Table of Contents

Cover
Table of Contents
Title Page
Copyright
Dedication
List of Figures
List of Algorithms
About the Authors
Foreword
Preface
Acknowledgements
Acronyms
1 Introduction
1.1 IoT in 5G
1.2 IoT Networks
1.3 Characteristics of IoT Signals
1.4 Outline
2 Background of IoT Signal Detection
2.1 Random Access
2.2 Signal Detection Methods
2.3 Conclusion and Remarks
3 Sparse Signal Detection for Multiple Access
3.1 System Model
3.2 Sparse Signal Detection
3.3 Performance Analysis
3.4 Simulation Results
 3.5 Conclusion and Remarks
4 Collided Signal Detection for Multiple Access
4.1 System Model
4.2 Automatic Modulation Classification-based Detection
4.3 Performance Analysis
4.4 Simulation Results
4.5 Conclusion and Remarks
5 Multiple Delay Estimation for Collided Signals
5.1 System Model
5.2 Multiple Delay Estimation
5.3 Signal Number Estimation and Channel Estimation
5.4 Simulation Results
5.5 Conclusion and Remarks
Notes
6 Detection and Division for Backscatter Signals
6.1 System Model
6.2 Central Limit Theorem-based Signal Detection
6.3 Simulation Results
6.4 Conclusion and Remarks
7 Analysis and Optimization for NOMA Signals
7.1 System Model
7.2 Throughput and Power Consumption Analysis
7.3 Energy Efficiency Performance Optimization
7.4 Simulation Results
7.5 Conclusion and Remarks
Note
8 Signal Design for Multicluster Coordination
8.1 Multi-cluster Coordination in IoT
8.2 Multi-cluster Coordination with NOMA
 8.3 CI-aided Multi-cluster Coordination with Interference
Management
8.4 Future Works
8.5 Conclusion and Remarks
9 Conclusion of the Book
References
Index
End User License Agreement

List of Tables
Chapter 4
Table 4.1 Parameters Setting.
Chapter 7
Table 7.1 Simulation Setup.

List of Illustrations
Chapter 1
Figure 1.1 Architecture of IoT.
Chapter 2
Figure 2.1 The contention-based RA.
Figure 2.2 The contention-free RA.
Figure 2.3 The grant-free RA.
Figure 2.4 The compressed sensing.
Figure 2.5 A MIMO system.
Figure 2.6 The constellation of - system with ML.
Figure 2.7 BER of ML detector in - MIMO
system.
 Figure 2.8 The constellation of - system with ZF.
Figure 2.9 BER of ZF detector in - MIMO
system.
Figure 2.10 The process of MMSE detection.
Figure 2.11 The constellation of - system with MMSE.
Figure 2.12 BER of MMSE detector in - system.
Figure 2.13 Comparison between ZF detector and MCMC
detector when .
Figure 2.14 BER of MCMC detector in - MIMO
system.
Figure 2.15 BER of VI detection in MIMO system.
Figure 2.16 The process of compressive sensing.
Figure 2.17 The reconstruction residual with different observation
vector si...
Chapter 3
Figure 3.1 Structure of NGMA systems.
Figure 3.2 The tree search for the presence of (sparse) signals in a
block....
Figure 3.3 Complexity ratio: (a) as a function of with
; (b) as a functio...
Figure 3.4 -divergence and its approximation in Eq. (3.59)
when , , and

Figure 3.5 Performance of Stage 1 for various SNRs with

, , , , and
: (...
Figure 3.6 Performance of Stage 1 for various values of with
, , ,
,...
 Figure 3.7 Performance of Stage 1 for various values of with
, , , ,...
Figure 3.8 Performance of Stage 1 for various values of with
, , ,
,...
Figure 3.9 Performance of TS approach and SS approach for
various values of
Figure 3.10 Performance of TS approach and SS approach for
various values of...
Figure 3.11 Performance of TS and SS approach for various
values of with
Chapter 4
Figure 4.1 A typical uplink massive MIMO System.
Figure 4.2 Flowchart for classifying Cases 1–3.
Figure 4.3 Feature versus with and
.
Figure 4.4 Probability of correct estimation of versus
with and .
Figure 4.5 Probability of resolving a two-UE collision
versus with CB, v...
Figure 4.6 Success probability versus the total number
of UEs in the cell ...
Figure 4.7 Probability of resolving a two-UE collision
versus with CB,
Figure 4.8 Probability of resolving a two-UE collision
versus with ZFB, ...
Figure 4.9 Success probability versus the total number
of UEs in the cell ...
 Figure 4.10 Probability of correctly estimating via SORTE
versus with ZF...
Figure 4.11 Achievable probability of resolving a three-UE
collision versus
Chapter 5
Figure 5.1 System model of two active devices with different
RTDs.
Figure 5.2 Normalized frequencies of of the CAVI algorithm
and the correla...
Figure 5.3 Performance of the CAVI algorithm to estimate RTDs
for different ...
Figure 5.4 Performance of the CAVI algorithm to estimate RTDs
for different ...
Figure 5.5 Performance of the CAVI algorithm and the ML
approach for differe...
Figure 5.6 Performance of the CAVI algorithm and the ML
approach for differe...
Figure 5.7 Performance of the CAVI algorithm and the ML
approach for differe...
Figure 5.8 Conditional normalized MSE of the estimated
composite CIR (withou...
Figure 5.9 Conditional probabilities of correct estimation of
for given ...
Figure 5.10 Conditional probabilities of correct estimation of
for given
Figure 5.11 with numbers of samples under ,
, , , and .
Figure 5.12 with numbers of mutations under ,
, , , and .
 Figure 5.13 with under , , ,
, and .
Chapter 6
Figure 6.1 Backscatter communication system.
Figure 6.2 Ambient backscatter communications for parasite
devices.
Figure 6.3 versus number of host devices, , with
, , , , ,
and .
Figure 6.4 versus , with , ,
, , and .
Figure 6.5 versus number of antennas, , with
, , , , and .
Figure 6.6 versus channel correlation coefficient, , with
, , , , and...
Figure 6.7 versus number of host devices, , with
, , , , ,
and .
Figure 6.8 versus , with , ,
, , and .
Figure 6.9 versus number of antennas, , with
, , , , and
.
Figure 6.10 versus channel correlation coefficient, ,
with , , , ,
,...
Chapter 7
Figure 7.1 NOMA system with two IoT devices.
 Figure 7.2 The stylized relationship of EE maximization, power
minimization,...
Figure 7.3 Impact of the normalized distance between the BS and
the SU on th...
Figure 7.4 Impact of the maximum constraint at the SU
on the SIC’s failure...
Chapter 8
Figure 8.1 Low energy consumption, high energy efficiency, and
low complexit...
Figure 8.2 Transmission power consumption and EE performance
with 5 MHz band...
Figure 8.3 (a) An elementary example of CI exploitation with
BPSK constellat...
Figure 8.4 The performance gain of the FJT-CI and PBF-CI
designs over the co...
Figure 8.5 The symbol error rate and execution time of different
precoders i...

OceanofPDF.com
 IEEE Press
445 Hoes Lane
Piscataway, NJ 08854
IEEE Press Editorial Board
Sarah Spurgeon, Editor-in-Chief

Moeness Amin Ekram Desineni Subbaram

Hossain Naidu
Jón Atli
Benediktsson Brian Tony Q. S. Quek
Johnson
Adam Drobot Behzad Razavi
James Duncan Hai Li Thomas Robertazzi
James Lyke Diomidis Spinellis
Joydeep
Mitra

OceanofPDF.com
IoT Signal Detection

Rui Han
Beihang University

Jingjing Wang
Beihang University

Lin Bai
Beihang University

Jianwei Liu
Beihang University
OceanofPDF.com
Copyright © 2024 by The Institute of Electrical and Electronics Engineers, Inc. All rights reserved.
Published by John Wiley & Sons, Inc., Hoboken, New Jersey.
Published simultaneously in Canada.
No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any
form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise,
except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either
the prior written permission of the Publisher, or authorization through payment of the appropriate
per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923,
(978) 750-8400, fax (978) 750-4470, or on the web at www.copyright.com. Requests to the Publisher
for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111
River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or online at
http://www.wiley.com/go/permission.
Trademarks: Wiley and the Wiley logo are trademarks or registered trademarks of John Wiley &
Sons, Inc. and/or its affiliates in the United States and other countries and may not be used without
written permission. All other trademarks are the property of their respective owners. John Wiley &
Sons, Inc. is not associated with any product or vendor mentioned in this book.
Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts
in preparing this book, they make no representations or warranties with respect to the accuracy or
completeness of the contents of this book and specifically disclaim any implied warranties of
merchantability or fitness for a particular purpose. No warranty may be created or extended by sales
representatives or written sales materials. The advice and strategies contained herein may not be
suitable for your situation. You should consult with a professional where appropriate. Further, readers
should be aware that websites listed in this work may have changed or disappeared between when
this work was written and when it is read. Neither the publisher nor authors shall be liable for any
loss of profit or any other commercial damages, including but not limited to special, incidental,
consequential, or other damages.
For general information on our other products and services or for technical support, please contact
our Customer Care Department within the United States at (800) 762-2974, outside the United States
at (317) 572-3993 or fax (317) 572-4002.
Wiley also publishes its books in a variety of electronic formats. Some content that appears in print
may not be available in electronic formats. For more information about Wiley products, visit our web
site at www.wiley.com.
Library of Congress Cataloging-in-Publication Data applied for:
Hardback ISBN: 9781394183081
Cover Design: Wiley
Cover Image: © Yuichiro Chino/Getty Images

OceanofPDF.com
To our families and friends

OceanofPDF.com
List of Figures
Figure 1.1 Architecture of IoT
Figure 2.1 The contention-based RA
Figure 2.2 The contention-free RA
Figure 2.3 The grant-free RA
Figure 2.4 The compressed sensing
Figure 2.5 A MIMO system
Figure 2.6 The constellation of - system with ML
Figure 2.7 BER of ML detector in - MIMO system
Figure 2.8 The constellation of - system with ZF
Figure 2.9 BER of ZF detector in - MIMO system
Figure 2.10 The process of MMSE detection
Figure 2.11 The constellation of - system with MMSE
Figure 2.12 BER of MMSE detector in - system
Figure 2.13 Comparison between ZF detector and MCMC detector
when
Figure 2.14 BER of MCMC detector in - MIMO
system
Figure 2.15 BER of VI detection in MIMO system
Figure 2.16 The process of compressive sensing
Figure 2.17 The reconstruction residual with different observation
vector sizes
Figure 3.1 Structure of NGMA systems
Figure 3.2 The tree search for the presence of (sparse) signals in a
block
Figure 3.3 Complexity ratio: (a) as a function of with ;
(b) as a function of with
Figure 3.4 -divergence and its approximation in Eq. (3.59) when
, , and : (a) as a function
of with ; (b) as a function of with
Figure 3.5 Performance of Stage 1 for various SNRs with ,
, , , and
: (a) the probabilities of FA and
MD of the CAVI algorithm as functions of ; (b) -
divergence as a function of
Figure 3.6 Performance of Stage 1 for various values of with
, , ,
, and : (a) the
probabilities of FA and MD of the CAVI algorithm as
functions of ; (b) -divergence as a function of
Figure 3.7 Performance of Stage 1 for various values of with
, , , , and
: (a) the probabilities of FA and
MD of the CAVI algorithm as functions of ; (b) -
divergence as a function of
Figure 3.8 Performance of Stage 1 for various values of with
, , ,
, and : (a) the
probabilities of FA and MD of the CAVI algorithm as
functions of ; (b) -divergence as a function of
Figure 3.9 Performance of TS approach and SS approach for various
values of with , ,
, , and
Figure 3.10 Performance of TS approach and SS approach for various
values of with , ,
, , and
Figure 3.11 Performance of TS and SS approach for various values of
with , ,
, , and
Figure 4.1 A typical uplink massive MIMO System
Figure 4.2 Flowchart for classifying Cases 1–3
Figure 4.3 Feature versus with and
Figure 4.4 Probability of correct estimation of versus
with and
Figure 4.5 Probability of resolving a two-UE collision
versus with CB, various , ,
, and
Figure 4.6 Success probability versus the total number of
UEs in the cell with CB, , ,
and
Figure 4.7 Probability of resolving a two-UE collision
versus with CB, , , and

Figure 4.8 Probability of resolving a two-UE collision

versus with ZFB, various , ,
, and
Figure 4.9 Success probability versus the total number of
UEs in the cell with ZFB, , ,
and
Figure 4.10 Probability of correctly estimating via SORTE versus
with ZFB, , , , and

Figure 4.11 Achievable probability of resolving a three-UE collision

versus with and
Figure 5.1 System model of two active devices with different RTDs
Figure 5.2 Normalized frequencies of of the CAVI algorithm and
the correlator-based detector when (
), , , , ,
, and
Figure 5.3 Performance of the CAVI algorithm to estimate RTDs for
different numbers of iterations when , ,
, , , , and
Figure 5.4 Performance of the CAVI algorithm to estimate RTDs for
different values of step size , when , ,
, , , , and

Figure 5.5 Performance of the CAVI algorithm and the ML approach

for different values of when , ,
, , , and
Figure 5.6 Performance of the CAVI algorithm and the ML approach
for different numbers of multipaths, , when ,
, , , , and

Figure 5.7 Performance of the CAVI algorithm and the ML approach

for different numbers of antennas, , when ,
, , , , and

Figure 5.8 Conditional normalized MSE of the estimated composite

CIR (without CIR overlapping) when , ,
, and : (a) as a function of
with and ; (b) as a function of
with and ; (c) as a function of
with and
Figure 5.9 Conditional probabilities of correct estimation of for
given as functions of when
 , , , , and

Figure 5.10 Conditional probabilities of correct estimation of for

given when , ,
, and : (a) as functions of the
number of multipaths, , when ; (b) as functions
of the length of CP, , when
Figure 5.11 with numbers of samples under , ,
, , and
Figure 5.12 with numbers of mutations under , ,
, , and
Figure 5.13 with under , , , ,
and
Figure 6.1 Backscatter communication system
Figure 6.2 Ambient backscatter communications for parasite devices
Figure 6.3 versus number of host devices, , with ,
, , , , and

Figure 6.4 versus , with , , ,

, and
Figure 6.5 versus number of antennas, , with ,
, , , and
Figure 6.6 versus channel correlation coefficient, , with
, , , , and

Figure 6.7 versus number of host devices, , with ,

, , , , and
Figure 6.8 versus , with , ,
, , and
Figure 6.9 versus number of antennas, , with
, , , , and
Figure 6.10 versus channel correlation coefficient, , with
, , , ,
, and
Figure 7.1 NOMA system with two IoT devices
Figure 7.2 The stylized relationship of EE maximization, power
minimization, and throughput maximization problems
Figure 7.3 Impact of the normalized distance between the BS and the
SU on the value of EE and throughput, with
. (a) normalized distance
between the BS and the SU and (b) normalized distance
between the BS and the SU
Figure 7.4 Impact of the maximum constraint at the SU on the
SIC’s failure probability, where the SU is in the middle of
the BS and the WU
Figure 8.1 Low energy consumption, high energy efficiency, and low
complexity techniques are preferable to enable an energy-
efficient multi-cluster coordinated IIoT system
Figure 8.2 Transmission power consumption and EE performance with
5 MHz bandwidth. (a) transmission power consumption
performance and (b) EE performance with different
coordination designs
Figure 8.3 (a) An elementary example of CI exploitation with BPSK
constellation; (b) A quadrature phase shift keying (QPSK)
constellation example with CI precoding exploits
interference as a beneficial element; (c) A schematic
representation of - constellation points
Figure 8.4 The performance gain of the FJT-CI and PBF-CI designs
over the conventional coordination techniques
Figure 8.5 The symbol error rate and execution time of different
precoders in the FJT scenario. (a) Symbol error rate of
different precoders and (b) execution time of different
precoders

OceanofPDF.com
List of Algorithms
Algorithm 2.1 ML Detection
Algorithm 2.2 ZF Detection
Algorithm 2.3 MMSE Detection
Algorithm 2.4 Gibbs Sampler
Algorithm 2.5 VI Detection
Algorithm 2.6 CS Detection
Algorithm 3.1 VI Detection for Sparse Signal
Algorithm 5.1 The GA-MCMC Gibbs Sampler
Algorithm 7.1 EE Oriented FD Cooperative NOMA Algorithm

OceanofPDF.com
About the Authors
Rui Han received the PhD degree in cyber security from Beihang
University, Beijing, China, in 2022. From 2022 to 2024, Dr. Han was a
research fellow at National Research Center, Tsinghua University, Beijing,
China. Her current research interests include the Internet of things (IoT),
unmanned aerial vehicle (UAV) communications, and satellite
communications.
Jingjing Wang received his BSc degree in electronic information
engineering from the Dalian University of Technology, Liaoning, China, in
2014 and the PhD degree in information and communication engineering
from the Tsinghua University, Beijing, China, in 2019, both with the highest
honors. From 2017 to 2018, he visited the next-generation wireless group
chaired by Prof. Lajos Hanzo in the University of Southampton, UK. Dr.
Wang is currently a professor at the School of Cyber Science and
Technology, Beihang University, Beijing, China. His research interests
include AI-enhanced next-generation wireless networks, UAV networking,
and swarm intelligence. He has published over 100 IEEE
Journal/Conference papers. Dr. Wang was a recipient of the Best Journal
Paper Award of IEEE ComSoc Technical Committee on Green
Communications & Computing in 2018, and the Best Paper Award of IEEE
ICC and IWCMC in 2019. He is currently serving as an editor for the IEEE
Wireless Communications Letter and the IEEE Open Journal of the
Communications Society. He has served as a guest editor for IEEE Internet
of Things Journal.
Lin Bai received the BSc degree in electronic and information engineering
from the Huazhong University of Science and Technology, Wuhan, China,
in 2004, the MSc degree (Hons.) in communication systems from the
University of Wales, Swansea, UK, in 2007, and the PhD degree in
advanced telecommunications from the School of Engineering, Swansea
University, UK, in 2010. Since 2011, he has been with Beihang University
(Beijing University of Aeronautics and Astronautics, BUAA), Beijing,
China, where he is currently a professor at the School of Cyber Science and
Technology. His research interests include the security of space-air-ground
integrated network (SAGIN), broadband wireless ad hoc network,
unmanned aerial vehicle (UAV) communications, and Internet of Things
(IoT). He has authored two books published by Springer in 2012 and 2014.
He was the Symposium Co-Chair of IEEE GLOBECOM 2019, IEEE VTC
2021, and IEEE/CIC ICCC 2024, the Tutorial Co-Chair of IEEE/CIC ICCC
2019. He is the founding chair of IEEE ComSoc Wireless Communications
Technical Committee Special Interest Group (SIG) on Space Air Ground
Integrated (SAGI) Communications. He has served as an editor for IEEE
Transactions on Signal Processing and IEEE Wireless Communications
Letters, a lead guest editor for IEEE Wireless Communications, and a guest
editor for IEEE Internet of Things Journal. He is currently serving as an
editor for IEEE Transactions on Wireless Communications and IEEE
Transactions on Mobile Computing, and the managing editor for Journal of
Communications and Information Networks. He is a distinguished lecturer
of the IEEE Communications Society and the IEEE Vehicular Technology
Society.
Jianwei Liu received the BSc and MSc degrees in electronic and
information from Shandong University, Shandong, China, in 1985 and
1988. He received the PhD degree in communication and electronic system
from Xidian University, Shaanxi, China, in 1998. Currently, he is a
professor with the School of Cyber Science and Technology, Beihang
University, Beijing, China. His research interests include wireless
communication network, cryptography, and network security.

OceanofPDF.com
Foreword
Over the past decades, Internet of Things (IoT) has been well developed to
become one of the most important technologies in the 21st century, which
aims to provide heterogeneous services for massively connected devices.
Evidently, massive connectivity in IoT causes severe access congestion, and
signal collision and signal superposition occur frequently. Therefore, signal
detection becomes crucial in IoT communication systems. This book
provides a range of key techniques to support massive IoT devices, and
various signal detection techniques are explained in the context in terms of
sparse signal detection, collided signal detection, round-trip delay
estimation, backscatter signal division, etc. It makes an easy-to-follow
presentation from the elementary to the profound level with a carefully
balanced blend of theoretical elements and applications.
My colleagues, Dr. Han, Prof. Wang, Prof. Bai, and Prof. Liu, have worked
on this topic for many years. They have made good achievements and
published a number of papers within this topic. This book provides
fundamentals of signal detection and estimation together with new results
that have been developed for IoT applications, which is ideal for both
graduate students and practicing engineers in wireless communications.
Academician Quan Yu
Chinese Academy of Engineering
Beijing

OceanofPDF.com
Preface
Machine-type communication (MTC) is expected to play a crucial role in
supporting a number of devices for Internet of Things (IoT). Due to the fact
that most IoT devices have sparse activity and low signaling overhead,
random access (RA) can be employed for MTC to provide an efficient way
to support massive IoT devices with minimized network overload.
However, there exist many problems in IoT RA, e.g., signal collision, signal
superposition. In order to face these challenges, we focus on the signal
detection for IoT in terms of sparse signal detection, collided signal
detection, round-trip delay estimation, and backscatter signal division.
Our book mainly focuses on the signal detection for RA in IoT, which
covers the fundamentals of signal detection with two chapters dedicated to
important background materials. Besides, various signal detection and
estimation techniques are explained, e.g., variational inference algorithm,
compressive sensing reconstruction algorithm, and we include a number of
recent research outcomes that are useful for those experts in this area. In
addition, the techniques are then analyzed using performance analysis tools,
and simulation results are also given to help readers to understand the
theorem and algorithm.
So far, there are many existing books related to signal detection. To be
different from those books, our book focuses on signal detection for RA in
IoT systems. Although our book is very specific, we have adopted an easy-
to-follow presentation from the elementary to the profound level. We
include a number of recent research outcomes that are useful for those
experts in this area, where the methods of signal detection have low
complexity and are suitable for IoT massive access scenarios. Meanwhile,
signal processing approaches to detect the event of collision or to estimate
the arrival time of multiple signals are also considered in this book.
Besides, our book includes more illustrations of the signal detection results
to help readers to understand the theorem and algorithm.
Our group has worked on the design of signal detection for many years and
has produced various new research outcomes on IoT signal detection. In
this book, the fundamentals of signal detection are covered, which is ideal
for both graduate students and practicing engineers in wireless
communications. In addition, this book includes not only our research
outcomes but also other recent research outcomes that could be very useful
to practitioners and postgraduate students who want to learn new outcomes
of IoT signal detection in the field of signal processing.
Beijing Rui Han
October 2024 Jingjing Wang
Lin Bai
Jianwei Liu

OceanofPDF.com
Acknowledgements
We would like to thank many people for supporting this work, in particular:
Q. Yu (Chinese Academy of Engineering), J. Choi (Deakin University), W.
Zhang (University of New South Wales). They helped us by providing
valuable comments and useful discussions.
Special thanks go to those people who inspire and encourage us all the time:
J. Lu (Tsinghua University) for guidance and encouragement, C. Jiang
(Tsinghua University) for generous support, J. Wang for long-term
friendship, and many others including our students, J. Ma, Q. Zeng, Y.
Shan, Y. Yu, M. Liu, C. Liu, Y. Li, Y. Wang, for careful proofreading.
Then, we want to express our appreciation to our parents, families, and
friends. Without their support, we can barely make the achievement.
Finally, we deeply thank Editor B. Kurzman and S. Indirakumari, who were
always there with us, for their wonderful help during the completion of the
book.

