Spatial Multiplexing in Multiuser Networks With Limited Feedback - 1 PDF

SPATIAL MULTIPLEXING IN MULTIUSER NETWORKS WITH
LIMITED FEEDBACK
DISS . ETH NO.
18307
SPATIAL MULTIPLEXING IN MULTIUSER

NETWORKS WITH LIMITED FEEDBACK
A dissertation submitted to
ETH ZURICH
for the degree of
Doctor of Sciences
presented by
JATIN THUKRAL
M.Sc., ETH Zurich
born 10.02.1982
citizen of India
accepted on the recommendation of
Prof. Dr. Helmut B
olcskei, examiner
Prof. Dr. David Gesbert, coexaminer
2009
Abstract
The recent increase in demand for high data-rate wireless applications
has resulted in a need for wireless communication systems with high
spectral efficiencies. One of the most promising ways to achieve such
high spectral efficiencies is a transmission technique called spatial
multiplexing. Over the recent years, spatial multiplexing has been
extensively studied for single-user wireless systems as well as for multiuser wireless networks such as broadcast networks and interference
networks. In particular, it has been shown that achievability of full
spatial multiplexing gain in both broadcast and interference networks
depends critically on the channel knowledge available at the sources.
Since sources often obtain channel knowledge through limited capacity
feedback links, it is important to develop transmission schemes that
can achieve full spatial multiplexing gain in broadcast and interference networks with limited feedback. In this thesis, we consider three
important classes of interference networks and a class of broadcast
networks, and propose, for each class, a (limited capacity) feedback
and transmission scheme that achieves full spatial multiplexing gain.
We first consider a large interference network that models the
second hop of the multiuser relaying protocol proposed by Bolcskei
et al. (2006). The results in B
olcskei et al. (2006) imply that full
spatial multiplexing gain is achievable in this network, provided that
each source has perfect knowledge of its channel to its corresponding
destination. We show that full spatial multiplexing gain is achievable
even if the perfect channel knowledge assumption is relaxed to having
an error-free 1-bit broadcast feedback link from each destination
to its corresponding sources. Specifically, we provide an iterative

feedback-based (multiuser) beamforming algorithm and show that
spatial multiplexing based on the beamforming weights produced
by the algorithm achieves full spatial multiplexing gain, provided
the training period length of the algorithm scales linearly with the
number of sources in the network. We furthermore show that the
iterative algorithm is asymptotically optimal in the sense that full
spatial multiplexing gain cannot be achieved with training period
length scaling sublinearly with the number of sources.
We second consider an M source-destination pair single-antenna
interference network where all channels are frequency-selective with L
taps and each source has an average transmit power of P/M . Recent
results by Cadambe and Jafar (2008) and Grokop et al. (2009) imply
that full spatial multiplexing gain is achievable in this network with
a novel transmission scheme called interference alignment, provided
that each node (that is, each source and each destination) knows
all the channels in the network perfectly. We show that full spatial
multiplexing gain is achievable with interference alignment, even if
the perfect channel knowledge assumption is relaxed to having an
error-free broadcast feedback link of rate M (L 1) log P from each
destination to all the nodes in the network.
We then consider an M source-destination pair interference network
where each source is equipped with at least M antennas and has an
average transmit power of P/M , and each destination is equipped
with a single antenna. Assuming that all the channels in the network
are frequency-selective with L taps, we provide a feedback and transmission scheme, which we term spatial interference-nulling, and show
that the scheme achieves full spatial multiplexing gain of M , provided
that each destination can feedback to its non-corresponding sources
at the rate of (M 1)(M L 1) log P .
We finally consider broadcast networks where a source, equipped
with M antennas and having an average transmit power of P , communicates with K (where K M ) non-cooperating single-antenna
destinations and all channels are frequency-selective with L taps. Yoo
et al. (2007) showed that full spatial multiplexing gain is achievable in
vi
such a network for L = 1, provided that each destination can feedback

to the source at the rate of (M 1) log P log K. We generalize the
results of Yoo et al. (2007) to the L > 1 case and provide a feedback
and transmission scheme that achieves full spatial multiplexing gain
of M , under the assumption that at least M of the K destinations
can feedback to the source at the rate of (M L 1) log P log K.
vii
Kurzfassung
Die j
ungst steigende Nachfrage nach drahtlosen Geraten mit hoher
Datenrate hat zu einem Bedarf an drahtlosen Kommunikationssystemen mit hoher spektraler Effizienz gef
uhrt. R
aumliches Multiplexen ist
eine Ubertragungsart, welche eine der vielversprechendsten Methoden

darstellt, solch eine hohe spektrale Effizienz zu erreichen. Wahrend
der vergangenen Jahre wurde das r
aumliche Multiplexen sowohl f
ur
drahtlose Systeme mit einem Nutzer als auch f
ur drahtlose Netzwerke
mit mehreren Nutzern wie zum Beispiel Rundfunk- und Interferenznetzwerke eingehend untersucht. Es wurde gezeigt, dass die Erzielbarkeit
des vollen raumlichen Multiplexgewinns sowohl in Rundfunk- als
auch Interferenznetzwerken entscheidend von der an den Sendern
verf
ugbaren Kanalkenntnis abh
angt. Da Sender die Kanalkenntnis oft
durch R
uckkan
ale mit begrenzter Kapazit
at erhalten, ist es wichtig,
raumliche Multiplexverfahren zu entwickeln, die den vollen dadurch
moglichen Gewinn in Rundfunk- und Interferenznetzwerken erreichen
konnen. In dieser Dissertation betrachten wir drei wichtige Klassen von
Interferenz- und eine Klasse von Rundfunknetzwerken und schlagen
f
ur jede dieser Klassen ein R
uckmeldungs- und Ubertragungsverfahren (mit begrenzter Kapazit
at) vor, das den vollen raumlichen
Multiplexgewinn erzielt.
Zuerst betrachten wir ein grosses Interferenznetzwerk, das den
zweiten Ubertragungsteil
des von B
olcskei et al. (2006) vorgeschlagenen Relaisprotokolls f
ur mehrere Nutzer darstellt. Die Resultate
in Bolcskei et al. (2006) implizieren, dass der volle raumliche Multiplexgewinn in diesem Netzwerk erzielbar ist, gegeben, dass jeder
ix
Sender u
ber vollkommene Kenntnis des Kanals zum zugehorigen
Empfanger verf
ugt. Wir zeigen, dass der volle raumliche Multiplex
gewinn selbst dann erzielbar ist, wenn die Annahme, die Kanale vollkommen zu kennen, dahingehend gelockert wird, einen fehlerfreien
1-bit R
uckkanal von jedem Empf
anger zu seinen zugehorigen Sendern
zu haben. Insbesondere stellen wir einen schrittweise arbeitenden
Algorithmus zur Strahlausrichtung (f
ur mehre Nutzer) vor, der auf
R
uckinformation basiert. Wir zeigen, dass r
aumliches Multiplexen,
welches auf den Strahlausrichtungsgewichten, die vom Algorithmus
erzeugt werden, basiert, den vollen r
aumlichen Multiplexgewinn erzielt,
gegeben, dass die Dauer der Lernperiode des Algorithmus linear mit
der Anzahl der Sender im Netzwerk w
achst. Dar
uber hinaus zeigen
wir, dass der schrittweise arbeitende Algorithmus in der Hinsicht
asymptotisch optimal ist, dass der volle r
aumliche Multiplexgewinn
nicht mit einer Lernperiode erreicht werden kann, die sublinear mit
der Anzahl von Sender anw
achst.
Als Zweites betrachten wir ein Netzwerk, das aus M Sender /
Empfanger-Paaren mit jeweils einer Antenne besteht, in dem alle
Kanale frequenzselektiv mit Impulsantworten der Lange L sind und
jeder Empfanger eine durchschnittliche Sendeleistung von P/M hat.
K
urzlich publizierte Resultate von Cadambe and Jafar (2008) und
Grokop et al. (2009) implizieren, dass in diesem Netzwerk der volle
raumliche Multiplexgewinn mit einem neuartigen Ubertragungsschema, der sogenannten Interferenzausrichtung, erzielbar ist. Damit
dies moglich ist, muss jeder Knoten (das heisst, jeder Sender und jeder
Empfanger) alle Kan
ale im Netzwerk vollst
andig kennen. Wir zeigen,
dass der volle r
aumliche Multiplexgewinn selbst dann erzielbar ist,
wenn die Annahme, die Kan
ale vollkommen zu kennen, dahingehend
gelockert wird, einen fehlerfreien Rundfunkr

uckkanal mit Ubertragungsrate M (L 1) log P von jedem Empf
anger zu allen Knoten im
Netzwerk zu haben.
Anschliessend betrachten wir ein Interferenznetzwerk, das aus M
Sender/Empfanger-Paaren besteht, in dem jeder Sender mit mindestens M Antennen ausgestattet ist und eine durchschnittliche
Sendeleistung von P/M hat. Jeder Empf
anger ist mit einer Antenne
ausgestattet. Unter der Annahme, dass alle Kanale in diesem Netzwerk frequenzselektiv mit einer Impulsantwort der Lange L sind,
geben wir ein R

uckmeldungs- und Ubertragungsverfahren
an, welches
wir als raumliche Interferenzausl
oschung bezeichnen. Unter der Voraussetzung, dass jeder Empf
anger Information zu allen nicht zu ihm
gehorenden Sendern mit einer Datenrate von (M 1)(M L 1) log P
r
uckf
uhren kann, erzielt das Verfahren den vollen raumlichen Multiplexgewinn der Gr
osse M .
Zum Schluss betrachten wir Rundfunknetzwerke, in denen ein
Sender, der mit M Antennen ausgestattet ist und u
ber eine durchschnittliche Sendeleistung von P verf
ugt, mit K (wobei K M )
nichtkooperierenden Empf
angern, die mit jeweils einer Antenne ausgestattet sind, kommuniziert. Alle Kan
ale seien frequenzselektiv mit
Impulsantworten der L
ange L. Yoo et al. (2007) zeigten, dass der volle
raumliche Multiplexgewinn in einem solchen Netzwerk mit L = 1 unter
der Voraussetzung erzielbar ist, dass jeder Empfanger Information
zum Sender mit einer Datenrate von (M 1) log P log K r
uckf
uhren
kann. Wir verallgemeinern die Resultate von Yoo et al. (2007) f
ur den
Fall, dass L > 1 und geben ein R

uckmeldungs- und Ubertragungsverfahren an, das den vollen r
aumlichen Multiplexgewinn der Grosse
M erzielt, gegeben, dass mindestens M der K Empfanger Information zum Sender mit einer Datenrate von (M L 1) log P log K
r
uckf
uhren konnen.
xi
Acknowledgments
As far as learning is concerned, the last four years constitute the
most productive period of my life, so far. The lessons of patience,
self-discipline, paying attention to details, professional ethics and a
confidence to question conventional wisdom that I have developed
over these years, form an integral part of my personality today. And
no other person deserves more credit for this learning than Prof. Dr.
Helmut Bolcskei. Without his guidance and support, neither this
learning nor this dissertation would have been possible. In particular,
I can never forget his painstakingly going through each and every line
of our Allerton paper and converting it from a mathematical mess
that only I understood to a piece of research that any person with
suitable background can understand. I am also extremely thankful to
him for his numerous teachings, including his belief that it is better to
have one great contribution than to have ten mediocre contributions.
An old saying in my home country is that the best way to thank
a teacher is to follow his teachings. I therefore hope to follow Prof.
Bolcskeis teachings for the rest of my life.
I would also like to thank the members of the communication
theory group, whose support was critical to the accomplishment of
this thesis. In particular, I would like to thank Dr. Ulrich Schuster
and Dr. Giuseppe Durisi for teaching me the importance of being
precise in my formulations. I would like to thank Cemal Akcaba for his
friendship and support as a colleague throughout the last four years. I
am grateful to Veniamin Morgenshtern for all the diverse discussions
we had during our dinners at Mensa, on topics ranging from biology
xiii
and quantum physics to intelligence and the future of human society.

I am thankful to Dr. Pedro Coronel and Patrick Kuppinger, whose
disarming ingenuity makes them not only wonderful colleagues but
also great friends. I would moreover like to express my appreciation
for Michael Lerjen, Davide Cescato and Dr. Dominik Seethaler for
their support.
In addition, I would like to thank my friends Shubha Pandey,
Rajneesh Verma, Mayank Mittal, Sidharth Jha, Harish Kaushik and
Shilpi Minocha, whose presence in my life turned the last four years
into such a wonderful experience.
Finally and most importantly, I would like to thank my family,
whose unconditional support allows me to undertake challenges far
beyond my capacity.
Jatin Thukral
April 29, 2009
xiv
Contents
1. Introduction and Outline

1.1. Interference networks . . . . . . . . . . . . . . .
1.1.1. Large network analysis approach . . . .
1.1.2. Large transmit power analysis approach
1.2. Broadcast networks . . . . . . . . . . . . . . . .
1.3. Outline of the thesis . . . . . . . . . . . . . . .
.
.
.
.
.
1
2
3
5
7
11
2. Large Interference Networks with 1-bit feedback

2.1. System and signal model . . . . . . . . . . . . . . . . .
2.2. Upper bound on sum capacity . . . . . . . . . . . . . .
2.3. Iterative distributed beamforming . . . . . . . . . . .
2.4. Convergence of the iterative algorithm . . . . . . . . .
2.4.1. Convergence in probability to -level . . . . . .
2.4.2. Convergence in expectation to -level . . . . . .
2.4.3. Numerical results . . . . . . . . . . . . . . . . .
2.5. Achievability of full spatial multiplexing gain . . . . .
2.6. Training period length scaling sub-linearly in K . . . .
2.7. Extension to complex-valued channels . . . . . . . . .
2.7.1. Modified iterative beamforming algorithm . . .
2.7.2. Achievability of full spatial multiplexing gain .
2.7.3. Training period length scaling sub-linearly in K
2.8. Observations . . . . . . . . . . . . . . . . . . . . . . .
13
16
18
20
24
26
33
39
42
47
55
55
56
61
62
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xv
CONTENTS
3. SISO Interference Networks with Broadcast Feedback

3.1. System model . . . . . . . . . . . . . . . . . . . . . . .
3.2. Interference alignment with perfect CSI at all nodes .
3.3. Interference alignment with limited feedback . . . . .
3.3.1. The vector quantization and feedback scheme .
3.3.2. Transmission scheme and achievability of full
spatial multiplexing gain . . . . . . . . . . . . .
3.4. Extension to MIMO interference networks . . . . . . .
3.5. Observations . . . . . . . . . . . . . . . . . . . . . . .
65
68
70
74
74
77
82
83
4. MISO Interference Networks

4.1. System model . . . . . . . . . . . . . . . . . . . . . .
4.2. The frequency-flat case . . . . . . . . . . . . . . . . .
4.2.1. SIN with perfect transmit CSI . . . . . . . .
4.2.2. The quantization and feedback scheme . . . .
4.2.3. The transmission scheme . . . . . . . . . . .
4.2.4. Achievability of full spatial multiplexing gain
4.3. The frequency-selective case . . . . . . . . . . . . . .
4.3.1. Spatial interference-nulling in time-domain .
4.3.2. Frequency-domain interpretation of SIN . . .
4.3.3. Feedback design criterion . . . . . . . . . . .
4.3.4. The feedback scheme . . . . . . . . . . . . . .
4.3.5. Computation of SIN vectors . . . . . . . . . .
4.4. Observations . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
85
87
89
90
92
93
94
96
97
99
101
103
104
107
5. MISO Broadcast Networks with Shared Feedback

5.1. System model . . . . . . . . . . . . . . . . . . . . . .
5.2. The K = M case . . . . . . . . . . . . . . . . . . . .
5.2.1. Feedback scheme . . . . . . . . . . . . . . . .
5.2.2. Transmission scheme . . . . . . . . . . . . . .
5.3. The K M case . . . . . . . . . . . . . . . . . . . .
5.3.1. Quantization scheme . . . . . . . . . . . . . .
5.3.2. Iterative feedback scheme . . . . . . . . . . .
5.4. Observations . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
109
112
114
115
115
116
117
124
127
129
xvi
CONTENTS
6. Conclusions and Outlook
131
Appendices
134
A. Bounds and Proofs

A.1. Union bounds . . . . . . . . . . . . . . . . .
A.2. Large deviations bounds . . . . . . . . . . .
A.3. Intersection bound . . . . . . . . . . . . . .
A.4. Interference alignment with perfect CSI . .
A.5. Linear independence of vkm , m . . . . . . .
A.6. Line packing bound . . . . . . . . . . . . .
A.7. Surface area of M -dimensional hyper-sphere
A.8. Area of a spherical cap . . . . . . . . . . . .
A.9. Non-overlap of spherical caps . . . . . . . .
A.10.Proof of Property 3 . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
135
135
135
137
138
142
144
145
148
150
151
B. Wireless Systems and Channel Feedback: A primer

B.1. Single-user system . . . . . . . . . . . . . . .
B.1.1. Single-input single-output channel . .
B.1.2. Multi-input multi-output channel . . .
B.2. Multiuser networks . . . . . . . . . . . . . . .
B.2.1. Multi-access networks . . . . . . . . .
B.2.2. Broadcast networks . . . . . . . . . .
B.2.3. Interference networks . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
155
155
156
159
161
162
163
167
.
.
.
.
173
173
174
176
176
C. Notation
C.1. Miscellaneous . . . . . . .
C.2. Linear Algebra . . . . . .
C.3. Probability and Statistics
C.4. Information Theory . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
D. Acronyms
177
References
179
xvii
CONTENTS
Curriculum Vitae
xviii
185
C HAPTER
Introduction and Outline

he recent increase in demand for high data-rate wireless
applications has resulted in a need for wireless communication
systems with high spectral efficiencies. A promising approach
to achieve such high spectral efficiencies is a transmission technique
called spatial multiplexing, which can be intuitively understood as
a simultaneous transmission of independent messages on the same
frequency band. A key aspect of any spatial multiplexing scheme is
the presence of multiple source and destination antennas, which may
be cooperating or non-cooperating. If both source antennas as well
as destination antennas in a wireless system are fully cooperating,
the system is modeled in information theory as a single-user system.
Telatar (1999) showed that if all the channels in a single-user system
are Rayleigh-fading, full spatial multiplexing gain (that is, maximum
pre-log term in capacity expression) can be achieved irrespective of the
availability of channel knowledge at the sources. If the source antennas
or the destination antennas or both are non-cooperating, the system is
most efficiently modeled as a multiuser network. In contrast to a singleuser system, achieving full spatial multiplexing gain in two important
classes of multiuser networks, namely, interference networks and
broadcast networks, depends critically on the availability of channel
state information (CSI) at the sources. Assuming that the sources
obtain CSI through limited capacity feedback links, our focus in this
thesis is on the feedback and transmission schemes that achieve full
1.
INTRODUCTION AND OUTLINE
spatial multiplexing gains in these two classes of multiuser networks.
1.1.
I NTER FER ENC E N ETWOR KS
General interference networks, as defined by Carleial (1978), Gamal

and Cover (1980) and van der Meulen (1994), consist of M sourcedestination pairs such that each source communicates information
only to its corresponding destination and does not care about what
the other destinations receive or understand. There is no cooperation
between the sources or the destinations; and hence, transmission
from each source creates interference at all the destinations that
do not intend to receive information from it. Such a situation occurs frequently in communications and is characterized by M principal links between the corresponding source-destination pairs and
M (M 1) interference links. Since there are only M communication
rates of direct interest, the corresponding capacity region is naturally defined as the M -dimensional closure of the set of rate vectors
R = {R1 , R2 , . . . , RM } for which jointly reliable communications are
possible over the M principal links, with independent information
symbols at the sources. While a network of this type was already
considered by Shannon (1961) in the context of his analysis of the
two-way channel, the capacity region is unfortunately not known till
now even for the Gaussian case. Nevertheless, some intricate results
have been reported in the literature regarding the achievable rate
regions for interference networks under different assumptions. For the
general two-user discrete memoryless interference network, Ahlswede
(1971), Sato (1977) and Carleial (1978) obtained inner bounds on the
capacity region, while Sato (1977) obtained a useful outer bound as
well. Han and Kobayashi (1981) provided a new achievable rate region
for the general two-user discrete memoryless interference network,
which includes the achievable rate regions established in Ahlswede
(1971), Sato (1977) and Carleial (1978). For the Gaussian interference
network, Carleial (1975) showed interestingly that in case of very
strong interference, the capacity region is the same as if there were no
1.1.
INTERFERENCE NETWORKS
interference whatsoever. Carleial (1978) also established an achievable

rate region for the Gaussian case. Certain useful bounds were given
by Sato (1978a, 1981), Carleial (1983) and Costa (1985). The state of
the art results didnt change following the work by Costa (1985) for
the next decade and a half. Thereafter, Kramer (2001, 2004) provided
two outer bounds on the capacity region, again for the two-user case.
The first bound unified and improved the bounds of Sato and Carleial,
while the second bound was a direct extension of Costas and Satos
bound and was shown to possess certain optimality properties for
weak interference. More recently, Tuninetti (2007) studied cooperative
communication strategies for interference networks with generalized
feedback and Etkin et al. (2008) characterized the capacity of the two
user interference network up to an accuracy of 1-bit for all channel
parameters. A limitation common to most of the above results is that
they do not easily extend to more general settings encountered in
the current wireless systems and therefore cannot be used by system
designers to derive practical guidelines for better spectral efficiency.
Consequently, the focus of research has shifted over the last few years
to (more tractable) asymptotic analysis of the capacity region, either
in the limit of large number of nodes (that is, source and destinations)
in the network or in the limit of large sum (across all sources) transmit
power P . This shift has resulted in two distinct approaches being
followed:
1.1.1. Large network analysis approach

This approach was initiated by Gupta and Kumar (2002), who let
the number of source-destination pairs go to infinity and established
a capacity scaling law governing the behavior of the sum capacity
of the network with respect to the number of nodes in the network.
Their work resulted in a flurry of further such remarkable capacity
scaling results over the last few years (see Xie and Kumar (2004),
Jovicic et al. (2004), Gastpar and Vetterli (2005), Leveque and Telatar
(2005), Ahmad et al. (2006), Dana and Hassibi (2006), Bolcskei et al.
ur et al. (2006), Morgenshtern and Bolcskei (2007) and
(2006), Oyg
1.
ur et al. (2007)). Most of these scaling results however assume

Oyg
either perfect CSI at the sources (this requires infinite feedback) or an
inherent path loss component in the channel models. To the best of our
knowledge, capacity scaling laws for large interference networks with
no significant path loss in their channels have not been investigated
for limited feedback scenarios. The first major contribution of this
thesis is to determine a capacity scaling law that governs, under the
assumption of limited feedback, the sum capacity behavior of a special
class of large interference networks, as discussed below.
Large interference networks with 1-bit feedback
Bolcskei et al. (2006) proposed a multiuser relaying protocol for a
network where M single-antenna sources S1 , , SM , communicate
with M single-antenna destinations D1 , , DM , with the help of
M K non-cooperating single-antenna relays. The relays are divided
into M groups G1 , , GM of K relays each, such that the group Gi
is dedicated to facilitate transmission between the source-destination
pair {Si , Di }. The K relays in the group Gi are denoted by Si1 , , SiK .
Transmission from Si to Di occurs over two hops. During the first
hop, each Si transmits its message to all the relays Sik , and each Di
stays silent. During the second hop, each relay Sik forwards, after appropriate processing, its received symbol to Di with all Si being silent.
Under the assumption of perfect knowledge of both channel Si Sik
and channel Sik Di at each Sik , B
olcskei et al. (2006) showed that
for M finite and K , full spatial multiplexing gain of M is
achievable in the sum rate of the network. While perfect knowledge of
channel Si Sik at the relay Sik is a reasonable assumption (since Sik
is the destination for the link Si Sik ), perfect knowledge of channel
Sik Di at the relay Sik requires infinite-capacity feedback links from
each destination Di to the relays Sik , k. It is therefore critical to
develop a feedback and transmission scheme that achieves full spatial
multiplexing gain in the second hop with limited-capacity feedback
links from each destination Di to the relays Sik , k. In fact, assuming the first hop transmission to be error-free, we indeed propose in
1.1.
INTERFERENCE NETWORKS
Chapter 2 a limited-feedback based iterative (multiuser) beamforming

algorithm for the second hop and show that spatial multiplexing based
on the beamforming weights produced by the algorithm results in full
spatial multiplexing gain of M . The algorithm requires a broadcast
feedback of 1-bit from each destination Di to the relays Sik , k, and
the training period length of the algorithm scales linearly in K.
1.1.2. Large transmit power analysis approach

This approach entails investigation of capacity scaling laws governing
the behavior of the capacity region of an interference network with
respect to the sum (across all sources) transmit power P , in the
limit of P going to infinity. Host-Madsen and Nosratinia (2005) and
Cadambe and Jafar (2008) considered a single-antenna M sourcedestination pair interference network and showed that, in the large P
regime, spatial multiplexing gain is upper-bounded by M/2. Cadambe
and Jafar (2008) also presented a novel transmission scheme, called
interference alignment, and showed that the scheme achieves full
spatial multipexing gain of M/2, provided that all the channels in
the network are time-selective (also see Gou and Jafar (2009) and
Gomadam et al. (2009)). Generalization of interference alignment to
large frequency-selective interference networks was investigated by
Grokop et al. (2009). All these results however assume each node
(that is, each source and each destination) to have perfect CSI of all
the channels in the network. The second major contribution of this
thesis is to analyze the spatial multiplexing gain achievable, in the
large P regime, in the interference networks where the sources have
only partial CSI, obtained through limited capacity feedback links.
Specifically, we consider two distinct classes of frequency-selective
interference networks, namely, single-input single-output (SISO) interference networks and multi-input single-output (MISO) interference
networks, and demonstrate that a channel vector quantization scheme
proposed by Mukkavilli et al. (2003) can be employed to achieve full
spatial multiplexing gain in both the classes of interference networks,
as discussed below.
1.
SISO interference networks with broadcast feedback

SISO interference networks refer to the interference networks where
M single-antenna sources communicate with M single-antenna destinations concurrently and in the same frequency band, and each source
has an average transmit power of P/M . Cadambe and Jafar (2008)
showed that performing interference alignment in a SISO interference
network operating over time-selective channels results in full spatial
multiplexing gain, provided that each source and each destination has
perfect CSI of all the channels in the network. In this thesis, we show
that full spatial multiplexing gain is achievable even with partial CSI
at the sources and the destinations, obtained through limited capacity
broadcast feedback links. Specifically, assuming that all channels in
the network are time-flat but frequency-selective with L taps, we
show that naive interference alignment based on vector-quantized
impulse responses, employing the vector quantization scheme proposed by Mukkavilli et al. (2003) and Love et al. (2003), achieves full
spatial multiplexing gain of M/2, provided that each destination can
broadcast at least M (L 1) log P bits in an error-free fashion to all
the sources and destinations in the network. Conceptually this result
is shown to amount to the notion of rather than aligning interference perfectly by creating completely interference-free dimensions, it
suffices to ensure that, as the sum transmit power P increases, the
interference power in these dimensions remains bounded. Our results
hold for multi-input multi-output (MIMO) frequency-selective interference networks where each source and each destination is equipped
with an equal number of antennas.
MISO interference networks

MISO interference networks refer to a special class of MIMO interference networks where each source is equipped with multiple antennas
and each destination is equipped with a single antenna. While such
networks have been studied by Karakayali et al. (2006) under the
assumption of all sources being partially cooperating, feasibility of
1.2.
BROADCAST NETWORKS
spatial multiplexing in such networks has not been investigated for

the case where sources are not cooperating at all and where each
source obtains its CSI through limited capacity feedback links. In
this thesis, we consider a MISO interference network where M noncooperating sources, each equipped with at least M antennas and
having an average transmit power of P/M , communicate with M
single-antenna destinations, and all the channels in the network are
frequency-selective with L taps. Assuming that each source obtains its
CSI from the destinations through limited capacity feedback links, we
propose a feedback and transmission scheme, which we term spatial
interference-nulling (SIN), and show that this scheme, along with
the channel vector quantization scheme proposed by Mukkavilli et al.
(2003) and Love et al. (2003), results in full spatial multiplexing gain
of M , provided the number of feedback bits from each destination to
its non-corresponding sources scales with the sum (across all sources)
transmit power P as (M 1)(M L 1) log P . In contrast to interference alignment where both coding length and the number of taps L
in channel impulse responses need to be large for achievability of full
spatial multiplexing gain, spatial interference-nulling, by employing
multiple antennas at each source, achieves full spatial multiplexing
gain without any such constraints.
1.2.
B ROADCAST N ETWOR KS
A broadcast network consists of a single source and K non-cooperating

destinations such that the source has an independent message for
each of the K destinations. There are K links in the network and
the capacity region is defined as the K-dimensional closure of the
set of rate vectors R = {R1 , R2 , . . . , RK } for which jointly reliable
communications are possible over the K links. Although broadcast
networks were introduced by Cover (1972) almost four decades ago
and some upper bounds on its achievable sum rate were given by
Sato (1978b) and Marton (1979) a few years later, the capacity region
for general broadcast networks is still not known. The special case
1.
of MISO Gaussian broadcast networks where the source is equipped

with two antennas, the number of single-antenna destinations is two
and both the source and the two destinations have perfect CSI, was
solved for the sum capacity by Caire and Shamai (1999), who showed
that full spatial multiplexing gain of two can be achieved. Their result
was extended by Yu and Cioffi (2001), Viswanath and Tse (2003) and
Vishwanath et al. (2003) to the case where the source is equipped with
M antennas and the number of single-antenna destinations is M , and
it was established that the full spatial multiplexing gain of M can be
achieved, provided the source has perfect CSI of all the channels in the
network. The complete capacity region of multi-input multi-output
(MIMO) broadcast networks was eventually established by Weingarten
et al. (2006) and constitutes one of the few classes of non-degraded
broadcast networks for which the capacity region of individual rates
is fully characterized. When only the single-antenna destinations have
perfect CSI, the multi-antenna source has no CSI, and the channels
to the different destinations are symmetric, then it was noticed
by Caire and Shamai (1999) that the multiplexing gain collapses to
one. This result was extended to a general isotropic model by Jafar
and Goldsmith (2005). Motivated by the huge difference between the
sum capacities of the broadcast networks with perfect CSI and those
with no CSI at the source, there has been a flurry of contributions in
recent years that investigate achievability of full spatial multiplexing
gain in broadcast networks with partial CSI at the source (see the
references in Gesbert et al. (2007) and Love et al. (2008)). Most of
these contributions focus however on the case where all channels are
frequency-flat, with relatively little emphasis on frequency-selective
channels. Specifically, to the best of our knowledge, there is no result in
the literature on feedback rate scaling required for achievability of full
spatial multiplexing gain in frequency-selective broadcast networks
in the large source power regime (that is, in limit of source power
going to infinity). Our third and final major contribution in this thesis
is that we partially plug this gap by investigating achievability, in
the large source power regime, of full spatial multiplexing gain in a
frequency-selective MISO broadcast network where the source obtains
1.2.
BROADCAST NETWORKS
channel knowledge through a limited capacity (shared) feedback link,

as discussed below.
MISO broadcast networks with (shared) limited feedback
MISO broadcast networks refer to a class of broadcast wireless networks where a source, equipped with M antennas, communicates
with K single-antenna destinations. We consider a MISO broadcast
network where the source has an average transmit power of P , all the
channels are frequency-selective with L taps and the source obtains
its CSI through a limited-capacity (shared) feedback link from the
destinations. We investigate the network under two scenarios, namely,
for K = M (motivated by Jindal (2006)) and for K M (motivated
by Kountouris et al. (2008) and Yoo et al. (2007)).
Specifically, Jindal (2006) considered a MISO broadcast network
under the assumption that all channels are frequency-flat (L = 1)
and all channel coefficients are drawn independently according to a
Rayleigh-dsitribution. Employing random isotropically distributed
channel quantization vectors along with zero-forcing transmission is
shown to achieve full spatial multiplexing gain of M in terms of the
average (over the channel realizations) sum rate, provided the feedback
rate from each destination to the source scales with the sum (across all
source antennas) transmit power P as (M 1) log P . We show that the
feedback and transmission scheme that yields full spatial multiplexing
gain in MISO interference networks, when applied to the broadcast
network considered by Jindal (2006) (that is, for K = M ), achieves
full spatial multiplexing gain of M , provided the number of feedback
bits from each destination to the source scales with the sum (across
all source antennas) transmit power P as (M L 1) log P . This result
generalizes the main result by Jindal (2006), who assumed all channels
to be frequency-flat, to the case where all channels are frequencyselective with L taps. Interestingly, in contrast to Jindal (2006), our
result does not require the Rayleigh-distribution assumption and is
valid for any continuous channel-coefficient distribution with finite
energy. Moreover, while the findings of Jindal (2006) pertain to average
1.
sum rate, our statements are valid on a per-realization basis.

