ABSTRACT

ZHANG, HONGYUAN. MIMO Communications Systems: Antenna Selection and

Interference Mitigation. (Under the direction of Professor HUAIYU DAI).

Multiple-input multiple-output (MIMO) techniques have evolved as one of the key

enabling technologies to address the ever-increasing demand for high-speed wireless data
access on current and emerging wireless communication networks. In this dissertation, we
explore some important problems in point-to-point and multiuser MIMO systems, making
certain contributions to both theory and practice of this timely research topic
Antenna selection provides a new form of diversity in MIMO systems with low cost and
relatively small overhead increase. In the first part of this dissertation, we investigate this
technique, concerning (1) associated diversity property analysis in MIMO spatial
multiplexing systems with practical coding and decoding schemes; (2) fast algorithm design
with good tradeoff between performance and complexity, especially in correlated fading
channels; (3) practical implementation in commercial high-throughput wireless local area
networks.
The first two problems above are addressed via novel and effective geometric tools. These
tools are also extended to analyze some open problems in MIMO study, in particular the
diversity-multiplexing tradeoff in V-BLAST and SDMA systems employing ordered
successive interference cancellation (SIC).
While for some systems it is sufficient to consider single-user performance, many others
need to address the impairments due to the coexistence of multiple users, especially for
cellular systems with universal frequency reuse, which usually undergo severe interference in

the downlink. Conventional interference mitigation strategies are typically not effective in
high data rate and MIMO scenarios, especially for the downlink. Then in the latter part of
this dissertation, we focus on the study of interference mitigation in multiuser multicell
MIMO downlink, and investigate the potential of cooperative transmission among adjacent
base stations (BS) for effectively combating co-channel interference. Our study starts with a
quasi-synchronous model to obtain performance upper bounds, by which we also explore
some other advantages like channel rank/conditioning improvement and macro-diversity
protection. When considering a more practical channel model, in which signals from
different BSs in the downlink arrive asynchronously at each MS, we propose some novel
and effective pre-coding algorithms achieving different levels of tradeoffs between
interference mitigation and computational complexity. These algorithms are extended to the
scenarios where timing information cannot be perfectly acquired at the cooperative BSs.
In summary, we have tackled some open and interesting problems in MIMO study, in
particular, the diversity analysis for MIMO spatial multiplexing systems with antenna
selection and practical coding and decoding schemes, and the impact of ordering on the
performance of SIC (V-BLAST) receivers. The underlying geometric tools for these analyses
may find applications in other relevant fields as well. We have also endeavored to improve
the performance of MIMO systems in real operating scenarios and accelerate their
employment in future wireless networks, with contributions in fast and practical antenna
selection algorithms, and co-channel interference mitigation with base-station cooperation
explicitly considering the effect of signal asynchronism and timing errors.

MIMO Communications Systems: Antenna Selection and Interference Mitigation

by
Hongyuan Zhang
A dissertation submitted to the Graduate Faculty of
North Carolina State University
In partial fulfillment of the
Requirements for the degree of
Doctor of Philosophy
Electrical Engineering
Raleigh, NC
August, 2006

Approved by:

_________________________

_________________________

Dr. Huaiyu Dai

Dr. Alexandra Duel-Hallen

Chair of Advisory Committee

_________________________

_________________________

Dr. Brian L. Hughes

Dr. J. Keith Townsend

UMI Number: 3247105

UMI Microform 3247105

Copyright 2007 by ProQuest Information and Learning Company.
All rights reserved. This microform edition is protected against
unauthorized copying under Title 17, United States Code.

ProQuest Information and Learning Company

300 North Zeeb Road
P.O. Box 1346
Ann Arbor, MI 48106-1346

To my wife, my parents, and my friends......

ii

Biography
Hongyuan Zhang received his Bachelor of Science degree in Electronic Engineering in 1998 from
Tsinghua University, Beijing, China; and Master of Science degree in Electrical Engineering in
2001 from the Graduate School of Chinese Academy of Sciences, Beijing, China. In August 2001,
he started his work towards the PhD degree in Electrical Engineering at North Carolina State
University, under the guidance of Prof. Huaiyu Dai. His research topic is focused on the general
area of MIMO wireless communication systems including antenna selection and multiuser
interference mitigation technologies. From May 2005 to Dec. 2005, and from May 2006 to Aug.
2006, he worked as an intern at Mitsubishi Electric Research Labs (MERL), Cambridge, MA,
where he was dedicated in the research of base station cooperation for MIMO multiuser
interference mitigation, and made contributions to the emerging IEEE 802.11n wireless LAN
standard by proposing antenna selection training techniques. He has made publications or
submissions in the form of book chapter, major transactions, and major conference proceedings.
Nine patents were filed based on his research outcomes in MERL.

iii

Acknowledgements
First, I would like to express my deepest gratitude to my advisor, Dr. Huaiyu Dai, for his
guidance and continuous support through my PhD study. His wisdom, vision, and hard work
benefited me not only in enhancing my research skills, but also in the way I look at any problem
faced in my life. Working with him was really a great experience for me.
I am also grateful to the professors serving in my PhD committee: Dr. Brian Hughes, who
guided me in an interesting summer research project, and provided me insightful viewpoints in
my research on the diversity analysis for antenna selection and V-BLAST SIC receivers; Dr.
Alexandra Duel-Hallen and Dr. Keith Townsend, who provided valuable feedback on my
research, and useful knowledge in the general area of wireless communications, in the courses
they taught.
My gratitude is also given to my supervisors and colleagues at Mitsubishi Electric Research
Labs: Dr. Andreas Molisch, Dr. Neelesh Mehta, Dr. Daqing Gu (who is currently with NTT
Docomo Labs, Beijing, China), Dr. Koon Hoo Teo, Dr. Dong Wang (who is currently with
Philips Research North America, Briarcliff Manor, NY), and Mr. Yves-Paul Nakache. Without
the constructive discussions with them, it was not possible for me to accomplish the research at
MERL (in particular, Sections 2.4 and 4.4 in this dissertation). Special thanks should be given to
Dr. Jin Zhang, my manager at MERL, who helped me out in difficult times.
Finally, I appreciate the support from my family and my friends.

iv

Contents
List of Figures...............................................................................................................................viii
Abbreviations and Acronyms .......................................................................................................... x
Chapter 1 Introductions ................................................................................................................... 1
1.1 MIMO Communications Systems ......................................................................................... 1
1.2 Point-to-Point MIMO Systems: Antenna Selection............................................................... 2
1.3 Multiple User MIMO Systems: Interference Mitigation ....................................................... 4
1.4 Dissertation Outline ............................................................................................................... 7
Chapter 2 Antenna Selection in Point-to-Point MIMO ................................................................... 9
2.1 Introduction of Antenna Selection......................................................................................... 9
2.2 Performance Analysis of Antenna Selection: Diversity Order of MIMO-SM with Transmit
Antenna Selection - A Geometrical Approach .......................................................................... 12
2.2.1 Introduction................................................................................................................... 12
2.2.2 System Model and Problem Formulation ..................................................................... 14
2.2.3 Diversity Order and Diversity-Multiplexing Tradeoff when L = 2 ............................ 20
2.2.4 Extension to General L ................................................................................................. 38
2.2.5 Summary....................................................................................................................... 44
2.3 Antenna Selection Algorithm Designs: Fast Transmit Antenna Selection in Correlated
MIMO Channels ........................................................................................................................ 44
2.3.1 Introduction................................................................................................................... 44
2.3.2 System model and Problem Formulations .................................................................... 46
2.3.3 Gram-Schmidt (GS) Based Algorithm: A Geometric Analysis.................................... 49
2.3.4 The Gerschgorin Circles (G-circles) Based Algorithm................................................. 51
2.3.5 Numerical Results......................................................................................................... 56
2.3.6 Summary....................................................................................................................... 60
2.4 Implementing Antenna Selection in Realistic Wireless Networks ...................................... 61
2.4.1 Introduction................................................................................................................... 61
2.4.2 Introduction of 802.11n PHY and MAC....................................................................... 64
2.4.3 Mathematical Model for Antenna Selection ................................................................. 66
2.4.4 MAC-based AS Training Protocol................................................................................ 71
v

2.4.5 AS with RF Imbalance.................................................................................................. 74

2.4.6 Numerical Results......................................................................................................... 78
2.4.7 Summary....................................................................................................................... 82
2.5 Conclusions.......................................................................................................................... 82
Chapter 3........................................................................................................................................ 84
The Diversity-Multiplexing Tradeoff of Ordered SIC Receivers in MIMO Systems ................... 84
3.1 Introduction.......................................................................................................................... 84
3.2 System Model and Problem Formulation ............................................................................ 86
3.2.1 Point-to-Point MIMO Channels.................................................................................... 86
3.2.2 Multiple Access MIMO Channels ................................................................................ 92
3.3 Ordered V-BLAST SIC Receiver ........................................................................................ 95
3.4 Ordered SDMA SIC Receiver ........................................................................................... 100
3.5 Conclusions........................................................................................................................ 106
Chapter 4 Interference Mitigation in Multi-User MIMO ............................................................ 107
4.1 Introduction........................................................................................................................ 107
4.2 System Models................................................................................................................... 113
4.3 BS Cooperation with Synchronized Interference .............................................................. 121
4.3.1 Pre-coding Designs ..................................................................................................... 121
4.3.2 Other Advantages of Base Station Cooperation ......................................................... 127
4.3.3 Numerical Results....................................................................................................... 130
4.3.4 Summary..................................................................................................................... 132
4.4 BS Cooperation with Asynchronized Interference ............................................................ 133
4.4.1 Pre-coding Designs ..................................................................................................... 133
4.4.2 Generalization to Imperfect Timing-Advance Case ................................................... 140
4.4.3 Numerical Results....................................................................................................... 144
4.4.4 Summary..................................................................................................................... 149
4.5 Conclusions........................................................................................................................ 150
Chapter 5 Conclusions and Future Work..................................................................................... 152
5.1 Conclusions........................................................................................................................ 152
5.2 Achievements .................................................................................................................... 153
5.3 Future Work....................................................................................................................... 154
Appendix A Proof of Lemma 2.1 ................................................................................................ 165
vi

Appendix B Lemmas Used for Proving Theorem 2.1 ................................................................. 170

Appendix C The Derivation of (2.38) ......................................................................................... 177
Appendix D Lemma Used for Proving Proposition 2.5............................................................... 180
Appendix E Derivation of JWF Solution in (4.25) ...................................................................... 182
Appendix F Derivations of JWF and JLS Solutions in (4.35), (4.36), and (4.37) ....................... 183

vii

List of Figures
Figure 2-1 Antenna Selection System Model ........................................................11
Figure 2-2 The Illustration of (2.6) .........................................................................17
Figure 2-3 The Illustration of the Proof for Proposition 2.1: A Graph with Size n ..25
Figure 2-4 The Exponential Behaviors of the Outage Probabilities for the
NT = N R = 3, L = 2 Scenario...............................................................................30
Figure 2-5 BER Performance of the ZF-DF Receiver with NT = N R = 3, L = 2 .......37
Figure 2-6 The Exponential Behaviors of the Outage Probabilities for the
N T = N R = 4, L = 3 Scenario .................................................................................43
Figure 2-7 Two and three dimensional examples of Gram-Schmidt process .........50
Figure 2-8 An example of G-circles.......................................................................52
Figure 2-9 Capacity performances for well- and ill-conditioned channels .............56
Figure 2-10 Capacities of RT -based selections w.r.t. different .......................57
Figure 2-11 BER performances for well- and ill-conditioned channels ..................59
Figure 2-12 MIMO-OFDM PHY .............................................................................63
Figure 2-13 11n PHY Preamble ............................................................................65
Figure 2-14 Illustration of an AS Cycle ..................................................................66
Figure 2-15 Time Variation in WLAN Channel Model B...........................................72
Figure 2-16 Transmit AS Training Example ............................................................73
Figure 2-17 Transmit AS Calibration Process .........................................................75
Figure 2-18 Results of Channel B ..........................................................................79
Figure 2-19 Results of Channel E ..........................................................................80
Figure 2-20 Results of Channel B under RF Imbalance ..........................................81
Figure 3-1 Geometric Illustration of ZF-SIC Receiver with NT = 2 ........................91
Figure 3-2 Geometric Illustration of (3.20).............................................................97
Figure 3-3 Bounds on the Diversity Multiplexing Tradeoff for SDMA-SIC Detectors
........................................................................................................................105
Figure 4-1 A Simple BS Cooperation Scenario ...................................................115
Figure 4-2 Asynchronous Interference :Symbol Overlap Diagram ......................117
Figure 4-3 Simulations of BS Cooperation Shcemes in Rayleigh Fading Channels
........................................................................................................................131
viii

Figure 4-4 Simulations of BS Cooperation Shcemes in Rank Deficient Channels132

Figure 4-5 Comparison of Pre-coding Schemes for NT = N R = L = 2 ...................145
Figure 4-6 Comparison of Pre-coding Schemes for N T = 3, N R = L = 2 .................146
Figure 4-7 Comparison of Pre-coding Schemes for NT = N R = 3, L = 2 ...............147
Figure 4-8 The Effectiveness of CISVD in NT = 3, N R = L = 2 Scenario ...............148
Figure 4-9 JWF and JLS with Timing Inaccuracy in the NT = 3, N R = L = 2 Scenario
........................................................................................................................149
Figure B-1 Independence of ki1ki and ki+1ki ...171

ix

Abbreviations and Acronyms

AD/DA

analog-to-digital/digital-to-analog

AGC

automatic gain controls

AP

access point

AS

antenna selection

BER

bit error rate

BF

beamforming

BS

base station

CBS

correlation-based selection

CCI

co-channel interference

CDF

cumulative distribution function

CDMA

code-division multiple access

CDMA/CA

Carrier Sense Multiple Access/Collision

Avoidance

CISVD

controlled iterative singular value

decomposition

CSI

channel state information

DF

decision-feedback

DOD

direction of departure

DPC

dirty-paper coding

FDMA

frequency-division multiple access

GI

guard interval

GS

Gram-Schmidt

HS-MRC

hybrid selection-maximum ratio combining

HT-LTF

high-throughput long training field

HT-STF

high-throughput short training field

i.i.d.

independent and identically distributed

ISI

Inter-symbol interference

JLS

joint leakage suppression

JT

joint transmission

JWF

joint Wiener filtering

LOS

line-of-sight

MAC

medium-access control

MCS

modulation/coding sets

MEMS

micro-electro-mechanical systems

MIMO

multiple-input multiple-output

MMSE

minimum mean square error

MMSE-DF

MMSE decision-feedback

MS

mobile station

MSE

mean square error

MUD

multiuser detection

MUI

multi-user interference

NLOS

non-line-of-sight

OFDM

orthogonal frequency division multiplexing

OFDMA

orthogonal frequency division multiple

access

PBS

power-based selection

PDF

probability density function

PER

packet error rate

PHY

physical

PU

processing units

RF

radio frequency

RX

receive(r)

SD

spatial diversity

SDMA

Spatial-division multiple access

SIC

successive interference cancellation

SIFS

short inter-frame interval

SINR

signal to interference plus noise ratio

SISO

single-input single-output

SLNR

signal to leakage plus noise ratio

xi

SM

spatial multiplexing

SNR

signal to noise ratio

STA

station

STBC

space-time block coding

STDS

space-time data streams

SVD

singular value decomposition

TDMA

time-division multiple access

TD-SCDMA

time division synchronous CDMA

THP

Tomlinson-Harashima precoding

TX

transmit(er)

V-BLAST

vertical Bell Labs layered space-time

architecture

WLAN

Wireless local area network

WMAN

wireless metropolitan area network

ZF

zero-forcing

ZF-DF

zero-forcing decision-feedback

xii

Chapter 1
Introductions

1.1 MIMO Communications Systems

In the past few years, the demand for broadband wireless data access in mobile
communication networks has grown exponentially. For example, existing standards for third
generation (3G) provide up to 2 Mbps indoors and 144 Kbps in vehicular environments;
while the minimum speed currently targeted for 4G systems is 10-20 Mbps indoors and 2
Mbps in moving vehicles.
Recently, the remarkable capacity potential of MIMO systems has been unveiled
[20][21][25][86][94]. A (flat fading) MIMO channel, typically modeled as a matrix with i.i.d.
complex Gaussian entries, provides multiple spatial dimensions for communications and
yields a spatial multiplexing gain. At high SNR, Shannon capacity can increase linearly with
the minimum of the number of transmit and receive antennas. Meanwhile, multiple
transmit/receive antennas in MIMO systems also provide diversity protections for data
streams in fading channels [53][68][79]. Spatial multiplexing and diversity gains typically
1

compete with each other, and the tradeoff between them is an active research area after the
pioneering work of [104].
Built on intensive theoretical investigations, the development of MIMO systems has
already entered the stage of commercial applications: they are being considered as a key
enabling technology for boosting the effective data rate or throughput in systems such as:
3GPP [45], IEEE 802.11 WLANs [105], and IEEE 802.16 WMANs [106].
In this dissertation, both theoretical and practical aspects of MIMO systems are taken into
considerations in our study of two major topics: antenna selection in point to-point MIMO
systems, and interference mitigation in cellular multiuser MIMO networks, as will be
specified in the subsequent sections.

1.2 Point-to-Point MIMO Systems: Antenna Selection

The remarkable ability of MIMO wireless communication systems to provide spatial
diversity or multiplexing gains was mainly studied in the single-user scenario, so called
point-to-point MIMO. This scenario includes multi-user wireless communications with
orthogonal channelizations [69], e.g. TDMA or FDMA, and orthogonal CDMA. Another
example is the IEEE 802.11n high-throughput WLAN [105], where the MAC protocol
CSMA/CA is applied, such that theoretically there is only one device in transmission in each
frequency band at any instance.
In general there are two major categories of point-to-point MIMO systems:

1.

Open-loop: the transmitter does not have CSI, so data streams are usually
transmitted in a balanced manner, in which power and modulation/coding sets (MCS)
are allocated equally [20][22].

2.

Closed-loop: the transmitter knows full or partial CSI and adapts the data streams
accordingly to achieve performance improvements. Examples of closed-loop
systems include transmitter beamforming [53][79], link adaptation [105], and
transmit antenna selection [52], one of the major focuses of this dissertation.

In general, antenna selection techniques provide a new form of diversity in MIMO systems.
In this dissertation, we focus on three aspects:
First, we study the best diversity order achieved by transmit antenna selection for MIMO
spatial multiplexing systems employing practical coding and decoding schemes. This
problem is difficult by nature and rarely addressed in literature, except some conjectures
based on numerical experiments. We have proposed a new geometrical framework for this
type of analysis and obtained some interesting results, which confirm and generalize some of
the conjectures in literature (see Section 2.2). This approach has also been extended to tackle
some previously unsolvable problems in general MIMO systems without antenna selection
(see Chapter 3).
Secondly, design of antenna selection algorithm is another important problem, which
actually is the focus for most of the papers in the area. We also propose a fast transmit
antenna selection method based on some geometrical properties of MIMO channel matrices.
This algorithm achieves significant performance improvement with very affordable

computational effort, and is amendable to different levels of CSI at the transmitter (see
Section 2.3).
Finally, we look into the issue of how to implement antenna selection techniques in a
realistic commercial system, more specifically, the emerging IEEE 802.11n high-throughput
WLAN. In this scenario, practical issues should be jointly considered, such as the increased
training overhead, distortions caused by channel variations, cost and performance loss due to
the deployment of antenna switches, and impairments caused by RF mismatches. Our goal
here is to accommodate antenna selection with the least influence from the above aspects.
With this in mind, based on the specific characteristics of WLAN channels, we propose
effective antenna selection training/calibration protocols for the emerging 802.11n standard,
which are successfully adopted in the draft specifications [105] (see Section 2.4).

1.3 Multiple User MIMO Systems: Interference Mitigation

Multi-user MIMO refers to the scenario where multiple multi-antenna devices are
communicating in the same frequency-time slot simultaneously. Consequently, data streams
from different users introduce multi-user interference (MUI), and it is assumed that the only
way for separating them is to utilize their spatial signatures. Specifically, if different mobile
users are located far apart, their corresponding channel vectors/matrices provide a distinct
spatial signature for each of them. Whenever their spatial signatures are "orthogonal" enough
(i.e. with low correlation), their signals can be effectively separated by well-designed precoding or detection algorithms, an idea equivalent to randomly coded CDMA systems.
However, compared with CDMA, since the spatial signatures cannot be artificially optimized
4

and there are usually insufficient spatial dimensions due to the hardware limitations, more
effort on the transmitter/receiver design is required for interference mitigation. This scheme
is sometimes named as SDMA.
Our study mainly targets the infrastructure networks, or cellular networks [26][69][87],
which, compared with ad hoc networks, utilize base stations or access points to provide
access for mobile stations to a backbone network. Here a "cell" is defined as a specific area
serviced by a BS, through which a subscriber in the area can directly access the backbone
network a single-hop scheme. Base stations are usually connected by reliable wirelines to
centralized processing units to facilitate advanced functionalities such as handoff and
resource allocations. Therefore the performances of infrastructure based wireless networks
are usually much better than those without infrastructure, which explains why they will
continuously be relied on for fulfilling the stringent wideband requirements.
In the uplink, where the signals are transmitted from multiple MSs to a single BS,
multiple antennas at the receiver are used for detecting different data streams of different
users, exploiting multiuser detection techniques. The studies on MUD were well developed
in the 1990s, a comprehensive review of which can be found in [89] and references therein.
Note that MUD was originally discussed under the CDMA settings. Due to the similarity
between CDMA and SDMA as mentioned earlier, they can be naturally extended to the
multi-antenna systems. Some recent development on MIMO MUD can be found in [14][90].
Despite its great potential in MUI mitigation and performance enhancement, the applications

of MUD are often limited by many factors including complexity and timing accuracy (i.e.
synchronization).
In the downlink, where the signals intended for multiple users are transmitted together
from the BS, both receiver processing (MUD) [16] and transmitter pre-processing (assuming
transmitter side CSI) are explored in the literature. A detailed survey will be given in Chapter
4.
Some simple multi-user MIMO architectures already appear in realistic systems. For
example, the Chinese 3G standard TD-SCDMA employs multiuser detectors in its
synchronized uplink [47]; and one optional mode of the new IEEE 802.16e WMAN standard
applies uplink joint detection and downlink beamforming in its OFDMA PHY layer, under
multiuser MIMO settings [106].
In this dissertation, we develop some new methods in transmitter side pre-processing for
the MIMO cellular downlink. With the development of RF and baseband IC circuitry, and in
some low Doppler environment, the acquisition of full CSI at the transmitter(s) in the
downlink is made possible through, e.g. a joint training process and feedback. In cellular
systems, universal frequency reuse is beneficial regarding the network throughput, as long as
the interference can be effectively managed. Moreover, in cellular systems, intra-cell and
inter-cell interferences are often modeled and mitigated in different ways. While the former
one is well investigated, inter-cell interference is far less explored, due to its random and
uncontrollable nature in conventional cellular architectures. We propose to tackle this
problem through base station cooperation schemes (see Chapter 4). To show the potential

benefit of BS cooperation, we start with a synchronous and strong inter-cell interference

environment, where all the involved MSs are located close to cell boundaries, so that the
associated BSs act like a distributed giant antenna array; later on we extend the study to
arbitrary MS locations, where the asynchronous channel model is indispensable. The preprocessing designs under this setting are then more technically involved. We have proposed
some robust designs that lead to significant performance enhancement with small
computational overhead, and some of them are even amendable with additional timinginaccuracy.

1.4 Dissertation Outline

Our research outcomes on antenna selection are discussed in detail in Chapter 2, which
contains the following parts: the introduction of antenna selection (Section 2.1), the diversity
analysis of transmit antenna selection with practical coding and decoding schemes (Section
2.2), the geometry-based fast antenna selection methods (Section 2.3), and implementation of
antenna selection in IEEE 802.11n WLAN (Section 2.4). Finally concluding remarks are
given in Section 2.5.
The extension of the same geometrical analytical tool to the diversity analysis of general
MIMO SIC receivers is given in Chapter 3, where Section 3.1 and 3.2 give general
introductions and system models; Section 3.3 presents the diversity analysis of SIC receivers
in V-BLAST point-to-point systems, while the counterpart in MIMO multiple access systems
is addressed in Section 3.4; finally Section 3.5 summarizes the chapter.

In Chapter 4, our focus moves to interference mitigation in multi-user MIMO. At first, a

detailed literature survey is given in Section 4.1, followed by the multiuser MIMO system
models in Section 4.2; then Section 4.3 proposes the idea of BS cooperation in a quasisynchronized scenario; Section 4.4 extends the discussions to asynchronous settings; finally
Section 4.5 concludes the whole chapter.
Conclusions and some perspectives on future work are given in Chapter 5.

Chapter 2
Antenna Selection in Point-to-Point MIMO

2.1 Introduction of Antenna Selection

As introduced in Chapter 1, MIMO techniques can significantly increase system capacity
in a rich scattering wireless environment. However, an accompanying problem is increased
hardware complexity and cost, since in a typical communications system, each
transmit/receive antenna requires a separate RF chain including a modulator/demodulator, an
AD/DA converter, an up/down converter, and a power amplifier. In addition, the processing
complexity at the baseband also increases with the number of antennas.
Antenna selection comes as a remedy for this problem [52]. In many scenarios, judicious
antenna selection may incur little or no loss in system performance, while significantly
reducing system cost. Although it is argued that by Moores Law, and by the deployment of
new efficient design architectures (e.g. the direct-conversion technique, or zero intermediate
frequency), the cost of RF chains may be reduced to the same magnitude as antenna elements

in the near future, and the impairment caused by cross-talks between multiple integrated RF
chains may be even smaller than that caused by antenna switches used for antenna selection,
antenna selection is still beneficial in improving the system performance, since it is operated
in the RF domain, and essentially doesnot interfere with any other performance-enhancement
techniques. In another word, antenna selection is not proposed as an alternative, but rather as
an enhancement to other advanced processing (e.g. the transmit beamforming), to achieve
better performances. We will re-stress this idea in Section 2.4.
The idea of antenna selection refers to selecting an optimal sub-matrix from a complete
MIMO channel matrix according to some predetermined criterion, and switching the selected
antennas to the limited numbers of RF chains. To perform antenna selection, the complete
channel matrix is estimated by sending training frames to measure the complete channel state
information. On transmitting or receiving these AS training frames, the device conducting
antenna selection switches different antenna subsets to the RF chains so that the receiver can
estimate the corresponding subchannels matrices.
Figure 2-1 shows a point-to-point MIMO communication system, including a transmitting
station (STA A), and a receving station (STA B). Either station can operate in the receive or
transmit mode. Each station includes a set of receive RF chains, and a set of transmit RF
chains both connected to a set of antennas by switches. Generally, in MIMO devices, the
number of antennas is larger than that of RF chains. Therefore, a subset of antennas is
selected from the set of total available antennas by an antenna selection method during a

10

training phase. The selection can be initiated by either the transmitter or the receiver, and the
selection can be conducted at the transmitter and/or at the receiver.

Figure 2-1 Antenna Selection System Model

In literature, research on antenna selection is focused on two aspects: performance
analysis and selection algorithms. A brief survey on these topics will be given in Section 2.2
and Section 2.3 respectively.
Previous work either investigates the performance of antenna selection for MIMO spatial
diversity schemes (e.g. STBC [30]), or for MIMO spatial multiplexing schemes with
idealized encoding and decoding [32]. In contrast, we propose an effective geometrical
analysis tool in Section 2.2 to exploit the exact diversity performance of transmit antenna
selection in MIMO-SM systems with practical coding and decoding schemes. While good in
theoretical analysis, the antenna selection algorithms used in Section 2.2 are generally
infeasible for realistic implementations. Tradeoffs between performance and complexity play
an important role in practice. From a different geometric view, we propose fast antenna
selection algorithms, which may achieve near optimal performance in some fading
11

environments, particularly with high fading correlations. In Section 2.4 we take one step
further to investigate antenna selection in commercial wireless networks. Specifically, we
propose the training and calibration methods for antenna selection in high-throughput WLAN
systems, a topic receiving little attention so far.

2.2 Performance Analysis of Antenna Selection: Diversity Order of MIMOSM with Transmit Antenna Selection - A Geometrical Approach
2.2.1 Introduction
As mentioned earlier, a major potential problem with MIMO is increased hardware cost
due to multiple analog/RF front-ends, which has recently motivated the investigation of
antenna selection techniques for MIMO systems [52].
MIMO systems can be exploited for spatial diversity or spatial multiplexing gains [104].
The majority of analytical work on MIMO antenna selection focuses on the former, including
selection combining, HS-MRC [37][79], and antenna selection with space-time coding
[1][30]. Essentially in these works, with i.i.d. Rayleigh fading, the system error performance
or outage probability can be readily analyzed through order statistics [17]. It has been shown
that the diversity order of the original full-size system can be maintained through the SNR
maximization selection criterion.
By contrast, antenna selection for MIMO systems with multiple data streams, i.e., spatial
multiplexing systems, has received less attention. The few existing analytical results
generally assume capacity-achieving joint space-time coding and optimal decoding.
Capacity-maximizing receive antenna selection is analyzed in [32] and shown to achieve the
12

same diversity order as the full-size system. In [33], it is shown that the fundamental tradeoff
between diversity and spatial multiplexing of the full-size system, obtained in [104], holds as
well for MIMO systems with antenna selection.
In practice, the multiple streams in a SM system may be uncoded or separately encoded
and sub-optimally decoded due to complexity and channel feedback overhead concerns. In
[36], several transmit antenna selection algorithms for SM with linear receivers are proposed,
and some conjectures on the achieved diversity orders are made from numerical results. To
the best of our knowledge, the exact diversity order achieved by antenna selection for
practical SM systems has not been rigorously obtained. In contrast to MIMO diversity
schemes, the key challenge that hinders accurate performance analysis is that selection is
performed among a list of inter-dependent random quantities, which are correlated in a
complex manner.
In this Section, we propose a new framework to analyze the diversity order achieved by
transmit antenna selection for SM systems with separately (and independently) encoded data
streams and linear or decision-feedback receivers (i.e., the V-BLAST structure [20][95]). In
particular, we rigorously show that the optimal diversity order is ( NT 1)( N R 1) for an
N R NT separately encoded SM system employing linear or DF receivers, when L = 2
antennas are selected from the transmit side. This should be compared with the diversity
order of a two-stream SM system without antenna selection: N R 1 . Such a diversity gain

can be tremendous for downlink high-data-rate communications, where there may be a large
number of transmit antennas at the base stations but few receive antennas at the mobiles (e.g.,
13

N R = 2 ). For DF receivers, it is also clarified that, the antenna selection rule that maximizes
the performance of the first decoded data stream is not optimal, as the freedom to choose
transmit antennas for subsequent data streams is restricted. Furthermore, following the same
geometrical approach, we give an upper bound, ( NT L + 1)( N R 1) , and a lower bound,
( NT L + 1)( N R L + 1) , on the diversity order for general L , which coincide when L = 2 .
The corresponding diversity-multiplexing tradeoff curves are also derived. Generally
speaking, our results confirm and generalize some of the conjectures in [36], thus verifying
that transmit antenna selection can achieve high data rates and robust error performance in
practical SM systems without complex coding. Furthermore, the proposed geometrical
approach may be used to solve other open problems related to MIMO communication
systems.
Among feasible sub-optimal receivers for separately encoded SM systems, we first
investigate the linear ZF receiver [89]. The analysis for the ZF-DF receiver relies heavily on
the former. Furthermore, their diversity order analyses, a study at high SNR regimes, hold for
linear MMSE and MMSE-DF receivers as well.
2.2.2 System Model and Problem Formulation

We consider a frequency non-selective block Rayleigh fading channel model, in which the
N R NT channel matrix is represented by H = [h1 , h 2 ,

, h NT ] , modeled with i.i.d.

normalized complex Gaussian entries with zero mean and unit variance. With transmit

14

N
antenna selection, there are a total of NU = T possible antenna subsets, defined as
L
U1 ~ U NU 1:
U1 = {h1 , h 2 ,

, hL}

U 2 = {h1 , h 2 ,

h L 1 , h L +1}

U NU = h NT L +1 ,

(2.1)

, h NT .