OceanofPDF.com
Acronyms
ACB access class barring
AMC automatic modulation classification
AMP approximate message passing
AWGN additive white Gaussian noise
BER bit error rate
BP basis pursuit
BPSK binary phase shift keying
CAVI coordinate ascent variational inference
CB conjugate beamforming
CDMA code division multiple access
CLT Central Limit Theorem
CS compressive sensing
CSCG circular symmetric complex Gaussian
CSI channel state information
DA device association
EAB extended access barring
FBMC filter bank multicarrier
FD full-duplex
FJT fully-coordinated joint transmission
F-OFDM filter-based orthogonal frequency division multiplexing
FPC fractional power control
GA Genetic Algorithm
GLRT generalized likelihood ratio test
HD half-duplex
HOC high-order cumulant
IA interference alignment
ISI inter-symbol interference
LASSO least absolute shrinkage and selection operator
LDPC low-density parity check
MAP maximum a posteriori
MCMC Markov chain Monte Carlo
MIMO multiple-input multiple-output
ML maximum likelihood
MMSE minimum mean square error
NFV network function virtualization
NOMA non-orthogonal multiple access
OFDM orthogonal frequency division multiplexing
OMP orthogonal matching pursuit
OOK on–off keying
PA power amplifier
PBF partially-coordinated beamforming
pdf probability density function
PEP pairwise error probability
QAM quadrature amplitude modulation
QPSK quadrature phase shift keying
SER symbol error rate
SIC successive interference cancellation
SINR signal-to-interference-plus-noise ratio
SNR signal-to-noise ratio
SM spatial modulation
SORTE second-order statistic of the eigenvalues
VI variational inference
ZF zero forcing
ZFB zero-forcing beamforming
4G fourth generation
5G fifth generation

OceanofPDF.com
1
Introduction
In the present era, as Internet applications continue to evolve, studies
on Internet of Things (IoT) have represented a growing field. This is
particularly driven by the advancements in the latest generation of
information technology, which leads to various innovative paradigms
such as smart agriculture, smart health, and smart logistics. As a result,
the concept of intelligent interconnectivity between all objects has
become a reality. This has led to a profound transformation in the way
people live, work, and travel, ushering in a new era of possibilities.

1.1 IoT in 5G
1.1.1 What Is IoT
IoT refers to the technique of using various devices, such as radio frequency
identification (RFID), sensors to collect necessary information of things or
progresses, then transmitting the collected data through the network, to
achieve the ubiquitous connection between objects and people, objects and
objects [1].
The concept of IoT was initially proposed by the Massachusetts Institute of
Technology in 1999 [2]. At its early stages, IoT referred to the network
created by the RFID technology and equipment. By incorporating the
Internet and adhering to agreed-upon communication protocols, this
network facilitated intelligent identification and management of product
information, enabling seamless interconnectivity [3].
However, with ongoing technological advancements and application
development, the scope of IoT has been broadened. In its modern aspect,
IoT encompasses the integration of perception, identification, and control of
interconnected objects. This integration, combined with networking and
intelligent processing capabilities, enables the formation of highly
intelligent decision-making systems [4].
As outlined in the White Paper on IoT published by the China Academy of
Information and Communications Technology (CAICT) [5], IoT represents
an expansive application and network extension of existing communication
networks and the Internet. Through the utilization of perceptual technology
and intelligent devices, IoT enables the perception and identification of the
physical world. By leveraging network transmission and interconnectivity,
IoT facilitates computation, processing, and knowledge mining. As a result,
it enables seamless interaction and connectivity between individuals and
objects, as well as between objects themselves. This, therefore, enables real-
time control, precise management, and informed decision-making
processes.
Regarding the fifth generation (5G) of wireless technology, advancements
of IoT primarily stem from innovations in wireless and network
technologies [6]. Within the field of wireless technology, the industry has
placed particular emphasis on large-scale antenna arrays, ultra-dense
networking, innovative multiple access techniques, and full spectrum
access. Network technology has also witnessed significant progress, with
the widespread recognition of a new network architecture based on
software-defined networking (SDN) and network function virtualization
(NFV) [7].
Furthermore, several key technologies have been exploited as important and
promising contributors to 5G, including filter-based orthogonal frequency
division multiplexing (F-OFDM), filter bank multicarrier (FBMC), full
duplex and flexible duplex, multivariate low-density parity check (LDPC)
codes, network coding, and polarization codes.
From a network architecture perspective, 5G inherits the overall
characteristics of the fourth generation (4G), encompassing access
networks, core networks, and upper-layer applications. However, to
accommodate the diverse requirements of IoT, 5G has introduced new key
technologies in both the core network and access network domains,
bringing on technological innovations and network advancements.

1.1.2 Applications of IoT

Currently, there has been a successful combination between 5G
communication technology and IoT [8]. This integration has brought on
widespread applications in various domains including intelligent
manufacturing, wireless healthcare, and smart logistics. The inherent
advantages of 5G and IoT, such as enhanced efficiency and reliability, have
significantly propelled the improvement of practical applications in these
areas. The utilization of 5G-based IoT is expected to incredibly enhance the
overall quality of life and transform our work and lifestyle practices
fundamentally.
In recent years, the widespread adoption of IoT, particularly in the industrial
area, has become increasingly prevalent [9]. A prime example of the
combination between 5G communication technology and IoT is the
application of wireless healthcare, which has greatly facilitated the life of
individuals. The integration of IoT technology into medical devices has
brought on an escalating demand for superior communication quality in
remote control scenarios [10].
During the era of 4G communications, exposure to issues of signal quality
has been shown to be related to adverse effects on the continuity and safety
of the entire diagnostic and treatment process. Thankfully, the advent of 5G
networks has significantly improved signal quality, offering low-latency
and reliable connections. This enhancement enables medical professionals
to accurately and promptly assess patients’ conditions in remote diagnosis,
thanks to high-resolution images and videos, and provide precise feedback
in wireless healthcare. Besides, the utilization of remote medical devices in
scientific surgical procedures ensures a high-quality service for patients.
Moreover, 5G communications have emerged as powerful platforms for
autonomous driving systems. With the benefits of 5G, autonomous vehicles
can strengthen their perception, decision-making, and control functions to
the edge cloud. This architecture allows the processing and dissemination of
data to be carried out utilizing the computational capabilities of the edge
cloud, thus reducing the reliance on vehicle sensors [11].
Furthermore, the remote driving within the field of Internet of Vehicles
(IoV) becomes a key instrument in efficient vehicle control and monitoring
from production scheduling centers. The incorporation of remote driving
not only ensures the driving safety and well-being of individuals but also
gives significant advancements within the automotive industry, which
promotes convenient transportation for individuals and generates economic
benefits for the automotive sector.

1.1.3 Future of IoT

The Narrowband Internet of Things (NB-IoT) technique has been a major
contributor in driving the widespread implementation of IoT within various
domains, e.g., industrial interconnection, and improving people’s quality of
life. This advancement has led to the establishment of a innovative
industrial ecosystem and substantial commercial expansion [12].
Looking ahead, the future of IoT promises to experience significant
transformations in the next two decades due to the continuous development
of innovative technologies. One prominent example is the forthcoming
realization of large-scale vehicle and utility automation. This progress will
encompass various facets, such as smart grids for energy production,
efficient waste management systems, and intelligent environmental
monitoring, all aimed at reducing greenhouse gas emissions and pollution.
While the utilization of 5G network in IoT offers numerous benefits and
shows remarkable potential for future advancements, it is crucial to
acknowledge the issues that arise during the practical implementation of a
5G-based IoT architecture. These issues include inadequate communication
security, significant investment costs, increasing energy consumption, and
limited availability of high-frequency resources, all of which contribute to
the complexities associated with deploying IoT networks.
These difficulties primarily stem from the expansion of network
infrastructure, effective management of IoT equipment, and efficient
processing of the vast amount of data generated by the extensive
deployment of mechanical devices within the IoT ecosystem. Furthermore,
interoperability and heterogeneity pose additional challenges within the
context of IoT. For instance, seamless integration of heterogeneous
networks remains a persistent issue, inhibiting effective connectivity among
these networks. In particular, the interoperability problem is increasingly
recognized as a serious concern in the communication and exchange of
information between numerous mechanical devices and intelligent
networks, as well as in establishing connections with various applications.
The lack of standardized protocols and compatible interfaces hinders
smooth interoperability. Moreover, there are potential concerns regarding
information security. Ensuring data integrity and confidentiality within the
IoT communication system presents a significant challenge, as vulnerability
in the network infrastructure and device ecosystem may compromise the
security of the transmitted data.
To effectively tackle the challenges arising from the application of IoT
technology in the 5G network and ensure the efficient utilization of IoT
systems, it is imperative to establish unified standards for wireless
communications, data collection, and security management. These
standards should satisfy the specific requirements of 5G IoT applications
and facilitate the advancement of robust security management and control
mechanisms.
Currently, researches on 5G-based IoT networks are in their early stages
and lack a mature research framework and content. Given this, it is crucial
to dedicate efforts towards exploring and developing practical technologies
for 5G-based IoT systems to satisfy the growing demand of IoT
communications. Simultaneously, addressing existing challenges requires
enhancing the heterogeneity and mobility of 5G network applications to
cater to the scale and characteristics of IoT devices. By doing so, we can
effectively tackle the existing obstacles.

1.2 IoT Networks

The IoT network architecture consists of the sensor/perception layer, the
network layer, the communication layer, the architecture layer, and the
application layer, as shown in Fig. 1.1, which are introduced as follows:

The sensor/perception layer realizes intelligent perception, information

collection and processing, automatic control of the physical world, and
connects the physical entities to the network layer and the application
layer through the communication. It includes various IoT terminals
composed of IoT chips, modules, sensors, IoT cards, and integrated
circuits;
Figure 1.1 Architecture of IoT.
The network layer mainly has the functions of transmission, routing,
and controlling of information, including the extended network, the
access network, and the core network. It mainly consists of the
Internet, wireless broadband networks, wireless low-speed networks
(e.g., ZigBee and Bluetooth), mobile communication networks, and
many emerging wireless access techniques such as low power wide
area networks (LPWAN) (e.g., long range radio (LoRa) and NB-IoT);
 The communication layer transmits the entire information between
layers, which can rely on the Internet and the public
telecommunication network, or the industrial private network. It can be
considered as the backbone of the IoT architecture;
The architecture layer, including cloud computing, big data analysis,
and other architectures, provides information processing, computing,
and other general infrastructure services, capabilities, and resource
invocation interfaces for IoT applications. It is the basis for various
applications of IoT in various fields;
The application layer uses the Internet to integrate all devices, sensors,
and information through wireless connections, which can realize
different IoT applications, such as smart factories, smart homes, smart
agriculture, smart transportation, etc [13].

Meanwhile, the IoT network consists of core network, transmission bearer

network, and wireless network [14]. The core network needs to support IoT
functions and provide IoT service license to meet the writing requirements
of users’ data. The transmission bearer network uses the existing IP
metropolitan area network for data transmission. Among them, the
transmission bearing part can be divided into backhaul propagation and
fronthaul propagation, where the backhaul propagation refers to the two-
layer virtual private network (VPN) docking interaction between the core
network and the baseband unit (BBU) through the router, while the forward
propagation refers to the direct connection between BBU and radio remote
unit (RRU) through optical fiber. The wireless network part is combined
with the existing site resources.
Moreover, the network deployment modes can be categorized into
independent deployment, protection band deployment, and in-band
deployment. The independent deployment mode involves utilizing new
bands or unused bands exclusively for deployment purposes. This mode
ensures that there is no interference with the existing networks, thereby
maintaining optimal network performance. Additionally, the protection
band deployment mode primarily utilizes the protection band within the 5G
edge spectrum for deployment. By leveraging weak signals, this mode
maximizes the utilization of spectrum resources and ensures efficient
network operation. Lastly, the in-band deployment mode is predominantly
implemented within a specific frequency band of an existing carrier. This
mode allows for seamless integration within the carrier’s network
infrastructure, utilizing the available frequency bands efficiently.

1.3 Characteristics of IoT Signals

In recent years, academia and industry have begun to study the 5G-based
IoT. Compared with IoT currently being deployed, 5G-based IoT has more
stringent requirements for power consumption, connectivity performance,
and coverage. Therefore, it is challenging for traditional technologies to
meet these performance requirements. The main challenges are shown as
follows:

1. Signal transmitting with low-time delay: In order to meet the usage

requirements, 5G-based IoT needs to solve the problems of low
transmission rate and high delay of 4G networks, which severely limits
the applications of IoT and affects the use of various IoT applications.
Taking smart home as an example, users need to send information
through voice or remote control terminals. After smart home
recognizes, receives, and processes the user’s information, it will
conduct automatic and intelligent operations to achieve the intended
purpose of users. All of these processes rely on the high-speed
communication network support. Once the network speed is poor or
the network time delay is high, the smart home cannot identify,
receive, and process the user’s information timely, and the entire
intelligent operation will have a huge delay, which will seriously affect
the actual experience of users. Besides, the automated driving
technology, in recent years, also has increasing demands for higher
information transmission rate and lower delay of the communication
network. If the intelligent vehicle cannot receive terminal instructions
or quickly process relevant information of operations due to the low
transmission speed, it will cause traffic chaos, and even threaten the
life safety of passengers.
2. Signal detection with low-power users: IoT continues to evolve from
a niche market to a vast network that connects almost every aspect of
our lives, thus the power consumption is critical for such a wide range
 of applications. In the field of IoT, many network devices are sensors,
wireless devices, and brakes equipped with data acquisition nodes. In
the usual case, these nodes may be power-limited or battery-free
devices, which can transmit their data through obtaining and reflecting
the energy of the radio frequency (RF) sources. Especially in the
industrial installations, these nodes are often placed in areas that are
difficult to access or are inaccessible, which means they must operate
and transmit data for years on a single button battery.
Therefore, 5G-based IoT needs to extend the life of the battery in the
device by exploiting low-power technologies. It is also necessary to
balance the relationship between transmission power, data processing,
data storage, and energy consumption, as well as research low-power
signal transmission and processing technologies. In addition, new
signal detection technologies are supposed to be explored, so that the
receiver can successfully detect and receive the signal of terminals.
3. Signal collision caused by large number of users: The traditional
random access (RA) scheme faces challenges, e.g., up to ten million
idle users, thousands of simultaneous active users, short packet data
transmission, and low energy consumption constraints. In the
traditional communication systems, RA is a necessary process to
establish a wireless link between the terminal and the network. Only
after the RA is completed, the data transmission between the terminal
and the network can be carried out normally.
However, when the number of users in a cell increases, the
performance of RA begins to degrade significantly. Even if the
overload control technologies, such as access class barring (ACB),
adaptive ACB (A-ACB), extended access barring (EAB) are adopted,
with the continuous increase of the number of supported users in the
cell, the access success probability of the whole system will be greatly
decreased.
4. Active user sensing of large-scale burst signals: In a complex IoT
ecosystem, there are various subsystems. For example, different drones
can be used to serve thousands of industries, e.g., mining and
exploration industries can use them for unmanned inspection, media
and entertainment industries can use them for aerial photography.
 Besides, drones can also act as mobile base stations (BS) to provide
high-capacity coverage and high-precision positioning services on
demand, and large drones play crucial roles in the logistics industry.
The common feature of the above 5G-based IoT scenarios is that a
large number of terminals carry scattered services, while the device
activation is usually sporadic. In each time slot, only a small number
of devices are active and can communicate with the BS. Other devices
are temporarily dormant to save energy, and they are only activated
when triggered by an external event. Therefore, the fundamental
challenge of 5G for IoT is that the BS needs to dynamically identify
activated devices and complete data transmission in an efficient and
timely manner.

1.4 Outline
The remaining themed chapters of this book are introduced as follows:
Chapter 2 considers the background of IoT detection, including RA and
some signal detection methods. Among them, RA is divided into two types
of introductions, namely grant-based and grant-free, which are
distinguished by whether to obtain the reserved channel information. Some
traditional signal detection methods, such as maximum likelihood (ML),
zero forcing (ZF), etc., are also mentioned. Other signal detection
algorithms in use, such as variational inference (VI) detection, Markov
Chain Monte Carlo (MCMC) detection, etc., are described in detail.
Chapter 3 is concerned with the sparse signal detection for multiple access,
with specific technologies like VI detection and compressive sensing (CS)
detection. In this chapter, the algorithm of VI detection is firstly introduced
with corresponding analysis of its performance. Then, the theory of CS
detection is discussed, including Bayesian CS algorithm and structured
subspace pursuit algorithm with the simulation results of them.
In Chapter 4, collided signal detection for multiple access is introduced.
With the displayed system model, we analyze the theory of automatic
modulation classification-based detection through the simulation results.
Accordingly, the simulation results of collided signal detection for multiple
access are offered.
Chapter 5 analyses the round-trip delay estimation for collided signals with
ML detection, VI detection, and MCMC detection. Firstly, the system
model of the round-trip delay estimation for collided signals is preformed.
Then, the ML detection is described and we compare it with the VI
detection and MCMC detection. Finally, the simulation results of all these
detection algorithms are brought together to show their detection
performance.
Chapter 6 presents the signal detection for backscatter signal. After giving
the explanation of the system model of backscatter signal, we describe the
algorithm of the central limit theorem-based detection and analyze its
performance by comparing its simulation results with the ML detection.
Chapter 7 firstly explains the concept of non-orthogonal multiple access
(NOMA) and adopts the technique of NOMA in IoT systems. Then, this
chapter focuses on throughput and power consumption of NOMA systems.
Finally, the puzzle of energy efficiency maximization of a full-duplex
cooperative NOMA system is investigated.
The purpose of Chapter 8 is to present the signal design for multi-cluster
coordination, including multi-cluster coordination with NOMA,
constructive interference-aided multi-cluster coordination, and performance
analysis. Then, discussions about successive interference cancellation
design, device grouping, and power control to achieve multi-cluster
coordination with NOMA are included.
In the end, Chapter 9 gives a brief conclusion for this book.

OceanofPDF.com
2
Background of IoT Signal Detection
The aim of this chapter is to introduce random access (RA) in Internet
of Things (IoT) systems. Then, we present several well-known
detection approaches, including the maximum likelihood (ML), linear,
variational inference (VI), Markov chain Monte Carlo (MCMC), and
compressive sensing (CS) detectors.

2.1 Random Access

2.1.1 Grant-based Random Access
For grant-based RA, each device needs to collaborate with other devices or
base stations (BSs) to obtain a grant for access, which can be further
divided into contention-based RA and contention-free RA according to the
access resource acquisition methods. For contention-based RA, devices
need to initiate access requests on the specified RA resources using the
preamble allocated by the BS, where devices obtain the preamble in a
competitive way to initiate access requests [15]. Although the contention-
free RA allocates specific access resources to access requests that have a
high probability of success, the communication between devices should
allow the delay caused by the signal collision.

1. Contention-based random access: The process of contention-based

RA consists of four phases [16], as shown in Fig. 2.1:
Phase 1 – Preamble transmission: Each device randomly selects
an available preamble from the preamble pool upon the arrival of
a new RA time slot, and transmits the preamble in the physical
random access channel (PRACH) for identification of the device
by the BS.
Phase 2 – RA response: After receiving the preamble transmitted
from the device, the BS decodes the received preamble and sends
the RA response to the device through the physical downlink
shared channel (PDSCH), where the RA response mainly includes
the messages contained in the preamble, the uplink time schedule
used for synchronization, the parameters of the retreat, and
resource scheduling information, etc.

Figure 2.1 The contention-based RA.

Phase 3 – Connection request: After receiving the RA response
from the BS, the device transmits the connection request on the
physical uplink shared channel (PUSCH) allocated to it. If the
device does not receive the RA response, it retries the access after
a certain period of time.
 Phase 4 – Collision resolution: The BS transmits collision
resolution information to the device. When the device receives a
message corresponding to its identity, the connection between the
device and the BS can be established and the device can begin to
transmit data. If the device does not receive the corresponding
message, it indicates that the device has collides with other
devices and needs to resend the access request.
2. Contention-free random access: The contention-free RA procedure
consists of three phases [16], as shown in Fig. 2.2:
Figure 2.2 The contention-free RA.
Phase 1 – Preamble assignment: The BS assigns the preamble to
the device.
Phase 2 – Preamble transmission: The device sends the
dedicated preamble on the PRACH.
Phase 3 – RA response: After receiving the preamble from the
device, the BS generates the RA response and sends it to the
device.
The main advantage of grant-based RA is that it is easy to be operated at the
BS and does not require complex access algorithms. However, the grant-
based RA has the following drawbacks:

The number of IoT devices is large while the number of orthogonal

preamble is limited, which leads to a high probability that the same
preamble is repeatedly selected by multiple devices. Therefore, the
access failure probability due to the collision is high.
Since the grant-based RA scheme requires four handshakes, the
signaling overhead is considerable.

To solve the above problems, there have been a number of studies involving
the methods of collision resolution in RA protocol, e.g., backoff
mechanism, access class barring, time slot access mechanism, priority
access mechanism, and other mechanisms, etc [17]. The access class
barring mechanism and the backoff mechanism are two common congestion
control strategies, which are introduced as follows:

1. Access class barring: Access class barring is an important congestion

control strategy, which directly controls the traffic by adjusting the
probability of RA requests initiated by the grant-based devices on the
PRACH. When using the access class barring mechanism, the BS will
set an access class barring probability parameter in advance, which is
also known as the access class barring factor, and broadcast it before
the access time slot. Each device generates a random number and
compares it with the access class barring factor. If the generated
random number is less than the access class barring factor, the devices
will send preamble on the PRACH, otherwise, it will be prevented
from access and need to request RA again. Although the access class
barring mechanism is promising in congestion control, it increases the
time delay and can only be used to alleviate short-term network
congestion. Besides, the update speed of the access class barring factor
is limited by the broadcast update period, so it cannot timely respond
to a large number of access devices in a short time.
2. Backoff mechanism: In the backoff mechanism, the BS sets backoff
parameters for all devices, including access time and backoff time.
When the device fails in the RA process, it first determines whether the
 maximum number of allowed backoffs has been reached. If not, the
waiting time will be selected according to the backoff indicator for the
next access attempt, where the backoff time follows a uniform
distribution. Otherwise, the device fails to access and waits for the next
activation. However, the collision probability is still high because of
the large-scale devices, and the number of access attempts before
successful access may be quite high, which increases not only the time
delay, but also the energy consumption of the signal transmission. In
this case, the traditional grant-based RA is challenging to meet the
requirements of the large-scale IoT system.

