Full-Duplex Communications and Networks

Learn about the key technologies and understand the state of the art in
research for full-duplex communication networks and systems with this
comprehensive and interdisciplinary guide. Incorporating physical, MAC,
network, and application layer perspectives, it explains the fundamental
theories on which full-duplex communications are built, and lays out the
techniques needed for network design, analysis, and optimization.
Techniques covered in detail include self-interference cancellation and
signal processing algorithms, physical layer algorithms, transceiver design,
resource allocation, networking and game theory for full-duplex systems.
Potential applications and networking schemes are discussed, including
full-duplex cognitive radio networks, cooperative networks, and
heterogeneous networks.

The first book to focus exclusively on full-duplex communications, this

is an indispensable reference for both researchers and practitioners
designing the next generation of wireless networks.

Lingyang Song is Professor of Wireless Communications and Signal

Processing in the School of Electrical Engineering and Computer Science at
Peking University, China, and an IEEE Distinguished Lecturer. He is a
coauthor of Wireless Device-to-Device Communications and Networks
(Cambridge, 2015).

Risto Wichman is a Professor in the Department of Signal Processing and

Acoustics at Aalto University, Finland. He was previously a Senior
Research Engineer at the Nokia Research Center, Finland.

Yonghui Li is a Professor in the School of Electrical Engineering at the

University of Sydney, Australia. He is a coauthor of Cooperative
Communications: Hardware, Channel and PHY (Wiley 2010).

www.TechnicalBooksPDF.com
Zhu Han is a Professor in the Department of Electrical and Computer
Engineering at the University of Houston, Texas. He has coauthored several
books on wireless communications, including Game Theory in Wireless and
Communication Networks (Cambridge, 2012) and Resource Allocation for
Wireless Networks (Cambridge, 2008). He is a Fellow and a Distinguished
Lecturer of the IEEE Communications Society.

www.TechnicalBooksPDF.com
Full-Duplex Communications and
Networks
LINGYANG SONG
Peking University

RISTO WICHMAN
Aalto University, Finland

YONGHUI LI
University of Sydney

ZHU HAN
University of Houston

www.TechnicalBooksPDF.com
University Printing House, Cambridge CB2 8BS, United Kingdom

One Liberty Plaza, 20th Floor, New York, NY 10006, USA

477 Williamstown Road, Port Melbourne, VIC 3207, Australia

4843/24, 2nd Floor, Ansari Road, Daryaganj, Delhi – 110002, India

79 Anson Road, #06–04/06, Singapore 079906

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of education, learning
and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781107157569
DOI: 10.1017/9781316662106

© Cambridge University Press 2017

This publication is in copyright. Subject to statutory exception and to the provisions of relevant
collective licensing agreements, no reproduction of any part may take place without the written
permission of Cambridge University Press.

First published 2017

Printed in The United Kingdom by TJ International Ltd. Padstow Cornwall

A catalogue record for this publication is available from the British Library

Library of Congress Cataloging-in-Publication data

Names: Song, Lingyang, author.
Title: Full-duplex communications and networks / Lingyang Song, University of Peking, Risto
Wichman, Aalto University, Finland, Yonghui Li, University of Sydney, Zhu Han, University of
Houston.
Description: New York, NY, USA : Cambridge University Press, [2017] |
Includes bibliographical references and index.

www.TechnicalBooksPDF.com
Identifiers: LCCN 2017000005 | ISBN 9781107157569 (alk. paper)
Subjects: LCSH: Wireless communication systems.
Classification: LCC TK5103.2 .S665 2017 | DDC 384.5–dc23
LC record available at https://lccn.loc.gov/2017000005

ISBN 978-1-107-15756-9 Hardback

Cambridge University Press has no responsibility for the persistence or accuracy of URLs for
external or third-party internet websites referred to in this publication and does not guarantee that any
content on such websites is, or will remain, accurate or appropriate.

www.TechnicalBooksPDF.com
To my wife, Tingting Li
— Lingyang Song To
my family
— Risto Wichman
To my three girls, Na Zou, Essie Li, and Eunice Li
— Yonghui Li
To my family
— Zhu Han

www.TechnicalBooksPDF.com
Contents

Preface

1 Basics of Communication Systems

1.1 Advanced Wireless Communication Technology
1.1.1 OFDM/OFDMA Technology
1.1.2 Multiple Antenna System and Massive MIMO
1.1.3 Mmwave Communication
1.1.4 Cognitive Radios
1.1.5 Relay Networks
1.2 Wireless Heterogeneous Networks
1.2.1 3G Cellular Networks
1.2.2 4G Cellular Network: LTE and WiMAX
1.2.3 5G Cellular Networks and Beyond
1.2.4 Femtocell/Small Cell Networks
1.2.5 Device-to-Device Networks
1.2.6 LTE-Unlicensed
1.2.7 WiFi Networks
1.2.8 Wireless Personal Area Networks
1.2.9 Wireless Ad Hoc Networks
1.2.10 Wireless Sensor Networks
1.3 Full Duplex
1.3.1 Self-Interference Cancellation
1.3.2 Current State of the Art
1.3.3 Key Application Scenarios

2 Signal Processing and Theoretical Limits

www.TechnicalBooksPDF.com
 2.1 System Modeling
2.1.1 Non-Orthogonal Multicarrier Modulation
2.1.2 Time-Domain Modeling
2.1.3 Modeling of Self-Interference Channel
2.2 Modeling of Self-Interference
2.2.1 Dynamic Range
2.2.2 RF Impairments
2.2.3 System Calculations
2.3 Self-Interference Cancellation Mechanisms
2.4 Algorithms for Self-Interference Cancellation
2.4.1 Time-Domain Cancellation
2.4.2 Frequency-Domain Cancellation
2.4.3 Spatial Suppression
2.4.4 Spatial Suppression and Time-Domain Cancellation
2.4.5 Channel Estimation
2.4.6 Adaptive Interference Cancellation
2.5 Key Challenges

3 Full-Duplex System Hardware Implementation

3.1 Full-Duplex Technology Used in Wireless Systems
3.2 Antenna Design
3.2.1 Antenna Geometry
3.2.2 Polarization
3.2.3 Wavetraps
3.3 Analog Cancellation Techniques
3.3.1 Shared Antenna Design
3.3.2 RF Cancellation Techniques
3.4 Advanced Digital Cancellation
3.5 System Architecture
3.5.1 Pure RF and Digital Cancellation
3.5.2 Hybrid RF and Digital Cancellation
3.5.3 Analog Baseband and Digital Cancellation
3.6 Key Challenges

4 Full-Duplex MIMO Communications

4.1 FD MIMO Signal Processing
4.1.1 Mode Switching between Full-Duplex and Half-Duplex

www.TechnicalBooksPDF.com
 4.1.2 Antenna Pairing
4.1.3 FD MIMO Communications
4.2 X-Duplex: Flexible Switching between SISO and FD
4.2.1 Maximum Sum-Rate Criterion
4.2.2 Performance Analysis
4.2.3 Simulation Results
4.3 Summary and Key Challenges

5 Full-Duplex OFDMA Communications

5.1 FD-OFDMA Model
5.1.1 System Model
5.2 Centralized Resource Management: User Pairing, Subcarrier and
Power Allocation
5.2.1 Problem Formulation
5.2.2 Resource Management with Full Channel State
Information (CSI)
5.2.3 Resource Management with Local CSI
5.2.4 Simulation Results
5.3 User Pairing and Resource Management Based on Matching
Theory
5.3.1 Matching Theory Basics
5.3.2 Joint User Pairing, Subcarrier Allocation and Power
Control
5.3.3 Matching Algorithm for FD-OFDMA Resource
Allocation
5.3.4 Simulation Results
5.4 Concluding Remarks

6 Full-Duplex Heterogeneous Networks

6.1 Full-Duplex Small Cell Networks
6.1.1 System Model
6.1.2 Problem Formulation
6.1.3 Numerical Results
6.2 Full-Duplex HetNet and MU-MIMO Switching
6.2.1 System Model
6.2.2 Switching Criterion between Full-Duplex and Half-
Duplex

www.TechnicalBooksPDF.com
 6.2.3 Numerical Analysis
6.3 Summary

7 Full-Duplex Cooperative Networks

7.1 Cooperative Communication Basics
7.2 Full-Duplex Amplify-and-Forward Relay Systems
7.2.1 System Model
7.2.2 Performance Analysis of FD AF Relaying
7.2.3 Relay Location Optimization
7.2.4 Simulation Results
7.3 X-Duplex Relay Systems
7.3.1 System Model of X-Duplex Systems
7.3.2 Performance Analysis of X-Duplex Relay Systems
7.3.3 Average SER Analysis
7.3.4 Average Sum Rate
7.3.5 Simulation Results of X-Duplex Relay Systems
7.4 Full-Duplex Relaying with Joint Relay and Antenna Mode
Selection
7.4.1 System Model
7.4.2 Performance Analysis of RAMS
7.4.3 Simulation Results
7.5 Concluding Remarks

8 Full-Duplex Cognitive Radio Networks

8.1 Cognitive Radio Basics
8.2 Listen-And-Talk Protocol
8.2.1 System Model
8.2.2 Protocol Description
8.2.3 Key Parameter Design
8.2.4 Performance Analysis: Power–Throughput Trade-Off
8.2.5 Numerical Evaluation
8.3 Cooperative Spectrum Sensing
8.3.1 Cooperation Scheme
8.3.2 Analysis of FD Cooperative Spectrum Sensing
8.4 Dynamic Spectrum Access
8.4.1 Infrastructureless Spectrum Access
8.4.2 Infrastructure Spectrum Access

www.TechnicalBooksPDF.com
 8.5 Key Challenges
8.5.1 Signal Processing Techniques
8.5.2 Coexistence of Multiple Systems

9 Full-Duplex Random Access Networks

9.1 Spectrum Access Models
9.1.1 Exclusive-Use Model
9.1.2 Shared-Use Model
9.1.3 Spectrum Commons Model
9.2 Traditional Half-Duplex MAC
9.2.1 Carrier Sensing Multiple Access
9.2.2 Hidden Terminal/Exposed Terminal Problems and
RTS/CTS
9.2.3 IEEE802.11 MAC
9.3 Full-Duplex MAC Protocol Design
9.3.1 System Model
9.3.2 MAC Protocol Design
9.3.3 Performance Analysis
9.3.4 Simulation Results
9.4 Summary

References
Index

www.TechnicalBooksPDF.com
Preface

Overview

With more and more new multimedia-rich services being introduced and
offered to a rapidly growing population of global subscribers, there is an
ever-increasing demand for higher data rate wireless access, making more
efficient use of this precious resource a crucial need. As a consequence,
new wireless technologies such as Long Term Evolution (LTE) and LTE-
Advanced have been introduced. These technologies are capable of
providing high-speed, large-capacity, and guaranteed quality-of-service
(QoS) mobile services. With the technological evolution of cellular
networks, new techniques, such as small cells, have also been developed to
further improve the network capacity by effectively reusing the limited
radio spectrum. However, all existing wireless communication systems
deploy half-duplex (HD) radios which transmit and receive the signals in
two separate/orthogonal channels. They dissipate the precious resources by
employing either time-division or frequency-division duplexing.

Full-duplex (FD) systems, where a node can send and receive signals at
the same time and frequency resources, can offer the potential to double
spectral efficiency; however, for many years it has been considered
impractical. This is because the signal leakage from the local output to
input, referred to as self-interference, may overwhelm the receiver, thus
making it impossible to extract the desired signal. How to effectively
eliminate self-interference has remained a long-standing challenge.
Recently, there has been significant progress in self-interference

www.TechnicalBooksPDF.com
cancellation in FD systems, which presents great potential for realizing FD
communications for the next generation of cellular networks.

This book provides state-of-the-art research on FD communications and

cellular networks covering the physical, MAC, network, and application
layer perspectives. The book also includes fundamental theories based on
which FD communications will be built. In addition to the self-interference
cancellation signal processing algorithms, the book discusses physical layer
algorithms, radio resource allocation and network protocols in the practical
design and implementation of centralized and distributed FD wireless
networks. Main applications such as FD cognitive radio networks, FD
cooperative networks, and FD heterogeneous networks are explored.

The key features of this book are as follows:

A unified view of FD communications and networking;

A comprehensive review of the state-of-the-art research and key
technologies of FD communications networks;
Coverage of a wide range of techniques for design, analysis,
optimization, and application of FD communications networks;
An outline of the key research issues related to FD communications
and networking.

To the best of the authors’ knowledge, this is the first book on FD

communications and networking. This book intends to provide background
knowledge on FD communications, discuss the research challenges, review
the existing literature on FD communications, and present techniques for
analysis, design, optimization, and application of FD communications
systems under given objectives and constraints.

Motivation and Objectives

Mobile data traffic, especially mobile video traffic, has dramatically

increased in recent years with the emergence of smart phones, tablets, and
various new devices and applications. It is therefore crucial to increase

www.TechnicalBooksPDF.com
wireless network capacity to accommodate these bandwidth-consuming
applications and services. FD communication is a promising concept to
improve link efficiency, user experience, and resource utilization in cellular
networks. However, design, analysis, and optimization of FD
communications and networking require multidisciplinary knowledge,
namely, knowledge of signal processing, communication theory, wireless
communications and networking, artificial intelligence (e.g., for learning),
decision theory, optimization, and economic theory. Therefore, a book
containing the basic concepts/theories for addressing the research advances
that enable FD communications for wireless networks, as well as the state-
of-the-art research and development and related information, will be very
useful for researchers and engineers. This is the primary motivation for
writing this book.

There are three main objectives in writing this book:

1. The first is to provide a general introduction to FD-based wireless

communications and networking from physical, MAC, networking layer
requirements, and system design.
2. The second is to introduce the key techniques in enabling FD
communications systems, and to present the related design, algorithms,
analysis, and optimization problems in a comprehensive way.
3. The third is to present the state of the art of FD communications and
networking schemes along with possible applications. This will include
classifications of the different schemes and the technical details in each
scheme.

The above objectives will be achieved as follows:

1. Starting with an introduction to wireless communications (including

radio propagation and channel models), different wireless access
technologies (e.g., cellular wireless, WLAN, WMAN, and WPAN
technologies) will be briefly reviewed. This will include their basic
components, features, and potential applications. Then, advanced

www.TechnicalBooksPDF.com
 wireless communications technologies in the physical layer, such as
cooperative communications, network coding, and cognitive radio, will
be also discussed.
2. The requirements and main characteristics of FD communications will
be discussed. Specifically, we will briefly describe the history,
requirements, and physical-layer techniques, which enable interference
cancellation, resource allocation, and possible applications.
3. Different self-interference cancellation and signal-processing techniques
for FD communications will be discussed and analyzed. In this context,
various cancellation methods, including antenna design, analog and
digital methods, will be discussed.
4. Other physical-layer techniques are also critical for FD
communications. To achieve FD transmission, the state-of-the art
physical-layer approaches, such as channel estimation, synchronization,
etc., should be adopted, and theoretical analysis on the performance
limits will be investigated. FD system design issues will be explained in
detail.
5. We will describe resource management techniques, which are required
by FD communications systems for distributed spectrum sharing.
Specifically, we will discuss methods for realizing efficient resource
allocation, such as time, frequency, space, and device allocation, which
can be developed based on optimization and game-theoretic models.
6. Game theory techniques will be discussed in the context of optimizing
radio resources. Game theory is an attractive tool for modelling the
spectrum-sharing problem in an FD communications network. The
basics of different game theoretic models, namely, non-cooperative
game, auction game, matching game and coalition game models, will be
explained and modeled for FD communications and networks.
7. Then we discuss major research issues and possible extensions of FD
communication and networks. Starting with FD MIMO networks,
heterogeneous networks and cooperative networks, we then elaborate
FD cognitive radio networks from physical to network layers: for
example, FD-enabled WiFi, which realizes better co-existence of
multiple distributed systems in the ISM bands.

www.TechnicalBooksPDF.com
Market and Readership

The reasons why this book is timely are based on the following
observations:

Boosts of mobile data traffic and scarcity of spectrum: Mobile data

traffic has dramatically increased in recent years with the emergence
of smart phones, tablets, and various new applications. It is hence
crucial to increase network capacity to accommodate bandwidth-
consuming applications and services. However, the frequency
spectrum is the scarcest radio resource, and its efficient usage is
crucial in the next generation of wireless communications systems.
FD communication is a promising concept to improve link efficiency,
user experience and resource utilization in wireless networks. This
new type of communication allows an additional communication
dimension, and thus significantly improves network capacity and
resource utilization.
Interdisciplinary principles: Designing FD communications networks
requires knowledge of multiple science and engineering disciplines
to achieve the design objectives. These disciplines include traditional
signal processing, wireless communications and networking,
optimization and game theory, and network economics. Therefore, a
unified treatment of this subject area is required.
Emergence of FD-based wireless applications and services: Emerging
wireless applications, a few of which are described below, can take
advantage of FD communications:
1. Future-generation cellular cooperative networks: Next-generation
wireless networking is likely to include the features of device
cooperation to enhance users’ quality of experience (QoE). FD
communication will facilitate provisioning of these future-generation
wireless cooperative networks.
2. FD cognitive radio networks: In traditional cognitive radio networks,
secondary users (SUs) typically access the spectrum of primary users
by a two-stage “listen-before-talk” protocol, i.e., SUs sense the
spectrum holes in the first stage before transmitting in the second.

www.TechnicalBooksPDF.com
 With a FD radio, SUs can simultaneously sense and access the vacant
spectrum. As a result, research topics such as spectrum-sensing
algorithms, dynamic spectrum access, communication protocol design,
etc., need to be redeveloped. Understanding the key concepts and
design techniques for FD communications networks is fundamental for
researchers, communications engineers, and application developers to
implement the above FD-based applications.
Integration of FD concepts in traditional wireless systems: FD
communications techniques can be integrated into traditional wireless
communications systems to achieve better flexibility of radio
resource usage so that system performance can be improved: for
example, load balancing/dynamic channel selection in traditional
cellular wireless systems and WLANs, distributed subcarrier
allocation in orthogonal frequency division multiplexing (OFDM)
systems, etc.

The primary audience for this book consists of:

Researchers and engineers interested in studying the new paradigm

of FD communications overlaid with traditional cellular
communications.
Researchers and engineers interested in the state-of-the-art research
on FD communications.
Engineers in the field of LTE/LTE-Advanced for anticipating the
emerging standards of FD communications.
Graduate students interested in obtaining comprehensive information
on design, evaluation, and applications of FD communications
networks.

We sincerely hope this book can bring audiences new perspectives on

the design of future communication and network systems, based on the new
paradigm of full-duplex communications.

Finally, we would like to thank our collaborators for contributing to our

full-duplex research: Yun Liao, Tianyu Wang, Kaigui Bian, Radwa Sultan,
Karim G. Seddik, Hongyu Cui, Bingli Jiao, Mingxin Zhou, Boya Di, Kun

www.TechnicalBooksPDF.com
Yang, Yunxiang Jiang, Francis Lau, He Chen, Siavash Bayat, Taneli
Riihonen, Mikko Vehkaperä, Mikko Valkama, Dani Korpi, Lauri Anttila,
Katsuyuki Haneda, Clemens Icheln, Emilio Antonio Rodríguez, and Stefan
Werner.

www.TechnicalBooksPDF.com
1 Basics of Communication Systems

A wireless network refers to a telecommunications network in which

interconnections between nodes are implemented without the use of wires.
Wireless networks have experienced unprecedented growth over the past
few decades, and they are expected to continue to evolve in the future.
Seamless mobility and coverage ensure that various types of wireless
connections can be made anytime, anywhere. In this chapter, we first
introduce some advanced types of wireless communication techniques.
Next, we study the heterogeneous wireless networks, and finally we provide
the reader with some necessary background on the state of the art in full-
duplex development.

1.1Advanced Wireless Communication

Technology

In this section, we discuss some of the current state of the art in wireless
technologies related to D2D communications.

1.1.1 OFDM/OFDMA Technology

Orthogonal Frequency-Division Multiplexing (OFDM) is a technique for

transmitting multiple digital signals simultaneously over a large number of
orthogonal subcarriers. Based on the fast Fourier transform algorithm to
generate and detect the signal, data transmission can be performed over a

www.TechnicalBooksPDF.com
large number of carriers that are spaced apart at precise frequencies. The
frequencies (or tones) are orthogonal to each other. Therefore, the spacing
between the subcarriers can be reduced and hence high spectral efficiency
can be achieved. OFDM transmission is also resilient to interference and
multipath distortion which causes Inter-Symbol Interference (ISI).

OFDM transmitter and receiver block diagrams are shown in Fig. 1.1
and Fig. 1.2, respectively. is a serial stream of binary digits to transmit.
After serial-to-parallel conversion, the data is split into N streams. Each
stream is then coded to with possibly different modulation
methods (such as PSK and QAM) depending on the subchannel condition.
An inverse FFT is computed on each set of symbols, giving a set of
complex time-domain samples. These samples are then quadrature-mixed to
passband in the standard way: the real and imaginary components are first
converted to the analog domain using digital-to-analog Converters (DACs);
the analog signals are then used to modulate cosine and sine waves,
respectively, at the carrier frequency, . These signals are then summed to
yield the transmission signal, . The receiver picks up the signal ,
which is transmitted through radio channels and contaminated by noise.
Then is quadrature-mixed down to baseband using cosine and sine
waves at the carrier frequency. The baseband signals are then sampled and
digitized using analog-to-digital converters (ADCs), and a forward FFT is
used to convert back to the frequency domain. This returns N parallel
streams, each of which is converted to a binary stream using an appropriate
symbol detector. These streams are then re-combined into a serial stream,
, which is an estimate of the original binary stream at the transmitter.

Figure 1.1 Illustration of OFDM transmitter

www.TechnicalBooksPDF.com
 Figure 1.2 Illustration of OFDM receiver

For a mathematical description, the low-pass equivalent OFDM signal

is expressed as

(1.1)

where is the modulated data symbol for the kth data stream and T is the
OFDM symbol time.

To avoid inter-symbol interference in multipath fading channels, a guard

interval of length is inserted prior to the OFDM block. During this
interval, a cyclic prefix is transmitted such that the signal in the interval
equals the signal in the interval . Cyclic prefix
is often used in conjunction with modulation in order to retain sinusoids’
properties in multipath channels. It is well known that sinusoidal signals are
eigenfunctions of linear and time-invariant systems. Therefore, if the
channel is assumed to be linear and time-invariant, then a sinusoid of
infinite duration would be an eigenfunction. However, in practice, this
cannot be achieved, as real signals are always time-limited. So, to mimic
the infinite behavior, prefixing the end of the symbol at the beginning
makes the linear convolution of the channel appear as though it were a
circular convolution, and thus preserves this property in the part of the
symbol after the cyclic prefix.

www.TechnicalBooksPDF.com
 Orthogonal Frequency Division Multiple Access (OFDMA) is a
multiple-access scheme implementation based on OFDM. In OFDMA,
different users are allocated different subcarriers, and hence multiple users
can transmit their data simultaneously. QoS can be achieved in OFDMA by
allocating different numbers of subcarriers to the users with different QoS
requirements. In OFDMA, different users can occupy different time-
frequency slots so as to fully utilize the diversity. OFDM can be combined
with the CDMA scheme (i.e., Multi-Carrier Code Division Multiple Access
(MC-CDMA) or OFDM-CDMA). In this case, different codes are assigned
to different users for concurrent transmissions.

Dynamic spectral management for OFDMA can be categorized into two

groups based on the network control architecture:

1. Centralized control – A Spectral Management Center (SMC) monitors

and controls the transmit spectra of all the users in the system. There
must be a considerable amount of coordination and communication
between the users and the SMC. This greatly increases the control
signaling overhead but leads to better performance when compared to
the distributed control. Centralized control can be further categorized as
follows [1, 2]:
Level 1 – The data rate and transmission power of the user is
reported to a central controller and corresponding control signals
are generated to control the rate and transmission power.
Level 2 – The noise spectra and the received signal spectra are
monitored and the transmission power is controlled by the central
controller.
Level 3 – Complete coordination of real-time control of
transmission power while monitoring the signals and noise
spectra.
2. Distributed control – The users have the capability of sensing the
channel conditions and adjusting their transmission spectra accordingly.
This may be efficient from the perspective of signaling and
coordination, but it has the drawback of converging to a suboptimal
point. It is also referred to as Level 0 [2], [1], where there is no Dynamic

www.TechnicalBooksPDF.com
 Spectrum Management (DSM) involved and the control is fully
distributed.

The control exhibited by the SMC over the users determines the
computational complexity of the overall system. A centralized control is
efficient but consumes a lot of bandwidth in control messaging between the
users and the base station. On the other hand, distributed algorithms
increase the complexity of the receiver of the user. The resource allocation
of OFDMA can be broadly categorized into three areas:

Sub-carrier assignment – The sub-carriers with the best channel gains

as seen by the user are allocated to the particular user.
Rate allocation – The data rate is allocated depending on the user
application requirements.
Power control – Optimal transmission power is to be allocated to the
user in order to meet the rate requirements while not interfering with
the other users.

OFDM and OFDMA are used in the emerging standards, including

IEEE 802.11a/WiFi, IEEE 802.15/WiPAN, IEEE 802.16/WiMAX, IEEE
802.20/MobileFi, IEEE 802.22/WiRAN, digital audio broadcasting (DAB),
terrestrial broadcasting of digital television (DVB-T, DVB-H), Flash-
OFDM (Fast Low-latency Access with Seamless Handoff Orthogonal
Frequency Division Multiplexing), SDARS for satellite radio, G.DMT (ITU
G.992.1) for ADSL, and ITU-T G.hn for power line communication.

1.1.2 Multiple Antenna System and Massive MIMO

For spatial diversity, transceivers employ antenna arrays and adjust their
beam patterns such that they have good channel gain toward the desired
directions, while the aggregate interference power is minimized at their
output. Antenna array processing techniques such as beamforming, MIMO
technology and space-time coding can be applied to receive and transmit
multiple signals that are separated in space. Hence, multiple co-channel
users can be supported in each cell to increase the capacity by exploring the

www.TechnicalBooksPDF.com
spatial diversity. In this subsection, we briefly discuss various multiple
antenna technologies.

Beamforming Technique
Beamforming is a technique in signal processing used for directional data
transmission and reception. Using beamforming, the signal is
received/transmitted in a particular direction to improve the
receive/transmit gain. When transmitting, a beamformer controls the phase
and amplitude of the signal to create a pattern of constructive and
destructive interference in the wavefront. When receiving, these signals are
combined in such a way that the expected pattern of radiation is
preferentially observed. Different weights can be assigned to the different
signals so that the decoded information has the smallest error probability. In
a conventional beamforming system, these weights are fixed and can be
obtained according to the location of the receive antenna and the direction
of the signal. Alternatively, the weights can be adaptively adjusted by
considering the characteristics of the received signal to mitigate the
interference from unwanted sources. In Fig. 1.3, we show an illustration of
beamforming technology.

www.TechnicalBooksPDF.com
 Figure 1.3 Illustration of beamforming

MIMO Technology and Space–Time Coding

To enhance the performance of wireless transmission, Multiple-Input
Multiple-Output (MIMO) or multiple antennas can be used to transmit and
receive the radio signals. Data transmitted from multiple antennas will
experience different multipath fading, and at the receiver these different
multipath signals are received by multiple antennas. By using advanced
signal processing techniques, multipath signals at the receiver can be
combined to reconstruct the original data. MIMO systems take advantage of
this spatial diversity to achieve higher data rates or lower bit error rates.
There are two basic types of MIMO systems, namely, the space–time
coding MIMO system (for diversity maximization) and spatial multiplexing
MIMO system (for data rate maximization).

In a space–time coding MIMO system, a single data stream is

redundantly transmitted over multiple antennas by using suitably designed
transmit signals. At the receiver, multiple copies of the signal are received
and are used to reconstruct the original data. Space–time coding can be
categorized into Space-Time Trellis Coding (STTC), in which trellis code is
transmitted over multiple antennas and multiple time-slots, and Space–Time
Block-Coding (STBC), in which a block of data is transmitted over multiple
antennas. Both STTC and STBC can achieve diversity gain, which
improves the error performance. While STTC can also achieve coding gain
(i.e., resulting in a lower error rate), STBC can be implemented with less
complexity.

Instead of transmitting the same data over multiple antennas, a spatial

multiplexing MIMO system transmits different data streams over multiple
antennas. In this case, the number of transmit antennas is equal to or larger
than the number of receive antennas and the data rate can increase by a
factor of the number of transmit antennas [3]. There is a fundamental trade-
off between diversity gain and multiplexing gain. In simple terms, the
diversity gain improves the BER performance, while multiplexing gain will
increase the rate but on the other hand the BER performance might be
reduced. Detailed analyses can be found in the literature [4].

www.TechnicalBooksPDF.com
 One simple example of an STBC is the Alamouti code [5], which was
designed for a two-transmit antenna system and has the coding matrix:

(1.2)

where and are two symbols, denotes complex conjugate, and

components of the matrix are sent on two antennas and in two time slots,
respectively. The coding rate is 1, and using the optimal linear decoding
scheme the BER of this STBC is equivalent to maximal ratio combining
(MRC).

Multi-user MIMO (MU-MIMO) was proposed to support data

transmission from multiple users simultaneously. In this case, data from
different users can be transmitted over different antennas. In MU-MIMO,
MIMO broadcast channels and MIMO multiple access channels are used
for downlink and uplink transmissions, respectively. Alternatively, Space-
Division Multiple Access (SDMA) can exploit the information of users’
locations to adjust the transmission and reception parameters to achieve the
best path gain in the direction of each user. Phased array antenna techniques
are generally used for SDMA. MIMO is an optional feature in the IEEE
802.16/WiMAX standard, and is to be used in the IEEE 802.11n standard.

Massive MIMO
Massive MIMO (also known as Large-Scale Antenna Systems, Very Large
MIMO, Hyper MIMO, Full-Dimension MIMO and ARGOS) makes a clean
break with current practice through the use of a large excess of service
antennas over active terminals and time division duplex operation. Extra
antennas help by focusing energy into ever-smaller regions of space to bring
huge improvements in throughput and radiated energy efficiency. Other
benefits of massive MIMO include the extensive use of inexpensive low-
power components, reduced latency, simplification of the media access
control (MAC) layer, and robustness against intentional jamming. The
anticipated throughput depends on the propagation environment providing
asymptotically orthogonal channels to the terminals, but so far experiments

www.TechnicalBooksPDF.com
have not disclosed any limitations in this regard. While massive MIMO
renders many traditional research problems irrelevant, it uncovers entirely
new problems that urgently need attention: the challenge of making many
low-cost, low-precision components that work effectively together;
acquisition and synchronization of newly-joined terminals; the exploitation
of extra degrees of freedom provided by the excess of service-antennas;
reducing internal power consumption to achieve total energy efficiency
reductions; and finding new deployment scenarios. There are many exciting
recent research discoveries concerning massive MIMO [6–23].

Illustrations of the potential of Massive MIMO are summarized in [24]

and listed as follows: increase capacity and improve radiated energy-
efficiency significantly; build with inexpensive, low-power components;
enable a great latency reduction in the air interface; simplify the multiple-
access layer; and increase robustness regarding both interference and
malicious jamming.

On the other hand, there are also many limiting factors of Massive
MIMO [24], as follows:

1. Channel Reciprocity: hardware chains in the base station and terminal

transceivers may not be reciprocal between the uplink and the downlink.
2. Pilot Contamination: performance is limited when the number of
antennas is increased without bounds, at least in terms of receivers that
rely on pilot-based channel estimation.
3. Radio Propagation and Orthogonality of Channel Responses: Massive
MIMO relies to a large extent on a property of the radio environment
called favorable propagation, which means the propagation channel
responses from the base station to different terminals are sufficiently
different. Favorable propagation is valid in practice when the number of
antennas is much greater than the number of users.

Also in [24], different future research directions are also listed as: fast
and distributed, coherent signal processing; challenge of low-cost hardware;

www.TechnicalBooksPDF.com
hardware impairments; internal power consumption; channel
characterization; cost of reciprocity calibration; pilot contamination; non-
CSI at Tx operation; new deployment scenarios; system studies and relation
to small-cell and HetNet solutions; and prototype development.

1.1.3 Mmwave Communication

Millimeter wave (such as 60 GHz) communication is an emerging

technology for high-speed short-range wireless communications. Recently,
there has been increasing interest in providing broadband communications
in the 60 GHz unlicensed band for Wireless Personal Area Networks
(WPANs). A wide available spectrum, as shown in Fig. 1.4, surrounding the
60 GHz operating frequency has the ability to support high-rate broadband
wireless systems such as voice, data, and full motion real-time video for
home entertainment applications. There are many existing works related to
Mmwave communication [25–41].

www.TechnicalBooksPDF.com
 Figure 1.4 60 GHz Communication Spectrum Worldwide

Mmwave Wireless Propagation

Mmwave wireless propagation has the following new properties, which are
different from those of lower frequency wireless communications:

Higher path loss compared to 5 GHz (e.g., 21 dB worse than 5 GHz,

29 dB worse than 2 GHz);
Oxygen/water absorption and rain loss not an issue for radius 200 m
and mm-wave frequencies between 70 and 100 GHz;
Foliage loss is severe;
Channel variations 30 worse than at 2 GHz;
Transmission through most objects is reduced at Mmwave but
reflection is amplified;
Delay spread with narrow beam antennas is less than 1 ns in LOS
conditions

www.TechnicalBooksPDF.com
 With narrow beam antenna the RMS delay spread in non-LOS
conditions has a mean of 7.39 ns and a maximum of 36.6 ns;
Reflective paths exist (between three and five paths), which can be
used for diversity combining;
Body loss at mmwave bands is quite high, ~20dB, due to the high
absorption of energy in these bands to water.

Mmwave Wireless Techniques

Currently, Mmwave is considered as a major component for 5G (or Beyond
4G, B4G) networks. Several developments have been spotlighted, as
follows:

1. 60 GHz UWB: Due to special propagation characteristics at this

frequency range, such as high absorption loss, there is improved
frequency reuse compared to the lower frequencies. However, multipath
effects still exist for indoor applications and they must be precisely
characterized for diversity transmission purposes. It was shown that
coherent carrier modulation schemes suffer from multipath effects for
indoor wireless communication applications. However, ultra-wide-band
(UWB) signals are much less sensitive to these effects. Hence, to take
the benefits of both 60 GHz and UWB communications, the UWB
signal is upconverted to 60 GHz carrier.
2. IEEE 802.11ad (WiGig): “WiGig” is a published standard that is already
seeing a major push from hardware manufacturers. On 24 July 2012,
Marvell and Wilocity announced a new partnership to bring a new tri-
band WiFi solution to market. Using 60 GHz, the new standard can
achieve a theoretical maximum throughput of up to 7 Gbit/s.
3. WirelessHD: An industry group created the first standard 60 GHz
protocol, WirelessHD, specifically to support high-definition video
streaming. The 1.0 version of the standard completed in 2008 supported
data rates of 4 Gbps, while version 1.1 improved support to a maximum
of 28 Gbps. UltraGig is a specific brand name for WirelessHD standard-
based technology from a company called Silicon Image.

1.1.4 Cognitive Radios

www.TechnicalBooksPDF.com
The Federal Radio Act under FCC allows predetermined users the right to
transmit at a given frequency. Unlicensed users are regarded as “harmful
interference,” and in most cases side bands were implemented to ensure that
interference is not an issue. As technology advanced, higher frequency
bands were sold at auction, bringing considerable revenue to the
government. For example, since the 1994 PCS auction, over $30 billion has
been generated. As the demands for wireless communication become more
and more pervasive, the wireless devices must find a way for the right to
transmit at frequencies in the limited radio band. However, there exist a
large number of frequency bands that have considerable, and sometimes
periodic, dormant time intervals. For example, TV stations often do not
work at night. There exist some spectrum holes at a given time over
different spectrum bands. So there is a dilemma that, on the one hand,
mobile users have no spectrum to transmit, while on the other hand some
spectrums are not fully utilized.

In order to cope with the dilemma, cognitive radio is a paradigm for

wireless communication in which either a network or a wireless node
changes its transmission or reception parameters to communicate efficiently
without interfering with the licensed users. This alteration of parameters is
based on the active monitoring of several factors in the external and internal
radio environment, such as radio frequency spectrums, user behaviors and
network states.

Depending on the set of parameters taken into account in deciding on

transmission and reception changes, and for historical reasons, we can
distinguish certain types of cognitive radio as:

Full Cognitive Radio (“Mitola radio”), in which every possible

parameter observable by a wireless node or network is taken into
account;
Spectrum Sensing Cognitive Radio, in which only the radio
frequency spectrum is considered;
Licensed Band Cognitive Radio, in which cognitive radio is capable
of using bands assigned to licensed users;

www.TechnicalBooksPDF.com
 Unlicensed Band Cognitive Radio, which can only utilize unlicensed
parts of the radio frequency spectrum.

Cognitive radio can be designed as an enhancement layer on top of the

Software Defined Radio (SDR) concept. An SDR system is a radio
communication system which can tune to any frequency band and receive
any modulation across a large frequency spectrum by means of a
programmable hardware which is controlled by software.

An SDR performs significant amounts of signal processing in a general

purpose computer, or a reconfigurable piece of digital electronics. The goal
of this design is to produce a radio that can receive and transmit a new form
of radio protocol just by running new software. The hardware of a software-
defined radio typically consists of a superheterodyne RF front end which
converts RF signals from (and to) analog IF signals, and analog-to-digital
and digital-to-analog converters which are used to convert a digitized IF
signal from and to analog form, respectively.

Software radios have significant utility for the military and cell phone
services, both of which must serve a wide variety of changing radio
protocols in real time. Software-defined radio can currently be used to
implement simple radio modem technologies. In the long run, software-
defined radio is expected by its proponents to become the dominant
technology in radio communications. It is the enabler of the cognitive radio.

The cognitive transmitter and receiver can adapt to 3G, WiFi, and
WPAN networks. By sensing the available spectrum, the cognitive radio
can adapt to the most suitable available communication links. For example,
if the user is at home, then he/she can communicate via Bluetooth; if the
user travels to the airport, WiFi communication can be available; if the user
drives on the highway, the cellular phone system can provide reliable
communication links. Another analogy for cognitive radio is as follows: the
licensed users are legally entitled to use the spectrum, or, like High
Occupancy Vehicles (HOV), can drive on the HOV lanes. However, the
HOV lanes are not always occupied. So other than during rush hours, the

www.TechnicalBooksPDF.com
other vehicles can also drive in the HOV lanes. In this sense, the cognitive
radios are similar to the other non-HOV vehicles.

Two major objectives of cognitive radio are to reliably communicate,

whenever and wherever, and to efficiently utilize the radio spectrum. To
achieve these objectives, three fundamental cognitive tasks [42] for
cognitive radio must be fulfilled, as follows:

1. Radio-Scene sensing analyzes the interferences and detects the

spectrum holes;
2. Spectrum analysis, such as channel-state estimation and predictions of
channel capacity;
3. Transmission power control and dynamic spectrum management.

In Fig. 1.5, we show the cognitive cycle, in which three tasks interact
with each other to handle the outside world so that the best strategy can be
calculated and implemented.

Figure 1.5 Cognitive cycle

www.TechnicalBooksPDF.com
 The IEEE 802.22 working group on Wireless Regional Area Networks
(WRAN) is a group of the IEEE 802 LAN/MAN standards committee. A
standard for WRAN Part 22 (Cognitive Wireless RAN Medium Access
Control (MAC) and Physical Layer (PHY) Specifications) regulates
policies and procedures for operation in the TV bands. The standard focuses
on constructing a consistent, fixed point-to-multipoint WRAN that will
utilize UHF/VHF TV bands between 54 and 862 MHz. Specific TV
channels, as well as the guard bands of these channels, are planned to be
used for communication in IEEE 802.22.

Table 1.1 WRAN coverage and capacity

RF channel bandwidth 6 MHz
Average spectrum efficiency 3 bit/s/Hz
Downlink user capacity 1.5 Mbit/s
Uplink user capacity 384 kbit/s
Over-subscription ratio 50
Number of users per downlink 600
Minimum no. of users 90
Assumed early take-up rate 3 bit/s/Hz
Potential no. of users 1800
No. of user per household 2.5
No. of user per coverage area 4500
WRAN base station power 98.3 W
Coverage 30.7 km
Min population density 1.5/km2

The IEEE, together with the FCC, is pursuing a centralized approach for
available spectrum discovery. Specifically, each Access Point (AP) would
be armed with a GPS receiver which would allow its position to be
reported. This information would be sent back to centralized servers (in the
USA, these would be managed by the FCC), which would respond with the
information about available free TV channels and guard bands in the area of
the AP. Other proposals would allow local spectrum sensing only, where the
AP would decide by itself which channels are available for communication.
A combination of these two approaches is also envisioned. Table 1.1
describes the coverage and capacity of IEEE 802.22 WRAN.

www.TechnicalBooksPDF.com
 Overall, cognitive radio can bring in a variety of benefits: for a
regulator, it can lead to a significant increase in spectrum availability for
new and existing applications. For a license holder, cognitive radio can
reduce the complexity of frequency planning, facilitate the secondary
spectrum market agreements, increase system capacity through access to
more spectrums, and avoid interference. For an equipment manufacturer,
cognitive radio can increase demands for wireless devices. Finally, for a
user, cognitive radio can provide a higher level of capacity per user,
enhance interoperability and bandwidth-on-demand for Public Safety and
Emergency Response operations, and provide ubiquitous mobility with a
single user device across disparate spectrum access environments.

1.1.5 Relay Networks

Cooperative communications has attracted significant recent attention as a

transmission strategy for future wireless networks. Cooperative
communications efficiently takes advantage of the broadcast nature of
wireless networks to allow network nodes to share their messages and
transmit cooperatively as a virtual antenna array, thus providing diversity
that can significantly improve system performance. Cooperative
communications can be applied in a variety of wireless systems and
networks. In the research community, a considerable amount of work has
been done in this area for networks such as cellular, WiFi, ad hoc/sensor
networks, and ultra-wideband (UWB). These ideas are also working their
way into standards; e.g., the IEEE 802.16 (WiMAX) standards body for
future broadband wireless access has established the 802.16j Relay Task
Group to incorporate cooperative relaying mechanisms into this technology.

To illustrate the basic idea of cooperative transmission, a highly

simplified topology with one source node, two relay nodes, and one
destination node is shown at the bottom of Figure 1.6. Cooperative
transmission is conducted in two phases. In phase one, the source
broadcasts a message to the destination and relay nodes. In phase two, relay
nodes send information to the destination (in different time slots or on
different orthogonal channels), and the destination combines messages from
the source and relays. It has been shown that the capacity region of this

www.TechnicalBooksPDF.com
communication channel can be significantly increased by such techniques,
and that the performance gain of cooperative transmission is proportional to
the number of relays in the Rayleigh fading case [43]. Here we denote the
source node as s, the several relays nodes as , and the destination nodes as
d. Now we can describe the process more precisely. In phase one, the
received signals and at the destination d and relay , respectively, can
be expressed as

(1.3)

and

(1.4)

where represents the transmit power to the destination from the source, X
is the unit-energy information symbol transmitted by the source in phase
one, and are the channel gains from s to d and respectively, and
and are samples from independent (discrete-time) additive white
Gaussian noise (AWGN) processes, independent of X. Without loss of
generality, we assume that the noise power is the same for all the links, and
is denoted by .

www.TechnicalBooksPDF.com
 Figure 1.6 Cooperative Communication System Model

A number of different types of cooperative communication protocols

have been developed in the literature. Here we describe several of the most
relevant of these briefly as follows:

1. Direct Transmission:
Without the relay nodes’ help, the signal-to-noise ratio (SNR) that
results from s to d can be expressed as

(1.5)

and the capacity of the direct transmission channel is

(1.6)

where W is the bandwidth used for information transmission.

www.TechnicalBooksPDF.com
2. Amplify-and-Forward (AF) Cooperative Transmission:
In phase two, relay i amplifies and forwards it to the destination
with transmitted power . The received signal at the destination is

(1.7)

where

(1.8)

is the energy-normalized transmitted signal from the source to the

destination in phase one, is the channel gain from relay i to the
destination, and is the received noise in phase two. Substituting (1.4)
into (1.8), we can rewrite (1.7) as

(1.9)

Using (1.9), the relayed SNR at the destination for the source, which is
assisted by relay node i, is given by

(1.10)

Therefore, by (1.6) and (1.10), the channel capacity assuming maximal

ratio combining (MRC) at the destination is given by

(1.11)

www.TechnicalBooksPDF.com
 If multiple relay nodes (say, , which ) are available to help
the source, then we have

(1.12)

Here, the in (1.11) and the in (1.12) are due to the fact that the
relays need extra orthogonal channels for transmission.

3. Decode-and-Forward (DF) Cooperative Transmission:

Here, the relay decodes the source information in phase one and relays it
to the destination in phase two. The destination combines the direct
transmission information with the relayed information. The achievable
rate in this case can be calculated by the following maximization:

(1.13)

where

(1.14)

and

(1.15)
4. Estimate-and-Forward (EF) Cooperative Transmission:
Here, in phase two, the relay sends an estimate of the received signal of
phase one. The destination uses the relay’s information as side

www.TechnicalBooksPDF.com
 information to decode the direct transmission of phase one. From [44]
and [45], the channel capacity resulting from this approach can be
written as

(1.16)

where

(1.17)

5. Coded Cooperation:
This type of cooperative transmission protocols integrate relay
cooperation with channel coding [46]. Instead of exactly repeating the
received information, the relay decodes the partner’s transmission and
transmits additional parity symbols (e.g., incremental redundancy)
according to a certain overall coding scheme. The destination receiver
conducts channel decoding by concatenating the data from the direct
transmission and relay transmission, so that the channel gain can be
obtained.

6. Distributed Space–Time Coded Cooperation:

Space–time coding has been shown to significantly improve the link
performance in multiple-input/multiple-output (MIMO) systems.
Distributed space-time cooperative diversity protocols [47] exploit
spatial diversity available among a collection of distributed nodes that
relay messages for one another, in such a manner that the destination
terminal can combat the fading. Those relays can fully decode the
transmission from the source and then code the data by using a space–
time code to cooperatively relay to the destination. At the destination,
the space–time code implemented by the source and relays can achieve
full diversity.

7. Incremental Relaying [43]:

www.TechnicalBooksPDF.com
 Here, the destination will broadcast ACK or NACK information after
the first stage. The relay retransmits only after receiving a NACK. By
doing this, the bandwidth efficiency can be greatly improved, since the
only bandwidth increase occurs when the direct transmission link fails.
But the gain comes with additional implementation costs for the
feedback mechanism.

8. Cognitive Relaying [48, 49]:

Cognitive radio is a revolutionary paradigm with high spectral
efficiency involving wireless communication in occupied spectrum
without interfering with existing band occupants. Cooperative
communication protocols can help cognitive users to reduce the
detection time for clear spectral band, and thus to increase their agility.
On the other hand, relays can monitor the spectrum cognitively so as to
improve the source-to-destination link.

Cooperative transmission began as a physical layer protocol, and most

work in this area focuses on its merits in the physical layer. However, the
impact of cooperative transmission on the design of the higher layers is
obviously also of importance, although it is not yet well understood. We
now briefly discuss the impact of cooperative communications on different
layers. Specifically, we divide the discussion as follows.

Capacity analysis and new cooperative communication protocols:

The major concerns here are to analyze how much gain cooperative
transmission can bring to a link and to the overall network [50–52],
or how to implement cooperative transmission under practical
constraints [45, 46, 53].
Relay selection and power control: When there are several relays, a
question arises as to which one to select for a given retransmission.
After such relay selection, the next issue is how limited power
resources should be distributed over sources and relays [54, 55].
These questions have also been addressed in systems using multiuser
detection [56] and orthogonal frequency division multiplexing
(OFDM) [57].

www.TechnicalBooksPDF.com
 Routing protocols: Cooperative transmission can provide extra routes
for network protocols so that the network performance can be
significantly improved. These extra routes can be found through
traditional routes [58, 59] or from cooperative routes [60]. It has been
shown that network lifetime can be significantly improved via such
considerations [61]. Multi-hop cooperative transmission can be
considered as a special case of routing in [62–65].
Distributed resource allocation: Game theoretic approaches are
natural for distributed cooperative resource allocation problems, as
the individual nodes can use only local information to optimize
cooperative communications [66]. Moreover, as shown in [67],
cooperative game theory and cooperative transmission can be used to
improve packet forwarding networks with selfish nodes.
Others: Cooperative transmission has been considered jointly with
other problems such as source coding [68] and energy-efficient
broadcast [69].

1.2 Wireless Heterogeneous Networks

We list several different standards in Fig. 1.7 for different communication

rates and ranges. Those standards will fit different needs of various
applications. In the following, we categorize wireless networks and provide
some specifics.

www.TechnicalBooksPDF.com
 Figure 1.7 Comparison of different wireless networks

1.2.1 3G Cellular Networks

Third-generation (3G) mobile communication systems based on the

wideband code-division multiple access (WCDMA) and CDMA2000 radio
access technologies have seen widespread deployment around the world.
The applications supported by these commercial systems range from
circuit-switch services, such as voice and video telephony, to packet-
switched services, such as video streaming, email, and file transfer. As more
packet-based applications are developed and put into service, there is
increased need for better support for different quality of service (QoS),
higher spectral efficiency, and higher data rate for packet-switched services,
in order to further enhance user experience while maintaining efficient use
of the system resources.

The development of 3G will follow a few key trends, and its evolution
following these trends will continue as long as the physical limitations or

www.TechnicalBooksPDF.com
backward compatibility requirements do not force development to move
from evolution to revolution. The key trends include:

Voice services will continue to be important in the foreseeable future,

which means that capacity optimization for voice services will
continue.
Along with increasing use of IP-based applications, the importance
of data as well as simultaneous voice and data will increase.
Increased need for data means that the efficiency of data services
needs to be improved.
When more and more attractive multimedia terminals emerge in the
markets, the usage of such terminals will spread from offices, homes,
and airports to roads, and finally everywhere. This means that high-
quality high-data-rate applications will be needed everywhere as
well.
When the volume of data increases, the cost per transmitted bit needs
to decrease in order to make new services and applications affordable
for everybody. The other current trend is that in the 3G evolution
path, very high data rates are achieved in hot spots with WLAN
rather than via cellular-based standards.

For 3G cellular systems, there are two camps: 3G Partnership Project

(3GPP) [70] and 3G Partnership Project 2 (3GPP2) [71], which is based on
different 2G technologies. We list three major 3G standards as follows.

CDMA2000
The CDMA family of cellular networks grew out of work undertaken by
Qualcomm, a California-based company. Working on direct sequence
spread spectrum techniques, by using different spreading codes, a large
number of users could occupy the same channel at the same time, which
could provide a multiple access scheme for cellular telecommunications.
The first standard was the IS-95 and the first network was launched in Hong
Kong in 1996 under the brand name CDMAOne.

The CDMA system also has the following standards in its

developmental stages: IS-95, IS-95A, IS-95B, and CDMA2000 (1x/EV-DO,

www.TechnicalBooksPDF.com
1xEV-DV, 1xRTT, and 3xRTT). The first version of system standard IS-95
has never been launched for commercial purposes due to its prematurity.
The IS-95A has been applied for businesses since then and is still used
widely nowadays. IS-95B was a short version since the CDMA2000
standard was announced six months after it came out. The original IS-95A
standard only allowed for circuit switched data at 14.4 kbit/s, and IS-95B
provided up to 64 kbit/s data rates as well as a number of additional
services. A major improvement came later with the development of 3G
services. The first 3G standard was known as CDMA2000 1x, which
initially provided data rates up to 144 kbit/s. With further developments, the
systems hold the promise of allowing a maximum data rate of 307 kbits/s.

WCDMA/UMTS
WCDMA was developed by NTT DoCoMo as the air interface for their 3G
network FOMA. Later, NTT DoCoMo submitted the specification to the
International Telecommunication Union (ITU) as a candidate for the
international 3G standard known as IMT-2000. The ITU eventually
accepted WCDMA as part of the IMT-2000 family of 3G standards, as an
alternative to CDMA2000, EDGE, and the short range DECT system. Later,
WCDMA was selected as the air interface for the Universal Mobile
Telecommunications System (UMTS), the 3G successor to GSM.

TD-SCDMA
Transmit Diversity (TD) is one of the key contributing technologies to
defining the ITU-endorsed 3G systems WCDMA and CDMA2000. Spatial
diversity is introduced into the signal by transmitting through multiple
antennas. The antennas are spaced far enough apart1 that the signals
emanating from them can be assumed to undergo independent fading. In
addition to diversity gain, antenna gain can also be incorporated through
channel state feedback. This leads to the categorization of TD methods into
open loop and closed loop methods. Several methods of transmit diversity
in the forward link have been either under consideration or adopted for the
various 3G standards.

1.2.2 4G Cellular Network: LTE and WiMAX

www.TechnicalBooksPDF.com
Looking at the development of the Internet and applications, it is clear that
the complexity of the transferred content is rapidly increasing and will
increase further in the future. Generally, it can be said that the more
bandwidth that is available, the more bandwidth applications will consume.
In order to justify the need for a new air interface, goals need to be set high
enough to ensure that the system will be able to serve us long into the
future. A reasonable approach would be to aim at 100 Mb/s full-mobility
wide area coverage and 1 Gb/s low-mobility local area coverage with a
next-generation cellular system. Also, the future application and service
requirements will bring new requirements to the air interface and new
emphasis on air interface design. One such issue, which already strongly
impacts the 3G revolution, is the need to support IP and IP-based
multimedia. If both technology and spectrum to meet such requirements
cannot be found, the whole discussion of 4G may become obsolete. In Table
1.2, we compare key parameters of 4G with 3G. There are also some pre-
4G systems such as Long-Term Evolution (LTE), WiMAX, Ultra Mobile
Broadband (UMB, formerly EV-DO Rev. C) and Flash-OFDM.

Table 1.2 Comparison of 3G and 4G

3G 4G
Major Requirement Voice Driven Data/Voice
Driving Architecture Data add-on over IP
Network Architecture Wide Area Cell Based Hybrid with
WiFi and WPAN
Speed 384 kbps to 2 Mbps 20 to 10 Mbps
Frequency Band 1.8–2.4 GHz 2–8 GHz
Bandwidth 1.25–5–20 MHz 100 MHz
Switching Design Circuit and Packet Packet
Access DS-CDMA OFDM/MC-CDMA
FEC Convolution/Turbo Code Concatenated Coding
Component Design Antenna, Smart Antennas,
Multiband Adapter Software Radios

The following key features can be observed in all suggested 4G

technologies:

www.TechnicalBooksPDF.com
1. Physical layer transmission techniques are as follows:
MIMO: To attain ultra-high spectral efficiency by means of
spatial processing, including multi-antenna and multi-user
MIMO;
Frequency-domain equalization, for example multi-carrier
modulation (OFDM) in the downlink or single-carrier frequency-
domain equalization (SC-FDE) in the uplink: To exploit the
frequency selective channel property without complex
equalization;
Frequency-domain statistical multiplexing, for example, OFDMA
or single-carrier FDMA (SC-FDMA, a.k.a. linearly precoded
OFDMA, LP-OFDMA) in the uplink: variable bit rate by
assigning different sub-channels to different users based on the
channel conditions;
Turbo principle error-correcting codes: To minimize the required
SNR at the reception side;
2. Channel-dependent scheduling: To use the time-varying channel;
3. Link adaptation: Adaptive modulation and error-correcting codes;
4. Mobile-IP utilized for mobility;
5. IP-based femtocells (home nodes connected to fixed Internet broadband
infrastructure).

As opposed to earlier generations, 4G systems do not support circuit-

switched telephony. IEEE 802.20, UMB and OFDM standards lack soft-
handover support, also known as cooperative relaying.

Two 4G candidate systems are commercially deployed: the Mobile

WiMAX standard (first used in South Korea in 2006), and the first-release
Long-Term Evolution (LTE) standard (in Oslo, Norway and Stockholm,
Sweden since 2009), as discussed in the following two subsections.

Long-Term Evolution (LTE) and LTE-Advanced Networks

LTE, an initialism of Long-Term Evolution, marketed as 4G LTE, is a
standard for wireless communication of high-speed data for mobile phones

www.TechnicalBooksPDF.com
and data terminals. It is based on the GSM/EDGE and UMTS/HSPA
network technologies, increasing the capacity and speed using a different
radio interface together with core network improvements. The standard is
developed by the 3GPP (3rd Generation Partnership Project) and is
specified in its Release 8 document series, with minor enhancements
described in Release 9.

LTE is a standard for wireless data communications technology and an

evolution of the GSM/UMTS standards. The goal of LTE was to increase
the capacity and speed of wireless data networks using new DSP (digital
signal processing) techniques and modulations that were developed around
the turn of the millennium. A further goal was the redesign and
simplification of the network architecture to an IP-based system with
significantly reduced transfer latency compared to the 3G architecture. The
LTE wireless interface is incompatible with 2G and 3G networks, so that it
must be operated on a separate wireless spectrum.

Much of the LTE standard addresses the upgrading of 3G UMTS to

what will eventually be 4G mobile communications technology. A large
amount of the work is aimed at simplifying the architecture of the system,
as it transits from the existing UMTS circuit + packet switching combined
network, to an all-IP flat architecture system. E-UTRA is the air interface of
LTE. Its main features are:

1. Peak download rates up to 299.6 Mbit/s and upload rates up to 75.4

Mbit/s depending on the user equipment category (with 4–4 antennas
using 20 MHz of spectrum). Five different terminal classes have been
defined from a voice centric class up to a high end terminal that
supports the peak data rates. All terminals will be able to process 20
MHz bandwidth.
2. Low data transfer latencies (sub-5 ms latency for small IP packets in
optimal conditions), and lower latencies for handover and connection
setup time than with previous radio access technologies.
3. Improved support for mobility, exemplified by support for terminals
moving at up to 350 km/h (220 mph) or 500 km/h (310 mph)

www.TechnicalBooksPDF.com
 depending on the frequency band. [25]
4. OFDMA for the downlink, SC-FDMA for the uplink to conserve
power
5. Support for both FDD and TDD communication systems as well as
half-duplex FDD with the same radio access technology
6. Support for all frequency bands currently used by IMT systems by
ITU-R.
7. Increased spectrum flexibility: 1.4 MHz, 3 MHz, 5 MHz, 10 MHz, 15
MHz, and 20 MHz wide cells are standardized. (W-CDMA has no
option for other than 5 MHz slices, leading to some problems with
rollouts of the technology in countries where 5 MHz is a commonly
allocated amount of spectrum which is frequently already in use with
legacy standards such as 2G GSM and cdmaOne.)
8. Support for cell sizes from tens of meters radius (femto and picocells)
up to 100 km (62 miles) radius macrocells. In the lower frequency
bands to be used in rural areas, 5 km (3.1 miles) is the optimal cell
size, 30 km (19 miles) having reasonable performance, and up to 100
km cell sizes supported with acceptable performance. In city and urban
areas, higher frequency bands (such as 2.6 GHz in EU) are used to
support high-speed mobile broadband. In this case, cell sizes may be 1
km (0.62 miles) or even less.
9. Supports at least 200 active data clients in every 5 MHz cell. [26]
10. Simplified architecture: The network side of E-UTRAN is composed
only of eNode Bs
11. Support for interoperation and coexistence with legacy standards (e.g.,
GSM/EDGE, UMTS and CDMA2000). Users can start a call or
transfer of data in an area using an LTE standard, and, should coverage
be unavailable, continue the operation without any action on their part
using GSM/GPRS or W-CDMA-based UMTS or even 3GPP2
networks (such as cdmaOne or CDMA2000) packet-switched radio
interface. Support for MBSFN (Multicast-Broadcast Single Frequency
Network). This feature can deliver services such as Mobile TV using

www.TechnicalBooksPDF.com
 the LTE infrastructure, and is a competitor for DVB-H-based TV
broadcasting.

LTE Advanced is a mobile communication standard, formally submitted

as a candidate 4G system to ITU-T in late 2009, approved into ITU,
International Telecommunications Union, IMT-Advanced and was finalized
by 3GPP in March 2011. It is standardized by the 3GPP as a major
enhancement of the Long-Term Evolution (LTE) standard. The LTE
Advanced standard formally satisfies the ITU-R requirements to be
considered IMT-Advanced. Major technical considerations include the
following:

1. Continual improvement to the LTE radio technology and architecture;

2. Scenarios and performance requirements for working with legacy radio
technologies;
3. Backward compatibility of LTE-Advanced with LTE: an LTE terminal
should be able to work in an LTE-Advanced network and vice versa;
any exceptions will be considered by 3GPP;
4. Consideration of recent World Radiocommunication Conference (WRC-
07) decisions regarding frequency bands to ensure that LTE-Advanced
accommodates the geographically available spectrum for channels
above 20 MHz; also, specifications must recognize those parts of the
world in which wideband channels are not available.

WiMAX Networks
Wireless Metropolitan Area Network (WMAN) technology is a relatively
new field that began in 1998. From that time, a new standard has emerged
to handle the implementation, IEEE 802.16. The equivalent of 802.16 in
Europe is HIPERMAN. The WiMAX Forum is working to ensure that
802.16 and HIPERMAN interoperate seamlessly. This standard has helped
to pave the way for WMAN technology globally, and since its first
inception it has received six expansions onto the standards. WMAN differs

www.TechnicalBooksPDF.com
from other wireless technologies in that it is designed for a broader
audience, such as a large corporation or an entire city.

The 802.16 MAC uses a scheduling algorithm for which the subscriber
station needs to compete once (for initial entry into the network). After the
competition, the subscribe station is allocated an access slot by the base
station. The time slot can enlarge and contract, but it remains assigned to
the subscriber stations, which means that other subscribers cannot use it.
The 802.16 scheduling algorithm is stable under overload and over-
subscription (unlike 802.11). It can also be more bandwidth efficient. The
scheduling algorithm also allows the base station to control QoS parameters
by balancing the time slot assignments among the application needs of the
subscriber stations. Moreover, the MAC layer is also in charge of protocol
data unit (PDU) assembly and disassembly. A detailed illustration for
different layers’ protocols of 802.16 is shown in Fig. 1.8.

Figure 1.8 WIMAX protocol stacks

The operation standards for WMANs are regulated under IEEE standard
802.16 [72], and the WMANs are allowed the operating frequency range of
10–66 GHz. With such a broad spectrum to work with, WMANs have the
ability to transmit over previous wireless frequencies such as IEEE

www.TechnicalBooksPDF.com
802.11b/g, causing less interference with other wireless products. The only
downside to using such high frequencies is that WMAN needs a line of
sight between the transmitters and receivers, much like a directional
antenna. Line of sight, however, decreases multi-path distortion, allowing
higher bandwidths to be achieved, and it can attain up to 75 Mbps for both
uplink and downlink on a single channel [73]. Some extensions of 802.16
standards are listed as follows and in Table 1.3.

IEEE 802.16a: The IEEE has developed 802.16a for use in licensed
and license-exempt frequencies from 2 to 11 GHz. Most commercial
interest in IEEE 802.16 is in these lower frequency ranges. At the
lower ranges, the signals can penetrate barriers and, thus, do not
require a line of sight between transceiver and antenna. This enables
more flexible WiMax implementations while maintaining the
technologies, data rate and transmission range. IEEE 802.16a
supports mesh deployment, in which transceivers can pass a single
communication on to other transceivers, thereby extending basic
802.16s transmission range.
IEEE 802.16b: This extension increases the spectrum the technology
can use in the 5 and 6 GHz frequency bands and improves quality of
service. WiMax provides QoS to ensure priority transmission for
real-time voice and video and to offer differentiated service levels for
different traffic types.
IEEE 802.16c: IEEE 802.16c represents a 10 to 66 GHz system
profile that standardizes more details of the technology. This
encourages more consistent implementation and, therefore,
interoperability.
IEEE 802.16d: IEEE 802.16d includes minor improvements and
fixes to 802.16a. This extension also creates system profiles for
compliance testing of 802.16a devices.
IEEE 802.16e or Mobile WiMAX: This technology will standardize
networking between carriers’ base stations and mobile devices, rather
than just between base stations and fixed recipients. IEEE 802.16e
would enable the high-speed signal hand-offs necessary for
communications with users moving at vehicular speeds. It offers peak
data rates of 128 Mbit/s downlink and 56 Mbit/s uplink over 20
MHz-wide channels.

www.TechnicalBooksPDF.com
 IEEE 802.16m or WirelessMAN-Advanced: The IEEE 802.16m or
WirelessMAN-Advanced evolution of 802.16e is under development,
with the objective to fulfill the IMT-Advanced criteria of 1 Gbit/s for
stationary reception and 100 Mbit/s for mobile reception.
Table 1.3 Comparison of 802.16 Standards
802.16 802.16a/802.16d 802.16e
Date Dec. 2001 Jan. 2003/Q3 2004 Q3, 2004
Spectrum 10–66 GHz 11 GHz 6 GHz
Channels line of sight only non–line of sight non–line of sight
Modulation QPSK,16QAM, OFDM256, QPSK, same as 802.11a
64QAM 16QAM, 64QAM
Mobility Fixed Fixed Pedestrian Mobility
regional roaming
Bandwidth 20.25.28 MHz 1.25–20 MHz same as 802.16a
Throughput up to 75 Mbps up to 75Mbps up to 30 Mbps
Cell radius 1–3 miles 3–5 miles 1–3 miles

In addition to IEEE 802.16, IEEE 802.20 (IEEE 802.20) and the Mobile
Broadband Wireless Access (MBWA) Working Group aim to prepare a
formal specification for a packet-based air interface designed for IP-based
services. The goal is to create an interface that will allow the creation of
low-cost, always-on, and truly mobile broadband wireless networks,
nicknamed Mobile-Fi. IEEE 802.20 will be specified according to a layered
architecture, which is consistent with other IEEE 802 specifications. The
scope of the working group consists of the following layers: the PHYsical
(PHY), Medium Access Control (MAC), and Logical Link Control (LLC)
layers. The air interface will operate in bands below 3.5 GHz and with a
peak data rate of over 1 Mbit/s. The goals of 802.20 and 802.16e, the so-
called “mobile WiMAX,” are similar.

WiMAX can be viewed as “last mile” connectivity at high data rates.

This could result in lower pricing for both home and business customers as
competition lowers prices. In areas without pre-existing physical cable or
telephone networks, WiMAX may be a viable alternative for broadband
access that has been economically unavailable. Prior to WiMAX, many
operators have been using proprietary, fixed wireless technologies for

www.TechnicalBooksPDF.com
broadband services. For this reason, WiMAX has its significant markets in
rural areas and developing countries.

1.2.3 5G Cellular Networks and Beyond

5G denotes the next major phase of mobile telecommunications standards

beyond the current 4G/IMT-Advanced standards. As of 2013, 5G does not
describe any particular specification in any official document published by
any telecommunication standardization body. Moreover, still no
transnational 5G development projects have officially been launched, and
there is still a large extent of debate on what 5G is exactly about.

If 5G appears, the major difference from a user point of view between

4G and 5G techniques must be something other than increased maximum
throughput; for example, higher system spectral efficiency, lower battery
consumption, lower outage probability, better coverage, high bit rates in
larger portions of the coverage area, lower latencies, higher number of
supported devices, lower infrastructure deployment costs, higher versatility
and scalability, or higher reliability of communications. Key concepts
suggested and discussed for 5G are:

1. New data coding and modulation techniques, including filter bank

multicarrier or non-orthogonal multiple access schemes;
2. Modified physical layer numerology, for instance, to enable strongly
reduced end-to-end latency;
3. Massive Dense Networks, also known as Massive Distributed MIMO,
providing green flexible small cells 5G Green Dense Small Cells: a
transmission point equipped with a very large number of antennas that
simultaneously serve multiple users; with massive MIMO, multiple
messages for several terminals can be transmitted on the same time-
frequency resource, maximizing beamforming gain while minimizing
interference;
4. Advanced interference and mobility management, achieved with the
cooperation of different transmission points with overlapped coverage,

www.TechnicalBooksPDF.com
 and encompassing the option of a flexible usage of resources for uplink
and downlink transmission in each cell, the option of direct device-to-
device transmission and advanced interference cancellation techniques;
5. Efficient support of machine-type devices to enable the Internet of
Things with potentially higher numbers of connected devices, as well
as novel applications such as mission-critical control or traffic safety,
requiring reduced latency and enhanced reliability;
6. The usage of millimeter wave frequencies (e.g., up to 90 GHz) for
wireless backhaul and/or access (IEEE rather than ITU generations);
7. Pervasive networks providing ubiquitous computing: the user can
simultaneously be connected to several wireless access technologies
and seamlessly move between them (e.g., 2.5G, 3G, 4G, or 5G mobile
networks, WiFi, WPAN, or any other future access technology); in 5G,
the concept may be further developed into multiple concurrent data
transfer paths;
8. Multi-hop networks, cognitive radio technology and D2D
communications, which will be discussed later;
9. Dynamic Adhoc Wireless Networks (DAWN), essentially identical to
Mobile Adhoc NETwork (MANET), Wireless Mesh Network (WMN)
or wireless grids, combined with smart antennas, cooperative diversity
and flexible modula- tion;
10. Vandermonde-subspace Frequency Division Multiplexing (VFDM): a
modulation scheme to allow the co-existence of macro-cells and
cognitive radio small-cells in a two-tiered LTE/4G network;
11. IPv6, where a visiting care-of Mobile IP address is assigned according
to location and connected network;
12. User-centric (or cell phone developer initiated) network concept
instead of operator-initiated (as in 1G) or system developer-initiated
(as in 2G, 3G and 4G) standards;
13. World Wide Wireless Web (WWWW), i.e., comprehensive wireless-
based web applications that include full multimedia capability beyond
4G speeds.

www.TechnicalBooksPDF.com
 Table 1.4 shows a comparison as well as other competing technologies.

Table 1.4 Comparison of different standards

Common Primary Use Radio Downstream Upstream Notes
Name Tech
HSPA+ Used in 4G CDMA/ 21 42 84 672 5.8 11.5 22 168 HSPA+ is widely
3GPP FDD deployed. Revision 11 of
MIMO the 3GPP states that
HSPA+ is expected to have
a throughput capacity of
672 Mbit/s.
LTE 3GPP General 4G OFDMA/ 100 Cat3 150 50 Cat3/4 75 LTE-Advanced update
MIMO/ Cat4 300 Cat5 Cat5 (in 20 expected to offer peak rates
SC- (in 20 MHz MHz FDD) up to 1 Gbit/s fixed speeds
FDMA FDD) and 100 Mb/s to mobile
users.
WiMax rel 1 WirelessMAN MIMO- 37 (10 MHz 17 (10 MHz With 2 2 MIMO.
802.16 SOFDMA TDD) TDD)
WiMax rel WirelessMAN MIMO- 83 (20 MHz 46 (20 MHz With 2 2 MIMO.
1.5 802.16- SOFDMA TDD) 141 (2 TDD) 138 (2 Enhanced with 20 MHz
2009 20 MHz FDD) 20 MHz FDD) channels in 802.16-2009.
WiMAX rel WirelessMAN MIMO- 2 2 MIMO 110 2 2 MIMO 70 Also, low mobility users
2 SOFDMA (20 MHz TDD) (20 MHz TDD) can aggregate multiple
802.16 m 183 (2 20 MHz 188 (2 20 MHz channels to get a download
FDD) 4 4 FDD) 4 4 throughput of up to 1
MIMO 219 (20 MIMO 140 (20 Gbit/s.
MHz TDD) 365 MHz TDD) 376
(2 20 MHz (2 20 MHz
FDD) FDD)
Flash- Mobile Flash- 5.3 10.6 15.9 1.8 3.6 5.4 Mobile range 30 km (18
OFDM Internet OFDM miles) extended range 55
mobility up to km (34 miles).
200 mph (350
km/h)
HIPERMAN Mobile OFDM 56.9 56.9
Internet
WiFi 802.11 Mobile OFDM/ 288.8 (using 288.8 (using Antenna, RF front end
(n) Internet MIMO enhancements and minor
configuration in configuration in protocol timer tweaks have
20 MHz 20 MHz helped deploy long range
bandwidth) or bandwidth) or P2P networks
600 (using 600 (using compromising on radial
coverage, throughput
configuration in configuration in and/or spectra efficiency.

www.TechnicalBooksPDF.com
 40 MHz 40 MHz
bandwidth) bandwidth)
iBurst Mobile HC- 95 36 Cell Radius: 312 km
802.20 Internet SDMA/ Speed: 250 km/h Spectral
TDD/ Efficiency: 13
MIMO bits/s/Hz/cell Spectrum
Reuse Factor: “1.”
EDGE Mobile TDMA/ 1.6 0.5 3GPP Release 7.
Evolution Internet FDD
GSM
UMTS W- General 3G CDMA/ 0.384 14.4 0.384 5.76 HSDPA is widely
CDMA FDD deployed. Typical
HSDPA + CDMA/ downlink rates today 2
HSUPA FDD/ Mbit/s, 200 kbit/s uplink;
MIMO HSPA+ downlink up to 56
Mbit/s.
UMTS-TDD Mobile CDMA/ 16 16 Reported speeds according
Internet TDD to IPWireless using
16QAM modulation
similar to
HSDPA+HSUPA.
EV-DO Rel. Mobile CDMA/ 2.45 3.1 4.9 N 0.15 1.8 1.8 N Rev B note: N is the
0 EV-DO Internet FDD number of 1.25 MHz
Rev.A EV- chunks of spectrum used.
DO Rev.B EV-DO is not designed for
CDMA2000 voice, and requires a
fallback to 1 RTT when a
voice call is placed or
received.

1.2.4 Femtocell/Small Cell Networks

Femtocell (or small cell) is a small low-power cheap cellular base station,
typically designed for use in a home or small business. Femtocell connects
to the service provider’s network via broadband (such as DSL or cable), as
illustrated in Fig. 1.9. Femtocell allows wireless service providers to extend
service coverage indoors or at the cell edge, especially where access would
otherwise be limited or unavailable. In addition to coverage, quality of
service is also improved.

www.TechnicalBooksPDF.com
 Figure 1.9 Illustration of Femtocell/Small Cell

In 3GPP terminology, a Home Node B (HNB) is a 3G femtocell. A

Home eNode B (HeNB) is an LTE femtocell. Typically the range of a
standard base station may be up to 35 kilometers (22 miles), a microcell is
less than two kilometers wide, a picocell is 200 meters or less, and a
femtocell is on the order of 10 meters, although AT&T calls its product,
with a range of 40 feet (12 m), a microcell. Many operators have launched
femtocell service, including Vodafone, SFR, AT&T, Sprint Nextel, Verizon,
Zain, Mobile TeleSystems, and Orange.

Femtocells can also be viewed as the alternative for Wi-Fi, provided

instead by wireless service providers that can offer attractive tariffs. The
major difference from Wi-Fi is that the wireless service provider has some
limited control, instead of ad hoc mode. The major difference from regular
cellular cell is that the backhaul is not dedicated lines but via broadband
networks such as DSL or cable.

Typically, two to four active mobile phones are supported in a

residential setting, and eight to sixteen active mobile phones in enterprise
settings. There are two femtocell access control models: Closed Subscriber
Group (CSG), where only pre-registered mobile users can use a certain
femtocell, and Open Subscriber Group (OSG), where any mobile can use
any femtocell; or, at least, one that is “open.”

www.TechnicalBooksPDF.com
 Femtocells must be self-organized networks and support an essentially
plug-and-play operation. There are many technique challenges for
deploying femtocell networks, such as Interference Management; Cell
Association and Biasing; Mobility and Handover; and Economic &
Regulatory Issues.

Femtocell Standardization
The governing body for standardization is the Femto Forum
(www.femtoforum.org). The mission is to advance the development and
adoption of small cells for the provision of high-quality 2G/3G/4G
coverage and services within residential, enterprise, public, and rural access
markets. The Femto Forum is active in two main areas: standardization,
regulation & interoperability, and marketing & promotion. There are
different working groups: the marketing & promotion group; the radio &
physical layer group; the networks & interoperability group; and the
regulatory group. Different standards have the following tasks

1. 3GPP standards for UMTS femtocells:

Interface between the femtocell (Home Node B, HNB) and the
femto network gateway (HNB Gateway, HNB-GW);
Security protocols to authenticate femtocell (HNBs) and secure
communications across the untrusted Internet;
Management protocols for touch-free Operations, Administration,
and Management (OA&M) of femtocells (HNB devices).
2. 3GPP2 standards for CDMA femtocells:
SIP/IMS-based circuit services architecture;
Packet data architecture;
Security framework;
Enhancements to mobile devices to make them more femto-
aware;
Foundations of femtozone services (Local IP Access and Remote
IP Access);
Femtocell management architecture.
3. LTE/LTE-Advanced:

www.TechnicalBooksPDF.com
 There is significant need for LTE femtocells, since there is a limit to
how many outdoor cell sites can be built; the spectrum available to any
particular operator is limited; cell site backhaul is expensive. LTE is the
first cellular technology which will be able to take full advantage of
femtocell, due to the large quantity of dynamically allocated time and
frequency slots. In Fig. 1.10, the femtocell architecture in LTE networks
is illustrated.