If the subset U j is selected, of which the L column vectors compose the channel matrix H j ,
the SM system can then be expressed as:
y=

0
L

H js + n ,

(2.2)

where the N R T matrix y is the received signal block with T the block length; n is the
circularly symmetric complex Gaussian background noise with zero mean and unit variance;
and the L T matrix s represents the transmitted signal block, with independent rows each
assuming unit average energy per symbol. Therefore, 0 is the average SNR per receive
antenna. As in [104], we assume the fading channel keeps constant for the entire data frame.
Each stream independently adopts a code of fixed data rate R0 2. As will be shown (see proof
of Lemma 2.1), channel coding in each data stream can provide coding gain but not diversity
gain. Throughout the paper we assume NT L and N R L . Therefore given the assumption

1
2

Without loss of generality, it is assumed that in each subset the selected columns are placed by index in increasing order.
We are mainly interested in the maximum diversity gain here, corresponding to a zero multiplexing gain scenario in [104].

15

of i.i.d. Rayleigh fading, the selected channel matrix H j has full column rank with
probability one [55].
1

For a ZF receiver, a spatial equalizer G ZF = H j = H Hj H j H Hj is applied to the received

signal y to obtain an estimate of the transmitted signal:
s = G ZF y =

where

0
L

s + G ZF n ,

(2.3)

denotes the pseudo-inverse of a matrix. Therefore L equivalent decoupled SISO

data links are formed as

sk =

sk + [G ZF ]k * n, 1 k L ,

(2.4)

where [G ]k * is the kth row in matrix G . The instantaneous channel capacity and error
performance of these subchannels are determined by the corresponding post-processing
SNRs, which can be directly derived from (2.4) as
0
[G ZF ]k *
L

)
k( ,jZF
=

where

1

= 0 H Hj H j , 1 k L ,
kk
L

(2.5)

stands for conjugate transpose, and []kk represents the kth diagonal element of a

matrix. Let h (k j ) be the kth column of H j , and H (k j ) the N R ( L 1) matrix resulting from
leaving out h (k j ) from H j . It is known that [89], 1/ H Hj H j

16

1
kk

is equal to Rk( j ) , the square of

the projection height3 from h (k j ) to the range of H (k j ) , or alternatively, the square of the norm
of the projection of h (k j ) onto the null space of ( H (k j ) ) . That is
T

)
k( ,jZF
=

0
L

Rk( j ) =

0
L

h (k j ) sin 2 k( j ) ,

(2.6)

where h (k j ) is the norm of h (k j ) , while k( j ) is the angle between h (k j ) and its projection on
the range of H (k j ) , defined as

( j)
k

= sin

Rk( j )
h

( j)
k

, 0 < k( j ) <

The following figure gives a geometrical view of the relationship in (2.6).

Figure 2-2 The Illustration of (2.6)

We further define

Projection height refers to the norm of the error vector, i.e., the difference between a vector and its projection onto a subspace.

17

( j)
Rmin
= min{Rk( j ) } , 1 j NU

(2.7)

k =1 L

for each subset U j , which essentially determines the system performance at high SNR
[66][89].
For

linear

MMSE

receiver,

the

spatial

equalizer

is

given

by

G MMSE = [H Hj H j + L / 0 I ]1 H Hj , and the post-processing SNRs are given by [89][64]:

)
k( ,jMMSE
=

where (k j ) =

L[H Hj H j + L / 0 I ]kk1

1 =

0 (k j )
L

1, 1 k L ,

(2.8)

1
. Similarly, we define
[H H j + L / 0 I ]kk1
*
j

j)
(min
= min{ (k j ) } , 1 j NU .

k =1 L

(2.9)

In this section, we adopt the antenna selection rule that maximizes the post-processing
SNR of the worst data stream (as in [36]). That is, we choose the subset among (2.1) such
( j)
j)
in (2.7) or (min
in (2.9) is maximized, and we denote
that Rmin
( j)
j)
RSL = max {Rmin
} and SL = max { (min
}.
1 j NU

1 j NU

(2.10)

We will show that this selection rule is optimal for linear receivers with respect to diversity
order.
Before proceed, we first introduce some notations. The diversity order is defined as the
slope of the average frame error probability (an error is declared if any substream is decoded
unsuccessfully. By assuming coding over a single block with constant fading, this is the true error

18

probability of a code averaged over the transmitted codewords, channel fading, and additive noise.)

Pe ( 0 ) in log-scale in the high SNR regime [104]:

log Pe ( 0 )
log Pe ( 0 )
= lim
.
0 log( )
0 log(1/ )
0
0

d = lim
We adopt the operator
f ( 0 )

(2.11)

as defined in [104], to denote exponential equality, i.e. we write

0b to represent
log f ( 0 )
=b .
0 log( )
0

lim

Equivalently (for the convenience of analysis in this paper), we use f ( x)

lim
x 0

log f ( x)
=b.
log x

xb to represent
(2.12)

The operators , , < , > are similarly defined. Note that according to our notation,
f ( x) g ( x) indicates f ( x) g ( x) for sufficiently small x .

Lemma 2.1: For separately encoded spatial multiplexing systems with linear ZF/MMSE

receivers, the antenna selection rule that chooses the antenna subset with the strongest
weakest data link achieves the optimal diversity order among all antenna selection rules.
Furthermore, the optimal diversity order can be evaluated as
log Pr( RSL x)
x 0
log( x)

d L = lim
opt

for both receivers.

Proof: See Appendix A.

19

(2.13)

Remarks: Lemma 2.1 indicates that we only need to evaluate (2.13) for the optimal diversity
order of transmit antenna selection with linear ZF/MMSE receivers. We also make no claims
on the practicality of this selection algorithm, as the main focus of this section is on
theoretical analysis. Note that efficient antenna selection algorithms exist in literature (e.g.,
[37][108] and references therein), and will be further discussed in Section 2.3.
2.2.3 Diversity Order and Diversity-Multiplexing Tradeoff when L = 2

In this subsection, we discuss the main idea of our geometrical approach for L = 2 , which
also admits an exact result. Extension to the general L scenario is not trivial, as will be seen
in Section 2.2.4.
Theorem 2.1: In an N R NT spatial multiplexing system with linear ZF/MMSE or ZF/MMSE

decision-feedback receivers satisfying NT 2 and N R 2 , if separately encoded data

streams are transmitted from two selected antennas, the optimal achievable diversity order is

d opt = ( NT 1)( N R 1) .

(2.14)

The optimal diversity-multiplexing tradeoff curve is the piecewise function of r connecting

the points (r , d opt (r )) for 0 r 2 , with
d opt (r ) = ( NT 1)( N R 1)(1 r / 2) + ,

(2.15)

where ( x) + = max( x, 0) .

This theorem will be proved in three steps: we start with linear ZF/MMSE receivers by
evaluating (2.13) in Subsection 2.2.3.1. Then we extend the analysis to ZF/MMSE-DF

20

receivers in Subsection 2.2.3.2. Finally we analyze the corresponding diversity-multiplexing

tradeoff curves in Subsection 2.2.3.3.
As L = 2 , each possible antenna subset in (2.1) contains two selected column vectors
from the original channel matrix H . For convenience, we will use Rkj to denote the squared
projection height from h k to h j . In another word, following the denotations in Section 2.2.2,
given that U s = {h k , h j } is selected, Rkj

R1( s ) and R jk

R2( s ) (c.f. (2.6)). Clearly

Rkj = h k sin 2 kj ,
2

where it can be shown that [18][49][55] h k

(2.16)

is 2 (2 N R ) distributed with probability

density function4

f||h ||2 ( x) =
k

1
1
x N R 1e x =
x N R 1e x ,
( N R 1)!
( N R )

(2.17)

and kj assumes a PDF of

fkj ( ) = ( N R 1) sin 2 (sin ) 2 N R 4 ,
Furthermore, h k

, hj

0, .
2

(2.18)

and kj are mutually independent. Also, Rkj is a 2 (2( N R 1))

distributed random variable, which leads to

Pr( Rkj x)

Chi-square distribution with 2 N R degrees of freedom and mean N R .

21

x N R 1 .

(2.19)

From (2.19) we can see that without transmit antenna selection ( NT = L ), any data stream
with a ZF/MMSE receiver bears a diversity order of N R 1 , so does the overall error
probability. When NT > L , through the diversity-maximization transmit antenna selection
(c.f. (2.7) and (2.10))
RSL =

max

k j{1, , NT }

{min {R , R }} ,
kj

(2.20)

jk

we will show that a product gain of NT 1 on the diversity order can be achieved.
2.2.3.1 Linear ZF/MMSE Receivers
L
In the following, the optimal diversity order (2.13) for linear ZF/MMSE receivers d opt
will

be explicitly explored. Note that neither the exact PDF of RSL nor its polynomial expansion
near zero seems tractable, which motivates us to solve the problem through tight upper and
lower bounds.
By definition
Pr( RSL x) = Pr(min( R12 , R21 ) x ,min( R13 , R31 ) x,

, min( R( NT 1) NT , RNT ( NT 1) ) x)

(2.21)

N
= Pr Ai ,
i =1

where the N = 2 NU events { Ai } are defined as

A1 = {R12 x, R13 x,

, R1NT x, R23 x,

, R( NT 1) NT x}

A2 = {R12 x, R13 x,

, R1NT x, R23 x,

, RNT ( NT 1) x}

AN = {R21 x, R31 x,

, RNT 1 x, R32 x,

, RNT ( NT 1) x},

22

(2.22)

each of which corresponds to the case that one data stream from each of the NU subsets is in
outage (i.e., close to zero for sufficiently small x). For example, event A1 picks the first

element from each possible subset U1 ~ U NU , i.e., A1 = { Rkj x} . Then from (2.21)
1 k < j NT

we have:
N

Pr( A1 ) Pr( RSL x) Pr( Ai ) ,

(2.23)

i =1

which indicates

lim
x 0

log max Pr( Ai )

i

log( x)

log Pr( A1 )
.
x 0
log( x)

L
d opt
lim

(2.24)

Clearly the lower bound in (2.24) actually is tight. However, it is generally difficult to
identify the dominant terms at high SNR from (2.22), since the cardinality grows
exponentially with NT2 . Alternatively, we take the following approach. First, we find a
common upper bound for Pr( Ai ), i , i.e., obtaining such a PU that
max Pr( Ai ) PU ,
i

(2.25)

L
which determines a lower bound for d opt
(c.f. the lower bound in (2.24))5. We then obtain a

lower bound for Pr( A1 ) :

Pr( A1 ) PL ,

Recall that an error probability with lower diversity order corresponds to an exponentially larger one, and vice versa.

23

(2.26)

L
and evaluate its error exponential, which gives an upper bound for d opt
(c.f. the upper bound

in (2.24)). It turns out that these two bounds coincide and represent the best achievable
diversity order, given by ( NT 1)( N R 1) .
Diversity Lower Bound
L
Proposition 2.1: d opt
( NT 1)( N R 1) .

N
Proof: Define Si as the set consisting of the NU = T random variables in Ai (see (2.22)).
2

For example, S1 = R12 , R13 ,

, R1NT , R23 ,

, R( NT 1) NT = { Rkj }

1 k < j NT

. First we claim that, in

any Si we can always find a subset of NT 1 random variables in the form of

Si _ indep = Rk1k2 , Rk2 k3 ,

, Rk N

k
T 1 NT

} , where k ~ k
1

1 ~ NT . For example, S1_ indep = R12 , R23 ,

NT

is some permutation of the integer array

, R( NT 1) NT , i.e., ki = i, 1 i NT . By Lemma B.1

given in Appendix B, we obtain the following common upper bound for Pr( Ai ) :
Pr( Ai ) PU = [Pr( Rkj x)]( NT 1) , k j , Ai .

(2.27)

With (2.19) we have

.

max Pr( Ai ) x ( NT 1)( N R 1) ,

i

and from (2.24) Proposition 2.1 follows.

24

(2.28)

Figure 2-3 The Illustration of the Proof for Proposition 2.1: A Graph with Size n

The claim is proved as follows. A graph is drawn in Figure 2-3 to visualize Si , where the
nodes represent the transmit antenna elements and an arrow directed from node k1 to
k2 appears in the graph if Rk1k2 exists in Si . Based on the definition of Si , there is one and
only one arrow between any two nodes. For such a graph of size n, a complete path is defined
as one that goes through a series of n 1 arrows, passing each of the n nodes one and only
one time. Therefore we can find a subset Si _ indep in Si if at least one complete path exists in
the graph of size NT , passing the nodes with the order k1 k2

k NT . It is easily

shown that all graphs of size 3 contain one complete path. Now suppose that all graphs of
size n 1 contain at least one complete path. Then in any graph of size n , we can always
find a path passing n 1 nodes one and only one time, assumed in the order
k1 k2

kn 1 . From Figure 2-3, we see that if there is no complete path existing in the

25

graph of size n, the arrow between k1 and kn should be in the direction of k1 kn , otherwise
kn k1

kn 1 forms a complete path. For the same reason, the arrow between k2 and

kn should be in the direction of k2 kn , otherwise k1 kn k2

kn 1 forms a complete

path. By repeating the same reasoning, we finally reach kn 1 kn . However, in this case
k1 k2

kn is a complete path, and we conclude that a complete path always exists

for any graph of size n . Thus the claim is proved for any NT by induction.

Diversity Upper Bound

L
Proposition 2.2: d opt
( NT 1)( N R 1) .

Proof: This proof is technically involved, so we divide it into three parts for ease of
illustration: (1) by some geometrical analysis Pr( A1 ) is further lower bounded (see (2.31));
(2) by Lemma B.2~B.5 in Appendix B, the probability lower bound is transformed into an
exponentially equivalent form with known statistics (see (2.35)); (3) the diversity upper
bound is explicitly evaluated (see (2.38)).
Part (1):

Given

2
Pr( A1 ) = Pr Rkj x = Pr h k sin 2 kj x ,
1 k < j N

1 k < j N

T
T

26

and by defining z =

NT 1

k =1

hk

, which is distributed as 2 (2( NT 1) N R ) , and

0 = ( / 2) /( NT 1) , we have:

Pr( A1 ) Pr z sin 2 kj x Pr z sin 2 kj x, max 1k < 0 ,

1 k < j N

1 k < j N

2 k NT
T
T

(2.29)

where for the second inequality we have further restricted the ranges of the NT 1 i.i.d.
random variables {1k }k =T 2 within (0, 0 ) (c.f. Corollary B.1 in Appendix B).
N

Based on the geometric structure involved, given 1k and 1 j , 2 k < j NT , kj is

constrained as kj 1k + 1 j , where the equality holds only when h1 , h k , h j are linearly
dependent (located in the same subspace with a dimension less than 3) [65]. Then within the
range (0, 0 ) , we have
sin 2 kj sin 2 (1k + 1 j ) sin 2 (12 + 13 +

+ 1 NT ) = sin 2 ,

(2.30)

NT

where = 1k is still in the range of (0, / 2) . Therefore (2.29) can be further lower
k =2

bounded as:

Pr( A1 ) PL = Pr z sin 2 x, max 1k < 0

2 k NT

Pr( z sin 2 ' x),

(2.31)

where we define a new set of i.i.d. random variables '12 ~ '1NT with PDF of
f '1i ( x) =

f1i ( x)
f1i ( x)dx

f1i ( x)
C

, 0 < x < 0 , 2 i NT ,

27

(2.32)

NT

i.e., the restriction of 12 ~ 1NT in the range of (0, 0 ) , and ' = '1i .
i =2

Part (2):

With Lemma B.2~B.5 on exponential equivalence given in Appendix B, the evaluation of

(2.31) can be further simplified. Specifically, by defining '0 = max '1k and m( x) = sin 2 ( x)
k

for x (0, 0 ) , from Lemma B.3 (whose proof requires Lemma B.2) we have
Pr(sin 2 ' x)

Pr(sin 2 '0 x) ,

(2.33)

i.e., with respect to monotonic functions, replacing the sum of a set of independent random
variables with the maximum of them does not change the associated exponential behavior.
Further by Lemma B.4, we have
Pr( z sin 2 ' x)

Pr( z sin 2 '0 x) ,

(2.34)

i.e., the exponential behavior is still not changed by multiplication of an independent random
variable.
Finally by Lemma B.5, we have
Pr( z sin 2 '0 x )

Pr( z sin 2 0 x ) ,

(2.35)

where 0 = max 1k , i.e. without the constraint of (0, 0 ) on 12 ~ 1NT .

k

Part (3):

We are then left to evaluate the smallest exponential in Pr( z sin 2 0 x) in (2.35). Note that

z is a 2 (2( NT 1) N R ) distributed random variable with cumulative distribution function:

28

Fz ( x) = 1 e x

M + NT 2

k =0

xk
,
k!

M = ( NT 1)( N R 1) ,

(2.36)

while from (2.18) the PDF of 0 can be easily derived as:

f0 ( ) = M [sin 2 ]M 1 sin 2 , M = ( NT 1)( N R 1), (0, ) .

2

(2.37)

After some algebra presented in Appendix C, we can get the following equivalent polynomial
form as x 0 :
NT 3
1
M
k!
M
Pr(z sin 2 0 x) =
M
x + o( x ), M = ( NT 1)( N R 1) ,
M
!
(
M
k
1)!
+
+
k =0

(2.38)

where the coefficient of x M in (2.38) is always positive (also shown in Appendix C), which
completes the proof.

For example, when NT = 3, N R = 3 , and L = 2 , from (2.38) we get as x 0

x4
Pr(z sin 0 x) =
+ o( x 5 ) .
120
2

To illustrate the soundness of our approach, Figure 2-4 presents by simulations the
exponential behavior of Pr(z sin 2 0 x) , together with the outage probability in (2.21) and
the probabilities representing the diversity bounds as in (2.23)(2.24). The results show a
diversity order of 4, verifying our derivations above.
With Proposition 2.1 and 2.2, (2.14) is proved for the linear receiver case.

29

Figure 2-4 The Exponential Behaviors of the Outage Probabilities for the
NT = N R = 3, L = 2 Scenario

2.2.3.2 Decision-Feedback Receivers

In this subsection we continue to explore the optimal achievable diversity order for
separately encoded SM systems with transmit antenna selection and DF receivers, denoted as
DF
d opt
. As expected, the performance analysis for DF receivers will rely heavily on that for

their linear counterparts in Section 2.2.3.1.

For DF receivers with L = 2 , the frame error probability is given by

Pe = Pe1 + Pe 2 (1 Pe1 ) ,

30

(2.39)

where Pe1 is the error probability of the first decoded data stream, and Pe 2 is that of the
second stream assuming perfect feedback. Therefore Pe

max{Pe1 , Pe 2 } . For a fixed-order

ZF/MMSE DF receiver without antenna selection, Pe1 Pe 2 is always fulfilled, so that we can
investigate its diversity order solely from the first decoded data stream, which is processed
by just a linear ZF/MMSE receiver (see, e.g. [104][66]). This may entice one to consider an
antenna selection rule that maximizes the performance of the first decoded data stream.
However, as will be shown, this antenna selection rule is generally not optimal.
Instead of seeking an optimal antenna selection rule for DF receivers like Lemma 2.1, we
DF
take the following simpler approach to evaluate d opt
, thanks to the close connection between

DF receivers and their linear counterparts. First, since for any antenna selection rule j ,
DF
DF
Pe( j ) Pe(1j ) , we have d (DF
j ) d1,( j ) , where d1,( j ) is the diversity order of the corresponding first

decoded data stream. Consequently,

DF
d opt
d1,DF
opt ,

(2.40)

where d1,DF
opt is the best achievable diversity order of the first decoded data stream. Note that
there may be multiple choices of antenna selection rules to achieve the optimum for both
sides of (2.40), and a selection rule that is best for one may fail for the other. Nonetheless,
following a procedure similar to what we have discussed above for linear receivers, we will
DF
explicitly evaluate d1,DF
opt and derive a diversity upper bound for d opt . We then continue by

constructing a specific antenna selection algorithm whose diversity order is easy to assess
DF
and serve as a lower bound for d opt
.

31

Diversity Upper Bound - Maximizing the SNR of the First Decoded Data Stream

We investigate an antenna selection algorithm that achieves d1,DF

opt : selecting the antenna
subset that maximizes the post-processing SNR of the first decoded data stream. We
distinguish two scenarios with respect to detection order: arbitrary but fixed ordering and
optimal ordering [20]. Our main result is summarized below.
DF
Proposition 2.3: d opt
d1,DFopt = ( NT 1)( N R 1) .

Proof: For fixed ordering, without loss of generality, we assume that decoding starts from
the signal transmitted from the antenna with the smallest index number in the selected
antenna subset. The antenna selection rule indicates that the post-processing SNR of the first
decoded data stream is given by (c.f. (2.20))
RSL1 = max

1 k < j NT

{R } .

(2.41)

kj

Clearly we have (c.f. (2.22))

Pr( RSL1 x) = Pr( A1 ) ,

(2.42)

whose upper bound can be derived from (2.27), while its lower bound exponential behavior
evaluation directly follows (2.29)~(2.38). Therefore for arbitrary but fixed ordering
d1,DF
opt = ( NT 1)( N R 1) .

(2.43)

For optimal ordering, our selection rule is reformulated with

RSL 2 =

max

k j{1, , NT }

{max {R , R }} ,

and we have:

32

kj

jk

(2.44)

Pr( RSL 2 x) = Pr max( Rkj , R jk ) x

1 k < j N

2
2
= Pr max( h k , h j ) sin 2 kj x .
1 k < j N

(2.45)

It is straightforward to upper and lower bound (2.45) as (c.f. (2.27) and (2.31)):
Pr( RSL 2 x) Pr( A1 ) Pr( Rkj x)

NT 1

k j ,

(2.46)

and

2
2
Pr( RSL 2 x) Pr max h1 , , h NT sin 2 x

NT

2
Pr hi sin 2 x .

i =1

(2.47)

The evaluation of the lower bound in (2.47) is similar as in (2.31), except that z is re-defined
NT

as z = h k

. Following a similar approach as (2.31)~(2.38), we get

k =1

NT + N R 3
1
M
k!
M +1
x)
M
x + o( x ), M = ( NT 1)( N R 1) .
( M + k + 1)!
k =0
M!
.

Pr( RSL 2

(2.48)

So the same result as in (2.43) is obtained with optimal ordering.

Remarks: Even though RSL 2 RSL1 RSL with probability 1, no advantage in diversity order
can be achieved. Of course, some coding gain in SNR is naturally expected, which is beyond
DF
L
the scope of this paper. So d opt
d1,DF
opt = d opt . On the other hand, we anticipate the

33

performance of DF receivers is no worse than their linear counterparts at high SNR, which
will be verified below.
In [104][49], the authors have shown that optimal ordering will not increase the diversity
order in the first decoded data stream of a separately encoded SM system with DF receivers
in the L = 2 case6. As a side product, here we reach the same conclusion in the antenna
selection context.
Finally we note that this antenna selection rule is in general sub-optimal with respect to
the diversity order for the whole system. The reason is that although Pe1 in (2.39) is
minimized, Pe 2 is not affected by the selection process. Rather, it behaves the same as in a
non-selection scheme with N R -order diversity. Therefore Pe

max{Pe1 , Pe 2 } is mostly

dominated

order

by

the

second

stream,

and

the

diversity

is

given

by

d = min{( NT 1)( N R 1), N R } = N R , for NT 3, N R 2 .

Diversity Upper Bound - QR Decomposition-Based Antenna Selection Algorithm
Here we analyze a simple yet effective antenna selection algorithm, for which the first
decoded data stream achieves a diversity order of ( NT 1)( N R 1) , while the second stream
performs better than the first one. This antenna selection algorithm is based on QR
decomposition, which was originally proposed for capacity maximization [24][108].
Compared with brute force methods, this algorithm greatly reduces the computational
complexity while achieving a near optimal performance with respect to channel capacity.

Extension to the general L scenario was done recently in [37][117].

34

Here we apply it in SM systems with DF receivers with the goal of minimizing the error rate,
and show that it also maximizes the diversity order, therefore verifying our observations in
[108] and revealing its great potential.
DF
Proposition 2.4: d opt
( NT 1)( N R 1) .

Proof: For simplicity we mainly focus on ZF-DF receivers. From a geometrical viewpoint,
this incremental antenna selection procedure starts by seeking the column vector with the
largest norm (or the largest projection height to the null-space); and in each of the following
steps, one column with the largest projection height to the space spanned by the selected
column vectors is chosen until the Lth antenna is selected. The detection order is the reverse
of the selection order, i.e., the stream decoded at the lth step is transmitted from the antenna
selected at the ( L l + 1)th step. The corresponding post-processing SNR for the stream
decoded at the lth step is then proportional to the maximum value among NT L + l
independent random variables distributed as 2 (2( N R L + l )) , resulting in a diversity order
of ( NT L + l )( N R L + l ) . Therefore for L = 2 , the first decoded stream achieves a diversity
order of d1 = ( NT 1)( N R 1) and the second decoded stream achieves d 2 = NT N R , and the
diversity order for the whole system equals ( NT 1)( N R 1) .
For MMSE-DF receivers, the algorithm is slightly modified: at the lth step, a column h ( l )
is selected such that:
l = arg max
l

1
[[H (l 1) , hl ] [H ( l 1) , hl ] + 1/ 0 I ]ll1
H

35

where H ( l 1) is the matrix formed by the previous selected l 1 column vectors, and
[H ( l 1) , h l ] means appending hl to the last column. Following the proof of Lemma 2.1 (in

Appendix A), the same diversity lower bound is achieved.

Simulation results are presented in Figure 2-5 for the NT = 3, N R = 3, L = 2 scenario,

where the BER of the ZF-DF receiver with the above two selection rules are shown, together
with that for the DF system without antenna selection. The achieved diversity orders of
different schemes are also provided for a clearer comparison. We see that although simpler,
the QR based method (bearing a diversity of 4) outperforms the algorithm that optimizes only
the first decoded data stream, which only achieves a diversity order of 3, due to the
unaffected second data stream, as discussed above.
With Proposition 2.3 and 2.4, (2.14) is also proved for ZF/MMSE-DF receivers.

36

Figure 2-5 BER Performance of the ZF-DF Receiver with NT = N R = 3, L = 2

2.2.3.3 The Diversity-Multiplexing Tradeoff

Using the same geometrical approach, we can also obtain the diversity-multiplexing tradeoff
curve introduced in [104] for the separately encoded SM systems with antenna selection.
With quasi-static fading assumption, a family of codes { ( 0 )} over a block length shorter
than fading coherence time is employed, one at each SNR level. We further assume that the
rate of the code increases with SNR, so a scheme achieves a multiplexing gain r if the rate

R ( 0 ) = r log 0 . Based on the diversity order analysis in Section 2.2.3.1 and 2.2.3.2 (see
especially proof of Lemma 2.1 in Appendix A), we get the following equation:

Pe

Pe _ max

37

Pout _ max ,

where Pout _ max can be viewed as the outage probability of the worst stream for linear
receivers, and that of the first decoded data stream for DF receivers, respectively. The
diversity-multiplexing tradeoff curve d (r ) can be evaluated directly from Pout _ max as

0 :7

Pout _ max = Pr log(1 + 0 RSL ) log 0

L
L

r
(1 )

Pr RSL L 0 L

r
M (1 )
(1 Lr )
L

,
=
0

where M = ( NT 1)( N R 1) .
Therefore (2.15) is proved and we finalize the proof of Theorem 2.1.
2.2.4 Extension to General L
Analyses in Section 2.2.3 can be partially extended to the general scenario. In particular, the
diversity lower bound can be obtained following a similar approach. In contrast, a tight
diversity upper bound could not be achieved with straightforward extension. Evaluations
become more involved as now the post-processing SNR is proportional to the squared
projection height from a column vector to a non-degenerated space. In the following, we
present our results for general L , highlight the key challenges, and finally give our
conjecture and future direction.
Our main result for general L is given below, which should be compared with Theorem
2.1.

Replace RSL with RSL1 or RSL 2 for DF receivers.

38

Theorem 2.2: In an N R NT spatial multiplexing system with linear ZF/MMSE or ZF/MMSE

decision-feedback receivers satisfying NT L and N R L , if separately encoded data

streams are transmitted from L selected antennas, the optimal achievable diversity order is
bounded as
M L dopt M U ,

(2.49)

where M L = ( NT L + 1)( N R L + 1) , and M U = ( NT L + 1)( N R 1) . The optimal diversitymultiplexing tradeoff curve is bounded as:
r
r
M L (1 ) + d opt (r ) M U (1 ) + .
L
L

(2.50)

We will mainly focus on the extension for linear receivers; those for diversity order of DF
receivers and diversity-multiplexing tradeoff are relatively straightforward and will be briefly
discussed.
Employing the antenna selection method (2.10), we can then derive a similar outage
probability expression as (2.21):

Pr( RSL x) = Pr(min{Rk(1) } x, min{Rk(2) } x,

U1

U2

, min{Rk( NU ) } x)
U NU

(2.51)

N
= P r Ai ,
i =1

( j)
k

where U1 ~ U NU are defined in (2.1), R

in (2.6), and the N = L

similarly defined with a generic form (c.f. (2.22)):

39

NT

events { Ai } are

Aj = R (ji()i ) x , 1 j N ,

1i NU

where 1 j (i ) L denotes a pick from subset U i . Clear (2.23) still holds and we proceed
with the evaluation of upper and lower bounds.
Proposition 2.5: For general L , the optimal diversity order for linear receivers can be lower
L
bounded as dopt
( NT L + 1)( NR L + 1) .