2.1.2 Grant-free Random Access

Since it is a tough problem for grant-based RA to meet the needs of the
large-scale IoT devices, a new RA mechanism should be introduced.
Recently, some scholars have proposed grant-free RA to ensure low delay
and high reliability of access to IoT devices [18].
As shown in Fig. 2.3, the core idea of grant-free RA is that the activated
device directly sends data signals without waiting for the grant of BS after
sending the preamble, which can significantly reduce the access delay and
the signaling overhead. The crucial assumption in the grant-free
transmission is that the BS has completed the ideal active user identity
detection, channel estimation, and signal synchronization before data
transmission. Thus, how to complete the above detection has become one of
the core problems to be solved in the grant-free RA. It can be seen that
grant-free RA sacrifices computational complexity to reduce access delay
and signaling overhead. Researches on the grant-free RA can be divided
into the following categories:

1. Grant-free random access algorithm based on compressed sensing:

Compressed sensing is a signal processing method further developed
by the Stanford University professor based on the theory of signal
sparse decomposition and approximation, which is widely used in
digital image processing, communication channel estimation, and other
fields. If the signal is a sparse signal containing only non-zero
elements and the measurement matrix is used for observation, the
measured signal can be written as
Figure 2.3 The grant-free RA.

(2.1)
where denotes the noise in the transmission process of the
observation signal. If the measurement matrix satisfies the criteria
of restricted isometry property (RIP), the receiver can accurately
reconstruct the sparse signal based on the signal , which is
shown in Fig. 2.4.
In general, the complexity of the massive grant-free RA is
proportional to the total number of devices. Therefore, although the
 above-mentioned grant-free RA based on compressed sensing can
reduce the signaling overhead in IoT, the high computational
complexity of these algorithms is still an urgent problem to be solved.
2. Grant-free random access algorithm based on statistics: In general,
the number of allowed activated devices is proportional to the
quadratic of the preamble length, which severely limits the number of
activated devices that can be held in the system. If the BS is equipped
with large-scale antennas, it is of great research value to design grant-
free RA based on the covariance matrix of received signals. However,
large antenna arrays lead to high array dimension and hardware cost,
and centralized processing of massive array signals may become
impractical. Thus, distributed multiple-input multiple-output (MIMO)
maybe a feasible method to support large-scale RA in the future [19].

Although the above access algorithms based on compressed sensing and

statistical information provide some specific solutions for grant-free RA,
the characteristics of IoT devices (e.g., high-dense and diversification)
present some new challenges, which are summarized as follows:

Figure 2.4 The compressed sensing.

Signal asynchronous: The models in most of the researches assume

that devices can achieve ideal signal synchronization. However, in the
 actual grant-free RA, different devices have different timing
deviations, which causes the preambles received by the BS from the
devices to be not fully synchronized. The phenomenon of preamble
asynchronism can cause the performance of active user detection
algorithm to deteriorate sharply or even fail completely, which is a
major challenge in the practical application of the grant-free RA. The
research of large-scale RA in asynchronous scenarios is still in the
exploratory stage, and there are few related studies.
Preamble shortage: The RA scheme needs to assign the
corresponding preamble to potential devices so that they can be
activated and connected at any time. However, since the number of
active devices is large, assigning a dedicated preamble to each device
causes the sequence length of preamble to be excessively long, which
reduces the transmission efficiency. In addition, due to the limitation of
the coherence time of channels, the preamble cannot be designed too
long. Therefore, how to design an RA algorithm to reduce the
preamble overhead is an urgent problem to be solved.
High computational complexity: The larger the number of antennas,
the higher the detection accuracy of grant-free RA. However. the
computational complexity of the algorithm will increase sharply. The
technique of multi-antenna is an important technology to support
future large-scale IoT applications. Due to the deployment of large-
scale antenna arrays and the access of large-scale IoT devices, the
signal received by the BS is ultra-high dimensional in the spatial
domain and frequency domain, which makes the operation complexity
extremely high and affects the practicability of the algorithm.
Moreover, the high computational complexity leads to the long
processing time, which results in the high access delay. Therefore, how
to design a low complexity algorithm for jointly activated device
detection and channel estimation is a troublesome issue.

In this book, we mainly focus on the signal detection for IoT systems,
where some traditional signal detection methods will be introduced first.
2.2 Signal Detection Methods
In this section, we first introduce the system model of IoT communications.
Then, we talk about the fundamental signal detection methods, which are
ML detection, zero forcing (ZF) detection, minimum mean square error
(MMSE) detection, and the modern signal detection methods, such as
MCMC detection, VI detection, and CS detection. These methods focus on
the detection of a group of signals, i.e., the signal vector, instead of a single
signal.

2.2.1 System Model

As the proliferation of IoT devices escalated in forthcoming wireless
networks, the next generation multiple access (NGMA) schemes are being
contemplated to facilitate extensive connectivity for a vast multitude of
devices. In IoT scenarios, it is expected to support a tremendous amount of
communicating machines, thus machine-type communication (MTC)
becomes essential.
In order to establish communication with a BS, an IoT device must initiate a
procedure of RA. This involves transmitting a preamble, randomly chosen
from a predefined set, through a shared PRACH. In scenarios where a large
number of devices are attempting to access the network simultaneously, it is
common for multiple IoT devices to select the same preamble, resulting in
preamble collisions. Unfortunately, when collisions occur, the data packet
transmitted by the devices involved may fail to be decoded successfully at
the BS. Therefore, mitigating congestion and overload on the PRACH is of
utmost significance, particularly in the context of supporting massive
connectivity to facilitate MTC.
Recent years have witnessed a growing academic interest in MIMO
technique due to its crucial importance in providing extensive spatial
freedom and improving spectral efficiency, MIMO has become a significant
technique. As a result, many IoT devices are now equipped with multiple
antennas, and a variety of transmission schemes have been proposed to
exploit the capabilities of MIMO channels. These schemes aim to achieve a
harmonious balance between the advantages of diversity and multiplexing.
A typical MIMO system is shown in Fig. 2.5, and its transmission model
can be written as

(2.2)
where is a vector of dimensions, which represents the received signal;
is a vector of dimensions, which represents the signal to be detected;
is a channel matrix, and denotes the noise
vector with dimensions.

Figure 2.5 A MIMO system.

Let , where is a discrete random variable. The
set represents the signal constellation. We denote as the a priori
probability of , and assume independence among all . To estimate ,
we can address the maximum a posteriori probability (MAP) problem
illustrated as

(2.3)

where represents the a posteriori probability of .

The problem outlined in Eq. (2.3) is known as a MIMO detection problem.
Generally, finding the MAP solution for this problem is challenging due to
its computational complexity, which may be unfeasible. Consequently,
various low-complexity approaches have been explored to tackle the
following problems.
2.2.2 ML Detection
ML detection searches all candidate vectors exhaustively and selects the
largest possible vector with the smallest error probability as the received
signal vector [20]. Generally, the method calculates the Euclidean distance
of the received signal vector, and finds out the smallest Euclidean distance
among all possibilities and the corresponding possible vector is chosen as
the target transmission signal vector. In this case, the ML detection has the
best detection performance.
Let denote the likelihood function of for a given received signal
vector , then the estimated vector of ML detection can be expressed as

(2.4)

where is the set of all signal candidates on the constellation and

denotes the 2-norm.
All transmit signal vectors in the constellation will be traversed to deal with
Eq. (2.4). If the transmitter has antennas and the modulation order is
, then the traversal search space is . In other words, the
computational complexity increases exponentially with the number of
antennas and the modulation order. Therefore, although ML detection has
good performance, it is generally not applied in massive MIMO detection.
Through the ML detection, the fully utilized spatial diversity gain is
achieved, which equals the number of receive antennas . This diversity
gain is derived by considering the pairwise error probability (PEP). If
denotes the transmit signal, denotes the erroneously detected signal, the
PEP can be expressed as
 (2.5)

where and denotes the real-part operation. The noise

vector is a circularly symmetric complex Gaussian (CSCG) random
vector, then also becomes a CSCG random vector, which has the
following properties:

(2.6)

where represents the expectation for the element in brackets and

denotes the trace of the matrix in brackets.
According to Eq. (2.6), it can be obtained that

(2.7)

Since for any matrix ,

(2.8)

the variance of becomes

 (2.9)

where denotes the Frobenius norm. Then, the PEP can be expressed
by the -function [21] as

(2.10)

where . According to Eq. (2.10), the

upper bound of PEP can be computed as

(2.11)

where , and

(2.12)

According to the Chernoff boundary, the average PEP’s upper bound can be
given by
 (2.13)

where denotes the determinant of the matrix, and represents the

variance of the channel matrix . Equation (2.13) shows that a complete
receive diversity gain of can be achieved by the ML detection with the
given system model.
The process of ML detection is presented in Algorithm 2.1.

Example 2.1 Consider a MIMO system with 64 quadrature

amplitude modulation (QAM) modulation method. Assume that the channel

matrix is , and the noise matrix is

. With the signal vector

transmitted in 64-QAM, the received signal is

 (2.14)

Let

(2.15)

where is the candidate vector. Note that the power is normalized. We

know that there are 4096 candidate vectors in 64-QAM, then we substitute
the candidate vector into Eq. (2.15) and find the minimum value of .
Through substituting every candidate vector and calculating, the minimum
value of is 0.0045 and the corresponding vector is

For a more generalized scenario, Fig. 2.6 demonstrates the constellation

diagram of the estimated signal obtained by ML detection in a 64-QAM
communication system with the same response matrix. Therefore, this
shows that the ML detector can correctly detect the signal, but because of
its exhaustive search, it trades high time complexity for good detection
performance.
In order to numerically analyze the performance of the ML detector, we will
simulate this detection method, where the bit error rate (BER) is considered
as a performance metric in simulation at a certain signal-to-noise ratio
(SNR).
In Fig. 2.7, we show the BER of ML detector in 4-QAM 2 2 MIMO
system. What stands out in this figure is that the BER of ML detector drops
from 0.3085 to , when SNR increases from 0 to 20 dB.
Figure 2.6 The constellation of - system with ML.
Figure 2.7 BER of ML detector in - MIMO system.

2.2.3 ZF Detection
ZF detection is one of the linear detection methods. The signal received by
a linear detector is filtered by a well-designed matrix which enables the
transmitted multiple signals to be detected separately without involving
interference. The main idea of ZF detection can be summarized as using
linear transformation to reduce the mutual interference between diverse
antennas, to obtain useful signals, where the estimator matrix can be
expressed as

(2.16)

Besides, the estimated signal can be given by

 (2.17)

Thus, the covariance matrix of the estimated error of the transmit signal is

(2.18)

where denotes the variance of noise.

It can be seen that the operation of ZF detection is simple, but it ignores the
influence of noise. In particular, the noise is amplified after being filtered
and detected by ZF estimator, and the estimation error of the transmit signal
becomes larger. Thus, ZF detection cannot offer a satisfied BER
performance when SNR is low.
The process of ZF detection is presented in Algorithm 2.2.

Example 2.2 Consider a MIMO system with 64-QAM modulation

method. Assume that the channel matrix is
 , and the noise matrix is

. With the signal vector

transmitted in 64-QAM, the received signal is

(2.19)

The estimator matrix is

(2.20)

The estimated signal becomes

(2.21)

Then it can be determined that the estimated signal vector is by

hard judgment. Therefore, this shows that ZF detection can correctly detect
signal.
Besides, the covariance matrix of the estimated error of the transmit signal
is
 (2.22)

from which we can see that the estimated error of the transmit signal is
small.
To give a more generalized scenario, Fig. 2.8 reveals the constellation
diagram of the estimated signal obtained by ZF detection in a 64-QAM
communication system with the same response matrix. It is apparent from
this figure that ZF detection can correctly detect signal.
Figure 2.8 The constellation of - system with ZF.
Figure 2.9 BER of ZF detector in - MIMO system.
In order to numerically analyze the performance of ZF detector, we will
simulate this detection method. Figure 2.9 shows the BER of ZF detector in
4-QAM 2 2 MIMO system. There is a clear trend in Fig. 2.9 that the
BER of ZF detector drops from 0.3671 to , when SNR
increasing from 0 to 20 dB.

2.2.4 MMSE Detection

The MMSE detection makes up for the defect of noise amplification in the
ZF detection. The main idea of MMSE detection is to calculate the mean
square error between the detected signal and the actual transmitted signal
, to identify the estimated value that minimizes the mean square error as
the detection result.
There are mainly two aspects during the MMSE detection process: channel
estimation and signal estimation, as shown in Fig. 2.10.

Figure 2.10 The process of MMSE detection.

In the first step, the scheme to obtain the MMSE filter is to minimize the
mean-square error (MSE). The solution of this minimization problem, as
well as the corresponding MMSE estimator matrix , can be
computed as

(2.23)

where denotes the MMSE estimator matrix, is the -order

identity matrix, and represents the symbol energy.
In the second step, the resulting estimated signal vector is

(2.24)

The covariance matrix of the estimated error of the transmit signal can be
given by
 (2.25)

Compared with Eq. (2.18), the MMSE detection suppresses noise

amplification and the estimation error of the transmit signal is smaller.
The specific steps of MMSE detection are shown in Algorithm 2.3.

Example 2.3 Consider a MIMO system with 64-QAM

modulation method. Assume that the channel matrix is

, and the noise matrix is

. With the signal vector

transmitted in 64-QAM, the received signal is

(2.26)

The estimator matrix is

 (2.27)

The estimated signal is

(2.28)

Similarly, the vector is decided as the estimated signal vector

by hard judgment.
For a more generalized scenario, Fig. 2.11 presents the constellation
diagram of the estimated signal obtained by MMSE detection in a 64-QAM
communication system with the same response matrix. Therefore, it can be
proved that MMSE detection can correctly detect signal.
Figure 2.11 The constellation of - system with MMSE.
Figure 2.12 BER of MMSE detector in - system.
To conduct a numerical analysis of the MMSE detector’s performance, the
simulation of this detection method is provided. In Fig. 2.12, we show the
BER of MMSE detector in 4-QAM 2 2 MIMO system with SNR from 0
to 20 dB. This figure illustrates that as SNR increases, the BER of MMSE
detector decreases. When SNR reaches around 18 dB, the BER is
approximately constant with the increased SNR. This result highlights that
the performance of MMSE detector can be enhanced by increasing SNR.
However, the performance is approximately stable when SNR reaches a
specific value.

2.2.5 MCMC Detection

To state the system model, suppose that a MIMO system is
equipped with transmit antennas and receive antennas. Based on the
system model in Section 2.2.1, the signal received by the receiver can be
assumed as

(2.29)
where , denote the channel matrix
and the transmitted signal vector, respectively, and
represents the noise vector.
The detection algorithm based on MCMC obtains the list of transmitted
sequences by statistical sampling method, and estimates the posterior
probability of each encoded bit based on statistical principle [22]. Since the
performance of MCMC detection is determined by the number of samples
and independent of the dimension of the random variable to be estimated, it
can effectively solve the problem that the algorithm complexity increases
exponentially with the number of transmit antennas and modulation orders.
The MCMC algorithm was first applied to physics by Metropolis in 1950
[23]. Its main idea is to obtain a sample of the desired distribution by
means of the Markov chain. By choosing the appropriate transfer kernel, the
sampling value of the Markov chain with any initial distribution can
approach to the expected distribution after enough sampling times.
In MIMO systems, each transmitted symbol sequence is defined as a
state of the Markov chain. Since the number of states of the Markov chain
increases exponentially with the dimension of in MIMO systems, Gibbs
Sampler is applied to reduce the sampling complexity, which is one of the
most practical MCMC sampling methods [24]. Let denote the prior
probability of , represents the variable at the th iteration. The
transfer kernel is obtained as

(2.30)

where denotes the complement set of

. Analyzing Eq. (2.30), we can obtain that
 (2.31)

(2.32)

From , it can be derived that

(2.33)

where denotes the set of symbols for all possible

values of .
When , in the form of the log-likelihood ratio
can be expressed as

(2.34)

where . According to
the posterior probability of each symbol, the posterior probability of each
coded bit can be computed as
 (2.35)

where and denotes the mapping relationship

between encoded bits and symbols. In addition, and
, where means the constellation diagram size.
The specific steps of Gibbs Sampler are shown in Algorithm 2.4.

Example 2.4 Consider a MIMO system with 4-QAM modulation

method. Assume that the initial signal is a 100-byte signal vector with the
corresponding value space , and the signal passes
through a randomly generated Rayleigh channel with

system noise . Assume that the number of

iterations is 100 and one Markov chain is used for a single iteration.

To conduct a numerical analysis of the MCMC detector’s performance, the

simulation of this detection method is provided. To start with, a 100-byte
signal vector is generated as the initial detection signal. Subsequently,
the acceptance probability of the th bit signal taking the value
is calculated. Then, if the acceptance probability satisfies the demands, the
th bit signal is updated to the corresponding variable , otherwise it is
unchanged. This process is then repeated until all iterations have been
completed.
Figure 2.13 Comparison between ZF detector and MCMC detector when
.
The comparison between ZF detector and MCMC detector in 4-QAM 2
2 MIMO system is shown in Fig. 2.13. From the figure, it can be observed
that the BER of MCMC is higher than that of ZF at the beginning.
However, as the number of iterations increases, the BER of MCMC
decreases. After about 50 iterations, the BER of MCMC remains stable,
which illustrates that MCMC detector has better performance than ZF
detector.
Figure 2.14 shows that the BER of MCMC detector in 4-QAM 2 2
MIMO system decreases stably with SNR from 0 to 20 dB, which reveals
that the performance of MCMC detector can be improved under the
increased SNR. There have different approaches to improve the
performance of MCMC detector, which will be introduced in Chapter 5.
2.2.6 VI Detection
In the MIMO detection problem, in order to detect the signal vector s from
the received signal vector r, it may be simpler to obtain an approximation
of the distribution of s using the VI detection method.
Generally, for those unknown distribution of some random variables,
sometimes their probability density functions (pdf) are intractable to
compute. In this case, the basic idea of the VI method is to construct a
distribution , namely the variational distribution, which can be easily
computed, and continuously decrease the error between the variational
distribution and the actual distribution of s until the variational distribution
is convergent. Once the convergence is achieved, can be regarded as
an approximation of .
Figure 2.14 BER of MCMC detector in - MIMO system.
The algorithm of VI detection mainly consists of two parts: assumption of
variational distribution and computation of Kullback–Leibler (KL)
divergence. The first procedure is the assumption of variational distribution.
Since the inference of unknown variable can be regarded as the problem of
computing the posterior distribution of that variable, the detection of signal
vector s can be replaced by the computation of the distribution of s.
Additionally, sometimes it is challenging to calculate the distribution of s.
In this case, the VI detection can provide a method to suppose a variational
distribution to fit the distribution of s, which can simplify the computation.
To select one family for the appropriate distribution, one effective type is
the mean-field variational family, which supposes that the elements of s are
independent of each other. Suppose the variational distribution in the mean-
field approximation of each is , then can be expressed as
[25]

(2.36)

The second procedure is the computation of KL divergence. Generally, the

error and distance between the designed variational distribution of s and its
original distribution are supposed to be decreased to improve the similarity
between these two distributions, so that the accuracy of variational
distribution can be improved. In this case, the KL divergence, defined as the
difference between the information entropy of variational distribution and
that of the original distribution, represents the comparability between the
variational distribution of s and its initial distribution. If one of two
distributions is the actual distribution, and the other is the fitted distribution,
then the KL divergence is equal to the difference between the cross entropy
and information entropy of the actual distribution as

(2.37)

Under a lower value of KL divergence, the variational distribution of s can

fix and approximate its original distribution better.
For instance, if the KL divergence is defined as , where
represents the initial distribution of , then the problem is to
minimize the KL divergence, as well as to consider the optimization
problem expressed as

(2.38)

where denotes the collection of all possible distributions of variable .

The KL divergence can be computed as [26]
 (2.39)

Since there are no relationships between and , then

. Therefore, Eq. (2.39) can be computed as

(2.40)

If is regarded as a variable, then is a constant. Therefore,

the aforementioned optimization problem to minimize the value of KL
divergence is equivalent to the problem to maximize the value of evidence
lower bound (ELBO), which is expressed as

(2.41)

Thus, the final aim is to find out the optimal solution as

(2.42)

With the maximum ELBO and the minimum KL divergence, the variational
distribution of s can fix the original distribution with less error.
The specific steps of VI detection are shown in Algorithm 2.5.
Example 2.5 Consider a MIMO system and the channel is an
additive white Gaussian noise (AWGN) channel. Let be the transmit
signal vector with , where and represent the transmitted
binary bit, i.e., 0 or 1, of the first and the second transmit antenna,
respectively. The channel matrix is generated using random complex
Gaussian distribution and the noise vector is assumed to be
. The probability that is
. The symbol energy is set to 2.

For each experiment, the main process is to compute ELBO in each

iteration and compare the difference between each two consecutive values
of ELBO to obtain the optimal value of ELBO. In this case, the accuracy
is set to . If the difference between two consecutive values of
ELBO is lower than , then the algorithm will be stopped to obtain the
optimal values of ELBO, and the final result of variational distribution of s
can be computed as well.
Figure 2.15 shows the result of BER with the change of SNR in dB, where
SNR is varied from 0 to 20 dB. It can be obtained that BER decreases with
the increased SNR. This means that with SNR increases, the outcome that
using the variational distribution of s to fix its real distribution becomes
more accurate, and the variational distribution of s can represent its true
distribution more effectively. Thus, with a relatively high SNR, the VI
detection can provide a strategy to estimate the real distribution of the target
signal s.
2.2.7 CS Detection
The CS theory was first proposed by Donoho [27]. As a new sampling
theory breaking Nyquist’s sampling theory, it has been widely concerned by
researchers. Based on the sparsity of signals, CS theory requires that signals
can be sparsely expressed in a certain variation domain. In CS theory, the
sampling rate of the signal, which is no longer determined by the highest
frequency of the signal, is closely related to the information structure of the
signal. Compared with Nyquist’s theory, CS theory usually reconstructs the
original sparse signal from a small number of observed values. Due to these
characteristics, CS can be used for signal detection in MIMO systems.
Besides, in some IoT scenarios, the device activation is usually sporadic
and only a small number of devices are active in each time slot, and it is
more effective to utilize the CS algorithm in IoT systems than the
traditional algorithms.
Figure 2.15 BER of VI detection in MIMO system.
CS detection mainly includes three aspects: signal sparse representation,
compression measurement, and signal reconstruction, which are shown in
Fig. 2.16.

Figure 2.16 The process of compressive sensing.