Kountouris et al. (2008) considered a MISO broadcast network
under the assumption that all channels are frequency-flat (L = 1)
and all channel coefficients are drawn independently according to a
Rayleigh-distribution. The number of destinations K is assumed to
be finite but large (that is, K M ). For this network and under
certain assumptions, the results in Yoo et al. (2007) imply that if
each destination can feedback its (unquantized) SINR along with
(M 1) log P log K bits for its channel direction to the source,
employing spatial zero-forcing beamforming transmission based on the
received quantized CSI to a subset of destinations, chosen according
to semi-orthogonal user-selection proposed by Yoo and Goldsmith
(2006), achieves full spatial multiplexing gain. We propose an iterative
feedback scheme in which M destinations are selected from the K
destinations and only these M destinations feedback their CSI through
a shared (by all destinations) feedback link to the source. We show
that the scheme, along with the transmission scheme proposed for
the MISO interference networks, achieves full spatial multiplexing
gain of M , provided the number of feedback bits from each of the
M selected destinations scales with P and K as (M L 1) log P
log K. This result generalizes the result in Yoo et al. (2007), where
all channels are frequency-flat, to the case where all channels are
frequency-selective with L taps. In contrast to Yoo et al. (2007) where
each of the K destinations needs to feedback (M 1) log P log K
bits, our scheme requires only M of the K destinations to feedback
(M L 1) log P log K bits. Conceptually, this reduction in feedback
is achieved by using a two-stage channel quantization scheme of a
flavor similar to Kountouris et al. (2008) and Zakhour and Gesbert
(2007). In addition, while the findings in Yoo et al. (2007) pertain to
average sum rate, our statements are valid on a per-realization basis.
10
1.3.
1.3.
OUTLINE OF THE THESIS
OUTLI N E OF TH E TH ESIS
The remainder of this thesis is organized as follows. In Chapter 2, we

present our results on large interference networks with 1-bit feedback.
Our results on spatial multiplexing in SISO and MISO interference
networks are presented in Chapters 3 and 4, respectively. In Chapter 5,
we analyze the feasibility of spatial multiplexing in frequency-selective
MISO broadcast networks with limited feedback. Finally, Chapter 6
concludes the thesis.
11
C HAPTER
Large Interference Networks with 1-bit

feedback
wireless system with a large number of non-cooperating
nodes is commonly modeled in information theory as a large
interference network. The capacity region of such a network
is however not known in general. A powerful technique to characterize
this capacity region is to analyze the sum capacity in the limit of the
number of nodes going to infinity. This technique was recently adopted
by Bolcskei et al. (2006) to derive a scaling law governing the sum
capacity of an M source-destination pair network where the source
Si , i = 1, . . . , M, communicates with the destination Di , with the
help of K distinct relays. Each relay is dedicated to only one sourcedestination pair so that there are M K relays in total. Bolcskei et al.
(2006) proposed a two-hop transmission protocol for this network
where the sources transmit their signals to the relays during the
first hop and the relays forward, after appropriate processing, their
received signals to the destinations during the second hop. It was
assumed that the number of source-destination pairs M is finite but
the number of per source-destination pair relays K goes to infinity.
The main result of B
olcskei et al. (2006) is that the proposed protocol
results in a sum capacity scaling of (M/2) log(1 + cK), with c a
constant independent of K, in the limit of large K. This achievability
depends critically, however, on the assumption that each relay knows
13
2.
LARGE INTERFERENCE NETWORKS WITH
1- BIT
FEEDBACK
both its channel from its corresponding source and its channel to its
corresponding destination, perfectly. While a relay can easily acquire
close-to-perfect channel knowledge about its channel from the source,
it requires an infinite capacity feedback link from its destination to
know its channel to the destination perfectly. A natural question to
ask therefore is how the sum capacity scaling behavior changes if the
capacity of the feedback links from the destinations to the relays is
limited. To investigate the effect of this limited feedback on the sum
capacity scaling law, we focus in this chapter only on the second hop
of the protocol, that is, on the communication between the M K relays
and the M destinations. Specifically, we assume that the first hop
transmissions are error-free, so that the second hop is itself equivalent
to a large interference network with M K sources and M destinations,
as shown in Fig. 2.1. The details of this network are as follows.
The M K non-cooperating sources communicate with the M noncooperating destinations over slow-fading frequency-flat channels.
Each source and each destination is equipped with a single antenna.
The sources are divided into M mutually exclusive groups G1 , , GM
of K sources each so that every group Gi is dedicated to transmit a
common message to its unique destination Di . All transmissions occur
concurrently and in the same frequency band. The K non-cooperating
single-antenna sources in group Gi are denoted by Si1 , , SiK . The
results of Bolcskei et al. (2006) imply that if there exist infinite
capacity feedback links from each destination to its corresponding
group of sources, then the sum capacity of this network scales as
M log(1 + cK), with c a constant independent of K, for large values
of K. We assume, in contrast to the perfect feedback assumption
of Bolcskei et al. (2006), that there exists an error-free dedicated
1-bit broadcast feedback channel from each destination Di to its
corresponding group of sources in Gi . Our main contribution in this
chapter is to show that even with this relaxed assumption, the sum
capacity of the network scales with K as M log(1 + cK), that is, full
spatial multiplexing gain of M and a per-stream (distributed) array
Recall
14
that the relays are the sources for the second hop transmission.
S11
G1
S12
D1
S1K
S21
G2
S22
D2
S2K
1
SM
GM
2
SM
DM
K
SM
Fig. 2.1: A large interference network where M groups of K single-antenna

sources each, communicate with M single-antenna destinations. This network
models the second hop of the multiuser relaying protocol of B
olcskei et al.
(2006).
gain proportional to K is achievable with this relaxed assumption.

The rest of the chapter is organized as follows. We provide a system and signal model for the considered network in Section 2.1. In
Section 2.2, we derive a simple upper bound on the sum capacity of
the network and show that spatial multiplexing gain and per-stream
array gain achievable in the network are upper-bounded by M and K,
respectively. Thereafter, in Section 2.3, we propose a feedback-based
iterative distributed (multi-user) beamforming algorithm to learn
the channels between each group of sources and its assigned destination. This algorithm is a generalization, to the multi-user case, of the
feedback-based iterative distributed beamforming algorithm proposed
by Mudumbai et al. (2006, submitted) for networks with a single
15
2.
1- BIT
FEEDBACK
group of sources and a single destination. Making the simplifying

assumption, compared to Mudumbai et al. (2006, submitted), of the
channel coefficients as well as all the signals being real-valued allows
us to put the iterative algorithm into a Markov chain context, thereby
setting the stage for a simple convergence proof in Section 2.4. In
Section 2.4, we also investigate the convergence rate of the algorithm
and show that convergence, both in probability and in expectation,
occurs within a training period length scaling linearly in K. We then
show in Section 2.5 that, for M finite and K , spatial multiplexing based on the beamforming weights produced by the iterative
algorithm achieves full spatial multiplexing gain of M and a perstream (distributed) array gain proportional to K, that is, full spatial
multiplexing gain is achieved with limited feedback. We furthermore
demonstrate that the M effective links Gi Di in the network not
only decouple (reflected by full spatial multiplexing gain) but also
converge to non-fading links as K , that is, in the terminology
of Morgenshtern and B
olcskei (2007), the network crystallizes. In
addition, we quantify the impact of the performance of the iterative
algorithm on the crystallization rate, and we show that the multiuser nature of the network leads to a significant reduction in the
crystallization rate, when compared to the M = 1 case. Employing a
cut-set argument, we show in Section 2.6 that the iterative distributed
beamforming algorithm, which requires training period length to scale
linearly in K, is asymptotically optimal in the sense that full spatial
multiplexing gain of M and an array gain proportional to K cannot be
achieved with training period length scaling sub-linearly in K. Finally,
in Section 2.7, we relax the assumption of all channel coefficients as
well as signals being real-valued to complex-valued channel coefficients
and signals, and prove that all the above results are still valid.
2.1.
SYSTEM AN D SIGNAL MODEL
We assume that xi [n] is the common message of the sources in Gi to

be transmitted to Di . In the remainder of the chapter, we distinguish
16
2.1.
SYSTEM AND SIGNAL MODEL
between a training phase during which the K scalar channels between

Sik , k = 1, . . . , K, and Di are learned for each i {1, . . . , M }, and
a data transmission phase following the training phase. During the
training phase for Gi the feedback broadcast channel Di Gi is used
once every frame of Tf time slots. The l-th frame (l = 0, 1, ...), denoted
by Fl , consists of the time slots n = lTf , lTf + 1, . . . , (l + 1)Tf 1. In
the l-th frame, each source Sik multiplies the sequence xi [n] (which
is a training sequence during the training phase) by a corresponding beamforming weight ik [l] {1, 1} before transmission; these
beamforming weights are kept constant during the entire frame. We
furthermore assume that all the channels in the network are frequencyflat and remain constant throughout the entire time-interval of interest,
that is, during training and data transmission phases. Throughout the
chapter, we assume that M is finite. The corresponding input-output
relations are now given by
yi [n] =
X
K
hki,i ik [l]
xi [n] +
M X
K
X
r=1
r6=i
k=1
where l =
n
Tf
xr [n]
k=1
|
+ wi [n],
hki,r rk [l]
{z
interference
i = 1, . . . , M,
}
(2.1)
k
, yi [n] C denotes the symbol received at Di in the
n-th time slot, hki,r C stands for the channel coefficient between Srk
and Di and wi [n] denotes the CN (0, 2No ) i.i.d. noise sequence at Di .
The channel coefficients hki,r , for all i, r, k, are assumed i.i.d. CN (0, 2).
The signals xi [n] obey the average power constraint

P
E |xi [n]|2 ,
K
i = 1, . . . , M,
so that the average power transmitted by each group Gi is limited

by P . During the training phase, the frame-rate sequences ik [l] are
updated based on the 1-bit feedback received at the end of each frame.
17
2.
1- BIT
FEEDBACK
The sources in Gi are assumed to be non-cooperating and therefore

Di cannot measure channel coefficients hki,r individually. Consequently,
feedback from Di to the sources in Gi is based on its received signal
yi [n] only. The goal of the training
process is to find the beamforming

k
k
weights i = sign R{hi,i } . While it is in general difficult to put the
individual groups into perfect beamforming configuration, so that
K
X
R{hki,i }ik
k=1
K
X

R{hki,i },
=
for group Gi ,
k=1
we will show that if in each group a sufficient number of sources

is in beamforming configuration, full spatial multiplexing gain can
be achieved. In what follows, we shall say that in a given group Gi ,
sources satisfying R{hki,i }ik > 0 are aligned, whereas sources with
R{hki,i }ik < 0 are reverse-aligned. We denote the final beamforming
weights ik produced during the training phase as
ik . In the data
transmission phase, the groups Gi , i = 1, . . . , M, perform co-channel
data transmission (that is, spatial multiplexing) based on
ik so
that the corresponding input-output relation is given by (2.1) with
ik [l] =
ik .
2.2.
U PPER B OU N D ON SUM CAPAC ITY
An upper bound on the sum capacity of the network can be obtained

by simply considering a single link Gi Di , under the assumption
that all the sources in the other groups, that is, in Gk , k 6= i, are
silent. The input-output relation for the link Gi Di during the data
transmission phase is then given by
yi [n] =
X
K
hki,i
ik

xi [n] + wi [n].
(2.2)
k=1
If the sources are allowed to cooperate, the destination D can estimate each
i
hki,i (by simply silencing all Sil , l 6= k, in a sequence) and feedback its phase to Sik .
Distributed beamforming would then be achieved in a straightforward manner.
18
2.2.
UPPER BOUND ON SUM CAPACITY
Assuming that the destination Di knows the effective channel coeffiPK

cient k=1 hki,i
ik , the rate of communication over the link Gi Di
is then upper-bounded according to
!
K
X k k

hi,i
i
(2.3)
Ri max
max
I yi [n]; xi [n]
f (xi [n])
k ,k
i
E[(xi [n])2 ]P/K
k=1
PK
2
!
k

ik E |xi [n]|2
k=1 hi,i
= max log 1 +
2No
k
i ,k
!

PK
2
k
k=1 |hi,i | P
log 1 +
2KNo
(since |
ik | 1, i, k)

2 !
K E |hki,i | P
= log 1 +
2KNo
(since
K
X

|hki,i | = K E |hki,i | for K )
(2.4)
(2.5)
(2.6)
k=1

2 !
K E |hki,i | P
= log 1 +
2No
(2.7)
= log(1 + cK)
(2.8)
where
c=

2
E |hki,i | P
2No
(2.9)
is independent of K. Since communicating in presence of interference

from other groups of sources cannot increase the capacity of a given
link Gi Di , the sum capacity of the network is upper-bounded
according to
M
X
Ri M log(1 + cK)
(2.10)
i=1
19
2.
1- BIT
FEEDBACK
for some positive constant c independent of K. It therefore follows

that spatial multiplexing gain and per-stream array gain achievable by
any transmission scheme (even with infinite capacity feedback links)
are upper-bounded by M and a term proportional to K, respectively.
In the rest of this chapter, we show that the sum capacity scaling
(2.10) is achievable if there exists a 1-bit broadcast feedback link
between each destination Di and its corresponding group of sources
Gi .
Real-value assumption
For ease of exposition, we shall first establish all our results in Section 2.3, 2.4, 2.5 and 2.6 under the simplifying assumption of all the
channel coefficients as well as all the signals being real-valued, that is,
we will assume xi [n] R, yi [n] R, hki,r R (so that R{hki,r } = hki,r )
and wi [n] N (0, No ). We will then show in Section 2.7 that all our
results can be extended to complex-valued channel coefficients and
signals in a (relatively) straight-forward fashion.
2.3.
ITERATIVE DISTR I BUTED B EAMFORMI NG
We shall next describe the algorithm carried out during the training
phase. The overall training phase is assumed to consist of Ttr = no M K
frames (each of which contains Tf time slots) divided into M blocks of
no K frames each. The role of the parameter no will become clear later.
During each of these M blocks, precisely one of the groups Gi follows
the three-step iterative distributed beamforming algorithm, described
below, while all the other groups of sources Gr , r 6= i, remain silent.
At the end of the training phase of no M K frames, each of the groups
Gi is in (close-to) beamforming configuration with respect to (w.r.t.)
its assigned destination. The order in which the groups follow the
three-step procedure below can be decided offline and communicated
to all the nodes in the network. Without loss of generality (w.l.o.g.),
we assume that the group Gi is being processed during the i-th block
20
2.3.
ITERATIVE DISTRIBUTED BEAMFORMING
defined as the set of frames l = (i1)no K, (i1)no K +1, . . . , ino K 1.

Since much of the analysis in the current and the next section deals
with a single group Gi only, we shall consider, w.l.o.g., the group G1 ,
drop the index i = 1 and, wherever appropriate, use the convention
G G1 , S k S1k , D D1 , k = 1k ,
k =
1k and hk = hk1,1 (or
k
k
equivalently h = R{h1,1 } since we are considering only real-valued
channel coefficients). The three steps carried out by the iterative
distributed beamforming algorithm can now be summarized as follows.
Step 1. Initialization of the received signal level: This step pertains to the zeroth frame in the block. Each of the sources S k
initializes its beamforming weight according to k [0] = 1, initializes an auxiliary beamforming weight
as
k [0] = k [0] and
p
starts transmitting the pilot symbol P/K. The corresponding
received signal at destination D is given by
r
y[n] =
K
P X k k
h [0] + w[n]
K
k=1
r
=
K
P X k
h + w[n],
K
n F0 .
k=1
The destination D then estimates the received signal level by

averaging y[n] over the entire frame, resulting in
rx
1 X
y[n] =
=
Tf
nF0
K
P X k
h .
K
k=1
Here, we assumed that the estimate of Lrx is perfect, which

requires that Tf be sufficiently large. Finally, D initializes Lmax =
Lrx .
Step 2. Iterative distributed beamforming: This is the iterative
step that is performed for each frame, except for the zeroth one,
that is, for l = 1, . . . , no K 1 (recall that the zeroth frame is
used to carry out the previous initialization step). The details
of this step are as follows. At the beginning of each frame, each
21
2.
1- BIT
FEEDBACK
of the sources S k , k = 1, . . . , K, chooses its beamforming weight

k [l], independently across k, according to:
(
1
k [l 1],
w.p. 1 K
k [l] =
.
(2.11)
1
k [l 1], w.p. K
p
Each of the sources S k then transmits the pilot symbol P/K
throughout the frame, using thepbeamforming weight k [l], that
is, S k transmits the signal k [l] P/K. At the end of the frame
under consideration, D estimates the corresponding received
signal level, as in the initialization step, according to
rx
1 X
=
y[n] =
Tf
nFl
K
P X k k
h [l]
K
(2.12)
k=1
and, through the 1-bit broadcast feedback channel, informs all the
sources in G whether Lrx > Lmax or not. Based on the received
feedback, the sources S k , k = 1, . . . , K, update their auxiliary
beamforming weights
k [l] as follows:
(
k [l 1], if Lrx Lmax

k
, k = 1, . . . , K.
(2.13)
[l] =
k [l],
if Lrx > Lmax
Finally, if Lrx > Lmax , D updates Lmax = Lrx .

Step 3. Silencing: At the end of the block, the sources in G
store the current values of their respective auxiliary beamforming
weights as
k =
k [no K 1] and go silent.
In the proposed protocol, only one group of sources is active during a

given block. It is therefore natural to ask whether the groups that have
finished their training phase could start their data transmission phase
while the remaining groups are learning their channels through the
three-step procedure above. At first sight, one would be tempted to
conclude that such a modified protocol would result in higher spectral
Again,
22
we assume the estimate to be perfect.
2.3.
ITERATIVE DISTRIBUTED BEAMFORMING
efficiency. We note, however, that performing Steps 1 and 2 above in

the presence of interference created by the groups already transmitting
data would require a longer Tf , say TfI , in order to achieve the same
quality of the received signal level estimate as in the original protocol,
that is, in the absence of interference. The resulting tradeoff can be
illustrated roughly by assuming maximum-likelihood estimation of the
parameter Lrx . The corresponding variance of the estimation error is
(No +I2 )
2
o
given by 2 = N
in
Tf in the original protocol and by M =
TI
f
the modified protocol, where, considering the link Gi Di , we have

2
i1 K
P X X k k
2
hi,r
r .
(2.14)
I =

K
r=1 k=1
Consequently, if we require the variance of the estimation error in the

two protocols to be equal, then
TfI
2
=1+ I .
Tf
No
(2.15)
Assuming that in each of the links already in (close-to) beamforming

configuration the corresponding group of sources transmits at a rate
of R bits per time slot, the total number of bits transmitted in the
modified protocol during the first no M KTfI time slots is given by
Rk0 KTfI M (M 1)/2 [bit] whereas it is Rno M 2 K(TfI Tf ) [bit] in
the original protocol. Using (2.15), this implies that the modified
protocol would have a higher spectral efficiency if
2
I2
<1
.
No
M +1
(2.16)
Now, since the weights

ik have been optimized to be in beamforming
k
configuration w.r.t. hi,i , they are independent of hki,r for r 6= i.
Furthermore,
ik = sign(hki,i ) and
ik = sign(hki,i ) for Sik in the set of
aligned and reverse-aligned sources, respectively. Since P (sign(hki,i ) =
1) = P (sign(hki,i ) = 1) = 1/2, it follows that P (
ik = 1) = P (
ik =
k
1) = 1/2, for all k. Along with the assumption hi,r N (0, 1), for
23
2.
1- BIT
FEEDBACK
all i, r, k, we therefore get hki,r

k hki,r , for all k and r < i, which
2
P r

K
allows us to conclude that k=1 hki,r
rk , r < i, scales linearly in
K. Consequently, it follows from (2.14) that I2 is proportional to
P , which implies that the condition (2.16) would not be met for any
reasonable choice of SNR = P/No . We can therefore conclude that,
provided the variance of the signal level estimation error is the relevant
performance measure, the original protocol would in practice always
outperform the modified protocol in terms of spectral efficiency.
2. 4 .
CONVERGENC E OF TH E ITERATIVE
ALGOR ITHM
We shall next show that the iterative algorithm described in the

previous section converges to the beamforming configuration for no
large. The proof is rather straightforward and consists of the following
sequence of arguments:
We start by recognizing that through (2.11) and (2.13), the

= [
coefficient vector [l]
1 [l] . . .
K [l]]T depends only on the
1]. The
coefficient vector in the previous frame, that is, on [l
therefore follows a Markov chain.
sequence of vectors [l]
can take on the values 1 or
Each of the elements in the vector [l]
1, which implies that the total number of states of the Markov
chain is given by 2K . Let us denote these states by Q1 , , Q2K .
PK
k .
To each of the states, we can associate a value k=1 hk
The Markov chain has one special state, namely, when
k =
k
sign(h ), for all k. W.l.o.g. we let this state be Q1 and note that
it corresponds to the beamforming configuration, that is,
K
X
k=1
hk
k =
K
X
k=1
hk sign(hk ) =
K
X
|hk |
(2.17)
k=1
which implies that once the system is in state Q1 the parameter

Lmax is maximized over all states Qi , i = 1, . . . , 2K . From the
24
2.4.
CONVERGENCE OF THE ITERATIVE ALGORITHM
update rule (2.13), we can therefore conclude that once in Q1

the Markov chain will remain in Q1 , which implies that Q1 is an
absorbing state (see (Grinstead, 1997, Def. 11.1)).
Each of the states Qr , r 6= 1, corresponds to a coefficient vector
with at least one reverse-aligned element. Let us denote the
number of reverse-aligned elements corresponding to Qr by ro ,

where ro 1. Then, the transition probability from Qr to the
absorbing state Q1 is (1/K)ro (1 1/K)Kro . We can therefore
conclude that the absorbing state can be reached from all the
states Qr so that the Markov chain is an absorbing Markov chain
(see (Grinstead, 1997, Def. 11.1)). It then follows from (Grinstead,
1997, Th. 11.3) that the system moves to the absorbing state
Q1 w.p. 1 as the length of the training phase becomes large.
Since Q1 corresponds to the beamforming configuration, we can
conclude that the iterative algorithm eventually converges to the
beamforming configuration.
We emphasize that due to our simplifying assumptions, compared

to Mudumbai et al. (2006, submitted), of the channel coefficients
and the signals being real-valued, we were able to prove convergence
of the iterative algorithm to the beamforming configuration in a
simple fashion using basic results from Markov chain theory. The
proof above, however, does not reveal anything about the rate of
convergence of the algorithm. It seems difficult to obtain results on
the actual convergence rate because the state-transition probabilities
in the Markov chain, implicitly, depend on the actual realizations
of the channel coefficients hk . However, interesting insights into the
behavior of the convergence rate, as a function of K, can be obtained
by considering convergence in probability according to
P
X
Sk G
h
[t] > zo
P
tno K
X
|h | > zo ,
Sk G
zo (, )
(2.18)
25
2.
1- BIT
FEEDBACK
and convergence in expectation according to

X k k
X k k
E
h
[t] = Ehk E{k [l]}tl=0 |hk
h
[t]
Sk G
Sk G
E
tno K
X
hk

|hk | .
(2.19)
Sk G
In particular, we shall show that convergence, both in probability and

in expectation, can be obtained if the length of the training phase
Ttr scales linearly in K, that is, no is independent of K. To be precise,
we shall establish convergence to the beamforming configuration up
to a certain level in the sense that we allow for a (small) fraction of
the sources to be reverse-aligned. The corresponding concepts and
the associated convergence proofs are provided next.
2.4.1. Convergence in probability to -level

In the following, we shall be interested in convergence of the iterative
algorithm in the sense of (2.18) up to a certain level. Concretely, we
shall allow that approximately a (small) fraction of the sources in
a given group is reverse-aligned, which trivially implies that (1 ) is
approximately the fraction of sources that are aligned. Denoting the
sets of aligned and reverse-aligned sources after the update (2.13) at
the end of the t-th frame as A[t] and A[t], respectively, so that
X
X
X
hk
k [t] =
|hk |
|hk |
(2.20)
Sk G
S k A[t]
S k A[t]
Strictly speaking, it can be shown that convergence in probability implies

convergence in expectation. We will however provide an independent derivation for
convergence in expectation, partly because of the insights the derivation provides
on the rate of convergence, and partly because we believe that the corresponding
proof techniques are of independent interest.
For a given K, this requires that be an integer multiple of 1/K.
26
2.4.
and defining ak , k = 1, . . . , K, to be i.i.d. random variables drawn

according to N (0, 1), so that
ak hk ,
k = 1, . . . , K
(2.21)
we say that convergence in probability to -level is achieved in the

t-th iteration if
X

k k
P
h
[t] > zo
Sk G
=P
X
|hk |
S k A[t]
(1)K
X
|ak |
k=1
|hk | > zo
S k A[t]
K
X
|ak | > zo

(2.22)
k=(1)K+1
holds for all zo (, ).

We will later see that convergence according to (2.22) is sufficient
for our proof of achievability of full spatial multiplexing gain, provided
that is a small positive constant independent of K. We also note
that although hk , for all k, is drawn according to N (0, 1), neither hk ,
for S k A[t], nor hk , for S k A[t], follows the N (0, 1) distribution
for t > 0. This anomaly can be illustrated by a simple example.
Let us consider two sources S p and S q that are aligned and reversealigned, respectively, at the end of the (t 1)-th iteration, that is,
S p A[t 1] and S q A[t 1]. Let us furthermore assume that
only S p and S q alter their beamforming weights in the t-th iteration
according to the random perturbation (2.11) as p [t] =
p [t 1]
q
q
and [t] =
[t 1], respectively, while all the other sources retain
k [t 1], for
their respective beamforming weights, that is, k [t] =
k 6= p, q. Keeping everything else same, this perturbation will result
in different outcomes, at the end of the t-th iteration, for the cases
|hp | > |hq | and |hp | < |hq |. Specifically, if |hp | > |hq | the perturbation
results in S p A[t] and S q A[t] (that is, no change from the
(t 1)-th iteration), and if |hp | < |hq | the update rule (2.13) leads
27
2.
1- BIT
FEEDBACK
to S p A[t] and S q A[t]. Thus, the sources S k corresponding to

relatively large values of |hk | are biased to stay in the set of aligned
sources A[t] compared to the sources corresponding to small values of
|hk |. Generally speaking, this bias may result, over a large number of
iterations, in concentration of sources with relatively large values of
|hk | in A[t] and of sources with relatively small values of |hk | in A[t].
However, in what follows, we will ignore this anomaly for the sake of
tractability and will assume that both hk , S k A[t] and hk , S k A[t]
are distributed according to N (0, 1). Our convergence result can now
be formalized as follows.
Theorem 1 For large K and any positive constant o independent

of K, the number of iterations required in the second step of the
distributed beamforming algorithm in Section 2.3 to achieve convergence in probability to o -level is at most no K, where no is a constant
independent of K.
Proof: Let [t] {0, 1/K, 2/K, , 1} denote a minimum convergence in probability level that is guaranteed at the end of the t-th
iteration, assuming that the above training algorithm is followed. We
trivially note that at the end of the zeroth iteration,
X

X

k k
k
P
h
[0] > zo = P
h > zo
Sk G
Sk G
X

K
k
P
|a | > zo
k=1
(since ak hk , k)
(2.23)
and hence, convergence in probability to [0] = 1 level is guaranteed.

We furthermore note that since
X
X
hk
k [t]
hk
k [t 1],
t,
(2.24)
Sk G
28
Sk G
2.4.
we have
hk
k [t] > zo
Sk G
hk
k [t 1] > zo
Sk G
(1[t1])K
X
K
X
|ak |
k=1
|ak | > zo

(2.25)
k=(1[t1])K+1
for t = 1, . . . , no K, for zo (, ), and hence

[t] [t 1],
t = 1, . . . , no K
(2.26)
which, in turn, implies that the decrease in convergence level in the

t-th iteration, given by [t] = [t 1] [t], is always non-negative.
For the rest of the proof, we therefore just need to show that the
convergence level [t] decreases from 1 to (less than) o within no K
iterations, that is,
n
oK
X
[t] (1 o )
(2.27)
t=1
with no independent of K, where [t] 0, for t = 1, . . . , no K.

To this end, we start by noting that since [t], for all t, is an integer
multiple of 1/K, the difference [t] will also be an integer multiple
of 1/K, that is,

1 2
[t] 0, , , ,
t.
(2.28)
K K
We shall focus on a special event [t], namely, when precisely one
of the sources in A[t 1] alters its beamforming weight from 1
to 1 or vice-versa according to (2.11) at the beginning of the t-th
frame and none of the sources in A[t 1] alters its beamforming
29
2.
1- BIT
FEEDBACK
weight. This naturally results in [t] = 1/K (note that may

not be the only event that results in [t] = 1/K). Assuming that
|A[t1]| = 0 [t1]K, for some 0 [t1], the probability P ([t] = 1/K)
can then be lower-bounded according to

1
P [t] =
P ()
(2.29)
K

K1
0
1
[t 1]K 1
1
(2.30)
=
1
K
K

K1
0 [t 1]K
1
=
1
(2.31)
K
K
0 [t 1]
=
(for large K)
(2.32)
e
o
(2.33)
e
where (2.33) follows because 0 [t 1] < o would imply that at least
(1 o )K sources are aligned, and hence, convergence in probability
to o -level is already achieved.
Using the fact that
o
eP [t] =
1
K
1,
t,
(from (2.33))
(2.34)
we then consider a set of no K auxiliary random variables, denoted

by [t], t = 1, . . . , no K, and defined as
o
1
1
, w.p. eP [t]=
, if [t] = K
(
K
K)
o
1
[t] , 0, w.p. 1
, if [t] = K
1
eP ([t]= K
)
0, if [t] {0, 2 , . . . , 1}
K
so that [t], t = 1, . . . , no K, are i.i.d. Bernoulli distributed random
variables (Papoulis and Pillai, 2002, Sec. 4.3) with probabilities

1
o
P [t] =
= ,
t
K
e
30
2.4.
P ([t] = 0) = 1
o
,
e
t.
(2.35)
Moreover, since [t] [t], t, we have

n
oK
X
[t]
n
oK
X
[t]
(2.36)
t=1
t=1
and hence,
P
nX
oK

nX

oK
[t] (1 o ) .
[t] (1 o ) P
t=1
t=1

Pno K
Finally, noting that
t=1 [t] K is equal to the number of times
[t] takes on the value 1/K out of a total of no K Bernoulli trials
and therefore follows a binomial distribution (Papoulis and Pillai,
2002, Sec. 4.3) with mean

o
= no K
(2.37)
e
and standard deviation
r
=
no K

o
o
e
and assuming
that no > (1o )e/o , the probability P

1 o can be lowerbounded according to
P
n
oK
X
(2.38)
Pno K
t=1
[t]
!
[t] 1 o
t=1
n
oK
X
!
[t] 1 o
(2.39)
t=1
=P
n
oK
X
!
[t]K (1 o )K
(2.40)
t=1
31
2.
=1P
n
oK
X
1- BIT
FEEDBACK
!
[t]K < (1 o )K
(2.41)
t=1
!
no K
no Ko X
no Ko
=1P
[t]K >
(1 o )K
(2.42)
e
e
t=1

!
no K
n K
n K

o o X
o
o
1P
[t]K >
(1 o )K
(2.43)
e

e
t=1
no K

X
no Ko

=1P
[t]K
e }
| {z
t=1
=
!
r

(1 o )K
o
o
>q
1

no K e
e
no K eo 1 eo |
{z
}
no Ko
e
(2.44)
q
1
no Ko
e
o
e
no K
o
e
2
(1 o )K
{z
(2.45)
0 as K
=1
(2.46)
where (2.45) follows because Chebyshevs inequality implies that

P (|x | > r) 1/r2 , for all r > 0. Consequently for
no >
(1 o )e
o
(2.47)
we have
P
n
oK
X
!
[t] 1 o
=1
for K
(2.48)
t=1
and from (2.27), convergence in probability to o -level is achieved

with probability 1.
32
2.4.
2.4.2. Convergence in expectation to -level

In this section, we shall be interested in convergence of the iterative
algorithm in expectation, in the sense of (2.19), once again up to a
certain level. Concretely, we shall allow that, on average, a (small)
fraction of the sources in a given group is reverse-aligned, which
trivially implies that (1 ) is the fraction of sources that are aligned,
on average. For the rest of the chapter, we shall assume that is
independent of K and K Z+ . We say that convergence in the sense
of (2.19) up to an -level has been achieved if

X

Ehk E{k [l]}tl=0 |hk
hk
k [t]
Sk G

X
= Ehk E{k [l]}tl=0 |hk
|hk |
S k A[t]

|hk |
(2.49)
S k A[t]

K(1 2) E |hk | .
(2.50)
We can now formalize our convergence (in expectation) result as

follows.
Theorem 2 For any o > 0 and large K, the number of iterations
required in the second step of the distributed beamforming algorithm
in Section 2.3 to achieve convergence in expectation to o -level is at
most no K, where no is a constant independent of K.
Proof: We start by defining
X
X
S new [t] =
hk
k [t]
hk
k [t 1]
Sk G
(2.51)
Sk G
so that S new [t] 0, for all t. With this definition, we have

X

Ehk E{k [l]}no K1 |hk
hk
k
l=0
Sk G
noX
K1

X k
Ehk E{k [l]}tl=0 |hk S new [t] + Ehk
h
t=1
(2.52)
Sk G
33
2.
1- BIT
FEEDBACK
noX
K1

Ehk E{k [l]}tl=0 |hk S new [t] .
(2.53)
t=1
Let there be (1 )K and K sources in A[t 1] and A[t 1],

respectively, with 0 < o < < 1. Moreover, let qa [t] and q a [t] denote
the number of sources in A[t 1] and A[t 1], respectively, that alter
their beamforming weights from 1 to 1 or vice-versa according to
(2.11) at the beginning of the t-th frame. Then, we get

Ehk E{k [l]}tl=0 |hk [S new [t]]

= P (qa [t] = 0) Ehk E{k [l]}tl=0 |hk [S new [t]qa [t] = 0]

+ P (qa [t] > 0) Ehk E{k [l]}tl=0 |hk [S new [t]qa [t] > 0]

P (qa [t] = 0) Ehk E{k [l]}tl=0 |hk [S new [t]qa [t] = 0]
= P (qa [t] = 0)
K
X
P (q a [t] = s)
s=0

Ehk E{k [l]}tl=0 |hk [S new [t]qa [t] = 0, q a [t] = s] .
(2.54)
We next show that the expected value inside the summation in (2.54)
can be lower-bounded by co s, where co is a constant independent of K.
To this end, we start by noting that qa [t] = 0, q a [t] = s corresponds to
the case where precisely s sources move from A[t 1] to A[t] and none
of the sources moves from A[t 1] to A[t]. We denote the sets of these
s sources as R, and note that since the s sources in R
are chosen from
the K sources in A[t 1], there are precisely K
possible choices
s
for R, with the corresponding sets denoted as R1 , R2 , . . . , R(K ) .
s
Furthermore, each source in A[t 1] alters its beamforming weight

independently and with equal probability, and hence, each of the K
s
choices is equally likely. From (2.51) it therefore follows that for a
given set Ri , we have
X

E{k [l]}tl=0 |hk S new [t]Ri = 2
|hk |
(2.55)
S k Ri
34
2.4.
which implies

E{k [l]}tl=0 |hk [S new [t]qa [t] = 0, q a [t] = s]
X

1
E{k [l]}tl=0 |hk S new [t]Ri
= K
s
K
s
(a)
Ri {R1 ,...,R
|hk |

Ri {R1 ,...,R
K
s
2s
K
(K
s )

S k A[t1]
S Ri
(K
s )

K 1
|hk |
s1
|hk |
S k A[t1]
where Step (a) is a resultof the fact that each source in A[t 1] is
present in precisely K1
of the sets Ri . We therefore get
s1

Ehk E{k [l]}tl=0 |hk [S new [t]qa [t] = 0, q a [t] = s]
X

2s
k
=
E k
|h |
K h
k
S A[t1]
(a)

= 2s Ehk |hk |S k A[t 1]
Z
= 2s xf k
(x) dx
(2.56)
|h | A[t1]
where Step (a) follows from |A[t 1]| = K and the fact that the hk
are identically distributed. Next, we need to show that the integral
on the right hand side (RHS) of (2.56) can be lower-bounded by a
constant independent of K. To this end, we start by noting that
f|hk | (x) = P (A[t 1])f
|hk |A[t1]
+ P (A[t 1])f
(x)
|hk |A[t1]
(x)
35
2.
P (A[t 1])f
1- BIT

FEEDBACK
|hk |A[t1]
=f
|hk |A[t1]
(x)
(x)
which, using > o , implies

f
|hk |A[t1]
(x)
f|hk | (x)
.
o
We can then argue that finding the minimum of

class of functions f (x) that satisfy
f (x) = 0,
x < 0,
f (x) 0,
x 0,
f|hk | (x)
f (x)
o
Z
f (x) dx = 1
(2.57)
R
x
xf (x) dx over the
and
(2.58)
R
(x) dx
guarantees that this minimum is a lower bound on x xf k
|h | A[t1]
as f k
(x) is a member of this class of functions. Concretely,
|h | A[t1]
we want to determine
Z
min xf (x) dx
(2.59)
f ()
x
where the minimization is under the constraints (2.58). This minimization problem can be solved as follows. We start by setting

1 o
xo = Q1
(2.60)
2
q
R
u2
x2
where Q(v) = v 12 e 2 du, so that using f|hk | (x) = 2 e 2 , x
36
2.4.
0, and f|hk | (x) = 0, x < 0, we have

Zxo
f|hk | (x)
dx = 1.
o
Next, we note that

Z
Zxo
Z
xf (x) dx = xf (x) dx + xf (x) dx
0
Zxo
Z
xf (x) dx + xo f (x) dx
(2.62)
xo
0
Zxo

Zxo
xf (x) dx + xo 1 f (x) dx
(2.61)
xo
(2.63)

Zxo
Zxo
Zxo
f|hk | (x)
dx f (x) dx
= xf (x) dx + xo
o
0
Zxo
= xo
Zxo
(xo x)f (x) dx
f|hk | (x)
dx
o
Zxo
f|hk | (x)
= x
dx +
o
0
(2.64)
0
Zxo
(xo x)
f|hk | (x)
dx
o
(2.65)
(2.66)
Zxo
(xo x)f (x) dx
(2.67)

Zxo
Zxo
f|hk | (x)
f|hk | (x)
= x
dx + (xo x)
f (x) dx.
o
o
0
(2.68)
Since the second term in (2.68) is positive due to the constraint
37
2.
1- BIT
FEEDBACK
f (x) (1/o )f|hk | (x), it follows that setting

(f
(x)
|hk |
o
f (x) =
0,
0 x xo
(2.69)
otherwise
yields the desired lower bound. Further, corresponding to f (x) in

(2.69), we get
2
o
r
(Q1 ( 1
2 ))
Z
2
co
1 e
2
, .
(2.70)
xf (x) dx =
o
2
x
Substituting (2.70) into (2.56) and the result thereof into (2.54),
we obtain

Ehk E{k [l]}tl=0 |hk [S new [t]]
co P (qa [t] = 0)
K
X
sP (q a [t] = s)
(2.71)
s=0
(1)K

1
= co 1
K
Ks

K
X K 1 s
1
s
1
K
K
s
s=1

K X

s
K
1
1
K
K
= co 1
s
1
K
s
1 K
s=1

K X

s
K
1
K
1
= co 1
s
K
s
K 1
s=1
K

K1

1
1
1
K 1 +
= co 1
K
K 1
K 1

K
X

K
kxk1 = K(1 + x)K1
since
k
k=1
38
(2.72)
(2.73)
(2.74)
(2.75)
2.4.