Proof: From the definition of Ai and U1 ~ U NU , in any Si (the set of random variables in Ai )

we can always find a subset bearing the form Si _ indep = Rk(1j1 ) , Rk(2j2 ) ,

, Rk N NTL+L1+1 , with
(j

NT L + 1 random variables. In this form, Rk( j ) is slightly modified from Rk( j ) with k
indicating the index of column vector in the original channel matrix H , instead of H j , i.e.,
the squared projection height from h k (instead of h (k j ) ) to the subspace spanned by the
remaining column vectors in subset U j , with the implication that h k U j . Here k1 ~ k NT L +1
are some NT L + 1 distinct integers within [1, NT ] , and U j1 ~ U jN

T L +1

is an ordered list of

NT L + 1 different subsets from U1 ~ U NU . It is further required that h ki U ji , and h ki U jl

for l > i , i.e., h ki belongs to subset U ji but not those after. Therefore by Lemma D.1 in
Appendix D we can get the following upper bound for any Pr( Ai ) :
Pr( Ai ) Pr( Rk( j ) x)

NT L +1

Furthermore, since Rk( j ) is 2 (2( N R L + 1)) distributed, we have

40

(2.52)

Pr( Ai ) x ( NT L +1)( N R L +1) ,

(2.53)

and Proposition 2.5 is proved.

On the other hand, for L > 2 scenarios, the derivation of a tight lower bound for Pr( A1 ) is
much more involved as compared to the L = 2 case, because the angles { k(1) } are correlated
in a complicated manner, and a general form of their joint PDF expressions is not accessible.
Nonetheless, we have the following result.

Proposition 2.6: For general L, the optimal diversity order for linear receivers can be upper
L
bounded as d opt
( NT L + 1)( N R 1) .

Proof: Since the projection height from a vector to a subspace represents the shortest distance
from the vector to any point in the subspace, we have Rk( j ) Rkl( j ) , for any h (k j ) , h l( j ) U j ,
where Rkl( j ) denotes the squared projection height from h (k j ) to hl( j ) . It is then not difficult to
build up the following lower bound:

Pr( A1 ) = Pr { R1( i ) x} Pr { R12(i ) x} .

1i NU

1i NU

(2.54)

Carefully examining the first two elements in all subsets (see (2.1)) reveals that the last
expression in (2.54) bears a similar form as the Pr( A1 ) in L = 2 case (see (2.22)), replacing

NT 1 with NT L + 1 . Following the same lines as in Section 2.2.3, we have for general L

41

Pr( RSL x) Pr( A1 ) x ( NT L +1)( N R 1) .

(2.55)

Combining the above two propositions, (2.49) is proved for linear receivers. Also, as for
the L = 2 case it is straightforward to show that (2.49) applies for the first decoded data
stream of DF receivers. Since the diversity order of the first decoded data stream is upper
.

bounded by ( NT L + 1)( N R 1) and Pe Pe1 , the upper bound ( NT L + 1)( N R 1) also

applies for Pe with DF receivers. On the other hand, by employing the QR based selection
algorithm for DF receivers, a diversity order lower bound ( NT L + 1)( N R L + 1) is
achieved. Therefore (2.49) also holds for DF receivers. Finally the derivation of (2.50)
follows the same method as given in Section 2.2.3.3, and Theorem 2.2 follows.

Remarks: Note that when L = 2 , the two bounds in (2.49) and (2.50) coincide and conform to
the results obtained in Section 2.2.3. A conjecture on the diversity order of separately
encoded SM systems with transmit antenna selection and linear ZF receivers was made in
[36] based on numerical results, which actually has motivated our research: for linear ZF

receivers, when N R = L , the achievable diversity order is NT L + 1 . Our results prove its
correctness and further extend it to general N R .

42

Figure 2-6 The Exponential Behaviors of the Outage Probabilities for the
N T = N R = 4, L = 3 Scenario
In Figure 2-6 we investigate the exponential behavior of Pr( RSL x) and Pr( A1 ) for an exemplary
scenario N T = N R = 4, L = 3 through simulations. By comparing with the reference curve x 4 , we
find that both of them present a diversity of M L = 4 , verifying its achievability as we claim in
Theorem 2.2. On the other hand, it shows that in this scenario the probability lower bound Pr( A1 )
also achieves the diversity lower bound M L , revealing that the diversity orders between M L + 1 and
M U may actually not be achievable. We thus have the following conjecture, which constitutes part of
our future work. The key seems to lie on finding a better lower bound of Pr( A1 ) than (2.54).
Conjecture: for general L , the optimal transmit antenna selection achieves a diversity order exactly
equal to M L .
43

2.2.5 Summary

In Section 2.2 we have analyzed the diversity order achieved by transmit antenna selection
for separately encoded SM systems with linear and DF receivers. Using a geometric
approach, we have rigorously derived their achievable diversity order for the L = 2 scenario.
We have also used the same geometrical approach to obtain bounds on the achievable
diversity order for general L . Our results prove and extend the previous conjectures in
literature drawn from simulations, and verify the predicted potential of antenna selection for
practical spatial multiplexing systems. Furthermore, the proposed geometrical approach may
be also used to solve other open problems related to MIMO.
Besides verifying the conjecture for general L , the analysis for maximum-likelihood
receivers, joint transmitter and receiver selection, and multiuser MIMO scenarios also direct
our future research.

2.3 Antenna Selection Algorithm Designs: Fast Transmit Antenna

Selection in Correlated MIMO Channels
2.3.1 Introduction

In the previous section, we investigate the achievable performance improvements by

deploying antenna selection in MIMO systems with practical transmitter and receivers. As
we mentioned earlier, we do not claim the practicality of the selection algorithm used for
diversity analysis. In another word, in some scenarios (e.g. large antenna array dimensions)
N
the AS method brute forcing all the T possible antenna subsets for the strongest weakest
L

44

link (or other similar brute forcing methods like in [29][36]), may be too cost-inefficient to
be applied.
In this section, we try to look into the other important topic of antenna selection: the
selection algorithm designs in MIMO-SM systems.
The most common antenna selection algorithm design comes from the one that
maximizing the Shannon capacity for the selection subchannel. Hence the optimal selection
comes from the exhaustive search over all possible antenna subsets [29][56]. On the other
extreme, the simplest algorithm called power-based selection, selects the antennas with the
largest channel gains and performs well only at low SNR regime [52][54] (also note that,
PBS is optimal for AS in MIMO spatial diversity schemes, as discussed in Section 2.2.1 and
[1][30][37][79]). Some other algorithms are proposed with good trade-offs between
performance and complexity, for example, the novel near optimal iterative algorithms in
[24][31], and the fast algorithm called correlation-bases selection in [6]. All the above
algorithms require instantaneous channel state information, and may present some difficulty
in channel environments with fast fading or high mobility. On the other hand, in realistic
outdoor channels, the channel correlation is a function of the local scattering environment
and thus varies on a much slower time scale than instantaneous channel coefficients.
Compared with instantaneous CSI-based AS, there is relatively little work in the literature
done on antenna selection for capacity maximization based on channel statistic information.
For error rate minimization in spatial multiplexing systems, we can find antenna selection
algorithms based on both instantaneous CSI (e.g. [36]), and channel correlation matrices (e.g.

45

[28][57]). An important figure of merit for error rate minimization is the maximization of the
post-detect SNR in the weakest sub-stream, as discussed in the previous section. It is also
noteworthy that in most of the time the antenna subset for capacity maximization and error
probability minimization are derived almost simultaneously, as will be explained in detail
later on in this section. Finally, a noticeable work comes from [84], which is sometimes
called beam selection, and essentially generalizes antenna selection systems as mapping from
RF chains to antenna elements by a matrix composed by arbitrary phase-shift-only values,
which is realizable in the RF domain. Then the AS algorithms in [84] boil down to adapting
the mapping matrix based on full or partial CSI (i.e. the channel correlations).
In this section, our attention is paid on some fast AS algorithm designs under realistic
MIMO channels, which usually present fading correlations. Motivated by the matrix
determinant properties of Hermitian positive definite matrices, fast transmit antenna selection
algorithms are explored considering both capacity and error rate, based on both instantaneous
CSI and channel correlation matrix. Specifically, Gram-Schmidt algorithm performs near
optimal, while the novel G-circles algorithm achieves many advantages over all the other
existing schemes.
2.3.2 System model and Problem Formulations

As in the previous section, we suppose that there are NT transmit and N R receive
antennas in a spatial multiplexing MIMO system. In a block fading channel model, we select

L from NT transmit antennas and connect them to the L available RF chains. No CSI is
available at the transmitter, so the selection might be implemented at the receiver, and the
46

selected antenna indices are fed back so that the transmitter can equally allocate its power
among L selected antennas. We denote H as the N R N T complete channel matrix, and H SL
the selected N R L sub-channel. The capacity after antenna selection can be expressed as:

C = log 2 det I L + SL
SL .
L

(2.56)

where I L is an L L identity matrix, is the average received SNR at each receive antenna,
and the operator

represents transpose conjugate. When the transmitter side is under fading

correlation (e.g. urban outdoor downlink channel), the corresponding channel matrix can be
modeled as [37]:

H = HW R1/T 2

(2.57)

where R T = E H H H is the NT NT transmitter correlation matrix, and HW is an N R NT

normalized white Gaussian matrix. Note that in (2.57), HW models the fast Rayleigh fading,
while R T is mainly determined by the geometrical structure of the propagation channel and
can be assumed to be unchanged over a much longer time-scale. The fading correlations can
be expressed as:

[ RT ]ik = exp 2 j (i k )
S

s =1

cos s ,

(2.58)

where is a real positive scalar, T is transmit antenna spacing, is the wavelength, S is

the number of major far-field scatterers at the transmitter side, and s represents the direction
of departure for the sth far-field scatterer. Suppose that S >> rank ( H ) , and T is chosen to

47

be large enough compared with , the key factor influencing the channel conditioning is the
range of DOD, or the angle spread of transmit scatterers.

H
SL . We can then
Maximizing (2.56) is equivalent to maximizing det ( L / )I L + SL

build a composite matrix:

G=
,
L
/

NT

(2.59)

and let G SL be the corresponding sub-matrix after transmit antenna selection, whose first N R
rows form the N R L channel matrix H SL . By further defining
H
Z SL = G SL
G SL ,

(2.60)

we can easily get that optimal selection is actually the exhaustive search for L columns in G
that maximize det ( Z SL ) . Assuming that L rank (H ) , then Z SL is a Hermitian positive
definite square matrix, which contains real positive eigen-values [51].
Also det ( Z SL ) bears the following properties:

Property 2.1:

det(Z SL ) = ( Rll )2 , where {Rll } are diagonal values of the upper

l

triangular matrix R in the QR decomposition of G SL .

Property 2.2: det(Z SL ) = Z _ l , where {Z _ l } are the eigen-values of Z SL .

l

In the following parts of this section, fast antenna selection algorithms will be explored based
on these two properties. These algorithms target on either capacity maximization or error

48

probability minimizations. Similar as in Section 2.2, these algorithms are proposed or

explained from a novel geographic viewpoint.

2.3.3 Gram-Schmidt (GS) Based Algorithm: A Geometric Analysis

From Property 2.1, to maximize the capacity, L transmit antennas with maximum Rll
l

are selected. The GS method, or named as QR-based method as discussed in Subsection

2.2.3.2 is suboptimal with greatly reduced complexity, which was first discussed in [24]. In
each step, one antenna with the largest Rll is selected. The QR decomposition G SL = QR , is
actually the complete road map of the Gram-Schmidt process, in which R is an L L upper
triangular matrix, whose lth diagonal value Rll is a real positive number representing the
projection height of the projection from lth column vector g l , to the space spanned by

g1 ~ g l 1 , denoted as span {g1...g l 1} . We can further explain the GS algorithm from a

geometric standpoint (which is partly stated in Subsection 2.2.3.2), based on the following
observation:
Observation 2.1: The optimal transmit antenna selection for capacity maximization is

equivalent to selecting L antennas such that the volume of the L-dimensional parallelotope
generated by the L selected column vectors in G SL is maximized.

From Property 2.1 and the analysis of Rll above, we can easily get that such volume equals to
Rll . See Figure 2-7 for two and three dimensional examples.
l

49

Figure 2-7 Two and three dimensional examples of Gram-Schmidt process

In step l of the GS method, we select one antenna whose corresponding channel column
vector has the largest projection height to the space spanned by the previously selected
column vectors, so that the volume of l-dimensional parallelotope can be maximized. Also
note that a large projection height in step l generally requires selecting g l with both a large
norm and small correlations with g1 ~ g l 1 , which results in a relatively well-conditioned

H SL (i.e. the ratio of the largest singular value to the smallest one is not large) with large
channel power, so that the capacity in (2.56) can be maximized.
On the other hand, if H SL approximated maximizes the channel capacity, with a high
probability it also achieves minimum error rate for SM-MIMO schemes (V-BLAST structure
[22][95]), because intuitively the channel matrix is well-balanced on each data stream, so that
the weakest link, which usually determines the system error probability, is strong.
Specifically for an V-BLAST SIC receiver, which essentially takes QR decomposition of the
(ordered) selected channel matrix H SL before the successive interference cancellation

50

procedure [57][95], the GS method is a good choice for the error rate minimization, because
in general the smallest Rll , which represents the weakest data link, can be approximately
maximized. This algorithm is fast because only vector/scalar multiplications and additions
are involved. However, it requires instantaneous CSI and cannot be implemented only based
on channel statistics.

2.3.4 The Gerschgorin Circles (G-circles) Based Algorithm

From Property 2.2, to maximize the capacity, L transmit antennas with the largest Z _ l
l

H
are to be selected. Denoting {l } as the eigen-values of the square matrix T( L ) = H SL
H SL , it is

easy to prove that Z _ l = l + L / . Therefore, we shall select H SL with L eigen-values

{l }l =1... L as large as possible to maximize Z _ l . As in [6][24], we develop a successive
l

selection strategy, in which one antenna is selected at each step. Denoting T( l ) = (H ( l ) ) H H ( l ) ,

where H ( l ) = h (1) , h (2) ,..., h ( l ) with

(l )

the selected antenna index in step l, and H SL = H ( L ) .

Let 1(l ) 2(l ) ... l(l ) be the eigen-values of T( l ) . We have [51]:

1(l ) 1(l 1) 2(l ) 2(l 1) ... l(l11) l(l ) .

(2.61)

H
H
H SL ) is upper bounded by max ( H H H ) , and a large min ( H SL
H SL )
Therefore max ( H SL

basically results in a small channel conditioning number: = max / min , desirable for
capacity maximization, especially at high SNR. To avoid investigating the eigen-values of all
H
H SL ) . It can be
possible antenna subsets, we will focus on the approximation of min ( H SL

51

proved by (2.61) that min is decreased after each selection. Our strategy is: in each step we

select one column in H so that the decrease of min can be approximately minimized.
The G-circles theorem gives us an approximation of the eigen-value distributions of T( L )
[51]:
H
Theorem 2.3 (G-circles): The L eigen-values of T( L ) = H SL
H SL are trapped in the circles

centered at T( L ) with radii given by the following expression:

ll

l [T( L ) ]ll rl = min [T( L ) ]lk , [T( L ) ]kl .

k l

Since T( L ) is Hermitian, we get

[T
k l

( L)

k l

(2.62)

]lk = [T( L ) ]kl . From the definition of T( L ) , we

k l

can further simplify (2.62) by:

l h( l )

rl = h(Hk ) h( l ) .
k l

Figure 2-8 An example of G-circles

52

(2.63)

An example of G-circles for T(3) can be found in Figure 2-8. Since the eigen-values are all
real positive numbers, they are actually distributed on the real axis within the range of Gcircles. From (2.63), a large center of the lth G-circle represents a large channel gain of
transmit antenna (l), while a small radius means that antenna (l) has low correlations with all
the other selected antennas. A lower bound of min is found from (2.63):

min min h( l ) h(Hk )h( l ) ,

k ,l =1... L
2

k l

(2.64)

which is the left most point among all the L G-circles. To approximately minimize the
decrease of min in each step, we maximize the lower bound of min in (2.64), motivating the
following algorithm, where (l ) and (l ) represent selected and remaining antenna sets at
step l respectively:
Algorithm 2.1 (G-circles):

Select h (1) from H with the largest norm h (1) , (1) = {(1)}, (1) = {i}i (1)
For l=2:L
For i (l 1)

temp = (l 1) , i

LBi = min h( k )
j ,ktemp

H
( j ) h( k )

j k

End
Select (l ) = arg max LBi

(l )

( l 1)

,(l ) , (l ) (l 1) {(l )}

End

53

Initially, we select the antenna with the largest channel gain. In the following steps, selecting
one more antenna results in adding one more G-circle and the expansion of the radii of
existing G-circles. From a geometric viewpoint, the maximization of (2.64) requires selecting
one antenna with large norm and small fading correlations with all the other selected
antennas. G-circles method is simple compared with existing algorithms, because only vector
multiplication and scalar additions are involved. In particular, it is also simpler than the CBS
algorithm in [6], which requires the calculations of the correlation between any two possible
transmit antenna candidates.
On the other hand, in fast fading channels, instantaneous CSI-based transmit antenna
selection is hard to be implemented, with the concerns of both channel variation and training
overhead. Therefore it is desirable to develop antenna selection algorithms based on slowly
varying channel statistics. At high SNR, we can approximate (2.56) by the equation
H
H
det( L / I L + H SL
H SL ) det(H SL
H SL ) . If only fading correlation information is available for

H
antenna selection, we need to maximize det(H SL
H SL ) based on the knowledge of R T . The

selected channel is then H SL = HW ( R1/T _2 SL ) , where R T _ SL is a certain principle minor of

H

R T corresponding to the selected antennas. From matrix determinant property [51], we can

get the following expression:

H
det H SL
H SL = det R1/T _2 SL HWH HW R1/T _2 SL

((
= det ( R

= det R1/T _2 SL
T _ SL

54

R1/T _2 SL HWH HW

) det ( H

H
W

HW

).

(2.65)

When only R T is available for antenna selection to maximize (2.65), it is natural to get the
following fact, which is also formally proved in [84]:
Fact 2.1: To maximize (2.65) if only R T is available, L transmit antennas are chosen

such that det ( R T _ SL ) is maximized.

The G-circles method is also applicable in this scenario. From Fact 2.1 and the property

det(R T _ SL ) = RT _ l , one possible method for capacity maximization is to maximize

l

min (R T _ SL ) by the exhaustive search over all possible L L principle minors of R T . To

avoid such exhaustive search, we further simplify it by the G-circles of R T _ SL :

Algorithm 2.2 (G-circles - RT ) :

a = [R T ]ii i (i.e. identical diagonals in R T )

Select h (1) and h (2) such that [RT ](1),(2) is the minimum
Follow the same procedure as Algorithm III. 1, to select antennas (3)~(L), using
[RT ]( j ),( k ) instead of h (Hj ) h ( k ) , and a instead of h ( k )

in (2.64).

Note that in step 1, since all the diagonals of R T are identical, we do not have the freedom to
select the antenna with the largest channel gain. Therefore initially we simply select two
antennas with the smallest correlation.

55

2.3.5 Numerical Results

We simulate a MIMO system with N R = 3, NT = 9 and L = 3 . In (2.58), we assume that

T = 2 , S = 10 , and 1... S are uniformly distributed in the range ( / 2, / 2) .

Therefore, the channel conditioning number largely depends on .

Figure 2-9 Capacity performances for well- and ill-conditioned channels

In the first simulation (Figure 2-9), capacities achieved by instantaneous CSI-based

antenna selection is investigated. Two extreme conditions: = 180o , and = 15o , are
simulated representing well- and ill- conditioned channels, respectively. We see that the Gcircles method is near optimal for well-conditioned channels. For ill-conditioned channels,
the G-circles method preserves the spatial multiplexing gain, and presents a reasonable

56

performance loss (about 1dB). It also yields uniformly better performances over the CBS in
[6], especially for well-conditioned channels.

Figure 2-10 Capacities of RT -based selections w.r.t. different

In the second simulation, we investigate RT -based antenna selection algorithms. The

transmit SNR is fixed to be 20dB and the capacity curves are drawn with respect to .
Also, we simulate the selection scheme that makes the exhaustive search for maximum

det R T _ SL in (2.65), which in general represents the optimal RT -based antenna selection for
capacity maximization (c.f. Fact 2.1). From Figure 2-10, we see that all the RT -based
algorithms perform near optimal for ill-conditioned channels. The RT -based G-circles
method even outperform its counterpart based on instantaneous CSI for channels with
57

 < 20o , since for ill-conditioned channels RT _ SL dominates the eigen-value distributions
H
in H SL
H SL .

Finally we investigate the BER for the proposed antenna selection schemes with SIC
receivers in V-BLAST systems. QPSK modulation is employed. Generally speaking, the
error rate performances have trends similar to their capacity counterparts. From Figure 2-11,
we can see that in well-conditioned channels, the G-circles method greatly outperforms CBS.
H
The GS method performs close to the exhaustive search for the largest min (H SL
H SL ) for any

channel conditioning situation. For ill-conditioned channels, the RT -based G-circles method
performs close to the exhaustive search for the largest min (RT _ SL ) and achieves considerable
selection gain over the MIMO system without antenna selection.

58

(a) A well-conditioned channel

(b) An ill-conditioned channel

Figure 2-11 BER performances for well- and ill-conditioned channels
59

2.3.6 Summary

In this paper, motivated by some matrix determinant properties of Hermitian positive definite
matrices, we develop fast MIMO transmit antenna selection algorithms considering both
capacity and error rate. The GS method is shown to achieve near optimal performances for
any channel conditioning situation. For the simpler G-circles method, compared with optimal
selection it reduces the complexity significantly with reasonable performance loss; compared
with the GS method, it is simpler and can be implemented only based on RT ; compared with
CBS, it is simpler with uniformly better performances and can be implemented for RT -based
selection.
Compared with Section 2.2, this section looks into the practical AS algorithm design
problems. Although not achieving optimal capacity/diversity/error probability performances,
the proposed algorithms provide good tradeoffs between performance and complexity. On
the other hand, we stress that in many practical MIMO systems with relatively small number
of antenna elements, optimal performance achieving methods (e.g. those mentioned in
Section 2.2, or in [29][36]) can be readily applied without imposing much computational
burden.
Until this point, we have explored the potential of antenna selection, and how to design
effective AS computation algorithms. Then naturally we are facing the following question:
how to encompass AS into commercial wireless systems, or more specifically, into the
wireless standards. As one particular solution in WLAN, we will present the training and
calibration methods of AS in the emerging IEEE 802.11n standard.

60

2.4 Implementing Antenna Selection in Realistic Wireless Networks

2.4.1 Introduction

As discussed above, antenna selection is a promising technique which explores the diversity
gain of MIMO channels while maintaining spatial multiplexing gain. It is then of particular
interest for us to address the problem of how to implement it into realistic wireless systems.
On the downside, to perform antenna selection the complete channel matrix is estimated
by sending training frames to measure the complete CSI, which causes increased overhead.
Therefore the core issue of antenna selection implementations is how to make the balance
between reducing AS training overhead and maintaining appreciable performance
enhancement caused by AS. The basic answer is: the AS training protocol designs should be
closely related to the baseline PHY and MAC protocols and the wireless environment
properties. To illustrate this idea, we present our AS training/calibration protocol proposed
for the emerging IEEE 802.11n High-Throughput WLAN standard [105].
WLANs based on the IEEE 802.11 standard [53][60], currently are one of the hottest
sectors of the wireless market. While the current IEEE 802.11a standard, which is based on
OFDM, is limited to data rates of 54 Mbit/s , the emerging IEEE 802.11n high-throughput
WLAN standard combines OFDM with MIMO techniques to achieve effective data rate in
excess of 100Mbps, as observed at MAC layer service access point [105].
The modem design of 802.11n employs closed-loop schemes, which require channel-state
information at the transmitter. Since the channels in WLANs exhibit slow time variations
61

(low Doppler spread around 5Hz [107]), the transmitter CSI, obtained by feedback or
channel estimation in the reverse link, does not get stale at the instance of transmission,
making it well suited for implementing closed-loop MIMO algorithms. One good example is
transmit beamforming with fast link adaptation techniques as foreseen in the current standard
draft [105]. Another example for closed-loop MIMO is transmit AS, which is the focus of
this section.
Antenna selection is particularly beneficial for WLANs for the following reasons:
1.

The obtained diversity gain may significantly boost the performance

especially at high SNR, i.e., the SNRs that are typical for WLANs.

2.

The slowness of the variations of WLAN channels greatly reduces the

overhead required for antenna selection training.

A potential problem of antenna selection is the increased hardware effort that is required
by the antenna selection training. If all possible antennas are trained at the beginning of each
packet, fast solid-state switches, and fast AGCs are required, leading to large switching loss
[83]; furthermore, the PHY-layer protocols have to modified considerably from the nonselection case.
In the subsequent subsections, we introduce a MAC-based AS training protocol that
eliminates all of these problems. In our scheme, different antenna subsets are trained in
multiple data packets (burst), and signaled only in the MAC headers. Numerical results show
that switching antenna subsets between packets does not significantly reduce the
performances, thanks to the low Doppler spread of WLAN channels.

62

Another important issue of implementing AS, largely ignored previously in the literature,
is the RF imbalance caused by antenna switching, because essentially different combinations
of RF chains and antenna elements may induce non-identical channel gains in the equivalent
baseband channels. We then address this problem by proposing a novel calibration
procedure.
By our efforts, the proposed training/calibration protocol has already been accepted in the
current IEEE 802.11n draft specifications [105].
In the following, after short descriptions of 802.11n MIMO-OFDM PHY and MAC
structures, we introduce the MAC-based WLAN AS training protocol; then we provide the
solutions used for addressing AS with RF imbalance, followed by numerical results.

Figure 2-12 MIMO-OFDM PHY

63

2.4.2 Introduction of 802.11n PHY and MAC

The MIMO-OFDM PHY structure of 802.11n WLAN is illustrated in Figure 2-12:

The MAC layer data payload (in frequency domain) is at first scrambled to avoid consecutive
0s and 1s, so that better coding performances can be achieved; then after coding and
puncturing, the original data is de-multiplexed into parallel data streams, each of which is
interleaved (in both time and frequency) and modulated into QAM symbols; the optional
orthogonal space-time block coding provides extra diversity gain on each of the streams, and
convert the input data streams into output space-time data streams; then a spatial mapper
maps STDS onto the RF chains. The mapper can be a simple spatial spreading matrix in
open-loop systems, or TX BF in closed-loop systems. After mapping, IFFT, guard interval
insertion, time-domain windowing and up-conversion are applied in sequence on each RF
chain; finally the RF signals are mapped to antenna elements for transmitting, via the antenna
selection switches.
As we repeatedly stressed and can be seem in Figure 2-12, AS is not an alternative with
other diversity achieving schemes like STBC and TX BF. It can be jointly employed with
them to improve the performances, as it is operated in the RF domain, while the other two are
in the baseband. Therefore, at least AS is beneficial regarding performance improvements,
even though the hardware improvements will finally make RF chains as cheap as antenna
elements.
The 802.11n PHY preamble in each PHY layer packet is illustrated in Figure 2-13, in
which the legacy portion is used for both backward compatibility with legacy 11 a/g devices,
and the time/frequency synchronizations at the receiver; while the HT portion are used for
64

achieving high-throughput, which contains both PHY layer transmission feature signaling,
and HT-STFs, HT-LFTs for receiver AGC setting and MIMO channel estimations in each
subcarrier. Finally the data payload follows the preambles in each PHY layer packet.

Figure 2-13 11n PHY Preamble

The MAC layer signal of 802.11n contains a MAC header and data payload. Except the
conventional contents like transmitting and receiver address, the most PHY layer packets
encompass a HT control field in their MAC headers. Such a field include the necessary
information used for achieving high-throughput features, e.g. TX BF, link adaptation and
other advanced MAC improvements. Our proposed AS training/calibration protocol will also
utilize this field. The CSMA/CA protocol in MAC theoretically guarantees that at any time
instance one station only communicates with a single peer station (like TDMA), so that we
only need to build up the system model based on point-to-point MIMO-OFDM.

65

2.4.3 Mathematical Model for Antenna Selection

Now we move back to antenna selection in high-throughput WLAN. For the ease of analysis
in this section, we re-define the mathematical model:
In the MIMO-OFDM system applying AS (Figure 2-1), the transmit STA A has a set of

N A antennas with nA transmit RF chains, while N B and nB are similarly defined at the
receive station B. In general each AS training cycle consists of an AS training phase and a

data transmission phase, as can be seen in Figure 2-14 below:

Figure 2-14 Illustration of an AS Cycle

Several AS training fields are transmitted in each AS training phase, each of them is
transmitted from and/or received by one subset of antenna elements to be selected. The
antenna selection computation is based on the complete channel matrix composed of the
subchannels estimated from all the AS training fields. In the data transmission phase, a
relationship between a transmitted signal and a received signal in one subcarrier (for
denotation simplicity we omit the subcarrier index here) can be expressed as:

rB = FBH [ H AB FAs A + n] ,

66

(2.66)

where rB is a n B 1 received signal vector, s A is a nA 1 transmitted signal vector, and

H AB is a N B N A equivalent channel matrix containing complete physical channel

responses and the effect of transmit and receive RF responses. A noise vector n has N B 1
entries that are i.i.d. zero-mean circular complex Gaussian random variables with
variance N 0 . FA is a N A nA transmit antenna selection matrix, and FB is a N B nB receive
antenna selection matrix. Both FA and FB are submatrices of an identity matrix, representing
antenna selection. The equivalent channel matrix after antenna selection is a nB nA matrix
H eq = FBH H AB FA , which is a submatrix of the channel matrix H AB . Different from our

previous discussions and from those in the literature, now we consider the realistic case, in
which the equivalent channel H AB also includes the impact of the RF responses:

H AB = C B , Rx (FB )H AB C A,Tx (FA ) ,

(2.67)

where H AB is the actual propagation channel, C A,Tx (FA ) is a N A N A diagonal matrix whose

i-th diagonal element [C A,Tx (FA )]ii collects the RF response corresponding to the i-th transmit
antenna element, which is a function of the antenna selection matrix FA : If the i-th row in FA
contains all zeros, the i-th antenna is not selected, so [C A,Tx (FA )]ii = 0 ; If the element at the i-

th row and l-th column of FA is one, the i-th antenna is selected and is connected to the l-th
transmit RF chain during the data transmission phase. Then [CA,Tx (FA )]ii = li(Tx ) , which is a
complex number characterizing both the amplitude and phase shift of the RF response (seen

67

at baseband) corresponding to the connection between transmit RF chain l and antenna

element i. C B , Rx (FB ) is similarly defined: [C B , Rx (FB )] jj = li( Rx ) if the element at the j-th row
and l-th column of FB is one.
On the other hand, in the m-th AS training field, a relationship between a transmitted
signal and a received signal can be expressed as:

rB ,t ( m) = TBH ( m)[ H AB TA ( m)s A ,t + n ] ,

(2.68)

where s A,t and rB,t are the training and received vectors; TA (m) and TB (m) are the
predetermined antenna mapping matrices in the m-th AS training field, indicating the
connections of all the available RF chains to the m-th antenna subset. All these antenna
subsets

are

typically

exclusive

with

each

other.