The first step is the signal sparse representation. Assume that is a column
vector of dimension in space, and any signal in space
can be written as a linear combination of an orthonormal basis .
Then, the signal can be written as
 (2.43)

where is the matrix formed by the orthonormal basis , and is

the dimensional column vector generated by projecting the
coefficients of the signal into the domain. If there are at most
nonzero elements in and is much less than , the signal is sparse
or compressible in the domain with sparsity . Note that CS theory is
based on the premise that signals can be expressed sparsely. Some common
transform bases include Fourier transform basis, discrete cosine basis,
wavelet transform basis, Gabor basis, etc.
The second step is the compressive measurement. The receiver projects the
received signal from the high-dimensional space to the low-dimensional
space by using the observation matrix , which can be shown as

(2.44)
Here, the observation matrix is a flat matrix of dimension and
is a column observation vector of length . The matrix of
dimension is called the sensing matrix, where the sensing matrix
computes a sparse vector of length to an observation vector of
length . Since , it maps the signal from high-dimensional space
to low-dimensional space. To losslessly reconstruct or from , must
contain all useful information.
It is also pointed out that the sensing matrix should satisfy the RIP to
ensure that the CS algorithm can accurately reconstruct the original sparse
signal. The RIP criterion states that, for any sparse signal , if there
exists a constant such that the sensing matrix can satisfy

(2.45)

then the sensing matrix is said to satisfy the RIP property, and is
called the constrained isometry constant. The RIP criterion ensures that the
mapping has uniqueness. That is, the sensing matrix will not map
various sparse signals into the same observation vector. Most uniformly
distributed random matrices can satisfy the RIP property well. Common
sensing matrices include Bernoulli random matrices, partial Fourier
matrices, independent and equally distributed Gaussian random matrices,
Rademacher matrices, and partial Hadamard matrices. Generally, a
necessary condition for CS to guarantee the uniqueness of the
reconstruction is that any columns of the sensing matrix are linearly
independent.
The last step is the signal reconstruction. If the sensing matrix satisfies
certain conditions, then the compressed low-dimensional vector can
reconstruct the original high-dimensional signal. The process of solving the
original high-dimensional signal is called the signal reconstruction. When
the signal is sparsely expressed, since the orthogonal basis matrix is
known, the reconstruction of the original signal and the sparse vector
are equivalent. Equation (2.44) is a linear equation and solving the
unknown vector from the linear equation is originally a linear regression
problem. However, because , Eq. (2.44) is an underdetermined
equation. Since CS restricts the sparsity of the vector , the multi-solution
problem is further limited and the CS algorithm can use the sparsest
solution to estimate the original signal, so that Eq. (2.44) is eventually
evolved into solving the optimization problem with sparse constraints as

(2.46)

However, the optimization problem based on the minimum norm

constraint is a non-deterministic polynomial (NP) hard problem. Solving
Eq. (2.46) requires considering all possible linear combinations of the
positions of the non-zero coefficients in , which is, . As it is
troublesome to directly solve Eq. (2.46), scholars generally convert it into
the problem of finding a sub-optimal solution, where the methods mainly
include greedy algorithm, convex relaxation algorithm, and statistical
optimization algorithm.
The specific steps of CS detection are shown in Algorithm 2.6.
Example 2.6 Suppose a 256-bit signal vector has the sparsity of 50,
i.e., there are 50 bits with a value of 1. Assume that the length of low-
dimensional observation vector is varied from 64 to 128. An identity
matrix is used to be sparse matrix, i.e., , while
represents the identity matrix of order . The measurement matrix is
assumed to be a Gaussian matrix. Then the observation vector can be
obtained according to Eq. (2.44). We consider the basis pursuit (BP)
algorithm, one of the classical reconstruction algorithms in CS, to obtain the
reconstructed signal . The reconstruction residual is shown in Fig. 2.17.
Figure 2.17 The reconstruction residual with different observation vector
sizes.
As can be seen from Fig. 2.17, the larger the size of the observed vector, the
smaller the reconstruction residual. That is, the smaller the gap between the
reconstructed vector and the original vector. This shows that although CS
algorithm can reduce the size of the transmitted data, in order to ensure the
accuracy of recovery, the compressed vector should not be too small.

2.3 Conclusion and Remarks

In this chapter, the background of IoT signal detection is introduced,
including RA strategies, traditional signal detection methods, and some new
signal detection algorithms. Among them, RA strategies can be divided into
grant-based RA and grant-free RA, which are distinguished by whether to
obtain the reserved channel information. Except for some traditional signal
detection methods such as ML, ZF, MMSE, and some popular signal
detection methods such as MCMC detection, VI detection, CS detection,
are discussed in detail as well. In Chapters 3–6, we will apply these
methods into IoT signal detection.

OceanofPDF.com
3
Sparse Signal Detection for Multiple Access
As the number of Internet of Things (IoT) devices grows explosively
in wireless networks, the next generation multiple access (NGMA)
schemes are considered to support massive access for a large number
of devices. Since multiple-input multiple-output (MIMO) has become
one of the essential techniques for providing a sufficient degree of
spatial freedom and improving the spectral efficiency, there have been
various transmission schemes over MIMO channels to achieve the
tradeoff between diversity and multiplexing gains, in which spatial
modulation (SM) [28] is a newly proposed technique to select among
multiple transmit antennas to obtain high data rates by introducing
spatial domain. Considering that the signal vector of SM systems
contains a number of zero elements, the information vector being
transmitted becomes sparse. Thus, compressive sensing (CS)
reconstruction algorithms [29] can be applied to SM systems to detect
sparse signals, which enable the reconstruction of compressible or
sparse signals. In this chapter, we introduce some sparse signal
detection approaches for multiple access of IoT devices [30].

3.1 System Model

Consider a MIMO system with inputs and outputs. The received
signal at a receiver, denoted by , is given by

(3.1)
where is an matrix, is the signal to be detected,
and is the background noise vector. Let
, where is a discrete random variable. Here,
is the signal constellation. Denote by the a priori probability of
and assume that the ’s are independent with each other. Denote by
 the likelihood function of . To estimate , we consider either
the maximum likelihood (ML) or maximum a posteriori probability (MAP)
problem as

(3.2)

respectively, where represents the a posteriori

probability of . The problem in Eq. (3.2) is referred to as a MIMO
detection problem. In general, it is challenging to find the ML or MAP
solution due to prohibitively high computational complexity. Thus, a
number of low-complexity approaches are investigated to tackle the issues
[31, 32], e.g., zero forcing (ZF) detection and minimum mean square error
(MMSE) detection. However, those sub-optimal approaches require that
in general.
In MIMO communications with multiple antennas, the signal vector
becomes sparse in SM or generalized SM, where a transmitter equipped
with antennas can transmit signals by activating one or few antennas at
a time [33, 34]. In this case, the signal constellation is to be extended to
in order to take into account the case that no signals are
transmitted from a transmit antenna. As for code division multiple access
(CDMA)-based random access (RA) with nodes where each column of
is a dedicated signature sequence for a node [35–37]. Let and
be the signature sequence and the signal of the th node, respectively.
Since a few nodes have signals to transmit at a time, becomes sparse.
Note that according to Eq. (3.1), is not only the signature sequence but
also the spreading sequence for node as in CDMA.
Let be the sparsity of , i.e., the number of non-zero elements is or
. For simplicity, let . In this case, the number of
the possible signal vectors for is . If is sufficiently small (e.g.,

), since , the ML or MAP detection may be carried out

using an exhaustive search. However, in general, if is not small, it is
troublesome to perform the ML or MAP detection due to the high
computational complexity.
Since , becomes sparse and CS-based approaches can be used
as in SM [38] for the signal detection and in [37, 39] for RA. In particular,
in Eq. (3.1) becomes the measurement matrix (that has more columns
than rows) and is a sparse vector. A low-complexity CS algorithm [40]
can be used to find from by exploiting the sparsity of . However,
although the complexity of CS-based approaches is lower than that of ML
or MAP, it is still proportional to . Thus, for a large , it would be
desirable to decrease the complexity.
For NGMA in IoT, we assume that a base station (BS) is equipped with a
number of antennas, saying antennas. In addition, there are a large
number of users, saying users, where each user is equipped with
multiple antennas, saying antennas. Then, there are inputs.
Each user can be active or inactive, and an active user uses SM to transmit
his/her information as shown in Fig. 3.1. In general, it is expected that
, while only a fraction of inputs (saying inputs) are not
zero (thanks to the sparse user activity and SM), which makes the input
vector sparse. CDMA-based NGMA can also be considered, where each
user has spreading codes of length for SM. With multiple antennas
and spreading codes, a large system for NGMA can also be formed. As a
result, in NGMA, it is necessary to deal with a large MIMO matrix , and
a low-complexity algorithm to solve a large MIMO detection problem by
effectively exploiting the sparsity is highly desirable.
Figure 3.1 Structure of NGMA systems.

3.2 Sparse Signal Detection

In this section, we introduce a tree search-based approach to detect sparse
signal by exploiting the sparsity with .

3.2.1 Tree Search-based Approach

Firstly, we introduce the method of tree search with a binary tree. At the
root node, the presence of any active signals is to be decided with as
shown in Fig. 3.2. If any signal is presented in , the signal index set,
, is divided into two subsets, namely and
. Then, the decision on the absence or presence is
carried out with the subsets. The same process is repeated with smaller
subsets.
Figure 3.2 The tree search for the presence of (sparse) signals in a block.
As shown in Fig. 3.2, there can be four subsets. If any of them do not have
any signals, they can be excluded. For example, if the third and fourth
subsets do not have any signals, the signal detection can be run with the
submatrix of , . Thus, the complexity can be reduced.
Note that, an -ary tree can also be considered, where .
Motivated by tree search, we propose a two-stage (TS) approach for sparse
signal detection, where TSs are summarized as follows:

Stage 1: Active variable detection using coordinate ascent variational

inference (CAVI) algorithm to decide active groups or inactive groups.
Stage 2: CS-based algorithms (e.g., the orthogonal matching pursuit
(OMP) algorithm, the least absolute shrinkage and selection operator
(LASSO) algorithm, and the approximate message passing (AMP)
algorithm) are carried out with the signals belonging to active groups,
while the signals associated with inactive groups are removed.

Then, we consider a special case with TSs with details. We assume that
columns are divided into groups and the aggregated signal for group
becomes
 (3.3)

where is the index set of the columns for group . For convenience, let
for all , where is a positive integer. It is assumed
that and , . In order to
detect the activity of group , we define the activity variable of group
as

(3.4)

It is indicated that there are no signals in group if , and vice

versa. In addition, we assume that is the independently identically
distribution with and , where
is the activity probability. The (conditional) signal variance when is
active, i.e., , can be given by

(3.5)

where represents the symbol energy for , with

.
Clearly, if , we have . On the other hand, if , under
the Gaussian approximation, we assume that

(3.6)

where it is assumed that and

(3.7)

It can be shown that

 (3.8)

where is a matrix consisting of the columns of corresponding to the

index set and

(3.9)

Note that the probability that is and is

. In addition, since is same for all , let
. Consequently, if we assume that

(3.10)

then the received signal can be re-written as

(3.11)

with the aggregated signals, , and the activity variables, .

As discussed above, the TS approach can be shown as follows:

In the first stage, with Eq. (3.11), we decide active and inactive groups
with the activity variables, .
In the second stage, the signals associated with inactive groups are
removed, and the sparse signal detection is carried out with the signals
belonging to active groups.
Thus, the computational complexity of the second stage can be lower if
there are more inactive groups.

Example 3.1 Let , , and with . In addition,

let . Then, , while . If
all the ’s are correctly detected, in Stage 2, we have

In this example, becomes a matrix. If we consider a

conventional approach without TSs, is a matrix. Thus, compared
with the conventional single-stage (SS) approach, the complexity of the TS
approach can be lower and the performance can be improved as well (as the
number of columns of the measurement matrix is smaller), provided that the
’s are correctly detected in Stage 1.

3.2.2 VI Detection Algorithm

In the following, we discuss the detection of aggregated signals with the
model in Eq. (3.11) under the Gaussian approximation [41].
For convenience, let

To decide , using Eq. (3.11), it can be shown that

(3.12)
where . In addition, we have two hypotheses:
for and for . For each hypothesis, under the Gaussian
approximation in Eq. (3.11) or (3.12), it can be shown that
 (3.13)

where . As in [42], the test statistic is given

by

(3.14)

where

(3.15)

For a given threshold , we accept if or if

according to the Neyman–Pearson test [32, 42]. However, in order to find
the test statistic in Eq. (3.14), it is required to know , which is not
available. As a result, the hypothesis testing with is considered to see
an achievable performance.
Another approach is based on the MAP criterion [32], which is given by

(3.16)

where . In Eq. (3.16), since all the ’s are jointly detected,

we do not have the difficulty with , i.e., we do not need in
detecting . However, the computational complexity can be high for a
large .
In order to find an approximate solution to Eq. (3.16), an iterative
variational inference (VI) approach is presented as follows. Suppose that
there is a parameter vector , that is to be estimated from an observation
. In VI, instead of detecting , we consider the estimation of the
distribution of [25]. Furthermore, the mean-field approximation can be
employed, where the ’s are assumed to be independent. Let the
variational distribution for each be and for the mean-field
approximation, let
 (3.17)

Then, we consider the following optimization problem to approximate

with as

(3.18)

where denotes the collection of all the possible distributions of

subject to the constraint in Eq. (3.17), and is the Kullback–Leibler
(KL) divergence that is defined as

(3.19)

Here, represents the distribution of with for all .

As shown in Eq. (3.18), we need to find that is close to the a
posteriori probability as an approximation. Then, the
minimization of KL divergence can be transformed to maximize the
evidence lower bound (ELBO) [25], which becomes

(3.20)
Let . The CAVI algorithm [25, 43] is to update
one variational distribution at a time with the other variational distribution
fixed in each iteration, which can be shown as

(3.21)

where denotes the expectation with respect to or with the

distribution .
It can be shown that
 (3.22)

Since the last two terms are independent of , it can be shown that

(3.23)

where is a constant. Let

and

(3.24)

It can be shown that

(3.25)

and then,

(3.26)

where and are constants. Let and

with a slight abuse of notation. It can be shown that
 (3.27)

where the approximation becomes tight when , ,

approaches either 0 or 1. For convenience, let and
. We also define as the subvector of , that is
obtained by removing .
To obtain the updating rule for in Eq. (3.21), suppose that is to
be updated. Let

(3.28)

Then, using the matrix inversion lemma [44], we can show that

(3.29)

where , , and

Then, it can be shown that

 (3.30)

(3.31)

Clearly, the updated value of is given by .

Note that, can be updated with a new as

(3.32)

where . It is also straightforward to find in Eq.

(3.28) from using the matrix inversion lemma. In Eqs. (3.29) and
(3.32), it can be seen that we only need matrix inversions.
The process of VI detection is presented in Algorithm 3.1, where
denotes the number of iterations and denotes the accuracy requirement of
detection.
3.3 Performance Analysis
3.3.1 Complexity Analysis
In this section, we discuss the complexity of the tree search-based approach
under the assumption that the activity variables, , are correctly detected.
For comparison, the conventional sparse signal detection based on CS
algorithm is discussed. In conventional approaches, the sparse signal
detection is carried out at all groups of received signals using CS algorithm,
e.g., OMP, LASSO, and AMP. The conventional approach has the total
complexity as [40, 45, 46]:

(3.33)
where denotes the complexity of CS algorithm (referred as Stage 2) per
group and signal.
Because the average number of inactive groups (associated with )
is given by

(3.34)

the average number of signals that are removed in Stage 2 becomes

, which increases with . However, as increases,
the complexity of Stage 1 increases. The total complexity of the TS
approach as a function of for given can be given by

(3.35)

where denotes the complexity of activity variable detection (referred as

Stage 1) per group and signal.
As a result, the complexity ratio is given by

(3.36)

It is indicated that the complexity ratio is influenced by the number of

signals per group and the complexity ratio of Stages 1 and 2. From this, it
can be seen that it is necessary to have to keep the complexity
ratio low, i.e., less than 1. In other words, for a computational saving using
the TS approach, the complexity of Stage 1 has to be lower than the
times of that of Stage 2.
In Fig. 3.3, the complexity ratio is shown as functions of the number of
signals per group, , and the activity probability . In general, we can see
that the complexity ratio becomes low for a low activity probability and
a low value of , while the complexity ratio decreases and then increases
as increases (as shown in Fig. 3.3 (a)). With or 8, as a rule, the
TS approach can reduce the complexity (compared with the conventional
one). It is observed that if , there is no complexity saving using
the TS approach as mentioned earlier.

Figure 3.3 Complexity ratio: (a) as a function of with ; (b) as

a function of with .

3.3.2 VI Detection Performance Analysis

In this section, we mainly focus on the detection performance of Stage 1
based on the Gaussian approximation using the Neyman–Pearson test with
the test statistic in Eq. (3.14). As mentioned earlier, the resulting
performance can be seen as an achievable one as is required when
is detected. Note that the performance analysis of CAVI algorithm is not yet
well known, while the analysis of the Neyman–Pearson test is to be used to
see the behaviors of the performance of CAVI algorithm.
Since is positive semidefinite, we can consider the following
factorization as

(3.37)

Under , define

(3.38)

Then, after some manipulations, the test statistic in Eq. (3.14) is re-written
as

(3.39)

where

(3.40)

In Eq. (3.40), the equality is the eigenvalue decomposition of , where

is the matrix consisting of the eigenvectors and is the diagonal
matrix consisting of the eigenvalues that are denoted by
.
Consequently, the characteristic function under each hypothesis is given by
 (3.41)

As a result, the test statistic can also be expressed (in terms of the
distribution) as

(3.42)

where the ’s are independent chi-squared random variables with 2

degrees of freedom and

(3.43)

We can find the probabilities of missed detection (MD) and false alarm
(FA) in Stage 1 for a given threshold using the distributions from the
characteristic functions in Eq. (3.41).
For an alternative performance measure that is independent of , the -
divergence [42] can be used, which is a measure of the distance between the
distributions associated with and . From Eq. (3.41), the -
divergence is given by
 (3.44)

In general, as the -divergence increases, a better detection performance is

expected. Since , from Eq. (3.40), it can be obtained

that

(3.45)

Lemma 3.1 Suppose that , i.e., for

. In addition, denote by for that is
associated with . Then, we have

(3.46)

Proof: Since , it can be shown that

 (3.47)

where , which is a positive semidefinite matrix.

Thus, using the matrix inversion lemma, it can be shown that

(3.48)

From this, since is positive semidefinite, it can be

shown that

(3.49)

which completes the proof.

From Lemma 3.1, we can easily see that the performance can be improved
if a small number of groups are active and the -divergence can be
bounded as

(3.50)

As a special case, suppose that there is no inter-group interference or each

is orthogonal to the others. Then, we have
 (3.51)
From Eq. (3.51), we have

(3.52)

and

(3.53)

if for all . As a result, it can be shown that

(3.54)

This indicates that as increases (or decreases) for a fixed ,a

better performance can be obtained. Note that in order to satisfy Eq. (3.51),
the rank of might be

(3.55)

which means that the ’s in can be highly correlated (if ),

which can lead to a poor performance of Stage 2.

Remark 3.1 In general, without any additional properties of or ,

it is difficult to find the -divergence or the trace of . However, when
and are sufficiently large, we can consider a large-system analysis
to find the trace of as follows.

Lemma 3.2 Denote by the number of non-zero elements of , i.e.,

. In addition, let . Assume that each element of is
an independent random variable with zero-mean and variance . For a
given , suppose that increases with a fixed ratio . Then, we have

(3.56)

where

(3.57)

Proof: From Eq. (3.45), it can be shown that

(3.58)

Each term in Eq. (3.58) can be seen as the signal-to-interference-plus-noise

ratio (SINR) of MIMO systems, where it is shown that the asymptotic SINR
(for ) depends on and . In particular, the asymptotic SINR
becomes Eq. (3.57) with . In addition, since each is
independent of the others, we have , which is
given in Eq. (3.56). This completes the proof.

Theorem 3.1 Suppose that the ’s are i.i.d. with

and . For a large
system, where with a fixed ratio , we can show that

(3.59)

where .

Proof: Since is fixed, is also fixed when . Here,

is a binomial random variable with parameters and ,
because the ’s are independent binary random variables. In particular,
. Thus, we can show that

(3.60)

Then, substituting Eq. (3.56) into Eq. (3.60), Eq. (3.59) can be obtained,
which completes the proof.
After some manipulations, we can see that is convex in (and
decreases with ). Thus, thanks to Jensen’s inequality, we can also have

(3.61)
where

(3.62)

From the first and second derivatives, which are given by

(3.63)

and

(3.64)

we can see that is concave in and has a -shape. Consequently,

there exists a maximizing , which leads to the minimum of
, i.e., the worst performance. In other words, in order to keep a good
performance (or a large -divergence), has to be close to or 1. In
addition, according to Eq. (3.62), the increase of leads to the increase of
as there are more active signals. From this, we can clearly see that the
approximation of -divergence decreases with .
In Fig. 3.4, the -divergence and its approximation in Eq. (3.59) are shown
as functions of (with ) and (with ), when
, , and dB. Note that as mentioned earlier,
is actually an analysis of the -divergence, while the
behavior of the -divergence in terms of can be seen using the
approximation that has a closed-form expression as in Eq. (3.59). Since we
use the scaling transform and the large-system analysis to approximate the
-divergence, the simulations and analysis performance have a gap. The
-divergence and its approximation become -shape functions of as
shown in Fig. 3.4 (a), which indicates that the increase of in general
degrades the performance. In addition, the performance becomes worse as
increases (that increases the interference level) as shown in Fig. 3.4 (b).
Consequently, we can see that a large is not desirable in terms of the
performance as well as the complexity, while a small, but not too small
(e.g., or 8), might be a good choice.