K

K1
K
1
K
= co 1
K
K 1 K 1

(1)K
1
= co 1
K
co e(1)
(2.76)
(2.77)
(for large K)
(1o )
> co o e
(2.78)
(since > o ).
(2.79)
Substituting (2.79) into (2.53) finally yields

X
noX
K1
co o
k k
h
>
(1o )
l=0
e
k
t=1
S G
= (no K 1)
no K
co o
(1
o)
e
co o
e(1o )
(2.80)
which upon setting

no =
(1 2o )e1o k
E |h |
co o
yields

X

hk
k
l=0
Sk G

> K(1 2o ) E |hk |
(2.81)
that is, convergence in expectation to o level, with o independent of

K, is achieved within no K iterations, with no independent of K.
2.4.3. Numerical results

We next present simulation results related to the convergence behavior of the iterative distributed beamforming algorithm described
39
2.
100
90 E
80
"
k
1- BIT
FEEDBACK
|h |
Magnitude
70
60
50
40
"$
!"
"#
"
"
k k "
E "
h [t]"
"
"
30
20
10
0
100
200
300 400 500 600 700

Iteration index t
800
900 1000
Fig. 2.2: Convergence of the iterative algorithm for K = 100.

500
450 E
400
"
k
|hk |
Magnitude
350
300
250
200
150
"$
!"
"#
"
"
k k "
E "
h [t]"
"
"
k
100
50
0
500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Iteration index t
in Section 2.3. In particular, Fig. 2.2, 2.3, 2.4 and 2.5 show how
P
E S k G hk
k [t] (averaged over 50 realizations of the channel co-
40
2.4.
1000
900 E
800
"
|hk |
Magnitude
700
600
500
"$
!"
"#
"
"
k k "
E "
h [t]"
"
"
400
300
200
100
0
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Iteration index t

5000
4500 E
4000
"
k
|h |
Magnitude
3500
3000
2500
"$
!"
"#
"
"
"
E "
hk k [t]"
"
"
2000
1500
1000
500
0
10000
20000
30000
Iteration index t
40000
50000
efficients hk ) evolves as a function of t for K = 100, 500, 1000 and

5000, respectively. In all four cases, convergence occurs within approximately 10K iterations, thereby corroborating the fact that the
41
2.
1- BIT
FEEDBACK
convergence time is linear in K.
2.5.
AC H I EVAB I LITY OF FU LL SPATIAL

MU LTI PLEXI NG GAI N
The aim of this section is to show that performing data transmission,

that is, spatial multiplexing, based on the beamforming weights
ik
obtained from the iterative distributed beamforming algorithm described in Section 2.3 results in a spatial multiplexing gain of M and
a per-stream (distributed) array gain proportional to K. Moreover,
using the framework introduced in Morgenshtern and Bolcskei (2007),
we prove that the network crystallizes, that is, the individual fading
links Gi Di converge to non-fading links as K . Finally, we
quantify the impact of M and o on the crystallization rate, that
is, the rate, as a function of K, at which the fading links Gi Di
converge to non-fading links.
Consider the group Gi and let Ai and Ai denote the set of aligned
and reverse-aligned sources, respectively, corresponding to the beamforming weights
ik . Then, the corresponding input-output relations
for the individual links Gi Di are
X
X
yi [n] = xi [n]
|hki,i |
|hki,i |
i
A
M X
K
X
r=1
r6=i
i
A
hki,r
rk xr [n] + wi [n], i = 1, . . . , M.
k=1
{z
interference
In the following, we assume that for any given realization of the

channel coefficients hki,i , precisely (1 0 )K of the sources are aligned.
Specifically, motivated by convergence in probability of the iterative distributed beamforming algorithm according to (2.22) to o level, we assume that |Ai | = (1 o )K, |Ai | = o K and that both
42
2.5.
ACHIEVABILITY OF FULL SPATIAL MULTIPLEXING GAIN
hki,i , Sik Ai and hki,i , Sik Ai are distributed according to N (0, 1).
We furthermore assume Gaussian codebooks. Under the assumption of the destination Di having perfect knowledge of its effective
P
P
channel coefficient ( Ai |hki,i | A |hki,i |) and of the coefficients
i
PK
k
rk , for all r 6= i, corresponding to the effective interference
k=1 hi,r
channels, the outage probability for the link Gi Di is given by

1
out
log(1 + SINRi ) < Ri
Pi (Ri ) = P
(2.82)
2

= P SINRi < 22Ri 1
(2.83)

P
P
2
P
k
k
i |hi,i |
i |hi,i |
A
K
A
=P
< 22Ri 1
(2.84)
P

P
P
k
k 2 + N
h
o
r
k i,r
r6=i K
P
2
P
k
k
|h
|
|h
|
i,i
i,i
Ai
A
2R
(2.85)
= P P
P
2i KN < 2 i 1 .
k
o
k

h
+
r
r6=i
k i,r
P
To upper-bound the outage probability, we use the (union) bounding techniques introduced in Morgenshtern and Bolcskei (2007). In
particular, we note that employing the union bounds provided in Appendix A.1 for any positive n1 , n2 and n3 with n1 > n2 , the following
upper bound holds:

2
P
P
k
k
i |hi,i |
i |hi,i |
A
(n1 n2 )2
A
P P
P
2 KN <
k k +
o
n3

r
r6=i
k hi,r
P
2
X
X
P
|hki,i |
|hki,i | < (n1 n2 )2
i
A
i
A

2

X X
KNo

hki,r
rk +
+ P
> n3

P
r6=i
(2.86)
43
2.
1- BIT
FEEDBACK
2
X
X k
P
|hki,i | < (n1 n2 )2
|hi,i |
i
A
i
A
2

X
KNo
X
n
3

P
+
P
hki,r
rk >

M 1
r6=i
k

X

X
|hki,i | < n1 + P
|hki,i | > n2
P
i
A
(2.87)
i
A
s
X

KNo
X
n
3

P
+
P
hki,r
rk >

M 1
r6=i
(2.88)
where we implicitly assumed that M > 1 and n3 (KNo )/P . Next,

we use the fact, noted previously at the end of Section III, that
hki,r
rk hki,r , for r 6= i. Consequently, we have
2
P
2
|hki,i | A |hki,i |
(n1 n2 )
P P
P
2i KN <
k
o
k
n3

r + P
r6=i
k hi,r
X

X
k
k
P
|hi,i | > n2
|hi,i | < n1 + P
P
i
A
i
A
i
A

s
X

KNo
n
3

P
.
+ (M 1)P
hki,r >

M 1
(2.89)
Each of the three terms in (2.89) can be upper-bounded using

large deviations bounds (see Shwartz and Weiss (1995)). In particular,
employing
P
S
X
!
|hki,i |
<m
k=1
S
X
k=1
44
!
|hki,i | > m
em
S
r !S
2

S
m2
2e 2S2
2.5.
P
ACHIEVABILITY OF FULL SPATIAL MULTIPLEXING GAIN

S
X
!
hki,r m
m2
e 2S
k=1
from Bounds 1, 2 and 3 provided in Appendix A.2 into (2.89), we get

Piout (Ri )
en1
(1 o )K
r !(1o )K
o K
n2
2
2( 2K)2
o
+ 2e
n3
KNo
P
+ 2(M 1)e 2K(M 1) .
(2.90)
The key to obtaining meaningful upper bounds from (2.90) lies

in a judicious choice of the constants n1 , n2 and n3 ensuring that,
for Ri = (1/2) log(1 + (n1 n2 )2 /n3 ), Piout (Ri ) 0 as K
while
the conditions n1 > n2 and
n3 (KNo )/P . Since
P satisfying

P
E Ai |hki,i | K and VAR Ai |hki,i | K, motivated by the
above large deviations bounds, it is sensible to set
r
(1 o )(K K)
n1 =
.
(2.91)
e
2
P
Similarly, since A |hki,i | deviates around a mean value proportional
i
P
to K with a variance proportional to K and | k hki,r | deviates around
a mean value proportional to K with a variance proportional to K,

again motivated by the above large deviations bounds, we set
p
n2 = 2(1 + ln 2)(o K + K)
KNo
and n3 = (M 1)K 1+ +
(2.92)
P
with the constant > 0. Note that the condition n1 > n2 implies
41
that 0 <
= 0.2419. With the above choices for the
1+e
(1+ln 2)
parameters n1 , n2 and n3 , in the limit K , we get

22Ri 1 =
(n1 n2 )2
n3
45
2.
K 1
(M 1)
1 o
e
1- BIT
FEEDBACK
p
2(1 + ln 2)o
2
, co K 1
2
(2.93)

so that Ri = (1/2) log 1 + co K 1 . Substituting (2.91) and (2.92)
into (2.90), in the large-K limit, we finally obtain

1
Piout
log 1 + co K 1
2

(1o )K
K
1
1
+ eo K K + 2(M 1)e 2
K
= e(1o )
0
+ eo K
+ 2(M 1)e
(as K ).
K
2
(2.94)
(2.95)
We can therefore conclude that Piout (Ri ) 0 as K for any

rate Ri (1/2) log(1 + co K) (recall that can be arbitrarily small).
Since this holds true for all groups Gi , we can choose
Ri =
1
log(1 + ci K)
2
(2.96)
with ci < co , for all i, and get Piout (Ri ) 0, i, as K , so that

M
X
i=1
Ri =
1X
log(1 + ci K)
2 i=1
(2.97)
is achievable, which implies full spatial multiplexing gain of M and a

per-stream array gain proportional to K. Moreover, each of the links
Gi Di converges to a non-fading link.
In summary, in the language of Morgenshtern and Bolcskei (2007),
we can conclude that the network crystallizes as K . The third
term on the RHS of (2.94) nicely reflects the impact of interference on
the crystallization rate. Specifically, for < 1/2, this term dominates
Note that the factor of 1/2 reflects the fact that we considering, for now,
only real-valued signals and channel coefficients.
46
2.6.
TRAINING PERIOD LENGTH SCALING SUB - LINEARLY IN K
the decay rate as a function of K. A smaller corresponds, through

Ri = (1/2) log(1 + co K 1 ), to higher data rates, but results in a
reduced crystallization rate. In the single-user case, that is, for M = 1,
the third term equals zero reflecting the absence of interference. We
can therefore conclude that the crystallization rate in the presence
of interference is significantly smaller than in the single-user case
M = 1. Finally, regarding the proportionality constant co , it can be
observed that the smaller o (that is, the smaller the fraction of reversealigned sources) the larger co , and hence the larger the individual
rates Ri = (1/2) log(1 + ci K) still guaranteeing crystallization. On
the other hand, for smaller o the second term on the RHS of (2.94)
becomes larger, again reflecting that a higher data rate comes at the
cost of increased outage probability.
2.6.
TRAI N I NG PER IOD LENGTH SCALI NG

SU B - LI N EAR LY I N K
We shall next investigate if the iterative beamforming algorithm

presented in Section 2.3, which requires the training period length
to scale linearly in K, can be optimized so that even with a training
period length scaling sub-linearly in K the full spatial multiplexing
gain of M and an array gain proportional to K could be achieved.
It will turn out that a training period length scaling sub-linearly in
K precludes simultaneous achievability of full spatial multiplexing
gain and an array gain proportional to K irrespective of the training
algorithm, and the iterative beamforming algorithm presented in
Section 2.3 is therefore asymptotically optimal. This result can be
formalized as the following theorem:
Theorem 3 In the system model described in Section 2.1, if the
training period length scales as K r with 0 < r < 1, for large K,
then the sum rate achieved with Gaussian codebooks is asymptotically
47
2.
upper-bounded according to
(
P
1
i Ri
lim
K 1 log K
2M r M
2
1- BIT
FEEDBACK
if 0 < r <
if
1
2
1
2M
1
2
1
2M
<r<1
(2.98 )
and therefore full spatial multiplexing gain of M and a per-stream

array gain proportional to K can not be achieved.
Proof: We start by noting that the outage probability for the link
Gi Di is lower-bounded, for any positive n1 , n2 , according to

1
out
Pi (Ri ) = P
log(1 + SINRi ) < Ri
(2.99)
2

= P SINRi < 22Ri 1
(2.100)
!

P
2
P
k
ik
k hi,i
K
=P P
< 22Ri 1
(2.101)

P
P
k
k 2 + N
h
o
r6=i K
k i,r r
!
P k k 2

i
k hi,i
2Ri
<2
1
(2.102)
=P P
P

k k 2 + KNo

r
r6=i
k hi,r
P
P k k 2
2

i
(n1 + n2 )
k hi,i
= P P
(2.103)
P
2 KN < (n1 +n2 )2
k
o
k

h
+
2R
r
i
r6=i
k i,r
(2
1)
P
X

P
hki,i
ik < n1 + n2
k
2
XX

(n1 + n2 )
KNo
2

> 2Ri
P
hki,r
rk +
P
2
1
r6=i
k
(
)

X k k
min P
hi,i
i < n1 + n2
k
i ,k
2
XX

KN
(n
+
n
)
2
o
1
2

> 2Ri
hki,r
rk +
P
P
2
1
r6=i
48
(2.104)
(2.105)
2.6.
where (2.104) follows from the intersection bound for division provided
in Appendix A.3 and the minimization of the first probability term in
(2.105) is over all possible feedback-based training algorithms, where
each training algorithm results in a selection of beamforming weights
ik , k.
We shall next focus on the minimization
(
)
X k k

min P
hi,i
i < n1 + n2
(2.106)
k
i ,k
over all possible feedback-based training algorithms with training

period length scaling as K r , 0 < r < 1. We assume without loss of
generality that
|h1i,i | > |h2i,i | > > |hK
i,i |.
(2.107)
Noting that the feedback rate from Di to all the sources in Gi over
a training period length scaling as K r is upper-bounded, according
to the cut-set bound with cut-set shown in Fig. 2.6, by K r , we then
have
(
)
X k k
hi,i
i < n1 + n2
min P
k
i ,k
r
X
K k
P
hi,i sign(hki,i ) +
K
X

hki,i
ik < n1 + n2

(2.108)
k=K r +1
k=1
r
X
K
=P
|hki,i | +
K
X
k=1
k=K r +1

hki,i
ik < n1 + n2 ,
n1 , n2
(2.109)
where P (
ik = +1) = P (
ik = 1) = 1/2, for k = K r + 1, , K and
(2.108) follows because at most K r realizations of sign(hki,i ) can be
communicated to the sources in Gi by employing K r bits of feedback
Note that the second probability term in (2.105) corresponds to interference
from other groups of sources and would therefore not depend on the training
algorithm.
49
2.
1- BIT
FEEDBACK
Nr
S1
S2
D
SN
Fig. 2.6: Cut-Set upper bound on the feedback rate from Di to all the sources in
Gi .
(recall that P (sign(hki,i ) = +1) = P (sign(hki,i ) = 1) = 1/2, k).

Using the intersection bound for summation, provided in Appendix A.3, (2.109) can be further lower-bounded as
(
)
X k k

min P
hi,i
i < n1 + n2
k
i ,k
r
X
K

P
|hki,i | +
K
X
k=1
k=K r +1
K
X
k=1
!
|hki,i |
< n1
hki,i
ik

< n1 + n2
K
X
k=K r +1
hki,i
ik
(2.110)
!
< n2
(2.111)
Our final goal is to show that any scaling behavior of Ri that violates
(2.98) leads to an outage probability of 1 in (2.105) and thereby
precludes achievability of full spatial multiplexing gain of M along
with a (per-stream) array gain proportional to K. The key to this
result is a judicious choice of n1 and n2 that drives both the probability
50
2.6.
terms in (2.111) to 1. Specifically, setting n1 = K r

with a small positive constant, we obtain
!
Kr
X
P
|hki,i | < n1
2(1 + ) ln K,
k=1
r
=P
K
X
!
|hki,i | < K
p
r
2(1 + ) ln K
(2.112)
k=1

p
P K r |h1i,i | < K r 2(1 + ) ln K

p
= P |h1i,i | < 2(1 + ) ln K

p
= P |hki,i | < 2(1 + ) ln K, k = 1, . . . , K

p
K
= 1 2Q 2(1 + ) ln K

K
1
1 (1+) ln K
e
1 l2
(using Q(l) e 2 , l > 0)
2

K
1
= 1 1+
K
K 1+

K1
1
1
= 1 1+
K
K1
1
=
(for large K)
e
1
as
K .
(2.113)
(2.114)
(2.115)
(2.116)
(2.117)
(2.118)
(2.119)
(2.120)
(2.121)
(2.122)
PK
Moreover, since k=K r +1 hki,i
ik follows a Gaussian distribution with
mean 0 and a standard deviation that scales as K K r K (for

1
0 < r < 1), setting n2 = K 2 + , with small, we obtain
!
K

X

k
k

P
hi,i
i < n2
k=K r +1
51
2.
=P
K
X
k=K r +1
=1P
FEEDBACK
!

1
hki,i
ik < K 2 +
K
X
k=K r +1
(2.123)
!

1
hki,i
ik K 2 +
(2.124)
K
X
= 1 2P
1- BIT
hki,i
ik K
1
2 +
(2.125)
k=K r +1
1
K 2 +
K Kr
= 1 2Q
K 1+2
1 e 2(KK r )
1 l2
(using Q(l) e 2 )
2
=1e
1
K 2
1
K 1r
2 1
as
K .
Substituting (2.122) and (2.129) into (2.111), we obtain

(
)
X k k

min P
hi,i
i < n 1 + n2
1
k
i ,k
(2.126)
(2.127)
(2.128)
(2.129)
(2.130)
which, upon substitution into (2.105), yields

Piout (Ri )
2
XX

(n1 + n2 )
KNo
2

> 2Ri
P
hki,r
rk +
P
2
1
r6=i
k
XX
2
(n1 + n2 )2

P
hki,r
rk > 2Ri
2
1
r6=i
k
1
r
+ 2
2(1+)
XX

2
(K ln K
+K
)
2

=P
hki,r
rk >
22Ri 1
r6=i
52
(2.131)
(2.132)
(2.133)
2.6.
!!M 1
1
X k k 2
(K r ln K 2(1+) + K 2 + )2

hi,r
r >
P
(2.134)
(22Ri 1)(M 1)
k
(from intersection bound for summation, Appendix A.3)

!!M 1
1
X
K r ln K 2(1+) + K 2 +
k
k
= 2P
hi,r
r > p
(22Ri 1)(M 1)
k
!!M 1
1
K r ln K 2(1+) + K 2 +
= 2Q
1
K 2 M 1 22Ri 1
(since
K
X
(2.135)
(2.136)
hki,k
kk N (0, K), for k 6= i)
k=1
= 2Q
ln K 2(1+)
1
2Ri K 2
+K
q
M 1 1
1
2 +
M 1
(2.137)
22Ri
However, for any Ri with limK

< , so that
Ri
log K
= > 0, we can choose
lim Piout (Ri )

M 1
r 12
2(1+)
K
ln
K
q
lim 2Q
K
1
R
i
2
M 1 1 22Ri
(2.138)
and therefore if
r 12
2Ri
as
ln K 2(1+)
(2.139)
the outage probability is equal to one according to

lim Piout (Ri ) (2Q(0))M 1 = 1.
(2.140)
53
2.
1- BIT
FEEDBACK
Consequently, the outage probability is less than one only if
r 12
2Ri
c
ln K 2(1+)
(for large K)
(2.141)
for some finite constant c independent of K. We shall now consider

the following three cases:
Case 1 0 < r < 12 : The outage probability in this case will always
Ri
be one because (2.141) cannot be satisfied for limK 1 log
>
K
2
0. The best strategy might be to simply perform time-sharing so
that
PM
Ri
=1
(2.142)
lim 1 i=1
K
log
K
2
that is, spatial multiplexing gain of one and an array gain proportional to K is achieved.
1
: In this case, Piout (Ri ) is less than one
Case 2 12 < r 12 + 2M
only if
2Ri
c
K ln K 2(1+)

p
1
Ri r
log K + log(c ln K 2(1+) )
2
Ri
lim 1
< 2r 1
K
2 log K
PM
Ri
lim 1 i=1
< 2M r M 1
K
2 log K
r 12
(2.143)
(2.144)
that is, at most a spatial multiplexing gain of one can be achieved

with an array gain proportional to K.
1
Case 3 12 + 2M
< r < 1: As in Case 2, Piout (Ri ) is less than one
only if
PM
lim
K 1
2
54
Ri
< 2M r M.
log K
i=1
(2.145)
2.7.
EXTENSION TO COMPLEX -VALUED CHANNELS
1
Therefore, if 12 + 2M
< r < 1, the spatial multiplexing gain in
the sum-rate is upper-bounded by 2M r M for an array gain
proportional to K.
Combining the above three cases completes the proof.
2.7.
EXTENSION TO COMPLEX - VALU ED

C HAN N ELS
We next generalize the results derived so far under the assumption of

both channel coefficients as well as signals being real-valued to the case
where both channel coefficients and signals are complex-valued. The
generalization entails suitable modifications to the iterative distributed
beamforming algorithm and a corresponding proof of achievability
of both full spatial multiplexing gain of M and a per-stream array
gain proportional to K. The key idea is that coherent combining
along the real part of the effective channel Gi Di is sufficient for
achievability of both full spatial multiplexing gain and an array gain
proportional to K. Specifically, the iterative distributed beamforming
algorithm discussed for real-valued channel coefficients is employed
for complex-valued channel coefficients as well by simply ignoring the
imaginary parts of the received signals. The details are as follows.
2.7.1. Modified iterative beamforming algorithm

We present only the ways in which the modified iterative distributed
beamforming algorithm differs from the algorithm discussed previously. The key differences are as follows:
Although all channel coefficients are complex-valued, the pilotsymbol
used during the training phase is still real-valued, namely,
p
P/K. We shall also recall that beamforming weights in the
system model are also real-valued, that is, ik {1, +1}.
As in Section 2.3, we consider, w.l.o.g., the group G , drop the index i = 1
1
and, wherever appropriate, use the convention G G1 , S k S1k , D D1 , k =
k1 ,
k =
k1 and hk = hk1,1 .
55
2.
1- BIT
FEEDBACK
In Steps 1 and 2 of the iterative algorithm, the destination Di

measures only the real part of the received signal level, that is,
(
)
1 X
rx
L =R
y[n]
Tf
nFl
(r
)
K
P X k k
=R
h [l] .
K
k=1
It trivially follows from our analysis in Section 2.4 that the

modified algorithm achieves the following convergence
R
nX
k
o
hk
k [t]
X
tno K
|R{hk }|.
(2.146)
We can furthermore infer that the modified iterative algorithm

converges (both in probability and in expectation) within a
training period length scaling linearly in K. Performing data
transmission based on the beamforming weights obtained by
this algorithm results in the desired transmitted signals combining coherently along the real part of the effective channel with
incoherence restricted to the interval [/2, /2], as shown in
Fig. 2.7.
The modified iterative algorithm is performed by each of the
Gi Di pairs, so that at the end of no M K iterations, each of
the groups Gi is in beamforming configuration in the sense that
desired signals combine coherently along the real part of the
effective channel Gi Di .

The aim of this section is to show that performing data transmission,
that is, spatial multiplexing, based on the beamforming weights
ik
produced by the modified iterative distributed beamforming algorithm results in a spatial multiplexing gain of M and a per-stream
(distributed) array gain proportional to K. Moreover, we prove that
56
2.7.
Fig. 2.7: Residual channel incoherence is limited to [/2, /2]
the network crystallizes, that is, the individual fading links Gi Di

converge to non-fading links as K . To this end, we first note
that considering a single rate of transmission, say on the link Gi Di ,
and denoting by Ai and Ai the sets of aligned and reverse-aligned
sources respectively, we get the effective input-output relation
X
X
|R{hki,i }|
yi [n] =
|R{hki,i }| xi [n]+
i
A
K
XX
r6=i
hki,r
rk
i
A

xr [n] +
k=1
X
K
I{hki,i }
ik

xi [n] + wi [n],
(2.147)
k=1
{z
interference
where |Ai | = (1 o )K and |Ai | = o K. We implicitly consider, in

(2.147), the symbol received along the imaginary part of the effective
channel Gi Di to be interference because doing so cannot increase
the achievable rate. Under the assumption of the destination Di having
P
perfect knowledge of its effective channel coefficient ( Ai |R{hki,i }|
57
2.
1- BIT
FEEDBACK
|R{hki,i }|) and of the effective interference channel coefficients

PK
k
rk , r 6= i, and k=1 I{hki,i }
ik , the outage probability for
k=1 hi,r
the link Gi Di is given by
P
Ai
PK
Piout (Ri )
= P (log(1 + SINRi ) < Ri )
(2.148)

Ri
= P SINRi < 2 1
2
P
P
!
P
k
k
i |R{hi,i }|
i |R{hi,i }|
A
K
A
=P P
< 2Ri 1
P
P

P
k
k }
k 2 + 2N
k 2 + P
h
I{h
o
i,i
i
r6=i K
k i,r r
k
K
2
P
P
!
k
k
i |R{hi,i }|
i |R{hi,i }|
A
A
Ri
<2 1 .
=P P
P
P

k k 2 +
k k 2 + 2KNo

r
i
r6=i
k hi,r
k I{hi,i }
P
In a manner similar to what we did for the real-valued case, for any
positive n1 , n2 and n3 such that n1 > n2 , we can obtain the following
upper bound:
2
P
!
|R{hki,i }| A |R{hki,i }|
2
(n
n
)
1
2
i
<
P P
P

P

k
k }
k 2 + 2KNo
k 2 +
n3

I{h
h
i,i i
r6=i
k i,r r
k
P
X

X

|R{hki,i }| > n2
P
|R{hki,i }| < n1 + P
P
i
A
i
A
i
A

2
2
X

X X
2KN

o
(2.149)
+ P
hki,r
rk +
I{hki,i }
ik > n3

P
r6=i
where we implicitly assumed that n3 (2KNo )/P . Next, using the

fact that hki,r
rk hki,r , for r 6= i and I{hki,r }
rk I{hki,r }, we obtain
2
P
|R{hki,i }| A |R{hki,i }|
P

P i

P
k
k k 2 +
k 2 +

h
i
r6=i
k i,r r
k I{hi,i }
P
58
i
A
2KNo
P
(n1 n2 )2
<
n3
2.7.
P
|R{hki,i }|

< n1
+P
i
A
X
!
|R{hki,i }|
> n2
i
A

2

2

X X
X X

+P
R{hki,r } +
I{hki,r }

r6=i
k
r6=i
k
2

!

