Interference Management in MIMO Networks

A Thesis
Presented to
The Academic Faculty
by
Sudhanshu Gaur
In Partial Fulllment
of the Requirements for the Degree
Doctor of Philosophy in Electrical and Computer Engineering
School of Electrical and Computer Engineering
Georgia Institute of Technology
August 2008
Interference Management in MIMO Networks
Approved by:
Professor Mary Ann Ingram, Advisor
School of Electrical and Computer
Engineering
Georgia Institute of Technology
Professor Raghupathy Sivakumar
School of Electrical and Computer
Engineering
Georgia Institute of Technology
Professor Georey Li
School of Electrical and Computer
Engineering
Georgia Institute of Technology
Professor Gregory Durgin
School of Electrical and Computer
Engineering
Georgia Institute of Technology
Professor Prasad Tetali
School of Mathematics
Georgia Institute of Technology
Date Approved: April 2008
To the memory of my mother, Rama Gaur
ACKNOWLEDGEMENTS
I would like to express my deep gratitude to Prof. Mary Ann Ingram for her invaluable
advice and encouragement at every step of my PhD program. Without her unfailing
support, patience, and belief in me, this thesis would not have been possible. Her
contribution to this thesis goes well beyond her role as an academic supervisor and
includes constant support on a personal level without which this journey may never
have been completed. And for this, I am truly grateful.
I also thank the members of my thesis committee, Prof. Raghupathy Sivakumar
and Prof. Georey Li for being on my thesis reading committee. Their encouragement
and enlightening suggestions have greatly improved my research and this dissertation.
I express my appreciation to Prof. Gregory Durgin and Prof. Prasad Tetali for being
on my dissertation committee.
A big part of my PhD learning experience has been the interaction and collab-
oration with fellow students at Georgia Tech. I have had wonderful lab-mates at
Smart Antenna Research Lab. In particular my discussions with Hemabh Shekhar,
Vikram Anreddy, Jeng-Shiann Jiang, and Vijay Ganugapati have benetted me a lot.
I also thank great friends and former colleagues Shantidev Mohanty, Chirag Patel,
Ghurumuruhan Ganesan, Arnab Choudhury, Rajesh Luharuka, and Manas Bajaj, for
their everlasting friendship and support. In addition, I thank all my friends outside
Georgia Tech including Jasvinder Singh, Gaurav Sinha, Bhupinder Sooch, Kamesh-
war Chandrasekar, Ramanathan Viswanathan, and so many others for making my
life in US a pleasant experience.
And foremost, I oer my heartfelt thanks to my parents and my brothers Himan-
shu, Nalinaksh, and Abhishek, for their everlasting encouragement, faith, support
iv
and love. Thank you for everything. To you, I dedicate this work.
v
TABLE OF CONTENTS
DEDICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . iv
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
SUMMARY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
I INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation and Challenges . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
II BACKGROUND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 MIMO Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Channel Capacity . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Transmission Strategies . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Orthogonal Space-Time Block Coding . . . . . . . . . . . . . 11
2.2.2 Spatial-Multiplexing . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Optimization of Interference-Limited MIMO Links . . . . . . . . . . 14
2.3.1 Isolated Links . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.2 Co-channel Links . . . . . . . . . . . . . . . . . . . . . . . . 15
III STREAM CONTROL WITH ANTENNA SELECTION . . . . . 17
3.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Capacity-Optimal Stream Control . . . . . . . . . . . . . . . . . . . 20
3.2.1 CL-SDMA System . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2.2 OL-SDMA System with Optimal Antenna Selection . . . . . 20
3.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3.1 Throughput Performance . . . . . . . . . . . . . . . . . . . . 21
vi
3.3.2 Number of Streams and Stream Control . . . . . . . . . . . . 24
3.3.3 Eect of Path-Loss Exponent . . . . . . . . . . . . . . . . . . 25
3.4 Performance Over the Measured Channel . . . . . . . . . . . . . . . 25
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
IV STREAM CONTROL FOR FINITE COMPLEXITY RECEIVERS 29
4.1 System Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.1 Linear MMSE Decoder . . . . . . . . . . . . . . . . . . . . . 30
4.1.2 Rate Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2.1 Throughput Performance . . . . . . . . . . . . . . . . . . . . 33
4.2.2 Number of Streams and Stream Control . . . . . . . . . . . . 35
4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
V MSE OPTIMAL ANTENNA SELECTION . . . . . . . . . . . . . 38
5.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2 Antenna Selection Methodology . . . . . . . . . . . . . . . . . . . . 40
5.3 Transmit Antenna Selection . . . . . . . . . . . . . . . . . . . . . . . 41
5.3.1 No Interference . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.3.2 Co-channel Interference . . . . . . . . . . . . . . . . . . . . . 45
5.4 Receive Antenna Selection . . . . . . . . . . . . . . . . . . . . . . . 47
5.4.1 No Interference . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.4.2 Co-channel Interference . . . . . . . . . . . . . . . . . . . . . 49
5.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
VI MULTIUSER DETECTION FOR STBC USERS . . . . . . . . . . 56
6.1 Real STBC and IC of Two Co-channel Users . . . . . . . . . . . . . 57
6.2 IC with Rate-1/2 Complex OSTBCs . . . . . . . . . . . . . . . . . . 60
6.3 Simulation Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
vii
VII IC OF ALAMOUTI INTERFERENCE FOR A V-BLAST USER 65
7.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.2 IC for a SIMO Link . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
7.2.1 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 68
7.3 IC for a V-BLAST Link . . . . . . . . . . . . . . . . . . . . . . . . . 70
7.3.1 Sub-optimal IC . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.3.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 74
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
VIIISUMMARY AND FUTURE WORK . . . . . . . . . . . . . . . . . 77
8.1 Research Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
8.2 Suggestions for Future Work . . . . . . . . . . . . . . . . . . . . . . 78
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
VITA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
viii
LIST OF FIGURES
Figure 1 Block diagram of MIMO communication system . . . . . . . . . . . 8
Figure 2 Capacity of SISO and (N
r
, N
t
) MIMO systems as a function of SNR,
averaged over Rayleigh-faded channels . . . . . . . . . . . . . . . . 10
Figure 3 Alamouti space-time encoder [1] . . . . . . . . . . . . . . . . . . . . 12
Figure 4 V-BLAST transmitter architecture [52] . . . . . . . . . . . . . . . . 13
Figure 5 Mutually dependent link capacities and transmission strategies in a
SDMA network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Figure 6 A simple 2-link network with spatial multiplexing transmissions. . . 19
Figure 7 Average improvement in the network throughput relative to closed-
loop TDMA, (T T
TDMA
/T
TDMA
100%, fair energy approach. SC
stands for stream control and OS stands for optimal antenna selection. 22
Figure 8 Histograms of number of streams used by one link with dierent
MIMO congurations for dierent nlog(R/D) values. Each layer
of bars is associated with a dierent number of streams used, as
indicated on the y-axis. . . . . . . . . . . . . . . . . . . . . . . . . 23
Figure 9 Layout of Residential Laboratory [28] . . . . . . . . . . . . . . . . . 26
Figure 10 Average throughputs of various SDMA schemes for various net-
work congurations. Equal-SNR normalization is assumed with SNR
corresponding to 20dB noise-normalized total transmit power for
TDMA and 17dB for schemes with interference . . . . . . . . . . . 27
Figure 11 Average improvement in the network throughput relative to CL-
TDMA for Average BER = 10
2
. SC stands for stream control and
OS stands for optimal antenna selection. The subscript ()* indicates
that whitened channel information is available at the transmitter. . 32
Figure 12 Average improvement in the network throughput relative to CL-
TDMA for Average BER = 10
5
. The legend is same as in Figure
11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Figure 13 Achievable bit rates for various MIMO schemes as a function of
target BER for SIR = 0dB. The subscript ()* indicates that whitened
channel information is available at the transmitter. . . . . . . . . . 35
Figure 14 Number of streams used by one link with dierent MIMO congu-
rations for dierent SIR values. Target BER = 10
5
. . . . . . . . . 36
Figure 15 MMSE error performance for (6,6) MIMO system with varying num-
ber active transmit antennas and SNR = 5dB. . . . . . . . . . . . 50
ix
Figure 16 Error performance of (4,4) MIMO system with ZF receiver in the
presence of transmit correlation. Three transmit antennas are selected. 51
Figure 17 MMSE error rate for (6,6) MIMO link in presence of co-channel in-
terference from 3 other streams. Three transmit antennas are selected. 53
Figure 18 MMSE error rate for (6,4) MIMO system. Four receive antennas are
selected. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Figure 19 BER performance of MMSE receiver for (6,2) MIMO system in the
presence of 2 interfering streams with SIR = 0dB. Four receive an-
tennas are selected. . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Figure 20 Average BER performance of User-A . . . . . . . . . . . . . . . . . 63
Figure 21 The cumulative density function of normalized SNR for various in-
terference scenarios with SIR = 0dB. . . . . . . . . . . . . . . . . . 69
Figure 22 Average bit error probability as a function of SNR with SIR = 0dB. 70
Figure 23 Average bit error probability for 2 4 V-BLAST link in presence of
Alamouti interferer with SIR = 0dB. . . . . . . . . . . . . . . . . . 74
Figure 24 Average bit error probability with varying mean squared deviation
(MSD) in channel during consecutive timeslots. . . . . . . . . . . . 75
x
ABBREVIATIONS
ABER