For

example,

if

N A = 4, nA = 2, N B = 2, nB = 2 , we have 2 training fields with the transmit antenna mapping

matrices:
1
0
TA (1) =
0

0
0

0
1
and TA (2) =
1
0

0
0

0
0
.
0

Then there are totally M = N A / nA N B / nB training fields, where x is the smallest

integer larger than or equal to x. STA B then can estimate the complete channel matrix
(which will be used for AS computations) by combining the M subchannels. Consequently,
by ignoring channel estimation errors, the estimated subchannel by training field m is

68

HAB (m) = TBH (m)C B , Rx (TB (m))H AB C A,Tx (TA (m))TA (m) ,

(2.69)

and the AS computation is conducted based on the following estimated complete channel
matrix:
HAB = CB , Rx H AB CA,Tx ,

where

the

diagonal

matrix

CA,Tx

contains

(2.70)
all

non-zero

diagonal

values:

[CA,Tx ]ii = [CA,Tx (TA (m))]ii , if the i-th antenna element is trained by the m-th training field, and
CB , Rx is similarly defined. Therefore AS computation is based on the estimated complete

matrix HAB , i.e. by a certain AS criteria X , the selection can be expressed as:
{FA,opt , FB ,opt } = arg max X (FBH HAB FA ) .

(2.71)

FA , FB

For example, if the criterion is the maximization of the capacity, X(A)=log|1+AAHSNR/nA|.

If FA,opt , FB ,opt are selected based on the training phase, the equivalent channel in the data
transmission phase becomes

H eq = FBH,opt C B , Rx (FB ,opt )H AB C A,Tx (FA,opt )FA,opt .

(2.72)

Then X (H eq ) may not be optimal, because the RF responses of the used RF chains are
different in the two phases. This effect is called the RF imbalance. In the example of

N A = 4, nA = 2, N B = 2, nB = 2 (i.e. only STA A conducts AS), the selection is determined by

69

11(Tx )

( Tx )
22
( Rx )

.
= C B H AB
H AB

13(Tx )

24(Tx )

(2.73)

where C B , Rx is always fixed given FB = TB = I . If antennas 1 and 3 are selected at STA A,

during data transmission phase,

H AB

11(Tx )

( Rx )
= C B H AB

23(Tx )

(2.74)

(Tx )
, and transmit antennas 1 and 3 may not be
there will be a distortion caused by 13(Tx ) 23

the optimal subset. For simplicity and without loss of generality, we henceforth use the
following constraint: for any selected antenna subset, a RF chain with smaller index number
always connects to an antenna with smaller index. With this constraint, in both the AS
training phase and the data transmission phase there are totally nA ( N A nA + 1) possible
connections of RF chain with antenna element at STA A, and all the possible RF responses
can be expressed as:

a11(Tx )
a12(Tx )

a22(Tx )
a23(Tx )

a1((TxN) n

a2((TxN) n

A +1)

A + 2)

70

an(Txn )
an(Tx( )n +1)
A A

an(TxN)
A

(2.75)

2.4.4 MAC-based AS Training Protocol

Given the above introductions on 802.11n PHY and MAC structure and AS models, we
are ready to present the proposed AS training protocol. The widely accepted method for AS
training is the one that sending consecutive training fields one right after another, so that the
channel variation on these training fields is minimized. This corresponds to increased number
of STFs and LTFs in the PHY preamble, as illustrated in Figure 2-13. However, this
scheme will definitely increase the overhead, and brings in burdens on PHY signaling, which
explains the reason why it is hard to be accepted in the standard [105]. Moreover, this
scheme will lead to the deployment of fast solid state antenna switches (in nano seconds),
which will cause appreciable switching loss [83]. On the other hand, the low Doppler spread
of WLAN channels (see Figure 2-15 for a WLAN channel realization [107], where channel
model B means one specific model with relatively low frequency selectivity) allows us to
propose a MAC-based AS training, in which the AS training phase is formed by a sequence
of M consecutive training packets, each containing one of the M AS training fields
transmitted from and/or received by one of the M disjoint antenna subsets. All the training
information is signaled in MAC headers, where the dedicated high-throughput control field is
already defined for signaling the new MIMO high-throughput features such as transmit
beamforming and fast link adaptations [105], as introduced in Subsection 2.4.2. Therefore the
proposed training method greatly reduces the required modifications in the standard.

71

Figure 2-15 Time Variation in WLAN Channel Model B

Specifically, these training packets should be sent in burst, as illustrated in Figure 2-16, an
example of transmit AS training only. This protocol can be described as follows: the receiver
may choose to initiate the AS training cycle by sending a transmit AS request (TXASR),
whenever the current selection result gets stale. Or the transmitter can initiate its own AS
training cycle at a predetermined time, or when it observes more frequent re-transmissions of
packets. Then the transmitter sends out M = N A / n A consecutive AS training packets with
SIFS, equal to 16 s [105], each containing the regular long and short OFDM training fields
in its preamble as defined in [105], and transmitted from one subset of nA antennas. On
receiving these packets, the receiver conducts channel estimations to establish the complete
channel matrix in each subcarrier. Finally the receiver may either implement AS computation

72

and feedback the selected antenna indices, or directly feeds back the complete channel
matrices for the transmitter to conduct the selection.

Figure 2-16 Transmit AS Training Example

The receiver AS training process can be similarly defined, except that now different AS
training packets are received by different receive antenna subsets. When both sides conduct
antenna selection, the two training processes can be done one after another. Note that these
AS training packets may also contain data payload. In that case, some back-off strategies,
e.g. applying the lower level of modulation and coding schemes, are necessary, because the
link adaptation output regarding the previously selected antenna subset is not valid after
switching. The time available for switching the antennas is now one SIFS, allowing to
implement the MEMS based switches, which have much lower switching attenuation than
solid-state switches [74].

73

2.4.5 AS with RF Imbalance

To recover the possible performance degradation caused by the RF imbalance problem as

introduced Subsection 2.4.3, one solution is to train each possible antenna subset during each
of the AS training phases. However, this obviously is impractical for large numbers of
antenna elements and/or RF chains, and may induce more distortions due to channel Doppler
(longer AS training phase).
Alternatively we propose a simple calibration process. Since the RF responses vary with
the environment (e.g. carrier frequency drift, temperature variations), an over-the-air
calibration process is necessary. On the other hand, the overhead for calibration is negligible
because it needs to be conducted only at large time intervals, e.g., only upon station
association, or when the environment varies.
Figure 2-17 shows the calibration procedure for transmit AS. The transmitter (STA A)
sends consecutively N A nA + 1 AS calibration training packets, each transmitted with the RF
chain/antenna connections according to one single row of (2.75). For example, the first
training packet uses the connections:
RF 1 Ant 1, RF 2 Ant , ... ,RF nA Ant nA .

74

Figure 2-17 Transmit AS Calibration Process

On receiving these training packets, the receiver (STA B) estimates the corresponding
subchannels, denoted as HAB (1)...HAB ( N A nA + 1) , and feeds them back after receiving all the
training packets. The transmitter then determines its RF imbalance correction coefficients
based on all the estimated subchannel matrices fed back from STA B. When STA B also
conducts receive AS, i.e. N B > nB , it should use a predetermined subset of receive antennas,
each connected to a predetermined receive RF chain on receiving all the training packets in
Figure 4. The correction coefficients are determined as follows: by ignoring channel
estimation errors and assuming static channel during the transmission of all the AS
calibration training packets,

75

(11)
(22)
hAB
hAB
,11
,12
(11)
(22)
hAB ,22
h
HAB (1) = AB ,21

(11)
(22)
hAB
,n 1 hAB ,n 2
B

(n n )

hAB
,1n

(n n )
hAB
,2 n
,

(n n )
hAB ,n n
A A

A A

(2.76)

A A

B A

( li )
( Rx )
(Tx )
where h A B , j i = l j h A B , j i li stands for the equivalent channel coefficient involving
B

all the RF responses; hAB , jB i is the actual physical channel coefficient from transmit antenna

i to receive antenna j B , which is connected to the predetermined receive RF chain lB , with

l( Rjx ) the corresponding receive RF response. H AB (2)...HAB ( N A n A + 1) can be expressed

B

similarly based on different transmit RF chain and antenna element connections following
the corresponding rows in (2.75). For the i-th transmit antenna, we then do the following
calculation:

li =

h A(mB in, n{B Li i } i )

h

( li )
AB ,nB i

l( Rjx ) h A B , j i m( Tinx {) L } i m( Tinx {) L } i

=
=
,
l( Rjx ) h A B , j i li( T x )
li( T x )
B

for every l Li , where Li is the set of RF chain indices that are possible to be connected to
antenna i according to (2.75). Then, li is multiplied with the baseband signal transmitted
from RF chain l, whenever it is connected to antenna i. As a result, any transmission from
(Tx )
antenna i leads to a corresponding transmit RF response min{
Li }i . As special cases, transmit

antennas 1 and NA are always connected to RF chain 1 and nA , respectively following the
constraint in (2.75), so no correction is needed for the transmissions from them. By doing the

76

same calculations and by applying the results for all transmit antennas, at any time the
equivalent complete channel matrix can be expressed as:

H AB = C B ,Rx H AB diag{11(Tx ) , 12(Tx ) ,

( Tx )
, min{
,
L }i
i

, n(TxN ) } ,
A

there is no distortion between the AS training phase and the data transmission phase. Note
that these correction coefficients are applied in both the AS training phase and the data
transmission phase, and is equivalent to replacing the 1s in FA or TA(m) by the
corresponding correction coefficients { li } . The above calculations can be repeated nB times,
corresponding to jB = 1

nB respectively. The resultant nB sets of correction coefficients can

then be averaged to reduce the impact from channel estimation errors.

The receiver AS calibration process can be similarly defined, where the transmitter should
send ( N B nB + 1) calibration training packets from a fixed subset of antennas with fixed RF
connections. Then the calculation of receive AS correction coefficients is straightforward, as
long as we have a similar constraint of receive RF chain and antenna connections as in (2.75).
When both stations perform antenna selection, their calibrations can be conducted one after
the other. As a result, the equivalent complete channel matrix always contains fixed transmit
and receive RF responses, and the AS training protocol in Section 2 can be deployed without
distortions.
Note that the above calibrations should be conducted in each (or each subgroup of)
subcarrier(s) when applied in OFDM systems. Also, they can be straightforwardly applied

77

when the connection constraints in (2.75) does not hold (i.e. Li contains any RF chains for
any antenna i ).
Finally, in WLAN the calibration frame exchange sequence in Figure 2-17 can be
conducted by utilizing a normal AS training phase as in Figure 2-16, where the receiver
should feedback CSI only. Consequently, the peer station does not need to know the current
AS training packets are used for normal AS training or AS calibrations, therefore no extra
signaling needs to be defined for calibrations.
2.4.6 Numerical Results

In the first scenario where N A = 4, nA = 2, and N B = 2, nB = 2 , we investigate the

effectiveness of the MAC-based AS training method assuming that RF imbalance has no
significant impact. We use 64-QAM and rate convolution channel coding (1/2 convolution
code with puncturing [105]) as the modulation and coding set in each of the two transmitted
data streams, with 20MHz bandwidth and a 0.8 s guard interval in each OFDM symbol.
Therefore the burst data rate observed at the PHY layer is 117 Mbps. We also assume the
simplest least-square channel estimation in each subcarrier, and assume that there is no
further impairment from time synchronization errors and RF imperfections such as carrier
and sampling frequency offsets, phase noise, I/Q imbalance, AGC/ADC related issues, and
transmitter distortions. For each SNR we simulate 10000 packets, each containing 1000 bytes
of data payload plus the preamble as defined in [105]. The inter-packet interval is set to be 1

ms during the data transmission phase, so that the 10000 packets may experience sufficient
channel variations. For comparison we also simulate the AS training method where all of the
78

M=2 training fields are sent in one packet by extending its PHY preamble and ignoring the
switching loss (we call this scheme as PHY-based in the figures). The parameter TAS

Figure 2-18 Results of Channel B

defines the length of the AS training cycle. Since the channel encoding and interleaving are
conducted over all spatial data streams and all sub-carriers, it is natural to deploy the antenna
selection rule which maximizes the aggregated 2 2 MIMO channel capacity over all
subcarriers. From the PER results in channel model B (Figure 2-18), where the channel is
under relatively low level of frequency selectivity [107], we see that the proposed MACbased AS training method leads to almost the same results as PHY-based training method. It
is also noticeable that in reality the MAC-based method will even outperform the PHY-based
one by a few dBs, when considering the reduced switching loss by introducing MEMSbased antenna switches. From the same figure, we also see that the gains of applying AS in

79

WLAN are tremendous (5dB when TAS = 10ms , and more than 1dB when TAS = 100ms ). We
stress that many other effective AS rules have been developed in literature to achieve
different tradeoffs between performance gain and sensitivity to TAS , and the problem of
finding these AS rules, a topic beyond the scope of this paper, can be found in [37] and
references therein.

Figure 2-19 Results of Channel E

In channel model E (Figure 2-19), where the channel is much more frequency selective
[107], the relative gains of AS is reduced (although they are still as high as 3 dB
for TAS = 10ms ), because the less correlated sub-carriers make different antenna subsets look
more even with respect to the performance criterion (aggregated capacity or PER).
In the second scenario, RF imbalance is taken into considerations in channel B, where
the PER of 2-data stream WLAN system without AS, and MAC-based AS with and without
80

calibration (both setting TAS = 10ms ), are simulated. Each RF chain and antenna element
connection results in a baseband equivalent RF response li

( Tx )

with its magnitude uniformly

distributed in 3dB , and phase uniformly distributed in . We can then see from Figure
2-20 that calibration alleviates the impairment caused by RF imbalance. Hence the proposed
AS calibration method, a process imposing negligible training overhead, will buy us about 2
dB gain in this scenario. It is also noteworthy that RF imbalance will degrade the
performance of 2x2 MIMO without AS, hence the gains achieved by applying AS is even
larger than in Figure 2-18.

Figure 2-20 Results of Channel B under RF Imbalance

Again we do not stress the practicability of the brute forcing AS algorithm. One the one
hand, it is not so costly in the N A = 4, nA = 2 scenario (6 possible subsets). On the other hand,

81

efficient AS methods in MIMO-OFDM achieving good tradeoffs between performance and

complexity exist, and are beyond the scope of this section.
2.4.7 Summary

In this section we address two important issues for employing antenna selection techniques in
emerging high throughput WLAN systems: AS training protocol and RF imbalance
impairment mitigations. The proposed MAC-based AS training method minimizes the
amount of amendments required for accommodating AS in the new standard, and leads to
several other advantages such as ability to use switches with reduced switching loss. The
novel calibration process effectively alleviates the potential impairments caused by RF
imbalance, with negligible overhead. In general, the proposed techniques move a step closer
to the practical implementation of MIMO antenna selection techniques in high throughput
WLAN systems, and have been adopted in the recently adopted preliminary version of the
IEEE 802.11n baseline draft.
The above AS training/calibration protocol in WLAN is just one example of the
implementation of AS in commercial wireless networks. We believe that AS is also
applicable in some other similar systems, e.g. the MIMO-OFDM based WiMax wireless
access networks [106], DVB/DAB, and the 3GPP cellular networks [45].
2.5 Conclusions
In Chapter 2 we explore almost all aspects of antenna selection technology: we first investigate its
potential on achieving diversity increase; then we show that there exist practical antenna selection

82

algorithms achieving good tradeoffs between complexity and performance improvements; finally we
try to encompass antenna selection into a practical commercial wireless system, with particular
considerations on training overhead reductions and the alleviation of the impairments caused by RF
mismatches.

Based on our work, we believe that antenna selection is a promising technique in

multiple-antenna systems, as in principle it can lead to significant performance improvements
in fading channels with little expense on hardware cost and feedback overhead. It is expected
to be widely exploited in real wireless communication systems in the foreseeable future. In
additon, similar geometric tools can be extended to the study of general MIMO systems
without antenna selection, as will be shown in the next chapter.

83

Chapter 3
The Diversity-Multiplexing Tradeoff of Ordered SIC
Receivers in MIMO Systems
3.1 Introduction
In the previous Chapter, we use geometric analytical tools to derive the diversity orders and
effective computation algorithms for antenna selection in MIMO systems. In this chapter, we
extend this approach to some other scenarios: the point-to-point or multiple access MIMO
systems deploying SIC receivers.
It is well known that MIMO fading channels can be explored to provide either spatial
multiplexing gain or diversity gain. The tradeoff between them is expressed by the diversitymultiplexing tradeoff curve [104]. In [88], the discussion of diversity-multiplexing tradeoff is
extended to multiple access channels, where the fundamental tradeoff among diversity gain,
multiplexing gain and multiple access gain are effectively characterized. The tradeoff
discussions in [104] and [88] mainly deal with the performance limits among all possible
MIMO schemes, i.e. optimal joint encoding and decoding are employed so that the channel
capacity is assumed to be achievable. There are also some discussions on sub-optimal and

84

practical encoding and decoding schemes, for example, the V-BLAST architecture [95], in
which different layers are separately encoded and transmitted without any transmitter side
channel state information [104]. The discussion on V-BLAST is naturally extended to
multiple access systems (or SDMA systems) in [88], where independent transmitters are
separately encoded, but can be jointly detected (through MUD).
This chapter is mainly focused on the impact of optimal ordering on the performance of
V-BLAST and SDMA schemes with SIC detectors, where ZF or MMSE processing is
applied in each stage for the residue interference cancellation. In [104][88], the tradeoff
curves of these schemes with fixed ordering are accurately derived. However, when ordering
is involved, only some loose performance upper bounds are provided (using genie-aided
approaches), which are shown to be still away from the optimal tradeoff curves. The
difficulty lies in that, the explicit distributions of the ordered channel gains (or capacities in
the SDMA case) are no longer accessible: it is in general a problem of order statistics among
inter-dependent random variables, an under-developed topic itself [17]. Recently, the
diversity property of the ordered SIC receivers with fixed data rate (i.e. the spatial
multiplexing gain r = 0 ) is partly analyzed for some simplified scenarios, or conjectured
from numerical results. In particular, in [66][67][49], the diversity order for a two-layer VBLAST scheme with an ordered SIC receiver is rigorously shown to be equal to that with
fixed ordering; while in [50][66] numerical results are provided to show that the diversity
order of SIC receivers in a V-BLAST system with more than two layers is not increased by
the ordering rule proposed in [95]. To the best of our knowledge, the diversity achieved by

85

ordered SIC receivers with general system settings has not been rigorously analyzed thus far.
In this chapter, using a similar geometrical approach as that proposed in Section 2.2, we
rigorously prove that ordering will not improve the performance of V-BLAST SIC receivers,
with respect to the diversity-multiplexing tradeoff, thus versifying the conjectures in [50][66].
Extending the study to SDMA SIC detectors, we give tighter bounds than those provided in
[88]. Particularly, when the data streams for different users are transmitted with fixed rate
( r = 0 ), the diversity order is not improved by applying optimal user ordering.
The rest of the chapter is organized as follows. The problem formulation is provided in
Section 3.2. Our result on the diversity-multiplexing tradeoff for ordered V-BLAST SIC
receivers with point-to-point channels is presented in Section 3.3, and is extended to multiple
access channels in Section 3.4. Finally Section 3.5 contains some concluding remarks.

3.2 System Model and Problem Formulation

3.2.1 Point-to-Point MIMO Channels

We consider a frequency non-selective block Rayleigh fading channel model, for which a
point-to-point system can be expressed as:

y = Hs + n ,

(3.1)

where the N R t matrix y is the received signal block; the NT t matrix s is the transmitted
signal block, each row representing one separately encoded data layer with equal power
allocation; H = [h1 , h 2 ,

, h NT ] is an N R NT channel matrix with the constraint N R NT ;

and n is the background noise matrix of size N R t . Because we focus on the non-ergodic
86

scenario in a quasi-static fading channel, where each codeword spans one fixed fading block
(i.e. in delay limited scenarios), the coding length t of each layer is actually immaterial in
our study on the diversity orders (thus can be chosen to be 1 for ease of analysis). As is
known, layered one-dimensional coding can only bring coding gain but not the diversity
gain.
Given the above channel model, following the same line as in [49] [50][66][67][88][104],
we at first apply the ZF-SIC algorithm at the receiver, with the understanding that its
diversity order analysis, a study at high SNR regimes, also applies to MMSE-SIC receivers
(as will be shown in Lemma 3.1). Specifically, in such a V-BLAST system [95] let
(l ) [1, NT ] be the transmit antenna index corresponding to the lth decoding layer. Following
the expressions listed in Subsection 2.2.2, at the first decoded layer, a spatial equalizer
g1 = H

(1)*

, where [ A](i )* is the (i)th row in the matrix A , is applied to the received signal

y to obtain an estimate of the transmitted symbol of the same layer:

s1 = g1y = g1Hs + g1n = 0 / NT s1 + g1n ,

(3.2)

where 0 stands for the average SNR per receive antenna; s1 is the signal transmitted in
layer 1. The instantaneous channel capacity and error performance of this layer are
determined by the corresponding post-processing SNRs, which can be directly derived from
(3.2) as
0
NT

1 =

0
0
2
H
R(1), span{(2),(3)....( NT )} , 1 k NT ,
/ g1 =
/ H H (1)(1) =
NT

NT

87

(3.3)

where R(1), span{(2),(3)....( NT )} = 1/ H H H

1
(1)(1)

is the square of the projection height from h (1) to the

space spanned by h (2) ...h ( NT ) . In the lth layer, assuming perfect decision feedback, the spatial
equalizer g l is designed to null out the interference from the yet to be detected layers, then

l =

0
NT

R(l ), span{(l +1),

,( NT )}

(3.4)

For a MMSE-SIC receiver, the spatial equalizer of the first layer is given by
1
g1 = ( H H H + NT / 0 I ) H H , and the post-processing SNR of the first layer is given by
(1)*

[64]:

1 =

where 1 =

N T [ H H + NT / 0 I ]

1
(1)(1)

1
. Similarly, we have l
1
[H H + NT / 0 I ](1)(1)
H

Again, we adopt the operators

0 1
NT

0 l
NT

1 , 1 k NT ,

(3.5)

1 for l = 2...NT .

and defined in [104] and Subsection 2.2.2. The

diversity order a communication system is defined as the slope of its joint error probability
Pe ( 0 ) , which indicates the probability that any layer declares erroneous reception.

R( 0 )
where R ( 0 )
0 log
0

When a non-zero multiplexing gain r , defined as [104] r = lim

stands for the current data rate, is concerned, a family of codes { ( 0 )} over a block length
shorter than fading coherence time is employed, one at each SNR level with the data rate

R ( 0 ) = r log 0 . We then have the following lemma:

88

Lemma 2.1: ZF-SIC and MMSE-SIC receivers in V-BLAST systems have the same

diversity order, which is determined by

r
(1
)

NT
log Pr R(1), span{(2),(3)....( NT )} NT 0

.
d (r ) = lim
0
log(1/ 0 )

(3.6)

Proof: At first we show that the diversity order is determined by the first decoded layer,
for either ZF- or MMSE- SIC receiver, and for any ordering rule. We define Pl as the
average error probability of the lth decoding layer assuming perfect feedback (no interference
from the previously decoded layers), Pl J as the joint error probability of layers 1 to l, and

Pe as the system joint error probability. Therefore, we can naturally get P1J = P1 , Pe = PNJT
and

Pl J = Pl (1 Pl J1 ) + Pl J1 .

(3.7)

While in the second layer, (3.7) becomes

P2J = P2 (1 P1 ) + P1 .

(3.8)

Since P2 P1 is an exponentially smaller term, we derive P2J

applying the above analysis in (3.7), we finally get Pe
rule, it is intuitive to have max( P1 , P2 ,

max( P1 , P2 ) . By iteratively

max( P1 , P2 ,

, PNT ) . For any ordering

, PNT ) = P1 , because by assuming perfect feedback,

together with identical noise statistics and transmit power across different layers, all the other
layers face less interference than layer 1. We then have Pe

P1 , i.e. for any ordering rule the

diversity order is always determined by that of the first decoded layer.

89

Secondly, it was clearly shown in [104][90] that in non-ergodic scenarios the error
probability is dominated by the outage probability, then (3.6) applies to ZF-SIC receiver.
Finally, the diversity equivalence between ZF- and MMSE- SIC receivers can be verified
by (A.11) in Appendix A, then (3.6) applies to MMSE-SIC receiver as well.

Furthermore, the following corollary appears to be useful in Section 3.3.

Corollary 2.1: The ordering rule in [95] that maximizes the post processing SNR at each

detection step of the V-BLAST SIC receivers, is optimal with respect to the diversity order,
which is determined by
r
(1
)

NT
log Pr max Rk , Span{k } NT 0

(V BLAST )
,
d opt
(r ) = lim
0
log(1/ 0 )

(3.9)

where Rk , Span{k } is the squared projection height from h k to the space spanned by all the
other NT 1 column vectors.
Proof: From Lemma 2.1, since the joint error probability is determined by the first
decoded layer (i.e. Pe

P1 ), while the ordering rule in [95] can guarantee that the

performance of the first layer is optimized. (3.9) is straightforward, since in (3.6)

R(1), span{(2),(3)....( NT )} = Rk , Span{k } .

Remarks: Lemma 2.1 implies two important facts: 1) for any V-BLAST SIC receiver, the
diversity order is only determined by the first decoded layer, even though error propagations
90

are present; 2) we only need to evaluate (3.6) for the diversity order properties of both ZFand MMSE- SIC receivers.

Figure 3-1 Geometric Illustration of ZF-SIC Receiver with NT = 2

For a fixed-order V-BLAST SIC receiver, it can be assumed [49][66][104] (without loss
of generality) that detection starts from the layer corresponding to the first transmit antenna
and ends at that of the NT th antenna. So R(1), span{(2),(3)....( NT )} = R1, Span{2,3,

, NT }

and

r
(1
)

NT
log Pr R1, Span{2,3, , NT } NT 0

= ( N R NT + 1)(1 r / NT ) .
d (r ) = lim
0
log(1/ 0 )

(3.10)

When the ordering rule in [95] is taken into considerations (Section 3.3), the problem
becomes more involved, since different Rk , Span{k } in (3.9) are inter-dependent, and the exact
distribution of the post-processing SNR of the first decoding layer is in general not accessible
[88][104]. An exception is for the NT = 2 case [49][95], for which one has (see Figure 3-1)

91

R1,2 =|| h1 ||2 sin 2 12 , and R2,1 =|| h 2 ||2 sin 2 12 ,

where Ri , j represents the squared projection height from hi to h j ; and ij is the angle
between the two vectors hi and h j , defined as ij = sin 1

Rij
hj

, 0 <ij <

. The ordering rule

of [95] is reduced to simply choosing the transmit antenna whose corresponding column
vector has a larger norm, therefore the asymptotic exponential behavior of its outage
probability can be explicitly explored. It is shown that in this case the diversity order is the
same as that of the fixed-order case [49][95]. However, the analysis for NT > 2 scenarios,
which will be rigorously derived in Section 3.3, cannot be solved with this approach.
3.2.2 Multiple Access MIMO Channels

For a multiple access (or SDMA) channel, multiple receive antennas can be used to
separate signals from different users transmitted in the same frequency-time slot, leading to
multiple access gains. We assume that there are K transmitters each equipped with N Tk
K

antennas, assuming that N TAll = N Tk N R . In this case (3.1) still applies, except that H is
k =1

now expressed as
H = [H1 , H 2 ,

, HK ] ,

(3.11)

where H k is an N R N Tk sub-matrix corresponding to the channel between the kth

transmitter and the receiver. Also s becomes
s = [s1T , sT2 ,
92

, sTK ]T ,

(3.12)

in which the N Tk t sub-matrix s k is the data block from transmitter k. Data streams of one
transmitter are allowed to be jointly encoded, while those from different transmitters are
independent (which can be viewed as a generalized layered structure, each general layer
containing more than one data streams). If N Tk = 1, k , SDMA coincides with V-BLAST.
For analytical simplicity, we only investigate the ZF-SIC detector as suggested in [88].
Specifically, with fixed detection order (from user 1 to K), when detecting the data streams
for the kth transmitter, we null out the interference from transmitters k+1 to K by projecting
each column vector in H k to the null space of the space spanned by the column vectors in
H k +1...H K , after canceling the interference contributed by transmitters 1 to k-1 (fed back

from previous detectors). According to [88], the fixed-order diversity-multiplexing tradeoff,

denoting the asymptotic exponential behavior of the probability that any user is erroneously
detected, can be expressed as

d (r1 , r2 ,..., rK ) = min d N* k , N

k

NTj

(rk ) ,

(3.13)

j >k

where d m* ,n (r ) = (m r )(n r ) stands for the optimal diversity-multiplexing tradeoff for a

n m Rayleigh fading MIMO channel [104], and rk is the spatial multiplexing gain of user k.

Here for the ease of analysis, we focus on the symmetric scenario suggested by [88]:

rk = r, N Tk = N T , k . The extensions to non-symmetric and MMSE-SIC cases direct our

future research. With these settings, (3.13) becomes

d (r ) = d N

, N R ( K 1) NT

(r ) = [ NT - r ][ N R ( K 1) NT r ] ,

93

(3.14)

and we have the following analogues of Lemma 3.1 and Corollary 3.1, denoting (k ) as the
user index decoded in the kth stage
Lemma 3.2: For ZF-SIC multiuser detectors in SDMA systems with the symmetric setting,

the system diversity order is determined by that of the first decoded user,
d ( r ) = lim

log Pr{C(1) r log 0 }

log(1/ 0 )

(3.15)

where Ck is the capacity of the kth user .

Proof: It can be easily shown that (3.7)(3.8) in the proof of Lemma 3.1 still hold for
SDMA, with Pk as the error probability of user (k ) , assuming perfect decision feedback.
Therefore we get Pe

max( P1 , P2 ,

, PK ) = P1 , where the second equality holds due to our

symmetric settings. Finally it is shown in [104] that (3.15) determines the diversity order of

P1 , so as to Pe .

