Background on DUSTM
The same channel model is considered as in Chapter 6. The encoding process of DUSTM
can be found in Section 2.5.4.
The non-coherent demodulation is based on two consecutive observations Rk-1 and
Rk. We group them into an N × 2M matrix as [ kkk RRR 1−= ] . Similarly we let
[ kkk CCC 1−= ] . It is assumed that the channel is constant within two consecutive time
blocks. Then, by applying (6.3) and using the unitary property of kC , the pdf of kR can
be simplified to
( )[ ]
( )( ) ,|
2
2tr1exp
|
)(
222
)()(
22
klk
MNs
MNMN
HkklH
kls
sk
kk
p
p
VR
RVIV
IIR
CR
=+
+−−
=ρσσπ
ρσρ
σ (7.1)
which is independent of Ck-1. Hence the decision metric of the M × M data matrix Vl(k)
based on the log function of (7.1) can be written, after ignoring irrelevant terms, as
( ) [ ]
{ }2
)(1
or
1)(
or
)()(
)(
trRe
tr,
klkk
kHkkl
HkklH
klkklkm
VRR
RRV
RVIV
IRVR
−
−
−−=
=
=
(7.2)
where Re(⋅) denotes the real part of a complex value. Notice that average transmitted
signal energy ρs and the variance σ2 are not present in the decision metric (7.1). That
implies the receiver does not need to estimate ρs and σ2. Later on, similar observations
87
will be made in the proposed TC-DUSTM. It can also be seen from (7.2) that the
implementation of the decision metric can be either in quadratic form, correlation form,
or minimum distance form. These are all equivalent.
At high SNR, the system performance is determined by the so-called diversity
product ζ of the signal constellation [57]
( ) MllLll
/1''0
detmin21 VV −=
<<≤ζ . (7.3)
Full transmit antenna diversity can be achieved as long as ζ > 0, and a larger ζ results in
better performance. As will be demonstrated in Section 7.2, it is not physically
straightforward to link ζ with the design of TC-DUSTM. Hence, in this thesis, we
reformulate ζ as a new relevant performance parameter, called the determinant distance,
denoted as D. Between any pair of elements Vl and Vl’ in the constellation, we have3
( ) ( ) ( )[ ] ( ) Mll
Mll
HllllD /2
'/1
'''2 detdet, VVVVVVVV −=−−≡ . (7.4)
Indeed, maximizing the minimum D is equivalent to maximizing ζ. Moreover when M =
1,
( )'22''
2 ,),( llllll vvdvvvvD =−= .
Consequently, the determinant distance can be viewed as a novel generalization of
Euclidean distance for signals in matrix form.
7.2 System Model and Performance Measure
7.2.1 System Model
The simplified transmission system diagram under investigation is depicted in Fig. 7.1.
For convenience, we break the trellis-coded unitary space-time modulation (TC-USTM)
into two parts: trellis encoder and USTM mapper. A sequence of information binary bits
b is passed through a rate m / n trellis encoder to generate an encoded bit stream c. Then,
the encoded bit stream c is divided into groups of n bits and each group is mapped at each
3 Based on matrix equations, the squared determinant distance between Vl and Vl’ is also the geometric mean of the eigenvalues of (Vl -Vl’)H(Vl -Vl’).
88
time block k into an element Vl(k) selected from a unitary space-time modulated
constellation with size L = 2n, according to a certain set-partition method. Hence each
transition branch in the trellis corresponds to one coded signal matrix. This is the
combined TC-USTM. Following the encoder, the coded matrix sequence {Vl(k)} is
reordered by a matrix-wise interleaver as {Vl(k’)}. The purpose of the matrix-wise
interleaver is to break the correlation of the fading that affects adjacent coded signal
matrices. The reordered coded matrix sequence from the interleaver is then differentially
encoded as {Ck’} and transmitted by M transmit antennas.
Trellis Encoder
USTM Mapper
Vl(k’) b Matrix-Wise Interleaver
Ck’ = Ck’-1Vl(k’) Ck’
TC-USTM
c Vl(k)
Fig. 7.1: Transmission system diagram of trellis-coded differential unitary space-time modulation
With respect to the interleaver design, we consider two extreme cases. 1) Ideal
interleaver: we assume the system can allow infinite delay such that the length of the
interleaver can be made sufficiently large, then the channel seen by different coded signal
matrices can be assumed to be statistically independent. 2) No interleaver: we assume the
system is delay sensitive and interleavers cannot be used. In the following we are going
to show the decision metrics and design criteria for these two cases respectively.
7.2.2 Ideal Interleaver
At the receiver, the received signal matrix sequence is partitioned
into groups of two blocks in such a way that two adjacent groups are overlapped by one
time block. The overlapping is necessary since the last block in previous group must be
used as a reference in current group for the differential decoding to be applied to each
trellis branch. Note that, due to overlapping, the noise terms in different groups are not
independent. The partitioned sequence is then deinterleaved group by group. The k
[ K10 RRR = ]
th
deinterleaved group and the associated transmitted signal matrices are denoted by the
89
same notations as those in Section 7.1, i.e., kR and kC . Therefore, under the assumption
of ideal interleaver/deinterleaver, both the channel coefficients and the noise terms
contained in different kR are made independent. Hence, given the coded matrix
sequence [ ]K)1( ll VVV = )2( , kR is independent of each other and the joint pdf of R
is simply given by
( ) =p | VR
( ) = pm )|(ln, VRVR
( ))(, klk VR
( )
( ) ( )∏∏ =k
klkk
kk pp )(|| VRCR
where the second equality holds because of (7.1). Hence the decision metric of selecting
the maximum-likelihood (ML) path through the trellis is
( )∑∑ ==k
klkklkk
mp )()( ,)|(ln VRVR (7.5)
where m is the corresponding branch metric. It can be written as any one of
the forms in (7.2). Thus the ML decoding can be efficiently implemented by the soft-
decision Viterbi algorithm.
Let us assume that the transmitted coded matrix sequence is V and the erroneously
decoded sequence is U. Then the pair-wise error probability (PWEP) is given by
( ) ( )
>−=→ ∑∑ VVRURUV 0,, )()(
kklk
kklk mmpP . (7.6)
Applying the Chernoff bound on (7.6) gives
( ) ( ) ( )( )[ ]{ }∏ −⋅−≤→k
klkklk mmEP VURVRUV )()( ,,exp λ (7.7)
where λ ≥ 0 is the Chernoff bound parameter to be optimized. It is observed that the kth
product term in (7.7) is exactly the Chernoff bound of
( ) ( )( ))()()( 0,, klklkklk mmP VVRUR >− ,
which corresponds to the PWEP of an uncoded system ( ))()( klklP UV → . As the explicit
expression of the Chernoff bound of ( ))()( klklP UV → is given by [56]
( ) ( ) ( )
−−+
+−''
2
214det ll
Hll
N VVVVIρ
ρ ,
we obtain the Chernoff bound of the PWEP (7.6) as
90
( ) ( ) ( ) (∏
−−+
+≤→ −
kklkl
Hklkl
NP )()()()(
2
214det UVUVIUV
ρρ ) . (7.8)
Assuming SNR is high enough, the PWEP can be further upper bounded by
( ) ( )∏∈
−−
−
≤→
η
ρk
Nklkl
vMN
P2
)()(det8
UVUV , (7.9)
where η is the set of all k that Vl(k) ≠ Ul(k) and v is the size of η, termed as the Hamming
distance between U and V. From (7.9) it is observed that the achieved diversity order is
the product of antenna diversity MN and the Hamming distance v. Hence, when the
Hamming distance increases by one, the order of diversity increases by MN. This
indicates the substantial advantage of using multiple antennas over single antenna. From
(7.9), it is also observed that the coding advantage is the vth root of the product of the
squared determinant distance
( )v
k
Mklkl
/1/2
)()(det
−∏
∈ ηUV = , ( )
v
kklklD
/1
)()(2
−∏
∈ ηUV
From the two observations, we reach the following design criteria.