= Average Bit Error Rate
AWGN

= Additive White Gaussian Noise
BER

= Bit-Error Rate
CDF

= Cumulative Distribution Function
CL

= Closed-Loop
CL-MIMO

= Closed-Loop MIMO
CSI

= Channel State Information
DOFs

= Degrees Of Freedom
FEC

= Forward Error Correction
IID

= Independent and Identically Distributed
INR

= Interference-to-Noise Ratio
LOS

= Line-of-sight
MAC

= Multiple-Access Control
MIMO

= Multiple-Input Multiple-Output
MISO

= Multiple-Input Single-Output
MMSE

= Minimum Mean-Square Error
MSE

= Mean Square Error
MUD

= Multiuser Detection
NBS

= Norm Based Selection
NLOS

= Non-line-of-sight
OL

= Open-Loop
OL-MIMO

= Open-Loop MIMO
OSTBC

= Orthogonal Space Time Block Code
PDF

= Probability Density Function
RF

= Radio Frequency
xi
SDMA

= Space Division Multiple Access
SIC

= Successive Interference Cancellation
SISO

= Single-Input Single-Output
SIMO

= Single-Input Multiple-Output
SIR

= Signal-to-Interference Ratio
SM

= Spatial Multiplexing
SNR

= Signal-to-Noise Ratio
STBC

= Space Time Block Code
SVD

= Singular-Value Decomposition
TDMA

= Time-Division Multiple-Access
V-BLAST

= Vertical - Bell Labs Layered Space-Time
WLAN

= Wireless Local Area Network
ZF

= Zero-Forcing
xii
SUMMARY
New interests in wireless communications driven by consumer electronics have
raised the bar in terms of throughput for wireless networks. Towards this goal, the
multiple-input multiple-output (MIMO) systems have emerged as a key technology
capable of delivering extremely high data rates. Extensive research in MIMO tech-
nology has led to its inclusion in several wireless standards including IEEE 802.11n
(WLAN) and IEEE 802.16e/m (WiMax). The objective of the research presented
in this dissertation is to develop ecient low complexity interference management
techniques for improving the performance of MIMO networks.
The rst half of this dissertation focuses on the interference problems arising in
the context of space-division multiple access (SDMA) based MIMO networks, which
more eectively exploit the available network resources than time-division multiple
access (TDMA) MIMO networks. Sub-optimal techniques for joint optimization of
co-channel MIMO links are considered by applying optimal antenna selection-aided
stream control algorithms. These algorithms are tested on both simulated and mea-
sured indoor MIMO channels. It is shown that the use of the SDMA scheme along
with partial channel state information (CSI) at the transmitters signicantly reduces
the signaling overhead with minimal loss in throughput performance. Next, a mean
squared error (MSE) based antenna selection framework is presented for developing
low complexity algorithms for nite complexity receivers operating in the presence of
co-channel interference. These selection algorithms are shown to provide reasonable
bit-error rate (BER) performance while keeping the overall system complexity low.
The later half of this dissertation considers interference problems for space-time
encoded transmissions. Despite the low data rates supported by various Orthogonal
xiii
Space-Time Block Codes (OSTBCs), they are attractive from a network point of
view as they cause correlated interference, which can be mitigated using only one
additional antenna without sacricing space-time diversity gains. These algebraic
properties of linear OSTBCs are exploited to facilitate a single-stage and minimum
MSE (MMSE) optimal detector for two co-channel users employing unity rate real
and derived rate-1/2 complex OSTBCs for 3 and 4 transmit antennas. Furthermore,
a single-stage multi-user interference suppression technique is proposed for OSTBC
interference that exploits the temporal and spatial structure of OSTBCs leading to
simple linear processing. Next, a sub-optimal space-time IC technique is developed
for a V-BLAST link subjected to Alamouti interference. The proposed IC technique
outperforms conventional IC techniques that do not take the structure of Alamouti
interference into account.
This research deals with the challenges posed by co-channel interference in MIMO
networks and provides practical means of managing interference at a marginal cost of
system performance. The physical layer algorithms provide low complexity solutions
for interference management in SDMA MIMO networks, which can be used for the
design of next generation multiple access control (MAC) layers.
xiv
CHAPTER I
INTRODUCTION
The eld of wireless communications has witnessed revolutionary developments in
the past decade. Tremendous research in this area has made it more realistic for
future generation wireless networks to match the data-rates of wired networks. The
key driving force behind these developments is the multiple-input multiple-output
(MIMO) technology, which has rapidly emerged as a reliable means of supporting
extremely high data rates over wireless channels. As a result, MIMO has been adopted
as the key PHY layer technology for the upcoming WLAN standard, IEEE 802.11n,
which is expected to oer 600 Mbps PHY rate. In addition, MIMO has also been
adopted in several other wireless standards including the WiMAX standard, IEEE
802.16, and next generation cellular networks such as UMTS.
Dierent from conventional links with single antenna transceivers, MIMO links
employ multiple antennas at both ends. These antennas can be used to create mul-
tiple spatial channels in the same bandwidth by partitioning a high signal-to-noise
ratio (SNR) channel into many low-SNR subchannels. Thus, a MIMO link can carry
multiple data streams in parallel on the same frequency band, leading to increased
spectral eciency. As a result, in a rich scattering environment the capacity of a
MIMO link scales linearly with the number of transmit and receive antennas [18, 49].
For this reason, MIMO transceivers are an obvious choice for next-generation wireless
networks, including WLANs. Apart from oering extremely high spectral eciencies,
MIMO links also oer an attractive diversity/rate trade-o. The additional degrees of
freedom (DOF) due to multiple antennas can be used to suppress interference and/or
lend diversity gains to protect data streams against transmission errors [56]. This
1
exibility due to multiple antennas, enables MIMO networks to tolerate co-channel
transmissions leading to better resource utilization than a time-division multiple ac-
cess (TDMA) scheme [14].
1.1 Motivation and Challenges
The multiple-access control (MAC) protocol used in current WLAN standards, includ-
ing the developing IEEE 802.11n standard with enhancements for higher throughput,
is based on the carrier sense multiple access/collision avoidance (CSMA/CA) proto-
col. As a result, simultaneous transmissions from two or more neighboring nodes that
might cause interference to each other are not allowed. This leads to sub-optimal
performance for networks comprising MIMO capable nodes [14].
With these networks as our motivating application, we develop sub-optimal phys-
ical layer techniques that allow interfering MIMO links to operate simultaneously
and provide a reasonable performance/complexity tradeo. These algorithms can be
viewed as resource allocation methods which might be used in next generation MAC
layer designs to improve network performance. In this dissertation, we focus on the
following techniques:
Stream control with partial channel state information (CSI): We pro-
pose an ecient stream control aided by optimal antenna selection for jointly
optimizing the network throughput for open-loop MIMO (OL-MIMO) systems.
The partial CSI at the transmitter is used to convey the subset of selected trans-
mit antennas. Next, the usefulness of this scheme is shown in the context of
nite complexity linear receivers.
Mean-squared error (MSE) optimal antenna selection: Several greedy
algorithms are developed to improve the error performance of linear receivers in
presence of co-channel interference. These algorithms aim to minimize the MSE
2
associated with the linear receivers and provide a good performance/complexity
trade-o.
Suppression of Space-time interference: Exploiting the rich algebraic
structure of orthogonal space-time block codes (OSTBC) interference, we show
computationally ecient ways to detect two co-channel users employing the
same rate-1/2 OSTBC codes. Furthermore, we present linear interference can-
cellation (IC) techniques for suppressing space-time interference.
An optimal transmission scheme for OL-MIMO systems in an interference-free
zone is to put independent data streams with equal power into dierent antennas
[18, 49]. However, in an interference-limited environment this may not be the best
strategy as it is likely overload the receiver with more streams than available antenna
elements. In such situations, network throughput may still be improved by MAC
layer regulation of the co-channel transmissions.
A distributed stream control mechanism was proposed for space-division multi-
ple access (SDMA) networks, which greatly improves the overall network throughput
compared to a TDMA network [13, 14]. This stream control mechanism works best
with the closed-loop MIMO (CL-MIMO) systems but the required overhead is signif-
icant, as the CSI for each pair of transmit-receive nodes has to be signaled back to
the transmitters. In addition, the implementation of stream control for CL-SDMA is
numerically intensive, involving matrix decompositions, thus real-time implementa-
tion would become a challenge with growing size of channel matrices. On the other
hand, stream control for OL-SDMA has signicantly less complexity but it performs
signicantly worse relative to CL-SDMA when the interference is strong [13, 14].
In this dissertation, we propose an ecient stream control strategy assisted by op-
timal antenna selection for jointly optimizing the network throughput for OL-MIMO
where only limited CSI is available at the transmitter. We show that a middle-path
3
approach of having a limited-feedback channel (used to convey the optimal subset of
selected transmit antennas) provides a good trade-o between the feedback signal-
ing load and the network throughput performance. Our results, for both simulated
and measured channels, show that the performance gap between CL- and OL-SDMA
with limited feedback can be substantially abridged if optimal antenna selection is
combined with stream control.
The main challenge in employing stream control with optimal antenna selection for
OL-SDMA systems is the search of optimal subset of transmit antennas. A straight-
forward approach is to evaluate the cost function (e.g., capacity, bit error rate) over all
possible antenna subsets. However, it quickly becomes computationally prohibitive
with increasing number of available antenna combinations. For instance, choosing
8 antennas out of 16 available antennas requires 12,870 computations of the cost
function to determine the optimal antenna subset.
Various suboptimal antenna selection schemes for MIMO systems have been stud-
ied in the recent literature. The most simple one is norm-based selection (NBS),
which selects the receive (transmit) antennas corresponding to the rows (columns)
of the channel matrix with the largest Euclidean norms [43, 19]. Indeed, the main
drawback of NBS is that it may lead to much lower capacity in the presence of receive
(transmit) correlation [19]. A better approach is to remove the antennas in a step-
by-step method, which causes minimal degradation in the capacity [36]. However,
this method still involves the complexity of matrix inversion, which is avoided by
performing incremental antenna selection starting with a null set of selected antennas
[22, 19].
Most of the above studies assume channel capacity as the cost function instead
of the more practical mean-squared error (MSE) or bit-error rate (BER) metrics,
which are dependent on the receiver complexity. In this dissertation, we develop an
MSE-based antenna selection framework for both minimum MSE (MMSE) and zero
4
forcing (ZF) receivers assuming presence of co-channel interference. For either the
transmitter or receiver, there are two sequential greedy algorithms, one that starts
with a full set of antennas and decrements, and the other that starts with an empty
set that increments. The choice of which is best depends on how many antennas
will ultimately be selected. These presented algorithms have low implementation
complexity while oering near optimal error performance.
Another part of this dissertation is devoted to space-time interference cancella-
tion techniques and multi-user detection of co-channel users employing Orthogonal
Space-Time Block codes (OSTBCs). Despite the low data rates supported by various
OSTBCs, they are attractive from a network point of view as they cause correlated
interference, which can be mitigated using only one additional antenna without sacri-
cing space-time diversity gains [37], [26]. A simple suboptimal 2-stage linear receive
processor can achieve IC of two co-channel users employing any rate-1/2 complex
STBC based on an orthogonal design while preserving the diversity gains [45].
We exploit special algebraic properties of linear OSTBCs to facilitate a single-
stage and MMSE-optimal detector for two co-channel users employing unity rate real
and derived rate-1/2 complex OSTBCs for 3 and 4 transmit antennas. Furthermore,
we propose a single-stage multi-user interference suppression technique for OSTBC
interference that exploits the temporal and spatial structure of OSTBCs leading to
simple linear processing. The algorithm requires only one additional antenna to cancel
co-channel OSTBC interference.
1.2 Dissertation Outline
The rest of the dissertation is organized as follows: Chapter 2 presents an overview of
MIMO technology and related concepts followed by a brief discussion of the stream
control concept for joint optimization of co-channel links. Chapter 3 proposes stream
control aided with optimal selection for OL-MIMO systems in the context of SDMA
5
networks. In Chapter 4, the stream control concept is extended to include nite com-
plexity linear receivers. Chapter 5 presents MSE optimal antenna selection framework
for ZF/MMSE receivers. These algorithms can be applied to interference limited en-
vironments such as MIMO ad hoc networks to assist stream control algorithms. In
Chapter 6, we turn our attention to simplied IC methods for OSTBC users employ-
ing rate-1/2 codes and present linear IC techniques for two co-channel OSTBC users.
Chapter 7 considers the problem of Alamouti interference for a co-channel V-BLAST
link and develops sub-optimal space-time IC techniques. Finally, in Chapter 8, we
conclude by summarizing our research contributions and open areas of research in
interference management in MIMO networks.
6
CHAPTER II
BACKGROUND
In this chapter, we provide an overview of MIMO communication systems along with
a brief discussion on their use in ad hoc networks and related challenges. The rst
section introduces some key concepts of MIMO systems. The modeling of a MIMO
link is considered along with brief discussions on the channel capacities of open-
loop (OL) and closed-loop (CL) MIMO systems. The next section presents various
transmissions techniques associated with MIMO links. In particular, system designs
for Spatial Multiplexing (SM) and Space-Time Block Codes (STBC) are considered.
The last section discusses the problem of interference in MIMO networks and presents
stream control algorithms for the joint optimization of co-channel MIMO links.
2.1 MIMO Architecture
A MIMO system consists of multiple antennas at both ends of the link. The use of
multiple antennas oers it N
r
N
t
degrees of freedom (DOFs), where N
r
and N
t
denote receive and transmit antennas, respectively. Unlike single antenna transceiver
systems, which are limited to a single degree of freedom, MIMO systems can exploit
multiple DOFs to alter various aspects of the underlying communication link such as
channel capacity, bit-error rate (BER), coverage, interference suppression etc. More
specically, at the receiver end, these DOFs can be used to provide power gain of
N
r
over white noise, null as many as N
r
1 interfering streams, or receive up to N
r
parallel data streams. Similarly, at the transmitter end these DOFs can be used to
provide transmit diversity, high power gain by concentrating the transmission power
in beams, to avoid transmitting interference to certain directions, or to transmit N
t
streams in parallel.
7
x
1
x
2
x
Nt
Transmitter Receiver
r
1
r
2
r
Nr
Figure 1: Block diagram of MIMO communication system
2.1.1 System Model
A generic MIMO communication system model is shown in Fig. 1. During each sym-
bol period, N
t
symbols are multiplexed for transmission over N
t
transmit antennas.
At the receiver, each antenna receives a linear combination of transmit symbols so
that received baseband vector r is represented as
r = Hx +n (1)
where Hdenotes the N
r
N
t
channel gain matrix whose (i, j)
th
element corresponds to
the complex channel gain between i
th
receiving antenna and j
th
transmitting antenna,
x is the transmit vector and n is the additive white Gaussian noise (AWGN) with
variance N
o
/2 per real and imaginary dimension. In this dissertation, the channel is
assumed to be slowly varying, Rayleigh faded, and xed for the duration of an entire
burst. It is further assumed that
E(xx
H
) = P; E(nn
H
) = N
o
I; E(xn
H
) = 0 (2)
where P is known as power allocation matrix, the subscript H denotes the conjugate
transpose and E(.) denotes the expected value of its argument.
At the receiver, the transmit vector can be estimated using well known linear
8
detectors such as ZF, or MMSE. In addition to ZF/MMSE solutions, more com-
plex non-linear techniques such as successive interference cancellation (SIC) can be
employed at the receiver to boost the BER performance.
2.1.2 Channel Capacity
A MIMO link is capable of creating multiple spatial channels to increase the data
rate while maintaining reliable data detection at the receiver. Shannons channel
capacity, dened as the maximum rate of information that can be sent essentially
error free across the channel, provides a useful measure of the data-carrying capability
of MIMO systems. In an interference free environment, the channel capacity of the
MIMO system can be expressed as
C = max
P
log
2

I +
1
N
o
HPH
H

. (3)
It is apparent that the channel capacity is dependent on the power allocation scheme
at the transmitter and can be maximized if the transmitter is made aware of the
underlying channel or its statistics. To achieve this, the input symbols at the trans-
mitter are passed through a linear precoder, which is optimized for given channel
information available at the transmitter. Thus the received baseband vector after
precoding can be represented as
r = HFx +n (4)
where F is the precoder matrix with complex elements. When the channel gain matrix
is available at the receiver but not at the transmitter, as in OL-MIMO systems,
the optimal precoder matrix F =
_
P
T
/N
t
I, where P
T
denotes the total available
transmit power. This leads to the best channel capacity for OL-MIMO systems given
by [49, 18]:
C =
K

k=1
log
2
_
1 +
P
T
N
t

2
k
_
(5)
9
0 5 10 15 20
2
4
6
8
10
12
14
16
18
20
22
SNR (dB)
A
v
e
r
a
g
e

C
a
p
a
c
i
t
y

(
b
i
t
s
/
s
/
H
z
)
(4,4)
(2,2)
(1,1)
SISO
OLMIMO
CLMIMO
Figure 2: Capacity of SISO and (N
r
, N
t
) MIMO systems as a function of SNR,
averaged over Rayleigh-faded channels
where
k
, k = 1, ...K denote the nonzero singular values (in decreasing order), which
are obtained via the Singular Value Decomposition (SVD) of the channel matrix H.
In CL-MIMO systems, the channel gain matrix is available both at the receiver
and at the transmitter, thus allowing the link to decompose the channel into a logical
collection of uncoupled parallel channels. In this case, the optimal precoder matrix
can be obtained as F = V
F
. Here
F
is a diagonal matrix whose k
th
non-zero
diagonal element is obtained using classical water-lling approach as [42]:

k
=
__

1

2
k
_
+
_1
2
(6)
where (.)
+
indicates that only non-negative values are acceptable, and is chosen
such that

2
k
= P
T
. Thus, the expression for the capacity for CL-MIMO systems
becomes [49]
C =
K

k=1
log
2
(1 +
2
k

2
k
). (7)
Next, in Figure 2 we compare the average capacities for single-input single-output
10
(SISO) and MIMO systems, where the average is taken over independent and iden-
tically distributed (i.i.d) Rayleigh fading channels. It is evident that OL/CL-MIMO
systems oer far greater capacity than SISO over a wide range of SNR. Also, the
capacity of MIMO system increases with increasing number of antennas. In fact,
it is well known that a rich scattering environment leads to a linear increase in ca-
pacity with min(N
r
, N
t
) [18]. It is also apparent that the capacity gains oered by
CL-MIMO systems over OL-MIMO systems are more prominent at low SNRs and
diminishes rapidly as SNR improves [49, 18].
2.2 Transmission Strategies
Unlike SISO links, multiple antennas at the transmitter and receiver allow MIMO
links to adapt transmissions to suite the link requirements such as increased spectral
eciency, transmit/receive diversity gains to reduce BER, improved range etc. This
can be achieved by coding the signals across space and/or time using well known
techniques such as Space-Time Coding, Spatial Multiplexing or Transmit Beamform-
ing [1, 48, 52]. In this dissertation, we focus on two of the most common MIMO
transmission strategies namely, Orthogonal Space-Time Block Codes (OSTBC) and
Spatial Multiplexing (SM).
2.2.1 Orthogonal Space-Time Block Coding
Space-Time Block Codes (STBCs) are well known to provide transmit diversity with-
out requiring explicit channel feedback from the receiver [1, 48]. The temporal and
spatial structure of certain STBCs, called orthogonal STBCs (OSTBCs), oer addi-
tional advantage of maximum likelihood detection based only on linear processing at
the receiver [48]. As a result, the OSTBCs are being widely adopted in various wire-
less communication standards such as the 3GPP cellular standard, WiMAX (IEEE
802.16), and IEEE 802.11n. An example of an OSTBC is the popular Alamouti code
[1]
11
symbols in
... x

2
, x
1
...x

1
, x
2
...x
3
, x
2
, x
1 Encoder
Alamouti
Figure 3: Alamouti space-time encoder [1]
X =
_
_
_
x
1
x

2
x
2
x

1
_
_
_
(8)
As shown in Figure 3, Alamoutis code encodes the input sequence of symbols
across time and space. The Alamouti code falls into a more general category of linear
STBC codes. These codes have relatively simple structure, as the transmitted code
matrix, X, is linear in the real and imaginary parts of the data symbols [32]:
X =
Ns

i=1
x
i
A
i
(9)
where {x
i
}
Ns
i=1
denotes the transmitted symbols and {A
i
}
Ns
i=1
are the N
t
N matrices
representing the code structure, where N is the duration over which N
s
symbols are
transmitted using N
t
antennas. From the denition of real orthogonal codes, it follows
that
A
i
A
T
j
=
_

_
I
N
if i = j
A
j
A
T
i
otherwise
(10)
For a MIMO link with N
t
transmit antennas and N
r
receive antennas, the received
space-time signal Y corresponding to X can be written as [32]:
Y = HX+V (11)
where H is the channel matrix and V denotes the complex additive white Gaussian
noise matrix with zero mean and unity variance. For a quasi-static at Rayleigh fading
12
symbols in
x
Nt
Converter
..., x
Nt
, ..., x
2
, x
1
x
2
x
1
Parallel
Serial to
Figure 4: V-BLAST transmitter architecture [52]
channel, the received baseband signal can be represented using vector notation as [32]:
y = vec(Y) = Fx +v (12)
where vec(.) denotes the standard vector representation of its argument, v = vec(V),
and F is dened as
F =
_
vec(HA
1
) . . . vec(HA
Ns
)
_
(13)
The vector representation for the received baseband signal in (12) can be expressed
dierently as y

= F

x + v

, where the operator (.)

is dened as

=
_
_
_
Re()
Im()
_
_
_
.
Therefore, the problem of detecting the transmitted data x given y amounts to min-
imizing the metric ||y

x||
2
. This can be solved using well known linear detection
techniques such as MMSE or ZF ltering.
2.2.2 Spatial-Multiplexing
Dierent from OSTBC transmission techniques, Spatial Multiplexing schemes intend
to increase the spectral eciency of MIMO systems. Several SM-MIMO schemes have
been developed including the most popular Vertical Bell Laboratories Layered Space
Time (V-BLAST) architecture illustrated in Figure 3 [52]. In this dissertation, we will
13
focus on V-BLAST transmission for SM-MIMO systems, which requires transmission
of N
t
input symbols simultaneously via the N
t
transmit antennas resulting in a much
higher data rate, N
s
= N
t
. However, unlike OSTBC, the associated symbol detection
using linear MMSE/ZF techniques is rather complex.
2.3 Optimization of Interference-Limited MIMO
Links
As mentioned in the previous sections, MIMO links are better adept at suppressing
interference compared to SISO links. In the following, we present a brief overview of
capacity optimization techniques for interference limited MIMO networks.
2.3.1 Isolated Links
Many studies on MIMO systems consider a single point-to-point link or MIMO net-
works, which require links to access medium in a time-division fashion [18, 49, 10, 38].
Such communication systems avoid introducing co-channel interference. However,
even then the links may still suer from co-channel interference arising from other
networks or jammers. For such networks, overall capacity optimization amounts to
maximizing individual link capacities.
In the presence of strong interference, it is important not to overload the receivers.
The optimal strategy for OL-MIMO systems is to excite as many as N
r
N
int
trans-
mit antennas, where N
int
denotes the number of incident interfering streams on the
intended receiver [51]. The precoder matrix in this case is:
F
i
=
_
P
T
N
r
N
int
diag
_
1, 1, ..., 1
. .
NrN
int
, 0, 0, ..., 0
. .
NtNr+N
int
_
(14)
where diag(.) denotes the diagonal matrix formed by the elements of its vector argu-
ment.
For CL-MIMO systems, the water-lling approach in (6)-(7) can be modied to
accommodate xed non-white interference at the receiver of a link (represented by a
14
1
3
2
4
C
12
= f(H
12
, P
1
, R
32
)
R
1
4
=
g
(
H
1
4
,
P
1
) R
3
2
=
g
(
H
3
2
,
P
3
)
C
34
= f(H
34
, P
3
, R
14
)
Figure 5: Mutually dependent link capacities and transmission strategies in a SDMA
network.
noise-normalized covariance matrix, R) by whitening the channel matrix rst. Ap-
plying a spatial whitening transform to the channel yields

H = [I +R]
1/2
H (15)
which reduces the capacity relation to the simple form in (3), with a substitution of
H