Figure 3.4 -divergence and its approximation in Eq. (3.59) when

, , and : (a) as a function of with
; (b) as a function of with .
3.4 Simulation Results
In this section, we present simulation results. It is assumed that each
element of is independent zero-mean circular symmetric complex
Gaussian (CSCG) with variance . In addition, we assume that is a
quadrature phase shift keying (QPSK) symbol, i.e., ,
where . The signal-to-noise ratio (SNR) is defined as SNR = ,
where the symbol energy is for QPSK.
Firstly, we mainly present simulation results for the performance of CAVI
algorithm for Stage 1. We consider the decision rule as

(3.65)

where is a decision threshold. In addition, we consider two

different values for , which are 0.5 and .
If , we decide that when .
On the other hand, if , we take into
account the a priori probability to make a decision, where is
decided when . For
simulations, the number of iterations for the CAVI algorithm is set to 4.
Probabilities of FA and MD (for given ) for Stage 1 are obtained by
simulations. Here, an event of FA or MD is obtained for each group.
Figure 3.5 shows the performance of Stage 1 for various SNRs with
, , , , and
. In general, we can see that the probabilities of
FA and MD of the CAVI algorithm decrease with the SNR in Fig. 3.5 (a).
This indicates that a better performance can be achieved with a higher SNR,
which is also confirmed by the -divergence that increases with the SNR
as shown in Fig. 3.5 (b).
Figure 3.6 shows the performance of Stage 1 with different activity
probabilities , when , , , SNR
dB, and . It shows that probabilities of FA
and MD of the CAVI algorithm increase with the activity probability as the
larger activity probability results in a higher interference level, which can
also be seen in Fig. 3.6 (b) that the -divergence decreases with
increases.
Figure 3.5 Performance of Stage 1 for various SNRs with ,
, , , and
: (a) the probabilities of FA and MD of the CAVI algorithm as functions of
; (b) -divergence as a function of .
Figure 3.6 Performance of Stage 1 for various values of with
, , , , and
: (a) the probabilities of FA and MD of the CAVI
algorithm as functions of ; (b) -divergence as a function of .
The impact of on the performance is shown in Fig. 3.7, where the
performance is shown for various values of with ,
, , SNR dB, and . As shown in
Fig. 3.7 (a), we see that the MD and FA probabilities of the CAVI algorithm
first increase and then tend to be stable as increases. Similarity, the -
divergence in Fig. 3.7 (b) has a -shape, which indicates that the
performance becomes worse and saturated as increases. Consequently,
according to Fig. 3.7 as well as Fig. 3.3, it is necessary to keep small,
but not too small for a low complexity. From them, in most cases, it seems
that is a reasonable choice.
Figure 3.8 shows the impact of on the performance of Stage 1 with
, , , SNR dB, and
. As we can see that a better performance can be
achieved with an increasing . As shown in Fig. 3.8 (b), the -divergence
increases as increases, which indicates that the performance becomes
better.
Figure 3.7 Performance of Stage 1 for various values of with
, , , , and
: (a) the probabilities of FA and MD of the CAVI
algorithm as functions of ; (b) -divergence as a function of .
As discussed above, we decide active and inactive groups with the activity
variables in Stage 1. For Stage 2, the signals associated with inactive groups
are removed and the sparse signal detection is carried out with the signals
belonging to active groups. In this way, the complexity at the receiver
becomes lower. In order to show the performance of the TS approach, we
consider the OMP algorithm, the LASSO algorithm, and the AMP algorithm
to detect the sparse signal, respectively. For convenience, the TS approach
with OMP, LASSO, or AMP for Stage 2 is referred to as TS-OMP, TS-
LASSO, or TS-AMP, respectively. In addition, for performance
comparisons, we consider a SS approach with OMP, LASSO, or AMP,
which is referred to as SS-OMP, SS-LASSO, or SS-AMP, respectively.
Figure 3.9 shows the performance of the TS approach and the SS approach
for various SNRs with , , ,
, and . It is shown that bit error rate (BER) decreases
with SNR increases, which indicates that a high SNR is desirable for a good
detection performance. Since these approaches have different diversity, the
performance of detection may be different. However, the detection
performance of these approaches becomes indistinguishable at some points
as it is influenced by the noise. In general, the TS-LASSO and the TS-AMP
algorithms can achieve better detection performance than the TS-OMP
algorithm. Besides, we can see that the TS approach significantly improves
the detection performance when SNR becomes higher.
Figure 3.8 Performance of Stage 1 for various values of with
, , , , and
: (a) the probabilities of FA and MD of the CAVI
algorithm as functions of ; (b) -divergence as a function of .
Figure 3.9 Performance of TS approach and SS approach for various values
of with , , , , and
.
Figure 3.10 Performance of TS approach and SS approach for various
values of with , , ,
, and .
Figure 3.10 shows the performance of the TS approach and the SS approach
with different activity probabilities , when ,
, , SNR dB, and . In
general, the BER of both approaches increases with the activity probability
as the larger activity probability results in the higher interference level. In
addition, the performance of TS approach has improvement compared with
SS approach, when the activity probability is small. Since the activity
probability of sparse signal can be low, the TS approach can achieve a good
signal detection performance.
Figure 3.11 shows the performance of the TS approach and the SS approach
for various values of with , , SNR
dB, , and . We can see that BER decreases with the
number of outputs, which indicates that the larger leads to a better
detection performance. It also shows that the BER of three algorithms first
declines rapidly and then tends to be stable. Compared two approaches, we
can see that the algorithms of TS approach achieve lower BER than those of
SS approach especially at a large number of outputs.

Figure 3.11 Performance of TS and SS approach for various values of

with , , , , and
.

3.5 Conclusion and Remarks

In this chapter, we introduced a tree search-based approach to detect sparse
signals for NGMA schemes, where the received signals were divided into
groups. Then, the activity variables of aggregated signals were decided and
the sparse signal detection could be carried out at the signals belonging to
active groups, where the VI was applied to the activity variable detection. In
addition, the -divergence was introduced to measure the distance between
the distributions of activity variable detection, where the approximate
expression of -divergence was given.
To further improve the performance of IoT systems, different collided
signal detection and estimation methods are designed and then adopted,
which will be introduced in Chapters 4 and 5.

OceanofPDF.com
4
Collided Signal Detection for Multiple Access
To communicate with a base stations (BS), an Internet of Things (IoT)
device must initiate an random access (RA) procedure via sending a
preamble that is randomly selected from a predefined set through a
shared physical random access channel (PRACH) [47, 48]. For
massive access, the case that multiple IoT devices choose the same
preamble namely preamble collision can occur frequently, while the
data packet from the device experiencing collision may not be
successfully decoded at the BS. Therefore, overcoming congestion and
overload of PRACH is highly desirable, especially for massive
connectivity to support machine-type communication (MTC)
communications. In this chapter, we introduce collided signal detection
approaches for massive access-based grant-free RA. Based on the
channel hardening and favorable propagation characteristics of
massive multiple-input multiple-output (MIMO), collided signals
processed at the BS can be viewed as a variation of superposition
modulation, and are to be recovered by successive interference
cancellation (SIC) techniques [49].

4.1 System Model

We consider a single-cell massive MIMO network for massive access
depicted in Fig. 4.1. Suppose that a BS is equipped with antennas and
there are a large number of IoT devices that are equipped with a single
antenna for uplink transmissions, called user equipments (UEs). For a
typical massive access scenario, the total number of UEs in a cell can be
numerous. However, the number of UEs that attempt to access the BS may
not be large in a certain time interval [50], due to the sparse user activity in
most IoT and MTC scenarios. A UE with a data packet to transmit is called
an active UE. Let denote the number of active UEs attempting to initiate
an RA procedure over the same channel. In addition,
denotes the channel vector from the th active
UE to the BS, where is the large-scale path loss and
denotes the uncorrelated Rayleigh fading.

Figure 4.1 A typical uplink massive MIMO System.

Denote by the index of the preamble chosen by the th active UE.
Then, the received signal in the space-time domain at the BS is given by

(4.1)

Here, represents the background noise where each element is an

independent circularly symmetric complex Gaussian (CSCG) random
variable with zero-mean and variance , i.e., ,
and is the transmission power for the th UE.
Denote by the data packet transmitted by the th UE. We assume that
all the active UEs have data packet of the same length . Then, the
received signal at the BS is given by
 (4.2)

where is the background noise with .

Specifically, the large scale path loss coefficient is modeled as

(4.3)

where is a constant with respect to reference distance and carrier

frequency, is the distance between the th UE and the BS, and is
the path loss exponent.
Next, we model the location of the active UE as an independent
homogeneous Poisson point process and the distance follows, a
Rayleigh distribution with mean , where is the density of
BSs measured in [51, 52]. Therefore, the probability density function
of (in kilometer) is given by

(4.4)

and the probability density function of (in meter) is

.

The transmission power is further computed as

(4.5)

where denotes the open loop transmission power and

denotes the path loss compensation fraction. Fractional power control
(FPC) is used in the long-term evolution (LTE) systems in uplink
transmission to compensate for path loss, while it can also be employed in
the scenario where no coordination between UEs and the BS is needed,
since UEs can derive their own transmission power from the system
information broadcast by the BS [53].
Generally, attenuation of transmission power with distance is beneficial to
keep the colliding UEs from obtaining the same channel gains, while the
collided signals can hardly obtain similar receive power. If the
superimposed symbols are weighted by similar values formed by their
receive power, the detection is ambiguous, and correct recovery may not be
available [54]. Hence, full power control, i.e., , should be avoided
such that the receive power from different UEs at the BS can be
distinguishable.
For RA, we assume that there is a pool of orthogonal preambles denoted by
, where and represent the number and
the length of preambles, respectively. Therefore, we have , for
. Besides, near-orthogonal preambles can be generated by Zadoff–
Chu sequences such that , . At the BS, the
preamble is used to estimate the channel vector , that will be used for
coherent detection on the following data packet [55]. For example, to
estimate the channel vector of the active UE that chooses the th preamble,
we consider the output of the correlator with as

(4.6)

Since the preambles are assumed to be orthogonal, we have

(4.7)

where denotes the index set of the active UEs that choose the th
preamble.
Obviously, if is used for channel estimation, we have three different
cases as follows:
 Case 1: . In this case, is a noise vector given by
and the BS detects no active UEs that choose .

Case 2: . In this case, the preamble is uniquely chosen by

the th UE and is given by , where
. Then, the channel vector can be accurately
estimated using .
Case 3: . In this case, the preamble is chosen by two or

more UEs and is given by . Note

that includes a superposition of channel vectors, therefore, the
channel vectors cannot be correctly estimated using . The
corresponding channel estimate generated from thus becomes a
superposition of multiple channel vectors of the active UEs that choose
(including the noise).

Besides, with the beamforming vector generated from , the BS

detects data packets with the signal given by

(4.8)
Specifically, the beamforming vector of conjugate beamforming (CB) is
and the beamforming vector of zero-forcing beamforming (ZFB)
is the th row of the matrix , where is the channel
estimation matrix defined as [56,
57]. Here, denotes the number of preambles totally picked by the
co-channel UEs.

4.2 Automatic Modulation Classification-

based Detection
Under the circumstance that the th preamble is employed with the
received signal as an input, Cases 1–3 can be classified as depicted in
Fig. 4.2, where the procedure is introduced in detail as follows.

4.2.1 Preamble Sequence Detection

Upon a UE performing a RA procedure to access the network, the BS
detects the preambles in use by calculating the cross-correlation between
the received signal and available preambles, after which the
corresponding channels are to be observed by the BS. Here, we consider an
energy detector to decide Case 1, while preamble is detected if
is larger than a predefined threshold.
The decision of Case 1 can be formulated into a binary hypothesis testing
problem. and refer to the hypotheses of signal absence and
presence, respectively, which are given by
(4.9)
Figure 4.2 Flowchart for classifying Cases 1–3.
The statistic for energy detection is

(4.10)
When the required probability of false alarm is predetermined, the
threshold can be set as

(4.11)

By using the threshold , Case 1 can be identified.

Since the statistic can be approximated as a Gaussian distribution, the
detection probability given by the function is

(4.12)

which implies that the detection is particularly accurate with a sufficiently

large . Then, given the preambles in use, the channel vectors are
computed and buffered along with their associated preamble indices.

4.2.2 HOCs-based AMC Approach for Collision

Recognition
Thanks to the channel hardening and favorable propagation characteristics
of massive MIMO, estimation error due to preamble collision and
interference can be rapidly decreased as the number of receive antennas
grows [58], while the received collided signals processed at the BS can be
considered as superposition coded signals. Moreover, to differentiate Cases
2 and 3, we need to additionally take account of the signal constellation. If
quadrature amplitude modulation (QAM) is used, in Case 2, will show 4
constellation points clearly. However, in Case 3, will have more than 4
constellation points. Based on the above facts, a high-order cumulants
(HOCs)-based automatic modulation classification (AMC) approach is
adopted for classifying Cases 2 and 3.
We first describe the processed data packet in Eq. (4.8) of Cases 2 and 3.
Assume channel vector is estimated by a minimum mean square error
(MMSE) detector. Then it can be given by [59]

(4.13)

where is the index set of the UEs that select the th preamble. Due to
the orthogonality property of the MMSE estimate, decomposition of the
channel vector can be generated as , where
represents the uncorrelated estimation error following the distribution [52]

(4.14)

Without loss of generality, we take the first active UE as an example to

present the description, and mark the preamble picked by the first UE as the
first preamble, so that the data sequence in Eq. (4.8) can be rewritten as

(4.15)

where associated with and corresponds to Cases 2

and 3, respectively. Compared to the first term which is viewed as the signal
in Eq. (4.15), the estimation error, interference, and noise terms can be
increasingly weaken as grows.
Now, we present the HOCs-based AMC approach as CB is employed. We
first compute the second-order moment of Eq. (4.15) when , which
is given by
 (4.16)

Next, the signal is reorganized as . By

utilizing the statistic in Eq. (4.16), we can further obtain that

(4.17)

Besides, the same statistic can be obtained with when ZFB is

employed.
Note that associated with approaches 1 as signal-to-noise
ratio (SNR) grows while that associated with does not. The above
fact can be utilized to distinguish Case 3 from Case 2. This can be
attainable at a moderate SNR since the variance of noise is divided by the
length of preamble sequences, . Similarly, the approach can also be
feasible in the case where more than two UEs collide. Moreover, in order to
utilize the features of statistics and the constellation mentioned above to
improve the classification performance, we then employ HOCs including
the second-order cumulant , the fourth-order cumulant , the sixth-
order cumulant , and the eighth-order cumulant , which can further
mitigate the effect of noise. For the signal , the above HOCs can be
expressed as , ,
, and
, where
[60]. Upon that, we define the feature used
for classification as
 (4.18)
Then, Cases 2 and 3 can be classified according to via a predetermined
threshold .

4.2.3 Data Decoding with SIC

Case 3 is referred to as preamble collision as multiple active UEs choose
the same preamble. For example, let us consider the preamble picked by the
first UE. Then, the probability of that other UEs select the same
preamble is

(4.19)

In practice, with UEs out of 200 UEs active, i.e., each UE

attempts to access the BS with a probability of 0.1 [61], and 64 available
preambles, the probabilities of two UEs colliding, i.e., , and three
UEs colliding, i.e., , are approximately 0.2236 and 0.0319,
respectively. It can be foreseen that the cases of more UEs colliding can be
negligible in terms of the overall probability. Therefore, for Case 3, we
mainly consider two-UE and three-UE collisions, while the proposed
approach used to detect the existence of three-UE collisions is outlined as
follows.
First, all the signals associated with unique preambles are decoded and
removed from . That is, we find

(4.20)

where is the set of the preambles with and is the data

packet transmitted by the active UE that chooses the th preamble. Then,
only has the signals associated with . In addition, we use
to denote the set of preambles with .
By exploiting the channel hardening and favorable propagation
characteristics and taking the limit , we have
 (4.21)

where is the index set of active UEs that experience preamble collisions.
In general, the number of the active UEs that experience preamble
collisions, i.e., , could be small if .
Using the eigendecomposition of , can be estimated. Then the
existence of three-UE collisions can be generally discovered if
holds, where can be determined in the collision
recognition phase. In addition, using an appropriate SIC approach that is
applied to , it is able to recover the signals transmitted from the active
UEs that experience preamble collisions.
Here, we adopt the eigenvalue gap measure-based second-order statistic of
the eigenvalues (SORTE) approach to determine the number of colliding
UEs [62]. Computed from , the matrix of eigenvalues is given by

(4.22)

where is the th eigenvalue. The SORTE approach chooses the value

as the estimation of , which minimizes the criterion

(4.23)

where , and

(4.24)

It is noteworthy that signal cancellation in Eq. (4.20) will additionally

introduce interference due to the channel estimation error, which leads to a
complicated noise environment. Since the ZFB outperforms the CB with the
same receive antennas [63], the more favorable performance of SORTE can
be obtained with ZFB.

4.3 Performance Analysis

We present the derivation of the success probability of the collision
resolution protocol. The two-UE collision is studied, where the theoretic
analysis consists of two parts, i.e., no collision and two-UE collisions.
Without loss of generality, we take the first active UE as an example to
derive the approximate analytic expressions. For simplicity, another UE that
selects the same preamble as the first UE is referred to as the second UE.
Assuming that UEs simultaneously carry out the RA procedure over the
same channel, transmission of the first UE is viewed to be successful when
its data packet is successfully decoded. Specifically, in the absence of
collisions, the data packet of the first UE is considered to be successfully
decoded when the signal-to-interference-plus-noise ratio (SINR) of its
signal layer is greater than the threshold [64]. Alternatively, in the
presence of collisions, the data packet is considered to be successfully
decoded with SIC when the SINR of the signal layer from the first UE
(after stronger signal layers are detected and canceled) together with all the
stronger signal layers is greater than the threshold . In this section, the
threshold is determined by , where is the code
rate.
In the proposed RA procedure, given the length of preambles
and a sufficiently large , the preamble detection probability can be
nearly equal to 1. Besides, as shown in the simulation results, superimposed
signals can be identified accurately even at a low by setting an
appropriate threshold. Therefore, the probability of collision recognition can
also approach 1. It is noteworthy that even though Case 2 can be falsely
recognized as Case 3 at a low , signals transmitted from the first UE
as the strongest signal layer can still be decoded by SIC. As mentioned
above, the success probability we focus on mainly depends on the data
decoding phase, which is given by
 (4.25)

where denotes the number of other co-channel UEs that select the same
preamble with the first UE and represents the probability that the
data transmission for the first UE is successful. Note that in the special case
of two-UE collision, the probability reduces to

(4.26)

1. Success Probability with CB: Then we calculate the success

probability with CB. We first discuss the case associated
with , where no collision occurs, and derive the expression
of . Let be substituted by ,
denotes the SINR of the received signal from the first UE, the
superscript 0 indicates that there are no other UEs select the same
preamble as the first UE, and the subscript CB represents that CB is
applied. From Eq. (4.15), through substituting , the SINR
corresponding to the first UE is computed as

(4.27)

Note that the interference in the denominator in Eq. (4.27) contains

correlated terms since they have common path-loss terms as
multipliers, while deriving exact expression of turns
to be intractable. Therefore, channel estimation error and small-
scale fading are calculated with the help of expectation operators.
The SINR can be simplified as
 (4.28)

Derivation of Eq. (4.28): In the numerator of , we have the

signal power simplified as

(4.29)

where is the normalized vector of which each entry has a unit

variance. Furthermore, we denote for convenience.
Since can be rewritten as an average of terms, where each
term is the square of the independent sample of standard normal
distribution, has the probability density function of Gamma
distribution . Then, with the help of the strong law of
large numbers, we have .

Next, we compute the terms in the denominator of Eq. (4.27) as

 (4.30)

(4.31)

From the central limit theorem, we have , for

, which converges to the standard normal distribution with a
sufficiently large . Therefore, converges to Gamma
distribution . Then, is approximated with its
expectation such that .

Lastly, the noise term in the denominator of Eq. (4.27) is simplified as

(4.32)

Then in Eq. (4.28) is computed by substituting Eqs. (4.29)–

(4.32) into Eq. (4.27), which completes the derivation.
From [[52], Appendix B], the approximate probability of
can be derived as
 (4.33)

where denotes the number of terms used in the calculation,

, ,

, and

Next, we consider the case of two-UE collision associated with

and derive the probability , which can be
expressed as

(4.34)

where represents the probability that the signal

transmitted from the first UE, as the strong signal layer, is decoded
successfully. denotes the SINR of the strong signal, where
the subscript UE1 stands for the signal transmitted from the first UE
is the strong signal, the subscript CB indicates that the CB is applied,
and the superscript 1 shows that there is an arbitrary UE colliding
with the first UE. Recall that the UE that chooses the same preamble
as the first UE is referred to as the second UE. Then the SINR
of the first UE is given by
 (4.35)

where

. Conditioning on , , and ,
can be rewritten as

(4.36)

where and the

calculation of follows the same line as in

[[52], Appendix B].

Let represent the probability

that the signal transmitted from the first UE, as the weak signal layer,
is decoded successfully. Let be the SINR of the strong
signal, where the subscript UE2 denotes that the signal transmitted
from the second UE is the strong signal, the subscript CB indicates
that the CB is applied, and the superscript 1 means that there is
another UE, i.e., the first UE, colliding with the second UE. Let
be the SINR of the weak signal layer, where the subscript
UE1 shows that the signal transmitted from the first UE is the weak
signal. The SINR of the second UE and the SINR of
the first UE are given by

(4.37)

(4.38)

respectively.
Similar to that in Eq. (4.36), the probability
can be modified as

, where

. Then,

is equivalent to

. Conditioning on ,

, and , given that , , and

are calculated in Derivation of Eq. (4.28), we can obtain

that
 (4.39)

As demonstrated above, can

be decomposed as

(4.40)

where the calculation of and

are also similar to that in [[52], Appendix

B]. Finally, substituting Eqs. (4.36) and (4.40) into Eq. (4.34), the
probability is obtained.
In the case of CB and no collision, the channel estimation error term,
interference terms, and noise term decay at a rate of compared to
the signal term. Therefore, the performance of CB can be near-
optimal in massive MIMO system with a sufficiently large in the
case of no collision. However, in the presence of collisions, the
performance of CB will heavily deteriorate due to the weak co-
existing signal that becomes a prominent interference term since the
weak signal term will not decay as grows, as in Eqs. (4.35) and
(4.37). Therefore, the performance of CB in the presence of collisions
 mainly depends on the signal power gap due to the different path
losses besides the estimation error, interference, and noise.
2. Success Probability with ZFB: The success probability with ZFB can
be different from that with CB. Since the vector utilized to decode data
is computed from the whole channel estimation vectors, the collision
among the other UEs is supposed to be considered. Let us calculate the
success probability in Eq. (4.26) with ZFB. We first discuss
the case associated with , where no collision occurs, and
derive the expression of , which can be expressed as

(4.41)

where the subscript ZF indicates that ZFB is employed.

Let represent the probability that preambles, except the
one picked by the first UE, are totally selected by the other
UEs. Similar to that in [[61], Section III-B], is given by

(4.42)

where denotes the

Stirling numbers of the second kind. Note that

refers to the probability of given and , where
preambles, except the one picked by the first UE, are totally selected
by the other UEs and denotes the SINR corresponding to
the first UE.
When ZFB is employed, is the data sequence processed by the
vector . Recall that is the first column of
 and is the channel estimation matrix defined as
. Besides,
is the nonempty index set, where
denotes the th preamble selected by the UEs.
To simplify the derivation, the transformation from [[55], Section III-
B] is applied to in Eq. (4.15), and then the SINR can be
expressed as

(4.43)

where is an arbitrary element in set , represents the set

excluding the element , and .

Since the exact expression of is intractable to derive, channel

estimation error and small-scale fading are calculated with
the help of expectation operators. The SINR can be simplified as

(4.44)

Derivation of Eq. (4.44): In the numerator of , since

, we have the signal power simplified as
 (4.45)

From [54], follows the Chi-square distribution and

we have via the

strong law of large numbers. Approximating the values with their
expectations as shown in [52], the terms can be computed
as

(4.46)

Then, we consider the second term in the denominator in Eq. (4.43),

which can be approximated as its upper bound such that

(4.47)

From [55], the second term is approximated as

(4.48)

Finally, in Eq. (4.44) is obtained by substituting Eqs. (4.45)–

(4.48) into Eq. (4.43), which completes the derivation.
As mentioned in above, given and , similar to that with CB, the
approximate probability of can be calculated as

(4.49)

where ,

, and

Next, we consider the case of two-UE collision associated with

, and derive the probability which can be
expressed as

(4.50)

Here, represent the probability that preambles, except the

one picked by the first UE, are totally selected by the other
UEs, which is given by
 (4.51)

where denotes the

Stirling numbers of the second kind.
Similar to that with CB, , , and with ZFB
can be rewritten as

(4.52)

(4.53)

(4.54)

respectively, where

. Note that the calculations of the probabilities

and

follow the same line as in

Eqs. (4.36) and (4.40), respectively. As a result, by substituting the
above calculations into Eq. (4.50), the probability is
obtained.
 Similar to CB, in the absence of collision, the channel estimation
error term, interference terms, and noise term decay at a rate of
compared to the signal term. Since can generally hold in
massive MIMO systems and ZF can have insignificant interference
terms, the performance of ZF outperforms that of CB. In the presence
of collisions, the weak co-existing signal that becomes a prominent
interference term cannot be further suppressed by employing ZF.
However, since other interference terms after ZF become insignificant
compared to CB, the performance of ZF in the presence of collisions
is preferable.