X
2KNo

k
I{hi,i } > n3
+

P
k
X

X
k
k
|R{hi,i }| > n2
P
|R{hi,i }| < n1 + P
i
A
(2.150)
i
A
s
X

2KNo
n3 P

+ (2M 1)P
I{hki,r } >

2M 1
(2.151)
where (2.151) follows from the union bound for summation in Appendix A.1 and the fact that I{hki,i } I{hki,r } R{hki,r }, r 6= i. Each
of the three terms in (2.151) can be upper-bounded using large deviations bounds of Shwartz and Weiss (1995). In particular, employing
(from (A.10),(A.17) and (A.18))
S
X
!
|R{hki,i }|
<m
i=1
S
X
!
|R{hki,i }|
>m
em
S
r !S
2
S

m2
2e 2S2
i=1
S
X
!
I{hki,i } m
m2
e 2S
i=1
in (2.151), we get
Piout (Ri )
en1
(1 o )K
r !(1o )K
o K
n2
2
2( 2K)2
o
+ 2e
59
2.

n3
1- BIT
FEEDBACK
KNo
P
+ 2(2M 1)e 2K(M 1) .
(2.152)
Choosing the same n1 and n2 as in Section 2.5, choosing n3 =

(M 1)K 1+ +(2KNo )/P (rather than n3 = (M 1)K 1+ +(KNo )/P,
as in Section 2.5) and following the same analysis, we get

Piout log 1 + co K 1
e
0
+ eo K
+ 2(2M 1)e
(as K ).
K
2
(2.153)
(2.154)
We can therefore conclude that Piout (Ri ) 0 as K for any

rate Ri log(1 + co K) (recall that can be arbitrarily small). Since
this holds true for all groups Gi , we can choose
Ri = log(1 + ci K)
(2.155)
with ci < co , for all i, and get Piout (Ri ) 0, for all i, as K , so
that a sum rate of
M
X
i=1
Ri =
M
X
log(1 + ci K)
(2.156)
i=1
is achieved, which is equivalent to full spatial multiplexing gain of M ,

a per-stream array gain proportional to K, and convergence of each
of the links Gi Di to a non-fading link.
In summary, in the language of Morgenshtern and Bolcskei (2007),
we can conclude that the network crystallizes as K . As in
the analysis of the real-valued case, the third term on the RHS of
(2.153) nicely reflects the impact of interference on the crystallization
rate. Specifically, for < 1/2, this term dominates the decay rate as a
function of K. A smaller corresponds, through Ri = log(1+co K 1 ),
to higher data rates, but results in a reduced crystallization rate. Furthermore, the dominant term in the outage probability in (2.153) has
a pre-factor of 2(2M 1). It can be shown that if coherent combining
takes place along the imaginary part of the effective channels as well
60
2.7.
(apart from the real part), then this pre-factor would be reduced
to 2(2M 2). However, to achieve coherent combining along the
imaginary part as well as the real part, the training period length
would have to be doubled. Thus, there is a clear tradeoff between
outage probability and length of the training period.
2.7.3. Training period length scaling sub-linearly in K

Our results in Section 2.6 for training period length scaling sub-linearly
in K extend to the case of complex-valued channels and signals in a
straightforward manner. Specifically, the outage probability for the
link Gi Di is lower-bounded according to
Piout (Ri )
= P (log(1 + SINRi ) < Ri )
(2.157)

Ri
= P SINRi < 2 1
(2.158)
!
P k k 2
P
i
k hi,i
K
=P P
< 2Ri 1
(2.159)

2
P
P
k
k

h
+
2N
o
r
i,r
r6=i K
k
!
P k k 2

i
k hi,i
Ri
<2 1
(2.160)
=P P
P

k
k 2 + 2KNo

h
r
i,r
r6=i
k
P
!
P

P
2
k
k 2
k

i +
ik
k R{hi,i }
k I{hi,i }
Ri
<2 1
=P
P

P
k
k 2 + 2KNo

h
r
r6=i
k i,r
P
!
P

k
k 2

i
2Ri 1
k R{hi,i }
P P
<
P

k
k 2 + 2KNo
2

h
r
r6=i
k i,r
P
!
P

2
k

ik
2Ri 1
k I{hi,i }
P P
<
(2.161)
P

k
k 2 + 2KNo
2

h
r
r6=i
k i,r
P
!
P

k
k 2

i
1
k R{hi,i }
Ri 1
P P
<2
P
2
2

hk
k + 2KNo
r6=i
i,r
61
2.
1- BIT
FEEDBACK
!
P
2
k

ik
1
k I{hi,i }
< 2Ri 1
P

P
k
k 2 + 2KNo
2

h
r6=i
k i,r r
P
(2.162)
where (2.161) follows from the intersection bound for summation,

provided in Appendix A.3. Setting 2Ri = Ri 1, neglecting the
additive term of 1/2, and noting that

2
2

2

X X
X X
X X
k
k
k
k
k
k
I{hi,r }
r
R{hi,r }
r +
hi,r
r =

r6=i
r6=i
r6=i
P
with R{hki,r }
rk k I{hki,r }
rk , r 6= i, our analysis in Section 2.6
(from (2.102) onwards) implies that with a training period length
scaling as K r (for r < 1), each of the two probability terms in (2.162)
is lower-bounded by 1 (that is, outage probability is one) unless the
sum rate satisfies
(
P
1
1
if 0 < r < 21 + 2M
i Ri
lim
(2.163)
1
K log K
2M r M if 12 + 2M
<r<1
and hence full spatial multiplexing gain along with a per-stream array
gain proportional to K cannot be achieved with a training period
length scaling sub-linearly in K.
2.8.
OBSERVATIONS
In this chapter, a large interference network was considered where M

groups of K non-cooperating sources each are dedicated to transmit
M independent messages to M non-cooperating destinations. An iterative feedback-based multiuser beamforming algorithm was proposed
and it was shown that with this algorithm, the signals received at
each destination from its corresponding group of sources combine
coherently and the interference from the non-corresponding groups
of sources combine incoherently. This coherent combining, in turn,
results in the signal power received at each destination scaling as K 2 ,
62
2.8.
OBSERVATIONS
approximately, in the large K regime while the interference power

scales only as K, approximately. The signal-to-interference and noise
ratio at each destination therefore scales as K, which implies that full
multiplexing gain of M and a per-stream array gain proportional to K
is achieved. The M links between the M groups of sources and their
corresponding destinations, moreover, turn from fading to non-fading
links. Interestingly, coherent combining of the desired signals along
the real part was found to be sufficient for all the above results to
hold. Finally, the iterative algorithm was found to be asymptotically
optimal in the sense that the feedback rate scaling required for the
algorithm is the minimum required for achievability of full spatial
multiplexing gain.
63
C HAPTER
SISO Interference Networks with

Broadcast Feedback
ingle-input single-output (SISO) interference networks

consist of M single-antenna sources, S1 , , SM , communicating with M single-antenna destinations, D1 , , DM , as shown
in Fig. 3.1. Each source Si wants to communicate an independent
message to its corresponding destination Di and does not care about
what the other destinations receive or understand. The symbols transmitted by each source Si obey an average power constraint of P/M ,
so that the sum transmit power of the sources is P . Denoting the rates
of communication over the link Si Di by Ri , the capacity region is
defined as the convex closure of the set of rates R = {R1 , , RM }
for which jointly reliable communication is possible.
Despite the apparent simplicity of this network, the capacity region
is not known in general. In the large P regime, however, the capacity
region is more tractable. Specifically, it is trivial to note that the
sum capacity of the network scales at least as log P (achieved with
time-division multiplexing) and at most at M log P (achieved if full
cooperation is allowed among all the sources and all the destinations), in the large P regime. The maximum spatial multiplexing
gain achievable in this network therefore lies somewhere between 1
and M (including both the numbers). In a preliminary result, Host-
65
3.
SISO INTERFERENCE NETWORKS WITH BROADCAST FEEDBACK
S1
D1
S2
D2
DM
SM
Fig. 3.1: A SISO interference network with M source-destination pairs where

each source and each destination is equipped with a single antenna. Principal
links are represented in dark lines and interference links in dotted lines.
Madsen and Nosratinia (2005) showed that spatial multiplexing gain

is actually upper-bounded by M/2. The proof of this upper bound
is relatively straightforward. The authors first consider two random
(distinct) source-destination pairs, say {Si , Di } and {Sk , Dk }. Assuming that all the other source-destination pairs do not communicate, it
was shown that the following bound holds
lim
Ri + Rk
1.
log P
(3.1)
It then follows that if all other source-destination pairs communicate,

the bound (3.1) still holds because interference from other pairs can
only reduce the rate of communication. Moreover, summing up the
bound (3.1) for all possible {Ri , Rk } pairs yields
PM P
lim
66
i=1
k6=i (Ri
log P
+ Rk )
M (M 1)
(3.2)
PM
2(M 1) i=1 Ri
M (M 1)
P
log P
PM
M
i=1 Ri
lim
P log P
2
lim
(3.3)
(3.4)
that is, spatial multiplexing gain is upper-bounded by M/2 in the

limit of large P .
The fact that spatial multiplexing gain of M/2 is indeed achievable,
was established by Cadambe and Jafar (2008). To be precise, Cadambe
and Jafar (2008) proposed an innovative transmission scheme, called
interference alignment (IA), and showed that performing spatial multiplexing according to this scheme results in full spatial multiplexing
gain of M/2 in the large P regime. The key idea behind IA is that
transmission is coded over N time-domain dimensions and each source
Si precodes its transmission in such a manner that all the interference
at any given destination Di is restricted to only (approximately) N/2
of the N dimensions. Thus, each link Si Di sees at least N/2
interference-free dimensions so that its rate of communication scales
as (1/2) log P and the sum rate of communication of the network
scales as (M/2) log P .
The results of Cadambe and Jafar (2008) depend, however, critically
on the assumptions that all channels in the network are time-selective
and that each node (that is, each source and each destination) knows
all the channels in the network perfectly. However, to satisfy the
assumption of perfect channel knowledge at all the nodes in a timeselective channel would require not only infinite-capacity feedback
links, but also non-causal feedback. IA, nonetheless, does not distinguish between time and frequency dimensions. Therefore, although IA
was originally developed for time-selective frequency-flat channels, it
can easily generalized for time-flat frequency-selective channels. This
generalization was provided by Grokop et al. (2009), thereby obviating
the need for non-causal feedback, but still requiring infinite-capacity
For definitions of time-flat, time-selective, frequency-flat and frequencyselective channels, please refer to Appendix B.
67
3.
feedback links. In this chapter, we show that the assumption of

infinite-capacity feedback links can be relaxed to having an error-free
broadcast feedback link of rate M (L 1) log P from each destination to all the other nodes (that is, all the sources and all the other
destinations) in the network.
The rest of this chapter is organized as follows. After providing a
system model for SISO networks in Section 3.1, we provide a review of
the IA scheme in Section 3.2. We then show in Section 3.3 that naive
IA, based on vector quantized channel impulse responses obtained
using the vector quantization scheme proposed by Mukkavilli et al.
(2003) and Love et al. (2003), results in full spatial multiplexing gain
of M/2, provided that each destination can feedback (broadcast) at a
rate of M (L 1) log P to all the other nodes in the network. Finally,
we extend this result to MIMO interference networks in Section 3.4.
3.1.
SYSTEM MODEL
The L-tap impulse response of the frequency-selective single-input

single-output (SISO) channel between source Sk , k = 1, . . . , M, and
destination Di , i = 1, . . . , M, is given by hi,k [l], l = 0, . . . , L 1.
The channel coefficients hi,k [l] remain constant throughout the time
interval of interest (outage setting) and are drawn independently
(across i, k, l) from a single continuous probability density function
such that 0 < |hi,k [l]| < , i, k, l, with probability 1. We use a cyclic
signal model (such as in orthogonal frequency division multiplexing) to
convert the frequency-selective channel Sk Di , k, i, into N (with
N L) parallel frequency-flat channels, where
hi,k (r) = Fr {hi,k [n]},
r = 0, , N 1
(3.5)
is the channel coefficient between source Sk and destination Di , corresponding to the r-th tone. The input-output relation between Si
68
3.1.
SYSTEM MODEL
and Di , for the r-th tone, is then given by

X
yi (r) = hi,i (r)xi (r) +
hi,k (r)xk (r) + zi (r)
(3.6)
k6=i
{z
interference
where yi (r) is the symbol received at destination Di , xk (r) denotes

the transmit symbol for source Sk and zi (r) is CN (0, No ) noise at Di ,
all for the r-th tone. Defining
y
i , [yi (0) . . . yi (N 1)]T
x
i , [xi (0) . . . xi (N 1)]T
zi , [zi (0) . . . zi (N 1)]T

i,k , diag{hi,k (0), . . . , hi,k (N 1)}
H
the input-output relation (3.6) can be rewritten as
X
H
H
y
i = H
i +
H
k +
zi .
i,i x
i,k x
(3.7)
k6=i
The transmit signals obey the power constraints

P
, k = 1, . . . , M, r = 0, . . . , N 1.
E |xk (r)|2
M
(3.8)
Finally, we shall also need the channel vector hi,k and the normalized
channel vector wi,k corresponding to the link Sk Di , defined as
hi,k = [hi,k [0] hi,k [L 1]]T CL1 and wi,k = hi,k /khi,k k
CL1 , respectively.
We assume that each destination Di knows its channels from each
of the sources Sk perfectly, that is, Di knows hi,k , k. We further
assume that there exist dedicated non-interfering error-free broadcast
feedback links from each destination Di to all the other nodes in
the network, that is, to the sources Sk , k, and to the destinations
Dk , k 6= i. In the remainder of this chapter, we distinguish between a
We
conjugate the channel coefficients for notational simplicity later on.
69
3.
channel feedback phase during which Nf bits of feedback are broadcast

by each destination and a data transmission phase following the
channel feedback phase. The channels, being deterministic, are fed
back only once during the entire time interval of interest so that
the transmission rate loss due to the channel feedback phase can be
PM
assumed to be negligible. Defining Rsum , i=1 Ri , we say that full
spatial multiplexing gain is achieved if
lim
Rsum
M
=
.
log P
2
(3.9)
Recall that spatial multiplexing gain is upper-bounded by M/2, even

if each node knows all the channels in the network perfectly. Our
main result in this chapter can now be formalized as follows.
Theorem 4 In the SISO interference network defined above, naive
interference alignment, in conjunction with the channel vector quantization scheme proposed in Mukkavilli et al. (2003), achieves full
spatial multiplexing gain of M/2, provided that the number of feedback
bits Nf broadcast by each destination to all the sources and all the
other destinations is at least M (L 1) log P .
Before proceeding to the proof of the above theorem, we first briefly
review the concept of interference alignment (IA) by adapting the
main results of Cadambe and Jafar (2008) to our setup.
3.2.
I NTER FER ENC E ALIGNMENT WITH

PER FECT CSI AT ALL NODES
IA entails that each source Sk transmits a linear combination of dk

scalar symbols, x1k , . . . , xdkk , in N frequency slots by modulating the
symbols onto the transmit direction vectors vk1 , . . . , vkdk , that is,
x
k =
dk
X
m=1
70
vkm xm
k ,
k = 1, . . . , M
(3.10)
3.2.
INTERFERENCE ALIGNMENT WITH PERFECT CSI AT ALL NODES

m
N 1
2
=
where xm
with kvkm k2 = 1, and E |xm
k C, vk C
k |
P/(M dk ), k, m. Setting Q = (M 1)(M 2) 1, the number of
data symbols dk (corresponding to Sk ) and the number of tones N
are chosen according to (Cadambe and Jafar, 2008, Appendix III)
(
(t + 1)Q , k = 1
(3.11)
dk = Q
t ,
k = 2, . . . , M
N = (t + 1)Q + tQ
(3.12)
where t is an auxiliary variable and the choice of S1 to transmit

(t + 1)Q symbols in N frequency slots, in contrast to tQ symbols
for the other sources, is without loss of generality. Each destination
Di computes the projections of its received signal y
i onto di receive
P
direction vectors u1i , . . . , udi i resulting in a total of i di effective
input-output relations given by
X
H
H H m m
H H p p
(um
i = (um
(um
i ) y
i ) Hi,i vi xi +
i ) Hi,i vi xi
p6=m
|
+
dk
XX
{z
interference
m H
H H p p
i
(um
i ) Hi,k vk xk +(ui ) z
}
(3.13)
k6=i p =1
{z
interference
N 1
2
for m = 1, . . . , di , i = 1, . . . , M, where um
with kum
i C
i k =
1, i, m. Choosing xm
k , k, m, to be i.i.d. Gaussian, treating the two
interference terms in (3.13) as additional noise, and assuming that
H H m
Di knows the effective channel coefficient (um
i ) Hi,i vi , m, perfectly, the rate of communication over link Si Di is lower-bounded
We will later let t and see that the difference between the multiplexing
gain achieved by IA and the maximum possible multiplexing gain can be made
arbitrarily small.
71
3.
according to
di
1 X
Ri
log 1 +
N m=1
P
m H H m 2
M di |(ui ) Hi,i vi |
Ii,1 + Ii,2 + No
(3.14)
with
Ii,1 =
X P

H H p 2
(um
i ) Hi,i vi
M di
(3.15)
p6=m
Ii,2 =
dk
XX
k6=i p
P m H H p 2
(ui ) Hi,k vk
M dk
=1
(3.16)
where we assumed that Ii,1 and Ii,2 are also known at Di .

Based on their respective channel knowledge, each source Sk computes transmit direction vectors vkm , m = 1, . . . , dk , and each destination Di computes receive direction vectors um
i , m = 1, . . . , di , that
together satisfy the following three sets of conditions:
H H m
|(um
i ) Hi,i vi | c > 0,
H H p
(um
i ) Hi,i v = 0,
i
H H p
(um
)
H
v
i
i,k k
= 0,
i, m
(3.17)
i, m 6= p
(3.18)
k 6= i, m, p
(3.19)
with the constant c independent of P . It then follows that Ii,1 =

Ii,2 = 0, i, and the spatial multiplexing gain achieved by IA is
lower-bounded according to

m H H m 2
P
di
M P
P
M di |(ui ) Hi,i vi |
log 1 +
No
Rsum
m=1
lim i=1
lim
P
P log P
N log P
PM
di
= i=1
(3.20)
N
(t + 1)Q + (M 1)tQ t M
=
(3.21)
(t + 1)Q + tQ
2
that is, full spatial multiplexing gain, in the sense of (3.9), is achieved.
72
3.2.
INTERFERENCE ALIGNMENT WITH PERFECT CSI AT ALL NODES
Under the assumption of every node (i.e., every source and every destination) knowing all the channels in the network perfectly, one way
p
to find vectors um
i , vi satisfying (3.17), (3.18) and (3.19) is provided
in (Cadambe and Jafar, 2008, Appendix III) (or see Appendix A.4
of this thesis). The basic idea is that each Sk computes, based on
its knowledge of all the channels in the network, a set of linearly
independent transmit direction vectors vk1 , . . . , vkdk such that all the
vectors corresponding to interference from Sk , k 6= i, at Di (that is, the
H vp , k 6= i, p = 1, . . . , dk ) span an (N di )-dimensional
vectors H
i,k k
complex subspace of CN . Consequently, di dimensions remain completely interference-free. Each Di , in turn, computes, based on its
knowledge of all the channels in the network, a set of di unit-norm
receive direction vectors u1i , . . . , udi i that spans the di -dimensional
interference-free subspace corresponding to the link Si Di , thereby
satisfying (3.19). Moreover, it was shown in (Cadambe and Jafar,
H vm , m, along with the
2008, Appendix III) that if the vectors H
i,i i
p
H v , k 6= i, p, span CN , then Di can choose the di receive
vectors H
i,k k
direction vectors um
i , m, such that along with (3.19), both (3.17) and
(3.18) are satisfied as well. It turns out that, in the frequency-selective
case, this is possible provided that L > ((t + 1)Q 1)/(3tQ) (the
proof of this statement is similar to (Grokop et al., 2009, Th. 6.4)
and the details are provided in Appendix A.5 of this thesis).
The developments in the remainder of this chapter are based on
the simple observation that if the interference power terms Ii,1 and
Ii,2 , for all i, are not equal to zero, but upper-bounded by a constant
independent of P , full spatial multiplexing gain is still achieved. The
key to realizing this will be a vector quantization scheme, which
H H p 2
satisfies (3.17) and ensures that both |(um
i ) Hi,i vi | , i, m 6= p,
m H H p 2
and |(ui ) Hi,k vk | , k 6= i, m, p, scale as 1/P when P .
It will turn out that the vector quantization scheme developed in
Mukkavilli et al. (2003) and Love et al. (2003) for beamforming in
single-user frequency-flat multi-input multi-output (MIMO) channels
satisfies this condition.
73
3.
3.3.
I NTER FER ENC E ALIGNMENT WITH

LIMITED FEEDBAC K
We start by recalling that each destination Di knows the channel

coefficient vectors hi,k , k, (and hence, wi,k , k) perfectly. Knowledge
of wi,k , k, at all the sources and all the other destinations is obtained
through feedback. Specifically, each Di broadcasts, during the channel
feedback phase, quantized versions of wi,k , k, to Sk , k, and Dk , k 6=
i. We shall next describe the vector quantization and feedback scheme.
3.3.1. The vector quantization and feedback scheme

The vector quantization scheme works on unit norm vectors and
quantizes the vector wi,k CL1 to the unit norm vector w
i,k CL1
using Nd bits. The corresponding quantizer codebook A therefore
contains 2Nd vectors, that is, A = {w
1, . . . , w
2Nd }. The quantization
policy is as follows:
H

l wi,k | .
(3.22)
w
i,k = arg max |w
w
l A
The quantizer codebook is chosen as the solution of the following

Grassmannian line-packing problem by Conway et al. (1996): Find
the maximum number of unit-magnitude vectors in CL such that the
absolute value of the inner product between any two of the vectors is
less than cos(), where (0, /2]. Here, is an auxiliary variable
whose role will become clear later.
The line-packing problem: If Npack is the solution to the line-packing
problem and {p1 , . . . , pNpack } is a set of vectors corresponding to this
solution, we choose
2Nd = Npack
and set
{w
1, . . . , w
2Nd } = {p1 , . . . , pNpack }.
(3.23)
For simplicity of exposition, we assume that N can take on non-integer

d
values.
74
3.3.
INTERFERENCE ALIGNMENT WITH LIMITED FEEDBACK
This approach was used in Mukkavilli et al. (2003) and Love et al.
(2003) for beamforming in single-user MIMO channels.
Quantization error
We define the quantization error d (wi,k , w
i,k ) as
q
H w
d (wi,k , w
i,k ) , 1 |wi,k
i,k |2 .
The maximum quantization error max
is then given by
d
q
max
=
max
1 |xH w
x |2
d
xCL ,kxk=1
(3.24)
where w
x A is the unit-magnitude quantized version of x CL
obtained according to (3.22).
We will need an upper bound on max
in terms of Nd . While such
d
a bound is known in the literature (see Mukkavilli et al. (2003)), we
will provide a derivation, partly for completeness, and partly to get
the bound in a form required for our proof. We start by noting the
following two key properties of the chosen set of quantization vectors:
i) The following relation holds between Nd and sin() (see (Love
et al., 2003, Th. 3) or Appendix A.6 of this thesis):
2Nd
sin() 2
sin()
2
2(L1)
1
Nd
!
.
(3.25)
2 2(L1)
ii) The maximum quantization error max
is upper-bounded by
d
sin(). This can be proved by contradiction as follows:
Let us assume that max

> sin().
d
Further, let xo be the unit-magnitude vector, quantized to
w
xo , that corresponds to the maximum quantization error,
75
3.
that is,
q
xo = arg max
1 |xH w
x| .
(3.26)
xCL ,kxk=1
Then, we have
q
2
1 |xH
xo | > sin()
o w

2
xo < 1 sin2 () = cos2 ()
xH
o w

xH
xo < cos()
o w

Nd

xH
.
o pl < cos(), l = 1, . . . , 2
(3.27)
(3.28)
(3.29)
(3.30)
However, (3.30) implies that in the line-packing problem,

there exists a solution set of vectors {p1 , . . . , p2Nd , xo }, such
that the absolute value of the inner product between any
two vectors is less than cos(). In other words, 2Nd (and
consequently Npack ) is not the maximum possible number
of unit-magnitude vectors with the absolute value of the
inner product between any two vectors being less than
cos() and can hence not be the solution of the line-packing
problem, which results in a contradiction. Thus, our premise
max
> sin() must be incorrect and we must necessarily
d
have
max
sin().
d
(3.31)
Inserting (3.31) into (3.25), we get the desired upper bound:

!
1
max
.
(3.32)
d 2
Nd
2 2(L1)
Number of feedback bits During the channel feedback phase, each

destination Di broadcasts Nd bits for each w
i,k to all sources and
destinations (except to itself, of course) in the network, resulting in a
76
3.3.
total of
Nf = M Nd
bits
(3.33)
being broadcast by each destination. Each source Sk therefore receives

a total of M 2 Nd bits of (error-free) feedback from all the destinations
and each destination Di receives a total of (M 1)M Nd bits from the
destinations Dk , k 6= i. Each source and each destination can therefore
recreate the quantized normalized channel vectors w
i,k , i, k, and the
key is to determine a value of Nd that ensures the achievability of full
spatial multiplexing gain with naive IA based on w
i,k , i, k.
3.3.2. Transmission scheme and achievability of full

spatial multiplexing gain
Each source and each destination first converts its received quantized
L-dimensional vectors w
i,k , i, k, into N -dimensional vectors [w
i,k [0]
... w
i,k [N 1]]T through zero-padding. It then computes the N -point
DFTs
w
i,k (r) = Fr {w
i,k [n]},
r = 0, . . . , N 1
(3.34)
and organizes the results into the quantized channel matrices

c i,k = diag{w
W
i,k (0), . . . , w
i,k (N 1)}.
(3.35)
c i,k , i, k, that
i,k = W
IA is now performed naively assuming that H
is, each source Sk computes its transmit direction vectors v
km , m =
1, . . . , dk , and each destination Di computes its receive direction
c
vectors u
m
i , m = 1, . . . , di , from Wi,k , i, k (rather than from Hi,k ).
1
Sk transmits a linear combination of dk scalar symbols, xk , . . . , xdkk ,
in N frequency-slots by modulating the symbols onto the vectors
v
k1 , . . . , v
kdk , that is,
x
k =
dk
X
v
km xm
k ,
k = 1, . . . , M
(3.36)
m=1
77
3.

2
=
where xm
km CN 1 with k
vkm k2 = 1, and E |xm
k C, v
k |
P/(M dk ), k, m. The number of data symbols dk (corresponding
to Sk ) and the number of tones N are chosen according to (3.11) and
(3.12), respectively. Each destination Di computes the projections of
its received signal y
i onto the receive direction vectors u
1i , . . . , u
di i
P
resulting in a total of i di effective input-output relations given by
X
H
H H m m
H H p p
(
um
i = (
um
i xi +
(
um
i xi
i ) y
i ) Hi,i v
i ) Hi,i v
p6=m
dk
X
H H p p
H
i ,
(
um
k xk + (
um
i ) Hi,k v
i ) z
i, m
(3.37)
k6=i p =1
N 1
2
where u
m
with k
um
i C
i k = 1, i, m. Defining hi,k , [hi,k (0) . . .
m,p
p
T
m
hi,k (N 1)] and bi,k , (

ui ) v
k , we can rewrite the input-output
relations (3.37) as
X
H
H
m,m m
H
m,p p
(
um
i = h
h
i ) y
i,i bi,i xi +
i,i bi,i xi
p6=m
dk
X
H
H
m,p p
i ,
h
um
i,k bi,k xk + (
i ) z
i, m.
(3.38)
k6=i p =1
Choosing xm
i , i, m, to be i.i.d. Gaussian, treating the two interference
terms in (3.38) as additional noise and assuming that the destination
m,m
H b
Di knows the effective channel coefficients h
i,i i,i , m, perfectly, the
rate of communication over the link Si Di is then lower-bounded
according to
!
di
P H m,m 2
1 X
M di |hi,i bi,i |
(3.39)
log 1 +
Ri
N m=1
Ii,1 + Ii,2 + No
with
Ii,1 =
2
X P
m,p
H
b
h

M di i,i i,i
p6=m
78
3.3.
Ii,2 =
dk
XX
P
M
dk
p =1

H m,p 2
hi,k bi,k .
k6=i
H H m
Recall that in IA with perfect CSI, we had |(um
i ) Hi,i vi | c >
m H H p
0, i, m (with c independent of P ), (ui ) Hi,i vi = 0, i, m 6= p, and
m,p
p
H H p
m
(um
i ) Hi,k vk = 0, k 6= i, m, p. Defining bi,k , (ui ) vk , these
conditions are equivalent to
m,m
H
|h
i,i bi,i | c > 0, i, m
m,p
H
h
= 0, i, m 6= p
i,i b
i,i
H m,p
hi,k bi,k
= 0, k 6= i, m, p.
(3.40)
(3.41)
(3.42)
Defining w
i,k , [w
i,k (0) w
i,k (1) . . . w
i,k (N 1)]T , naive IA therefore
p
m
entails finding vectors u
i , v
i satisfying the following conditions:
H m,m
|w
i,i
bi,i | c > 0,
H m,p
w
i,i
bi,i
H m,p
w
i,k
bi,k
i, m
(3.43)
= 0,
i, m 6= p
(3.44)
= 0,
k 6= i, m, p.
(3.45)
As noted earlier, one way to find vectors u

m
ip satisfying (3.43), (3.44)
i ,v
and (3.45) is provided in (Cadambe and Jafar, 2008, Appendix III).
The key point here is that this choice of u
m
ip , i, m, p, results in
i ,v
full spatial multiplexing gain. This can be proved as follows.
m,p
Since kw
i,i k = kw
i,i k = 1 (from Parsevals theorem) and w
i,i , b
i,i
CN 1 , i, m, p, we can always find vectors q1 , . . . , qN 2 such that

)
(
m,p
b
i,i
(3.46)
w
i,i , m,p , q1 , . . . , qN 2 , m 6= p
k
kb
i,i
i,i into this orthonormal
is an orthonormal basis for CN . Expanding h
basis, we get

2

m,p 2 NX
H 2

b
2

i,i

i,i k2 =
H
hi,i ql
kh
hH
i,i + h
i,i w
i,i m,p +

kbi,i k
l=1
79
3.

m,p 2
H
2 H b

i,i

hi,i w
i,i + hi,i m,p ,
k

kb
i,i
i, m 6= p
which yields
P H m,p 2
h b
M di i,i i,i
2
P m,p 2 2 H
i,i
kbi,i k khi,i k hi,i w
M di

2
h
H
i,i
P m,p 2 2
i,i w
=
kb k khi,i k 1

khi,i k
M di i,i
2

hH w
P m,p 2 2
i,i i,i
kbi,i k khi,i k 1
=

khi,i k
M di
2
P m,p 2
kbi,i k
hi,i (max
)2
d
M di
!
2
4P m,p 2
1
,
kb k
hi,i
Nd
M di i,i
2 (L1)
(3.47)
(3.48)
(3.49)
(3.50)
i, m 6= p
(3.51)
where (3.49) follows from Parsevals theorem and (3.51) is obtained

by invoking (3.32).
i,k into the orthonormal basis

A similar expansion of h
)
(
m,p
b
i,k
w
i,k , m,p , q1 , . . . , qN 2 , k 6= i
k
kb
(3.52)
i,k
yields
P H m,p 2
h b
M dk i,k i,k
2
4P m,p 2
kbi,k k
hi,k
M dk
80
1
Nd
2 (L1)
!
,
k 6= i, m, p.
(3.53)
3.3.
If we now choose Nd = (L 1) log P , we get from (3.51) that

!
2
P H m,p 2
4P m,p 2
1
kb k
hi,i
h b
Nd
M di i,i i,i
M di i,i
2 (L1)
| {z }
=1/P
2

hi,i
,
M di
{z
}
m,p k2
4kb
i,i
i, m 6= p
m,p
i,i
and from (3.53) that

2
P H m,p 2
4P m,p 2
kbi,k k
hi,k
hi,k bi,k
M dk
M dk
1
Nd
|
=
m,p k2
4kb
i,k
|

2

hi,k
M dk
{z
m,p
i,k
2 (L1)
{z }
=1/P
k 6= i, m, p
which implies that, with Nd = (L 1) log P , the overall interference

power in the rate lower bound (3.39) is upper-bounded by a constant
independent of P according to
Ii,1 + Ii,2
X
p6=m
m,p
i,i +
dk
XX
m,p
i,k .
(3.54)
k6=i p =1
Finally, we note that as P , the vector quantizer codebook size

2Nd also tends to infinity and the maximum quantization error tends
to zero according to (3.32). The resolution of the vector quantizer
therefore becomes arbitrarily high and we obtain w
i,i hi,i /khi,i k,
i,i /kh
i,i k. Substituting this into (3.43) yields
which implies w
i,i h
H b
m,m
|h
i,i i,i | khi,i kc > 0. Consequently, using (3.54), the spatial
81
3.
multiplexing gain achieved by naive IA is lower-bounded according to

Rsum
log P
lim
log 1 +
di
M X
X
lim
i=1 m=1
P
=
di
H b
m,m 2
|h
i,i i,i |
dk
P P
m,p
m,p
i,i +
i,k +No
p6=m
k6=ip=1
N log P
P
M di
M
2
(from (3.11) and (3.12))
(3.55)
which proves that full spatial multiplexing gain is achieved. We complete the proof by noting that the number of bits fed back (broadcast) by each destination for achievability of full spatial multiplexing
gain using naive IA is given, according to (3.33), by Nf = M Nd =
M (L 1) log P .
3. 4 .
EXTENSION TO MIMO I NTER FER ENC E

N ETWOR KS
Our analysis for single-antenna interference networks can be easily

extended to MIMO interference networks where each source and each
destination is equipped with precisely Mt antennas. In particular, if
each antenna of each source is considered as an independent source
and each antenna of each destination is considered as an independent
destination, then the MIMO interference network would be equivalent
to a single-antenna interference network with a total of M Mt sourcedestination pairs. It then follows from our discussion for the singleantenna case that spatial multiplexing gain of M Mt /2 is achievable in
the single-antenna interference network with M Mt source-destination
pairs if each destination can broadcast a total of M Mt (L 1) log P
bits to all the sources and destinations in the network. Since allowing
cooperation between the nodes cannot reduce the achievable rate, this
implies that the full spatial multiplexing gain of M Mt /2 is achievable.
82
3.5.
3.5.
OBSERVATIONS
OBSERVATIONS
An M source-destination pair frequency-selective SISO interference

network was investigated for feasibility of spatial multiplexing with
limited feedback. Cues were taken from a well-known transmission
scheme, called interference alignment, and it was shown that interference alignment is equivalent to a combination of transmit and receive
beamforming for each source-destination pair. For the case where
each source and each destination has perfect knowledge of all the
channels in the network, it is possible to choose the transmit and
receive (beamforming) direction vectors for interference alignment in
such a manner that every destination sees no interference in at least
half of the signal-space dimensions, and hence, full spatial multiplexing gain of M/2 is achieved. For the case where the sources and the
destinations have access to only quantized versions of all the channels,
it was shown that the transmit and receive (beamforming) direction
vectors for interference alignment can be chosen in such a manner
that interference terms at every destination are upper-bounded by a
finite constant (independent of P ) in at least half of the signal-space
dimensions, so that full spatial multiplexing gain is still achieved in
the limit of large sum transmit power P . Finally, it was shown that
the required quantized versions of all the channels can be obtained at
each source and each destination, provided that each destination can
broadcast M (L 1) log P bits to all the nodes in the network.
83
C HAPTER
MISO Interference Networks

n this chapter, we consider a special class of interference
networks, called multi-input single-output (MISO) interference
networks, where each source is equipped with multiple antennas
and each destination is equipped with a single antenna. In particular,
we consider a network where M sources, denoted by S1 , , SM and
equipped with M antennas each, communicate with M destinations,
denoted by D1 , , DM and equipped with a single-antenna each.
All the channels in the network are frequency-selective with L taps
and each source has an average power constraint of P/M . Moreover,
the destinations have perfect channel state information (CSI) of
its channels from each of the sources and each source obtains CSI
through limited capacity feedback links. Our goal is to investigate the
feasibility of spatial multiplexing in such a network. To this end, we
propose a feedback and transmission scheme, which we term spatial
interference-nulling (SIN), and show that spatial multiplexing based
on SIN precoding vectors results in full spatial multiplexing gain of M ,
provided that each destination can feedback to its non-corresponding
sources at the rate of (M 1)(M L 1) log P . The basic idea being
SIN is that each source precodes its transmission in such a manner
that the interference created by its transmission at each of its noncorresponding destinations has a finite power level, independent of P .
A natural question to ask is why should we be interested in having
85
4.
MISO INTERFERENCE NETWORKS
1
2
S1
D1
1
2
D2
S2
M
1
2
SM
DM
Fig. 4.1: A MISO interference network with M sources, each equipped with M
antennas, and M destinations, each equipped with a single antenna.
multiple antennas at the sources for spatial multiplexing, when full

spatial multiplexing gain is already known to be achievable with
a single antenna at each source (as shown in the chapter on SISO
interference networks). The answer to this question is two-pronged.
First, of course, is that the maximum achievable spatial multiplexing
gain in an M source-destination pair MISO interference network is
M , in contrast to SISO interference networks where it is M/2. Second
and perhaps more subtle reason is that interference alignment, that
is, the scheme that achieves full spatial multiplexing gain in SISO
networks, is rather difficult to implement in practice because of its
dependence on three critical assumptions: i) the number of OFDM
tones is extremely large, ii) the number of taps L in channel impulse
responses is very large and iii) there exist feedback channels from
each destination to all the other destinations. In this chapter, we show
86
4.1.
SYSTEM MODEL
that employing SIN results in full spatial multiplexing gain in MISO

interference networks without these three assumptions.
The rest of the chapter is organized as follows. After describing a
system model for MISO interference networks in Section 4.1, we first
present our results for the case where all channels in the network are
frequency-flat (that is, L = 1) in Section 4.2. We then extend our
analysis to the case where all channels are frequency-selective (that
is, L > 1) in Section 4.3.
4 .1.
SYSTEM MODEL
The impulse response of the frequency-selective MISO channel between the source Sk and the destination Di is given by the taps
T
hi,k [l] = [h1i,k [l] . . . hM
CM 1 , for l = 0, . . . , L 1, where
i,k [l]]
m
hi,k [l] denotes the l-th tap of the channel impulse response between
the m-th transmit antenna of Sk and the single antenna destination
Di . The channel coefficients hm
i,k [l] are assumed to remain constant
throughout the time interval of interest (outage setting) and are drawn
independently (across i, k, m, l) from a continuous probability density
function such that 0 < |hm
i,k [l]| < , for all i, k, l, m, with probability
1. We use a cyclic signal model (such as in orthogonal frequency division multiplexing (OFDM)) to convert the L-tap frequency-selective
MISO channel Sk Di to N parallel frequency-flat MISO channels
T
M 1
given by hi,k (r) = [h1i,k (r) . . . hM
, for r = 0, . . . , N 1,
i,k (r)] C
where
m
hm
i,k (r) = F{hi,k [n]}
(4.1)
is the channel coefficient between the m-th transmit antenna of Sk

and the single antenna destination Di corresponding to the r-th tone.
The input-output relation between Si and Di , for the r-th tone, is
87
4.
then given by
yi (r) = hH
i,i (r)xi (r) +
hH
i,k (r)xk (r) + zi (r)
(4.2)
k6=i
{z
interference
where yi (r) C is the symbol received at the destination Di , xk (r) =