H [5]. Thus, the capacity formula becomes

C = max
P
log
2

I +

HP

H
H

. (16)
The whitening operation assumes that the interfering symbols are unknown and only
exploits the spatial characteristics of the interference. With the whitening trans-
formation in (15)-(16), the optimum precoder becomes a function of the received
interference, R, in addition to the channel matrix, H.
2.3.2 Co-channel Links
MIMO networks based on space-division multiple access (SDMA), which allow con-
trolled co-channel transmissions, can outperform TDMA type networks [13, 14, 47].
Since these networks have multiple interfering links, the power allocation scheme, P
i
,
used at each transmit node aects the interference correlation matrices, R
ij
, seen
15
by the receiver nodes. Therefore, optimization of individual link capacities does not
necessarily translate to optimal network capacity as the link capacities have mutual
dependence because the whitened channel matrix for a given link is a function of the
interference. And the transmission strategy, in turn, is dependent on the whitened
channel matrix. Thus, a change in the power allocation matrix of one link induces
a change in the optimum power allocation matrix of the other co-channel links as
shown in Figure 5.
As a result, the optimum precoder matrix and the powers of interfering CL-MIMO
links cannot be calculated independently at each link and iterative methods are used
instead [16, 8, 13, 14]. Joint transmit and receive beamforming for interfering SISO
links is studied in [8], where the antenna weights and transmit powers are optimized
to achieve certain SIRs for each link. An iterative stream control algorithm for CL-
MIMO systems is proposed in [13, 14] to maximize the SDMA network throughput.
At each iteration, all transmit-receive pairs optimize their link capacities under mea-
sured interference and allocated transmit powers. Each links transmission strategy
is determined according to the water-lling solution given in (16).
16
CHAPTER III
STREAM CONTROL WITH ANTENNA
SELECTION
In the previous chapter, we highlighted the mutual dependence of optimal power
allocation schemes for interfering MIMO links and the interference covariance matri-
ces. In this chapter, we consider the joint optimization of these interfering MIMO
links in the context of an SDMA network. Dierent from [13], we propose a low
complexity stream control algorithm for OL-SDMA systems that provides a good
tradeo between feedback signaling load and the network throughput performance.
The proposed algorithm augments stream control approach through the use of optimal
antenna selection to jointly optimize the interfering links.
In the past, the joint optimization problem for interfering MIMO links has been
treated by several researchers [16, 17, 7, 11, 12]. For cellular systems, iterative meth-
ods were used to optimize the uplink in [16] and the downlink in [17]. In [11] and
[12], the authors explore ways to control the relative capacities of the interfering
CL-MIMO links. In [11], each link iteratively maximizes the closed-loop capacity
of its whitened channel under power constraints that generally dier among nodes,
and in [12], each link minimizes the interference it makes on its neighbors, subject
to capacity constraints. In [13], Demirkol et al. considered CL-MIMO systems and
proposed a distributed stream control mechanism wherein an additional stream is
added if it leads to an increment in the network throughput. The authors show that
MIMO nodes operating under this strategy improve the overall network throughput
compared to a TDMA protocol, in which MIMO links operate in succession.
Although stream control works best with CL-SDMA, the overhead is signicant,
17
as the CSI for each pair of transmit-receive nodes has to be signaled back to the trans-
mitters. Moreover, implementation of stream control for CL-SDMA is numerically
intensive, involving matrix decompositions, thus real-time implementation would be
a challenge as the channel matrix size grows. On the other hand, stream control for
OL-SDMA has signicantly less complexity. In [13], CL-SDMA was compared with
OL-SDMA, where both used stream control, but the antenna selection in OL-SDMA
was deterministic; in this case, OL-SDMA performed signicantly worse relative to
CL-SDMA when the interference was strong. In this chapter, we extend the analysis
of [13] for OL- SDMA systems and show that a middle-path approach of having a
limited feedback channel (used to convey the set of selected transmit antennas) pro-
vides a trade-o between the feedback signaling load and the network throughput
performance.
In the next section, we outline the system model followed by a brief overview of
optimal power allocation schemes for xed interference. In the next two sections we
discuss the simulation results and compare with the measured data. We show that
for both simulated and measured channels, the performance gap between CL- and
OL-SDMA with limited feedback can be substantially abridged if optimal antenna
selection is used in addition to stream control.
3.1 System Model
Consider a simple network as shown in Figure 6, consisting of two spatial multiplexing
MIMO links where each link is subjected to co-channel interference from the other
link. The average signal-to-interference ratio (SIR) varies linearly as 10 log(R/D)
n
on a logarithmic scale, where n denotes the path-loss exponent, R and D denote the
receiver-transmitter separation for the interfering and the desired link, respectively.
The transmitting nodes are equipped with N
t
antenna elements and receiver nodes
use N
r
antennas. Each transmitter uses a linear precoder to improve the system
18
Data Paths
l
12
l
34
1
4 3
2
Interference Paths
D
R
Figure 6: A simple 2-link network with spatial multiplexing transmissions.
performance. The received baseband vector corresponding to the i
th
link can be
represented as
r
i
= H
ii
F
i
x
i
+H
ji
F
j
x
j
+n (17)
where H
ij
denotes the channel gain matrix corresponding to the transmitter of the j
th
link and receiver of the i
th
link, F
i
is the precoder matrix with complex entries, x
i
is
the transmit vector corresponding to the i
th
link and n is the additive white Gaussian
noise (AWGN), having variance N
o
/2 per real and imaginary dimension. The channel
is assumed to be slowly varying, Rayleigh faded, and xed for the duration of an entire
burst. It is further assumed that
E(nn
H
) = N
o
I; E(x
i
n
H
) = 0 (18)
E(x
i
x
H
j
) =
_

_
I if i = j
0 otherwise
(19)
where the subscript H denotes the conjugate transpose and E(.) denotes the expected
value of its argument.
19
3.2 Capacity-Optimal Stream Control
In this section, we briey discuss the optimal precoder and decoder design with respect
to system throughput for OL- and CL-SDMA systems. Our goal is to maximize the
network throughput, which is dened as the sum of the link data rates.
3.2.1 CL-SDMA System
The traditional water-lling approach can be modied to accommodate xed non-
white interference at the receiver of a link by whitening the channel matrix [5]:

H
ii
= (N
o
I +H
ji
F
j
F
H
j
H
H
ji
)

1
2
H
ii
(20)
Let the SVD of the whitened channel matrix be denoted as

H
ii
= U
i

i
V
H
i
. The
optimal precoder matrix for the i
th
link can be written as F
i
= V
i

F
where
F
is a
diagonal matrix whose k
th
non-zero diagonal element is given by [42]:

k
=
__

1

2
k
_
+
_1
2
(21)
where (.)
+
indicates that only non-negative values are acceptable,
2
k
denotes the
SVD values of the whitened channel, P
T
denotes the available transmit power, and
is chosen such that

2
k
= P
T
. It may be noted that the capacity based precoder
design is not necessarily optimal for nite complexity transceivers as will be explained
in Chapter 4.
3.2.2 OL-SDMA System with Optimal Antenna Selection
For an OL-SDMA system, the best strategy is to allocate equal power to all transmit
antennas [7]. In the presence of strong interference, not all transmit antennas are
used, to avoid overloading the receiver. In this case, the optimal strategy is to excite
as many as N
r
N
int
transmit antennas, where N
int
denotes the number of incident
interfering streams on the intended receiver [51]. The precoder matrix in this case is:
F
i
=
_
P
T
N
r
N
int
diag
_
perm
_
1, 1, ..., 1
. .
NrN
int
, 0, 0, ..., 0
. .
NtNr+N
int
__
(22)
20
where diag(.) denotes the diagonal matrix formed by the elements of its vector argu-
ment and perm(.) denotes the permutation of the elements of the vector argument
based on antenna subset selection.
3.3 Simulation Results
In this section, we present simulation results for 2-link network as shown in Figure 6
and the 3-link network. The results are generated using Monte Carlo simulation of
1000 channel trials. For the CL-SDMA results, the algorithm of [13] was used. In [13],
authors demonstrate the usefulness of stream control for various SDMA techniques
with the exception of OL-SDMA with optimal antenna selection. This section mainly
considers SDMA schemes with stream control and draws performance comparisons
between OL-SDMA performance with deterministic antenna selection and optimal
antenna selection. For more detail about the contents of this section, with the ex-
ception of optimal selection, the reader is referred to [14]. We consider a fair energy
transmission approach, which requires both TDMA and SDMA networks use equal
transmit powers, to allow for a fair performance comparison. The noise-normalized
transmit power is xed at P
T
= 20dB for the TDMA scheme. For SDMA scheme,
the total transmit power is divided equally among all the transmitting nodes, i.e.
= P
T
/2 = 17dB for the 2-link network and = P
T
/3 = 15.2dB for the 3-link network.
3.3.1 Throughput Performance
Figure 7 shows the average percent throughput improvements for several SDMA
schemes relative to TDMA for a 2-link network. The horizontal axis is nlog(R/D),
where n is the path loss exponent. As a reference, the 802.11 MAC is likely to en-
force time multiplexing due to interference if R/D < 2. For n = 3, for example,
R/D < 2 corresponds to 3 log(R/D) < 0.9. Therefore, if an SDMA scheme has
positive throughput improvement for nlog(R/D) < 0.9, then a MAC that exploits
SDMA, such as the one in [47], would outperform the 802.11 MAC.
21
2 1 0 1 2 3 4 5 6 7
0
10
20
30
40
50
60
nlog(R/D)
%
T
h
r
o
u
g
h
p
u
t

I
m
p
r
o
v
e
m
e
n
t
CLSDMA, optimal SC
OLSDMA, SC with OS
OLSDMA, SC without OS
Figure 7: Average improvement in the network throughput relative to closed-loop
TDMA, (T T
TDMA
/T
TDMA
100%, fair energy approach. SC stands for stream
control and OS stands for optimal antenna selection.
From Figure 7, we observe that CL-SDMA with stream control yields the best
performance as expected. In [13], authors highlight the importance of stream control,
which strikingly improves the throughput of OL-SDMA in strong interference regions.
Yet, without optimal selection, stream control is not enough to make SDMA better
than TDMA for nlog(R/D) < 1/2, when the interference is strong. However, when
optimal antenna selection is used, the gap between CL-SDMA and OL-SDMA is
reduced and in fact OL-SDMA outperforms TDMA even in high interference regions,
oering an improvement of about 5%. For nlog(R/D) > 1, the interference from
neighboring nodes is weak enough to allow the use of 4 data streams, thus explaining
similar performances by various SDMA schemes. Thus stream control and optimal
selection are the key factors in throughput performance when the interference is
strong.
We also found that for a 3-link hexagonal network model, the average percent
throughput improvement curves for dierent SDMA schemes follow similar trends.
22
1
2
3
4
1.39
0.76
2.90
5.05
7.19
0
20
40
60
80
nlog(R/D)
C
o
u
n
t
(a) OL-SDMA, SC with deterministic
antenna selection
1
2
3
4
1.39
0.76
2.90
5.05
7.19
0
20
40
60
80
nlog(R/D)
C
o
u
n
t
(b) OL-SDMA, SC with optimal an-
tenna selection
1
2
3
4
1.39
0.76
2.90
5.05
7.19
0
20
40
60
80
nlog(R/D)
C
o
u
n
t
(c) CL-SDMA, SC
Figure 8: Histograms of number of streams used by one link with dierent MIMO
congurations for dierent nlog(R/D) values. Each layer of bars is associated with
a dierent number of streams used, as indicated on the y-axis.
Again CL-SDMA with stream control yields the best performance with an improve-
ment of about 20% over TDMA when the interference is strong. Also, OL-SDMA
scheme with deterministic antenna selection performs poorly when R D. With op-
timal antenna selection combined with stream control, OL-SDMA is able to provide
an improvement of about 8% over TDMA. In the weak interference environment,
all SDMA schemes exhibit superior performances against TDMA scheme with an
improvement of about 122%.
23
3.3.2 Number of Streams and Stream Control
We shall consider the previous 2-link topology, assuming the noise-normalized total
transmit power of each link is set to 17dB, and each node is assumed to have 4 an-
tennas. 100 channel trials are generated, and the link parameters are found using the
stream control method for 20 dierent values of nlog(R/D). Figure 8 demonstrates
the regulation of streams by dierent SDMA schemes as a function of the strength of
interference, which varies with nlog(R/D). We observe that, in accordance with [6],
all the SDMA schemes mostly use 4 streams when interference is weak, thus greatly
improving the throughput of the network. However, the similarity ends when the
interference is signicant (R D). Figure 8(a) shows histograms of the number of
streams used by link l
12
with OL-SDMA with stream control but with deterministic
antenna selection. It is apparent that when interference is strong, the link mostly
uses either one or two streams with about equal probability. Figure 8(b) shows that
with optimal antenna selection, OL-SDMA mostly uses 2 streams when the interfer-
ence is strong. This is because if both the links use single stream, it would leave
the victim receiver with two additional degrees of freedom, thus allowing each link
to add another stream. It is apparent that optimal antenna selection-aided stream
control enables OL-SDMA to exploit spatial multiplexing better than the determin-
istic selection. The transition occurs when nlog(R/D) 0.9 when both the schemes
use mostly three streams. It is interesting to note that after this transition, both
schemes perform almost identically. Finally, Figure 8(c) shows the optimal stream
control that could be achieved by CL-SDMA, the trends being very similar to those
of [13]. Unlike OL-SDMA, with and without optimal antenna selection, CL-SDMA
more often uses three streams when interference is relatively weak. Comparing dif-
ferent schemes, we see that optimal stream control, in consonance with [6] and [51],
eliminates the use of excessive numbers of streams when interference is strong. In par-
ticular, the algorithm penalizes additional streams for nlog(R/D) < 0.76. However,
24
when nlog(R/D) exceeds 0.9, stream control has less inuence.
3.3.3 Eect of Path-Loss Exponent
The average SIR varies as on the logarithmic scale, where n denotes the path-loss
exponent. Therefore, a change in path-loss exponent value reects as a change in
the received SIR at a given (R/D) location. Thus the value of n actually determines
the transition region from strong interference to weak interference. A change in the
value of n shifts this transition point (R/D) without aecting stream control before
or after the transition point.
3.4 Performance Over the Measured Channel
The experiments were conducted with our 3D MIMO measurement system at 5.8 GHz
[28] in the Residential Laboratory (RL) at the Georgia Institute of Technology. As
shown in Figure 9, there are two receive array (Rx) locations and eight transmit ar-
ray (Tx) locations. For each Tx-Rx pair, measurements were sequentially performed
to acquire the channel matrices for antenna spacing of 0.5 , where is the wave-
length. The matrices were measured at 51 frequencies (10MHz separation) to obtain
20 51 = 1020 realizations of (4,4) at fading channel matrices. Four representa-
tive congurations will be considered Conf. I: (T2-R1, T7-R2), Conf. II: (T8-R1,
T6-R2), Conf. III: (T3-R1, T4-R2), and Conf. IV: (T3-R1, T5-R2). The rst two
congurations represent channels with less correlated interference because the direc-
tions of data and interference paths are angularly separated for both links. For the
last two congurations, the data and interference for both links are spatially more
correlated because of the connement of the hallway. In order to clearly illustrate the
correlation between the links, the rst and third congurations are shown in Figure 9
with solid and dashed arrows, respectively. All four links (data and interferences) are
individually normalized such that the signal links have the same SNR, and the SIR
is equal to 1. This approach maintains the angular spread of the multipaths, while
25
T5
T8
T4
R2
T3
T7
T6
R1
T2
T1
Figure 9: Layout of Residential Laboratory [28]
removing the range-dependent eects [28].
Figure 10 illustrates the average throughput performance of dierent MIMO schemes
with stream control for dierent network congurations. As expected, CL-SDMA
with stream control provides the best throughput performance for all four network
congurations. Next we note that OL-SDMA with deterministic antenna selection
yields better performance than OL-TDMA for Confs. I and II, when interference is
less correlated. However, for highly correlated interference, as in Confs. III and IV,
26
Conf. I Conf. II Conf. III Conf. IV
0
5
10
15
20
25
30
T
h
r
o
u
g
h
p
u
t

(
b
p
s
/
H
z
)
OLTDMA
OLSDMA, SC without OS
OLSDMA, SC with OS
CLSDMA, SC
Figure 10: Average throughputs of various SDMA schemes for various network
congurations. Equal-SNR normalization is assumed with SNR corresponding to
20dB noise-normalized total transmit power for TDMA and 17dB for schemes with
interference
the performance of OL-SDMA without optimal selection degrades and is inferior to
OL-TDMA. In fact, the throughput performance for all schemes goes down for Confs
III and IV, when interference is highly correlated. This is because the correlated inter-
ference reduces the rank of the whitened channel matrix (19), thereby decreasing the
system capacity. Hence for highly correlated interference, antenna selection becomes
even more critical. From Figure 10, we observe that with optimal antenna selec-
tion, the performance of OL-SDMA improves signicantly for all four congurations.
However, OL-SDMA with optimal selection gives only comparable performance to
OL-TDMA when the interference is highly correlated. Since OL-SDMA with optimal
selection requires some CSI feedback, and TDMA requires none, TDMA is prefer-
able in the correlated interference environment. In the other environments, however
OL-SDMA with optimal selection oers an attractive alternative to TDMA.
27
3.5 Summary
In this chapter, we have analyzed the performance gains oered by the use of optimal
transmit antenna selection in conjunction with stream control algorithm for OL-
MIMO for both simulated data and measured data. Although CL-MIMO with stream
control oers the best throughput, it has substantial overhead of providing the CSI to
the transmit nodes. Our results for two- and three-link network models indicate that
OL-MIMO with optimal antenna selection is an attractive alternative to CL-MIMO
with stream control as it incurs a minimal overhead of specifying the chosen antenna
set.
28
CHAPTER IV
STREAM CONTROL FOR FINITE
COMPLEXITY RECEIVERS
In the previous chapter, we analyzed the capacity improvements oered by the use
of optimal transmit antenna selection assisted stream control for OL-MIMO systems.
The analysis presented therein along with the related work in [7, 11, 12, 13] is of the-
oretical importance as it evaluates the network performance in terms of Shannons
capacity assuming Gaussian inputs. However, in practice, discrete signalling constel-
lations are used and transceiver complexity is also rather limited. In this chapter,
we extend the analysis presented in the previous chapter along with the work in [13]
to investigate the impact of stream control on the performance of interfering MIMO
links assuming M-QAM constellations and linear MMSE receiver processing.
In the following sections, we rst review the rate adaptation techniques for MIMO
links given M-QAM constellations. Next, we present the simulation results for various
SDMA congurations. Our simulation results indiciate a similar trend in throughput
performances of CL- and OL-SDMA systems, as expected. Also, the use of optimal
antenna selection in conjunction with stream control is shown to be benecial for
OL-SDMA systems with limited feedback.
4.1 System Design
In this section, we briey discuss the optimal decoder design with respect to system
throughputs for OL- and CL-SDMA systems. We again refer to the network shown in
Figure 6, consisting of two SM-MIMO links where each link is subjected to co-channel
interference from the other link. Our goal is to maximize the network throughput,
which is dened as the sum of the link data rates.
29
4.1.1 Linear MMSE Decoder
Without loss of generality, let us consider the link l
12
. To keep the analysis general,
let us drop the link specic subscripts. The symbol estimation error vector associated
with a linear decoder C can be written as e = xCr, where x and r denote the trans-
mit vector and received baseband vectors, respectively. The MMSE decoder requires
that C be chosen such that the mean-squared error (MSE), E(||e||
2
), is minimized.
The MSE can also be expressed as tr(R
e
), where R
e
is the error covariance matrix
and tr(.) denotes the trace of its argument. Using simple matrix manipulations, R
e
can be rewritten as,
R
e
= (CF
H
H
H
R
r
1
)R
r
(CF
H
H
H
R
r
1
)
H
+I F
H
H
H
R
r
1
HF (23)
where R
r
is the receive covariance matrix. It is not dicult to show that the rst term
in (23) has non-negative diagonal entries. The MMSE solution, therefore, minimizes
the contribution of the rst term towards the MSE by choosing C = F
H
H
H
R
r
1
.
Thus, the MSE for the i
th
data stream is given by
mse
i
= (I F
H
H
H
R
r
1
HF)
ii
(24)
where (.)
ii
denotes the i
th
diagonal entry of its matrix argument. Now, the post-
processing signal to interference and noise ratio (SINR) for the i
th
data stream can
be obtained as sinr
i
=
1
mse
i
1, [42]. The data carrying capability of the link depends
on this value of SINR.
4.1.2 Rate Adaptation
In this subsection, we give a brief overview of rate adaptation techniques for interfering
MIMO links. We consider square as well as rectangular M-QAM modulations. The
probability of bit error for the i
th
data stream modulated using I J rectangular
30
QAM can be linked to the SINR in (24) as [9]:
P
b
(I, J) =
1
log
2
(I.J)
_
log
2
I

k=1

I,J
(k) +
log
2
J

l=1

J,I
(l)
_
(25)
where the function
P,Q
can be computed as

P,Q
(n) =
1
P
P(12
n
)1

j=0
(1)
m
_
2
n1
m+ 0.5
_
erfc
_
(2j + 1)