Figure 1.10 LTE architecture with deployed HeNB gateway

1.2.5 Device-to-Device Networks

As one of next-generation wireless communication systems, Third

Generation Partnership Project (3GPP) Long-Term Evolution (LTE) is
committed to provide technologies for high data rates and system capacity.
Further, LTE-Advanced (LTE-A) was defined to support new components
for LTE to meet higher communication demands [74]. In particular,
performance and Quality of Service (QoS) of local area services need to be
improved significantly by reusing the spectrum resources. However, reuse
of the unlicensed spectrum may not provide a stable controlled

www.TechnicalBooksPDF.com
environment [75]. Therefore, the approach to exploit the licensed spectrum
for local area services has attracted much attention. In this chapter, we
present the basic concepts of Device-to-Device (D2D) communications in
the licensed spectrum bands. We first provide an overview of D2D
communications underlaying cellular network. We then discuss about
access methods, device synchronization and discovery mechanisms. Next,
mode selection, spectrum sharing, power control, Multiple-Input-Multiple-
Output (MIMO) techniques, and signal and interference model are briefly
introduced. The concept of D2D Local Area Networks (LAN) is proposed,
and finally, we outline the issues and challenges in D2D communications.

Overview of D2D Communications

D2D communications commonly refers to the techniques that enable
devices to communicate directly without an infrastructure of access points
or base stations. D2D communications is a technology component for LTE-
A, where items of user equipment (UE) transmit data signals to each other
over a direct link/connection using the cellular resources instead of through
the eNB (i.e., a base station). As an underlay to the cellular network, D2D
communications allow the spectral efficiency to be increased [74, 76–82].
While D2D communications is considered as an add-on component in the
4G systems, it is expected to be a native feature supported by the next
generation (e.g., fifth generation [5G]) cellular networks.

D2D communications may be of three types as shown in Figure 1.11:

Peer-to-peer communication: This is a point-to-point communication,

and most studies on D2D communication consider this type of
transmission.
Cooperative communication: This uses mobiles as relays to extend
coverage, and exploits cooperative diversity through multiple
collaborative mobiles to obtain space diversity.
Multiple-hop (multihop) communication: This is similar to mobile ad
hoc network and mesh network, which may include complex data
superposition, and data routing, e.g., wireless network coding.

www.TechnicalBooksPDF.com
 Figure 1.11 Device-to-device signal transmissions

Although D2D communications brings improvement in spectral

efficiency and makes large benefits on system capacity, it also causes
interference with the cellular network as a result of spectrum sharing. Thus,
an efficient interference coordination must be formulated to guarantee a
target performance level of the cellular communication. There exist several
works about the of D2D UEs for restricting co-channel
interference [74, 76, 83, 84]. The authors in [85] utilize MIMO transmission
schemes to avoid interference from cellular downlink to D2D receivers
sharing the same resources, which aims at guaranteeing the required
performances for the D2D communications. Interference managements both
from cellular to D2D communications and from D2D to cellular networks
are considered in [86]. To further improve the gain from intra-cell spectrum
reuse, properly pairing the cellular and D2D users for sharing the same
resources are studied [87, 88]. The authors in [88] propose an alternative
greedy heuristic algorithm to lessen interference to the primary cellular
networks using channel state information (CSI). The scheme is easily
operated yet cannot avoid signaling overhead. In [89], the resource
allocation scheme avoids the harmful interference by tracking the near-far
interference, identifies the interfering cellular users, and makes the uplink
(UL) frequency bands efficiently used. Additionally, the goal is to prevent
interference from cellular to D2D communication. In [90], the authors
provide an analysis on optimal resource allocation and power control
between the cellular and D2D connections that share the same resources for
different resource sharing modes, and evaluate the performance of the D2D
underlay system in both a single cell scenario and the Manhattan grid

www.TechnicalBooksPDF.com
environment. Then, the schemes are to further optimize the resource usage
among users sharing the same resources.

The existing works in the literature show that, by proper resource

management, D2D communications can effectively improve the system
throughput with the interference between cellular networks and D2D
transmissions being minimized. However, the problem of allocating cellular
resources to D2D transmissions is not trivial. In the next subsection, we will
discuss different resource management methods for D2D communications.

Key Technologies for D2D Communications

Configuration of D2D Communications

The D2D networks can be configured in the three following ways to allow
or restrict their usage by certain users:

Network controlled D2D: In this scenario, the communication

signaling setup and thereafter resource allocation for both cellular
and D2D users are controlled by the Base Station (BS) and the core
network. This centralized configuration benefits from efficient
interference avoidance and resource management. However, when
the number of D2D links becomes large, this scheme incurs large
amounts of control signaling, which can increase the overhead and
reduce the spectrum efficiency. Therefore, this fully network-
controlled approach is particularly useful for scenarios with small
number of D2D links.
Self-organized D2D: In this scenario, D2D users themselves realize
the communication in a self-organizing way by finding the empty
spectrum hole. This configuration is similar to cognitive radio, which
allows the D2D users to sense a surrounding environment obtaining
CSI, interference, and cellular system information. This distributed
method can effectively avoid the controlling signaling overhead, and
the time delay, but the self-organized method may cause
communication chaos and unstability due to lack of control by the
operators in the licensed spectrum.
Network-assisted D2D: The D2D users operate in a self-organized
way, and for resource management, exchange with cellular system

www.TechnicalBooksPDF.com
 limited controlling information. The cellular network can use the
status of D2D communications for better control purposes. This
approach has the merits of the first two approaches.

Device Synchronization and Discovery

For D2D communications, synchronization between cellular networks and
D2D users and among D2D users themselves will be necessary to minimize
multiple access interference and for proper handoff. The approaches in
IEEE 802.11 or in LTE can be adopted to enable the synchronization among
mobiles. Typically, device synchronization and discovery are realized in a
joint way.

The fundamental problem of device discovery is that two devices have

to meet in space, time, and frequency without any coordination. This can be
made possible via some randomized procedure and one of the peers
performs the responsibility of sending the beacon. For traditional peer
discovery, in both the ad hoc and the cellular cases, the discovery is made
possible by one party transmitting a known synchronization or reference
signal sequence (the beacon). Based on whether there are responses from
the discovering UEs or not, the discovery approaches can be classified into
two categories: beacon-based discovery and request-based discovery.
According to whether there is network participation for identification
detection, discovery procedure could be categorized into types: network-
assistance detection and non-network-assistance detection.

In the case of network-assisted D2D, the network can mediate in the

discovery process by recognizing D2D candidates, coordinating the time
and frequency allocations for sending/scanning for beacons, and thereby
making the pairing process more energy efficient and less time consuming.
A typical procedure is as follows:

Use direct signal to discover peer;

Set transmission power so that UEs in certain distance can hear the
broadcast;
Whoever receives the broadcast confirms that with the eNB.

www.TechnicalBooksPDF.com
Mode Selection
In a D2D underlay communications system, one of the most challenging
problems is to decide whether communicating devices should use cellular
or direct communication mode. In the D2D mode, data is directly
transmitted to the receiver while the cellular communication mode requires
the source device to transmit to the eNB and then the destination device
receives from the eNB on downlink (DL). Here, three different mode
selection criteria are considered.

1. Cellular: All devices are in cellular mode.

2. Force D2D: D2D mode is always selected for all the communicating
devices.
3. PL D2D: D2D mode is selected if any of the path losses between a
source device and its serving eNB, destination device and its serving
eNB, are greater than the path loss in the direct link between the source
node and the destination node.

Spectrum Sharing and Resource Management

Spectrum sharing methods for D2D communications can be categorized as
follows:

Overlay D2D communications: The D2D users occupy the vacant

cellular spectrum for communication. This approach can completely
eliminate cross-tier interference by dividing the licensed spectrum
into two parts (i.e., orthogonal channel assignment). That is, a
fraction of the subchannels will be used by the cellular users while
another fraction would be used by the D2D networks. Although it is
optimal from a cross-tier interference standpoint, this approach is
inefficient in terms of spectrum reuse.
Underlay D2D communications: In this spectrum sharing scheme,
multiple D2D users are allowed to work as an underlay with cellular
users, and thus improve the spectrum efficiency. Co-channel
assignment of the cellular and D2D users will be more efficient and
profitable for operators, although this is far more intricate than the
overlay scheme from the technical point of view.

www.TechnicalBooksPDF.com
 The overlay approach is easy to realize, but might not be spectrally
efficient. While the underlay method incurs a relatively more signaling
overhead, it can achieve a better overall system performance. To optimize
the system performance over spectrum sharing of both D2D and cellular
modes, radio resource management is important. Radio resource
management can be performed either in noncooperative or cooperative
manner. In a noncooperative solution, each D2D user can manage its
spectrum so as to maximize the throughput and QoS. By contrast, in a
cooperative approach, the D2D users can gather partial information about
spectrum usage and perform its allocation taking into account the effect that
it would have on its co-channel neighbors. In this way, the average cellular
and D2D users’ throughput and QoS, as well as their performances, can be
locally optimized.

Power Control
Power control is an important and effective way to coordinate the co-
channel interference. Power control can be performed by two methods:

Self-organized power control: The D2D users make power changes

in a self-organized way according to a predefined signal-to-
interference-plus-noise ratio (SINR) threshold in order to meet the
QoS, and meanwhile not affect the cellular users.
Network managed power control: Both cellular and D2D users
adaptively adjust their transmit power according to the SINR report.
Typically, the D2D users can control the transmit power first, and
then the cellular users make changes afterward. This iterative process
terminates when all the users meet their SINR requirements.

Obviously, the first method is not going to change the behaviors of

cellular users since the D2D users are treated invisibly. This method is
simple, but less efficient than the second method, which allows all the users
to adjust their transmit powers. However, the network-controlled approach
requires some information exchange among cellular users, D2D users, and
the eNB.

Uplink and Downlink Transmission with MIMO

www.TechnicalBooksPDF.com
The use of multiple-input and multiple-output (MIMO) antennas can
improve system capacity by multiplexing signals in the spatial domain and
increase robustness by exploiting space diversity. Specifically, by
performing transmit or receive beamforming, the use of multiple antennas
at the eNB and the UE can reduce the co-channel interference to other users
and thus to improve spectrum efficiency.

The different MIMO-based methods are as follows:

eNB beamforming: This type of multi-user MIMO like approach can

be performed at the cellular downlink to reduce interference to the
D2D users so that D2D communications can be allowed.
D2D beamforming: This avoids any harmful interference caused by
the D2D transmissions to the cellular and other D2D users.
Virtual D2D beamforming: This borrows the ideas of cooperation
among mobile nodes such that multiple D2D users collaboratively
form the beamforming matrices to improve the system performance.

1.2.6 LTE-Unlicensed

Currently, technologies such as 802.11(WiFi), 802.15.1(Bluetooth), and

802.15.4 (ZigBee) are implemented in 2.4 GHz ISM (Industry, Scientific,
and Medical) and 5 GHz U-NII (Unlicensed National Information
Infrastructure) bands, or more commonly referred to as the unlicensed
bands. So far, some operators have deployed a large number of WiFi access
points (APs) to offload the cellular traffic to the unlicensed spectrum. The
WiFi offloading is the trade-off of capacity and performance. However,
such efforts are not achieving their expectations in terms of reducing cost
and improving performance, for reasons such as the extra cost of the
backhaul and core network in addition to the existing cellular
infrastructures, the lack of good coordination between the cellular and WiFi
systems, and so on [91].

Carrier Aggregation (CA), as an LTE innovation, has enabled operators

to make the best use of the unlicensed bands. By aggregating the unlicensed

www.TechnicalBooksPDF.com
bands with an LTE anchor in the licensed spectrum, better coverage and
higher spectral efficiency can be achieved compared to WiFi offloading,
while seamless flow of data across licensed and unlicensed in a single core
network is also guaranteed. Generally, there are two approaches in terms of
integrating LTE into the unlicensed spectrum, as shown in Fig. 1.12: (1)
LTE-WiFi link aggregation (LWA),2 to leverage their WiFi networks using
both 2.4 GHz and 5 GHz bands; or (2) LTE Unlicensed carrier aggregation,
also referred to as the LTE Unlicensed, where LTE operating in the
unlicensed 5 GHz spectrum [92]. The link aggregation needs support from
both device and network sides. On the device side, the aggregation
happpens deep at the modem level. On the network side, it can either be
between collocated or seperate WiFi and LTE-WiFi APs. The eNBs control
the amount of traffic scheduled to LTE and WiFi, which ensures proper load
balancing between the LTE and WiFi links. The link aggregation is defined
in 3GPP Rel. 13 [93], and is demonstrated by QUALCOMM at MWC
2015. The LTE Unlicensed offers the tightest possible aggregation, as
shown in Fig. 1.13. It comes in multiple flavors: LTE-Unlicensed [94],
LAA(Licensed Assisted Access) [93] and MuLTEfire3 [95], which depends
on the specific deployment scenarios and the regions. The LTE-U, based on
3GPP Rel. 10/11/12, targets early mobile operators deployments in USA,
Korea, India and for marketers without Listen Before Talk (LBT)
requirements. For example, the US deployment of LTE Unlicensed in the
U-NII radio bands ranges from 5.15 GHz to 5.85 GHz as regulated by the
FCC [96]. Defined in 3GPP Rel. 13, LAA targets deployments in Europe,
Japan, and beyond, and for regions with specific access procedures with
LBT. The performance of LTE-U and LAA, and how they coexist fairly
with WiFi were also demonstrated at MWC 2015. The MuLTEfire broadens
the LTE ecosystem to new deployment opportunities by operating solely in
the unlicensed spectrum without a licensed anchor channel.

www.TechnicalBooksPDF.com
Figure 1.12 LTE Unlicensed (LTE-U/LAA) deploy scenario

Figure 1.13 LTE-U carrier aggregation

www.TechnicalBooksPDF.com
Benefits of Extending LTE-A to the Unlicensed Spectrum
In this section, we dicussed the benefits of ecxtending LTE-A to the
unlicensed spectrum, i.e., the LTE Unlicensed, in forms of LTE-U and
LAA. For both LTE-U and LAA, the signaling and control message is sent
through a reliable licensed anchor, and the unlicensed link is used only for
data. Both options use a single, unified core network that provides cost
efficiency and simplicity of management to operators, while offering
seamless service continuity and a better broadband experience to users. Due
to the low power resitriction imposed by the regulations on unlicensed
transmissions, the transmission range of unlicensed bands is relatively
small. Therefore, small cell deployment is more suited for LTE Unlicensed.

1. Better Network Performance

By aggregating unlicensed carriers with licensed FDD or TDD carriers
in small cells, LTE Unlicensed augments existing LTE mobile networks
to provide the best mobile broadband performance. By using the LTE
technology, which supports coordinated, synchronized scheduling of
resources, and has an efficient radio link with features such as scaling to
lower date rates, handling larger delay spreads, Hybrid ARQ (HARQ)
and so on, the LTE Unlicensed can provide or more capacity in dense
deployments. By trading off part of the capacity increase, the network
coverage can be improved.

2. Enhanced User Experience

As an integration of the LTE system, the unlicensed secondary carrier
function can be enabled or disabled seamlessly by the LTE control
without the notice of users. Thus, mobile users can experience an
enhanced date rate and seamless transmission. In addition, in
applications that have strict QoS requirements and important signaling
traffic, data are routed through the licensed spectrum, thus ensuring high
reliability.

3. Unified LTE Network

For cellular operators that already have an LTE network, LTE
Unlicensed can be bundled seamlessly with LTE carrier in the licensed

www.TechnicalBooksPDF.com
 spectrum using Supplemental DownLink (SDL)4 or both downlink and
uplink CA. The licensed and unlicensed carriers can be managed
conveniently as one unified network, and data traffic can be routed
flexibly through both carriers.

4. Fair Co-Existence with WiFi

The criterion for good co-existence is that the impact of LTE Unlicensed
interference on a given neighboring WiFi node is no more than the
impact from other WiFi nodes in the same setup. LTE, as a coordinated
system, uses sensing-based channel occupancy can effectively achieve
interference coordination with existing WiFi users. Interestingly, LTE
Unlicensed in many cases can be a better neighbor to WiFi than WiFi
itself [92].

1.2.7 WiFi Networks

IEEE 802.11 denotes a set of Wireless Local Area Network (WLAN)

standards developed by working group 11 of the IEEE LAN/MAN
Standards Committee (IEEE 802). WiFi is a brand originally licensed by the
WiFi Alliance to describe the underlying technology of WLAN based on
the IEEE 802.11 specifications. It was developed to be used for mobile
computing devices, such as laptops, in LANs, but it is now increasingly
used for more services, including Internet and VoIP phone access, gaming,
and basic connectivity of consumer electronics such as televisions, DVD
players, and digital cameras.

In the physical layer, 802.11b operates within the 2.4 GHz industrial,
scientific, and medical (ISM) band. The original 802.11b defines data rates
of 1 Mbps and 2 Mbps via radio waves using Frequency Hopping Spread
Spectrum (FHSS) or Direct Sequence Spread Spectrum (DSSS). For FHSS,
2.4 GHz band is divided into 75 1-MHz subchannels. The sender and
receiver agree on a hopping pattern, and data is sent over a sequence of the
subchannels. Each conversation within the 802.11 network occurs over a
different hopping pattern. Due to Federal Communications Commission

www.TechnicalBooksPDF.com
(FCC) regulations that restrict subchannel bandwidth to 1 MHz, FHSS
techniques are limited to speeds of no higher than 2 Mbps. DSSS divides
the 2.4 GHz band into 14 22-MHz channels. Adjacent channels overlap one
another partially, with three of the 14 being completely non-overlapping.
The spreading code is an 11-bit Barker sequence. Binary Phase Shift
Keying (BPSK) and Quadrature Phase Shift Keying (QPSK) are used to
provide different rates. To increase the data rate to 5.5 Mbps and 11 Mpbs
in the 802.11b standard, an advanced coding technique, Complementary
Code Keying (CCK) is employed. A complementary code contains a pair of
finite bit sequences of equal length, in which the number of pairs of
identical elements (1 or 0) with any given separation in one sequence is
equal to the number of pairs of unlike elements having the same separation
in the other sequence. A network using CCK can transfer more data per unit
time for a given signal bandwidth than a network using the Barker code,
because CCK makes more efficient use of the bit sequences. CCK consists
of a set of 64 8-bit code words. The 5.5 Mbps rate uses CCK to encode four
bits per carrier, while the 11 Mbps rate encodes eight bits per carrier. Both
speeds use QPSK as the modulation technique and signal at 1.375 MSps.
Table 1.5 shows the different rates for 802.11b.

Standard 802.11a adopts orthogonal frequency division multiplexing

(OFDM) at 5.15–5.25 GHz, 5.25–5.35 GHz, and 5.725–5.825 GHz to
support multiple data rates up to 54 Mbps. Standard 802.11g utilizes the 2.4
GHz band with OFDM modulation and is also backward compatible with
802.11b. For OFDM, the FFT has 64 subcarriers. There are 48 data
subcarriers and four carrier pilot subcarriers, for a total of 52 nonzero
subcarriers defined in IEEE 802.11a, plus 12 guard subcarriers. The IEEE
802.11a/g physical layer provides eight PHY modes with different
modulation schemes and different convolutional coding rates, and it can
offer various data rates.

To achieve higher data rates in the PHY layer, in January 2004, IEEE
announced that it had formed a new 802.11 Task Group (TGn) to develop a
new amendment to the 802.11 standard for wireless local-area networks.
802.11n builds upon previous 802.11 standards by adding MIMO (multiple-
input multiple-output). MIMO uses multiple transmitter and receiver

www.TechnicalBooksPDF.com
antennas to allow for increased data throughput through spatial
multiplexing and increased range by exploiting the spatial diversity.
Channels operating with a width of 40 MHz are another feature
incorporated into 802.11n; this doubles the channel width from 20 MHz in
previous 802.11 PHYs to transmit data, and provides twice the PHY data
rate available over a single 20 MHz channel. Data rates up to 600 Mbit/s are
achieved only with the maximum of four spatial streams using one 40
MHz-wide channel. 802.11n is backward compatible with 802.11b and
802.11g.

Table 1.5 802.11b Rates

Data Rate Code Length Modulation Symbol Rate Bits/Symbol
1 Mbps 11 (DSSS) BPSK 1 MSps 1
2 Mbps 11 (DSSS) QPSK 1 MSps 2
5.5 Mbps 8 (CCK) QPSK 1.375 MSps 4
11 Mbps 8 (CCK) QPSK 1.375 MSps 8

Recently, the 802.11ac standard has been under development. This

specification has expected multi-station WLAN throughput of at least 1
gigabit per second and a single link throughput of at least 500 megabits per
second (500 Mbit/s). This is accomplished by extending the air interface
concepts embraced by 802.11n: wider RF bandwidth (up to 160 MHz),
more MIMO spatial streams (up to eight), multi-user MIMO, and high-
density modulation (up to 256-QAM). Some new technologies are listed as
follows:

1. Extended channel binding: Mandatory 80 MHz channel bandwidth for

stations (vs. 40 MHz maximum in 802.11n), 160 MHz available
optionally.
2. More MIMO spatial streams: Support for up to eight spatial streams (vs.
four in 802.11n).
3. Multi-user MIMO (MU-MIMO): Multiple STAs, each with one or more
antennas, transmit or receive independent data streams simultaneously,
and “Space Division Multiple Access (SDMA)”: streams not separated
by frequency, but instead resolved spatially, analogous to 11n-style

www.TechnicalBooksPDF.com
 MIMO. Downlink MU-MIMO (one transmitting device, multiple
receiving devices) included as an optional mode.
4. Modulation: 256-QAM, rate 3/4 and 5/6, added as optional modes (vs.
64-QAM, rate 5/6 maximum in 802.11n). Controversy has been raised
with the suggestion that this modulation rate is not suitable for any
architecture other than extremely small cells and would be completely
useless to 3GPP.
5. Other elements/features: Beamforming with standardized sounding and
feedback for compatibility between vendors (non-standard in 802.11n,
making it hard for beamforming to work effectively between different
vendor products). MAC modifications (mostly to support above
changes). Coexistence mechanisms for 20/40/80/160 MHz channels,
11ac and 11a/n devices.

In the following, we list some other design topics.

Security (IEEE 802.11i)

Most concentration for WiFi is on free public access. However, the
eavesdroppers and hackers can fully take advantage of the WiFi
system. Currently, all 802.11a, b, and g devices support WEP (Wired
Equivalent Privacy) encryption, which has flaws. IEEE 802.11i, also
known as WiFi Protected Access 2 (WPA2), is an amendment to the
802.11 standard specifying security mechanisms for wireless
networks. The 802.11i specification defines two classes of security
algorithms: Robust Security Network Association (RSNA), and Pre-
RSNA. Pre-RSNA security consists of Wired Equivalent Privacy
(WEP) and 802.11 entity authentication. RSNA provides two data
confidentiality protocols, called the Temporal Key Integrity Protocol
(TKIP) and the Counter-mode/CBC-MAC Protocol (CCMP). The
RSNA establishment procedure includes 802.1X authentication and
key management protocols. Beyond IEEE 802.11i, it is worth
mentioning that WAPI (WLAN Authentication and Privacy
Infrastructure) is a Chinese National Standard for Wireless LAN (GB
15629.11-2003).
Mobility (IEEE 802.11p)
Mobility is an important attribute of wireless networks. Current

www.TechnicalBooksPDF.com
Wireless LAN standards provide mobility through roaming
capabilities. IEEE 802.11p also referred to as Wireless Access for the
Vehicular Environment (WAVE) defines enhancements to 802.11
required to support Intelligent Transportation Systems (ITS)
applications. This includes data exchange between high-speed
vehicles and between the vehicles and the roadside infrastructure in
the licensed ITS band of 5.9 GHz (5.85–5.925 GHz). 802.11p will be
used as the groundwork for DSRC (Dedicated Short Range
Communications), a U.S. Department of Transportation project that
will be emulated elsewhere, looking at vehicle-based communication
networks, particularly for applications such as toll collection, vehicle
safety services, and commerce transactions via cars. The ultimate
vision is a nationwide network that enables communications between
vehicles and roadside access points or other vehicles.
QoS support (IEEE 802.11e)
802.11e is the first wireless standard that spans home and business
environments. It adds quality-of-service (QoS) features and
multimedia support to the existing 802.11 wireless standards, while
maintaining full backward compatibility with these standards. QoS
and multimedia support are critical to wireless home networks where
voice, video, and audio will be delivered. Broadband service
providers view QoS and multimedia-capable home networks as an
essential ingredient for offering residential customers video on
demand, audio on demand, voice over IP, and high-speed Internet
access. 802.11e introduces two enhancements, Enhanced DCF
(EDCF) and Hybrid Coordination Function (HCF). In EDCF, a
station with high priority traffic waits a little less before it sends its
packet, on average, than a station with low priority traffic, so that
high priority traffic has a higher chance of being sent than low
priority traffic. In addition, each priority level is assigned a Transmit
Opportunity (TXOP) which is a bounded time interval during which
a station can send as many frames as possible. HCF works more like
PCF. With the PCF, QoS can be configured with great precision.
QoS-enabled stations have the ability to request specific transmission
parameters (data rate, jitter, etc.), which should allow advanced
applications like VoIP and video streaming to work more effectively
on a WiFi network.

www.TechnicalBooksPDF.com
 Integration of 3G and WLAN
The third-generation cellular networks and 802.11 local area wireless
networks possess complementary characteristics. 3G cellular
networks promise to offer always-on, ubiquitous connectivity and
mobility with relatively low data rates. 802.11 offers much higher
data rates, comparable to the cellular networks, but can cover only
smaller areas without mobility, suitable for hot-spot applications in
hotels and airports. The performance and flexibility of wireless data
services would be dramatically improved if users could seamlessly
roam across the two networks. By offering integrated 802.11/3G
services, 3G operators and Wireless Internet Service Providers
(WISP) can attract a wider user base and ultimately facilitate the
ubiquitous introduction of high-speed wireless services. Users can
also benefit from the enhanced performance and lower overall cost of
such a combined service. For a network node changing the type of
connectivity between 3G cellular phone and WLAN, the concept of
vertical handoff will be discussed in the latter chapters.

1.2.8 Wireless Personal Area Networks

A Wireless Personal Area Network (WPAN) is a computer network used for

wireless communication among devices (including telephones and personal
digital assistants) close to a person. The reach of a WPAN is typically a few
meters. WPANs can be used for communication among the personal
devices themselves (intrapersonal communication), or for connecting to a
higher level network and the Internet (an uplink). 802.15 is a
communications specification that was approved in early 2002 by the IEEE
Standards Association (IEEE-SA) for wireless personal area networks
(WPANs). Specifically, we list the following three sub-standards:

1. Bluetooth: The IEEE Standard 802.15.1 was approved as a new

standard for Bluetooth by the IEEE-SA Standards Board on 15 April
2002. The Bluetooth standard enables wireless communication between
multiple electronic devices within ten meters of each other. Bluetooth
devices are organized in piconets, which include one master device and

www.TechnicalBooksPDF.com
 up to seven slave devices. The Bluetooth devices communicate in the
2.4 GHz radio frequency band, enabling devices to communicate
without line-of-sight spacing, such as through walls or through a
person’s body. Bluetooth piconets utilize frequency-hopping spread
spectrum in seventy-nine 1 MHz bands, reducing the likelihood of
interference with other Bluetooth piconets.
2. ZigBee: IEEE 802.15.4-2003 (Low Rate WPAN) deals with low data
rate but very long battery life (months or even years) and very low
complexity. The first edition of the 802.15.4 standard was released in
May 2003. In March 2004, after forming Task Group 4b, Task Group 4
put itself in hibernation. The ZigBee set of high-level communication
protocols is based upon the specification produced by the IEEE 802.15.4
task group.
3. Ultra Wide Band (UWB): 802.15.3 is the IEEE standard for high data
rate WPAN designed to provide Quality of Service (QoS) for real-time
distribution of multimedia content, like video and music. It is ideally
suited for a home multimedia wireless network. The original standard
uses a “traditional” carrier-based 2.4 GHz radio as the physical layer
(PHY). A follow-on standard, 802.15.3a, defines an alternative PHY,
based on UWB, that will provide in excess of 110 Mbps at a 10 m
distance and 480 Mbps at 2 m. This will allow applications requiring
streaming of high-definition video between media servers and flat
screen HD monitors and extremely fast transfer of media files between
media servers and portable media devices.

1.2.9 Wireless Ad Hoc Networks

An ad hoc network is an autonomous collection of mobile users that

communicate over bandwidth-constrained wireless links. The network is
decentralized, so that all network activity including discovering the
topology and delivering messages must be executed by the nodes
themselves. Ad hoc networks need efficient distributed algorithms to
determine network organization, link scheduling, and routing. In a special
case of ad hoc network, Mobile Ad Hoc NETworks (MANETs), since the

www.TechnicalBooksPDF.com
nodes are mobile, the network topology may change rapidly and
unpredictably over time.

The first generation of ad hoc networks were initiated in the early

1970s, when Packet Radio NETworks (PRNET) was proposed from
Defense Advanced Research Projects Agency (DARPA) for multihop
networks in a combat environment and Areal Locations of Hazardous
Atmospheres (Aloha) was proposed in Hawai’i for distributed channel
access management. The second generation of ad hoc networks emerged in
1980s, when the ad hoc network systems were further enhanced and
implemented as a part of the Survivable Adaptive Radio Networks
(SURAN) program. SURAN provided a packet-switched network to the
mobile battlefield in an environment without infrastructure so as to be
beneficial in improving the performance of radios by making them smaller,
cheaper, and resilient to electronic attacks. In the 1990s, the concept of
commercial ad hoc networks arrived with notebook computers and other
viable communications equipment. For example, the IEEE 802.11
subcommittee had adopted the term “ad hoc networks.”

The advantages of ad hoc networks are the ease and speed of

deployment, which are important requirements for military applications.
For civil applications, ad hoc networks decrease dependence on expensive
infrastructure. The set of applications for ad hoc networks is diverse,
ranging from small, static networks that are constrained by power sources,
to large-scale, mobile, highly dynamic networks. Some typical applications
are personal area network, emergency operations such as policing and fire
fighting, civilian environments such as taxi networks, and military use on
the battle fields.

In contrast to the traditional wireless network with infrastructure, an ad

hoc network needs its own design requirements so as to be functional. Some
important aspects are distributed operation and self organization, dynamic
routing, connectivity, and mobility.

www.TechnicalBooksPDF.com
 Finally, we discuss two ad hoc standards in 802.11 for wireless local
networks and 802.15 for wireless personal area networks. One future design
goal of ad hoc networks is to let mobile users be able to form connections
and perform basic functions. These mobile users belong to different types of
networks such as cellular, WiFi, and WPAN.

Most installed wireless LANs today utilize “infrastructure” mode that

requires the use of one or more access points. With this configuration, the
access point provides an interface to a distribution system (e.g., Ethernet),
which enables wireless users to utilize corporate servers and Internet
applications. As an option, the 802.11 standard specifies “ad hoc” mode,
which allows the radio network interface card (NIC) to operate in what the
standard refers to as an independent basic service set (IBSS) network
configuration. With an IBSS, there are no access points. User devices
communicate directly with each other in a peer-to-peer manner.

For wireless personal area network applications such as Bluetooth, the

ad hoc network is set up by forming piconets. Within each piconet, only one
master device and possibly several slave devices form connections. A slave
device can belong to different piconets and serve as connection between
piconets. The major functionality of piconets are piconet forming and
maintenance, packet forwarding, and intra-piconet/inter-piconet scheduling.

1.2.10 Wireless Sensor Networks

A Wireless Sensor Network (WSN) is a wireless network consisting of

spatially distributed autonomous devices using sensors to cooperatively
monitor physical or environmental conditions, such as temperature, sound,
vibration, pressure, motion, or pollutants, at different locations. The goals
and tasks of sensor networks are to determine the value of some parameter
at a given location, to detect the occurrence of events of interest and
estimate parameters of the detected events, to classify a detected object, or
to track an object. The development of wireless sensor networks was
originally motivated by military applications such as battlefield

www.TechnicalBooksPDF.com
surveillance. However, wireless sensor networks are now used in many
civilian application areas, including environment and habitat monitoring,
healthcare applications, home automation, and traffic control.

In Fig. 1.14, we show the typical structure of one unit of sensor

networks. In addition to one or more sensors, each node in a sensor network
is typically equipped with a radio transceiver or other wireless
communications device, a small microprocessor, some memory, and an
energy source, usually a battery. For the sensors, different sensing
applications can include temperature, light, humidity, pressure,
accelerometers, magnetometers, chemical, acoustics, and image/video. The
microprocessor has significant constraint in computational power.
Currently, the devices typically have the component-based embedded
operating system. The available memory is also very limited. The current
radio transceivers for sensor networks are low-rate and short-range. Some
sensors can be powered by wired power source, while most of widely
deployed sensors are powered by battery. Exchanging the energy depleting
sensors is a challenging job, so power saving is critical in the design of such
wireless sensor networks.

Figure 1.14 Sensor network structure

The size of a single sensor node can vary from shoe-box-sized nodes
down to devices the size of a grain of dust. The cost of sensor nodes is

www.TechnicalBooksPDF.com
similarly variable, ranging from hundreds of dollars to a few cents,
depending on the size of the sensor network and the complexity required of
individual sensor nodes. Size and cost constraints on sensor nodes result in
corresponding constraints on resources such as energy, memory,
computational speed, and bandwidth. Basically, to design a wireless sensor
network, the following requirements should be considered [97]: large
number of (mostly stationary) sensors; low energy use to extend sensor
network lifetime; network self-organization; collaborative signal
processing; and querying ability.

In practice, there are some other design issues [98], such as Sensor
Network Deployment, Localization, Time Synchronization, Sleeping
Mechanism, Energy-Efficient Routing, Fusion/Aggregation, and Security.

1.3 Full Duplex

The enormous growth in wireless communication makes the efficient use of

the available and valuable resources a crucial need. However, most of the
wireless communication devices are operating in half-duplex mode, which
dissipates the precious resources by either employing time-division or
frequency-division. A full-duplex system allows communication at the
same time and frequency resources. Theoretically, operating in full duplex
will diminish the spectrum demand of a given communication system by
half.

1.3.1 Self-Interference Cancellation

The prime challenge in full duplex is to suppress the interference caused by

the node’s self transmission, as shown in Fig. 1.15. The self interference
can be 50–110 dB larger than the received signal, depending on the
internode distance. In Fig. 1.16, we can see that the desired received signal
is dominated by the self-interference. This causes further problems in
Analog-to-Digital Converters (ADC), since most dynamic range is for the
self-interference. Consequently, the effective bits for the desired received

www.TechnicalBooksPDF.com
signal is much smaller, and the resulting SNR is low. As a result, the self-
interference needs to be reduced before ADC in the analog circuit.

Figure 1.15 Full-duplex communication

Figure 1.16 Full-duplex self-interference

There are three major techniques for self-interference cancellation, as

follows:

1. Antenna Placement
In Fig. 1.17, we show examples of how to use multiple antennas to
achieve self-interference cancellation. The basic idea is that the signals
from the multiple transmitting antennas cancel with each other. The
drawback is that only narrowband signals are effective.

www.TechnicalBooksPDF.com
 Figure 1.17 Full duplex using multiple antenna

2. Active Analog Cancellation

As mentioned before, self-interference causes a significant problem for
ADC output SNR. To overcome this problem, an active analog
cancellation circuit (as illustrated in Fig. 1.18) has be implemented. The
main idea is to adaptively duplicate the propagation channel from the
transmit antenna to the receive antenna, and consequently the self-
interference can be removed.

Figure 1.18 Full duplex using active analog cancellation

3. Active Digital Cancellation

After analog self-interference cancellation, the residual interference can
be reduced by active digital cancellation. It can be seen as an adaptive
filter design procedure. A training sequence is needed to train the filter
tap weights so as to reduce the residual interference energy. After the
filter coefficients are obtained, the data are transmitted, and the self-
interference is minimized.

www.TechnicalBooksPDF.com
1.3.2 Current State of the Art

Reducing the self-interference level has recently been studied in the

literature. In [99], several interference mechanisms have been proposed
and, by experimental results, it has been shown that it is possible to
adequately reduce the self-interference level, which will make full duplex
attainable by achieving higher rates than those achieved by half duplex. In
[100], a three-node full-duplex network is studied and a decode-and-cancel
scheme proposed for improved interference cancellation, which is able to
deal with both self-interference at the full-duplex node and inter-node
interference at the other half-duplex nodes. In [101], two models for
loopback or self-interference cancellation have been proposed in full-
duplex wireless communication. In the first model, it is assumed that the
self-interference is random and unknown; while in the seconde model, it is
assumed that the self-interference is precisely known. In [102], an in-band
full-duplex WiFi radio design and implementation is presented, which is
capable of reducing the self-interference level to the receiver noise floor by
including new analog and digital cancellation techniques. In [103], a
balanced/unbalanced (Balun) transformer is used in signal inversion in
addition to adaptive cancellation in designing and implementing a full-
duplex radio that can support high-power and wideband systems.
Furthermore, investigating the effect of full duplex on resource allocation
problems have been underscored in many research works. In [104], the
optimal resource allocation schemes for both full-duplex and half-duplex
relay wireless networks have been proposed; in addition, dynamic resource
allocation strategies were developed under both full-duplex and half-duplex
modes to maximize the network throughput under a certain delay constraint.
In [105], optimum power allocation schemes subject to individual power
constraints have been analytically obtained for a full-duplex decode-and-
forward relay channel, and lower and upper bounds on the capacity of the
channel model were found; in addition, a systematic technique to solve
min–max problems in closed form was proposed. In [106], a power
allocation scheme for maximizing full-duplex wireless link throughput
given a certain delay constraint is proposed and analyzed. In addition, two
different models have been proposed to examine the full-duplex
transmission. In [107], optimal dynamic power allocation schemes have

www.TechnicalBooksPDF.com
been proposed which intend to maximize the sum rate of wireless full-
duplex bidirectional transmissions.

Furthermore, the increased number of users of wireless communication

has generated a continuous demand for new schemes that are able to
decrease traffic congestion. Accordingly, two-tier networks were introduced
as a solution that will assist in traffic offloading from the macro-base
stations to the small cell access points. In [108], several techniques for
traffic offloading are discussed, clarifying the merits and disadvantages of
each offloading technique. In [109], two traffic offloading techniques are
proposed, and it is proven that high energy gains can be achieved when
applying these algorithms in traffic offloading.

1.3.3 Key Application Scenarios

FD technology, with its key feature of enabling simultaneous transmission

and reception and its potential to double spectral efficiency, has ignited
great interest in both academia and industry. We introduce the following
three key application scenarios: FD bidirectional communications, FD
cooperative communications, and FD cognitive radio networks.

Fig. 1.19 depicts FD bidirectional communication between a pair of FD

transceivers. The two FD nodes transmit and receive data to and from each
other over the same frequency band simultaneously. The two nodes A and B
are equipped with and antennas, respectively. We use to denote
the transmission link from node A to B, and to denote the link from B
to A. Each node receives a combination of its desired signal from the other
node and SI from its own transmit antennas. In this scenario, if the SI can
be removed completely, then the spectral efficiency will be doubled by
using FD rather than HD communications.

www.TechnicalBooksPDF.com
 Figure 1.19 Full-duplex bidirectional communications

The next application of FD to wireless communication came in the

context of cooperative communications, illustrated in Fig. 1.20, which
generally consists of three types of node: a source node, a relay node, and a
destination node. The source node wants to transmit its data signal to the
destination node. The relay node is used to increase the coverage of the
source node, i.e., it receives, amplifies and retransmits the wireless signal.
Let us denote the transmission links from s to r as and r to d as ,
respectively. When FD transmission is applied to the relay node, however,
the source and relay nodes can simultaneously transmit their signals via
and in the same frequency band. Similar to the FD bidirectional
scenario, the FD cooperative communications can theoretically double the
spectral efficiency.

Figure 1.20 Full-duplex cooperative communications

Finally, we consider FD Cognitive Radio Networks (CRNs). In CRNs,

the Secondary (unlicensed) Users (SUs) are allowed to share the licensed
spectrum with the Primary (licensed) Users (PUs) as long as the SUs do not
cause harmful interference with the PUs. Spectrum sensing is essential in

www.TechnicalBooksPDF.com
identifying spectrum holes that can be utilized by SUs. Conventionally, SUs
need to periodically suspend their ongoing transmissions to perform
spectrum sensing, which leads to spectrum waste and tardy reaction to the
PU’s state change. When SUs have FD capability, they can perform
simultaneous sensing and transmission, as shown in Fig. 1.21, and thus to
bypass the intrinsic problems in conventional HD CRNs.

Figure 1.21 Full-duplex cognitive radio network

Apart from the aforementioned application scenarios, full-duplex

wireless can be applied to more communication scenarios, such as cellular
networks, ad hoc networks, local area networks, satellite communications,
vehicle communications, etc. Taking FD cellular network as an example,
the base station can operate in FD, then it can support simultaneous uplink
and downlink communications, potentially doubling the spectral efficiency.

1 A typical spacing is half the wavelength.

2 The LTE-WiFi link aggregation is beyond the scope of this book.

www.TechnicalBooksPDF.com
3 MuLTEfire does not fall within the scope of this book, but is still a good choice of traffic
offloading.

4 Supplemental Downlink: The unlicensed spectrum is used for downlink traffic only.

www.TechnicalBooksPDF.com
2 Signal Processing and Theoretical Limits

Mitigation of self-interference is the prime challenge in making full-duplex

technology feasible in wireless communications. However, self-interference
poses a difficult problem in general and is not limited to full-duplex
communications. A wireless transceiver may suffer from self-interference
even when transmit and receive frequencies are not the same, because
signal from the transmitted chains leaks into the receiver chain inside the
device. For this reason, 2G GSM handsets do not transmit and receive
during the same time slot, although uplink and downlink frequencies are
different, as in Frequency Division Duplex (FDD) systems. This is possible
because one frame in GSM is divided into eight time slots, each slot is used
by a different user, and uplink and downlink transmissions by one user take
place in different slots.

Limiting the number of available slots for communication naturally

limits the data rates of a user, and in the next generation this limitation was
relaxed. In 3G WCDMA, handsets are able to transmit and receive
simultaneously in FDD mode. This causes self-interference inside a device
when the transmitted signal leaks in the circuit board to the receiver chain.
As a matter of fact, 3G handsets require a duplex filter, making the
implementation more complex and expensive than that of GSM. Thus,
mitigation of self-interference is required also when using different
frequencies for transmission and reception in handsets.

On the base station side, co-siting of WCDMA-TDD and WCDMA-

FDD base stations operating on adjacent frequencies is not considered

www.TechnicalBooksPDF.com
commercially viable [110] due to the mutual interference. This has been
further observed in [111], concluding that 4G LTE-TDD adjacent band
interference is harmful to uplink LTE-FDD received signals when the LTE-
TDD network operates in the 1900–1920 MHz band and LTE-FDD above
the 1920 MHz band. Thus, self-interference is also a problem in base
stations when transmit and receive bands are different but close to each
other. Therefore, self-interference cancellation techniques developed for in-
band full-duplex communication can be useful for wireless transceivers in
general.

In this chapter, we first present a system model of a wireless

communication link employing a full-duplex transceiver. Then we present
signal processing techniques in digital baseband that aim to mitigate self-
interference. In addition to digital cancellation, other necessary techniques
to mitigate self-interference are antenna isolation between transmitter and
receiver antennas, and different analog cancellation techniques. As antenna
and Radio Frequency (RF) techniques are closely related to hardware
architecture, they are presented in more detail in Chapter 3.

2.1 System Modeling

Figure 2.1 depicts a generic communication link involving a full-duplex

multiantenna transceiver. When the source and the destination are the same
physical device, the scenario refers to bidirectional communication. When
source and destination are different devices, the full-duplex transceiver may
either relay the transmission from source to destination, or source and
destination are interpreted as different users in uplink and downlink and the
full-duplex access point serves both users simultaneously.

www.TechnicalBooksPDF.com
 Figure 2.1 System model of full-duplex communications

The source and destination have transmit and receive antennas,

respectively, and the transceiver is equipped with receive and
transmit antennas. In practice, full-duplex transceivers are likely to be
implemented with spatially separated receive and transmit arrays, which
yields physical isolation (cf. Section 2.3). This is not viable for handsets
due to their small form factor, but is more amenable when the transceiver is
part of a wireless infrastructure, i.e., a relay or an access point. However,
the following signal model is also applicable for full-duplex transmission
with a shared antenna array by setting and assuming high gain for
the self-interference channel .

A common practice is to build the signal model upon frequency-flat

block-fading channels. The model can be interpreted as a representation of
a single narrowband subcarrier within broadband orthogonal frequency
division multiplexing (OFDM) transmission. This model is viable when
subcarriers remain approximately orthogonal in the receiver. In OFDM, this
requires, i.e., that phase noise is small enough, carrier frequency offset has
been estimated and compensated, channel coherence time exceeds the
length of the OFDM symbol such that the channel does not change within
symbol, and the length of cyclic prefix (CP) exceeds the length of the
multipath channel.

Matrices , and represent combined

MIMO channels in digital baseband from all sources to the transceiver, and
from the transceiver to all destinations. All channels may vary between
transmitted symbols. At time instant n, the source transmits combined
signal vector , and the transceiver transmits signal vector
while it simultaneously receives signal vector . This
creates an unavoidable feedback loop from the transceiver output to the
transceiver input through channel .

Finally, the respective received signals in the transceiver and in a

destination node in digital baseband can be expressed as

www.TechnicalBooksPDF.com
 (2.1)

where and are additive noise vectors in the

transceiver and in the destination, respectively. The signal model and
Fig. 2.1 is purpose-built for full-duplex relays, but by reinterpreting the
channel matrices they apply to full-duplex access points and bidirectional
communications as well.

Transmit Signal Noise

The transceiver is obviously able to know the digital baseband signal
it has generated, but the corresponding transmitted analog signal is not
perfectly known any more, because signal conversion between baseband
and radio frequencies is prone to various distortion effects such as carrier
frequency offset, oscillator phase noise, digital-to-analog conversion
imperfections, I/Q imbalance, and high-power amplifier nonlinearity among
others. These distortion effects and their contribution to self-interference are
further elaborated in Section 2.2.2. In (2.1) signal represents this
analog signal in digital baseband after down-conversion and sampling in the
receiver.

Joint effect of RF impairments in digital baseband is often modeled as

additional noise source by introducing additive transmit distortion noise
vector based on which

(2.2)

where time index has been dropped to simplify notation. Herein the
elements of are modeled as independent identically-distributed (i.i.d.)
circularly symmetric complex Gaussian random variables. Consequently,
the covariance matrix of transmit signal noise can be expressed as

www.TechnicalBooksPDF.com
 (2.3)

in which and the variance of is defined with relative

distortion level . In addition, and are assumed to be uncorrelated,
which implies that . The level of transmit distortion is
usually well below the actual data signal in typical wireless communication
systems, i.e., , although it has a fundamental effect on full-duplex
transceivers. This is because self-interference signals can be much stronger
than the signal of interest, and consequently the distortion level may be
relatively large as well.

Analogous distortion effects occur also at the receiver side when

transforming radio-frequency signal into digital baseband signal .
Often, the joint effect of all receiver imperfections is implicitly included in
the conventional thermal noise vector and the amount of transmit noise
is given by Error-Vector Magnitude (EVM) value (2.4). Thus, one may
effectively denote .

Another approach is to model transmission imperfections in more detail

using behavioral models in digital baseband and presented in Section 2.2.2.
After imperfections have been modeled, their effect is included in the
replica of the self-interference signal and cancelled from received signal.
This reduces the power of but increases the complexity of signal
processing. This approach is studied in Section 3.4. If the imperfections are
not modeled in detail, transmitter distortion can be modeled by assuming
that the elements of are i.i.d. complex Gaussian random variables and
setting their powers to comply with realistic EVM levels.

In conformance tests of wireless transceivers, distortion in the

transmitted signal is typically quantified as Error-Vector Magnitude
(EVM) that sums up the effects of different impairments. The EVM is
experimentally measured by

www.TechnicalBooksPDF.com
 (2.4)

where refers to transmitted symbol sequence, denotes the

demodulated symbol sequence, and N is the number of samples.
Specifications define maximum EVM limits for certified devices depending
on modulation, because higher order modulation requires smaller EVM
levels. For example, in LTE, maximum allowed EVM levels for base station
transmitters are 17.5%, 12.5%, and 8% for QPSK, 16 QAM, and 64 QAM,
respectively [112], corresponding to – dB power levels for the
distortion. LTE small cell enhancement targeting home, local area and
medium-range base stations specifies even 256 QAM requiring better than
3.5% EVM level, i.e., better than dB. As a reference, high-end spectral
analyzers achieve EVM levels of the order of dB. Therefore, even if
thermal noise did not degrade the transmitted signal, EVM level in the
transmitter defines the maximum SNR that receiver can ever achieve. In
addition, EVM in the receiver degrades the received SNR even more.

When transmitter and interferers are located within comparable

distances from the receiver, EVM levels as specified in, e.g., LTE, are
tolerable. However, when the interfering signal is much stronger than the
signal of interest, as may happen within full-duplex transceivers, these
kinds of EVM levels may seriously degrade the performance of a full-
duplex transceiver.

Channel Estimation Error

The full-duplex transceiver may use any regular channel estimation
technique or techniques developed specifically for full-duplex transceivers
to obtain respective estimates and of and . Such processes
are non-ideal in practice, which can be modeled with additive error
coefficients and so that estimates differ from physical channels
as

www.TechnicalBooksPDF.com
 (2.5)

All elements of and are assumed to be mutually independent

circularly symmetric complex Gaussian random variables. In case of
minimum mean-square error (MMSE) channel estimation, the channel
estimates and the errors are uncorrelated conditioned on received signals
and pilot sequences. The variance of the estimation error components is
defined with relative error level such that

(2.6a)

for all and . Thus, under the assumptions,

, which means that the Frobenius norm of the error
matrix becomes a gamma random variable with shape and scale
. Analogous relations hold also between and , i.e.,

(2.6b)

for all and .

2.1.1 Non-Orthogonal Multicarrier Modulation

In case of OFDM, a narrowband subcarrier signal model is often enough;

signal processing can be performed on subcarrier bases, and the frequency
flat or single path signal model (2.1) is adequate. OFDM has its limitations,
though, and alternative waveforms have been actively studied that could
replace OFDM in wireless communication systems and obtain a better
performance–complexity trade-off than OFDM.

www.TechnicalBooksPDF.com
 Ideally, a waveform for wireless communications should have low
generation and detection complexity, limited time–frequency overhead,
good localization in time, good spectral containment, a straightforward
extension to MIMO, and robustness to hardware impairments. Pure OFDM
signal using a rectangular window in the time domain (OFDM signal
without filtering) is well-localized in time but consequently spread in
frequency the attenuation of the first sidelobe being only dB. This
would cause severe co-channel interference to an unsynchronized
communication link in an adjacent band, and in practice filtering or
windowing of the transmitted OFDM signal is used to limit the sidelobes.
Furthermore, overhead from cyclic prefix in OFDM may become too large
when the number of subcarriers is small. This may be typical in, e.g., a
machine-type communications scenario.

These features of OFDM have sparked several research efforts to

develop improved waveforms for 5G systems. Several different non-
orthogonal multicarrier techniques have been promoted as alternatives to
OFDM for 5G systems, like Filter Bank Multi-Carrier (FBMC) [113] and
Generalized Frequency Division Multiplexing (GFDM) [114], and
properties of waveforms have been compared and their suitability to 5G
have been evaluated, e.g., in [115, 116]; see also EU projects PHYDYAS
and 5GNOW.1

The promise of non-orthogonal multicarrier waveforms is to relax the

overhead due to cyclic prefix and facilitate better localization in the
frequency domain than OFDM waveform. This is achieved by sophisticated
filter design that controls the distribution of energy in time and frequency,
and good localization in frequency is achieved by extending and
overlapping the multicarrier symbols in time. Thus, multicarrier modulation
schemes trade off time and frequency localization, and this trade-off is
controlled by prototype filter design. Eventually, this trade-off in time and
frequency cannot be avoided because of the well-known Heisenberg
Uncertainty Principle from quantum mechanics [117]. Exploiting the trade-
off requires more complicated signal processing than within OFDM and
mitigation of the interference from neighboring subcarriers in the receiver.

www.TechnicalBooksPDF.com
 Several comparisons between OFDM and FBMC have been presented
in the literature, with a common conclusion: nominally, FBMC outperforms
OFDM in terms of spectral containment and system overhead in link level.
At the same time it is significantly more complex due to the usage of long
prototype filters and equalization in the receiver. It remains to be seen
whether performance vs. complexity is sufficient that FBMC or other non-
orthogonal multicarrier waveforms find their way into future wireless
systems.

Generalized Frequency Division Multiplexing (GFDM) is another

multicarrier waveform [118] that has been proposed for 5G scenarios like
machine type communication and cognitive radio. In contrast to FBMC,
GFDM uses Cyclic Prefix (CP) for better synchronization and improved
channel robustness [119]. In the following, we use GFDM to illustrate the
effect of non-orthogonal modulation formats on self-interference in full-
duplex transceivers [120]. Similar treatment applies to FBMC as well. All
non-orthogonal waveforms require more involved processing for
interference cancellation in the frequency domain than OFDM due to
inherent inter-carrier interference.

Let us consider a single transmitter and a single receiver antenna to

simplify notation, and suppose that the number of subsymbols and
subcarriers in a single block are denoted by M and N, respectively. The
representation of a transmitted GFDM signal in the time domain before
adding CP is given by

(2.7)

where is an mth subsymbol vector, and is a data symbol

vector, W is an inverse Discrete Fourier Transform (DFT) matrix,
and P is an M-fold expander that consists of M identity matrices as

www.TechnicalBooksPDF.com
 (2.8)

where is a identity matrix and is the transpose operator. A

diagonal matrix of size represents a time-domain real-valued
window function at mth subsymbol time, which corresponds to filtering in
the frequency domain. The diagonal part of is a circular shift of the
diagonal part of by mN symbols. Transmission pulse response at mth
subsymbol time is denoted by . After adding a CP, the GFDM
block in (2.7) is transmitted through a wireless channel.

Denote the received signal vector after removing CP by . The

frequency-domain signal vector at the receiver for subsymbol
is reconstructed by passing the received signal through
GFDM demodulator

(2.9)

Finally, transmitted symbols are decoded by applying a channel equalizer

to .

The reconstructed signal inherently contains interference due to the non-

orthogonal signaling. Figure 7 in [119] shows the pattern of inherent
interference composed of inter-subsymbol interference (ISI), inter-carrier
interference (ICI) from the same time slot’s subsymbols, and ICI from the
other time slot’s subsymbols. Thus, creating a replica of self-interference is
not straightforward but requires taking into account the inherent
interference components.

The model of the received signal on subcarrier level can be rewritten as

(2.10)

www.TechnicalBooksPDF.com
where and now refer to channel convolution matrices of the signal
of interest and self-interference, respectively, in a SISO system.
Furthermore, the vectors for refer to self-interference and
desired signal for subsymbol m, respectively, and models additional
noise. When the channel delay is shorter than the CP length, the channel
matrix is circulant. If the channel delay is longer than the CP length, the
channel matrix is not circulant any more and the interference from the
previous GFDM block occurs, similar to OFDM.

2.1.2 Time-Domain Modeling

Time domain interference cancellation is agnostic to the modulation

technique used, and it can be applied to OFDM, FBMC, and GFDM alike.
Thus, it applies also to the case when the transmitted signal and the signal
of interest use different modulations. This happens, for example, when a
full-duplex access point is operating on LTE signals such that the signal of
interest (LTE uplink signal) obeys single carrier modulation while the
transmitted signal (LTE downlink signal) and self-interference signal use
OFDM. The frequency flat channel model is not valid in general, and
multipath channel should be included in the system model instead. In the
received signal model, multiplications in (2.1) are substituted for
convolutions and the received signal in the full-duplex transceiver becomes

(2.11)

(2.12)

where refers to convolution. As before, the dimensions of the channel

matrices are , and , and and refer
to the order of the channels. Thus, the number of multipath components is
and , respectively. The signal model can be further cast in
the form

www.TechnicalBooksPDF.com
where is th-order column expansion of
, and are column expansions of and , respectively, and
and refer to row-diagonal-expanded matrices of
and . For simplicity and without losing generality, we assume
that . The latter formulation of the multipath received signal
makes it possible to apply standard linear algebra tools when developing
interference cancellation techniques, as shown in Section 2.4.6.

2.1.3 Modeling of Self-Interference Channel

In a realistic environment, transmitted signals from a full-duplex transceiver

may be reflected back to the receiver chain through different paths which
are each characterized by an attenuation and a delay. Different signal paths
are characterized by different attenuations and phase shift giving rise to a
frequency-selective channel, in general. In the following we present
measurement results of physical isolation between transmitter and receiver
antennas at 2.6 GHz carrier frequency and 100 MHz bandwidth.

Early channel measurement results on full-duplex MIMO transmission

have been presented, e.g., in [121, 122]. The study reported in [121]
compares channel gain with and without repeaters in coverage-extension
and outdoor-to-indoor forwarding scenarios. The self-interference
phenomenon is not observed in [121], though, probably because
measurement scenarios guarantee large passive physical isolation and
repeater gain is set to a low level. The study reported in [122] builds the
(perhaps) first-ever prototypes of antenna arrays to be used as a full-duplex
MIMO transceiver and concentrates explicitly on measuring loopback
channels and physical isolation achieved with the prototypes.

Measurement Data on MIMO Self-Interference Channels

The experimental antenna arrays built in [122] are targeted for full-duplex
outdoor-to-indoor transmission at 2.6 GHz band. The main design criteria
were to implement a real MIMO transceiver with multiple mutually

www.TechnicalBooksPDF.com
uncoupled receive and transmit feeds, to support at least 100 MHz
bandwidth, and to achieve compact size but with as high isolation as
possible. As a validity check, the performance of the prototype antenna
arrays was measured in an anechoic chamber.

A schematic of the relay station is depicted in Fig. 2.2. The antenna

arrays comprise two dual-polarized square ( mm) patch antennas
placed over a rectangular ( mm) ground plate. They are identical
except that transmit polarizations are slanted 45 degrees with regard to
receive polarizations to slightly improve physical isolation. A more detailed
description and photo of the antenna arrays can be found in [122]: there are
four antenna feeds in each array. The mutual coupling between the feeds is
low enough to allow the forming of four orthogonal spatial beams at both
receive and transmit sides. Since the transmitter and receiver antennas are
on the opposite sides of the box, the natural use case for the design is a
relay providing wireless backhaul to a donor cell on one side and serving
user equipment on the other side.

www.TechnicalBooksPDF.com
 Figure 2.2 Overall view of full-duplex relay

The experimental data covers measurements for two antenna array

configurations from [122] and a set of simulated channel samples as
follows:

In compact array configurations the arrays are attached side by side

(while facing in opposite directions) using plastic spacers. The
distance between the arrays is only mm. By placing the
transceiver electronics between the arrays, the complete transceiver
device could be integrated into the box that is about the same size as
typical WLAN routers. The samples of the channel are measured
by deploying the transceiver at several locations in a meeting room
such that the receive array points outdoors (toward an imaginary base
station) through windows and the transmit array provides indoor
coverage.
In separate array configurations the receive array is deployed on the
outer surface of a window pointing outwards to an imaginary outdoor

www.TechnicalBooksPDF.com
 base station, and the transmit array is deployed indoors. The distance
between the arrays, , varies roughly from 20 centimeters to 12
meters in the experiment. The samples of the channel are
measured by deploying the transmit array at various locations in the
same meeting room and in the adjacent corridor rendering LOS and
NLOS channels, respectively. The measurements include four
horizontal array orientations: In Orientations 1 and 2 the transmitter
array is perpendicular to the receiver array. Orientation 3 points the
main beams of the transmit antenna elements toward the receive
array, and Orientation 4 points the two arrays in opposite directions.

Physical Isolation
Given the covariance matrix of the transceiver transmit signal and the
transceiver transmit power , the power of the self-interference
term becomes

(2.13)

assuming that the transmitted signal is spatially white, i.e., .

Consequently, the physical isolation is given by the ratio of the transmitted
power to the received interference power per antenna:

(2.14)

Channel fluctuation over frequency bins and different transceiver positions

makes a random variable.

www.TechnicalBooksPDF.com
 Figure 2.3 illustrates the average physical isolation as a function of the
distance between the arrays in the separate array configuration. Thus, these
results reveal the additional propagation loss that can be achieved by
moving arrays further away from each other. In fact, the average isolation
closely follows a linear trend in decibel scale as a function of the distance
between the arrays: the isolation is improved roughly 2–3 dB per meter.

Figure 2.3 Average physical isolation in terms of the distance from the transmit antenna array to
the receive antenna array in the separate array configuration. The lines show linear least-squares fit
to the measurement results.

The isolation of non-line-of-sight (NLOS) channels fluctuates slightly

more than that of line-of-sight (LOS) channels, but there is no statistical
difference in the average isolation. On the other hand, the transmit antenna
orientation affects significantly the average physical isolation: Orientations
2 and 3 result in around 10 dB lower isolation than Orientations 1 and 4

www.TechnicalBooksPDF.com
with the same distance between the arrays. This can be expected since the
antenna elements are not omnidirectional.

By extrapolating the average physical isolation of Orientation 4 with

LOS channels down to 20 mm antenna separation, roughly 68 dB isolation
is observed with the separate array configuration when it is made equivalent
to the compact array configuration that achieves only 48 dB isolation. This
additional 20 dB isolation originates partly from the coated glass
windowpanes which attenuate the propagation path between the outdoor
receive array and indoor transmit array. However, the two array
configurations are measured in different buildings, which implies that the
differences of multipath richness in the propagation environment may be
another significant reason for the higher isolation.

Overall, empirical results suggest that mere physical isolation may be

insufficient for high-rate full-duplex transmission with high power and long
range (especially with compact arrays), even if it is a necessary prerequisite
to avoid receiver saturation. This gives general motivation to develop active
analog and digital schemes for additional isolation.

Like physical isolation, the self-interference channel can be

characterized as a random variable. Measurement campaigns typically
conduct several measurements in different environments giving rise to
different self-interference channels. Even when the full-duplex transceiver
does not move, the signal paths change depending on the location of the
device under test. Self-interference observed in digital baseband is the one
remaining after physical isolation and active cancellation. It is a random
variable that can be modeled, e.g., as Rayleigh, Additive White Gaussian
Noise (AWGN), or Rician distribution. However, it is difficult to give a
definitive answer about the statistics of the residual self-interference,
because it depends on passive isolation and active RF cancellation that vary
depending on the implementation.

Measurement results using the antenna design in Fig. 2.2 and presented
in [123] showed that physical isolation has a larger spread than an

www.TechnicalBooksPDF.com
exponential random variable. Since there was no interference cancellation
besides antenna isolation, the physical isolation is the same as the power of
self-interference. After analog cancellation, the distribution of the
remaining self-interference would change again. It was shown in [124] that
analog cancellation changes the K-factor in Rician fading. The
measurement campaign assumed a LOS path between transmitter and
receiver antennas, and when the strongest (LOS) path of the self-
interference was cancelled, the multipath channel profile was changed
accordingly. Since the strongest path is largely suppressed, the K-factor
becomes smaller.

When analyzing the performance of full-duplex systems on digital

baseband it is often assumed that the known part of the self-interference has
been removed. Furthermore, if Channel State Information (CSI) is assumed
to be perfect, in (2.5) becomes zero and the residual self-interference
can be assumed as Gaussian. Perfect CSI is a common assumption
when calculating mutual information, because capacity is in general
unknown when assuming imperfect CSI and mismatched decoding [125].
Even when CSI is assumed perfect, generalized mutual information can be
used to derive lower bounds for capacity in full-duplex communications
[126]. In this case, the mismatched decoding is due to unknown distortions
in the transmitted signal. This distortion is analyzed in more detail in the
next section.

2.2 Modeling of Self-Interference

In the signal model at digital baseband (2.1) self-interference was modeled

as the sum of transmitted information signal and RF distortion. The latter
was assumed to comprise a single noise term in the transceiver input.
Assuming that the transmitted signal in digital baseband is known and
channel estimation is ideal, the known information signal can be subtracted
from the received signal. The unknown noise part due to RF distortion
remains, and the components that make up the residual self-interference are
elaborated in this section. When the RF distortion is modeled in detail in

www.TechnicalBooksPDF.com
digital baseband, its estimate can be subtracted from the received signal
decreasing the power of the remaining self-interference.

Modern wireless transceivers use predominantly direct-conversion

architecture, depicted in Fig. 2.4. Compared to superheterodyne
transceivers, Intermediate-Frequency (IF) stages, which are typically off-
chip components, are replaced by baseband components. Furthermore, the
RF Image Rejection (IR) filter becomes unnecessary. This makes it possible
to integrate the whole receiver or transmitter (with the exception of the
power amplifier) on a single chip, using cheap integrated circuit technology.

Figure 2.4 Direct conversion architecture

Following the signal flow in Fig. 2.4, the received signal is first
bandpass filtered, amplified with a Low-Noise Amplifier (LNA), and down-
converted directly to the baseband and sampled. The in-phase (I) and
quadrature (Q) components of the complex signal are then separately low-
pass filtered, amplified and sampled. Signal level before the Power
Amplifier (PA) and Analog-to-Digital Converter (ADC) is adjusted with a
Variable Gain Amplifier (VGA).

www.TechnicalBooksPDF.com
 The direct-conversion architecture is simple, but it suffers from some
characteristic impairments. The main problems are Direct-Current (DC)
offset, I/Q imbalance and second-order intermodulation. In addition, power
amplifier nonlinearity is the main source of the distortion in the transmitted
signal. Together these different components contribute to the transmit
distortion term .

2.2.1 Dynamic Range

Typically, transmitter antennas in full-duplex transceivers are much closer

to the receiver antennas than the transmitter antennas of the source signal.
Thus, without physical isolation and cancellation, self-interference would
completely saturate the front end of the receiver. If the analog cancellation
signal is injected into the receiver chain after a Low-Noise Amplifier
(LNA), the self-interference signal should fit to the dynamic range of the
LNA. Alternatively, the cancellation signal can be injected before LNA,
which relaxes the requirements of the dynamic rage of the LNA, but
increases the noise figure of the receiver.

Figure 2.5 sketches the requirements of self-interference cancellation in

a full-duplex-capable transceiver. To detect the signal of interest, its level
has to be larger than the receiver noise floor given by

where B refers to signal bandwidth and is receiver noise figure.

Depending on the modulation and coding in use, the receiver sensitivity
in dBm is given by

(2.15)

where SNR refers to the required SNR in the input of the detector given the
selected modulation and coding scheme.

www.TechnicalBooksPDF.com
 Figure 2.5 Power budget for a full-duplex transceiver

The noise figure is caused by receiver electronics and describes the

degradation of the signal-to-noise ratio in a real receiver when compared to
an ideal receiver. The noise figure is given by where the
noise factor is given by Friis’ formula [127]. In the direct conversion
architecture depicted in Fig. 2.4, the noise factor becomes

(2.16)

where , , and refer to noise factors of LNA, IQ mixer and

VGA, respectively. Linear gains of LNA and IQ mixer are denoted by
and . This assumes that the Band-Pass Filter (BPF) is passive and
linear component has 0 dB gain so that the distortion can be ignored. The
contribution of these different components to the received signal is
presented in more detail in Section 2.2.3.

Suppose that the transmitter power is 23 dBm, corresponding to the

transmitter power of actual user equipment in LTE. Typical assumptions for
the noise figure , e.g., in LTE systems, are 5 dB in base stations and 7 dB
in handsets. For example, assuming 10 MHz signal bandwidth and a 7 dB
noise figure, the noise floor becomes dBm and there is 120 dB
difference between the noise floor and the transmitted signal level within
the same full-duplex transceiver. The goal of the self-interference
cancellation is to reduce the interference close to the noise floor, preferably

www.TechnicalBooksPDF.com
even below it. Therefore, self-interference should be suppressed by 120 dB,
which is far from a trivial task. If the full-duplex transceiver is a handset,
the difficulty factor increases even more, because spatial isolation between
transmit and receive antennas is limited due to the small form factor of the
device.

In general, cancellation of self-interference requires a combination of

three complementary techniques: passive isolation, Radio Frequency (RF)
or analog cancellation, and digital cancellation in baseband. The first two
steps should attenuate the interference as much as possible, and at least the
signal of interest and the self-interference should fit together to the dynamic
range of the ADC. If self-interference saturates the ADC, the signal of
interest is lost and self-interference cancellation in digital baseband is not
able to recover it. The dynamic range of the ADC is typically smaller than
that of the LNA and therefore it presents the limiting factor in the receiver
chain.

The dynamic range of the ADC is given by

(2.17)

where b is the number of bits in the ADC and the last term refers to the
estimated Peak-to-Average Power Ratio (PAPR) of the input signal, be it
due to the self-interference or the signal of interest. The equation assumes
that the input signal is not clipped, and additional headroom has been
reserved to accommodate signal peaks. Typically one bit is reserved for
this, although PAPR depends on the modulation and waveform used. In
case of OFDM signal that can be assumed Gaussian when the number of
subcarriers is large, optimized dynamic range in terms of quantization range
and clipping noise can be approximated as [128]

www.TechnicalBooksPDF.com
 In addition to reserving headroom to avoid excessive clipping, another
quantization bit is typically reserved to avoid the system from being
quantization limited but noise limited. For example, if the ADC has 14 bits,
the effective number of bits being 11, the dynamic range of the ADC
becomes 54 dB. Therefore, passive isolation together with RF cancellation
should attenuate the transmitted signal at least by 66 dB when assuming
120 dB difference between receiver noise floor and the transmit power.

In case of an LTE base station with 46 dBm transmit power, passive

isolation and RF cancellation should provide up to 89 dB attenuation of
self-interference. Therefore, full-duplex operation is more promising in
local area networks and small cells, where the transmit powers are limited
when compared to macro cellular networks. On the other hand, in local area
networks, nodes may be close to each other, and co-channel interference
may already start to dominate over self-interference. This far, 70–100 dB
overall attenuation levels of self-interference have been reported in the
literature [102, 124, 129, 130].

2.2.2 RF Impairments

Once the self-interference and signal of interest fit the dynamic range of the
ADC, the known part of the self-interfence can be subtracted from the
received signal. Conversion of the digital baseband signal to RF signal in
the transmitter is subject to several impairments due to RF components that
make up the unknown part of self-interference, which was modeled as
additive noise component . The effect of the distortion can be reduced
by better hardware, compensating for the distortion in the transmitter,
modeling the sources of the distortion in the cancellation process, or by
spatial suppression. In the following, the sources of the distortion are
described in more detail.

DC-Offset
Direct current-offset arises when the Local Oscillator (LO) signal leaks
from its mixer input to the second mixer input. The LO signal mixes with
itself, effectively down-converting itself to DC [131]. Even worse, the LO
signal can leak to the input of the amplifier prior to the mixer so that the

www.TechnicalBooksPDF.com
leaked LO signal is amplified. DC offset can also occur when the amplified
received signal leaks from its mixer port to the mixer port of the LO signal.
The resulting self-mixing product is the signal squared, which results in a
time-varying signal-dependent low-frequency signal with a strong DC
component.

Static DC offset is easy to estimate and remove from the received

signal, and cancellation is usually done already before the ADC to avoid
ADC saturation, usually by high-pass or notch filtering. In digital baseband,
remaining static DC offset is simply removed by subtracting the mean of
the signal from the signal itself. Therefore, performance studies usually
assume that there is no DC offset in the signal.

IQ imbalance
Direct conversion works perfectly when In-phase (I) and Quadrature (Q)
branches have 90-degree phase shift and the frequency responses of the two
branches are the same. However, this is not the case in practice, and
amplitude and phase mismatch cause IQ imbalance in up-converted and
down-converted analog signals. Due to the mismatch, the signals from
negative (positive) frequencies are mixed to the baseband on top of the
desired signal. The signals do not combine constructively, and they interfere
with each other.

Imbalance imposes a correlation between I and Q branches that would

otherwise be uncorrelated. Assuming that the transmitted signal is random,
this makes the signal a non-proper or non-circular complex random signal.
A random vector is proper if its pseudo-covariance matrix
. Thus, for a complex random scalar x to be
proper, its in-phase and quadrature components must be uncorrelated and
have the same variance. This is not the case in the presence of I/Q
imbalance. Depending on the LTE release, the specification limits the
maximum in-band image component to 25–28 dB below the payload signal
for mobile handset transmitters. Such attenuation is sufficient when the
signal is transmitted to a destination far away. In case of full-duplex
transceivers, such levels give rise to significant residual self-interference.

www.TechnicalBooksPDF.com
Phase Noise
In addition to imperfections characteristic of a direct-conversion
architecture, wireless transmitter and receiver chains are subject to other
kinds of distortions like carrier frequency offset, phase offset, phase noise,
quantization noise, sampling jitter, and symbol synchronization error, to
name a few. Phase noise is the most complex of these, caused by the
imperfections of the oscillator. The effect of phase noise for full-duplex
transceivers has been studied, e.g., in [132–134].

In the block diagram of the transmitter in Fig. 2.4, the transmitter and
receiver share the same oscillator. This kind of architecture is preferred
because using the same oscillator in transmitter and receiver chains saves in
the necessary quantity of hardware components when compared to using
separate oscillators for transmitter and receiver. In addition, using the same
oscillator decreases the contribution of the phase noise in the self-
interference signal.

Up-conversion of the baseband signal to carrier frequency is performed

by mixing the baseband signal with a carrier. At the receiver, down-
conversion uses the signal

where refers to the carrier frequency and denotes the phase noise
process. The total phase noise distortion in the self-interference signal is
given by

where τ refers to the round-trip delay from transceiver output to its input. In
this case there is one shared oscillator and the effect of the
phase noise is negligible, because τ is in nanosecond scale for full-duplex
transceivers with a small form factor. With two independent oscillators, the
phase noise term would be a sum of two independent phase noise processes,

www.TechnicalBooksPDF.com
which would limit the performance more when compared to the shared
oscillator.

Quantization Noise
Uniform distribution is generally considered a good model for quantization
noise [135]. Assuming that the signal is not saturated and its dynamic range
equals , a uniform quantizer with b bits partitions the signal in
levels, and the width of each level becomes . Uniformly
distributed quantization error is bounded in and the variance of
the error is given by . However, for analysis purposes it is often
more convenient to model the quantization noise as Gaussian
distributed [136, 137]. This has been justified by assuming that, in addition
to quantization, the transceiver is subject to several other impairments,
making their joint contribution Gaussian.

Power Amplifier Nonlinearity

Most severe nonlinear effects in the transmitter are caused by a Power
Amplifier (PA) after the RF mixer. Power amplifiers should be operated
close to the nonlinear region, because power efficiency and linearity are
contradicting requirements in power amplifier design. Power efficiency
measures a PA’s capability of converting the power of the DC supply to RF
power, and it is considered an important design parameter in wireless
communication systems.

The trade-off between PA efficiency and linearity is determined by

power class. For example, a class-A power amplifier is highly linear, but its
power efficiency is only 50% [138]. In contrast, the efficiency of a class-C
amplifier reaches 85% but its linearity is poor. In addition to in-band
distortion causing self-interference in full-duplex transceivers, nonlinear
distortion causes spectral regrowth and therefore adjacent channel
interference. However, outband distortion is the same whether the
transceiver operates in half-duplex or in full-duplex mode. The only
exception to this is full-duplex amplify-and-forward relay. It amplifies the
residual nonlinear distortion after self-interference cancellation. Therefore,
the transmitted signal is different when compared to half-duplex amplify-
and-forward relay.

www.TechnicalBooksPDF.com
 Classical figures of merit to measure the linearity of the PA are the 1 dB
compression point and the third-order intercept point (IP3). Typically, the
gain of a PA decreases as the input level increases, producing a compressive
AMplitude-to-AMplitude (AM/AM) response. In a similar way, the phase
conversion at the output does not stay constant when the input amplitude
changes. The 1 dB compression point is defined as the output level at which
the gain of the actual PA output is 1 dB below the output if the PA were
linear.