Corollary 3.2: With the settings in Lemma 3.2, the ordering rule that maximizes the per-

user capacity at each detection stage (or general layer) is optimal with respect to diversity
order, which is determined by:
( SDMA )
d opt
(r ) = lim

log Pr{max Ck r log 0 }

k

log(1/ 0 )

(3.16)

The proof follows straightforwardly that of Corollary 3.1, and is omitted here.
Suppose user k is chosen in the first decoding stage, its equivalent channel after nulling
can be expressed by
94

H kEq = h1,k span{Hk } , h 2,k span{Hk } ,...., h kN ,span{H } ,

T
k

(3.17)

where hik,span{Hk } is the projection from the ith column vector in H k to the null space of the
space spanned by the channel matrices for all the other users. As indicated in [88], H kEq is
equivalent to a point-to-point MIMO channel with N T transmit and N R ( K 1) N T receive
antennas. Then Ck in (3.16) can be expressed by Ck = C ( H kEq , 0 ) , meaning the MIMO
channel capacity given the channel H kEq and SNR level 0 (assuming equal power
allocations). It was claimed in [88] that the diversity analysis, when ordering is involved, is
in general not accessible. We are able to solve it in Section 3.4, by applying the technique
proposed in the following section.
3.3 Ordered V-BLAST SIC Receiver

In this section, by applying the novel geometrical approach similar as that in Subsection
2.2.2, we rigorously prove the following Theorem:
Theorem 3.1: For general N R NT V-BLAST systems with SIC receivers, applying any

ordering rule will not change the diversity-multiplexing tradeoff of the system.
As indicated above, [49][66] verify Theorem 3.1 only for the simplest NT = 2 case, while
we will investigate the general scenario with arbitrary values of NT . As we know, an
exhaustive search among all the NT ! possible permutations targeting the minimum system
joint error probability represents the optimal ordering rule (and leads to the optimal diversity)
[66]. However, its diversity property is generally hard to analyze. Through Corollary 3.1, the
95

ordering rule proposed in [95], although not optimal in minimizing the error probability, is
optimal w.r.t. diversity order. We can then prove Theorem 3.1 by analyzing this specific
ordering rule.
We start from the fixed data rate ( r = 0 ) case. The outage probability of the first decoded
layer (c.f. (3.9)) can be upper bounded as
Pr{max Rk , Span{k } x} Pr{R1, Span{1} x}
k

for arbitrary positive value x . When x 0 , the asymptotic exponential behavior of the above
upper bound represents the diversity of non-ordered V-BLAST SIC receiver, which equals to

N R NT + 1 (c.f. (3.10)). We then get8

Pr{max Rk , Span{k } x} x N R NT +1 .
k

(3.18)

In the following we try to derive a tight probability lower bound bearing the same
diversity property.
Lemma 3.3: With the above settings, we have
Pr{max Rk , Span{k } x} x N R NT +1 .
k

(3.19)

Proof: For the ease of illustration and without loss of generality, we analyze the NT = 3
case, and the extension to general values of NT is straightforward, as will be explained at the
end of the proof. In this scenario, we can decompose h1 by Gram-Schmidt orthogonalization
and coordinate transformation [37] as (see Figure 3-2):

Here we stress again that f ( x) g ( x) indicates f ( x) g ( x) for sufficiently small x .

96

h1 = 3h3 + 2h 2,3 + R1, span{2,3} h1, span{2,3}

(3.20)

= 3h3 + 2h 2 + R1, span{2,3} h1, span{2,3} ,

Figure 3-2 Geometric Illustration of (3.20)

where h3 = h3 / h3 ; h 2,3 is the unit vector along the projection of h 2 on null (h3 ) , the null
space of the space spanned by h3 ; and h 2 , h1, span{2,3} are similarly defined. 1 , 2 and
R1, span{2,3} , the coordinates of h1 on the orthonormal basis {h3 , h 2,3 , h1, span{2,3}} , can be shown
jointly independent [49][55][66]. The second equality in (3.20) is obtained with coordinate
transformation

from

{h3 , h 2,3 }

to

{h 2 , h3 } in

the

space

span{2,3} such

that

3h3 + 2h 2,3 = 3h3 + 2h 2 . Through geometrical analysis, we have

2 =

cos 23
2
, and 3 = 3 2
.
sin 23
sin 23
97

(3.21)

It is known that the angle 23 is independent with R1, span{2,3} , as indicated in [55], so 2
and 3 are independent with R1, span{2,3} . We then have the following claim:
Pr{max Rk , Span{k } x} = Pr{R1, Span{2,3} x, R2, Span{1,3} x, R3, Span{1,2} x}
k

Pr{R1, Span{2,3} x, R2 b, R3 b, 2 b , 3 b },

(3.22)

where b is an arbitrary constant, R2 =|| h 2 ||2 and R3 =|| h3 ||2 . To prove this claim, we define
the following events:
A : R1, Span{2,3} x
B : R2 b, R3 b, 2 b , 3 b

(3.23)

C : R1, Span{2,3} x, R2, Span{1,3} x, R3, Span{1,2} x.

Since R2, Span{1,3} is the magnitude of the projection from h 2 to the null space of span{1,3} , it
represents the shortest distance from vector h 2 to any points in span{1,3} . By choosing
{h1 , h3} as one basis of span{1,3} , together with (3.20) , it can be upper bounded as:
R2 h 2 ( 23h3 + 21h1 )

R2, Span{1,3} = min

23 , 21

= min || R2 h 2 23h3
23 , 21

21[ 2h 2 + 3h3 + R1, span{2,3} h1, span{2,3} ] ||2

212 R1, span{2,3} =

R2

22

R1, span{2,3} ,

(3.24)

where the inequality is derived by taking specific values of 21 and 23 , so that the
magnitudes in directions h 2 and h3 are zeros, i.e. 21 = R2 / 2 and 23 = 3 R2 / 2 . In

98

another word, we upper bound R2, Span{1,3} by (non-orthogonally) projecting h 2 onto span{1,3}
along the direction h1, span{2,3} . Similarly, we have
R3, Span{1,2}

R3

32

R1, span{2,3} .

(3.25)

It is then easy to see that given the event A B , event C is true, which proves the claim
(3.22). Also note that the random variables R2 , R3 , 2 and 3 are independent with R1, Span{2,3} ,
thus A and B are independent with each other, and the lower bound in (3.22) satisfies:
Pr{R1, Span{2,3} x, R2 b, R3 b, 2 b , 3 b}
= Pr{R1, Span{2,3} x}Pr{R2 b, R3 b, 2 b , 3 b}

(3.26)

x N R NT +1 ,

Pr{R1, Span{2,3} bx}

since Pr{R2 b, R3 b, 2 b , 3 b} will only introduce a constant factor, as it is a

function of x .
.

Therefore, we get Pr{max Rk , Span{k } x} x N R NT +1 for NT = 3 case. From the derivations

k

above, it is not difficult to see that the extension to the general NT scenario is straightforward.
Generally speaking, by the theorem of coordination transformation [37], (3.21) can be reshaped as:

= Q ,
where is the coordinates of h1 s projection on span{1} along the basis {h 2 , h3 ,..., h NT } ,
is its coordinates along the orthonormal basis obtained by Gram-Schmidt procedure, and

the transfer matrix Q is only associated with the directional relationships among
99

{h 2 , h3 ,..., h NT } . Both and Q are then independent with R1, span{1} [49][55], so is .

Therefore we can build up a probability lower bound as in (3.22), whose proof follows the
same line as in (3.23)(3.26).

Finally by combining (3.18) and (3.19), it is shown that the diversity of the ordered VBLAST SIC receiver is the same as the non-ordered case.
The analysis is easily extendable to r > 0 cases, where the outage is now defined as the
event that the instant code rate R ( 0 ) = r log 0 is larger than the channel capacity
conditioned on the current channel state. Therefore, from (3.9) and the analysis above, we
can easily derive the same result as (3.10), so that Theorem 3.1 is proved.

3.4 Ordered SDMA SIC Receiver

In multiple access channels, multiple data streams for one user are allowed to be jointly
encoded in space and time. The diversity analysis with ordering is much more involved than
in V-BLAST, since different capacities are correlated in a more complicated manner, making
the analysis of (3.16) difficult. This section presents our method for getting around this
problem, and gives the proof of the following Theorem:
Theorem 3.2: For symmetric SDMA systems defined in Section II-B, the following lower

and upper bounds of the diversity-multiplexing tradeoff hold for any ordering rule:
d L (r ) d (r ) dU ( r ) ,

where d L (r ) = [ NT - r ][ N R ( K 1) NT r ] , and dU (r ) = NT [ N R ( K 1) NT ] (1 r / NT ) .
100

(3.27)

Proof: We only need to investigate the ordering rule defined in Corollary 3.2. The
diversity lower bound in (3.27) is then easy to derive, since given a data rate of r log 0 we
get (c.f. (3.16)(3.17)):
Pr{max C (H kEq , 0 ) r log 0 } Pr{C (H1Eq , 0 ) r log 0 } 0 d L ( r ) .
k

(3.28)

We are left to show a tight lower bound of the outage probability.

Because C (H , 0 ) = log 1 + 0 nk , where nk stands for the nth largest eigen-value

n =1
NT

NT

Eq
k

of the Hermitian matrix ( H kEq ) H kEq , and

H

NT

n =1

we get nk H kEq

2
F

k
n

= Trace ( H kEq ) H kEq = H kEq

H

2
F

, n , so

C (H kEq , 0 ) NT log 1 + 0 H kEq

NT

2
F

and

Pr{max Ck r log 0 } Pr NT log 1 + 0 max H kEq

k

NT k

Pr max H
k

Eq 2
k F

( r / NT 1)

}.

2
F

r log 0

(3.29)

Therefore

Pr{max Ck r log 0 } Pr max H kEq

k

101

2
F

0( r / NT 1) ,

(3.30)

and to derive the diversity upper bound in (3.27) we are left to investigate the right hand side
of (3.30).
Considering the length limitation on this paper and without loss of generality, we
investigate the simplest case, where K = 2 , NT = 2 and the channel matrices are denoted by:
H1 = [h1 , h 2 ] , and H 2 = [h3 , h 4 ] . So after nulling in the first stage we get the equivalent
channels as:
H1Eq = [h1,span{3,4} , h 2,span{3,4} ]
H 2Eq = [h3,span{1,2} , h 4,span{1,2} ].

(3.31)

The outage probability in (3.30) can then be rewritten as

Pr(max H kEq
k =1.. K

2
F

x) = Pr( R1, span{3,4} + R2, span{3,4} x,

R3, span{1,2} + R4, span{1,2} x),

(3.32)

where x = 0( r / NT 1) , which satisfies lim x = 0 .

0

Similar as in (3.20), we decompose h1 and h 2 by:

h1 = 13h 3 + 14h 4 + R1, span{3,4} h1, span{3,4}

h 2 = 23h 3 + 24 h 4 + R2, span{3,4} h 2, span{3,4} .
So following the same method as in (3.24), we get

102

(3.33)

R3, Span{1,2} = min

31 , 32

R3 h3 ( 31h1 + 32h 2 )

= min || R3 h3
31 , 32

31[ 13h3 + 14h 4 + R1, span{3,4} h1, span{3,4} ]

(3.34)

32 [ 23h3 + 24h 4 + R2, span{3,4} h 2, span{3,4} ] ||

31 R1, span{3,4} + 32 R2, span{3,4}

312 R1, span{3,4} + 322 R2, span{3,4} ,

where the last step comes from the triangle inequality, and 31 and 32 are the solutions for
zero coefficients of h 3 and h 4 in the second equality of (3.34), expressed by:
1

31 13 23 R3
.
=

14
24
32

(3.35)

As what we argued in Section 3.3 (see (3.21)), since 13 and 14 are independent with
R1, span{3,4} , 23 and 24 are independent with R2, span{3,4} , and R3 is independent with any of the

projection heights, we have that 31 and 32 are independent with R1, span{3,4} + R2, span{3,4} . With
the same procedure, we can further derive:
R4, Span{1,2} 412 R1, span{3,4} + 422 R2, span{3,4} ,

(3.36)

where
1

41 14 24 R4
,
=

42 13 23 0

which is also independent with R1, span{3,4} + R2, span{3,4} . So if the following event E is true:

103

(3.37)

E : {R1, span{3,4} + R2, span{3,4} x / a;

312 , 322 , 412 , 422 a / 2; a 1},

from (3.34) and (3.36), we get that R1, span{3,4} + R2, span{3,4} x and R3, span{1,2} + R4, span{1,2} x are
both fulfilled. So we get the following lower bound of the outage probability in (3.32):
Pr( R1, span{3,4} + R2, span{3,4} x,
R3, span{1,2} + R4, span{1,2} x) Pr( E ).

(3.38)

Given
Pr( E )

Pr( R1, span{3,4} + R2, span{3,4} x / a)

Pr( H1Eq

x)

x NT [ N R ( K 1) NT ] = 0 NT [ N R ( K 1) NT ](1 r / NT ) ,

(3.39)

when x 0 , from (2.11)(3.30) we get

d (r ) NT [ N R ( K 1) NT ](1 r / NT ) .

(3.40)

From the analysis in (3.31)(3.40), it is easily seen that the extension to general values of
K and NT is straightforward except for more complicated mathematical expressions.

Combining (3.28) and (3.40)concludes the proof of Theorem 3.2.

Remarks: Figure 3-3 shows the upper and lower bounds of the tradeoff curve given by

Theorem II. Note that the point d (0) upper bound in [88], expressed as NT (N R NT ) , is
derived by a genie-aided method in [88]. Therefore our result provides a tighter upper bound.
More importantly, in the scenario with fixed data rates ( r = 0 ), Theorem 3.2 tells us that the
optimal ordering will not increase the diversity order of the joint error probability among the

104

K users deploying SIC detector. In another word, the diversity order under fixed data rate is
unchanged by applying any user ordering.

Note that the tradeoff lower bound coincides with that of the non-ordering case in (3.14),
which indicates that when applying adaptive coding schemes to maintain a non-zero
multiplexing gain r , the SDMA system diversity order with SIC detector might be increased
by user ordering. However, given that the two extreme points, corresponding to r = 0 and

d (r ) = 0 respectively, are unchanged from non-ordering case, we predict that the whole
diversity-multiplexing tradeoff curve will not be improved by optimal ordering, although the
accurate analysis is still challenging for the intermediate points with r > 0 .

Figure 3-3 Bounds on the Diversity Multiplexing Tradeoff for SDMA-SIC

Detectors
105

3.5 Conclusions
In this chapter, we utilize a geometry-based method, which is essentially the same as that
proposed in Section 2.2, to analyze the diversity-multiplexing tradeoff in point-to-point VBLAST and SDMA systems employing SIC receivers. Our results rigorously show that the
tradeoff curve for V-BLAST SIC receivers is not changed by ordering; while for SDMA, we
derive a tighter upper bound of the tradeoff curve than that in [88], and we prove that the
diversity order with fixed data rate is not changed by ordering.
Finally we stress that the diversity-multiplexing tradeoff is sometimes a loose indication
of the error probability performance, since it only characterizes the decay rate (exponential
behavior) at high SNR. Although optimal ordering does not improve the diversity order of
SIC receivers, it still provides an SNR gain (coding gain), compared with fixed-order SIC
receivers.
Note that the SIC detector in SDMA systems in our discussion is essentially a multiuser
detector applied in the simplest multiuser MIMO scenario: the synchronized uplink multiple
access channel. In the next chapter, we will consider to some more general scenarios, where
multiple users (each with multiple antennas) are located in a realistic cellular network with
universal frequency reuse. Therefore, the problem of inter-cell interference mitigation will be
our major focus.

106

Chapter 4
Interference Mitigation in Multi-User MIMO

4.1 Introduction
In the previous chapters, we mainly investigate point-to-point MIMO-SM systems. From
another standpoint, multiple data streams in SM systems can be viewed as multiple
(synchronous) users, therefore the MIMO receiver is equivalent to a multiuser detector. It is
then natural to investigate the problems of multiuser MIMO in the same frequency-time slot,
in which each user may transmit multiple data streams, and different multi-antenna users may
be located far apart and not synchronized with each other.
The receiver processing for interference mitigation is well-developed in literature (see [89]
and references therein). Some recent development on MIMO MUD can be found in [14][90].
In this chapter, we are interested in interference mitigation at the cellular downlink. In
particular, the inter-cell interference can be effectively addressed by base station cooperative
transmissions.
107

As introduced earlier, a MIMO channel in rich scattering environments can provide

multiple spatial dimensions for communications and yield a spatial multiplexing gain. At
high SNR, Shannon capacity can increase linearly with the minimum of the number of
transmit and receive antennas.
However, achieving the predicted enormous capacity gains in realistic cellular multiuser
MIMO networks could be problematic. First, for realistic cellular systems, the sharing of
common system resources by multiple users and the frequency reuse among adjacent cells
will bring in co-channel interference, which may greatly diminish the advantages of MIMO
systems [3][6]. Another problem in achieving these dramatic capacity gains in practice,
especially for outdoor deployment, is the rank deficiency and ill-conditionness of the MIMO
channel matrix. This is mainly caused by the spatial correlation due to insufficient scattering
and antenna spacing [38][78], and sometimes by the keyhole effect even though the fading
is essentially uncorrelated at each end of the channel [8]. Finally, the effect of the large-scale
fading, largely neglected in the current MIMO study, may also induce negative impact on the
anticipated performance.
The study on the performance of interference-limited multicell multiuser MIMO attracted
research attention only recently [3][6]. The study in [6] indicated the ineffectiveness of a
MIMO system in an interference-limited environment, when CCI is treated as background
noise. This is especially true in the downlink of cellular systems [16][25][26]. In a
conventional cellular system, CCI is primarily handled by careful radio resource management
techniques such as power control, frequency reuse, and spreading code assignments [53].

108

However, these techniques limit the achievable spectral efficiency gains and/or result in
insufficient suppression of CCI.
In recent years, several advanced techniques have been proposed to better combat the
effect of inter-cell CCI in cellular MIMO systems. Motivated by [6], MUD and turbo
decoding have been explored to significantly improve the performance of MIMO systems in
a multicell structure [13][16]. Such advanced receiver techniques improve the system
performance at the cost of increased receiver complexity. While they are readily applicable
today at base stations or access points for uplink processing of wireless networks, they still
impose challenges for the design of MSs in downlink communications, which is considered
to be the bottleneck for next generation high-speed wireless systems. Furthermore, it is also
found that there is a significant performance gap between the obtained MUD capacity and the
single-cell interference-free capacity upper bound, especially in environments with strong
CCI. This advocates a need to devote more system resources for performance enhancement
in the downlink multicell multiuser MIMO networks. The idea naturally arises to move the
CCI mitigation to the transmitter (BS) side on the downlink, where complex structure and
advanced processing can be more easily accommodated, if channel state information can be
obtained at the transmitter side either through uplink estimation or through a feedback
channel, for low user mobility scenarios such as indoor or outdoor pedestrian environments.
Joint optimization of the pre-coding designs among coordinated BSs has been proposed
[77][63] where the signal intended for an MS is transmitted from only one serving BS.
However, its very stringent dimension constraints severely limit the number of users that can

109

be handled by the system and lead to limited gains notwithstanding the significant additional
system complexity.
As multiple users in multiple cells are involved, cooperative processing at relevant base
stations can be exploited. This approach is feasible, as in the current infrastructure that is
common to both cellular communications and indoor wireless access, the base stations and
access points in the system are connected by a high-speed wired backbone that allows
information to be reliably exchanged among them. This approach is also reasonable, as in
environments with strong interference a mobile usually experiences several comparable and
weak links from surrounding radio ports, where soft handoff typically takes place in current

CDMA networks. In this scenario, cooperative processing among relevant ports transforms
the obstructive interference into constructive signals, which should offer large performance
improvement. Base station cooperation approach is proposed in [76] to enhance the downlink
sum capacity (throughput) with SISO systems employed in each cell, implementing dirtypaper coding proposed in [12]. While DPC or Tomlinson-Harashima precoding [93] can
effectively eliminate the effect of inter-cell CCI [25][53][76], both techniques are nonlinear
and may be prohibitively complicated for multi-BS systems. Linear pre-coding at the
cooperative BSs is, therefore, an attractive candidate given its good performance and the
relatively lower complexity it entails at both the BSs and the MSs [2][23][59].
Meanwhile,

transmitter

pre-coding

for

single-cell

downlink

multiuser

MIMO

communications has aroused much research interest recently (see, e.g. [19][96]). Assuming
the full set of transmitter CSI, such linear pre-coding designs combating intra-cell

110

interference in single cell downlink channels have been well developed. Based on a certain
criterion (e.g. maximum sum rate, minimum SINR or minimum power), iterative
optimizations deploying gradient methods [48] or alternate & maximize algorithms
[7][73][97] were proposed, which, however, in general suffer from non-convexity issues that
result in local suboptimum solutions, and from the complexity in the search for appropriate
initial points. To reduce the complexity introduced by the iterations, [98][85] try to optimize
some suboptimal criteria such that closed form pre-coding designs can be derived. For the
same purpose, some other authors try to design the Wiener filtering pre-coding matrices to
minimize the aggregate mean square errors (MSE) for all the users, e.g. [44][75][103].
Finally, it is a broadly accepted idea that the nullification designs which guarantee zero
interference at the MSs as in [2][9][61][80] are effective due to their simplicity and relatively
good performances. These single cell pre-coding designs then serve as a good starting point
for our proposed BS cooperative transmission.
In the first part of this Chapter, both information-theoretic DPC and some more practical
joint transmission schemes will be studied with cooperative multicell base stations for
downlink multiuser MIMO communications, assuming perfect synchronization of the signals
transmitted from cooperative BSs at each MS. Therefore the analysis here in general provides
an upper bound for achievable performance with BS cooperation, which defines a common
benchmark to gauge the efficiency of any practical scheme. Such an assumption also allows
us to explore other advantages of base station cooperative processing, e.g. channel
rank/conditioning advantage, and macro-diversity protection.

111

On the other hand, however, all the schemes proposed above and in [2][23][25][59][76]
for BS cooperation invariably assume that both the desired and the interfering signals from
different BSs arrive at each of the MSs synchronously (at the same time). While this
assumption enables the well-studied single-cell downlink transmission model to be applied in
a straightforward manner, it is fundamentally unrealizable especially for high data rate
systems. The BSs can align their transmissions so that the signals intended for any MS arrive
at that MS synchronously. However, even under the assumption of perfect BS cooperation,
the BSs cannot simultaneously control when these signals are received as interference by
other MSs.

This severely degrades the performance of the designs proposed in

[2][23][25][59][76], which do not account for it, especially, in high data rate regimes. To the
best of our knowledge, the problem of asynchronous interference has not been addressed in
the literature.
Given the above concerns, the second part of this chapter develops a framework for BS
cooperation in a multi-user multi-cell MIMO cellular network that explicitly accounts for
the asynchronous interference described above. We focus on linear-precoding designs given
their relatively lower complexity and satisfactory performance. Doing so also enables us to
effectively utilize the techniques developed for combating intra-cell interference in single
cell downlink channels, although the extensions are not straightforward. Three linear precoding-based algorithms are proposed to enable BS cooperation in the presence of
asynchronous interference. These algorithms provide different trade-offs between complexity
and performance in different SNR regimes. We also show that the proposed algorithms can

112

be extended to handle the under-explored scenario where even the desired signals from
different BSs do not arrive perfectly synchronously at the MSs due to practically unavoidable
timing errors or inaccuracies.
The rest of this chapter is organized as follows: in Section 4.2 we provide the system
models; and we give the analysis for the quasi-synchronous BS cooperative schemes in
Section 4.3; we then extend the discussions to asynchronous scenarios in Section 4.4; finally
we conclude this chapter in Section 4.5.

4.2 System Models

We consider a cellular system with B BSs (each with NT antennas) and K MSs (each with
N R antennas). The cooperative BSs together transmit Lk data streams to MS k. The links

between different transmit and receive antenna pairs are assumed to undergo frequency-flat
Rayleigh fading. The baseband channel between BS b and MS k is denoted by H (kb ) . For any
MS, the BSs cooperate and jointly transmit linearly precoded signals intended for it. As
mentioned, we focus on joint linear pre-coding among the BSs. The transmit vector for MS k
from BS b is linearly pre-coded by the matrix Tk(b ) (of size NT Lk ) and takes the form
x(kb ) (m) = Tk(b ) s k (m) , where s k (m) denotes zero-mean data vector, of size Lk 1 , meant for MS

k, in the m-th symbol interval. As in [2][23][25][59][76], we assume that each BS has

complete channel state information for all the channels between itself and all the MSs. This
can be achieved, for example, by an initial joint training phase that involves all the BSs, or by
means of adaptive tracking and feedback algorithms implemented at the MSs.

113

We assume a block-fading channel model with a large enough coherence time such that
the channel state remains the same over the duration during which Tk(b )

is used (the

frequency selective channels with OFDM will be discussed in Section 4.4). The coherence
time is much larger than the propagation delay between any BS-MS pair. In order to
maximize the information transmission rate for each MS, a Gaussian code book is used for
transmit data vectors. As transmit power constraints can be imposed on pre-coding matrices,
we assume, without loss of generality, that E {s k (m)s kH (m)} = I Lk , where the operator H
stands for matrix conjugate transpose, and I Lk denotes an Lk Lk identity matrix.
Furthermore, it is assumed that the transmit data vectors for different users are independent
of each other, i.e., E {s k (m)s Hj (m)} = 0 , for k j .
We assume perfect inter-BS synchronization and that the BSs know the propagation delay
from each BS to each of the MSs. (We will relax this assumption in Section 4.3) Inter-BS
synchronization can be realized by GPS, over-the-air approaches [4][70], or by a wired
backbone. Such infrastructure is already in place in current CDMA2000 and IS-95 cellular
networks to facilitate soft handoffs [53]. Timing-advance mechanisms, employed currently in
the uplink of GSM and 3GPP cellular networks [45], can then be used in the downlink to
achieve synchronization. Specifically, the BSs advance their respective transmission times to
ensure that the signals arrive at the desired MS (say, k) synchronously. However, due to the
broadcast nature of the wireless channel, MS k also inevitably receives signals intended for
other MSs. As detailed below, contrary to what is assumed in [2][23][25][59][76], these

114

signals arrive at MS k with different delay offsets and not synchronously with the data
streams intended for it.

Figure 4-1 A Simple BS Cooperation Scenario

Let k(b ) denote the propagation delay from BS b to MS k, as shown in Figure 4-1. To
guarantee simultaneous reception of the signals {x(kb ) }b =1,..., B at MS k, the BS b advances the
time at which x(kb) (m) is transmitted by the interval k(b) = k(b ) k(bk ) , where bk denotes the
index of the BS closest to MS k. This ensures that {x(kb ) }

b =1,..., B

arrive at MS k with the same

delay, k(bk ) . Assuming a linear modulation with baseband signature waveform g (t ) , the
equivalent received baseband signal at MS k, is given by
rk (t ) =

g (t mT

k( bk ) ) H k x k ( m )

(b )
(b ) (b )
(b)
+ g (t mTS k + j ) H k x j ( m ) + n k (t ),
m j
b

( j k )

115

( B)
H
(2)
( B)
(1) H
, and n k (t ) is additive white

where Hk = H(1)
k , Hk ,..., Hk , xk (m) = xk (m) , , xk (m)

Gaussian noise. At MS k, the received baseband signal rk (t ) is passed through the filter
matched to g (t mTS k(bk ) ) (which is also delayed by k(bk ) ) to generate sufficient statistics.
The corresponding received signal component at MS k is H k Tk s k (m) , where
Tk = Tk(1) H , Tk(2) H ,...Tk( B ) H

is the collection of all transmit pre-coding matrices used by the B

BSs for MS k. The discrete-time received signal after matched filtering then takes the form
y k ( m) = H k Tk s k ( m) +

(b )
k

Tj( b ) i (jkb ) + n k ,

(4.1)

j
b
( jk )

where n k is the discrete-time noise vector satisfying E {n k n kH } = N 0 I , and i (bjk) is the

asynchronous interference at user k from the signal transmitted by BS b for MS j. Its strength
depends on the difference, (jkb ) , between the timing-advances used by BS b for MS j and MS
k:
(jkb ) = ( k(b ) (jb ) ) k(bk ) = k(b ) (jb ) .

As shown in Figure 4-2, the asynchronous interference to MS k from the signal transmitted
by BS b for user j arises due to two adjacent symbol transmissions with indices m(jkb ) and

m(jkb ) + 1 , where m(jkb ) = m (jkb ) / TS and x denotes the smallest integer larger than or
equal to x . Let 0 jk(b ) TS represent the delay offset (jkb ) modulo the symbol duration, TS .

116

Denoting the autocorrelation of g (t ) as ( ) =

g (t ) g (t )dt , with (0) = 1 , we then

have
i (jkb ) = ( jk(b ) TS )s j (m(jkb ) ) + ( jk(b ) )s j (m(jkb ) + 1) .

(4.2)

Figure 4-2 Asynchronous Interference : Symbol Overlap Diagram

(b )
We shall need the first and second moments for i jk , which are calculated as follows.

{ }

(b)
From (4.2), it follows that E i jk = 0 . Assuming the information signals intended for two

different users j1 and j2 are independent of each other, E {i (jbk1) i (jbk2) H } = 0, for j1 j2 k . Finally,
1

it can be shown that the correlation between i (jkb1) and i (jkb 2) , for j k , is
E {i (jkb1) i (jkb 2) H } = (jkb1,b 2) I Lk ,

0,
m(jkb2) > m(jkb1) + 1

, for j k .
where (b1,b2) =
m(jkb2) = m(jkb1) + 1
( jk(b1) ) ( jk(b2) TS ),
jk
b
b
b
b
b
b
(
1)
(
2)
(
1)
(
2)
(
2)
(
1)
( ) ( ) + ( T ) ( T ), m = m
jk
jk
S
jk
S
jk
jk
jk

(4.3)

When b1=b2=b, we have (jkb,b ) = ( jk(b ) ) 2 + ( jk(b ) TS ) 2 . Furthermore, kk(b1,b 2) = 1 for all b1
and b2.
117

As in [16], our goal is to jointly optimize the transmitter pre-coding matrices, {Tk }k =1...K ,
so as to maximize the sum of information rates over all the K users, given the channel states
H1 ,..., H K . From (4.1), the bandwidth-normalized information rate, Rk , of MS k is given by
[3][25]
Rk = log I + k 1H k Tk TkH H kH ,

(4.4)

where
k = N0I +

(jkb1,b 2) H (kb1) T(j b1) T(j b 2) H H (kb 2) H

j ( b1,b 2)
( jk )

is the covariance matrix of noise plus interference term in (4.1). Since all the K users use the
same waveform g (t ) , the asynchronous interference correlation terms,

( b1,b 2)
jk

corresponding to different timing parameters, can be pre-calculated and stored in a look-up

table.
On the other hand, following the assumptions in [2][23][25][59][76], when all the
considered MSs are located near the cell boundaries, the delay offsets are negligible so that
we can get a synchronized model, which will be used in Section 4.3 for exploring the
potential of BS cooperation. Then (4.1) can be rewritten by:
y k ( m) = H k Tk s k ( m) +

H Ts
k

+ nk .

(4.5)

j
( j k )

The model in (4.5) is similar to a downlink intra-cell interference model

[25][39][40][42][43][100][102], except for the different power constraints as will be
discussed below. These observations then advocate the extension of single cell downlink pre-

118

coding designs, which target for mitigating intra-cell interference, to the BS cooperation case
with some minor modifications, as will be seen in Section 4.3. In general, for the
synchronous case the following expression holds true for the spectral efficiencies of user k
with linear pre-coding schemes:
Rk = log I + [N0I + Hk ( j k Tk TkH )HkH ]1 Hk Tk TkH HkH ,

(4.6)

Typically the optimization has three kinds of power constraints, and we will use different
constraints in different scenarios (asynchronous and synchronous) for simplifying the
discussions, although all of them are by nature extendable to both cases. For the first one
(which will be used in Section 4.4 for the asynchronous case), we limit the power dedicated
to the transmission of each MS. This type of power constraint is motivated by power
fairness for the different users, and leads to analytically tractable solutions based on the
asynchronous model in (4.1). As a consequence, the transmit power at each BS is limited as
long as the number of users it serves is finite. Therefore, the optimization problem at hand is:

max Rk ,
{Tk _ opt }k =1..K = arg
{T }
k k =1... K

H
s.t. Trace Tk Tk PT for k = 1,..., K .