Design criteria for the case of ideal interleaver:
• In order to achieve the diversity of order vMN in flat Rayleigh fading channels with
ideal matrix-wise intereleaver, the Hamming distance between any two distinct
coded matrix sequences must be at least v.
• The minimum product of the squared determinant distance between any pair of
coded matrix sequences having the minimum Hamming distance, denoted as
( )min2∏ D , must be maximized.
7.2.3 No Interleaver
In the case without an interleaver, we assume the channel is constant within a certain
period of time, say K + 1 time blocks4. Without the knowledge of CSI at the receiver, the
4 If K is large enough, i.e., slow fading, coherent schemes with chanel estimation may be preferred. Nevertheless, even in slow fading channels, the differential scheme is still reasonable choice in practical systems when the receiver does not want to estimate the channel.
91
pdf of the received matrix sequence R over K + 1 blocks conditioned on the coded matrix
sequence [ ])()2()1( Klll VVV L=V is [53]
( )( )
( )( )MNs
KMNMNK
HH
s
s
K
Kp
ρσσπ
ρσρ
σ
1
1tr1exp
|22)1(
22
++
++−−
=+
RCCIR
VR ,
where ( )[ ])()1()1( Klll VVVIC KL= is the differentially transmitted matrix
sequence. Then we can obtain the ML decision metric as
( ) ( )HHm CRRCVR tr, = . (7.10)
It is seen that (7.10) cannot be simplified to the sum of each branch metric ( ))(, klk VRm
defined in (7.2). A direct implementation of such ML sequence decoding becomes
infeasible. We are seeking a suboptimal decoder that is much less complicated and can be
easily implemented in practice. For each trellis branch, we take into consideration two
consecutive received blocks [ kkk RRR 1−= ] , and compute the total decision metric as
the sum of the branch metrics at all time blocks, i.e.
( ) ( )∑=
=K
kklkmm
1)(,,' VRVR . (7.11)
Therefore, the suboptimal decoder is to select the path through the trellis that has the
largest total decision metric, and can be implemented by the soft-decision Viterbi
algorithm in an efficient manner. Compared with the ML decision metric (7.10), the
suboptimal decision metric (7.11) is a simplified version of (7.10) in which all cross
terms between different Vl(k) are ignored. Nevertheless, the performance gap may be
narrowed by using an approach similar to the multiple-symbol differential detection in
[64] for DPSK in single-antenna systems.
Based on (7.11), we are unable to derive the PWEP due to the memory channel and
non-white Gaussian noise among different kR . However, because we know that DPSK
modulation can be viewed as PSK modulation but with a 3-dB higher noise power, it is
reasonable to expect that the trellis code structure designed for TC-PSK systems should
work well in TC-DPSK systems but with around 3 dB performance degradation. As
DUSTM is an extension of DPSK, we will also expect that a good trellis code designed
92
for TC-USTM with coherent decoding is also a good code in TC-DUSTM with non-
coherent decoding but with some performance degradation. Simulation results in Section
7.4 will help to verify this statement. Thus, in the following, we will derive the design
criterion of coherent TC-USTM when CSI is available at the reciver and then apply it to
the non-coherent case.
In TC-USTM without differential encoding, the transmitted signal matrix sequence C
is the same as the coded matrix sequence V. Thus the coherent ML decision metric of V
is clearly given by 2
)|( HVRHVR, sm ρ−−= . (7.12)
Based on (7.12), the Chernoff bound of the PWEP of transmitting V but detecting U is
given by [8]
( )( )N
HP−
−−+≤→ UVUVIUV
4det)( ρ (7.13)
Assuming ρ is sufficiently high, (7.13) is further upper bounded by
( )( )([ ] NHMN
P−
−
−−
≤→ UVUVUV det
4)( ρ )
)]
. (7.14)
From (7.14), we can see that the TC-USTM achieves a full antenna diversity order of MN
and a coding advantage of
( )( )([ MH /1detmin UVUV
UV−−=
≠δ . (7.15)
However (7.15) is not helpful enough in designing good trellis codes. Hence a simpler
but slightly looser criterion will be derived instead as follows.
Recall the Minkowski’s determinant inequality [74], which states that
( )[ ] ( ) ( ) MMM /1/1/1 detdetdet BABA +≥+
for M × M positive definite matrices A and B and equality holds if and only if B = cA for
c ≥ 0. Applying it in (7.15) yields a lower bound of δ
2free
1
/2
)()( )det(min DK
k
M
klkl ≡
−≥ ∑=
≠UV
UVδ , (7.16)
93
which is defined as the minimum free squared determinant distance . Hence an
alternative design criterion for TC-USTM is to maximize between any two distinct
coded matrix sequences.
2freeD
2freeD
Now let us come back to TC-DUSTM when CSI is unavailable at the receiver. We
arrive at the following criterion.
Design criterion for the case of no interleaver:
• In order to maximize coding advantage, the minimum free squared determinant
distance between any pair of distinct coded matrix sequences must be made as large
as possible.
Note that we design trellis codes that work well in the coherent case and expect them
to work well too in the non-coherent case only by adding an inner differential encoder at
the transmitter.
7.3 Code Construction
The design criteria of TC-DUSTM derived in the previous section for both ideal
interleaver case and no interleaver case are quite close to the well-known design criteria
of TCM in Rayleigh fading channels and AWGN channels, respectively [76]. Thus, we
propose relatively easy code construction methods in this section. We focus our attention
on the design of Ungerboeck type rate m / (m + 1) trellis codes. The coded matrix is
selected from a unitary space-time modulated constellation of size L = 2m + 1.
The first step before designing any code is to do set-partition for the employed
constellation based on its determinant distance profile. In the following we shall describe
a general set-partitioning method with idea brought over from conventional set-partitions
for PSK or QAM type constellations.
The set-partition starts from the minimum determinant distance. Consider the whole
set as the zeroth level. It is split into two subsets at the first level with inter-distance equal
to the minimum determinant distance. At the second level, each subset from the first level
is partitioned into two smaller subsets with inter-distance equal to the intra-distance of
the parent subset in the first level. This procedure continues for the succeeding levels and
94
stops at a desired level. The subsets after partition should try to exhibit similar distance
properties and provide a maximum of symmetry.