3log
2
(P.Q.sinr)
P
2
+ Q
2
2
_
(26)
where m =
j.2
n1
P
and . returns the greatest integer less than or equal to the argu-
ment. Now, the maximum rate supported by the i
th
data stream with SINR =sinr
i
can be determined as
R
i
= log
2
(I
i
J
i
), (27)
where {I
i
, J
i
} are chosen such that M
i
= I
i
J
i
is the largest constellation, which
meets the target BER requirement. In the stream control algorithm, we use the rate
expression in (27) instead of Shannons capacity to compute the link throughputs.
The maximum possible link rate R
i
is a stair-case function of the associated SINR
and target BER. It is apparent that the MMSE decoder doesnt necessarily optimize
the link throughput as it tries to minimize the sum

1
sinr
i
instead of maximizing the
sum

R
i
.
4.2 Simulation Results
In this section, we present throughput results for the 2-link network shown in Figure 6.
We consider several SDMA schemes with optional features such as stream control,
optimal selection, channel whitening and draw performance comparisons among them
and CL-TDMA scheme. In [13], the authors demonstrated the usefulness of stream
control for various SDMA techniques, and in the previous chapter we further demon-
strated the eectiveness of optimal selection-aided stream control for OL-SDMA; both
31
10 5 0 5 10 15 20
100
80
60
40
20
0
20
40
60
SIR/symbol (dB)
%
T
h
r
o
u
g
h
p
u
t

I
m
p
r
o
v
e
m
e
n
t
CLSDMA, optimal SC*
CLSDMA, optimal SC
CLSDMA, all streams*
CLSDMA, all streams
OLSDMA, SC with OS
OLSDMA, SC without OS
OLSDMA, all streams
Figure 11: Average improvement in the network throughput relative to CL-TDMA
for Average BER = 10
2
. SC stands for stream control and OS stands for optimal
antenna selection. The subscript ()* indicates that whitened channel information is
available at the transmitter.
using channel capacity metric to quantify the network throughputs. In this section,
dierent from the past work, we assume linear MMSE receiver processing and eval-
uate the network throughput performance using the input signals drawn from both
square and rectangular M-QAM constellations.
Each node is assumed to have 4 transmit/receive antennas. The results are gener-
ated using Monte Carlo simulation of 1000 channel trials. For the CL- and OL-SDMA
results, the algorithm of [13] is used with one exception of using rate expression in
(27) for computing link capacities. We consider a fair energy transmission approach,
which requires both TDMA and SDMA networks use equal transmit powers, to allow
for a fair performance comparison. The noise-normalized transmit power is xed at
P
T
= 20dB for the TDMA scheme. For SDMA scheme, the total transmit power is
divided equally among all the transmitting nodes, i.e.

P
T
= P
T
/2 = 17dB for the
2-link network.
32
4.2.1 Throughput Performance
Figure 11 shows the average percent throughput improvements of several SDMA
schemes relative to CL-TDMA as a function of SIR with a target average BER = 10
2
.
It is apparent that CL-SDMA with stream control yields the best performance as ex-
pected. For example, when the interference is strong (SIR < 9dB), CL-SDMA with
stream control using the whitened channel information, oers an improvement of
about 18% over CL-TDMA. This gain can be explained as resulting from multiuser
diversity because joint-optimization with stream control oers 8 channel modes to
choose from rather than just 4 as in CL-TDMA. On the other hand, OL-SDMA with-
out stream control gives extremely poor performance. This can be attributed to the
fact that a linear decoder like MMSE is overwhelmed with interfering streams in the
absence of stream control mechanism. When stream control is used with determinis-
tic antenna selection, the performance of OL-SDMA is signicantly improved but is
still worse than that of CL-TDMA. However, optimal antenna selection-aided stream
control improves the throughput of OL-SDMA by almost 18%, enabling it to match
the performance of CL-TDMA. Thus OL-SDMA systems using a limited amount of
feedback information oers an attractive alternative to CL-TDMA systems, which
have signicant signaling overhead. From Figure 11, we also observe that as the in-
terference becomes weak, the performance of all SDMA schemes, barring OL-SDMA
without stream control, observes signicant improvement over CL-TDMA. We also
note that the signicance of whitened channel information at the transmitter reduces
with reducing interference strength, as expected. It is also interesting to note that,
contrary to the capacity results of [13], there is substantial gap between several SDMA
schemes even in the high SIR zone.
Figure 12 shows the SDMA throughput improvements for a network with which
permits far less packet errors, with average BER = 10
5
. The throughput curves for
various SDMA schemes follow similar trends as in Figure 11. However, it can be seen
33
10 5 0 5 10 15 20
100
80
60
40
20
0
20
40
60
SIR/symbol (dB)
%
T
h
r
o
u
g
h
p
u
t

I
m
p
r
o
v
e
m
e
n
t
Figure 12: Average improvement in the network throughput relative to CL-TDMA
for Average BER = 10
5
. The legend is same as in Figure 11.
that optimal antenna selection and stream control further improve the throughput
performance of various SDMA schemes. On one hand CL-SDMA without stream con-
trol observes a fall in throughput, on the other hand CL-SDMA with stream control
observes a further gain relative to CL-TDMA, which now amounts to about 21%. For
OL-SDMA, the gain due to optimal antenna selection-aided stream control further
increases to about 3% in the high interference zone. It can also be observed that
OL-SDMA systems without stream control always perform worse than CL-TDMA
systems even when interference is negligible.
Figure 13 shows the achievable bit rates for dierent MIMO schemes employing
an MMSE decoder as a function of target BER for xed SIR = 0dB. The performance
curves for CL-SDMA with stream control using whitened channel information and
OL-TDMA form the upper and lower bounds, respectively. The CL-SDMA scheme
with stream control oers a gain of about 2.5 bps/Hz relative to CL-TDMA over a
wide range of target BER. The availability of whitened channel information improves
the performance of stream control, rendering a gain of about 1.25 bps/Hz. It is
34
5 10 15 20 25
10
6
10
5
10
4
10
3
10
2
10
1
Data Rate (bits/s/Hz)
T
a
r
g
e
t

B
E
R
CLTDMA
OLTDMA
CLSDMA, optimal SC*
CLSDMA, optimal SC
OLSDMA, SC with OS
OLSDMA, SC without OS
Figure 13: Achievable bit rates for various MIMO schemes as a function of target
BER for SIR = 0dB. The subscript ()* indicates that whitened channel information
is available at the transmitter.
interesting to note that the relative gap between any two MIMO schemes is almost
independent of the target BER. From Figure 13, we also observe that optimal selection
for OL-SDMA with stream control improves its performance by about 2.5 bps/Hz
relative to deterministic selection. Infact, OL-SDMA with optimal selection, which
requires a limited-feedback channel, matches the performance of CL-TDMA, which
requires full CSI at the transmitter. Thus, it can be concluded that stream control
along with optimal selection achieves a good tradeo between performance required
and feedback overhead.
4.2.2 Number of Streams and Stream Control
Figure 14 illustrates the regulation of streams by dierent SDMA schemes as a func-
tion of SIR, assuming a target BER = 10
5
. A hundred channel trials are generated,
and the link parameters are found using the stream control method. Figure 14(a)
shows the relative frequency of the number of streams used by link l
12
for OL-SDMA
with stream control but deterministic antenna selection. It is apparent that when
35
10 5 0 5 10 15 20
0
10
20
30
40
50
60
70
SIR/symbol (dB)
%
S
e
l
e
c
t
i
o
n
1 Stream
2 Streams
3 Streams
4 Streams
3
2
1
4
2
3
1
(a) OL-SDMA, stream control with deter-
ministic antenna selection.
10 5 0 5 10 15 20
0
10
20
30
40
50
60
70
SIR/symbol (dB)
%
S
e
l
e
c
t
i
o
n
1 Stream
2 Streams
3 Streams
4 Streams
2
1
3
3
2
4
(b) OL-SDMA, stream control with optimal
antenna selection.
10 5 0 5 10 15 20
0
10
20
30
40
50
60
70
SIR/symbol (dB)
%
S
e
l
e
c
t
i
o
n
1 Stream
2 Streams
3 Streams
4 Streams
3
1
3
4
2
2
(c) CL-SDMA, stream control using
whitened channel information.
Figure 14: Number of streams used by one link with dierent MIMO congurations
for dierent SIR values. Target BER = 10
5
interference is strong, SIR < 9dB, the link mostly uses one or two streams with about
equal probability. This leads to receive diversity gain, which is critical in satisfying
the target BER. The use of optimal antenna selection along with stream control re-
sults in higher selection probability for two streams compared to one stream. This
is because with optimal antenna selection, transmit diversity gain is also available
and as the link can aord two streams more often. Thus optimal antenna selection-
aided stream control enables OL-SDMA to exploit spatial multiplexing better than
36
the deterministic selection.
From 14(b), we also note that at high SIR, OL-SDMA with optimal selection-
aided stream control rarely chooses 4 streams, thus resulting in lower throughput
compared to CL-SDMA with stream control, as shown in 12. Finally, Figure 14(c)
shows the optimal stream control that could be achieved by CL-SDMA using whitened
channel information. We observe that unlike OL-SDMA, both with and without
optimal antenna selection, CL-SDMA rarely uses one or two streams when interference
is relatively weak. It is important to note that the link throughput is dependent not
only on the number of streams used, but also on the number of bits carried by each
stream.
4.3 Summary
In this chapter, we have analyzed the performance gains oered by the use of stream
control for interfering MIMO links with linear MMSE receiver processing. Although,
CL-SDMA with stream control using whitened channel information oers the best
throughput, it has substantial overhead of providing the CSI to all the transmit
nodes. However, the use of optimal transmit antenna selection in conjunction with
stream control algorithm for OL-MIMO is an attractive alternative to CL-MIMO as
it incurs a minimal overhead of specifying the chosen antenna set.
37
CHAPTER V
MSE OPTIMAL ANTENNA SELECTION
In Chapters 3 and 4, we demonstrated the importance of optimal antenna selec-
tion in conjunction with stream control for improving the performance of OL-SDMA
networks. Though, one major challenge in employing stream control with optimal
selection is the search of optimal subset of antennas. This chapter presents an opti-
mal MSE-based selection framework for developing computationally ecient antenna
selection algorithms that work particularly well in terms of the BER of SM-MIMO
receivers that use linear processing.
Antenna subset selection problems involve the choice of appropriate cost function
(e.g., capacity, bit error rate) and a methodology to determine the antenna subset
that optimizes the cost function. One of the most straightforward selection approach
is to evaluate the cost function over all possible antenna subsets. However, it quickly
becomes computationally prohibitive with increasing number of combinations. Vari-
ous suboptimal antenna selection schemes for SM-MIMO systems have been studied
in the recent literature [36, 22, 19, 43, 44, 35, 24, 23, 20, 27, 40]. A suboptimal algo-
rithm that does not need to perform an exhaustive search over all possible subsets is
proposed in [36]. Another promising approach for fast antenna selection was proposed
by Gorokhov [22], and was later improved in [19]. In addition to these schemes, the
norm-based selection (NBS) method is another popular algorithm [43, 19]. The NBS
method selects the receive antennas corresponding to the rows of the channel matrix
with the largest Euclidean norms. Indeed, the main drawback of the NBS is that
it may lead to much lower capacity in the presence of receive correlation [19]. Most
of these studies assume channel capacity as the cost function instead of the more
practical bit error rate (BER) metric, which is dependent on the receiver processing.
38
Since the receivers with practical complexity cannot achieve the channel capacity,
the antenna selection methods based on the capacity optimization are bound to be
inecient with respect to the error performance of these receivers.
Antenna selection algorithms to minimize the BER in linear receivers has been
addressed recently [21, 4, 3, 25, 33, 54]. A transmit antenna selection mechanism for
ZF receivers is proposed in [21] to mitigate the eect of transmit antenna correlation.
The proposed selection algorithm uses the transmit correlation matrix to determine
the antenna subset. Since the proposed algorithm does not exploit the information
based on instantaneous channel matrix, it cannot achieve full transmit diversity. In
[4] Berenguer et al. proposed a transmit antenna selection algorithm based on the
geometrical interpretation of the decoding process for ZF receivers. This work was
later extended to lattice-reduction-aided (LRA) ZF receivers in [3]. The authors
also proposed an approximate selection rule for LRA MMSE receivers based on the
minimization of maximum MSE [3, Eq (24)]. A multimode antenna selection approach
for improving the error performance of SM-MIMO is reported in [25]. The proposed
scheme dynamically controls both the number of substreams transmitted and the
mapping of substreams to a subset of transmit antennas, resulting in additional array
gain.
In the following sections, we rst outline the system model for developing antenna
selection framework in the following section. Next, we develop sub-optimal trans-
mit/receive antenna selection algorithms and present simulation results considering
both isolated and interference-limited channels. These algorithms are greedy in na-
ture and attempt to minimize the MSE at each step when an antenna is added (or
removed) from the selected subset.
39
5.1 System Model
We consider a spatial multiplexing system with N
t
transmit antennas and N
r
receive
antennas .The channel is at-fading and slowly time varying. It is unknown at the
transmitter but is known at the receiver. A low bandwidth, zero-delay, error-free
feedback link provides limited channel information from the receiver to the transmit-
ter.
During each symbol time, L
t
input symbols are multiplexed for transmission over
L
t
transmit antennas. The subset of transmit antennas is determined by a selection
algorithm operating at the receiver. Via the feedback path, the receiver sends the
transmitter the optimal subset S S, where S is the set of all possible
_
Nt
Lt
_
subsets
of transmit antennas.
Let H
NtNt
denote the N
r
N
t
channel matrix and let H
S
denote the N
r
L
t
submatrix corresponding to transmit antenna subset S. The corresponding vector of
received complex baseband samples after matched ltering is given by
r = H
S
a +n, (28)
where a denotes the transmitted vector and n is the additive white noise. The noise
components are assumed to be uncorrelated with complex variance
2
n
. The inputs are
chosen from same unit-energy constellation and are uncorrelated so that E[aa
H
] = I,
where E[.] denotes the expected value of its argument.
5.2 Antenna Selection Methodology
Let the receiver employ a linear detector, C, so that the mean-square error (mse) is
given by
mse = E[||Cr a||
2
]
= tr
_
E[(Cr a)(Cr a)
H
]
. .
W
_
, (29)
40
where ||.|| denotes the Euclidean norm of the vector argument and tr(.) denotes the
trace of its argument. Now the matrix W in (29) can also be written as [2]
W = (CH
H
S
R
1
)R
1
(CH
H
S
R
1
)
H
+I H
H
S
R
1
H, (30)
where R = N
o
I + H
S
H
H
S
. Since only the rst term in (30) depends on C, MMSE
solution chooses C = H
H
S
R
1
to make it zero. The term tr(IH
H
S
R
1
H
S
) represents
the mse. Optimal antenna selection must choose the transmit antennas so as to min-
imize the mse. Observing that mse is minimized when tr(H
H
S
R
1
H
S
) is maximized
and using the trace identity tr(XY) = tr(YX), optimal antenna selection solution
can be obtained as
S
mmse
= arg max
H
S
H
N
t
tr(H
S
H
H
S
R
1
), (31)
where S
mmse
denotes the subset of selected antennas and H
Nt
denotes the N
r