(4.7)

In this asynchronous case, an additional constraint follows from the constraints on channel
spatial dimensions: BNT Lk . This optimization problem is non-linear and not even
k

convex. Brute-force numerical optimization is not a viable option as it involves searching

over an extremely large space of dimension BNT Lk . Alternative sub-optimal techniques
k

therefore need to be developed in order to determine the pre-coding matrices. One such
solution is the nullification method [2][23] (also will be mentioned in Section 4.3), which is
119

widely applied in single-cell multiuser scenarios due to its simplicity and relatively good
performance.

It

makes

all

the

transmitters

satisfy

the

zero-forcing

constraint: H k Tj = 0, for all k j . However, with asynchronous interference, this constraint

can no longer null out the interference term in (4.1) and leads to significant interference
leakage. Another option, as put forth in [77], is to force a stronger per-BS
constraint, H (kb ) Tj(b ) = 0 , for all k j . While this completely cancels interference, it severely
restricts the maximum number of users to K NT / N R , which is clearly undesirable.
The second and third power constraints, which will be used in Section 4.3 for the
synchronous cases, assume a pooled and per-BS power constraint respectively, expressed as:
K

Tr(

T T
k

H
k

) KPT , and

(4.8)

k =1

Tr(

[b ] [b] H
k Tk

) PT(b ) , b = 1, 2, ...., B

(4.9)

k =1

where PT(b ) is the power constraint for BS b.

In the following, we first investigate the synchronous case to explore the performance
upper bounds of BS cooperation. The for the asynchronous case we propose three joint
transmitter linear pre-coding approaches, which lead to closed-form or simplified numerical
solutions. They all allow a controllable amount of interference leakage at the input of all the
MSs, and are based on different sub-optimal optimization criteria.

120

4.3 BS Cooperation with Synchronized Interference

In this section, we investigate BS cooperation with the model (4.5). As stated above, we
assume the involved MSs are all located near the cell boundaries, so that signals transmitted
from different BSs arrive at each of them simultaneously. This assumption also represents
the case for strong CCI at each MS, which is operated with relatively low data rate (similar at
the soft handoff in CDMA).
4.3.1 Pre-coding Designs
With base station cooperation, the system resources can be pooled together for more
efficient use. In particular, the severe CCI problem can be effectively controlled and
significant performance improvement can be achieved. Meanwhile, the complexity at MS
can be significantly reduced. In the following, the potential of cooperative processing for
system capacity enhancement is first studied from a theoretical standpoint, by extending the
dirty paper coding approach with a pooled power constraint (4.8) to the downlink multicell
multiuser scenario. Then some more practical suboptimal transmission schemes, with more
practical per-base power constraints (4.9), will also be investigated.
A. Performance Upper Bound: Dirty Paper Coding with Pooled Power Constraint
For a system with perfect data and power cooperation among B base stations, we can
implement the throughput-achieving dirty paper coding [12][25], at the vector broadcast
channel formed by cooperative BSs, to obtain a system performance upper bound. Note that
with the vector BC model (4.5), we apply DPC in a slightly different way, where the encoder
121

for the current user needs to non-causally know not only the encoding of previous users
and associated CSI, but also the corresponding propagation delays, to pre-cancel the
interference from previous users. The rate of each user and the (optimal) sum rate can then
be expressed as:

R ( k ) _ DPC (T) = log

SRDPC =

max

N 0 I + H ( k ) ( j k T ( j ) TH( j ) )HH( j )
N 0 I + H ( k ) ( j > k T ( j ) TH( j ) )HH( j )

( R ( j ) _ DPC (T)) ,

(4.10)

(4.11)

Tr ( TT H ) = KPT , ( j )

where T = [T1 , T2 ,..., TK ] , and (1), (2), ....., ( K ) represents a certain user ordering. By
applying the duality of the broadcast and multiple-access channels (MAC) [43], (4.11) can be
obtained by calculating the sum rate of a dual multiple-access channel with the same total
power constraint KPT . Iterative numerical methods that jointly optimize (4.11) on the dual
uplink were proposed (see e.g. [42] and [102]) based on the iterative water-filling algorithm
proposed in [101]. Furthermore, it is shown that (4.11) is actually the saddle point (with
worst-case colored noise) of the Satos bound, which is the sum rate of a heuristic
cooperative system where both transmitters and receivers can cooperate with each other,
given by [71]:
SRSato =

max log

n + HT TT H HTH

Tr( TT H ) Pt

(4.12)

where HT = [H1T , HT2 , ... , HTK ]T and n is the noise covariance matrix.
DPC with per-base constraints is much more involved, as the MAC/BC duality does not
hold any more. DPC has not been shown to achieve the capacity region or even the sum
122

capacity with constraints (4.9). Complex iterative multistage numerical methods for
cooperative DPC with constraints (4.9) are proposed in [25][39][40], in which a small piece
of power is invested to a certain selected user until one of the BS reaches its power
constraint. The path gains corresponding to this BS are then set to zero in subsequent stages
so that no further power is allocated to these antennas. In general, (4.9) is more strict than
(4.8), so a performance degradation is expected.
Although the DPC scheme with a pooled power constraint gives us a simple performance
upper bound for BS cooperation, some unrealistic assumptions are made, such as the noncausal knowledge of interference sequence at the transmitter, which motivates the
exploration of more practical joint transmission schemes, as will be discussed in the
following.
B. Suboptimal Joint Transmission (JT) Schemes for BS Cooperation with Per-Base
Power Constraints
In this part, several suboptimal but more practical joint transmission schemes with the more
practical per-base power constraints are exploited for better understanding of the achievable
performance gains of BS cooperative processing in practice. Essentially these techniques are
counterparts of corresponding multiuser detectors, some of which have been revisited in the
co-located MIMO context recently [10][80]. These suboptimal joint transmission schemes
will be compared with both the receiver MUD approaches [13][16] and DPC approach
above. Intuitively, compared with DPC, linear pre-coding may induce more transmit power
inefficiency, as the pre-coding matrices are responsible for the mitigation of interference

123

from both previous and subsequent users, and per-base power constraints are
implemented instead of a pooled power constraint.
Before discussing these suboptimal joint transmission schemes in detail, we first propose
a simple algorithm for designing T with per-base power constraints. Let LT = k =1 Lk be the
K

overall number of data streams of K users. Suppose that a preliminary joint linear transmit
matrix G BNT LT is given, whose designs will be introduced in the sequel. Our design of T
with per-base power constraints (4.9) is given by
T = G. ,

(4.13)

where is a LT LT diagonal matrix with diagonals { l }lL=T1 each representing the allocated
power for the corresponding original data stream. Since typically LT >> B and there are only
B per-base power constraints in (4.9), we can further divide { l }lL=T1 into B groups each with
LT / B elements having the same value:

= blockdiag ( 1I LT / B , 2I LT / B , ..... , B I LT / B } .

(4.14)

Further define

Q B B

G [1] 2
1
[2] 2
G
= 1
:

G[ B ] 2
1

G[1]
2

G[2]
2

:
G

[B] 2
2

.....
.....
:
.....

G[1]
B

[2] 2
GB
,

:
2
G[BB ]

(4.15)

where G[kb ] is an NT LT / B submatrix in G , corresponding to the transmit weights at BS b

for the kth group of data streams as defined above. Let P = [ PT(1) , PT(2) , ...... , PT( B ) ]T be the
124

per-base power constraint vector, then we can calculate by solving the linear system
equation:
= [ 12 , 22 , ....., B2 ]T = Q 1P .

(4.16)

When an infeasible solution ( does not have all positive entries) is obtained, we can refine
it as:
(b )
P
= I , = min T 2 ,
b =1,2,..., B
G[b ]

(4.17)

where G [b ] is the rows of G corresponding to transmit antennas at BS b. Note that (4.16) can
utilize the full power at each BS, while in (4.17) only the BS satisfying the minimum value
can transmit with full power and any other BS transmits with a power less than its power
constraint. Nonetheless, it has been shown that designs with per-base power constraints incur
insignificant performance loss for linear joint transmission schemes compared to the
corresponding designs with pooled power constraint, thus demonstrating the applicability of
these joint transmission schemes in realistic cellular communications. These suboptimal
schemes are illustrated as follows.
Joint Transmitter Zero Forcing (JT-ZF): This approach can be viewed as the counterpart
of

the

multiuser

decorrelator,

which

employs

pseudo-inverse

of

HT = [H1T , HT2 , ... , HTK ]T with

G = HTH ( HT HTH ) 1 ,

125

Tk = G k . k ,

(4.18)

where k is the corresponding kth Lk Lk diagonal block of . With this approach, we can
k I , j = k
. Then interference from both other users and other sub 0, j k

easily get H k .Tj =

streams of the same user are eliminated simultaneously. Just as ZF receivers eliminate the
interference at the expense of noise enhancement, ZF transmitters generally increase the
average transmit power by the same factor. Furthermore, in the rank deficient scenario where
H T H TH is singular, this approach cannot even be applied. These observations naturally

motivate the joint transmission design based on MMSE criterion.

Joint Transmitter MMSE (JT-MMSE): Recalling the trade-off between interference
cancellation and noise enhancement for its receiver counterpart, JT-MMSE makes a good
trade-off between interference cancellation and transmitter power efficiency, whose
preliminary joint transmission matrix is given as
G = HTH (HT HTH +

N 0 1
I) .
KPT

Joint Transmitter Null-space Decomposition (JT-Decomp):

(4.19)
For this approach, the

preliminary transmission matrix for user j can be described as:

G k = Vk Vk' ,

(4.20)

where Vk includes the right singular vectors corresponding to the null space (zero singular
values) of HT _ k = [H1T , ..., HTk 1 , HTk +1 , ..., HTK ]T , and Vk' includes the first Lk right singular
vectors of the virtual channel H 'k = H k Vk . HT _ k usually has a non-zero null space, so

126

H j Vk = 0, j k can be guaranteed. Interference from all other users is cancelled, and we

obtain a set of equivalent parallel interference-free subchannels.

JT-Decomp approach generally outperforms JT-MMSE and JT-ZF at the cost of extra
complexity, as will be seen in our numerical results. There is again a potential problem with
this approach in the rank deficient scenario. If H k is rank deficient, there will be a higher
chance that null{HT _ k } overlaps with null{H k } , which means that the equivalent channel may
have reduced channel rank or even null rank (when null{HT _ k } = null{H k } ). Fortunately, as
we will illustrate in the next section, this will rarely happen in our distributed multicell
MIMO setting, as compared to the traditional co-located MIMO.
4.3.2 Other Advantages of Base Station Cooperation
With synchronous base station cooperation, all antennas of the relevant multicell base
stations can be jointly employed for data processing and transmission. In some sense, it is
equivalent to a single giant base station or processing center. The key differences from the
single-cell scenario with co-located antennas at one BS, are the widely distributed antennas
and independent large-scale fading experienced for each link between a mobile-port pair.
Besides its potential in CCI cancellation, base station cooperative processing assumes other
advantages in multicell MIMO communications, as indicated below.
A. Channel Rank and Conditioning Advantages
The capacity of a MIMO system is determined by the characteristics of the channel matrix
H. Clearly r = rank (H ) is one such important factor, which determines the number of parallel
127

sub-channels that can be opened up for independent communications. In rich-scattering

environments, full rank can be assumed and essentially r = min( NT , N R ) more bits/s/Hz are
obtained for every 3 dB increase in power. However, in some extreme environments (e.g.,
the keyhole problem [8]), a MIMO system will lose its capacity advantage (spatial
multiplexing gain) over a SISO system, even though other advantages like diversity and array
gains may still preserve. Another important factor influencing the MIMO capacity is the
channel condition number =

max i i
, or more generally the singular value distribution of
min i i

the channel matrix. Noting that equal power allocation achieves optimal performance at high
SNR in the full channel rank case, we conclude by Jensens inequality that a channel with
= 1 has the largest capacity, with the same power constraint. In rich scattering environment,

channel matrix H is assumed to have normalized i.i.d. complex Gaussian entries and thus is
well-conditioned. In realistic environments, H may get ill-conditioned due to fading
correlation, resulting from the existence of few dominant scatterers, small angle spread, and
insufficient antenna spacing [78]. Those sub-channels with l2 << 1 are essentially of no use,
even though the channel still has a full rank.
In a typical outdoor urban scenario, antenna arrays at the base stations are elevated above
urban clusters and far away from local scattering, while mobile terminals are surrounded by
rich scatterers, and the number of independent paths is limited by few far-field reflectors
[38]. Therefore, the downlink channel matrix with co-located transmit array can be modeled
by (2.57). In this case, the channel matrix may be both rank-deficient and ill-conditioned,
determined by the propagation and system parameters.
128

For BS cooperation in multicell MIMO, the overall transmit array is distributed among
cooperative base stations, so in the equivalent channel for user k, {H (kb ) }b =1...B are independent
B

with each other. The overall number of independent links is then given by rank (H (kb ) ) ,
b =1

which is guaranteed to be at least equal to B. If B N R , H k will always have a full rank.

Furthermore, the channel conditioning will not be greatly degraded even if transmit fading
correlation happens at each base station, as the fading between different transmit antennas at
different base stations are still uncorrelated.
Because of the channel rank and conditioning advantages with BS cooperation, joint
transmission schemes in Subsection 4.3.1.B will not significantly degrade, compared to
single-cell signaling with co-located antennas. In particular, the chance that JT-ZF cannot be
implemented, or the chance that null{HT _ k } overlaps with null{H k } for JT-Decomp, will be
greatly reduced. These suboptimal joint transmission schemes will be shown to even
outperform the single-cell upper bound at high SNR for rank deficient channels.
B. Macro-diversity Protection for Shadowing Channels
Shadowing in wireless channels is a position sensitive factor, which means that signals
from transmit antennas co-located at the same base station are generally subject to the same
shadowing, while those from different base stations subject to independent shadowing. Under
severe shadowing, the capacity of a single cell MIMO with co-located antennas will degrade
significantly. On the other hand, cooperative BSs can provide the macro-diversity protection
for shadowing impairment, because of their independency. Intuitively, the probability that all
129

sub-channels of H k are under deep shadow fading is much lower than a co-located MIMO
channel. Macro-diversity cannot increase the mean of the received SNR, but will greatly
reduce its variance.
4.3.3 Numerical Results
We simulate a synchronous BS cooperation scenario with B=K=3, and NT = N R = 2 .
Averaged spectral efficiencies for different pre-coding schemes are compared with each
other, where we provide the baseline schemes including the conventional single cell
signaling scheme treating inter-cell interference as noise [3][6], the 2 2 point-to-point
MIMO capacity, and the receiver MUD at each MS [16].
From Figure 4-3, which simulates the Rayleigh fading channels, we see that the
performance of MUD in [16] is still significantly away from the single-cell interference-free
case (2x2 MIMO). On the other hand, the DPC schemes for BS cooperation result in a
significant performance gain over the multiuser receivers, and they even outperform the 2x2
interference-free MIMO. As discussed above, suboptimal joint transmission schemes reduce
interference at the expense of transmitter power inefficiency, which may compromise other
gains and may result in a performance worse than 2x2 MIMO. This can be observed in
Figure 4-3. Nonetheless, all these suboptimal joint transmission schemes outperform the
multiuser receiver, thus verifying the capability of BS cooperation in practical deployment.
Furthermore, the JT-Decomp approach outperforms JT-MMSE at the cost of extra
complexity, and that JT-ZF converges to JT-MMSE at high SNR, which is a well-known
result.
130

Figure 4-3 Simulations of BS Cooperation Shcemes in Rayleigh Fading Channels

Now we assume that the channel model of (2.57) is used for {H (kb ) } with one dominant
path, so the ranks of {H (kb ) } are one. From Figure 4-4, we can see that the spectral
efficiencies of conventional single-cell schemes degrade (cf. Figure 4-3) because of the
reduced channel rank. Notice that different rank results in different slopes of the rate curves.
From Figure 4-4, we can see that there is a 2 more bits/s/Hz spectral efficiency gain for every
3dB transmit SNR increase for BS cooperation schemes, while there is only a 1 bits/s/Hz
gain with the same transmit SNR increase for the 2x2 MIMO. That is why JT-ZF and JTDecomp can even outperform the 2x2 MIMO at high SNR, as the spatial multiplexing gain
compensates the power inefficiency. Thanks to channel rank and conditioning advantages
131

with BS cooperation, performances of all joint transmission schemes dont deteriorate

significantly in this deficient channel (cf. Figure 4-3), including the otherwise problematic
JT-ZF and JT-Decomp schemes.

Figure 4-4 Simulations of BS Cooperation Shcemes in Rank Deficient Channels

4.3.4 Summary
In this section, cooperative processing at multicell base stations is introduced under
synchronous assumptions. In particular, the capability for co-channel interference
cancellation, channel rank/conditioning advantage, and macro-diversity protection for BS
cooperation are illustrated and verified. Although these advantages may not be achieved
132

simultaneously and may compete with each other, there is still an optimistic prediction on
overall system performance improvement, which reveals the great potential of base station
cooperative processing on meeting the ever-increasing capacity demands for wireless
communications.
On the other hand, however, we will find in the next section that when releasing the
synchronous assumption, the above mentioned nullification (JT-ZF and JT-Decomp.) or JTMMSE methods are not effective anymore, because the channel model in (4.5) should now
be replaced by (4.1) due to its asynchronous interference nature. We will then propose some
more technique involved pre-coding schemes to address the problem of interference
mitigation in asynchronous BS cooperation systems.

4.4 BS Cooperation with Asynchronized Interference

4.4.1 Pre-coding Designs
We learn from Section 4.3 the BS cooperation has the great potentials of interference
mitigation, channel rank/conditioning and shadow fading improvements. Now we look back
to the asynchronous channel model (4.1), which is obviously more realistic than (4.5), and
particularly suitable for general high data rate communication systems, where mobile users
may be arbitrarily located in the whole system. In this subsection, we assume that the timingadvance mechanisms work perfectly so that the desired signals (but not the interference
signals) arrive synchronously at each MS. Imperfect timing-advances, which affect even the
desired signals, are treated in Subsection 4.4.2.
133

A. Joint Wiener Filtering (JWF)

As discussed in the Section 4.2, it is difficult to apply direct optimization and nullification
approaches to the signal model in (4.1), and it is also not appropriate to direct apply the
nullification or JT-MMSE methods in Section 4.3. We therefore aim to design the transmitter
pre-coders, {Tk }k =1,..., K , using the Wiener smoothing criterion such that the overall mean
square error for all the K users, seen at the input of receivers, can be minimized. This
approach, while sub-optimal for maximizing the sum information rate, has the important
advantage of yielding closed-form solutions that can realize the large gains from BS
cooperation. While similar approaches have been proposed in the context of intra-cell
interference mitigation [44][75][103], extending them to our case with asynchronous
interference is not straightforward.
For ease of notation, we drop the symbol index m in (4.1). Let y = y1H , y 2H ,..., y HK
the received vector over all users, and s = s1H , s 2H ,..., s HK

denote

denote the data vector over all

users. When Lk = N R , for all k , the vectors y and s have the same dimensions, and the
overall system-wide MSE can be expressed as:

MSE = E y s

} {
=

E y k sk

k =1

} MSE
=

(4.21)

k =1

where MSEk stands for the MSE of MS k, and the expectation, E{.} is taken over the random
data vectors, {s k }k =1...K , and the noise, {n k }k =1...K . The optimization criterion then becomes:

134

min
(b )

{Tk }k =1... K ,b =1... B

MSE

k =1

s. t. Trace(Tk*Tk ) = Trace Tk(b )*Tk(b ) PT , for k = 1,..., K . (4.22)

B

b=1

Denoting the MUI term in (4.1) as J k =

(b) (b) (b)

k T j i jk

, MSEk follows:

j
b
( j k )

{(

MSEk = E H k Tk s k s k + J k + n k H k Tk s k s k + J k + n k ,

B
B
B
= Trace H (kb ) Tk(b ) Tk(b ) H H (kb ) H H (kb ) Tk(b ) Tk(b ) H H (kb ) H
b =1
b =1
b =1

+ ( N 0 + 1)I N R +

(4.23)

(jkb1,b 2) H (kb1) T(j b1) T(j b 2) H H (kb 2) H ,

j k ( b1,b 2)

where we use the identity E {J kH J k } = Trace E J k J kH . To solve (4.22) in closed-form, we

apply the following Lagrange objective function, A({Tk(b ) }k =1..K ,b =1...B ) ,:
A({Tk(b ) }k =1.. K ,b =1... B ) =

MSEk +

k =1

k =1

b =1

k Trace( Tk(b) H Tk(b) ) PT ,

(4.24)

where 1 ,..., K are the Lagrange multipliers associated with the power constraints for MSs
1,,K, respectively. As shown in Appendix A, we get the following closed-form solution for
Tk :
Tk = [Ck + k I ] H kH ,
1

(4.25)

where Ck , which corresponds to MS k, is an NT B NT B block matrix of the form

Ck = C(kb1b 2)
, with
b1,b 2 =1,..., B

135

 K (b1b 2) (b1) H (b 2)
kj H j H j , for b1 b 2

= ( j k )
K (bb ) (b ) H (b )
kj H j H j ,
for b1 = b 2 = b.

j =1

C(kb1b 2)

(4.26)

The expression for k is given in Appendix E, and is not repeated here to save space.
B. Joint Leakage Suppression (JLS)
An alternative sub-optimal objective function, which has the advantage of allowing a userby-user optimization, can be obtained by considering leakage suppression. For MS k, we
design Tk to maximize the ratio of the desired signal power received by MS k and the total
power of the noise and the leakage of x k at other MSs. We call it the signal to leakage plus
noise ratio. Unlike SINR, which considers the interference that arrives at an MS, SLNR
considers interference that stems from the data streams intended for the MS. This is a
generalization of the SLNR approach proposed in [15][85][98] which only addressed the
simple case of a single data stream per user and did not model asynchronous interference
to the general BS cooperation scenario considered in this paper.
For analytical tractability, we restrict the space of Tk to scaled versions of semi-unitary
matrices. Thus, Tk =

PT
Q k , where the NT B Lk semi-unitary matrix Q k contains
Lk

orthonormal columns. Orthonormality ensures that there is no cross-talk among the Lk data
streams for MS k, and simplifies the receiver at MS k. Then the received signal is

136

yk =

PT
P
H k Qk sk + T
Lk
Lk

(b )
k

Q (jb ) i (jkb ) + n k ,

j
b
( jk )

where Q(kb) consists of the rows in Q k that are associated with the bth BS. The signal
component power is then
PT
P
E {s kH Q kH H kH H k Q k s k } = T Trace Q kH H kH H k Q k .
Lk
Lk

The power of the (asynchronous) interference leakage at MS j due to x k , the signal meant for
MS k, can be shown to be
P
E T
Lk

kj( b1,b 2)Trace ( Q (kb1) H H (jb1) H H (jb 2) Q (kb 2) ) .

( b1, b 2)

The noise power at MS k is PNk = N 0 N R . Therefore, the SLNR for MS k is given by:
Trace Q kH M k Q k

SLNRk =
N 0 N R Lk +

j
( j k )

PT Trace kj(b1,b 2) Q(kb1) H H (jb1) H H (jb 2)Q(kb 2)

(b1,b 2)

(4.27)
Lk

Trace Q kH M k Q k
Trace Q kH N k Q k

q klH M k q kl
l =1
Lk

l =1

q klH N k q kl

where q kl is the lth column of Q k , M k = PT H kH H k , and N k = N 0 N R I +

PA
T

kj

. Here, A kj is

j
( j k )

given by
H
kj(11) H (1)
H (1)
j
j
(21) (2) H (1)
H
H

j
j
A kj = kj

( B1) ( B ) H (1)
kj H j H j

H
H (2)
kj(12) H (1)
j
j
H
(2)
H
kj(22) H (2)
j
j

kj( B 2) H (jB ) H H (2)

j

137

H
H (jB )
kj(1B ) H (1)
j

H
H (jB )
kj(2 B ) H (2)
j

( BB )
(B) H
( B)
kj H j H j

(4.28)

Notice that T1 ,..., TK can now be independently optimized. Even so, directly maximizing
(4.27) with respect to q k1...q kL is still analytically intractable. Therefore, we derive a lower
k

bound that can be maximized analytically. The SLNRk in (4.27) can be lower bounded as
[34]:
qH M q
SLNRk min klH k kl
l =1... Lk
q kl N k q kl

(4.29)

The following Lemma then states the optimal solution.

Lemma 4.1: The lower bound of SLNRk in (4.29) is maximized when:
q kl = v l (N k 1M k ) , for 1 l Lk ,

(4.30)

where v l ( A) is the eigenvector of matrix A corresponding to its lth largest eigenvalue.

Proof: Since q k1 ~ q kL are orthonormal vectors, the vector space V = span{q k1...q kL } has
k

dimension dimV = Lk . Since M k is Hermitian and N k is positive-definite, the CourantFischer Max-Min Theorem [72] applies. The maximum value of (4.29) is

qH M q
max min klH k kl
q k 1 ~ q kLk l =1... L q N q

kl k kl

q H M q

min H k = Lk ( N k1M k ) ,

= max
qV
V
q Nk q

dimV = Lk

(4.31)

where L ( A) is the Lth largest eigenvalue of A . By inspection, (4.30) satisfies (4.31) in

equality.

It can be seen that the single closed-form solution in (4.30) is less complex than the JWF
solution in Section 4.4.1.A. As a special case, when Lk = 1 , (4.27) can be directly maximized
by applying the Rayleigh-Ritz quotient theorem [27]. This reduces to the solution in [85]:

138

SLNRk =

q kH1M k q k1
1 ( N k 1M k ) . Finally, the fact that the overall amount of interference power
q kH1N k q k1

equals that of the leakage power suggests that the JLS algorithm is still effective for
interference mitigation. This intuition is verified numerically in Section 4.4.4.
C. Controlled Iterative Singular Value Decomposition (CISVD)
The previous two approaches, while simple, optimized metrics that were different from
the sum-rate spectral efficiency metric in (4.7). In this section, we directly optimize this
metric of interest; however, this is analytically intractable and prohibitively complicated
numerically. To directly improve the spectral efficiency in our BS cooperation scenario while
keeping acceptable computational complexity, we propose an iterative optimization
algorithm called controlled iterative SVD to determine (4.7). CISVD applies the alternate &
maximize algorithm [7][98] in a greedy manner. At each iterative step, Tk is optimized by
keeping the other pre-coding matrices, {Tj } j k , fixed. Thus, each step is a water-filling power
allocation on the equivalent matrix k 1/ 2 H k with unit additive noise power. Given the nonlinear nature of the problem, the iterations are continued only as long as the target function
increases. The iterations are initialized with the JLS solution in (4.30).
CISVD consists of the following three steps:
Step 1:

For k = 1, , K , calculate T1 ,

Step 2:

For k = 1, , K , fix {Tj } j k , set H k _ equ = k 1/ 2 H k , and update the precoding matrix Tk by

, TK according to (4.30)

eigen-beamforming and water-filling power allocation based on H k _ equ with noise power 1.

139

Step 3:

Repeat step 2 until the target function in (4.7) does not increase or the increase is less than a
pre-defined threshold.

The above three steps trivially guarantee that the algorithm stops. As we shall see in Section
4.4.4, CISVD provides significant performance gains with only a small number of iterations
and is significantly faster than random, non-intelligent search methods.
4.4.2 Generalization to Imperfect Timing-Advance Case
The three joint BS pre-coding designs of the previous section assume accurate timing
advance, i.e., the desired signal components are assumed to arrive synchronously. However,
imperfect timing advances are inevitable in practical systems because of imperfect delay
estimation, user mobility, inaccurate cross-BS synchronization, and MS synchronization
errors.
Let k(b ) denote the timing advance error (jitter) of BS b in sending the signal for MS k.
Therefore, BS b now advances the time when the signal for user k, x(kb ) (m) , is transmitted by
the interval
k(b) = k(b) + k(b ) .

(4.32)

Consequently, the new delay offset at MS k due to the signal transmitted by BS b for MS j is
given by
(jkb ) = k(b ) (j b ) .

(4.33)

In addition to the MUI term, J k , in (4.1), imperfect timing advance also results in ISI. By
applying (4.32), (4.1) is modified to

140

y k ( m) =

(b )
k

H (kb ) Tk( b )s k (m) + k( b ) H (kb ) Tk( b )s k ( mk( b ) ) +

(b )
k

Tj( b ) i (jkb ) + n k

j
b
( jk )

(4.34)

= H k k Tk s k (m) + O k (m) + J k + n k ,

where k(b ) = ( k(b ) ) 1 , O k (m) = k(b ) H (kb ) Tk(b ) s k (mk(b ) ) is the ISI term with k(b ) and
b

mk(b ) given by

(b )
(b )
k , mk

{
} {

(TS k(b ) ), m + 1 , if k(b ) > 0

= (TS k(b ) ), m 1 , if k(b ) < 0 .

if k(b ) = 0
{0, m} ,

The block diagonal matrix k = blockdiag { k(1) I NT

k( B ) I NT can be interpreted as the power

degradation matrix due to imperfect signal synchronization (it equals I N B for perfect timing
T

advance).
From (4.34), the information rate of MS k is
Rk = log I + k 1H k k Tk TkH kH H kH ,

where the covariance of the noise plus interference term now becomes:
k = N 0I +

k( b1) k( b 2) H (kb1) Tk( b1) Tk( b 2) H H (kb 2) H 1 sgn ( k( b1) ) sgn ( k( b 2) )

( b1, b 2 )

(jkb1,b 2 ) H (kb1) T j( b1) T j( b 2) H H (kb 2) H

)
.

j
( b1, b 2 )
( jk )

Here the indicator function 1(.) equals 1 if the input argument is 0, and equals 0, otherwise.
The function sgn (.) is the signum function. The new asynchronous interference coefficients,

( b1,b 2)
jk

} , which incorporate the inaccurate timing advances, are now determined by { }

(b)
jk

141

in (4.33), in the same way that (jkb ) was determined in (4.3). As the cooperative BSs can no
longer determine { (jkb ) } exactly, they cannot calculate { (jkb1,b 2) } exactly either. As we see
from (4.34), timing advance inaccuracy degrades performance in three ways: the power
degradation term, k , the imperfect knowledge of { (jkb1,b 2) } , and the additional ISI term, O k .
While the exact value of the jitter is, by assumption, unknown, its statistics, if known, can
certainly be determined and exploited by the cooperating BSs to mitigate the performance
degradation.
The new JWF and JLS designs incorporating the timing inaccuracy in addition to
asynchronous interference leakage are derived in Appendix F. For JWF, the joint pre-coder
of MS k is:
Tk = [Ck + k I ]1 k H kH ,

(4.35)

where k = blockdiag{ k(1) I NT , , k( B ) I NT } . The matrix Ck has submatrices

K

kj(b1b 2) H (jb1) H H (jb 2) + k(b1) k(b 2) Pk(b1b 2) H (kb1) H H (kb 2) ,

( jk )
=
K (bb ) (b ) H (b )
( b )2
( b )2
H (kb ) H H (kb ) ,
kj H j H j + ak + k
j =1

Ck(b1b 2)

for b1 b2,

(4.36)
for b1 = b 2 = b,

where

k(b ) = E

(b)
k

{ ( )} ,
(b )
k

k(b ) = Pr (k(b ) 0 ) E

(b )
k

{ (T

Pk( b1b 2) = Pr sgn (k( b1) ) = sgn (k( b 2) ) ,

k(b1) k(b 2) = E

( b1) ( b 2 )
k k

k( b ) ) + Pr (k(b ) < 0 ) E ( b ) ( TS k( b ) ) ,

k(b 2) } ,

( b1)
k

142

 k(b )2 = E
k(b )2 = E

(b)
k

(b)
k

{ }
{ } ,
{ } .
( b )2
k

( b )2
k

and kj( b1,b 2) = E ( b )

k

( b1,b 2)
jk

Given the knowledge of jitter-statistcs, the BSs can calculate the jitter-averaged
asynchronous interference leakage value, kj( b1,b 2) .
Determining the first moment of { (jkb1,b 2) } is difficult due to the modulo TS operation on
(jkb ) . However, the jitters are typically considerably smaller than the symbol duration.