Of all existing signal constellations, the diagonal cyclic group constellations proposed
in [57] are the most systematic ones and can be designed for any number of transmit
antennas and any spectral efficiency. Thus we will take this class of constellations as an
example and briefly review it here. Other constellations can be easily extended.
The lth element Vl, with l = 0, 1, …, L−1, in the considered diagonal cyclic group
constellation is generated by , with ( ) ll 1VV =
( )LujLuj Mee ππ 221
1diag K=V
and u1, …, uM taken from the integer set {0, …, L−1} and chosen to maximize the
minimum determinant distance. Hence the constellation is fully specified by the
parameters u1, …, uM and size L. We denote such constellation by G = ([u1, …, uM]; L)
for simplicity. Later on, we will also simply use integer l, l = 0, 1, …, L−1, to represent
Vl in the constellation.
The group properties of diagonal cyclic group constellations characterize that the
determinant distance between any two elements l and l’ depends only on the absolute
value of l − l’ and, thus, allows efficient set-partitions. Two examples are provided to
demonstrate this procedure.
Example 1. Set-partition of G = ([1 3]; 8)
The determinant distance profile of G = ([1 3]; 8) is listed in Table 7.1.
Table 7.1: Determinant distance profile of G = ([1 3]; 8)
| l − l’ | D2 1, 3, 5, 7 1.4142
2, 6 2 4 4
The resulting set-partition is shown in Fig. 7.2
95
{0,1,2,…,7}
{0,2,4,6} {1,3,5,7}
{0,4} {2,6} {1,5} {3,7} Fig. 7.2: Set-partition of G = ([1 3]; 8).
Example 2. Set-partition of G = ([1 7]; 32)
The determinant distance profile of this example is given in Table 7.2.
Table 7.2: Determinant distance profile of G = ([1 7]; 32)
| l − l’ | D2 | l − l’ | D2 1, 31 0.2487 6, 10, 22, 26 1.8478 9, 23 0.3031 8, 24 2 5, 27 0.5474 7, 25 2.5254 4, 28 0.5858 15, 17 3.0772
2, 14, 18, 30 0.7654 11, 21 3.3758 3, 29 1.0240 12, 20 3.4142 13, 19 1.8044 16 4
The set-partition is illustrated in Fig. 7.3 {0,1,2,…,31}
{0,2,4,…,30} {1,3,5,…,31}
{0,2,8,10,16,18,24,26}
{0,8,16,24}
{4,6,12,14,20,22,28,30}
{1,3,9,11,17,19,25,27}
{5,7,13,15,21,23,29,31}
{2,10,18,26}
{4,12,20,28}
{6,14,22,30}
{1,9,17,25}
{3,11,19,27}
{5,13,21,29}
{7,15,23,31}
{0, 16}
{8, 24}
{2, 18}
{10, 26}
{4, 20}
{12,28}
{6, 22}
{14,30}
{1, 17}
{9, 25}
{3, 19}
{11,27}
{5, 21}
{13, 29}
{7, 23}
{15, 31}
Fig. 7.3: Set-partition of G = ([1 7]; 32).
From the above two examples, we observe that all subsets in the same partition level
have the same size and identical intra-distance. Such partition is called a fair partition.
For other constellations fair partition may not be guaranteed.
Using the two design criteria derived in Section 7.2, we propose two corresponding
design rule sets in the following subsections.
96
7.3.1 Ideal Interleaver
Set the target minimum Hamming distance to be v. We have the following design rules:
• All elements in the constellation should be equally probable.
• The minimum number of trellis states is 2m (v - 1). Parallel transitions are not
allowed.
• The 2m branches departing from a common state or converging to a same state
must be assigned with elements from one subset at the first level of set
partitioning.
According to the above design rules, we provide three code examples based on the
diagonal cyclic group constellations with two transmit antennas. The code parameters are
listed in Table 7.3, where R denotes the spectral efficiency in terms of information
bit/s/Hz. The trellis diagrams are illustrated in Fig. 7.4. All the three codes can achieve
the minimum Hamming distance of 2.
Table 7.3: Parameters of the codes designed for ideal interleaver case
code R [bit/s/Hz] code rate state number G v ( )min2∏ D
(a) 1 2 / 3 4 ([1 3]; 8) 2 2.8284
(b) 1 2 / 3 8 ([1 3]; 8) 2 8
(c) 2 4 / 5 16 ([1 7]; 32) 2 0.232
As the employed constellations in our examples have a diagonal structure and each
diagonal element is chosen from L-PSK type symbols, a layered coding strategy, i.e. the
coding is done on each diagonal element individually and independently, is also possible.
However, in order to ensure the same diversity advantage at the same spectral efficiency
as that of the proposed joint coding strategy, higher decoding complexity is required. To
demonstrate this point, we measure the decoding complexity using the average number of
branch metrics that need to be calculated through the soft-decision Viterbi decoder for
each information (info) bit. For the joint coding strategy with the minimum Hamming
distance be v, we have the decoding complexity as
97
state per branches2m × / =
statesmin
)1(2 −vm
bitsinfom
m
mv2 . (7.17)
For the layered coding strategy, the minimum Hamming distance for each component
trellis code has to be v’ = vM. Hence, the total decoding complexity is
state per branches
/2 Mm ×statesmin
)1(2
−vMMm
× / = layers
M bitsinfo
mm
M mv2⋅ (7.18)
which is M times larger than (7.17). Moreover, for the joint coding strategy, the coding
advantage within each matrix block is already maximized, as the matrix constellation
itself is optimal to maximize the minimum determinant distance. Hence it is relatively
easier to maximize the overall coding advantage compared with the layered coding
strategy. In fact, in the layered coding strategy, the M component codes are independent
and since they do not interfere with each other due to the diagonal transmission, the
system performance should be identical to that of the system using one component code
and transmitting with only one antenna continuously. Therefore, there is no point in using
the layered coding strategy.
7.3.2 No Interleaver
Following Ungerboeck’s rules we have the following criteria for assigning element to the
trellis branches.
• All elements should be used equally often.
• Parallel transitions should be assigned with elements from the subset with the
greatest intra-subset distance.
• The 2m transitions that diverge from a common state or remerge into a same state
must be assigned with elements from one subset at the first level of set
partitioning.