4.4 Simulation Results

We present numerical results to evaluate the collision resolution protocol,
where each UE is assumed to be randomly located in the cell and its
distance from the BS follows a Rayleigh distribution with mean
.
In practice, although the total number of UEs in a cell can be large for
massive access scenarios, the number of UEs that attempt to access the
network may not be large in a certain time interval, as the sparse user
activity is considered in most IoT and MTC applications. Thus, for
example, given , it is assumed that there are 200 UEs in the
cell and each UE is active with a probability of 0.05, i.e., , within a
given time interval. The assumption is applied in the simulations. Besides,
since the number of active UEs may not always remain constant at different
uplink transmission time intervals, the impact of various , i.e.,
, on performance is taken into consideration.
For the first active UE, let us recall that its transmission power
consists of open loop power and FPC factor
derived from the measurement of the system information,
where and the constant term is set to be 1 for
simplicity. Assuming that each UE has the capability to implement the FPC,
we focus on the open loop transmission signal-to-noise ratio (OSNR), i.e.,
 . It is noteworthy that the arrived SNR at the BS is lower than
OSNR. Particularly, simulation parameters are given in Table 4.1.
Table 4.1 Parameters Setting.
Modulation 4QAM
Number of simultaneous UEs 2 –30 UEs
Number of antennas 100, 400, 800
Number of orthogonal preambles 64
Length of preamble sequences 139
Length of data packets 60
Path loss exponent 4
BS density
SINR threshold , 0 dB
Figure 4.3 Feature versus with and .
Figure 4.4 Probability of correct estimation of versus with
and .
Prior to evaluate the success probability of the RA protocol, we present the
results of the HOCs-based approach for collision recognition. Feature
designed in Eq. (4.18) as a function of OSNR with and
is illustrated in Fig. 4.3. As we can see, the feature of the signal
with a two-UE collision approaches 0, while that of the signal without a
collision tends to converge on a much lower value as OSNR grows. Based
on the facts above, the cases of preamble collisions can be recognized
through a predetermined threshold. However, due to the randomness of the
transmission power of UEs, the threshold should be set to a relatively low
value to deal with some extreme cases, i.e., the transmission power of the
weak signal is much smaller than that of the strong signal, where the feature
of a signal with a collision may become similar to that without a collision.
Figure 4.5 Probability of resolving a two-UE collision versus
with CB, various , , , and .

Figure 4.4 presents the probability of correct estimation of versus

OSNR with and , where and
correspond to Cases 2 and 3, respectively, while the threshold to decide the
two cases is set to be . As shown in the figure, Case 2 is easy to
be recognized as Case 3 over the range of 25 , which means
that the noise is considered as a useful signal. However, the noise can be
distinguished due to its different frame structure, and the data from the first
UE is still able to be successfully decoded. Besides, the probability of
correctly identifying Case 2 gradually increases to 1 as OSNR grows.
Furthermore, Case 3 can be efficiently identified by the AMC approach at a
broad range of OSNR.
1. Performance on resolving a two-UE collision with CB: The
probabilities of resolving a two-UE collision as a function of OSNR
with CB under different values of are presented in Fig. 4.5. The
observation shows that the analytic and simulation results match well
with each other. Besides, we can see that the probabilities tend to get
saturated as OSNR increases, in which the performance primarily
depends on . As grows, the probabilities can reach the
saturation point at a lower OSNR for better performance. Thus, the
analysis can be verified that the impact of estimation error,
interference, and noise decays as increases.

Figure 4.6 Success probability versus the total number of

UEs in the cell with CB, , , and
.
Success probabilities as a function of the total number of UEs in the
cell with CB are presented in Fig. 4.6. Recall that the success
probability in the case of two-UE collision is given in Eq. (4.26). If
the two-UE collision is perfectly resolved, we can obtain
and . Therefore, the upper bound of RA with two-UE
collision resolution is given by .
However, if the collision resolution is not employed, the first UE can
complete RA successfully only when there are no other UEs that
select the same preamble, i.e., , . Thus, the upper
bound of RA without collision resolution is given by .
The observation shows that the proposed protocol can effectively
improve the success probability of RA when the number of available
preambles is limited. As becomes larger, we can see that the
success probability decreases gradually, while the performance can be
improved as increases.
Figure 4.7 Probability of resolving a two-UE collision
versus with CB, , , and .
Focusing on the impact of FPC, Fig. 4.7 illustrates the probability
as a function of with adequate open loop transmission
power and . As shown in the figure, the
probability becomes nearly optimal over the range of
and drops for or . When is
small, the signals received at the BS become weak, which can be
easily drowned in the co-channel interference. Thus, when the two
UEs collide, the strong one can be successfully recognized and
decoded while the weak one may not. Alternatively, when
approaches 1, since the difference between the channel gains of the
two colliding UEs reduces, the decreased SINR of the strong signal
from the superimposed signals may lead to the unsuccessful SIC.
2. Performance on resolving a two-UE collision with ZFB: In Fig. 4.8,
the probabilities of resolving a two-UE collision as a function of
OSNR with ZFB are illustrated. As we can see, the analytic results
coincide with the simulation ones, in which the success probabilities
are asymptotically saturated as OSNR grows. Moreover, compared
with the results of CB, ZFB can achieve the same performance by
utilizing fewer receive antennas, which reveals that ZFB is more
effective in suppressing the co-channel interference than CB.
Figure 4.8 Probability of resolving a two-UE collision
versus with ZFB, various , , , and
.
Figure 4.9 Success probability versus the total number of
UEs in the cell with ZFB, , , and
.
Success probabilities as a function of the total number of UEs in the
cell with ZFB are presented in Fig. 4.9. Similar to the observation in
Fig. 4.6, the success probability is effectively improved relative to the
RA without collision resolution. As becomes larger, we can see
that the success probability approaches the upper bound of RA with
collision resolution. It is clear that ZFB is more effective than CB in
accommodating more UEs.
Figure 4.10 Probability of correctly estimating via SORTE versus
with ZFB, , , , and .
Figure 4.10 shows the probability of correctly estimating the number
of collided signals via SORTE approach as a function of OSNR. The
results show that based on SORTE method, can be accurately
estimated over the range of . Hence, the existence
of three-UE collisions can be effectively discovered by employing the
proposed approach.
Figure 4.11 Achievable probability of resolving a three-UE collision
versus with and .
Figure 4.11 plots the achievable probabilities of resolving a three-UE
collision with ZFB as a function of OSNR. In this case, we assume
the BS always attempts to decode the third signal layer, if the two-UE
collision has been discovered. Besides, the cases considering SORTE
may cause performance degradation at a relatively low OSNR, due to
the incorrect estimation of . As the figure shows, the success
probability gets saturated at around with , while
the performance can be slightly improved when increases from
400 to 800. For the received signal with a three-UE collision, i.e., the
signal superimposed by three collided signals after ZFB detection, the
SINR of the superimposed signal to be canceled should be sufficiently
high, which may not be practical as there is no coordination among
the UEs to optimize their distances from the BS. However, the
 probability can still be improved at the expense of sacrificing the code
rate , i.e., decreasing .

4.5 Conclusion and Remarks

In this chapter, we introduced a massive MIMO-based grant-free RA with a
collision resolution protocol to support massive access for MTC
communications in IoT networks. To resolve preamble collisions, a HOCs-
based AMC approach was studied to effectively recognize the collision
cases. By exploiting the channel hardening and favorable propagation
characteristics of massive MIMO, collided signals can be first processed as
superposition coded signals by employing the beamforming, and then were
successfully decoded using the SIC. In addition, theoretical analyses on the
success probability of the collision resolution with CB and ZFB were
studied.
For collision resolution, there are some other issues for future work. For
example, the path loss compensation fraction can be optimized to
maximize the success probability and the analysis to see the impact of inter-
cell interference on the success probability is another meaningful extension.
Besides, in order to significantly improve the performance of collided
signal detection, we will introduce some important parameter estimation
methods in Chapter 5.

OceanofPDF.com
5
Multiple Delay Estimation for Collided
Signals
In machine-type communication (MTC), active devices randomly
choose resource blocks (RB) and access to an access point (AP) [65,
66]. The arrival times of multiple signals can be different as devices
are randomly located in a cell. Although multiple devices may choose
the same RB, their signals can be detected by taking into account the
difference between the arrival times of multiple signals, which is
called non-orthogonal random access (NORA) [67]. NORA utilizes the
different arrival times to identify multiple devices with the identical
preamble, while the successive interference cancellation (SIC) can be
employed to decode multiple signals when they coexist in the same RB
[68–70]. Thus, the detection of the event of collision via the estimation
of round-trip delay (RTD) in NORA becomes important. In this
chapter, we introduce different approaches to estimate of RTD of
multiple signals and determine the number of multiple signals [71, 72].

5.1 System Model

Consider an random access (RA) system consisting of an AP and multiple
devices in a single-cell, with radius , and assume that devices are
randomly located in the cell.
We consider an orthogonal frequency division multiplexing (OFDM)
system where subcarriers are divided into multiple groups. In particular, we
assume that there are RBs and each RB consists of subcarriers
(thus, the total number of subcarriers is )), where is the
length of the pilot signal and is the length of cyclic prefix (CP), which
is used to avoid the inter-symbol interference (ISI). A device with data
packets, which is called an active device, sends a pilot signal of length
in a randomly chosen block among RBs, and then the AP is able to
estimate the channel state information (CSI) for coherent detection of the
following data transmissions.
In MTC, the RTD varies from one device to another. The maximum RTD,
denoted by , is proportional to the cell radius . That is,

where is the speed of light. In addition, denote by the maximum

delay spread and by the sampling interval. Then, the maximum length
of the effective channel impulse response (CIR), which includes the delay
spread and RTD, is given by

where . For example, if the cell radius is 1 km, the maximum RTD
becomes 6.67 s. If the maximum delay spread is 5 s (which is a value
according to the extended typical urban (ETU) model in [73]) and
s (for a bandwidth of 1 MHz), becomes 12.
Next, we consider the detection of collision based on RTD estimation, we
mainly focus on one RB for RA, saying RB 1. In addition, considering the
transmission efficiency, we assume that . Suppose that there are
active devices that choose RB 1. For , within the RB, the
interference is inevitable and we have NORA among active devices.
Denote by the RTD of active device (or the arrival time of the
transmitted signal by active device ) in the discrete time, i.e., after
sampling, where . Here, is quantized in time1
and is the maximum RTD in the discrete time. In addition,
denotes the maximum tap number of active device ’s CIR due to delay
spread in the discrete time, and assume that
In Fig. 5.1, we consider a scenario of an AP and multiple devices in an RA
system. For example, MTC devices for the industrial manufacturing
monitoring can be considered. Suppose that there are two active devices.
The effective CIRs with different RTDs are illustrated.
Let denote the pilot sequence of length transmitted
by a device to allow the AP to estimate its CSI. Suppose that the AP has
receive antennas. Then, the received signal at the AP (through RB 1)
through the th antenna in the frequency-domain after removing CP
becomes

(5.1)

where is the (frequency-domain)

diagonal channel matrix from device to the th antenna at the AP and
is the background noise. Here, the channel
coefficient corresponding to the th subcarrier is given by
Figure 5.1 System model of two active devices with different RTDs.

where is the effective CIR of device (to the th antenna at

the AP), and . For convenience, let
. Then, can also be expressed as

(5.2)

where . In order to detect collision, the AP

is to determine the value of from with known .
Unfortunately, the collision detection approaches in the frequency-domain
may not provide a reliable performance, because the devices can have
different RTDs. That is, for an active device with a long RTD,
becomes 0, which means that the signal power from this active device is not
taken into account by the correlator-based detector whose output is
unless the device’s RTD is 0. To avoid this problem, the approach in the
time-domain can be considered as in [74].
Let . Then, can be expressed as

(5.3)

where for and ,

and the received signal becomes

(5.4)

where and , which is referred to as the

composite channel vector. The arrival time of the transmitted signal by
active device is , and let denote the CIR of length without
RTD from device to the th antenna at the AP. Then, can be
given by

(5.5)

For convenience, let , i.e., denotes

the th element of . From Eq. (5.4), we have
 (5.6)

where is the pseudo-inverse of , which is given by

.
Then, we can consider an optimal estimation of RTDs, { }, based on the
maximum likelihood (ML) approach.

5.2 Multiple Delay Estimation

5.2.1 ML Detection Algorithm
In the following, we focus on the ML approach to obtain the estimates of
RTDs, which can be used for the determination of the number of active
devices or signals, , under the assumptions below:

A1) For , multipath Rayleigh fading channels are considered,

i.e.,

(5.7)

where is the variance of . Furthermore, the ’s are

independent.
A2) Since the RTD, , is proportional to the distance between the AP
and active device , is given by

where is the normalized power delay profile (PDP), is the

path loss exponent, and is the distance function of the RTD, ,
i.e., represents the distance from an MTC device to the AP when
the RTD of the MTC device is .

From Eq. (5.4), the conditional probability density function of for

given can be written as
 (5.8)

where is the normalization constant. From Eq. (5.5), we can see that
is determined by . In addition, according to A1 and A2, we
have

(5.9)

where is constant, is given by

and is the pseudo-inverse of , which is given by

(5.10)

Note that depends on . Then, since the ’s are independent (for

given ) under A1, the likelihood function of
becomes
 (5.11)

where

(5.12)

To derive , we define the diagonal matrix as

(5.13)

In addition, define

(5.14)

Note that depends on according to A2.

Lemma 5.1 Let

(5.15)

where . Then, we have

 (5.16)

Proof: For convenience, we omit the antenna index . It can be shown

that

where and . Then, we have

(5.17)

which completes the proof.

According to Lemma 5.1, the ML estimation of RTDs can be formulated as

(5.18)

where is a matrix that is a function of the RTDs, , as in Eqs.

(5.10) and (5.13). An exhaustive search can be considered to find the ML
estimates of the RTDs in Eq. (5.18). In this case, since
, the complexity order becomes .
Thus, if is large, the complexity can be high.

5.2.2 CAVI Detection Algorithm

As a method of machine learning for approximating probability densities,
variational inference (VI) is widely used to approximate posterior densities
for Bayesian models. The principle of VI is to perform optimization rather
than sampling, where a family of densities is first posited, and then a
member of the family that is close to the target density can be found.
Therefore, we apply VI to the ML estimation problem in Eq. (5.18) in order
to obtain approximate solutions with low computational complexity.
We assume that the RTDs, , are independent random variables, which
are seen as the latent variables. Let , which is
referred to as the variational distribution, and let ,
where . In VI [25, 75], the estimation of becomes the
optimization problem formulated as

(5.19)

where represents the collection of all possible distributions of ,

is the posteriori probability of , and is the
Kullback–Leibler (KL) divergence defined as

Then, from , the estimate of from can be

found as
 (5.20)

The minimization of KL divergence in Eq. (5.19) is equivalent to the

maximization of the evidence lower bound (ELBO) [25], which is given by

(5.21)
where the expectation is carried out over . Let
. Then, for given , it can be
shown that

(5.22)

where the expectation, denoted by , is carried out over . The

coordinate ascent variational inference (CAVI) algorithm [25, 43] is to
update , where and , with fixing
other variational distributions, , where , through iterations.
In Eq. (5.22), the term on the right-hand side (RHS) can be further shown as

(5.23)

Note that if is equally likely, becomes a constant term.

Since , it can be shown that

(5.24)

Thus, for the CAVI algorithm, we need to find a closed-form expression for
. To this end, we can find an expression of
as follows.
Lemma 5.2 Suppose that the ’s are constant-modulus signals, i.e.,
is same for all . Let , . In
addition, we define

(5.25)

which is also the th diagonal element of if . Then, we have

(5.26)

where

(5.27)

Here, (for constant-modulus signals).

Proof: From Eq. (5.16), we have

(5.28)

From Eq. (5.14) and according to A2, we can show that

 (5.29)

where the constant term is the term that does not depend on .
Using the matrix inversion lemma, it can be shown that

(5.30)

where

(5.31)

Since , it can be further shown that

From this, after some manipulations, we can show that

 (5.32)

Substituting Eq. (5.32) into Eq. (5.30), and then into Eq. (5.28), we have

(5.33)

where the second equality is due to the fact that is diagonal.

Substituting Eqs. (5.29) and (5.33) into Eq. (5.28), we can obtain Eq. (5.26),
which completes the proof.
For given variational distributions of , where , we need to
carry out the expectation . From Eqs. (5.23), (5.24), and
(5.26), if is equally likely, we have
 (5.34)

Unfortunately, although the ’s are assumed to be independent, it is

troublesome to obtain . Thus, we resort to an approximation
as

(5.35)

where and the approximation is based on

for . In Eq. (5.35), with the variational distribution of
, i.e., , we can have
 (5.36)

Thus, an approximation of the term on the RHS in Eq. (5.22) can be found
and the CAVI algorithm can be carried out with the updating rule as

(5.37)

where . Note that the term in

due to is independent
of and can be treated as a constant in the updating rule in Eq. (5.37). In
particular, with the step size , the CAVI algorithm can be
implemented using the sequential updating rule as

(5.38)

where

(5.39)

Here,

(5.40)

where represents the expectation with the th variational distribution

.
To check the complexity of the CAVI algorithm, we mainly focus on the
number of multiplications. The complexity of the CAVI algorithm in Eq.
(5.38) depends on the complexity to find and that to obtain
 , which is to carry out the expectation in Eq. (5.36) with the

updated variational distribution, . The number of multiplications of

the former is proportional to provided that is available, while
that of the latter is proportional to . To update the variational
distribution for each in an iteration, we need divisions. Note
that since is obtained prior to iterations, we ignore the complexity to
obtain . Consequently, during each iteration, it can be shown that the
complexity order is . Thus, we can
see that the complexity of the CAVI algorithm per iteration is proportional
to .

5.2.3 MCMC Detection Algorithm

According to the ML function, the posterior probability of is given by

(5.41)
Based on the ML estimation, we use the Markov chain Monte Carlo
(MCMC) algorithm to find an approximate solution by obtaining samples
of from the target distribution [76, 77].
We use the Gibbs sampler [78] and the sampling operation can be carried
out recursively as [71]

(5.42)

where denotes the th sample. Letting

, the sampling rule can be
formulated as
 (5.43)

where follows the uniform distribution, i.e., ,

According to Eq. (5.43), the update of can be found as

(5.44)

where . After taking the th sampling, the

samples can converge to a peak of the likelihood function. The
Markov sequence can be obtained with
.
However, since the proposed MCMC algorithm may get stuck at a local
optimum, the mutation operator in the Genetic Algorithm (GA) is beneficial
to protect against this problem by maintaining diversity in the population
[79]. Thus, we apply the principle of the mutation operator in GA to the
MCMC algorithm to improve the probability to converge to the global
optimum. The sampling rule and the sampling process of can be
regarded as the fitness function and selection process, respectively. Note
that the fitness function is used to evaluate the quality of samples in each
step, which is
 (5.45)

In the proposed GA-MCMC approach, the value of and its fitness

value are recorded. Then, a mutation operation is carried out to ,
which is then updated entirely from the previous ones. After mutation,
can be used as for next samplings. Finally, with the largest
fitness value of the estimation will be selected through times of
iterations. Let represent the set of and represent the fitness
value . The algorithm of GA-MCMC is presented in Algorithm 5.1.
Then, we compare the complexity of the GA-MCMC approach with the ML
approach.
In the GA-MCMC algorithm, since the complexity of depends on
, considering the number of iterations and mutations, the
complexity order becomes . It can be observed that the
complexity of ML algorithm grows exponentially while the complexity of
GA-MCMC grows linearly with . Thus, if is large, the complexity
saving can be obviously significant.

5.3 Signal Number Estimation and Channel

Estimation
In the field of signal processing, the most classical algorithms to estimate
the number of targets are information theoretic criteria, including the
Akaike Information Criterion (AIC) [80], the Bayesian Information
Criterion (BIC) [81], and the Minimum Description Length (MDL)
criterion. The MDL criterion is proposed to find the model most likely in
the Bayesian sense, which minimizes the total description code-length,
comprised of the data likelihood and the model complexity.
To address the problem of finding the number of active signals in RTD
estimation, we focus on the determination of the number of active devices
based on the MDL criterion. According to the MDL principle, the model
which yields the shortest description length of the data, should be selected
from a given data set and a family of probabilistic models.
Suppose that the RTDs are estimated for a given . Let denote the
estimate of . Then, from , the support of can be obtained,
which is denoted by . From Eq. (5.6), the non-zero elements of can
be estimated by taking the elements of corresponding to the
support, i.e., , or
 (5.46)

where represents the th element of . If the inter-arrival time

differences or differences of RTDs are greater than or equal to , i.e.,
for , from , the CIRs of active devices, ,
can be obtained as there is no overlapping between the supports of .
When the inter-arrival time differences are smaller than , the CIR of each
active device may not be clearly estimated from . However, we are still
able to determine the number of active devices, , since the ’s are
required to form an objective function to determine . With , and
(for given ), denote by the
generalized likelihood function of . Then, based on the MDL criterion
[82], the problem to determine can be formulated as

(5.47)

where and represent the number of parameters for given

and the number of data samples, respectively, and is the
maximum number of active devices. For given , since there are RTDs
and channel coefficients (for , where each element is
complex-valued), we have

In addition, since the elements of are complex-valued, becomes

. The second term on the RHS in Eq. (5.47) is the penalty function
and increases with , which is seen as the model order, whereas the first
term on the RHS in Eq. (5.47), i.e., the negative likelihood function, is
generally a non-increasing function of .
5.4 Simulation Results
5.4.1 CAVI Simulation Results
We present simulation results with the distance function as

(5.48)

The gain of this distance function becomes unity when for

normalization and the path loss exponent is set to 3. For the PDP, we
consider the exponential decay function as

where the exponent is set to 0.8. It is also assumed that is an

independent binary phase shift keying (BPSK) symbol, i.e.,
, and the signal-to-noise ratio (SNR) is defined as

(5.49)

Thus, from Eq. (5.48), this SNR is normalized for the device of the
maximum RTD, i.e., . We consider a random initialization for the
CAVI algorithm where is randomly chosen from .
Prior to presenting the main simulation results, we compare the
performance of the CAVI algorithm with that of the conventional method in
[74]. We consider the case of with , , , and
SNR dB. Furthermore, we assume that the RTDs are fixed as
and so that there is no CIR overlapping. For the CAVI
algorithm, we have . For the correlator-based detector
[74], we choose the largest elements of to estimate RTDs.
In particular, we obtain the estimates of RTDs of the correlator-based
detector as
 (5.50)

The normalized frequencies2 of are presented in Fig. 5.2, where we can

clearly see that the performance of the CAVI algorithm outperforms that of
the conventional correlator-based detector. It is also shown that the CAVI
algorithm can successfully estimate the correct RTDs ,
with a high probability, i.e., 0.98.
Figure 5.2 Normalized frequencies of of the CAVI algorithm and the
correlator-based detector when ( ), ,
, , , , and .
We now present the main simulation results with random RTDs to evaluate
the performance of the CAVI algorithm under various conditions. The RTD,
, is assumed to be uniformly and independently chosen from
at random, i.e., ,
. In addition, in order to analyze the RTD estimation
performance under the complex condition of different number of active
signals, we consider the mean of the number of successfully recovered
RTDs MNSRD as a performance measure for the RTD estimation, which is
given by
where represents the set difference ( ).
Clearly, if , we can see that the RTDs are correctly
estimated with a high probability. In order to take into account the impact of
the inter-arrival time difference on the MNSRD for the case of , we
consider two cases:

Large inter-arrival time difference: The case of ,

i.e., no CIR overlapping;
Small inter-arrival time difference: The case of ,
i.e., CIR overlapping.