T
M 1
[x1k (r) . . . xM
denotes the transmit vector symbol for
k (r)] C
source Sk and zi (r) is CN (0, No ) noise at Di , all for the r-th tone.
The equivalent relation in time-domain is given by
X
yi [n] = (hH
(hH
(4.3)
i,i ~ xi )[n] +
i,k ~ xk )[n] + zi [n]
k6=i
for n = 0, . . . , N 1, where yi [n] = F 1 {yi (r)} C, xk [n] =

T
M 1
1
11
[x1k [n] . . . xM
with xm
{xm
and
k [n]] C
k [n] = F
k (r)} C
1
zi [n] = Frn
{zi (r)} C11 is CN (0, No ) noise. The transmit signals
obey the power constraints

P
, k = 1, . . . , M, n = 0, . . . , N 1.
E kxk [n]k2
M
(4.4)
We shall also need the channel vector hi,k and the normalized channel vector wi,k corresponding to the link Sk Di defined as hi,k =
[hTi,k [0] hTi,k [L 1]]T CM L1 and wi,k = hi,k /khi,k k CM L1 ,
respectively. Finally, defining wi,k [l] = hi,k [l]/khi,k k CM 1 , for all
T
T
i, k, l, we have wi,k = [wi,k
[0] . . . wi,k
[L 1]]T .
We assume that each destination Di knows its channel from each
of the sources perfectly, that is, Di knows hi,k , k. We further assume
that there exist dedicated non-interfering error-free broadcast feedback
links from each destination Di to its non-corresponding sources Sk ,
k 6= i. In the remainder of this section, we distinguish between a
channel feedback phase during which Nf bits of feedback are broadcast
by each destination and a data transmission phase following the
We
In
88
conjugate the channel coefficients for notational simplicity later on.

contrast to IA, SIN does not require Di to feedback CSI to Si or Dk , k 6= i.
4.2.
THE FREQUENCY- FLAT CASE
channel feedback phase. The channels, being deterministic, are fed

back only once during the entire time interval of interest so that
the transmission rate loss due to the channel feedback phase can be
assumed to be negligible. Denoting the rate of communication over
PM
the link Si Di by Ri , and letting Rsum = i=1 Ri , we say that full
spatial multiplexing gain is achieved if
lim
Rsum
= M.
log P
(4.5)
Recall that the spatial multiplexing gain of an M source-destination

pair MISO interference network is upper-bounded by the total number
of destination antennas, that is, M .
4 .2.
TH E FR EQU ENCY - FLAT CASE
In this section, we consider a simplified setup for MISO interference

networks where all channels are frequency-flat (that is, L = 1) so
that hi,k = hi,k [0] and wi,k = wi,k [0], for all i, k, and hence the
input-output relation (4.3) simplifies to
X
yi [n] = hH
hH
n.
(4.6)
i,i xi [n] +
i,k xk [n] + zi [n],
k6=i
Our key contribution for this simplified setup is to propose a vector

quantization scheme and a novel transmission scheme, namely spatial
interference-nulling (SIN), that achieves full spatial multiplexing gain.
Our main result can be formalized as follows.
Theorem 5 In the MISO interference network defined in Section 4.1,
if all the channels are frequency-flat (L = 1) and the number of
feedback bits Nf broadcast by each destination to its non-corresponding
For the trivial case of two-source two-destination frequency-flat MISO interference networks, spatial interference-nulling would be equivalent to spatial
zero-forcing proposed in Jafar and Fakhereddin (2007).
89
4.
sources scales with the average sum transmit power P as

Nf = (M 1)2 log P
(4.7 )
full spatial multiplexing gain is achievable.

Before proceeding to the proof of the above theorem, we first present
only SIN by considering a case where all sources have perfect transmit
CSI.
4.2.1. SIN with perfect transmit CSI

Assuming that each source Sk knows the channels Sk Di , for all
i 6= k, perfectly (this requires infinite feedback), SIN entails that each
source beamforms its transmission along a direction that is orthogonal
to its channels to each of its non-corresponding destinations. In
particular, each source Sk transmits the vector signal
xk [n] = vk xk [n]
(4.8)
where xk [n] denotes a scalar symbol with variance P/M and vk

CM 1 , with kvk k2 = 1, is a precoding vector for Sk that is chosen
such that
hH
i,k vk = 0,
i 6= k
H
wi,k
vk
i 6= k
= 0,
(4.9)
(since wi,k = hi,k /khi,k k).
(4.10)
That a non-trivial vk (that is, vk 6= 0M ) always exists follows because

there are only M 1 orthogonality constraints in (4.10), while vk
is M -dimensional. We denote by Wk CM M the matrix with
columns equal to the normalized channel vectors at source Sk , that is,
Wk = [w1,k . . . wM,k ] CM M with wk,k replaced by a (dummy)
all-one vector (recall that we did not assume Sk to know the channel
Sk Dk ). The vector vk is then computed as
H
vk =
90
[Cof[{Wk }1,k ] Cof[{Wk }M,k ]]

.
k[Cof[{Wk }1,k ] Cof[{Wk }M,k ]]k
4.2.
Thanks to

Cof[{Wk }m,k ] = {adj(Wk )}k,m = det(Wk ) Wk1 k,m
we have vkH wi,k = 0, and hence, hH
i,k vk = 0, for all i 6= k. The received
symbol at Di is given by
X
yi [n] = hH
hH
(4.11)
i,i vi xi [n] +
i,k vk xk [n] + zi [n].
k6=i
Choosing xk [n], for all k, n, to be i.i.d. Gaussian, and assuming that

each destination Di knows its effective channel coefficient hH
i,i vi and
P
P
2
its interference power level k6=i M
|hH
v
|
perfectly,
the
rate of
i,k k
communication Ri over the link Si Di is lower-bounded according
to
P
H
2
M |hi,i vi |

2

P H
v
h
k6=i M
i,k k
Ri log 1 + P
+ No
(4.12)
where (4.12) is a lower-bound because there may exist transmission

schemes with higher rates of communication than SIN. Consequently,
the spatial multiplexing gain is lowerbounded according to

H
2
P
PM
M |hi,i vi |
H
2
P
i=1 log 1 + P
k6=i M |hi,k vk | +No
| {z }
Rsum
=0
lim
lim
P log P
P
log P
=M
(4.13)
and hence the full spatial multiplexing gain is achieved.
We next proceed to the more realistic case where the sources have
partial CSI obtained through limited capacity feedback links. In this
case, the source Sk , k, cannot compute the SIN precoding vector
vk , but only its approximation v
k based on its received feedback.
P
While the resulting interference power k6=i (P/M )|hH
k |2 will not
i,k v
be zero, it is possible, using an appropriate quantization and feedback
91
4.
P
scheme, to upper-bound k6=i (P/M )|hH
k |2 by a constant, say co ,
i,k v
independent of P . The resulting lower bound on multiplexing gain is
then given by

H
2
P
PM
M |hi,i vi |
i=1 log 1 +
co +No
Rsum
lim
=M
(4.14)
lim
P
P log P
log P
and hence the full spatial multiplexing gain is achieved. In the rest of
this section we analyze this case in detail and indeed show that SIN,
along with the vector quantization scheme of Mukkavilli et al. (2003),
results in full spatial multiplexing gain with the number of feedback
bits as specified in (4.7).
4.2.2. The quantization and feedback scheme

vectors hi,k (and hence the direction vectors wi,k ), for all k, perfectly.
It quantizes and broadcasts M 1 of the direction vectors, that is, wi,k ,
for all k 6= i. In particular, the direction vector wi,k is quantized, using
Nd bits, to w
i,k . Replacing CL by CM , the same vector quantization
scheme as discussed for IA in Chapter 3, Section 3.3.1, is employed,
so that the maximum quantization error is upper-bounded according
to
q
H w
max
, max 1 |wi,k
i,k |2
(4.15)
d
i,k
q
1 |xH w
x |2
(4.16)
max
xCM 1 ,kxk=1
!
1
2
.
(4.17)
Nd
2 2(M 1)
Total feedback rate Each destination Di feeds back Nd bits to each Sk ,
k 6= i. The total number of feedback bits broadcast by each destination
is therefore given by Nf = (M 1)Nd . Consequently, each source
Sk receives Nd bits of feedback from each of its non-corresponding
92
4.2.
destinations. Since the feedback links are assumed to be error-free,

the source Sk can recreate the quantized direction vectors w
i,k , for
all i 6= k.
4.2.3. The transmission scheme

SIN is performed by each of the M sources Sk , naively assuming
that the quantized normalized channel vectors are equal to the actual
k CM M the matrix
normalized channel vectors. Let us denote by W
with columns equal to the quantized normalized channel vectors at
k = [w
source Sk , that is , W
1,k . . . w
M,k ] CM M . Substituting
w
k,k by a (dummy) all-one vector (recall that Sk does not know w
k,k ),
the source Sk computes its SIN precoding vector v
k as
iH
k }1,k ] Cof[{W
k }M,k ]
Cof[{W
h
i .
v
k =

k }1,k ] Cof[{W
k }M,k ]

Cof[{W
h
Thanks to
k }m,k ] = {adj(W
k )}k,m = det(W
k ) {W
1 }k,m
Cof[{W
k
we have v
kH w
i,k = 0, for all i 6= k. The source Sk transmits the vector
signal
xk [n] = v
k xk [n]
(4.18)
where xk [n] denotes a scalar complex symbol with variance P/M .

The received symbol at Di is finally given by
X
yi [n] = hH
i xi [n] +
hH
k xk [n] + zi [n].
(4.19)
i,i v
i,k v
k6=i
93
4.

Choosing xk [n], for all k, n, to be i.i.d. Gaussian, and assuming
that each destination Di knows its effective channel coefficient hH
i
i,i v
(corresponding to the link Si Di ) and its interference power level

P
H 2
(P/M
)
v
h
, the rate of communication Ri over the link
k
i,k
k6=i
Si Di is lower-bounded according to
P
H
i |2
M |hi,i v

2

P H
k
k6=i M hi,k v
Ri log 1 + P
+ No
(4.20)
where (4.20) is a lower bound because there may exist transmission

schemes with higher rates of communication than SIN. Consequently,
the sum rate Rsum is lower-bounded according to
!
M
P
H
X
i |2
M |hi,i v
log 1 + P
Rsum
.
(4.21)
P
H |2 + N
k
o
k6=i M |hi,k v
i=1
The key to proving achievability of full spatial multiplexing gain is to
upperbound the interference power
X P
|hH v
k |2
M i,k
(4.22)
k6=i
by some constant independent of P . More concretely, we want to

find an Nd such that the interference power in (4.22) is, indeed,
upperbounded by a constant independent of P . To this end, we start
by noting that
v
kH w
i,k = 0,
i 6= k.
(4.23)
Furthermore, since kw
i,k k = 1, k
vk k = 1 and w
i,k , v
k CM 1 , for
all i 6= k, we can always find q1 , . . . , qM 2 such that
{w
i,k , v
k , q1 , . . . , qM 2 }
94
4.2.
is an orthonormal basis for the M -dimensional complex inner product

space. Expanding hi,k , i 6= k, into this orthonormal basis, we get
2
X
2 M
H 2

2
hi,k qt
k +
i,k + hH
khi,k k2 = hH
i,k v
i,k w
(4.24)
t=1

2
2
hH
i,k + hH
k
i,k w
i,k v
(4.25)
which yields the bound

2
P H 2
P H
P
hi,k v
khi,k k2
hi,k w
k
i,k
M
M
M
2
hH w

P
i,k
i,k

khi,k k2 1
=

khi,k k
M
P
2
khi,k k (max
)2
d
M
!
1
4P
2
khi,k k
.
Nd
M
2 (M 1)
If we now choose Nd = (M 1) log P , we get

!
2
P H 2
4P
1
4 khi,k k
2
hi,k v
k
khi,k k
=
Nd
M
M
M
2 (M 1)
| {z }
(4.26)
(4.27)
(4.28)
(4.29)
(4.30)
=1/P
which is the desired result. Upon substituting (4.30) into (4.21), we

obtain the lower bound
!
lim
Rsum
log P
M
X
i=1
=M
lim
|hH
i |2
i,i v
4khi,k k2
+No
k6=i
M
P
M
log 1 +
log P
(4.31)
which proves that full spatial multiplexing gain is achieved. We com-
95
4.
plete the proof by noting that Nf = (M 1)Nd = (M 1)2 log P .
4 .3.
TH E FR EQU ENCY - SELECTIVE CASE
We next extend our analysis for frequency-flat MISO interference

networks to frequency-selective MISO interference networks. A naive
approach to achieving full spatial multiplexing gain in the frequencyselective case would be to apply the feedback and transmission scheme
discussed for the frequency-flat case to each of the N OFDM tones
independently. The problem with this approach is that the corresponding feedback of N (M 1)2 log P bits is unlikely to be fundamental
because of its dependence on the number of tones N . In fact, a fundamental approach should achieve full spatial multiplexing gain with a
much less feedback that, rather than being proportional to N (M 1)2 ,
is proportional to the number of relevant channel coefficients per
destination (also called degrees of freedom), which is approximately
M 2 L. The following theorem concretizes this result:
Theorem 6 In the MISO interference network defined in Section 4.1,
if all the channels are frequency-selective (L > 1) and the number of
feedback bits Nf broadcast by each destination to its non-corresponding
sources scales with the average sum transmit power P as
Nf = (M 1) (M L 1) log P
(4.32 )

In the rest of this section, we first generalize SIN to the frequencyselective case. We then provide a frequency-domain representation
of SIN and derive the requirements on the accuracy of feedback for
The exact degrees of freedom is (M 1)(M L 1), where M 1 is the number
of sources to which each destination should feed back and M L 1 is the degrees of
freedom per Sk Di , k 6= i, link. Intuitively, the degrees of freedom per Sk Di
link is M L 1, rather than the number of fading coefficients M L, because the
magnitude (or the power contained in the channel Sk Di ) is not required for
a direction-based precoding scheme such as SIN.
96
4.3.
THE FREQUENCY- SELECTIVE CASE
achieving full spatial multiplexing gain. Finally, we generalize the

feedback scheme discussed for the frequency-flat case to the frequencyselective case and show that with number of feedback bits as specified
in (4.32), the generalized feedback scheme meets the feedback accuracy
requirements and hence full spatial multiplexing gain is achieved.
4.3.1. Spatial interference-nulling in time-domain

We start by restating the input-output relation (4.3) as
H
yi [n] = khi,i k(wi,i
~ xi )[n]
X
H
+
khi,k k(wi,k
~ xk )[n] + zi [n].
(4.33)
k6=i
{z
interference
The source Sk transmits the vector signal

xk [n] = (
vk ~ xk )[n],
n = 0, . . . , N 1
(4.34)
where xk [n], n = 0, . . . , N 1, denotes a scalar symbol sequence

with variance P/M and v
[n] = [
vk1 [n] . . . vkM [n]]T CM 1 , n =
PN 1 k
2
vk [n]k = 1, is the SIN precoding vector
0, . . . , N 1, with n=0 k
sequence at Sk . The received symbol at Di is then given by
H
yi [n] = khi,i k(wi,i
~v
i ~ xi )[n]
X
H
+
khi,k k(wi,k
~v
k ~ xk )[n] + zi [n].
(4.35)
k6=i
Defining
y
i , [yi [0] . . . yi [N 1]]T
CN 1
x
i , [xi [0] . . . xi [N 1]]T
CN 1
zi , [zi [0] . . . zi [N 1]]T
CN 1
97
4.
H
H
(wi,k
~v
k )[0]
. . . (wi,k
~v
k )[1]
H
H
(wi,k
~v
k )[1]
. . . (wi,k
~v
k )[2]
,
CN N
..
..
..
.
.
.
H
H
(wi,k ~ v
k )[N 1] . . . (wi,k ~ v
k )[0]
Ui,k
the input-output relation between Si and Di can be rewritten as

X
y
i = khi,i kUi,i x
i +
khi,k kUi,k x
k +
zi .
(4.36)
k6=i
Choosing xk [n], for all k, n, to be i.i.d. Gaussian, and assuming that

each destination Di knows its effective channel matrix Ui,i (corresponding to the link Si Di ) and its interference power (covariance)
P
matrix k6=i (P/M )khi,k k2 Ui,k UH
i,k , the rate of communication Ri
over the link Si Di is lowerbounded according to
Ri
det
P
M
P
2
H
k=1 M khi,k k Ui,k Ui,k
+ N o IN
P
H
2
k6=i M khi,k k Ui,k Ui,k
+ N o IN
1
P
log
N
det

.
(4.37)
Substituting the eigenvalue decomposition of the circulant matrix

Ui,k , that is, Ui,k = FH Di,k F where Di,k CN N is the diagonal
H
matrix with r-th diagonal entry as di,k (r) = F{(wi,k
~v
k )[n]}, into
H
(4.37), and noting that FF = IN , we obtain
Ri
1
N
1
N

PM
P
2 H
H
det
kh
k
F
D
D
F
+
N
I
i,k
i,k i,k
o N
k=1 M

log P
P
2 FH D DH F + N I
det
kh
k
i,k
i,k i,k
o N
k6=i M

PM
P
2
H
det
k=1 M khi,k k Di,k Di,k + No IN

log P
P
2 D DH + N I
det
kh
k
i,k
i,k
o
N
k6=i M
i,k
(since det(IN + AB) = det(IN + BA) and FFH = IN )
98
(4.38)
(4.39)
4.3.
1
=
log
N

2
P
2

k=1 M khi,k k di,k (r)

P
P
2
2
k6=i M khi,k k di,k (r)

2
P
2

M khi,i k di,i (r)

2
P
P
2

k6=i M khi,k k di,k (r)
QN 1 PM
r=0
QN 1
r=0
N 1
1 X
=
log 1 +
N r=0
!
+ No

+ No
!
.
(4.40)
(4.41)
+ No
Therefore, if each Sk can determine, based on quantized feedback

of wi,k [l], for all i 6= k, for all l, a precoding-vector sequence v
k [n]
such that
X P
2

khi,k k2 di,k (r)
M
k6=i
X P

2
H
khi,k k2 F{(wi,k
~v
k )[n]}
M
(4.42)
k6=i
is upper-bounded, for all i 6= k, for all r, by a constant independent

of P , then full spatial multiplexing gain is achieved. Thus, we need to
develop a feedback and transmission scheme (that is, a set of precodingvector sequence v
k [n] based on quantized feedback of wi,k [l], for
l = 0, . . . , L 1, for all i 6= k) such that the interference power (4.42)
is indeed upperbounded by a constant independent of P .
4.3.2. Frequency-domain interpretation of SIN

The source Sk transmits the vector signal
xk (r) = v
k (r)xk (r),
r = 0, . . . , N 1
(4.43)
on the r-th tone, where xk (r) denotes a scalar symbol with variance
P/M and v
k (r) = [
vk1 (r) . . . vkM (r)]T , with vkm (r) = F{
vkm [n]}, is
the SIN vector for the r-th tone. The received symbol at Di on the
We will later assume that x (r), for all r, are i.i.d, and thus ensure that the
k
power constraint (4.4) is satisfied.
99
4.
r-th tone is then given by

yi (r) = hH
vi (r)xi (r) +
i,i (r)
hH
vk (r)xk (r)
i,k (r)
k6=i
+ zi (r),
r = 0, . . . , N 1.
(4.44)
Chosing xk (r), for all k, r, to be i.i.d. Gaussian, and assuming that

each destination Di knows its effective channel coefficient hH
vi (r)
i,i (r)
corresponding to the link Si Di and its interference power level
P
H
vk (r)|2 , for all r, the rate of communication Ri
k6=i (P/M )|hi,k (r)
over the link Si Di is lower-bounded according to
!
N 1
P
H
2
|h
(r)
v
(r)|
1 X
i
i,i
M
log 1 + P
Ri
.
(4.45)
P
H
N r=0
vk (r)|2 + No
k6=i M |hi,k (r)
Therefore, if each Sk can determine, based on quantized feedback of
wi,k [l], for all i 6= k, for all l, a precoding-vector sequence v
k (r) such
that the interference power
X P
|hH (r)
vk (r)|2
M i,k
k6=i
X P
2
F{(hH
k )[n]}
i,k ~ v
M
(4.46)
X P

2
H
khi,k k2 F{(wi,k
~v
k )[n]}
M
(4.47)
k6=i
k6=i
is upper-bounded, for all i 6= k, for all r, by some constant independent

of P , then full spatial multiplexing gain is achievable. At the risk of
stating the obvious, the interference power (to be upperbounded) in
(4.47) is the same as the interference power (4.42) in the time-domain
analysis.
This assumption is equivalent to our assumption on destination CSI in the

time-domain analysis.
100
4.3.
4.3.3. Feedback design criterion

We now focus on the interference power (4.47) (or equivalently (4.42))
to derive guidelines for our feedback design. We start by noting that
(4.47) can be upper-bounded as
X P

2
H
khi,k k2 F{(wi,k
~v
k )[n]}
M
k6=i

2
N
1
X
X P

j2rn
2 1
H
khi,k k
(wi,k ~ v
k )[n]e N
=
M
N n=0
(4.48)
k6=i
N 1
X P
j2rn
1 X
H
|(wi,k
~v
k )[n]|e N
khi,k k2
M
N n=0
k6=i
X
X
(using |
ai bi |
|ai | |bi |, for all ai , bi C)
i
!2
(4.49)
2

X P
N
H
~v
k )[n]|
khi,k k2 max |(wi,k
M
N n
k6=i
js
(since e = 1, for all real scalars s)
X PN
H
=
khi,k k2 (max |(wi,k
~v
k )[n]|)2 .
n
M
(4.50)
(4.51)
k6=i
Therefore, in order to upperbound the interference power by a constant

co (independent of P ), we need a quantization and feedback scheme
under which each source Sk can compute, based on quantized feedback
of wi,k [l], i 6= k, l = 0, . . . , L1, the SIN precoding vectors v
k [n], n,
that satisfy
X
k6=i
H
khi,k k2 |(wi,k
~v
k )[n]|2
co M
,
PN
n.
(4.52)
A natural consideration while designing such a feedback scheme is

whether to quantize the normalized channel vectors wi,k [l] independently for each tap (this involves a feedback of (M 1)L degrees of
freedom) or to quantize them jointly as normalized channel vector
101
4.
wi,k (this involves a feedback of M L 1 degrees of freedom). The

constraints (4.52) imply, although in a subtle manner, that quantizing
the taps independently, which ignores the information about the relative magnitudes kwi,k [l1 ]k/kwi,k [l2 ]k, for all l1 6= l2 , is not sufficient
H
~v
k )[n] by a term that goes to 0 as P ,
to upperbound (wi,k
and hence, is not sufficient to upperbound the interference power by a
constant independent of P . The subtlety can be intuitively understood
by rewriting, with a slight abuse of notation, the constraints (4.52) as
X
H
khi,k k2 |(wi,k
~v
k )[n]|2
k6=i
co M
0, n
PN
(4.53)
X

L1
H
H
wi,k
[l]
vk [n l] 0, k 6= i, n (4.54)
|(wi,k
~v
k )[n]| =
l=0
where we have implicitly defined v

k [s] , v
k [N s], for s = 1, . . . , L
1. Consider now the case where Sk knows the relative magnitudes
kwi,k [l1 ]k/kwi,k [l2 ]k, l1 , l2 6= l1 . The source Sk then needs to satisfy
PL1 H
only the summation constraint l=0 wi,k
[l]vk [n l] = 0, while the
H
individual terms wi,k
[l]vk [n l] can be anything (that is, they need
not be zero). This implies that in (4.54), we have only (M 1)N
constraints, while the vector v

k CM N 1 , defined as
v
k , [
vkT [0] . . . v
kT [N 1]]T
(4.55)
k 6=
is M N -dimensional and hence a non-trivial solution (that is, v
0M N ) exists. Next, consider the case where a random tap, say lo -th
tap, is not jointly quantized with other taps. Then, the information
about the relative magnitudes kwi,k [lo ]k/kwi,k [l]k, for all l 6= lo , will
be lost. In such a case, the only way to satisfy (4.54) would be by
P
H
H
choosing v
k such that both wi,k
[lo ]vk [nlo ] and l6=lo wi,k
[l]vk [nl]
in (4.54) are individually zero. However, this would imply that we
have 2(M 1)N constraints and v

k , being M N -dimensional, will
not have any non-trivial solution, and hence, the interference power
cannot be upperbounded by a constant independent of P . Since the
same argument holds for any lo , all the taps must be jointly quantized.
102
4.3.
Also note that in the above arguments, we deliberately assumed that
the dimension of v
k is fixed to be M N . Therefore, one may argue
that if we, for example, double the number of antennas at Sk to 2M ,
then we can do without one of the relative magnitudes. The problem
is that the resulting decrease of one degree of freedom in feedback
is outnumbered by the increase of M L degrees of freedom because
of each normalized channel vector now being 2M dimensional (recall
the L channel vector taps with M antennas have L(M 1) degrees of
freedom, but if Sk has 2M antennas, each destination Di , i 6= k will
have L(2M 1) degrees of freedom to feedback to Sk ).
In summary, the taps wi,k [l], for all l, must be quantized jointly
as normalized channel vector wi,k in order to satisfy (4.52). We next
present such a quantization scheme (that is, a quantization scheme
that satisfies (4.52)) and show that SIN, in conjunction with this
scheme, results in full spatial multiplexing gain.
4.3.4. The feedback scheme

vectors hi,k [l], for l = 0, . . . , L 1, for all k, perfectly. It computes the channel vector hi,k CM L1 and the normalized channel vector wi,k CM L1 as hi,k = [hTi,k [0] . . . hTi,k [L 1]]T and
wi,k = hi,k /khi,k k, respectively, and quantizes and broadcasts M 1
of the normalized channel vectors, that is, wi,k , for all k 6= i. In
particular, the normalized channel vector wi,k is quantized, using Nd
bits, to w
i,k . Replacing CL by CM L , the same vector quantization
scheme as discussed for IA in Chapter 3, Section 3.3.1, is employed,
so that the maximum quantization error is upper-bounded according
to
q
H w
max
,
max
1 |wi,k
i,k |2
(4.56)
d
i,k
q
max
1 |xH w
x |2
(4.57)
xCM L1 ,kxk=1
103
4.
Nd
!
.
(4.58)
2 2(M L1)
Note that in contrast to our analysis in Section 3.3.1 where a given
normalized channel vector is quantized over L-dimensional complex
space, the normalized channel vectors in the MISO frequency-selective
case are quantized over M L-dimensional complex space, resulting in a
factor of M L 1 (rather than L 1 as in (3.32)) in the denominator
of the power in (4.58).
Total feedback rate As in the frequency-flat case, each destination Di feeds back Nd bits to each Sk , k 6= i, resulting in a total
feedback of Nf = (M 1)Nd bits per destination. Each source Sk
therefore receives Nd bits of (error-free) feedback from each of its
non-corresponding destinations. Using this feedback, Sk recreates the
quantized normalized channel vectors w
i,k , for all i 6= k. Furthermore,
it computes the quantized normalized channel vectors w
i,k [l] CM 1 ,
for l = 0, . . . , L 1, for all i 6= k, according to
T
T
[w
i,k
[0] . . . w
i,k
[L 1]]T = w
i,k .
(4.59)
4.3.5. Computation of SIN vectors

We define
gi,k [n] , CyclicnM

T
T
T
T
[L 1] . . . wi,k
[1]
wi,k [0] 0TM . . . 0TM wi,k
{z
}
|
N L times
for n = 0, . . . , N 1, so that
H
H
(wi,k
~v
k )[n] = gi,k
[n]v
k .
(4.60)
Each source Sk computes the cyclically-shifted vector sequences

g
i,k [n] CM N 1 , for n = 0, . . . , N 1, for all i 6= k, using the
received feedback of w
i,k [l], for l = 0, . . . , L 1, for all i 6= k, accord-
104
4.3.
ing to
g
i,k [n] = CyclicnM

T
T
T
T
w
i,k [0] 0TM . . . 0TM w
i,k
[L 1] . . . w
i,k
[1]
.
|
{z
}
N L times
k CM N M N is then computed as
The matrix G
k = [
G
g1,k [0] . . . g
1,k [N 1] g
M,k [0] . . . g
M,k [N 1]]
with g
k,k [n], for n = 0, . . . , N 1, replaced by (dummy) all-one vectors
(recall that Sk does not know w
k,k [l], for l = 0, . . . , L 1, and hence
cannot compute g
k,k [n]). Each source Sk computes v
k (and hence
M 1
v
k [n] C
, for n = 0, . . . , N 1,) as
iH
k }1,kN ] Cof[{G
k }M N,kN ]
Cof[{G
h
i .
v
k =

k }1,kN ] Cof[{G
k }M N,kN ]

Cof[{G
h
Thanks to
k }n+1,kN ] = {adj(G
k )}kN,n+1
Cof[{G
k ) {G
1 }kN,n+1
= det(G
k
(4.61)
H
we have v
i,k [n] = 0, and hence

k g
H
g
i,k
[n]v
k = 0,
n = 0, . . . , N 1, i 6= k
(4.62)
Furthermore, since k
gi,k [n]k = 1, kv
k k = 1 and g
i,k [n], v
k CM N 1 ,
for n = 0, . . . , N 1, we can always find unit-norm vectors q1 , . . . ,
qM N 2 such that
{
gi,k [n], v
k , q1 , . . . , qM N 2 }
Although we choose the index kN in the computation of v

k , any of the
indices (k 1)N + 1, (k 1)N + 2, . . . , kN can be chosen without loss of generality.
105
4.
is an orthonormal basis for the M N dimensional complex inner product space. Expanding gi,k [n], i 6= k, into this orthonormal basis, we
get
2
2 H
H
[n]v
k
[n]
gi,k [n] + gi,k
kgi,k [n]k2 = gi,k
+
MX
N 2
H

gi,k [n]qt 2
(4.63)
t=1
H
2 H
2
gi,k
[n]
gi,k [n] + gi,k
[n]v
k
r

2
H

H

gi,k [n]v
k kgi,k [n]k2 gi,k
[n]
gi,k [n]
r

2
H

= 1 wi,k
, i 6= k, n
w
i,k max
d
(4.64)
(4.65)
(4.66)
where (4.66) follows because

kgi,k [n]k2 = kwi,k k2 = 1
and
H
gi,k
[n]
gi,k [n]
H
wi,k
w
i,k .
(4.67)
(4.68)
H
H
Furthermore, substituting gi,k

[n]v
k = |(wi,k
~v
k )[n]| into (4.66), we
get
!
1
H
max
|(wi,k ~ v
k )[n]| d 2
, i 6= k, n
(4.69)
Nd
2 2(M L1)
which, along with (4.51), yields