98
0 4 2 6
1 5 3 7
4 0 6 2
5 1 7 3
0 4 2 6
1 5 3 7
4 0 6 2
5 1 7 3
2 6 0 4
3 7 1 5
6 2 4 0
7 3 5 1 (a) (b)
0 16 8 24 2 18 10 26 4 20 12 28 6 22 14 30
1 17 9 25 3 19 11 27 5 21 13 29 7 23 15 31
16 0 24 8 18 2 26 10 20 4 28 12 22 6 30 14
17 1 25 9 19 3 27 11 21 5 29 13 23 7 31 15
8 24 0 16 10 26 2 18 12 28 4 20 14 30 6 22
24 8 16 0 26 10 18 2 28 12 20 4 30 14 22 6
2 18 10 26 0 16 8 24 6 22 14 30 4 20 12 28
18 2 26 10 16 0 24 8 22 6 30 14 20 4 28 12
10 26 2 18 8 24 0 16 14 30 6 22 12 28 4 20
26 10 18 2 24 8 16 0 30 14 22 6 28 12 20 4
9 25 1 17 11 27 3 19 13 29 5 21 15 31 7 23
25 9 17 1 27 11 19 3 29 13 21 5 31 15 23 7
3 19 11 27 1 17 9 25 7 23 15 31 5 21 13 29
19 3 27 11 17 1 25 9 23 7 31 15 21 5 29 13
11 27 3 19 9 25 1 17 15 31 7 23 13 29 5 21
27 11 19 3 25 9 17 1 31 15 23 7 29 13 21 5 (c)
Fig. 7.4: Trellis diagrams of the codes designed for ideal interleaver case.
When comparing this design rule set for the no interleaver case with the previous one
for the ideal interleaver case, it is found that they differ only in the second rule. In
particular, parallel transitions can be used in the no interleaver case but are not allowed in
the ideal interleaver case. This is because in the case with an ideal interleaver, the
minimum Hamming distance of a trellis code quantifies the order of time diversity and
should be made as large as possible. If parallel transitions exist, the minimum Hamming
distance becomes one. However in the case with no interleaver, the dominating parameter
is the minimum free squared determinant distance. Parallel transitions can help to
increase this value especially when there are only a few trellis states.
99
Two code examples based on the same signal constellation as before are given in this
subsection. The trellis diagrams are shown in Fig. 7.5. Both of the codes have parallel
transitions denoted by { , }. For code (a), the code rate is 2/3 with the spectral efficiency
R = 1 bit/s/Hz. It is not difficult to see that the error path within the parallel transitions
achieves the minimum free squared determinant distance of 4. In this case, the
equality in (7.16) holds. Hence, the actual coding advantage δ is also 4. With a coherent
receiver, its asymptotic coding gain over uncoded constellation G = ([1 1]; 4) with =
2 at the same spectral efficiency is expected to be 10 = 3.01 dB. For
code (b), the code rate is 4/5 with R = 2 bit/s/Hz. It is examined that the parameter
and δ are equal to 1.5999 and 1.7631, respectively, where the former is slightly smaller
than the latter. The asymptotic coding gain over uncoded constellation G = ([1 7]; 16)
with = 0.5858 at the same spectral efficiency is 4.78 dB when the receiver knows
CSI.
2freeD
2free10 D
2minD
2freeD
( 2min/log D )
2minD
{0,4} {2,6}
{1,5} {3,7}
{2,6} {0,4}
{3,7} {1,5}
{0,16} {2,18} {8,24} {10,26} {4,20} {6,22} {12,28} {14,30}
{1,17} {3,19} {9,25} {11,27} {5,21} {7,23} {13,29} {15,31}
{2,18} {0,16} {10,26} {8,24} {6,22} {4,20} {14,30} {12,28}
{3,19} {1,17} {11,27} {9,25} {7,23} {5,21} {15,31} {13,29}
{8,24} {10,26} {0,16} {2,18} {12,28} {14,30} {4,20} {6,22}
{10,26} {8,24} {2,18} {0,16} {14,30}{12,28} {6,22} {4,20}
{4,20} {6,22} {12,28} {14,30} {0,16} {2,18} {8,24} {10,26}
{6,22} {4,20} {14,30} {12,28} {2,18}{0,16} {10,26} {8,24}
{12,28} {14,30} {4,20} {6,22} {8,24} {10,26} {0,16} {2,18}
{14,30} {12,28} {6,22} {4,20} {10,26} {8,24} {2,18} {0,16}
{9,25} {11,27} {1,17} {3,19} {13,29} {15,31} {5,21} {7,21}
{11,27} {9,25} {3,19} {1,17} {15,31}{13,29} {7,23} {5,21}
{5,21} {7,23} {13,29} {15,31} {1,17} {3,19} {9,25} {11,27}
{7,23} {5,21} {15,31} {13,29} {3,19}{1,17} {11,27} {9,25}
{13,29} {15,31} {5,21} {7,23} {9,25} {11,27} {1,17} {3,19}
{15,31} {13,29} {7,23} {5,21} {11,27} {9,25} {3,19} {1,17} (a) (b)
Fig. 7.5: Trellis diagrams of the codes designed for no interleaver case: (a) rate 2/3 4-state with G
= ([1 3]; 8); (b) rate 4/5 16-state with G = ([1 7]; 32).
100
7.4 Simulation Results
In Fig. 7.6 we provide the bit error rate (BER) performance of the two codes with R = 1
bit/s/Hz designed for the ideal interleaver case shown in Table 7.3 and Fig. 7.4. The
channel is modeled as flat Rayleigh fading with an ideal interleaver/deinterleaver. The
channel coefficients are kept constant within each time block and are not known to either
the transmitter or the receiver. The derived ML differential decoder (7.5) is employed.
For comparison, the BER performance of uncoded constellation G = ([1 1]; 4) at the same
spectral efficiency is also provided. Gray mapping is used from the binary information
bits to the elements in the constellation. As can be seen in this figure, a significant
improvement is observed. In particular, at a BER of 10-3 and with one receive antenna,
the 4-state code achieves around 5 dB gains compared with the uncoded case, while the
8-state code is seen to provide more than 6.5 dB improvement over the uncoded case.
When the number of receive antennas is increased to two, the improvement is not so
large, but a gain of 3~4 dB at a BER of 10-3 can still be achieved by these two codes.
For the code with R = 1 bit/s/Hz designed for the no interleaver case shown in
Fig.7.5(a), we first demonstrate its performance with direct transmission and coherent
receiver, and then show its performance with differential transmission and suboptimal
non-coherent receiver (7.12). This is done to confirm our statement in Section 7.2.2 that a
good TC-USTM also works well in the non-coherent case if an inner differential encoder
is employed. The channel is assumed to be constant during a frame of 100 transmission
time intervals, and an additional two time intervals are required to send an initial signal
matrix in the differential transmission case. As the errors caused by fading tend to occur
in bursts, frame error rate (FER) rather than BER is simulated.
Fig. 7.7 illustrates the results of direct transmissions with a coherent receiver for this
rate 2/3 4-state code. The results of the corresponding uncoded transmission at the same
spectral efficiency are given for comparison. It can be observed in Fig. 7.7 that there is
no diversity improvement but a coding gain of 3 dB over a wide range of FER is
achieved by this code, as expected, compared with the uncoded case at both one and two
receive antennas.