N
r
channel matrix. To select the antennas in the optimal way, the trace involving
matrix inversion has to be computed for
_
Nt
Lt
_
possible combinations. This could
be computationally burdensome when the number of antenna combinations is large.
In the following section, we present several reduced complexity antenna selection
algorithms. Looking at (31), we observe that as N
o
0, C H
H
S
(H
S
H
H
S
)
1
, which
is the Zero-Forcing (ZF) lter. Therefore, all the algorithms presented in this chapter
for MMSE receiver can also be applied to ZF receiver by choosing a suciently small
value for N
o
.
5.3 Transmit Antenna Selection
As described in the previous section, (31) can be used to determine the optimal subset
of transmit antennas. However, (31) involves matrix inversion and large number of
choices will make the search prohibitive. In the following subsections, we present
two greedy solutions that are suboptimal but ecient. Both greedy algorithms try to
41
maximize the trace in (31) at each intermediate step. An antenna is selected/removed
so that doing so leads to maximum increase in trace.
5.3.1 No Interference
5.3.1.1 Incremental Transmit Antenna Selection (ITAS)
The algorithm begins with an empty set of selected antennas and selects one antenna
in each step. After n steps, there are n selected antennas and the corresponding
channel matrix is denoted by H
n
. The matrix H
n
consists of the columns of H
Nt
in
the same order as they appear in H
Nt
. We will denote the k
th
column of by h
k
and
will use the following notations
B
n
=
_
I +
1
N
o
H
n
H
H
n
_
1
(32)

n
= tr(H
n
H
H
n
B
n
), (33)
where the subscript n denotes the n
th
step. Now if in the (n + 1)
st
step, transmit
antenna corresponding to the k
th
column of H
Nt
is selected, then h
k
is inserted in
proper position in H
n
to obtain the channel matrix H
n+1
. The matrices in (33) are
updated as
H
n+1
H
H
n+1
= H
n
H
H
n
+h
k
h
H
k
(34)
B
n+1
= B
n

1
N
o
+h
H
k
B
n
h
k
B
n
h
k
h
H
k
B
n
, (35)
where (35) follows as a direct consequence of the matrix inversion lemma [34, (3.5.2.2)].
Using (34) and (35), the updated trace can be computed as

n+1
=
n
+ tr[h
k
h
H
k
B
n
] (36)

tr[H
n
H
H
n
B
n
h
k
h
H
k
B
n
] + tr[h
k
h
H
k
B
n
h
k
h
H
k
B
n
]
N
o
+h
H
k
B
n
h
k
.
42
Since B
n
is a Hermitian matrix, we can use the identity tr(XY) = tr(XY) to further
simplify (36) as
tr[h
k
h
H
k
B
n
] = h
H
k
B
n
h
k
tr[H
n
H
H
n
B
n
h
k
h
H
k
B
n
] = ||h
H
k
B
n
H
n
||
2
(37)
tr[h
k
h
H
k
B
n
h
k
h
H
k
B
n
] = (h
H
k
B
n
h
k
)
2
.
Finally, substituting (37) in (36) and after some algebraic simplications, we can
express the change in the trace after the (n + 1)
st
step as

n+1,k
=
n+1

n
=
N
o
h
H
k

n,k
||
H
n,k
H
n
||
2
N
o
+h
H
k

n,k
, (38)
where
n,k
= B
n
h
k
, and
n+1,k
represents the increase in the trace if k
th
transmit
antenna is selected in the (n+1)
st
step. Since our objective is to maximize the trace,
we select the antenna whose contribution to the trace is the maximum. Thus the
selection criteria is
P = arg max
k

n+1,k
. (39)
To reduce the number of computations,
n+1,k
can be computed recursively as

n+1,k
= R
n+1
h
k
= B
n
h
k

B
n
h
P
h
H
P
B
n
h
k
N
o
+h
H
P
B
n
h
P
=
n,k

n,P
h
H
P

n,k
N
o
+h
H
P

n,P
(40)
assuming that h
P
is selected in the n
th
step. The proposed algorithm is summarized
in Table 1 with the right column showing order of complexity of each step.
5.3.1.2 Decremental Transmit Antenna Selection (DTAS)
As the name suggests, the algorithm begins with full set of selected antennas and
removes one antenna in each step. Thus at end of n
th
step, n antennas are rejected
43
and H
Ntn
represents the channel matrix corresponding to remaining N
t
n antennas.
The matrix B
n
is dened as
B
n
=
_
I +
1
N
o
H
Ntn
H
H
Ntn
_
1
. (41)
If in the (n + 1)
st
step, transmit antenna corresponding to k
th
column of H
Nt
is
removed, then R
n+1
is given by
B
n+1
= B
n
+
1
N
o
h
H
k
B
n
h
k
B
n
h
k
h
H
k
B
n
, (42)
where (42) is obtained using matrix inversion lemma [34, (3.5.2.2)]. Following the
development similar to (34)-(38), the change in trace can be expressed as

n+1,k
=
n+1

n
=
||
H
n,k
H
n
||
2
N
o
h
H
k

n,k
N
o
h
H
k

n,k
, (43)
where
n,k
is dened as in (38). Now the elimination criteria is to remove antenna
whose removal leads to the maximum increase in trace, i.e,
P = arg max
k

n+1,k
. (44)
It is not dicult to show that the vectors
n+1,k
can be updated as

n+1,k
=
n,k
+
n,P
h
H
P

n,k
N
o
h
H
P

n,P
. (45)
The algorithm along with the complexity analysis is outlined in Table 2. Let us com-
pare the complexity of the two algorithms. For the single user case, ITAS doesnt
require matrix inversion and has the lower complexity of O
_
max(N
r
, L
2
t
)N
r
N
t
_
com-
pared to O(N
r
N
2
t
L
t
) of transmit selection algorithm for LRA-ZF in [4, Table.1].
On the other hand, DTAS requires matrix inversion and has a complexity order of
O
_
max(N
3
t
, N
2
r
)N
r
_
. Thus the ITAS algorithm has signicantly lower complexity
compared to the DTAS algorithm.
When only one antenna is selected, ITAS always nds the optimal solution as it
requires just one step. Whereas when N
t
1 transmit antennas are selected, DTAS
44
Algorithm 1 ITAS Algorithm
Require: N
t
, N
r
, L
t
,
2
n
, H
Nt
Ensure: N
r
N
t
L
t
Complexity
1: S := {1, 2, ..., N
t
}, H := [ ]
00
2: B := I
3: for k S do
4:
k
:= Bh
k
5: end for
6: for n = 1 to L
t
do
7: for k S do
8:
k
:=
Noh
H
k

k
||
H
k
H||
2
No+h
H
k

k
O(N
r
N
t
L
2
t
)
9: end for
10: P := arg max
k

k
O(N
t
L
t
)
11: S := S {P}
12: H := [H, h
p
]
13: for k S do
14:
k
:=
k

P
h
H
P

k
No+h
H
P

P
O(N
r
N
t
L
t
)
15: end for
16: end for
return {1, 2, ...N
t
} S
yields the optimal solution. It is intuitive to think that if ITAS takes fewer steps
compared to DTAS, then it is more likely to yield optimal solution than the DTAS
algorithm and vice versa.
5.3.2 Co-channel Interference
Let us consider a system where cochannel interference is present from K other users.
Let G
i
denote the normalized channel matrix between the i
th
interferer and the
desired receiver. Now, the vector of received complex baseband samples after matched
ltering can be expressed as
r = Ha +
K

i=1
G
i
x
H
i
+n, (46)
where x
i
represents the normalized transmitted signal from the i
th
interferer. The
noise vector and the channel matrices G
i
have independent and identically distributed
zero means and unit variance complex Gaussian entries. For simplicity, we assume
45
Algorithm 2 DTAS Algorithm
Require: N
t
, N
r
, L
t
, N
o
, H
Nt
Ensure: N
r
N
t
L
t
Complexity
1: S := {1, 2, ..., N
t
}, H := H
Nt
2: B :=
_
I
Nr
+
1
No
H
Nt
H
H
Nt
_
1
O(N
3
r
)
3: for k S do
4:
k
:= Bh
k
O(N
2
r
N
t
)
5: end for
6: for n = 1 to N
t
L
t
do
7: for k S do
8:
k
:=
||
H
k
H||
2
Noh
H
k

k
Noh
H
k

k
O(N
r
N
3
t
)
9: end for
10: P := arg max
k

k
O(N
2
t
)
11: S := S {P}
12: H := H{h
p
}
13: for k S do
14:
k
:=
k
+
P
h
H
P

k
Noh
H
P

P
O(N
r
N
2
t
)
15: end for
16: end for
return S
all of the interfering signals x
i
, i (1, K), are unknown to the receiver. This model
is well suited to the case where each user chooses his signaling without knowing the
exact interference environment he will face. If we assume that the power covariance
matrix of x
i
is E[x
i
x
H
i
] = I, then the covariance matrix of received vector r can be
obtained as
R = E[rr
H
] = HH
H
+
K

i=1
G
i
G
H
i
+ N
o
I. (47)
Now, the matrix B used in Table 1 and Table 2 can be obtained as
B =
_

_
N
o
_
RHH
H
_
1
for ITAS
N
o
R
1
for DTAS
(48)
It is apparent that in presence of cochannel interference, the computational complex-
ity associated with the Steps 2 and 4 of the ITAS algorithm becomes the same as
in the DTAS algorithm, thus removing any computational advantage oered by the
46
former.
5.4 Receive Antenna Selection
For spatial multiplexing systems, receive antenna selection is an eective, low-complexity
means to provide diversity gain to the data streams. Antenna selection at the receiver
plays an even more important role in interference-limited channels. In the following
subsections, we present greedy solutions for receive antenna selection.
5.4.1 No Interference
5.4.1.1 Incremental Receive Antenna Selection (IRAS)
The algorithm works in a similar fashion as the incremental transmit antenna selection
algorithm presented in the previous section. In each step, we select the row vector
that leads to minimal increase in the trace given by (31). Let us use the following
notations
B
n
=
_
I +
1
N
o
H
H
n
H
n
_
1
(49)

n
= tr(B
n
), (50)
where B
0
= I is the initial condition for the algorithm. Now if in the (n + 1)
st
step,
the k
th
row of H, r
k
, is selected then B
n+1
can be updated using matrix inversion
lemma [34, (3.5.2.2)] as
B
n+1
= B
n

1
N
o
+r
k
B
n
r
H
k
B
n
r
H
k
r
k
B
n
. (51)
Thus the change in trace after the (n + 1)
st
step can be easily computed as

n+1,k
=
n+1

n
=
||
n,k
||
2
N
o
+
n,k
r
H
k
, (52)
where
n,k
= r
k
B
n
. The selection rule is to choose the antenna that leads to minimum
increase in trace, i.e,
P = arg min
k

n+1,k
. (53)
47
Algorithm 3 IRAS Algorithm
Require: N
t
, N
r
, L
r
, N
o
, H
Nr
Ensure: N
r
L
r
N
t
Complexity
1: S := {1, 2, ..., N
r
}
2: B := I
Nt
3: for k S do
4:
k
:= r
k
B
5: end for
6: for n = 1 to L
r
do
7: for k S do
8:
k
:=
||
k
||
2
No+
k
r
H
k
r
H
k
O(N
r
N
t
L
r
)
9: end for
10: P := arg min
k

k
O(N
r
L
r
)
11: S := S {P}
12: for k S do
13:
k
:=
k

P

k
r
H
P
No+
P
r
H
P
O(N
r
N
t
L
r
)
14: end for
15: end for
return {1, 2, ...N
r
} S
The vectors
n+1,k
can be recursively computed as

n+1,k
=
n,k

n,P

n,k
r
H
P
N
o
+
n,P
r
H
P
. (54)
The algorithm is summarized in Table 3 with the right column showing the complexity
corresponding to each part of the algorithm.
5.4.1.2 Decremental Receive Antenna Selection (DRAS)
The algorithm is quite similar to the DTAS algorithm presented in the previous sec-
tion. For the sake of brevity we omit the development and present the implementation
in Table 4. From Tables 3 and 4, it is apparent that IRAS has signicantly lower
complexity, O(N
t
N
r
L
r
), compared to O
_
max(N
2
t
, N
t
N
r
, N
2
r
)N
t
_
, that of the DRAS
algorithm. Observing that N
t
N
r
is always satised to avoid overloading the re-
ceiver, complexity order of DRAS algorithm can be expressed as O(N
t
N
2
r
). The
performance relationship between the two algorithms is same as between ITAS and
DTAS algorithms.
48
Algorithm 4 DRAS Algorithm
Require: N
t
, N
r
, L
r
, N
o
, H
Nr
Ensure: N
r
L
r
N
t
Complexity
1: S := {1, 2, ..., N
r
}
2: B :=
_
I
Nt
+
1
No
H
H
Nr
H
Nr
_
1
O(N
3
t
)
3: for k S do
4:
k
:= r
k
B O(N
r
N
2
t
)
5: end for
6: for n = 1 to N
r
L
r
do
7: for k S do
8:
k
:=
||
k
||
2
No
k
r
H
k
r
H
k
O(N
2
r
N
t
)
9: end for
10: P := arg min
k

k
O(N
2
r
)
11: S := S {P}
12: for k S do
13:
k
:=
k
+
P

k
r
H
P
No
P
r
H
P
O(N
2
r
N
t
)
14: end for
15: end for
return S
5.4.2 Co-channel Interference
Unlike transmit antenna selection, it is dicult to obtain optimal solution for receive
antenna selection using the receive covariance matrix, R. However, an approximate
solution can be obtained using the simple channel whitening approach of [5] as

H = T
1/2
H, (55)
where T represents the noise plus interference covariance matrix. It is interesting to
note that such an approach leads to fairly good antenna selection as will be shown
below.
5.5 Simulation Results
In this section we present the simulation results on the BER performance of the
proposed antenna selection algorithms and compare them against some of the existing
49
1 2 3 4 5 6
10
5
10
4
10
3
10
2
10
1
L
t
A
v
e
r
a
g
e

B
E
R
Deterministic Selection
NBS Algorithm
DTAS Algorithm
ITAS Algorithm
Optimal Selection
Figure 15: MMSE error performance for (6,6) MIMO system with varying number
active transmit antennas and SNR = 5dB.
antenna selection techniques [24, 23, 20, 21, 4]. For the sake of completeness, in all the
gures, we also show the performance curves corresponding to deterministic and NBS
selection. All simulation results assume an independent and identically distributed
(i.i.d) Rayleigh fading channel unless specied otherwise. Simulations results are
generated using Monte Carlo simulation of 10
6
channel realizations. Throughout our
simulations, we assume 16-QAM modulation.
In the rst example, we consider an isolated MIMO link with linear MMSE re-
ceiver. Figure 15 shows the BER performance versus L
t
for SNR = 5 dB. It is
apparent that all antenna selection algorithms perform signicantly better than the
deterministic selection, which cant exploit transmit diversity. In particular, the pro-
posed ITAS and DTAS algorithms oer close to optimal performance. The ITAS
and NBS algorithms are both optimal for L
t
= 1. Whereas, for L
t
= N
t
1, DTAS
algorithm yields optimal performance. As a general rule of thumb, ITAS algorithm
50
0 5 10 15 20 25 30
10
4
10
3
10
2
10
1
10
0
SNR (dB)
A
v
e
r
a
g
e

B
E
R
Deterministic Selection
NBS Algorithm
Gores Algorithm [21]
Berenguers Algorithm [4]
ITAS Algorithm
DTAS Algorithm
Optimal Selection
Figure 16: Error performance of (4,4) MIMO system with ZF receiver in the presence
of transmit correlation. Three transmit antennas are selected.
is preferable when L
t
< N
t
/2 whereas for L
t
> N
t
/2, DTAS is desirable.
From Figure 15, it is also interesting to note that as number of selected antennas
increases, the BER performance of deterministic selection deteriorates less rapidly
compared to other antenna selection algorithms. This can attributed to the fact that
the diversity gain associated with the deterministic selection comes only from the
receive diversity, which reduces as more streams are added. Contrary to this, the
proposed transmit antenna selection algorithms help achieve transmit diversity gain
resulting in an overall diversity gain of (N
t
L
t
)(N
r
L
t
), which decreases more
rapidly with increasing L
t
.
Next, we consider a (4,4) MIMO link with ZF receiver and assume transmit cor-
relation but no correlation at the receiver. Figure 16 compares the BER performance
of the proposed algorithms against Gores Algorithm [21] and Berenguers algorithm
[4]. The transmit correlation matrix, R
t
, used in the simulations is same as in [21,
51
Eq. (15)]. Gores algorithm uses the transmit correlation information to select anten-
nas that minimize the error probability of the worst stream. The selected antennas
are {1, 2, 4} for the transmit correlation matrix, R
t
. It, obviously, is not an optimal
strategy because all the streams and not just the worst stream determine the error
performance of a linear receiver.
From Figure 16, we observe that although Gores algorithm performs almost 3
dB better than non-optimal deterministic selection, which selects {1, 2, 3}, it fails to
exploit the full transmit diversity that can be achieved by choosing optimal set of
transmit antennas for each channel realization. For example, when the channel is in
deep fade, it is better to choose two strongly correlated but un-faded streams and an
uncorrelated stream in deep fade rather than all three uncorrelated streams with two
being in deep fade. It can also be noted that Berenguers transmit antenna selection
algorithm for ZF receiver performs very close to the ITAS algorithm but is still 1 dB
worse than the DTAS algorithm, which is optimal in this case.
Figure 17 shows the BER performance of various sub-optimal selection approaches
in the presence of xed co-channel interference assuming i.i.d Rayleigh faded chan-
nels. The interference is assumed to be resulting from another MIMO transmitter
using 3 antennas. The desired transmitter is equipped with 6 antennas and chooses 3
antennas to improve the BER performance of the desired link. From Figure 17, it is
apparent that the BER curves for both deterministic selection and NBS selection are
relatively at. This is because the receiver, in the absence of transmit diversity, uses
additional 3 antennas to suppress the interference and thus no diversity gain is ren-
dered to the desired streams. On the other hand, the proposed algorithms, ITAS and
DTAS, are able to exploit transmit diversity better and provide reasonably good BER
performance. The robustness of the proposed algorithms is further highlighted when
the interference power increases to 0 dB, which causes only a minimal degradation in
the BER performance.
52
0 5 10 15 20 25
10
4
10
3
10
2
10
1
10
0
SNR (dB)
A
v
e
r
a
g
e