X P khi,k k2
k6=i
H
|F{(wi,k
~v
k )[n]}|2
X 4P N khi,k k2
2 M L1
k6=i
106
Nd

.
(4.70)
4.4.
OBSERVATIONS
Upon choosing Nd = (M L 1) log P , we get

X P khi,k k2
M
k6=i
H
|F{(wi,k
~v
k )[n]}|2
X 4P N khi,k k2
k6=i

(4.71)
Nd
M L1
| 2 {z }
=1/P
X 4N khi,k k2
k6=i
(4.72)
which is the desired result. Upon substituting (4.72) into (4.41), we

obtain the lower bound
Rsum
P log P
lim
M
X
i=1
1
N
PN 1
r=0
log 1 +
lim
=M
P
M
H
khi,i k2 |F {(wi,i
~
vi )[n]}|2
P
4N khi,k k2
+No
k6=i
M
log P
(4.73)
which proves that full spatial multiplexing gain is, indeed, achieved.
Finally, we complete the proof by noting that the number of bits
broadcast by each destination to its non-corresponding sources is
given by Nf = (M 1)Nd = (M 1)(M L 1) log P .
4.4.
OBSERVATIONS
A frequency-selective multi-input single-output (MISO) interference

network was investigated for feasibility of spatial multiplexing gain
with limited feedback. It was observed that having multiple antennas
at each source provides extra degrees of freedom, compared to SISO
interference networks, that can be employed by each source to nullify
(or cancel) the interference its transmission generates at its non-
107
4.
corresponding destinations. For the case where the source knows its
channels to each of its non-corresponding destinations perfectly, the
cancellation is perfect and each destination receives an interferencefree signal from its corresponding source. This naturally results in
full spatial multiplexing gain. On the other hand, if each source has
only partial CSI of its channels to its non-corresponding destinations,
then the cancellation can still be made near-perfect by employing
the vector quantization scheme of Mukkavilli et al. (2003) and a
transmission scheme called spatial interference-nulling. It was shown
that this near-perfect cancellation yields full spatial multiplexing gain,
provided that the number of feedback bits used to quantize each
channel scales appropriately with the sum transmit power P .
In contrast to spatial multiplexing with interference-alignment,
which depends critically on the assumptions of extremely large number of OFDM tones, large number of taps L in the frequency-selective
channels and feedback from each destination to all the other destinations, full spatial multiplexing gain is achievable with spatial
interference-nulling without any such constraints. The tradeoff, of
course, is that multiple antennas are required at each source for feasibility of spatial interference-nulling while interference alignment
yields full spatial multiplexing gain even with a single antenna at each
source.
108
C HAPTER
MISO Broadcast Networks with Shared

Feedback
multi-input single-output (MISO) broadcast network

consists of a single source S, equipped with M antennas,
and K single-antenna destinations D1 , , DK . The source
obeys an average power constraint of P and has independent messages
for each destination. The capacity region is defined as the convex
closure of the set of K rates that are jointly achievable over the K
source-to-destination links. Weingarten et al. (2006) recently showed
that transmission based on dirty-paper coding achieves the complete
capacity region of this network. Dirty-paper coding is however difficult
to implement in practice because of its dependence on the assumption
of source knowing all the channels in the network perfectly. Several
other transmission schemes have therefore been proposed for achievability of full spatial multiplexing gain in the case where the source
has only partial channel knowledge, obtained through limited capacity
feedback links (see reviews in Gesbert et al. (2007), Love et al. (2008)
and Heath Jr. et al. (2008)). Our focus in this chapter is on the
schemes that achieve full spatial multiplexing with limited feedback
in the large P regime (that is, for P and M, K finite). To be
sure, Lapidoth et al. (2005) showed that full spatial multiplexing gain,
in the large P regime, cannot be achieved with any feedback scheme
109
5.
MISO BROADCAST NETWORKS WITH SHARED FEEDBACK
D1
1
2
D2
DK
Fig. 5.1: A MISO broadcast network with a source equipped with M antennas,
and K single-antenna destinations.
of finite rate for the case where M = K = 2. Lapidoth and Shamai

(2006) further conjectured that spatial multiplexing gain collapses
to one for the general values of M and K, provided the feedback
rate is finite. This conjecture, if true, implies that feedback rate must
scale with P for achievability of full spatial multiplexing gain. In this
chapter, we therefore consider schemes where feedback rates indeed
scale with P . Specifically, we consider two different scenarios.
The K=M case
We first consider the case where the number of destinations is equal
to the number of antennas at the source, that is, K = M . For such a
case, Jindal (2006) showed that full spatial multiplexing gain of M
is achievable in the sum rate, provided that all the channels in the
110
network are frequency-flat and that each destination can feedback to

the source at the rate of (M 1) log P . We generalize the main result
of Jindal (2006) to the case where all channels are frequency-selective
with L taps and show that full spatial multiplexing gain is achievable,
provided that each destination can feedback to the source at the rate
of (M L 1) log P .
The K M case
We next consider the case where the number of destinations is much
larger than the number of source antennas (that is, K M ) but does
not scale with the transmit power P (that is, K is finite). We assume
that all channel coefficients are drawn independently according to a
Rayleigh-distribution. Such a network has been considered independently by Yoo et al. (2007) and Kountouris et al. (2008). Specifically,
Yoo et al. (2007) showed that if all the channels in the network are
frequency-flat, zero-forcing beamforming transmission to a subset
of destinations selected according to semi-orthogonal user-selection
proposed by Yoo and Goldsmith (2006), results in full spatial multiplexing gain of M , provided that each destination can feedback its
(unquantized) SINR along with (M 1) log P log K bits for its
channel direction to the source. Kountouris et al. (2008) proposed,
under the assumption of all channels being frequency-flat (L = 1), a
two-stage feedback framework that can lead to feedback rate reduction
in random multiuser beamforming schemes. In this chapter, we show
that employing a two-stage channel feedback scheme similar to Kountouris et al. (2008) and the transmission scheme proposed for MISO
frequency-selective interference networks in Chapter 4, results in full
spatial multiplexing gain in the case where all channels are frequencyselective with L taps, provided that at least M of the K destinations
can feedback at the rate of (M L 1) log P log K to the source.
We hence generalize the result of Yoo et al. (2007), who assumed all
the channels to be frequency-flat, to the case where all channels are
frequency-selective with L taps. Moreover, we show that in contrast
to the analysis by Yoo et al. (2007), where each of the K destinations
111
5.
needs to feedback (M 1) log P log K bits, our scheme requires

only M of the K destinations to feedback (M L 1) log P log K
bits. Conceptually, this reduction in feedback is achieved using the
ideas of two-stage feedback framework proposed by Kountouris et al.
(2008).
We next present our system model, followed by our results for the
K = M and the K M cases.
5.1.
SYSTEM MODEL
The impulse response of the frequency-selective MISO channel between

the source S and the destination Di is given by the taps hi [l] =
T
M 1
[h1i [l] . . . hM
, for l = 0, . . . , L 1, where hm
i [l]] C
i [l] denotes
the l-th tap of the channel impulse response between the m-th transmit
antenna of S and the single antenna destination Di . The channel
coefficients hm
i [l] are assumed to remain constant throughout the time
interval of interest (outage setting) and are drawn independently from
a continuous probability density function such that 0 < |hm
i [l]| < ,
for all i, l, m, with probability 1. We use a cyclic signal model to convert
the L-tap frequency-selective MISO channel S Di to N parallel
T
frequency-flat MISO channels given by hi (r) = [h1i (r) . . . hM
i (r)]
CM 1 , for r = 0, . . . , N 1, where
m
hm
i (r) = F{hi [n]}
(5.1)
is the channel coefficient between the m-th transmit antenna of S

and the single antenna destination Di , corresponding to the r-th tone.
The input-output relation between S and Di , for the r-th tone, is
then given by
yi (r) = hH
i (r)x(r) + zi (r)
(5.2)
where yi (r) C is the symbol received at the destination Di , x(r) =

[x1 (r) . . . xM (r)]T CM 1 denotes the transmit vector symbol for
source S and zi (r) is CN (0, No ) noise at Di , all for the r-th tone. The
112
5.1.
Iterative Feedback
Scheme
Data Transmission
TF
TTX
SYSTEM MODEL
Fig. 5.2: Feedback and data transmission phases
equivalent relation in time-domain is given by

yi [n] = (hH
i ~ x)[n] + zi [n]
(5.3)
for n = 0, . . . , N 1, where yi [n] = F 1 {yi (r)} C, x[n] = [x1 [n] . . .

xM [n]]T CM 1 with xm [n] = F 1 {xm (r)} C, and zi [n] =
F 1 {zi (r)} C is CN (0, No ) noise. The transmit signals obey the
power constraints

E kx(r)k2 P, r = 0, . . . , N 1.
(5.4)
We shall also need the channel vector hi and the normalized channel vector wi corresponding to the link S Di defined as hi =
[hTi [0] hTi [L 1]]T CM L1 and wi = hi /khi k CM L1 , respectively.
In the remainder of this section, we distinguish between a channel
feedback phase comprising TF time slots during which an iterative
feedback scheme is followed and a data transmission phase that lasts
TT X time slots following the channel feedback phase, as shown in
Fig. 5.2. The purpose of iterative feedback scheme is to select M of
the K destinations so that only these M destinations need to feedback
their CSI to the source. We assume TF TT X so that
1
TF
1
TF + TT X
and the transmission rate loss due to the iterative feedback scheme
can be neglected in our discussion. We assume that each destination
Di knows its channel from the source perfectly, that is, Di knows
113
5.
hi . We furthermore assume that there exists a shared feedback link

from all the destinations D1 , . . . , DK to the source S such that any
destination can feedback Nf bits to the source without interference .
Finally, denoting the rate of communication over the link S Di by
PK
Ri and letting Rsum = i=1 Ri , we say that full spatial multiplexing
gain is achieved if
lim
Rsum
= M.
log P
(5.5)
Recall that spatial multiplexing gain in a MISO broadcast network

with K M is upper-bounded by M .
5.2.
TH E K
M CASE
In this section, we consider a simplified case where the number of

destinations is equal to the number of source antennas, that is, K = M .
This case corresponds to the setup considered in Jindal (2006), with
the following two differences: i) in contrast to Jindal (2006), where all
channels are assumed to be frequency-flat, we assume all channels to
be frequency-selective with L taps and ii) in our setup, the channel
coefficients hm
i [l] can be drawn according to any continuous probability
density function satisfying 0 < |hm
i [l]| < with probability 1, rather
than being restricted to be drawn according to Rayleigh distribution
as assumed by Jindal (2006). Our main result for this section can be
formalized as follows.
Theorem 7 In the MISO broadcast network defined in Section 5.1,
if K = M and the number of feedback bits Nf from each destination
to the source scales with the average transmit power P as
Nf = (M L 1) log P
(5.6 )

In practice, this can be done through a MAC-layer protocol which allows a
destination to feedback only when other destinations are silent
114
5.2.
THE K
M CASE
The proof for the above result is refreshingly simple. In fact, with
slight modification, the feedback and transmission scheme discussed for
the MISO interference network case achieves full spatial multiplexing
gain in MISO broadcast networks as well. The details are as follows:
5.2.1. Feedback scheme

vectors hi [l], for l = 0, . . . , L1, perfectly. Since K = M, the selection
of M destinations out of K destinations is trivial, and hence, there
are no iterations in the iterative feedback scheme. Each Di simply
computes the channel vector hi as hi = [hTi [0] . . . hTi [L 1]]T and
the normalized channel vector wi as wi = hi /khi k. It quantizes wi
to w
i using Nd bits and employing the same quantization scheme
as discussed for MISO frequency-selective interference networks in
Section 4.3.4, so that the maximum quantization error is upperbounded according to
!
q
1
H
2
i| 2
.
(5.7)
max 1 |wi w
Nd
i
2 2(M L1)
Note however that in contrast to the MISO interference networks,
where each Di needs to feedback M 1 normalized channel vectors
(that is, w
i,k , k 6= i), only one normalized channel vector (that is,
w
i ) is fed back by Di in the K = M MISO broadcast network case,
and hence, Nf = Nd .
5.2.2. Transmission scheme

The source S simply transmits
x(r) =
M
X
v
k (r)xk (r)
(5.8)
k=1
where xk (r) C is a scalar symbol with variance P/M and v

k (r)
M 1
C
is the same as the SIN precoding vector v
k (r) computed for
115
5.
frequency-selective MISO interference networks under the assumption

hi,k [l] = hi [l], for all k, for l = 0, . . . , L 1. It then follows from
our analysis for MISO interference networks that if we choose Nd =
(M L 1) log P, we obtain
X P
X 4N khi k2
|hH
vk (r)|2
i (r)
M
M
k6=i
k6=i
4N (M 1)khi k2
.
M
(5.9)
Choosing xk (r), k, r, to be i.i.d. Gaussian and assuming that destination Di knows its effective channel coefficient hH
vi (r) and
i (r)
P
2
its interference power k6=i (P/M )|hH
(r)
v
(r)|
,
the
rate
of comk
i
munication over the link S Di is then lower-bounded according
to
!
N 1
P
H
vi (r)|2
1 X
M |hi (r)
log 1 + P
Ri
(5.10)
P
H
N r=0
vk (r)|2 + No
k6=i M |hi (r)
!
N 1
P
H
vi (r)|2
1 X
M |hi (r)
log 1 + 4N (M 1)kh k2
(5.11)
i
N r=0
+ No
M
PM
Rsum
i=1 Ri
lim
= lim
=M
P log P
P log P
(5.12)
and hence, the full spatial multiplexing gain is achieved. Note that
this result generalizes the result in Jindal (2006) in three aspects,
namely i) to any fading distribution with finite energy; ii) to the
frequency-selective case and iii) to a per channel realization basis
rather than just in the sense of averages over the random channel.
5.3.
TH E K
M CASE
Our main contribution in this section is to show that if the number

of destinations K in the considered broadcast network is large (specif-
116
5.3.
THE K
M CASE
ically, K M ), then the number of feedback bits required to achieve

full spatial multiplexing gain is measurably less than what is required
in the K = M case. The result can be formally stated as follows.
Theorem 8 In the MISO broadcast network defined in Section 5.1,
if all the channels are frequency-selective (L > 1), all the channel
coefficients hm
i [l] are drawn according to a Rayleigh distribution, the
number of destinations K is much larger than the number of source
antennas M and the number of feedback bits Nf from at least M
of the K destinations to the source scales with the average transmit
power P as
Nf = (M L 1) log P log K
(5.13 )
then full spatial multiplexing gain is achievable .

We next present a quantization and feedback scheme which, along
with the transmission scheme discussed for the K = M case, achieves
full spatial multiplexing gain with a number of feedback bits as
specified in (5.13).
5.3.1. Quantization scheme

vectors hi [l], for l = 0, . . . , L 1, perfectly. It computes the channel
vector hi CM L1 and the normalized channel vector wi CM L1
as hi = [hTi [0] . . . hTi [L 1]]T and wi = hi /khi k, respectively. Since
the channel coefficients hm
i [l], for all i, m, l, are drawn according to
a given Rayleigh distribution, each normalized channel vector wi
is uniformly distributed over the surface of the unit hypersphere
in the M L-dimensional complex space. This surface, which can be
mathematically defined as
, {r : r CM L1 and krk2 = 1}
(5.14)
is therefore the quantization space for quantizing wi . In our scheme,

the quantization space is divided into K mutually exclusive and
117
5.
Gi
Qi1
Qi2
Gj
gi
gj
Fig. 5.3: Quantization scheme in three dimensions: Sample space (surface of

the unit sphere) is divided into K groups of 2Nf equal-sized Voronoi regions
each. Two such groups Gi and Gj are shown in detail in the above figure, where
the smallest circles indicate the divisions of a group into 2Nf Voronoi regions
and the dark central lines are the angle-bisectors gi and gj of the solid angles
traversed by the two groups.
non-exhaustive regions, denoted by G1 , , GK and centered around

unit-magnitude vectors g1 , . . . , gK , respectively. Each Gs is further
divided into 2Nf Voronoi regions, denoted by Qs,1 , , Qs,2Nf and
centered around unit-magnitude vectors qs,1 , . . . , qs,2Nf , respectively
(see Fig. 5.3 for a hypothetical illustration in three real dimensions).
Choice of g1 , . . . , gK and G1 , , GK The vectors g1 , . . . , gK are
chosen as a solution to the following line-packing problem: Find the
maximum number of unit-magnitude vectors in C M L such that the
absolute value of the inner product between any two of the vectors is
less than cos(), where (0, /2]. Here, is an auxiliary variable
whose role will become clear later.
118
5.3.
THE K
M CASE
In particular, if Npack is the solution to the line-packing problem

and {p1 , . . . , pNpack } is a set of vectors corresponding to this solution,
we choose
K = Npack
and
{g1 , . . . , g2Nf } = {p1 , . . . , pNpack }.
(5.15)
The region Gs is then chosen as

(
Gs =
r: rC
M L1
)
, krk and |r gs | cos

.
2
2
(5.16)
The above choice of Gs , s, and gs , s, allows us to compute a lower

bound, which we will later need in our proof:
2(M L1)
1
Bound 1 :
sin 2
K22(M
L1) .
To prove this bound, we first note that for any unit-magnitude
vector u in the M L-dimensional complex space (that is, u
CM L1 and kuk2 = 1), the following holds
max |uH gr | cos().
gr
(5.17)
This can be easily proved using the method of contradiction as

follows. Let us assume that there exists a unit-magnitude vector
u CM L1 such that maxgr |uH gr | < cos(). Then, there exists
a solution set of unit-magnitude vectors {p1 , . . . , pNpack , u} such
that the absolute value of the inner-product between any two
vectors is less than cos(). In other words, K (and consequently
Npack ) is not the maximum possible number of unit-magnitude
vectors with the absolute value of the inner product between
any two vectors being less than cos() and can hence not be
the solution of the line-packing problem, which results in a
To be precise, the choice K = N
pack implies that as long as the number
of destinations K is greater than Npack , we simply set K = Npack by randomly
choosing any Npack destinations and ignore rest of the destinations for our analysis.
119
5.
contradiction.
We can now obtain the Bound 1 as follows. The condition (5.17)
implies that each unit-magnitude vector u CM L1 lies in at
least one of the extended regions G1 , . . . , GK , where

Gs , r : r CM L1 , krk2 = 1 and |rH gs | cos () . (5.18)
Therefore, the sum of the surface areas of all the extended regions Gs must be greater than the total surface area of the
unit hypersphere in the M L-dimensional complex space. Since
each of the extended regions Gs has the same surface area
2 M L (sin())2(M L1) /(M L1)! (see Appendix A.8) and the total
surface area of the unit hypersphere is given by 2 M L /(M L 1)!
(see Appendix A.7), we obtain
2 M L
2 M L (sin())2(M L1)
(M L 1)!
(M L 1)!
1
(sin())2(M L1)
K

2(M L1)
1
2 sin
cos
2
2
K
2(M L1)
1
sin
2(M
L1)
2
K2
K
(5.19)
(5.20)
(5.21)
(5.22)
which is the desired lower bound.
Choice of qs,1 , . . . , qs,2Nf and Qs,1 , , Qs,2Nf The vectors qs,1 , . . . ,

qs,2Nf within the group Gs are chosen as a solution to the following
line-packing problem: Find the maximum number of unit-magnitude
vectors in C M L such that the absolute value of the inner product between any two of the vectors is less than cos(1 ), where 1 (0, /2),
and the inner product
of each of the vectors with gs is greater than

cos ( 1 )/2 . Here, 1 is an auxiliary variable whose role will
become clear later.
120
5.3.
THE K
M CASE
Once again, if Npack is the solution to the line-packing problem

and {p1 , . . . , pNpack } is a set of vectors corresponding to this solution,
we choose
2Nf = Npack
and
{qs,1 , . . . , qs,2Nf } = {p1 , . . . , pNpack }.
(5.23)
The region Qs,p is then chosen as

1
M L1
2
H
Qs,p = r : r C
, krk = 1 and |r qs,p | cos
.
2
(5.24)
The above choice of Qs,p and qs,p allows us to compute two bounds,
which we will later need in our proof:
2(M 1L1)

Bound 2 :
sin 21 K21Nf
.
To prove this bound, we first note that the following three properties of Gs and Qs,p :
Property 1: The quantization regions Qs,p1 and Qs,p2 do not
overlap for p1 6= p2 .
Property 2: The regions Gs1 and Gs2 do not overlap for
s1 6= s2 .
Property 3: If the unit-magnitude vector u CM L1 is an
element of Qs,p , for some p {1, . . . , 2Nf }, then u should
also be an element of Gs . That is,
u Qs,p u Gs .
(5.25)
Both Property 1 and Property 2 follow from Appendix A.9 and

Property 3 is proved in Appendix A.10.
We then note that the above three properties together imply
that no two quantization regions Qs,p1 and Qs,p2 , for p1 6= p2 ,
overlap; the quantization region Qs,p , p, lies in Gs ; and no
two regions Gs1 and Gs2 , for s1 6= s2 , overlap. Consequently,
121
5.
none of the quantization regions Qs,p , s, p, overlap with another quantization region. Since there are a total of K2Nf quantization regions Qs,p , each Qs,p has a surface area equal to
2 M L (sin(1 /2))2(M L1) /(M L 1)! (see Appendix A.8), and the
total surface area of the M L-dimensional unit hypersphere is
2 M L /(M L 1)! (see Appendix A.7), we obtain

ML
(sin 21 )2(M L1)
2 M L
Nf 2
(5.26)
K2
(M L 1)!
(M L 1)!

2(M 1L1)
1
1
sin
(5.27)
2
K2Nf
and hence, Bound 2 is proved.

2(M L1)
Bound 3 :
2Nf (sin (1 ))
1
.
K24(M L1)
To prove this bound, we first define the region Gs , for s =

1, . . . , K, as

1
Gs , r : r CM L , krk2 = 1 and |rH gs | cos
2
and note that for any unit-magnitude vector u Gs , the following
holds
max |uH qs,r | cos(1 ).
qs,r
(5.28)
This can be easily proved by contradiction. In particular, let us

assume that there exists a unit-magnitude vector u Gs such that
maxqs,r |uH qs,r | < cos(1 ). Then, there exists a solution set of
unit-magnitude vectors {p1 , . . . , p2Nf , u} such that the absolute
value of the inner-product between any two vectors is less than
cos(1 ) and the absolute value of the inner
product of each vector

with gs is greater than cos ( 1 )/2 . In other words, 2Nf is not
the maximum possible number of unit-magnitude vectors with the
absolute value of the inner product between any two vectors being
122
5.3.
THE K
M CASE
less than cos(1 ) and the absolute value of the inner product of
each vector with gs being greater than cos ( 1 )/2 . Therefore,
2Nf can not be the solution of the line-packing problem, which
results in a contradiction.
We can now obtain the Bound 3 as follows. The condition (5.28)
implies that each unit-magnitude vector u Gs lies in at least
s,p , p = 1, . . . , 2Nf , where
one of the extended regions Q

s,p , r : r CM L1 , krk2 = 1 and |rH qs,p | cos (1 ) .
Q
Therefore, the sum of the surface areas of the extended regions
s,p , p, must be greater than the total surface area of Gs . Since
Q
s,p has the same surface area
each of the extended regions Q
ML
2(M L1)
2 (sin(1 ))
/(M L 1)! and the total surface area of
Gs is given by 2 M L (sin( 1 ))2(M L1) /(M L 1)! (see Appendix A.8), we obtain the following lower-bound:
2
Nf
2(M L1)
1
2 M L sin
2 M L (sin(1 ))2(M L1)
2
(M L 1)!
(M L 1)!
2(M L1)

1
Nf
2(M L1)
2 (sin(1 ))
sin
(5.29)
2
2(M L1)
sin
(5.30)
4
(since 1 < /2)
!2(M L1)
sin 2
(5.31)
(since sin(a/2) sin(a)/2, a (0, /2))

1
(from Bound 1 )
4(M
L1)
K2
and hence Bound 3 is proved.
123
5.
Quantization policy and maximum quantization error We employ a

quantization policy where a given channel vector hi with normalized
channel vector wi Qs,p is quantized to w
i = qs,p . The destinations
with wi Gs but wi
/ Qs,1 Qs,2 Qs,2Nf are not quantized
at all, that is, such destinations are ignored in our scheme and we
will later see that this does not affect our achievability proof. With
this quantization policy, if the source S is aware that wi Qs,1
Qs,2 Qs,2Nf , then the destination Di needs to feedback only Nf
bits to communicate the index of its precise quantization region Qs,p .
Furthermore, since |wiH w
i | > cos(1 /2), following a similar definition
of maximum quantization error as in the frequency-selective MISO
interference networks, we obtain the following upper bound:
q
max
, max 1 |wiH w
i |2
(5.32)
d
i

1
(5.33)
sin
2

2(M 1L1)
1
(from Bound 2 ).
(5.34)
K2Nf
This upper bound will be later used to prove achievability of full
spatial multiplexing gain.
5.3.2. Iterative feedback scheme

We follow an iterative feedback scheme where the source S selects
M of the K destinations and identifies, for each of the M selected
destinations Di , the regions Gs and Qs,p that satisfy wi Gs and
wi Qs,p . The scheme can be elaborated as the sequence of following
four steps:
124
Step 1: The source S sets the number of identified destinations

Ko = 0.
Step 2: The source then randomly chooses one of the regions
1
. The set of vectors
G1 , , GK , say Gs , with equal probability K
M 1
{gs [0], . . . , gs [L 1]}, with gs [l] C
, for l = 0, . . . , L 1, is
5.3.
THE K
M CASE
then computed according to

[gsT [0] . . . gsT [L 1]]T = gs
(5.35)
where gs is the central vector corresponding to the chosen Gs .

Over the next L transmissions, the source transmits
r
P
x[n] =
gs [L 1 n], n = 0, . . . , L 1
(5.36)
M
so that the received symbol at the destination Di for n = L 1
is given by
r
yi [L 1] =
L1
P X H
hi [l]gs [l] + zi [L 1]
M
(5.37)
P
khi kwiH gs + zi [L 1].
M
(5.38)
l=0
r
=
Step 3: Each destination Di computes its normalized received

power as
Pirecv =
1
P
|y 2 [L 1]|2
|wH gs |2
khi k2 i
M i
(5.39)
where the noise variance is assumed to be negligible, for large

values of P , in the computation of Pirecv . Now, if there are ko
receivers with normalized channel vectors lying in one of the
quantization regions Qs,1 , , Qs,2Nf of the chosen Gs , then
these ko destinations will detect a normalized received power
greater than (P/M ) cos2 (o /2). Each of these ko destinations will
then feedback Nf bits corresponding to the index of its precise
quantization region Qs,p in the chosen region Gs (recall that there
are a total of 2Nf quantization regions in Gs and hence Nf bits are
sufficient to communicate its index). Note that since the source
already knows its chosen region Gs , the above feedback ensures
Recall
that Di is assumed to know khi k.
125
5.
that the source accurately identifies the regions Gs and Qs,p

corresponding to each of the ko destinations. The destinations
with normalized channel vector wi
/ Qs,1 Qs,2 Qs,2Nf
remain silent (that is, they do not feedback anything).
Step 4: The source increases Ko by ko and repeats Steps 2 and
3 until Ko M . The iteration of Steps 2 and 3, however, is
performed with two caveats: first, any randomly chosen Gs in
a particular iteration is not chosen for future iterations, and
second, in Step 3 of the last iteration (that is, when Ko M ),
only ko (Ko M ) (rather than ko ) destinations feedback their
Nf bits, where these ko (Ko M ) destinations are randomly
chosen from the ko destinations with wi Qs,1 Qs,2 Qs,2Nf .
It is evident that by following the above four steps, the source

identifies the regions Gs and Qs,p corresponding to M of the K
destinations. We emphasize that we ignored the possibility of two
destinations having their normalized channel vectors in the same
quantization region Qs,p of the chosen region Gs . This is because
the probability of such an event is of the order 1/2Nf and since we
are interested in the large Nf regime, this event can be neglected.
Furthermore, we implicitly assumed in the above scheme that a finite
number of iterations (independent of K and P ) is sufficient for the
source to identify M destinations. This assumption can be justified
as follows:
Since each wi is uniformly distributed over the surface of the M Ldimensional unit hypersphere, the probability of wi lying in one of
the 2Nf quantization regions of the chosen region Gs is equal to the
ratio of the sum of the surface areas of Qs,p in Gs to the total surface
area of the unit hypersphere, that is,
P (wi Qs,1 Qs,2 Qs,2Nf )
2Nf 2
=
ML
(sin(1 ))2(M L1)

(M L1)!
2 M L
(M L1)!
= 2Nf (sin(1 ))2(M L1)
126
(5.40)
(5.41)
5.3.
1
K24(M L1)
THE K
(from Bound 3 ).
M CASE
(5.42)
Therefore, the probability that none of the normalized channel vectors

wi satisfies wi Qs,1 Qs,2 Qs,2Nf in a given iteration is given
by
KKo
1 P (wi Qs,1 Qs,2 Qs,2Nf )
KKo

1
(5.43)
1
K24(M L1)
4(M1L1)
1 2
(for large K and Ko M ) .

(5.44)
e
Consequently, the probability that none of the normalized channel
vectors wi satisfies wi Qs,1 Qs,2 Qs,2Nf in To iterations is
given by
o
4(MTL1)
1 2
.
e
(5.45)
Since the above probability goes to zero for large but finite (independent of K and P ) values of To , a finite number of iterations is
sufficient to identify at least one destination with arbitrarily high
probability, and hence, a finite number of iterations is sufficient to
identify M destinations with arbitrarily high probability.

Our feedback scheme implies that the source knows the quantization
regions Qs,p , and hence the quantization vectors qs,p , corresponding to
the M chosen destinations. Furthermore, the maximum quantization
error for the M selected destinations is upper-bounded according to
(from (5.34))
q
max
=
max
1 |wiH w
i |2
(5.46)
d
i
127
5.
1
K2Nf
2(M 1L1)
.
(5.47)
Ignoring the K M destinations that are not selected, the source

employs the same transmission scheme for the M selected destinations
as was employed for the K = M case discussed earlier. Without
loss of generality, we assume that the M chosen destinations are
D1 , , DM . Following the same analysis as for the frequency-selective
MISO interference networks, the rate of communication over the link
S Di , i = 1, , M, is then lowerbounded according to
!
N 1
P
H
2
|h
(r)
v
(r)|
1 X
i
i
M
log 1 + P
Ri
(5.48)
P
H
N r=0
vk (r)|2 + No
k6=i M |hi (r)
!
N 1
P
H
vi (r)|2
1 X
M |hi (r)
log 1 + P
(5.49)
PN
max )2 + N
2
N r=0
o
k6=i M khi k (d
N
1
P
H
2
X
vi (r)|
1
M |hi (r)
log 1 +
(5.50)
1
(M L1)

P
N r=0
PN
1
2
+ No
k6=i M khi k
K2Nf
which, upon choosing Nf = (M L 1) log P log K, yields
N 1
P
|hH
vi (r)|2
1 X
i (r)
log 1 + PM 4N
Ri
2
N r=0
k6=i M khi k + No
PK
PM
i=1 Ri
i=1 Ri
lim
= lim
=M
P log P
P log P
!
(5.51)
(5.52)
and hence, the full spatial multiplexing gain is achieved.

We emphasize that this result is stronger than the result in Yoo
et al. (2007) because i) only M destinations need to feedback (M L
1) log P log K bits, in contrast to Yoo et al. (2007) where all K
destinations need to feedback (M L 1) log P log K bits; ii) the
result is valid for frequency-selective channels and iii) the result holds
on a per channel realization basis rather than just in the sense of
averages over the random channel.
128
5.4.
5. 4 .
OBSERVATIONS
OBSERVATIONS
A multi-input single-output (MISO) broadcast network was investigated for feasibility of spatial multiplexing with limited feedback. It
was assumed that the source is equipped with M antennas, that there
are K single-antenna destinations and that all the channels in the network are frequency-selective with L taps. Spatial multiplexing to M of
the K destinations by employing spatial interference-nulling precoding
vector sequences, computed according to Section 4.3.5 of Chapter 4,
was shown to result in full spatial multiplexing gain. For the case
where the number of destinations K is equal to the number of source
antennas M , it was observed that a feedback rate of (M L 1) log P
from each destination to the source guarantees the achievability of
full spatial multiplexing gain. However, if the number of destinations
K is much larger than the number of source antennas M , then a
measurably smaller feedback rate of (M L 1) log P log K from M
of the K destinations was found to be sufficient for achievability of full
spatial multiplexing gain. The key to this reduction of feedback rate
by log K is a two step channel vector quantization scheme, where the
quantization space of the normalized channel vectors is first divided
into K non-overlapping regions and each region is then divided into
a number of Voronoi regions. The source can identify M sources and
the corresponding M non-overlapping regions with a very few bits of
feedback using an innovative iterative algorithm, and the M sources
therefore need to feedback only the indices of their Voronoi regions
within their respective non-overlapping regions. Provided that each
of the K non-overlapping regions is approximately (1/K)-th of the
total quantization space, each of the M destination requires log K
less bits to feedback the index of any of its Voronoi region compared
to the case where the iterative algorithm is not followed.
Throughout this chapter, two important considerations were ignored. First, fairness of rates of communication to different destinations was not considered, that is, some destinations may get very high
rates of communication while others may have a zero rate. Second,
multiuser diversity gain that is inherent in the system due to the
129
5.
presence of a large number of destinations K is not exploited by

the transmission scheme. Because these two considerations are of
immense importance in the systems operating in the finite P regime,
the benefits of the scheme proposed in this chapter may be limited in
practice.
130
C HAPTER
Conclusions and Outlook
Spatial multiplexing is one of the most promising transmission techniques to achieve high communication rates in future wireless systems.
In some wireless systems such as interference and broadcast networks
however, spatial multiplexing is feasible only if the source(s) have
at least some channel knowledge. Since channel knowledge is often
obtained at the sources through limited capacity feedback links, it
is important to develop transmission schemes that can achieve full
spatial multiplexing gain in broadcast and interference networks with
limited feedback. In this thesis, we proposed (limited) feedback and
transmission schemes for three different classes of interference networks and a class of broadcast networks and showed that the schemes
achieve full spatial multiplexing gain.
Large interference networks

We first considered a large interference network that models the
second hop of the multiuser relaying protocol proposed by Bolcskei
et al. (2006). Assuming that first hop transmissions are error-free
and that 1-bit broadcast feedback links between each destination and
its corresponding sources exist, we proposed an iterative feedback
scheme and a transmission scheme that together achieve full spatial
multiplexing gain. A logical next step in this research would be to
relax the assumption of error-free first hop transmission and analyze
131
6.
CONCLUSIONS AND OUTLOOK
the effects of this relaxed assumption on the spatial multiplexing

gain achievable in the network. Another issue worth investigation
is the performance of the feedback and transmission scheme in the
case where all the channels in the network, rather than being static,
are slowly changing over time. Finally, the scheme requires very low
feedback rate and it should therefore be possible to generalize it to
other types of large networks than the one considered in this thesis.
We believe that a deeper look into these generalizations can throw
open a number of potential research problems.
SISO interference networks

We second considered an M source-destination pair frequency-selective
SISO interference network and showed that naive interference alignment along with an appropriate quantization and feedback scheme can
provide full spatial multiplexing gain. We believe that our contribution
is a significant step towards the understanding of achievable spatial
multiplexing gains in interference networks with limited feedback. It
should however be cautioned that practical benefits of interference
alignment might be less spectacular. This is because the capacity
gains from interference alignment are apparent only at very large
values of signal-to-noise ratio, which are rarely encountered in practice.
Moreover, spatial multiplexing in interference alignment depends critically on the assumption of a very large number of OFDM tones and a
large number of channels taps L. A challenging research problem is to
adapt interference alignment so that it can yield high communication
rates for practical values of signal-to-noise ratio and other system
parameters.
MISO interference networks

We then considered an M source-destination pair frequency-selective
interference network where each source is equipped with at least M
antennas and each destination is equipped with a single antenna.
We proposed a feedback and transmission scheme, which we termed
132
spatial interference-nulling, and showed that the scheme achieves full

spatial multiplexing gain of M . In contrast to interference alignment,
spatial interference-nulling does not require the assumptions of large
OFDM tones or number of taps in the channel. In fact, we believe that
it is much easier to adapt spatial interference-nulling so that it yields
tangible performance gains for practical values of signal-to-noise ratio,
than is the case with interference alignment. All these benefits come,
of course, at the cost of having multiple antennas at each source.
MISO broadcast networks

MISO broadcast networks have been extensively and almost exhaustively studied over the last few years. Several remarkable transmission
schemes have been proposed in the literature. Feasibility of spatial
multiplexing in frequency-selective MISO broadcast networks in the
large source power regime (that is, for P ) is not however completely understood. We proposed a feedback and transmission scheme
that achieves full spatial multiplexing gain in frequency-selective
broadcast networks with limited feedback. We moreover quantified
the required feedback rate scaling with P . In addition, we showed
that if the number of destinations is very large and all the channel
coefficients are drawn according to a Rayleigh distribution, the feedback rate required for achievability of full spatial multiplexing gain
is measurably less than the case where the number of destinations is
small. The scheme however does not capture the multiuser diversity
gain that is inherently present in broadcast networks with large number of destinations. We believe that to capture full multiuser diversity
gain, destinations also need to feedback their channel magnitudes
(apart from the channel direction feedback which is considered in
our scheme) to the source, and investigation of this issue should be a
part of future research. We also believe that the Rayleigh-distribution
assumption is not critical to our result and it should be possible to
generalize the proposed quantization and feedback schemes to the
case of more general channel coefficient distributions. Finally, an
interesting problem would be to derive guidelines from the proposed
133
6.
CONCLUSIONS AND OUTLOOK
scheme for developing practical wireless systems operating at finite

values of P .
134
APPEN DIX A
Bounds and Proofs
A .1. U N ION B OU N DS
Assume the real-valued random variables X1 , X2 , X3 are such that
P (|Xi | Ci ) Pi ,
i = 1, 2
and
P (|X3 | C3 ) P3
(A.1)
where C1 , C2 and C3 are positive constants.