101
2 4 6 8 10 12 14 16 18 20 22 24
10−5
10−4
10−3
10−2
10−1
SNR [dB]
Bit
Err
or R
ate
(2,1) uncoded G=([1 1];4)(2,1) 4−state G=([1 3];8)(2,1) 8−state G=([1 3];8)(2,2) uncoded G=([1 1];4)(2,2) 4−state G=([1 3];8)(2,2) 8−state G=([1 3];8)
Fig. 7.6: BER of rate 2/3 4-state and 8-state G = ([1 3]; 8) TC-DUSTM compared with uncoded
G = ([1 1]; 4) DUSTM, ML differential decoder and ideal interleaver, R = 1 bit/s/Hz.
102
0 2 4 6 8 10 12 14 16 1810
−3
10−2
10−1
100
SNR [dB]
Fra
me
Err
or R
ate
(2,1) uncoded G=([1 1];4)(2,1) 4−state G=([1 3];8)(2,2) uncoded G=([1 1];4)(2,2) 4−state G=([1 3];8)
Fig. 7.7: FER of rate 2/3 4-state G = ([1 3]; 8) TC-USTM compared with uncoded G = ([1 1]; 4)
USTM, ML coherent decoder and no interleaver, frame length = 100, R = 1 bit/s/Hz.
The results of differential transmission with the non-coherent receiver are illustrated
in Fig. 7.8. The simulated coding gain over the uncoded system is about 2.4 dB at both
one and two receive antennas. 0.6 dB gain is lost compared with direct transmission and
coherent receiver. A comparison with Fig. 7.7 shows that the differentially non-coherent
results have a performance loss of 3.5 dB.
In Fig. 7.8 this code is also compared with the traditional TC-DPSK designed for
single transmit antenna. For a fair comparison, both the spectral efficiency and the trellis
state number are set the same. The TC-DPSK for R = 1 bit/s/Hz is constructed using the
rate 1/2 4-state convolutional code with Gray mapped QPSK. The antenna diversity gain
enhancement is obvious by the use of TC-DUSTM. This yields a substantial
improvement in FER especially when there is only one receive antenna.
103
2 4 6 8 10 12 14 16 18 20 22 2410
−3
10−2
10−1
100
SNR [dB]
Fra
me
Err
or R
ate
(2,1) uncoded G=([1 1];4)(2,1) 4−state G=([1 3];8)(1,1) 4−state QPSK(2,2) uncoded G=([1 1];4)(2,2) 4−state G=([1 3];8)(1,2) 4−state QPSK
Fig. 7.8: FER of rate 2/3 4-state G = ([1 3]; 8) TC-DUSTM compared with uncoded G = ([1 1]; 4)
DUSTM and rate 1/2 4-state TC-DPSK, suboptimal differential decoder and no interleaver, frame
length = 102, R = 1 bit/s/Hz.
7.5 Extensions to Trellis-Coded Differential Space-Time Block Codes
All the code examples provided in this chapter are based on the diagonal cyclic group
constellations. Nevertheless, the proposed design methods are suitable for any unitary
space-time modulated constellations. One special class is the space-time block codes.
Due to its orthogonal structure, the trellis code design correspondingly has particular
rules. They are discussed in this section.
Let the coded matrix during the kth transmission block Vl(k) consist of D (D ≤ M)
independent PSK type symbols c(k−1)D+1, …, ckD, with unit energy. For any pair of distinct
coded matrices at block k, Vl(k) and Ul(k), we have [9]
104
( ) ( ) M
D
iiDkiDkklkl
Hklkl ec
DIUVUV ∑
=+−+− −=−−
1
2)1()1()()()()(
1 , (7.19)
where e(k−1)D+i , i = 1, …, D, are the symbols contained in Ul(k).
We first consider the case without an interleaver. The coding advantage δ in (7.15)
and the minimum free squared determinant distance in (19) are equal because of the
orthogonality shown in (7.19). We rewrite (7.15) as
2freeD
∑∑∑∈≠∈ =
+−+−≠−=−=
ςηδ
ttt
k
D
iiDkiDk ec
Dec
D2
1
2)1()1(
1min1minecUV
(7.20)
where c = {ct} and e = {et} are the symbol sequences contained in the coded matrix
sequences V and U, respectively, and ς is the set of all t that ct differs from et. Using
(7.20), the design criterion becomes maximizing the minimum free squared Euclidean
distance between any distinct symbol sequence pairs. Therefore, using the conventional
set-partition method, we simply map the binary coded bits to the symbols ct directly,
rather than the matrix Vl(k).
Next, we consider the case when an ideal interleaver/deinterleaver is used. We rewrite
the PWEP in (7.9) as
( ) ∏ ∑∈
−
=+−+−
−
−
≤→
ησρ
k
MND
iiDkiDk
vMN
ecD
P1
2)1()1(2
18
UV .
We further propose a layered structure for this system. The information bits are divided
into D substreams (or layers), each of which is then encoded using an individual
component trellis code with outputs mapped to symbols c(k-1)D+i at the kth time block for
the ith layer. All the component codes are set to be identical for simplicity and have the
minimum Hamming distance of v. Then, the whole system achieves vMN order of
combined antenna and time diversity. At the receiver, since every coded data symbol is
decoupled without the loss of any information, all the component codes can be decoded
individually. To better understand the benefit of having a layered structure, we compare
its decoding complexity with that of the general joint coding structure. The decoding
complexity is measured in the same way as introduced in Section 7.2.1, i.e., the average
number of branch metric calculations for each info bit through the soft-decision Viterbi
decoder. Let the spectral efficiency R be m / M bit/s/Hz and each component code has the
105
minimum trellis states to ensure the minimum Hamming distance is v. Then the
complexity of the layered structure is
Dmv /)1(2 −
stateper branches
/2 Dm × × layers
/ = statesmin
/)1(2 Dmv− Dbits info
mm
D Dmv /2 . (7.21)
However, at the same spectral efficiency and diversity order, the complexity of the joint
coding structure has to be
stateper branches2m × / =
statesmin
)1(2 −vm
bits infom
m
mv2 . (7.22)
Comparing (7.21) with (7.22), it can be observed that the proposed layered structure
significantly reduces the decoding complexity. As an example, let M = 2 (D = 2 in this
case), R = 4 bit/s/Hz and v = 2, the layered strategy has a complexity of 26 = 64 branch
metric calculations per info bit, while the joint coding strategy has that of 213 = 8192
branch metric calculations per info bit.
7.6 Summary
In this chapter we proposed trellis-coded differential unitary space-time modulation. No
channel estimation is required in this scheme. Therefore, it reduces the receiver
complexity and relaxes the design requirement on the pilot signals. We derived decision
metrics for differential decoding and proposed design criteria for the trellis codes. We
also described a general set-partition method for unitary space-time modulated
constellations based on their determinant distance profile, and code constructions for
Ungerboeck type trellis codes. Several code examples, based on diagonal cyclic group
constellations, were provided. Simulation results exhibited excellent performance with
computationally efficient non-coherent receivers.
We also considered the extensions to trellis-coded differential space-time block
codes. Therein, the decoupling property retained by space-time block codes allows
efficient and simple encoding and decoding structures.
106
Chapter 8
CONCLUSION AND FURTHER WORK
In this chapter we give some concluding remarks on this work and discuss several directions for further work.