The IP3 is the power level at the interception between the linear gain of
the fundamental signal and linear gain of the third-order component.
Among intermodulation products, the third-order products are usually the
strongest ones. The intercept point is a purely mathematical concept to
characterize the behavior of PAs, and does not correspond to physical
power levels. Often, the intercept point lies beyond the damage threshold of
the PA.

These PA characteristics can be used to fit a behavioral model to capture

the operation of a real power amplifier. Behavioral models can then be used
to develop pre-distortion and post-distortion algorithms that aim to mitigate
the nonlinear effects in digital baseband, i.e., to reduce the power of .

Depending on the signal bandwidth, the characteristics of a PA can be

categorized into a memoryless nonlinear system or a nonlinear system with
memory. The envelope characteristics of the memoryless PA are constant
over the band of interest, and the PA responds to an excitation
instantaneously without any delay. A strictly memoryless nonlinear PA
exhibits only AM/AM distortion, and ways to capture memoryless PA
distortion have been presented in the literature. Memoryless PA models
include, e.g., polynomial models [139], orthogonal polynomials [140] and
Saleh model [141]. However, memoryless assumption is only an
approximation, and in practice these kinds of PAs do not exist.

A memoryless PA model is useful, though, for analyzing the behavior

under OFDM in case of a nonlinear PA. If the number of subcarriers in

www.TechnicalBooksPDF.com
OFDM is sufficiently large, the transmitted signal becomes Gaussian
and the PA output signal can be modeled according to the extended version
of the Bussgang theorem [142] as

(2.18)

where is the nonlinear response of the PA and is a constant.

Coefficient scales the up-converted signal , whereas is an
additive distortion term. When the PA operates sufficiently close to its
saturation level, the distortion noise term becomes approximately
Gaussian.

Memoryless solid-state power amplifier (SSPA) model describing the

nonlinear PA response is given by [141]

(2.19)

where is the saturation voltage of the PA, p is a model parameter that

adjusts the smoothness of the transition from the linear region to the
saturation region and is the Input Back-Off (IBO). Then, we can calculate
the parameters in the equivalent linear model (2.18) as [143]

(2.20)

Thus, given the parameters of the PA model, it is possible to relate

nonlinear distortion to additional Gaussian noise in the linear model. This
simplifies the analysis of PA distortion effects significantly. The approach
can be used, e.g., to determine optimum IBO for full-duplex amplify-and-
forward relays [144].

www.TechnicalBooksPDF.com
 A large class of narrowband systems fall into the quasi-static category.
The memory time constant of a quasi-memoryless nonlinear PA is of the
order of the RF carrier period giving rise to short-term memory effects.
Quasi-memoryless PAs are characterized by both AM/AM and AM/PM
conversions. When the signal bandwidth is comparable to the carrier
frequency of the system, nonlinear PAs exhibit memory effects which are
manifested as frequency-dependent or bandwidth-dependent nonlinear
responses. Although memory effects normally do not cause huge changes to
the PA nonlinearity, e.g., to the level of inter-modulation distortion
components, they affect the amount of self-interference in full-duplex
transceivers. Wideband PAs have been modeled, e.g., using Volterra series
[145], memory polynomial [146], piecewise linear functions [147], and
Wiener and Hammerstein models that divide the wideband PA model into
static nonlinear blocks and dynamic linear blocks [148].

2.2.3 System Calculations

Having characterized various sources of RF distortions, we now move on to

quantify the effect of various imperfections following the approach
in [149]. The aim is to find out what the most significant distortion
components in the self-interference signal are. Later on, in Section 3.4, we
study how these features can be used to develop behavioral models and
improve the performance of self-interference cancellation in digital
baseband.

For simplicity, we assume that distortion caused by IQ mixer and VGA

in the direct-conversion architecture 2.4 are omitted from the analysis,
because their effect is typically small compared to distortion by PA. The
signal at the input of the PA in the transmitter can be written as ,
where is the combined gain of the IQ mixer and VGA. The PA
distorts the signal in a nonlinear fashion, and in the time domain its output
can be written as

(2.21)

www.TechnicalBooksPDF.com
where is the power gain of the PA, and represents the nonlinear
distortion produced by the PA. The received signal in the full-duplex
transceiver (without cancelling self-interference) can be written as

(2.22)

where refers to the self-interference channel, denotes convolution,

and represent receiver noise. Thereafter, the signal propagates through
the receiver chain, where it is amplified and distorted again in nonlinear
fashion. Modeling the nonlinear distortion again in an additive form, the
signal at the input of the ADC can be expressed as

(2.23)

where is the gain of the receiver chain and is thermal noise.

Second-order and third-order nonlinear distortions produced by the receiver
chain are denoted by and , respectively. Additional noise produced
in the receiver chain is characterized by the receiver noise figure.
SINR in the input of the quantizer is given by

(2.24)

where is the power of the signal of interest, is the receiver noise

factor (2.16), and is the transmitted power. The effect of interference
cancellation is not yet included in the expression. The quantities , , ,
refer to powers of thermal noise, PA nonlinearities and second-order and
third-order nonlinearities. The equation quantifies the useful signal power
with regard to noise and interference power entering the ADC. With a fixed
ADC voltage range, and assuming that the receiver gain is properly
controlled, the total input power in the ADC is the sum of the nominator

www.TechnicalBooksPDF.com
and the denominator. It is assumed that the total power is always matched to
the maximum allowed power in ADC input.

After quantization and sampling with rate , the signal in the input of
the detector is given by

(2.25)

where refers to the quantization noise term. Corresponding Signal-to-

Interference-and-Noise Ratio (SINR) expression becomes

(2.26)

where refers to power of the quantization noise calculated according

to (2.17).

It is necessary to know the receiver gain to assess the powers of

individual signal components. Assuming that the level of the self-
interference signal is much higher than the signal of interest and taking into
account interference cancellation mechanism, the receiver gain can be
approximated as

(2.27)

Now, interference cancellation is included in the gain through , and

that refer to attenuation of the signal power by passive isolation and active
RF cancellation, respectively. Plugging the attenuation values into the SINR
expression (2.26),

www.TechnicalBooksPDF.com
 (2.28)

where refers to attenuation due to digital cancellation. This assumes

linear digital interference cancellation, because the attenuation does not
affect PA third-order nonlinearity .

The second-order and third-order distortion terms in (2.28) are

characterized by using their Input Intercept Points (IIP). Thus, total powers
of the second-order and third-order distortion terms in the receiver gain can
be approximated by [149]

where all parameters are in linear scale, and , , and refer to

gains of LNA, mixer, and GCA, respectively. Furthermore, the power of the
PA-induced nonlinear distortion in the receiver, without taking into account
any cancellation mechanism, is given by

(2.29)

Now the power of different signal components can be calculated, and it

remains to assign values for different noise sources to quantify the level of
self-interference in full-duplex receiver. Table 2.1 gives an example of
system parameters used to illustrate the contribution of nonlinearities to
self-interference. The parameters are chosen to represent a state-of-the-art
wideband transceiver [149]. With the selected parameters, the sensitivity
level can be calculated as –88.9 dBm; that is close to the reference
sensitivity levels in LTE specifications. Passive isolation due to antenna

www.TechnicalBooksPDF.com
design is set to 50 dB, which approximately corresponds to the isolation
achieved with the compact antenna array design in Section 2.1.3. Analog
and digital cancellation capabilities are set to 30 dB and 35 dBm,
respectively. These kinds of figures represent a good cancellation
performance, but such numbers have been demonstrated in the literature.

Table 2.1 Parameters for full-duplex transceiver

SNR Requirement 10 dB
Bandwidth 12.5. MHz
Receiver noise figure 4.1 dB
Sensitivity dBm
Received signal power dBm
Passive isolation 50 dB
RF cancellation 30 dB
Digital cancellation 35 dB
ADC bits 8
ADC voltage range 4.5 V
PAPR 10 dB
Allowed SINR loss 3 dB

The components of the direct-conversion receiver are given in Table 2.2

together with the total IIP2 and IIP3 values calculated using the component
values. The components values are selected with the aim to model a typical
receiver chain. The ADC input is controlled by VGA such that the full
voltage range is properly used.

Table 2.2 Parameters of the receiver components

Component Gain [dB] IIP2 [dBm] IIP3 [dBm] NF [dB]
BPF 0 – – 0
LNA 25 43 –9 4.1
Mixer 6 42 15 4
LPF 0 – – 0
VGA 0–69 43 14 4

The transmitter components are given in Table 2.3, again representing

values of a typical transmitter chain. Assuming that the power of the

www.TechnicalBooksPDF.com
feeding amplifier input signal is approximately 35 dBm, the power of the
transmitted signal varies between –8 dBm and 27 dBm.

Table 2.3 Parameters of the transmitter components

Component Gain [dB] IIP2 [dBm] IIP3 [dBm] NF [dB]
PA 27 – 20 5
Mixer 5 – 5 9
LPF 0 – – 0
VGA 0–35 – 5 10

The power of the different signal components with these parameters is

depicted in Fig. 2.6 as a function of varying transmitted power. The
received signal of interest is assumed to be only slightly, i.e., 5 dB, above
the receiver sensitivity level, because this is the most challenging case from
the self-interference cancellation point of view. The power of the signal
entering the ADC is kept approximately constant using an automatic gain
controller. Thus, the total gain of the receiver chain reduces when the
transmission power increases. Maximum allowed SINR loss is set to 3 dB,
meaning that the performance loss due to full-duplex operation becomes too
high when the total noise-plus-interference power more than doubles.
Signal-to-noise requirement in Table 2.1 refers to the SNR requirement in
(2.15) that depends on the modulation and coding scheme in use. Maximum
transmitted power, marked with the vertical line, is 13.5 dBm given the
allowed SINR loss. Thus, under these parameters, the transceiver is not able
to use its full power. As expected, the contribution from second-order and
third-order nonlinearities increases when the transmitted power increases.
With small transmission power, the linear part of the self-interference,
remaining after a fixed amount of self-interference attenuation, dominates.
Principles of the self-interference mitigation techniques, modeled by , ,
and , will be outlined next.

www.TechnicalBooksPDF.com
 Figure 2.6 Power levels of different signal components

2.3 Self-Interference Cancellation Mechanisms

Self-interference cancellation techniques can be classified as passive and

active. The former refers to spatial separation and placement of transmit
and receive antennas, and antenna design techniques to minimize over-the-
air leakage of the transmitted signal to the receiver chain. Active
cancellation techniques can be further divided into RF or analog
cancellation and digital baseband cancellation.

These cancellation mechanisms are sketched in Fig. 2.7, where

transmitter and receiver antennas are spatially separated for clarity. In case
of MIMO, self-interference can also be mitigated with spatial suppression
using spatial filters and in addition to subtracting the estimate of
the transmitted signal from the received signal.

www.TechnicalBooksPDF.com
 Figure 2.7 Interference cancellation mechanisms in a full-duplex transceiver

Passive Isolation
In practice, the dynamic range of any front-end circuitry is finite
(see Fig. 2.5), and large difference in power levels may saturate the receiver
as noticed before. One way to improve dynamic range in the receiver would
be to implement several parallel receivers and choosing the output of the
ADC that has the largest linear signal [150] when the input signal to
different ADCs is attenuated with different factors. Dynamic range is
increased compared to a single ADC, but implementing parallel receiver
chains is costly, especially in the case of consumer products targeted at the
wireless communication market. Parallel receiver chains are used in radars,
though [151], where Automatic Gain Control (AGC) is too slow to follow
rapidly changing input signal levels.

When parallel receiver chains are not an option, physical isolation and
analog interference cancellation should be applied instead. By partitioning
transmit and receive antennas into separate arrays, physical isolation arises
from the path-loss between the transmit and receive antenna arrays, as
observed in Fig. 2.3. Furthermore, it is possible to place transmit and
receive antennas in such a way that isolation is improved beyond the path-
loss of the antenna design in Fig. 2.2. Given the signal model (2.1), the
residual self-interference after spatial isolation is expressed as

(2.30)

www.TechnicalBooksPDF.com
RF Cancellation
Polarization, directivity, antenna design, and geometry of transmit and
receive antenna arrays provide the first frontier in the cancellation of self-
interference. Next, self-interference should be mitigated further in the RF
domain to avoid saturation of the ADC by the self-interference signal. The
principle of self-interference cancellation in analog domain is the same as in
digital baseband: a replica of the interfering signal is created and then
subtracted from the received signal. This calls for active electronic circuits
which can cancel a time-varying self-interference signal. As this is more of
an implementation issue, different system architectures for analog
cancellation are presented in the next chapter.

Analog cancellation changes the statistics of self-interference, but from

developing algorithms in digital baseband, it is often assumed that after
passive isolation and interference mitigation in the RF domain, the
remaining self-interference has a certain power.

Digital Cancellation
After passive isolation and RF cancellation, the self-interference signal
should be within the dynamic range of the ADC and the remaining self-
interference is mitigated in digital baseband. As long as this is true,
cancellation of self-interference in digital baseband is agnostic to physical
isolation and RF cancellation. In practice, the effect of the cancellation
techniques is not strictly additive, though. For example, if the interfering
signal is already weak when converted into digital baseband, the estimation
of the self-interference channel becomes more difficult. Consequently, self-
interference cancellation in digital baseband is compromised due to the
noisy channel estimate.

Digital signal processing makes it possible to deploy more sophisticated

algorithms than in the analog domain. In Particular, different multipath
components of the self-interference channel are easier to handle in the
digital domain than in the analog domain, because tunable analog delays are
difficult to implement. In the following we present different algorithms for
interference mitigation in the digital domain.

www.TechnicalBooksPDF.com
2.4 Algorithms for Self-Interference Cancellation
2.4.1 Time-Domain Cancellation

Cancellation is based on the reasonable presumption that the transceiver

always knows its own transmitted signals at least approximately; if the
transceiver can also emulate the self-interference channel, interfering
signals can be replicated and subtracted from received signals.

Typically, cancellation is performed either in the time domain or in the

frequency domain. In case of OFDM, mitigation in the frequency domain
processes each subcarrier signal individually, which may become
computationally demanding if the number of subcarriers is large. Non-
orthogonal multicarrier modulation schemes like FBMC and GDMA
inherently cause inter-carrier interference that must be taken into account
when subtracting interference in the frequency domain. Thus, the single
carrier model does not apply to non-orthogonal multicarrier modulation.

Cancellation in the time domain processes the signal samples

independently of the number of subcarriers, but, due to the different
interference paths between antennas, it requires the delay spread of the self-
interference channel to be gauged [152]. Mitigation takes place after the
ADC and demodulation of the received signal, usually being the first
operation within the digital pipeline. As a result, the signal must be sampled
above the Nyquist rate, which demands the use of special techniques to deal
with arbitrary signal spectra [152]. Given the signal model (2.1) the residual
self-interference power after time-domain cancellation is expressed as

(2.31)

2.4.2 Frequency-Domain Cancellation

www.TechnicalBooksPDF.com
In case of OFDM and assuming that the channel length is shorter than
cyclic prefix, cancellation in the frequency domain can be done
independently on a subcarrier basis, and a frequency-flat signal model (2.1)
is sufficient. With non-orthogonal multicarrier modulation, Inter-Carrier
Interference (ICI) must be taken into account as well as noticed; see
Section 2.1.1.

Let us again use GFDM as an example of a non-orthogonal modulation

scheme and self-interference cancellation. Self-interference cancellation
should be applied before channel equalization of the desired signal to avoid
amplifying the self-interference. The cancellation process is executed on a
per-subcarrier basis after GFDM demodulation by subtracting a replica of
the transmitted signal. When taking into account ISI and ICI components,
self-interference cancellation becomes

(2.32)

where is the coefficient matrix for replica generation.

The main task of the replica generation process is to build the

coefficient matrix based on the GFDM signaling and the estimated self-
interference channel. To reproduce the subsymbol interference of GFDM
signaling, the replica generation sums up all data symbols with its own
coefficient matrix. The coefficient matrix is constructed by convolving the
estimated self-interference channel with the ICI pattern. Assuming that the
self-interference channel is perfectly estimated per subcarrier, the
coefficient matrix for replica generation is given by

www.TechnicalBooksPDF.com
 (2.33)

where is the frequency-domain filter matrix,

, and refers to the frequency-domain self-interference
channel matrix interpolated by M. The channel matrix is interpolated,
because the self-interference is passed through the channel being upsampled
in the frequency domain; see (2.7). By using the interpolated channel, the
canceller is able to duplicate the upsampled self-interference signal. After
cancellation, the receiver performs channel equalization.

Due to non-orthogonal signaling, self-interference cancellation in

GFDM system has larger computational complexity than that of orthogonal
signaling like OFDM. Even when the filter response is deterministic,
calculating full coefficient matrix requires high computational complexity.
This is because the coefficient matrix should be updated whenever the
self-interference channel is changing. Thus, computational complexity
should be reduced, and one option for this is to include only significant
interference terms in whose power exceeds a given threshold.

Significant interference terms are determined by examining cross-

correlation terms of transmission pulse responses . When , the
diagonal part indicates the self-interference due to transmitted symbol
vector . When the diagonal part indicates the ISI of self-
interference pattern. The non-diagonal parts of indicate the ICI from
its own subsymbol or different subsymbols. The diagonal part of matrices
cannot be ignored, but off-diagonal elements are small compared to the
diagonal elements, and interference power becomes smaller the further
away from the diagonal of the matrix. This observation is valid for GFDM
signaling in general regardless of the channel.

Computational complexity can be reduced in replica generation by

using only the coefficients near the diagonal and ignoring the rest of the off-

www.TechnicalBooksPDF.com
diagonal part. Denoting the element at the ith row and jth column of
, the coefficient matrix applying the valid set is expressed as

(2.34)

where threshold γ determines the number of ICI components included in the

cancellation. The canceller reproduces only the ICI components within the
distance γ from the center subcarrier while the remaining weaker
components are neglected.

The performance of the interference cancellation is simulated with the

following parameters:

1. The sampling frequency and the carrier frequency are 15.36MHz and
2GHz, respectively.
2. The number of subsymbols and subcarriers in each block is and
, while the subcarrier bandwidth is 15/256 MHz. The number of
sequentially transmitted blocks is 10, the length of cyclic prefix is 32
samples, and the modulation is 4-QAM.
3. Convolutional coding is applied with a constraint length of 7 and a
coding rate of 1/3.
4. The pulse-shaping filter in the frequency domain is the raised cosine
filter with a roll-off factor 0.8.
5. The channel model self-interference and signal of interest is ITU
Pedestrian A (“Ped-A”) [71].
6. A matched filter with successive interference cancellation is used as an
equalizer for the signal of interest.

Figure 2.8 shows the coded and uncoded BER of the signal of interest
when applying the SIC architecture with different γ. The x-axis indicates the

www.TechnicalBooksPDF.com
energy per bit-to-noise power spectral density ratio (EbNo) of the signal of
interest. The self-interference power is fixed at 40 dB. When the number of
subcarriers included in the cancellation increases, the uncoded BER
performance becomes better and gets close to the perfect SIC case that uses
the full coefficient matrix. The behavior of the coded BER is similar to that
of the uncoded BER. The BER curves of the case almost match the
perfect case with .

Figure 2.8 BER curves in GFDM receiver corresponding to different number of interference
components when self-interference power is fixed to 40dB. The channel model is ITU PED-A

Accordingly, for the given environment, the SIC architecture with

is the best choice to improve system performance with reduced
complexity.

2.4.3 Spatial Suppression

While only cancellation is possible in SISO transceivers, a new dimension

opens up for mitigation when considering modern MIMO transceivers. In
addition to subtracting the replica signal, the full-duplex MIMO transceiver
may also apply spatial suppression in case transmitter and receiver antennas

www.TechnicalBooksPDF.com
are spatially separated. Precoder at the transmitter side and beamformer at
the receiver can be designed to simultaneously attenuate the self-
interference signal and receive the signal of interest. The beamformer is
able to attenuate the effect of RF imperfections as well, because
the imperfections pass through the same channel together with the known
part of the self-interference.

The price of spatial suppression is the loss of dimensions for

information transfer. Spatial suppression techniques transform the physical
transceiver to an equivalent “interference-free”
transceiver. Here and represent the input and output dimensions, i.e.,
the number of spatial streams, reserved for the actual multiplexing protocol.
Without loss of generality, the concepts presume that and
.

Again, the target is to push residual self-interference to receiver noise

level. The signal model (2.1) becomes

(2.35)

where and are the respective receive and transmit signal

vectors of the equivalent interference-free transceiver, and
represent respective equivalent MIMO channels from
sources to the interference-free transceiver and from the interference-free
transceiver to destinations, and is an equivalent noise vector
including all residual self-interference after mitigation, receiver distortion
and thermal noise. The covariance matrix of is defined as .

The transceiver applies MIMO receiver filter and MIMO

transmitter filter for suppression as illustrated in Fig. 2.7.
The signal models (2.2), and (2.35) are linked together according to

www.TechnicalBooksPDF.com
 (2.36)

Equivalent filtered channels used in (2.35) become and

while the latter equation implies .

To be used in (2.35), the equivalent receiver noise vector becomes

(2.37)

in which the residual self-interference channel is given by

(2.38)

Consequently, the residual self-interference power can be expressed as

(2.39)

Self-interference can be now suppressed by appropriately designing

and to minimize the first term of and/or by designing only to
minimize the second term that is due to transmit signal noise. The spatial
filters can be designed, independently, separately, or jointly, where the last
option is supposed to provide the best performance.

These procedures are developed next for antenna selection (AS),

general eigenbeam selection (ES), null-space projection (NSP), and
minimum mean squared error filtering [153]. Here we assume that the
channel of the signal of interest and the self-interference channel
are known or estimated in the full-duplex transciver, but the forward

www.TechnicalBooksPDF.com
channel in the signal model (2.1) is unknown. Spatial suppression
techniques may also embrace the forward channel [154] when it can be
assumed known by channel reciprocity or channel state feedback.

Antenna Selection
In case of antenna selection, the respective receive and transmit filters are
constructed from row and column subset selection matrices with scaling for
gain normalization:

(2.40)

where and denote binary matrices for

which and with all j as well as and
with all i.

To reduce the gain of the residual self-interference channel given in

(2.38), the general objective for spatial-domain suppression is defined as

(2.41)

for decreasing the known part of . With (2.40), this becomes

(2.42)

specifically for antenna selection.

Optimal joint filter design can be solved by calculating the Frobenius

norm for all combinations and choosing the lowest one.

www.TechnicalBooksPDF.com
Eigenbeam Selection
Signals sent from each antenna propagate to all receive antennas. Thus,
Eigenbeam Selection (ES) makes MIMO subchannels effectively
orthogonal, allowing subset selection over virtual antennas with less
coupling.

Eigenbeamforming is based on the Singular Value Decomposition

(SVD):

(2.43)

in which and are unitary matrices from

singular vectors while diagonal matrix comprises singular
values , , in descending order. In
the last form, submatrices and collect basis vectors associated with
the null-space of the matrix, i.e., zero singular values.

After realizing that the singular vectors represent the set of available
eigenbeams, the spatial filters can be constructed by combining selection
and beamforming matrices with scaling for gain normalization as follows:

(2.44)

This transforms the general objective from (2.41) to

(2.45)

www.TechnicalBooksPDF.com
as and by definition. Thus, filter design remains
conceptually similar to antenna selection but row and column subset
selection is now based on the effective diagonal channel instead of
.

Null-Space Projection
In Null-Space Projection (NSP), the spatial filters and are selected
such that the transceiver receives and transmits in different subspaces, i.e.,
transmit beams are projected to the null-space of the self-interference
channel combined with the receive filter and vice versa. Such a condition
can be formalized for joint or separate filter design as

(2.46)

to eliminate the known part of the first term in (2.39). Similarly, for
suppressing the transmit signal noise, the condition becomes to
partly eliminate the second term in (2.39).

In particular, the ES algorithm yields null-space projection whenever

(2.47)

This condition also defines the general existence of joint null-space

projection, if and are additionally constrained to have full rank.
Even if may be rank-deficient, is of full rank in practice due to
estimation errors which also cause residual self-interference.

In the case of , the total number of antennas ( ) is

minimized for null-space projection by choosing
or when (full rank), or by
choosing or
when (minimum rank).

www.TechnicalBooksPDF.com
Selecting may be preferable due to transmit signal noise as
discussed later.

For separate filter design, one should recall that the Moore–Penrose
pseudoinverse is unique, always exists and satisfies by
definition [155]. For designing separately given , null-space
projection can, thereby, be implemented by applying projection matrix

(2.48)

if . Separate design for is given by a similar

projection matrix which is obtained by replacing above with
. Unfortunately, the case of having full rank for or
in above leads to or , respectively.

Minimum Mean Squared Error Filtering

All the above spatial-domain suppression schemes aim at minimizing the
effect of self-interference with the cost of spatially shaping the useful
signals, which does not happen with time-domain cancellation. In order to
reduce the effect of this drawback, a Minimum Mean-Square Error
(MMSE) scheme is developed next to both minimize the distortion and
attenuate self-interference.

Now and as with cancellation. Thus, the Mean-

Square Error (MSE) matrix of the transceiver input signal is given by

(2.49)

in which .

www.TechnicalBooksPDF.com
 On the one hand, for separate filter design given , the minimum
MSE receive filter is derived from the condition , yielding

(2.50)

which needs to be additionally scaled to satisfy . One can note

that MMSE filtering requires knowledge of and signal covariance
matrices as opposed to the other mitigation schemes.

On the other hand, the converse condition

(2.51)

to minimize MSE at the transmit side can be reduced to the condition of

null-space projection given in (2.46). Therefore, the order for joint spatial
filter design is to firstly minimize interference at the transmit side using any
scheme, and then secondly design the receive filter using (2.50).

2.4.4 Spatial Suppression and Time-Domain Cancellation

Time domain and frequency domain cancellation suffer from residual

interference that is mainly due to the transmit signal noise. On the other
hand, spatial-domain suppression requires more antennas than data streams
for effective mitigation of self-interference. These two approaches are not
mutually exclusive, and a smart hybrid combination may harness the
advantages of both schemes for achieving enhanced isolation. Then the
question arises as to how cancellation and suppression should be combined
together.

Four options to combine time (or frequency) domain cancellation and

spatial domain suppression are shown in Fig. 2.9. With independent or

www.TechnicalBooksPDF.com
separate filter design, the design (and performance) of these options differs
according to the residual self-interference channels:

(2.52)

However, these become equivalent if all three filters are designed jointly.

Figure 2.9 Different options for combining subtractive cancellation with spatial suppression

Options 1) and 4) allow the most straightforward implementation,

because the two layers of mitigation can be decoupled and joint design is
still feasible for and : in option 1), spatial suppression is first
applied, and the residual interference is mitigated with cancellation while
option 4) is the converse case. In options 2) and 3), all three filters must be
implemented sequentially, leading to suboptimal separate design. Moreover,
these options cannot exploit the full potential of the spatial transmit filter,

www.TechnicalBooksPDF.com
because cancellation has already eliminated the known part of the
interference. Thus, option 1) turns out to be the best choice from spatial
suppression point of view: as cancellation is blind to the spatial domain, it
is better to first design jointly and for exploiting all possible
degrees of freedom.

Performance Evaluation

The main objective is to study how much isolation is improved with spatial
suppression and time-domain cancellation and why it is typically limited,
leading to residual self-interference. To that end, specific factors to be
explored are the antenna configuration, the rank of the self-interference
channel, the distance and orientation of antenna arrays when there is (or is
not) line-of-sight present, and the effect of RF distortion and channel
estimation errors.

Channels are modeled with Rayleigh fading such that the elements of
and are identically distributed, but not
necessarily independent, circularly-symmetric complex Gaussian random
variables. And all experiments are iterated in a Monte Carlo manner over
independent trials.

While will always have full rank, many of the results illustrate
cases where is rank-deficient, i.e., . In
order to both control the rank and yield identically distributed elements with
proper correlation, channel matrices are generated as follows:

(2.53)

where, denoting ,

is an complex Gaussian random matrix where all

elements are independent and identically distributed,

www.TechnicalBooksPDF.com
 is an constant all-zero matrix, and
is an random unitary matrix, independent from , and
uniformly distributed according to the Haar measure.

When in the full-rank case, matrix disappears and

(resp. ), if (resp. ), which is
statistically equivalent to (resp. ) omitting the unitary
rotation. Thus, full-rank channel matrices can be generated by simply
picking all elements independently from identical distributions.

The transmitted signals are assumed to be spatially white with equal

power per transmitted stream, thus

(2.54)

where and denote the total transmit power in the source and in the
transceiver, respectively. The thermal receiver noise is white and Gaussian
with , where denotes noise power per antenna. Consequently,
the quality of the channels is parametrized with

(2.55)

which can be referred to as the channel Signal-to-Noise Ratios (SNRs).

The imperfect side information available for mitigation is generated as

explained in Section 2.1. The associated key parameters and are given
by

www.TechnicalBooksPDF.com
 (2.56)

controlling the relative level of additive error included in the respective

channel and signal estimates. The case of perfect side information can be
obtained by setting .

Additional Isolation from Mitigation

In principle, interference mitigation manifests itself as a shift in Bit-Error

Rate (BER) (or similar measures), which is herein referred to as the
additional isolation obtained with signal processing with respect to physical
isolation. The observed shift is unique for each mitigation scheme, but
similar for all multiplexing protocols, system setups and performance
metrics. Thus, the performance of the mitigation schemes can be compared
by studying directly the statistics of additional isolation, which is a versatile
metric and even the intuitive choice since the schemes are formulated to
minimize self-interference power in the first place. Given any level of
physical isolation, the statistics of additional isolation show
comprehensively how different schemes are affected by errors in side
information (parametrized with and ), configurations for antennas (
) or streams ( ) and the loop channel rank ( ), to
name but a few key factors.

With the above motivation, the analysis is conducted in terms of

additional isolation, , obtained by implementing active mitigation after
passive physical isolation:

(2.57)

www.TechnicalBooksPDF.com
for which the numerator is given by (2.30) and the denominator is given by
(2.31) or (2.39) with time-domain cancellation or spatial-domain
suppression, respectively. Average channel gains and total transmit powers
do not actually affect (2.57) because such parameters scale equivalently
both the numerator and the denominator. As is a random variable, the
performance of the mitigation schemes can be comprehensively quantified
by the cumulative distribution function (CDF), .

Figure 2.10 illustrates the benefit of combining Time-Domain

Cancellation (TDC) and spatial suppression in terms of cumulative
distribution function when the rank of the self-interference channel varies
between 1 and 4. Joint time-domain cancellation is implemented using
Option 1). Antenna selection was not evaluated because its performance is
clearly inferior to eigenbeam selection and null-space projection.

Figure 2.10 The joint performance of the combination of Time-Domain Cancellation (TDC) and
spatial suppression in comparison to their separate performance when ,
, , and

www.TechnicalBooksPDF.com
 The lower the rank of the self-interference channel, the better the
additional isolation. The combination of TDC and spatial suppression can
indeed offer better performance than either alone. However, when the rank-
deficient self-interference channel enables the usage of null-space
projection (NSP), adding TDC on top of suppression may actually reduce
the overall performance due to transmit signal noise. This is because NSP
can be made immune to transmit signal noise such that it becomes
eliminated together with the actual interference signal while TDC
effectively acts as an extra interference signal due to channel estimation
error. Without channel estimation error the self-interference would be
completely eliminated when eigenbeam selection becomes null-space
projection, i.e., when the rank of the channel is 1 or 2.

In general, the link-level simulation results show that spatial-domain

suppression is better than Time-Domain Cancellation (TDC) whenever
there is a sufficient number of antennas compared to the number of spatial
streams. In other words, time-domain processing is preferred if the spatial
domain is congested. However, the level of isolation is more stable in TDC
than in suppression.

2.4.5 Channel Estimation

In the simulation above, channel estimation was characterized with a

normalized variance in (2.56). A strong self-interference component
may create additional problems when estimating the channel of
the signal of interest and the self-interference channel. Treating self-
interference as additional noise when estimating the channel of the signal of
interest would result in a large estimator variance. The same happens when
estimating self-interference channel in the presence of the signal of interest
. The worse is the estimate of the self-interference channel, the
more residual self-interference there is after cancellation and suppression as
seen in (2.56). Unless a priori information of the transmitted signal is
utilized in the estimation, the problem is similar to channel estimation in the
presence of strong interferer, and channel estimators developed for
interferer scenario can be used in full-duplex transceivers as well.

www.TechnicalBooksPDF.com
 Joint estimation of the two channels is a reasonable choice if timing
between the two signals is unknown, or if the signal of interest does not
have any control signal structure resulting in blind channel estimation
[156]. It is justified to assume that transmitted and received signals are
synchronized at least at symbol level, because transmitter and receiver
should employ the same local oscillator to reduce distortion and residual
self-interference due to phase noise. Even then, this assumption may not
always hold, though, e.g., when baseband processing is done in software
running on a general purpose computer. In this case, the operating system
can cause unexpected delays, driving input and output streams out of
synchronism. This causes problems with self-interference cancellation and
may lead to a sudden increase in self-interference levels.

When baseband is implemented in hardware, synchronization at symbol

level can be enforced, though. Typically, wireless communication systems
adopt a frame and slot structure such that received and transmitted frames
can be synchronized within a full-duplex transceiver. Frame and symbol
synchronization make it possible to employ orthogonal pilot sequences in
the received and transmitted signals for channel estimation. This kind of
arrangement is known to be optimal in MIMO systems [157]. Most wireless
systems employ a pilot signal structure for channel estimation, so
orthogonal pilot sequences should be applied in full-duplex
communications as well whenever possible. Large self-interference levels
still deteriorate the channel estimation, because even when the
orthogonality of the pilot signals is easily achieved, transmitter distortion is
not orthogonal to received pilot sequence.

Given the noise due to transmitter imperfections it is more useful to

employ orthogonal pilot sequences in the time and/or frequency domain
instead of in the code domain. In the latter case, transmitter distortions
deteriorate channel estimation of the signal of interest even when the pilot
sequences themselves are orthogonal. In the former case, the self-
interference signal is not present when estimating the channel of the signal
of interest so residual self-interference does not pose a problem for channel
estimation. Time slots reserved for orthogonal pilot signals decrease the

www.TechnicalBooksPDF.com
throughput in uplink and downlink, but this overhead is the same in all
domains–time, frequency, and code–when pilot sequences are orthogonal.

The next section presents an adaptive self-interference cancellation

algorithm that does not need to estimate the self-interference channel and
does not require synchronization between the signal of interest and the self-
interference signal.

2.4.6 Adaptive Interference Cancellation

The building blocks for adaptive digital baseband cancellation according to

Option 1) in Fig 2.7 are depicted in Fig. 2.11. Cancellation is done in the
time domain following the signal model in (2.11). The effects of transceiver
impairments are modeled by noise vector and quantization noise and
limited dynamic range are captured in noise term . Let us first assume
that to simplify notation. Linear time-domain
cancellation aims at estimating the self-interference channel
, where refers to the order of the channel. Then the
canceller subtracts from the received signal,
where refers to the estimate of the self-interference channel.

Figure 2.11 Baseband interference cancellation

It is important to ensure that and are uncorrelated so that the

signal of interest is not accidentally suppressed in the process. However,

www.TechnicalBooksPDF.com
this condition is already satisfied by default in the case of bidirectional
communication and full-duplex access point, because the uplink and
downlink data are different. The condition is also true when the full-duplex
transceiver is a decode-and-forward relay, because decoding and re-
encoding the received signal necessarily imposes a sufficient delay, longer
than the delay due to the self-interference channel. In case of full-duplex
amplify-and-forward relay, signal can be delayed in the transceiver to
ensure the condition is met.

Let us arrange the MIMO matrices of the interference

canceller into the matrix . In
order to obtain a sufficient estimate of , the order of should
satisfy . Thus, we may write

(2.58)

where collects the samples .

If the source and the full-duplex transceiver employ orthogonal pilot

sequences for channel estimation, and the transmit and receive signals are
synchronized to maintain orthogonality, then the estimation process for
is similar to conventional multiuser channel estimation. Here we
assume that the source and the full-duplex transceiver are continuously
transmitting. Due to the decorrelation property of the transmit and receive
signals, we may use a power minimization algorithm without resorting to
pilot signals for the estimation of the self-interference channel.
Furthermore, by oversampling the received signal, the cancellation
algorithm may operate without synchronizing the transmit and receive
signals. This kind of architecture makes it possible to plug the digital
cancellation block to an existing transceiver when, e.g., more attenuation is
needed after antenna isolation and RF cancellation.

To this end, we minimize with respect to by the

stochastic gradient descent (SGD) algorithm

www.TechnicalBooksPDF.com
 (2.59)

where is the adaptation step size controlling stability, convergence

speed, and misadjustment. Note that the algorithm operates in the time
domain and is independent of modulation. In contrast, the operation domain
of interference cancellation using channel estimation and pilot symbols is
determined by the structure of the pilot symbols. For example, scattered
pilot symbols in a time–frequency grid as in LTE would imply frequency-
domain processing. In the time domain, data symbols would appear as noise
in the channel estimation while in the frequency domain, data and pilot
symbols do not interfere.

It can be shown [158] that the stationary points of the algorithm provide
perfect cancellation of the self-interference signal (in the sense that
converges to ) if the order of is sufficient and is
persistently exciting. Furthermore, the SGD algorithm can be combined
with spatial suppression such that, after the convergence of the algorithm,
the spatial filters are alternatively optimized [152]. A Recursive Least
Squares (RLS) algorithm to minimize has also been presented
in [158].

Figure 2.12 shows an example of the performance of the SGD and

SGD-RLS within a MIMO-OFDM system. The figure depicts mean
estimation error of the SI channel ( ) per iteration for SGD and SGD-
RLS adaptation algorithms. Parameter κ defines the power ratio between
information signal transmitted by the source node and SI, i.e., when ,
the power of the SI and information signal are the same, and when
there is no SI. The antenna configuration of the transceiver is of size
for transmitting two data streams simultaneously. The SI channel has three
taps and SNR is 15 dB. The number of subcarriers is 8192 and the length of
the cyclic prefix is samples so the algorithms converge within few
symbols depending on κ. The step size of SGD and the
forgetting factor of SGD-RLS and is used to initialize
the autocorrelation matrix in SGD-RLS algorithm.

www.TechnicalBooksPDF.com
 Figure 2.12 Convergence of LMS and RLS adaptation algorithms toward 25 dB suppression of
self-interference

Figure 2.13 shows the trade-off between convergence time and self-
interference channel estimation error for LMS and RLS [159] algorithms
and for different numbers of antennas in the transceiver: , , .
The OFDM symbol has 8192 subcarriers so that the algorithms converge
quickly with regard to symbol length. Convergence time is defined as the
number of samples needed for the algorithm to attenuate the self-
interference 25 dB. The figure shows the performance of the algorithms
when varying the forgetting factor in RLS algorithm or step size μ in LMS
algorithm. Parameter ε is used to initialize the autocorrelation matrix in the
RLS algorithm as . When the number of antennas in the transceiver
increases, there are more parameters to estimate resulting in longer
convergence time and larger residual interference. Here the power ratio
between the signal transmitted from the source and self-interference signal
is 30 dB and SNR is 15 dB.

www.TechnicalBooksPDF.com
 Figure 2.13 Convergence time in samples for different adaptation steps in LMS and RLS
algorithms

As mentioned before, linear cancellation by subtracting the contribution

of the SI in the digital baseband domain is not able to remove the effects of
transmitter imperfections. Instead of modeling the imperfections as in
Section 2.2, another option is to combine the time-domain cancellation and
spatial-domain suppression [137, 152, 153], where the latter is able to
suppress transmitter distortions as well.

Let us model the total receiver noise as

, where models the RF
imperfections as Gaussian noise, whose power depends on EVM value,
models the quantization noise, and denotes noise in the receiver. By
modifying it is possible to model the effects of limited dynamic range
of the analog-to-digital converter. The received SINR is given by

(2.60)

Assuming that the SGD algorithm has converged, the term containing
vanishes. However, given the limited dynamic range of the receiver, it is

www.TechnicalBooksPDF.com
nevertheless important to minimize . To this end,
we can apply alternative optimization such that

where refers to the Frobenius norm, and is the maximum

transmit power. Linear constraints can be set on the basis of an effective
post-processing target channel between the transceiver and the
destination. The target channel can be defined, e.g., by the maximum
number of channel taps so that performs channel shortening [137].
Once is determined can be solved by casting Eq. (2.60) to a
generalized eigenvalue problem. This alternating optimization is repeated
until convergence.

2.5 Key Challenges

Nonlinear techniques are essential to further improve the performance of

self-interference cancellation when transmitter distortions are large. The
drawback of nonlinear cancellation techniques is that their performance and
convergence is considerably more difficult to analyze than with linear
techniques. Furthermore, nonlinear cancellation schemes utilizing
behavioral models make several assumptions about hardware design, and
their performance should eventually be verified with protype
implementations.

Modern wireless communication systems employ multi-antenna

techniques, and full-duplex transceivers should be able to make use of
emerging MIMO techniques, like massive MIMO. Self-interference
cancellation and suppression algorithms should be scalable in such a way as
to allow a large number of transmitter and receiver antennas without
excessive complexity.
 Non-orthogonal multicarrier modulation like Filter Bank Multi-Carrier
(FBMC) and Generalized Frequency Division Multiplexing (GFDM)
require more complicated signal processing for self-interference
cancellation. Therefore, self-interference cancellation techniques with low
complexity are required to make full-duplex and non-orthogonal
multicarrier modulation a feasible combination.

1 www.ict-phydyas.org and www.5gnow.eu

3 Full-Duplex System Hardware Implementation

In this chapter we outline recent hardware development of full-duplex

transceivers. The focus is on the development of compact full-duplex
transceivers suitable for wireless communication systems. Isolation
between the transmitter and receiver chains is limited when using the same
antenna for transmitting and receiving. This kind of antenna sharing
requires excellent performance of analog and digital cancellation techniques
to mitigate self-interference. Therefore, full-duplex technology is more
likely to be implemented first in access points and relays that may
accommodate separate transmitter and receiver antenna arrays to improve
passive isolation. This makes the task of analog and digital cancellation
easier, although they are still needed because passive isolation is generally
not enough in the case of compact devices. Wireless relays may be
preferred, because they have inherent traffic symmetry irrespective of
uplink and downlink traffic ratio, and the gains from full-duplex operation
are largest when the traffic is symmetric.

3.1 Full-Duplex Technology Used in Wireless

Systems

In-band full-duplex technology has been utilized in wireless

communications systems for a long time in continuous-wave radars and in-
channel repeaters in broadcast and cellular systems. Not surprisingly, self-
interference has been one of the key challenges in these systems.
 Continuous Wave (CW) radar systems use either two separate antennas
or one shared antenna to transmit and receive simultaneously [160],
whereas pulsed radar systems switch off the transmitter while radar returns
are collected. The conventional CW radars of the 1940s and 1950s achieved
isolation between the transmitter and receiver through antenna separation-
based path-loss in separate-antenna systems, or through the use of
circulators in shared-antenna systems. Because only mild levels of isolation
could be achieved using these techniques, keeping self-interference to a
manageable level required the transmission power to be greatly limited,
which then limited the detectable range of targets. This restriction of CW
radar to nearby targets (i.e., short ranges) turned out to be useful, because
detecting nearby targets with a pulsed radar system would require on/off
switching times that are impractically small. Thus, the operational range of
CW radars matched their needs.

In the 1960s, an analog circuit-based form of self-interference

cancellation was proposed to increase the dynamic range of CW
radars [161]. Although a total isolation of 60 dB was reported, the canceller
required a 60+ kg block of precisely machined ferrite rotators and was very
expensive. In 1990, an improved analog canceller was proposed in [162]
that allowed adaptability to non-constant channel conditions.

In wireless communication systems, full-duplex, or on-frequency,

repeaters have been used for coverage extension, to fill gaps in network
coverage, to provide more capacity in hot spots, and to improve indoor
coverage [163]. The motivation is to save in costs and to densify network
by deploying repeaters instead of full-blown base stations. Repeaters are
essentially amplify-and-forward relays that transparently repeat all signals
in the donor area and may change, e.g., link balancing and co-channel
interference. Therefore, careful network planning, set-up, and parameter
tuning are important factors to get advantage from full-duplex repeaters.
Implementation faces the same problems as full-duplex transceiver design
in general: isolation between transmitter and receiver chains should be high
enough to facilitate high transmission power. In broadcast networks,
repeaters have typically been used as gap fillers to cover shadowed areas
when the number of available frequencies is limited [164, 165].
 Recently, a provision for in-band full-duplex repeaters has been
included in the LTE specification [166]. Three different relay types are
specified to establish a wireless backhaul link between a relay node and its
donor base station. Type 1 and type 1b in-band relays use the same
spectrum resource for both the relay backhaul and the access link between
the relays and user equipments. Type 1a is an out-band relay, which uses a
different frequency bands for access and backhaul links. Type 1 relays have
physical cell ID and functionalities for radio resource management of their
own, thereby creating a new cell. In addition, the specification includes the
type 2 relay, which implements the protocol stack only partly and does not
generate a new cell. Type 1 relays operate the two links in time-duplex
manner, but the type 1b relay is targeted for simultaneous operation of both
links, i.e., it is essentially an in-band full-duplex relay. Type 1, 1a and Type
1b relays have less impact on standardization than Type 2, which makes
them a preferred choice for deployment. Even then, wireless relays have not
been widely deployed in LTE.

This may change in the future evolution of 3GPP specifications toward

5G. In-band full-duplex transmission has been one of the technologies that
have been considered for 5G communication systems as well in a broader
scope than mere relays. Operators have been conducting field trials on full-
duplex transceivers in realistic network conditions, and the operators’
alliance Next Generation Mobile Networks Ltd. (NGMN) has listed full-
duplex technology as one of the building blocks for 5G systems in their
2015 White Paper [167].

3.2 Antenna Design

Self-interference mitigation consists of physical isolation and RF and digital

baseband cancellation. In the case of broadcast systems, the distance
between the transmitter and receiver antennas of the full-duplex transceiver
operating as a gap-filler can often be made rather large, which creates path-
loss between the antennas and thereby improves physical isolation. In case
of wireless local area networks the form factor of a full-duplex transceiver
should be rather small, which makes it difficult to achieve sufficient
physical isolation, though. Thus, the techniques that can be applied to
achieve enhanced antenna isolation are different depending on the form
factor of a radio device. When the form factor is extremely small, such as in
a mobile phone, it is not desirable to have separate transmitter and receiver
antennas but to share the same antenna. In this case, it is necessary to resort
to the RF and digital cancellation entirely. When the space limitation is
more relaxed, e.g., in the case of access points and relays, the use of
separate receiver and transmitter antennas offers the ability to increase the
electromagnetic isolation between the antenna ports before they are
connected to the RF front-ends. Different techniques to improve physical
isolation by antenna design include at least the following techniques:

Antenna placement and geometry to create destructive interference

between transmitter and receiver antennas;
Orthogonal polarization in transmitter and receiver antennas;
Wavetraps and slots in the ground plane that suppress ground plane
currents;
Loop structures that cancel local reactive magnetic fields to improve
isolation between transmitter and receiver antennas;
High impedance bandgap structures that suppress surface currents;
RF absorber materials that can be placed in electric and magnetic
field maxima to decrease coupling.

These different options are outlined in the following section.

3.2.1 Antenna Geometry

Probably the first attempt to exploit antenna geoemetry to cancel self-

interference is depicted in Fig. 1.17. Transmit antennas are located
asymmetrically around the receive antenna with the distance d and ,
respectively, where is the wavelength of the carrier. Ideally, the phase
difference between the two identical transmitted signals is 180 in the
receiver antenna and the signals cancel each other. On the other hand, this
only works for narrowband signals and the scheme is limited to a single
receive antenna. In the far-field, the antenna pattern by two transmit
antennas is not omnidirectional but directional, and the pattern depends on
the distance d. Thus, the antenna placement may cause unexpected behavior
at the receiving end.

This kind of self-interference cancellation by exploiting antenna

geometry has been generalized in [168] for an arbitrary number of transmit
and receive antennas. Transmit and receive antennas are located
symmetrically on lines perpendicular to each other. Two transmit and two
receive antennas operate in pairs, and the other branch is phase shifted by
180 . This facilitates the interference cancellation in the transmitter and in
the receiver. MIMO with data streams and receiver streams is
obtained by placing and antennas symmetrically on the two lines.
Total isolation of 35 dB was achieved for a MIMO system over 20
MHz bandwidth at 2.5 GHz carrier frequency. This is not due to antenna
geometry alone but requires phase shifters implemented by passive or
active analog techniques. Antenna geometry also affects to transmitted and
received beam-patterns, and may have further implications to link and
system level throughput in wireless systems.

Instead of employing multiple omnidirectional antennas, antenna

elements can be directional pointing to different directions [122, 123, 169].
An example of this kind of design with corresponding results from channel
measurements were presented in Section 2.1.3.

In general, line-of-sight path should be blocked between transmitter and

receiver antenna arrays to improve spatial isolation and decrease self-
interference. For this purpose, the installation may exploit surrounding
structures or add, e.g., a shielding plate between transmitter and receiver
antenna arrays as in the back-to-back antenna design, but already a
thick glass window can be rather useful even though the optical path is not
blocked. Experimental results have shown that physical isolation remains
rather low when implementing compact full-duplex transceivers, such as
the Single-Input Single-Output (SISO) prototype of [124], and the design
challenges become even more pronounced with MIMO with coupling
between multiple antennas.

3.2.2 Polarization
In theory, the self-interference signal could be blocked completely when all
the elements of the receive antenna array have the same polarization which
is orthogonal to the polarization of the transmit antenna elements. In
practice, real-world antenna elements are not perfectly polarized and the
polarization is preserved only in ideal line-of-sight channels without
reflections. Depolarization is typically expressed as a cross-polarization
ratio, which is a ratio between the expected powers when using ideal
horizontally polarized receiver antennas and the signal is transmitted
employing an ideal vertically polarized antenna. Cross-polarization power
ratio has been studied in many measurement campaigns, indicating that in
urban and suburban environments the value of cross-polarization ratio
varies between 0 and 12 dB [170]. This may give some indication of the
expected ratios in the case of full-duplex transceivers, although the ratio
depends strongly on the environment and location of reflectors close to full-
duplex devices. Thus, the additional isolation from different polarizations is
likely to be rather modest in real environments. Moreover, in wireless
communications, polarization is often used for diversity or spatial
multiplexing, and dedicating polarization to mitigation of self-interference
loses available degrees of freedom in a similar manner to using separate
transmit and receive antennas.

The usage of different polarizations to yield higher isolation has been

demonstrated using the MIMO antenna design in Fig. 3.1, where the
transmit antenna elements are slanted 45 degrees with respect to the receive
antenna elements. Similar arrays with orthogonal polarizations were
tested [122] and it was observed that the improvement from orthogonal
polarization is minimal. This matches expectations based on the
measurement results on cross-polarization ratios.
 Figure 3.1 Overall view of full-duplex relay with wavetraps to improve spatial isolation

3.2.3 Wavetraps

To improve isolation between transmitter and receiver antennas it is

necessary to reduce ground-plane currents and electromagnetic fields that
cause the transmitted signal to appear at the receiver antenna. To this end,
high-impedance surfaces [171] and slots in the ground planes [172] have
been considered. However, high-impedance surfaces usually need to be
fairly large (larger than quarter-wavelength) and therefore they are not
feasible in compact devices. On the other hand, slots in the ground plane
usually cause prohibitive problems when trying to meet electromagnetic
compatibility requirements.

Wavetraps are resonant structures consisting of a planar quarter-

wavelength patch short-circuited at one end to the ground plane. The input
impedance of a short-circuited quarter-wave transmission line is very high.
This results in small currents at the open end of the patch with large
currents concentrated within the patch. When placed between two antennas
in a suitable manner, wavetraps reduce the surface currents flowing between
the antennas and thereby effectively improve the electromagnetic isolation
between the antenna ports.

In order to increase spatial isolation, the electromagnetic isolation

between the antennas on opposite sides of the box is improved with
wavetraps. Figure 3.1 shows the schematic of the relay antenna structure
with wavetraps on the ground plane. The design has been modified from
Fig. 2.2, but now the design has instead of antennas. Wavetraps
are shown on the edges of the box.

The patch antennas of the relay are resonant at 2.56 GHz with dB
matching across a 139 MHz band, and dB matching across 238
MHz [129, 173, 174]. In the relay structure, transmitter ports 1 and 2 are on
the front side and receiver ports 3 and 4 are on the back side (not shown in
the figure). The dimensions, the number of the wavetraps, and the gap
between the wavetraps are optimized with electromagnetic-field
simulations. The goal was to optimize the worst-case isolation between the
antenna ports, i.e., among the four different combinations of all the receiver
and transmitter ports. The wavetraps on both sides of the relay are arranged
in a similar order, but oriented in opposite directions, i.e., the open ends of
the wavetraps are located toward the center on the transmitter side and
away from the center on the receiver side. This asymmetry resulted in better
overall isolation between the antenna ports.

The designed relay antenna was manufactured, and then measured in an

anechoic chamber. The isolation was evaluated by measuring the S-
parameters of the relay antenna as a four-port network. In the relay without
the wavetraps, the minimum isolation between receiver and transmitter
ports is 50 dB at the resonance frequency of the patch antenna as with
Fig. 2.2. After adding the wavetraps, the minimum isolation improved by
21 dB to 71 dB at the resonance frequency. In addition, the wavetraps offer
also a wideband isolation improvement around the designed frequency,
even though they are narrow-band structures by design. The isolation across
a 167 MHz band around the center frequency is better than 65 dB, which is
an improvement of 15 dB over the 50 dB minimum isolation without
wavetraps. Within the given isolation bandwidth, the corresponding antenna
port matching is better than dB. The gain of the relay antenna was also
measured in the anechoic chamber to determine the impact of the wavetraps
on the radiation pattern. The impact was minor with a maximum gain of
dBi for all ports of the relay without wavetraps and 9.6 dBi for
ports 1 and 2 and dBi for ports 3 and 4 for the relay with
wavetraps.

The wavetraps efficiently concentrate the ground plane currents and

reduce the electromagnetic coupling to the other side of the relay box,
thereby improving the antenna isolation. In some cases, the isolation
achieved by wavetraps and back-to-back patch antennas may already be so
large that the signal of interest and self-interference fit together to the range
of ADC converter, and only digital cancellation may be sufficient [174].
This requires separate transmitter and receiver antennas, though, and analog
cancellation is vital in transceivers sharing the same antenna between
transmitter and receiver chains.

3.3 Analog Cancellation Techniques

A shared antenna solution enables small form factors and in principle

facilitates full-duplex operation in mobile handsets and tablets. Consumer
equipments is typically used in environments where scatterers are in the
vicinity of the transceiver, and due to movement and changes in
surroundings, self-interference channel and the amount of self-interference
may change. For example, if a user happens to cover the antenna with their
hand, antenna impedance experiences an abrupt change. This variability in
conditions makes the cancellation of self-interference more complicated
than in devices having separate transmit and receive antennas. Nevertheless,
several designs for full-duplex transceivers using a shared antenna have
been presented in the literature, because the implementation is evidently a
challenging and scientifically interesting problem.

On the other hand, full-duplex technology is most useful with

symmetrical traffic, which is often not the case with handsets and tablets
that are mostly used to consume downlink traffic. Symmetrical traffic is
inherently satisfied when the transceiver works as a full-duplex relay. Full-
duplex relays can be useful building blocks to densify networks when
providing wireless backhaul to donor base stations. In this case a, full-
duplex relay is part of the infrastructure, and it can allow larger form factors
and separate transmit and receive antennas.

3.3.1 Shared Antenna Design

In shared antenna transceivers there are two sources of self-interference:

leakage through the circuit board inside the transceiver and reflection from
the antenna. To cancel the latter, similar RF cancellation techniques can be
employed in shared antenna and separate antenna transceivers. To cancel
the former, one option to create the cancellation signal is to couple part of
the transmitted signal to the cancellation network, as in [102, 130, 175–
177], and another is to use a separate transmitter chain to create the
cancellation signal [178]. In the former case, isolation is typically obtained
by using a circulator and Electric Balance Duplexer (EBD). This kind of
implementation is useful when the aim is to minimize the form factor of the
full-duplex device and share the antenna between transmitter and receiver
chains.

A circulator that can be active or passive is a three-port device that

ideally allows the signals to rotate in one direction only. In practice, signal
leaks in the unwanted direction as well, which causes additional self-
interference together with reflections of the transmitted signals from objects
nearby. The circulator in [102] offers 15 dB attenuation of self-interference.
Circulators with improved isolation have been developed, e.g., in [179]
offering up to 30 dB isolation, but the isolation is still moderate and
narrowband when compared to using different transmitter and receiver
antenna arrays as in Fig. 3.1 that is able to provide 70 dB passive isolation.
Therefore, additional RF and digital cancellation techniques are vital for
full performance.

Typical EBD implementations integrate four core elements: a hybrid

junction, an antenna, a balancing impedance and a control system for tuning
the balancing impedance. The balancer consists of a hybrid transformer and
a balance network which is essentially a tunable dummy load impedance.
The transmitted signal is split in two, and by tuning the balance network
impedance, the magnitude and phase of the two transfer paths are made
equal, such that they destructively interfere at the receiver port. Interference
from the receiver chain to the transmitter chain is handled symmetrically.
This hybrid junction circuit can be densely implemented in CMOS
technology and can be co-integrated on-chip, which makes it an attractive
technique for a compact full-duplex transceiver.

By controlling the balancing impedance such that it reflects the antenna

impedance as closely as possible, one can directly impact the level of
transmitter–receiver isolation, and hence the level of self-interference in a
full-duplex radio unit. Splitting the signals is inherently lossy, because part
of the transmitted and received power is dissipated in the balancing
network. In a symmetric case the loss for both transmitter and receiver
chain is 3 dB. By tuning antenna and balancing impedances, the insertion
loss in the transmitter and receiver noise figure can be made comparable to
the cancellation using splitter, circulator and coupler. In general, several
alternative topologies can be used to realize a hybrid junction, such as auto-
transformer [180], 90 or 180 hybrid couplers [178], center-tapped
transformer [178] and coupled line hybrid junction [181], each of them
having different relative benefits and drawbacks.

3.3.2 RF Cancellation Techniques

The principle of self-interference cancellation in the analog domain is the

same as in digital baseband in Section 2.4. The transmitted signal is tapped
somewhere in the transmitter chain, then the copy of the transmitted signal
is convolved with the estimate of the self-interference channel in the analog
domain and subtracted from the received signal. In narrowband systems the
self-interference channel can be assumed to be frequency flat, and can be
cancelled using a phase shifter and attenuator. It is necessary to know the
delay between the transmitted and received self-interference signal for
cancellation to work. With separate transmit and receive antennas, the
distance between antennas is known by design, and therefore the delay or
the required phase shift is fixed. Only the attenuation of the replica signal
needs to be adjusted, but since the transmitted power is known, the range of
attenuation is known as well. In a shared antenna system, where the
strongest path is due to reflections in the radio propagation environment,
unknown a priori, the phase shift and required attenuation must be
estimated.

The design in [182] used the antenna geometry in Fig. 1.17 and off-the-
shelf noise canceller chip QHx220 to remove the self-interference in the RF
domain. The QHx220 chip takes the known self-interference and received
signals as inputs, and outputs the received signal with the self-interference
subtracted out. The chip allows the amplitude and phase of the interference
reference signal to be changed to match the interference in the received
signal. The amplitude and phase of the canceller signal are adjusted
manually, and an RF splitter is used to supply the transmit signal to the
cancellation circuit as the interference reference. The RF interference
cancellation achieves roughly 20 dB reduction in the received self-
interference on top of the reduction achieved by passive isolation. Other
designs with a single RF tap are reported, for example, in [130, 176, 177].

This design was further improved in [103] by introducing a balanced-to-

unbalanced (balun) converter and adaptive tuning of amplitude and delay.
The transmit antenna of a node transmits the positive signal from the balun.
To cancel self-interference, the radio combines the negative signal from the
balun with its received signal after adjusting the delay and attenuation of
the negative signal to match the self-interference. Delay and attenuation
were adjusted by using QHx220 and a gradient descent algorithm.

Multipath Cancellation
Things get considerably more complicated with wideband transmission and
consequently with a multipath self-interference channel. Even with separate
transmitter and receiver antennas, attenuation of the first signal path is not
sufficient with wideband signals any more due to multipath components
caused by radio propagation environment that was not known beforehand.
Building the cancellation signal in the analog domain requires multipath
analog filters that are challenging to implement. Furthermore, even when
the transceiver stays in a fixed position, movement in the environment
changes the reflected paths.
 Cancelling a multipath signal requires delays to be implemented in the
RF domain. This can be done by using transmission lines or delay line
components. The transmission lines, such as coaxial cables, are rather ideal
delays, but consume a significant amount of space. The delay line
components are physically small in size and can be slightly dispersive, as
described in [183]. Amplitude and phase within the taps can be controlled
by separate chips or using a vector modulator. Commonly, these
components are analog devices but can be controlled digitally. They show
reasonable linearity and noise properties, especially when amplitude control
is implemented as a variable attenuator. The components used in the
implemented designs in [102] and [175] include a Peregrine PE43703
variable attenuator and Hittite HMC631LP3 vector modulator, respectively.
The linearity of the used components should be as high as possible, since
otherwise the already distorted signal from the Power Amplifier (PA) output
is further distorted. This results in double distortions which can be difficult
to cancel in the digital domain.

In general, the delay profile of the multipath channel does not match
exactly to the fixed delays, but programmable analog delays with
sufficiently high tuning resolution are hard to build. Therefore, [102] uses
fixed delays and implements an offline interpolation algorithm that
calculates amplitude weights used in the analog cancellation circuit to
weight the cancellation signal for each delay. The algorithm is run in the
digital domain and the parameters are updated in 100 ms resolution.
Effectively, the algorithm implements a fractional delay filter [184] in the
analog domain instead of the digital domain. In digital signal processing,
fractional delay filters are used to implement an arbitrary delay while
keeping the sampling rate and timing fixed. Up to 16 analog taps are used to
create the cancellation signal. Such a large number of taps is required as
only the amplitude is controlled in the taps.

In the design of [175], implemented in [185], both the amplitude and

phase are controlled within the taps. This allows a significant reduction in
the number of delay lines: from 16 to 2 in [185], which was considered
sufficient to cancel leakage from circulator and antenna reflection before
digital cancellation. Thus, vector modulation that is able to tune amplitudes
and phases simultaneously is able to improve the modeling accuracy in RF
domain drastically.

Accurate regeneration and cancellation of the self-interference still

requires careful optimization and control of the amplitudes and phases of
the parallel analog filter branches. In [185], automatic self-tuning is applied
by using an analog Least-Mean Squares (LMS)-type adaptive filtering
algorithm, which seeks to minimize the instantaneous squared error at the
canceller output given by

where the number of channel taps is L and and refer to amplitude

weights and delays, respectively, in the RF domain. Signal is the
transmitted signal at PA output and is the received signal
after passing through the self-interference channel .

For the analog tracking using LMS algorithm, the instantaneous error
signal , is I/Q demodulated to baseband, and the weight update equation
is given by

where and refer to I/Q demodulated signals in the baseband and μ

denotes the step size controlling the learning and tracking rate as well as the
steady-state variance. Thus, the control loop is fully analog and thereby
faster compared with implementing the adaptive algorithm in digital
baseband. Furthermore, it does not require ADC and DAC.

The analog cancellation techniques above make use of off-the-shelf and

discrete components. Fully-integrated implementations of analog
cancellation techniques have also been presented recently [186–188].
Instead of tuning the cancellation signal given fixed analog delay lines,
[188] employs frequency-domain equalization to implement a frequency-
selective cancellation signal. A bank of reconfigurable second-order
bandpass filters is used to tune amplitude and phase scaling and group delay
implementation and match the frequency responses of the interference and
cancellation signals. A bank of such filters with independently controllable
parameters enables replication of a frequency-selective self-interference
channel at multiple points in different sub-bands, enhancing cancellation
bandwidth.

A similar approach could be exploited in digital baseband as well, i.e.,

to model the self-interference channel with an Infinite Impulse Response
(IIR) filter instead of Finite Impulse Response (FIR) filters. However,
unless channel impulse response is very long, it is more straightforward to
model the channel as an FIR filter due to its inherent stability when
computational complexity is not an issue. Realistic channels also have finite
impulse responses, so the FIR model is a natural choice.

3.4 Advanced Digital Cancellation

Time or frequency domain self-interference cancellation and spatial

suppression in digital baseband employing a linear signal model were
discussed in Section 2.4. Transmitter distortions can be suppressed with
spatial receiver filter, but cancellation is unable to reduce the interference
due to unknown distortions. The contribution of different impairments to
self-interference was analyzed in Section 2.2.3. In the following, we present
a digital baseband algorithm to model the effect of transmitter distortion
and further cancel the nonlinear self-interference from the received signal.

Nonlinear Cancellation
The largest distortions are caused by amplifiers in the transmitter and
receiver, I/Q imbalance or phase noise of the local oscillator, and ADC
quantization noise. The most significant of these distortions is typically
caused by the power amplifier in the transmitter. To mitigate this adverse
effect, the transfer function of the power amplifier in RF should be first
identified in the digital baseband, which requires some a priori knowledge
of PA characteristics.

When parameters of the selected model have been identified, the

transmitted baseband symbol sequence is passed through the PA model and
subtracted from the received signal. In a similar way, the effect of I/Q
imbalance can be modeled digitally as an additional distortion component,
which consists of the complex conjugate of the original baseband transmit
signal with memory coefficients [189]. Amplifier distortion and mixer
nonlinearities can be modeled using polynomial-based
systems [102, 190, 191], whereas I/Q imbalance can be modeled using
widely linear filters [189], both of which have been extensively studied in
the literature.

Widely Linear Digital Canceller

As a starting point, the baseband signal in the transmitter j ( )
is denoted by . The first component distorting the transmitter signal (cf.
Fig. 2.4) is the I/Q modulator, which will inherently introduce some I/Q
imbalance. The output signal of the I/Q modulator model (frequency-
independent model for simplicity) is [191]

(3.1)

with and , where are

the gain and phase imbalance parameters of transmitter j. Note that for any
practical transmitter front-end, . The strength of the induced
I/Q image component, represented by the conjugated signal term in (3.1), is
typically characterized with the image rejection ratio (IRR) as
.

Widely linear filtering is typically used to handle signals distorted by

I/Q imbalance. Widely linear time-invariant filtering with filters and
is defined by
where sequences are complex-valued and refers to convolution. The
operation is called widely linear because it involves filtering and its
conjugate. Using a widely linear canceller, the estimate of the self-
interference signal at the receiver antenna port i in digital baseband
becomes

(3.2)

where refers to the channel estimates for the linear part of self-
interference ( ) and its I/Q mirror image ( ) [189]. The index pair ij
refers to a channel from transmitter antenna j to receiver antenna i in the
self-interference channel matrix in the signal model (2.1). The time
domain canceller then uses this estimate when subtracting interference from
the received signal. Assuming the same memory length for both estimates,
the number of parameters is doubled with respect to the linear canceller.
However, potential improvement in self-interference cancellation can be
significant.

Nonlinear Digital Canceller

After the I/Q modulator, the signal is next fed to the transmitter PA, which
will further distort it. Often, it is not even necessary to model the
nonlinearities produced by the other components, because usually the Tx
PA produces most of the nonlinear distortion [189]. Therefore, the next step
in increasing the accuracy and complexity of the self-interference model is
the nonlinear digital canceller, which takes into account the nonlinear
distortion produced by the Tx PA but ignores the I/Q imbalance for the
moment.

There are several alternatives for behavioral models to model the

nonlinear distortion produced by PAs as discussed in Section 2.2.2. A
common approach is to use polynomials to model nonlinearities by
processing higher-order terms of the input signal. In the following we
employ a polynomial Parallel Hammerstein (PH) model to include high-
order distortion terms in the signal. The PH nonlinearity is a widely used
nonlinear model for direct as well as inverse modeling of power amplifiers
and has been observed, through RF measurements, to characterize the
operation of various power amplifiers accurately [192].

Using the PH model, the output signal of the PA can be modeled in

digital baseband as

(3.3)

where M and P denote the memory depth and nonlinearity order of the PH
model, denote the FIR filter impulse responses of the PH branches
for transmitter j, and the basis functions are defined as

(3.4)

The nonlinear model of the self-interference signal at receiver antenna

port i in digital baseband becomes

(3.5)

where contains the self-interference channel estimates for the

different basis functions. The number of basis functions, whose
corresponding channel coefficients must be estimated, is per
signal path, which shows the increase in computational complexity when
compared to the linear and widely linear canceller.

Augmented Nonlinear Digital Canceller

In the case of compact full-duplex transceivers sharing their antennas
between transmitter and receiver chains, the power of impairments can be
relatively high compared to the signal of interest, and modeling PA
distortion may not be sufficient any more. Naturally, if I/Q imbalance and
PA distortion are jointly modeled accurately enough and subsequently
cancelled, the performance of self-interference cancellation may be
improved when compared to modeling only I/Q imbalance or PA distortion.
In fact, augmented nonlinear digital canceller modeling transmitter and
receiver I/Q imbalance is able to improve cancellation with high transmitted
powers [189, 191]. The model of the self-interference signal is of the
similar form as in (3.5) and is given by

(3.6)

with again containing the self-interference channel estimates.

The full augmented nonlinear SI signal model further increases the number
of basis functions to be estimated to functions per signal
path. Thus, the computational complexity of the augmented nonlinear
canceller can be rather high, but the cancellation performance is also likely
to be better compared to more simple cancellers.

Figure 3.2 shows the power density spectrum of one experiment using
the antenna design with wavetraps, active RF cancellation and nonlinear
digital cancellation [129] The design is able to attenuate self-interference
more than 110 dB measured in anechoic chamber. The next evolution of the
transceiver design is presented in [174] in which the measurements were
conducted in real multipath environment. It was observed that indoor delay
spread was too large for the designed RF canceller. However, spatial
isolation and nonlinear digital cancellation were able to attenuate self-
interference approximately 100 dB.

Figure 3.2 Self-interference after isolation from antenna design, active RF cancellation and
nonlinear digital cancellation

Predistortion
Nonlinear digital cancellation techniques model transmitter impairments in
digital baseband to better match the cancellation signal to the actual
received signal. A question may arise: why not reduce transmitter
distortions instead of compensating for them in digital baseband, because in
addition to improving the cancellation of self-interference it would improve
the link quality in the destination as well? Instead of using the copy of the
transmitted signal to cancel self-interference, the signal can be used for
digital predistortion [138, 147]. Digital predistortion first identifies the
nonlinear characteristics of the power amplifier based on the original
baseband signal and its distorted copy, and then compensates these effects
to reduce distortion in RF. Identification step requires the assumption of the
PA model (cf. Section 2.2.2) to learn the parameters.
 The distorted copy for system identification is obtained by tapping the
transmitted signal at the PA output and down-converting it back to the
digital baseband. A similar operation is required with RF cancellation
techniques, in which the copy of the transmitted signal is used to cancel the
self-interference signal in the RF domain instead of converting it back to
the baseband. In the nonlinear digital cancellation techniques above,
channel estimation is done from the received signal, and an additional copy
of the transmitted signal is not required as with predistortion.

3.5 System Architecture

Apart from academia, different companies have developed prototypes of

full-duplex transceivers to study their feasibility in 5G communication
systems. The most popular prototyping platform is probably Wireless open-
Access Research Platform (WARP), originating from Rice University, that
have been used, e.g., in [102, 124]. In addition, National Instruments (NI)
FlexRIO platform [193] and proprietary platforms [194] have been
employed for prototyping.

Universal Software Radio Peripherals (USRPs) are popular

development platforms in academia. Implementation of full-duplex
transceiver using USRPs is possible but challenging without Field
Programmable Gate Array (FPGA) programming. If baseband signal
processing is done with a general-purpose computer instead of using FPGA,
the timing of the transmitted data packets cannot be fully controlled.
Therefore, arrival times of self-interference signal to the receiver side are
varying, which degrades the performance of self-interference cancellation.
This applies to software-defined radios in general, and therefore
implementation of digital baseband using FPGA is recommendable. Full-
duplex protyping requires rather good hardware, because transmitter
distortions may cause large self-interference levels due to the large power
difference between the transmitted and received signals.

Let us consider the direct-conversion transceiver sketched in Fig. 2.4.

Cancellation of the self-interference is based on subtracting a replica of the
transmitted signal from the received signal, and the location where the
replica is tapped and injected makes a difference in performance.
Alternatives for interference cancellation architecture in full-duplex
transceiver are at least the following:

1. Place to tap the signal in the transmitter chain:

Analog RF at PA output
Analog RF at PA input
Analog baseband
Digital baseband
2. Place to inject the signal in the receiver chain:
RF at LNA input
RF at LNA output
Analog baseband
Digital baseband

There are already 16 options for cancellation, and many of the options
can be combined together to obtain even more variations. There are
solutions that span multiple domains: for example, if the transmitted signal
is tapped in the digital baseband and injected at RF in the receiver, the
cancellation signal needs to pass through a Digital-to-Analog Converter
(DAC) into analog baseband, and subsequently up-converted to RF. In this
case, the self-interference channel can be modeled in either domain. To
ensure self-interference cancellation in the RF domain, the amplitude, phase
and delay of the cancellation signal need to match the self-interference
signal as closely as possible. Mitigation of self-interference in the RF
domain is not perfect, and the analog canceller may create additional
distortion depending on how well the generated replica signal is matched.
System bandwidth plays a decisive role in the complexity of RF
cancellation because the number of multipath components that are required
in the cancellation architecture depends on the bandwidth. Some system
architectures presented in the literature are introduced in the following
discussion.

3.5.1 Pure RF and Digital Cancellation

Figure 3.3 shows the architecture that implements multipath cancellation
using first a coarse channel model in the RF domain and then multipath
cancellation in the digital baseband [102, 174, 185]. The channel model in
the digital baseband is more accurate than that in the RF domain, because
the digital baseband allows more extensive signal processing. The
cancellation signal is tapped at the output of the transmitter PA, which
ensures that the cancellation signal contains all the impairments so that
most of those imperfections are cancelled, depending on the accuracy of the
channel model. At the transmitter side, this reduces the issue of DAC
quantization noise, upconverter phase noise and PA nonlinearity. At the
receiver side, cancellation occurs at the LNA input, and therefore the
dynamic range of the entire receiver chain is relaxed, alleviating the
requirements on LNA nonlinearity, downconverter phase noise and ADC
quantization noise.

Figure 3.3 Self-interference cancellation architecture tapping the transmitted signal at PA output
and injecting the replica at LNA input

The coarse RF channel model is designed to include direct crosstalk of

the antenna solution as well as moderately delayed self-interference signal
(up to approximately 1.4 ns). The RF channel model consists of very linear
passive delay lines that do not introduce new sources of nonlinearity. The
remaining cancellation is handled by the digital cancellation path, including
self–interference that was delayed by more than 1.4 ns through the
environment. Power-amplifier nonlinearity is also still present in the
remaining self-interference, and therefore it is accounted for in the digital
channel model by applying nonlinear self-interference cancellation.

The combination of these techniques achieves 100–110 dB of total self-

interference attenuation over MHz bandwidths [102, 174]. Also, it
provides enough cancellation already at RF to be operated with
commercially available transceivers. The major drawback of the solution
in [102] is that the RF channel model requires analog delay lines and a
circulator, resulting in a rather bulky solution. Scalability to multiple
antennas is another issue. A full-duplex MIMO transceiver would require
an RF cancellation block for each transmitter and receiver antenna pair, i.e.,
cancellation blocks. Therefore, the complexity becomes
prohibitive with a large number of antennas.

3.5.2 Hybrid RF and Digital Cancellation

Another popular self-interference cancellation architecture [124] is shown

in Fig. 3.4. Two cancellation paths are implemented. The first path taps the
signal in the digital domain, and injects the cancellation signal at RF at the
LNA input using a channel model in the digital domain that accounts for
delayed self-interference. The second path is the usual digital cancellation
path that taps and then injects the signal in the digital baseband.
 Figure 3.4 Hybrid RF/digital self-interference cancellation architecture in which the RF
cancellation signal is first constructed in the digital baseband

The attractive point of this topology is the hybrid RF/digital

cancellation signal: an accurate channel model including multipath
components can be implemented in the digital domain, and since the signal
is injected at RF, it can provide a large amount of cancellation right at the
LNA input. This reduces the dynamic range requirements on the entire
receiver chain drastically. As a result, commercially available receiver
components provide sufficient dynamic range to perform full-duplex
operation.