Note that in the case of small inter-arrival time difference or CIR

overlapping, the estimation error of RTD can be high due to the
interference.
In Fig. 5.3, with , the performance of the CAVI algorithm for
different numbers of iterations, denoted by , is shown when
and , , , dB, and . It is shown
that the (conditional) MNSRD of CAVI for the case of large inter-arrival
time difference or no CIR overlapping decreases with the number of
iterations and approaches that of ML, i.e., the CAVI algorithm can provide a
near ML performance if with a sufficiently large number
of iterations (e.g., ). On the other hand, the (conditional)
MNSRD of CAVI for the case of CIR overlapping (after convergence) is
slightly larger than that of ML. This shows that the CAVI algorithm cannot
provide a near ML performance if .
Figure 5.4 shows the performance of the CAVI algorithm to estimate RTDs
for different values of step size , when , , ,
, , dB, and . It is shown that the
(conditional) MNSRD of CAVI for the case of large inter-arrival time
difference decreases with and approaches that of ML.
From Figs. 5.3 and 5.4, we can see that the CAVI algorithm can provide a
reasonably good performance with and , and its
performance can be close to that of ML for the case of no CIR overlapping,
i.e., . Furthermore, even if , the CAVI
algorithm has a slightly worse performance than ML.
The impact of SNR on the performance of the CAVI algorithm and the ML
approach is shown in Fig. 5.5 when , , , ,
, and . In general, the MNSRD decreases with
SNR, and the conditional MNSRD for the case of large inter-arrival time
difference approaches 0 in both the ML approach and the CAVI algorithm.
However, for small inter-arrival time difference, i.e., , the
MNSRD does not approach 0 although the SNR is high (e.g., 30 dB) due to
the interference from the other active device, although the background noise
is negligible.
Figure 5.3 Performance of the CAVI algorithm to estimate RTDs for
different numbers of iterations when , , , ,
, , and .
Figure 5.6 shows the performance of the CAVI algorithm and the ML
approach for various numbers of paths, , when , ,
, dB, , and . It is shown
that the MNSRD increases with . Note that since , if ,
CIR overlapping cannot be avoided. Thus, when , the conditional
MNSRD for the case of large inter-arrival time difference cannot be
obtained, which is however shown to be 0 in Fig. 5.6 for convenience.
Furthermore, since the probability that two CIRs are overlapped, increases
with , the average MNSRD becomes the conditional MNSRD for the
case of small inter-arrival time difference as increases. Clearly, we can
see that a large is not desirable for a reliable estimate of RTD.
Figure 5.4 Performance of the CAVI algorithm to estimate RTDs for
different values of step size , when , , , ,
, , and .
The impact of the number of antennas, , on the performance, is shown in
Fig. 5.7 when , , , dB, , and
. In Fig. 5.7, we can see that the conditional MNSRD
for the case of large inter-arrival time difference can approach 0 as
increases, whereas the conditional MNSRD for the case of small inter-
arrival time difference is almost independent of . Since the interference
due to CIR overlapping exists regardless of the number of antennas, the
increase of does not significantly help reduce the MNSRD in the case
of small inter-arrival time difference.
From Figs. 5.5–5.7, we can see that the performance can vary considerably
depending on CIR overlapping. In general, the RTD estimation is quite
reliable if there is no CIR overlapping, and the channel estimation can be
successfully carried out as the CIR of each active device can be recovered
without any interference. In Fig. 5.8, the conditional normalized mean-
squared error (MSE) of the estimated composite CIR in Eq. (5.46), defined
as

Figure 5.5 Performance of the CAVI algorithm and the ML approach for
different values of when , , , ,
, and .
is shown for the case of no CIR overlapping when , ,
, and . The normalized MSEs of the channel
estimate as functions of the number of multipaths, , SNR, and the
number of antennas, . In general, we can see that the channel estimation
can be reliably performed as the normalized MSE is close to with
moderate SNRs (e.g., greater than or equal to 10 dB). As expected, the
normalized MSE of channel estimate decreases with SNR, , and .
As mentioned earlier, we have considered the RTD estimation to determine
the number of active devices or signals, . Based on the MDL criterion in
Eq. (5.47), simulations are carried out to find the conditional probability of
correct estimation of for given , i.e., , for
different values of SNR when , , , and .
The CAVI algorithm is used with to obtain the
estimates of RTDs with 4 possible values of , i.e., .
From the simulation results presented in Fig. 5.9, we can see that can be
reliably estimated when , i.e., there is no active device, as a
reference, for a wide range of SNR. However, as increases, the
conditional probability of correct estimation of decreases, because the
event of CIR overlapping becomes more frequent as increases. In
addition, we observe that for a larger , a higher SNR is required to
successfully estimate with a high probability. It is noteworthy that in
practice, since the number of active devices per RB might be small with a
high probability,3 it is desirable to have a high (conditional) probability of
correct estimation of when is small [67]. From this perspective, we
can see that the results in Fig. 5.9 can be reasonable because the average
probability of an accurate estimate of can be high.
Figure 5.6 Performance of the CAVI algorithm and the ML approach for
different numbers of multipaths, , when , , ,
, , and .
Figure 5.10 (a) shows the conditional probabilities of correct estimation of
for given as functions of the number of multipaths,
, when , , dB, , and
. While the conditional probability of correct
estimation of decreases with as mentioned earlier, we can observe
that the conditional probability of correct estimation of decreases with
. Since the probability of the event of CIR overlapping increases with
for a fixed , it is more difficult to correctly decide RTDs, which leads
to the increase in the conditional probability of correct estimation of
with . To overcome this problem, we can increase the length of CP,
which can effectively decrease the probability of the event of CIR
overlapping for a fixed , at the cost of more bandwidth, i.e., with more
subcarriers. We can confirm the performance improvement by increasing
in Fig. 5.10 (b).

Figure 5.7 Performance of the CAVI algorithm and the ML approach for
different numbers of antennas, , when , , ,
, , and .

5.4.2 MCMC Simulation Results

Since the GA-MCMC can provide the optimal solution to the ML approach
asymptotically, we analyze the performance of the GA-MCMC method.
Suppose that is an independent BPSK symbol, i.e., . Let
and we define
 (5.51)

Figure 5.8 Conditional normalized MSE of the estimated composite CIR

(without CIR overlapping) when , , , and
: (a) as a function of with and ;
(b) as a function of with and ; (c) as a function
of with and .
In addition, in order to analyze the RTD estimation performance, we
consider the probability of the number of correct recovered RTDs ( ),
which is given by
 (5.52)

where denotes the estimate of .

There are two cases for the estimation of :

the completely correct estimation of ( ),

the partially correct estimation of ( and ).

For the first case, let denote a random vector in , where

. For convenience,
we use the logarithm of the likelihood function, while the correct estimation
probability is shown as
Figure 5.9 Conditional probabilities of correct estimation of for given
as functions of when , , ,
, and .

(5.53)
where ,
, and , represent the th element of , ,
respectively. For the second case, let denote the set of with
and denote the number of elements in the intersection (
), then we have
Figure 5.10 Conditional probabilities of correct estimation of for given
when , , , and
: (a) as functions of the number of multipaths, ,
when ; (b) as functions of the length of CP, , when .

(5.54)

Thus, can be simplified as

(5.55)

where is the weight parameter of and . For

example, while , we have .

Assuming that is randomly chosen from , we present

the simulation results with respect to to illustrate the performance of the
proposed method. For comparisons, the following methods are considered:

MCMC: RTD estimation using the MCMC approach with Gibbs

Sampler;
 GA-MCMC: RTD estimation using the GA-based MCMC approach
with Gibbs Sampler;
ML: RTD estimation using the ML approach with Exhaustive Search.

Figure 5.11 shows of the three methods with iterations. It is shown that
of GA-MCMC for the case of large inter-arrival time difference first
increases dramatically and then becomes saturated as the number of
samples increases. It also indicates that GA-MCMC has a higher than
MCMC, which provides the near ML performance.
The impact of the number of mutations on is shown in Fig. 5.12. It
can be observed that of GA-MCMC for the case of large inter-arrival
time difference approaches that of ML as increases. It shows that by
carrying out the mutation operations, GA-MCMC is able to find a global
optimal solution with high probability.
Figure 5.13 shows the comparison of obtained from analytical results
and numerical results. It is shown that the analytical results match well with
the numerical ones. It can be shown that increases with SNR in general.
In particular, approaches 1 in GA-MCMC and ML for large inter-arrival
time difference. Therefore, the GA-MCMC method can improve the
performance, especially under the condition of large inter-arrival time
difference, i.e., no CIR overlapping.
Figure 5.11 with numbers of samples under , ,
, , and .
Figure 5.12 with numbers of mutations under , ,
, , and .
Figure 5.13 with under , , , ,
and .

5.5 Conclusion and Remarks

In this chapter, we introduced signal estimation methods to provide the
estimates of the arrival time of multiple signals in NORA. We first
considered the ML approach to estimate RTDs, and then derived low-
complexity approaches based on VI and GA-MCMC, respectively. With the
estimated RTDs, the determination of the number of active devices or
signals has been studied with the MDL criterion.
While we only focused on the determination of the number of multiple
signals, , through the estimation of RTDs, for collision detection, there
are other issues to be studied in the future. For example, the impact of the
error of RTD estimates on the performance of NORA can be investigated.
To mitigate CIR overlapping, access control that assigns devices to RBs
depending on their RTDs might be an interesting topic.
After discussing Internet of Things (IoT) multiple access detection with
signal collision, we will introduce some special cases in Chapter 6, e.g.,
backscatter communications, where IoT devices transmit their data through
harvesting and reflecting the energy of radio frequency sources, which may
cause the signal superposition.

Notes
1 For example, is given by , where is the actual RTD of
device .

2 The normalized frequency of is .

3 The maximum probability of is 0.056, when the total number of

arrived devices is .

OceanofPDF.com
6
Detection and Division for Backscatter
Signals
In machine-type communication (MTC), there might be power-limited
or battery-free devices (namely parasite devices) such as radio
frequency identification (RFID) tags, which can use backscatter
communications to transmit their data through harvesting and
reflecting the energy of radio frequency sources (namely host devices).
Two types of backscatter communication techniques are most widely
applied in MTC systems, which are the bistatic backscatter
communication system (BBCS) and the ambient backscatter
communication system (ABCS). Since radio frequency sources of
BBCS are dedicated carrier emitters, e.g., power beacons, the cost can
be relatively high. While ABCS uses the already-available radio
frequency (RF) sources for backscatter communications, it can reduce
costs and improve spectrum efficiency as no dedicated energy supply
is deployed in the system. However, the unpredictability and dynamics
of the ambient signal make it a direct interference to the receiver,
enabling the detection of signals to play a pivotal role in the ABCS. In
this chapter, we introduce an approach to detect the presence of
parasite devices and separate parasite signals from host signals [83].

6.1 System Model

In most cases, frequency ranges for backscatter communications with
passive RFID tags are not high (e.g., high frequency (HF) of 13.56 MHz or
ultrahigh frequency (UHF) of 856–960.56 MHz) [84], while there is also
microwave RFID operating at 2.45 GHz [85]. Thus, to support backscatter
communications for battery-free devices such as RFID tags within MTC
[86, 87], it would be desirable for 5G systems to operate at a sub-gigahertz
band. In fact, in long-term evolution (LTE), Narrowband Internet of Things
(NB-IoT) operates at sub-gigahertz frequency ranges, i.e., 700, 800, and
900 MHz. It is also noteworthy that long-range backscatter communications
become possible using multiple antennas or multiple-input multiple-output
(MIMO) [88–92].

Figure 6.1 Backscatter communication system.

Firstly, we describe the structure of backscatter communication systems.
Suppose that a backscatter communication system consists of a base
stations (BS) and multiple devices. It is assumed that the BS is equipped
with antenna elements, where is usually large for massive MIMO
[18]. As illustrated in Fig. 6.1, there are two different types of devices: host
devices and parasite devices. Considering the security of information, host
devices help multiple parasite devices transmit signals to the BS directly
[93, 94].
In order to minimize signaling overhead and access delay, the grant-free
random access (RA) scheme is used in backscatter communication systems,
which skips the resource scheduling of BS and enables the simultaneous
transmissions of multiple signals [95]. The process is presented as follows.
Firstly, the BS broadcasts the system information, from which each host
device or each parasite device estimates its channel gain, and then employs
fractional power control (FPC) to compensate for the path loss, where FPC
is an open-loop power control technology used in the LTE cellular
networks. It is noteworthy that to facilitate the subsequent analysis, the path
loss compensation factor of FPC is set to 1. Once a host device or a parasite
device attempts to access the network, it randomly picks a preamble from a
predefined preamble pool assigned to the corresponding cell. Followed by
the data, the preamble is transmitted to the BS through the physical random
access channel (PRACH) [55].
In grant-free RA, there are two phases. The first one is the phase of
preamble transmission and the second one is the phase of data transmission.
In the following, we present a transmission model with host devices and
parasite devices for grant-free RA in detail. The first one is the phase of
preamble transmission and the second one is the phase of data transmission.

1. Grant-free RA with host devices: In the phase of preamble

transmission, a host device with data to send firstly becomes active,
and then selects a preamble from the preamble pool randomly, which
is shown as

(6.1)
Here, denotes the th preamble, which is an orthonormal
sequence of length . Thus, we assume that . After sending a
preamble, the host device sends its data packet in the phase of data
transmission. For convenience, we assume that devices are
synchronized and data packets of host devices have the same length.
In the phase of preamble transmission, the signal received at the BS
can be given by

(6.2)

where denotes the number of active host devices,

denotes the channel vector from the th host device to the BS,
represents the index of the preamble chosen by the th host device,
represents the transmit power of the th host device, and
is the background noise at the th antenna
and the th preamble symbol duration.
In the phase of data transmission, each active host device transmits its
data packet of length to the BS. The signal received at the BS
becomes

(6.3)

where denotes the data packet from the th host device and
denotes the background noise at the th
antenna in the th data symbol duration. Throughout the paper, we
consider the following assumptions.
A1) As in [96], is decided to be inversely proportional to the
distance between the BS and the th active host device via
power control so that

(6.4)

where represents the (average) received signal power and

is independent for all , i.e., Rayleigh fading
is assumed for small-scale fading. This power control policy is
known as statistics-aware power control, which is easier to be
implemented as it only needs to know the distance between the
device and the BS. Since host devices may meet local scatterers
and obstacles, we assume that represents a Rayleigh fading
channel, which can be expressed as

(6.5)

where is the distance between the host device and the BS,
is the reference channel gain, and is the path-loss
exponent.
 A2) A symmetric signal constellation is used for modulation. As a
result, we have and , i.e., the symbol
energy is normalized.
Then, the BS can use the correlator to estimate the channel vector,
which is shown as

(6.6)

where is the Kronecker delta (i.e., if , and 0

otherwise) and represents the index set for active host devices
that choose preamble . If the th active host device is the only
device that chooses preamble , i.e., or , thanks to the
orthogonality of preambles, it can be shown that

(6.7)

where .
In order to decode the data packet from the th active host device,
conjugate beamforming can be applied and the output of the
beamformer can be given by

(6.8)

Therefore, if is sufficiently larger than , the data

packet from the th active host device can be successfully decoded.
2. Grant-free RA with parasite devices: As mentioned earlier, suppose
that a host device can serve a number of parasite devices, where
parasite devices can be power-limited devices. We assume that when a
host device does not have data to send, it can help a parasite device
with data to send to the BS.
For convenience, we ignore the index of parasite devices in this part.
Then, as illustrated in Fig. 6.2, the received signal at the BS during
the preamble transmission phase can be re-written as

Figure 6.2 Ambient backscatter communications for parasite devices.

(6.9)

Here, is the activity variable for the th parasite

device, which becomes 1 if the parasite device has data to send, or 0 if
it does not, is the reflection coefficient of the parasite device
corresponding to the th host device, denotes the small-scale
fading channel vector of the parasite device (to the BS) associated
with the th host device, which is referred to as the th host device-
 aided parasite device, and represents the Rayleigh fading channel
from the parasite device to the BS, which can be expressed as

(6.10)

where is the distance between the parasite device and the BS.
Note that the BS does not know whether or not the th host device’s
signal is used as an ambient signal for its parasite device. Thus, is
unknown to the BS.
The channel estimate for the device that uses the th preamble is
given by

(6.11)

where the last equality is valid if .

It is assumed that when a host device helps its parasite devices, it transmits
a void signal, which is fixed and known to the BS. In particular, let the void
signal be an unmodulated signal, which is . With parasite
devices, the received signal during the data transmission phase is re-written
as

(6.12)

where
 (6.13)

Here, represents the data block transmitted from the parasite device
corresponding to the th host device. For backscatter communications, on–
off keying (OOK) is used by parasite devices and we have
in this case. It is also assumed that is independent and

(6.14)

In addition, there are a few remarks as follows:

A parasite device can send its data to the associated host device so that
the host device can forward the data as a relay node to the BS.
However, in terms of data privacy, this approach may not be desirable,
since the information of the parasite device is known to its host device.
Furthermore, in backscatter communications, it is not necessary to use
a relay protocol between parasite devices and host devices, as the
parasite node sends its data directly to the BS.

6.2 Central Limit Theorem-based Signal

Detection
In the backscatter communication system, when parasite devices transmit
data, the host devices transmit a void signal instead of data. After data
transmissions by host devices and parasite devices, the BS needs to perform
the following operations [97, 98].

1. Channel estimation: At the BS, a bank of correlators with preamble

sequences are used to estimate the channel vector associated with each
preamble as shown in Eq. (6.11).
2. Detection of preamble collision: As for the preamble collision
detection, an elementary HOCs-based automatic modulation
classification (AMC) approach is introduced in Chapter 4.
 3. Detection of parasite devices: Since the variables are unknown
to the BS, hypothesis testing is to be carried out at the BS with in
Eq. (6.12) for the decision of .
4. Detection of signals: If , a conventional approach can be used
to detect the signals from the th host device. On the other hand, if
, it is necessary to use central limit theorem-based approach to
detect signals.

In the following, we show signal detection methods in backscatter systems.

6.2.1 Activity Detection Algorithm

For convenience, we assume that when detecting the presence
of a generalized parasite device’s signal in the output of the th correlator.
In addition, let and assume that the host device is the device of
interest.

1. Hypothesis testing: As indicated in Eq. (6.12), the interference-plus-

noise at the th antenna of the BS during the th symbol interval
becomes

(6.15)

where

(6.16)

Let and denote the variance of and the

variance of , respectively. Then, we have

(6.17)

For tractable analysis, we assume that is a circularly

symmetric complex Gaussian (CSCG) random variable.
For convenience, we omit the index . Then, the hypothesis test can
be given by

(6.18a)

(6.18b)

Under , the likelihood function of is given by

(6.19)

where and

(6.20)

Here, denotes the data symbol constellation.

Under , the likelihood function of and is given by

(6.21)

where and
 (6.22)

Then, the generalized likelihood ratio test (GLRT) can be carried out
using the test statistic as

(6.23)

Unfortunately, although it is possible to form the GLRT as in Eq.

(6.23), it is difficult to carry out the optimization with
. Thus, we need to resort to a suboptimal
approach.
2. A suboptimal approach based on the Central Limit Theorem
(CLT): To efficiently detect the presence of parasite devices, the
detection statistic is proposed by considering the differences of
channel information and data together with the correlation between
host devices and parasite devices, where and are used. Thus,
let

(6.24)

Moreover, the decision statistic is considered, which is given as

(6.25)

In order to characterize , we need to consider the relationship

between and . As is also a small-scale fading channel
coefficient, according to the assumption of A1, it can be assumed that
, while can be correlated with due to a short
distance between the host device and its parasite devices. Thus, we
consider the following assumption.
 A3) The pair of is independent of other pairs,
, where , and it is given by

(6.26)

where .
Under , the statistical expectation and variance of can be
given by

(6.27)
and

(6.28)

respectively, under the assumption of A2. According to the

assumption of A1, we have . In addition, since
has an exponential distribution, it can be shown that .
Thus, we have

(6.29)

where

(6.30)

Based on the CLT for a large and , the decision statistic

can be approximated using a Gaussian random variable, i.e.,
 (6.31)

Let denote the variance of . can be calculated

by

(6.32)

Since not all the variables are independent and identically distributed
to , we consider the distribution of each variable. Then, can
be written as

(6.33)

Under , it can be shown that

(6.34)

and

(6.35)

Therefore, the decision statistic under can be shown as

 (6.36)

where

(6.37)
and

(6.38)

Finally, the maximum likelihood (ML) detection can be used to detect

the presence of parasite devices, which is

(6.39)

where

(6.40a)

(6.40b)
6.2.2 Signal Detection Algorithm
In this section, we discuss the signal detection of parasite devices when
. Let

(6.41)

and

(6.42)

Then, the test statistic for the detection of parasite devices’ signal is

(6.43)
which can be approximated by a Gaussian random variable.
If , we have

(6.44)

and

(6.45)

If , it can be shown that

(6.46)
and
 (6.47)

Therefore, the ML detection can be shown as

(6.48)

Note that the channel estimation error affects the variance of decision
statistic as less channel estimation error results in a smaller variance, which
is easier to differentiate host signals and parasite signals.

6.2.3 Performance Analysis

In the following, we analyze the detection performance of the presence of
parasite devices and the bit error rate (BER) of the signal detection.
Firstly, we define the probability of error detection, denoted by , in the
presence detection as total probability of an incorrect decision, which
consists of the probability that host device’s data signal is misjudged to the
combination of parasite device’s data signal and host device’s void signal,
and vice versa. Therefore, the probability of error detection of the presence
detection in Eq. (6.39) is given by
 (6.49)

where and denote the probability that the th

parasite device is inactive and the th parasite device is active,
respectively, and denote the probability of the condition that
the host device is misjudged to the parasite device and vice versa,
respectively, and can be written as:

(6.50)

Furthermore, for convenience, let ,

, , and
. The expression of BER of the signal detection is
shown as follows:
 (6.51)

where and denote the probability that the

th symbol of parasite signal is 0 and the th symbol of parasite signal is
1, respectively, and can be written as

(6.52)

6.3 Simulation Results

We assume that there are host devices in the backscatter communication
system and each host device is to help parasite devices. The background
noise is additive white Gaussian noise (AWGN) and the signal-to-noise
ratio (SNR) is defined as the ratio of the received signal at the BS to the
background noise. Note that the multiple-preamble scheme can be used to
ensure partial orthogonality between preambles and greatly reduces the
probability of collision [99]. Thus, we assume that no collisions happen in
our system.
Figure 6.3 versus number of host devices, , with ,
, , , , and .
Firstly, we show the probability of error detection in terms of the number of
host devices, , the number of antennas, , SNR, and the channel
correlation coefficient, , to verify the effectiveness of detection method.
In Fig. 6.3, the probability of error detection of the CLT-based approach for
different numbers of hose devices, , is shown when , SNR
dB, , , , and [100]. It is shown
that the probability of error detection based on CLT increases as the number
of host devices increases because the interference between signals of
different devices increases when there are more host devices. Thus, it
becomes more difficult to differentiate between host devices and parasite
devices.
Figure 6.4 shows the probability of error detection of the CLT-based
approach with SNR when , , , ,
, and . We notice that the detection performance
improves as SNR increases, which indicates that at a sufficiently high SNR,
the BS can carry out successful detection. It also indicates that the CLT-
based approach provides a good presence detection performance when the
number of host devices is up to 15.