8.1 Conclusion
In this thesis, we have studied the design of space-time transmission techniques mainly
from the coding perspective in multiple-antenna wireless communication systems. The
main contributions of this thesis are summarized as follows.
We proposed in Chapter 3 a set of improved design criteria for space-time trellis-
coded modulation taking care of different operating SNRs in quasi-static channel
environments. We essentially divided the SNR into three regions: high, moderate
and low SNR regions and proposed individual design criteria for each of them. It
was shown that the full-rank criterion is no longer necessary at both moderate
and low SNRs. It was also shown that the minimum Euclidean distance of a code
is the dominating factor at low SNR. Based on these improved design criteria, we
provided several space-time trellis code examples using two transmit antennas
via computer search, which are expected to work optimally at moderate SNR
regions. Simulation results demonstrated that they outperform the codes based on
the traditional criteria over a wide SNR range.
To avoid the prohibitively large time complexity of searching for good space-
time trellis codes with a larger number of transmit antennas and high modulation
level, we proposed in Chapter 4 diagonal block space-time coding, a systematic
code structure. The key of this approach is to separate the traditional design
method into two parts. It first encodes the information symbols by a one-
dimensional nonbinary block code, and then transmit the coded symbols
diagonally across the space-time grid. It was shown that this scheme can achieve
107
full transmit diversity and good coding advantage and, hence, is particularly
suitable for high SNR. This coding approach is suitable for an arbitrary number
of transmit antennas with arbitrary signal constellations. Some code examples
with 2 ~ 6 bit/s/Hz spectral efficiency and 2 ~ 6 transmit antennas were provided
and demonstrated excellent performance in computer simulation. Though it was
proposed for flat fading channels, this coding scheme can be easily extended to
frequency-selective fading to achieve the maximum possible combined antenna
and frequency diversity.
We presented in Chapter 5 the generalized layered space-time architecture which
provides a tradeoff between diversity order and transmission rate. This
architecture can be viewed as a compromise between traditional space-time
coding and traditional layered space-time architecture and both of them appear as
an extreme case of GLST. A number of important aspects, including signal-to-
antenna mapping, decoding order and power allocation were discussed.
Moreover, a low complexity iterative decoding algorithm was proposed that
efficiently exploits full receive diversity and, hence, dramatically improves the
overall system performance.
We proposed in Chapter 6 differential space-time block codes that do not require
channel knowledge for decoding at the receiver. This differential scheme allows
multiple-amplitude modulation for uncoded data symbols and can be designed
for any number of transmit antennas where orthogonal space-time block code
with square codeword matrices exists. We also did a comparison between this
scheme and DUSTM with diagonal cyclic group constellations, and significant
advantages have been demonstrated in both decoding complexity and error
performance.
We proposed in Chapter 7 the design of outer trellis channel coding in front of
the general differential unitary space-time modulation to further enhance the
system performance without requiring channel knowledge at the receiver. It
brought considerable coding gain and possibly diversity gain in a time-varying
fading environment if a matrix-wise interleaver is applied. It also demonstrated
108
significant performance improvement at the same complexity over traditional
coded differential scheme with only a single transmit antenna.
8.2 Further Work
There are several possible directions that can follow the work done in this thesis.
Rate-diversity space-time codes
Our previously studied generalized layered space-time architecture is just at the early
stage of a new research direction, that is, rate-diversity space-time codes. As a matter of
fact, the tradeoff between rate and diversity only exists when the modulation for the
transmitted symbols is fixed [8]. Recent efforts in [68, 69, 71, 73] revealed that, if (linear)
precoding or constellation rotation in the complex field is applied, the tradeoff
disappears. This is ensured by the information-theoretic study in [72]. Nevertheless, the
techniques on this “rate-diversity” topic are far from mature and there is still a lot of
research potential. For example, the peak-to-average power ratio, the resolution of
complex numbers, and the joint design and decoding with outer channel coding are the
issues.
Higher-rate trellis-coded differential space-time modulation
Similar to the traditional TCM concept, our previously proposed TC-DUSTM requires
expansion on the space-time modulation in order to achieve the same spectral efficiency
as uncoded systems. It may also be possible to achieve the same spectral efficiency
without expansion due to the additional degree of freedom introduced by multiple
transmit antennas. In fact, a superset technique [20, 21, 22] has already been applied in
the design of concatenated STBC with TCM to achieve similar objectives. It is, therefore,
expected that a similar technique can be developed for our TC-DUSTM to avoid
constellation expansion and, hence, be more spectrally efficient. The detailed superset
technique, the set-partition of unitary space-time modulation, and the design of trellis
codes need to be carefully addressed.
109
MIMO in future wireless networks
The MIMO systems we considered so far are point-to-point wireless links with flat fading
channels. When a multi-access or broadcasting channel is considered, MIMO can provide
interference suppression gain besides diversity and spectral efficiency gain. In that case,
the development of combined signal processing and space-time coding schemes to exploit
these gains is necessary. In addition, when a broadband wireless connection is
considered, it is necessary to extend this work to frequency-selective channels.
110
Bibliography
[1] G. J. Foschini and M. J. Gans, “On limits of wireless communications in a fading
environment when using multiple antennas,” Wireless Personal Communications,
vol.6, no.3, March 1998.
[2] I. E. Telatar, “Capacity of multi-antenna Guassian Channels,” European Trans. on
Telecom., vol.10, no.6, Nov/Dec 1999.
[3] G. J. Foschini, “Layered space-time architecture for wireless communication is a
fading environment when using multiple antennas,” Bell Labs Technical Journal,
vol.1, no.2, Autumn 1996.
[4] G. D. Golden, C. J. Foschini, R. A. Valenzuela and P. W. Wolniansky, “Detection
algorithm and initial laboratory results using V-BLAST space-time communication
architecture,” Electronics Letters, vol.35, no.1, Jan. 1999.
[5] P. W. Wolniansky, G. J. Foschini, G. D. Golden and R. A. Yalenzuela,“V-BLAST:
an architecture for realizing very high data rates over the rich-scattering wireless
channel,” Proc.1998 URSI International Symposium on Signals, Systems, and
Electronics, New York, 1998.
[6] G. J. Foschini, G. D. Golden, R. A. Valenzuela, and P. W. Wolniansky, “Simplified
processing for high spectral efficiency wireless communications employing multi-
element arrays,” IEEE Journal on Selected Areas in Comm., vol.17, no.11, Nov.
1999.
[7] J. C. Guey, M. P. Fitz, M. R. Bell and W. Y. Kuo, “Signal design for transmitter
diversity wireless communication systems over Rayleigh fading channels”, IEEE
Trans. on Comm., vol. 47, no. 4, pp. 527-537, April 1999.
[8] V. Tarokh, N. Seshadri, and A. R. Calderbank, “Space-time codes for high data rate
wireless communication: performance criterion and code construction,” IEEE
Trans. on Infor. Theory, vol.44, no.2, March 1998.