Transmission and cancellation are performed using two separate DACs

and upconverters, so the noise and distortion produced by these transmitters
is partly uncorrelated. At the cancellation point, these noise contributions
add up as explained in Section 2.2.2. This limits the amount of cancellation
that can ultimately be achieved with this topology. The practical limit is
reported to be 70–80 dB when using the WARP platform, mainly limited by
the uncorrelated phase noise of the two oscillators. More cancellation could
be achieved if both upconverters used the same clock signal. In [195], 10
dB theoretical improvement is reported, but since delay in the self-
interference path decorrelates the phase noise, the practical improvement
dependends highly on the characteristics of self-interference channel.

The advantage of this design is that the number of additional transmitter

chains and cancellation blocks is proportional to , because the
interference from antennas can be added up in the digital baseband, and
the multipath interference channel is easier to emulate in digital baseband
than in RF. The disadvantage is that the cancellation signal does not include
RF imperfections, which makes cancellation in digital baseband more
complicated and may limit the achieved performance. Including the
transmitter impairments in the RF cancellation and still building the
cancellation signal in the digital baseband would be possible when
implementing an additional receiver chain and converting the transmitted
RF signal back to digital baseband. This kind of receiver chain would be
needed for digital predistortion as well.
3.5.3 Analog Baseband and Digital Cancellation

Figure 3.5 shows the architecture, in which the signal is tapped at PA output
and injected in the receiver chain after down-conversion to analog baseband
[130]. The channel model in the RF domain has one tap only, i.e., it models
on phase shift and amplitude. The idea behind the architecture is that
attenuation, phase shift and down-conversion can be combined into one
functional component: a vector modulator mixer. Since the subtraction
point is located in the analog baseband, and high linearity under strong self-
interference is required, mixer and LNA are interchanged. To prevent
drowning the weak underlying desired signal in distortion at these stages,
extreme linearity is required. This can be accomplished by using a mixer-
first architecture [196] for both the main mixer and the vector modulator.
Also, the phase noise that mixes with the strong signals in both
downconverters can be an issue, but this is expected to be largely mitigated
by clocking the two down-converters with the same local oscillator signal.

Figure 3.5 Mixer-first self-interference cancellation architecture

3.6 Key Challenges

By now several protypes have demonstrated that making a compact full-

duplex transceiver with a shared antenna between transmitter and receiver
chains is indeed possible, let alone a full-duplex transceiver having separate
transmit and receive antennas. Pushing self-interference down to noise level
requires sophisticated active analog and digital cancellation techniques, and
it remains to be seen whether protypes will evolve into commercial devices.
There are several alternatives for system architecture, and the best
implementation in terms of performance and complexity may not have been
seen yet.

One way to reduce the implementation costs is to drop active Radio

Frequency (RF) cancellation and resort to passive isolation and nonlinear
digital cancellation in full-duplex devices having separate transmit and
receive antennas. A recent prototype without RF cancellation reaches
roughly 100 dB attenuation of self-interference, so this may become
feasible.

New radio systems aim to achieve ever higher carrier frequencies and
wider bandwidths, so keeping up with self-interference cancellation is
challenging. Wideband cancellation is considerably more difficult to
achieve than narrowband cancellation in the RF domain, and the
implementation of electronics may become expensive. Scalability of full-
duplex technology to large-scale multi-antenna systems should be verified
as well. Thus far, prototypes have employed up to four transmitter and
receiver antennas.
4 Full-Duplex MIMO Communications

Thanks to the various state-of-the-art approaches for self-IC schemes, SI is

no longer a critical bottleneck when implementing a practical FD system. In
this chapter, we introduce FD MIMO communications, including some key
techniques and some performance analysis. The FD MIMO advantages can
be summarized as: efficient and flexible utilization of wireless
communication resources; increasing the capacity of the communication
networks; and guaranteeing reliable communication. These have all become
crucial claims for the next generation of cellular networks. Full Duplex
(FD) is a very promising technique that allows for efficient use of wireless
communication resources, given that the self-interference level can be
suppressed to an acceptable level.

In the following, we describe a few signal processing techniques of the

FD MIMO system where two nodes bidirectionally communicate with each
other. Firstly, we describe the mode switching between FD and half-duplex
spatial-multiplexing (HD-SM). By configuring the antennas as either
transmit or receive antennas, the MIMO system can be configured as either
an FD system or an HD-SM system; these are considered two important
techniques for improving the spectral efficiency of MIMO systems. FD
transmission suffers from self interference, while the performance of an SM
system is greatly affected by spatial correlation. Therefore, there is an
optimal trade-off between FD and SM, depending on the system setting and
channel conditions. Then, we introduce the antenna pairing strategy, where
each node is equipped with two antennas, used for either transmission or
reception. Specifically, one transmit and receive antenna combination is
selected based on two system performance criteria: 1) maximum sum-rate
(Max-SR), and 2) minimum symbol-error-rate (Min-SER). We further
extend our strategy to the scenario where each node is equipped with an
arbitrary number of antennas. We describe bidirectional link selection
schemes by selecting a pair of transmit and receive antenna at both ends for
communications in each direction, to maximize the weighted sum-rate or
minimize the weighted sum SER. Then, we introduce an X-Duplex scheme,
where the antenna is adaptively configured based on the channel conditions.
The X-Duplex scheme aims to maximize the instantaneous sum-rate of the
system. Finally, we conclude this chapter and discuss some of the key
challenges in FD MIMO communications.

4.1 FD MIMO Signal Processing

4.1.1 Mode Switching between Full-Duplex and Half-Duplex

Typically, there are two kinds of baseline for FD and HD mode switching: a
fixed number of antennas and a fixed number of RFs. When the number of
antennas is fixed, the system can be adaptively switched between FD and
HD-SM, where both modes have the same degree of freedom. On the other
hand, when the number of transmit and receive RFs is fixed, the number of
active RFs in FD can be twice that in HD,1 thus the degree of freedom of
FD is potentially doubled. In this section, we only consider the first kind;
details of the second kind can be found in [197].

The system model is illustrated in Fig. 4.1, where each node is equipped
with two antennas. In FD, one antenna is used for transmission and the
other is used for reception at each node, while, in HD-SM, both antennas at
one node are used for transmission and those at the other node are used for
reception. In FD, the SI can be partially cancelled by the aforementioned SI
cancellation techniques, and the residual interference can generally be
modeled to follow Rayleigh distribution. On the other hand, the HD-SM
system can use all of the communication links and thus avoid the self-
interference problem; however, it may suffer from the sum-rate loss caused
by the spatial correlation. As depicted in Fig. 4.1, there are six links overall,
comprising four communication links and two self-interference links. The
communication links between the two nodes can be denoted as ,
, the entry of which follows independent identically
distributed (i.i.d.) Rayleigh distribution with zero mean and unit variance,
while the residual self-interference channel is , , with zero mean
and unit variance.

Figure 4.1 Mode switching between FD and HD-SM ([199] Reproduced with permission from the
IEEE.)

In FD, the two nodes employ two communication links to exchange

information. After self-interference cancellation, both nodes receive a
combination of the signal transmitted by the other node, the residual self-
interference, and noise

(4.1)
where ρ and η denote the average SNR and INR, respectively. , denote
the signal transmitted from nodes A and B, and denote the AWGN
with zero mean and unit variance at the corresponding nodes.

At the same time, when the HD-SM system is employed, assuming time
division duplex here, each node processes one half of the time resources for
transmission through a channel H. We first consider thee received
signal at node B, which is given by

(4.2)

where is an AWGN vector with unit variance. is the signal vector

transmitted from node A.

The HD-SM system can avoid the self-interference problem, but, it

suffers from the sum-rate loss caused by spatial correlation. Under the
separable correlation model [198], it is possible to express the channel
matrix as , where is an i.i.d. zero-mean complex
Gaussian random matrix with unit variance, and and represent the
normalized correlation matrices at receiver and transmitter, respectively.
Let’s assume a block-fading reciprocal channel, where the channel varies
independently at different time slots and the forward channel is reciprocal
with the reverse channel. Because the two nodes process one half of the
time resource, transmit/receive correlation matrices at one node are
receive/transmit correlation matrices at the other node. Thus, the received
signal at node A is symmetric, and the average achievable rates for the two
nodes are equal. For simplicity, in the following sections, we only consider
node B as the receiver.

Average Sum-Rates
Due to symmetry, the average rates of the two directions are equal, thus the
average sum-rate is a doubled rate of one direction
 (4.3)

After some manipulation, the closed form expression can be expressed

as [199]

(4.4)

where is the exponential integral function of the first order.

Next, we give the performance of HD-SM. We consider the case of HD-

SM with power constraint ρ and equal power allocation at the transmitter.
The eigenvalues of the transmit and receive correlation matrices, and
, are represented as and , respectively. Therefore, the
sum-rate of the correlated HD-SM system can be represented as [198]

(4.5)

where , , are matrices with entries given by

(4.6)
 As a result, combining the exact sum-rates of the FD mode (4.4) and the
HD-SM mode (4.5), we can define

(4.7)

The system can adaptively select between FD and HD-SM modes

depending on . That is, the system selects HD-SM, if ;
otherwise, it selects FD. The criterion is determined by three statistical
factors: the average INR η in FD, the spatial correlation , in HD-SM,
and the average SNR ρ. However, it is rather complicated to determine the
exact criterion through these three factors. In the following, we provide the
approximate results to facilitate the mode switching.

Approximate Results
We first calculate the approximate average sum-rate of the FD system at
high SNR. Using the approximation and
for [200], the approximate sum-rate of the FD system in (4.4)
can be rewritten as

(4.8)

where is Euler’s constant [200]. Furthermore, at low or

moderate INR, i.e., , (4.8) can be well approximated by
 (4.9)

The first term is the high-SNR sum-rate of the FD system without residual
self-interference, and the second term denotes the rate degradation caused
by the residual self-interference.

On the other hand, at high SNR, the average sum-rate of the HD-SM
system can be well approximated by [201]

(4.10)

The first term is the high-SNR sum-rate for independent fading HD-SM.
The second and third terms represent the rate degradation due to the spatial
correlations at transmit and receive sides, respectively. It is worth noting
that and are less than one, and a larger correlation
implies a smaller .

Similarly, we can further derive an approximate selection criterion by

calculating the difference between the approximate average sum-rates in
(4.9) and (4.10)

(4.11)
Given the high-SNR assumption, is just determined by two factors:
monotonically increasing with spatial correlation in the HD-SM system,
and monotonically decreasing with the average INR η, but independent of
SNR ρ.

Let , and we can derive the approximate selection threshold.

We consider the exponential correlation model [198] with equal correlation
coefficients at both sides

(4.12)

Then, the selection threshold of spatial correlation can be given by

(4.13)

Given , the system selects HD-SM, if ; otherwise, it selects

FD. If , is always positive, i.e., FD is always superior to HD-
SM even if the MIMO channel is independent. is the value that satisfies

(4.14)

By using [200] for , we can

obtain
 (4.15)

Or equivalently, we have .

The derivative of can be written as

(4.16)

where . It is obvious that , and can be given

by

(4.17)

where the first inequality is obtained by using [200] for

. Thus, we can obtain , which implies that the spatial
correlation threshold in (4.13) monotonically increases with self-
interference.

Fig. 4.2 depicts the exact and approximate switching thresholds between
FD and HD-SM, where we assume without loss of
generality. The exact threshold is obtained by comparing the exact sum-
rates of the FD and HD-SM systems in (4.4) and (4.5). The approximate
threshold is given in (4.13). Above the threshold is the area where the FD
system outperforms the HD-SM system, whereas the HD-SM system
achieves a better performance in the region below the curve. Moreover,
through simulation, we can obtain , which accords with the
aforementioned analysis . Through this figure, we
find that at high SNR, the approximate threshold is effective to determine
whether the FD system or the HD-SM system should be employed in a
specific condition.

Figure 4.2 Exact and approximate switching thresholds, with SNR ([199] Reproduced
with permission from the IEEE.)

4.1.2 Antenna Pairing

In the following, we present an antenna pairing strategy [202, 203], where

the functions of the antennas are not predefined, and the system can
adaptively configure the antennas for transmission or reception based on the
given criteria. In this subsection, we first introduce the antenna pairing
scheme in a bidirectional FD system, then we extend it in a more
general system in the next subsection.
 As shown in Fig. 4.3, the bidirectional FD system consists two
nodes, and , where each node is equipped with two antennas, which
can be dynamically configured for transmission and reception. There exist
four kinds of possible combinations (patterns), as depicted in Fig. 4.3. The
channels between the two nodes exhibit independent Rayleigh fading. We
use to denote the channel of pattern from node to , ,
, , in Fig. 4.3. We assume that they are all complex-
Gaussian random variables with zero mean and unit variance, i.e.,
. The transmit powers of the two nodes are equal, denoted by
.

Figure 4.3 System Model of the Antenna Pairing Scheme ([202] Reproduced with permission
from the IEEE.)

The received signal at node can be expressed as

(4.18)

where denotes the transmitted signal from node in pattern j. The

second term represents the residual interference after cancellation.
follows , and η denotes the interference cancellation factor,
indicating the cancellation capability. A small η indicates a good
performance of interference cancellation. denotes the additive white
Gaussian noise (AWGN) at node i, with zero mean and unit variance.
Therefore, the instantaneously received signal-to-interference-plus-noise
ratio (SINR) of node in pattern j can be expressed as

(4.19)

The probability density function (PDF) of is given by

(4.20)

Furthermore, we have , . Thus, pattern ,

and pattern , have the same instantaneous SINR at each node,
respectively. For simplicity, we only consider pattern and for the rest
of this subsection. In the following, we present the maximum sum-rate and
minimum SER criteria.

Maximum Sum-Rate Criterion

In the Max-SR approach, the pattern is selected with a higher sum-rate of
the two nodes

(4.21)

where represents the rate of node in pattern

. Thus, the sum-rate of the Max-SR approach can be given by
 (4.22)

By using , for
2, we derive an analytical upper bound

(4.23)

After some manipulation, (4.23) can be rewritten as

(4.24)

In the high-SNR regime, we can obtain

(4.25)

The first term of (4.25) is the sum-rate of traditional FD in the high-SNR

regime, while the second term, which is easily proved to be positive,
represents the improvement of the sum-rate ceiling by using Max-SR
approach.

We define , and we have . It

indicates that when the residual self-interference gets smaller, the
improvement of the sum-rate increases. Furthermore, when the self-
interference is cancelled properly, i.e., , an upper bound of can be
obtained, by using ,
 (4.26)

Substituting (4.26) into (4.25), we obtain that the improvement of the sum-
rate increases with the improvement of the interference cancellation ability,
i.e., the decrease of η, and it is upper bounded by 2.

Fig. 4.4 plots the sum-rate versus the transmit power for the FD with
or without the antenna pairing scheme. The exact sum-rate curves are given
by Monte Carlo simulations with the interference cancellation factor
. The upper bound of the Max-SR is depicted
by (4.24). Fig. 4.4 shows that the Max-SR achieves a sum-rate gain, and
converges to a higher sum-rate ceiling than that of the traditional FD. In
addition, with the improvement of interference cancellation, i.e., the
decrease of η, the sum-rate gain increases correspondingly.
 Figure 4.4 Analytical and simulated sum-rate versus transmit power of Max-SR and traditional
FD, with ([202] Reproduced with permission from the
IEEE.)

Minimum SER Antenna Criterion

In the Min-SER criterion, the pattern is selected with a lower sum SER of
the two nodes

(4.27)

where denotes the SER at from the other node in pattern . The
average SER of the Min-SER approach is conditioned on the instantaneous
SINR and can be written as

(4.28)

where is the Gaussian-Q function, and b is a constant determined by the

modulation format, e.g., for binary phase-shift keying (BPSK).

In most cases, the sum SER of the two nodes is dominated by the larger
one, and thus we consider a simplified selection criterion for which the
larger instantaneous SER is minimized so as to minimize the average sum
SER in a suboptimal way, referred to as a Min-Max selection
criterion [204]. The selected pattern is denoted by , and we have

(4.29)
Since is a monotonic decreasing function, the Min-Max selection
criterion in (4.29) can be further written as

(4.30)

As the performance of the Min-Max criterion is almost the same as the

optimal criterion, which will be verified through the simulations, we then
analyze the average SER based on the Min-Max criterion. Let
denote the instantaneous SER at node , based on the Min-Max criterion.
The average SER based on the Min-Max criterion can be written as

(4.31)

Furthermore, as earlier discussed, the sum SER (4.31) can be lower-

bounded by using the larger SER

(4.32)

where we define , and based on (4.20), the pdf of

is given by the order statistics [205]

(4.33)

Its corresponding cdf can be written as

 (4.34)

By substituting (4.34) into (4.32), we can obtain

(4.35)

At high SNR, we have

(4.36)

From (4.36), it indicates that the SER of the Min-Max criterion also
converges to an error floor.

We also provide the average SER of traditional FD without antenna

pairing for comparison

(4.37)

At high SNR, the SER of the traditional FD without antenna pairing can be
rewritten as

(4.38)
By using Taylor expansion, the SER lower bound of the Min-Max criterion
(4.36) and the average SER of traditional FD (4.38) can be rewritten as

(4.39)

It clearly indicates in (4.39) that the error floor of the antenna pairing
scheme based on the Min-Max criterion is limited by , while that of
traditional FD is limited by η. Assuming , we have ,
and the improvement of the average SER significantly increases as η gets
smaller.

Fig. 4.5 compares the Min-SER approach with the traditional FD. The
simulations are performed for a BPSK modulation, with
. The simulated SER based on the Min-Max
criterion is also included, showing that the Min-Max criterion achieves
almost the same performance as the optimal criterion. We also include the
analytical approximation using only the larger node’s SER (4.35), which is
tight at medium to high SINR. Another important observation is that the
SER curves converge to different error floors. The Min-SER approach has
much lower error floor than the traditional FD, and the average SER gain
gets larger with the decrease of the residual self-interference.
 Figure 4.5 Analytical and simulated SER versus transmit power of Min-SER and traditional
FD, with ([202] Reproduced with permission from the
IEEE.)

4.1.3 FD MIMO Communications

In this subsection, we extend the scheme in the previous subsection. We

consider a general FD MIMO system with and antennas equipped at
two nodes. Such an FD MIMO system will create possible links
between the two FD MIMO nodes, with one possible link representing the
channel from a transmit antenna of a node to a receive antenna of the other
node. Since FD radios enable simultaneous bidirectional information
exchanges between two FD MIMO nodes, a fundamental question arisen in
such a system is how to select the link for each direction to optimize the
system performance. In this chapter, we consider two performance metrics:
weighted sum-rate maximization and weighted sum SER minimization. The
optimal3 approach requires an exhaustive search from all possible antenna
links: however, as the number of antennas increases, such a brute-force
search bears very high complexity in the selection process. To resolve this
issue, in this subsection we introduce a simple Serial-Max selection
algorithm by selecting the link with optimal performance for each direction
in a two-step serial way, which can achieve asymptotically optimal
performance. By using the law of total probability and order statics, the
probability distribution functions of the two selected links are calculated, on
the basis of which the closed-form expressions on average weighted sum-
rate and sum-SER are derived. We show that the Serial-Max method
approaches the brute-force search method in terms of the average weighted
sum-rate and sum SER as the number of antennas increases.

The system model is depicted in Fig. 4.6, where node A and B are
equipped with and antennas, respectively. Both nodes use the same
frequency band at the same time for FD operation. Each node employs only
one transmit and one receive RF chains, and any antenna can be configured
to connect to either the transmit or receive RF chain. In the simultaneous
bidirectional link selection (SBLS) scheme, two antenna links are selected
for simultaneous bidirectional communication by selecting a pair of
transmit and receive antennas at both ends. Within each antenna pair, one
antenna is selected for transmission and the other is for reception.

Figure 4.6 Full-duplex MIMO systems with simultaneous bidirectional link selection ([203]
Reproduced with permission from the IEEE.)

We assume that the links between the two nodes are reciprocal and
subject to independent Rayleigh fading, and together denoted by a
channel matrix . The entry
represents the fading coefficient from the i-th antenna at node A to the j-th
antenna at node B, and it follows the circularly symmetric Gaussian
distribution . All the possible communication channels are
assumed to follow non-selective independent block fading, where the
channel coefficients remain constant during a time slot and vary from one to
another independently. In the beginning of each time slot, all the possible
communication links can be estimated perfectly.

Since the two nodes are operated in FD mode, there exists self-
interference caused by signal leakage from the transmit antenna to the
receive antenna at the same node. We assume that passive propagation
suppression and active analog/digital cancellation techniques are employed
to cancel the self-interference. The direct link of the self-interference can be
effectively suppressed [124, 206], and the residual self-interference can be
approximated to follow the Rayleigh distribution [207]. Due to practical
constraints, we assume that only the distribution of the residual
interference, including the mean and variance, can be obtained and used in
the selection process. Therefore, the selection process is based on the
“obtainable SINR.” This is equivalent to that based on the instantaneous
SNR, and it will later be shown that the former is a scaled version of the
latter.

In the SBLS, the bidirectional links, i.e., , from the -th antenna
at node A to the -th antenna at node B, and , from the -th antenna
at node B to the -th antenna at node A, are selected. The received signals
at node A and B, denoted by and , can be expressed as

(4.40)

where and denote the links corresponding to the selected antenna

pair from node A to B and that from node B to A. is the transmit power at
each node. The second term denotes the residual interference, at nodes A
and B, respectively. We assume that both residual interference links, and
, are subject to Rayleigh fading with zero mean and variance . The
AWGN at nodes A and B are denoted by and , which both follow
.

Without the knowledge of the instantaneous residual interference, the

selection procedure is based on the “obtainable SINR” matrix . The
entry in the “obtainable SINR” matrix is defined as , where
is the instantaneous SNR and is the average INR. follows
an exponential distribution with mean , and the average INR is
given by , where η denotes the cancellation ability. The
“obtainable SINR” is a scaled version of the instantaneous SNR ,
because we assume that all the links have the same average INR .

In the following, we first calculate the instantaneous performance based

on the instantaneous residual interference, and then average it with the
distribution of fading channel and residual interference channel. We define
the instantaneous SINR and for the two selected
links. where the instantaneous INR and are both exponential
random variables with mean .

Simultaneous Bidirectional Link Selection (SBLS)

SBLS Based on Weighted Sum-Rate Maximization Criterion (Max-

WSR)
We describe SBLS based on Weighted Sum-Rate Maximization criterion
(Max-WSR) under the Gaussian input assumption. Under this criterion, two
communication links from the -th transmit antenna at
node A to the -th receive antenna at node B and the -th transmit antenna
at node B to the -th receive antenna at node A, are selected to maximize
the weighted bidirectional sum-rate
 (4.41)

where denotes the rate under the “obtainable SINR”

, and is the given weight of the transmission from node
A to B, depending on the rate requirement or quality of service (QoS) of
each user.

SBLS Based on Weighted Sum SER Minimization (Min-WSER)

In SBLS based on weighted sum SER minimization criterion (Min-WSER),
under the assumption that the input signal is modulated with finite
constellations, the bidirectional antenna links are selected to minimize the
weighted sum SER

(4.42)

where represents the SER under the “obtainable

SINR” .

Serial-Max SBLS
The aforementioned SBLS schemes for the Max-WSR or Min-WSER
criteria both require a brute-force search in order to find the optimal
antenna pairs. This will become highly complex in selection process as the
number of antennas increases. We introduce a low-complexity selection
method, referred to as the Serial-Max method, which selects the antenna
pairs with maximum “obtainable SINR,” or equivalently, the maximum
SNR, in a two-step serial way.

In the first step, the best link with the maximum “obtainable SINR” is
selected

(4.43)
We use to denote the “obtainable SINR” of the selected link
, and to denote the SNR of this link. By removing the -th
column and -th row from the “obtainable SINR” matrix (and the
corresponding SNR matrix ), we can obtain a
submatrix (and a corresponding pruned SNR matrix ).

In the second step, the link with the maximum SNR is then
selected from the pruned submatrix

(4.44)

We use to denote the “obtainable SINR” of the second selected

link by the Serial-Max method, which is the maximum element of , and
we use to denote the SNR of this link, which is the maximum element
of the corresponding pruned SNR matrix .

To maximize the weighted sum-rate or minimize the weighted sum

SER, the time-shared scheme can be employed, which allocates a fraction α
of the time to use the best link for node A’s transmission, and the rest (
) to use the best link for node B’s transmission. Given w, the weighted
maximization sum-rate problem can be solved by calculating the allocation
fraction α. We have

(4.45)

To maximize R, we first calculate the derivative

(4.46)
It is obvious that is positive, therefore the allocation
factor α depends on wether or not. Specifically, we have

(4.47)

Then, the weighted sum-rate can be rewritten as

(4.48)

The weighted sum SER minimization problem can be solved in a similar

way, and we have

(4.49)

It is shown from (4.48) and (4.49) that the best link selected in the Serial-
Max method will be allocated with the greater weight in the weighted sum-
rate and SER expression.

It can be proved that the Serial-Max method can achieve the optimal
weighted sum-rate and sum SER performances simultaneously, if the
“obtainable SINR” of the second selected link is the second or third
largest element of the “obtainable SINR” matrix .

We use and to denote the respective rate and SER of the link
with the i-th largest “obtainable SINR.” If the second selected link
corresponds to the second largest element in the “obtainable SINR” matrix
, the weighted sum-rate is given by
. It is obvious that in this case the Serial-Max method can achieve the
optimal performance in terms of weighted sum-rate. Similarly, it can be
easily proved that the Serial-Max method can achieve the optimal
performance of SER.

Recall that the pruned matrix is obtained by removing the row and
column where the largest element, i.e., the first selected link, is located.
Meanwhile, the second selected link is the largest element of the pruned
matrix . Therefore, if the second selected link corresponds to the third
largest element in the original “obtainable SINR” matrix, it implies that the
second largest element in is removed in the aforementioned manipulation.
In other words, the two links associated with the first and second largest
elements in share the same antenna, which cannot be selected
simultaneously. In this case, the largest and third largest elements provide
the best option for the two selected links. Therefore, the Serial-Max
method, which selects the largest and third largest elements in , achieves
the optimal performances of weighted sum-rate and sum SER.

In addition, the probability that the Serial-Max method does not achieve
the same performance as the optimal methods, denoted by , is upper-
bounded by

(4.50)

is upper-bounded by the probability, denoted by , that the

“obtainable SINR” of the second selected link is not the second nor the
third largest element of the “obtainable SINR” matrix . This implies that
both the second and third largest elements are in the same row or column as
the largest element, and they are both removed in the process of obtaining
the pruned “obtainable SINR” matrix . Note that the elements of are
independent and identically distributed. Due to symmetry, each element of
has the same probability of being the i-th largest element. Thus, we have

(4.51)
Combining the fact that is no more than , (4.50) can be proved.

It is shown from (4.50) that the upper bound of decreases

quadratically as the number of antennas increases, which implies the
probability that the Serial-Max method selects the same pairs as the optimal
one will increase, and thus approaches the optimal performance in terms of
weighted sum-rate and sum SER asymptotically.

In addition to the asymptotically-optimal performance, the complexity

of the Serial-Max method is much simpler than the exhaustive search
approach, as shown in Table 4.1. For the Serial-Max method, in the first
step, the maximum “obtainable SINR” is selected from a matrix,
and comparisons are required. Similarly, for the second step
comparisons are needed, and the Serial method overall
needs comparisons. By contrast, the optimal method
requires an exhaustive search in order to find the optimal antenna pairs,
leading to comparisons. Therefore, the proposed Serial-
Max algorithm can approach the optimal algorithm with significantly
reduced complexity.

Table 4.1 Complexity comparison ([203] Reproduced with permission from the IEEE.)
Optimal Selection Approach Serial-Max Approach
Complexity

To compare the computational complexity of the optimal and Serial-

Max methods, the number of required floating-point operations (flops) per
antenna is presented in Fig. 4.7. It is clear that the optimal method has a
high complexity due to the brute-force search operation, while the Serial-
Max method can provide a significant complexity reduction, especially
when the number of antennas is large.
 Figure 4.7 Complexity comparison between the optimal method and the Serial-Max method

Performance Analysis
Next, we analyze the performance of the Serial-Max algorithm. It will be
shown later in simulations that the Serial-Max method can achieve near-
optimal performance in terms of average weighted sum-rate and SER.

Probability Distributions of Two Selected Links

To analyze the performance of the Serial-Max method, the distributions of
the real instantaneous SINR corresponding to the two selected links and
are required. For simplicity and without loss of generality, we consider
the case of for the following analysis, which can be easily extended
to .

According to the description of the Serial-Max method in (4.43), the

SNR of the first selected link, is the largest order statistic among
i.i.d. exponentially distributed random variables . The corresponding link
 is used for transmission from node A to B. Thus, the CDF of can
be given by

(4.52)

The second selected link in the Serial-Max method is used for

transmission from node B to A. According to (4.44), we can obtain that the
CDF expression of the instantaneous link SINR of the second selected link,
i.e., , is given by

(4.53)

where is expressed as

(4.54)

Average Weighted Sum-Rate

Based on the CDF expressions of the two selected links and , in this
section the average weighted sum-rate of the two links is obtained. Firstly,
the instantaneous rate is calculated for given realizations of the selected
communication channel and the corresponding self-interference channel.
Then the average rate is obtained by averaging the instantaneous rate with
respect to the distributions of the channels. Under the Gaussian input
assumption, the average rate of the link with SINR γ can be obtained as
 (4.55)

where is the CDF of γ.

Specifically, the average weighted sum-rate of the Serial-Max method,

denoted by is given by

(4.56)

where and are the average rates of the two selected links,
respectively, which can be expressed as

(4.57)

and

(4.58)

where
 (4.59)

and

(4.60)

It is shown from (4.56)–(4.60) that the average weighted sum-rate of the

Serial-Max method is only determined by the average SNR and
cancellation ability η. When goes to infinity, we have
and , for , where is Euler’s constant [208].
After some manipulation, converges to a rate ceiling of .

Fig. 4.8 depicts the average weighted sum-rate of the Max-WSR and
Serial-Max methods versus SNR with different levels of self-interference
for and . It can be seen that the weighted
sum-rate expression in (4.56) perfectly matches with the simulation results.
In addition, the weighted sum-rate performance is limited by the rate
ceilings. From the figure, we find that at low the weighted sum-rate
performance for different η is quite similar, because the weighted sum rate
performance at low is SNR-limited. However, at large , the residual
self-interference will dominate the performance, and the performance is
limited by the rate ceiling caused by the residual self-interference. The
figure also reveals that the Serial-Max method achieves almost the same
average weighted sum-rate as the Max-WSR method across all SNR
regions.
 Figure 4.8 Weighted sum-rate performances of the Max-WSR and Serial-Max methods, where
, and ([203] Reproduced with permission from the
IEEE.)

Average Weighted Sum SER

The average weighted sum SER of the Serial-Max method, denoted by
, is

(4.61)

where and are the average SER of the two selected links, and
can be calculated as

(4.62)
and

(4.63)

where

(4.64)

and

(4.65)

Similarly, the SER performance converges to an error floor, when the

average SNR increases to infinity.

On the other hand, when , i.e., the self-interference is perfectly

cancelled, we can further calculate the asymptotic SER of the Serial-Max
method at high SNR. In this case, the CDF expression of can be
rewritten as

(4.66)

Also, when the self-interference cancellation is perfect, the CDF of is

rewritten as
 (4.67)

where

(4.68)

Thus, the asymptotic weighted sum SER with perfect interference

cancellation is expressed as

(4.69)

It is implied by (4.69) that, given , the diversity order of the

Serial-Max method is with .

In Fig. 4.9, the weighted sum SER performance of the Serial-Max

method is provided for different , where large η means
severe self-interference whereas small η means slight self-interference
levels. In particular, means perfect interference cancellation. Based on
that figure, it can be observed that the simulated SER performance for
tightly matches with the asymptotic result given by (4.69) at high SNR. It
can also be seen that when the self-interference is perfectly cancelled, the
diversity order of the Serial-Max method is . However,
when residual self-interference exists, the Serial-Max method has a zero
diversity order.
 Figure 4.9 Weighted sum SER performances of the Serial-Max method, where and
([203] Reproduced with permission from the IEEE.)

Fig. 4.10 compares the simulated weighted sum SER performance of

the Min-WSER and Serial-Max methods with different numbers of
antennas. Combinations of different SNR and self-interference levels (
, ) are provided. It can be observed that the gaps
between the Min-WSER and Serial-Max methods are reduced as N
increases. It also shows that the SER performance of the Serial-Max
method approaches the Min-WSER method when the self-interference is
large or SNR is small. This is because, in these cases, these two factors
dominate the SER performances of both methods.
 Figure 4.10 Simulated weighted sum SER performances of the Min-WSER and Serial-Max
methods, where and ([203] Reproduced with permission from
the IEEE.)

4.2X-Duplex: Flexible Switching between SISO

and FD

In this subsection, we introduce a X-Duplex scheme, where the antenna is

adaptively configured based on the channel conditions. The X-Duplex
scheme aims to maximize the instantaneous sum-rate of the system. By
adaptively configuring the antenna as either transmit or receive antenna, the
system can operate in either HD or FD mode. X-Duplex can select the
optimal mode to maximize the system sum-rate and avoid the serious self-
interference regions.

As shown in Fig. 4.11, we consider a two-node wireless bidirectional

communication system. We assume each node is equipped with one radio
frequency (RF) chain of the transmitter (Tx) and one RF chain of the
receiver (Rx), and has only one shared antenna, which is shared by the Tx
and Rx modules [102]. The block Rayleigh fading is considered in all links
[207], where the channels remain unchanged during a transmit slot, and
change independently from one slot to another.

Figure 4.11 The two-node bidirectional communication system mode for X-Duplex. The system
can select between these two pure duplex modes.

In the FD system, both RF chains of Tx and Rx are active. The received

signal at node i, denoted by , can be expressed as

(4.70)

where , denotes the flat fading channel for transmitter–

receiver pair between the two nodes, denotes the residual self-
interference of node i, and are the transmission power of the
nodes i and j, respectively, in the FD mode. The instantaneous system sum-
rate can be written as

(4.71)
where , , , and .

On the other hand, the received signal from the node j to the node i in
the HD mode is given by

(4.72)

The instantaneous sum-rate in the HD system can be written as

(4.73)

where .

4.2.1 Maximum Sum-Rate Criterion

In the proposed X-Duplex scheme, when the self-interference is weak, the

FD mode is superior to the HD mode. Meanwhile, the strong residual self-
interference results in severe performance degradation in FD mode, and
thus HD mode becomes a better option. To maximize the instantaneous
system sum-rate, the selection criterion can be expressed as

(4.74)

4.2.2 Performance Analysis

Next, we derive the achievable sum-rate for two nodes wireless X-duplex
bidirectional transmission. Without loss of generality, we assume =
= = P and , and the mode selection criterion in
Eq. (4.74) can be simplified to
 (4.75)

The variables γ, , follow the exponential distributions with the

expectations , , . We use the variables , .

Average Sum-Rate
The sum-rate of X-Duplex system, , can be calculated as

(4.76)

where , denote the sum-rate when the system operates in HD and

FD mode, respectively.

HD Part of Average Sum-Rate,

Based on the basic definition of sum-rate, , can be calculated as

(4.77)

where denotes the integral area where the

system

chooses the HD mode. , , are the PDFs of γ, , .

After some manipulation, the integral of , Eq. (4.77), can be

rewritten as

(4.78)
where , denotes the
conditional probability of the system choosing HD mode when and
. According to the subsection integral
method, the integral Eq. (4.78) can be further calculated as

(4.79)

where
denotes the probability of whether the system chooses the and .
And .

The probability, , can be further calculated as

(4.80)

Substituting Eq. (4.80) into Eq. (4.79), can be calculated as

(4.81)

To our best knowledge, does not have a close-form expression.

FD Part of Average Sum-Rate,

According to the symmetry of the nodes A and B, the sum-rate of X-
Duplex, can be calculated as

(4.82)

where , ,
denotes the integral
area where the system chooses FD mode. And is the probability
distributed function of , and are the sum-rate of the nodes
A and B when the system chooses the mode.

We derive the integral, , as follows,

(4.83)

where denotes the probability density of

when the system chooses the FD mode and
.

According to subsection integral method, the integral Eq. (4.83) can be

rewritten as

(4.84)
where denotes the
probability of which the system chooses the FD mode and , and
.

The probability can be further calculated as

(4.85)

Substituting Eq. (4.85) into Eq. (4.84), the sum-rate can be

derived as

(4.86)

Based on the symmetry of the two bidirectional nodes, can be

obtained from the same derivation as . The sum-rate can be
respectively calculated as
 (4.87)

Lower Bound of the Sum-Rate

Due to the complicated mode selection criterion, the integral of these two
parts, and , cannot be derived into a close-form solution. Therefore,
based on the Jensen’s inequality, we derive an analytical lower bound of the
average X-Duplex sum-rate, ,

(4.88)

where

(4.89)

The lower bound of X-Duplex sum-rate can be calculated as

 (4.90)

where and denote the HD and FD part of the lower bound.

can be further calculated as

(4.91)

And can be further calculated as

(4.92)

4.2.3 Simulation Results

As shown in Fig. 4.12, X-Duplex achieves optimal system performance and

the lower bound approaches the simulated curve as the transmission power
increases. In the whole transmit power region, the sum-rate of X-Duplex
system is superior to FD and HD. In particular, X-Duplex achieves a
significant rate gain over the pure FD and HD modes in the middle transmit
power region. Both the analysis and the simulations indicate that X-Duplex
could achieve the optimal system sum-rate.
 Figure 4.12 The sum-rate for the different duplex modes, ([247] Reproduced
with permission from the IEEE.)

4.3 Summary and Key Challenges

In this chapter, we first presented the mode switching between FD and HD-
SM, including the exact and approximate performances of the FD system
and HD-SM system in the presence of residual self-interference and spatial
correlation. The exact performance comparison showed that the system
selection criterion is determined by three statistical factors: the average INR
η in FD, the spatial correlation , in HD-SM, and the average SNR ρ.
We derived the approximate expressions, based on which an approximate
threshold at high SNR was obtained for the system selection. It showed
that, in the high SNR region, the selection criterion is related to the spatial
correlation and residual self-interference, but independent of SNR. We also
presented the case that, when the residual self-interference is weaker than a
certain value , FD can always outperform HD-SM. Through analysis, we
find . We also introduced the antenna pairing
strategy into FD MIMO systems, and proposed two effective selection
approaches, Max-SR and Min-SER, to obtain maximum sum-rate and
minimum SER, respectively. It was shown, by analysis and simulation, that
the proposed approaches achieve significant sum-rate gain and SER gain,
respectively. Then, the scenario was extended to the situation where each
node has an arbitrary number of antennas. We presented a simple Serial-
Max method with near-optimal performance. Finally, we described the X-
Duplex approach that adaptively configures the antenna as either transmit
or receive to maximize the sum-rate.

The FD MIMO system we have presented in this chapter is considered

to be equipped with only one transmit and one receive RF chain for use at
each node. To achieve more flexible and efficient FD bidirectional
communications, an FD system with more general multiple RF chains at
each node is a possible extension. Specifically, when it comes to a system
consisting of two FD nodes equipped with N antennas, M transmit RF
chains, and K receive RF chains, the system can adaptively configure the
antennas based on instantaneous channel conditions. It can configure the
connection between the RF chain and the antenna; it also can use some
other advanced MIMO techniques for each direction, such as STBC code
and beamforming. In these cases, the interference cancellation needs to be
carefully designed to incorporate the MIMO techniques.

1 In FD, all of the RFs are active, while only transmit-RFs at the transmitter and receive-RFs at the
receiver in HD are active

2 This inequality can be proved by classified discussion on the different relationships of A, B, C,

and D.

3 “Optimal” in this subsection means that this scheme can achieve optimal performance under the
practical constraint that only the distribution of the residual interference, rather than the
instantaneous one, can be obtained at each node.
5 Full-Duplex OFDMA Communications

In this chapter, we study the resource allocation and scheduling problem for
a Full-Duplex (FD) Orthogonal Frequency-Division Multiple-Access
(OFDMA) network where an FD Base Station (BS) simultaneously
communicates with multiple pairs of UpLink (UL) and DownLink (DL)
Half-Duplex (HD) users bidirectionally. We aim to find the optimal pairing
of UL and DL users for each FD link and the allocation of the OFDM
subcarriers and power to different pairs such that the sum rate of the network
is maximized. This is a traditional combinatorial problem and the optimal
approach requires an exhaustive search, which becomes prohibitively
complicated as the number of users and subcarriers increase. In this chapter,
we introduce two alternative ways of solving such a complex problem. In the
first approach we formulate the problem as a mixed-integer nonlinear
programming problem and solve it by using the dual method and Sequential
Parametric Convex Approximation (SPCA). In the second approach, we
introduce a low-complexity distributed approach based on matching theories
and present an efficient near-optimal matching algorithm for resource
allocation in FD-OFDMA systems.

5.1 FD-OFDMA Model

5.1.1 System Model

We consider a multi-user FD-OFDMA system, as shown in Fig. 1, consisting

of one FD BS and multiple uplink (UL) and downlink (DL) half-duplex
(HD) users, each with a single antenna. The UL and DL users are paired to
form an FD transceiver unit in which the UL user acts as a Tx (transmit) user
and the DL one acts as an Rx (receive) user, which communicate with the
BS simultaneously. Note that each subcarrier is assigned to at most one
transceiver unit only, but each transceiver unit can utilize more than one
subcarrier. Let denote the set of subcarriers,
the set of UL Tx users, the set
of DL Rx users, and a transceiver unit consisting of and . We
assume that the numbers of Tx users, Rx users, and the subcarriers are not
necessarily the same.

We consider a block fading channel, for which the channel remains

constant within a time slot, but varies independently from one to another.
The channel coefficient from the BS to user , and that from user to the
BS, on subcarrier k are denoted by and , respectively. The residual
self-interference coefficient after interference cancellation at the BS, and the
co-channel interference coefficient from user to user , on subcarrier k
are denoted by and , respectively. We assume that they follow the
Rayleigh distribution. Therefore, for a transceiver unit containing and ,
over subcarrier k denoted by , the received signal at the BS consists of
the signal transmitted by , the residual self-interference at the BS, and
noise, and is given by

(5.1)

where is the transmitted power of user , is the transmitted

power of the BS to user over , and are,
respectively, the propagation loss from to the BS and that of self-
interference signal at the BS. denotes the distance between and the
BS and β is the decay factor. And and are the transmitted information
symbols of unit energy from to the BS and that from the BS to ,
respectively. is the additive white Gaussian noise (AWGN)
at the BS, and is the noise variance.
 Figure 5.1 System model: full-duplex OFDMA networks ([223] Reproduced with permission from
the IEEE.)

The received signal at over subcarrier k can then be written as

(5.2)

where and are, respectively, the propagation

loss from to and from BS to . and are, respectively, the
distance between and , and that between the BS and , and
is the AWGN at .

We consider the following policy for subcarrier allocation and user

pairing.

Policy: Since the BS operates in FD mode, it can communicate with an UL

user and a DL user over each subcarrier simultaneously. Therefore, UL
users can form pairs with DL users. Then the UL user and DL user in each
pair form a FD link and communicate with BS simultaneously. In our
OFDMA system, each subcarrier can only be assigned to exactly one UL
user and/or one DL user. However, each UL user or DL user can be
assigned with multiple subcarriers.

Let us assume that the UL user and the DL user are paired over
subcarrier k. The corresponding UL rate, DL rate and bidirectional sum rate
of the OFDM transmission, denoted by , , and , respectively,
can be expressed as

(5.3)

(5.4)

(5.5)

Our objective is to maximize the overall system transmission sum-rate

subject to a set of system constraints, such as the BS and user power
constraints and exclusive subcarrier allocation constraint that each subcarrier
can only be assigned to exactly one UL user and/or one DL user. This is a
traditional combinatorial problem and the optimal exhaustive search is
prohibitively complicated as the number of users and subcarriers increase. In
next two sections, we will introduce two alternative ways of solving such a
complex problem.

5.2 Centralized Resource Management: User

Pairing, Subcarrier and Power Allocation
5.2.1 Problem Formulation

We assume that the BS and UL users are power-limited. Denoting and

as the maximum transmission powers of the BS and UL user ,
respectively, we then have the following power constraints.

BS power constraint

(5.6)
 UL user power constraint

(5.7)

To represent the Policy as a mathematical equation, we denote

as the subcarrier and user-pairing assignment. If the subcarrier
k is assigned to the user pair , then ; otherwise . Then
the policy can be formulated as the following constraint.

Joint subcarrier-and-user-pairing-assignment constraint

(5.8)

The above constraint indicates that each subcarrier should be allocated to

one and only one user-pair. However, the UL/DL user may or may not use
the assigned subcarrier, depending on the allocated power being positive or
zero. In other words, each subcarrier may practically be used by (i) one UL
user and one DL user; (ii) only one UL user; (iii) only one DL user; and (iv)
no UL/DL users. Thus, the allocated powers and jointly indicate the
transmission status of the UL/DL users over each subcarrier k. Table 5.1
summarizes the detailed UL/DL transmission status over each subcarrier by
jointly considering allocated powers, and subcarrier and user-pairing
assignments. Note also that means subcarrier k has not been
assigned to UL user AND DL user simultaneously. It only guarantees
that the transmit powers of UL user and DL user cannot both be
nonzero, i.e., AND is not allowed. For example, it is
still possible that AND because the m-th uplink user
may have paired up with the -th downlink user over the k-th subcarrier for
.

Table 5.1 The Transmission Status of and over subcarrier k when or 0 ([300]
Reproduced with permission from John Wiley and Sons.)
 Status Status
Nonzero Nonzero On On Possible (when SI is low and IUI is low) Not possible
Nonzero 0 On Off Possible (when SI is very high or DL Possible (when subcarrier k
rate is very low after BS transmit power is assigned to UL user m and
is turned on) DL user where )
0 Nonzero Off On Possible (when IUI is very high or UL Possible (when subcarrier k
rate is very low after UL user is assigned to UL user and
transmit power is turned on) DL user n where )
Zero Zero Off Off Possible only when in our Possible (when subcarrier k
system and all transmit powers have is assigned to UL user and
been used on other subcarriers; and not DL user where
possible when and )

LEMMA 1: Given in the optimal subcarrier allocation scheme that

maximizes the system sum-rate, there always exists an UL transmission or a
DL transmission or both over each subcarrier.

Proof: We prove by contradiction. Assume that there is neither UL nor DL

transmissions in a certain subcarrier , i.e., and
. As each subcarrier can be allocated to at most one UL
user, there exists at least one UL user which is not associated with any
subcarrier (because ). Since there is no transmission in subcarrier ,
can use the subcarrier for UL transmission and hence to increase the
system sum-rate. Such a scenario is contradictory to the assumption and
Lemma 1 is hence proved.

To maximize the system sum-rate, we need to optimize the power

allocation among subcarriers subject to the constraints
(5.6) and (5.7); and to optimize the joint subcarrier and user-pairing
assignments subject to the constraint (5.8). Such an
optimization problem will be solved by the BS and the resulting allocation
information will be sent to each user.

Defining as and as
, we seek the joint optimization of and to
maximize the system transmission rate subject to the power and subcarrier
assignment constraints, i.e.,

s.t. BS power constraint (5.6);

s.t. UL user power constraint (5.7) and
s.t. Joint subcarrier-and-user-pairing-assignment constraint (5.8).
(5.9)

LEMMA 2: The optimization problem in OP1 is an NP-hard problem.

Since the optimization problem in OP1 is NP-hard, it is impossible to

find both efficient and optimal algorithm for it. In the next sections, we will
introduce several efficient and near-optimal algorithms to find the solutions
to optimization problem OP1.

5.2.2 Resource Management with Full Channel State

Information (CSI)

We first consider the scenario when the BS has the access to all global CSI,
including CSI of UL, DL, BS self-interference (SI) CS and UL to DL Inter-
User Interference (IUI). The optimal solution to (5.9) requires solving
a Mixed Integer NonLinear Programming (MINLP) problem. The total
possibilities of subcarrier and user-pairing assignment significantly
complicate the problem when K, M, and N are large. Fortunately, our
optimization problem is a multiple subcarriers allocation problem and can be
solved by using the dual method [209]. Before describing the proposed
algorithms, let’s first briefly introduce the dual optimization method.

(1) Introduction to Dual Optimization Method

In many communication system designs, we often need to optimize some
objective functions subject to various system constraints. The optimization
problem becomes even more numerically difficult when the objective or
constraint functions are not convex. In this subsection, we introduce a
Lagrangian dual method, proposed in [209], for non-convex spectrum
optimization problems.

Let’s consider an optimization problem

(5.10)

where are vectors of optimization variables, are that

are objective functions, are that are the constraint functions,
and are not necessarily convex.

The dual method solves the above problem by forming its Lagrangian
dual

(5.11)

where is a vector of Lagrangian dual variables.

Then we can define a dual objective function as an unconstrained

maximization of the Lagrangian

(5.12)

The dual optimization problem is to

 (5.13)

If and are convex, then the original problem (5.10) and the
dual problem (5.13) have the same solution. However, when they are not
convex, the solution to the dual problem only provides an upper bound to the
solution of the original problem (5.10). However, as shown in [209], when
the optimization problem satisfies some conditions, the dual problem (5.13)
will have the same solution as the original problem (5.10), even and
are not convex. The condition is called the time-sharing condition.

Definition: time-sharing condition [209]: Let and be optimal

solutions to the optimization problem (5.10) with and ,
respectively. An optimization problem of the form (5.10) is said to satisfy
the time-sharing condition if, for any and for any , there
always exists a feasible solution , such that

(5.14)

THEOREM 1: Zero Gap Dual Method [209]: If the optimization problem

satisfies the above time-sharing property, then the original problem (5.10)
and the dual problem (5.13) have the same solution.

To numerically calculate the optimal , [209] developed a subgradient

method to update iteratively in the subgradient direction by designing a
step-size sequence as follows:
 (5.15)

where l is the iteration number, is a sequence of scalar step sizes, and

is defined as . The algorithm can guarantee to converge to the
optimal as long as is sufficiently small. A common criterion for choosing
the step sizes is that they must be square summable, but not absolute
summable. When the norm of the subgradient is bounded, the following
choice:

(5.16)

for some constant β can guarantee the convergence to the optimal .

Next, let’s apply the dual method to solve the previous non-convex
resource allocation problem. We show that the duality gap is negligible when
the number of subcarriers becomes sufficiently large, i.e., the solution is
asymptotically optimal when K is large. We introduce a two-layer iterative
algorithm called Dual-SPCA algorithm, which includes three main steps: (i)
problem transformation; (ii) optimizing the dual function at given primal
variables; (iii) optimizing primal variables at a given dual point. Details of
the algorithm are described as follows.

(2) Problem Transformation

The problem under consideration is a mixed combinatorial and non-convex
optimization problem. The combinatorial nature comes from the integer
constraint for the subcarrier and user-pairing assignment strategy. In general,
a brute force approach is needed to obtain the global optimal solution. To
tackle this problem, we first relax the integer constraint in (5.8) and show
that the relaxed problem has the same solution as the original optimization
problem in (5.9). This is summarized in the following lemma.
LEMMA 3 (Optimality of the time-sharing relaxation): If we relax the
subcarrier and user-pairing variable in (5.8) to be a real value
between zero and one, i.e., , instead of a Boolean value, then
the relaxed problem has the same solution as the original optimization
problem in (5.9).

Proof: The above lemma can be proved by using a similar approach as in

[209, 210].

To facilitate time-sharing relaxation of subcarrier and user-pairing

variable , we introduce two “virtual powers” variables and ,
defined as

(5.17)

The reason for introducing virtual powers is to eliminate the integer factor
and to decompose the optimization problem OP1 into a solvable dual
problem.

The UL rate and the DL rate of the transmission over subcarrier k can be
rewritten as

(5.18)

(5.19)

Then the BS and UL power constraints can be rewritten as follows:

 BS power constraint

(5.20)

UL user power constraint

(5.21)

Hence the problem OP1 can be transformed equivalently to

(5.22)

s.t. (5.8), (5.20) and (5.21)

where is defined as . It can be seen that

the integer constraint for is eliminated.

(3) Optimizing the Dual Function

We define as the set of possible subcarrier and user-pairing assignments
satisfying (5.8). In addition, we define as the set of possible power
allocations under the given . Then, the Lagrange dual problem can be
readily written as [211]

(5.23)

where the Lagrangian is

 (5.24)

and represents the vector of the dual variables associated

with the individual power constraints .

Since a dual function is always convex by definition [211], gradient or

subgradient-based methods can be used to minimize with a guaranteed
convergence. We let be the optimal power allocation at a dual point (to
be discussed in the next subsection). Then a subgradient of can be
derived using a similar method as in [209] and the dual variables can be
updated as

(5.25)

(5.26)

where denotes , the superscript (l) denotes the iteration number,

and represents the step size. When the step size follows the
diminishing policy in [212], the subgradient method above is guaranteed to
converge to the optimal dual variables. Here, we just take the simple
diminishing step, i.e., , where is the initial step size.
The computational complexity of such an updating method is polynomial in
the number of dual variables [209].

(4) Optimizing Primal Variables at a Given Dual Point

Computing the dual function involves determining the optimal
at the given dual point . In this subsection, we present the detailed
derivation of the optimal primal variables in three phases. Before that, we
rewrite (5.23) as

(5.27)

where

(5.28)

A. Optimal Power Allocation for a Given Subcarrier and User-Pair

Assignment
Let’s analyze the optimal power allocation for a given subcarrier and
user-pairing assignment . We suppose that a user-pair is assigned with
the subcarrier k, i.e., . Then the optimal power allocation over this
combination can be determined by solving the following problem:

(5.29)

(5.30)
Since is non-convex, it is very hard to obtain a closed-from expression
of and based on the Karush–Kuhn–Tucker (KKT) conditions
[211]. Since the power control algorithm is not our main concern, we simply
utilize a local optimization method, such as Successive Parametric Convex
Approximation (SPCA), to maximize and solve the problem
iteratively. The SPCA method that maximizes is shown in the
appendix at the end of this chapter. The SPCA method has been proved to
satisfy the KKT conditions [213], which form a sufficient condition to apply
the dual method [209]. Finally, the optimal value of the Lagrangian function
for a given subcarrier and user-pairing assignment is set as .

B. Optimal Subcarrier and User-Pairing Assignment

For a given subcarrier k, we select a user-pair as
through exhaustive search of all possible combinations. Then we set
and for or . It is not hard to verify that
(5.8) is satisfied subsequently. Finally, the virtual powers and
are set as zeros.

(5) Optimality Analysis

In the outer iteration of Dual-SPCA algorithm, the dual method is used to
optimize the dual variables based on a gradient-based method. In addition, in
the inner iteration of our proposed Dual-SPCA algorithm, the SPCA method
is guaranteed to converge to a KKT point [213]. Since a multi-subcarrier
allocation problem solved by using the dual method is asymptotically
optimal given the inner problem reaches a KKT point [209], we can
conclude that the Dual-SPCA algorithm is near-optimal.

5.2.3 Resource Management with Local CSI

In the previous section, we have assumed that the BS has the global CSI
information. The UL CSI and SI CSI can be estimated by the BS directly.
However, the DL CSI and Inter-User Interference (IUI) CSI are measured by
the DL users and then fed back to the BS. Each of the N DL users is required
to measure K DL CSI and MK IUI CSI. Since , measuring the IUI
CSI and sending the information back to the BS requires considerable
overhead. To overcome this problem, in this section we investigate a
suboptimal IUI estimation algorithm. Instead of estimating and uploading
the instantaneous IUI CSI, each user only has to report its location – based
on the global positioning system (GPS) or some other means – to the BS
periodically. We further assume that the BS has the prior knowledge of the
Probability Density Function (PDF) of the channel fading model between
users. Then, the BS can make use of the user-locations information to
estimate the IUIs and to allocate resources.

(1) Problem Relaxation

We first introduce a set of slack variables denoting the constraint of the
self-interference of BS over subcarrier k, and denoting the constraint of
the interference suffered by over subcarrier k. The vector
is referred to as the Interference-to-Noise Ratio
Temperature (INRT) vector. Compared to the problem OP1, the following
problem decouples the subcarrier and user-pairing assignment into separable
UL and DL user scheduling problem with the use of the INRT vector :

(5.31)
s.t. (5.20) and (5.21)

(5.32)

(5.33)

(5.34)

where
 (5.35)

(5.36)

, and are, respectively,

Boolean variables indicating scheduling of UL user and DL user over
subcarrier k. Specifically, indicates that the DL user is
scheduled over subcarrier k; otherwise . is also defined in a
similar way. Here, the constraint (5.32) ensures that the subcarrier
assignment and user scheduling satisfy the Policy; the constraint (5.33)
guarantees that the actual self-interference-to-noise ratio over each
subcarrier k is below the INRT level ; and the constraint (5.34) indicates
that for any DL user who is scheduled over each subcarrier k, the
corresponding interference-to-noise ratio should be below the INRT .

Since there is no instantaneous IUI CSI available at the BS side, the

constraint (5.34) may not be satisfied all the time. Hence, we consider a
more realistic interference constraint called chance-constrained IUI, which
specifies the minimum probability that the IUI interference constraint in
(5.34) is satisfied. We introduce the concept of chance-constrained resource
allocation which has been recently proposed in [214, 215]. The constraint
(5.34) is then rewritten as

(5.37)

The above inequality enforces that if DL user is scheduled over

subcarrier k, the probability of the interference suffered by the DL user being
smaller than INRT is no less than . As a result, denotes the
maximum probability that the INRT is exceeded.
LEMMA 4: Assuming that the long-term fading and the PDF of short-term
fading of the IUI channel are known, the constraint (5.37) can be rewritten
as

(5.38)

where , and is the pathloss over distance

between users and over subcarrier k.

Subsequently for a given , the problem of (5.31) with the additional

constraint (5.38) can be formulated as follows:

s.t. (5.20), (5.21), (5.32), (5.33), and (5.38)

Obtaining the optimal solution of problem requires solving

a mixed integer programming problem. In next two sections we show how,
by using the dual method, the problem can be solved iteratively with a fixed
. After solving the problem , the main problem in (5.31) can be solved
by updating the INRT vector , and we will explain this later.

(2) Optimizing Dual Function

We define as the set of possible subcarrier and user scheduling
satisfying (5.32). In addition, we define as the set of all
power allocations that satisfy (i) for , (ii) for
, (iii) for , and (iv) for . Then,
the dual problem can be readily written as
 (5.39)

where the Lagrangian is

(5.40)

represents the vector of the dual variables associated

with the individual power constraints ;
represents the vector of the dual variables associated with the individual
INRT constraints ; and represents the
vector of the dual variables corresponding to the individual INRT constraints
. Similar to Section 5.2.2, the updating rules of these dual
variables , , and are based on the subgradient method.

(3) Optimizing the Primal Variables at a Given Dual Point

By dual decomposition, the dual problem is decomposed into
sub-problems, i.e.,
 (5.41)

where

(5.42)

(5.43)

From (5.42), we can see that the sub-problem highly depends on the DL
user scheduling . However, the subproblem is independent of UL
user scheduling. Hence, we can first optimize the power allocation and user
scheduling of DL users and then optimize the power allocation and user
scheduling of UL users. The optimal solution is further described in Lemma
4.

LEMMA 5: The optimal power allocation and user scheduling for DL users
and UL users over any subcarrier k are given as follows:

(5.44)

(5.45)

(5.46)
 (5.47)

and are defined as the optimal values of and with respect to

the powers and , respectively. is the optimal DL user over
subcarrier k.

(4) Optimizing Interference-to-Noise Ratio Temperature

Having solved the problem , we then define the master problem as
the function of updating the INRT vector , i.e.,

(5.48)

(5.49)

The master problem can be solved iteratively using a subgradient method

and the following lemma suggests the subgradients for each and .

LEMMA 6: The subgradient of and in the master problem are given

by, respectively,

(5.50)

(5.51)

where is the selected UL user over subcarrier k; and are

given by (5.47) and (5.45), respectively; and are the optimal dual
multipliers corresponding to the constraints (5.33) and (5.34), respectively.
 According to Lemma 6, the BS has to update INRTs after
solving the relaxed . and are updated, respectively, using

(5.52)

(5.53)

where is a positive step size. The updating process is stopped

when and , where ζ is a
sufficiently small convergence tolerance.

REMARK: In the Dual-SPCA algorithm, all CSI information is available. The

BS can therefore exhaustively consider all possible combinations of paired
UL-DL users and determines the best user-pair over each subcarrier in order
to achieve near-optimal performance. Compared with the Dual-SPCA
algorithm with full CSI, the location-aware algorithm with limited CSI does
not need the exact IUI-CSI of each subcarrier but requires the long-term
fading and short-term fading PDF of the IUI channel. Only UL users of
those whose interference to the DL users are under the interference threshold
with a certain probability are treated as potential candidates for pairing.
Then, the best paired UL and DL users over each subcarrier are selected
under the given interference threshold. Subsequently, by updating the
interference threshold iteratively, we can find the ultimate paired users over
each subcarrier. In summary, by introducing the interference threshold in the
location-aware algorithm, we can first select a DL user over a given
subcarrier and then find a UL user that can maximize the sum-rate over the
subcarrier. This fact also leads to another advantage of the location-aware
algorithm, i.e., we can easily adopt a semi-distributed way to implement the
location-aware algorithm.

5.2.4 Simulation Results

In this section, we evaluate the performance of algorithms presented in the
previous sections under the 3GPP LTE specifications for both urban macro
(UMa) cell and small cell deployments. The simulation parameters of an
UMa cell and a small cell are taken from [216, 217] and listed in Table 5.2.
The cell coverage area is assumed to be circular. For an UMa cell, the radius
is set to 2 km and all the channels are considered to be under the non-line-of-
sight (NLOS) environment. By setting the heights of the BS and users to
31 m and 1.5 m, respectively, above ground [217], we obtain the path-loss
models of the UL channels, DL channels, and the IUI channels1 . Also,
according to the suggestion in [216], the peak power constraints for UL
users are the same and set to be 20 dB below the peak power constraint of
the BS. For a small cell, the radius is set to 100 m. Moreover, the DL and UL
channels are assumed to experience a path-loss model for line-of-sight
(LOS) communications while IUI channels are assumed to encounter a path-
loss model for NLOS transmissions [218]. Furthermore, the peak power
constraint for UL users is the same and set to be 3 dB below the peak power
constraint of the BS [216].

Table 5.2 System simulation parameters ([300] Reproduced with permission from John Wiley and
Sons.)
Basic Parameters
Cellular layout Isolated Cell, 1-sector
Noise spectral density dBm/Hz
Bandwidth 10 MHz
Number of subcarriers 32
Center frequency 2 GHz
Noise figure 9 dB
Short-term fading Standard Rayleigh fading

Parameters for UMa Cell

Cell area The radius is 2 km

Path-loss of UL and DL channels 122.5+35
Path-loss of IUI channels 146.2+39.8
20 dB

Parameters for Small Cell

Cell area The radius is 100 m

Path-loss of UL and DL channels 103.8+20.9
Path-loss of IUI channels 145.4+37.5
3 dB

In addition to Dual-SPCA algorithms, the following four approaches are

simulated:

Half-Duplex Single User Selection (HD-SUS): For each subcarrier,

only one uplink or downlink user is selected optimally based on the
water-filling algorithm [219, 220]. In this approach, even if the total
number of users is larger than the number of subcarriers, at most K
users are allowed to access to the network simultaneously.
Equal Power Allocation (EPA): In this approach, we firstly set equal
power for all subcarriers and select the best user-pair over each
subcarrier. Supposing the m-th UL user is assigned with a total of
subcarriers, we then divide the total transmit power equally
among these subcarriers, i.e., we set if .
Random User-Pair Selection (RUPS): For each subcarrier, a random
uplink and downlink user-pair is selected. Then power allocation is
optimized at the given subcarrier and user-pairing assignment.
Separately Maximizing UL transmissions and Maximizing DL
transmissions (SMUMD): For each subcarrier, one uplink user and
one downlink user are selected based on the water-filling algorithm.
Then the UL transmissions and DL transmissions are maximized
separately without taking IUI into account. In other words, SMUMD
optimizes subcarrier assignment and power allocation without taking
IUI into account.

A total of 5000 different channel realizations have been used. The

number of subcarriers is set to . The number of UL and DL users is
set to , which is the same as the number of subcarriers.

(1) Average Sum-Rate Performance of the Dual-SPCA Algorithm with

Full CSI at the BS
We first study the average sum-rates achieved by different schemes with
different SI attenuations τ under both UMa cell scenario and small cell
scenario. The results under the UMa cell scenario and the small cell scenario
are plotted in Figs. 5.2 and 5.3, respectively. Based on the curves, we have
the following observations.

The average sum-rates of all algorithms increase with the BS transmit

power.
As expected, the average sum-rate of the HD-SUS algorithm is
independent of SI attenuation.
Under FD transmissions, the average sum-rates of Dual-SPCA, EPA,
RUPS and SMUMD algorithms increase with SI attenuation.
Dual-SPCA algorithm outperforms the other four baseline approaches,
i.e., HD-SUS, EPA, RUPS and SMUMD algorithms. Among the four
FD transmission algorithms, their relative performance in decreasing
order is as follows: Dual-SPCA, EPA, SMUMD, RUPS.
When SI is large (e.g., 120 dB), the Dual-SPCA algorithm can
outperform the HD-SUS algorithm substantially. In particular, with
25 dBm BS transmit power and 120 dB SI attenuation, the proposed
Dual-SPCA algorithm achieves about 60% average sum-rate
improvement compared with the HD-SUS approach under an UMa
cell environment and about 20% improvement under a small cell
environment. The results also show that besides SI attenuation, other
parameters such as cell environment, also affect the performance of
FD transmissions.
When SI attenuation is not large (e.g., 80 dB), Dual-SPCA algorithm
only slightly outperforms the HD-SUS algorithm. When the SI
attenuation is not large, the transmitted DL signal at the BS creates
substantial interference in the receive signal in the UL channel. Under
such circumstances, the effectiveness of FD transmissions is reduced
and using FD transmissions becomes not much beneficial.
Among the four FD transmission algorithms, the Dual-SPCA is the
best. Moreover, EPA and SMUMD are the second and third best,
respectively. The results show that ignoring IUI in the optimization
process (as in SMUMD) produces a larger degradation (compared
with Dual-SPCA) compared to the case without the power allocation
(as in EPA). Thus, pairing up UL and DL users appropriately by
considering IUI is an important step in FD transmissions.
 Furthermore, depending on the BS transmission power, SI attenuation
and channel environment, EPA and SMUMD may or may not
outperform HD-SUS.
For a given SI attenuation, the RUPS algorithm always provides the
lowest average sum-rate among the five algorithms. The results
indicate that randomly pairing UL and DL users followed by
optimized power allocation gives the worst performance – even worse
than HD-SUS – and hence should never be used in practice.

Figure 5.2 Average sum-rate versus BS transmit power with different SI attenuations (τ) under a
UMa cell scenario ([300] Reproduced with permission from John Wiley and Sons.)
 Figure 5.3 Average sum-rate versus BS transmit power with different SI attenuations (τ) under a
small cell scenario ([300] Reproduced with permission from John Wiley and Sons.)

Based on the above observations, we can conclude that without both

appropriate user-pairing and optimized power allocation, FD transmissions
may be worse than HD transmissions in an OFDMA multi-user cellular
system. To gain further insights into the system performance, we plot the
percentage of sum-rate for the Dual-SPCA algorithm, relative to the HD-
SUS approach in Fig. 5.4. The results show that when SI attenuation is low (
80 dB), Dual-SPCA has a little gain compared with HD-SUS. This is
because SI causes strong interference to the UL transmissions and FD
transmissions cannot work effectively. When SI attenuation increases from
90 dB to 120 dB, we can observe a steep rise in the gain. The Dual-SPCA
algorithm can provide substantial sum-rate improvements over HD-SUS.
When SI attenuation is beyond 130 dB, the gain becomes flat but less than
100%. In this case, the FD system’s achievable rate will be mainly limited
by noise instead of SI. If there were no IUI, FD transmissions can almost
double the average sum-rates of HD transmissions. However, due to the
effect of IUI, FD cannot achieve the double rate of HD systems.

Figure 5.4 Average sum-rate gain of Dual-SPCA algorithm over HD-SUS approach versus SI
attenuation for different BS transmit powers under both UMa cell and small cell scenarios ([300]
Reproduced with permission from John Wiley and Sons.)

In Figs. 5.5 and 5.6, we plot the average sum-rate of the Dual-SPCA
algorithm and SMUMD approach under a UMa scenario and a small cell
scenario, respectively. Given a BS transmission power, we observe that the
average sum-rate increases with SI attenuation and reaches a maximum after
the SI attenuation is above a certain threshold. When SI attenuation
increases, FD transmissions become more effective and hence the sum-rate
increases. Above a certain SI attenuation, the effect of SI becomes very
minimal and the change in sum-rate is negligible. Figs. 5.5 and 5.6 also
show that the Dual-SPCA algorithm always outperforms the SMUMD
approach under the same conditions. The main reason is that IUI has been
considered in the Dual-SPCA algorithm but ignored in the SMUMD
algorithm. Thus, we can conclude that IUI plays a major role in determining
the performance of a FD multiuser OFDMA cellular system.

Figure 5.5 Average sum-rate of Dual-SPCA algorithm and SMUMD approach versus SI attenuation
for different BS transmit powers under a UMa cell scenario ([300] Reproduced with permission
from John Wiley and Sons.)
 Figure 5.6 Average sum-rate of Dual-SPCA algorithm and SMUMD approach versus SI attenuation
for different BS transmit powers under a small cell scenario ([300] Reproduced with permission
from John Wiley and Sons.)

(2) Performance Comparison of the Dual-SPCA Algorithms with Global

and Local CSI
In this section, we further consider the case when the BS does not receive
any IUI-CSI information from the users. Instead, the BS makes use of the
location-aware sub-optimal algorithm to estimate the IUIs and to assign the
subcarriers to users. Furthermore, the IUI for each user over each subcarrier
should lie below the INRT threshold with a chance-constrained probability
of . The value of in fact determines the confidence level of the
system sum-rate. A larger value increases the confidence probability but at
the same time reduces the system sum-rate. Thus, the effect of the value of ϵ
needs to be investigated.

Since we focus on the effect of IUI in this section, we evaluate the

system performance at a SI attenuation of 120 dB such that the effect of SI
becomes negligible.
 We also use the “goodput,” which is defined as bits/s/Hz successfully
received [221, 222], as the performance metric when comparing the
location-aware algorithm with limited CSI and the Dual-SPCA algorithm
with full CSI. Figs. 5.7 and 5.8 plot the goodput of the two algorithms under
an UMa cell scenario and a small cell scenario, respectively. From the
results, we observe that the location-aware sub-optimal algorithm can
approach the performance of Dual-SPCA near-optimal algorithm when the
value of ϵ in the location-aware algorithm is around dB . In
Tables 5.3 and 5.4, we further plot the performance loss when dB.
Results show that the performance losses under all scenarios range between
4.7% and 8.4%. Thus we conclude that the location-aware algorithm, which
does not require precise IUI-CSI but estimates IUI-CSI based on user
locations, suffers from a small degradation in goodput compared with the
Dual-SPCA near-optimal algorithm.

Figure 5.7 Goodput versus ϵ for Dual-SPCA near-optimal algorithm and location-aware sub-
optimal algorithm under a UMa cell scenario ([300] Reproduced with permission from John Wiley
 and Sons.)

Figure 5.8 Goodput versus ϵ for Dual-SPCA near-optimal algorithm and location-aware suboptimal
algorithm under a small cell scenario ([300] Reproduced with permission from John Wiley and
Sons.)

Table 5.3 Goodput of Dual-SPCA algorithm and location-aware algorithm under an UMa cell scenario
([300] Reproduced with permission from John Wiley and Sons.)
Average Goodput (bits/s/Hz)
(dBm) Loss
Dual-SPCA Location-Aware
5 3.44 3.15 8.4%
15 8.40 7.83 6.8%
25 13.53 12.72 6%
35 18.20 17.21 5.5%
45 21.66 20.65 4.7%

Table 5.4 Goodput of Dual-SPCA algorithm and location-aware algorithm under a small cell scenario
([300] Reproduced with permission from John Wiley and Sons.)
 (dBm) Average Goodput (bits/s/Hz) Loss
Dual-SPCA Location-Aware
–15 9.29 8.54 8.1%
–5 13.84 12.89 6.9%
5 17.82 16.68 6.3%
15 21.44 20.26 5.5%
25 24.09 22.88 5.1%

5.3User Pairing and Resource Management

Based on Matching Theory

In the previous section, we formulate the user pairing and resource allocation
problem as a mixed integer nonlinear programming problem and solve the
problem by using the dual method and the sequential parametric convex
approximation (SPCA). In this section, we introduce a low-complexity
approach based on matching theories and present an efficient near-optimal
matching algorithm for resource allocation in FD-OFDMA systems [223].

5.3.1 Matching Theory Basics

In economics, matching theory is a mathematical framework attempting to

describe the formation of mutually beneficial relationships over time. It has
been especially influential in labor economics, where it has been used to
describe the formation of new jobs, as well as to describe other human
relationships like marriage. It has recently been used widely in wireless
communications. Matching theory is one general and practical tool for
analyzing wireless markets that studies the interaction of competitive agents
in a network.

The matching problems can be classified as follows: 1) Bipartite

matching problems with two-sided preferences: here, the participating agents
can be partitioned into two disjoint sets, and each member of one set ranks a
subset of the members of the other set in order of preference. Example
applications include: assigning junior doctors to hospitals, pupils to schools,
and school leavers to universities.
 2) Bipartite matching problems with one-sided preferences: here, the
agents are partitioned into two disjoint sets, but only one set of the agents
has preference over the other set. Example applications iclude: campus
housing allocation, DVD rental markets, and assigning papers to reviewers.

3) Non-bipartite matching problems with preferences: Here, all agents

form one single homogeneous set, and each agent ranks a subset of the
others in order of preferences. Example applications include: forming pairs
of agents for a chess tournament, finding kidney exchanges involving
incompatible patient–donor pairs, and creating partnerships in P2P networks.

Besides the above means of classification, we can also classify

matchings by the number of agents in each side, i.e., one-to-one matching,
many-to-one matching and many-to-many matching. In the next section, we
will apply matching theory to solve resource allocation problems in FD
communications.

5.3.2 Joint User Pairing, Subcarrier Allocation and Power

Control
(1) Problem Reformulation
We consider the same system model and setup as in Section 5.1.1. Without
loss of generality, in this section we assume that
and . We also assume that the transmitted power of the BS is
uniformly allocated to all paired subcarriers, i.e.,
.

To facilitate the representation of the pairing of users and subchannels,

we define a Tx–Rx matrix ϒ where the binary entry denotes
whether is paired with . Similarly, we introduce a Tx-
subchannel matrix Γ where the binary entry denotes whether subchannel
k is allocated to . Our objective is to maximize the sum-rate of the system
by jointly optimizing the pairing variables . We reformulate the
optimization problem
 (5.54a)

(5.54b)

(5.54c)

(5.54d)

(5.54e)

Since each subchannel can only be assigned to one UL-DL transceiver

unit, the binary variables in matrix must satisfy the constraint (5.54b).
Each transceiver unit can utilize at most subchannels, and thus, the binary
variables in matrix Γ must satisfy the constraint (5.54c). Also, each Tx user
can only be paired with one Rx and vice versa, so the binary variables in
matrix ϒ must satisfy the constraints (5.54d) and (5.54e).

Since the optimization problem in (5.54) is NP-hard, it is impossible to

find both efficient and optimal algorithm for it [224]. However, we can
observe that the optimization problem in (5.54) is equivalent to a three-
dimensional matching (3-DM) problem, which inspires us to develop a low-
complexity distributed algorithm based on the matching theory.

(2) Formulation of Resource Allocation as a One-to-One Matching

Problem
Now we formulate the optimization problem in by utilizing the
matching theory. When the BS allocates a subchannel to one transceiver
unit , we say , and are matched with each other, and the
three together compose a matching triple, denoted by .
Thus, one matching triple is formed by one UL user, one DL user, and one
subchannel.

While performing the user pairing and subchannel allocation, the BS

assumes that the UL Tx users, DL Rx users and subchannels2 are all selfish
and rational individuals who have preferences over the pairs in ,
and , respectively. We denote the set of preference lists
of the Tx users, Rx users, and subchannels as

(5.55)

where are the preference lists of , , and ,

respectively. The preference list of each UL Tx user, DL Rx user or
subchannel can be decided by evaluating the sum-rate of each potential
matching triple. The relations of preference can be represented using the
following expression:

(5.56)

which means that prefers the Rx-subchannel pair to

because the former pair can bring higher utility to it.