Then, we can have the following union bounds:
Union bound for summation

P (|X1 + X2 | C1 + C2 ) P1 + P2 .
Union bound for division

|X1 |
C1
P
P1 + P3 .
|X3 |
C3
(A.2)
(A.3)
A .2. LARGE DEVIATIONS B OU N DS

Let hi , i = 1, 2, . . . , M be i.i.d. N (0, 1) and let s > 0 be an optimization variable. Then, for k > 0, we can have the following bounds:
135
A . BOUNDS AND PROOFS
Bound 1:
P
M
X
!
|hi | < k
i=1

PM
= P es( i=1 |hi |) > esk
h
i
PM
E es( i=1 |hi |)
(by Chebyshevs Inequality) (A.4)

esk
h
iM
= esk E es|hi |
(since hi are i.i.d.)
(A.5)
iM
h s2
(A.6)
= esk e 2 2Q(s)
#M
" r
s2
2 1 s2
sk
2
2
e
(from Mckean (1969), p.5)
e
e
s
(A.7)
"r
#M
2
1
min esk
s>0
"r
#M
2
1
= esk
(A.8)
(A.9)
s= M
k
Bound 2:
P
M
X
i=1
136
r !M
2
.
ek
M
!
|hi | > k
(A.10)
A .3. INTERSECTION BOUND
PM

= P es( i=1 |hi |) > esk
h PM
i
E es( i=1 |hi |)
(by Chebyshevs Inequality)
esk
h
iM
= esk E es|hi |
(since hi are i.i.d.)
h s2
iM
= esk e 2 2(1 Q(s))
h s2 i M
esk 2e 2
h
iM
s2
sk
M
2
min 2e
e
s>0
h
iM
s2
sk
M
2
= 2e
e
(A.11)
(A.12)
(A.13)
(A.14)
(A.15)
(A.16)
k
s= M
M

k2
.
= 2e 2M 2
(A.17)
Bound 3: (Shwartz and Weiss, 1995, Section 1.3)

!
M
X
k2
P
hi k e 2M .
(A.18)
i=1
A .3. I NTERSECTION B OU N D
Assume the real-valued random variables X1 , X2 , X3 are such that
P (|Xi | Ci ) Pi ,
P (|X3 | C3 ) P3 ,
i = 1, 2
and
(A.19)
137
where C1 , C2 and C3 are positive constants.

Then, we can have the following intersection bounds:
Intersection bound for summation

P (|X1 + X2 | C1 + C2 ) P1 P2 .
Intersection bound for division

|X1 |
C1
P
P1 P3 .
|X3 |
C3
(A.20)
(A.21)
A . 4 . I NTER FER ENC E ALIGNMENT WITH

PER FECT CSI
We present a constructive scheme, originally proposed by Cadambe
and Jafar (2008), to determine the transmit direction vectors vkm , k,
for m = 1, . . . , dk , such that all the directions corresponding to
H vm , k 6= i,
interference symbols at each destination Di , that is, H
i,k k
for m = 1, . . . , dk , span an N di dimensional complex space. We
start by defining
xk , [x1k . . . xdkk ]T
Vk ,
[vk1
...
vkdk ]
Cdk 1
(A.22)
N dk
(A.23)
In this appendix, we assume, for ease of notation, that H

H vm and
i,k k
H
m
H
m
Hi,k vk /kHi,k vk k are the same because existence of one in a N di dimensional

complex space implies the same for the other.
138
A .4. INTERFERENCE ALIGNMENT WITH PERFECT CSI
so that
x
k = Vk xk
(A.24)
and the input-output relation for the link Si Di can be stated as

y
i =
M
X
H
H
zi .
i,k Vk xk +
(A.25)
k=1
Now, the interference at destination Di , i, from the sources Sk , k 6= i

will be restricted to N di dimensions (recall N di = tQ , for i = 1,
and N di = (t + 1)Q , i 6= 1) out of a total of N = (t + 1)Q + tQ
dimensions corresponding to y
i if the following two sets of conditions
are satisfied:
H
H
H
H
H
1,2 V2 = H1,3 V3 = H1,4 V4 = = H1,M VM
H
H
H
k
/ {1, i}, i 6= 1
i,k Vk Hi,1 V1 ,
(A.26)
(A.27)
where (A.26) implies that the interference symbols from sources

S2 , , SM at destination D1 are perfectly aligned and the (M
1)(M 2) conditions in (A.27) imply that interference symbols from
sources Sk , k
/ {1, i} at destination Di , i 6= 1, lie in a (t + 1)Q dimen H V1 .
sional complex space spanned by the columns of the matrix H
i,1
Q
Q
i,k have a full rank of (t + 1) + t almost
Since channel matrices H
surely, equations (A.26) and (A.27) can be equivalently expressed as
Vk = Sk E,
Di,k E V1 ,
k = 2, . . . , K
k
/ {1, i}, i 6= 1
(A.28)
(A.29)
139
where
1 H
H
E = (H
H2,3 V3 CN d3
2,1 )
1 H H 1 H
H
Sk = (H
H1,3 (H2,3 ) H2,1 CN N ,
1,k )
Di,k
(A.30)
k = 2, . . . , M
1 H
H
= (H
Hi,k Sk , CN N ,
i,1 )
(A.31)
i = 2, 3, . . . , M, k = 2, 3, . . . , i 1, i + 1, . . . , M.
(A.32)
We now choose V1 and E so that they satisfy the (M 1)(M 2)

conditions in (A.29) and then use (A.28) to determine V2 , , VM .
In particular, defining the set

S , (i, k) : i {2, 3, . . . , M }, k {2, 3, . . . , M },

i 6= k, (i, k) 6= (2, 3)
(A.33)
the tQ = t(M 1)(M 2)1 column vectors of E are chosen to be of the
form
Y

CN 1
(Di,k )i,k 1N
(A.34)
(i,k)S
where i,k {0, 1, . . . , t 1}, (i, k) S. Notice that since each i,k
can take t values and there are Q = (M 1)(M 2) 1 terms in
the product in (A.34), precisely tQ column vectors of the form (A.34)
exist. The (t + 1)Q column vectors of V1 are chosen to be of the same
form as the column vectors of E, that is,
Y

(Di,k )i,k 1N
(A.35)
(i,k)S
140
A .4. INTERFERENCE ALIGNMENT WITH PERFECT CSI
with the difference that i,k {0, . . . , t}, (i, k) S, in contrast to

i,k {0, . . . , t 1}. Finally, the matrices Vk , for k = 2, . . . , M, are
computed in a straightforward fashion using (A.28). The fact that
the above choice of E and Vk , k, satisfies the interference alignment
equations (A.29) can be easily verified as follows. For the special case
(i, k) = (2, 3), we have Di,k = IN , which implies that
Di,k E = E V1
(A.36)
and hence, (A.29) is satisfied. For the general case where (i, k) S, let
us consider any column vector of the matrix E, say e, that corresponds
to a given set {i,k : (i, k) S}. Then,
Y

p,q
Di,k e = Di,k
(Dp,q )
1N
(A.37)
(p,q)S
=g
(A.38)
where g is a column vector of V1 corresponding to

(
p,q + 1 if (p, q) = (i, k)
p,q =
p,q
if (p, q) 6= (i, k), (p, q) S
(A.39)
and hence,
Di,k E V1
(i, k) S.
(A.40)
It therefore follows that with the above construction of Vk , k, the

conditions (A.28) and (A.29) are satisfied and the interference at
destination Di , i, is restricted to N di dimensions. However, we
still need to check that the direction vectors corresponding to the
H Vi ),
desired signal at destination Di (that is, the column vectors of H
i,i
141
are linearly independent of the direction vectors corresponding to

H Vk where
the interference at Di (that is, the column vectors of H
i,k
k = 1, i 6= 1 and k = 2, i = 1). This can be verified by showing that
the matrix
H
[H
i,i Vi
H
H
i,k Vk ]
(A.41)
is full rank for all i, where k = 1, i 6= 1 and k = 2, i = 1. Since the

i,k , i, k, are drawn independently from a contindiagonal entries of H
uous distribution, the non-singularity of (A.41) is intuitive (a rigorous
proof can be found in (Cadambe and Jafar, 2008, Appendix III))
as long as Vk , k, has full rank dk (that is, vkm , m, are linearly
independent), which is proved in Appendix A.5.
A .5. LI N EAR I N DEPEN DENC E OF
vkm , m
We derive conditions on L under which v1m , m = 1, . . . , (t + 1)Q , are

linearly independent, that is, V1 has rank d1 = (t + 1)Q . To this end,
N 1
1
we start by defining F = diag{1, ej2 N , . . . , ej2 N } so that
H =
H
i,j
L1
X
hi,j [l]Fl ,
i, j.
l=1
We then expand v1m according to

Y
(from (A.35))
v1m =
(Di,k )i,k 1N
(i,k)S
Y
(i,k)S
142
1 H
H
(H
Hi,k Sk
i,1 )
i,k
1N
(A.42)
A .5. LINEAR INDEPENDENCE OF
Y
H
H H
H
i,1 H1,k H2,3
i,k
H
H H
H
i,k H1,3 H2,1
vkm , m
i,k !
1N
(i,k)S
i,k , i, k)
(using commutative property of diagonal matrices H
Y

H
H H t
=
H
i,1 H1,k H2,3
(i,k)S
Y
H
H H
H
i,1 H1,k H2,3
ti,k
H
H H
H
i,k H1,3 H2,1
i,k

1N
(A.43)
(i,k)S
where i,k {0, 1, . . . , t}, (i, k) S, so that every distinct choice of

set {i,k : (i, k) S}, upon substitution into (A.43), yields one of the
(t + 1)Q vectors v1m , m = 1, . . . , (t + 1)Q . Using (A.42), (A.43) can
then be rewritten as
Y
v1m =
H
H H
H
i,1 H1,k H2,3
t
3tQ(L1)
X

g
m
F
1N
1,g
(A.44)
g=0
(i,k)S
where m
1,g , g = 1, . . . , 3tQ(L 1), are some polynomial functions in
hi,j [l], i, j, l with powers of hi,j [l] being functions of i,j . Defining
fg CN 1 as fg , Fg 1N , we can further simplify (A.44) as
Y
v1m =
H H
H
H
i,1 H1,k H2,3
t
3tQ(L1)
X
(A.45)
g=0
(i,k)S

m
f
1,g g .
{z
(a)
}|
{z
(b)
Since term (a) is the same for all m and term (b) is a linear combination of vectors fg , g = 0, . . . , 3tQ(L 1), the vectors v1m , m =
1, . . . , (t + 1)Q , will be linearly independent only if
3tQ(L 1) + 1 (t + 1)Q
(A.46)
143
(t + 1)Q 1
+1
3tQ
(t + 1)Q 1
L>
3tQ
(A.47)
(A.48)
that is, L must be greater than ((t + 1)Q 1)/(3tQ) for feasibility
of interference alignment. Also note that satisfying (A.48) indeed
guarantees that v1m , m = 1, . . . , (t+1)Q , are linearly independent. This
is because hi,j [l], i, j, l, are drawn independently from a continuous
probability density function, and hence, the probability of vectors
m
m
[m
1,1 1,2 . . . 1,(t+1)Q ], m, being linearly dependent is almost surely
0.
Finally, we note that the above analysis can be easily extended
to show that Vk , k 6= 1, will be of rank tQ (that is, vkm , m, will be
linearly independent) if and only if
L>
tQ 1
.
3tQ
(A.49)
Since (A.48) is a stricter condition than (A.49), it is the dominant

condition.
A .6. LI N E PAC KI NG B OU N D
We use a bounding technique from Mukkavilli et al. (2003) to upperbound 2Nd in terms of sin(). In particular, we first define a
spherical cap around each unit-magnitude quantization vector pr , r =
1, . . . , Npack , as

2
2
(pr ) = x CM : kxk2 = 1 and |pH
x|
cos
.
r
2
144
A .7. SURFACE AREA OF
m- DIMENSIONAL
HYPER- SPHERE
The spherical caps (pr ) have the following two key properties:
1. There is no overlap between any two of the spherical caps. This
follows from Appendix A.9.
2. The surface area of the spherical caps satisfies (see Appendix A.8)
|(pr )| =
2 M (sin( 2 ))2(M 1)
,
(M 1)!
r = 1, . . . , 2Nd .
Noting that the total surface area of the M -dimensional unit hypersphere is given by || = 2 M /(M 1)! (see Appendix A.7) and
using the fact that the spherical caps are non-overlapping, we get the
following upper-bound:

2(M 1)
M
2 M
Nd 2 (sin 2 )
(A.50)
2
(M 1)!
(M 1)!
2(M 1)
Nd
2 sin
(A.51)
2

2(M 1)
sin()
.
(A.52)
2
A .7. SU R FAC E AR EA OF
m- DIMENSIONAL
HYPER - SPH ER E
In this appendix, we elaborate the computation of the surface area
of an M -dimensional complex hypersphere as given in the Appendix
I of Mukkavilli et al. (2003) and based on a methodology given in
Kendall (1961).
We will first compute the volume of the complex sphere khk r,
145
where h CM . Let h = [h1 hM ]T , where hj C, j = 1, . . . , M .

Further, for each j, let hj = xjr + jxji , that is, the real and imaginary
parts of hj are given by xjr and xji respectively. Now consider the
transformation of variables given by
xjr = rj cos(j )
xji = rj sin(j )
(A.53)
so that hj = rj ejj for all j.

The Jacobian for the transformation in (A.53) is given by
|J| = det (diag (T1 , . . . , TM ))
(A.54)
where

Tj =
cos(j )
sin(j )
rj sin(j ) rj cos(j )
det(Tj ) = rj cos2 (j ) + rj sin2 (j ) = rj .
(A.55)
(A.56)
Therefore,
|J| = det (diag (T1 , T2 , . . . , TM ))
=
M
Y
(A.57)
det(Tj )
(A.58)
= r1 r 2 r M .
(A.59)
j=1
As a result, the volume elements across the transformation are related

as
dx1r dx1i dx2r dx2i dxM r dxM i
= r1 r2 rM dr1 dr2 drM d1 d2 dM
146
(A.60)
A .7. SURFACE AREA OF
m- DIMENSIONAL
HYPER- SPHERE
and the volume of the hypersphere khk r can be computed as

Z
Z
VM (r) =
dx1r dx1i dx2r dx2i dxM r dxM i

(A.61)
PM
j=1
x2jr +x2ji r 2
r1 r2 rM dr1 dr2 drM d1 d2 dM (A.62)
PM
2
2
j=1 rj r
0i 2
M
= (2)
r1 r2 rM dr1 dr2 drM
(A.63)
PM
2
2
j=1 rj r
where we have integrated over the M variables 1 , 2 , . . . , M in (A.63).

Now consider another transformation given by
r1 = uc1 c2 cM 3 cM 2 cM 1
r2 = uc1 c2 cM 3 cM 2 sM 1
r3 = uc1 c2 cM 3 sM 2
rM 1 = uc1 s2
rM = us1
(A.64)
where we have used the notation ci = cos(i ) and si = sin(i ), for

i = 1, . . . , M 1. Note that i goes only from 0 to /2 since ri 0 for
each i. Further, u goes from 0 to r. The Jacobian of this transformation
is given by (Kendall, 1961, pp 35)
2 M 3
uM 1 cM
c2
cM 2
1
(A.65)
147
and the new volume element is given by

2 M 3
uM 1 cM
c2
cM 2 dud1 d2 dM 1 .
1
(A.66)
Substituting (A.64) and (A.66) in (A.63), we get after simplification
Zr
VM (r) = (2)
u
u=0
2t1
Z/2 Z/2
du
1 =0 2 =0
Z/2
M 1 =0
3
5
(c2M
s1 )(c2M
s2 ) (c3M 2 sM 2 )(cM 1 sM 1 )
1
2
d1 d2 dM 1
r2 M
1
1
11
2M 2M 2 2M 4
42
(2)M r2M
=
.
2M M !
(A.67)
= (2)M
(A.68)
Finally, the surface area is obtained by differentiating the volume

of the sphere as follows:
dVM (r)
dr
2 M r2M 1
=
.
(M 1)!
AM (r) =
(A.69)
A .8. AR EA OF A SPH ER ICAL CAP

Let AM (r, ro ) denote the area of the spherical cap formed by the
intersection of the subspace |h1 | > ro and the t-dimensional complex
hypersphere given by khk = r, where h CM and h = [h1 hM ]T
148
A .8. AREA OF A SPHERICAL CAP
such that hj Cj = 1, 2, . . . , M . We will use the same technique

as in the Appendix A.7, where we first calculate the volume of a
solid object and treat the surface area as a differential element of
the volume at the required radius. Now consider the region formed
by the intersection of the subspace |h1 | > ro with the interior of the
hypersphere given by khk r. In terms of the transformation given
in (A.53), this region is given by
(
)
M
X
M
jj
2
2
h C : hi = ri e ,
ri r , r 1 r o .
i=1
The volume of this region, denoted by VM (r, ro ), is given by the

integral in (A.63) where the limits of ri are now different for each i.
In particular, we have
Zr
Z
Z
VM (r, ro ) = (2)M
r2 rM dr2 drM r1 dr1 .
r1 =ro
PM
j=2
rj2 r 2 r12
(A.70)
Note that the multiple integral in brackets in (A.70) is nothing but
the
p volume of (t 1)-dimensional complex hypersphere of radius
r2 r12 with a scaling factor of (2)M 1 , which is given by (A.68).
On substitution, we get
VM (r, ro ) =
(2)M
M
1
2
(M 1)!
Zr
r1 (r2 r12 )M 1 dr1
(A.71)
r1 =ro
149
(2)M
= M 1
2
(M 1)!
Zr
(r2 u)M 1
du
2
where
u = r12
u=ro2
(A.72)
M
(2)
(r2 ro2 )M .
2M M !
(A.73)
Finally the surface area of the spherical cap is obtained by differentiating VM (r, ro ) as follows:
dVM (r, ro )
dr
2 M
=
r(r2 ro2 )M 1 .
(M 1)!
AM (r, ro ) =
(A.74)
A .9. NON - OVER LAP OF SPH ER ICAL CAPS

Lemma 1 If a spherical cap around the unit magnitude vector pi
CM 1 , for i = 1, is defined as

M
2
H 2
2
(pi ) , r C : krk = 1 and |pr r| cos
2
and pH
1 p2 < cos(), then (p1 ) and (p2 ) do not overlap.
This can be proved by contradiction as follows:
Let us assume that the spherical caps (p1 ) and (p2 ) overlap.
Then, the intersections of (p1 ) and (p2 ) with the linear span of
p1 and p2 (i.e. the two-dimensional plane containing all possible
linear combinations of p1 and p2 ) must also overlap. Let zo be
a unit-magnitude vector that lies on the region where the two
150
A .10. PROOF OF PROPERTY
intersections overlap. Then, since zo (p1 ) and zo (p2 ),

we have

H
pH
z
cos
and
p
z
cos
.
1 o
2 o
2
2
However, since zo , p1 and p2 lie on the same two dimensional

plane, this can happen only if pH
1 p2 cos(), which is not
possible. Therefore, our assumption is wrong and by contradiction,
we can conclude that the two spherical caps do not overlap.
A .10. PRO OF OF PROPERTY
Property 3 can be proved as follows. Since u Qs,p , we have

1
|uH qs,p | cos
(A.10.75)
2
1
arccos(|uH qs,p |).
(A.10.76)
2

Furthermore, since |qH
s,p gs | cos ( 1 )/2 , we have
1
arccos(|qH
s,p gs |).
2
(A.10.77)
Adding (A.10.76) and (A.10.77), we obtain
arccos(|uH qs,p |) + arccos(|qH

s,p gs |)
2

cos
cos(arccos(|uH qs,p |) + arccos(|qH
s,p gs |))
2
(A.10.78)
(A.10.79)
= |uH qs,p | |qH

s,p gs |
151
1 |uH qs,p |2
2
1 |qH
s,p gs | .
(A.10.80)
Let {qs,p , t1 , . . . , tM L1 } be a set of orthonormal vectors in the M Ldimensional complex space such that u and gs can be expanded on
this set as
u = u1 qs,p +
MX
L1
u r tr
(A.10.81)
r=1
gs = gs,1 qs,p +
MX
L1
gs,r tr .
(A.10.82)
r=1
Then, (A.10.80) can be rewritten as

v
v
uM L1
uM L1

u
u X
X
t
2
cos
|u1 | |gs,1 |
|ur | t
|gs,r |2
2
r=1
r=1
v
v
uM L1
uM L1
u X
u X
2
|ur | t
|gs,r |2
= |u1 | |gs,1 |
r=1
|u1 gs,1 |
MX
L1
(A.10.83)
(A.10.84)
r=1
|ur gs,r |
r=1
(using Cauchy-Schwarz inequality)
(A.10.85)
MX
L1

u1 gs,1 +
ur gs,r
r=1
(since|a| |b| |a + b|, a, b C11 ) (A.10.86)

= |uH gs |
u Gs
152
(A.10.87)
(A.10.88)
A .10. PROOF OF PROPERTY
and hence, Property 3 is proved.
153
APPEN DIX B
Wireless Systems and Channel Feedback:

A primer
General wireless systems consist of one or more sources communicating
with one or more destinations over wireless channels. Each source
and each destination is equipped with at least one antenna and all
antennas of a given source and all antennas of a given destination
cooperate fully among themselves. If there is only one source and only
one destination in a wireless system, it is called a single-user system.
A system with more than one sources or more than one destinations
is naturally called a multiuser system.
B .1. SI NGLE - USER SYSTEM

A single-user system consists of a single source S communicating with
a single destination D. Depending on the number of antennas the
source and the destination are equipped with, a single-user system can
be broadly classified as a single-input single-output (SISO) channel
or a multi-input multi-output (MIMO) channel.
155
B. WIRELESS SYSTEMS AND CHANNEL FEEDBACK : A PRIMER
B.1.1. Single-input single-output channel

A simple discrete-time single-input single-output (SISO) channel is
represented by the input-output (IO) relation
y[n] = h[n]x[n] + z[n]
(B.1.1)
where x[n] denotes the symbol transmitted by the source S, y[n] is

the symbol received at the destination D, h[n] denotes the channel
coefficient between source S and destination D and z[n] is CN (0, No )
noise at destination D, all for n-th time index. The transmit symbol
x[n] obeys the average power constraint
E[|x[n]|2 ] P.
(B.1.2)
A. Perfect channel knowledge at the destination assumption

In many cases of practical interest, it can be assumed that the destination D knows the channel coefficients h[n] perfectly. With this
assumption and with the choice of x[n] being i.i.d. Gaussian, the
rate of communication over the link S D is given by (Tse and
Viswanath, 2005, Sec. B.7.1)

|h[n]|2 P
R = E log 1 +
.
(B.1.3)
No
B. Time-selective and time-flat channels
The channel model represented by the IO relation (B.1.1) corresponds
to a time-selective channel, that is, the channel coefficient h[n] changes
with time index n. If the channel coefficient h[n] does not change with
n, we say the channel is time-flat and the IO relation (B.1.1) can be
156
B.1. SINGLE - USER SYSTEM
restated as
y[n] = hx[n] + z[n].
(B.1.4)
Assuming that D knows h perfectly and choosing x[n] to be i.i.d.

Gaussian, the rate of communication over the link S D is given by

|h|2 P
R = log 1 +
.
(B.1.5)
No
In this thesis, we have considered time-flat channels only.
C. Frequency-selective and frequency-flat channels
The channel model represented by the IO relation (B.1.4) corresponds
to a frequency-flat channel. A more general channel model is frequencyselective channel, which is represented by the IO relation
y[n] =
L1
X
h[l]x[n l] + z[n]
(B.1.6)
l=0
where h[l], l = 0, . . . , L1, denotes the l-th tap of the channel impulse
response between S and D.
D. Orthogonal frequency division multiplexing
Orthogonal frequency division multiplexing (OFDM) is a common
transmission technique to convert an L-tap frequency-selective channel
to N parallel frequency-flat channels with N L. Specifically, OFDM
entails that the source S computes its transmit symbol x[n], n =
0, . . . , N 1, as the inverse discrete Fourier transform (IDFT) of an
N -point frequency-domain sequence x(r), r = 0, . . . , N 1, according
157
to
x[n] = F 1 {x(r)}
(B.1.7)
and the destination D computes an N -point frequency-domain sequence y(r), r = 0, . . . , N 1, as the discrete Fourier transform (DFT)
of its received symbol y[n] according to
y(r) = F{y[n]}.
(B.1.8)
It can then be shown that x(r) and y(r) follow the input-output
relation
y(r) = h(r)x(r) + z(r),
r = 0, . . . , N 1
(B.1.9)
where h(r) = F{h[n]} is the channel coefficient between S and D,

and z(r) is CN (0, No ) noise at D, both for the r-th tone (or r-th
frequency index). Choosing x(r) to be i.i.d. Gaussian and under the
assumption of perfect channel knowledge at D and no feedback from
D to S, the rate of communication over the link S D is given by
R=

N 1
1 X
|h(r)|2 P
log 1 +
.
N r=0
No
(B.1.10)
E. Channel feedback and spatial multiplexing

Even with availability of perfect channel knowledge at S, spatial multiplexing gain (or pre-log term in capacity expression) of a SISO channel
is upper-bounded by one (Tse and Viswanath, 2005, Sec. 5.4.6). Since
Strictly speaking, the sequence x[n] includes a cyclic prefix of L 1 symbols,
which we assume to be negligible and will therefore neglect for the rest of our
discussions in this thesis.
158
B.1. SINGLE - USER SYSTEM
a spatial multiplexing gain of one is any case achievable without channel feedback in (B.1.5) and (B.1.10), it trivially follows that channel
feedback is not required for achievability of full spatial multiplexing gain in a SISO channel. To be sure, channel feedback can be
used in the frequency-selective case to increase the term inside the
log() in the capacity expression (B.1.10) through water-filling (Tse
and Viswanath, 2005, Sec. 5.4.6). However, in this thesis, we have
restricted our analysis to the role of channel feedback in increasing
spatial multiplexing gain, and have ignored other potential benefits
of feedback.
B.1.2. Multi-input multi-output channel

A. Frequency-flat case
A frequency-flat multi-input multi-output (MIMO) channel with M
antennas at S and K antennas at D is represented by the input-output
relation
y[n] = Hx[n] + z[n]
(B.1.11)
where y[n] CK1 denotes the vector received along the K antennas
of D, x[n] CM 1 is the input vector transmitted along the M
antennas of S and z[n] CK1 denotes the noise vector at D, all for
the n-th time index. H CKM is the channel matrix between S
and D with its (i, k)-th element {H}i,k being the channel coefficient
between the k-th antenna of S and the i-th antenna of D. The input
vector x[n] obeys the power constraint

E kx[n]k2 P.
(B.1.12)
159
Choosing each element of x[n] to be i.i.d. Gaussian, assuming that D

knows H perfectly and that there is no channel feedback from D to S,
the rate of communication over the link S D is given by (Paulraj
et al., 2003, Sec. 4.3)

P
R = log det IK +
HHH
(B.1.13)
M No

min{M,K}
X
i P
log 1 +
(B.1.14)
=
M No
i=1
where i is the i-th eigenvalue of the matrix HHH . That is, a
spatial multiplexing gain of min{M, K} is achievable, provided the
min{M, K} eigenvalues of HHH are non-zero. Since spatial multiplexing gain of a MIMO channel with M antennas at S and K antennas at
D is upper-bounded by min{M, K} in the presence of perfect channel
knowledge at S (Paulraj et al., 2003, Sec. 4.4), it follows that full
spatial multiplexing gain is achievable even without channel feedback.
B. Frequency-selective case
An L-tap frequency-selective multi-input multi-output (MIMO) channel with M antennas at S and K antennas at D is represented by the
input-output relation
y[n] =
N
1
X
H[l]x[n l] + z[n]
(B.1.15)
l=0
where y[n], x[n] and z[n] are the same as defined for the frequency-flat
case and H[l], l = 0, . . . , L 1, is the channel matrix between S and
D with its (i, k)-th element {H[l]}i,k being the l-th tap of the channel
impulse response between k-th antenna of S and i-th antenna of D.
160
B.2. MULTIUSER NETWORKS
If a cyclic signal model such as OFDM is employed to convert all the

L-tap frequency-selective channels to N (with N K) frequency-flat
channels, the IO relation (B.1.15) can be rewritten as
y[n] = (H ~ x)[n] + z[n],
n = 0, . . . , N 1
(B.1.16)
where H[n], n = 0, . . . , N 1, is obtained by zero-padding the sequence

H[l], l = 0, . . . , L 1. Defining y[n] , [y1 [n] . . . yK [n]]T , x[n] ,
[x1 [n] . . . xM [n]]T , z[n] , [z1 [n] . . . zM [n]]T , xi (r) , F{xi [n]},
yi (r) , F{yi [n]}, zi (r) , F{zi [n]}, y(r) , [y1 (r) . . . yK (r)]T , x(r) ,
[x1 (r) . . . xM (r)]T and z(r) , [z1 (r) . . . zM (r)]T , the time-domain
IO relation (B.1.16) can be written in frequency-domain as
y(r) = H(r)x(r) + z(r),
r = 0, . . . , N 1
(B.1.17)
where H(r) is the channel matrix between S and D for the r-th tone
and its (i, k)-th element is computed as {H(r)}i,k = F{{H[n]}i,k },
for all i, k. An L-tap frequency-selective MIMO channel is therefore
equivalent to N parallel frequency-flat MIMO channels (represented
by (B.1.17)) and since full spatial multiplexing gain is achievable
in a frequency-flat MIMO channel without any channel feedback, it
follows that channel feedback is not required for achieving full spatial
multiplexing gain in frequency-selective MIMO channels as well.
B .2. MU LTI USER N ETWOR KS

Multiuser networks can be broadly classified into three categories:
Multi-access networks, Broadcast networks and Interference networks.
161
B.2.1. Multi-access networks
We shall consider only single-input multi-output (SIMO) multi-access

networks in this section. A SIMO multi-access network consists of
M non-cooperating single-antenna sources, denoted by S1 , , SM ,
communicating with a single destination D, equipped with K antennas. If all channels in the network are frequency-flat, the network is
represented by the IO relation
y[n] =
M
X
hm xm [n] + z[n]
(B.2.18)
m=1
where y[n] CK1 is the symbol vector received along the K antennas
of D, xm [n] C denotes the symbol transmitted by Sm and z[n]
CK1 is the noise vector along the K antennas of D, all for n-th
time index. The channel coefficient from Sm to the i-th antenna of
T
D is denoted by him so that hm = [h1m hK
m ] is the single-input
multi-output channel vector between Sm and D. The input symbols
xm [n] obey the power constraint