111
[9] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space-time block codes from
orthogonal designs,” IEEE Trans. on Infor. Theory, vol.45, no.5, July 1999.
[10] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space-time block coding for
wireless communications: performance results,” IEEE Journal on Selected Areas in
Commu., vol. 17, no. 3, March 1999.
[11] S. M. Alamouti, “A simple transmit diversity technique for wireless
communications,” IEEE Journal on Selected Areas in Comm., vol.16, no.8, Oct.
1998.
[12] G. Ganesan and P. Stoica, “Space-time diversity using orthogonal and amicable
orthogonal designs”, in Proc. Int. Conf. Acoustics, Speech, Signal Processing
(ICASSP), Istanbul, Turkey, 2000.
[13] G. Ganesan and P. Stoica, “Space-time block codes: A maximum SNR approach”,
IEEE Trans. on Infor. Theory, vol. 47, no. 4, pp. 1650-1656, May 2001.
[14] H. Jafarkhani, “A quasi-orthogonal space-time block code”, IEEE Trans. on
Commu., vol. 49, no. 1, pp. 1-4, Jan. 2001.
[15] O. Tirkkonen and A. Hottinen, “Complex space-time block codes for four Tx
antennas”, in Proc. GLOBECOM’00, 2000.
[16] O. Tirkkonen, A. Boariu, and A. Hottinen, “Minimal non-orthogonality rate 1 space-
time block codes for 3+ Tx antennas”, in Proc. IEEE 6th Symp. On Spread-Spectrum
Tech. & Application, New Jersey, USA, Sept. 6~8, 2000.
[17] W. Su and X.-G. Xia, “Quasi-orthogonal space-time block codes with full
diversity”, in Proc. IEEE Global Telecom. Conference (GLOBECOM’02), 2002.
[18] Il-Min Kim and V. Tarokh, “Variable-rate space-time block codes in M-ary PSK
systems”, IEEE Journal on Selected Areas in Comm., vol. 21, no. 3, pp. 362-373,
April 2003.
[19] V. Tarokh, A. Naguib, N. Seshadri, and A. R. Calderbank, “Combined array
processing and space-time coding,” IEEE Trans. on Infor. Theory, vol.45, no.4, May
1999.
112
[20] S. Siwamogsatam and M. Fitz, “Improved high rate space-time codes via
orthogonality and set partitioning”, in Proc. WCNC’02, 2002.
[21] S. Siwamogsatam and M. Fitz, “Improved high rate space-time TCM via
concatenation of expanded orthogonal block codes and MTCM”, in Proc ICC’02,
2002.
[22] H. Jafarkhani and N. Seshadri, “Super-orthogonal space-time trellis codes”, IEEE
Trans. on Infor. Theory, vol. 49, no. 4, pp. 937-950, April 2003.
[23] T. H. Liew, J. Pliquett, B. L. Yeap, L.-L. Yang, and L. Hanzo, “Comparative study
of space-time block codes and various concatenated turbo coding schemes”, in Proc.
IEEE International. Symp. on Personal, Indoor and Mobile Radio Comm.
(PIMRC’00), 2000.
[24] G. Bauch, “Concatenation of space-time block codes and ‘turbo’-TCM”, in Proc.
ICC’99, 1999.
[25] S. Baro, G. Bauch and A. Hansmann, “Improved codes for space-time trellis coded
modulation,” IEEE Comm. Letters, vol. 4, no. 1, Jan. 2000.
[26] Q. Yan and R. S. Blum, “Optimum space-time convolutional codes,” in Proc.
WCNC, Sept. 2000.
[27] Q. Yan and R. S. Blum, “Improved space-time convolutional codes for quasi-static
slow fading channels”, IEEE Trans. on Wireless Comm., vol. 1, no. 4, pp. 563-571,
Oct. 2002.
[28] X. Lin and R. S. Blum, “Systematic design of space-time codes employing multiple
trellis coded modulation”, IEEE Trans. on Comm., vol. 50, no. 4, pp. 608-615, April
2002.
[29] E. Biglieri, G. Taricco, and A. Tulino, “Performance of space-time codes for a large
number of antennas”, IEEE Trans. on Infor. Theory, vol. 48, no. 7, 1794-1803, July
2002.
[30] Z. Cheng, J. Yuan, and B. Vucetic, “An improved space-time trellis coded
modulation scheme on slow Rayleigh fading channels”, in Proc. ICC 2001.
113
[31] Z. Chen, B. S. Vucetic, J. Yuan and K. L. Lo, “Space-time trellis codes for 4-PSK
with three and four transmit antennas in quasi-static flat fading channels”, IEEE
Comm. Letters, vol. 6, no. 2, pp. 67-69, Feb. 2002.
[32] Z. Chen, B. S. Vucetic, K. L. Lo and J. Yuan, “Space-time trellis codes for 8-PSK
with two, three and four transmit antennas in quasi-static flat fading channels”,
Electronics Letters, vol. 38, no. 10, pp. 462-464, May 2002.
[33] Z. Chen, B. S. Vucetic, J. Yuan and K. L. Lo, “Space-time trellis codes with two,
three and four transmit antennas in quasi-static flat fading channels”, in Proc.
ICC’02, May 2002.
[34] H. El Gamal, A. R. Hammons, Jr., “On the design and performance of algebraic
space-time codes for BPSK and QPSK modulation”, IEEE Trans. on Comm., vol.
50, no. 6, pp. 907-913, June 2002.
[35] A. R. Hammons, Jr. and H. El Gamal, “On the theory of space-time codes for PSK
modulation”, IEEE Trans. on Infor. Theory, vol. 46, no. 2, pp. 524-542, March
2000.
[36] D. M. Ionescu, “New results on space-time code design criteria,” in Proc.
WCNC’99, Sept. 1999.
[37] R. van Nee, A. van Zelst and G. Awater, “Maximum likelihood decoding in a space
division multiplexing system,” in Proc. IEEE VTC 2000, Tokyo, Japan, May 2000.
[38] Da-shan Shiu, “Iterative decoding for layered space-time codes,” in Proc. IEEE ICC
2000, New Orleans, Louisiana, USA, June 2000.
[39] S. L. Ariyavisitakul, “Turbo space-time processing to improve wireless channel
capacity,” in Proc. IEEE ICC 2000, New Orleans, Louisiana, USA, June 2000.
[40] S. Baro, G. Bauch, A. Pavlic, and A. Semmler, “Improving BLAST performance
using space-time block codes and turbo decoding,” in Proc GLOBECOM’00, 2000.
[41] Da-shan Shiu and J. M. Kahn, “Layered space-time codes for wireless
communications using multiple transmit antennas,” in Proc. IEEE ICC ’99, 1999.
114
[42] B. Hassibi, “An efficient square-root algorithm for BLAST”, in Proc. Acoustics,
Speech, and Signal Processing (ICAPP’00), 2000.
[43] N. Seshadri and J. Winters, “Two signaling schemes for improving the error
performance of frequency-division duplex (FDD) transmission systems using
transmitter antenna diversity”, in Proc. VTC’93, 1993.