Therefore, we can then assume that symbol index differences {m(jkb ) } in (4.2) do not change.
We then have jk(b ) = (jkb ) mod TS jk(b ) + k(b ) . This simplifies the calculation of the jitter-

averaged asynchronous interference coefficients, jk(b1,b 2) .

For JLS, Lemma 4.1 still holds, but with the following modified expressions of M k and
N k (denoted by M k and Nk , respectively):
H
k(1)2 H (1)
H (1)
k
k

M k = PT
( B ) (1) ( B ) H (1)
k k H k H k

H
k(1) k( B ) H (1)
H (kB )
k

k( B )2 H (kB ) H H (kB )

(4.37)

Nk = N 0 N R I + PT Akj ,
j =1

where Akj bears the same form as in (4.28), with kj(b1b 2) replacing kj(b1b 2) for j k , and
k(b1) k(b 2) Pk(b1b 2) replacing kk(b1b 2) .

143

4.4.3 Numerical Results

We simulate the downlink scenario of Figure 4-1, which consists of an urban micro-cell
network with two cells consisting of one BS and one MS each. The inter-BS distance is
500m. As BS cooperation results in performance gains when the signal from one BS does not
completely dominate the signal from the other BS, we consider scenarios in which the MSs
are uniformly distributed in the shaded area shown in Figure 4-1, in which the MSs are at
least 150m away from their closest BSs. The results average over 100 MS locations, with
100 channel realizations generated per MS location. The path-loss coefficient, , for all the
BS-MS channels is 2 (free-space) up to a distance of 30m, and is 3.7, for larger distances [69].
Without loss of generality, the path-losses are normalized with respect to the largest in-cell
path loss. The NLOS small-scale fading is assumed to be independent identically distributed
(i.i.d.) at the antenna elements, and the channels for different BS-MS pairs are independent
with each other. The symbol pulse, g (t ) , is rectangular with a duration, TS , of 1 s , which
corresponds to a downlink system bandwidth of 1 MHz that is shared by both BSs. Since the
delay spread in micro-cells is typically on the order of 100 ns [53], the flat-fading assumption
is justified. In all the considered scenarios we assume that L1 = L2 = L .
Similar as in Subsection 4.3.3, besides comparing the proposed schemes with the
conventional nullification method, which does not account for asynchronous interference, we
also compare them with the following two useful benchmarks: (i) a conventional SVD precoding beamformer that treats all interference as additive Gaussian noise [6][16][21], and (ii)
single-cell (interference-free) point-to-point MIMO.
144

Figure 4-5 Comparison of Pre-coding Schemes for NT = N R = L = 2

Figure 4-5 considers the case of two transmit antennas at the BS and two receive antennas
at the two MSs ( L = 2, NT = N R = 2 ) with perfect timing advance. It can be seen that in the
presence of asynchronous interference leakage, all the three proposed approaches, JLS, JWF,
and CISVD, outperform conventional nullification at all SNRs. Also shown in the figure is
the optimal joint pre-coding solution of (4.7), called BS Coop: Optimal, that is obtained
after intensive numerical computations using the Nelder-Mead method [46]. Since there are
B

no additional spatial dimensions providing diversity ( NT B = Lk ), even the optimal prek =1

coder performs worse than the single-cell point-to-point interference-free MIMO. To achieve

145

a spectral efficiency of 4bps/Hz, JWF requires 5 dB lower transmit power per user than JLS,
and CISVD requires 8 dB lower transmit power per user than JLS. Note, however, that JLS
involves the smallest computational effort among the three proposed schemes.

Figure 4-6 Comparison of Pre-coding Schemes for N T = 3, N R = L = 2

Figure 4-6 and Figure 4-7 consider scenarios in which redundant spatial dimensions are
K

present, i.e., NT B > Lk . Figure 4-6 considers N T = 3, N R = L = 2 , and Figure 4-7 considers
k =1

NT = N R = 3, L = 2 . JLS and CISVD now achieve much higher gains than the conventional

nullification method. Over a large SNR range, CISVD even achieves a higher spectral
efficiency than the single-cell point-to-point interference-free MIMO scenario. In Figure 4-6,
JWF outperforms JLS and the conventional nullification method at low to intermediate SNRs.
146

JWF does not apply to Figure 4-7 as N R > L . Extensive investigations (not shown here)
reveal that CISVD typically stops after 5 to 7 iterations.

Figure 4-7 Comparison of Pre-coding Schemes for NT = N R = 3, L = 2

In Figure 4-8, we verify the intelligence of the proposed CISVD method in the
N T = 3, N R = L = 2 scenario. Specifically, we compare the CISVD algorithm, where the hill

climbing procedure starts from the JLS result (4.30), with the following two algorithms. The
first algorithm is a random search in which we randomly generate many pre-coding matrix
pairs of T1 and T2 , each of which satisfies the power constraints, and choose the pair that
maximizes (4.7). Assuming that 10 iterations are required before the CISVD converges, we
generate 20 matrix pairs for random search to compare the two approaches fairly. The second

147

algorithm is the CISVD algorithm, but with 10 different randomly selected initial starting
points. The one with the best performance is selected. From the figure, we see that the
proposed CISVD greatly outperforms random search. Despite its ten-fold lower complexity,
CISVD with JLS as the starting point also outperforms the randomly initiated CISVD. Using
the JLS solution as the initial point is thus an important factor contributing to the
computational efficiency and superior performance of CISVD.

Figure 4-8 The Effectiveness of CISVD in NT = 3, N R = L = 2 Scenario

Figure 4-9 considers the case where timing advance inaccuracies are present (c.f.
Subsection 4.4.2). The timing advance inaccuracies, {J k(b ) } , are

each assumed to be

uniformly distributed in the interval [0.1TS ,0.1TS ] and are independent of each other. The

148

figure shows that the modified JWF solution (in (4.35) and (4.36)), marginally improves the
performance over the original design (in (4.25)). However, the modified JLS solution (in
(4.37)) reduces the transmit power per user by 2.5 dB compared to the original JLS solution
(in (4.30)).

The conventional nullification method is inapplicable as timing advance

inaccuracy was not at all modeled by it.

Figure 4-9 JWF and JLS with Timing Inaccuracy in the NT = 3, N R = L = 2

Scenario

4.4.4 Summary
In Section 4.4, we investigated the impact of asynchronous interference on the downlink
performance of multi-user MIMO systems that use linear pecoding and exploit BS

149

cooperation. We saw that when cooperative BSs jointly transmit data intended for multiple
users, the timing advance can only be chosen such that the desired signal components are
perfectly aligned at their intended MSs. However, as a consequence, the different
propagation delays of the interfering signals then ensure that they are asynchronous. Not
accounting for this fact in the BS cooperation design (e.g. Section 4.3) degrades the system
performance considerably. We developed three linear pre-coding algorithms, namely JWF,
JLS, and CISVD, that markedly outperform conventional designs that ignored the
asynchronicity of interference. CISVD realizes significant spectral efficiency gains, JLS
achieves a good trade-off between interference mitigation and complexity, and JWF performs
well at low to intermediate SNRs. JLS and CISVD perform exceptionally well in channels
that have redundant spatial dimensions. By utilizing jitter-statistics, the JWF and JLS
solutions were extended to handle the practical scenario in which the timing-advance
inaccuracies exist.

4.5 Conclusions
In this chapter, we investigate the problem of downlink cooperative transmissions among the
cooperative BSs for interference mitigation. Specifically, the performance bounds are
explored in the synchronous cases, where some other advantages of BS cooperations, other
than intereference mitigations are also analyzed. Further considering more practical scenarios,
where the inter-cell interfering signals are by nature asynchronous at each MS, we present a
fundamental mathemeatical framework, and propose some novel joint pre-coding algorithms

150

for BS cooperation. These algorithms are also partly extendable to the senario with timing
errors.
Essentially, our work in this chapter constitutes a solid step towards realizing the great
potential of BS cooperation in practical interference-limited settings for multi-user MIMO
systems.

151

Chapter 5
Conclusions and Future Work

5.1 Conclusions
With the ever increasing demands for high data rate wireless communication systems, MIMO
systems have been considered as one of the key enabling techniques for current and emerging
wireless systems [45][105][106]. In this dissertation, we have studied some important
problems in point-to-point and multiuser MIMO, making certain contributions to both theory
and practice of this timely research topic.
As a promising technique that introduces diversity improvement with low cost and
relatively small overhead, antenna selection is investigated in the first part of this dissertation,
particularly on: (1) diversity property analysis in MIMO-SM systems with practical coding
and decoding schemes; (2) fast selection algorithm designs; and (3) application in
commercial wireless systems. The first two problems are addressed via novel and effective
geometric tools, while for the third a training/calibration protocol for IEEE 802.11n WLANs
152

is proposed. These geometric tools are also extended to analyze some open problems in
MIMO, the diversity-multiplexing tradeoff in V-BLAST and SDMA systems employing
ordered SIC receivers.
While some systems involve only point-to-point MIMO, many others need to address the
impairments induced by multiuser MIMO, especially for cellular systems, which undergos
severe inter-cell interference in the downlink. In the second part of this dissertation, we are
mainly concerned with the multiuser multicell downlink MIMO scenario, and investigate the
potential of cooperative transmission among adjacent base stations for effectively combating
the inter-cell interference. Our idea of interference mitigation through BS cooperation is first
exposed with a quasi-synchronous model, by which we also explore some other advantages
like channel rank/conditioning improvement and macro-diversity protection. Further
considering a more practical channel model, where signals from different BSs in the
downlink are by nature asynchronous at each MS, we propose some novel and effective
algorithms achieving different levels of tradeoffs between interference mitigation effects and
computational complexity. These algorithms are also extended to the scenarios with
additional timing errors.

5.2 Achievements
Overall, we have tackled some open and interesting problems in MIMO study, and have
also endeavored to improve the performance of MIMO systems in real operating scenarios
and accelerate their employment in future wireless networks

153

Part of this PhD work has been published or submitted for publication [108][123], which
include one book chapter, two journal papers, eight conference papers and four more journal
papers under review. The protocols in Section 4.4 (proposed during the authors internship at
MERL) have been accepted in the current draft specifications of IEEE 802.11n standard
[105]. Meanwhile, seven patents were filed at MERL.

5.3 Future Work

Finally we list below some of the interesting topics that desearve further study: the exact
diversity order of antenna selection in SM-MIMO with arbitrary numbers of RF chains (L);
the exact optimal diversity-multiplexing tradeoff for SDMA systems with ordered SIC
receivers; alternative asynchronous BS cooperative pre-coding designs with better tradeoff
between performance and complexity; and the BS cooperation with insufficient transmitter
CSI.

154

Bibliography
[1]

I. Bahceci, T. M. Duman, and Y. Altunbasak, Antenna selection for multiple-antenna

transmission systems: performance analysis and code construction, IEEE Trans. Info.
Theory, vol. 49, no. 10, pp. 2669-2681, Oct. 2003.

[2]

P. W. Baier, M. Meurer, T. Weber and H. Troeger, Joint transmission (JT), an alternative

rationale for the downlink of time division CDMA using multi-element transmit antennas,
Proc. 2000 IEEE 6th Int. Symp. Spread Spectrum Techniques, vol. 1, pp. 1-5, Parsippany, NJ,
Sept. 2000.

[3]

R. S. Blum, J. H. Winters and N. R. Sollenberger, On the capacity of cellular systems with

MIMO, Proc. 2001 Fall IEEE Vehicular Technology Conference, pp. 1220 -1224, vol.2,
Atlantic City, 7-11 Oct. 2001.

[4]

M. Capaccioli, and D. Rispo, A technique to realize base stations on-air frame

synchronization in TD-SCDMA system, Proc. IEEE Vehicular Technology Conference,
2003. VTC-Fall 03, vol. 2, pp. 982-986, Oct. 2003.

[5]

G. Caire, S. Shamai, On the achievable throughput of a multiantenna Gaussian broadcast

channel, IEEE Trans. Information Theory, vol. 49, no. 7, pp. 1691-1706, July 2003.

[6]

S. Catreux, P. F. Driessen, and L. J. Greenstein, Attainable throughput of an interferencelimited multiple-input multiple-output (MIMO) cellular system, IEEE Trans.
Communications, vol. 49, no. 8, pp. 479-493, Aug 2001.

[7]

J.-H. Chang, L. Tassiulas, and F. Rashid-Farrokhi, Joint transmitter receiver diversity for
efficient space division multiaccess, IEEE Trans.Wireless Communications, vol. 1, no. 1,
pp. 16-27, Jan. 2002.

[8]

D. Chizhik, G. J. Foschini, M. J. Gans and R. A. Valenzuela, Keyholes, correlations, and

capacities of multielement transmit and receive antennas, IEEE Trans. Wireless
Communications, vol. 1, no. 2, pp. 361 -368, April 2002.

[9]

R. L. Choi, and R. D. Murch, "A downlink decomposition transmit preprocessing technique

for multi-user MIMO systems," Proc. IST Mobile and Wireless Telecommunications Summit,
June 2002.

[10]

R. Choi, and R. Murch, MIMO transmit optimization for wireless communication systems,
Proc. 1st IEEE International Workshop on Electronic Design, Test and Applications, pp. 3337, Jan. 2002.

[11]

Y. S. Choi, A. F. Molisch, M. Z. Win, and J. H. Winters, Fast algorithms for antenna

selection in MIMO systems, Proc. IEEE Vehicular Technology Conference, Fall 2003,
Orlando, FL., Oct. 2003.

[12]

M. Costa, Writing on dirty paper, IEEE Transactions on Information Theory, vol. 29 no. 3,
pp. 439 -441, May 1983.

155

[13]

H. Dai, and H. V. Poor, Asymptotic spectral efficiency of multicell MIMO systems with
frequency-flat fading, IEEE Trans. Signal Processing, vol. 51, no. 11, pp.2976-2989, Nov.
2003.

[14]

H. Dai and H. V. Poor: "Iterative Space-Time Processing for Multiuser Detection in

Multipath CDMA Channels," IEEE Trans. Signal Processing, vol. 50, no. 9, pp. 2116-2127,
Sept. 2002.

[15]

H. Dai, L. Mailaender, and H. V. Poor, CDMA Downlink Transmission with Transmit

Antenna Arrays and Power Control in Multipath Fading Channels, European Journal on
Wireless Communications and Networking, 3rd quarter, 2004.

[16]

H. Dai, A. F. Molisch, and H. V. Poor, "Downlink capacity of interference-limited MIMO

systems with joint detection," IEEE Trans. Wireless Communications, vol. 3, no.2, pp. 442453, Mar. 2004.

[17]

H. A. David, Order Statistics, 3rd edition, John Wiley & Sons, 2003.

[18]

A. Edelman, Eigenvalues and Condition Numbers of Random Matrices, PhD dissertation,

MIT, 1989. Availalble at http://math.mit.edu/~edelman/thesis/thesis.ps

[19]

R. F.H. Fischer, C. Windpassinger, A. Lampe, and J. B. Huber, MIMO precoding for

decentralized receivers, Proc. 2002 IEEE International Symposium on Information Theory,
pp. 496, June 30~July 5, 2002.

[20]

G. J. Foschini, "Layered space-time architecture for wireless communication in a fading

environment when using multi-element antennas," Bell Labs Tech. J., vol. 2, no. 2, pp. 41-59,
Fall 1996.

[21]

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, pp. 311-335, Mar. 1998.

[22]

G. J. Foschini, G. D. Golden, R. A. Valenzuela and P. W. Wolniansky, Simplified

processing for high spectral efficiency wireless communication employing multi-element
arrays, IEEE J. Selected Areas Comm., vol. 17, pp. 1841-1852, 1999.

[23]

G. J. Foschini, H. Huang, K. Karakayali, R. A. Valenzuela, and S. Venkatesan, The value of

coherent base station coordination, Proc. 2005 Conference on Information Sciences and
Systems (CISS 05), The Johns Hopkins University, Mar. 16-18, 2005.

[24]

M. Gharavi-Alkhansari, and A. B. Gershman, Fast antenna subset selection in MIMO

systems, IEEE Trans. Signal Processing, vol. 52, no. 2, pp. 339-347, Feb. 2004.

[25]

A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, Capacity limits of MIMO

channels, IEEE J. Select. Areas Commun., vol. 21, no. 5, pp. 684-702, June 2003.

[26]

A. Goldsmith, Wireless Communications, Cambridge University Press, 2005.

[27]

G. Golob, and C. V. Loan, Matrix computations, 3rd edition, The John Hopkins University
Press, 1996.

156

[28]

D. A. Gore, R. W. Health, Jr. and A. J. Paulraj, Statistical antenna selection for spatial
multiplexing systems, Proc. 2002 IEEE International Conference on Communications, vol.
1, pp. 450 -454 , New York City, 28 April- 2, May, 2002.

[29]

D. A. Gore, R. U. Nabar, and A. J. Paulraj, Selecting an optimal set of transmit antennas for
a low rank matrix channel, Proc. Int. Conf. Acoust., Speech and Signal Processing, Istanbul,
Turkey, June 2000.

[30]

D. A. Gore, and A. J. Paulraj, MIMO antenna subset selection with space-time coding,
IEEE Trans. Signal Processing, vol. 50, no. 10, pp. 2580-2588, Oct. 2002.

[31]

A. Gorokhov, Antenna selection algorithms for MEA transmission systems, Proc. Int.
Conf. Acoust., Speech and Signal Processing, vol. 3, pp. 2857-2860, May 2002.

[32]

A. Gorokhov, D. A. Gore, and A. J. Paulraj, Receive antenna selection for MIMO spatial
multiplexing: theory and algorithms, IEEE Trans. Signal Processing, vol. 51, no. 11, pp.
2796-2807, Nov. 2003.

[33]

A. Gorokhov, D. Gore, and A. Paulraj, Diversity versus multiplexing in MIMO systems with
antenna selection, Proc. Allerton Conf. on Communication, Control, and Computing,
Monticello, IL, Oct. 2003.

[34]

I. S. Gradsbteyn, and I. M. Ryzbik, Table of Integrals, Series, and Products, 6th Edition,
Academic Press, 2000.

[35]

B. Hassibi, Random matrices, integrals and space-time systems, DIMACS Workshop on

Algebraic Coding and Information Theory, Rutgers University, Dec. 2003. Available at
http://dimacs.rutgers.edu/Workshops/ CodingTheory/slides/hassibi.ppt

[36]

R. W. Heath, and A. Paulraj, Antenna selection for spatial multiplexing systems based on
minimum error rate, Proc. IEEE International Conference on Communications 2001, ICC
01, vol. 7, pp. 2276-2280, June 2001.

[37]

H. Huang, H. Viswanathan, and G.J. Foschini, Multiple antennas in cellular CDMA

systems: transmission, detection, and spectral efficiency, IEEE Trans. on Communications,
vol. 1, no. 3, pp. 383-392, July 2002.

[38]

M. T. Ivrlac, W. Utschick, and J. A. Nossek, Fading correlation in wireless MIMO

communication systems, IEEE Journal on Selected Areas in Communications, vol. 21, no. 5,
pp. 819-828, June 2000.

[39]

S. A. Jafar, G. J. Foschini , and A. Goldsmith, PhantomNet: exploring optimal multicellular

multiple antenna systems, Proc. 2002 Fall IEEE Vehicular Technology Conf. (VTC), vol. 1,
pp. 261-265, Vancouver, Canada, Sept. 2002.

[40]

S. A. Jafar, and A. Goldsmith, Transmitter optimization for multiple antenna cellular

systems, Proc. 2002 IEEE International Symposium on Information Theory, pp. 50, 30 June5 July 2002.

[41]

Y. Jiang, X. Zheng and J. Li, Asymptotic performance analysis of V-BLAST, Proc. 2005
IEEE Global Communications Conference (GLOBECOM), St. Louis, MO, Nov. 2005.

[42]

N. Jindal, S. A. Jafar, S. Vishwanath, and A. Goldsmith, Sum power iterative water-filling

for multi-antenna gaussian broadcast channels, Conference Record of the Thirty-Sixth
157

Asilomar Conference on Signals, Systems and Computers, 2002, vol. 2, pp. 1518 -1522, Nov.
3-6, 2002.
[43]

N. Jindal, S. Vishwanath, A. Goldsmith, On the duality of Gaussian multiple-access and

broadcast channels, Proc. 2002 IEEE International Symposium on Information Theory, pp.
500, 30 June-5 July 2002.

[44]

M. Joham, K. Kusume, M. H. Gzara, W. Utschick, and J. A. Nossek, "Transmit Wiener filter

for the downlink of TDD DS-CDMA systems," Proc. IEEE 7th Int. Symp. on SpreadSpectrum tech. and Appl., Prague, Czech Republic, vol. 1, pp. 9-13, Sept. 2002.

[45]

H. Kaaranen, A. Ahtiainen, L. Laitinen, S. Naghian, and V. Niemi, UMTS Networks:

Architechture, Mobility and Services, Wiley, 2005.

[46]

J. C. Lagarias, J. A. Reeds, M. H. Wright, and P. E. Wright, "Convergence Properties of the

Nelder-Mead Simplex Method in Low Dimensions," SIAM Journal of Optimization, Vol. 9,
Number 1, pp.112-147, 1998.

[47]

B. Li, D. Xie, S. Cheng, J. Chen, P. Zhang, W. Zhu, and B. Li, "Recent advances on TDSCDMA in China," IEEE Communications Magazine, Jan. 2005.

[48]

T. M. Lok and T. F. Wong, Transmitter and receiver optimization in multicarrier CDMA

systems, IEEE Trans. Communications, vol. 48, no. 7, pp. 1197-1207, July 2000.

[49]

S. Loyka, and F. Gagnon, Performance analysis of the V-BLAST algorithm: an analytical

approach, IEEE Trans. Wireless Communications, vol. 3, no. 4, pp. 1326-1337, July 2004.

[50]

S. Loyka, and F. Gagnon, Analytical framework for outage and BER analysis of the VBLAST algorithm, Proc. International Zurich Seminar on Communications (IZS), pp. 120123, Feb. 2004.

[51]

C. D. Meyer, Matrix analysis and applied linear algebra, SIAM, 2000.

[52]

A. F. Molisch, and M. Z. Win, MIMO systems with antenna selection-an overview, IEEE
Microwave Magazine, vol. 5, no. 1, pp 46-56, Mar. 2004.

[53]

A. F. Molisch, Wireless Communications, IEEE Press-Wiley, 2005.

[54]

A. F. Molisch, M. Z. Win, and J. H. Winters, Capacity of MIMO systems with antenna

selection, Proc. IEEE Int. Conf. on Commu., ICC 01, vol. 2, pp. 570-574, June 2001.

[55]

R. J. Muirhead, Aspect of Multivariate Statistics Theory, New York: Wiley, 1982.

[56]

R. U. Nabar, D. A. Gore, and A. J. Paulraj, Optimal selection and use of transmit antennas in
wireless systems, Proc. Int. Conf. on Telecommunications, Acapulco, Mexico, May, 2000.

[57]

R. Narasimhan, Spatial Multiplexing with transmit antenna and constellation selection for
correlated MIMO fading channels, IEEE Trans. Signal Processing, vol. 51, no. 11, pp.
2829-2838, Nov 2003.

[58]

Nelder, J.A. and R. Mead, "A Simplex Method for Function Minimization," Computer J.,
Vol.7, pp 308-313, 1965.

158

[59]

B. L. Ng, J. S. Evans, S. V. Hanly, and D. Aktas, Transmit beamforming with cooperative

base stations, Proc. IEEE International Symposium on Information Theory, ISIT 05, pp.
1431-1435, Adelaide, Australia, Sept. 2005.

[60]

B. O'Hara and A. Petrick, "The IEEE 802.11 Handbook: A Designer's Companion", 2nd ed.,
IEEE Standards Publications (2005).

[61]

Z. Pan, K.-K. Wong, and T.-S. Ng, "Generalized multiuser orthogonal space-division
multiplexing," IEEE Trans. Wireless Commun., vol. 3, no. 6, pp. 1969-1973, Nov. 2004.

[62]

A. Papoulis, and S. U. Pillai, Probability, Random Variables, and Stochastic Processes.

Boston, MA: McGraw Hill, 2002.

[63]

A. Pascual-Iserte, A. I. Perez-Neira, and M. A. Lagunas, An approach to optimum joint

beamforming design in a MIMO-OFDM multiuser system, European Journal on Wireless
Communications and Networking, 2004, no. 2, pp. 210-221, 4th Quarter, 2004.

[64]

A. Paulraj, R. Nabar, and D. Gore, Introduction to Space-Time Wireless Communications,

Cambridge, 2003.

[65]

D. Pedoe, A Course of Geometry for Colleges and Universities, Cambridge: University Press,
1970.

[66]

N. Prasad and M. K. Varanasi, Analysis of decision feedback detection for MIMO Rayleigh
fading channels and optimization of power and rate allocations, IEEE Trans. Info. Theory,
vol. 50, no. 6, pp. 1009-1025, June 2004.

[67]

N. Prasad and M. K. Varanasi, Outage analysis and optimization for multiaccess/V-BLAST

architecture over MIMO Rayleigh fading channels, Proc. 41st Annual Allerton Conf. on
Comm. Control, and Comput., Monticello, IL, Oct., 2003.

[68]

J. G. Proakies, Digital Communications, 3rd Ed, McGraw-Hill, New York, 1995.

[69]

T. S. Rappaport, Wireless Communications: Principles and Practice, Prentice-Hall PTR, NJ,

1996.

[70]

M. Rudlf, and B. Jechoux, Design of concatenated extended complementary sequences for

inter-base station synchronization in WCDMA TDD mode, Proc. IEEE Global
Telecommunications Conference, 2001. GLOBECOM 01, vol. 1, pp. 674-679, Nov. 2001.

[71]

H. Sato, An outer bound on the capacity region of broadcast channel, IEEE Trans. on Info.
Theory, vol. 24, pp. 374-377, May 1978.

[72]

J. R. Schott, Matrix analysis for statistics, 2nd ed, Wiley, 2004.

[73]

M. Schubert, and H. Boche, Solution of the multiuser downlink beamforming problem with
individual SINR constraints, IEEE Trans. Vehicular Technology, vol. 53, no. 1, pp. 18-28,
Jan. 2004.

[74]

S. D. Senturia, Microsystem Design, Springer, 2000.

[75]

S. Serbetli, and A. Yener, Transceiver optimization for multiuser MIMO systems, IEEE
Trans. Signal Processing, vol. 52, pp. 214-226, Jan. 2004.

159

[76]

S. Shamai and B. M. Zaidel, Enhancing the cellular downlink capacity via co-processing at
the transmission end, Proc. 2001 Spring IEEE Vehicular Technology Conf., pp. 1745-1749,
Rhodes, Greece, May 2001.

[77]

L. Shao, and S. Roy, Downlink Multicell MIMO-OFDM: An Architecture for Next

Generation Wireless Networks, Proc. IEEE Wireless Communications and Networking
Conference (WCNC), New Orleans LA, March, 2005.

[78]

D. Shiu, G. J. Foschini, M. J. Gans and J. M. Kahn, Fading correlation and its effect on the
capacity of multielement antenna systems, IEEE Trans. on Communications, vol. 48, no. 3,
pp. 502-513, March 2000.

[79]

M. K. Simon, and M. S. Alouini, Digital Communications over Fading Channels, New York:
John Wiley, 2000.

[80]

Q. H. Spencer, and M. Haardt, Capacity and downlink transmission algorithms for a multiuser MIMO channel, Conference Record of the Thirty-Sixth Asilomar Conference on
Signals, Systems and Computers, vol. 2, pp. 1384-1388, Nov. 2002.

[81]

Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, "Zero-forcing methods for downlink

spatial multiplexing in multiuser MIMO channels," IEEE Trans. Signal Processing, vol. 52,
no. 2, pp. 461-471, Feb. 2004.

[82]

W.-H. Steeb, Matrix calculus and Kronecker product with applications and C++ programs,
World Scientific, 1998.

[83]

G. L. Stuber, J. R. Barry, S. W. McLaughlin, Y. Li, M. A. Ingram, and T. G. Pratt,

Broadband MIMO-OFDM wireless communications, Proceedings of the IEEE, vol. 92, no.
2, pp 271-294, Feb. 2004.

[84]

P. Sudarshan, N. B. Mehta, A. F. Molisch, and J. Zhang, Spatial multiplexing and channel

statistics-based RF pre-processing for antenna selection, Proc. IEEE Global
Communications Conference (GLOBECOM), Nov. 2004.

[85]

A Tarighat, M. Sadek, and A. H. Sayed, A multi user beamforming scheme for downlink
MIMO channels based on maximizing signal-to-leakage ratios, Proc. IEEE International
Conference on Acoustics, Speech, and Signal Processing (ICASSP), vol. 3, pp. 1129-1132,
Philadelaphia, PA, March 2005.

[86]

E. Telatar, "Capacity of multi-antenna Gaussian channels," European Trans. Telecomm., vol.

10, no. 6, pp. 585-595, Nov.-Dec. 1999.

[87]

D. N. C. Tse, and P. Viswanath, Fundamentals Of Wireless Communication, Cambridge

University Press, 2005.

[88]

D. N. C. Tse, P. Viswanath, L. Zheng, Diversity-multiplexing tradeoff in multiple-access

channels, IEEE Trans. Inform. Theory, vol. 50, no. 9, pp. 1859-1874, Sept. 2004.

[89]

S. Verd, Multiuser Detection, Cambridge, UK: Cambridge University Press, 1998.

[90]

H. Viswanathan, S. Venkatesan, and H. Huang, Downlink capacity evaluation of cellular

networks with known-interference cancellation, IEEE J. Select. Areas Commun., vol. 21, no.
5, pp. 802-811, June 2003.