P
E |xm [n]|2
.
M
(B.2.19)
Defining H = [h1 hM ], choosing xm [n] to be i.i.d. Gaussian and

assuming that D knows hm , m, perfectly, the sum of the M rates
of communications over the links Sm D, m = 1, . . . , M, is given by
(Tse and Viswanath, 2005, Sec. 10.1.2)

P
Rsum = log det 1 +
HHH
(B.2.20)
M No
162

min{M,K}
X
i=1
i P
log 1 +
M No

(B.2.21)
where i is the i-th eigenvalue of the matrix HHH . That is, a

spatial multiplexing gain of min{M, K} is achievable, provided the
min{M, K} eigenvalues of HHH are non-zero. By allowing cooperation among the sources and noting that spatial multiplexing gain
of the resulting MIMO channel is upper-bounded by min{M, K},
it follows that spatial multiplexing gain of this SIMO multi-access
network is also upper-bounded by min{M, K} (recall that allowing
cooperation cannot decrease capacity). Since we did not assume any
channel feedback from D to Sm , m, to arrive at the rate expression
(B.2.21), it follows that full spatial multiplexing gain is achievable in
the network without any channel feedback.
If all channels in the network are frequency-selective with L-taps,
a cyclic signal model such as OFDM can be employed to convert all
the channels to parallel frequency-flat channels and it can be shown
(following a similar argument as for single-user MIMO channels) that
full spatial multiplexing gain is achievable without channel feedback.
B.2.2. Broadcast networks

A broadcast network consists of a single source S communicating
with K destinations, denoted by D1 , , DK . We will restrict our
analysis in this section to multi-input single-output (MISO) broadcast
networks. A MISO broadcast network consists of a source S, equipped
with M antennas, communicating with K single-antenna destinations,
denoted by D1 , , DK . It will turn out that in contrast to single-user
systems and multi-access networks where channel feedback does not
play much role in spatial multiplexing, achievability of full spatial
163
multiplexing gain in broadcast networks depends critically on channel

feedback from Dk , k, to S.
A. Frequency-flat case
A frequency-flat broadcast network is represented by the IO relations
yk [n] = hTk x[n] + zk [n],
k = 1, . . . , K
(B.2.22)
where yk [n] C is the symbol received at Dk , x[n] CM 1 denotes

the input vector along the M antennas of S, zk [n] is CN (o, No ) noise
T
m
at Dk , all for n-th time index; and hk = [h1k hM
k ] with hk being
the channel coefficient between the m-th antenna of S and Dk . The
input vector x[n] obeys the power constraint

E kx[n]k2 P.
(B.2.23)
In a frequency-flat MISO broadcast network, if each Dk can feedback
the channel vector hk to S perfectly (this requires infinite-capacity
feedback links), then full spatial multiplexing gain of min{M, K} is
known to be achievable in the sum rate (Tse and Viswanath, 2005,
Sec. 10.3.1). On the other hand, if there is no channel feedback
and all MISO channels hk are isotropic fading, then multiplexing
gain is upper-bounded, in the limit of large P , by one (see Jafar
and Goldsmith (2005)). It is therefore evident that achievability of
full spatial multiplexing gain depends critically on channel feedback
in the network. In fact, Lapidoth et al. (2005) showed that for the
M = K = 2 case, even if there exist finite-capacity feedback links from
each Dk to S, full spatial multiplexing gain is not achievable in the
limit of large P . Moreover, Lapidoth and Shamai (2006) conjectured
that even for general values of M and K, full spatial multiplexing
gain is not achievable with finite-capacity feedback links. For finite
164
values of M and K, the conjecture by Lapidoth and Shamai (2006)

implies, implicitly, that for achievability of full spatial multiplexing
gain, capacity of feedback links should scale with the sum transmit
power P . Jindal (2006) and Yoo et al. (2007) formalized this idea.
Specifically, Jindal (2006) showed that for K = M , full spatial
multiplexing gain of M is achievable in the sum rate, provided that
all channels are frequency-flat and that the feedback rate rf from
each Dk to S scales with sum (across all source antennas) transmit
power P according to
rf = (M 1) log P.
(B.2.24)
For the case K M , Yoo et al. (2007) showed that full spatial multiplexing gain is achievable, provided that all channels are frequency-flat
and that the feedback rate rf from each Dk to S scales with sum
(across all source antennas) transmit power P and number of destinations K according to
rf = (M 1) log P log K.
(B.2.25)
The results by Yoo et al. (2007) imply that presence of a large number
of destinations can reduce the feedback load required for achievability
of full spatial multiplexing gain.
B. Frequency-selective case
A frequency-selective broadcast network is represented by the IO
relations
yk [n] =
L1
X
hTk [l]x[n l] + zk [n],
k = 1, . . . , K
(B.2.26)
l=0
165
where yk [n], x[n] and zk [n] are the same as defined for the frequencyT
m
flat case and hk [l] = [h1k [l] hM
k [l]] with hk [l] being the l-th tap of
the channel impulse response between the m-th antenna of S and Dk .
If a cyclic signal model such as OFDM is employed to convert all the
L-tap frequency-selective channels to N (with N K) frequency-flat
channels, the time-domain IO relations (B.2.26) can be rewritten as
yk (r) = hTk (r)x(r) + zk (r),
k = 1, . . . , K
(B.2.27)
where yk (r) C is the symbol received at Dk , x(r) CM 1 denotes

T
the input vector along the M antennas of S, hk (r) = [h1k (r) hM
k (r)]
m
with hk (r) being the channel coefficient between the m-th antenna
of S and Dk and zk (r) is CN (o, No ) noise at Dk , all corresponding to
the r-th tone. A frequency-selective (with L taps) broadcast network
can therefore be considered equivalent to N parallel frequency-flat
broadcast networks.
Consequently, the results by Jindal (2006) imply that for K =
M and all channels frequency-selective (with L taps), full spatial
multiplexing gain of M is achievable in the sum rate, provided that
the feedback rate rf from each Dk to S scales with sum (across all
source antennas) transmit power P according to
rf = N (M 1) log P.
(B.2.28)
Similarly, the results by Yoo et al. (2007) imply that for K M and
all channels frequency-selective (with L taps), full spatial multiplexing
gain of M is achievable in the sum rate, provided that the feedback
rate rf from each Dk to S scales with sum (across all source antennas)
transmit power P according to
rf = N ((M 1) log P log K) .
166
(B.2.29)
The feedback scaling laws in (B.2.28) and (B.2.29) are however unlikely to be fundamental because of their dependence on the number
of tones N . In fact, it has been proved in this thesis (in Chapter 5)
that the results by Jindal (2006) and Yoo et al. (2007) can be generalized to the frequency-selective case with feedback rates scaling as
(M L 1) log P and (M L 1) log P log K, respectively.
B.2.3. Interference networks

A general interference network consists of M non-cooperating sources
S1 , , SM , communicating, concurrently and on the same frequency
band, with M non-cooperating destinations D1 , , DM . The capacity
region (and even the sum capacity) of such networks is not known
in general. However, full spatial multiplexing gain is known to be
achievable in certain important cases of interest. We have considered
three such cases in this thesis.
A. Large relay-based interference networks

Morgenshtern and B
olcskei (2007) considered a large relay-based
frequency-flat single-antenna interference network where M groups
of relays (supporting nodes), denoted by G1 , , GM , are dedicated
to facilitate spatial multiplexing between the M source destination
pairs {S1 , D1 }, , {SM , DM }. Each group Gi consists of K noncooperating relays, denoted by Si1 , , SiK . Each Si , each Sim and
each Di is equipped with a single antenna and transmission from Si
to Di occurs over two hops. During the first hop, each Si transmits its
message to its dedicated group of relays Si1 , , SiK , and each Di stays
silent. During the second hop, each Sim forwards, after appropriate
processing, its received symbol to Di with Si being silent. The IO
167
relations for the first hop are given by

M
X
yik [n] =
k
gi,m
xm [n] + zik [n],
k = 1, . . . , K, i = 1, . . . , M
m=1
(B.2.30)
where yik [n] C is the symbol received at Sik , xm [n] denotes the input
symbol for Sm and zik [n] is CN (0, No ) noise at Sik , all for n-th time
k
index. In addition, gi,m
stands for the channel coefficient between the
single-antenna source Sm and the single-antenna relay Sik and the
input symbols obey the power constraints

P
E |xm [n]|2
,
M
m = 1, . . . , M.
(B.2.31)
The IO relations for the second hop are given by

yi [n] =
K
M X
X
hki,m xkm [n] + zi [n],
i = 1, . . . , M
(B.2.32)
m=1 k=1
where yi [n] C is the symbol received at Di , xkm [m] denotes the input
k
symbol for Sm
and zi [n] is CN (0, No ) noise at Di , all for n-th time
index. In addition, hki,m stands for the channel coefficient between the
k
single-antenna relay Sm
and the single-antenna destination Di and
the input symbols obey the power constraints

P
E |xkm [n]|2 ,
N
k = 1, . . . , K, m = 1, . . . , M.
(B.2.33)
k
Under the assumption of perfect knowledge of both gi,i
and hki,i
k
at each Si , Morgenshtern and B
olcskei (2007) showed that for M
finite and K , full spatial multiplexing gain of M is achievable
168

k
in the sum rate of the network. While perfect knowledge of gi,i
at
k
k
Si is a reasonable assumption (recall that Si is the destination for
the link Si Sik ), perfect knowledge of hki,i at Sik requires infinitecapacity feedback links from Di to Sik , k. It is therefore imperative to
develop a feedback and transmission scheme that achieves full spatial
multiplexing gain in the second hop transmission with limited-capacity
feedback links from Di to Sik , k. In fact, assuming the first hop
transmission to be error-free, we have proposed in Chapter 2 a limitedfeedback based iterative distributed beamforming algorithm for the
second hop transmission and have shown that spatial multiplexing
based on the beamforming weights produced by the algorithm results
in full spatial multiplexing gain of M .
B. Single-antenna interference networks

Cadambe and Jafar (2008) considered a frequency-flat interference
network where each Si and each Di is equipped with a single antenna
and the IO relations are given by
yi [n] =
M
X
hi,k [n]xk [n] + zi [n],
i = 1, . . . , M
(B.2.34)
k=1
where yi [n] denotes the symbol received at Di , xk [n] is the input

symbol for Sk , hi,k [n] denotes the channel coefficient between the
single-antenna source Sk and the single-antenna destination Di , and
zi [n] is CN (0, No ) noise at Di ; all for n-th time index.
Motivated by a proof from Host-Madsen and Nosratinia (2005),
Cadambe and Jafar (2008) showed that spatial multiplexing gain of
such a network is upper-bounded by M/2. Cadambe and Jafar (2008)
furthermore showed that the full spatial multiplexing gain of M/2 is
achievable with interference alignment, provided that i) each source
169
and each destination knows all the channels in the network and that
ii) all the channels are time-selective. Satisfying these two conditions
however require not only infinite capacity feedback links, but also
non-causal feedback. Although Grokop et al. (2009) generalized the
results of Cadambe and Jafar (2008) to time-flat (but frequencyselective) single-antenna interference networks and therefore obviated
the need for non-causal feedback, results of Grokop et al. (2009) still
depend critically on availability of infinite-capacity feedback links. The
achievability of full spatial multiplexing gain with limited-capacity
feedback links is not known in the literature so far. Nevertheless,
we have shown in this thesis (in Chapter 3) that naive interference
alignment, along with a vector quantization scheme proposed for
single-user beamforming by Mukkavilli et al. (2003) and Love et al.
(2003), achieves full spatial multiplexing gain of M/2, provided that
each destination can broadcast at least M (L 1) log P bits in an
error-free fashion to all the sources and destinations in the network.
C. MISO interference networks
MISO interference networks are defined as the interference networks
where each Si is equipped with multiple antennas and each Di is
equipped with a single antenna. Denoting the number of antennas at
Si , i, by Mi , the IO relations for frequency-selective MISO interference networks are given by
yi [n] =
M L1
X
X
hTi,k [l]xk [n l] + zi [n],
i = 1, . . . , M
(B.2.35)
k=1 l=0
where yi [n] C denotes the symbol received at Di , xi [n] CMi 1 is

the input vector along the Mi antennas of Si , and zi [n] is CN (0, No )
noise at Di ; all for the n-th time index. In addition, hi,k [l] = [h1i,k [l]
170

T
m
k
hM
i,k [l]] with hi,k [l] C being the l-th tap of the channel impulse
response between the m-th antenna of Sk and Di . Finally, the input
vectors obey the power constraints

P
E kxi [n]k2
.
M
(B.2.36)
While such networks have been studied under the assumption of

all Si being cooperating Karakayali et al. (2006), feasibility of spatial
multiplexing in such networks has not been investigated for the case
where Si , i, are not cooperating and where each Si obtains channel
knowledge through limited capacity feedback links. We have shown in
this thesis (in Chapter 4) that full spatial multiplexing gain is indeed
achievable in such networks, provided that Mi M, i, and that each
Di can feedback at a rate of (M 1)(M L 1) log P to Si , i 6= k.
171
APPEN DIX C
Notation
C .1. MISC ELLAN EOUS
SD
GD
x[n]
x(r)
|A|
R
C
RN M
CN M
Z+
log(x)
ln(x)
arccos(a)
link between source S and destination D

link between group of sources G and destination D
value of a time-domain sequence x at discrete-time
index n
value of a frequency-domain sequence x at
discrete-frequency index r
cardinality of the set A
set of real numbers
set of complex numbers
set of real matrices with N rows and M columns
set of complex matrices with N rows and M
columns
set of all positive integers
logarithm of x to the base 2
natural logarithm of x
inverse cosine of a
173
C . NOTATION
C .1.
MISC ELLAN EOUS ( CONTI N U ED )
R{a}
I{a}
|x|
j
ab
ab
bac
AB
(A ~ B)[n]
Cyclics {a}
AB
AB
real part of a complex-valued quantity

imaginary part of a complex-valued quantity
absolute value of the scalar x
1
a is much greater than b
a is much less than b
greatest integer that is smaller than or equal to the
real number a
the sets A and B are equal
PN 1
computed as s=0 A[n s]B[s], is circular
convolution between the N -point matrix sequences
A[n] CP Q and B[n] CQR , n = 0, . . . , N 1,
where A[m] = A[N m], for m = 1, . . . , N 1
vector obtained by cyclically shifting the elements
of the vector a downwards by s positions
set of column vectors of A CP Q is a subset of
the set of column vectors of B CP R
union of sets A and B
C .2. LI N EAR ALGEB RA

a, A, b, B, . . .
a, b, . . .
A, B, . . .
aT , AT
a , a , A
174
scalars
vectors
matrices
transpose of the vector a and the matrix A
element-wise complex conjugate of the scalar a, the
vector a, and the matrix A
C .2. LINEAR ALGEBRA
C .2.
LI N EAR ALGEB RA ( CONTI N U ED )
aH , AH
Hermitian transpose of the vector a and the

matrix A
A1
inverse of the square matrix A
IN
N N identity matrix
1M
all-ones column vector of size M
0M
all-zeros column vector of size M
diag(a1 , . . . , aN ) diagonal matrix with diagonal elements a1 , . . . , aN
|x|
absolute value of the complex scalar x
kak
Euclidean norm of the complex vector a
det(A)
determinant of the square matrix A
adj(A)
adjoint of the square matrix A
{A}i,k
element in the i-th row and j-th column of the
matrix A
Ai,k
matrix obtained by removing the i-th row and the
k-th column of A
Cof[{A}i,k ]
cofactor of {A}i,k , computed as (1)i+k det(Ai,k )
AB
Hadamard product of the matrices A and B
H
a b
inner product of the column vectors a and b
PN 1
n
F{a[n]}
, (1/ N ) n=0 a[n]ej2r N , is the discrete
Fourier transform (DFT) of the N -point sequence
a[n], n = 0, . . . , N 1
PN 1
j2rn
F 1 {a(r)}
, (1/ N ) r=0 a(r)e N , is the inverse DFT of
the frequency-domain sequence
a(r), r = 0, . . . , N 1
F
N N unitary DFT matrix with {i, k}-th element
j2(i1)(k1)
N
{F}i,k = (1/ N )e
, i = 1, . . . , N, k =
1, . . . , N
175
C . NOTATION
C .3. PROBAB I LITY AN D STATISTICS

P ()
P(1 |2 )
A, B, . . .
a, b, . . .
A, B, . . .
N (0, 2 )
CN (0, 2 )
fX ()
E[X]
VAR[X]
XY
FA (a|b)
fA (a|b)
w.p.1
Q(v)
probability of the event

conditional probability of the event 1 given the
event 2
random variables
random vectors
random matrices
real normal distributions with mean zero and
variance 2
circularly symmetric complex normal distributions
with mean zero and variance 2
probability density function (p.d.f.) of the random
variable X
expected value the random variable X
variance of the random variable X
equivalence in distribution of random variables X
and Y
conditional CDF of A, given the realization B = b
conditional PDF of A, given the realization B = b
with probability 1
R
u2
right tail of N (0, 1), computed as v 12 e 2 du
C . 4 . I N FORMATION TH EORY
I(x, y)
h(x)
176
mutual information between the random vectors x

and y
differential entropy of the continuous random
variable x
177
D. ACRONYMS
APPEN DIX D
Acronyms
AWGN
CDF
CSI
DFT
IDFT
IEEE
IA
IO
LHS
MIMO
MISO
MMSE
OFDM
PDF
RHS
SIN
SIMO
SISO
SNR
SINR
178
additive white Gaussian noise

cumulative distribution function
channel state information
discrete Fourier transform
inverse discrete Fourier transform
Institute of Electrical and Electronics Engineers
interference alignment
input-output
left-hand side
multi-input multi-output
multi-input single-output
minimum mean-squared error
orthogonal frequency division multiplexing
probability density function
right-hand side
spatial interference-nulling
single-input multi-output
single-input single-output
signal-to-noise ratio
signal-to-interference and noise ratio
References
Ahlswede, R. (1971), Multi-way communication channels, in Proc. of

2nd Int. Symp. on Inf. Th., pp. 2352, Thakadsor, Armenian SSR:
Akademiao Kiado, Budapest.
Ahmad, S., Jovicic, A., and Viswanath, P. (2006), Outer bounds to the
capacity region of wireless networks, IEEE Trans. Inf. Th., vol. 52,
no. 6, pp. 27702776.
and Paulraj, A. J. (2006), Capacity
B
olcskei, H., Nabar, R. U., Oyman, O.,
scaling laws in MIMO relay networks, IEEE Trans. Wireless Comm.,
vol. 5, no. 6, pp. 14331444.
Cadambe, V. R. and Jafar, S. A. (2008), Interference alignment and the
degrees of freedom for the K user interference channel, IEEE Trans.
Inf. Theory, vol. Vol. 54, no. 8, pp. 34253441.
Caire, G. and Shamai, S. (1999), On the capacity of some channels with
channel state information, IEEE Trans. Inf. Th., vol. 45, no. 6, pp.
20072019.
Carleial, A. B. (1975), A case where interference does not reduce capacity,
IEEE Trans. Inf. Th., vol. 21, no. 5, pp. 569570.
(1978), Interference channels, IEEE Trans. Inf. Th., vol. 24, no. 1,
pp. 6070.
(1983), Outer bounds on the capacity of interference channels,
Conway, J. H., Hardin, R. H., and Sloane, N. J. A. (1996), Packing lines,
planes, etc.: Packings in Grassmannian spaces, Exper. Math., vol. 5,
no. 2, pp. 139159.
179
REFERENCES
Costa, M. (1985), On the gaussian interference channel, IEEE Trans. Inf.

Th., vol. 31, no. 5, pp. 607615.
Cover, T. M. (1972), Broadcast channels, IEEE Trans. Inf. Th., vol. 18,
no. 1, pp. 214.
Dana, A. F. and Hassibi, B. (2006), On the power efficiency of sensory
and ad hoc wireless networks, IEEE Trans. Inf. Th., vol. 52, no. 7,
pp. 28902914.
Etkin, R., Tse, D., and Wang, H. (2008), Gaussian interference channel
capacity to within one bit, IEEE Trans. Inf. Th., vol. 54, no. 12, pp.
55345564.
Gamal, A. E. and Cover, T. M. (1980), Multiple user information theory,
Proc. IEEE, vol. 68, no. 12, pp. 14661483.
Gastpar, M. and Vetterli, M. (2005), On the capacity of large Gaussian
relay networks, IEEE Trans. Inf. Th., vol. 51, no. 3, pp. 765779.
Gesbert, D., Kountouris, M., Heath Jr., R. W., Chae, C., and Salzer, T.
(2007), Shifting the MIMO paradigm, IEEE Sig. Proc. Magazine,
vol. 24, no. 5, pp. 3646.
Gomadam, K. S., Cadambe, V. R., and Jafar, S. A. (2009), Approaching
the capacity of wireless networks through distributed interference
alignment, Available: http://arxiv.org/pdf/0803.3816v1.
Gou, T. and Jafar, S. A. (2009), Degrees of freedom of the
K user M x N MIMO interference channel, Available:
http://arxiv.org/pdf/0809.0099v1.
Grinstead, C. M. (1997), Introduction to Probability, American Mathematical Society, 2nd ed.
Grokop, L., Tse, D. N. C., and Yates, R. D. (2009), Interference alignment
for line-of-sight channels, Available: http://arxiv.org/abs/0809.3035.
Gupta, P. and Kumar, P. R. (2002), The capacity of wireless networks,
Han, T. S. and Kobayashi, K. (1981), A new achievable rate region for the
interference channel, IEEE Trans. Inf. Th., vol. 27, no. 1, pp. 4960.
Heath Jr., R. W., Love, D. J., Rao, B. D., Lau, V., Gesbert, D., and Andrews, M. (2008), Exploiting limited feedback in tomorrows wireless
communication networks, IEEE J. Sel. Areas Commun., vol. 26, no. 8,
pp. 13371340.
Host-Madsen, A. and Nosratinia, A. (2005), The multiplexing gain of
180
REFERENCES
wireless networks, in IEEE Int. Symp. on Inf. Th., pp. 20652069,

Adelaide, Australia.
Jafar, S. and Goldsmith, A. (2005), Isotropic fading vector broadcast
channels: The scalar upperbound and loss in degrees of freedom,
Jafar, S. A. and Fakhereddin, M. J. (2007), Degrees of freedom for the
MIMO interference channel, IEEE Trans. Inf. Theory, vol. 53, no. 7,
pp. 26372642.
Jindal, N. (2006), MIMO broadcast channels with finite rate feedback,
IEEE Trans. Inf. Theory, vol. 52, no. 11, pp. 50455060.
Jovicic, A., Viswanath, P., and Kulkarni, S. R. (2004), Upper bounds
to transport capacity of wireless networks, IEEE Trans. Inf. Th.,
vol. 50, no. 11, pp. 25552565.
Karakayali, M. K., Foschini, G. J., and Valenzuela, R. A. (2006), Network
coordination for spectrally efficient communications in cellular systems,
IEEE Wireless Comm. Magazine, vol. 13, no. 4, pp. 5661.
Kendall, M. G. (1961), A course in the geometry of n dimensions, in
Griffins Statistical Monographs and Courses, London, U.K.: Charles
Griffin and Co. Ltd.
Kountouris, M., Gesbert, D., and S
alzer, T. (2008), Enhanced multiuser
random beamforming: dealing with the not so large number of users
case, IEEE J. Sel. Areas Commun., vol. 26, no. 8, pp. 15361545.
Kramer, G. (2001), Genie-aided outer bounds on the capacity of interference channels, in IEEE Int. Symp. on Inf. Th., p. 103, Washington,
D.C., USA.
(2004), Outer bounds on the capacity of gaussian interference
channels, IEEE Trans. Inf. Th., vol. 50, no. 3, pp. 581586.
Lapidoth, A. and Shamai, S. (2006), Collapse of Degrees of freedom in
MIMO Broadcast with Finite Precision CSI, in Open Problems
Session, Inf. Th. and its Applications - Inaugural Workshop, UCSD,
San Diego, CA, USA.
Lapidoth, A., Shamai, S., and Wigger, M. (2005), On the capacity of
Fading MIMO Broadcast Channels with Imperfect Transmitter SideInformation, in Proc. 43rd Allerton Conf. on Commun., Contr. and
Comput., pp. 9961005, Monticello, IL, USA.
Leveque, O. and Telatar, E. (2005), Information theoretic upper bounds
181
REFERENCES
on the capacity of large, extended ad-hoc wireless networks, IEEE

Trans. Inf. Th., vol. 51, no. 3, pp. 858865.
Love, D., Heath Jr., R. W., and Strohmer, T. (2003), Grassmannian
beamforming for multiple-input multiple-output wireless systems,
Love, D. J., Heath Jr., R. W., Lau, V. K. N., Gesbert, D., Rao, B. D.,
and Andrews, M. (2008), An overview of limited feedback in wireless
comunication systems, IEEE J. Sel. Areas Commun., vol. 26, no. 8,
pp. 13411365.
Marton, K. (1979), A coding theorem for the discrete memoryless broadcast
channel, IEEE Trans. Inf. Th., vol. 25, no. 3, pp. 306311.
Mckean, H. P. (1969), Stochastic Integrals, Academic Press.
van der Meulen, E. V. (1994), Some reflections on the interference channel,
in R. E. Blahut, D. J. Costello, and T. Mittelholzer (Eds.), Communications and Cryptography: Two Sides of One Tapestry, pp. 409421,
Kluwer.
Morgenshtern, V. I. and B
olcskei, H. (2007), Crystallization in large
wireless networks, IEEE Trans. Inf. Th., vol. 53, no. 10.
Mudumbai, R., Hespanha, J., Madhow, U., and Barriac, G. (2006, submitted), Distributed transmit beamforming using feedback control,
IEEE Trans. Inf. Th.
Mukkavilli, K. K., Sabharwal, A., Erkip, E., and Aazhang, B. (2003), On
beamforming with finite rate feedback in multiple-antenna systems,
ur, A., Leveque, O., and Tse, D. (2006), How does the information
Oyg
capacity of ad hoc networks scale? in Proc. 44th Allerton Conf. on
Commun., Contr. and Comput., pp. 354363, Monticello, IL, USA.
(2007), Hierarchical cooperation achieves linear capacity scaling
in ad hoc networks, in IEEE Int. Conf. on Computer Commun.
INFOCOM, pp. 382390.
Papoulis, A. and Pillai, S. U. (2002), Probability, Random Variables and
Stochastic Processes, McGraw-Hill, 4th ed.
Paulraj, A., Nabar, R., and Gore, D. (2003), Introduction to Space-Time
Wireless Communications, Cambridge University Press.
Sato, H. (1977), Two-user communication channels, IEEE Trans. Inf.
Th., vol. 23, no. 3, pp. 295304.
182
REFERENCES
(1978a), On degraded Gaussian two-user channels, IEEE Trans.

Inf. Th., vol. 24, no. 5, pp. 637640.
(1978b), An outer bound on the capacity region of broadcast
channels, IEEE Trans. Inf. Th., vol. 24, no. 3, pp. 374377.
(1981), The capacity of the Gaussian interference channel under
strong interference, IEEE Trans. Inf. Th., vol. 27, no. 6, pp. 786788.
Shannon, C. E. (1961), Two-way communications channels, in 4th Berkeley Symp. Math. Stats. and Prob., vol. 1, pp. 611644, University of
California Press.
Shwartz, A. and Weiss, A. (1995), Large Deviations For Performance
Analysis, Chapman and Hall, London, UK.
Telatar, E. (1999), Capacity of multi-antenna Gaussian channels, European Trans. Telecomm., ETT, vol. 10, pp. 585596.
Tse, D. and Viswanath, P. (2005), Fundamentals of Wireless Communications, Cambridge University Press.
Tuninetti, D. (2007), On interference channel with generalized feedback,
in IEEE Int. Symp. on Inf. Th., pp. 28612865.
Vishwanath, S., Jindal, N., and Goldsmith, A. (2003), Duality, achievable
rates and sum rate capacity of gaussian mimo broadcast channels,
Viswanath, P. and Tse, D. (2003), Sum capacity of vector gaussian broadcast channel and downlink-uplink duality, IEEE Trans. Inf. Th.,
vol. 49, no. 8, pp. 19121921.
Weingarten, H., Steinberg, Y., and Shamai, S. S. (2006), The capacity region of the Gaussian multiple-input multiple-output broadcast
channel, IEEE Trans. Inf. Th., vol. 52, no. 9, pp. 39363964.
Xie, L. L. and Kumar, P. R. (2004), A network information theory for
wireless communications: Scaling laws and optimal operation, IEEE
Trans. Inf. Th., vol. 50, no. 5, pp. 748767.
Yoo, T. and Goldsmith, A. (2006), On the optimality of multiantenna
broadcast scheduling using zero-forcing beamforming, IEEE J. Sel.
Areas Commun., vol. 24, pp. 528541.
Yoo, T., Jindal, N., and Goldsmith, A. (2007), Multi-antenna downlink
channels with limited feedback and user selection, IEEE J. Sel. Areas
Commun., vol. 25, no. 7, pp. 14781491.
Yu, W. and Cioffi, J. (2001), Trellis precoding for the broadcast channel,
183
REFERENCES
in IEEE Global Commun. Conf., vol. 2, pp. 13441348, San Antonio,

USA.
Zakhour, R. and Gesbert, D. (2007), A two-stage approach to feedback
design in multi-user MIMO channels with limited channel state information, in IEEE Int. Symp. on Pers., Indoor and Mob. Radio
Commun., PIMRC, pp. 15.
184
Curriculum Vitae
Jatin Thukral
born 10 February 1982
Education
05/200504/2009
10/200304/2005
08/199905/2003
Doctoral studies, ETH Zurich, Switzerland

M.Sc. in Electrical Engineering and Information
Technology, ETH Zurich, Switzerland
Bachelor of Technology in Electrical
Engineering, Indian Institute of
Technology (I.I.T) Madras, India
Professional Experience
05/200504/2009
06/200309/2003
06/200208/2002
Research Assistant, Communication Technology

Laboratory, ETH Zurich, Switzerland
Research and Design Engineer, Tejas Networks
Limited, Bangalore, India
Research Assistant, Illinois Institute of
Technology, Chicago, USA
185

Spatial Multiplexing in Multiuser Networks With Limited Feedback - 1 PDF

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Spatial Multiplexing in Multiuser Networks With Limited Feedback - 1 PDF

Загружено:

Авторское право:

Доступные форматы

SPATIAL MULTIPLEXING IN MULTIUSER NETWORKS WITH

DISS . ETH NO.

SPATIAL MULTIPLEXING IN MULTIUSER

to its corresponding sources. Specifically, we provide an iterative

such a network for L = 1, provided that each destination can feedback

eine Ubertragungsart, welche eine der vielversprechendsten Methoden

gelockert wird, einen fehlerfreien Rundfunkr

geben wir ein R

Fall, dass L > 1 und geben ein R

and quantum physics to intelligence and the future of human society.

1. Introduction and Outline

2. Large Interference Networks with 1-bit feedback

3. SISO Interference Networks with Broadcast Feedback

4. MISO Interference Networks

5. MISO Broadcast Networks with Shared Feedback

6. Conclusions and Outlook

A. Bounds and Proofs

B. Wireless Systems and Channel Feedback: A primer

Introduction and Outline

INTRODUCTION AND OUTLINE

spatial multiplexing gains in these two classes of multiuser networks.

I NTER FER ENC E N ETWOR KS

General interference networks, as defined by Carleial (1978), Gamal

interference whatsoever. Carleial (1978) also established an achievable

1.1.1. Large network analysis approach

INTRODUCTION AND OUTLINE

ur et al. (2007)). Most of these scaling results however assume

Chapter 2 a limited-feedback based iterative (multiuser) beamforming

1.1.2. Large transmit power analysis approach

INTRODUCTION AND OUTLINE

SISO interference networks with broadcast feedback

MISO interference networks

spatial multiplexing in such networks has not been investigated for

A broadcast network consists of a single source and K non-cooperating

INTRODUCTION AND OUTLINE

of MISO Gaussian broadcast networks where the source is equipped

channel knowledge through a limited capacity (shared) feedback link,

INTRODUCTION AND OUTLINE

sum rate, our statements are valid on a per-realization basis.

OUTLINE OF THE THESIS

The remainder of this thesis is organized as follows. In Chapter 2, we

Large Interference Networks with 1-bit

LARGE INTERFERENCE NETWORKS WITH

Fig. 2.1: A large interference network where M groups of K single-antenna

gain proportional to K is achievable with this relaxed assumption.

LARGE INTERFERENCE NETWORKS WITH

group of sources and a single destination. Making the simplifying

SYSTEM AN D SIGNAL MODEL

We assume that xi [n] is the common message of the sources in Gi to

SYSTEM AND SIGNAL MODEL

between a training phase during which the K scalar channels between

so that the average power transmitted by each group Gi is limited

LARGE INTERFERENCE NETWORKS WITH

The sources in Gi are assumed to be non-cooperating and therefore

we will show that if in each group a sufficient number of sources

U PPER B OU N D ON SUM CAPAC ITY

An upper bound on the sum capacity of the network can be obtained

UPPER BOUND ON SUM CAPACITY

Assuming that the destination Di knows the effective channel coeffiPK

E[(xi [n])2 ]P/K

is independent of K. Since communicating in presence of interference

LARGE INTERFERENCE NETWORKS WITH

for some positive constant c independent of K. It therefore follows

2.4.1. Convergence in probability to -level

we say that convergence in probability to -level is achieved in the

Theorem 1 For large K and any positive constant o independent

and hence, convergence in probability to [0] = 1 level is guaranteed.

with no independent of K, where [t] 0, for t = 1, . . . , no K.

weight. This naturally results in [t] = 1/K (note that may

Moreover, since [t] [t], t, we have

and from (2.27), convergence in probability to o -level is achieved

2.4.2. Convergence in expectation to -level

Let there be (1 )K and K sources in A[t 1] and A[t 1],

which, using > o , implies

f (x) (1/o )f|hk | (x), it follows that setting