B
E
R
Deterministic Selection
NBS Algorithm
DTAS Algorithm
ITAS Algorithm
Optimal Selection
(a) SIR = 10 dB
0 5 10 15 20 25
10
4
10
3
10
2
10
1
10
0
SNR (dB)
A
v
e
r
a
g
e

B
E
R
Deterministic Selection
NBS Algorithm
DTAS Algorithm
ITAS Algorithm
Optimal Selection
(b) SIR = 0 dB
Figure 17: MMSE error rate for (6,6) MIMO link in presence of co-channel inter-
ference from 3 other streams. Three transmit antennas are selected.
53
0 5 10 15 20 25
10
4
10
3
10
2
10
1
10
0
SNR (dB)
A
v
e
r
a
g
e

B
E
R
Deterministic Selection
NBS Algorithm
Gores Algorithm [20]
ITAS Algorithm
DTAS Algorithm
Optimal Selection
Figure 18: MMSE error rate for (6,4) MIMO system. Four receive antennas are
selected.
Figure 18 illustrates the error performance of various receive antenna selection al-
gorithms for the MMSE receiver. We notice that Gores algorithm [20], which chooses
antennas that maximize the capacity, is only slightly better than the NBS algorithm
and performs much worse than the optimal solution. It is evident that sub-optimal
receive algorithms yield near-optimal performances. In particular, DRAS algorithm
performs better than the ITAS algorithm as it requires 2 iterations compared to 4 for
the latter.
Figure 19 shows the BER performance of dierent receive antenna selection algo-
rithms in the presence of xed co-channel interference. The (6,2) MIMO system is
subjected to co-channel interference from 2 other streams with equal power as the de-
sired streams. The selections algorithms choose 4 receive antennas that minimize the
link error. It is evident that optimal selection approach, which assumes interferers
channel knowledge, performs the best. Whereas the IRAS and DRAS algorithms, in
54
0 5 10 15 20 25
10
4
10
3
10
2
10
1
10
0
SNR (dB)
A
v
e
r
a
g
e

B
E
R
Deterministic Selection
NBS Algorithm using whitened channel
ITAS Algorithm using whitened channel
DTAS Algorithm using whitened channel
Optimal Selection using interferers channel
Figure 19: BER performance of MMSE receiver for (6,2) MIMO system in the
presence of 2 interfering streams with SIR = 0dB. Four receive antennas are selected.
the absence of interferers channel knowledge, provide signicant improvement over
deterministic and NBS selection. It was also found that as the SIR is increased, the
gap between the proposed algorithms and the optimal selection approach is consid-
erably reduced.
5.6 Summary
In this chapter, we have presented new computationally ecient algorithms for trans-
mit and receive antenna selection for linear receivers. The presented algorithms are
also applicable to handle xed co-channel interference and oer near optimal BER
performance both in presence of transmit correlation and co-channel interference. In
absence of co-channel interference, incremental selection algorithms have much lower
complexity compared to the decremental selectional algorithms and are therefore more
attractive.
55
CHAPTER VI
MULTIUSER DETECTION FOR STBC USERS
In the previous chapters we considered closed-loop or transmitter selection Spatial
Multiplexing (SM) MIMO systems, which oer extremely high spectral eciencies.
These systems require explicit feedback paths for communicating full or partial CSI
to the transmitter for computing the power allocation matrices. However, the trans-
mission errors in the feedback channel may adversely aect the performance of such
systems. Also, the bandwidth requirements for the feedback channel along with errors
in the feedback path may altogether reduce the eective throughput gains oered by
SM-MIMO systems.
In this chapter, we focus our attention to Space-Time Block Codes (STBCs), which
oer transmit diversity gains without requiring any explicit channel feedback [1, 48].
The temporal and spatial structure of certain STBCs, called orthogonal STBCs (OS-
TBCs), oer the additional advantage of maximum likelihood detection based only
on linear processing at the receiver [48]. As a result, the OSTBCs, especially the
Alamouti code [1], are being widely adopted in various wireless communication stan-
dards such as the 3GPP cellular standard, WiMAX (IEEE 802.16), and IEEE 802.11n.
These standards do not allow co-channel interference among dierent users. However,
in a multiuser environment, the network performance can be signicantly improved if
spatial diversity is exploited by allowing controlled co-channel transmissions [41]. In
this chapter, we explore co-channel interference problem and propose IC techniques
for OSTBCs with 3 and 4 antennas which can apply to an SDMA MAC.
Despite the low data rates supported by various STBCs, they are attractive from
a network point of view as they cause correlated interference, which can be mitigated
using only one additional antenna without sacricing space-time diversity gains. In
56
[37], Naguib et al. developed a minimum mean square error (MMSE) interference
suppression technique for two co-channel users employing the Alamouti code [1]. It
was shown that the impact of co-channel interference can be completely eliminated
by adding one more receive antenna without sacricing space-time diversity gains.
This scheme was later extended to frequency-selective channels in [53]. In [29] au-
thors extend the work of [37] to the interference cancellation (IC) and detection for
users with 4 transmit antennas using quasi-orthogonal space-time block coding [26].
In another related work, Stamoulis et al. [45] presented a simple suboptimal linear
processing scheme that achieves IC of two co-channel users employing any rate-1/2
complex STBC based on an orthogonal design. Like [37], [45] cancels the correlated
STBC interference with the aid of additional receive antenna while preserving the
diversity gains. However, the method is based on 2-stage receive processing and is
not optimal in terms of mean square error (MSE).
In contrast to [45], the algorithm presented in this chapter is single-stage and
MMSE-optimal. In contrast to [37], the proposed algorithm treats unity rate real
and derived rate-1/2 complex OSTBCs for 3 and 4 transmit antennas [48]. Like [37]
and [45], the algorithm requires only one additional antenna to cancel an OSTBC
interference. Some special algebraic properties of linear OSTBCs are exploited to
make the algorithm have low complexity.
6.1 Real STBC and IC of Two Co-channel Users
In this section, we focus on the MMSE-based multiuser detection of two synchronous
co-channel users, each employing the same real STBCs. We will show for n
t
= 3 or
4, the linear receiver processing can be implemented without matrix inversions, and
after the IC-stage, the symbols are recovered with space-time diversity gain.
Let us use subscripts 1 and 2 to dierentiate between the corresponding channel
matrices and symbol vectors of the two users. Thus the received baseband vector is
57
given by
y

= F

1
s
1
+F

2
s
2
+v

= F

s +v

(56)
where F
i
s are dened as in (13), F =
_
F
1
F
2
_
, and s =
_
s
T
1
s
T
2
_
T
. Let us assume
that the receiver employs the MMSE lter given by C = (F
T
F

+N
o
I)
1
F
T
, where N
o
denotes the noise variance. It is not dicult to see that F
T
F

= Re(F
H
F). Exploiting
the channel decoupling property of orthogonal STBCs, Re(F
H
i
F
i
) = ||H
i
||
2
I [32,
(7.4.1.4)], we can simplify F
T
F

as
F
T
F

=
_
_
_
||H
1
||
2
I R
12
R
T
12
||H
2
||
2
I
_
_
_
, (57)
where R
12
= Re(F
H
1
F
2
) and ||.|| denotes the Frobenius norm of the matrix argument.
Now the matrix inversion component of MMSE lter can be simplied using (57)
with the aid of block matrix inversion [34, (3.5.3.1)], as
(F
T
F

+ N
o
I)
1
=
_
_
_
(||H
2
||
2
+ N
o
) R
12

R
T
12
(||H
1
||
2
+ N
o
)
_
_
_, (58)
where
=
_
(||H
1
||
2
+ N
o
)(||H
2
||
2
+ N
o
)I R
12
R
T
12
_
1
. (59)
Next, by multiplying both sides of (58) on the right by (F
T
F

+ N
o
I), substituting
F
T
F

with (57), and equating the (2, 1)

th
sub-matrix element of the resulting product
to 0, we get R
T
12
= R
T
12
. Exploiting the self-adjoint nature of , it can be further
shown also commutes with R
12
, i.e.,
R
12
= R
12
. (60)
Using (58)-(60), the resulting MMSE lter can be expressed in a particularly useful
form as
C =
_
_
_
Re(W
1,2
) Im(W
1,2
)
Re(W
2,1
) Im(W
2,1
)
_
_
_
, (61)
58
where
W
i,j
=
_
(||H
j
||
2
+ N
o
)F
H
i
R
ij
F
H
j
_
. (62)
It is interesting to note that the design matrices associated with the unity rate real
orthogonal STBCs for n
t
= 3 and 4 [48, (37)-(38)], have some unique properties
in addition to (10), namely A
3
A
T
1
= A
4
A
T
2
and A
3
A
T
2
= A
4
A
T
1
[32]. For such
OSTBCs, it is not dicult to show that R
12
R
T
12
= I, where
=
4

k=1
_
tr
_
Re(H
H
1
H
2
A
k
A
T
1
)
_
2
. (63)
It may be noted that the product A
k
A
T
1
is a sparse matrix and hence the computation
cost associated with is very insignicant. On an average, the computation of
involves n
r
n
t
multiplication operations.
After substituting the simplied expression for R
12
R
T
12
into (59), the MMSE lter
expression in (61) for real orthogonal codes with n
t
= 3 or 4, can be simplied as
W
i,j
=
(||H
j
||
2
F
H
i
R
ij
F
H
j
)
(||H
1
||
2
+ N
o
)(||H
2
||
2
+ N
o
)
, (64)
where is dened as in (63). Now the decoded streams for the two users can be
easily obtained as
r = Cy

=
_
_
_
Re(W
1,2
y)
Re(W
2,1
y)
_
_
_
, (65)
where Re(W
i,j
y) corresponds to the decoded stream for the i
th
user. It may be
noted that the expression W
i,j
in (64) does not involve any matrix inversion. The
computation of W
i,j
with the aid of equation (64) only requires 3n
r
n
t
+ 2nn
r
n
2
s
multiplication operations compared to 2n
r
n
t
+ 3nn
r
n
2
s
+ n
2
s
+ n
3
s
needed if W
i,j
is
computed using (59) and (62). As a numerical example, for n
t
= 3 and n
r
= 2,
computation of W
i,j
in (62) and (64) requires 274 and 476 multiplication operations,
respectively. Thus, the proposed MMSE-based IC of two co-channel users with real
STBCs can be achieved using simple linear processing.
59
In contrast to the presented algorithm, Stamoulis et al. [45] constrain the matrix
lter to have identity blocks on the main diagonal and require that the o-diagonals
remove all interference:
r
i
= y
i
Z
i
y
j
, (66)
where i, j (1, 2), j = i and r
i
indicates the post-IC recovered data stream corre-
sponding to the i
th
user. The matrix Z
i
denotes the space-time lter for the i
th
user
as dened in [45, (11)], and y
i
indicates the baseband signal corresponding to the i
th
receive antenna. In other words, the i
th
antenna is intended to receive only the i
th
user; the other antenna is used only to estimate and subtract the interference that
is in the i
th
antenna. While this is a zero-forcing (ZF) solution, it does not have the
most general (i.e. full matrix) formulation, and as we will show, it suers a signi-
cant SNR degradation relative to the MMSE solution as well as the full-matrix ZF
solution.
6.2 IC with Rate-1/2 Complex OSTBCs
Let us consider two co-channel users, each employing the same rate-1/2 complex or-
thogonal code. A rate-1/2 generalized complex orthogonal design can be constructed
from a real orthogonal design X
nt
as X
C
nt
=
_
X
nt
conj(X
nt
)
_
and can also be rep-
resented as
X
C
nt
=
ns

i=1
Re(s
i
)
_
A
i
A
i
_
+ jIm(s
i
)
_
A
i
A
i
_
, (67)
where X
C
nt
is the n
t
(2n) STBC matrix, {s
i
}
ns
i=1
denotes the transmitted complex
symbols, {A
i
}
ns
i=1
are the n
t
n matrices representing the code structure of parent
real codes, and the operator conj(.) denotes the conjugate of its argument. Following
the development similar to (11)-(56), and observing that n
s
symbols are transmitted
using n
t
antennas in 2n timeslots, it is not dicult to show that the 2n 1 vector
60
representing the received baseband signal for the rst link can be expressed as
_
_
_
y
1:n
y
n+1:2n
_
_
_

= G

+v

, (68)
where G =
_
G
1
G
2
_
, and G
k
s are dened as
G
k
=
_
_
_
F
k
jF
k
F
k
jF
k
_
_
_
. (69)
where the subscript k (1, 2). Following the development similar to (56)-(63) and
after some simplications, the linear MMSE detector expression can be shown to be
C =
1
2
_
_
_
_
_
_
Re(W
1,2
) Re(W
1,2
) Im(W
1,2
) Im(W
1,2
)
Im(W
1,2
) Im(W
1,2
) Re(W
1,2
) Re(W
1,2
)
Re(W
2,1
) Re(W
2,1
) Im(W
2,1
) Im(W
2,1
)
Im(W
2,1
) Im(W
2,1
) Re(W
2,1
) Re(W
2,1
)
_
_
_
_
_
_
(70)
where W
i,j
is dened in (62). After some simplications, the decoded streams corre-
sponding to the two users can be obtained as
r

=
1
2
_
_
_
_
_
_
_
_
_
Re(W
1,2
(y
1:n
+y
n+1:2n
))
Im(W
1,2
(y
1:n
y
n+1:2n
))
Re(W
2,1
(y
1:n
+y
n+1:2n
))
Im(W
2,1
(y
1:n
y
n+1:2n
))
_
_
_
_
_
_
_
_
_
(71)
or expressed dierently as
r =
1
2
_
_
_
W
1,2
y
1:n
+ conj(W
1,2
y
n+1:2n
)
W
2,1
y
1:n
+ conj(W
2,1
y
n+1:2n
)
_
_
_ (72)
The above equation shows that the IC of two co-channel users employing the derived
rate-1/2 complex OSTBCs can be implemented using the same MMSE lter designs,
W
i,j
, as for the IC of two users employing real STBCs.
It may be noted that for n
t
= 3 and 4 there exist higher rate complex OSTBCs
than rate-1/2 codes considered in this paper. For example, there exist rate-3/4 codes
61
called truncated Octonion and Octonion for n
t
= 3 and 4, respectively [32, (7.4.8)-
(7.4.10)]. We further note that the above IC simplication cannot be achieved using
these codes as the corresponding matrices R
12
R
T
12
are not diagonal as required to
simplify (59).
6.3 Simulation Examples
Figure 20 shows the average BER performance of the proposed interference suppres-
sion technique. The simulation results assume two co-channel users, A and B, each
equipped with 3 transmit antennas and employing rate-1/2 complex OSTBC [48,
(37)]. All channels are assumed to be i.i.d Rayleigh fading channels. QPSK modula-
tion is used and the receiver is assumed to have 2 antennas.
Fig. 20(a) shows the case where the two users have equal power. We also include
the zero-forcing (ZF) solution, which can be obtained from the MMSE solution by
setting N
o
= 0. We observe that Stamoulis IC technique for two users and two
receive antennas delivers about the same ABER performance as if there were only
one user and one receive antenna; this is consistent with the results in [4]. On the
other hand, both the ZF and MMSE solutions have a signicant SNR advantage
over the Stamoulis solution; for example, at ABER = 10
3
, the SNR advantage
is approximately 3.5 dB. Therefore, although Stamoulis solution separates the two
users, it is not optimal with respect to the MSE criterion as it requires 2-stage receive
processing and uses one receive antenna specically to cancel out the interference
from the other user.
Fig. 20(b) shows the case of a 10 dB power imbalance between the users. We
see that for the higher-powered user, full matrix ZF results are about the same as in
Fig. 20(a), but the MMSE case is about 1 dB better at ABER = 10
3
. It is well
known that the lower-power user will suer in a linear multi-user detector, and this
is conrmed here for all three cases (the top three curves), although the full-matrix
62
0 1 2 3 4 5 6 7 8
10
4
10
3
10
2
10
1
SNR (dB)
A
v
e
r
a
g
e

B
E
R
One User, 1 Rx
One User, 2 Rx
Two Users, 2 Rx [Stamoulis]
Two Users, 2 Rx [ZF, Eq. (72)]
Two Users, 2 Rx [MMSE, Eq. (72)]
Stamoulis
ZF
MMSE
(a) Equi-powered Users, P
T
(A) = P
T
(B)
0 1 2 3 4 5 6 7 8
10
4
10
3
10
2
10
1
10
0
SNR (dB)
A
v
e
r
a
g
e