[44] A. Wittneben, “A new bandwidth efficient transmit antenna modulation diversity
scheme for linear digital modulation”, in Proc. ICC’ 93, 1993.
[45] J. H. Winters, “The diversity gain of transmit diversity in wireless systems with
Raleigh fading”, IEEE Trans. on Vehicular Technology, vol. 47, no. 1, pp. 119-123,
Feb. 1998.
[46] S. Li, X. Tao, W. Wang, P. Zhang, and C. Han, “Generalized delay diversity code: a
simple and powerful space-time coding scheme”, in Proc. International Conference
on Communication Technology, 2000.
[47] Z. Safar and K. J. R. Liu, “Systematic design of space-time trellis codes for diversity
and coding advantage”, EURASIP Journal on Applied Signal Processing, pp. 221-
235, March 2002.
[48] A. Matache, R. D. Wesel and J. Shi, “Trellis coding for diagonally layered space-
time systems”, in Proc. ICC’02, May 2002.
[49] G. Caire and G. Colvalope, “On space-time coding for quasi-static multiple-antenna
channels”, in Proc. Globecom’01, San Antonio TX, Nov. 2001.
[50] D. Gore, S. Sandhu and A. Paulraj, “Delay diversity code for frequency selective
channels”, Electronics Letters, vol. 37, no. 20, Sept. 2001.
[51] T. L. Marzetta and B. M. Hochwald, “Capacity of a mobile multiple-antenna
communication link in Rayleigh flat fading,” IEEE Trans. on Infor. Theory, vol. 45,
no. 1, Jan. 1999.
[52] L. Zheng and D. N. C. Tse, “Communication on the Grassmann manifold: A
geometric approach to the noncoherent multiple-antenna channel”, IEEE Tran. on
Infor. Theory, vol. 48, no. 2, pp. 359-383. Feb. 2002.
115
[53] B. M. Hochwald and T. L. Marzetta, “Unitary space-time modulation for multiple-
antenna communications in Rayleigh flat fading”, IEEE Trans. on Infor. Theory,
vol, 46, no. 2, pp. 543-564, March 2000.
[54] B. M. Hochwald,T. L. Marzetta, T. J. Richardson, W. Sweldens, and R. Urbanke,
“Systematic design of unitary space-time constellations”, IEEE Trans. on Infor.
Theory, vol. 46, no. 6, pp. 1962 -1973, Sept. 2000.
[55] B. Hassibi, B. W.Hochwald, A. Shokrollahi and W. Sweldens, “Representation
theory for high-rate multiple-antenna code design,” IEEE Trans. on Infor. Theory,
vol. 47, no. 6, pp. 2335-2367, Sept. 2001.
[56] B. L. Hughes, “Differential space-time modulation,” IEEE Trans. on Infor. Theory,
vol. 46, no. 7, Nov. 2000.
[57] B. M. Hochwald and W. Sweldens, “Differential unitary space-time modulation,”
IEEE Trans. on Comm., vol. 48, no. 12, Dec. 2000.
[58] V. Tarokh and H. Jafarkhani, “A differential detection scheme for transmit
diversity,” IEEE Journal on Selected Areas in Comm., vol. 18, no. 7, July 2000.
[59] G. Ganesan and P. Stoica, “Differential modulation using space-time block codes”,
IEEE Signal Processing Letters, vol. 9, no. 2, pp. 57-60, Feb. 2002.
[60] E. G. Larsson, P. Stoica and J. Li, “On maximum-likelihood detection and decoding
for space-time coding systems”, IEEE Trans. on Signal Processing, vol. 50, no. 4,
pp. 937-944, April 2002.
[61] J. Wang, M. P. Fitz, and K. Yao, “Differential unitary space-time modulation for a
large number of receive antennas”, in Proc. 36th Asilomar Conference on Signals,
Systems and Computers, 2002.
[62] X.-B. Liang and X.-G. Xia, “Unitary signal constellations for differential space-time
modulation with two transmit antennas: parametric codes, optimal designs, and
bounds”, IEEE Trans. on Infor. Theory, vol. 48, no. 8, pp. 2291-2322, Aug. 2002.
[63] X.-G. Xia, “Differentially en/decoded orthogonal space-time block codes with
APSK signals”, IEEE Comm. Letters, vol. 6, no. 4, pp. 150-152, April 2002.
116
[64] D. Divsalar and M. K. Simon, “Multiple-symbol differential detection of MPSK”,
IEEE Trans. on Comm., vol. 38, no. 3, pp. 300-308, March 1990.
[65] J. B. Anderson and S. Mohan, “Sequential decoding algorithms: A survey and cost
analysis”, IEEE Trans. on Comm., vol. 32, pp. 169-176, Feb. 1984.
[66] E. Viterbo, J. Boutros, “A universal lattice code decoder for fading channels”, IEEE
Trans. on Infor. Theory, vol. 45, no. 5, pp. 1639-1642, July 1999.
[67] O. Damen, A. Chkeif and J-C Belfiore, “Lattice code decoder for space-time codes”,
IEEE Commn. Letters, vol. 4, no. 5, pp. 161-163, May 2000.
[68] B. Hassibi and B. M. Hochwald, “High-rate codes that are linear in space and time”,
IEEE Trans. on Infor. Theory, vol. 48, no. 7, pp. 1804-1824, July 2002.
[69] R. W. Heath, Jr and A. J. Paulraj, “Linear dispersion codes for MIMO systems
based on frame theory”, IEEE Trans. on Signal Proecessing, vol. 50, no. 10, Oct.
2002.
[70] M. O. Damen, K. Abed-Meraim, and J-C. Belfiore, “Diagonal algebraic space-time
block codes”, IEEE Trans. on Infor. Theory, vol. 48, no. 3, pp. 628-636, March
2002.
[71] M. O. Damen, A. Tewfik, and J.-C. Belfiore, “A construction of a space-time code
based on number theory”, IEEE Trans. on Infor. Theory, vol. 48, no. 3, pp. 753-760,
March 2002.
[72] L. Zheng and D. N. C. Tse, “Divesity and multiplexing: a fundamental tradeoff in
multiple antenna channels”, IEEE Trans. on Infor. Theory, vol. 49, no. 5, pp. 1073-
1096, May 2003.
[73] H. El Gamal and M. O. Damen, “Universal space-time coding”, IEEE Trans. on
Infor. Theory, vol. 49, no. 5, pp. 1097-1119, May 2003.
[74] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press 1985.
[75] G. L. Nemhauser, A. H. G. Rinnooy Kan and M. J. Todd, Optimization, North-
Holland, Amsterdam, 1989.
117
[76] E. Biglieri, D. Divsalar, P. J. McLane and M. K. Simon, Introduction to trellis-
coded modulation with applications, Macmillan Publishing Company, 1991.
[77] T. S. Rappaport, Wireless Communications Principles & Practice, Prentice Hall,
Inc., a Simon & Schuster Company, 1996.
[78] J. G. Proakis, Digital Communications, Third Edition, McGraw-Hill, Inc., 1995.
118