160

[91]

X. Wang and H. V. Poor, "Space-time multiuser detection in multipath CDMA channels,"

IEEE Trans. Signal Processing, vol. 47, no. 9, pp. 2356-2374, Sept. 1999.

[92]

Z. Wang, and G. B. Giannakis, A simple and general parameterization quantifying

performance in fading channels, IEEE Trans. Communications, vol. 51, no. 8, pp. 13891397, Aug. 2003.

[93]

C. Windpassinger, R. F. H. Fischer, T. Vencel, and J. B. Huber, "Precoding in multiantenna

and multiuser communications," IEEE Trans. Wireless Commun., vol. 3, no. 4, pp. 13051316 July 2004.

[94]

J. H. Winters, On the capacity of radio communications systems with diversity in Rayleigh

fading environments, IEEE J. Sel. Area Commun. (JSAC), vol. 5, no. 5, pp. 871-878, June
1987.

[95]

P.W. Wolniansky, G. J. Foschini, G. D. Golden, R. A. Valenzuela, V-BLAST: an

architecture for realizing very high data rates over the rich-scattering wireless channel, 1998
URSI International Symposium on Signals, Systems, and Electronics, pp. 295 300, 29 Sept.2 Oct. 1998.

[96]

K.-K. Wong; R. D. Murch, R.S.-K. Cheng, and K. B. Letaief, Optimizing the spectral
efficiency of multiuser MIMO smart antenna systems, Proc. IEEE Conference on Wireless
Communications and Networking, vol. 1, pp. 426-430, Chicago, IL Sept. 2000.

[97]

K.-K. Wong, R. S.-K. Cheng, K. B. Letaief, and R. D. Murch Adaptive antennas at the
mobile and base stations in an OFDM/TDMA system, IEEE Trans. Communications, vol.
49, no. 1, pp. 195-206, Jan. 2001.

[98]

K.-K. Wong, R. D. Murch, and K. B. Letaief, Performance enhancement of multiuser

MIMO wireless communication systems, IEEE Trans. Communications, vol. 50, no. 12, pp.
1960-1970, Dec. 2002.

[99]

A. D. Wyner, Shannon-theoretic approach to a Gaussian cellular multiple-access channel,

IEEE Trans. on Inform. Theory, vol. 40, no. 6, pp. 1713-1727, Nov. 1994.

[100]

W. Yu, and J. Cioffi, Trellis precoding for the broadcast channel, Proc. IEEE Global
Telecommunications Conference, 2001. GLOBECOM 01, vol. 2, pp. 1344-1348, Nov. 2001.

[101]

W. Yu, W. Rhee, S. Boyd, and J. M. Ciofli, Iterative water-filling for Gaussian vector
multiple access channels, Proc. 2001 IEEE International Symposium on Information
Theory, pp. 322, Washington, DC, 24-29 June. 2001.

[102]

W. Yu, A dual decomposition to the sum power Gaussian vector multiple access channel
sum capacity problem, Proc. 2003 IEEE Conference on Information Sciences and Systems,
the Johns Hopkins University, March, 2003.

[103]

J. Zhang, Y. Wu, S. Zhou, and J. Wang, Joint linear transmitter and receiver design for the
downlink of multuser MIMO systems, IEEE Communication Letters, vol. 9, no. 11, pp. 991993, Nov. 2005.

[104]

L. Zheng, D. N. C. Tse, Diversity and multiplexing: A fundamental tradeoff in multiple

antenna channels, IEEE Trans. Info. Theory, vol. 49, no. 5, pp. 1073 -1096, May 2003.

161

[105]

IEEE P802.11n/D1.0: Draft Amendment to Wireless LAN Media Access Control (MAC) and
Physical Layer (PHY) Specifications: Enhancements for Higher Throughput, Mar. 2006.

[106]

Draft IEEE Standard for LAN/MAN, Part 16: Air Interface for Fixed and Mobile Broadband
Wireless Access Systems, IEEE P802.16e/D12, Dec. 2005.

[107]

IEEE P802.11 Wireless LANs TGn Channel Models, IEEE 802.11-03/940r4, May 2004.

[108]

H. Zhang, and H. Dai, Chap 6: Cellular Networks: Design fundamentals and interference
mitigation, Modeling and Simulation of Wireless Networks: analysis, evaluation and
enhancement of QoS for wireless multimedia, Nova Science, to appear 2006.

[109]

H. Zhang, H. Dai, Q. Zhou, and B. L. Hughes, Transmit antenna selection for spatial
multiplexing systems: a geometrical approach, IEEE Trans. Info. Theory, submitted and
revised.

[110]

H. Zhang, and H. Dai, Cochannel interference mitigation and cooperative processing in

downlink multicell multiuser MIMO networks, European Journal on Wireless
Communications and Networking, 2004, no. 2, pp. 222-235, 4th Quarter, 2004.

[111]

H. Zhang, N. B. Mehta, A. F. Molisch, J. Zhang, and H. Dai, Joint transmission by

cooperative base stations in multiuser cellular downlink with asynchronous interference,
IEEE Trans. Wireless Communications, submitted.

[112]

H. Zhang, H. Dai, and B. L. Hughes, Analysis on the diversity-multiplexing tradeoff for

ordered MIMO SIC receivers, IEEE Trans. Communications, submitted.

[113]

H. Zhang, and H. Dai, Fast MIMO transmit antenna selection algorithms: a geometric
approach, IEEE Communication Letters, to appear.

[114]

H. Zhang, A. F. Molisch, and J. Zhang, Applying Antenna Selection in WLANs for

Achieving Broadband Multimedia Communications IEEE Trans. Broadcasting, submitted.

[115]

H. Zhang, A. F. Molisch, D. Gu, D. Wang, and J. Zhang, Implementing Antenna Selection

in Wireless LAN, to appear IEEE International Symposium on Broadband Multimedia
Systems and Broadcasting 2006, Las Vegas, NV, April 6-7, 2006.

[116]

H. Zhang, and H. Dai, Fast transmit antenna selection algorithms for MIMO systems with
fading correlation, Proc. Vehicular Technology Conference, Fall 2004, VTC Fall 04, Sept.
2004.

[117]

H. Zhang, H. Dai and B. L. Hughes, On the diversity-multiplexing tradeoff for ordered SIC
receivers over MIMO channels, 2006 IEEE International Conference on Communications
(ICC), Istanbul, Turkey, June 2006.

[118]

H. Zhang, and H. Dai, On the diversity order of transmit antenna selection for spatial
multiplexing systems, Proc. IEEE Global Telecommunications Conference, GLOBECOM
2005, Nov. 2005.

[119]

H. Zhang and H. Dai, A Geometrical Analysis on Transmit Antenna Selection for Spatial
Multiplexing Systems with Linear Receivers, Proc. 2005 IEEE International Symposium on
Information Theory (ISIT), Adelaide, Australia, Sept. 2005.

162

[120]

H. Zhang, N. B. Mehta, A. F. Molisch, J. Zhang, and H. Dai, Base station cooperation with
asynchronous interference in downlink multiuser MIMO cellular networks, submitted to ICC
07.

[121]

H. Zhang, and H. Dai, Delay constrained multiuser scheduling schemes based on upperlayer performance, Proc. IEEE Wireless Communications and Networking Conference
2006, WCNC 06, Apr. 2006.

[122]

H. Zhang, H. Dai and Q. Zhou, Base station cooperation for multiuser MIMO: joint
transmission and BS selection, Proc. 2005 Conference on Information Sciences and
Systems, CISS05, John Hopkins University, Mar. 2005.

[123]

H. Zhang, and H. Dai, On the capacity of distributed MIMO systems, Proc. 2004
Conference on Information Sciences and Systems, CISS 04, Princeton University, Mar. 17-19,
2004.

163

Appendix

164

Appendix A
Proof of Lemma 2.1
Proof: Suppose subset U j is selected following an arbitrary antenna selection rule, which
may be channel dependent. Given any channel realization H , let the corresponding selected
( j)
antenna subset be H j (as in (2.2)). Let random variable Rmin
denote the minimum squared

projection height for this antenna selection rule (c.f. (2.7)). We claim that the optimal
diversity order achieved for this antenna selection rule is given by
( j)
log Pr( Rmin
x)
.
x 0
log( x)

d (Lj ) = lim

(A.1)

( j)
with probability 1, the lemma
Since our antenna selection rule (2.10) dictates RSL Rmin

readily follows.
The claim of (A.1) is proved as follows. Let us consider a linear ZF receiver first. For
block fading channels, given any channel realization H , let Pe( j ) ( H) and Pel( j ) (H) denote
the conditional error probability of the whole frame and the lth substream, respectively,
assuming optimal one-dimensional coding for each substream. The worst link, with
)
conditional error probability Pe(_jmax
(H ) = max Pel( j ) (H) , has instantaneous post-detection
l

SNR

0 ( j )
Rmin (H ) 9. The corresponding average error probabilities are denoted as Pe( j ) , Pel( j )
L

)
and Pe(_jmax
. Clearly we have

( j)
( j)
Rmin
(H ) refers to the realization of Rmin
with respect to H .

165

)
Pe(_jmax
(H ) Pe( j ) (H) Pel( j ) (H ) ,

(A.2)

l =1

which leads to
L

)
max Pel( j ) Pe(_jmax
Pe( j ) Pel( j ) .
l

(A.3)

l =1

Since the upper and lower bounds above bear the same diversity order [104][89], we have
)
)
Pe( j )
Pe(_jmax

E { Pe(_jmax
(H )} .

(A.4)

)
It is known from Lemma 5 of [104] that the error probability Pe(_jmax
is (uniformly) lower

( j)
bounded by the corresponding outage probability, given by Pr log 1 + 0 Rmin
R0 ,
L

where R0 is the fixed data rate (see Section 2.2.2). Therefore the diversity order is upper
bounded by

d (ZFj)

( j ) L R0

log Pr Rmin
(2 1)
( j)
0
log Pr( Rmin
x)

lim
= lim
.
0
x
0

log( 0 )
log( x)

(A.5)

The above error probability lower bound is obtained without specifying any coding
scheme. On the other hand, the error probability with uncoded square M-QAM signaling can
serve as an upper bound [68][104]:
)
)
Pe(_jmax
Pe(_jmax_
QAM 4 P ( s0 s1 ) ,

(A.6)

where P( s0 s1 ) denotes the pairwise error probability between two closest QAM
constellation points. By Eq. (21) in [104], we have

166

3
( j)

P( s0 s1 )
P 0 Rmin

< 1 .
M 1
L

(A.7)

We therefore obtain
( j)
log Pr( Rmin
x)
,
x 0
log( x)

d (ZFj) lim

(A.8)

and thus have proved the claim of (A.1) for a linear ZF receiver. An immediate conclusion
from analysis above is that, one-dimensional channel coding has no impact on system
diversity order.
Note that the above analysis (see (A.3)) also indicates that
( j)
log Pr( Rmin
x)
x 0
log( x)

d (ZFj) = lim

)
log Pr( min { k( ,jZF
} 1)

k =1
L

= lim

log(1/ 0 )

)
log Pr( k( ,jZF
1)

= min lim

k =1
L 0

(A.9)

(A.10)

log(1/ 0 )

As for a linear MMSE receiver, following a similar approach, we have

MMSE
( j)

= lim

)
log Pr( min { k( ,jMMSE
} 1)

k =1
L

log(1/ 0 )

)
log Pr( k( ,jMMSE
1)

= min lim

log(1/ 0 )

k =1
L 0

In the following we complete the proof by showing that

)
)
Pr( k( ,jZF
1)
Pr( k( ,jMMSE
1) , k .

(A.11)

We alternatively have expressions of post-processing SNRs for ZF and MMSE receivers as

(see, e.g., [89])
)
k( ,jZF
=

0 ( j ) H
(hk )
L

( j)
I H k

(( H )

( j) H
k

167

H (k j )

(H )

( j) H
k

( j)
hk ,

(A.12)

( j)
k , MMSE

1
H
0 ( j ) H
L
( j)
( j) H
( j)
= ( h k ) I H k ( H k ) H k + I ( H (k j ) ) h (k j ) .
L
0

By matrix inversion Lemma [37], with A (k j ) = ( H (k j ) ) H (k j )

( j)
k , MMSE

( j)
k , ZF

= (h

(A.13)

, we have

( j) H
k

H A
( j)
k

( j)
k

L ( j)

( j)
( j) H
( j)
( j)
A k + I A k ( H k ) h k
k ( 0 ) .

(A.14)

Clearly k( j ) ( 0 ) 0 and thus

)
)
Pr( k( ,jMMSE
1)
Pr( k( ,jZF
1) .

(A.15)

)
Another key observation is that, k( j ) ( 0 ) is statistically independent with k( ,jZF
, as the latter

is proportional to the squared norm of the projection of h (k j ) onto the null space of ( H (k j ) ) ,
T

while the former is the correlation of h (k j ) with a vector in the range of H (k j ) [55][35].
Furthermore, as 0 , it can be shown that
a .s .

k( j ) ( 0 ) k( j ) ,

(A.16)

i.e., it converges almost surely to a positive random variable with finite mean. This
observation was also made in [37] where the PDF of k( j ) was explicitly given. For our
purpose, it is sufficient to know that Pr(k( j ) 1/ 2) is a positive probability of O(1) (the
same order as 1). We thus have

168

)
log Pr( k( ,jMMSE
1)

log(1/ 0 )

)
log Pr( k( ,jZF
+ k( j ) ( 0 ) 1)

log(1/ 0 )

)
log Pr( k( ,jZF
1/ 2)

log(1/ 0 )

log Pr(k( j ) ( 0 ) 1/ 2)
+
log(1/ 0 )

(A.17)

with
log Pr( k( j ) ( 0 ) 1/ 2)
log Pr( k( j ) 1/ 2)
= lim
=0,
0
0
log(1/ 0 )
log(1/ 0 )
lim

(A.18)

where (A.18) is due to the fact that almost sure convergence always leads to convergence in
distribution. (A.17) and (A.18) together gives rise to
)
)
Pr( k( ,jMMSE
1)
Pr( k( ,jZF
1) ,

(A.19)

and (A.11) follows with (A.15) and (A.19).

169

Appendix B
Lemmas Used for Proving Theorem 2.1
Lemma B.1: For any permutation of the integer array 1 ~ NT , denoted as k1 ~ k NT , we have

Pr Rk1k2 x, Rk2 k3 x,
, Rk( N

k
T 1) NT

i.e. random variables Rk1k2 , Rk2 k3 ,
, Rk( N

k
T 1) NT

x = [Pr( Rk1k2 x)]( NT 1) ,

(B.1)

are jointly independent.

Proof: Essentially Rki ki+1 is only a function of h ki and h ki+1 , denoted as Rki ki+1 = g (h ki , h ki+1 ) ,
therefore the conditional PDF of

Rki ki+1 given those variables appearing earlier in the

sequence admits:
f ( Rki ki+1 | Rki1ki ,
, Rk1k2 ) = f ( g (h ki , h ki+1 ) | g (h ki1 , h ki ),
, g (h k1 , h k2 ))

= f ( g (h ki , h ki+1 ) | g (h ki1 , h ki )) = f ( Rki ki+1 | Rki1ki ),

where the second equality holds because the states of h k1 ~ h ki1 do not affect h ki and h ki+1 .
Consequently, the above sequence forms a Markov chain.
We are left to prove the independence between
Rki ki+1 = h ki
2

Rki ki+1 and

Rki1ki . Given

sin 2 ki+1ki and Rki1ki = h ki1 sin 2 ki1ki , together with the independence between
2

h ki and h ki1 , and between vector norms and directions (angles) [55], we only need to

show that ki+1ki and ki1ki are independent.

170

Figure B-1 Independence of ki1ki and ki+1ki

Following a similar rotation approach as in [49], we define e1 ~ e N R as a fixed orthonormal

basis (e.g., Cartesian coordinates) of the vector space N R . We rotate [h ki1 , h ki , h ki+1 ] as a
whole so that h ki is parallel to e1 , denoted as [h
ki1 , h
ki+1 ] = Q( ki )[h ki1 , h ki+1 ] , where ki is the
angle between h ki and e1 , and Q( ki ) is the corresponding unitary rotation matrix. Since
[h ki1 , h ki+1 ] is an i.i.d. Gaussian matrix (therefore the joint distribution is rotationally

invariant) and is independent with ki , [h
ki1 , h
ki+1 ] is still i.i.d. Gaussian. Because ki+1ki and

ki1ki are unchanged after the rotation, and equal to the angles between h
ki+1 and e1 , and
between h
ki1 and e1 , respectively (see Figure B-1), given the fact that h
ki+1 and h
ki1 are
independent, it is straightforward to show that ki+1ki and ki1ki are independent, so are
Rki ki+1 and Rki1ki , and Lemma B.1 follows.

171


CorollaryB.1: 12 , 13 ,
, 1NT are jointly independent.

Proof: Using the same rotation approach as above, if we rotate [h1 , h 2 ,
, h NT ] as a whole
such that h1 is parallel to e1 , h
2 , , h
NT are jointly independent vectors; their angles with e1 ,
which are equal to 12 , 13 ,
, 1NT , respectively, are also jointly independent.

Lemma B.2: If 1 ,
, K are independent positive random variables, whose cumulative

distribution functions admit Fk ( x)
x nk , we have

K

nk
K

Pr k x
x k =1 .
k =1

(B.2)

Proof: At first we evaluate the exponential behavior of Pr(1 + 2 x) . Note that the
following expression assumes x 0 :
x 1

Pr(1 + 2 x) = f2 ( 2 )d 2 f1 (1 )d1
0
0

f ( ) F ( x )d
1

n1 1
1

x

n2

+ o(1n1 1 ) . ( x 1 ) n2 + o(( x 1 ) n2 ) d1

n1 1
1

x n1 + n2 ,
172

+ o(1n1 1 ) d1

where o( x) denotes a higher-order function of x such that lim o( x) / x = 0 . Next, we can

x 0

treat 1 + 2 as a new random variable exponentially equivalent to x n1 + n2 and evaluate

(1 + 2 ) + 2 following the same approach, whose CDF asymptotically behaves as x ( n1 + n2 )+ n3 .
Lemma B.2 follows after such repeated operations.

Lemma B.3: Let m( ) be a positive function of , which monotonically increases with ,

satisfying m(0) = 0 and m 1 ( x)
x n0 . If 1 ,
, K are independent positive random

variables, whose cumulative distribution functions admit Fk ( x)
x nk , we have

n0 nk
K

Pr m k x
Pr m max k x
x k =1 .

k
k =1

(B.3)

Proof: Since 1 ,
, K are independent, we get:

K

nk

Pr[max k x] = Fk ( x)
x k =1 .
k

(B.4)

k =1

By Lemma B.2 we have

K

nk
K

Pr k x
x k =1 .
k =1

(B.5)

Because both and m( ) are positive, and m( ) monotonically increases with , m 1 ( x)

is also a positive function of x, which monotonically increases with x. Therefore,
K

n0 nk
nk
Pr[m(max k ) x] = Pr[max k m ( x)]
[m ( x)] k =1
x k =1 .
1

173

(B.6)

Similarly,
K

n0 nk
K

Pr m k x
x k =1
k =1

(B.7)

and Lemma B.3 follows.

Lemma B.4:

For independent continuous random variables a , b1 and b2 satisfying

a 0 with Pr(a x)
x na , and 0 b1 , b2 1 with Pr(b1 x)
Pr(b2 x)
x nb , we have

.

Pr(ab1 x)
Pr(ab2 x) x na .

(B.8)

Proof: We have
1

x
Pr(ab1 x) = Pr(a ) f b1 ( y ).dy
y
0
=

x
x
0 Pr(a y ) fb1 ( y).dy + Pr(a y ) fb1 ( y ).dy

(B.9)

= t11 ( x) + t12 ( x),

and similarly,
Pr(ab2 x) =

x
x
0 Pr(a y ) fb2 ( y).dy + Pr(a y ) fb2 ( y).dy

(B.10)

= t21 ( x) + t22 ( x),

where > 0 is a fixed small positive number, small enough to make the PDF approximations
for b1 and b2 :
f b1 ( x) = c1 x nb 1 + o( x nb 1 ),
hold true.
174

f b2 ( x) = c2 x nb 1 + o( x nb 1 )

(B.11)

x
Since Pr(a ) is positive, and is a decreasing function of y, when x 0 we have
y
1

x
x
t12 ( x) Pr(a ) fb1 ( y ).dy Pr(a )
x na .

(B.12)

In another word, t12 ( x) x na .

On the other hand, since 0 b1 1 , we have
Pr(ab1 x) Pr(a x)
x na ,

(B.13)

which means that Pr(ab1 x) x na .

The same inequalities as in (B.12)(B.13) also hold true for t22 ( x) and Pr(ab2 x) ,
respectively.
Meanwhile, since is small enough, by applying (B.11), we get that

x
t11 ( x)
t21 ( x)
Pr(a ) y nb 1.dy .
0
y

(B.14)

If t12 ( x ) >
x na or t22 ( x ) >
x na , t11 ( x )  t21 ( x )
x na dominate and (B.8) follows. We are left
to check the case t12 ( x )
t22 ( x )
x na , which together with (B.14) leads to (B.8) as well.

Lemma B.5: With z , '0 , and 0 as defined in Section 2.2.3.1, we have

Pr( z sin 2 '0 x)
Pr( z sin 2 0 x) .

175

Proof: From (2.18)(2.32) the PDF of '0 = max '1k is derived through results in order
k

statistics:
f '0 ( ) =

where C =

M
[sin 2 ]M 1 sin 2 ,
NT 1
C

M = ( NT 1)( N R 1), (0, 0 ) ,

(B.15)

f1i ( x)dx, 2 i NT . Therefore, we have

Pr( z sin '0 x) =

F sin

M
= NT 1
C

F sin
z

t = sin
2

M
=
C NT 1

t1 =

f ' ( ).d
0

[sin 2 ]M 1 sin 2 .d

a0

x M 1
0 Fz t t .dt

(B.16)

t
a0

1
( x / a0 ) M 1
Ma0M
= NT 1 Fz
t1 .dt1 ,
C
t1
0

where a0 = sin 2 0 is a positive real number. On the other hand, by applying (2.37) we have

Pr( z sin 2 0 x) =

/2

=M

x
Fz 2 f0 ( ).d
sin

/2

t = sin 2

x
Fz 2
sin

M 1
2
[sin ] sin 2 .d

(B.17)

x
= M Fz t M 1.dt.
t
0

By comparing (B.16) and (B.17), we get Pr( z sin 2 '0 x)
Pr( z sin 2 0 x) .

176

Appendix C
The Derivation of (2.38)
We continue the evaluation of (B.17) to derive the polynomial expansion of Pr( z sin 2 0 x) :
1

x
Pr( z sin 0 x) = M Fz t M 1.dt
t
0
2

m =1/ t

= M Fz ( mx )
1

m M +1

.dm

(C.1)

M + NT 2

mk xk 1
= M 1 e mx
.dm

k ! m M +1
k =0
1
= 1 P1 ( x) P2 ( x),

where

P1 ( x) = M
k =0

xk
EM +1k ( x)
k!

and

P2 ( x) = M

M + NT 2

k = M +1

xk
E ( k M 1) ( x)
k!

with

e xm
Ek ( x) = k dm the integral exponential function. From [34], we have the following
m
1

recursive rules:
Ek +1 ( x) =

1 x
(e xEk ( x))
k

Ek +1 ( x) =

e x
k!

k 1

(1)i (k i 1)! xi +
i =0

(1) k k
x E1 ( x).
k!

Therefore, after some involved mathematical manipulations, we have

M 1

P1 ( x) = Me x ck x k ,
k =0

177

where ck =
i =0

(1) k i ( M k 1)!
, for k M 1 . We can expand e x by its Taylors series
( M i )!i !

and get the polynomial equivalent form: P1 ( x) = an x n . For n M 1 ,

n=0

(1) n k k (1)k i ( M k 1)!

an = M
( M i)!i !
k = 0 ( n k )! i = 0
n

=M

1 n (1)n n !( M k 1)! k (1)i M !

(n k )!M !
n ! k =0
i = 0 ( M i )!i !

n
(1)n k
1
k (1)i M
=
i
n ! k =0 M 1 i =0

k
n

n n

(
1)

1 n
k (1) k M 1
=

n ! k =0 M 1
k

(1) n
=
n!

(1)
k =0

where we use the equality [34]:

n=0
n 1,
=
k 0, 0 < n M 1,

M
k M 1
= (1)
. With a similar approach, we
i
k

(1)
i

i =0

can obtain aM =

1
, therefore
M!
P1 ( x) = 1

1 M
x + o( x M ) .
M!

(C.2)

On the other hand, P2 ( x) can be represented as

P2 ( x) = Me x

M + NT 2

k = M +1

(k M 1)! k M 1 x i + M

k!

i =0 i !

178

and after some manipulations we obtain

P2 ( x) = bM x M + o( x M ) ,

where bM = M

NT 3

(C.3)

k!

( M + k + 1)! . By combining (C.1), (C.2) and (C.3), we can derive (2.38).

k =0

Furthermore, from [34], we have

k!

( M + k + 1)! = M ( M + 1)! ,
k =1

therefore
NT 3

k =0

k!
k!
<
( M + k + 1)! k =0 ( M + k + 1)!

k!
1
1
1
1
,
+
=
+
=
( M + 1)! M ( M + 1)! ( M + 1)! M .M !
k =1 ( M + k + 1)!

so the coefficient of x M in (2.38) is always positive.

179

Appendix D
Lemma Used for Proving Proposition 2.5
Lemma D.1: For any NT L + 1 distinct integers within [1, NT ] , denoted as an ordered list
k1 ~ k NT L +1 , and any NT L + 1 different subsets from U1 ~ U NU , denoted as an ordered list

U j1 ~ U jN

T L +1

, suppose h ki U ji , and h ki U jl for l > i , i.e., h ki belongs to subset U ji but

not those after. We have:

(j
)
Pr Rk(1j1 ) x, Rk(2j2 ) x,
, Rk N NTL+L1+1 x = Pr( Rk(1j1 ) x)
T

( NT L +1)

(D.1)

where Rk( j ) is modified from Rk( j ) with k indicating the index of column vector in the original
channel matrix H , instead of H j 10.
Proof: From the definition, Rk(1j1 ) = Ph k1 , where P = I BB is the projection matrix to the
2

}{

null space of range U j1 \ h k1 ( U j1 \ h k1 represents the column vectors in U j1 except h k1 ), in

which B is composed of any basis of the subspace spanned by the remaining column vectors
in subset U j1 [21]. Since any projection matrix is idempotent, i.e. P 2 = P , its eigenvalues are
either 1 or 0. Noting that P is Hermitian, we can write the eigenvalue decomposition of P as
P = VV H , where V is unitary, and = diag (1N R L +1 , 0 L 1 ) . Therefore

Rk(1j1 ) = VV H h k1

180

= V H h k1 ,

(D.2)

where the second equality follows by the fact that a unitary transformation preserves length.
From the definition of P , the unitary matrix V H is independent of h k1 . Therefore the
conditional PDF

f Rk(1j1 ) | U j1 \ h k1

}) = f ( V

h k1

) = f ( h ) = f ( R ) ,
2

k1

( j1 )
k1

{U

where V0 H is a fixed unitary matrix dependent on the given realizations of

(D.3)

j1

\ h k1 , and

the second equality comes from the rotationally invariant property of the i.i.d. Gaussian

vector h k1 [55][35]. That is, Rk(1j1 ) is independent of any vector in range U j1 \ h k1 . It is

therefore straightforward to show that Rk(1j1 ) ~ Rk N NTL+L1+1
(j

are jointly independent and

identically distributed, and Lemma D.1 holds.

10

Lemma B.1 is a special case of Lemma D.1. However, the proof of Lemma B.1 bears some interesting
geometric elements, and is needed for Corollary B.1 and Proposition 2.2.

181

Appendix E
Derivation of JWF Solution in (4.25)
Using matrix calculus [82], the optimal precoding matrices, {Tk(b ) }k =1..K ,b =1... B , to be used by
each of the B BSs for all the K MSs are the solutions of the following set of Lagrangian
equations
A
=2
Tk(b )

j b=1
( jk )

( bb) ( b ) H
H j H (jb) Tk(b)
kj

+ 2H (kb ) H H (kb ) Tk(b ) + 2 k Tk(b ) 2H (kb ) H

, for b = 1,..., B .

(E.1)

=0

Doing so leads to Ck Tk + k Tk = H kH , and (4.25) follows. To calculate k , we apply the

eigenvalue decomposition on the Hermitian matrix Ck : Ck = U k k U kH , where
k = diag{k1 , k 2 ,..., k ( NT B ) } . Then,

where

bki = U kH H kH H k U k
ii
NT B

equation
i =1

NT B

} (

PT = Trace TkH Tk = Trace (Ck + k I ) 1 H kH H k (Ck + k I ) 1H =

i =1

Therefore

ki
= PT .
( x + bki ) 2

182

is

one

of

the

ki
,
2
k + bki )

roots

of

the

Appendix F
Derivations of JWF and JLS Solutions in (4.35), (4.36), and (4.37)
1. Modified JWF solution: In (4.34), the mean value of MSEk , obtained by averaging over
{ k(b ) } , can be shown to be:

MSEk = Trace k(b )2 H (kb ) Tk(b ) Tk(b ) H H (kb ) H k(b ) H (kb ) Tk(b ) + Tk(b ) H H (kb ) H + ( N 0 + 1)I N R

b =1
b =1
(F.1)

(b1,b 2) ( b1) ( b1) ( b 2) H

( b 2) H
( b1) ( b 2) ( b1b 2) ( b1) ( b1) ( b 2) H
( b 2) H
+
jk H k Tj Tj
Hk
+
k k Pk
H k Tk Tk
Hk
.

j k ( b1,b 2)
( b1,b 2)

In a manner similar to (4.35) and (4.36), maximizing the Lagrange objective function,

) MSE +

A {Tk(b ) }k =1..K ,b =1...B =

k =1

k =1

B (b ) H (b )
Trace Tk Tk PT ,

b =1

results in (E.1).

2. Modified JLS solution: From (4.34), the power of received desired signal component at
MS k, averaged over {k(b ) } , is given by
Pk =

PT
1
E ( b ) Trace ( Q kH kH H kH H k k Q k ) = Trace ( Q kH M k Q k ) ,
Lk {k }
Lk

where Mk follows from (4.37). When x k is transmitted for MS k, the jitter-statisticsaveraged MUI power it causes at MS j ( j k ) is given by
PL _ kj =

PT
Lk

( b1,b 2)

kj(b1,b 2)Trace ( Q (kb1) H H (jb1) H H (jb 2) Q (kb 2) ) ,

and the jitter-statistics-averaged ISI it causes at MS k itself is given by

PL _ kk =

PT
Lk

( b1,b 2)

k(b1) k( b 2) Pk(b1b 2)Trace ( Q (kb1) H H (kb1) H H (kb 2)Q (kb 2) ) .

183

Deriving a corresponding lower bound (as in (4.29)) for modified SLNR expression,
SLNRk =

Pk
K

PNk + PL _ kj

, and applying Lemma 4.1 to it, yields the solution in (4.37).

j =1

184