In the open literature, various properties of preferences have been

described and studied for various scenarios, such as the strict-ordered
preference [225], the acylic preferences [226], and the consistent
preference [227]. In our scenario, we can prove that the users’ and
subchannels’ preferences over pairs are strict, mutual, transitive, and
inconsistent, as explained below.

The potential sum-rate of any transceiver unit over any subchannel is

only determined according to equation (5.5) by the BS, so the preference of
each Tx user, Rx user or subchannel is only determined by the potential
sum-rate. That is, no object is indifferent3 between the objects in the other
two parties, and thus, we say the preference list of each UL Tx user, DL Rx
user or subchannel is strict. It is also worth noting that if
, then , because
for both and , indicating the mutuality of the preference lists. By
calling a preference list as transitive, we mean that if
and , then . The preference is
determined based on the sum-rate, so it is highly likely that prefers to
when paired with , but prefers to when paired with , i.e.,
. In this case, we say that the preference lists are
inconsistent.

With the notion of the preference list and the matching triple, we can
then formulate the optimization problem as a matching problem.

DENITION 1: Given a set of Tx users, a set

of Rx users and a set of subchannels, each Tx user, Rx
user and subchannel has a complete preference list over the pairs in
, and , respectively. Set , and
a matching Ψ is a disjoint set of at most L matching triples, denoted by
(i.e., each Tx user, Rx user and subchannel appears in
exactly one triple in Ψ) such that:

1) ;
3) ;
4) ;
5) and
;
6) and
;
7) and
,
where condition 1) indicates that , and subchannel are matched to
each other and form a matching triple. Since the numbers of three sets of
agents (i.e., Tx users, Rx users, or the subchannels) are not necessarily the
same, not all the agents will be matched. When an agent is not matched, we
say that it is matched to a pair of virtual nodes. For instance, is matched
to a virtual subchannel-RX pair if is not matched to any
subchannel or Rx, and the transmitted power of is zero. Similarly, if is
not matched to any Tx user or subchannel, is matched to a virtual Tx-
subchannel pair , and no signals will be transmitted from the BS to
. They are described in conditions 2) and 3). Note that each pair of virtual
nodes can be matched to multiple Tx users, Rx users or subchannels, since
these virtual nodes have no physical meanings. Item 4) implies that a Tx user
can be matched to either one subchannel-Rx pair or a virtual pair, and item
5) and item 6) both imply the same principle for the Rx users and
subchannels.

We now introduce an indicator variable , in which ,

, and , to better indicate the matching of the Tx
users, Rx users and the subchannels. The binary matching parameter, ,
is defined as

(5.57)

The utility of any triple is defined as the sum-rate of this

triple, , which is shown in equation (5.5). Therefore, we can define the
social welfare of the whole network as the total utility of all the matching
triples,

(5.58)
 Corresponding to the optimization problem in (5.54), the one-to-one
three-sided matching problem here can be described as below: we aim to
find as many matching triples of Tx users, Rx users and subchannels as
possible that maximize the social welfare U in (5.58) subject to the
constraints in (5.54b)–(5.54e) and (5.57).

Note that the matching model we formulated here is more complicated

than the traditional two-sided matching problem, since we need to jointly
match three sets of agents, i.e., the Tx users, Rx users, and the subchannels.
This results in an NP-hard problem instead of a polynomial time-solvable
one in the conventional stable-matching problems. The matching model we
formulate is different from the three-gender stable marriage problems
defined in [228], because in our model, each agent’s preference list is
mutual, causing different definitions of stability. Moreover, beside of the
agents’ own benefits, we also consider the social welfare, i.e., the total sum-
rate of the network, which will be explained in details in next Section.
Therefore, the traditional two-sided matching algorithms such as the Gale–
Shapley algorithm in which a stable and optimal matching can be achieved
cannot be applied here. To solve the one-to-one three-sided matching
problem, we will propose a modified matching algorithm in the next section.

5.3.3 Matching Algorithm for FD-OFDMA Resource Allocation

In the optimization problem (5.54), we aim to maximize the total sum-rate,

which can also be defined as the social welfare of the whole network.
Traditionally, this kind of social welfare optimization problem is solved by
utilizing centralized methods in which a centralized controller, i.e., the BS in
our scenario, finds the optimal allocation of the subchannels and pairs the Tx
users and the Rx users. However, since the problem is NP-hard, the
centralized methods have high computational complexity, rendering them
impractical, especially in a dense network. In this section, we will propose a
low complexity distributed matching algorithm.

Inspired by the Gale–Shapley algorithm that is designed for the two-

sided matching, we introduce a distributed low-complexity Tx–Subchannel–
Rx Matching Algorithm (TSRMA). By representing the UL Tx users, DL Rx
users and subchannels as three sets of independent agents pursuing their own
interests, the algorithm can be performed by the BS without extra signaling
cost compared to the traditional centralized methods.

(1) Design of Distributed Transmitter–Subchannel–Receiver Matching

Algorithm
To describe the proposed matching algorithm, let’s first define a subchannel-
Rx (SR) unit that consists of one subchannel and one Rx user. Since we have
K subchannels and N RX users in the network, there are NK different SR
units. Note that the SR units are not independent since they may contain the
same Rx or SC. As mentioned above, each Tx user is an agent trying to
maximize its own profits by matching with most profitable SR unit, and each
SR unit tends to choose a Tx user who can bring the largest utility. Here we
simply assume that if a Tx user wants to match with one SR unit, it proposes
to this SR unit, i.e., the Tx user makes an offer to its preferred SR unit in
order to get matched with the SR unit. In the matching process, Tx users
propose to the SR units. If all the unmatched Tx users propose to their
preferred SR units once, then we say one round of proposals is performed.
The key idea in the TSRMA is that, during each round of proposals in the
TSRMA, every unmatched Tx user proposes to its most preferred set of SR
units, and each SR unit has the right to reject or accept the Tx users’ offers
according to the potential utility gained by the proposing Tx users.

Before we describe the TSRMA in detail, we first introduce the concepts

of blocking individuals and blocking triples to explain how the SR units
choose among different offers from the proposing Tx users.

DENITION 2: Given any agent (or ) where

(or ), a matching Ψ is blocked by an
individual (or ) if (or ) prefers not to be matched than being
matched with its current pair of partners under Ψ.

During the proposing process of the Tx users, the conflicts between

different Tx users are very likely to appear, i.e., one Rx user or subchannel is
being proposed by two different Tx users. When the conflicts happen, the
conflicting party, i.e., an Rx user, subchannel or an SR unit, has the right to
choose one offer and reject all the other ones. This will exclude the existence
of the blocking triples, which can be defined as

DENITION 3: Given a triple , for any member

of this triple, we define to be the pair formed by
removing a from the triple . Under matching Ψ, is a blocking triple if it
satisfies the following constraints:

(i) ;

(ii) .

Item (i) guarantees that triple can bring positive utility to the related
transceiver unit. Item (ii) is also applied in the traditional definitions of
blocking triples [229] and [230]. It implies that if any agent a can obtain a
higher utility by forming a triple with two agents (say, b and c) other than its
original partners , then a just rejects and composes a blocking
triple with b and c, showing the selfishness of each individual.

Commonly, a member in a blocking triple is ready to switch its

current match with the members in as long as the utility it obtains, i.e.,
the sum-rate, increases by any (arbitrary small) amount. However, there
exists one special case in our matching model as explained below. Given a
blocking triple of the current matching Ψ (not stable), Tx a
tends to propose to , which have already been matched in two different
triples in Ψ. Note that changing the partners of b and c may reduce the
number of matching triples in Ψ, leading to the dissatisfaction of the other
agents as well as the decrease of the social welfare for the BS. Therefore, it
is more realistic that the players would switch their partners only when a
significant improvement can be made through the switching. To quantify this
observation, we then introduce the notion of the α-blocking triple as below.
Note that a similar definition of α-blocking is presented in the two-sided
matching [231].
DENITION 4: Given a matching Ψ (not stable) and a blocking triple
, we say is α-blocking if it satisfies:

(i) , , if , ,
, .

(ii) , if the condition in (i) is not applicable.

Item (i) implies that during the Txs proposing, if the conflict parts of a
blocking triple involve more than one matching triple in a matching, the
parties that are causing the conflict, i.e., b and c, evaluates their profits as a
group. That is, if the proposing Tx user along with b and c can bring a higher
utility than the proportional part of the sum of the matched triples where b
and c belong to, this blocking triple is also an α-blocking triple. Item (ii)
implies that if the conflict parties of a blocking triple only involve one
matching triple, then when the blocking triple brings higher utility
compared to the current matching triple, this blocking triple is α-blocking.
This is a more general condition compared to Definition 3, since most agents
wouldn’t bother to switch their current partners unless the gain that a
blocking triple brings is high enough. Besides, items (i) and (ii) guarantee
that the profits of the agents won’t decrease, and so does the social welfare
of the network. We can say that a blocking triple is acceptable for the agents
if it is not α-blocking.

With all the conditions above, we can then describe the action of each
SR unit when it makes decisions among multiple offers. Thus, if either or
receives more than one offers from the Tx users in a matching Ψ, it first
checks these offers one by one to examine whether there is a blocking triple,
and if the blocking triple is also α-blocking, the party causing conflict
accepts the blocking one and rejects its original partner(s) in Ψ.

(2) Description of the Transmitter–Subchannel–Receiver Matching

Algorithm
We now describe the TSRMA to solve the social welfare optimization
problem in (5.54) in a distributed way. The specific details of our proposed
algorithm are described in Table 5.5, consisting of an initialization phase in
Step 1, a pre-calculation phase in Step 2, and a matching phase in Step 3.
Note that all the three phases are performed not directly by the real Tx users,
Rx users, and the subchannels, but by the virtual processor units controlled
by the BS representing the users and the subchannels, which can be regarded
as virtual and independent agents created by the BS in order to perform the
whole matching process of the users and the subchannels. Thus, the BS can
then allocate the subchannels and pair the users according to the outcome of
the TSRMA.

Table 5.5 Transmitter–Subchannel–Receiver Matching Algorithm

Step 1: Initialization.
1. Construct the list of all unmatched transmitters, receivers, and subchannels, denoted by
, , and
, respectively.
2. Construct the list for all the subchannels to record each subchannel’s matching condition and
corresponding sum-rate, denoted by , , and .

3. Construct the list for all the Rx users to record each Rx user’s matching condition and
corresponding sum-rate, denoted by , , and .

Step 2: Pre-Calculation.
1. Obtain for all .

2. For each , sort all the SR units according to ’s preference over the sum-rate of the
corresponding transceivers, and obtain , with representing
’s t-th preferred SR unit.

3. For each subchannel , set , and for each , set

.

Step 3: Matching.
For each unmatched , it goes through the following steps:

1. ’S currently most preferred accessible SR unit is , and proposes to

.

2. If the proposed SR unit finds out that and

:
 (a) Match with the SR unit .

(b) Remove from , remove subchannel from , and

remove from .

3. Else if the SR unit finds out that it composes a α-blocking triple with when they’re matched:
(a) Remove from , from , and from
.

(b) Add to , to , to , and

to .

(c) Match with , and record the matching, i.e., ,

, , , , and
.
4. Else, is rejected by the SR unit .

This process continues until no new offers can be made by any , then the Tx
users and the SR units finalize all the matchings.

Step 4: End of the algorithm.

In the initialization phase, we first construct three matching lists,

, , and , respectively, to record the
current matching condition of Tx users, Rx users and subchannels, i.e.,
whether or not each Tx user, Rx user and subchannels are matched. Then for
each (or ), we construct a triple (or
) to record the detailed matching information in each
iteration, i.e., the pair that (or ) is matched to and its corresponding
sum-rate, so as to record the current matching triples in each iteration.

In the pre-calculation phase, we obtain the sum-rate of every

potential matching triple . And for each , we get its
preference list by sorting the sum-rate of all the
potential matching triples containing .
 In the matching phase, the matching process consists of multiple
iterations. In each iteration t, each unmatched starts by proposing to its
most preferred SR unit which has never rejected ’s offer in the previous
iterations. When the elements of an SR unit have never received any offers
in the previous iterations and only receive one offer from in iteration t,
will be temporarily matched, and this SR unit will record current
matching information in the corresponding triple as expressed in Step 3-2.
More often, an SR unit may be proposed by when other Txs have already
been matched with at least one member of this SR unit in earlier iterations,
and then the SR unit checks whether it can compose a potential blocking
triple4 with . If this is a potential blocking triple as well as an α-blocking
one, the SR unit keeps this offer and rejects the other offers received earlier,
otherwise, if this potential blocking triple is not α-blocking, then the SR unit
rejects this offer and do not switch its current partners, thereby eliminating
the existence of an α-blocking triple as described in Steps 3-3 and 3-4.

In each iteration, each unmatched Tx user makes new offers and each SR
unit that receives offers decides whether to reject the offers it keeps. Note
that each SR unit that temporarily keeps a Tx user’s offer will not actually
accept the offers, since it still has the right to choose a better offer in the next
iteration if other Tx users also propose to it. Each Tx user will not propose to
those SR units which have rejected their offers in the former iterations. The
iterations stop when there are no Tx users making new offers to the SR units,
which means that either a Tx user is matched to a SR unit, or it has been
rejected by all the SR units.

Note that the TSRMA gives the Tx users the priority to choose Rxs and
subchannels; we refer to it as a Tx-oriented matching algorithm. If it were
the Rx users, not the Tx users, proposing to the transmitter-subchannel (TS)
units consisting of the Tx users and the subchannels, then we could form
another algorithm, denoted as an Rx-oriented algorithm. It was shown in
[232] that the outcome of a matching algorithm will benefit the party which
makes the proposals more than the party which accepts the proposals. It is
notable that the Tx-oriented scheme may produce different outcomes from
the Rx-oriented scheme. With the former scheme, the outcome is best for the
Tx users, and vice versa. Here we choose the Tx-oriented scheme for the
following reason: the preference lists in our matching model is mutual, so
the preferences of the Tx users and the Rx users are mutually consistent in
the order of the sum-rate of the matching triples. Thus, the outcomes of the
afore prementioned schemes are quite similar in terms of the total sum-rate.
The comparison of these two schemes is presented in simulations.

(3) Analysis of the Transmitter–Subchannel–Receiver Matching

Algorithm
Different from the traditional two-sided matching problem, the 3D matching
problem is still under research, and various forms of preference lists and
stability definitions are being developed [230]. Usually, the definition of
stability depends on that of a blocking triple, such that its existence
determines whether a matching is stable, and different researchers may
introduce different versions of blocking triples and different levels of
stability. For example, with the traditional definitions of blocking triple as in
Definition 3, it is proved in [233] that the problem of deciding whether a
given 3DM instance has a stable matching is NP-complete, implying that
there are instances of the 3DM problem for which no stable matching exists.

However, in our matching model, we define the concept of α-blocking

triples as described before, because the 3DM problem we formulate is
different from those traditional ones in two aspects. First, the preference lists
in our 3DM problem are determined by the sum-rate of corresponding Tx–
Rx–subchannel triples, such that the preference lists of different members
are mutual. Second, in the traditional 3DM problems, all three sides of the
matching are selfish individuals chasing as large profits as possible.
Nevertheless, in our proposed 3D matching model, it is the BS that executes
the TSRMA while considering the social welfare as well as the rational
nature of the users, since we aim to maximize the total sum-rate.

DENITION 5: A matching Ψ is defined as -stable if it is not blocked by any

individual in Ψ or α-blocked by any triple which does not exist in Ψ.

Then for the convergence and stability of the above matching algorithm,
we have the following Lemma and Theorems.
LEMMA 7: If the TSRMA converges to a matching , then is a -stable
matching.

THEOREM 2: The TSRMA converges to a -stable matching after a limited

number of iterations.

Generally speaking, there may exist different outcomes of stable

matching due to different matching algorithms. Since we give complete and
strict preference lists in our matching model, we can always obtain the same
stable matching from the TSRMA given the initial conditions.

By performing the STRMA, we guarantee that there is no α-blocking

triple in the final matching. However, the optimization problem to be solved
is an NP-hard problem: it is likely that there still exist blocking triples in the
final matching. Some works have computed a c-factor (c is an absolute
constant) approximation to three gender maximally stable marriage problem
(3G-MSM) in terms of the number of blocking triples [234], in which the
minimum number of the stable triples is approximated, and the size of all
three sets of agents are the same. Now we give an upper bound of the
number of the blocking triples in our proposed STRMA given .

THEOREM 3: Let be the final matching found by our proposed TSRMA.

Without loss of generality, we assume that , and thus the number
of blocking triples is upper bounded by

(5.59)

Theorem 3 also provide an estimation on the gap between a super-stable

matching [229] (also known as an optimal matching in our scenario) and the
-stable matching found by the TSRMA. Normally, an optimal matching
contains MNK stable triples,5 while the TSRMA can provide at least
stable triples. Note that since the optimization problem we
formulate in (5.54) is NP-hard, we can’t get an optimal solution utilizing any
algorithm within polynomial time, so it is natural that the solution of the
TSRMA is not guaranteed to be an optimal one. This is because the
incentive of the agents is to maximize their own utility, so it is likely that
some of their choices aren’t the best strategies in terms of the social welfare.
Therefore, the -stable outcome that the BS utilizing the TSRMA is not an
optimal one sometimes.

Next, let’s discuss the relationship between the total sum-rate, total BS
transmitted power, , and the transmitted power of each Tx user, ,
respectively. According to equations (5.5) and (5.54), the total sum-rate is a
function of both and . Let be the largest distance between one Tx
user and one Rx user and represent the largest distance between the users
and the BS. We have the following two propositions.

PROPOSITION 1: The slope of the total sum-rate vs. can be determined by

the following expression:

(5.60)

where , , and .

For example, when dBm, m, m, we have

, , , and equation (5.60) can be written as
. When varies from 0 dB to
23 dB, this equation is smaller than zero, which means that the total sum-rate
decreases as increases.

PROPOSITION 2: The slope of the total sum-rate vs. can be determined by

the the following expression:

(5.61)
where , , .

For example, when dBm, m, m, we have

, , , and equation (5.61) can be written as
. When varies from 5 dB to
46 dB, this equation is larger than zero, which means that the total sum-rate
increases as grows.

5.3.4 Simulation Results

In this section, we evaluate the performance of the TSRMA and compare it

to the centralized solution and the random matching algorithm. The
centralized solution can be regarded as the upper bound of the total sum-rate
of the network. In the random matching algorithm, the Tx users are
randomly matched with the SR units, and when the number of the Tx users
is not equal to L, Tx users will be randomly omitted from the
matching, and the same is for the Rx users and the subchannels. We assume
that for all three algorithms mentioned above, the BS allocates the
transmitted power equally among the transceiver units, and all the mobile
phone users including the Tx users and Rx users are randomly distributed in
a square area with the size length of . Given the simulation parameters
listed in Table 5.6, we obtain our simulation results as shown below, and all
curves are generated based on averaging over 10,000 instances of the
algorithms.

Table 5.6 Main Simulation Parameters

Parameter Value
The BS’s peak transmitted power 23 ~ 46 dBm
Each Tx user’s transmitted power 0 ~ 23 dBm
Noise variance –90 dBm

Interference cancellation coefficient 0.5

The largest distance between one Tx user and one Rx user 400 ~ 2600 m
The largest distance between the users and the BS 100 ~ 1300 m
The distance between antenna a and b of the BS 1m
The attenuation factor 2
The stability factor 1

Fig. 5.9 shows the total sum-rate as a function of each Tx user’s

transmitted power , with dBm, m, and m.
From Fig. 5.9, the total sum-rate of the network decreases as each Tx user’s
transmitted power increases, which is consistent with Proposition 1. This
implies that it is necessary to limit the users’ transmitted power to a certain
level so as to achieve satisfactory total sum-rate. The performance of the
TSRMA is very close to that of the centralized one, and is significantly
better than that of the random matching algorithm. For example, when
, , , the TSRMA reaches 97.5 of the performance of the
centralized one, and is 225 above the random matching algorithm.

Figure 5.9 Total sum-rate vs. transmitted power of each user

In Fig. 5.10, two matching schemes, i.e., the Tx-oriented TSRMA in

which the Txs propose to the SR units and the Rx-oriented scheme in which
the Rxs propose to the Txs and the subchannels, as mentioned in Section
IV.B, are compared in terms of the total sum-rate. We set dBm,
m, and m. It can be noted from the figure that the
outcome of these two schemes are nearly the same. This makes sense
because the preference lists in our matching model are mutual, so the Tx
users and Rx users may have similar preference over the matching triples in
terms of the sum-rate, leads renders them to choose the same triples. Since
the outcome of two prementioned schemes are quite similar, it does not
make a great difference whether to utilize the TSRMA or the Rx-oriented
scheme in terms of the total sum-rate of the network.

Figure 5.10 Total sum-rate vs. transmitted power of each user with the Tx-oriented TSRMA and
the Rx-oriented scheme

Fig. 5.11 illustrates the total sum-rate as a function of the BS’s peak
transmitted power , with dBm, m, and m. We
present two cases with the different numbers of Tx users, Rx users, and the
subchannels. We observe that the total sum-rate grows as the BS’s peak
transmitted power increases, which is consistent with the results in
Proposition 2. As increases, the energy of the useful signals gets larger,
while the influence of the interference caused by the BS’s antennas on SINR
is limited in our scenario, and thus, the total sum-rate increases as well. Note
that this is different from the traditional point-to-point full-duplex model,
since we have multiple users to work in the full-duplex mode
simultaneously.

Figure 5.11 Total sum-rate vs. total transmitted power of the BS

Fig. 5.12 plots the total sum-rate as a function of the number of the
subchannels K, with , dBm, dBm, m,
and m. When the centralized algorithm and the TSRMA are
applied, the total sum-rate increases with the number of the subchannels K,
and grows slower after K becomes larger than . When the random
matching algorithm is utilized, the total sum-rate increases with K as well,
and then converges to a maximum value when K is larger than .
Comparing these three curves, we regard the gap between the TSRMA and
the random matching as the benefit brought by the user pairing and the
subchannel allocation. If , the number of matching triples increases
with K, so the total sum-rate increases rapidly in all the three curves. But
when , the number of matching triples is smaller than L, so for the
TSRMA, the sustained growth of the total sum-rate comes from the selection
of the matching triples. For the random matching algorithm, there is no such
selection mechanism and the users are randomly matched with each other, so
the total sum-rate remains the same as K increases.

Figure 5.12 Total sum-rate vs. the number of the subchannels K with four Tx users and four Rx
users

The total sum-rate as a function of the number of the Tx users M and the
number of the Rx users N, is illustrated in Fig. 5.13 where dBm,
dBm, m, and m. When , the total sum-rate
increases significantly with the number of the Txs (or Rxs) due to the
growing number of the matching triples. When , the total sum-rate of
the random matching algorithm converges to a fixed value, while for the
centralized algorithm and the TSRMA the total sum-rate continues to grow
with a slower rate. We can then infer that the total sum-rate is greatly
influenced by the number of potential matching triples: the larger the
number of potential matching triples, the higher the total sum-rate.
Figure 5.13 Total sum-rate vs. the number of the Tx users with four Rx users and four subchannels,
and the total sum-rate vs. the number of the Rx users with four Tx users and four subchannels
 Fig. 5.14 shows the total sum-rate as a function of , with ,
, , dBm, and dBm. From both subgraphs, it can
be seen that with different values of and , the total sum-rate decreases
as increases. This can be explained by observing the expression of the
sum-rate in equation (5.5). In a transceiver unit , the interference of
to , i.e., , is much larger than the noise, i.e., , so
the second part decreases as increases. The distance between two
antennas of the BS is fixed, and thus, the first part of the sum-rate decreases
as increases. This implies that, to maximize the total sum-rate, the BS
pairs those Tx users and Rx users which are far from each other and can
communicate with the BS through good channels so as to obtain as large a
total sum-rate as possible. Besides, Fig. 5.14 also shows that when the value
of is fixed, varying the values of and does not significantly
affect the total sum-rate of the network. This implies that when evaluating
the affecting factors of the total sum-rate, we should focus on the radio
, but not or alone.
Figure 5.14 Total sum-rate vs. with varying, and the total sum-rate vs. with
varying
5.4 Concluding Remarks

In this chapter, we have studied the resource allocation problems for a

single-cell FD-OFDMA network. In the considered system, an FD Base
Station (BS) simultaneously communicates with multiple pairs of UpLink
(UL) and DownLink (DL) Half-Duplex (HD) users bidirectionally. We
aimed to find the optimal pairing of UL and DL users for each FD link and
the allocation of the OFDM subcarriers and power to different pairs such
that the sum-rate of the network is maximized. We introduced two
alternative ways of solving such a complex problem. We first present a
method to solve the problem by using the dual method and Sequential
Parametric Convex Approximation (SPCA). We then introduced a low-
complexity approach based on matching theories. In these works, we do not
consider the power control at the user sides, which can further improve the
network performance. It will be an interesting but challenging problem,
which can be considered for future research. Also, here we have only
considered the single-cell FD-OFDMA system. In practical systems, we also
need to consider the multi-cell cellular networks, where the inter-cell
interference and cross-cell resource allocations need to be considered: the
resource allocation of the FD-OFDMA system in such a multi-cell scenarios
will be very important but also very challenging. In such networks, not only
will the intra-cell resource allocation, but inter-cell resource allocation, also
need to be considered in the optimization, which make the problem very
complicated.

APPENDIX: SPCA Algorithm to Maximize

The basic principle of the SPCA method is to replace the non-convex
function by its convex upper bound and to iteratively solve the resulting
problem by judiciously updating the variables in the convex approximation
until convergence. This convex upper bound should also have two properties
[213].

LEMMA 8: Consider a non-convex function . If the function

replaces in the SPCA method, should have the following two
properties:
i) for any , ;

ii) for a given feasible point , there exists satisfying

and .

Proof: Refer to [213].

Additionally, it has been proven that the convergence of the SPCA

method always exists and the convergent point satisfies the KKT conditions
theoretically [213].

Now, we describe the solution process to maximize . For simplicity,

and are replaced by and , respectively. Similarly ,
, , and are replaced by , , , and , respectively.
and are replaced by and , respectively. We use the first-order
inequality for where y is an
arbitrary operating point. Then, in (5.28) is lower bounded by

(5.62)

It is not hard to verify that this inequality satisfies the two properties defined
in Lemma 8. Also, the right-hand side of this inequality is a convex function
[211]. According to the SPCA method, in each iteration, and is
updated by and , respectively, where l represents the l-th iteration.
One can use the interior-point method (or other convex optimization
method) to obtain the optimal values and that maximize in each
iteration. However, we can give a closed-form solution of the lower-bounded
optimization problem alternatively. For simplicity, we omit the superscript in
the following derivation.

The partial derivative of over and are given, respectively, by

(5.63)

(5.64)

where and .

Defining and , we divide

the original problem that maximizes with respect to and into four
cases.

Case A: When and , is a decreasing function of

both and . Thus, we have .
Case B: When and , (i) is a decreasing function
of , and (ii) increases then decreases as increases. Thus, we obtain
(i) and (ii) a unique solution for by letting , i.e.,

(5.65)

Case C: When and , is a decreasing function of

. Using a similar analysis as in Case B, we obtain (i) and (ii)
as
 (5.66)
Case D: When and , increases then decreases
as both and increase. However, deriving the closed-form solution in
this case is not straightforward. Hence, we give an iterative solution in
this case. By letting , we obtain

(5.67)

(5.68)

where and are the values obtained in the previous iteration. In

practice, we do not have to wait for full convergence and a single ascent
step is sufficient before updating.

To sum up, the entire algorithm is listed in Algorithm 1.

Algorithm 1 Solution of
1: Set (maximum number of iterations)
(convergence tolerance)
2: Set arbitrary positive values for .
3: .
4: while and do
5: if and then
6: .
7: else if and then
8: and is given in (5.65).
9: else if and then
10: and is given in (5.66).
11: else
12: use (5.67) and (5.68) to obtain and , respectively.
13: end if
14: .
15: .
16: end while
17: The optimal solution of maximizing is . The optimal
value of is obtained as .

1 According to [217], the height of the BS can be from 0 to 50 m above ground. We therefore set the
height of the BS to 31 m in the UL and DL channel models. However, the path-loss model of IUI
channels has not been given specifically. Here we set the height of users to 1.5 m above ground in
the IUI channel models as well as UL and DL channel models.

2 Normally, subchannels are not individuals who can make decisions, but since the BS is in charge
of allocating the subchannels, we can regard the subchannels as “individuals” who are controlled
by the BS.

3 When we say an object is indifferent between two objects in the other parties, we mean that the
object might choose either one.

4 We say the blocking triples are potential since they are still not in the final matching.

5 A stable triple is a triple which is not blocking in a matching.

6 Full-Duplex Heterogeneous Networks

The continuous growth in the numbers of mobile users and the increase in
data communication make deploying new techniques with increased data
rates a crucial need. In other words, modern network topologies are
required to give the required performance boost. Since decreasing the
number of users in each cell, as well as decreasing the distance between the
users and the base station, increases the total achieved capacity, therefore,
small cells are required. However, it is well known that large cells are the
cost-effective way to serve fast-moving users. Therefore, one solution to
match these demands is to employ Heterogeneous Networks (HetNets). The
main idea of HetNets is to densify the existing macro-cell by adding a mix
of pico, femto and relay base stations to which users will be offloaded from
the macro base station. This will help in increasing the network capacity in
hot spots and in giving better coverage for both outdoor and indoor areas
not covered by the macro network.

Furthermore, both the efficiency of utilizing the available

communication resources and the capacity achieved by HetNets can be
further improved by deploying Full Duplex (FD). Theoretically speaking,
allowing the network nodes to simultaneously transmit and receive data at
the same channel and in the same time slot achieves double the capacity
achieved by Half Duplex (HD) communication when using the same
resources. However, it must be mentioned that FD was considered
unfeasible for a long time as the increase in networks’ capacity offered by
deploying FD is restricted by the ability to suppress the self-interference
which is caused by the node’s transmission on the node’s reception and
results in a great degradation in the received Signal-to-Interference Noise
Ratio (SINR). Recent evolution in self-interference cancellation techniques
revived the attention to FD. As mentioned, adding FD communication to
HetNets will improve the network efficiency and capacity. Accordingly,
efficient resource allocation techniques are needed to fully benefit from the
available resources and to overcome the challenges that arise from
deploying FD to HetNets.

In this chapter, we provide two examples for FD HetNet, which are

summarized as follows:

First, a study of the effect of FD on resource allocation for a small

cell network is presented. In our model, both the macro-base station
and femtocell access point operate in FD, where users attempt to
connect either to the base station or the femtocell access point. A
resource allocation problem is formulated to adjust the base station
and the femtocell access point transmitting powers to maximize the
downlink channel capacity while guaranteeing the uplink channel
QoS by maintaining a certain channel capacity. In addition, a solution
algorithm using the interior point method is proposed. Our results
show that, depending on the interference conditions, it can be decided
when it is preferable to employ FD and when it is more advantageous
to use HD.
Second, we consider a simple FD communication system, consisting
of one FD Access Point (AP), and two HD mobile users, and
investigate when it is more advantageous for the communication
network to operate in FD or HD Multi-User MIMO (MU-MIMO)
modes. Since FD transmission is degraded by self-interference, while
HD suffers from spatial correlation between MIMO antennas, which
cause rate loss, we study the effect of both the self-interference
cancellation parameter at the FD AP, and the mutual distance
between the HD users on FD performances, and the effect of the
spatial correlation coefficient on HD MIMO. Afterwards, a novel
switching criterion is proposed, which chooses the operation mode
that is capable of maximizing the downlink channel capacity while
maintaining the uplink channel capacity at a certain level.
Subsequently, based on our study of the system’s parameters that
 affect the performance of both FD and HD, theoretical thresholds for
these parameters are derived. Finally, numerical analysis is presented
verifying the validity of the optimization problem solution and the
derived thresholds.

6.1 Full-Duplex Small Cell Networks

The prime challenge in FD is to suppress the interference caused by the

node’s self-transmission. Moreover, the impact of FD on resource allocation
is not well investigated yet. Reducing the self-interference level has
recently been studied in the literature. In [99], several interference
mechanisms have been proposed and it was shown that it is possible to
make FD attainable. In [101], two models for self-interference cancellation
have been proposed in FD communication. In [102], an in-band FD WiFi
radio design is presented, which is capable of reducing the self-interference
to the receiver noise floor. In [103], balanced/unbalanced (balun)
transformer is used in implementing an FD radio that can support high-
power and wide-band systems. Furthermore, investigating the effect of FD
on resource allocation problems is recently underscored in research works.
In [104], the optimal resource allocation schemes for both FD and HD relay
wireless networks have been proposed, in addition dynamic resource
allocation strategies were developed under both FD and HD modes. In
[106], a power allocation scheme for maximizing FD wireless link
throughput given a certain delay constraint is proposed. In [107], optimal
dynamic power allocation schemes have been proposed which intend to
maximize the sum-rate of wireless FD transmissions.

Furthermore, the increased number of wireless communication users has

created a demand for new resource allocation schemes. Accordingly, two-
tier networks were introduced as a solution that will assist in traffic
offloading from the macro-base stations to the small-cell (femtocell) access
points. In [108], several techniques for traffic offloading are discussed. In
[109], two traffic offloading techniques are proposed, and it is proven that
high energy gains can be achieved when applying these algorithms in traffic
offloading. In [235], the technical challenges in the deployment of small
cells are presented. Different techniques for traffic offloading to small cells
are then explained and evaluated.

In this section, we are studying a two-tier network operating in FD. We

are concerned by the case in which there exist both transmitting and
receiving users which attempt to connect either to the macro Base Station
(BS) or to the Femtocell Access Point (FAP). Both BS and FAP are working
in the FD mode. We formulate a power allocation problem which
maximizes the channel capacity of the downlink communication while
keeping a certain QoS for the uplink communication represented by the
uplink channel capacity. In addition, we are going to study the effect of
limited backhaul between the BS and the FAP and its effect on the power
allocation problem. Finally, we validate the performance of our model
numerically, which proves that our model is able to allocate power
appropriately in order to maximize the downlink channel capacity while
granting a certain QoS for the uplink channel. Our contributions in this
work can be summarized as follows:

Studying the effect of FD on resource allocation in two-tier networks;

Formulating a power allocation problem which maximizes downlink
channel capacity while retaining a certain QoS for the uplink
channel;
Proposing a solution algorithm for the power allocation problem
using the interior point algorithm;
Providing results to show when to use FD and when to use HD.

The remainder of the section is organised as follows. In Section 6.1.1,

the system model will be presented. In Section 6.1.2, our problem
formulation will be presented and the solution algorithm will be
constructed. In Section 6.1.3, we validate our system performance by
numerical results.

6.1.1 System Model

We consider a single channel to be a single-cell network that includes a
single BS with two antennas and a single FAP with two antennas. The cell
has two users that attempt to connect either to the BS or the FAP. The
system model is shown in Fig. 6.1.

Figure 6.1 System model

Without loss of generality, we consider the case in which, in a given

time slot, one user is transmitting in the UpLink (UL), while the other user
is receiving in the DownLink (DL). The BS and the FAP are able to work in
FD mode. However, the main problem of supporting FD is the increased
interference level. Therefore, special FD radios are used in both BS and
FAP. In our model, we consider the radio design proposed in [103]. It is
able to cancel the self-interference up to 73 dB.

Furthermore, the users have the choice to connect either to the BS or to

the FAP. Therefore, we are going to have four different possible modes for
the given system:
1. Mode 1: Both UL and DL users are connected to the BS. In this case the
received Signal-to-INterference Ratio (SINR) at the BS receiving
antenna is given by

(6.1)

where P is the transmitted power from the UL user, gives the

large-scale propagation loss, with distance and path loss
exponent α, is the channel coefficient between the UL user and
the receiving antenna of the BS, all the channel coefficients are assumed
to be an i.i.d. zero-mean complex Gaussian random variables with unit
variance, i.e., Rayleigh fading; is the noise variance. Finally, is
the self-interference power introduced from the BS transmitting antenna,
where is the BS transmission power and is the self-interference
cancellation parameter due to using the FD radio circuit at the BS.

In addition, the received SINR at the DL user is given by

(6.2)

where is the distance between DL user and BS, is the

channel coefficient between the BS transmitting antenna and the DL
user, is the distance between UL user and DL user,
is the interference introduced by the UL user, where
is the distance between UL and DL users and is the
channel coefficient between UL and DL users.

2. Mode 2: The UL user is connected to BS but the DL user is connected

to FAP; in this case, the received SINR at the BS receiving antenna is
given by,
 (6.3)

where is the value of the interference introduced

from the FAP transmission, in which is the FAP transmission
power and is the channel coefficient between the FAP
transmitting antenna and the BS receiving antenna. Furthermore, the
received SINR at the DL user is given by

(6.4)

where is the distance between the DL user and the FAP.

is the channel coefficient between the DL user and the FAP.

3. Mode 3: Both UL and DL users are connected to the FAP; in this case,
the received SINR at the receiving antenna of the FAP is given by

(6.5)

where is the distance between the UL and the FAP, is

the channel coefficient between the UL user and the FAP and is the
self-interference introduced by the FAP transmission, where is the
self-interference cancellation parameter due to using the FD radio circuit
at the FAP.

In addition, the received SINR at the DL user is given by

 (6.6)

4. Mode 4: The UL user is connected to the FAP and the DL user is

connected to the BS. The SINR at the FAP receiving antenna and at the
DL user are given, respectively, by

(6.7)

(6.8)

where is the channel coefficient between the FAP receiving

antenna and the BS transmitting antenna.

From equations (6.2), (6.4), (6.6), and (6.8), it can be noticed that the
SINR of the DL user can be written as

(6.9)

where is a parameter indicates the selection of the DL user; if

then the DL user is connected to the BS; while if then the user is
connected to the FAP. Similarly, refers to the UL user selection. In the
same way, from equations (6.1), (6.3), (6.5), and (6.7), the received SINR of
the UL transmission could be written as
 (6.10)

Now, the channel capacity of the UL transmission and the DL

transmission are given respectively by

(6.11)

(6.12)

6.1.2 Problem Formulation

In our problem we are trying to find the values of , , , which

maximize the overall capacity calculated in (6.12), while keeping a certain
quality of service for the UL transmission represented by the overall
capacity calculated in (6.11). The optimization problem is given by

(6.13)

where is the minimum accepted value for the UL overall capacity,

is the maximum allowed transmission power for the BS and is the
maximum allowed transmission power for the FAP. It is worth mentioning
that both the second and the fourth modes are very challenging to realie in
practice, as they require coordination between the BS and the FAP which is
very difficult because of the limited backhaul between them. This will be
translated as an additional constraint on the selection process. Therefore,
the resource allocation problem formulated in (6.13) will be modified to:
 (6.14)

As mentioned, the values of and are constrained either to be

equal to zero or one.1 In order to be able to solve the problems given in
(6.13) and (6.14), first we relax the values of and from the integer
domain to the continuous domain . Then, we will use the interior point
algorithm [211] to solve the relaxed optimization problem, in which the
objective function to be minimized is with the following constraint
functions:

(6.15)

where . Therefore, based on the constraint functions, we

can calculate the value of the logarithmic barrier function, which is given
by

(6.16)

which will approximate the optimization problem to be

(6.17)
where t is the approximation parameter, which can be controlled in each
iteration of the interior point method. After solving the problem defined in
(6.17), the values of and are remapped to the integer domain from
the optimization solution. Finally, we construct the solution algorithm
shown in Fig. 6.2, where is an initial point in the feasible set, ε is the
tolerance value, and m is the number of constraint functions; and μ is
the step size.

Figure 6.2 Solution algorithm ([301] Reproduced with permission from the IEEE.)

6.1.3 Numerical Results

We will compare our system performance with the case where the system is
using HD. It must be noticed that in the case of using HD, maximizing the
DL user capacity will not be constrained by the UL channel capacity
because each user is active in a different time slot. The simulation
parameters used are shown in Table 6.1.

Table 6.1 Simulation Parameters ([301] Reproduced with permission from the IEEE.)
20 mW 10 mW
2 mW 2 Km
3 Km 3 Km
 2 Km 4 Km
8 Km [103]

Effect of Biasing
Biasing is an offloading technique used to offload users from the BS to the
FAP, by which the user will be willing to connect to the FAP even if the
received power from the FAP is less than that received from the BS. Biasing
can be translated to a scaling factor for the received power. It must be
mentioned that only the DL user is biased to connect to the FAP as the UL
user is constrained by a minimum channel capacity, . However, it must
also be clear that, even if the users are biased, the solution of the
optimization problems defined in (6.13) and (6.14) could be to connect the
users to the BS rather than to the FAP. The variation of the DL channel
capacity with the biasing value is shown in Fig. 6.3. It is clear that for the
small values of biasing, it is more reliable to apply FD. This consequence is
reasonable as large biasing will make the DL user connect to the FAP. In
addition, increasing the biasing value will cause an increase in the BS and
the FAP transmitting power which will increase the self-interference level.
Nevertheless, using HD is less vulnerable to biasing because maximizing
the DL channel capacity will be achieved by setting the transmission power
to its maximum value. Therefore, regardless of the biasing value, the
channel between the DL user and the FAP will be maintained at a certain
capacity as the transmission power is constant.
 Figure 6.3 Variation of downlink channel capacity with the biasing value ([301] Reproduced with
permission from the IEEE.)

Effect of Uplink Capacity Constraint

The variation of DL user channel capacity is shown in Fig. 6.4. It can be
noticed that for larger values of minimum UL channel capacity constraint, it
will be better to use HD because increasing the value of will require a
decrease in the interfering power represented by the BS or the FAP
transmitting powers. As a result, the value of the DL channel capacity will
decrease. In the case of HD, the DL channel capacity is independent of the
UL channel capacity constraint. Therefore it will remain constant whatever
the value of . Consequently, it can be concluded that a conservative
demand of high UL user channel capacity will force the system to operate
in HD mode in order to guarantee a reliable communication for the DL user.
 Figure 6.4 Variation of downlink channel capacity with UL capacity constraint value ([301]
Reproduced with permission from the IEEE.)

From Figs. 6.3 and 6.4, it can be decided when it is preferable to use FD
or to use HD. This decision is mainly controlled by the self-interference
power. Whenever the self-interference is adequately low, the FD mode will
be more able to utilize the given resources much better than the HD mode.

6.2Full-Duplex HetNet and MU-MIMO

Switching

As mentioned before, the main challenge that restrains FD operation is the

increased self-interference caused by the node transmission on the node
reception, especially in a multiuser (MU) communication system.
Accordingly, as the ability of a communication system increases to suppress
the self-interference, the gain that can be achieved from FD will increase.
Since the gain achieved by FD is controlled by the self-interference
cancellation, it is not always guaranteed that FD will outperform HD. In
addition, in the case of multiuser MIMO systems, HD may suffer from
spatial correlation between antennas [236], which results from the limited
spacing between antennas, and hence the MIMO channels will be
correlated. This correlation is expected to degrade the HD performance.
Therefore, in order to be able to optimize the network performance and
build a framework that optimizes the network’s rate by choosing the
transmission mode that offers the better performance, a clear investigation
for the parameters that affect the performance of both FD and HD
transmissions is needed. This investigation will give an insight on the
conditions required for the communication system to operate either in FD or
HD. Accordingly, an efficient switching algorithm between FD and HD can
be established based on the derived conditions. The switching algorithm
will be able to optimally utilize the available communication network
resources.

In this section, a complete study of the parameters that affect the

performance of both FD and HD operation is presented. We consider a
system with one AP such as a base station, and two users. Accordingly, a
switching criterion between FD and HD is proposed, which aims at
choosing the transmission mode that maximizes the total DownLink (DL)
capacity while maintaining the UpLink (UL) transmission’s total capacity at
a certain level. In addition, the switching criterion is formulated as a
resource allocation problem, which is solved using the Karush–Kuhn–
Tucker (KKT) conditions. Furthermore, theoretical expressions are derived
for the system parameters’ thresholds controlling the switching process
between FD and HD. Finally, the network’s operation regions are
determined based on the derived thresholds. The contributions can be
summarized as follows:

1. Presenting an investigation on the parameters affecting FD and HD

performance;
2. Proposing a switching criterion that chooses the optimal transmission
mode for the communication network that maximizes the DL capacity
and keeps the UL capacity at a certain level;
3. Deriving theoretical thresholds expressions for self-interference
cancellation, mutual distance between users and spatial correlation
coefficients between antennas that dominate the switching between FD
and HD, and deriving the network’s operation regions based on the
derived thresholds.

The rest of the section is organized as follows. In Section 6.2.1, the

system model is presented. In Section 6.2.2, the switching criterion is
formulated and the analytical solution is presented. In Section 6.2.3,
numerical results are presented.

6.2.1 System Model

In this subsection, we consider a single channel network with one AP

equipped with two antennas. The cell has two active users , that may
be both transmitting in the UL mode or receiving in the DL mode, and in a
special case we will have one user transmitting in the UL and the other
receiving in the DL. The AP is trying to choose the optimal transmission
mode that will maximize the DL channel capacity while keeping a certain
QoS for the UL transmission represented by the UL channel capacity. It
must be taken into consideration that all Channel State Information (CSI)
for the UL transmissions, DL transmission, and channels between users
must be available at the AP in order to be able to switch between FD and
HD. Channel estimation is out of our work scope; different techniques of
channel estimation are found in [237–240]. The system model is shown in
Fig. 6.5. Based on the above assumptions, the difference between FD and
HD operations will be illustrated.
 Figure 6.5 System model

FD Transmission
In case of FD transmission, in the first time slot, the AP will be dedicated to
serve the DL transmission of user m along with the UL transmission of user
n. However, in the second time slot, the AP will serve the DL transmission
of user n along with the UL transmission of user m. In addition, the AP is
required to have a special FD radio [99], [124] that helps in self-
interference cancellation. Accordingly, the DL received Signal-to-
Interference-Noise Ratio (SINR) of users m and n are given, respectively,
by

(6.18)

where is the AP transmission power, denotes the large-scale

propagation loss between user m and the AP, with distance and path
loss exponent α, is the channel coefficient between user m and the
AP transmitting antenna,2 is the AWGN noise variance, denotes
the large-scale propagation loss between users m and n. Finally,
denotes the interference introduced by the UL
transmission. From (6.18) it is clear that the interference from the UL
transmission on the DL transmission needs to be estimated; therefore, it is
sufficient to only estimate the SINR between the UL and DL user. Observe
that the SINR estimation is much simpler than the channel estimation, and
can be done using simple power estimation techniques. After SINR is
estimated for the DL user, the estimation result is reported to the AP in a
predetermined reporting period.3 It must be pointed out that the notation
means that we can get the expression for the nth user from (6.18) by
interchanging m and n in the equation. All channel coefficients are assumed
to be independent and identically distributed (i.i.d.) zero-mean complex
Gaussian random variables with unit variance, i.e., Rayleigh fading. On the
other hand, UL received SINR at the AP receiving antenna for users m and
n are given, respectively, by

(6.19)

where P is the UL user transmission power, denotes the large-scale

propagation loss with distance and path loss exponent α, and
represents the Residual Self-Interference (RSI) after using a FD radio at the
AP with a cancellation parameter C, and which reflect the quality
of SI cancellation [99], [124], [241].4 From (6.18)–(6.19), it can be noticed
that the FD DL-received SINR can be regulated by controlling the mutual
distance between users. Moreover, the quality of the UL transmission
depends on the FD radio used in the AP, and hence on its self-interference
cancellation parameter.

HD Transmission
In the case of HD transmission, the AP will serve two simultaneous DL
transmissions of users m and n in the first time slot. Then the AP will serve
two simultaneous UL transmissions of users m and n in the second time
slot. Therefore, the DL received SINR of users m and n are given,
respectively, by

(6.20)
where is the precoding vector applied by the AP to user m DL
data, R is a matrix which models the spatial correlation at the
AP with spatial correlation coefficient β, denotes
the interference caused by the AP transmission to user n, and is the
precoding vector applied by the AP to user n data. The value of is
given by

(6.21)

where is the hermitian transpose of that denotes the channel

vector between the AP and the DL user, and denotes the -norm.
Notice that, in the case of HD transmission, the DL user is receiving data
from both antennas of the AP. This makes . In addition, it is clear
that the channel coefficient defined in (6.18) represents an element in
the channel vector . On the other hand, the UL received SINR of
users m and n are given, respectively, by

(6.22)

where denotes the channel vector between user m and the AP.
Similarly, . In addition defined in (6.19) represents an
element in . denotes the interference caused by the
UL transmission of user n. From (6.20) and (6.22), it can be seen that both
the DL and UL received SINR, in the case of HD depending on the spatial
correlation between the antennas. Hence, the value of the spatial correlation
coefficient β will influence the quality of HD operation.

To summarize, the relation between the HD- and FD-achieved

capacities are mainly dependent on the self-interference cancellation, the
mutual distance between users and the spatial correlation between antennas,
as follows:

For the UL capacity, it is expected that HD will outperform FD for

small values of SI cancellation C. However, it is expected that FD
will offer higher UL capacity than HD for high values of spatial
correlation coefficient β.
For the DL capacity, it is anticipated that HD will outperform FD for
small mutual distances between the users . On the other hand,
FD will achieve higher DL capacity for large values of spatial
correlation coefficient β.

6.2.2 Switching Criterion between Full-Duplex and Half-

Duplex

From the received SINR expressions given by (6.18)–(6.22), the total DL

channel capacity per unit bandwidth in two successive time slots when
operating in FD transmission and HD transmission are given, respectively,
by

(6.23)

(6.24)

In addition, the total UL channel capacity per unit bandwidth in two

time slots when operating in FD and HD transmissions are given,
respectively, by

(6.25)

(6.26)
 We assume that the QoS of the UL transmission of each user is
constrained by a certain minimum channel capacity . In other words,
regardless of the operation mode, the UL transmission capacity for each
user must be greater than . The total DL channel capacity and UL
channel capacity per unit bandwidth for users m and n are given,
respectively, by

(6.27)

(6.28)

where denotes that the system will operate in the FD mode. While
denotes that the system will operate in the HD mode. The main idea
of the switching criterion is to choose the transmission mode that will be
able to:

1. Provide the UL transmission with the minimum required channel

capacity which is equal to ;
2. Maximize the DL user capacity as given by (6.27).

Based on the previous assumptions, the switching criterion is shown in

Fig. 6.6. In the beginning, both the DL and UL channel capacities for FD
and HD are calculated, respectively, from (6.23)–(6.26). Afterwards, the
system will compare the achieved UL capacity with . If FD is not able
to satisfy the UL channel capacity constraint, the system will operate in
HD. On the contrary, when HD does not meet the UL channel capacity
constraint, the system will operate in FD. On the other hand, if both FD and
HD satisfy the minimum channel capacity for both users, the system will
choose the transmission mode that will maximize the DL channel capacity
given by (6.27). It is worth mentioning that, if both FD and HD fail to
satisfy the minimum channel capacity constraint for the users, the network
will be considered in outage.
 Figure 6.6 Switching criterion

Switching Criterion Problem Formulation

In this section, we are going to provide the analysis of the proposed
switching criterion, which can be formulated as the following optimization
problem:

(6.29)

It must be mentioned that the second binary constraint on T makes the

problem hard to solve. Therefore, in order to solve this problem, we are
going to take two different approaches. The first is the exact solution, which
is to enumerate all possible cases of the AP’s operation. In other words, we
separately compares the system performance in the cases of FD and HD
operation, which correspond to and , respectively. The other
approach is to relax the second constraint on the T parameter to be in the
continuous domain instead of the binary domain .5 After relaxing
the T coefficient to be in the interval, the problem will be solved using
the KKT conditions. Afterwards, the value of (the optimum T) will be re-
mapped to the binary domain using a simple rounding rule as follows: if
0.5, it will be rounded to 0; otherwise, it will be rounded to 1.
Accordingly, the relaxed optimization problem is given by
 (6.30)

Thus, the solution of the optimization problem presented in (6.30) using

the KKT conditions will be as follows:

PROPOSITION 1: The resource allocation problem presented in (6.30) is

solved in three different cases:

1. System Operates in HD only

In this case, . This case occurs when FD transmission fails to
satisfy the UL channel capacity constraint while HD does satisfy the
constraint.
2. System Operates in FD only
In this case, . This case occurs when HD transmission fails to
satisfy the UL channel capacity constraint while FD does satisfy the
constraint.
3. System will switch between HD and FD
This case occurs when both FD and HD transmissions succeed in
maintaining the UL channel capacity at . Therefore, the system will
choose the mode that maximises the DL channel capacity. In that case,
is given by

(6.31)

Self-Interference Cancellation, Mutual Distance between Users and

Correlation Coefficient Thresholds for the Switching Criterion
Since the performance of FD is greatly affected by the RSI and the mutual
distance between the user transmitting in the UL and the user receiving in
the DL, and the HD operation is influenced by the value of the spatial
correlation between the channels, it will be useful to derive the thresholds
for the self-interference cancellation, the mutual distance between users,
and the correlation coefficient that will cause the system to switch between
FD and HD. In other words, the derived thresholds will define the regions
in which the system will operate in FD or HD.

Self-Interference Cancellation Threshold

Observe from (6.18) and (6.19) that the value of C affects the value of the
UL-received SINR only. Therefore, FD operation will be allowable as long
as the value of C does not make the UL-received SINR less than . In
other words, the minimum allowable value for C is that value that decreases
the UL channel capacity for FD to be equal to . If the value of C is less
than then the UL channel capacity of FD will be lower than , and
hence the system will operate directly in HD. This condition can be used to
obtain the threshold as follows:

PROPOSITION 2: The value of at which is given by

(6.32)

where

(6.33)

(6.34)

Mutual Distance between Users Threshold

Observe from (6.18) and (6.19) that the value of the distance between users
affects the value of the DL-received SINR only. So, if it is assumed that
both FD and HD satisfy the UL channel capacity constraint, the system will
select the mode that will maximize the DL capacity. Therefore, can be
defined by the minimum allowable distance between users m and n such
that is equal to . If the distance between m and n becomes less
than , then the interference introduced by the UL transmission will
increase. Therefore, the DL-received SINR given by (6.18) will decrease as
well. Accordingly, will be lower than . Therefore, the system
will operate in HD. This condition can be used to obtain as follows:

PROPOSITION 3: The value of at which is given by

(6.35)

where

(6.36)

(6.37)

(6.38)

(6.39)

Spatial Correlation Coefficient Threshold

Finding the spatial correlation coefficient threshold will be a little different.
The first difference is that the value of β . In other words, the value of
β is constrained to be greater than zero and smaller than one. In addition,
the value of β influences both the UL-and DL-received SINR values of HD
operation. Therefore, we will have two different thresholds for β that
control switching between HD and FD. The first threshold will be a
result of the necessity that the received UL SINR given by (6.22) equals
. The second threshold is the value of β that makes the HD DL
capacity given by (6.24) equal to the FD DL capacity given by (6.23).

Next, we are interested in finding the value of which makes the

achieved HD uplink capacity equal to . If the actual value of β exceeds
, then the HD uplink capacity will be less than . Therefore, the system
will directly operate in FD. This condition can be used to obtain as
follows:

PROPOSITION 4: The value of at which is given by

(6.40)

where is the minimum possible value of HD uplink capacity that

corresponds to the case β 1, and is the maximum possible value of
HD uplink capacity that corresponds to the case β 0. The constraint
is added to guarantee that the value of is
bounded in . In addition,

(6.41)

(6.42)
 (6.43)

(6.44)

(6.45)

(6.46)

where it is assumed that , and .

The second threshold is found by equating the HD DL capacity given

by (6.24) to the FD DL capacity given by (6.23). If the exact value of β
exceeds , then the HD DL capacity will be smaller than the FD DL
capacity. Therefore, the system will operate in FD. The value of will be
computed numerically.

Therefore, based on the derived thresholds, we can define the regions in

which the system will suffer from outage , operate in FD , or operate
in HD . These regions can be defined, respectively, by

(6.47)

(6.48)

(6.49)
where the values of , , and are summarized in Table 6.2.

Table 6.2 Switching Thresholds

Self-Interference Cancellation Threshold

Mutual Distance between Users Threshold

Spatial Correlation Coefficient Threshold

6.2.3 Numerical Analysis

In this section, we are going to validate our optimization problem solution

by comparing the performance of the relaxed optimization problem solution
presented in Proposition 1 and the exact solution. Furthermore, we are
going to validate the derived threshold formulas by numerical results. In
addition, we will show that our switching criterion is able to choose the
transmission mode that offers the optimal performance represented by the
maximum DL capacity. In addition, our results will study the effect of
different system parameters on the overall system performance.

Validating the Performance of the Relaxed Optimization Problem

In this section, we will validate our optimization problem solution proved in
Proposition 1. Fig. 6.7 shows the variation of the achieved DL capacity with
the AP transmission power. In addition, it compares between the achieved
DL capacity in the case of the exact solution and the relaxed KKT solution
given in Proposition 1. From the results shown in Fig. 6.7 it can be seen that
the biggest deviation between the relaxed and the exact solution
approximately equals at a value of approximately equals . This
indicates that the solution proposed using the KKT conditions gives a good
approximation for the exact solution. In other words, relaxing the T
parameter to the continuous domain is a near-optimal approximation.
 Figure 6.7 Comparing between the relaxed and the exact solution achieved DL capacity

Validating the Derived Thresholds for Self-Interference Cancellation,

Mutual Distance between Users, and Spatial Correlation Coefficient
In this section, we are going to validate the derived thresholds in
Propositions 2-4. Fig. 6.8.a shows the variation of the total achieved UL
capacity per unit bandwidth in the case of FD given by (6.25) with
changing the self-interference cancellation parameter C. It is expected from
(6.19) for there to be an increase in the received UL SINR by increasing C.
This increase is attributable to the decrease in the self-interference level by
increasing C. The increase in the UL-received SINR will result in an
increase in the value of the total UL-received channel capacity given by
(6.25). This behavior can be also verified from the results shown in Fig.
6.8.a. In addition, we can verify the value of calculated in (6.32). From
the results shown in Fig. 6.8.a, it can be seen that the FD-achieved UL
capacity will be equal to at the value of self-interference cancellation,
C approximately equals . This value matches the value
calculated in(6.32).
Figure 6.8 Verifying the validity of the theoretical expressions of (a) the self-interference
cancellation parameter; (b) the mutual distance between users
 Besides, Fig. 6.8.b shows the variation of the achieved DL capacity per
unit bandwidth with the mutual distance between users. From (6.20), it is
expected that the HD performance will not be affected by changing the
mutual distance between users. However, from (6.18), it can be expected
that the DL-received SINR in the case of FD will increase with increasing
the mutual distance between users. This increase is due to the decrease in
the interference level introduced by the UL transmission on the DL
transmission. Accordingly, the DL-achieved capacity in the case of FD
given by (6.23) will increase as well. It is obvious that the results obtained
in Fig. 6.8.b match the discussed performance behavior. Furthermore, our
switching criterion effectiveness can be verified from the results shown in
Fig. 6.8.b. It can be seen that, for small mutual distances between users, HD
was able to achieve better DL capacity than FD. In that case, the proposed
switching criterion chose to operate in HD. By contrast, for large values of
the mutual distance between users, FD transmission starts to out-perform
HD transmission. In that case, the proposed switching criterion chose to
operate in FD. Finally, we can verify the value of calculated in (6.35).
From the results shown in Fig. 6.8.b, it is clear that switching between HD
and FD occurs at equals . In other words, the value of the
achieved DL capacity in the case of FD given by (6.23) equals the value of
the achieved DL capacity in the case of HD given by (6.24) at a value of
equals . This value matches the value of calculated in (6.35).

Furthermore, Fig. 6.9 shows the variation of given in (6.26) with

the spatial correlation coefficient β. It is anticipated that increasing the
spatial correlation coefficient will cause a degradation in the performance of
the HD transmission. Accordingly, it is expected to have a decrease in the
value of by increasing β. This behavior can be verified from the
results shown in Fig. 6.9.6 In addition, we can verify the value of
calculated in (6.40). Notice that the value of becomes equal at a
value of β . This value matches the value of , calculated in (6.40).
In addition, observe that the value of approximately equals
, corresponding to β . In addition, the value
approximately equals , corresponding to β . Since
, we are able to find a valid solution for .
Otherwise, if is not bounded between and , there will be
no valid solution for . This result validates the added constraint on
in (6.40).

Figure 6.9 Validating spatial correlation coefficient threshold

Studying the Effects of Different System Parameters on the Proposed

System Performance
In this section, we are going to study the effect of the uplink capacity
constraint on the achieved DL network capacity and the outage
probability. In addition, the effects of the spatial correlation coefficient β
and the self-interference cancellation C on the total achieved network DL
capacity will be studied.

Fig. 6.10 shows the variation of the total DL capacity with . In the
beginning, for small values of , both HD and FD will be able to fulfill
the UL constraints, and the DL capacities for FD and HD are dependent on
the values chosen in simulation for and β, respectively. As can be
shown from the results, FD will always outperform HD for .
However, when the distance decreases to , FD will outperform HD only
for some values of . Afterwards, as the value of grows, the
achieved DL capacity decreases. This can be explained simply by knowing
that increasing requires the system to allocate more resources for the
UL transmission. However, if continues to increase, the probability of
fulfilling this QoS demand will decrease and hence the outage probability
will increase. Additionally, in the case of system outage, both the UL and
DL capacity will be equal to zero. Therefore, on average, the achieved DL
capacity will decrease. Accordingly, further increases in the value of
will make the outage probability approach one. In other words, the system
will always fail to provide the UL transmission with its QoS demand.
Accordingly, the DL capacity will be zero. In addition, the results shown in
Fig. 6.10 verify the effectiveness of the proposed switching criterion in
choosing the transmission mode that guarantees better performance for the
network. It can be observed that, in the case of and small
values of , our switching criterion is able to achieve a gain of .
However, this gain decreases as grows. The decrease in the achieved
gain is attributable to the fact that increasing will render both modes
unable to fulfill this strenuous demand. Accordingly, the outage probability
will increase. Similarly, in case of , our switching criterion is
able to choose the transmission mode that guarantees better performance of
the network.
 Figure 6.10 Variation of with

Actually, the behavior of the network outage probability with can

be verified from Fig. 6.11.a; it shows the variation of the outage probability
with , for different values of . As shown, it can be seen that the
outage probability increases by increasing . Furthermore, it can be seen
that by decreasing from to , the outage probability decreases.
Actually, it must be noticed from (6.19) and (6.22) that only will
increases by decreasing as a result of the decreased interference level;
is independent of . However, based on the outage probability
definition, decreasing will increase the chance of the FD mode to
satisfy the UL constraint. Consequently, the overall outage probability will
decrease. Additionally, Fig. 6.11.b shows the variation of the outage
probability with for different values of the uplink transmission power
P. Notice that increasing P from to will result in a decrease in
the system outage probability. Similarly, from (6.19) and (6.22), it is clear
that increasing the value of P will increase the value of the UL received
SINR for both FD and HD, respectively. Therefore, the system ability to
fulfill the constraint increases. Hence, the outage probability decreases.
Figure 6.11 Variation of outage probability with with different values of (a) AP transmission
power ; (b) uplink transmission power P
 Moreover, the effect of β on the average achieved downlink capacity is
shown in Fig. 6.12.a. It is expected that the FD-achieved DL capacity
calculated in (6.23) will be independent of β. However, it is also expected
that increasing the spatial correlation between the antennas will decrease
the HD-achieved DL capacity calculated in (6.24). This behavior can be
verified by the results shown in Fig. 6.12.a. In addition, it can be observed
that for , HD achieves better DL capacity than FD. However, for
, FD starts to outperform HD. Therefore, it can be concluded that the
value of at which calculated in (6.23) equals calculated in
(6.24) equals 0.8, which also corresponds to the switching point between
HD and FD. In addition, it is clear that, on average, the proposed switching
criterion is able to achieve a gain of about for . However, this gain
starts to decrease until it reaches for β 1. Finally, effect of the
varying the self-interference cancellation parameter on system performance
is shown in Fig. 6.12.b. As expected, the HD performance is independent of
C. However, increasing C will reduce the values of the self-interference
introduced on the UL transmission in the case of FD. Accordingly, FD
becomes more able to keep the constraint and the achieved DL capacity
will increase. In addition, our switching criterion efficacy can be verified
from Fig. 6.12.b. It can be observed that the proposed switching criterion
can achieve better DL capacity than FD and HD.
Figure 6.12 Variation of with (a) spatial correlation coefficient β; (b) self-interference
cancellation parameter C
6.3 Summary

In this chapter, first, a study of the effect of FD on resource allocation in

small-cell networks has been presented. A power optimization problem is
formulated which aims to maximize the DL channel capacity while keeping
a certain QoS for the UL channel capacity. In addition, a solution algorithm
using the interior point method has been developed. Finally, from numerical
results it is proven that in some cases using FD is more beneficial than
using HD, provided that the self-interference level is sufficiently low.
Second, a complete study of the parameters affecting the performance of
FD and HD operations is presented. In addition, we propose a novel
switching criterion that chooses the transmission mode that maximizes the
DL total channel capacity while keeping the UL channel capacity at a
certain level. Additionally, the switching criterion was formulated as a
resource allocation problem and then solved using the KKT conditions.
Furthermore, in order to identify the network’s operation regions, we
derived different thresholds for the self-interference cancellation parameter,
spatial correlation coefficient, and mutual distance between the users. This
allowed for a clear definition of outage, FD, and HD operation regions.
Besides, with the aid of numerical results, we validated our optimization
problem’s solution and the derived thresholds. Furthermore, our system’s
performance is assessed and compared with both FD and HD performance,
for different system parameters. It was found that our system is able to
achieve high performance gains for small values of UL channel capacity
constraint and spatial correlation coefficient. In addition, it was revealed
that increasing the AP transmission power increases the network outage
probability due to the increased interference introduced to the UL
transmission in FD and HD mode. However, increasing the mobile user’s
transmission power will decrease the network outage probability. This
decrease is due to the fact that increasing the value of mobile user
transmission power will increase the value of the UL-received SINR for
both FD and HD. Finally, the effects of varying both the spatial correlation
coefficient and the UL channel capacity constraint on the network
performance are discussed.
1 The binary variables in the optimization problem make it a nonlinear integer programming
problem; which is an NP-hard problem.

2 denotes the channel coefficient between receiver x and transmitter y.

3 Due to the centralized nature of the cellular network in which the AP fully control the users, the
AP is capable of adjusting the channel estimation procedures with the existing network’s
users.

4 It must be pointed out that the transmission power of the AP must be very carefully chosen in
order to avoid the receiver saturation problem. However, self-interference cancellation, RSI
model, and the receiver saturation problem are beyond the scope of this work. The only focus of
the proposed scheme is the resource allocation problem.

5 The “optimum” approach, which is simply an enumeration of all possible cases of the AP
operation, can easily be applied in our model due to its simplicity. However, the enumeration
method will be tedious if the number of access points increases. Therefore, we need a more
generalized approach to solve the proposed allocation problem.

6 Observe that changing β from to only changes the UL capacity from to

, which is a very small variation.
7 Full-Duplex Cooperative Networks

In this chapter, we introduce Full Duplex (FD) cooperative networks,

consisting of one source node, multiple relays, and one destination. We
assume that the relay works in a full duplex mode. Due to the residual self-
interference in FD relay nodes, the analysis of FD relay systems will
become essentially different from conventional relay systems. We will
analyze the performance of FD Amplify-and-Forward (FA) relay systems.
The performance of FD AF relay systems is limited by the residual self-
interference. To overcome such a limitation, we will introduce an effective
X-duplex relaying protocol, which switches adaptively between FD and
Half-Duplex (HD) spatial multiplexing modes based on the channel
conditions, such that the limitations of FD and HD can be overcome
through such an adaptive protocol. Finally, we consider a general setup of
multi-relay FD systems and introduce a joint antenna and relay selection
scheme for such a network.

7.1 Cooperative Communication Basics

Before introducing FD relay networks, in this section, we first briefly

introduce some basics of cooperative communications and relaying
protocols.

The concept of cooperative communications can actually be traced back

to the early work of Cover and ElGamal on achievable capacity of a relay
network in 1979; it was rediscovered recently in relation to great potential
applications in cellular and wireless sensor networks. The distributed nature
of wireless networks provides a unique opportunity for cooperation and
distributed signal processing. The design of efficient cooperative protocols
and distributed signal processing techniques has been an important issue in
implementing cooperative communications in wireless networks. Therefore,
recent research on cooperative wireless networks has focused on designing
relaying protocols, signaling and distributed coding. In particular, the
design of efficient relaying protocols and distributed coding schemes has
attracted significant attention and a number of novel relaying protocols and
distributed coding schemes have been developed in the past several years.
Capacity-approaching performance has been achieved by some elegantly
designed distributed coding schemes.

When we talk about “cooperative networks” and “relay networks,” they

have the following distinctions. In cooperative networks, each node acts as
both a source and a relay node. That is, each node not only transmits its
own information but also helps other nodes to transmit signals. In relay
networks, the relay nodes are explicitly built nodes only for the purpose of
relaying and forwarding information. They do not have their own
information to transmit. Despite these differences, as far as signal
processing at relays is concerned, they are almost the same. Therefore, in
this book, we use these two terms interchangeably.

Let’s first describe a relay network model. As an example, let us

consider a relay network consisting of one source, multiple relays, and one
destination. Due to the practical constraints in hardware implementation, it
is usually assumed that each node cannot transmit and receive at the same
time. The overall transmission can be divided into two phases. The source
node first broadcasts its messages to all relays and destination. Upon
receiving signals from the source, each relay processes the received signals
and then forwards them to the destination. There are various strategies,
which can be applied at relays to process the received signals. In the
simplest case, relays can just simply amplify the received signals. Relays
can also reconstruct the original messages from the received signals with or
without decoding operations. In a more complex protocol, relays can
calculate and transmit additional information related to source messages,
such as incremental redundancy, in order to improve the system
performance. We will discuss various relaying protocols shortly.

When relays transmit in the second phase, the source could either
transmit at the same time or remain silent. Furthermore, the source and
relays can transmit through orthogonal channels or non-orthogonal
channels. In orthogonal transmissions, all relays transmit to the destination
through orthogonal channels, such that the transmitted signals from each
relay can be separated at the destination, without any interference from
other relays. For non-orthogonal transmissions, relays transmit to the
destination at the same time and same frequency. Therefore the received
signal at the destination is the superposition of signals transmitted from the
source and relays. A silent source in the second phase and orthogonal
transmissions from relays simplify the receiver structures as no multiple
access interference suppression is needed. However, this will reduce the
channel capacity. It is up to a practical system designer to determine which
transmission mode should be used in the relay networks.

At the destination, depending on the transmission mode and relay

protocols, various processing methods will be used to recover the source
messages based on the received packets transmitted from the source and
relays, such as linear combinations, iterative decoding, etc. Since signals
received at the destination arrive via different paths from source and various
relays, a spatial diversity can be achieved at the destination.

Based on operations at the relays, there are several commonly used

relay protocols. In this section, we give an overview of these protocols
[242]:

Amplify-and-Forward (AF): AF is one of the simplest relay

protocols. In AF, upon receiving signals from the source, each relay
simply forwards to the destination a scaled version of the received
signals, including both information and noise. By properly
combining the received signals from the source and relays, the
destination node makes a final decision. Since the destination
receives two copies of signals transmitted from the source and relay
through two independent paths, a spatial diversity can always be
achieved by the AF protocol at high SNRs. Obviously, the major
drawback of AF protocols is noise amplification at the relays. The
outage probability and end-to-end BER performance of the AF
protocol have been widely investigated. Given a fixed transmission
power at the source, relay, and destination, it is shown that the
performance of AF depends on the position of the relay relative to the
source and destination. When the relay is positioned midway
between the source and destination, AF achieves its optimum
performance and worsens as the relay moves closer to or further from
the source.
Demodulate-and-Forward (DemF): To eliminate the effects of noise
amplification, several relay protocols have been proposed. The DemF
is one of the simplest solutions among them. In DemF, the relay
simply demodulates the received signals, with no decoding, and re-
modulates to reconstruct the symbols transmitted by the source. This
process can simply remove the noise components residing in the
received signals at the relay. Its performance variation with respect to
the relay position is similar to AF. When the relay is placed midway
between the source and destination, the lowest outage probability is
achieved.
Decode-and-Forward (DF): DF is another commonly used protocol
for eliminating the noise effect, especially for coded systems. The
relay decodes the received signals and re-encodes them before
forwarding to the destination. When the channel quality in the link
between the source and relay is good, the process of decoding and re-
encoding will provide more powerful error-correcting capabilities
compared with DemF. Thus, the method can considerably outperform
both AF and DemF. However, when the link from the source to the
relay suffers from a deep fading, decoding errors may occur at the
relay. In this case, if the relay re-encodes these incorrect bits, error
propagation will occur and lead to even worse performance.
Adaptive Relay Protocol (ARP): So far, we have seen that different
protocols mentioned above all have their advantages as well as
disadvantages. Intuitively, we would like to ask one question: are
there any possible protocols that can not only effectively mitigate the
noise amplification but also avoid error propagation? ARP is one
protocol developed to meet this need [243]. ARP has the advantages
of both AF and DF and minimizes their negative effects. In ARP,
each relay adaptively selects the AF or DF protocol based on whether
its decoding result is correct or not. All the relays that fail to decode
correctly use the AF protocol to amplify the received signals and
forward them to the destination. On the other hand, all the relays that
can successfully decode the received signals use the DF protocol.
The signals received at the destination, forwarded from all relays by
using either the AF or DF protocol, are combined into one signal to
recover the source information. It has been shown that the ARP
considerably outperforms the AF scheme and simultaneously avoids
error propagation due to the imperfect decoding at relays in the DF
protocol. Thus it outperforms both AF and DF protocols. The
performance gain grows as the number of relays increases, and it
approaches the perfect DF scheme at high SNRs.
Piecewise-and-Forward (PF) [244]: In PF, the received signal at each
relay is compared to an adaptive threshold [244]. If the amplitude of
the received signal is above the threshold, it is assumed to have a
high reliability and the relay will decode the signal; otherwise, the
relay will forward the received signal after linear processing. It is
shown that the PF protocol outperforms the existing AF, DF and
Estimate-and-Forward relaying protocols, and the gain increases as
the number of relays increases.
Soft Information Relaying (SIR) [245]: An alternative approach to
avoiding error propagation in DF is to calculate and forward the
corresponding soft information instead of making a decision on the
transmitted information symbols at the relay. Soft Information
Relaying provides additional information to the destination decoder
to make decisions, instead of making premature decisions at the relay
decoder. It has been shown that the SIR considerably outperforms AF
and DF protocols. Soft information is an analog signal. To transmit
such analog signals, compression and/or quantization should be
performed at the relays to convert analog soft information into digital
signals.
 Figure 7.1 System model of FD AF relay system

7.2 Full-Duplex Amplify-and-Forward Relay

Systems
7.2.1 System Model

In this section, we consider a two-hop FD relay system with one Source (S),
one Destination (D), and one AF Relay node (R). The source transmits the
information to the destination with the help of the relay. The relay works in
the FD transmission mode, and both source and relay use the same time-
frequency resource. For simplicity, we assume that the relay is equipped
with two antennas, one for reception and the other for transmission. All the
links, including the source–relay, relay–destination and self-loop
interference channel, are independent Rayleigh flat fading. We consider that
all the channel links are block Rayleigh fading channels for which channels
remain constant with each frame but change independently from one frame
to another.

The received signal at the relay is given as

(7.1)

where is the channel between the source and the relay, and is the self-
loop interference at the FD relay. and are the signal with the unit
power transmitted from the source and the relay, respectively. and are
the transmit power of the source and relay, respectively. is the additive
white Gaussian noise (AWGN) at the relay with variance .
 Upon receiving the signal from the source, the relay uses the AF
protocol to forward the following signal:

(7.2)

where β is the power amplification factor to ensure that the average power
of signal satisfies the following power constraint,

(7.3)

The received signal at the destination is given by

(7.4)

where is the channel from the relay to destination, and is AWGN

with mean zero and variance at the destination.

Therefore, the end-to-end Signal-to-Interference Noise Ratio (SINR)

can be expressed as

(7.5)

where , , and . We
assume that the distances between the source/destination and the relay are
denoted as and respectively, and the distance from the source to
the destination is denoted as D. The Signal to Noise Ratios (SNR) of the
channel link , and are exponentially distributed with mean ,
and , respectively. Thus the average channel SNRs can be expressed
as , , and , where v denotes the path loss
exponent of the wireless channel, and k denotes the self-loop interference
level.