B
E
R
One User, 1 Rx
One User, 2 Rx
Two Users, 2 Rx [Stamoulis]
Two Users, 2 Rx [ZF, Eq. (72)]
Two Users, 2 Rx [MMSE, Eq. (72)]
UserB
UserA
(b) Power imbalance, P
T
(A) = 10P
T
(B)
Figure 20: Average BER performance of User-A
63
solutions still outperform Stamoulis et al.s solution.
6.4 Summary
A simple single-stage interference suppression method is proposed that exploits the
algebraic structure of unity rate OSTBCs for 3 and 4 transmit antennas. The pro-
posed method provides an MMSE-optimal interference suppression technique that is
also low complexity because it avoids matrix inversion. The simulation results show
the advantages of the proposed method.
64
CHAPTER VII
IC OF ALAMOUTI INTERFERENCE FOR A
V-BLAST USER
In the previous chapter, we considered the problem of interference cancellation for
two co-channel OSTBC users. The analysis presented therein along with the related
work in [15, 45, 53, 29, 30] assume that all users employ the same OSTBC. However,
in a MIMO network dierent nodes can transmit using dierent transmission tech-
niques including V-BLAST, beam-forming or STBC to improve the overall network
performance. The BER of a single-input multiple-output (SIMO) or a MIMO link
subjected to STBC interference can be signicantly improved if the victim receiver
exploits the rich algebraic structure of the STBC interference [30, 31]. In [55], Zhang
et al. exploit the algebraic properties of Alamouti code to develop a selective co-
channel interference mitigation method for two users employing Alamouti codes. In
[46], MMSE-based IC is modied to adept to beamforming, which further improves
the BER performance.
In [39], Okazaki et al. consider symbol detection problem for a hybrid STBC
system that transmits both STBC symbols and Spatial Multiplexing (SM) symbols
simultaneously. In such systems, conventional SIC detects STBC symbols rst owing
to their inherent SNR advantage. As a result, STBC symbols suer more from resid-
ual interference compared to SM symbols, which are detected at a later stage. The
authors propose a soft decision SIC which changes detection order at every stage so
as to minimize the residual interference faced by each symbol. Dierent from com-
putationally intensive IC techniques presented in [39], we develop computationally
ecient but sub-optimal linear space-time IC techniques, which outperform conven-
tional linear IC techniques that do not take the structure of Alamouti interference
65
into account.
In the following, we rst consider a SIMO link subjected to Alamouti interference
and develop IC techniques that exploit the temporal and spatial structure of the
Alamouti code. In the subsequent sections, we extend the framework to include a
V-BLAST link and present simulation results to evaluate the various aspects of the
proposed IC technique. Based on our results, we can make the novel observation
that Alamouti interference is less severe than a single-input single-output (SISO)
interference.
7.1 System Model
We consider a system comprising two links (pairs of transmit and receive nodes),
where each link is subjected to co-channel interference from the other link. The
desired link is assumed to have n
t
transmit and n
r
receive antenna elements whereas
the interfering link is assumed to have m
t
transmit antenna elements. During each
symbol period, n
t
input symbols are spatially multiplexed for transmission over the
desired link whereas the interfering link transmits space-time block coded symbols.
The received baseband vector corresponding to the desired link is given by
r =

Hx +

Gy +n (73)
where {H, } and {G, } denote the channel gain matrices and noise normalized
powers corresponding to the desired and interfering links, respectively. Furthermore,
the channel gain matrices are assumed to be slowly varying, Rayleigh faded, and
xed for the duration of an entire burst. The entries of the channel matrices are
independent Gaussian distributed complex random variables with zero mean and
unit variance. The vectors {x, y} denote the transmit symbols for the desired and
interfering links, respectively. The noise n is assumed to be additive white Gaussian
noise (AWGN) with zero mean and unit variance.
66
7.2 IC for a SIMO Link
In this section, we focus on the IC of two synchronous co-channel users. The interferer
is assumed to be using the Alamouti OSTBC and the desired SIMO link consists
of consists of one transmit antenna and two receive antennas. In the following, we
present a linear SIMO receiver processor, which exploits the structure of the Alamouti
code to separate out the desired symbols.
Let h
i,j
and g
i,j
indicate the complex channel gain between the i
th
receive antenna
and j
th
transmit antenna for the desired and the interfering links, respectively. Now
the received baseband vector at the k
th
receive antenna, r
1
k
, can be written as
r
1
k
=

h
k,1
x
1
+

g
k,1
y
1
+

g
k,2
y
2
+ n
1
k
(74)
where the superscript denotes the timeslot index. During the next timeslot, assuming
that the channel gains for the desired and the interfering link remain the same, the
receive baseband signal r
2
k
can be expressed as
r
2
k
=

h
k,1
x
2
+

g
k,1
y

g
k,2
y

1
+ n
2
k
(75)
The received baseband signal at the k
th
antenna in (74)-(75) can be represented in
matrix notations to facilitate further analysis as
r
k
=
_
_
_
r
1
k
r
2
k
_
_
_
=

_
_
_
h
k,1
0
0 h

k,1
_
_
_
. .
H
k,1
_
_
_
x
1
x
2
_
_
_
. .
w
+

_
_
_
g
k,1
g
k,2
g

k,2
g

k,1
_
_
_
. .
G
k
_
_
_
y
1
y
2
_
_
_
. .
y
+
_
_
_
n
1
k
n
2
k
_
_
_
. .
n
k
(76)
or
r
k
=

H
k,1
w+

G
k
y +n
k
(77)
Using the orthogonality of Alamouti code, viz. G
H
k
G
k
=
1
2
||G
k
||
2
I, the interference
can be cancelled using the following linear combination of the received baseband
vectors as
= G
H
2
r
2
/||G
2
||
2
G
H
1
r
1
/||G
1
||
2
(78)
67
which is equivalent to
=

w+ (79)
where = G
H
2
H
2,1
/||G
2
||
2
G
H
1
H
1,1
/||G
1
||
2
, and = G
H
2
n
2
/||G
2
||
2
G
H
1
n
1
/||G
1
||
2
denotes the resultant noise term, which follows the distribution of Additive White
Gaussian Noise with zero mean and a variance (
1
||G
1
||
2
+
1
||G
2
||
2
). It can be easily
veried that the matrix follows the algebraic structure of Alamouti code. Hence,
the desired symbols in (79) can be easily decoupled using the matched lter
H
as
=
H
=

(
H
)w+
H
(80)
where
H
is a 22 diagonal matrix. It is interesting to note that unique structure
of
H
implies that the resultant noise term
H
is also AWGN.
7.2.1 Numerical Results
The diversity gain of a system is characterized by the symbol error probability in the
high SNR regime. A symbol error occurs when the overall channel gain is smaller
than the noise variance, which has a probability P{ <
1

}. In the high SNR regime,

this probability is proportional to
D
, where D denotes the diversity gain of the
system. Thus the diversity gain of a system may be computed as [50]
D = lim
inf
d
d ln
[ln F

(1/)] (81)
where F

(.) denotes the cumulative density function (CDF) of the normalized SNR
, which can be obtained from (80) after algebraic manipulations as
=
|h
1,1
|
2

2
i=1
|g
i,2
|
2
+|h
2,1
|
2

2
i=1
|g
i,1
|
2
2Re(g

1,1
h
1,1
h

2,1
g
1,2
+ g
2,1
h

1,1
h
2,1
g

2,2
)

2
i=1
(|g
i,1
|
2
+|g
i,2
|
2
)
(82)
Unfortunately, a closed-form expression for F

() appears dicult to obtain. We,

therefore, resort to Monte Carlo simulation for generating the CDF. Figure 21 shows
68
10
8
10
6
10
4
10
2
10
0
10
2
10
8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Normalized SNR = /
C
D
F
Rx=1, No interference
Rx=2, IC of SISO interference
Rx=2, IC of Alamouti interference (80)
Rx=2, No interference
Figure 21: The cumulative density function of normalized SNR for various interfer-
ence scenarios with SIR = 0dB.
the numerical F

() computed using 10
7
channel trials as a function of . A loga-
rithmic XY scale is chosen so that the diversity order can be reected by the slope
of respective CDF tails. It is not dicult to observe from Figure 21 that the slope
of CDF tails for no interference scenarios with 1 and 2 receive antennas is approxi-
mately 1 and 2, respectively. Therefore, the diversity order of these two scenarios is
1 and 2, respectively, as expected. Next, we observe from the slope of CDF tail of
the expression in (82) that the diversity order of the concerned system is closer to
2. However, it must be noted that this diversity gain is fully realized at much higher
SNR compared to the scenario with no interference and 2 receive antennas. On the
other hand, the scenario with single-input single output (SISO) interference and 2
receive antennas perform the same as a SISO link with no interference.
For a more practical comparison, Figure 22 shows the comparative BER perfor-
mance for the same scenario as in Figure 21. All transmitters are equi-powered and
69
0 5 10 15 20
10
5
10
4
10
3
10
2
10
1
10
0
SNR dB
A
B
E
R
Rx=1, No interference
Rx=2, IC of SISO interference
Rx=2, IC of Alamouti interference (80)
Rx=2, No interference
SIR = 0 dB
Figure 22: Average bit error probability as a function of SNR with SIR = 0dB.
employ QPSK modulation. The results are generated using Monte Carlo simulation
of 10
7
channel trials. It is apparent that a SIMO link subjected to Alamouti inter-
ference performs better than the SIMO link subjected to SISO interference. Also,
the gap between the BER curves for SISO and SIMO link with Alamouti interfer-
ence increases with increasing SNR, which is indicative of its higher diversity gain as
reected in Figure 21.
7.3 IC for a V-BLAST Link
In the previous section, we treated the case where the desired signal was transmitted
by a single antenna (i.e. the link was SIMO), and the interference was STBC. In this
section, we extend the analysis to include a V-BLAST link subjected to Alamouti
interference from another co-channel user.
Following the development in the previous section, the received baseband signal
70
at k
th
receive antenna can be expressed as:
r
k
=
_
_
_
r
1
k
r
2
k
_
_
_
=

_
_
_
h
k
0
0 h

k
_
_
_
_
_
_
x
1
x
2
_
_
_
+

_
_
_
g
k,1
g
k,2
g

k,2
g

k,1
_
_
_
_
_
_
y
1
y
2
_
_
_
+
_
_
_
n
1
k
n
2
k
_
_
_
(83)
where h
k
=
_
h
k,1
h
k,nt
_
denotes the k
th
row vector of the corresponding channel
matrix H, x denotes the transmitted vector, and the superscripts denote the asso-
ciated timeslot index. The rest of the variables follow the same denition as in the
previous section. The received baseband vector in (83) can be rewritten to better
exploit the structure of Alamouti code as
r
k
=

_
H
k,1
H
k,nt
_
_
_
_
_
_
_
w
1
.
.
.
w
nt
_
_
_
_
_
_
+

G
k
y +n
k
(84)
where, w
j
denotes the 2 1 vector comprising the desired symbols transmitted by
the j
th
antenna during the two consecutive timeslots, i.e.,
w
j
=
_
_
_
x
1
j
x
2
j
_
_
_
(85)
and H
k,j
is dened as
H
k,j
=
_
_
_
h
k,j
0
0 h

k,j
_
_
_
(86)
Now the desired symbols can be recovered employing the zero-forcing (ZF) based
linear block lter given by
Z = pinv
_
_
_
_
_
_
_
_
_
H
1,1
H
1,nt
G
1
H
2,1
H
2,nt
G
2
.
.
.
.
.
.
.
.
.
.
.
.
H
nr,1
H
nr,nt
G
nr
_
_
_
_
_
_
_
_
_
(87)
where pinv denotes the Moore-Penrose pseudoinverse. It is apparent that the numer-
ical complexity of computing Z can be prohibitive with increasing number of receive
71
and/or transmit antennas. We, therefore, look into sub-optimal ways of recovering
the desired symbols.
7.3.1 Sub-optimal IC
The received baseband signal at the n
th
receive antenna can be employed to cancel
out the Alamouti interference at other receive antennas using the following linear
combination,

k
= G
H
k
r
k
/||G
k
||
2
G
H
n
r
n
/||G
n
||
2
(88)
Without loss of generality, lets assume that 1
st
receive antenna is used to cancel out
the Alamouti interference, i.e. n = 1. Now the interference-free vectors,
k
, can be
stacked together to obtain the following representation
=
_
_
_
_
_
_
_
_
_

3
.
.
.

nr
_
_
_
_
_
_
_
_
_
=

_
_
_
_
_
_
_
_
_

2,1

2,2

2,nt

3,1

3,2

3,nt
.
.
.
.
.
.
.
.
.
.
.
.

nr,1

nr,2

nr,nt
_
_
_
_
_
_
_
_
_
. .

_
_
_
_
_
_
_
_
_
w
1
w
2
.
.
.
w
nt
_
_
_
_
_
_
_
_
_
. .
w
+
_
_
_
_
_
_
_
_
_

3
.
.
.

nr
_
_
_
_
_
_
_
_
_
. .

(89)
or
=

w+ (90)
where the 2 2 sub-matrices
k,j
are given by

k,j
= G
H
k
H
k,j
/||G
k
||
2
G
H
1
H
1,j
/||G
1
||
2
(91)
and the resultant noise components
k
are given by

k
= G
H
k
n
k
/||G
k
||
2
G
H
1
n
1
/||G
1
||
2
(92)
Again, it can be easily veried that the sub matrices
k,j
follow the algebraic structure
of Alamouti code, i.e. the product
H
k,j

k,j
is a diagonal matrix. Unfortunately, the
72
said property doesnt hold for as
H
is not diagonal if n
t
> 1. Therefore, the
symbol decoupling in a V-BLAST link is more involved compared to a SIMO link.
Since only one additional receive antenna is used to suppress Alamouti interfer-
ence, we need n
r
n
t
+ 1. For the following analysis, let us assume n
r
= n
t
+ 1.
Now, to recover the transmitted symbols, we follow the approach similar to [45] and
partition the matrix into four blocks as
=
_
_
_

2,1
B
C D
_
_
_
(93)
where B, C, and D denote the 2 2(n
t
1) upper-right, 2(n
t
1) 2 lower-left and
2(n
1
1) 2(n
t
1) lower-right submatrices of . Now, to recover the transmitted
symbols from the 1
st
antenna during the two timeslots, we design the following block
linear lter
=
_
_
_
I
2
BD
1
C
1
2,1
I
2nt2
_
_
_
(94)
which when applied to in (90) gives
_
_
_

1

nt1
_
_
_
=

_
_
_
(
2,1
BD
1
C)w
1
(DC
1
2,1
B)w
(nt1)
_
_
_
+ (95)
where w
(nt1)
denotes the lower (n
t
1) subvectors of w, which will be used for
the recovery of the symbols transmitted from other antennas, whereas the symbols
corresponding to the rst antenna will be recovered from
1
. Furthermore, we can
iteratively apply the same procedure to w
(nt1)
to recover the symbols transmitted
from the remaining n
t
1 antennas. For example, the transmitted vectors for a 2 3
V-BLAST link can be recovered using (95) as
_
_
_