7.2.2 Performance Analysis of FD AF Relaying

In this section, we present the asymptotic expressions of the outage

probability and symbol error rate (SER) performance by taking into account
the effect of the self-loop interference in FD AF relay systems.

(1) Outage Probability

Due to the self-loop interference at the relay, the SNR expressions at the
destination in the FD relay systems are essentially different from those in
the HD relay systems. In the high transmit power, the end-to-end SINR
in (7.5) can be rewritten as

(7.6)

where . The distribution of X is given by ,

where .

Theorem 1 gives a closed-form asymptotic CDF expression of SINRs.

THEOREM 1: The asymptotic CDF expression of the end-to-end SINR can be

calculated as

(7.7)

where is the first-order-modified Bessel function of the second

kind [200], and .
(2) Average SER
The average SER can be calculated by

(7.8)

where is the CDF of the end-to-end SINR given in Eq. (7.7), and is
the Gaussian Q-Function [200]. The parameters depend on the
modulation formats, e.g., , for BPSK modulation.

THEOREM 2: The asymptotic end-to-end SER expression can be

approximated as

(7.9)

where

(7.10)

and

(7.11)
where the function denotes the hypergeometric function [200], and
the parameters and can be computed recursively as follows,

(7.12)

7.2.3 Relay Location Optimization

When deploying the relay node, it is obvious that the distance from the
source to destination node is not larger than the sum of the distances from
the relay to the source/destination, . Therefore, the optimal
relay deployment must be located on the straight line between the source
and destination node, .

Problem Formulation: For any given transmit power at the source and
relay node ( and ), the self-loop interference level k of the FD relay
node, and the path loss exponent v of the wireless channel, the optimal relay
location can be determined by

(7.13)

By representing the outage probability as the function of distance, the

outage probability of FD relay system can be expressed as

(7.14)

where is PDF of the self-loop interference, and independent from the

distance d. The probability is a convex function in
term of the distance d based on the analytical results of the HD relay
systems [246]. Therefore, the second derivative with respect to d can be
expressed as

(7.15)

Since the second derivative of is positive,

Eq. (7.15) is also positive in term of the distance d. Therefore, is
convex in terms of d. Similarly, since the average SER can be expressed as
the integral of the outage probability (7.14), we know that the average SER
performance is also convex for the distance d.

At the high transmit power, (7.10) can be reduced as

(7.16)

Indicated by (7.16), . Therefore, the SER performance is mainly

determined by the item , and the problem can be simplified to calculate
the optimal location which maximizes the value of only.

(7.17)

THEOREM 3: In the high transmit power region, the optimal location can be
approximately obtained as

(7.18)
 Theorem 3 indicates that the optimal relay location tends to approach
the source node when the relay transmit power increases. This is
because the self-loop interference increases with the relay transmit power,
and the received SINR at the FD relay is degraded. To maintain the quality
of the received signal at the relay, the relay has to move towards to the
source node to enhance the channel . Similarly, the decrease of the
source transmit power also makes the optimal relay location move towards
to the source.

7.2.4 Simulation Results

In this section, we present the numerical results for the FD AF relaying

systems. The path-loss exponent , the distance from the source to the
destination is set as , and the transmit power at the source and relay is
set as equal, i.e., . Without loss of generality, a BPSK
modulation is used in the outage probability and SER performance
evaluations.

In Fig. 7.2, we plot the numerical and analytical results of the outage
probability for the relay system with one FD relay deployed in the middle
of the source and destination. The simulated outage probability curves
tightly match with the expressions in (7.7). We know from Theorem 1 that
the performance floor in the figure is only determined by the loop
interference level k and the average channel gains of and .
 Figure 7.2 The outage probability for different loop interference levels when the relay node is
deployed in the middle between the source and destination,

Fig. 7.3 illustrates the derived SER expression (7.9) for the fixed relay
system and the simulated SER performance for the optimal and the derived
suboptimal relay location results under different self-loop interference level
conditions. The performance of a relay system with the relay deployed at
the middle point is also illustrated for comparison. The simulated SER
curves tightly match with the expressions in Theorem 2. From this figure,
we can see that the derived suboptimal solution approaches nearly the same
performance as that of the optimal one in whole SNR regime and
considerably outperforms that with non-optimized relay location, and the
SER performance gain increases as the SNR increases.
 Figure 7.3 The average SER comparisons with and without the optimal relay locations

Fig. 7.4 investigates the impact of the relay location on SER

performance. The optimal relay location derived by Theorem 3 is a
suboptimal solution for the location optimization problem, and approaches
the optimal one at high transmit power. When the loop interference
increases, the optimized location of the FD relay needs to move toward to
the source node to get the better signal quality, and thus makes the channel
from the relay to destination worse. If the interference cancellation is
perfect at the relay, the optimized relay location will be the middle between
the source and the destination, which is consistent with our analysis in
Theorem 3.
 Figure 7.4 The average SER performances when the relay is deployed between the source and
destination node

7.3 X-Duplex Relay Systems

As shown in previous section, FD has the potential to enhance the spectral

efficiency via its efficient bandwidth utilization. However, in the strong
interference regions, FD performance can be worse than HD. Therefore,
there is a fundamental trade-off between FD and HD transmission modes.

In this section, we introduce a novel X-Duplex protocol [247] and apply

this protocol to relay networks where the relay is equipped with a shared
antenna and two radio frequency (RF) chains used for transmission or
reception. The X-Duplex protocol can be reduced to either FD or HD with
different RF chain configurations. In the X-duplex scheme, the relay
adaptively configures the connection between its RF chains and antenna to
optimize the end-to-end performance according to the instantaneous
channel conditions. We will analyze the overall performance of the X-
Duplex relay system, including outage probability, average SER, and
average sum rate.

7.3.1 System Model of X-Duplex Systems

As shown in Fig. 7.5, we consider a system which consists of one Source

node (S), one Destination node (D), and one Amplify-and-Forward Relay
node (R). The direct link between S and D is assumed to be strongly
attenuated and information can only be forwarded through the relay node.
All nodes in this network use the same frequency band and each of them is
equipped with one antenna. Node R is equipped with one transmit (Tx) and
one receive (Rx) RF chain which can transmit and receive signals at the
same time over the same frequency [248]. In the X-Duplex relay, node R
can adaptively switch between FD and HD modes based on the
instantaneous SNR of the channel from the source to the destination node
and the self-interference between the two RF chains of the relay node.
 Figure 7.5 System model of X-duplex relay

(1) End-to-End SINR

In FD mode, both Tx and Rx chains at node R are active. The received
signal at node R is given by

(7.19)

where denotes the channel between source and relay, denotes the
residual self interference of relay R, and denote the transmit signal of
source and relay with unit transmit power, and are the transmit power
of the source and relay node, and is the additive white Gaussian noise
with the power .
 The relay uses the AF protocol to process the received signal at RX RF
chain and the forwarding signal at relay R can be expressed as

(7.20)

where is the power amplification factor. Factor ensures that the

average power of fulfills the power constraint at relay R.

(7.21)

Factor can be derived as

(7.22)

The received signal at the destination is given by

(7.23)

where denotes the channel link between relay and destination, and is
the additive white Gaussian noise with the power .

The end-to-end SINR of FD mode can be obtained as

(7.24)

With the amplification factor , the SINR can be further simplified as

 (7.25)

where , , , represented the link gains and

.

In the HD mode, the relay R receives the signal from the source at the
first half of a time slot.

(7.26)

In the second half of a time slot, relay R forwards the received signal to the
destination using the AF protocol. The received signal at destination D can
be written as

(7.27)

(7.28)

where is the amplification factor. Under the transmit power constraint at

relay R, can be expressed as

(7.29)

The end-to-end SINR at destination D is thus given by

(7.30)
 We consider a block Rayleigh fading channel. The instantaneous SNR
, can be modeled as the exponential random variable with respective
mean and . We assume that the average chain gains between the
source/destination are the same, . The residual self-
interference at relay is modeled as a block Rayleigh fading channel [124].
The interference gain at the relay R, denoted by , follows the exponential
distribution with the mean value given by

(7.31)

(2) X-Duplex Relay

When the residual self-interference is severe, the HD mode outperforms the
FD mode. On the other hand, in a relatively weak self-interference region,
the FD mode becomes a better option. To optimize the system performance,
we introduce an X-Duplex relay which can be reduced to either FD or HD
with different RF chain configurations based on the SINR performance. In
this case, we assume that the source and relay operate under the same
power constraint, .

In the high SNR region, the system’s sum rate under FD and HD modes
can be expressed as

(7.32)

where and denote the end-to-end SINR of the FD and HD modes,

respectively.

In order to maximize the instantaneous sum rate, the instantaneous

SINR of X-Duplex relay scheme can be given by

(7.33)
(3) X-Duplex Relay with Adaptive Power Allocation
To improve the system performance, the adaptive power allocation (PA) is
incorporated into the X-Duplex relay scheme to maximize the end-to-end
SINR of the relay system under total transmit power constrains,
. The optimal power allocation scheme for FD based on the
instantaneous CSIs is given by [249],

(7.34)

The respective end-to-end SINR of FD and HD with PA can be obtained as

(7.35)

Therefore, the instantaneous SINR of X-duplex relay scheme with power

allocation can be given by

(7.36)

7.3.2 Performance Analysis of X-Duplex Relay Systems

In this section, we investigate the CDF of the X-Duplex relay, and then we
analyze the performance of the X-Duplex system, including outage
probability and SER.

LEMMA 1: The asymptotic complementary CDF (CCDF) expression of SINR

of FD, , is given by

(7.37)

where , and are the first- and zero-order Bessel

functions of the second kind [200].

LEMMA 2: The probability of can be given by

(7.38)

where

LEMMA 3: The asymptotic probability of is given by

(7.39)

where , , is expressed as
 (7.40)

(7.41)

(7.42)

, , is the Heaviside
step function [250].

(1) Distribution of the Received Signal

PROPOSITION 1: The asymptotic complementary CDF of at high SNR can
be derived as

(7.43)

where , , ,
, , are the first- and zero-order Bessel
functions of the second kind, and is the Heaviside step function [250].

(2) Outage Probability

The outage probability is defined as

(7.44)
where the threshold of the outage probability is set so that the transmit rate
is below bps/Hz, and is the CDF function of the end-to-
end SINR γ.

The X-Duplex relay configures the antenna to provide the maximum

sum rate of the relay network. The CDF of the X-Duplex relay system is
given in (7.43). Here we discuss the situation where .

From Lemma 1 and (7.44), the outage probability of the FD mode can
be obtained. In the high SNR conditions, when z approaches zero, the
function converges to , and the value of is comparatively small.
Therefore, in the high SNR scenarios, the outage probability of FD mode is
approximately given by

(7.45)

When the SNR comes to infinity, the outage probability of FD

mode will approach

(7.46)

Therefore, the outage probability of FD mode is limited by the error floor in

the high SNR scenarios.

By substituting (7.43) into (7.44), the outage probability of the X-

Duplex relay can be obtained. In the high SNR region, adopting the similar
approximation in (7.45), the outage probability can be written as

(7.47)
When the SNR approaches infinity, the outage probability approaches zero.

For the X-Duplex relay system, the finite-SNR diversity is defined as

[251]

(7.48)

where denotes the average SNR, and denotes the outage

probability of the relay system at SNR . We use this equation to evaluate
the diversity order of the finite SNR regime.

The finite-SNR diversity order of the X-Duplex relay is obtained as

(7.49)

The diversity order can be estimated using the Taylor expansion

(7.50)

It can be observed that when the transmit power increases to infinity, the
diversity order of X-Duplex relay to approaches one, indicating that there is
no error floor in the X-Duplex system.
 For HD mode, from equation (7.32), the equivalent SINR of HD mode
in one time slot is given as . Therefore, the outage probability of HD
mode can be obtained with the help of (7.38):

(7.51)

In high SNR scenarios, the finite-SNR diversity order of HD mode can be

written as

(7.52)

When the SNR increases to infinity, the diversity order of the HD mode
approaches one. This indicates that the outage probability curve of the X-
Duplex relay system parallels HD mode when the SNR reaches this region.

It is obvious that the outage probability curves of FD and HD systems

will intersect. Let denote the outage probability intersection point of
FD and HD mode and it can be calculated from (7.45) and (7.51) as

(7.53)

when , the outage probability of FD is lower than HD, and the

opposite is the other way around. The conjunction point is determined only
by self-interference level η. When η reaches zero, the conjunction point
increases to infinity, indicating that in ideal self-interference cancellation
scenarios, FD outperforms HD in all SNR circumstances.

7.3.3 Average SER Analysis

The average SER can be computed as

(7.54)

where is CDF function of the end-to-end SINR γ and is the

Gaussian Q-Function. The parameters are decided by the modulation
formats, e.g., for BPSK modulation.

PROPOSITION 2: The average SER of the X-Duplex relay can be expressed

approximatively as

(7.55)

where , ,
, , and , are the incomplete Gamma
Functions, is the Gamma Function, is the Whittaker function
[250].

According to Eq. (7.43), when SNR approaches infinity, the CDF of

is given as , and the SER of
X-Duplex relay scheme comes to zero.
 For FD mode, when the SNR approaches infinity, the CDF of FD mode
becomes . With Eq. (7.54), the SER of FD mode can
be obtained as

(7.56)

From Eq. (7.56), it can be seen that the SER of FD mode is limited by
the lower bound, which is determined by self-interference level η, , .
Compared with FD mode, the X-Duplex scheme removes the error floor
and achieves a lower SER in high SNR regions.

7.3.4 Average Sum Rate

Based on the CDF expressions of , in this section, we calculate the

average weighted sum rate of X-Duplex system.

(7.57)

where is the CDF of .

Before presenting the sum rate, let’s first introduce two integrals. Let
and denote the approximate value of integral
and and their expressions are
given in Lemmas 4 and 5, respectively.
LEMMA 4: When , the exact value of integral
is given by

(7.58)

where is the probability integral, and is the Whittaker function

[250], we use the first items of the third part of Eq. (7.58) for
approximation, denoted as .

LEMMA 5: The approximate value of integral is

given by

(7.59)

where , , , , , the first

items are used to approximate value.

PROPOSITION 3: The average sum rate of the X-Duplex system is given below
 (7.60)

According to (7.43) and (7.57), when the SNR goes to infinity, the CDF
of is given as and the
average sum rate of X-Duplex relay scheme can be derived.

(7.61)

It can be observed that the maximal achievable average sum rate of the X-
Duplex relay scheme is not restricted by the self-interference.

The average sum rate of FD mode can be obtained with Eq. (7.37) and
Eq. (7.57). When SNR approaches infinity, the upper bound of FD mode
can be derived with [250, Eq. (3.195)].

(7.62)

Eq. (7.62) is the upper bound of the average sum rate of FD mode. It
means the practical average sum rate cannot go above Eq. (7.62).

Comparing the Eq. (7.61) and Eq. (7.62), the X-Duplex relay scheme
overcomes the restriction of self-interference compared with the FD mode.
(3) Diversity Order of the X-Duplex Relay with PA
In this subsection, we give an upper bound and a lower bound for the end-
to-end SINR of the X-Duplex relay with power allocation and obtain the
CDF expression of these two bounds. Finally, we provide the diversity
order of the X-Duplex relay system with PA.

The end-to-end SINR (7.36) can be upper-bounded as

(7.63)

where , ,
, . The function
is a monotonically increasing function in the region .
Therefore the function is also a monotonically increasing function in
this region.

The CDF distribution of and is given as

(7.64)

With the help of [250, eq. (3.322)], the outage probability of the upper
bound is obtained as

(7.65)
where , , is the PDF of ,
.

The Taylor expansion of the upper bound is given by

(7.66)

It indicates that the diversity order of the X-Duplex relay scheme with PA is
at least one.

Similarly, the outage probability of the lower bound can be

calculated as

(7.67)

Using the Taylor’s series, we know that

(7.68)

where
 It indicates that the diversity order of the X-Duplex relay scheme with
PA is at most two.

7.3.5 Simulation Results of X-Duplex Relay Systems

In this section, we provide the performance of the X-Duplex relay. In

Fig. 7.6, we provide the outage probability of the X-Duplex relay with
different levels of self interference η = 0.2, 0.05, and 0.01. The threshold of
the outage probability is set to guarantee the transmit rate is 2 bps/Hz.
Based on this figure, we can observe that the exact outage probability given
by Monte Carlo simulation curves is tightly matched with the analytical
expression in (7.44). In the high transmit power regions, the effect of self-
interference is very small.

Figure 7.6 Outage probability of X-Duplex relay system versus the transmit power
 In Fig. 7.7, the outage probability is compared between X-Duplex, FD
and HD schemes with self interference . The figure reveals that the
outage probability of X-Duplex relay is lower than both FD and HD
schemes. In the high SNR scenarios, the FD scheme encounters the error
floor, which matches with analytical results in (7.46). When the SNR
approaches infinity, the X-Duplex relay eliminates the error floor caused by
self-interference and achieves the same diversity order of the HD scheme,
which is given in (7.50) and (7.52).

Figure 7.7 Outage probability of X-Duplex relay system versus the transmit power ; the
dashed lines correspond to the numerical evaluation of (7.46), the conjunction point accord with
analytical results in (7.53)

Fig. 7.8 plots the SER in the X-Duplex relay system with . The
SER performance of FD and HD is depicted for comparison. From the
figure, it can be observed that X-Duplex relay achieves a better
performance compared with the pure FD and HD scheme. Besides, X-
Duplex is not restricted by the error floor that FD scheme encounters in the
high SNR region. In addition, this figure verifies the average SER
expression given by (7.55).

Figure 7.8 Average SER of X-Duplex relay system versus the transmit power,

Fig. 7.9 depicts the average sum rate of the X-Duplex system versus
SNR with levels of self-interference . The approximate analytical
expression in (7.60) is very close to the exact average sum rate. It is
obvious that the X-Duplex protocol provides a higher sum rate than the FD
or HD modes.
 Figure 7.9 Average sum rate of X-Duplex relay system versus the transmit power,

In Fig. 7.10, the simulated average sum rate of the X-Duplex system
versus self-interference with different levels of transmit power is depicted.
In the weak self-interference region, the average sum rate of FD mode is
larger than that of HD mode. As self interference increases, the average sum
rate of FD mode significantly decreases and performs worse than the HD
mode. The average sum rate of the XD scheme is always better than the FD
and HD mode. In the high self-interference region, the sum rate
performance of XD scheme is obviously better than the FD mode and is
very close with the HD mode.
 Figure 7.10 Average sum rate of X-Duplex relay system versus self-interference

Fig. 7.11 shows the outage probability of the X-Duplex relay scheme
with power allocation. The total transmit power of the source and relay
node is limited. The performance of the X-Duplex scheme, in which the
source and relay have the same transmit power, is plotted for comparison.
This figure indicates that adaptive power allocation can further improve the
performance of the XD scheme. It can be seen that the diversity order of the
proposed X-Duplex scheme with PA changes within the region from one to
two.
 Figure 7.11 Outage probability of X-Duplex relay scheme with power allocation versus the
transmit power

7.4Full-Duplex Relaying with Joint Relay and

Antenna Mode Selection
7.4.1 System Model

In this section, we consider a multiple relay system which consists of one

Source (S), one Destination (D), and N AF relay nodes, as shown in
Fig. 7.12. We assume that there is no direct channel link between the source
and the destination due to the strong path-loss and attenuation in this link.
Both source and relay nodes use the same time-frequency resource, and the
relay nodes work in FD mode with two antennas (one for receiving and one
for transmitting), denoted by and . We will introduce a joint Relay and
Antenna Mode Selection (RAMS) scheme, where only the optimal relay
with the optimal configuration of FD Tx and Rx antenna mode is selected to
forward the signals from the source to the destination. For each relay, the
Tx/Rx mode of the two antennas is adaptively determined based on the
instantaneous channel conditions between the source/destination and the
two antennas of the FD relay nodes.

Figure 7.12 System model of the multiple AF relay system, and

(1) End-to-End SINR Analysis

Assuming that the ith relay node, denoted by , is selected to forward the
signal, and the relay node configures its antenna / to receive/transmit
the signal, , and .

The signal received by antenna of the relay is given as

(7.69)

where is the channel link between the source and the antenna of the
relay , and is the residual self-loop interference from the transmit
antenna to the receive antenna of the relay . and are the tth
signal with the unit power, transmitted from the source and the relay ,
respectively. and are the transmit power of the source and relay node.
is the AWGN at the receive antenna with the power .
 Upon receiving the signal from the source, the relay uses the AF
protocol to forward the following signal.

(7.70)

where is the power amplification factor to ensure that the average

power of signal satisfies the following power constraint:

(7.71)

From Eq. (7.71), we have

(7.72)

The received signal at the destination is given by

(7.73)

where is the channel link from the transmit antenna of the relay to
destination, and is the AWGN with mean zero and variance at the
destination.

The end-to-end SINR can be obtained as

 (7.74)

Substituting (7.72) into (7.74), the end-to-end SINR via the relay can
be calculated as

(7.75)

where , , and , representing

the link gains between the source/destination and the relay , and the
interference gain from the transmit antenna to the receive antenna at the
relay respectively.

We consider the block Rayleigh fading channels for all links. The mean
value of channel gains between the source/destination and the Tx/Rx
antenna of the FD relay nodes, and , are denoted by and ,
respectively. We also assume that the average channel gains between the
source/destination and the two antennas of the relay is the same,
and . The residual self-loop interference
between Tx and Rx antenna at relay is modeled as a block Rayleigh fading
channel. We assume that the mutual interferences between Tx and Rx
antenna at each relay node, and at the relay node , are the same
due to the same loop interference cancellation method with the same
configurations. Therefore, the interference gain of the relay ,
, follows the exponential distribution with the
expectation of . The interferences at different relay nodes are
independent.

(2) Joint Relay and Tx/Rx Antenna Mode Selection

We assume that each relay node is equipped with two antennas. One
antenna is used to receive the signal from the source node, and the other is
to forward the signal at the same time.
 For the two-antenna relay node, there are two possible channel link
pairs between one relay and source/destination nodes, depending on which
antenna of and is used as Tx and Rx antenna at relay. It is called
when relay configures the antenna as the receive antenna and
as the transmit antenna, and called when relay configures
as the receive antenna and as the transmit antenna. The optimal Tx/Rx
mode for relay is determined based on the maximal end-to-end SINR
criterion as follows:

(7.76)

where denotes the end-to-end SINR when the relay works under
, and can be computed via Eq. (7.75). Similarly, is the end-to-
end SINR via the relay under . The channel state information
(CSI) of and can be measured by the standard pilot-based channel
estimation. The CSIs of the loop interference channel can be obtained
through sufficient training.

In the multiple relay system, the optimal relay with the optimal Tx/Rx
antenna mode configuration is selected to optimize the end-to-end SINR
performance. Therefore, RAMS scheme can be formulated as

(7.77)

where . At the beginning of each transmission, all the

CSIs are measured and collected at a decision node (say source node). The
system computes the end-to-end SINR of the candidate relaying modes,
and selects the optimal relay with the optimal antenna mode
configuration based on the end-to-end SINR maximizing criterion
in the RAMS scheme as shown in equation above. The selected relay node
will be used to help to forward the signal from the source to the
destination under the optimal mode .
 The proposed RAMS scheme can be extended to the relay system with
more than two antennas at the relays. We assume that the FD relay node
equipped with one transmit module and one receive module and L antennas,
denoted as . The relay chooses one transmit antenna and one
receive antenna from L antennas, and there are possible
combinations at each FD relay. At the beginning of each transmission, all
the CSIs are measured and collected at a decision node. The system
computes the end-to-end SINR of the candidate relaying modes,
and selects the optimal relay with the optimal antenna mode
configuration based on the end-to-end SINR maximizing criterion
in the RAMS scheme. The selected relay node will help forward the
signal from the source to the destination through the selected Tx and Rx
antennas under the optimal mode .

(3) Joint Relay and Tx/Rx Antenna Mode Selection with Adaptive
Power Allocation
The above selection scheme is designed with the fixed transmit power
allocation, and suffers from the performance floor due to the loop
interference at the FD relay node. To improve the system performance, the
adaptive power allocation (PA) is imported into the RAMS scheme. In this
subsection, joint relay and Tx/Rx antenna mode selection scheme with
adaptive power allocation (RAMS-PA) is designed to maximize the end-to-
end SINR of the relay system under the total transmit power constraints,
. For the relay configured with Rx antenna and Tx antenna
, the adaptive power allocation problem is formulated as

(7.78)

The optimal power allocation scheme based on the instantaneous CSIs

can be calculated as [249]
 (7.79)

The end-to-end SINR with power allocation can be obtained as follows

in Eq. (7.80).

(7.80)

For the two-antenna relay node, the optimal Tx/Rx mode with power
allocation for relay is determined based on the maximal end-to-end SINR
criterion as follows:

(7.81)

where , and , denotes the end-to-end SINR when the

relay works under with adaptive power allocation, and can be
calculated via Eq. (7.80).

The optimal relay with the optimal Tx/Rx antenna mode configuration
and adaptive transmit power is selected to optimize the end-to-end SINR
performance in the multiple two-antenna relay systems. Therefore, the
modified RAMS scheme with adaptive power allocation can be formulated
as
 (7.82)

where . The system computes the end-to-end SINR

of the possible relaying candidates, and selects the optimal relay
with the optimal antenna mode configuration and the optimal power
allocation, and , based on the end-to-end SINR maximizing
criterion. The selected relay node will be used to forward the signal
from the source to the destination with the optimal mode and
transmit power .

7.4.2 Performance Analysis of RAMS

In this section, we analyze the performance of the RAMS scheme with

fixed transmit power, in terms of the outage probability, the average SER
and the ergodic capacity. The diversity order for the RAMS-PA scheme is
analyzed in the last part of this section, and a region of the diversity order is
given for the RAMS-PA scheme.

(1) The Distribution of the End-to-End SINR

Let us first calculate the distribution of the end-to-end SINR at the
destination via the ith relay node. The approximate CDF expression of the
end-to-end SINR via the ith FD relay node, denoted by , can be calculated
as [252]

(7.83)

where

(7.84)
and is the first-order modified Bessel function of the second
kind [200], and .

Based on the derivations, we know that when the link channel SNRs,
and , are very large, or the loop interference is very
small, the above approximation becomes very accurate. Therefore, the
approximate CDF expression (7.83) for the end-to-end SINR approaches
the exact CDF in high link SNR or the small self-loop interference
scenarios.

When the average link SNRs go to infinity, and

, the factor in Eq. (7.83)
approaches one [253], and thus . The distribution of the received
SINR via the ith relay in the high transmit power scenarios will approach

(7.85)

In the high link SNR conditions, the distribution of the end-to-end SINR
approaches a special distribution which is only determined by the self-loop
interference level at the relay node. This indicates that the performance of
the FD relay system is similar in the high link SNR scenarios, and
increasing the transmit power at the transmitter will not be useful to
improve the system performance.

(2) Outage Probability

In the RAMS scheme, the optimal relay with the largest end-to-end SINR is
selected to forward the signal. Therefore, the outage probability of RAMS
scheme of the N relay networks can be calculated as

(7.86)
 Under the fixed power allocation condition, we assume that the transmit
power of the source and the relay is same, . Moreover, in the
high transmit power scenario, the function approaches ,
the outage probability can be written as

(7.87)

For the FD relay system, at high SNR region, the diversity order
becomes zero due to their irreducible error floor caused by the self
interference of the FD relay [254]. The finite-SNR diversity is defined
as [251]

(7.88)

where denotes the average SNR of the channels from the source/relay to
the relay/destinaton, and denotes the outage probability of the FD
relay system at link SNR . It can be used to estimate and evaluate the
diversity order of the proposed scheme at finite SNR regime.

The finite-SNR diversity order of the RAMS scheme is obtained

as

(7.89)
 In the transmit power scenario, the Taylor expansion can be used to
simplify the expression . The diversity order can be estimated as

(7.90)

where . We can observe that when the transmit power

increases to infinity, the diversity order approaches zero, representing the
performance floor of the FD relaying system. It also shows that in the small
loop interference condition, , the diversity order of RAMS scheme
can reach almost before the reaching the performance floor.

The performance floor of the outage probability are expressed as

(7.91)

The performance floor of the RAMS scheme is determined only by the

self-loop interference level of the N relay nodes, and is lower than that
without the Tx/Rx antenna mode selection.

(3) Average SER Analysis

In the wireless communication system with linear modulations, the average
SER of the RAMS scheme can be calculated as follows [255]:

(7.92)

where is the CDF function of the end-to-end SINR given in Eq. (7.83),
and is the Gaussian Q-Function [200]. The parameters are
decided by the modulation formats, e.g., , for the orthogonal
binary frequency-shift keying (BFSK) modulation [255, eq. (6.31)], and
, for the binary phase-shift keying (BPSK) modulation [255, eq.
(6.6)].

To compute the average SER, let denote the N relay

nodes in the system, and the sets and are the subset of . Moreover, let
the sets , and be the subsets of , and the sets , , and be
the subsets of . The average SER of the proposed RAMS scheme can be
expressed as follows:

(7.93)

where

(7.94)
 The function denotes the size of one set, and denotes the empty
set. If the set or is empty, the value of the product operation is 1.
Since the average SER is calculated by the integral of the CDF
expression (7.86), the approximate SER expression (7.93) approaches the
exact SER performance in the high link SNR region.

When the link SNRs increase, and , the

parameter c approaches a constant value, and the items of the nonempty sets
, , and in Eq. (7.93) approach zero. Therefore, the error floor of the
average SER can be simplified as

(7.95)

where

Eq. (7.95) is independent of the link SNR, and indicates that there is an
error floor in the high SNR region representing the minimal value of the
achievable average SER for our proposed RAMS scheme. The error floor is
only determined by the loop interference of the relays in the FD relay
system.

(4) Capacity Analysis

Based on the CDF expression of the end-to-end SINR, the capacity of the
RAMS scheme can be calculated. Through integration by parts, the capacity
of the FD relay system with RAMS scheme can be expressed as [256]
 (7.96)

The ergodic capacity can be approximated as

(7.97)

where

(7.98)

Using the residue theorem [257], the function can be rewritten as

follows:

(7.99)
where

With the help of the formula [253, eq. (3.353.2)], we can then derive the
ergodic capacity of RAMS scheme.

(7.100)

where , and
denotes the exponential integral function [200].

When the link SNRs go to infinity, the CDF of the end-to-end SINR
approaches the distribution of Eq. (7.91). Thus, the upper bound of the
ergodic capacity can be expressed as

(7.101)

The fraction in the integral (7.33) can be rewritten as

(7.102)

and

(7.103)

Therefore, the upper bound of the ergodic capacity is given by

(7.104)

Eq. (7.36) is the upper bound of the ergodic capacity for RAMS
scheme. It represents the maximal achievable capacity of RAMS scheme,
which means that the ergodic capacity of RAMS scheme in practice cannot
be larger than the derived upper bound (7.36). Therefore, it provides an
achievable region of the ergodic capacity for our proposed RAMS scheme.
(5) Diversity Order of RAMS-PA Scheme
In this subsection, we provide an upper bound and a lower bound for the
end-to-end SINR of the proposed RAMS-PA scheme. Then we derive the
CDF expression of these two bounds, and provide the region of the
diversity order of RAMS-PA scheme.

Consider the scenario that the total transmission power of the relay
networks is P, and the relay is selected to forward the signal from the
source to the destination with Rx antenna and Tx antenna . There is an
upper bound and a lower bound for the end-to-end SINR (7.80)

(7.105)

where , and ,
. The function is a monotone increasing function in
the region .

Based on the assumption about the loop interference at the two-antenna

FD relay, we know that . Due to the monotonicity of the
function , the upper and lower bound for the end-to-end SINR after
the Tx/Rx antenna mode selection are obtained

(7.106)

where , and .
The CDF distribution of and is expressed, respectively, as
 (7.107)

The outage probability of the upper bound can be

calculated as

(7.108)

where , and denotes PDF of the variable ,

which can be calculated from Eq. (7.39).

With the help of the formula [253, eq. (3.322.1)], the outage probability
of is obtained as

(7.109)

where
 (7.110)

and the function is the error function defined in [200]. Using the Taylor
series, we know that

(7.111)

It indicates that the diversity order of the Tx/Rx antenna mode selection is
at least one. Therefore, the diversity order of RAMS-PA scheme for the N
FD relay system is at least N.

Similarly, the lower bound is calculated as follows:

(7.112)

The Taylor expansion of the lower bound is

 (7.113)

It indicates that the diversity order of the Tx/Rx antenna mode selection
is at most two, and the diversity order of RAMS-PA scheme for the N FD
relay system is at most .

The outage probability of the received SINR with the power allocation
via the ith relay of the FD relay networks can be expressed as

(7.114)

When the interference is very small, , the outage probability

approaches

(7.115)
 The diversity order of approaches one when the loop
interference is very small. Thus the diversity order of the RAMS-PA
scheme for a relay system with N relays will approach in low
interference conditions.

On the other hand, we can calculate the exact outage probability of

as

(7.116)

According to Eq. (7.46), we know that is a monotone

increasing function in terms of . When the average loop interference
increases, the outage probability performance becomes worse.

From the distribution of the upper bound Eq. (7.43), we know that the
diversity order of the upper bound is N. The outage probability of RAMS-
PA scheme is always smaller than the upper bound, which means that the
diversity order of RAMS-PA scheme is at lest N (bigger than zero), and thus
the error floor caused by the self-interference can be removed with the
power allocation.

7.4.3 Simulation Results

In this section, we illustrate the performance of the RAMS scheme for a FD

relay system. Without loss of generality, all the channel gains, and ,
and , are all set as one. Under the fixed power
allocation condition, the transmit power of the source and relay is set as
equal, . We consider a BPSK modulation in outage probability and
average SER performance evaluation. The power of the noise is considered
to be constant, .
 Fig. 7.13 plots the outage performance versus the transmit power for the
FD relay system with and the self-loop interference level
. The performance of the conventional optimal relay selection
(ORS) scheme, proposed in [207], is also illustrated for comparison. As can
be seen, the simulated outage probability curves tightly match with the
expression in (7.86). It can be observed from the figure that there is a
performance floor which coincides with the result computed by (7.91). In
the same self-loop interference environment, the proposed RAMS scheme
outperformes that with the fixed antenna configuration obviously. For
example, the RAMS in a single relay network can achieve a similar
performance as the conventional ORS scheme of the system with two relay
nodes.

Figure 7.13 Outage probability of RAMS scheme versus the transmit power, ([252]
Reproduced with permission from the IEEE.)

Fig. 7.14 shows the finite SNR diversity order of RAMS scheme with
different self-loop interference level for , and the finite SNR diversity
order with the fixed Tx/Rx antenna configuration is provided for
comparison. Although when the self-loop interference exists, the finite SNR
diversity order of RAMS scheme or the fixed antenna case approaches zero
in the high transmit power region, the diversity of RAMS scheme is larger
than that of the fixed antenna case in the medium transmit power region.
The simulation results of the case of are also plotted in this figure. As
expected, the finite SNR diversity order of RAMS scheme achieves
without the self-loop interference at the relay node, while the finite SNR
diversity order of the fixed antenna case is only N. Moreover, for the case of
, the finite SNR diversity order almost reaches before reaching
the error floor, which validates the analysis of the finite SNR diversity order
in the previous section.

Figure 7.14 Finite SNR diversity order of RAMS scheme and conventional optimal relay
selection scheme versus the link SNR, relays ([252] Reproduced with permission from the
IEEE.)
 Fig. 7.15 shows the outage probability of the RAMS-PA scheme with
the different self-loop interference levels in the single relay system. As
described in the RAMS-PA scheme, the transmit power of the source and
FD relay node varies with the instantaneous CSIs based on the total system
power constraint. For comparison, the performance of the RAMS scheme is
also plotted in this figure with equalized power allocation for the source and
relay node. As expected, the power allocation can provide a better
performance to the RAMS scheme than the fixed power allocation,
especially when the total system power constraint is very large. The error
floor is eliminated due to the adaptive power allocation. This figure
indicates that the outage probability of the RAMS-PA scheme increases
with the increase of the self-loop interference level η. It can be seen that the
diversity order of the proposed RAMS-PA scheme increases when the self-
loop interference level decreases, and changes within the region from N to
. Therefore, the diversity order depends on the self-loop interference.

Figure 7.15 Outage probability of the proposed RAMS-PA scheme with the different self-loop
interference levels in the single FD relay system ([252] Reproduced with permission from the
IEEE.)
 Fig. 7.16 compares the SER performance of the proposed RAMS with
the conventional relay selection scheme with fixed relay Tx and Rx antenna
configuration [207] for a system with fixed antenna configuration. The
average SER is plotted versus the transmit power with relays at
. The RAMS significantly outperforms the ORS scheme.
Specifically, the RAMS achieves twice the diversity order at low to medium
SNRs and a much lower error floor at high SNRs, compared to the ORS
scheme. In the figure, we also compare the analytical results with
simulation results. It can be seen from the figure that the analytical results
in (7.93) is asymptotically tight.

Figure 7.16 Average SER of RAMS scheme versus the transmit power, , the dashed
lines correspond to the numerical evaluation of (7.93) and (7.95) ([252] Reproduced with
permission from the IEEE.)

The ergodic capacity of RAMS scheme with and , 0.05,

and 0.01 is shown in Fig. 7.17. We can observe that the approximate
results (7.32) are asymptotically tight as SNR increases. Even at low SNR
region, the analytical approximation is also very close to the simulation
results.

Figure 7.17 Capacity of RAMS scheme versus the transmit power when the number of relays is
; the dashed lines correspond to the numerical evaluation of (7.32) and (7.33) ([252]
Reproduced with permission from the IEEE.)

Fig. 7.18 compares the simulated capacity performance of the FD and

HD relay system with the antenna mode selection. In the simulation, we
assume that the source and the destination are equipped with one antenna,
and the relay nodes are equipped with two antennas. The transmit power of
the source and relay is set as equal in both FD and HD antenna selection
schemes. The performance of the FD relay system with RAMS is simulated
for the different self-loop interference levels. The antenna mode and relay
selection scheme for HD relay system proposed in [258] is used in the
simulations, where the HD relay node chooses the optimal antenna from the
same antenna set based on the maximal link SNR criterion at the multi-
access phase and the broadcast phase respectively. The HD relay system
selects the optimal relay with the optimal antenna configuration to achieve
the maximal ene-to-end SNR. The simulation results indicate that when the
self-loop interference is small, the capacity performance of RAMS in the
FD relay systems outperforms that in the HD relay systems. With the
increase of the self-loop interference, the performance of the FD relay
systems decreases obviously, and becomes worse than the HD relay
systems.

Figure 7.18 Capacity of the full duplex and half duplex relay networks versus the self-loop
interference level η ([252] Reproduced with permission from the IEEE.)

Finally, Fig. 7.19 compares the ergodic capacity for RAMS scheme with
the conventional ORS scheme versus the number of relays N for different
SNR values ( = 0, 10, and 40 dB) at η = 0.05. It can be seen that the
system capacity increases with N and the capacity gain of RAMS relative to
the ORS also increases with SNR.
 Figure 7.19 Capacity of RAMS scheme versus the the number of the relays with the self-loop
interference level ([252] Reproduced with permission from the IEEE.)

7.5 Concluding Remarks

In this chapter, we have introduced FD cooperative networks. We have

focused on a single hop relay network consisting of one source node,
multiple relays and one destination. We assume that the relay works in an
FD mode. Due to the residual self-interference in FD relay nodes, the
analysis of FD relay systems has become essentially different from
conventional relay systems. We studied the FD Amplify-and-Forward relay
system. It is shown that its performance is limited by the residual self-
interference and thus there is an error floor (rate ceiling) in error (rate)
performance. To overcome such a limitation, we presented an effective X-
duplex relaying protocol, which switches adaptively between FD and HD
spatial multiplexing modes based on the channel conditions, such that the
limitations of FD and HD can be overcome through such an adaptive
protocol. Finally, we considered a general setup of multi-relay FD systems
and introduced a joint antenna and relay selection scheme for such a
network. The future research problems include FD two-way relay networks,
where all the nodes in the network, such as source, relays and destination,
are all operated in FD mode. Other FD relay network setups include (1) FD
multiple access relay network, where multiple HD source nodes
communicate with the HD destination via a FD relay; and (2) FD multi-way
relay networks, where multiple FD nodes exchange information via a
multiple antenna FD relay network. The design and analysis of these FD
systems will be essentially different from the conventional HD systems, but
some analytical tools presented in this chapter can be used.
8 Full-Duplex Cognitive Radio Networks

With the proliferation of wireless services and the ever increasing data rate
demands, spectrum resources have become more and more scarce. As a
promising technique to increase the efficiency of spectrum utilization,
Cognitive Radio (CR) has great potential to meet such a requirement by
allowing unlicensed users (Secondary Users, SUs) to coexist in licensed
bands allocated to licensed users (Primary Users, PUs). In conventional CR
systems, the spectrum sensing is performed at the beginning of each time
slot before the data transmission, which is also known as the “listen-before-
talk” protocol. While this protocol has worked well in CR networks, two
inherent problems still exist: 1) transmission time reduction due to sensing,
and 2) sensing accuracy impairment due to data transmission.

While a great many works have discussed the design of sensing interval
and duration to minimize the impact of the above problems, we, on the
other hand, manage to apply FD technology to CR networks to bypass the
problems. Specifically, FD technology enables simultaneous sensing and
transmission for SUs. In other words, neither sensing nor data transmission
needs to be interrupted by the other. In this way, SUs can react promptly to
the PUs’ access and departure, and possibly fully utilize spectrum
opportunities for data transmission.

In this chapter, we first provide some preliminaries of cognitive radio

and some existing works. Then, we elaborate the proposed FD CR protocol,
named the “Listen-And-Talk” (LAT) protocol, and provide detailed analysis
about the parameter design and a unique trade-off between secondary
transmit power and secondary throughput. Based on the basic LAT protocol
that evolves only one pair of SUs, we extend the scenario to cooperative
spectrum sensing and dynamic spectrum access, respectively, in the next
two sections. Finally, we list some key challenges in the design and
implementation of FD CRNs.

8.1 Cognitive Radio Basics

The existing and new wireless technologies, such as smart phones, wireless
computers, and WiFi home and business networks, are rapidly consuming
the radio spectrum. Unlike the wired Internet, the wireless world has a
limited number of links to distribute. The usage of radio spectrum resources
and the regulation of radio emissions are coordinated by national regulatory
bodies like the Federal Communications Commission (FCC). These bodies
assign spectrum to licensed holders, also known as Primary Users (PUs), on
a long-term basis for large geographical regions. However, a large portion
of the assigned spectrum remains under-utilized [259]. The inefficient usage
of the limited spectrum necessitates the development of Opportunistic
Spectrum Access (OSA) schemes, where users who have no spectrum
licenses, also known as Secondary Users (SUs), are allowed to use the
licensed spectrum without causing harmful interference for the PUs.

Cognitive radio is the key enabling technology that enables OSA

networks to utilize the spectrum more efficiently in an opportunistic fashion
without interfering with the PUs. It is defined as a radio that can change its
transmitter parameters according to the interactions with the environment in
which it operates. It differs from conventional radio devices in that
cognitive users have cognitive capability and reconfigurability [260].

Cognitive capability (spectrum sensing) refers to the ability to sense and

gather information from the surrounding environment, such as information
about transmission frequency, bandwidth, power, modulation, etc. With this
capability, SUs can identify the best available spectrum hole. A wealth of
literature on spectrum sensing strategies focuses on primary transmitter
detection based on the local measurements of SUs, since detecting the PUs
that are receiving data is in general very difficult. According to the priori
information they require, and the resulting complexity and accuracy,
spectrum sensing techniques can be categorized in the following three
types, namely energy detector, feature detector, and matched filter detector:

Energy detection is the most common type of spectrum sensing

strategy. Secondary users estimate the spectrum occupancy by
comparing the average received power with a fixed threshold. Energy
detection is easy to implement and requires no a priori information
about the primary signal.
When some specific features, especially cyclostationary features
associated with the PU’s signal, are known by SUs, a feature detector
can be used [261]. As in most communication systems, the
transmitted signals are modulated signals coupled with sine wave
carriers, pulse trains, hopping sequences, or cyclic prefixes, while the
additive noise is generally wide-sense stationary with no correlation;
cyclostationary feature detectors can be utilized to differentiate noise
from the PU’s signal [262] and distinguish between different types of
transmissions and primary systems [263].
If SUs know information about a PU’s signal a priori, then the
optimal detection method is matched filtering, since a matched filter
can correlate the already known primary signal with the received
signal to detect the presence of the PU and thus maximize the SNR in
the presence of additive stochastic noise.

Reconfigurability refers to the ability to rapidly adapt the operational

parameters according to the sensed information in order to achieve optimal
performance as shown in Fig. 8.1 [264].
 Figure 8.1 Channel sensing and dynamic spectrum access of CR users

In a common OSA network, the PUs, usually not equipped with CR,
have priority in using the spectrum band. Whenever they reclaim the
spectrum usage, SUs have to adjust their operating parameters, such as
power, frequency, and bandwidth, to avoid interrupting the PUs.

With cognitive capability, SUs are able to obtain an accurate estimation

of the interference temperature or spectrum occupancy status. After this
estimation, SUs need to address issues such as when and how to use a
spectrum band, how to coexist with PUs and other SUs, and which
spectrum band they should sense and access if the current one in use is not
available, which are in the scope of spectrum allocation and sharing
strategies.

For SUs, there are mainly two methods to perform spectrum access,
namely spectrum underlay and spectrum overlay:

1. Spectrum underlay: In spectrum underlay, SUs are allowed to transmit

their data in the licensed spectrum band when PUs are also transmitting.
The interference temperature model is imposed on SUs’ transmission
power so that the interference at a PU’s receiver is within the
interference temperature limit and PUs can deliver their packet to the
receiver successfully. Spread spectrum techniques are usually adopted
 by SUs to fully utilize the wide range of the spectrum. However, due to
the constraints on transmission power, SUs can only achieve short-range
communication. If PUs transmit data all the time in a constant mode,
spectrum underlay does not require SUs to perform spectrum detection
to find available spectrum band.
2. Spectrum overlay: Spectrum overlay is also referred to as opportunistic
spectrum access. Unlike spectrum underlay, SUs in spectrum overlay
will only use the licensed spectrum when PUs are not transmitting, so
there is no interference temperature limit imposed on SUs’
transmissions. Instead, SUs need to sense the licensed frequency band
and detect the spectrum white space in order to avoid harmful
interference to PUs.

In the rest of this chapter, we focus on spectrum overlay technique,

where SUs need to precisely identify the spectrum hole to perform
secondary transmission, and the term “cognitive radio” refers to overlay
cognitive radio.

Most existing work on overlay cognitive radio networks (CRNs)

employs the so-called “Listen-Before-Talk” (LBT) protocol by Half-
Duplex (HD) radio as illustrated in Fig. 8.2 [265–268]. In the LBT protocol,
SUs sense the target channel periodically, and begin transmission once a
spectrum hole is detected. For this protocol, the design of sensing duration
and sensing interval are of great importance, since there is a trade-off
between accurate sensing, which requires a long sensing time and short
sensing interval, and high throughput [267, 269]. Besides the optimization
of sensing and transmission duration, some other solutions, such as the
adaptive sensing and transmission duration [270], have also been proposed
to achieve higher spectrum efficiency. Though the conventional HD-based
LBT protocol is proved to be effective, it actually dissipates the precious
resources by employing a time-division duplexing, and thus unavoidably
suffers from two major problems.
 Figure 8.2 Illustration of the Listen-Before-Talk protocol

The first is that the SUs have to sacrifice the transmitting time for
spectrum sensing, and even if the spectrum hole is long and continuous, the
data transmission need to be split into small discontinuous fragments. This
wastes time and uses the spectrum inefficiently. The second problem is that
during SUs’ transmissions, the SUs cannot detect changes in the PU’s
states. This means that, on one hand, when the PU reclaims the spectrum,
the SUs cannot back off or adjust their transmission promptly, which may
lead to harmful interference in the primary transmission. On the other hand,
when the PU quits the spectrum, the SUs may not be able to access it
immediately, since they are not sensing the spectrum at that specific time,
and there will be a fraction of spectrum hole that remains unused.

Nowadays, the rapid development of FD communications has revealed

the possibility of simultaneous transmission and reception on the same
frequency band in wireless devices. Motivated by the realization of FD
transmission techniques, we try to explore a new way to bypass the above
problems by introducing FD SUs into CR Networks (CRNs).

In the rest of this chapter, we first present a “Listen-And-Talk” (LAT)

protocol that allows SUs to perform spectrum sensing and data transmission
simultaneously [271, 272]. Apparently, the proposed FD CR system is
totally different from the traditional HD-based one in many aspects,
including:

Spectrum sensing: in FD CRNs, sensing is continuous, but the

received signal for sensing is interfered with by the Residual Self-
Interference (RSI), which degrades the Signal-to-Interference-plus-
 Noise Ratio (SINR) in sensing. While in HD CRNs, there exists no
RSI in the received signal for sensing, but the sensing process is
discontinuous and only takes a small fraction of each slot. This leads
to unreliable sensing performance due to the inefficient number of
samples to make decisions.
Data transmission: in traditional HD CRNs, the SUs can only utilize
the remaining part of each slot after sensing for data transmission. By
contrast, in FD CRNs, SUs can continuously transmit as long as PUs
are absent. However, in FD CRNs, the data transmission affects the
sensing process, and thus there exists a constraint on transmission
power available to achieve acceptable sensing performance.

As shown above, the FD technology enables another dimension of the

network resources to be explored for increasing the capacity of CRNs. This
thus requires a new design of signal processing techniques, resource
allocation algorithms, and network protocols. For example, one of the major
challenges faced by FD-CR is how to optimize the transmission power for
the FD source node to maximize the system throughput.

Based on the LAT protocol, we investigate the collaborative spectrum

sensing and dynamic spectrum access scenarios in the FD CRNs. Also,
some key challenges in the implementation of FD CRNs are listed at the
end of the chapter.

8.2 Listen-And-Talk Protocol

This section elaborates the LAT protocol, the subject of which underlies the
remaining sections of this chapter. We start with a simple model, and
present the key parameter design. Some unique features brought by the FD
technique are investigated along with the performance analysis.

8.2.1 System Model

Consider a CR system consisting of one PU and one SU pair, as shown in
Fig. 8.3, in which SU is the secondary transmitter and SU is the receiver.
Each SU is equipped with two antennas, Ant and Ant . Ant is used for
spectrum sensing and Ant is used for secondary communication. The
transmitter SU uses both Ant and Ant for simultaneous spectrum sensing
and secondary transmission with the help of FD techniques, while the
receiver SU uses only Ant to receive signals from SU .1

Figure 8.3 System model of the LAT protocol ([272] Reproduced with permission from the IEEE.)

The spectrum band occupancy by the PU is modeled as an alternating

busy/idle random process where the PU can access the spectrum at any
time. We assume that the probabilities of the PU’s arrival and departure
remain the same across the time, and the holding time of either state is
distributed as the exponential distribution [268]. We denote the variables of
the idle period and busy period of the PU as and , respectively. And let
and represent the average idle and busy duration.

For SUs, on the other hand, only the idle period of the spectrum band is
available for use. To detect the spectrum holes and avoid collision with the
PU, SU needs to sample the spectrum at sampling frequency and make
decisions of whether the PU is present after every samples. This makes
the secondary traffic time-slotted, with slot length .
 Considering the common case that can be sufficiently high and the
state of the PU changes sufficiently slowly, we assume that and
is a sufficiently large integer. If we divide the PU traffic into slots in
accordance with the SU’s sensing process, the probabilities that PU arrives
and leaves the channel in a stochastic slot can be written as
and , respectively, where and are assumed
to be large integers. Note that when and are sufficiently large, we have
and .

With the help of FD technique, SU can detect the PU’s presence when
it is transmitting a signal to SU . However, as shown by the dotted arrow in
the system model in Fig. 8.3, the challenge of using FD technique is that the
transmit signal at Ant is received by Ant , which causes self-interference at
Ant . Note that, for Ant , the received signal is affected by the state of the
transmit antenna (Ant ): when Ant is silent, the received signal at Ant is
free of self-interference, and the spectrum sensing is the same as the
conventional half-duplex sensing. Thus, we consider the circumstances
when SU is transmitting or silent separately.

When SU is silent, the received signal at Ant is the combination of

potential PU’s signal and noise. The cases when the PU is busy or idle are
referred to as hypothesis and , respectively. The received signal at
Ant under each hypothesis can be written as

(8.1)

where denotes the signal of the PU, is the channel from the PU to Ant
of SU , and denotes the complex-valued Gaussian noise.
Without loss of generality, in this book, we assume that is PSK
modulated with variance , and is a Rayleigh channel with zero mean
and variance . Note that other modulation modes of the PU’s signal and
channel conditions, such as Gaussian channels, Rician channels, or path-
loss models for real conditions, may lead to different distributions of the
received signal for sensing, but they do not affect the main conclusion in the
remainder of this book.

When SU is transmitting to SU , RSI is introduced to the received

signal at Ant . Thus, the received signal varies from (8.1), and can be
written as

(8.2)

where and are the hypotheses under which the SU is transmitting

and the PU is busy or idle, respectively. Variable w in (8.2) denotes the RSI
at Ant , which can be modeled as the Rayleigh distribution with zero mean
and variance [103, 273], where denotes the secondary transmit
power at Ant and represents the degree of self-interference suppression,
which is defined as

The parameter is commonly expressed in dBs, and indicates how well

the self-interference can be suppressed.

Spectrum sensing refers to the hypothesis test in either (8.1) or (8.2).

Given that SU is aware of its own state (silent or transmitting), it can
automatically choose one pair of the hypotheses to test, i.e., when SU is
silent, it judges the PU’s state by testing ; and when SU is
transmitting, it chooses to test. If or is verified, the
spectrum is judged idle; otherwise, the spectrum is judged occupied by the
PU.
8.2.2 Protocol Description

Fig. 8.4 shows the sensing and spectrum access procedure of the LAT
protocol. SU performs sensing and transmission simultaneously by using
the FD technique: Ant senses the spectrum continuously while Ant
transmits data when a spectrum hole is detected. Specifically, SU keeps
sensing the spectrum with Ant with sampling frequency , which is shown
in the line with down arrows. At the end of each slot with duration T, SU
combines all samples in the slot and makes the decision of the PU’s
presence. The decisions are represented by the small circles, in which the
higher ones denote that the PU is judged active, while the lower ones
denote otherwise. The activity of SU is instructed by the sensing decisions,
i.e., SU can access the spectrum in the following slot when the PU is
judged absent, and it needs to back off otherwise.

Figure 8.4 The LAT protocol ([272] Reproduced with permission from the IEEE.)
 With the LAT protocol, ideally, the spectrum hole can be fully utilized
by SUs without any interference in the primary network due to the
following characteristics of the LAT protocol:

Secondary transmitters no longer need to sacrifice a fraction of each

time slot to sense the spectrum silently. Instead, they can sense
whether the PU is present continuously while transmitting.
Secondary transmitters can respond to the PU’s arrival and departure
promptly since there is no “blind duration” as in the conventional
half-duplex protocols, in which SUs either transmit or keep silent
without sensing the occupancy of the spectrum.

However, energy detection requires a minimum time duration, i.e., a

minimum slot length to make reliable decisions. Thus, SUs may not be able
to detect the PU’s state change immediately, which leads to potential
collision and spectrum wastage when the PU changes its state. Also, due to
the sensing noise and the RSI, sensing errors may occur. There are two
types of sensing errors, namely false alarm and missed detection. The
former refers to the case when SUs judge that the spectrum is occupied by a
PU when it is actually not, while the latter means that SUs fail to detect the
presence of the PU’s signal.

From the discussion above, it can be seen that there exist the following
four states of spectrum utilization:

State : the spectrum is occupied only by the PU, and SU is silent;

State : the PU is absent, and SU utilizes the spectrum;
State : the PU and SU both transmit, and a collision happens;
State : neither the PU nor SU is active, and there remains a spectrum
hole.

Among these four states, State and State are the normal cases, and
State and State are referred to as collision and spectrum waste,
respectively. There are two reasons leading to State and State : (a) the PU
changes its state during a slot, and (b) sensing error, i.e., false alarm and
missed detection.

8.2.3 Key Parameter Design

The most important constraint of the secondary networks is that their

interference to the primary network must be under a certain level. In this
book, we consider this constraint as the collision ratio between SUs and the
PU, defined as

The sensing parameters are designed according to the constraint of . In

the rest of this subsection, sensing performance is evaluated, based on
which we provide the analytical design of the sensing thresholds.

Energy Detection
We adopt energy detection as the sensing scheme, in which the average
received power in a slot is used as the test statistic O:

(8.3)

where denotes the nth sample in a slot, and the expression for is
given in (8.1) and (8.2).

With a chosen threshold ϵ, the spectrum is judged occupied when ,

otherwise the spectrum is idle. Generally, the probabilities of false alarm
and missed detection can be defined as,
 (8.4)

where and are the hypothesises when the PU is idle and busy,
respectively.

Considering the difference of the received signal caused by RSI, we can

achieve better sensing performance by changing the threshold according to
SU ’s activity. Let and be the thresholds when SU is silent and busy,
respectively, and now we have two sets of probabilities of false alarm and
missed detection accordingly, denoted as and
, respectively.

Design Procedure
Fig. 8.5 shows a sketch of the design procedure of the sensing thresholds,
and the derivation procedure of the secondary throughput. Specifically, the
sensing error probabilities ( , ), and ( , ) are determined by
thresholds and , respectively. Combining these sensing error
probabilities with the PU’s arrival and departure probabilities in each slot,
the state transition probabilities among State , State , State , and State can
be derived. Then we can obtain the overall missed detection and false alarm
and , which is defined as the probability that the system stays on State
and State , respectively. Further, we consider the average time duration of
collisions and unused spectrum holes, and the collision ratio and spectrum
waste ratio can be given. Here, similar to the definition of collision ratio,
the spectrum waste ratio is defined as

Taking the collision ratio as the system constraint, the sensing thresholds
can be derived in reverse, as shown by the dashed arrows in Fig. 8.5.
 Figure 8.5 The schematic design procedure of the sensing thresholds of the LAT protocol

In the remainder of this subsection, we first calculate the sensing error

probabilities ( , ), and ( , ), then show the state transition among the
four states of the system, based on which we calculate the collision ratio as
a function of sensing thresholds and PU’s state change probabilities. Then,
the sensing thresholds can be obtained in reverse from the constraint of
collision ratio.

Sensing Error Probabilities

With the statistical information of the received signal in (8.1) and (8.2), the
statistical properties of O under each hypothesis can be derived. We
consider the following two types of time slot:

Slots in which the PU changes its state: if the PU arrives in a certain

slot, the received signal power is likely to increase in the latter
fraction of the slot, and the average signal power (O) is likely to be
 higher than the previous slots when the PU is absent. Then the
probability of correct detection is higher than , with i denotes the
current activity of SU , i.e., means that SU is silent, while
means that SU is active. Similarly, if the PU leaves in a slot, the
probability of correct detection is higher than . Note that these
slots are rare in the whole traffic; we only consider the lower limits
of correct detection in these slots, i.e., we set without further
derivation the probabilities of correct detection to be and when
the PU arrives or leaves, respectively.
Slots in which the PU remains either present or absent: in these slots,
the received signal in the same slot is i.i.d., and as we assumed
in Section II-A, the number of samples is sufficiently large.
According to Central Limit Theorem (CLT), the PDF of O can be
approximated by a Gaussian distribution .

PROPOSITION 1: The statistical properties and the description under each

hypothesis are given in Table 8.1, in which denotes the Signal-to-
Noise Ratio (SNR) in sensing, and is the Interference-to-Noise
Ratio (INR).

Table 8.1 Properties of PDFs of LAT protocol

Hypothesis PU SU
Idle Silent

Busy Silent

Idle Active

Busy Active

Proof 1: We first provide the general properties of the test statistics. Given
that each in (8.3) is i.i.d., the mean and the variance of M can be
calculated as
Further, if the received signal y is complex-valued Gaussian with mean zero
and variance , we have

and

(8.5)

Then we consider the concrete form of the received signal under each
hypothesis. In the LAT protocol, given the PU signal, RSI, and i.i.d. noise,
the received signal y is complex-valued Gaussian with zero mean. The
variance of y under the four hypothesises are as follow:

(8.6)

By substituting them into (8.5), we can obtain the results in Table 8.1.

Based on Table 8.1, the sensing error probabilities can be derived.

When SU is silent and the test threshold is , the probability of

missed detection ( ) and the probability of false alarm ( ) can be
written as
 (8.7)
and

(8.8)
respectively, where is the complementary distribution function of the
standard Gaussian distribution.

Similarly, when SU is transmitting with the threshold , the missed

detection probability ( ) and the false alarm probability ( ) are,
respectively,

(8.9)
and

(8.10)

State Transition and Overall Collision Probability

The collision ratio is related to not only the sensing error probabilities, but
also the PU’s state. Thus, a joint consideration of all the four states of the
system is needed. Since the slots in which the PU changes its presence are
considered together with the other slots, the state transition of the system
can be simplified to a Discrete-Time Markov Chain (DTMC), in which the
system can be viewed as totally time-slotted with T as the slot length.
Fig. 8.6 shows the state transition diagram, where we denote State as
( ) for simplicity.
 Figure 8.6 State transition of the system ([271] Reproduced with permission from the IEEE.)

PROPOSITION 2: The probability that the system stays in the collision state
is

(8.11)

where , , , and .

Proof 2: The probability of the system staying in each state

can be calculated considering the steady-state distribution
of the Markov chain:

(8.12)

where is the vector of steady probabilities, and is the

state transition matrix abstracted from Fig. 8.6, which can be written as
 (8.13)

Combining the constraint that , we have

(8.14)

where , , , and .

To have a check on the result, we consider the probability that the PU is

busy and idle as and
, which are consistent with the results when
we consider the PU’s traffic only.

Collision Ratio
The collision of the PU and SU occurs in the following two kinds of
circumstances: (a) When the PU keeps occupying the spectrum and SU
fails to detect the presence of the PU’s signal in the previous slot, which is
depicted in Fig. 8.6 as with the probability of . The collision length is
T. (b) The certain slots in which PU arrives. SU is very likely to be
transmitting in these slots, since the PU is likely to be absent in the previous
ones. The occurrence probability of this circumstance is equal to the PU’s
arrival rate .
PROPOSITION 3:
The average collision length under circumstance (b), where
the PU changes state can be approximated by , when is large enough.

Proof 3: The average collision length in this case can be calculated as

(8.15)

where the approximation is valid when is large enough.

It is unavoidable in the LAT protocol that when the PU arrives, a short

head of the signal, with the length of a SU’s slot approximately, collides
with the SU’s signal. Combine the two circumstances, and the overall
collision rate can be given by

(8.16)

Design of Sensing Thresholds

For the parameter design, we have a maximum allowable as the system
constraint, and all the parameters of the sensing process should be adjusted
according to . Note that Δ and r are only related to the PU’s traffic, and
and are closely related via thresholds and ,
respectively. Thus, we actually have two independent variables of the
secondary network to design to meet the constraint of .
 We choose and as the independent variables. With (8.16) as the
only constraint, there are infinite choices of ( ) pair. For simplicity, we
set , i.e., to reduce the degree of freedom, and the
constraint can be simplified as

(8.17)

where Δ, r, μ, and ν are only relevant to the PU traffic, and

can be derived from via test thresholds and .

In the rest of this part, we calculate from the constraint of , from

which the sensing thresholds and can be obtained.

Combining (8.7) and (8.8), (8.9) and (8.10), we can obtain and as
functions of as, respectively,

(8.18)

(8.19)

From (8.18) and (8.19), we can find a rise of the false alarm probability
when the RSI exists. This result indicates that when interference increases,
the sensing performance gets worse.

With (8.18) and (8.19), ξ can be expressed as

. With given parameters of the PU’s traffic
and the slot length, can be solved from (8.17). Since the analytical
expression of is complicated, we only give some typical numerical
solution in Fig. 8.7, where the sensing SNR dB, INR dB,
number of samples , and r is set to be 6 to meet the real case that
the typical spectrum occupancy is less than [259].

Figure 8.7 Numerical solution of ; dB, = 5 dB, , and

It is shown in Fig. 8.7 that when μ goes down, becomes a fine

approximation of . With the large- assumption, we regard μ as
sufficiently small, and is determined by

(8.20)

This indicates that with the same constraint and parameters of the PU’s
traffic, the required gets squeezed when the SU’s slot length T increases.
 With , the thresholds and can be obtained from (8.7)
and (8.9), respectively:

(8.21)

(8.22)

A lift of sensing threshold due to the RSI ( ) can be found from (8.21) and
(8.22), which is in accordance with the previous analysis.

8.2.4 Performance Analysis: Power–Throughput Trade-Off

In this subsection, we first evaluate the sensing performance of the LAT by

the probabilities of spectrum waste ratio under constraint of collision ratio.
Then, with the closed-form analytical secondary throughput, a trade-off
between the secondary transmit power and throughput is elaborated
theoretically.

Spectrum Utilization Efficiency and Secondary Throughput

Similar to the analysis of the collision ratio, we combine the following two
kinds of time slots to derive the spectrum waste ratio: (a) the slots when the
spectrum remains idle; and (b) the slots of the PU’s departure. There exists
a waste of spectrum holes in (a) when the system is in the state in
Fig. 8.6, and the probability is given by in (8.14). Every time the SU fails
to find the hole, the waste length is T. In (b), the average waste length can
be derived from the PU’s traffic with the similar method in (8.15), and it
also yields of the average waste length. The probability of the PU’s
departure is , which is same as its arrival rate. The ratio of wasted
spectrum hole is then given by
 (8.23)

SU ’s throughput can be measured with the waste ratio. The achievable

sum rate of SU under perfect sensing is given as

(8.24)

where represents the SNR in transmission, with denotes the

pass loss of the transmit channel from SU to SU . SU ’s throughput is then
written as

(8.25)

Power–Throughput Trade-Off
In the expression of SU ’s throughput in (8.25), there are two factors: R and
. On the one hand, R is positively proportional to SU ’s transmit
power . On the other hand, however, the following proposition holds.

PROPOSITION 4: The spectrum waste ratio increases with the secondary

transmit power .

Proof 4: Firstly, the INR increases with the transmit power and in turn
lifts , which can be seen from (8.19).

Then, we can rewrite (8.23) as

 (8.26)

When increases, the third term decreases and increases

monotonically. Then the increase of SU ’s transmit power results in greater
waste of the vacant spectrum.

Thus, there may exist a power-throughput trade-off in this protocol:

when the secondary transmit power is low, the RSI is negligible, the
spectrum is used more fully with small , yet the ceiling throughput is
limited by R; when the transmit power increases, the sensing performance
get deteriorated, while at the same time, SU can transmit more data in a
single slot.

The analysis above indicates the existence of a mediate secondary

transmit power to achieve both high spectrum utilization efficiency in time
domain and high secondary throughput. To obtain this mediate value of
transmit power, we differentiate the expression of the throughput to find the
local optimal points of the secondary transmit power , which satisfies

(8.27)

PROPOSITION 5: The local optimal power2 satisfies

 (8.28)

where the notations are as follow:

Proof 5: The optimal power satisfies

(8.29)

The differentiation of the secondary throughput can be derived as

(8.30)

where , i.e., . With

, we have
 (8.31)

and therefore,

(8.32)

When μ is sufficiently small, the notations can be simplified as

(8.33)

and (8.32) becomes

(8.34)

i.e.,
 (8.35)

In (8.28), with as the only unknown variable, it can be calculated

numerically.

To obtain better comprehension about the properties of the local optimal

transmit power, we consider the case when μ is sufficiently small, and
(8.28) can be simplified as

(8.36)

The left side of (8.36) is a convex curve of with a single maximum.

When goes to zero or infinity, the value of the left side goes to zero. The
value of the right side changes from to . We can
roughly say that when the maximum of the left side is larger than either
or , there are two solutions to (8.36). When the
maximum of the left side is smaller than the minimum of the right, on the
other hand, no solution exists.

When there is no solution, the curves of transmit power on the left

and right sides never meet. Since the right side of (8.36) is always far
above zero and the left can go to zero when the transmit power is
extremely high or low, we can safely say that the left side is always
smaller than the right, i.e.,

(8.37)

Substituting the inequation to (8.30), we have , which indicates

that the secondary throughput would increase with the transmit power
monotonously.
 When the solutions of (8.36) exist, we discuss the sign of
piecewise. When the power is low or high enough, the left side is
small, while the right remains considerable. The solid red
curve (maximum of the left side of (8.36)) is below the dash-dotted
blue one (value of the right side of (8.36)), and . When the
power is between the two solutions, we have . Thus, at the
smaller solution, , and this is the local optimal transmit
power to achieve local maximum throughput. Similarly, the larger
solution denotes the local minimum of the throughput.

The maximum of the left side and the corresponding value of the right
side is shown in Fig. 8.8, in which the parameters are listed in Table 8.2. It
is shown that when is smaller than 0.8, the maximum of the left is larger
than the corresponding value of the right, and (8.36) will have solutions and
power-throughput is likely to exist. When is greater than 0.85, there may
be no trade-off between the transmit power and secondary throughput.
 Figure 8.8 Numerical verification of the existence of the power-throughput tradeoff

Table 8.2 Simulation Parameters

Parameters Value
The number of samples a whole time slot ( ) 300
The probabilities of the PUs arrival in the next slot (μ) 1/500
The probabilities of the PUs departure in the next slot (ν) 6/500
Normalized transmit power ( ) dB
SNR in sensing process ( ) dB
RSI factor (χ) 0.4
Probability of collision ( ) 0.1

8.2.5 Numerical Evaluation

In this section, simulation results are presented to evaluate the performance

of the proposed LAT protocol. Table 8.2 lists some default parameters in the
simulation.

Power–Throughput Relationship of the LAT Protocol

As is shown in Fig. 8.9, we consider the throughput performance of the
LAT protocol in terms of secondary transmit power. The solid and dotted
lines represent the analytical performance of the LAT protocol, and the
asterisks (*) denote the analytical local optimal transmit power. The small
circles are the simulated results, which match the analytical performance
well. The thin solid line depicts the ideal case with perfect RSI cancellation.
Without RSI, the sensing performance is no longer affected by transmit
power, and the throughput always goes up with the power. This line is also
the upperbound of the LAT performance. The thick dash-dotted, dotted and
dashed lines in the middle are the typical cases, in which we can clearly
observe the power-throughput trade-off and identify the local optimal
power, which is calculated from (8.36). With the decrease of RSI ( from
0.1 to 0.01 to 0.001), the local optimal transmit power increases, and the
corresponding throughput goes to a higher level. This makes sense since the
smaller the RSI is, the better it approaches the ideal case, and the
deterioration cause by self-interference becomes dominant under a higher
power. According to Fig. 8.8, when is sufficiently large (0.85 in the
figure), there exists no power-throughput trade-off. We verify this result by
the thick solid line denoting the cases when . No local optimal
point can be found in this curve, and the numerical results show that the
differentiation is always positive.

Figure 8.9 Power–Throughput curve in terms of different RSI factor , with the probability of
the PU’s arrival , departure , the collision ratio , the sample
number of a slot is 300, sensing SNR dB ([272] Reproduced with permission from
the IEEE.)

One noticeable feature of Fig. 8.9 is that when self-interference exists,

all curves approach the thin dotted line when the power
goes up. This line indicates the case that the spectrum waste is 0.5. When
the transmit power is too large, severe self-interference largely degrades the
performance of spectrum sensing, and the false alarm probability becomes
unbearably high. It is likely that whenever SU begins transmission, the
spectrum sensing result falsely indicates that the PU has arrived due to false
alarm, and SU stops transmission in the next slot. Once SU becomes
silent, it can clearly detect the PU’s absence, and begins transmission in the
next slot again. And the state of SU changes every slot even when the PU
does not arrive at all. In this case, the utility efficiency of the spectrum hole
is approximately 0.5, which is clearly shown in Fig. 8.9. Also, it can be
seen that the larger is, the earlier the sensing becomes unbearable and the
throughput approaches the orange line.

Sensing Performance
In this subsection, we use the receiver operating characteristic
curves (ROCs) to present the sensing performance. In Fig. 8.10, with the
sensing SNR fixed on dB, we have the relation between the collision
ratio and spectrum waste ratio. The thick lines denote the cases when the
PU changes its state very slowly, while the fine lines represent the cases
when the PU changes comparatively quickly. In Fig. 8.10, smaller areas
under a curve denote better sensing performance. It can be seen that the
thick lines are lower than the corresponding fine lines, which indicates that
when the PU changes its state slowly, the spectrum holes can be utilized
with higher efficiency. This is because the spectrum waste due to the state
change, i.e., re-access and departure of the PU happens less frequently.
Also, comparing the solid and dotted lines with the same μ, it can be found
that smaller RSI leads to better sensing performance, and the impact of the
RSI can be significant. It is noteworthy that the ratio of spectrum waste of
the LAT can be quite close to zero if the self-interference can be effectively
suppressed, and the constraint of collision ratio is not too strict. However,
recall the conventional listen-before-talk: the spectrum waste ratio can
theoretically never be suppressed lower than the sensing time ratio in a slot.
 Figure 8.10 ROCs in sensing. In this figure, the probability of the PU’s arrival and
1/100, departure , the sample number of a slot is 300, normalized secondary transmit
power dB, sensing SNR dB, and the RSI factor varies between 0.1
and 0.01.

8.3 Cooperative Spectrum Sensing

In this section, we extend the simple model shown in Fig. 8.3 to the
scenario with multiple cooperative SUs to further improve sensing
performance.

As shown in Fig. 8.11, we consider a CRN consisting of one PU, one

Fusion Center (FC), and M SUs denoted by SU , SU ,..., SU , each of
which equips two antennas Ant and Ant for . In each time slot
with duration T, the SUs sense the spectrum with local energy detectors and
then report their local sensing results to the FC. The FC then makes a final
decision on the presence of the PU by combining the local reports and
decides whether the PU is present in the current time slot.
 Figure 8.11 System model of the cooperative sensing scenario in which M SUs sense the same
spectrum channel cooperatively and report their local sensing results to a common Fusion
Center (FC). The FC makes the final judgement and decides which SU should transmit in the next
time slot.

8.3.1 Cooperation Scheme

Local Spectrum Sensing
For each SU in the network, the LAT protocol is adopted, i.e., Ant is
utilized to receive signal from the environment to sense the spectrum, and
Ant is utilized to transmit its own data. In each time slot, only one SU is
allowed to transmit if the PU is absent. Without loss of generality, we
assume only SU is allowed to transmit, and all the other SUs keep silent.

For the secondary transmitter SU itself, the spectrum sensing is almost

the same as introduced in the previous subsections. The received signal at
Ant is written as

(8.38)
where denotes the PU’s signal, is the Rayleigh channel
from the PU to Ant represents the complex-valued
Gaussian noise, and is the RSI term, with and
denote the power of SU ’s transmit power and the suppression level of self-
interference, respectively.

For any other SU ( ), the signal from SU is treated as interference.

We have

(8.39)

where is the SU ’s signal, and and denote

the channels from the PU and SU to SU , respectively.

Data Report and Data Fusion

We assume the SUs report their one-bit hard decisions to the FC, and no
error exists in the reporting process. In the FC, the OR fusion rule is
adopted,3 i.e., the FC decides the presence of the PU if at least one report
declares that the PU is detected, and vice versa. The missed detection and
false alarm probabilities of the cooperative decision is then given by

(8.40)
where and ( ) denote SU ’s local missed detection and false
alarm probabilities when SU is silent/active.