1

2
_
_
_
=

_
_
_
(
2,1

2,2

1
3,2

3,1
)w
1
(
3,2

3,1

1
2,1

2,2
)w
2
_
_
_
+ (96)
73
5 10 15 20
10
5
10
4
10
3
10
2
10
1
SNR dB
A
B
E
R
Cpnventional IC
Suboptimal spacetime IC (95)
Optimal spacetime IC (87)
SIR = 0dB
Figure 23: Average bit error probability for 2 4 V-BLAST link in presence of
Alamouti interferer with SIR = 0dB.
7.3.2 Numerical Results
Figure 23 shows the average BER performance of 2 4 V-BLAST link employing
various IC techniques for suppressing Alamouti interference. All transmitters are
equi-powered and employ QPSK modulation. The results are generated using Monte
Carlo simulation of 10
7
channel trials. The conventional IC technique does not take
the structure of Alamouti interference into account and as a result it perceives Alam-
outi interference as comprising of two independent data streams. As a result, con-
ventional IC techniques gives far inferior BER performance when compared to the
other two space-time IC techniques, which exploit the time-correlation in interference
data streams. It is apparent that space-time IC techniques in Figure 23 oer higher
diversity order compared to conventional IC technique. The optimal space-time IC
technique in (87) oers a gain of about 1 dB over sub-optimal IC technique presented
in (95) at much higher computation cost.
74
0 5 10 15 20
10
5
10
4
10
3
10
2
10
1
10
0
SNR dB
A
B
E
R
Channel MSD,
e
2
= 1
Channel MSD,
e
2
= 0.1
Channel MSD,
e
2
= 0
n
r
= 4
n
r
= 3
SIR = 0dB
Figure 24: Average bit error probability with varying mean squared deviation
(MSD) in channel during consecutive timeslots.
The results presented in this chapter assume that the channel for the V-BLAST
link remains unchanged for the duration of two timeslots over which Alamouti in-
terference is encoded. In Figure 24, we investigate the impact of V-BLAST channel
deviation on its BER performance assuming QPSK modulation. The mean standard
deviation (MSD) in consecutive channel gain matrices is measured as

e
=
_
1
n
r
n
t
E
_
||H
1
|| ||H
2
||
_
2
(97)
From Figure 24, we note that for a 2 3 V-BLAST link, the deviation in channel
aects the link BER quite adversely. For example, to achieve an average BER = 10
2
,
an MSD = 0.1 results in 2.3 dB SNR penalty whereas, an MSD = 1 causes 5.6 dB SNR
penalty. However, when an additional receive antenna is employed, the corresponding
SNR penalties dramatically drop down to 0.15 and 0.65 dB, respectively. This implies
that the proposed space-time IC techniques can still be eective when the channel
MSD is higher as long as sucient number of receive antennas are employed.
75
7.4 Summary
In this chapter, we have developed sub-optimal space-time IC techniques for a V-
BLAST link subjected to Alamouti interference. The proposed IC techniques outper-
form conventional IC techniques that do not take the structure of OSTBC interference
into account. Our results also show that the BER of the V-BLAST link depends on
the amount of mean-squared deviation in its channel for the duration over which an
Alamouti block encoded interference is transmitted. Although the BER performance
of V-BLAST link degrades rapidly with increasing channel deviation, employing suf-
ciently large number of receive antennas can counter the BER degradation due to
increasing channel deviation.
76
CHAPTER VIII
SUMMARY AND FUTURE WORK
In the following, we summarize the primary research contributions and suggested
directions for the future work.
8.1 Research Summary
The research reported in this dissertation has looked into the challenges posed by co-
channel interference in MIMO networks and provided practical means of managing
interference at a marginal cost of system performance.
We have addressed the interference problems arising in the context of SDMA based
MIMO networks, which more eectively exploit the available network resources than
TDMA based MIMO networks. Sub-optimal techniques for joint optimization of co-
channel MIMO links were developed by applying optimal antenna selection aided
stream control algorithms. These algorithms were tested on both simulated and
measured indoor MIMO channels. We also extended the analysis to investigate the
impact of stream control on the performance of interfering MIMO links assuming
M-QAM constellations and linear MMSE receiver processor. It was shown that the
use of SDMA scheme along with partial CSI (subset of best antennas) at the trans-
mitters signicantly reduces the signaling overhead with minimal loss in throughput
performance.
The main challenge in employing stream control with optimal antenna selection
for OL-SDMA systems is the search of optimal subset of transmit antennas. Towards
this, we developed an MSE-based antenna selection framework for implementing low
complexity algorithms for nite complexity receivers operating in the presence of
co-channel interference. For either the transmitter or receiver antenna selection, two
77
sequential greedy algorithms were developed, one that starts with a full set of antennas
and decrements, and the other that starts with an empty set that increments. The
choice of which is best depends on how many antennas will ultimately be selected.
These selection algorithms were shown to provide reasonable BER performance while
keeping the overall system complexity low.
Next, we addressed the interference problems for space-time encoded transmis-
sions which can be mitigated using only one additional antenna without sacricing
space-time diversity gains. These algebraic properties of linear OSTBCs were ex-
ploited to facilitate a single-stage and MMSE optimal detector for two co-channel
users employing unity rate real and derived rate-1/2 complex OSTBCs for 3 and 4
transmit antennas. Furthermore, a single-stage multi-user interference suppression
technique is proposed for OSTBC interference; this technique exploits the temporal
and spatial structure of OSTBCs leading to simple linear processing.
8.2 Suggestions for Future Work
The goal of the reported research is to develop reliable techniques for managing co-
channel interference in MIMO networks and jointly optimizing network throughput.
Some directions for research to follow the work presented herein are:
Finer tradeo between throughput gain and CSI overhead: The CSI
feedback creates a major overhead for CL-SDMA systems which provide optimal
network throughput. However, the computational and signalling complexity as-
sociated with CL-SDMA is prohibitive. The proposed OL-SDMA with optimal
selection provides a rough tradeo between network throughput and system
complexity. New SDMA techniques based on partial CSI such as second order
channel statics can allow ner trade-o.
Algorithms for nonlinear receiver architecture: The stream control al-
ogorithms as well as antenna selection algorithms presented in this thesis are
78
limited to linear receiver processing. However, the use of non-linear processing
techniques such as successive interference cancellation (SIC) can improve the
BER performance of MIMO links.
IC using the interference power covariance matrix: The IC techniques
presented in this dissertation assume the availability of interferers channel.
However, such an assumption is well suited only for multiuser detection. New IC
techniques are required which exploit the algebraic structure of power covariance
matrix associated with the space-time interference.
Design of Cross-layer PHY/MAC protocols: The sub-optimal MIMO-
SDMA algorithms introduced in this dissertation allow multiple co-channel
MIMO links to operate simultaneously with higher throughput than with time-
multiplexing schemes. However, for users with correlated channels time-multiplexing
is preferable. New cross-layer PHY/MAC protocols are needed to facilitate this
kind of operation.
Rician and Frequency selective channels: This dissertation addressed only
the quasi-static at fading Rayleigh channel. However, in many situations the
line-of-sight (LOS) signal component is also signicant resulting in Rician fading
channel. The algorithms presented in this thesis can be analyzed for Rician
channel and extended to include frequency selective channels which form the
basis on many standards including IEEE 802.11n, 802.16e etc.
79
REFERENCES
[1] Alamouti, S. M., A simple transmit diversity technique for wireless commu-
nications, IEEE J. Select. Areas Commun., vol. 16, pp. 14511458, Oct. 1998.
[2] Barry, J. R., Lee, E. A., and Messerschmitt, D. G., Digital Communi-
cations. Kluwer Academy Publishers, third ed., 2003.
[3] Berenguer, I. and Wang, X., Mimo antenna selection with lattice reduction
aided linear receivers, IEEE Trans. Vehicular Tech., vol. 53, pp. 12891302,
Sept. 2004.
[4] Berenguer, I., Wang, X., and Wassell, I. J., Transmit antenna selec-
tion in linear receivers: Geometrical approach, IEE Electronic Letters, vol. 40,
pp. 292293, Mar. 2004.
[5] Bliss, D., Forsythe, K. W., Hero, A. O., and Yegulalp, A. F., En-
vironmental issues for mimo capacity, IEEE Trans. Signal Processing, vol. 50,
pp. 21282142, Sept. 2002.
[6] Blum, R. S., Mimo capacity with interference, IEEE J. Select. Areas Com-
mun., vol. 21, pp. 793801, June 2003.
[7] Blum, R. S., Winters, J. H., and Sollenberger, N. R., On the capacity
of cellular systems with mimo, IEEE Communication Letters, vol. 6, pp. 242
244, June 2002.
[8] Chang, J.-H., Tassiulas, L., and Rashid-Farrokhi, F., Joint transmitter
receiver diversity for ecient space division multiaccess, IEEE Transactions on
Wireless Communications, vol. 1, pp. 1627, Jan 2002.
[9] Cho, K. and Yoon, D., On the general ber expression of one- and two-
dimensional amplitude modulations, IEEE Trans. Commun., vol. 50, pp. 1074
1080, July 2002.
[10] Chung, S. T., Lozano, A., and Huang, H. C., Approaching eigenmode
BLAST channel capacity using V-BLAST with rate and power feedback, Proc.
of the IEEE Vehicular Technology Conference, vol. 2, 2001.
[11] Demirkol, M. F. and Ingram, M. A., Power-controlled capacity for inter-
fering mimo links, IEEE Vehicular Technology Conference, vol. 1, pp. 187191,
Oct. 2001.
80
[12] Demirkol, M. F. and Ingram, M. A., Control using capacity constraints
for interfering mimo links, IEEE Int. Symp. on Personal, Indoor, and Mobile
Radio Communications, vol. 3, pp. 10321036, Sept. 2002.
[13] Demirkol, M. F. and Ingram, M. A., Stream control in networks with
interfering mimo links, IEEE Wireless Communications and Networking, vol. 1,
pp. 343348, Mar. 2003.
[14] Demirkol, M. F., Resource Allocation for Interfering MIMO Links. Atlanta,
GA: Ph.D Thesis, Georgia Institute of Technology, 2003.
[15] Diggavi, S. N., Al-Dhahir, N., and Calderbank, A. R., Algebraic prop-
erties of space-time block codes in intersymbol interference multiple-access chan-
nels, IEEE Trans. Inform. Theory, vol. 49, pp. 24032414, Oct. 2003.
[16] Farrokhi, F. R., Liu, K. J. R., and Tassiulas, L., Transmit beamforming
and power control for cellular wireless systems, IEEE J. Select. Areas Commun.,
vol. 16, pp. 14371450, Oct. 1998.
[17] Foschini, G. J. and Miljanic, Z., A simple distributed autonomous power
control algorithm and its convergence, IEEE Trans. Vehicular Tech., vol. 42,
pp. 641646, Nov. 1993.
[18] Foschini, G. and Gans, M., On the limits of wireless communications in a
fading environment when using multiple antennas, Wireless Personal Commu-
nicatios, vol. 6, pp. 311335, Mar. 1998.
[19] Gharavi-Alkhansari, M. and Gershman, A. B., Fast antenna subset se-
lection in mimo systems, IEEE Trans. Signal Processing, vol. 52, pp. 339347,
Feb. 2004.
[20] Gore, D., Gorokhov, A., and Paulraj, A., Joint mmse versus v-blast
and antenna selection, in Proc. Asilomar Conference on Signals, Systems and
Computers, vol. 1, pp. 505509, Nov. 2002.
[21] Gore, D. A., Heath, R. W., and Paulraj, A., Transmit selection in spatial
multiplexing systems, IEEE Communications Letters, vol. 6, pp. 491493, Nov.
2002.
[22] Gorokhov, A., Antenna selection algorithms for mea transmission systems,
in Proc. IEEE ICASSP, pp. 28572860, May 2002.
[23] Gorokhov, A., Collados, M., Gore, D., and Paulraj, A., Trans-
mit/receive mimo antenna selection, in Proc. IEEE ICASSP, vol. 2, pp. 1316,
May 2004.
[24] Gorokhov, A., Gore, D. A., and Paulraj, A. J., Receive antenna se-
lection for mimo at-fading channels: Theory and algorithms, IEEE Trans.
Inform. Theory, vol. 49, pp. 26672696, Oct. 2003.
81
[25] Heath, R. and Love, D., Multimode antenna selection for spatial multi-
plexing systems with linear receivers, IEEE Trans. Signal Processing, vol. 53,
pp. 30423056, Aug. 2005.
[26] Jafarkhani, H., A quasi-orthogonal space-time block code, IEEE Trans.
Commun., vol. 49, pp. 14, Jan. 2001.
[27] Jensen, M. and Morris, M., Ecient capacity-based antenna selection for
mimo systems, IEEE Trans. Vehicular Tech., vol. 54, pp. 110116, Jan. 2005.
[28] Jiang, J.-S. and Ingram, M., Comparison of beam selection and antenna
selection techniques in indoor mimo systems at 5.8 ghz, in Proc. of Radio and
Wireless Conference, pp. 179182, Aug. 2003.
[29] Kazemitabar, J. and Jafarkhani, H., Multiuser interference cancellation
and detection for users with four transmit antennas, in IEEE International
Symposium of Information Theory, ISIT, pp. 19141918, 2006.
[30] Klang, G. and Ottersten, B., Space-time interference rejection cancella-
tion in transmit diversity systems, The 5th International Conference on Wire-
less Personal Multimedia Communications, vol. 2, pp. 706710, Oct. 2002.
[31] Klang, G., On Interference Rejection in Wireless Multichannel Systems. Stock-
holm, Sweden: Ph.D Thesis, Royal Institute of Technology, 2003.
[32] Larsson, E. G. and Stoica, P., Space-Time Block Coding for Wireless Com-
munications. Cambridge, UK: Cambridge University Press, 2003.
[33] Lin, D., Sfar, S., and Letaief, K. B., Receive antenna selection for mimo
systems in correlated channels, in Proc. IEEE International Conf. on Comm,
vol. 5, pp. 29442948, June 2004.
[34] Lotkepohl, H., Handbook of Matrices. New York: Wiley, 1996.
[35] Molisch, A., Win, M., and Winters, J., Capacity of mimo systems with
antenna selection, in Proc. IEEE Int. Conf. Commun. (ICC01), (Helsinki, Fin-
land), pp. 570574, June 2001.
[36] Nabar, R., Gore, D., and Paulraj, A., Optimal selection and use of trans-
mit antennas in wireless systems, in Proc. International Conference on Telecom-
munications (ICT00), (Acapulco, Mexico), May 2000.
[37] Naguib, A. F., Seshadri, N., and Calderbank, A. R., Applications of
space-time block codes and interference suppression for high capacity and high
data rate wireless systems, in Proc. 32nd Asilomar Conf. Signals, Syst., Com-
put., pp. 18031810, 1998.
[38] Nam, S. H. and Lee, K. B., Transmit power allocation for an extended V-
BLAST system, Proc. of the IEEE Int. Symp. on Personal, Indoor and Mobile
Radio Communications, sep 2002.
82
[39] Okazaki, A., Fujikawa, M., Hasegawa, K., and Sasase, I., Soft deci-
sion iterative v-blast with detection of all symbols at the last stage by changing
detection order in hybrid stbc, in Proc. IEEE Pacic Rim Conference on Com-
munications, Computers and Signal Processing, pp. 194197, Aug. 2007.
[40] Park, J. and Park, D., A new antenna selection algorithm with low com-
plexity for mimo wireless systems, in IEEE International Conference on Com-
munications (ICC2005), vol. 4, p. 2308, May 2005.
[41] Runhua, C., Andews, J., and Health, R., Multiuser space-time block
coded mimo with downlink precoding, IEEE International Conference on Com-
munications, vol. 5, pp. 26892693, June 2004.
[42] Sampath, H., Stoica, P., and Paulraj, A., Generalized linear precoder
and decoder design for mimo channels using the weighted mmse criterion, IEEE
Trans. Commun., vol. 49, pp. 21982206, Dec. 2001.
[43] Sanayei, G. S., Antenna selection in mimo systems, IEEE Commun. Mag.,
vol. 42, pp. 6873, Oct. 2004.
[44] Sandhu, S., Nabar, R. U., Gore, D. A., and Paulraj, A., Selecting an
optimal set of transmit antennas for a mimo channel based on shannon capacity,
in Proc. Asilomar Conf. Signals, Syst., Comput., (Pacic Grove, CA), pp. 567
571, Nov. 2000.
[45] Stamoulis, A., Al-Dhahir, N., and Calderbank, A. R., Further results
on interference cancellation and space-time block codes, in Proc. Asilomar Conf.
Signals, Syst., Comput., pp. 257262, Oct. 2001.
[46] Sun, C. and Karmakar, N. C., Combining beamforming with mmse alam-
outi multiuser interference cancellation receiver, in Proc. Fourth IEEE Inter-
national Symposium on Signal Processing and Information Technology, pp. 254
257, Dec. 2004.
[47] Sundaresan, K., Sivakumar, R., Ingram, M. A., and Chang, T.-Y., A
fair medium access control protocol for ad-hoc networks with mimo links, IEEE
Trans. Mobile Comput., vol. 3, pp. 350365, Dec. 2004.
[48] Tarokh, V., Jafarkhani, H., and Calderbank, A. R., Space-time block
codes from orthogonal designs, IEEE Trans. Inform. Theory, vol. 45, pp. 1456
1467, July 1999.
[49] Telatar, I., Capacity of multi-antenna gaussian channels, AT&T Bell Labs,
Technical Report, pp. 311335, June 1995.
[50] Tse, D. and Viswanath, P., Fundamentals of Wireless Communication. New
York: Cambridge University Press, 2005.
83
[51] Winters, J. H., Salz, J., and Gitlin, R. D., The impact of antenna diver-
sity on the capacity of wireless communication systems, IEEE Trans. Commun.,
vol. 42, pp. 17401751, Feb./Mar./Apr. 1994.
[52] Wolniansky, P. W., Foschini, G. J., Golden, G. D., and Valenzuela,
R. A., V-blast: an architecture for realizing very high data rates over the rich-
scattering wireless channel, URSI International Symposium on Signals, Sys-
tems, and Electronics, pp. 295300, Sept. 1998.
[53] Younis, W. M., Sayed, A. H., and Al-Dhahir, N., Ecient adaptive
receivers for joint equalization and interference cancellation in multiuser space-
time block-coded systems, IEEE Trans. Signal Processing, vol. 51, pp. 2849
2862, Nov. 2003.
[54] Zhang, H., Dai, H., and Zhou, Q., A geometrical analysis on transmit
antenna selection for spatial multiplexing systems with linear receivers, in Proc.
International Symposium on Information Theory, pp. 15581562, Sept. 2005.
[55] Zhang, L., Li, S., Zheng, H., and Wu, M., A selective co-channel interfer-
ence mitigation method for alamouti code, in Proc. 11th IEEE Symposium on
Computers and Communications, pp. 149154, June 2006.
[56] Zheng, L. and Tse, D., Diversity and multiplexing: a fundamental tradeo in
multiple-antenna channels, IEEE Transactions on Information Theory, vol. 49,
pp. 10731096, may 2003.
84
VITA
Sudhanshu Gaur was born in 1979 in Meerut, India. He received Bachelor of Tech-
nology degree in Instrumentation Engineering from Indian Institute of Technology,
Kharagpur in 2000. Afterwards, he worked with the 2.5G Communications Group of
Sasken Communication Technologies, Bangalore. In Fall 2001, he joined the Mobile
and Portable Radio Research Group (MPRG) at Virginia Tech for graduate studies.
After completing his M.S. in Electrical Engineering in Fall 2003, he joined the Smart
Antenna Research Laboratory (SARL) at Georgia Tech to pursue his PhD. In Fall
2005, he joined the Wireless Systems Research Laboratory (WSRL) at Hitachi. His
research interests include MIMO signal processing, space-time coding, interference
management, resource allocation, random processes and number theory.
85