8.3.2 Analysis of FD Cooperative Spectrum Sensing

Local Sensing Error
Similar to the analysis in the previous section, the PDF of the average
received power in each slot can be approximated by a
Gaussian distribution according to the CLT. For simplicity, we assume that
all channels are independent, and both the PU’s and SU ’s signal is PSK
modulated, and the power of the PU’s signal is . Then the interference
term can be treated as random complex Gaussian distributed,
and the PDF of ( ) and can be written in similar forms. The
description and statistical properties are given in Table 8.3, where
denote the SNR from the PU to SU and the
interference-to-noise ratio (INR) caused by SU , respectively. Further, let
be the INR at SU , and the distribution of can be written
uniformly with as shown in Table 8.3, with which the local
probabilities of missed detection and false alarm at SU can be derived as,
respectively,

(8.41)
Table 8.3 Statistical Properties of ([271] Reproduced with permission from the IEEE.)
Hypothesis PU SU
Idle Silent
 Busy Silent

Idle Active

Busy Active

Collision Ratio and Spectrum Waste Ratio

Similar to the derivation in the previous section, we need to consider the
state transition among the four states of the system. For simplicity, we
assume that the probabilities that the PU changes its state within a slot
are sufficiently small that we omit it from our consideration of state
transition, and the collision ratio can be simplified as the conditional
probability that the system stays at given that the PU is occupying the
spectrum, i.e.,

(8.42)

Similarly, the spectrum waste ratio is given as

(8.43)

Considering the steady state probability, the collision ratio and spectrum
waste ratio can be derived as, respectively,

(8.44)

Local Sensing Thresholds

Substituting (8.41) to (8.40), and then to (8.44), we can obtain the collision
ratio and spectrum waste ratio. With the collision ratio as the system
constraint on the secondary network, the constraints of and can be
given by (8.44), and then the constraints of are given by (8.40). For
simplicity, we set all missed detection probabilities to be the same
without further optimization. So with a fixed , we have

(8.45)

The corresponding thresholds can be derived from (8.41) as

We can see a rise in the thresholds when interference is introduced to the

system. The local probabilities of false alarm can be derived accordingly:

(8.46)

Substituting (8.46) to (8.40) and (8.44), the spectrum waste ratio can be
obtained.

Now consider a comparison between the performance of CSS in the

LAT protocol and the non-cooperative sensing performance which could be
obtained from (8.41) and (8.44). With cooperation, much larger local
missed detection probabilities are allowed at each cooperating SU as is
shown in (8.45), and the corresponding local false alarm probabilities can
be reduced sharply. Suppose we have 10 cooperative SUs and the system
is set to be 0.01. Use (8.45) and we have , which is quite large,
and could be depressed to a very low level. Also, the interference
between two SUs can be much smaller than self-interference with careful
selection of cooperative SUs, so the assistant sensing results can be more
reliable than the result at the transmitting SU, and performance of CSS is
quite likely to be much better than non-cooperative sensing in the LAT
CRNs.

8.4 Dynamic Spectrum Access

In this section we consider the scenarios where multiple FD SUs contend

for the same licensed band. In these scenarios, how to design proper
Dynamic Spectrum Access (DSA) strategies to fully explore the advantage
of FD techniques becomes key problem. Regarding the network
architecture, there are mainly two kinds of DSA strategies, namely
infrastructureless and infrastructure-based strategies. In an infrastructureless
DSA network, no entity is in charge of the whole secondary network, and
each SU needs to sense and decide whether to access the spectrum
independently. For this scenario, a design of a new access and back-off
mechanism is presented in 8.4.1. In an infrastructure-based DSA network,
there exists a central control that allocates time and spectrum resources to
SUs. For this case, we discuss the power and spectrum allocation of a FD
cognitive AP system in 8.4.2.

8.4.1 Infrastructureless Spectrum Access

In many scenarios, such as in ad hoc CRNs where the SUs compete for
several PU channels, deploying a central controller is not always possible.
Therefore, distributed DSA mechanisms will be required, by which each
SU has to independently gather, exchange, and process the information of
the wireless environment. The commonly used CSMA/CA in distributed
DSA with HD users can effectively reduce collision probability [268], but
some problems still exist: (1) collision among the SUs can never be
detected if the SUs are synchronized, such that the secondary transmission
may fail to a large scale, and (2) SUs cannot abort transmission when
collision happens, which leads to long collision duration. With FD-CR, the
SUs can not only detect the presence of PUs, but also detect collision with
other SUs during transmission, such that the collision duration is reduced
significantly. But the RSI may degrade the collision detection accuracy,
which cannot be ignored. In the rest of this subsection, we present a feasible
distributed DSA protocol for FD-CRNs, and show its effectiveness of
achieving a high spectrum utilization ratio when the number of users goes
up.

System Model
Consider a CRN consisting of one PU with a licensed channel and M FD-
enabled SUs, where each SU can sense the channel and transmit
simultaneously in an independent manner, and they are allowed to access
the spectrum only when the PU is absent. The traffic model of the PU is the
same as the previous sections, and denotes the percentage that the
spectrum is free of the PU signal.

Each SU adopt LAT protocol for local spectrum sensing. The traffic of
SUs is slotted with duration T. At the end of each slot, each SU combines
the detected signal in the slot and makes a decision about whether or not to
transmit in the next slot. Since the RSI exists at the transmitting SUs, the
sensing of these SUs cannot be assumed perfect, which is different from the
common perfect-sensing assumption in conventional half-duplex scenarios.
For simplicity, we omit the noise in sensing. Thus, the sensing of a silent
SU becomes perfect, while for a transmitting SU, the sensing is imperfect
with false alarm probability and missed detection probability .
Furthermore, consider that if several (more than two) SUs collide in a slot,
the received signal in sensing would be much larger than the RSI: we
assume that the missed detection probability only applies to the case when
two SUs collide or one SU collides with the PU.

FD-DSA Protocol
In this part, we present a feasible distributed DSA mechanism for FD SUs
that can effectively eliminate collision between SUs and the PU as well as
collision among SUs.

Sensing with FD. With FD techniques, SUs can keep sensing during
transmission. The traffic of SUs is slotted with duration T, given that the
SUs cannot make any decision with a duration shorter than T. As shown in
Fig. 8.12, at the end of each slot, each SU combines the detected signal of
the slot and make a decision about whether to transmit or not in the next
slot. On the one hand, if SUs detect the PU’s signal or other SUs’ signals,
they keep silent and only sense the spectrum in the next slot. On the other
hand, if the spectrum is sensed idle, SUs do not access the spectrum
immediately. Instead, each of them randomly chooses one duration from the
set of Contention-Free Periods (CFP), and waits for the chosen duration
before they transmit. Note that all SUs keep sensing when they wait to
transmit, and once the PU’s signal or other SU’s signal is detected, the
waiting period suspends. The waiting period resumes when the spectrum is
sensed idle again.

Figure 8.12 DSA protocol for FD-CRNs, where there are four SUs, and the contention window
length is 3

Contention window. For simplicity, the CFP is also slotted with

duration T, and the contention window length is randomly chosen from
 , i.e., the size of the CFP is W. At the
beginning of a slot, if the spectrum is sensed idle in the previous slot, and
the countdown of the waiting time has reached zero, the transmission
begins. The transmission of any SU will not cease until other users are
detected utilizing the band. Then, if another spectrum hole is detected, the
SUs that have just performed transmission reselect a contention period
randomly and start countdown, while other SUs continue the uncompleted
countdown without another selection. This mechanism guarantees that
every SU can count to zero within slots and begins transmission.4

Performance Analysis
In this part, we analyse the performance of the proposed DSA protocol for
FD-CRNs.

As mentioned above, we omit noise in the analysis, and only the

transmitting SUs’ sensing is imperfect due to the RSI. We assume that the
false alarm probability of a SU is , and the missed detection probability is
. Furthermore, when several (more than two) SUs collide in the same
slot, the received signal in sensing would be much larger than the RSI, and
we assume that the collision can be detected with probability 1, i.e., the
missed detection probability only applies to the case when two SUs collide
or one SU collides with the PU.

Now consider a SU’s collision with the primary network. The situation
of SUs colliding with PUs only happens at the beginning of the PU’s
arrival, when some SUs are transmitting. And only when one SU collides
with the PU, the collision length may be longer than one slot. It can be seen
that the average collision length with the PU is slightly longer than one
secondary slot. With the assumption that the PU changes sufficiently slowly
compared with the secondary slot length, this collision length is quite
negligible.

The spectrum utilization efficiency is evaluated by the spectrum

utilization ratio ρ, defined by
 (8.47)

where is the average continuous transmit length, and is the

average spectrum waste length before a successful transmission. Since the
average continuous transmit length and waste length is independent, we
calculate them separately.

Duration of the continuous transmission: If a SU begins transmission

without collision with other SUs or the PU network, it suspends its
transmission only when (1) the PU arrives, or (2) a false alarm
happens. The probability that the PU comes in a slot is , and
the false alarm probability is . Thus, the probability that the SU
continues its transmission in the next slot is

(8.48)

and the average continuous transmit length can be calculated as

(8.49)

Spectrum waste: The spectrum waste mainly consists of three parts:

(i) pure sensing – after the transmission of the PU or SUs, all the SUs
need to sense for one slot to detect the idleness of the spectrum; (ii)
contending – when the spectrum is sensed idle, and all SUs are
waiting to transmit; (iii) collision between SUs leads to inefficient
transmission. First we consider cases (i) and (ii), where there is no
collision. Then we deal with the collision case (iii).

The probability that a SU waits for slots and begins transmission

without collision with other SUs is given by
 (8.50)

The average waste length in this case can be calculated as

(8.51)

If a collision happens, there are three circumstances: (i) more than two
SUs collide and they will stop transmission in the next slot; (ii) two SUs
collide and they stop together during some slots; (iii) two SUs collide, one
of them stops first, and the other performs a successful transmission after
that. For case (i), the probability that the waiting time equals is

(8.52)

and the average waste length is

(8.53)

If two SUs begin transmission together, the occurrence probability of

case (ii) is

(8.54)
and the occurrence probability of case (iii) is . The probability
of waiting for is

(8.55)

The average waste length for case (ii) can be expressed similarly as in
case (i):

(8.56)

where is the average collision length when two SUs collide, which can
be calculated as .

For case (iii), the expectation of the waste length is

(8.57)

Thus, the expectation of the overall waste length can be calculated as

(8.58)

Substituting (8.51), (8.53), (8.56), and (8.57) into (8.58), we can calculate
the waste length expectation as
 (8.59)

where .

The channel usage ratio can be readily derived by substituting (8.49)

and (8.59) into (8.47).

Now we consider the design of the contention window. If the contention

window is too small, the probability of collision among SUs is large, and
there may be long recurrence of “wait – collide – sensing – wait,” and the
spectrum usage efficiency is low. On the other hand, if the contention
window is too large, the collision ratio is small; yet the average contending
time (before the smallest CFP reaches zero) increases proportionally to the
maximum window length, and the usage efficiency also gets degraded.
Therefore, there exists an optimal contention window size that minimizes
the average waste time before a successful transmission.

Since the form of (8.59) is complicated, and the optimal point is

difficult to calculate, we approximate the summation by the following
integral:

The original summation approaches the above expression as the contention

window size increases.

Then, (8.59) can be simplified as

(8.60)
where A, B, C, D, E, and F are positive parameters, which can be derived
from (8.59). By solving , the optimal contention window length
can be obtained.

Comparison with DSA in HD-CRNs

In conventional DSA in HD-CRNs, an SU performs sensing before
transmission, and the sensing is commonly assumed to be perfect [268]. In
the conventional system, collisions typically happen in the following two
cases: (1) the PU comes while SUs are transmitting, and this collision
cannot be detected until the next sensing period; and (2) when SUs are
synchronized, they cannot detect collisions among them selves since all
active SUs stop transmission and sense the channel at the same time, and
the collision will last for the whole transmission. Consider the maximum
utilization of the conventional DSA protocol, which is under the constraint
of the sensing and transmission duration, can be written as

(8.61)

Thus, if we set the average collision duration of the proposed FD-CRN-

based DSA and the conventional HD-CRN-based DSA as the same, i.e., the
collision duration is about one slot, the transmitting period cannot be longer
than two slots, which means that the utilization ratio of the conventional
protocol cannot reach 70%. However, in the proposed protocol, the
utilization ratio does not have this kind of hard constraint. We will show in
the simulation section that the ratio can approach 90% with the proposed
protocol.

Also, in the conventional protocol, if all SUs are synchronized and they
have the slot structure, when several SUs collide with each other, the
collision will never be detected by themselves, and the collision may last
for quite long time and severely deteriorate the performance. While in the
DSA for FD-CRNs, whenever several SUs collide, they will detect the
collision in one slot with a large probability and stop transmission in the
next to avoid longer ineffective transmission. In this aspect, the proposed
protocol largely outperforms the conventional one.

Results
In Fig. 8.13, we set the PU’s arrival rate as 0.001, the false alarm
probability as 0.4 to shorten the average continuous transmit length, and the
total simulation length is 1,000,000 slots. We show in Fig. 8.13 the results
of the average waiting length. The analytical results (the solid and dotted
lines) are based on the simplified expression of in (8.60), which
matches the simulation results when the contention window length W is
sufficiently long. Also, we can find that the optimal contention window
length calculated from the simplified represents the real case well,
shown by the vertical lines. Comparing the solid and dashed lines, when
other parameters remain the same, the scenario with more SUs is more
adapted to larger contention windows, while short contention windows suits
the case with fewer SUs. Another interesting point is the comparison
between the solid and dotted lines. When missed detection probability rises,
the average wait time decreases. The reason is that when two SUs collide
with each other, the probability that one SU stops before the other, i.e.,
, increases with the missed detection probability, and this leads to shorter
waste time before successful transmission. However, even though it is not
shown in the figure, the rise of missed detection probability inevitably
results in a longer collision time with the PU network, which is harmful to
the whole system.
 Figure 8.13 DSA in FD-CRNs: average waiting length vs. contention window length

Fig. 8.14 shows the spectrum utilization ratio of the proposed protocol
under different contention window length, in which the PU’s arrival rate is
fixed on 0.0001, and missed detection probability is 0.05. Fig. 8.14, as well
as Fig. 8.13, shows that the optimal contention window length increases
with the number of SUs, and it is typically several slots longer than the
number of SUs. Also, it is shown in Fig. 8.14 that the spectrum utilization
ratio is affected by the false alarm probability more than the SU’s number
or contention window length, i.e., when false alarm probability rises from
0.05 to 0.15, the utilization ratio drops from the dashed line to the solid line
significantly. This indicates the importance of improving sensing
performance of the LAT protocol in the physical layer.
 Figure 8.14 DSA in FD-CRNs: spectrum utilization ratio vs. contention window length

8.4.2 Infrastructure Spectrum Access

In this subsection, we consider a Cognitive Cellular Network (CCN) in

which a Secondary Base Station (SBS) is deployed to control the secondary
transmissions between itself and multiple SUs. The SBS, as a central
controller, is in charge of sensing the primary spectrum and deciding
frequency bands and power of each secondary transmission. The SUs are
fully controlled by the SBS and no direct transmissions are allowed
between each other. Here, we only consider the downlink transmissions
from the SBS to the SUs. In traditional CCNs using the LBT protocol, the
spectrum sensing and spectrum access exclusively compete for the same
radio resource. Thus, most studies have focused on scheduling the time and
order of sensing and transmission, so as to achieve the optimal trade-off
[274, 275]. However, with the help of FD technique, i.e., when the SBS has
strong Self-Interference Suppression (SIS) capability, the whole scenario
can be significantly changed.

System Model
Consider a network with one PU, one SBS and M SUs, as shown in
Fig. 8.15. The primary network is an OFDM system in which the PU
transmits on K orthogonal channels. On each channel, the probabilities that
the PU is present and absent are given by and ,
respectively. The secondary network consists of one SBS and M SUs. The
SBS, equipped with two antennas, is a full-duplex device with strong SIS
capability. The SBS can simultaneously sense the primary signal using Ant
and transmit to the SUs using Ant . Once the idle channels are detected, the
SBS decides the power allocation and spectrum access strategy for the
downlink transmissions of the M SUs. At antenna Ant , the received signal
on channel k is given by:

(8.62)

where is the received signal of the PU, is the RSI, and is the
noise signal. We denote by as the noise power and
as the received SNR of the PU. The SIS factor of the
SBS is . On channel k with transmitting power , the INR is denoted by
.

Figure 8.15 System model of a full-duplex cognitive cellular network

Using an energy detector, the missed detection and false alarm

probabilities of the SBS on channel k are given by
 (8.63)

and

(8.64)

where is the threshold on channel k and is the number of samples.

Let be a binary variable denoting the presence and absence of the PU

on channel k, where represents that the PU is present and
represents the absence. Let be a binary matrix denoting
the spectrum access strategy of the SBS, where represents that
channel k is used to transmit to SU m and represents the opposite.
We assume that each SU can occupy at most one channel and the total
transmit power of the SBS is P. Thus, in order to maximize the total
throughput of the SBS, the optimization problem is described as follows:

(8.65a)

(8.65b)

(8.65c)

(8.65d)

(8.65e)
where denotes channel k for SU m, (8.65b) requests an
upper bound for the missed detection probability to ensure the PU’s
outage probability constraint, (8.65c) is the total power constraint of the
SBS, (8.65d) ensures that each SU can occupy at most one channel, and
(8.65e) ensures that each channel is exclusively occupied by at most one
SU.

As we see in (8.65), the joint spectrum access and power allocation

problem is to find an optimal combination of channels, users and power, so
as to maximize the total throughput, and at the same time, satisfy all
constraints. We will see that this problem can be seen as a three-
dimensional matching problem in graph theory.

Joint Spectrum Access and Power Allocation

In this part, we reconsider problem (8.65) from the perspective of three-
dimensional matching and prove its NP-hardness to find the optimal
solution. Then, we show the influence of power allocation and present an
approximate solution by extending a well-known two-dimensional matching
algorithm.

Firstly, we show the equivalence of the problem (8.65) and the three-
dimensional matching problem to prove the NP-hardness of the considered
optimization problem.

DENITION 1 (Three-Dimensional Matching): For a set of

ordered triples where X, Y and Z are disjoint sets, each element has an
associated weight . A subset is a three-dimensional matching
if, for any two distinct elements, , we have
, and . We denote by Ψ the set of all three-
dimensional matchings, and the problem is to find the matching with the
maximal weight; that is, .

PROPOSITION 6: The optimization problem (8.65) is not easier than a 3-

dimensional matching problem.
Proof 6: Let denote the K channels of the PU’s spectrum,
and let denote the M SUs of the SBS. We assume that the
optimal power allocation of channel k is given by and let
denote the set of all optimal powers. Consider a three-
dimensional matching where , and is the
maximal secondary throughput of SU m on channel k with transmit power
, given by

(8.66)

Thus, the optimal solution and of (8.65) provides the optimal

solution of the three-dimensional matching, where if and
only if and , as shown in Fig. 8.16. Therefore, the optimization
problem (8.65) is no easier than a three-dimensional matching problem.

Figure 8.16 The three-dimensional matching for the considered optimization problem

The three-dimensional matching problem is known to be NP-complete

[276], which means that the joint spectrum access and power allocation
problem (8.65) must be NP-hard, and it cannot be solved optimally in
polynomial time.
 Approximate Algorithm: Although the three-dimensional matching is
NP-complete, there are efficient polynomial-time algorithms for finding a
two-dimensional matching, for example, the Kuhn–Munkres algorithm
[277] and the Hopcroft–Karp algorithm [278]. We now propose an
approximate algorithm based on the two-dimensional matching, as in
Fig. 8.17. The idea is consider the spectrum access as a two-dimensional
matching problem between the channels and the SUs. The total transmit
power P is equally divided into M discrete units with each unit ,
and the discrete units are iteratively allocated by the SBS. In each iteration,
one power unit is added to the channel that generates the largest marginal
throughput of the current SU, and then, the three-dimensional matching
method is performed to optimize the spectrum access strategy with the new
power allocation. The algorithm stops when all power units are allocated or
there is no channel–user pair that can increase its secondary throughput by
increasing the transmit power.

Figure 8.17 The proposed algorithm based on two-dimensional matching

DENITION 2 (Two-Dimensional Matching): For a set of pairs

where X and Y are disjoint sets, each pair has an associated
weight . A subset is a two-dimensional matching if for any
two distinct elements , we have and
. We denote by Φ as the set of all two-dimensional matchings, and
the problem is to find the maximum weighted matching, that is,
. Note that the two-dimensional matching
problem can be seen as a graph theory problem. Given a bipartite graph
where and , and the weights of all
edges , a matching is defined as a set of edges that do not
coordinate with each other, and the problem is to find the maximum
weighted matching in graph G.

Consider a complete bipartite graph with and

, where denotes the K channels, and
denotes the M SUs. For any edge , the weight
is given by the secondary throughput of user m on channel k, which is a
function of the transmit power and the sensing threshold . Note that in
(8.63) and (8.64), the missed detection probability is increasing with ,
while the false alarm probability is decreasing with . Thus, equation
decides the optimal sensing threshold , which is
given by

(8.67)

Thus, the weight of edge can be written as

(8.68)

Note that in (8.68), the multiplier , which represents the

probability that the spectrum hole is correctly detected, is decreasing with
the transmit power , while the multiplier , which
represents the channel capacity, is increasing with . Therefore, we can
expect the total throughput first increases and then decreases with , and
we denote by as the turning point. The value of can be
numerically calculated by solving .
 To solve the above two-dimensional matching problem, we first present
some basic concepts. A feasible labeling l is a mapping from each vertex
to , where . And an l-equal problem is a two-
dimensional matching problem where
and . Briefly, the l-
equal problem considers a subgraph of the original problem, and it aims
to find the matching with the largest size. If the optimal matching of the
l-equal problem covers all vertexes in G, for which we call a perfect
matching for G, we have

(8.69)

On the other hand, for any matching S of G, we have

(8.70)

Therefore, we have , that is, is the maximum weighted

matching of G. Now, the two-dimensional matching problem can be seen as
two separate problems. The first one is to solve the l-equal problem for a
feasible labeling l, and the second one is to find a feasible labeling l such
that the solution of the l-equal problem is a perfect matching for G.
These two subproblems can be solved by the Hungarian algorithm and
label amendment algorithm, respectively.

Hungarian Algorithm: The Hungarian algorithm is to find the largest

matching of a bipartite graph for an l-equal problem. For any initial
matching , the algorithm iteratively increases the size of S by
finding an augmenting path, which is defined as a path
where the edges belong alternatively to
and S, and the starting point and the end point are not
matched by S, i.e., such that or . Note that the edges
that do not belong to S are one more than the edges that belong to S, i.e.,
. Therefore, we can get a larger matching
by replacing the edges in with the edges in
. The algorithm stops when there is no augmenting path for
the current matching and the output is the largest matching . We formally
present the Hungarian algorithm in Table 8.4.

Table 8.4 Hungarian Algorithm

Input:

Output: the largest matching

1: Set as the initial matching, and set as the vertexes in X that are not matched by S;
2: If , stop and output ; If , take , set , and turn to the
next step;
3: Set as the neighbors of U. If , there is no
augmenting path for S that starts from node x. Then set and turn to step 2; If
, take and turn to the next step;
4: If , then set and , and turn to step 3; If
, then we have an augmenting path P with the starting point x and the end point y.
Set and and turn to step 2.

Label Amendment Algorithm: To ensure the existence of a perfect

matching, some vertexes and zero-weight edges are added to to
make it a complete bipartite graph where and
. For any initial labeling l, we use the Hungarian
algorithm to calculate the largest matching for the corresponding l-equal
problem. If is a perfect matching for G, then output as the
maximum weighted matching for the original two-dimensional problem.
Otherwise, the Hungarian algorithm outputs , and we have
. Let
denote the “extra” labeling for
these edges, and we amend the vertex labeling as follows:
 (8.71)

We can verify that is still a feasible labeling, and the edges

will be included in the -equal problem . Then,
we find the largest matching of the -equal problem. The algorithm stops
when it outputs a perfect matching for G, and we formally present the label
amendment algorithm in Table 8.5.

Table 8.5 Label Amendment Algorithm

Input: and

Output: the maximum weighted matching

1: Add vertexes and zero-weight edges to make a complete bipartite graph with
;
2: Set an initial labeling for all , and for all ;
3: Find the largest matching of the l-equal problem using the algorithm in Table 8.4;
4: If is a perfect matching for G, then output ; otherwise, set the vertex labeling as in
(8.71) and return to step 3.

Proposed Joint Spectrum Access and Power Allocation Algorithm: The

above methods can find the optimal spectrum access strategy for a given
power allocation. In our problem, we iteratively allocate the power units to
the channels, and in iteration t, we assume the power allocation is given by
and the corresponding optimal spectrum access strategy is given by
matching with the final vertex labeling . Thus, we have
, where

(8.72)

is the channel with the largest marginal rate, and for the
rest channels. And the initial feasible labeling of the next iteration is
given by

(8.73)

where

(8.74)

The algorithm stops when all power units are allocated, or at an iteration t
such that

(8.75)

where the channel-user pair is the matched in . We formally

present our algorithm in Table 8.6.

Table 8.6 Joint Spectrum Access and Power Allocation Algorithm

1: Set for all and the initial spectrum access strategy ;

2: while and (8.75) are not satisfied do
3: Choose channel as in (8.72), and add one power unit Δ to this channel;
4: Set the labeling l as in (8.73), and calculate the optimal spectrum access strategy using
the algorithm given in Table 8.5;
5: Set ;
6: end while
7: The final power allocation is and the final spectrum access is where if
and only if .

Results

In this part, we present the performance of the proposed matching

algorithm, compared with the random algorithm and the greedy algorithm.
In the random algorithm, the SBS randomly chooses an SU for each idle
channel, while in the greedy algorithm, the SBS sequentially chooses the
optimal SU for each idle channel. The parameters of our simulations are
given in Table 8.7.

Table 8.7 Parameters for Simulation

Parameters Value
The probability of the PU’s transmission 0.5
The missed detection constraint 0.1
The received SNR of the PU signal dB
The number of samples in a sensing slot 200
The ratio of the transmit power and noise P 100 dB ~ 120 dB
The ratio of the interference and the power dB ~ dB
The number of subcarriers K 20
The number of SUs M
The upper bound of iteration number I

In Fig. 8.18, we show the throughput of the SBS as a function of the

number of SUs M for each algorithm. Due to the multi-user diversity, the
secondary throughput of the SBS increases with the number of SUs.
Compared with the greedy algorithm and the random algorithm, our
proposed matching algorithm increases the throughput of the SBS by
and , respectively, when the number of SUs . When the
number of SUs increases, the probability that the channels have different
optimal SUs also increases. Therefore, we can expect the greedy algorithm
also achieves a high multi-user diversity, and its performance can approach
the proposed matching algorithm, as seen in Fig. 8.18 when .
 Figure 8.18 Throughput of the SBS as a function of the number of SUs with dB and
dB

In Fig. 8.19, we show the trade-off between the performance and the
computation complexity of the proposed algorithm. In our proposed
algorithm, the total power is equally divided into I power units, and the
number of iterations is bounded by the value of I. When I is increased, the
computation complexity increases, while at the same time, the size of a
power unit decreases, which makes it possible to get a power
allocation that is closer to the optimal solution. Thus, the performance of
our proposed algorithm increases with the number of iterations, as seen in
Fig. 8.19.
 Figure 8.19 Throughput of the SBS as a function of the power unit size Δ with
dB, and dB

It can be seen from the discussion of this part that when we introduce
the full-duplex technology in a cognitive cellular network, the SBS can
simultaneously sense the spectrum and transmit to the SUs. In such a
network, the joint spectrum access and power allocation problem has been
formulated as a three-dimensional matching problem, and we have
proposed an approximate solution based on a two-dimensional matching
algorithm. Both the analysis and simulations show that the secondary
throughput of the SBS does not monotonically increase with the total
transmit power, and there exists an upper bound when the transmit power is
sufficiently high. Also, the secondary throughput is highly influenced by the
number of SUs and the SIS capability. Compared with the greedy algorithm
and the random algorithm, our proposed algorithm can greatly increase the
secondary throughput of the SBS in various conditions.

8.5 Key Challenges

In fact, regardless of all the promising features that the FD CRNs hold,
there still exist some challenges in the design and implementation of them.
In this section, we briefly summarize the main research problems for FD-
CR communication systems as well as the possible solutions. Also, some of
the inspiring future research topics are listed at the end. Similar to the
traditional wireless systems, multi-dimensional resources on space, time,
frequency, and power need to be properly managed to optimize the overall
system performance. Especially, FD communication provides another
dimension of resource and its performance is also greatly affected by the
RSI.

8.5.1 Signal Processing Techniques

Spectrum Sensing
Non-cooperative narrowband sensing has been elaborated in Section 8.2.
However, the degradation of sensing performance becomes unbearable
when transmit power is high. Cooperative sensing, as analyzed in
Section 8.3, is one of the promising solutions to this problem. However,
when employing cooperative spectrum sensing in FD-CRNs, there is
interference from the transmitting SU to other SUs, which degrades the
local sensing performance of the cooperative SUs to some degree, and the
selection of the cooperative nodes may be different from the conventional
HD scenario, i.e., the closest nodes may not be proper choices for
cooperation due to the strong interference, and some further nodes may be
better.

Also, since CRNs will eventually be required to exploit spectrum

opportunities over a wide frequency range, a FD-enabled wideband sensing
scheme is needed. With the impact of RSI, the original sparsity, which is the
basis for the conventional wideband sensing scheme, would change, and the
whole sensing scheme may be different.

Multi-Antenna Techniques
If multiple antennas are equipped at the FD-CRs, beamforming and antenna
selection can be employed to further improve the secondary network
performance:
 Transmit beamforming: transmit beamforming is used to control the
directionality of transmission in order to provide a large antenna
array gain in the desired directions. For a FD cognitive MIMO
system, the transmit antenna set at each FD-CR node can perform
transmit beamforming to simultaneously transmit information and
reduce the interference to its own received sensing signals. The
design is to jointly optimize the sum rate of the system. If the FD
cognitive AP node, which serves a group of users, is equipped with
multiple antennas, it may be able to support multiple downlink
transmissions while maintaining reliable sensing performance by
using certain structure of antennas to minimize the RSI.
Antenna selection: for a FD-CR node, especially a node with
multiple antennas, each antenna can be configured to sense (receive)
or transmit the signals. This creates an important problem to
optimally select the antenna configurations optimize the system
performance [202]. In a FD cognitive MIMO system, the problem is
to choose one group of antennas to sense the spectrum, one group to
transmit and the rest to receive signals from another SU
simultaneously. Such a combinatorial problem becomes much
complicated as the number of antennas increases. Similarly, in a
general FD cognitive relay system, each relay can adaptively select
its sense and transmit and receive antennas based on the
instantaneous channel conditions to achieve reliable sensing as well
as maximum SINR in transmissions.

8.5.2 Coexistence of Multiple Systems

Spectrum sharing has been recognized as a key remedy for the spectrum
scarcity problem, especially after the successful deployment of WLAN and
WPAN devices on an unlicensed band (e.g., ISM band). However, severe
performance degradation has been observed when heterogeneous devices
share the same frequency band due to mutual interference rooted in the lack
of coordination. The Cooperative Busy Tone (CBT) algorithm allows a
separate node to schedule a busy tone concurrently with the desired
transmission, and thereby improving the visibility among difference sorts of
devices [279]. But preventing the busy signal from interfering the data
packet still remains a problem. By deploying the FD techniques, the
coexistence between heterogeneous networks may become more flexible.
The research problem is to further reduce the RSI impact and realize
efficient spectrum access management.

1 In this model, only the secondary transmitter (SU ) performs spectrum sensing, while the
receiver (SU ) does not. This transmitter-only sensing mechanism has been widely adopted in
today’s cognitive radio, considering that secondary transmitters and receivers cannot continuously
exchange sensing results without interfering with the primary network. Besides, we assume that
SU has two antennas for fairness and generality, since SU does not always need to be the
receiver.

2 Note that the secondary throughput is not purely convex throughout the domain of transmit
power. There may exist local optimal points in low power region, while the throughput is
monotonically increasing in the high power region. The point of the discussion of the power-
throughput tradeoff and the calculation of the local optimal transmit power is that the secondary
throughput does not monotonously increase with the transmit power, which means that SUs may
not always transmit with their maximum transmit power to achieve highest throughput; instead, a
mediate value may lead to better performance.

3 Note that the FC can also use other fusion rules, such as the AND-rule or majority-rule.
However, the main conclusions remain the same regardless of the specific choice of fusion rules.
In this book, we adopt the OR-rule as a representative.

4 Note that the contention scheme does not need to be the scheme above. Other schemes, such as
that with changeable contention windows, can also be adopted. The expressions of performance
may be different but the major conclusions remain the same.
9 Full-Duplex Random Access Networks

Traditionally, the spectrum allocation policy grants fixed spectrum bands to

licensed users for exclusive access, which worked well over the past
decades. However, with the proliferation of wireless services and data
volume in recent years, spectrum scarcity, as a major drawback of this static
spectrum allocation policy, has been unveiled. Meanwhile, it is observed
that a significant amount of the licensed spectrum is rather underutilized.

As an opposite of the conventional static spectrum management policy,

the concept of dynamic spectrum access has been proposed to increase the
flexibility in spectrum usage, so as to alleviate the spectrum scarcity
problem and improve spectrum utilization. In the existing literature,
Dynamic Spectrum Access (DSA) models can be categorized as follows:
exclusive-use, shared-use, and commons models. In the exclusive-use
model, a licensed user can grant an unlicensed user the spectrum access
rights to have exclusive access to the spectrum. In a shared-use model, an
unlicensed user accesses the spectrum opportunistically without interrupting
a licensed user. In a commons model, an unlicensed user can access the
spectrum freely. DSA can be implemented in a centralized or a distributed
network architecture. DSA can be optimized globally if a central controller
is available in the network. On the other hand, when a central controller is
not available, distributed algorithms would be required for dynamic
spectrum access. Issues related to spectrum trading, such as pricing, will
also need to be considered for dynamic spectrum access, especially with the
exclusive-use model. For DSA-based radio networks, MAC protocols
designed for traditional wireless networks have to be modified to include
spectrum sensing, spectrum access, as well as spectrum trading between a
licensed user and an unlicensed user.

Most current dynamic spectrum access paradigms are designed for HD

devices with the basic assumption that data transmission and reception of
any devices must be separated in the time or frequency domain. Recently,
with the rapid development of self-interference suppression techniques, FD
communication rapidly extends its application to different scenarios. In FD
communications where co-channel simultaneous data transmission and
reception becomes possible, many more possible communication modes
among communication devices arise. For example, two FD devices can
perform bidirectional transmission to each other; one FD devices can
receive data from a source, and transmits data to another destination
concurrently on the same channel. However, it is found that the attractive
benefit that FD technology can bring to the PHY layer cannot be fully
exploited by directly applying current MAC protocols. In other words,
adaptations in the upper layers, especially the MAC layer and routing layers
are highly demanded.

In the design of FD MAC protocols, the devices’ ability to transmit and

receive data concurrently on the same channel can be exploited by
supporting more transmit links with more careful interference management,
or enhancing sensing performance by realizing real-time collision detection.
The former method may achieve nearly double throughput under light load,
but the performance gain may become limited in crowded networks. On the
other hand, the latter method, which is the main focus of this chapter, shows
superiority in crowded networks and high traffic load scenarios. With FD
capability, communication devices can perform continuous spectrum
detection while transmitting to their destinations. Upon collision detection,
they can suspend transmission immediately to reduce the time of failed
transmission, and thus enhance spectrum utilization. In this way, the FD
wireless devices resemble wired devices connecting to a common bus, and
CSMA/CD-like random access mechanisms can be utilized.

In this chapter, we first explain the spectrum access models, followed by

an explanation of half-duplex 802.11 random access. Then we study full-
duplex random access and show the trade-off. Finally, we summarize the
chapter.

9.1 Spectrum Access Models

As has been mentioned before, the traditional “command-and-control”

spectrum management policy results in spectrum under-utilization. In this
spectrum licensing scheme, the radio resource is statically assigned to a
particular licensee [280], while no flexibility is allowed to adjust the
operating frequency bands according to time-varying demands [281]. The
limitations of the “command-and-control”-based spectrum allocation
scheme can be summarized as follows:

The spectrum licensee cannot be changed. Consequently, even if the

radio spectrum is under-utilized, it cannot be reallocated to other
service providers needing to accommodate new users.
The type of wireless service in the licensed spectrum cannot be
changed. Some spectrum resource could be lightly utilized for a
certain period of time. For instance, it has been observed that the 6
MHz spectrum bands allocated to analog TV services remain largely
unused. However, under the current spectrum management policy,
using the TV bands for other services is forbidden.
The spectrum license is location-invariant. Consequently, the
spectrum can remain under-utilized in some locations. Take the
spectrum allocated to the cellular service as an example. While the
spectrum could be heavily used in urban areas, it would be under-
utilized in suburban/rural areas. In fact, it may not be profitable for a
service provider to extend its service to those areas. However, this
spectrum could be useful for other wireless systems (e.g., for
broadband wireless Internet access).
The granularity of spectrum usage is fixed. For example, spectrum
allocation for cellular service is made only in large chunks (e.g., 50
MHz in the 800 MHz band). Thus, it is not possible to allocate the
spectrum band in smaller chunks to meet traffic demands in special
scenarios (e.g., in a hotspot scenario).
 The spectrum is protected from use by unlicensed users.
Traditionally, only the licensed user has the right to use the allocated
radio spectrum exclusively, while unlicensed users are not allowed to
access it. However, this spectrum may not be used by a licensed user
all the time in all the areas. The spectrum utilization can be improved
if an unlicensed user is allowed to opportunistically access the
spectrum as long as the transmissions of the unlicensed users do not
interfere with the transmissions of licensed users.

To overcome the above limitations, the concept of open spectrum [282]

has been introduced. This concept defines a set of techniques and models to
support dynamic management of spectrum bands for wireless
communications, leading to new spectrum access models with the ability to
improve the flexibility and efficiency of spectrum access. These spectrum
access models includeexclusive-use, shared-use of the licensed spectrum,
andcommons models. A taxonomy of the different spectrum access models
is shown in Fig. 9.1.

Figure 9.1 Spectrum access models

9.1.1 Exclusive-Use Model

In the exclusive-use model, the licensed user/service can exclusively use the
radio spectrum resource under certain rules. In this model, a licenser (e.g.,
the government) allocates the spectrum to a licensee. However, the licensee
or the spectrum owner may not fully utilize the allocated spectrum in all
times and in all locations, and the spectrum access rights can be granted to
cognitive radio users. In this case, the spectrum owner is defined as the
licensed user while a cognitive radio user is an unlicensed user. Constrained
by the rules given by the spectrum owner, a cognitive radio user can
optimize its access strategy to achieve the best performance. There are two
variants of this exclusive-use model regarding the timescale of spectrum
allocation, namely long-term exclusive-use model and dynamic exclusive-
use model. In the dynamic exclusive-use model, the spectrum can be
divided into smaller chunks to be allocated to cognitive radio users for a
relatively short time period compared to the long-term exclusive-use model.

Long-Term Exclusive-Use Model

In a long-term exclusive-use model, radio spectrum is allocated exclusively
to a licensed user/service provider for some period of time. After the
spectrum is licensed, the type of wireless service using this spectrum may
or may not be changed (i.e., flexible-type sub-model or fixed-type sub-
model). In a fixed-type sub-model, the spectrum owner specifies the type of
a wireless service as well as its parameters for a cognitive radio user. On the
other hand, in the flexible-type sub-model, the service type of a cognitive
radio user can be various.

Dynamic Exclusive-Use Model

In a dynamic exclusive-use model, at any particular time, only one user can
exclusively utilize the spectrum. However, the permitted user and the type
of wireless service may change over time. In this model, a spectrum owner
can trade its spectrum to a cognitive radio user for a certain period of time
to earn revenue. This mechanism is also known as on-demand time-bound
spectrum lease [283], and the trading is referred to as a secondary
market [284]. Under this secondary market in a dynamic exclusive-use
model are defined three different sub-models: non-real-time secondary
market, real-time secondary market for homogeneous multi-operator
sharing, and heterogeneous multi-operator sharing.

Non-real-time secondary market: In this sub-model, the spectrum is

traded and allocated to a cognitive radio user for a certain period of
time before the spectrum is accessed. Two modes are defined for a
spectrum owner here – spectrum manager and de facto transfer. In
 the spectrum-manager mode, a spectrum owner controls the spectrum
access of a cognitive radio user. In other words, it has to enforce the
rules of spectrum access by a cognitive radio user. On the other hand,
in the de facto transfer mode, a spectrum owner does not control the
spectrum access by a cognitive radio user, and the cognitive radio
user itself must conform its transmission to the rules defined by the
licenser.
Real-time secondary market for homogeneous multi-operator
sharing: In this sub-model, the spectrum can be traded and allocated
to a cognitive radio user in an on-demand basis. However, the
wireless service type is fixed (i.e., homogeneous sharing). When
radio spectrum is needed, a cognitive radio user can request to buy
the spectrum from the spectrum owner. This spectrum trading is
known as spectrum leasing, since the agreement between a spectrum
owner and a cognitive radio user is temporary.
Real-time secondary market for heterogeneous multi-operator
sharing: Similar to homogeneous sharing, but the wireless service
type over the allocated spectrum can be changed. For example, the
spectrum can be traded between a TV broadcaster and a broadband
wireless service provider.

In the exclusive-use model, the spectrum is exclusively allocated by the

spectrum owner to the cognitive radio user, indicating the importance of the
economic issues (e.g., pricing). The spectrum owners would be able to gain
revenue by providing spectrum access rights to the cognitive radio users.

9.1.2 Shared-Use Model

In the shared-use spectrum access model, the spectrum can be shared

between a primary user (i.e., a licensed user) and a secondary user (i.e., an
unlicensed user) simultaneously. In this model, unlicensed users can
opportunistically access the radio spectrum as long as an unlicensed user
does not interrupt or interfere with a primary user (e.g., the collision
probability is maintained below the target level). In a shared-use model, the
spectrum can be accessed by an unlicensed user in two different modes,
namely, spectrum overlay and spectrum underlay modes (Fig. 9.2).
 Figure 9.2 Spectrum overlay and spectrum underlay

Spectrum Overlay
In the case of spectrum overlay, a primary user receives the exclusive right
of spectrum access. However, a secondary user is allowed to
opportunistically access the spectrum when it is not utilized by the primary
user. Therefore, to access a spectrum band, a secondary user needs to
perform spectrum sensing to detect the spectrum opportunities (Fig. 9.2). A
secondary user can access the spectrum only when a spectrum hole is
found. For a secondary user, whether to access the spectrum or not depends
on the constraints such as the collision probability, defined as the
probability that the transmission from a secondary user occurs at the same
time as that from a primary user. Spectrum overlay can be used for
cognitive radio in TDMA or OFDM wireless systems.

Spectrum Underlay

In the case of spectrum underlay, a secondary user can transmit

concurrently with a primary user with limited transmit power so that the
interference caused to the primary users remains below the interference
temperature limit. Spectrum underlay can be used for cognitive radio
systems using CDMA or UWB technology.

9.1.3 Spectrum Commons Model

In a spectrum commons model, all cognitive radio users have the same right
to access the radio spectrum. There are three variants of this model, namely,
uncontrolled, managed, and private-commons sub-models.
Uncontrolled-Commons Sub-Model
This is the simplest form of spectrum access, which is already being used in
the ISM (2.4 GHz) and U-NII (5 GHz) unlicensed bands. In this case, the
spectrum is not owned by any entity, and only the maximum transmit power
constraint applies to a cognitive radio user. However, since there is no
control over spectrum access, cognitive radio users suffer from interference,
which can be either uncontrolled or controlled interference (tragedy of the
commons [285]). Uncontrolled interference is commonly brought by the
devices outside the network. For example, for the IEEE 802.11 b/g
networks operating in the 2.4 GHz band, the uncontrolled interference
could result from microwave ovens [286, 287] or Bluetooth devices. In
contrast, controlled interference comes from the neighboring devices within
the network. For example, multiple IEEE 802.11 b/g networks may
interfere with each other if they decide to operate on the same channel.

Managed-Commons Sub-Model
The managed-commons sub-model avoids the tragedy of the commons in
the uncontrolled-commons spectrum access sub-model. This managed-
commons sub-model considers the spectrum as a resource which needs joint
control by a group of cognitive radios. The cognitive radios must follow the
rules/restrictions to access the spectrum [288] to mitigate interference. In
this case, a management protocol as well as a reliable and scalable
mechanism for a cognitive radio user is required. The management protocol
needs to satisfy the following objectives [289]:

Support advanced and efficient device design, services, and business

model,
Minimize communication and coordination overhead for a spectrum
access,
Provide flexibility for protocol changes in the future to support new
technology, and
Promote fair spectrum access among the cognitive radio users.

The basic rules for distributed spectrum access and management are as
follows [289]:
 The transmitter of all cognitive radio users must conform to the
maximum transmit power limit. The limit should be defined based on
aggregated transmit power from all users (e.g., interference
temperature limit).
All cognitive radio users must be able to detect the presence of both
transmitter-only and receiver-only licensed users. An example of a
transmitter-only user is a cellular base station, while that of a
receiver-only user is a TV receiver.
All cognitive radio users must be able to exchange information (e.g.,
transmission parameters) among each other to coordinate spectrum
access. This information exchange can be performed through out-
band signaling or in-band signaling.
Resource allocation and multiple access control are required in a
cognitive radio network. If multiple cognitive radio users access the
spectrum at the same time, collisions will occur. These collisions
need to be minimized to maximize the throughput of the cognitive
radio users. The resource allocation mechanism should guarantee the
performance requirements for the cognitive radios admitted into the
network.

An important issue in managed-commons spectrum access sub-model is

the enforcement of the spectrum access rules considering the possibility that
there could be some misbehaving cognitive radio users violating the
spectrum access agreement. Note that the issue of protocol misbehavior in
wireless networks has been studied extensively in the literature [290–293]
and different techniques (e.g., game theory) have been applied to avoid this
problem.

To avoid misbehavior in dynamic spectrum access, proactive or reactive

techniques can be used. The former are applied before the cognitive radio
users start misbehaving, including the rules (e.g., maximum power limit)
and the enforcement mechanisms (e.g., power allocation), while the latter
are applied as punishment to a misbehaving cognitive radio.

Private-Commons Sub-Model
In this model, a spectrum owner can specify a technology and a protocol for
the cognitive radio users to access the spectrum by delivering a command
containing the transmission parameters (e.g., time, frequency band, and
transmit power) to be used by a cognitive radio user. Alternatively, a
cognitive radio user may opportunistically sense and access the spectrum
without interrupting a spectrum owner. Note that the differences between
this private-commons sub-model and a shared-use model are as follows:

A cognitive radio user must receive an approval from the spectrum

owner before opportunistically accessing the spectrum;
A cognitive radio user can access the spectrum owned by a particular
owner; and
The sensing and coordination protocols must be approved by the
spectrum owner.

Currently, the application of this private-commons sub-model is limited

only for peer-to-peer communication between the cognitive radio users in a
flat network. Note that, in this model, a spectrum owner can control the
spectrum access of the cognitive radio user, while the cognitive radio user
has flexibility to access the spectrum. Therefore, a secondary market can be
established for spectrum trading as in an exclusive-use model.

9.2 Traditional Half-Duplex MAC

Multiple access means that multiple nodes may send and receive on the
medium. Transmissions by one node are generally received by all other
nodes connected to the medium. The most classic MAC protocol is
ALOHA, in which a mobile station transmits immediately whenever is has
data; it then waits for ACK or NACK; if ACK is not received, it waits a
random amount of time and retransmits. However, ALOHA has poor
performance, since it does not listen to the carrier before transmission and
consequently a lot of collisions can happen. To overcome this challenge,
there are many existing solutions based on the full duplex. In this section,
we first discuss several classic methods for random access, and then we
discuss a few challenges. Finally, we illustrate the IEEE 802.11 MAC
protocol.
9.2.1 Carrier Sensing Multiple Access

Carrier Sense Multiple Access (CSMA) is a probabilistic Media Access

Control (MAC) protocol in which a node verifies the absence of other
traffic before transmitting on a shared transmission medium, such as an
electrical bus, or a band of the electromagnetic spectrum. “Carrier sense”
means that a transmitter attempts to determine whether another transmission
is in progress before initiating a transmission. That is, it tries to detect the
presence of a carrier signal from another node before attempting to
transmit. If a carrier is sensed, the node waits for the transmission in
progress to end before initiating its own transmission. In other words,
CSMA is based on the principle “sense before transmit” or “listen before
talk.”

There are two important parameters that affect CSMA performance:

detection delay, which is the time required to sense the carrier and decide if
it is idle or busy, and propagation delay which is the time required for a bit
to travel from the transmitter to the receiver. If the detection delay is too
long, there will be more time wasted for detection after the channel is
vacant. If the propagation delay is too large, there will be higher
probabilities of packet collision. This is the reason why some situations
with long delays, like satellite cases, require reservation protocols instead of
directly using CSMA.

There are different types of CSMA, which are summarized as follows:

1-persistent CSMA: A station waits until a channel is idle. When it

detects that the channel is idle, it immediately starts transmission;
Non-persistent CSMA: When a station receives a negative
acknowledgement, it waits a random amount of time before
retransmission of the packet, although the carrier is idle;
P-persistent CSMA: P-persistent CSMA is applied to slotted
channels; when a station detects that a channel is idle, it starts
transmission with probability P in the first available time slot;
Carrier Sense Multiple Access/Collision Detection (CSMA/CD):
This is the same as CSMA, except that a station also listens to the
 carrier while transmitting to see if the transmission collides with
someone else’s transmission. CSMA/CD can be used in listen-before-
talk-capable channels (which require full duplex). In a single radio
channel, the transmission need to be interrupted in order to sense the
channel.

9.2.2 Hidden Terminal/Exposed Terminal Problems and

RTS/CTS

CSMA protocols sense the carrier, but sensing the carrier does not always
release true information about the status of the wireless channel. There are
two problems that are unique to wireless channels (different from wired
channels), that make CSMA useless in some cases. These problems are:
Hidden terminal problem and Exposed terminal problem.

Hidden Terminal Problem

Hidden nodes in a wireless network are nodes that are out of range of other
nodes or a collection of nodes. Take a physical star topology with an access
point with many nodes surrounding it in a circular fashion: each node is
within communication range of the access point, but the nodes cannot
communicate with each other, as they do not have a physical connection to
each other. As shown in Fig. 9.3 (a), it is likely that the node at the far edge
of the access point’s range, which is known as A, can see access point B,
but it is unlikely that the same node can see a node on the opposite end of
the access point’s range, C. These nodes are known as hidden. The problem
is when nodes A and C start to send packets simultaneously to the access
point B, because nodes A and C are out of range of each other and so
cannot detect a collision while transmitting.
 Figure 9.3 Two problems for random access networks: (a) Hidden terminal problem (b) Exposed
terminal problem

The hidden node problem can be observed easily in widespread (~50 m

radius) WLAN setups with many nodes that use directional antennas and
have high upload. This is why IEEE 802.11 is suited for bridging the last
mile for broadband access only to a very limited extent. Newer standards
such as WiMAX assign time slots to individual stations, thus preventing
multiple nodes from sending simultaneously and ensuring fairness even in
over-subscription scenarios.

Exposed Terminal Problem

In wireless networks, the exposed node problem occurs when a node is
prevented from sending packets to other nodes because of a neighboring
transmitter. Consider an example in Fig. 9.3 (b) of four nodes labeled A, B,
C, and D, where the two receivers (A, D) are out of range of each other, yet
the two transmitters (B, C) in the middle are in range of each other. Here, if
a transmission between A and B is taking place, node C is prevented from
transmitting to D as it concludes after carrier sense that it will interfere with
the transmission by its neighbor B. However note that D could still receive
the transmission of C without interference because it is out of range of B.

Request to Send/Clear to Send (RTS/CTS)

RTS/CTS (Request to Send / Clear to Send) is the optional mechanism used
by the 802.11 wireless networking protocol to reduce the problems
introduced by the hidden node problem and exposed terminal problem. The
protocol is described as following: When a station wants to transmit data, 1)
it sends an RTS (Ready-to-Send) packet to the intended receiver. The RTS
packet contains the length of the data that needs to be transmitted. Any
station other than the intended recipient hearing RTS defers transmission
for a time duration equal to the end of the corresponding CTS reception. 2)
The receiver sends a CTS (Clear-to-Send) packet back to the sender if it is
available to receive. The CTS packet contains the length of the data that the
original sender wants to transmit. Any station other than the original RTS
sender, hearing CTS, defers transmission until the data is sent. 3) The
original sender, upon reception of the CTS, starts transmitting.

The reasons why RTS/CTS can solve the two problems are illustrated in
Fig. 9.4. For the hidden terminal problem, the access point B will broadcast
an acknowledge signal, and node C will be notified for the hidden terminal
A. As a result, node C will not transmit. For the exposed terminal problem,
node B sends RTS and node C knows that node B is not talking to node D.
So node C is not prevented from transmission.

Figure 9.4 RTS/CTS Solutions for (a) Hidden terminal problem (b) Exposed terminal problem

9.2.3 IEEE802.11 MAC

IEEE 802.11 is a part of the 802 standard family for local area networks.
This family defines the physical and data link layers specified in the
International Organization for Standization Open Systems Interconnection
(OSI) basic reference model. More specifically, 802.11 defines the Medium
Access Control (MAC) layer and PHYsical (PHY) layer. Its relation with
other IEEE 802 standards is illustrated in Fig. 9.5.
 Figure 9.5 IEEE 802 standard family for local area networks

The 802.11 family consists of a series of half-duplex over-the-air

modulation techniques that use the same 802.11 MAC protocol. 802.11-
1997 was the first wireless networking standard in the family, but 802.11b
was the first widely accepted one, followed by 802.11a, 802.11g, 802.11n,
and 802.11ac. Other standards in the family (cCf, h, j) are service
amendments that are used to extend the current scope of the existing
standard, which may also include corrections to a previous specification.

The basic access method for 802.11 is the Distributed Coordination

Function (DCF) which uses the Carrier Sense Multiple Access/Collision
Avoidance (CSMA/CA), which requires each station to listen to other users
before access. If the channel is occupied by some stations, other stations
need to wait until the transmission stops, and then enter into a random back-
off procedure. This prevents multiple stations from seizing the medium
immediately after completion of the preceding transmission.

The CSMA/CA mechanism is very similar in operation to the

CSMA/CD in Ethernet. In both protocols, the availability of the
transmission medium is detected through carrier sensing. The distinguishing
feature between CSMA/CA and CSMA/CD is that the former requires the
receiver to send a positive acknowledgment (ACK) back to the transmitter
if a frame is received correctly. Retransmission is scheduled by the
transmitter if no ACK is returned. In CSMA/CD, the transmitter makes use
of collision detection to determine whether a data frame has been
transmitted successfully. The function of the MAC protocol is common to
all PHY layer options (i.e., FHSS, DSSS, etc.) and is independent of the
data rates [294].

In CSMA/CA, every active node which has a new packet to transmit

monitors channel activity before it accesses to the medium. A new
transmission is allowed only after the channel is free for a common DCF
Inter-Frame Spacing (DIFS) time. If a transmitter senses a busy medium, it
determines a random back-off period by setting an internal timer to an
integer number of slot times. Upon expiration of a DIFS, the timer begins to
decrement. If the timer reaches zero, the station may begin transmission.
The back-off time of a node is given by

(9.1)

where CW is the contention window, and aSlotTime is the length of a time

slot. However, collision is still possible in this protocol due to concurrent
transmission. After each unsuccessful transmission, CW is multiplied by σ
(which has the default value of 2), which is called the persistence
coefficient, and the back-off time selection and transmission process will be
repeated again. The process goes on till the packet is transmitted
successfully or the size of the contention window reaches its maximum
value , i.e., after each failed transmission,
. After a packet is successfully delivered, the CW
will be set back to .

Upon packet reception, acknowledgement is required, i.e., the receiver

transmits an ACK signal back after the interval of one Short Inter Frame
Space (SIFS) when transmission is finished. The interval SIFS is shorter
than the DIFS, meaning that the ACK has higher priority than other regular
transmissions. The fast acknowledgement is also a salient feature in the
802.11 standard.

9.3 Full-Duplex MAC Protocol Design

In this section, we elaborate a novel FD-MAC protocol based on FD
techniques and Carrier Sense Multiple Access with Collision
Avoidance (CSMA/CA) concepts, such that each user can sense the
spectrum and determine whether other users are occupying it while
transmitting its own data. This feature leads to the possibility of a new
MAC protocol design in which users are not assumed to be “blind” when
transmitting.

9.3.1 System Model

As show in Fig. 9.6, we consider an FD-WiFi network consisting of one

Access Point (AP) and N users (U ,…,U ), where the users are
independently and randomly distributed in the coverage area of the AP.
Each user is equipped with two antennas to realize FD communications. We
focus on the uplink traffic, in which data packets are transmitted from the
users to the AP, and each user is assumed to always have a packet to
transmit with the same transmission power.
 Figure 9.6 The CSMA/CD protocol for the uplink traffic of FD-WiFi and two types of sensing
errors (false alarm and missed detection) due to RSI

The channel can serve at most one user at a time, otherwise the collision
happens. Therefore, as shown in the upper part of Fig. 9.6, each user
performs carrier sensing to detect the channel state and contends for the idle
channel against each other by our proposed protocol. When a certain user,
say U , accesses the channel, it uses one antenna for
carrier sensing and the other antenna for data transmission simultaneously.
However, the RSI between those two antennas leads to imperfect sensing,
as shown in the lower part of Fig. 9.6. False alarm happens when the user
mistakenly judges the channel to be occupied by other users when it is not,
while missed detection means that the user fails to detect the occupation of
other transmitting users. Thus, the sensing errors should be considered for
analyzing the network performance. We further discuss the carrier sensing
in FD-WiFi networks as follows.

Full-Duplex Carrier Sensing

Since noise is negligible compared to collision signal and self-interference,
we omit it in this section. Thus, a silent user has a perfect sensing.
Furthermore, the probability of the case of more than two collided users is
negligible compared to the probability that only two users collide, and even
when it happens, the sensing performance is also better due to the
accumulated collision signal. Thus we assume perfect sensing for the case
that three or more users collide, and the sensing errors only exist in the
following two cases: (1) : the transmitting user singly occupies the
channel; (2) : the transmitting user has a collision with another user.

The received signal for sensing at the transmitting user can be given by

(9.2)
where denotes the transmitting user’s signal and is the collided user’s
signal, represents the collision channel, and denotes the equivalent RSI
channel indicating the SIS degree, which depends on the adopted SIS
techniques and network environment. We adopt a typical path loss Rayleigh
fading channel, thus, is zero-mean complex Gaussian distributed with
average power , where α is the path loss exponent and is the
reference received signal power at the reference distance , and d is the
distance between two users. Moreover, according to [295], is also a
complex Gaussian variable with zero mean and average power , where
denotes the SIS factor.

As for the sensing strategy, energy detection is adopted, and we assume

the process is time-slotted. Thus the sensing test statistics can be given by

(9.3)

where denotes the mth sample of the received sensing signal, and is
the sampling number in one slot.

The transmitting user compares M with the sensing threshold to decide

whether a collision happens or not. Two types of sensing errors exist, i.e.,
false alarm and missed detection. As shown in the lower part of Fig. 9.6,
false alarm wastes available channel slots, while missed detection causes
collisions. We need to balance the two types of sensing errors to ensure
network performance.

9.3.2 MAC Protocol Design

Fig. 9.7 shows the proposed protocol, which consists of the following
several parts.
 Figure 9.7 FD-MAC Protocol in decentralized FD networks, in which denotes the
residual back-off time and the back-off stage of user i

Sensing: All users keep sensing the channel continuously regardless of

their own activity, and make decisions about channel usage at the end of
each slot with duration τ, which is the required time to reliably detect the
transmission of any other user.

Back-off mechanism: Once the channel is judged idle without

interruption for a certain period of time as long as a Distributed
InterFerence Space (DIFS) (shown as the dotted area below each line),
users check their own back-off timers and generate a random back-off time
for additional deferral if their timers have counted down to zero. The
additional back-off time after a DIFS is also slotted by τ, i.e., the back-off
time is expressed as

(9.4)
where CW is the contention window length, and
is a random integer drawn from the uniform distribution
over the interval , where is the back-off stage
depending on the number of unsuccessful transmissions for a packet. The
countdown starts right after the DIFS, and suspends when the channel is
detected as occupied by others.

Channel access and transmission suspension: A user accesses the

channel and begins transmission when its timer reaches zero. During the
transmission, if it detects a signal from other users, it stops its transmission
and switches to the back-off procedure immediately. If the packet is
finished, the user resets the back-off state . Otherwise, it sets
.

9.3.3 Performance Analysis

In this part, we study the analytical performance of the proposed FD-MAC

protocol, and derive its saturation throughput. We first analyze the carrier
sensing performance and derive expressions for sensing error probabilities,
then we derive the throughput performance of our proposed MAC protocol
by taking the sensing errors. Note that when only one user is transmitting,
all other users can detect its transmission perfectly, which means that once a
collision-free transmission begins, it either completes the packet or
suspends it because of a false alarm. This process is independent of other
users’ sensing and contending, and thus, contention and transmission can be
considered separately.

Carrier Sensing Performance

We mainly derive the expressions of false alarm probability ( ) and missed
detection probability ( ). With Rayleigh fading channels, the sampling
signal power ( ) is Chi-square distributed. Furthermore, M is the sum
of sampling signal power in one slot, thus according to [296], M is gamma
distributed, the probability density function of which can be expressed as
 (9.5)

where , and for and , respectively.

With a certain sensing threshold ϵ, we can obtain the expressions of

and

(9.6)

(9.7)

where is the incomplete gamma function. By

deriving the expression of ϵ in terms of from (9.6) and substituting that
into (9.7), the following equations are presented:

(9.8)

(9.9)

where , , and , in which is the

inverse incomplete gamma function.

Furthermore, users are independently and randomly distributed in AP’s

coverage area, the radius of which is denoted by R. Then we can derive the
average missed detection probability
 (9.10)

where , are the distances of the transmitting user and collided user away
from the AP, θ is their included angle, and is the
distance between them. We can find that the expression of is related to
the path-loss exponent. Particularly, when free space channel is considered,
i.e., , we can derive an approximation of :

(9.11)

According to (9.6)–(9.11), is negatively related to ϵ, while is

positively related. Therefore, the sensing threshold should be well-designed
to balance false alarm and missed detection.

Collision Probability
We follow the assumption in [297] that each packet is collided with a
similar probability independent of the value of CW . Let denote the
state of the ith contending user. For each user, the state change can be
modeled as a discrete-time Markov chain, as illustrated in Fig. 9.8. The
nonzero transition probabilities are given as
 (9.12)

where denotes the probability that the considered user successfully

finishes its transmission without awareness of collision. Note that does
not equal the non-collision probability due to imperfect sensing.
Specifically, if two users collide, it is possible that only one user stops, and
when one user is transmitting without collision, it may cease the
transmission due to a false alarm.

Figure 9.8 Markov chain of the back-off window size

 Consider the steady-state distribution of the Markov chain: the
probability that one user stays in each state can be calculated. Let
denote the probability that one user is in the state of , and the
probability that a certain user begins transmission in the next slot is

(9.13)

Then, consider the relation between and p. For simplicity, we assume

the packet length L is fixed. The calculation of has two prerequisites:

1. The probability that one user starts collision-free transmission after

colliding with others for l slots, denoted as , which can be
expressed as

(9.14)
2. The probability of successfully finishing the current packet with
residual collision-free length of l, denoted as :

(9.15)

Successful transmission requires that at least one user transmits the entire
packet without the awareness of collision. Thus, can be calculated as
 (9.16)

Combining (9.13) and (9.16), the values of p and can be solved

numerically.

Throughput
We use the time fraction that the channel is occupied for successful
transmission as the normalized throughput, i.e., the throughput is defined
as,

(9.17)

where denotes the probability that a successful

transmission occurs, is the probability that the channel is
empty, represents the collision probability, and , and
denote the average length of successful transmission, empty state, and
collision, respectively. The average length of successful transmission and
collision can be calculated as, respectively,

(9.18)
 (9.19)

The throughput is readily obtained by substituting (9.18) and (9.19) into

(9.17).

Comparison with the Basic CSMA/CA Mechanism

We make a comparison between the proposed FD-MAC protocol with the
conventional CSMA/CA in this subsection. For fairness, we consider the
same system with M users, and omit the noise term. The analytical
performance of CSMA/CA is elaborated in [297], which is omitted here due
to the space limitation. Some main differences between the two protocols
are listed as follows.

Collision length. In conventional CSMA/CA, the “blindness” in

transmission results in a long collisions, which is typically a packet
length. FD allows users to detect collisions while transmitting. Thus,
the average collision length , as is derived in (9.19), is slightly
more than one slot, which is sharply reduced compared with
CSMA/CA.
Successful transmission length. In CSMA, once a collision-free
transmission begins, it can always be finished successfully without
interruption. However, in FD-MAC, the transmission may be
terminated due to a false alarm, especially for long packets.
According to (9.18), if L is sufficiently large, goes to . Also, a
false alarm leads to unnecessary back-off and increase of contention
window, which may further degrade the performance of FD-MAC.
Thus, in FD-MAC, the design of an appropriate packet length should
be carefully considered.

9.3.4 Simulation Results

In this section, simulation results are presented to show the performance of
our proposed FD-WiFi MAC protocol. We consider 20 users and AP’s
radius R is set as 10 m. We set and mW with unit. The
slot sampling number is set as 100. Furthermore, the packet length is
fixed to be 100 slots and DIFS is 2 slots. We run for transmission
attempts to fully develop the WiFi network and proceed with another
packet transmissions to obtain the simulated results by MATLAB. For
comparison, we provide the throughput performance of conventional HD-
WiFi and bidirectional FD transmission, which is similar to [298] and [299]
and named “Dual-Link FD-WiFi.” Specifically, for the dual-link simulation,
once a certain user transmits data to the AP, the AP also transmits data
packets back to this user. Moreover, due to RSI, the rate of dual-link at one
end is less than that of single-link. For simplicity, the single-link sum-rate is
normalized as 1 and the dual-link sum-rate is denoted as the relevant ratio.

In Fig. 9.9, we show the throughput of our proposed FD-WiFi protocol

versus a false alarm probability, which consists of two cases with the SIS
factor and , respectively. Note that the normalized
throughput is denoted as the time ratio that the channel is occupied for
successful transmissions, as shown in (9.17); thus it has no dimension.
Fig. 9.9 shows that there exists an optimal value of for FD-WiFi to
achieve the maximum throughput. Since is determined by ϵ, as shown in
(9.6), the sensing threshold should be well designed to achieve the
maximum throughput. We can also find that the optimal value of for the
case with is higher than that with , which can be
explained as follows. A higher SIS factor leads to worse sensing
performance. Setting the optimal value for the case with , the
collision is more frequent due to higher when increases to 0.3. Thus,
to obtain the maximum throughput, should be decreased by increasing
.
 Figure 9.9 Normalized throughput C versus false alarm probability , where the number of users
, the minimum contention window size , and the maximum contention
window size

In Fig. 9.10, by considering two cases with packet lengths and

, we present the relationship between network throughput and
maximum back-off stage, with the throughput of convention HD-WiFi and
dual-link FD-WiFi for comparison. According to Fig. 9.10, we can find
that, with proper parameters, our proposed FD-WiFi protocol has a better
throughput performance than HD-WiFi and dual-link FD-WiFi. For the HD-
WiFi and dual-link FD-WiFi, with higher , collision is less likely to
happen, so the throughput increases monotonously. However, the
throughput of our proposed FD-WiFi protocol may drop with large ,
which can be found in the dashed line. The reason is that the asymptotic
value of is 100 with , and for slots, false alarm is quite
likely to happen during transmission and the users are likely to enlarge their
contention windows up to due to the unsuccessful transmissions.
Thus, more time is spent in the back-off procedure and the throughput gets
smaller.
 Figure 9.10 Normalized throughput C of different protocols versus the maximum back-off stage
, where , SIS factor , and the false alarm probability in FD-WiFi
CSMA/CD protocol

9.4 Summary

In this chapter, a MAC protocol has been designed for FD random access
networks to improve spectrum utilization based on 802.11 DCF
(CSMA/CA) mechanism. Different from the 802.11 DCF mechanism, the
new FD MAC protocol exploits the benefits brought by FD technology by
allowing devices to perform simultaneous spectrum sensing and data
transmission, making the wireless devices resemble the wired devices
connected to a common bus. With continuous sensing, the FD devices can
response quickly to collisions without finishing a long packet, and collision
time can be significantly reduced. By considering the impact of the RSI in
PHY layer, we have analyzed two types of sensing error probabilities,
namely false alarm and missed detection probabilities, based on which the
detection threshold for energy detection has been derived, and analytical
spectrum utilization efficiency has been provided. We have shown by
analysis and simulation results the performance gains of the proposed FD
MAC protocol compared to the conventional CSMA/CA protocol.
References

[1] K. J. Kerpez, D. L. Waring, S. Galli, J. Dixon, and P. Madon, “Advanced DSL Management,”
IEEE Commun. Mag., vol. 41, no. 9, pp. 116–123, Sep. 2003.
[2] L. Wooyul, K. Youngjae, M. H. Brady, and J. M. Cioffi, “Band-Preference Dynamic
Spectrum Management in a DSL Environment,” in Proc. IEEE Global Telecommunications
Conference (GLOBECOM), San Francisco, CA, Nov.–Dec. 2006, pp. 1–5.
[3] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge University
Press, May 2005.
[4] L. Zheng and D. N. C. Tse, “Diversity and Multiplexing: A Fundamental Tradeoff in
Multiple-Antenna Channels,” IEEE Trans. Inform. Theory, vol. 49, no. 5, pp. 1073–1096,
May 2003.
[5] S. M. Alamouti, “A Simple Transmit Diversity Technique for Wireless Communications,”
IEEE J. Select. Areas Commun., vol. 16, no. 8, pp. 1451–1458, Oct. 1998.
[6] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors, and F. Tufvesson,
“Scaling Up MIMO: Opportunities and Challenges with Very Large Arrays,” IEEE Signal
Processing Mag., vol. 30, no. 1, pp. 40–60, Jan. 2013.
[7] J. Nam, J.-Y. Ahn, A. Adhikary, and G. Caire, “Joint Spatial Division and Multiplexing:
Realizing Massive MIMO Gains with Limited Channel State Information,” in Annual
Conference on Information Sciences and Systems (CISS), Princeton, NJ, Mar. 2012, pp. 1–6.
[8] C. Shepard, N. A. H. Yu, L. E. Li, T. L. Marzetta, R. Yang, and L. Zhong, “Argos: Practical
Many-Antenna Base Stations,” in ACM Int. Conf. Mobile Computing and Networking
(MobiCom), Istanbul, Turkey, Aug. 2012, pp. 53–64.
[9] X. Gao, F. Tufvesson, O. Edfors, and F. Rusek, “Measured Propagation Characteristics for
Very-Large MIMO at 2.6 GHz,” in Proc. of Annual Asilomar Conference on Signals,
Systems, and Computers, Pacific Grove, California, USA, Nov. 2012, pp. 295–299.
[10] J. Hoydis, C. Hoek, T. Wild, and S. ten Brink, “Channel Measurements for Large Antenna
Arrays,” in IEEE International Symposium on Wireless Communication Systems (ISWCS),
 Paris, France, Aug. 2012, pp. 811–815.
[11] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and Spectral Efficiency of Very Large
Multiuser MIMO Systems,” IEEE Trans. Commun., vol. 61, no. 4, pp. 1436–1449, Apr.
2013.
[12] A. Pitarokoilis, S. K. Mohammed, and E. G. Larsson, “On the Optimality of Single-Carrier
Transmission in Large-Scale Antenna Systems,” IEEE Wireless Commun. Lett., vol. 1, no. 4,
pp. 276–279, Aug. 2012.
[13] H. Yang and T. L. Marzetta, “Performance of Conjugate and Zero-Forcing Beamforming in
Large-Scale Antenna Systems,” IEEE J. Select. Areas Commun., vol. 31, no. 2, pp. 172–179,
Feb. 2013.
[14] C. Studer and E. G. Larsson, “PAR-Aware Large-Scale Multi-User MIMO-OFDM
Downlink,” IEEE J. Select. Areas Commun., vol. 31, no. 2, pp. 303–313, Feb. 2013.
[15] S. K. Mohammed and E. G. Larsson, “Per-Antenna Constant Envelope Precoding for Large
Multi-User MIMO Systems,” IEEE Trans. Commun., vol. 61, no. 3, pp. 1059–1071, Mar.
2013.
[16] P. Stenumgaard, D. Persson, K. Wiklund, and E. G. Larsson, “An Early-Warning Service for
Emerging Communication Problems in Security and Safety Applications,” IEEE Commun.
Mag., vol. 51, no. 5, pp. 186–192, Mar. 2013.
[17] F. Kaltenberger, J. Haiyong, M. Guillaud, and R. Knopp, “Relative Channel Reciprocity
Calibration in MIMO/TDD Systems,” in Proc. of Future Network and Mobile Summit,
Florence, Italy, Jun. 2010, pp. 1–10.
[18] T. L. Marzetta, “Noncooperative Cellular Wireless with Unlimited Numbers of Base Station
Antennas,” IEEE Trans. Wireless Commun., vol. 9, no. 11, pp. 3590–3600, Nov. 2010.
[19] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL of Cellular
Networks: How Many Antennas do We Need?” IEEE J. Select. Areas Commun., vol. 31,
no. 2, pp. 160–171, Feb. 2013.
[20] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A Coordinated Approach to Channel Estimation
in Large-Scale Multiple-Antenna Systems,” IEEE J. Select. Areas Commun., vol. 31, no. 2,
pp. 264–273, Feb. 2013.
[21] H. Q. Ngo and E. G. Larsson, “EVD-Based Channel Estimations for Multicell Multiuser
MIMO with Very Large Antenna Arrays,” in Proc. of the IEEE International Conference on
Acoustics, Speed and Signal Processing (ICASSP), Tokyo, Japan, Mar. 2012, pp. 3249–
3252.
[22] A. Ashikhmin and T. L. Marzetta, “Pilot Contamination Precoding in Multi-Cell Large Scale
Antenna Systems,” in IEEE International Symposium on Information Theory (ISIT),
Cambridge, MA, Jul. 2012, pp. 1137–1141.
[23] J. Zhang, X. Yuan, and L. Ping, “Hermitian Precoding for Distributed MIMO Systems with
Individual Channel State Information,” IEEE J. Select. Areas Commun., vol. 31, no. 2, pp.
241–250, Feb. 2013.
[24] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive MIMO for Next
Generation Wireless Systems,” arXiv:1304.6690, 2013.
[25] T. S. Rappaport, J. N. Murdock, and F. Gutierrez, “State of the Art in 60 GHz Integrated
Circuits & Systems for Wireless Communications,” Proc. of IEEE, vol. 99, no. 8, pp. 1390–
1436, Aug. 2011.
[26] Z. Pi and F. Khan, “An Introduction to Millimeter-Wave Mobile Broadband Systems,” Proc.
of IEEE Communication Magazine, vol. 49, no. 6, pp. 101–107, Jun. 2011.
[27] F. Gutierrez, S. Agarwal, K. Parrish, and T. S. Rappaport, “On-Chip Integrated Antenna
Structures in CMOS for 60 GHz WPAN Systems,” IEEE J. Sel. Areas Commun., vol. 27,
no. 8, pp. 1367–1378, Oct. 2009.
[28] T. S. Rappaport, E. Ben-Dor, J. N. Murdock, and Y. Qiao, “38 GHz and 60 GHz Angle-
Dependent Propagation for Cellular and Peer-to-Peer Wireless Communications,” in Proc. of
IEEE Int. Conf. Commun., Ottawa, Canada, Jun. 2012, pp. 4568–4573.
[29] S. Y. Seidel and H. W. Arnold, “Propagation Measurements at 28 GHz to Investigate the
Performance of Local Multipoint Distribution Service (LMDS),” in Proc. of Global
Telecommun. Conf., Singapore, Nov. 1995, pp. 754–757.
[30] Q. Zhao and J. Li, “Rain Attenuation in Millimeter Wave Ranges,” in Proc. of IEEE Int.
Symp. Antennas, Propag. EM Theory, Guangxi, China, Oct. 2006, pp. 1–4.
[31] C. R. Anderson and T. S. Rappaport, “In-Building Wideband Partition Loss Measurements at
2.5 and 60 GHz,” IEEE Trans. Wireless Commun., vol. 3, no. 3, pp. 922–928, May 2004.
[32] H. Zhao, R. Mayzus, S. Sun, M. Samimi, J. K. Schulz, Y. Azar, K. Wang, G. N. Wong,
F. Gutierrez, and S. T. Rappaport, “28 GHz Millimeter Wave Cellular Communication
Measurements for Reflection and Penetration Loss in and around Buildings in New York
City,” in Proc. of IEEE Int. Conf. Commun., Budapest, Hungary, Jun. 2013, pp. 5163–5167.
[33] Y. Azar, G. N. Wong, K. Wang, R. Mayzus, J. K. Schulz, H. Zhao, F. Gutierrez, D. Hwang,
and T. S. Rappaport, “28 GHz Propagation Measurements for Outdoor Cellular
Communications using Steerable Beam Antennas in New York City,” in Proc. of IEEE Int.
Conf. Commun., Budapest, Hungary, Jun. 2013, pp. 5143–5147.
[34] M. Samimi, Y. A. K. Wang, G. N. Wong, R. Mayzus, H. Zhao, J. K. Schulz, S. Sun,
F. Gutierrez, and T. S. Rappaport, “28 GHz Angle of Arrival and Angle of Departure
Analysis for Outdoor Cellular Communications using Steerable-Beam Antennas in New
York City,” in Proc. of IEEE Veh. Technol. Conf., Dresden, Germany, Jun. 2013, pp. 1–6.
[35] T. S. Rappaport, F. Gutierrez, E. Ben-Dor, J. N. Murdock, Y. Qiao, and J. I. Tamir,
“Broadband Millimeter Wave Propagation Measurements and Models using Adaptive Beam
 Antennas for Outdoor Urban Cellular Communications,” IEEE Trans. Wireless Commun.,
vol. 61, no. 4, pp. 1850–1859, Apr. 2013.
[36] J. N. Murdock, E. Ben-Dor, Y. Qiao, J. I. Tamir, and T. S. Rappaport, “A 38 GHz Cellular
Outage Study for an Urban Campus Environment,” in Proc. of IEEE Wireless Commun.
Netw. Conf., Paris, France, Apr. 2012, pp. 3085–3090.
[37] T. S. Rappaport, Y. Qiao, J. I. Tamir, J. N. Murdock, and E. Ben-Dor, “Cellular Broadband
Millimeter Wave Propagation and Angle of Arrival for Adaptive Beam Steering Systems,”
in IEEE Radio and Wireless Symposium (RWS), Santa Clara, CA, Jan. 2012, pp. 151–154.
[38] P. Smulders, “Exploiting the 60 GHz Band for Local Wireless Multimedia Access: Prospects
and Future Directions,” IEEE Commun. Mag., vol. 40, no. 1, pp. 140–147, Jan. 2002.
[39] P. F. M. Smulders, C. F. Li, H. Yang, E. F. T. Martijn, and M. H. A. J. Herben, “60 GHz
Indoor Radio Propagation – Comparison of Simulation and Measurement Results,” in Proc.
of IEEE Symp. on Commun. and Vehi. Tech., Benelux, Ghent, Belgium, Jan. 2004.
[40] B. Neekzad, K. Sayrafian-Pour, J. Perez, and J. S. Baras, “Comparison of Ray Tracing
Simulations and Millimeter Wave Channel Sounding Measurements,” in IEEE Int. Symp.
Personal, Indoor and Mobile Radio Commun., Athens, Greek, Sep. 2007, pp. 1–5.
[41] J. Meng, J. Ahmadi-Shokouh, H. Li, E. J. Charlson, Z. Han, S. Noghanian, and E. Hossain,
“Sampling Rate Reduction for 60 GHz UWB Communication using Compressive Sensing,”
in Asilomar Conference on Signals, Systems & Computers, Pacific Grove, CA, Nov. 2009,
pp. 1125–1129.
[42] B. Sharif, M. Hassibi, “On the Capacity of MIMO Broadcast Channels with Partial Side
Information,” IEEE Trans. Inform. Theory, vol. 51, no. 2, pp. 506– 522, Feb. 2005.
[43] J. N. Laneman, D. N. C. Tse, and G. W. Wornell, “Cooperative Diversity in Wireless
Networks: Efficient Protocols and Outage Behavior,” IEEE Trans. Inform. Theory, vol. 50,
no. 12, pp. 3062–3080, Dec. 2004.
[44] T. M. Cover and A. E. Gamal, “Capacity Theorems for the Relay Channel,” IEEE Trans.
Inform. Theory, vol. 25, no. 5, pp. 572–584, Sep. 1979.
[45] M. A. Khojastepour, A. Sabharwal, and B. Aazhang, “On the Capacity of ‘Cheap’ Relay
Networks,” in Proc. of Conference on Information Sciences and Systems (CISS), Baltimore,
MD, Mar. 2003, pp. 12–14.
[46] T. E. Hunter and A. Nosratinia, “Performance Analysis of Coded Cooperation Diversity,” in
Proc. of International Conference on Communications (ICC), Anchorage, AK, May 2003,
pp. 2688–2692.
[47] J. N. Laneman and G. W. Wornell, “Distributed Space-Time Coded Protocols for Exploiting
Cooperative Diversity in Wireless Networks,” IEEE Trans. Inform. Theory, vol. 49, no. 10,
pp. 2415–2525, Oct. 2003.
[48] A. K. Sadek, K. J. R. Liu, and A. Ephremides, “Cognitive Multiple Access via Cooperation:
Protocol Design and Performance Analysis,” IEEE Trans. Inform. Theory, vol. 53, no. 10,
pp. 3677–3696, Oct. 2007.
[49] T. E. Hunter and A. Nosratinia, “Outage Performance of Cognitive Wireless Relay
Networks,” in Proc. of IEEE Global Telecommunications Conference (GLOBECOM), San
Francisco, CA, Nov.–Dec. 2006, pp. 1–5.
[50] W. Su, A. K. Sadek, and K. J. R. Liu, “SER Performance Analysis and Optimum Power
Allocation for Decode-and-Forward Cooperation Protocol in Wireless Networks,” in Proc.
of IEEE Wireless Communications and Networking Conference (WCNC), New Orleans, LA,
Mar. 2005, pp. 984–989.
[51] A. Høst-Madsen, “Upper and Lower Bounds for Channel Capacity of Asynchronous
Cooperative Diversity Networks,” IEEE Trans. Inform. Theory, vol. 50, no. 4, pp. 3062–
3080, Dec. 2004.
[52] ——, “A New Achievable Rate for Cooperative Diversity Based on Generalized Writing on
Dirty Paper,” in Proc. of IEEE International Symposium Information Theory, Yokohama,
Japan, Jun. 2003, p. 317.
[53] T. E. Hunter, S. Sanayei, and A. Nosratinia, “Outage Analysis of Coded Cooperation,” IEEE
Trans. Inform. Theory, vol. 52, no. 2, pp. 375–391, Feb. 2006.
[54] J. Luo, R. S. Blum, L. J. Greenstein, L. J. Cimini, and A. M. Haimovich, “New Approaches
for Cooperative Use of Multiple Antennas in Ad Hoc Wireless Networks,” in Proc. of the
IEEE Vehicular Technology Conference (VTC-Fall), Los Angeles, CA, Sep. 2004, pp. 2769–
2773.
[55] Y. Zhao, R. S. Adve, and T. J. Lim, “Improving Amplify-and-Forward Relay Networks:
Optimal Power Allocation Versus Selection,” in Proc. of IEEE International Symposium on
Information Theory, Seattle, WA, Jul. 2006, pp. 3114–3123.
[56] Z. Han, X. Zhang, and H. V. Poor, “Cooperative Transmission Protocols with High Spectral
Efficiency and High Diversity Order Using Multiuser Detection and Network Coding,” in
Proc. of IEEE International Conference on Communications (ICC), Glasgow, UK, Jun.
2007, pp. 4232–4236.
[57] Z. Han, T. Himsoon, W. Siriwongpairat, and K. J. R. Liu, “Energy Efficient Cooperative
Transmission over Multiuser OFDM Networks: Who Helps Whom and How to Cooperate,”
in Proc. of IEEE Wireless Communications and Networking Conference, New Orleans, LA,
Mar. 2005, pp. 1030–1035.
[58] Z. Yang, J. Liu, and A. Høst-Madsen, “Cooperative Routing and Power Allocation in Ad-hoc
Networks,” in Proc. of IEEE Global Telecommunications Conference (GLOBECOM), St.
Louis, MO, Nov. 2005, pp. 2730–2734.
[59] A. E. Khandani, E. Modiano, L. Zheng, and J. Abounadi, Advannces in Pervasive Computing
and Networking. Kluwer Academic Publishers, 2004.
[60] A. S. Ibrahim, Z. Han, and K. J. R. Liu, “Distributed Energy-Efficient Cooperative Routing in
Wireless Networks,” in Proc. of IEEE Global Telecommunications Conference
(GLOBECOM), Washington, DC, Nov. 2007, pp. 4413–4418.
[61] Z. Han and H. V. Poor, “Lifetime Improvement in Wireless Sensor Networks via
Collaborative Beamforming and Cooperative Transmission,” IET Microwaves, Antennas and
Propagation, vol. 1, no. 6, pp. 1103–1110, Dec. 2007.
[62] J. Boyer, D. D. Falconer, and H. Yanikomeroglu, “Cooperative Connectivity Models for
Wireless Relay Networks,” IEEE Trans. Wireless Commun., vol. 6, no. 6, pp. 1992–2000,
Jun. 2007.
[63] F. Li, K. Wu, and A. Lippman, “Energy-Efficient Cooperative Routing in Multi-hop Wireless
Ad Hoc Networks,” in Proc. of the IEEE International Performance, Computing, and
Communications Conference, Phoenix, AZ, Apr. 2006, pp. 215–222.
[64] A. K. Sadek, W. Su, and K. J. R. Liu, “A Class of Cooperative Communication Protocols for
Multi-Node Wireless Networks,” in Proc. of IEEE International Workshop on Signal
Processing Advances in Wireless Communications (SPAWC), New York, NY, Jun. 2005, pp.
560–564.
[65] A. Bletsas, A. Lippman, and D. P. Reed, “A Simple Distributed Method for Relay Selection
in Cooperative Diversity Wireless Networks, Based on Reciprocity and Channel
Measurements,” in Proc. of the IEEE Vehicular Technology Conf. (VTC-Spring), vol. 3, May
2005, pp. 1484–1488.
[66] B. Wang, Z. Han, and K. J. R. Liu, “Distributed Relay Selection and Power Control for
Multiuser Cooperative Communication Networks Using Buyer/Seller Game,” in Proc. IEEE
Conf. on Comp. Comm. (INFOCOM), Anchorage, AK, USA, May 2007, pp. 544–552.
[67] Z. Han and H. V. Poor, “Coalition Game with Cooperative Transmission: A Cure for the
Curse of Boundary Nodes in Selfish Packet-Forwarding Wireless Networks,” in Proc. of
International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless
Networks (WiOpt), Limassol, Apr. 2007, pp. 1–8.
[68] D. Gunduz and E. Erkip, “Joint Source-Channel Cooperation: Diversity Versus Spectral
Efficiency,” in Proc. of IEEE International Symposium Information Theory, Chicago, IL,
Jun.–Jul. 2004, p. 392.
[69] I. Maric and R. D. Yates, “Cooperative Multihop Broadcast for Wireless Networks,” IEEE J.
Select. Areas Commun., vol. 22, no. 6, pp. 1080–1088, Aug. 2004.
[70] www.umtsworld.com.
[71] www.cdg.org.
[72] “IEEE Standard for Local and Metropolitan Area Networks Part 16: Air Interface for Fixed
Broadband Wireless Access Systems IEEE Std. 802.16-2004,” (Revision of IEEE Std.
802.16-2001), pp. 27–28, 2004.
[73] “Achieving Wireless Broadband with WiMAX,” IEEE Computer Society, Tech. Rep., 2004.
[74] K. Doppler, M. Rinne, C. Wijting, C. Ribeiro, and K. Hugl, “Device-to-Device
Communication as an Underlay to LTE-Advanced Networks,” IEEE Commun. Mag.,
vol. 47, no. 12, pp. 42–49, Dec. 2009.
[75] S. Basagni, M. Conti, S. Giordano, and I. Stojmenovic, Mobile Ad Hoc Networking. Wiley-
IEEE Press, 2004.
[76] C.-H. Yu, O. Tirkkonen, K. Doppler, and C. Ribeiro, “On the Performance of Device-to-
Device Underlay Communication with Simple Power Control,” in Proc. of IEEE Vehicular
Technology Conference VTC-Spring, Barcelona, Apr. 2009, pp. 1–5.
[77] T. Koskela, S. Hakola, T. Chen, and J. Lehtomaki, “Clustering Concept Using Device-to-
Device Communication in Cellular System,” in Proc. IEEE Wireless Communications and
Networking Conf., Sydney, Apr. 2010, pp. 1–6.
[78] K. Doppler, M. Rinne, P. Janis, C. Ribeiro, and K. Hugl, “Device-to-Device Communications;
Functional Prospects for LTE-Advanced Networks,” in Proc. of IEEE International
Conference on Communications Workshops, Dresden, Jun. 2009, pp. 1–6.
[79] K. Doppler, C.-H. Yu, C. Ribeiro, and P. Janis, “Mode Selection for Device-to-Device
Communication Underlaying an LTE-Advanced Network,” in Proc. IEEE Wireless
Communications and Networking Conf., Sydney, Apr. 2010, pp. 1–6.
[80] H. Min, W. Seo, J. Lee, S. Park, and D. Hong, “Reliability Improvement Using Receive Mode
Selection in the Device-to-Device Uplink Period Underlaying Cellular Networks,” IEEE
Trans. Wireless Commun., vol. 10, no. 2, pp. 413–418, Feb. 2011.
[81] S. Hakola, C. Tao, J. Lehtomaki, and T. Koskela, “Device-to-Device (D2D) Communication
in Cellular Network – Performance Analysis of Optimum and Practical Communication
Mode Selection,” in Proc. IEEE Wireless Communications and Networking Conf., Sydney,
Apr. 2010, pp. 1–6.
[82] C.-H. Yu, K. Doppler, C. Ribeiro, and O. Tirkkonen, “Performance Impact of Fading
Interference to Device-to-Device Communication Underlaying Cellular Networks,” in Proc.
of IEEE International Symposium on Personal, Indoor and Mobile Radio Communications
(PIMRC), Tokyo, Sep. 2009, pp. 858–862.
[83] C.-H. Yu, O. Tirkkonen, K. Doppler, and C. Ribeiro, “Power Optimization of Device-to-
Device Communication Underlaying Cellular Communication,” in Proc. Int. Conf. on
Communications, Dresden, Jun. 2009, pp. 1–5.
[84] H. Xing and S. Hakola, “The Investigation of Power Control Schemes for a Device-to-Device
Communication Integrated into OFDMA Cellular System,” in Proc. of IEEE International
 Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), Instanbul,
Sep. 2010, pp. 1775–1780.
[85] P. Janis, V. Koivunen, C. Ribeiro, K. Doppler, and K. Hugl, “Interference-Avoiding MIMO
Schemes for Device-to-Device Radio Underlaying Cellular Networks,” in Proc. of IEEE
International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC),
Tokyo, Sep. 2009, pp. 2385–2389.
[86] T. Peng, Q. Lu, H. Wang, S. Xu, and W. Wang, “Interference Avoidance Mechanisms in the
Hybrid Cellular and Device-to-Device Systems,” in Proc. of IEEE International Symposium
on Personal, Indoor and Mobile Radio Communications (PIMRC), Tokyo, Sep. 2009, pp.
617–621.
[87] P. Janis, V. Koivunen, C. Ribeiro, J. Korhonen, K. Doppler, and K. Hugl, “Interference-Aware
Resource Allocation for Device-to-Device Radio Underlaying Cellular Networks,” in Proc.
of IEEE Vehicular Technology Conference (VTC-Spring), Barcelona, Apr. 2009, pp. 1–5.
[88] M. Zulhasnine, C. Huang, and A. Srinivasan, “Efficient Resource Allocation for Device-to-
Device Communication Underlaying LTE Network,” in Proc. of IEEE International
Conference on Wireless and Mobile Computing, Networking and Communications, Niagara
Falls, Oct. 2010, pp. 368–375.
[89] S. Xu, H. Wang, T. Chen, Q. Huang, and T. Peng, “Effective Interference Cancellation
Scheme for Device-to-Device Communication Underlaying Cellular Networks,” in Proc. of
IEEE Vehicular Technology Conference (VTC-Fall), Ottawa, Sep. 2010, pp. 1–5.
[90] C.-H. Yu, K. Doppler, C. Ribeiro, and O. Tirkkonen, “Resource Sharing Optimization for
D2D Communication Underlaying Cellular Networks,” IEEE Trans. Wireless Commun.,
vol. 10, no. 8, pp. 2752–2763, Aug. 2011.
[91] HUAWEI, “U-LTE: Unlicensed Spectrum Utilization of LTE,” 2014, White Paper.
[92] Qualcomm, “Making the Best Use of Unlicensed Spectrum for 1000x,” Sep. 2014, White
Paper.
[93] [Online]. Available: www.lteuforum.org
[94] “LTE-U Forum,” 2014, Formed by Verizon in cooperation with Alcatel-Lucent, Ericsson,
Qualcomm Technologies, Inc., a subsidiary of Qualcomm Incorporated, and Samsung.
[Online]. Available: www.lteuforum.org
[95] Qualcomm, “Introducing MuLTEfire: LTE-Like Performance with Wi-Fi-Like Simplicity,”
Jun. 2015. [Online]. Available: www.qualcomm.com/news/onq/2015/06/ 11/introducing-
multefire-lte-performance-wi-fi-simplicity
[96] LTE-U Forum, “LTE-U SDL Coexistence Specifications,” Jun. 2015.
[97] http://w3.antd.nist.gov/wahn_ssn.shtml.
[98] B. Krishnamachari, Networking Wireless Sensors. Cambridge University Press, 2005.
 [99] M. Duarte and A. Sabharwal, “Full-Duplex Wireless Communications Using Off-the-Shelf
Radios: Feasibility and First Results,” in Proceedings of Asilomar Conference on Signals,
Systems and Computers (ASILOMAR), Pacific Grove, CA, Nov. 2010, pp. 1558–1562.
[100] J. Bai and A. Sabharwal, “Decode-and-Cancel for Interference Cancellation in a Three-Node
Full-Duplex Network,” in Proceedings of Asilomar Conference on Signals, Systems and
Computers (ASILOMAR), Pacific Grove, CA, Nov. 2012, pp. 1285–1289.
[101] A. Thangaraj, R. Ganti, and S. Bhashyam, “Self-Interference Cancellation Models for Full-
Duplex Wireless Communications,” in International Conference on Signal Processing and
Communications (SPCOM), Bangalore, Jul. 2012, pp. 1–5.
[102] D. Bharadia, E. McMilin, and S. Katti, “Full Duplex Radios,” in Proc. ACM SIGCOMM,
New York, NY, Aug. 2013, pp. 375–386.
[103] M. Jainy, J. I. Choiy, T. M. Kim, D. Bharadia, S. Seth, K. Srinivasan, P. Levis, S. Katti, and
P. Sinha, “Practical, Real-Time, Full Duplex Wireless,” in ACM MobiCom, Las Vegas,
Nevada, Sep. 2011, pp. 301–312.
[104] W. Cheng, X. Zhang, and H. Zhang, “Full/Half Duplex Based Resource Allocations for
Statistical Quality of Service Provisioning in Wireless Relay Networks,” in IEEE
INFOCOM, Orlando, FL, Mar. 2012, pp. 864–872.
[105] M. J. Emadi, A. G. Davoodi, and M. R. Aref, “Analytical Power Allocation for a Full-Duplex
Decode-and-Forward Relay Channel,” IET Communications, vol. 7, no. 13, pp. 1338–1347,
Sep. 2013.
[106] W. Cheng, X. Zhang, and H. Zhang, “QoS Driven Power Allocation over Full-Duplex
Wireless Links,” in IEEE International Conference on Communications (ICC), Ottawa, ON,
Jun. 2012, pp. 5286–5290.
[107] ——, “Optimal Dynamic Power Control for Full-Duplex Bidirectional-Channel Based
Wireless Networks,” in IEEE INFOCOM, Turin, Apr. 2013, pp. 3120–3128.
[108] H. ElSawy, E. Hossain, and S. Camorlinga, “Traffic Offloading Techniques in Two-Tier
Femtocell Networks,” in IEEE Conf. on Commun. (ICC), Budapest, Jun. 2013, pp. 6086–
6090.
[109] Q. Liu, G. Feng, and S. Qin, “Energy-Efficient Traffic Offloading in Macro-Pico Networks,”
in Wireless and Optical Communication Conference (WOCC), Chongqing, May 2013,
pp. 236–241.
[110] H. Holma and A. Toskala, Eds., WCDMA for UMTS: Radio Access for Third Generation
Mobile Communications, 3rd ed. Wiley, 2004.
[111] Joint Research Center, “Laboratory Tests on the Coexistence between 4G TDD and 3G/4G
FDD in the 1900–1920 MHz Spectrum Band,” Report EUR 25760 EN, 2012.
[112] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA); Base Station (BS) Radio
Transmission and Reception,” 3GPP TS 36.104 version 12.7.0 (Release 12), Sep. 2014.
[113] F. Schaich, “Filterbank Based Multi Carrier Transmission (FBMC) — Evolving OFDM:
FBMC in the Context of WiMAX,” in European Wireless Conference (EW), Lucca, Apr.
2010, pp. 1051–1058.
[114] N. Michailow, I. Gaspar, S. Krone, M. Lentmaier, and G. Fettweis, “Generalized Frequency
Division Multiplexing: Analysis of an Alternative Multi-Carrier Technique for Next
Generation Cellular Systems,” in International Symposium on Wireless Communication
Systems (ISWCS), Paris, Aug. 2012, pp. 171–175.
[115] G. Berardinelli, K. Pajukoski, E. Lahetkangas, R. Wichman, O. Tirkkonen, and P. Mogensen,
“On the Potential of OFDM Enhancements as 5G Waveforms,” in IEEE Vehicular
Technology Conference (VTC-Spring), Seoul, May 2014, pp. 1–5.
[116] C. Boyd, R.-A. Pitaval, O. Tirkkonen, and R. Wichman, “On the Time-Frequency
Localisation of 5G Candidate Waveforms,” in IEEE International Workshop on Signal
Processing Advances in Wireless Communications (SPAWC), Stockholm, Jun. 2015, pp.
101–105.
[117] G. B. Folland and A. Sitaram, “The Uncertainty Principle: A Mathematical Survey,”
J. Fourier Anal. and Appl., vol. 3, no. 3, pp. 207–238, May 1997.
[118] N. Michailow, M. Matthe, I. S. Gaspar, A. N. Caldevilla, L. L. Mendes, A. Festag, and
G. Fettweis, “Generalized Frequency Division Multiplexing for 5th Generation Cellular
Networks,” IEEE Trans. Commun., vol. 62, no. 9, pp. 3045–3061, Sep. 2014.
[119] B. M. Alves, L. L. Mendes, D. A. Guimaraes, and I. S. Gaspar, “Performance of GFDM over
Frequency-Selective Channels,” in Proc. Int. Workshop Telecommun., May 2013.
[120] W. Chung, D. Hong, T. Riihonen, and R. Wichman, “Interference Cancellation Architecture
for Full-Duplex System with GFDM Signaling,” in European Signal Processing Conference
(EUSIPCO), Aug. 2016.
[121] P. Persson, M. Coldrey, A. Wolfgang, and P. Bohlin, “Design and Evaluation of a 2 x 2
MIMO Repeater,” in Proc. European Conference on Antennas and Propagation (EUCAP),
Berlin, Mar. 2009, pp. 1509–1512.
[122] K. Haneda, E. Kahra, S. Wyne, C. Icheln, and P. Vainikainen, “Measurement of Loop-Back
Interference Channels for Outdoor-to-Indoor Full-Duplex Radio Relays,” in Proc. European
Conference on Antennas and Propagation (EUCAP), Barcelona, Spain, Apr. 2010, pp. 1–5.
[123] T. Riihonen, A. Balakrishnan, K. Haneda, S. Wyne, S. Werner, and R. Wichman, “Optimal
Eigenbeamforming for Suppressing Self-Interference in Full-Duplex MIMO Relays,” in
IEEE Annual Conference on Information Sciences and Systems (CISS), Baltimore, MD, Mar.
2011, pp. 1–6.
[124] M. Duarte, C. Dick, and A. Sabharwal, “Experiment-Driven Characterization of Full-Duplex
Wireless Systems,” IEEE Trans. Wireless Commun., vol. 11, no. 12, pp. 4296–4307, Dec.
2012.
[125] N. Merhav, G. Kaplan, A. Lapidoth, and S. S. Shitz, “On Information Rates for Mismatched
Decoders,” IEEE Transactions on Information Theory, vol. 40, no. 6, pp. 1953–1967, Nov.
1994.
[126] M. Vehkaperä, T. Riihonen, and R. Wichman, “Asymptotic Analysis of Full-Duplex
Bidirectional MIMO Link with Transmitter Noise,” in 2013 IEEE 24th Annual International
Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), Sep. 2013,
pp. 1265–1270.
[127] H. T. Friis, “Noise Figures of Radio Receivers,” Proc IRE, vol. 32, no. 7, pp. 419–422, Jul.
1944.
[128] T. Riihonen and R. Wichman, “Analog and Digital Self-Interference Cancellation in Full-
Duplex MIMO-OFDM Transceivers with Limited Resolution in A/D Conversion,” in Proc.
Asilomar Conference on Signals, Systems and Computers, Pacific Grove, CA, Nov. 2012,
pp. 45–49.
[129] M. Heino, D. Korpi, T. Huusari, E. Antonio-Rodriguez, S. Venkatasubramanian, T. Riihonen,
L. Anttila, C. Icheln, K. Haneda, R. Wichman, and M. Valkama, “Recent Advances in
Antenna Design and Interference Cancellation Algorithms for In-Band Full Duplex Relays,”
IEEE Commun. Mag., vol. 53, no. 5, pp. 91–101, May 2015.
[130] B. Debaillie, D.-J. van den Broek, C. Lavin, B. van Liempd, E.-A.-M. Klumperink,
C. Palacios, J. Craninckx, and B. Nauta, “RF Self-Interference Reduction Techniques for
Compact Full-Duplex Radios,” in Proc. Veh. Tech. Conf. (VTC-Spring), Glasgow, Scotland,
May 2015, pp. 1–6.
[131] B. Razavi, “Design Considerations for Direct-Conversion Receivers,” IEEE Trans. Circuits
Syst. II, vol. 44, no. 6, pp. 428–435, Jun. 1997.
[132] T. Riihonen, P. Mathecken, and R. Wichman, “Effect of Oscillator Phase Noise and
Processing Delay in Full-Duplex OFDM Repeaters,” in Proceedings of Asilomar Conference
on Signals, Systems and Computers (ASILOMAR), Pacific Grove, CA, Nov. 2012, pp. 1947–
1951.
[133] V. Syrjälä, M. Valkama, L. Anttila, T. Riihonen, and D. Korpi, “Analysis of Oscillator Phase-
Noise Effects on Self-Interference Cancellation in Full-Duplex OFDM Radio Transceivers,”
IEEE Trans. Wireless Commun., vol. 13, no. 6, pp. 2977–2990, Jun. 2014.
[134] A. Sahai, G. Patel, C. Dick, and A. Sabharwal, “On the Impact of Phase Noise on Active
Cancelation in Wireless Full-Duplex,” IEEE Trans. Veh. Technol., vol. 62, no. 9, pp. 4494–
4510, Sep. 2013.
[135] B. Widrow and I. Kollár, Quantization Noise: Roundoff Error in Digital Computation, Signal
Processing, Control, and Communications. Cambridge, UK: Cambridge University Press,
2008.
[136] B. P. Day, A. R. Margetts, D. W. Bliss, and P. Schniter, “Full-Duplex MIMO Relaying:
Achievable Rates under Limited Dynamic Range,” IEEE J. Select. Areas Commun., vol. 30,
no. 8, pp. 1541–1553, Sep. 2012.
[137] E. Antonio-Rodriguez, R. Lopez-Valcarce, T. Riihonen, S. Werner, and R. Wichman,
“Subspace-Constrained SINR Optimization in MIMO Full-Duplex Relays Under Limited
Dynamic Range,” in Proc. of IEEE Workshop on Signal Processing Advances in Wireless
Commun. (SPAWC), Stockholm, Jun. 2015, pp. 281–285.
[138] P. Kenington, High-Linearity RF Amplifier Design. Artech House, 2000.
[139] G. T. Zhou, H. Qian, L. Ding, and R. Raich, “On the Baseband Representation of a Bandpass
Nonlinearity,” IEEE Trans. Signal Processing, vol. 53, no. 8, pp. 2953–2957, Aug. 2005.
[140] R. Raich, H. Qian, and G. T. Zhou, “Orthogonal Polynomials for Power Amplifier Modeling
and Predistorter Design,” IEEE Trans. Veh. Technol., vol. 53, no. 5, pp. 1468–1479, May
2004.
[141] A. A. M. Saleh, “Frequency-Independent and Frequency-Dependent Nonlinear Models of
TWT Amplifiers,” IEEE Trans. Commun., vol. 29, no. 11, pp. 1715–1720, Nov. 1981.
[142] J. Minkoff, “The Role of AM-to-PM Conversion in Memoryless Nonlinear Systems,” IEEE
Trans. Commun., vol. 33, no. 2, pp. 139–144, Feb. 1985.
[143] D. Dardari, V. Tralli, and A. Vaccari, “A Theoretical Characterization of Nonlinear Distortion
Effects in OFDM Systems,” IEEE Trans. Commun., vol. 48, no. 10, pp. 1755–1764, Oct.
2000.
[144] G. González, F. Gregorio, J. Cousseau, T. Riihonen, and R. Wichman, “Performance Analysis
of Full-Duplex AF Relaying with Transceiver Hardware Impairments,” in European Signal
Processing Conference, Oulu, Finland, May 2016, pp. 1–5.
[145] G. Karam and H. Sari, “Analysis of Predistortion, Equalization, and ISI Cancellation
Techniques in Digital Radio Systems with Nonlinear Transmit Amplifiers,” IEEE Trans.
Commun., vol. 37, no. 12, pp. 1245–1253, Dec. 1989.
[146] L. Ding, G. Zhou, D. Morgan, Z. Ma, J. Kenney, J. Kim, and C. Giardina, “A Robust Digital
Baseband Predistorter Constructed Using Memory Polynomials,” IEEE Trans. Commun.,
vol. 52, no. 1, pp. 159–165, Jan. 2004.
[147] M. Y. Cheong, S. Werner, M. Bruno, J. Figueroa, J. Cousseau, and R. Wichman, “Adaptive
Piecewise Linear Predistorters for Nonlinear Power Amplifiers with Memory,” IEEE Trans.
Circuits Syst. I, vol. 59, no. 7, pp. 1519–1532, Jul. 2012.
[148] L. Ding, R. Raich, and G. Zhou, “A Hammerstein Predistortion Linearization Design Based
on the Indirect Learning Architecture,” in IEEE International Conference on Acoustics,
Speech, and Signal Processing (ICASSP), Orlando, FL, May 2002, pp. 2689–2692.
[149] D. Korpi, T. Riihonen, V. Syrjälä, L. Anttila, M. Valkama, and R. Wichman, “Full-Duplex
Transceiver System Calculations: Analysis of ADC and Linearity Challenges,” IEEE Trans.
 Wireless Commun., vol. 13, no. 7, pp. 3821–3836, Jul. 2014.
[150] V. Gregers-Hansen, S. Brockett, and P. Cahill, “A Stacked A-to-D Converter for Increased
Radar Signal Processor Dynamic Range,” in Proc. Radar Conf., Atlanta, GA, 2001, pp.
169–174.
[151] V. Bringi and V. Chandrasekar, Polarimetric Doppler Weather Radar: Principles and
Applications. Cambridge University Press, 2001.
[152] E. Antonio-Rodriguez, R. Lopez-Valcarce, T. Riihonen, S. Werner, and R. Wichman, “SINR
Optimization in Wideband Full-Duplex MIMO Relays Under Limited Dynamic Range,” in
Proc. of IEEE Sensor Array and Multichannel Signal Processing Workshop, A Coruna, Jun.
2014, pp. 177–180.
[153] T. Riihonen, S. Werner, and R. Wichman, “Mitigation of Loopback Self-Interference in Full-
Duplex MIMO Relays,” IEEE Trans. Signal Processing, vol. 59, no. 12, pp. 5983–5993,
Dec. 2011.
[154] U. Ugurlu, T. Riihonen, and R. Wichman, “Optimized In-Band Full-Duplex MIMO Relay
Under Single-Stream Transmission,” IEEE Trans. Veh. Technol., vol. 65, no. 1, pp. 155–168,
Jan. 2016.
[155] R. Penrose, “A Generalized Inverse for Matrices,” Mathematical Proceedings of the
Cambridge Philosophical Society, vol. 51, no. 3, pp. 406–413, Jul 1955.
[156] A. Koohian, H. Mehrpouyan, A. A. Nasir, S. Durrani, M. Azarbad, and S. D. Blostein, “Blind
Channel Estimation in Full Duplex Systems: Identifiability Analysis, Bounds, and
Estimators,” submitted to IEEE Trans. Signal Processing, Nov. 2015.
[157] B. Hassibi and B. Hochwald, “How Much Training Is Needed in Multiple-Antenna Wireless
Links?” IEEE Trans. Inform. Theory, vol. 49, no. 4, pp. 951–963, 2003.
[158] E. Antonio-Rodriguez, R. Lopez-Valcarce, T. Riihonen, S. Werner, and R. Wichman,
“Wideband Full-Duplex MIMO Relays with Blind Adaptive Self-Interference Cancellation,”
Accepted by Signal Processing, Jun. 2016.
[159] P. Diniz, Adaptive Filtering: Algorithms and Practical Implementation, 4th ed. Springer,
2013.
[160] M. Steer, Microwave and RF Engineering: A Systems Approach. Scitech Publishing Inc.,
2010.
[161] F. O’Hara and G. Moore, “A High Performance CW Receiver Using Feedthru Nulling,”
Microwave Journal, p. 63, Sep. 1963.
[162] P. Beasley, A. Stove, B. Reits, and B. As, “Solving the Problems of a Single Antenna
Frequency Modulated CW Radar,” in IEEE International Radar Conference, Arlington, VA,
May 1990, pp. 391–395.
[163] J. Shapira and S. Y. Miller, CDMA Radio with Repeaters. Springer, 2007.
[164] H. Suzuki, K. Itoh, Y. Ebine, and M. Sato, “A Booster Configuration with Adaptive
Reduction of Transmitter-Receiver Antenna Coupling for Pager Systems,” in IEEE
Vehicular Technology Conference (VTC-Fall), Amsterdam, Sep. 1999, pp. 1516 –1520.
[165] H. Hamazumi, K. Imamura, N. Iai, K. Shibuya, and M. Sasaki, “A Study of a Loop
Interference Canceller for the Relay Stations in an SFN for Digital Terrestrial Broadcasting,”
in IEEE Global Telecommunications Conference (GLOBECOM), San Francisco, CA, Nov.–
Dec. 2000, pp. 167–171.
[166] LTE spec, 3GPP TS 36.216 , “Physical Layer for Relaying Operation (release 10),” Jun 2011.
[167] N. Alliance, “NGMN 5G White Paper,” Feb. 2015.
[168] E. Aryafar, M. Khojastepour, K. Sundaresan, S. Rangarajan, and M. Chiang, “MIDU:
Enabling MIMO Full Duplex,” in Proceedings of ACM MobiCom, Istanbul, Turkey, Aug.
2012, pp. 257–268.
[169] E. Everett, M. Duarte, C. Dick, and A. Sabharwal, “Empowering Full-Duplex Wireless
Communication by Exploiting Directional Diversity,” in Proc. IEEE Annual Asilomar
Conference on Signals, Systems, and Computers (ASILOMAR), Pacific Grove, CA, Nov.
2011, pp. 2002–2006.
[170] B. Lindmark and M. Nilsson, “On the Available Diversity Gain from Different Dual-
Polarized Antennas,” IEEE J. Select. Areas Commun., vol. 19, no. 2, pp. 287–294, Feb.
2001.
[171] P. Lindberg and E. Ojefors, “A Bandwidth Enhancement Technique for Mobile Handset
Antennas Using Wavetraps,” IEEE Trans. Antennas Propagat., vol. 54, no. 8, pp. 2226–
2233, Aug. 2006.
[172] M. Knox, “Single Antenna Full Duplex Communications Using a Common Carrier,” in Proc.
of IEEE Annual Wireless and Microwave Tech. Conf., Cocoa Beach, FL, Apr. 2012, pp. 1–6.
[173] M. Heino, S. N. Venkatasubramanian, C. Icheln, and K. Haneda, “Design of Wavetraps for
Isolation Improvement in Compact In-Band Full-Duplex Relay Antennas,” IEEE Trans.
Antennas Propagat., vol. 64, no. 3, pp. 1061–1070, Mar. 2016.
[174] D. Korpi, M. Heino, C. Icheln, K. Haneda, and M. Valkama, “Compact Inband Full-Duplex
Relays with Beyond 100 dB Self-Interference Suppression: Enabling Techniques and Field
Measurements,” IEEE Trans. Antennas Propagat., submitted for review, 2016.
[175] Y.-S. Choi and H. Shirani-Mehr, “Simultaneous Transmission and Reception: Algorithm,
Design and System Level Performance,” IEEE Trans. Wireless Commun., vol. 12, no. 12,
pp. 5992–6010, Dec. 2014.
[176] B. Yang, Y. Dong, Z. Yu, and J. Zhou, “An RF Self-Interference Cancellation Circuit for the
Full-Duplex Wireless Communications,” in Proceedings of the International Symposium on
Antennas Propagation (ISAP), Nanjing, China, Oct. 2013, pp. 1048–1051.
[177] N. Phungamngern, P. Uthansakul, and M. Uthansakul, “Digital and RF Interference
Cancellation for Single-Channel Full-Duplex Transceiver Using a Single Antenna,” in IEEE
International Conference on Electrical Engineering/Electronics, Computer,
Telecommunications and Information Technology (ECTI-CON), Krabi, May 2013, pp. 1–5.
[178] L. Laughlin, M. Beach, K. Morris, and J. Haine, “Electrical Balance Duplexing for Small
Form Factor Realization of In-Band Full Duplex,” IEEE Commun. Mag., vol. 53, no. 5, pp.
102–110, May 2015.
[179] N. A. Estep, D. L. Sounas, J. Soric, and A. Alu, “Magnetic-Free Non-Reciprocity and
Isolation Based on Parametrically Modulated Coupled-Resonator Loops,” Nat Phys, vol. 10,
no. 12, pp. 923–927, Dec. 2014.
[180] M. Mikhemar, H. Darabi, and A. Abidi, “A Tunable Integrated Duplexer with 50dB Isolation
in 40nm CMOS,” in IEEE International Solid-State Circuits Conference – Digest of
Technical Papers, San Francisco, CA, Feb. 2009, pp. 386–387.
[181] A. Kumar, S. Aniruddhan, and R. K. Ganti, “Directional Coupler with High Isolation
Bandwidth Using Electrical Balance,” in IEEE MTT-S International Microwave Symposium
(IMS), Tampa, FL, Jun. 2014, pp. 1–3.
[182] J.-I. Choi, M. Jain, K. Srinivasan, P. Levis, and S. Katti, “Achieving Single Channel, Full
Duplex Wireless Communication,” in Proc. Annual Int. Conf. Mobile Computing and Netw.,
Chicago, IL, Sep. 2010, pp. 1–12.
[183] “Xinger XDL15-2-020S Delay Line,” Anaren, Data sheet revision B.
[184] T. Laakso, V. Välimäki, M. Karjalainen, and U. Laine, “Splitting the Unit Delay [FIR/All
Pass Filters Design],” IEEE Signal Processing Mag., vol. 13, no. 1, pp. 30–60, Jan. 1996.
[185] T. Huusari, Y.-S. Choi, P. Liikkanen, D. Korpi, S. Talwar, and M. Valkama, “Wideband Self-
Adaptive RF Cancellation Circuit for Full-Duplex Radio: Operating Principle and
Measurements,” in Proc. Veh. Tech. Conf. (VTC-Spring), Glasgow, Scotland, May 2015, pp.
1–7.
[186] B. Debaillie, D.-J. van den Broek, C. Lavin, B. van Liempd, E. Klumperink, C. Palacios,
J. Craninckx, B. Nauta, and A. Parssinen, “Analog/RF Solutions Enabling Compact Full-
Duplex Radios,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1662–1673, Sep. 2014.
[187] D. Yang and A. Molnar, “A Widely Tunable Active Duplexing Transceiver with Same-
Channel Concurrent RX/TX and 30dB RX/TX Isolation,” in IEEE Radio Frequency
Integrated Circuits Symposium, Tampa, FL, Jun. 2014, pp. 321–324.
[188] J. Zhou and H. Krishnaswamy, “Recent Developments in Fully-Integrated RF Self-
Interference Cancellation for Frequency-Division and Full-Duplex Radios,” in IEEE
Vehicular Technology Conference (VTC-Spring), Glasgow, May 2015, pp. 1–5.
[189] D. Korpi, L. Anttila, V. Syrjälä, and M. Valkama, “Widely Linear Digital Self-Interference
Cancellation in Direct-Conversion Full-Duplex Transceiver,” IEEE J. Select. Areas
 Commun., vol. 32, no. 9, pp. 1674–1687, Sep. 2014.
[190] L. Anttila, D. Korpi, V. Syrjärä, and M. Valkama, “Cancellation of Power Amplifier Induced
Nonlinear Self-Interference in Full-Duplex Transceivers,” in Proc. Asilomar Conf. Signals,
Systems and Computers, Pacific Grove, CA, Nov. 2013, pp. 1193 – 1198.
[191] L. Anttila, D. Korpi, E. Antonio-Rodriquez, R. Wichman, and M. Valkama, “Modeling and
Efficient Cancellation of Nonlinear Self-Interference in MIMO Full-Duplex Transceivers,”
in Proc. IEEE Globecom Workshops, Austin, TX, Dec. 2014, pp. 862–868.
[192] H. Tehrani, A. Cao, S. Afsardoost, T. Eriksson, M. Isaksson, and C. Fager, “A Comparative
Analysis of the Complexity/Accuracy Tradeoff in Power Amplifier Behavioral Models,”
IEEE Trans. Microwave Theory Tech., vol. 58, no. 6, pp. 1510–1520, Jun. 2010.
[193] A. Balatsoukas-Stimming, A. C. Austin, P. Belanovic, and A. Burg, “Baseband and RF
Hardware Impairments in Full-Duplex Wireless Systems: Experimental Characterisation and
Suppression,” EURASIP Journal on Wireless Communications and Networking, vol. 2015,
no. 1, pp. 1–11, 2015.
[194] R. Askar, T. Kaiser, B. Schubert, T. Haustein, and W. Keusgen, “Active Self-Interference
Cancellation Mechanism for Full-Duplex Wireless Transceivers,” in IEEE International
Conference on Cognitive Radio Oriented Wireless Networks and Communications
(CROWNCOM), Oulu, Jun. 2014, pp. 539–544.
[195] A. Sahai, G. Patel, and A. Sabharwal, “Phase Noise: Understanding the Bottleneck in Full-
duplex Designs,” in Proceedings of Asilomar Conference on Signals, Systems and
Computers, 2012.
[196] D. H. Mahrof, E. A. M. Klumperink, M. S. O. Alink, and B. Nauta, “A Receiver With In-
Band IIP3 20dBm, exploiting Cancelling of OpAmp Finite-Gain-Induced Distortion via
Negative Conductance,” in 2013 IEEE Radio Frequency Integrated Circuits Symposium
(RFIC), Jun. 2013, pp. 85–88.
[197] B. P. Day, A. R. Margetts, D. W. Bliss, and P. Schniter, “Full-Duplex Bidirectional MIMO:
Achievable Rates under Limited Dynamic Range,” IEEE Trans. Signal Processing, vol. 60,
no. 7, pp. 3702–3713, Jul. 2012.
[198] M. Kiessling and J. Speidel, “Mutual Information of MIMO Channels in Correlated Rayleigh
Fading Environments – A General Solution,” in IEEE International Conference on
Communications (ICC), Jun. 2004, pp. 814–818.
[199] M. Zhou, H. Cui, L. Song, and Y. Li, “Is Full Duplex Configuration Better than MIMO
Spatial Multiplexing?” in IEEE/CIC International Conference on Communications in China
(ICCC), Shenzhen, China, Nov. 2015, pp. 1–5.
[200] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with Formulas,
Graphs, and Mathematical Tables. NewYork: Dover, 1970.
[201] A. Paulraj, R. Nabar, and D. Gore, Introduction to Space-Time Wireless Communications.
Cambridge University Press, 2003.
[202] M. Zhou, H. Cui, L. Song, and B. Jiao, “Transmit-Receive Antenna Pair Selection in Full
Duplex Systems,” IEEE Wireless Commun. Lett., vol. 3, no. 1, pp. 34–37, Jan. 2014.
[203] M. Zhou, L. Song, Y. Li, and X. Li, “Simultaneous Bidirectional Link Selection in Full
Duplex MIMO Systems,” IEEE Trans. Wireless Commun., vol. 14, no. 7, pp. 4052–4062,
Jul. 2015.
[204] L. Song, “Relay Selection for Two-Way Relaying with Amplify-and-Forward Protocols,”
IEEE Trans. Veh. Technol., vol. 60, no. 4, pp. 1954–1959, Apr. 2011.
[205] H. A. David and H. N. Nagaraja, Order Statistics. Wiley Online Library, 1970.
[206] A. Sabharwal, P. Schniter, D. Guo, D. W. Bliss, S. Rangarajan, and R. Wichman, “In-Band
Full-Duplex Wireless: Challenges and Opportunities,” IEEE J. Select. Areas Commun.,
vol. 32, no. 9, pp. 1637–1652, Sep. 2014.
[207] I. Krikidis, H. Suraweera, P. Smith, and C. Yuen, “Full-Duplex Relay Selection for Amplify-
and-Forward Cooperative Networks,” IEEE Trans. Wireless Commun., vol. 11, no. 12, pp.
4381–4393, Dec. 2012.
[208] A. Jeffrey and D. Zwillinger, Table of Integrals, Series, and Products. Academic Press, 2007.
[209] W. Yu and R. Lui, “Dual Methods for Nonconvex Spectrum Optimization of Multicarrier
Systems,” IEEE Trans. Wireless Commun., vol. 54, no. 7, pp. 1310–1322, Jul. 2006.
[210] C. Y. Wong, R. S. Cheng, K. B. Lataief, and R. D. Murch, “Multiuser OFDM with Adaptive
Subcarrier, Bit, and Power Allocation,” IEEE J. Select. Areas Commun., vol. 17, no. 10, pp.
1747–1758, Oct. 1999.
[211] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, Mar. 2004.
[212] N. Z. Shor, K. C. Kiwiel, and A. Ruszcaynski, Minimization Methods for Non-Differentiable
Functions. New York: Springer, 1985.
[213] A. Beck, A. Ben-Tal, and L. Tetruashvili, “A Sequential Parametric Convex Approximation
Method with Applications to Nonconvex Truss Topology Design Problems,” Journal of
Global Optimization, vol. 47, no. 1, pp. 29–51, May 2010.
[214] W.-L. Li, Y. Zhang, A.-C. So, and M. Z. Win, “Slow Adaptive OFDMA Systems through
Chance Constrained Programming,” IEEE Trans. Signal Processing, vol. 58, no. 7, pp.
3858–3869, Jul. 2010.
[215] A.-C. So and Y. J. Zhang, “Distributionally Robust Slow Adaptive OFDMA with Soft QoS
via Linear Programming,” IEEE J. Select. Areas Commun., vol. 31, no. 5, pp. 947–958, May
2013.
[216] “TR 36.814-further Advancements for E-UTRA: Physical Layer Aspects (Release 9),” in 3rd
Generation Partnership Project Tech. Rep., Tech. Rep. TSGRA Network, 2010.
[217] “Guidelines for Evaluation of Radio Transmission Technologies for IMT-2000,” in
International Telecommunication Union. ITU-R.M Recommendation, 1997.
[218] D. Nguyen, L.-N. Tran, P. Pirinen, and M. Latva-aho, “On the Spectral Efficiency of Full-
Duplex Small Cell Wireless Systems,” IEEE Trans. Wireless Commun., vol. 13, no. 9, pp.
4896–4910, Sep. 2014.
[219] Y. J. Zhang and K. B. Letaief, “Multiuser Adaptive Subcarrier-and-Bit Allocation with
Adaptive Cell Selection for OFDM Systems,” IEEE Trans. Wireless Commun., vol. 3, no. 5,
pp. 1566–1575, Sep. 2004.
[220] J. Huang, V. G. Subramanian, R. Agrawal, and R. Berry, “Joint Scheduling and Resource
Allocation in Uplink OFDM Systems for Broadband Wireless Access Networks,” IEEE J.
Select. Areas Commun., vol. 27, no. 2, pp. 226–234, Feb. 2009.
[221] V. K. Lau, W. K. Ng, and D. S. W. Hui, “Asymptotic Tradeoff between Cross-Layer Goodput
Gain and Outage Diversity in OFDMA Systems with Slow Fading and Delayed CSIT,”
IEEE Trans. Wireless Commun., vol. 7, no. 7, pp. 2732–2739, Jul. 2008.
[222] Y. Huang and B. D. Rao, “Performance Analysis of Heterogeneous Feedback Design in an
OFDMA Downlink with Partial and Imperfect Feedback,” IEEE Trans. Signal Processing,
vol. 61, no. 4, pp. 1033–1046, Nov. 2013.
[223] B. Di, S. Bayat, L. Song, and Y. Li, “Radio Resource Allocation for Full-Duplex OFDMA
Networks Using Matching Theory,” in IEEE Conference on Computer Communications
Workshops (INFOCOM WKSHPS), Toronto, Canada, Apr. 2014, pp. 197–198.
[224] M. Sipser, Introduction to the Theory of Computation. Cengage Learning, 2012.
[225] V. Danilov, “Existence of Stable Matchings in Some Three-Sided Systems,” Mathematical
Social Sciences, vol. 46, no. 2, pp. 145–148, Oct. 2003.
[226] P. S. K. Eriksson, J. Sjostrand, “Three Dimensional Stable Matching with Cyclic
Preferences,” Mathematical Social Sciences, vol. 52, no. 1, pp. 77–87, Jul. 2006.
[227] T. Huang, “Two’s Company, Three’s a Crowd: Stable Family and Threesome Roommates
Problems,” Algorithms’ ESA 2007, the series Lecture Notes in Computer Science, vol. 4698,
pp. 558–569, Oct. 2007.
[228] D. Manlove, Algorithmics of Matching under Preferences. World Scientific, 2013.
[229] D. H. C. Ng, “Three-Dimensional Stable Matching Problems,” SIAM J. Discrete
Mathematics, vol. 2, pp. 245–252, 1991.
[230] S. M. K. Iwama, “A Survey of the Stable Marriage Problem and Its Variants,” in Int. Conf.
Informatics Education and Research for Knowledge-Circulating Society, Kyoto, Japan, Jan.
2008, pp. 131–136.
[231] e. a. E. M. Arkina, “Geometric Stable Roommates,” Info. Process. Lett., vol. 4, pp. 219–224,
2009.
[232] M. A. Roth, Digital CommunicationsTwo-Sided Matching: A Study in Game-Theoretic
Modeling and Analysis. McGraw Hill, fourth edition, Cambridge University Press, 1992.
[233] A. Subramanian, “A New Approach to Stable Matching Problems,” SIAM J. Comput.,
vol. 23, pp. 671–700, 1994.
[234] W. R. R. Ostrovsky, “It’s Not Easy Being Three: The Approximability of Three-Dimensional
Stable Matching Problems,” http://arxiv.org/abs/1412.1130, 2014.
[235] H. ElSawy, E. Hossain, and D. I. Kim, “HetNets with Cognitive Small Cells: User Offloading
and Distributed Channel Access Techniques,” IEEE Commun. Mag., vol. 51, pp. 28–36, Jun.
2013.
[236] A. Tulino, A. Lozano, and S. Verdu, “Impact of Antenna Correlation on the Capacity of
Multiantenna Channels,” IEEE Trans. Inform. Theory, vol. 51, no. 7, pp. 2491–2509, Jul.
2005.
[237] Y. El Hajj Shehadeh and S. Sezginer, “Fast Varying Channel Estimation in Downlink LTE
Systems,” in IEEE International Symposium on Personal Indoor and Mobile Radio
Communications (PIMRC), Turkey, Sep. 2010, pp. 608–613.
[238] H. Sahlin and A. Persson, “Aspects of MIMO Channel Estimation for LTE Uplink,” in IEEE
Vehicular Technology Conference (VTC-Fall), San Francisco, CA, Sep. 2011, pp.1–5.
[239] X. Xia, H. Zhao, and C. Zhang, “Improved SRS Design and Channel Estimation for LTE-
Advanced Uplink,” in IEEE International Symposium on Microwave, Antenna, Propagation
and EMC Technologies for Wireless Communications (MAPE), China, Oct. 2013, pp. 84–90.
[240] Q. Li, Y. Wu, S. Feng, P. Zhang, L. Xue, and Y. Zhou, “On Channel Estimation for Multi-
User MIMO in LTE-A Uplink,” in IEEE Vehicular Technology Conference (VTC-Spring),
Seoul, Korea, May 2014, pp. 1–5.
[241] K. Akcapinar and O. Gurbuz, “Full-duplex Bidirectional Communication under Self-
Interference,” in IEEE International Conference on Telecommunications (ConTEL), Graz,
Austria, Jul. 2015, pp. 1–7.
[242] Y. L. Mischa Dohler, Cooperative Communications: Hardware, Channel and PHY. Wiley and
Sons, 2010.
[243] Y. Li and B. Vucetic, “On the Performance of a Simple Adaptive Relaying Protocol for
Wireless Relay Networks,” in IEEE Vehicular Technology Conference (VTC-Spring),
Singapore, May 2008, pp. 2400–2405.
[244] S. Tian, Y. Li, and B. Vucetic, “Piecewise-and-Forward Relaying in Wireless Relay
Networks,” IEEE Signal Processing Lett., vol. 18, no. 5, pp. 323–326, May 2011.
[245] Y. Li, B. Vucetic, T. Wong, and M. Dohler, “Distributed Turbo Coding with Soft Information
Relaying in Multihop Relay Networks,” IEEE J. Select. Areas Commun., vol. 24, no. 11, pp.
2040–2050, Nov. 2006.
[246] H. Zhao and W. Su, “Cooperative Wireless Multicast: Performance Analysis and
Power/Location Optimization,” IEEE Trans. Wireless Commun., vol. 9, no. 6, pp. 2088–
2100, Jun. 2010.
[247] C. Yao, K. Yang, L. Song, and Y. Li, “X-Duplex: Adapting of Full-Duplex and Half-Duplex,”
in IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS),
Hong Kong, Apr. 2015, pp. 55–56.
[248] D. Bharadia and S. Katti, “Full Duplex MIMO Radios,” Usenix Conference on Networked
Systems Design and Implementation, vol. 2014, pp. 359–372, 2014.
[249] T. Riihonen, S. Werner, R. Wichman, and Z. Eduardo, “On the Feasibility of Full-Duplex
Relaying in the Presence of Loop Interference,” in IEEE Workshop on Signal Processing
Advances in Wireless Communications (SPAWC), Perugia, Italy, Jun. 2009, pp. 275–279.
[250] D. Zwillinger, Table of Integrals, Series, and Products. Elsevier, 2014.
[251] R. Narasimhan, A. Ekbal, and J. Cioffi, “Finite-SNR Diversity-Multiplexing Tradeoff of
Space-Time Codes,” in IEEE International Conference on Communications (ICC), vol. 1,
May 2005, pp. 458–462.
[252] K. Yang, H. Cui, L. Song, and Y. Li, “Efficient Full-Duplex Relaying With Joint Antenna-
Relay Selection and Self-Interference Suppression,” IEEE Trans. Wireless Commun.,
vol. 14, no. 7, pp. 3991–4005, Jul. 2015.
[253] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products. Academic Press,
1994.
[254] I. Krikidis, H. Suraweera, S. Yang, and K. Berberidis, “Full-Duplex Relaying over Block
Fading Channel: A Diversity Perspective,” IEEE Trans. Wireless Commun., vol. 11, no. 12,
pp. 4524–4535, Dec. 2012.
[255] A. Goldsmith, Wireless Communications. UK: Cambridge University Press, 2005.
[256] H. Suraweera, P. Smith, and M. Shafi, “Capacity Limits and Performance Analysis of
Cognitive Radio With Imperfect Channel Knowledge,” IEEE Trans. Veh. Technol., vol. 59,
no. 4, pp. 1811–1822, May 2010.
[257] D. Mitrinovic and J. D. Keckic, The Cauchy Method of Residues: Theory and Applications.
Springer, 1984.
[258] H. Suraweera, G. Karagiannidis, Y. Li, H. Garg, A. Nallanathan, and B. Vucetic, “Amplify-
and-Forward Relay Transmission with End-to-End Antenna Selection,” in IEEE Wireless
Communications and Networking Conference (WCNC), Sydney, NSW, Apr. 2010, pp. 1–6.
[259] FCC, “Report of the Spectrum Efficiency Group,” FCC Spectrum Policy Task Force, Tech.
Rep., Nov 2002, http://www.fcc.gov/sptf/files/SEWGFinalReport_1.pdf.
[260] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “Next Generation/Dynamic
Spectrum Access/Cognitive Radio Wireless Networks: A Survey,” Computer Networks,
vol. 50, no. 13, pp. 2127–2159, 2006.
[261] W. Gardner et al., “Signal Interception: A Unifying Theoretical Framework for Feature
Detection,” IEEE Trans. Commun., vol. 36, no. 8, pp. 897–906, 1988.
[262] M. Öner and F. Jondral, “Air Interface Recognition for A Software Radio System Exploiting
Cyclostationarity,” in IEEE International Symposium on Personal, Indoor and Mobile Radio
Communications (PIMRC), Barcelona, Spain, Sep. 2004, pp. 1947–1951.
[263] J. Lundén, V. Koivunen, A. Huttunen, and H. V. Poor, “Spectrum Sensing in Cognitive Radios
Based on Multiple Cyclic Frequencies,” in IEEE International Conference on Cognitive
Radio Oriented Wireless Networks and Communications (CrownCom), Orlando, FL, Aug.
2007, pp. 37–43.
[264] B. Wang and K. Liu, “Advances in Cognitive Radio Networks: A Survey,” IEEE J. Select.
Topics in Signal Processing, vol. 5, no. 1, pp. 5–23, 2011.
[265] Q. Zhao, L. Tong, A. Swami, and Y. Chen, “Decentralized Cognitive MAC for Opportunistic
Spectrum Access in Ad Hoc Networks: A POMDP Framework,” IEEE J. Select. Areas
Commun., vol. 25, no. 3, pp. 589–600, 2007.
[266] T. Yücek and H. Arslan, “A Survey of Spectrum Sensing Algorithms for Cognitive Radio
Applications,” IEEE Commun. Surveys & Tutorials, vol. 11, no. 1, pp. 116–130, 2009.
[267] Y. C. Liang, Y. Zeng, E. C. Y. Peh, and A. T. Hoang, “Sensing-Throughput Tradeoff for
Cognitive Radio Networks,” IEEE Trans. Wireless Commun., vol. 7, no. 4, pp. 1326–1337,
Apr 2008.
[268] S. Huang, X. Liu, and Z. Ding, “Opportunistic Spectrum Access in Cognitive Radio
Networks,” in IEEE INFOCOM, Phoenix, AZ, Apr. 2008, pp. 2101–2109.
[269] Q. Zhao, S. Geirhofer, L. Tong, and B. M. Sadler, “Optimal Dynamic Spectrum Access via
Periodic Channel Sensing,” in IEEE Wireless Communications and Networking Conference
(WCNC), Kowloon, Mar. 2007, pp. 33–37.
[270] W. Afifi, A. Sultan, and M. Nafie, “Adaptive Sensing and Transmission Durations for
Cognitive Radios,” in IEEE Symposium on New Frontiers in Dynamic Spectrum Access
Networks (DySPAN), Aachen, May 2011, pp. 380–388.
[271] Y. Liao, T. Wang, L. Song, and Z. Han, “Listen-and-talk: Full-Duplex Cognitive Radio
Networks,” in IEEE Global Communications Conference (GLOBECOM), Austin, TX, Dec.
2014, pp. 3068–3073.
[272] Y. Liao, L. Song, Z. Han, and Y. Li, “Full Duplex Cognitive Radio: A New Design Paradigm
for Enhancing Spectrum Usage,” IEEE Commun. Mag., vol. 53, no. 5, pp. 138–145, May
2015.
[273] E. Everett, A. Sahai, and A. Sabharwal, “Passive Self-Interference Suppression for Full-
Duplex Infrastructure Nodes,” IEEE Trans. Wireless Commun., vol. 13, no. 2, pp. 680–694,
Feb. 2014.
[274] H. Kim and K. G. Shin, “Efficient Discovery of Spectrum Opportunities with MAC-Layer
Sensing in Cognitive Radio Networks,” IEEE Trans. Mobile Comput., vol. 7, no. 5, pp. 533–
545, May 2008.
[275] A. Motamedi and A. Bahai, “MAC Protocol Design for Spectrum-Agile Wireless Networks:
Stochastic Control Approach,” in IEEE International Symposium on New Frontiers in
Dynamic Spectrum Access Networks, Dublin, Apr. 2007, pp. 448–451.
[276] M. R. Garey and D. S. Johnson, “Computers and Intractability: A Guide to the Theory of NP-
Completeness,” San Francisco, LA: Freeman, 1979.
[277] H. W. Kuhn, “The Hungarian Method for the Assignment Problem,” Naval Research
Logistics Quarterly, vol. 2, no. 1-2, pp. 83–97, 1955.
[278] J. E. Hopcroft and R. M. Karp, “An Algorithm for Maximum Matchings in Bipartite
Graphs,” SIAM Journal on Computing, vol. 2, no. 4, pp. 225–231, 1973.
[279] X. Zhang and K. G. Shin, “Enabling Coexistence of Heterogeneous Wireless Systems: Case
for ZigBee and WiFi,” in Proceedings of ACM International Symposium on Mobile Ad Hoc
Networking and Computing (MobiHoc), Paris, France, May 2011.
[280] “Report of the Spectrum Efficiency Group, FCC Spectrum Policy Task Force,” Report, Nov.
2002.
[281] M. M. Buddhikot, “Understanding Dynamic Spectrum Access: Models, Taxonomy and
Challenges,” in Proceedings of IEEE International Symposium on New Frontiers in
Dynamic Spectrum Access Networks (DySPAN), Dublin, Apr. 2007, pp. 649–663.
[282] R. J. Berger, “Open Spectrum: A Path to Ubiquitous Connectivity,” Queue, vol. 1, no. 3, pp.
60–68, May 2003.
[283] P. Kolodzy, “Spectrum Policy Task Force Report,” in FCC, Dec. 2002.
[284] J. M. Peha and S. Panichpapiboon, “Real-Time Secondary Markets for Spectrum,”
Telecommunications Policy, vol. 28, no. 7, pp. 603–618, Aug.-Sep. 2004.
[285] G. Hardin, “The Tragedy of the Commons,” Science, vol. 162, no. 3, pp. 1243–1248, Dec.
1968.
[286] G. Li, S. Srikanteswara, and C. Maciocco, “Interference Mitigation for WLAN Devices Using
Spectrum Sensing,” in Proceedings of IEEE Consumer Communications and Networking
Conference (CCNC), Las Vegas, NV, Jan. 2008, pp. 958–962.
[287] S. Srikanteswara, G. Li, and C. Maciocco, “Cross Layer Interference Mitigation Using
Spectrum Sensing,” in Proceedings of IEEE Global Telecommunications Conference
(GLOBECOM), Washington, DC, Nov. 2007, pp. 3553–3557.
[288] J. Brito, “The Spectrum Commons in Theory and Practice,” Stanford Technology Law
Review, 2007.
[289] W. Lehr and J. Crowcroft, “Managing Shared Access to a Spectrum Commons,” in
Proceedings of IEEE International Symposium on New Frontiers in Dynamic Spectrum
 Access Networks (DySPAN), Nov. 2005, pp. 420–444.
[290] P. Kyasanur and N. Vaidya, “Selfish MAC Layer Misbehavior in Wireless Networks,” IEEE
Trans. Mobile Comput., vol. 4, no. 5, pp. 502–516, Sep.–Oct. 2005.
[291] A. Toledo and X. Wang, “Robust Detection of Selfish Misbehavior in Wireless Networks,”
IEEE J. Select. Areas Commun., vol. 25, no. 6, pp. 1124–1134, Aug. 2007.
[292] J. Konorski, “A Game-Theoretic Study of CSMA/CA Under a Backoff Attack,” IEEE/ACM
Trans. Networking, vol. 14, no. 6, pp. 1167–1178, Dec. 2006.
[293] D. Niyato and E. Hossain, “Competitive Pricing for Spectrum Sharing in Cognitive Radio
Networks: Dynamic Game, Inefficiency of Nash Equilibrium, and Collusion,” IEEE J.
Select. Areas Commun., vol. 26, no. 1, pp. 192–202, Jan. 2008.
[294] B. Bing, “Measured Performance of the IEEE 802.11 Wireless LAN,” in Conference on Local
Computer Networks (LCN), Lowell, MA, Oct. 1999, pp. 34–42.
[295] W. Afifi and M. Krunz, “Exploiting Self-Interference Suppression for Improved Spectrum
Awareness/Efficiency in Cognitive Radio Systems,” in Proceedings IEEE INFOCOM,
Turin, Apr. 2013, pp. 1258–1266.
[296] H. V. Poor, An Introduction to Signal Detection and Estimation. Springer Science & Business
Media, 2013.
[297] G. Bianchi, “Performance Analysis of the IEEE 802.11 Distributed Coordination Function,”
IEEE J. Select. Areas Commun., vol. 18, no. 3, pp. 535–547, Mar. 2000.
[298] A. Sahai, G. Patel, and A. Sabharwal, “Pushing the Limits of Full-Duplex: Design and Real-
Time Implementation,” arXiv preprint arXiv:1107.0607, 2011.
[299] S. Goyal, P. Liu, O. Gurbuz, E. Erkip, and S. Panwar, “A Distributed MAC Protocol for Full
Duplex Radio,” in Asilomar Conference on Signals, Systems and Computers, Pacific Grove,
CA, Nov. 2013, pp. 788–792.
[300] R. Sultan, L. Song and Z. Han, “Impact of Full Duplex on Resource Allocation for Small Cell
Networks,” 2014 IEEE Global Conference on Signal and Information Processing
(GlobalSIP), Atlanta, GA, 2014, pp. 1257–1261.
[301] Y. Jiang, H. Chen, F. C. M. Lau, P. Wang, and Y. Li, “Full-Duplex OFDMA Multi-User
Cellular Systems: Resource Allocation and User Pairing,” Trans. Emerging Tel. Tech. doi:
10.1002/ett.3005.
Index

3GPP, 30
4G, 31
5G, 31
802.11, 4, 22, 38, 40, 42
802.15, 41
802.16, 21–23
802.20, 23
802.22, 11

Acess Point (AP), 35, 179

ACK, 15
Ad hoc network, 42–43
ADC, 2, 44
ALOHA, 42
Antenna selection, 244, 289
Audio on demand, 40
AWGN, 60, 111, 115, 139

Backhaul, 180
Balun, 46, 97
Base station (BS), 32, 138, 176
Baseband, 51, 60
Beacon-based discovery, 33
Beamforming, 4, 34, 137
BER, 5, 76, 83
Bluetooth, 41
BPSK, 38

Carrier sensing, 301, 305

CCK, 38, 39
CDF, 83, 126, 209
CCDF, 217
CDMA, 19
CDMA2000, 16–17
 IS-95, 17
MC-CDMA, 3
OFDM-CDMA, 3
TD-SCDMA, 18
WCDMA, 16–18
WCDMA-FDD, 50
Cellular mode, 33
Cellular network
3G, 16–18
4G, 18–23
Channel capacity, 10
Channel coding, 14
Channel estimation, 7, 189
Channel matrix, 56, 74, 111
Channel State Information (CSI), 31, 60, 142, 147, 153, 189, 230
Co-channel interference, 4
Cognitive cellular network, 279
Cognitive radio, 11, 247
cognitive radio networks, 47
Collision probability, 295, 308
Collision ratio, 255
Command-and-control model, 292
Commons model, 293
Contention free period, 274
Contention window, 274, 301
Continuous wave, 90
Convex optimization, 142
Cooperative Busy Tone (CBT), 289
Cooperative communications, 31
Amplify-and-Forward (AF), 13, 206
coded cooperation, 14
cognitive relaying, 15
cooperative networks, 205
Decode-and-Forward (DF), 14, 207
Demodulate-and-Forward (DemF), 207
distributed space–time coded cooperation, 14
Estimate-and-Forward (EF), 14
FD relay, 205
HD relay, 205
incremental relaying, 15
Piecewise-and-Forward (PF), 207
relay, 11
relay networks, 11
Soft Information Relaying (SIR), 208
x-duplex relay, 205, 214
Cooperative resource management, 34
Cooperative spectrum sensing, 268
Correlation matrix, 111
Cross-tier interference, 33
CSMA, 298
 1-persistent CSMA, 298
CSMA/CA, 302
CSMA/CD, 292, 298
non-persistent CSMA, 298
p-persistent CSMA, 298
CSMA/CA, 301
CTS, 298
Cyclic prefix, 2, 54

D2D beamforming, 34
D2D communication mode, 33
D2D communications, 30
DAC, 1
DCF, 301
DECT, 18
Device discovery, 32
Device synchronization, 32
DIFS, 301, 303
Discrete-Time Markov Chain (DTMC), 258
Distributed dynamic spectrum access, 273
Diversity
diversity gain, 18
diversity order, 129, 219, 223
spatial diversity, 4, 39
Diversity order
finite-SNR diversity order, 219, 233
Down-conversion, 52, 107
Downlink (DL), 138, 176, 181
Dual optimization, 142
Lagrangian dual, 143
Dynamic exclusive-use model, 294
Dynamic spectrum access, 272, 291

EDCF, 40
EDGE, 18
Electric balance duplexer (EBD), 96
Embedded operating system, 44
eNB beamforming, 34
End-to-end, 209, 214
Energy detection, 248, 254
Ergodic capacity, 236, 244
Ethernet, 43
Exclusive-use model, 293

False alarm, 255

FBMC, 54–56
FCC, 248
FDD, 50
Federal Communications Commission (FCC), 9, 38
Femtocell, 29, 179
 Femtocell Access Point (FAP), 181
FFT, 2
Field Programmable Gate Array (FPGA), 103
FIR, 99, 101
Flase alarm, 303
Frequency hopping, 38
Frequency-selective channel, 57
Full-duplex
in-band full-duplex, 46
Full-Duplex (FD), 50, 90, 109, 138, 179, 205, 247, 291
in-band full-duplex, 90, 91

GFDM, 54–56
GPS, 11
GSM, 18
Guard
guard band, 11
guard interval, 2
guard subcarrier, 38

Half-Duplex (HD), 109, 138, 205

Handoff, 32
HD-SM, 110, 136
HetNet, 179
High Occupancy Vehicles (HOV), 10
HIPERMAN, 21
Hopcroft–Karp algorithm, 282
Hotspot, 293
Hungarian algorithm, 284

IEEE 802.11, 32, 296

IMT-2000, 18
Independent Basic Service Set (IBSS), 43
INR, 111, 257
Inter-carrier interference, 55
Inter-cell interference, 176
Inter-subsymbol interference, 55
Inter-symbol interference, 1
Interference coordination, 31
Interference management, 31
Interference temperature limit, 295
Interference-to-Noise Temperature (INRT), 147, 151
Interior point algorithm, 184
ISM, 38, 289
ITU, 18

Jitter, 41

Karush–Kuhn–Tucker (KKT) conditions, 146, 177, 188

Kuhn–Munkres algorithm, 282

Label amendment algorithm, 284

Licensed spectrum, 30
Licensed user, 291
Line of sight, 22, 23
Listen-and-talk, 250
Listen-before-talk, 247
LMS, 87, 98
Logical link control, 23
Long-term exclusive-use model, 294
Low-Noise Amplifier (LNA), 61, 62, 104
Lower bound, 135, 223
LTE, 30
LTE-A, 30
LTE-FDD, 50
LTE-TDD, 50
LTE-Unlicensed, 35–37

MAC, 291
Macro base station, 181
Managed-commons sub-model, 296
Manhattan grid, 32
Matching theory, 160
2-dimensional matching (2-DM), 282
3-dimensional matching (3-DM), 161, 281
bipartite matching, 160
many-to-many matching, 160
many-to-one matching, 160
non-bipartite matching, 160
one-to-one matching, 160, 161
Maximum Sum-Rate (Max-SR), 109, 116, 132
Medium access control, 10, 23
Mesh network, 31
MIMO, 30, 34, 39, 109
massive MIMO, 7
MU-MIMO, 6, 39, 179
Minimum Symbol-Error-Rate (Min-SER), 109, 117
Miss detection, 255, 303
MObile ad hoc NETwork (MONET), 31
Mobility, 1, 11, 18, 40–41
Mode selection, 30
Multihop, 31
Multipath fading, 2
Multiple access interference, 32

NACK, 15
Network controlled D2D, 32
Network interface card, 43
Network managed power control, 34
Network-assistance detection, 33
Network-assisted D2D, 32
Noise floor, 46, 62, 180
Non-network-assistance detection, 33
Noncooperative resource management, 34
NP-complete, 282
NP-hard, 141

OFDM, 1–4, 18, 22

OFDM network, 138
OFDMA network, 1, 3, 38
Orthogonal channel assignment, 33
OSA, 248
Outage probability, 24, 200, 209, 223
Overlay
D2D overlay mode, 33

Packet Radio NETworks (PRNET), 42

Parallel Hammerstein (PH) model, 101
PCF, 40
PCS, 9
PDF, 116, 133, 147
Peer-to-peer
communication, 31
Phase noise, 65
PHY, 291
Picocell, 20, 29
Piconet, 41, 43
Power allocation, 47, 140, 180
Power amplifier (PA), 65, 68, 98
Power control, 4, 10, 30, 31, 34
Power-throughput tradeoff, 262
Primary users, 48, 247, 248
Private-commons sub-model, 297
Protocol data unit (PDU), 22

QAM, 23
QoS, 30, 179
QPSK, 38
Quality of service, 3

Radio frequency (RF), 50, 57, 91, 107

Random backoff, 301
Rayleigh distribution, 139
Rayleigh fading, 12, 81, 121, 131
Relay selection, 15, 240, 242
Request-based discovery, 33
Residue theorem, 236
Resource allocation, 3, 32, 138, 163, 176
Resource management, 160
Rician distribution, 60
Roaming, 23, 40
Routing protocols, 15
RTS, 298

Scheduling, 21, 42, 43

Secondary base station, 279
Secondary users, 48, 247, 248
Self-interference, 90, 139
residual self-interference, 190
self-interference cancellation, 44–46, 73–89, 91, 190
self-loop interference, 209
Self-organized D2D, 32
Self-organized power control, 34
Separable correlation model, 111
Sequential parametric convex approximation (SPCA), 138, 176
SER, 109, 209
Shared-use model, 295
SIFS, 301
SINR, 34, 115, 179, 209
SISO, 76, 93
Small cell, 180
SNR, 111, 257
Space-time coding, 5
Spatial correlation, 109
Spatial Multiplexing (SM), 109
Spectral efficiency, 11
Spectrum access, 291
Spectrum commons model, 295
Spectrum efficiency, 32, 250
Spectrum overlay, 249, 295
Spectrum reuse, 31
Spectrum sensing, 248
Spectrum sharing, 30
Spectrum trading, 291
Spectrum underlay, 249, 295
Spectrum waste ratio, 256
Spread spectrum, 17, 38, 41
STBC, 137
Subcarrier, 140, 176
Subchannel, 161
SUBR, 182
Sum rate, 138

Taylor’s series, 223

TDMA, 295
Throughput, 32, 247, 296
Transmitted power, 139, 169
Turbo code, 19
Tx–Subchannel–Rx matching algorithm (TSRMA), 164
Ultra-WideBand (UWB), 12
UMTS, 17–18
Uncontrolled-commons sub-model, 296
Underlay
D2D underlay mode, 33
Universal Software Radio Peripheral (USRP), 103
Unlicensed user, 291
UpLink (UL), 138, 176, 181
Upper bound, 143, 169, 223
User pairing, 140, 160, 173
Utility, 162, 165

Video on demand, 40
Virtual D2D beamforming, 35
Voice over IP, 40

WAPI (WLAN Authentication and Privacy Infrastructure), 40

WiFi, 19, 23–41
WiMAX, 21–23
Wired equivalent privacy, 40
Wireless Personal Area Network (WPAN), 41–43, 289
WLAN, 17, 38, 41
WMAN, 21–23
WSN, 43–44

X-duplex, 109, 130