Figure 6.4 versus , with , , ,

, and .
Figure 6.5 shows the probability of error detection of the CLT-based
approach with different numbers of antennas, , when SNR dB,
, , , , and . From Fig.
6.5, we can see that the probability of error detection decreases as the
number of antennas increases and the CLT-based approach can provide
a good presence detection performance. Since the advantage of a large
number of antennas can be taken to eliminate the effects of small-scale
fading, the variance decreases as the number of antennas increases.
Figure 6.6 shows the probability of error detection of the CLT-based
approach with the channel correlation coefficient, , when ,
SNR dB, , , , and . It can be
seen that the probability of error detection decreases rapidly, and then tends
to be steady as the channel correlation coefficient increases. Since host
devices transmit a void signal instead of data when parasite devices transmit
data, it is easier to detect the presence of parasite signals from received
signals through the energy of detection statistics when the channel state
information (CSI) of parasite and host devices is more relevant.
Figure 6.5 versus number of antennas, , with ,
, , , and .
Figure 6.6 versus channel correlation coefficient, , with ,
, , , and .
Figure 6.7 versus number of host devices, , with ,
, , , , and .
Then, we show the simulation results with respect to the BER to illustrate
the performance of signal detection from parasite devices. Since the number
of host devices is an important factor in our system, the BER of the signal
detection from parasite devices with different numbers of host devices, ,
is shown in Fig. 6.7, when , SNR dB, ,
, , and . It is shown that the numerical and
theoretical BER of CLT-based approach are monotonically increasing
functions of the number of host devices , which is caused by the
interference between devices.
Figure 6.8 shows the BER of the parasite signal detection with SNR when
, , , , , and
. It is observed that the curves of BER decline rapidly
and then tend to be stable. We also notice that with a sufficiently high SNR,
the BS can achieve reasonable detection performance.
Figure 6.9 shows the BER of the parasite signal detection with different
numbers of antennas, , when SNR dB, , ,
, , and . It shows that as the number of
antennas increases, the BER curves decline in a similar trend, where the
curve with a smaller has lower BER in general. It is noteworthy that
more antennas can reduce the error of channel estimation and make the
CLT-based approach better fit the Gaussian distribution, thereby reducing
the BER.
Figure 6.8 versus , with , , ,
, and .
Figure 6.9 versus number of antennas, , with ,
, , , and .
Figure 6.10 versus channel correlation coefficient, , with
, , , , , and .
Figure 6.10 shows the BER of the parasite signal detection with channel
correlation coefficient, , when , SNR dB, ,
, , and . It is indicted that the BER decreases
as the channel correlation coefficient increases and the in the
presence detection affects the performance in the signal detection.

6.4 Conclusion and Remarks

In this chapter, a backscatter communication system has been proposed,
where parasite devices were able to use the transmitted signals from host
devices as ambient signals for backscatter communications and send data
packets to a BS. Since massive MIMO has been used at the BS, a CLT-
based approach was introduced to detect the presence of parasite devices
and separate parasite signals from host signals.
Backscatter communications allow parasite devices use ambient signals of
host devices to send their data packets, while the technique of non-
orthogonal multiple access (NOMA) allows multiple devices to occupy the
same frequency, time, or code resource. In Chapter 7, we will introduce the
NOMA signal in IoT and investigate the optimization problem of NOMA
systems.

OceanofPDF.com
7
Analysis and Optimization for NOMA Signals
The key concept of non-orthogonal multiple access (NOMA) is that of
allowing multiple Internet of Things (IoT) devices to occupy the same
frequency-, time-, or code-resource [101, 102]. The device having
better channel conditions performs successive interference cancellation
(SIC) [103]. As a remedy, full-duplex (FD) can be invoked by
cooperative NOMA to avoid halving the rate by conventional half-
duplex (HD), where the strong user (SU) with good channel conditions
can act as an FD relay node for the weak user (WU) with poor channel
conditions [104–106]. However, the energy efficiency (EE) of
cooperative NOMA may be degraded due to additional power
consumption incurred at the SUs [107–109]. In this chapter, we
investigate the EE maximization problem of an FD cooperative
NOMA system. More importantly, we introduce the signal-to-
interference-plus-noise ratio (SINR) gap reversal problem of
cooperative NOMA systems, which imposes SIC performance
degradation at the SUs.

7.1 System Model

We consider a NOMA system supporting two IoT devices by a base stations
(BS) as shown in Fig. 7.1. The IoT users have different channel conditions
on the same frequency, which is a common scenario in NOMA systems
[103]. The BS, SU, and WU are equipped with a single antenna for the sake
of having a low hardware complexity. In particular, the SU acts as a decode-
and-forward relay, which cooperates with the WU. FD operation is enabled
at the SU, which can transmit and receive signals at the same time [110].
The channel state information (CSI) is obtained by channel estimation in
the training phase based on the channel’s reciprocity [108]. Explicitly,
during the training phase, the BS estimates the downlink CSI by inferring it
from the channel quality indicator feedback received from the SU and the
WU. Hence, the BS can regularly update the CSI of the BS-SU link and the
BS-WU link. Furthermore, since the SU is aware of its own channel quality
indicator and of the WU, the CSI of the self-interference (SI) leakage
channel and the SU-WU link can also be reported to the BS.

Figure 7.1 NOMA system with two IoT devices.

Since a three-node single-antenna scenario is assumed, based on the state-
of-the-art in overhead research, the signaling overhead (power, bandwidth,
and throughput) maintains low [111, 112]. Hence, we only focus our
attention on the subsequent transmission stage. Let us define the
transmission power allocated to the SU and to the WU at the BS by and
, respectively, where the maximum transmission power available at the
BS is upper bounded by . In addition, we define the transmission
power consumed at the SU by , which is upper bounded by the
maximum available transmission power constraint . Furthermore, to
guarantee that the SU can successfully decode the WU’s signal and subtract
it by SIC, we introduce the constraint of , where and
represent the WU’s SINR observed at the SU and the WU’s signal at its
own end, respectively. In addition, to guarantee the WU’s quality-of-service
(QoS), we have , where is the minimum SINR
requirement of the WU.
Let us define the EE (in bits/Joule/Hz) as the ratio of the system’s
throughput to the total power consumption . To maximize the
system’s EE, we jointly optimize the transmission power , , and .
Accordingly, the EE maximization problem is formulated as

(7.1)

Constraint (C1) indicates that the transmission power allocated at the BS

should be lower than the available transmission power . Constraint
(C2) means that the transmission power of the SU should be lower than its
total available transmission power . Constraint (C3) implies that all
the transmission power should be non-negative. Constraint (C4) guarantees
that the SU can decode the WU’s signal, while Constraint (C5) suggests that
the WU’s QoS should be satisfied.

7.2 Throughput and Power Consumption

Analysis
The optimization problem involves both throughput and power
consumption, which will be introduced in the following, respectively.

7.2.1 Throughput Analysis

At the WU, the received signal arrives both from the BS and the SU, which
is given by
 (7.2)

where and are the th data symbols intended for the SU and
the WU, respectively, is the additive white Gaussian noise (AWGN),
the symbol delay is caused by the processing delay at the SU, is the
channel response from the BS to the WU, and is the channel response
from the SU to the WU. All propagation channels capture the effects of
large-scale and small-scale fading. Since the phase of and is
different, the desired signal replicas arriving from the BS and the SU are
mis-aligned in phase. Hence, the phase of the transmitted signal at the SU
has to be shifted to align the pair of received signals. The channel spanning
from the BS to the WU can be represented as , where
and are the magnitude and phase of the channel , respectively.
Similarly, the channel spanning from the SU to the WU is represented as
, where and are the magnitude and phase of
the channel , respectively. Therefore, the phase difference between the
two channels is calculated as . To align the two signals, the
signal transmitted from the SU is pre-processed as . As a
result, the signal received at the WU becomes

(7.3)

As we can see, the pair of desired signals arriving from the BS and the SU
are now well aligned at the WU,1 and the SINR of the WU at its own end is
given by

(7.4)

The signal received at the SU is given by

 (7.5)

where and represent the channel spanning from the BS to the SU

and the self-interference (SI) leakage channel from the SU’s transmitter to
its receiver, respectively. Still referring to Eq. (7.5), is the SI reduction
factor defined as the ratio of the SI powers before and after SI suppression.
Therefore, the WU’s SINR observed at the SU is given by

(7.6)

where is the residual SI channel. To guarantee successful SIC at the SU,

we have to ensure that . Based on Eqs. (7.4) and (7.6), the
inequality is further derived into
. Since , we have

, As a result, the feasible domain of can be represented in the closed

form

(7.7)

where we have and

. Finally, the SU’s SINR after SIC is calculated as

(7.8)
7.2.2 Power Consumption Analysis
The total power consumption mainly consists of the power amplifier (PA)
power, circuit power of the transmit/receive chains, and the power
consumed by SI cancellation [113].

a. The PA power is closely related to the radiated transmit power and the
drain efficiency of the PA, which is given by [113],
where is the drain efficiency of the PA. Without loss of generality,
we assume that all PAs of the BS and of the users have the same drain
efficiency performance.
b. The circuit power consumption of the transmit/receive chains is
proportional to the number of active transmit and receive chains,
including the power consumed by the digital-to-analog converter, by
the filter and the synthesizer, etc. The dynamic power consumption can
be calculated as , where and denote the circuit
power consumed by the receive and transmit chains of the BS, of the
SU, and of the WU. The multiplier 2 indicates that we have two
transmit chains and two receive chains in the system.
c. For the SU acting as an FD relay, additional power is required for
SI cancellation, which can be modeled by a constant. Generally
speaking, complex SI cancellation schemes dissipate higher power by
their digital-to-analog converter, transmit radio unit, and adders [113].
d. The fixed power consumption represents power consumed by the
power supply, by the active cooling system, etc. This part is
independent of the state of the transmit/receive chains [113]. Finally,
the total power consumption is given by

(7.9)
7.3 Energy Efficiency Performance
Optimization
For simplicity, we collect the transmission power into a vector
and define as the feasible domain confined by the
constraints. Therefore, the EE maximization problem is re-formulated as:

(7.10)

where , , and . The

operator represents the 1-norm of vector. It can been seen that the
objective function is non-convex and (C4) (C5) represent quadratic
constraints, which impose a challenge in terms of obtaining a globally
optimal result within polynomial time. To strike an attractive balance
between the performance and complexity, a low-complexity near-optimal
solution is desirable. We observe that the numerator of the objective
function is the difference between two convex functions. To handle this
challenge, we first reformulate the objective function of Eq. (7.10) as
, where

and .
Therefore, the Frank–Wolfe method of [114] can be adopted, which
approximates by its first-order Taylor series, and iteratively updates
the first-order Taylor approximation along the specific direction that
approaches the original function. The first-order Taylor approximation of
at the th iteration is given by

, where is the value of at the th iteration. Now, the overall

throughput is the difference between a convex function and an affine
function . We state Theorem 7.1 to solve the problem.

Theorem 7.1 The reformulated problem, namely

is jointly quasi-convex with respect to the vector variable in the feasible

domain.

Proof: Let us define the sublevel set of as

. Recall from [113] that is jointly quasi-
convex with respect to the variables in , if is convex for any real
number . For , we have no physical interpretation. By contrast, for
, is equivalent to
. According to our
analysis above, is affine with respect to the variables, while
is strictly jointly convex with respect to the variables.
Therefore, the summation is strictly convex with respect to the variables,
and is quasi-convex with respect to the variables in .

Theorem 7.1 confirms the optimality of the re-formulated objective

function. For the fractional structured quasi-convex problem of Eq. (7.10),
can be associated with a subtract programming
formulation of [114]. Therefore, with the aid
of the equivalent subtract programming, the problem is reformulated as that
of solving with a given .
Now, we handle the constraints (C4) and (C5). Based on Eq. (7.7), (C4) is
equivalent to , where
and

Since the matrix is not a semi-positive definite matrix, its elements are
not confined to a convex set. Therefore, we introduce the Schur
complement [101] of to relax (C4) into

and constraint . On

the other hand, constraint (C5) is equivalent to

, where we have
. After a series of transformations, the
problem becomes

(7.11)

The problem P3 in Eq. (7.11), now, is a standard semi-definite

programming (SDP) problem with a convex set, which can be readily
solved by the CVX package of MATLAB. Finally, a so-called “EE oriented
FD cooperative NOMA” algorithm is introduced for optimizing , which
solves the SDP problem P3 of Eq. (7.11) in the inner layer and updates
in the outer layer. To tighten the first-order Taylor approximation ,
we update the value of in the specific direction that approaches the
original function. Since serves as the upper
bound of the original problem , the value of the
upper bound is reduced iteratively until convergence is reached, as
suggested by Line 4 of Algorithm 7.1. Furthermore, the value of is
squeezed by the bisection method in the outer layer according to the
accuracy factor , and the optimal can be found after a few updates.
Hence, the tightness and convergence of the algorithm is confirmed. The
associated procedure is summarized in Algorithm 7.1.

Let us now consider the complexity of the algorithm. Assume that

is the midpoint of the initial interval, and is the midpoint of
the interval in the th step. Then the difference between and is
bounded by . Given a tolerance factor , the required
number of iterations is given by . In the algorithm, the left
bound can be set to 0. Then, the value of is definitely positive.
Furthermore, a sufficient large value of can be chosen as the right bound
for making the value of negative. Therefore, the function has
opposite signs at the two bounds and thus the classic bisection method
readily leads to convergence. In the inner layer, the CVX solver invokes an
interior-point method to solve the SDP problem, which belongs to the class
of path-following methods and leads to a rapid convergence. Upon denoting
the complexity of the inner layer by , the complexity order of the
algorithm becomes .

Remark 7.1 Generally speaking, a throughput maximization problem

aims for achieving a high throughput regardless of the power consumption.
Hence, the BS only has to satisfy the WU’s SINR threshold, while
allocating all the remaining power to the SU. Then, the SU utilizes a high
transmission power to cooperate with the WU. By contrast, a power
minimization problem aims for reducing the total power consumption, and
thus merely satisfies the WU’s QoS constraint by a minimum transmission
power. As a better choice, EE maximization improves the ratio between the
throughput and power consumption, striking a trade-off between the
throughput maximization and power minimization problems. As suggested
by Theorem 7.1, is quasi-convex (thus the re-formulated EE is quasi-
concave) with respect to the variables in . For a quasi-concave
function, utilizing all transmission power may not lead to the most optimal
EE solution, because the power consumption in the denominator is also
boosted, hence the EE is decreased. For illustration, let us consider a simple
EE maximization problem in Fig. 7.2, where the EE curve is a quasi-
concave function of transmission power. Neither the power minimization
(point (a)) nor the throughput maximization solution (point (c)) achieves the
optimal EE.

Remark 7.2 As suggested by Eqs. (7.4), (7.6), and (7.7), the BS-SU, BS-
WU, and SU-WU distances also affect the SIC failure region. Given a
shorter BS-SU distance, the value of is increased due to the reduced
path loss (PL), and thus a higher cooperative power may be assigned to the
SU hence guaranteeing . A similar trend can be observed for a
shorter BS-WU distance. By contrast, a longer SU-WU distance makes
more likely. However, this also makes the cooperation between
the SU-WU less energy efficient, since additional power is consumed at the
SU without a reasonable throughput improvement. As expected, when the
SU-WU distance tends to infinity, the cooperative power assigned to the SU
tends to 0, as in non-cooperative NOMA systems.

Figure 7.2 The stylized relationship of EE maximization, power

minimization, and throughput maximization problems.

Remark 7.3 Given a lower drain efficiency, the system’s EE is reduced

owing to the boosted PA power consumption. Furthermore, the power
allocation of both the BS and of the SU becomes more conservative due to
the low drain efficiency, since the throughput improvement attained by
increasing the transmission power may not be in line with the increased
power consumption.

7.4 Simulation Results

Now, we discuss the numerical results to verify our analysis using the
parameters given in Table 7.1. The PL model of
[114] and log normal shadowing
model are adopted, which are featured in the 3GPP long-term evolution
(LTE) standards operating at 2 GHz. The small-scale fading is modeled by
Rayleigh fading except for the SI channel at the SU, which is modeled as
Rician fading and having a Rician factor of 5 dB [110]. A pair of typical
cooperative NOMA systems are selected for performance comparison. (a)
In HD cooperative NOMA systems [115], the BS transmits its downlink
signals to the two users (with transmission power
, and ) during the first
half of the time slot, while the SU helps the WU by assigning its full
transmission power in the second half of the time slot. (b) In
the pre-fixed FD cooperative NOMA systems [106], the SU acts as an FD
relay node for simultaneously helping the WU by assigning its full
transmission power , whereas the power allocation at the BS
is pre-fixed as , and .

Table 7.1 Simulation Setup.

Bandwidth 1.25 MHz
AWGN power spectral density dBm/Hz
Drain efficiency of PA 35%
and 100 mW
and 50 and 500 mW
and 100 and 20 mW
SI cancellation amount 80 dB
Distance between the BS and the WU 200 m
Figure 7.3(a) and (b) show the impact of the normalized distance between
the BS and the SU on the EE and the spectrum efficiency (SE), respectively.
As can be seen, the EE-oriented FD cooperative NOMA algorithm
outperforms the others in terms of its EE, whilst exhibiting a higher robust
against the SU’s location as well. This is because the benchmarking
algorithms consume all the available transmission power, which degrades
their EE. A further degrading factor is constituted by their low throughput.
In Fig. 7.3(b), the EE-oriented FD cooperative NOMA algorithm shows the
highest throughput among the three algorithms. This is because the EE-
oriented FD cooperative NOMA algorithm supports the effective
cooperation of the SU and of the WU, hence improving the WU’s
throughput. More importantly, no additional time slots are required for the
cooperation phase, and thus the EE-oriented FD cooperative NOMA
algorithm significantly outperforms the HD cooperative NOMA. As for the
pre-fixed FD cooperative NOMA, the power allocation is unaware of the
relative distance among the communication nodes. Furthermore, the SU
invokes its full transmission power to cooperate with the WU, which may
result in the “SINR gap reversal” and may lead to an unsuccessful SIC
operation at the SU. As a result, the throughput of the pre-fixed FD
cooperative NOMA is also inferior to the EE-oriented FD cooperative
NOMA algorithm. Finally, the throughput of all the algorithms degrades for
a longer distance between the BS and the SU due to the high PL.
Figure 7.3 Impact of the normalized distance between the BS and the SU
on the value of EE and throughput, with . (a)
normalized distance between the BS and the SU and (b) normalized
distance between the BS and the SU.
Figure 7.4 shows the impact of the constraint on the SIC’s failure
probability, which is defined as the ratio of the times of unsuccessful SIC
operation at the SU to the times of simulations. As we can see, the EE-
oriented FD cooperative NOMA algorithm is shown to guarantee that the
SU can successfully decode the WU’s signal by SIC. This is because an
additional constraint is imposed on the power control at the SU by the EE-
oriented FD cooperative NOMA algorithm. By comparison, the SIC failure
probability of the pre-fixed FD cooperative NOMA systems increases
rapidly with the transmission power constraint . This is because for
a higher transmission power used at the SU, the WU’s SINR increases,
while the WU’s SINR observed at the SU decreases due to the higher
level of residual SI. As a result, a higher value of leads to a higher
SIC failure probability at the SU. In case of poor SI cancellation, the SIC
failure probability is substantially increased by the strong residual SI. As for
the HD cooperative NOMA system, its SIC failure probability also
increases upon imposing a higher transmission power constraint due to the
“SINR gap reversal” problem. Since no residual SI is imposed by the HD
operation, its SIC failure probability is lower than that of the FD operation
[106]. However, its SIC failure probability still remains significantly higher
than that of the EE-oriented FD cooperative NOMA algorithm, because the
HD cooperative NOMA neglects that may be higher than owing to
the cooperation between the SU and WU. However, its SIC failure
probability is not affected by the SI cancellation factor in the presence of
HD operation.
Figure 7.4 Impact of the maximum constraint at the SU on the
SIC’s failure probability, where the SU is in the middle of the BS and the
WU.

7.5 Conclusion and Remarks

In this chapter, we introduced an EE-oriented algorithm for FD cooperative
NOMA systems, where the transmission power of both the BS and of the
SU is adaptively allocated, rather than being pre-fixed. More importantly,
the “SINR gap reversal” issue of cooperative NOMA systems is solved by
adaptively confining the transmission power at the SU.
Note that the intrinsic intra-cluster multi-device interference, also deemed
as an underlying resource, has not been fully exploited when designing
NOMA schemes. Considering the high device density in IoT systems, if the
rich multi-device interference can be utilized rather than being mitigated,
the system performance can be significantly enhanced. Therefore, in the
next chapter, we will outline the design principles of NOMA in partially-
coordinated beamforming and fully-coordinated joint transmission.

Note
1 The signal received at the WU has a low time delay, which can be
mitigated by an equalizer or sequence detector. Hereby, we assume that
the signal from the BS and the SU can be readily combined by the WU
[106].

OceanofPDF.com
8
Signal Design for Multicluster Coordination
The imminent Internet of Things (IoT) aims to provide massive device
connectivity and support ever-increasing data demands, making
today’s environment on the edge of a new era of innovations and
changes. In multi-cluster IoT, the devices may suffer severe inter-
cluster interference due to the intensive frequency reuse among
adjacent access points (APs), thus deteriorating their quality-of-service
(QoS). To address this issue, conventional multi-cluster coordination in
IoT provides orthogonal code-, frequency-, time-, or spatial-domain
multiple access for interference management, which yet results in a
waste of resources, especially in the context of explosively increased
number of devices. In this chapter, we review the recent advances in
energy-efficient solutions for multi-cluster coordination relying on
non-orthogonal multiple access (NOMA) and constructive interference
(CI) exploitation techniques. Moreover, their applications for handling
interference management are elaborated at different levels of multi-
cluster coordination. Departing from traditional orthogonal multiple
access (OMA), this line of research provides a new dimension for
inter- and intra-cluster multi-device interference management in IoT.

8.1 Multi-cluster Coordination in IoT

IoT networks aim to provide heterogeneous services for massive devices,
where the use cases can be roughly classified into massive machine-type
communications (MMTC), and ultra-reliable and low latency
communications (URLLC) [116]. Specifically, MMTC is featured by
massive connectivity, low power consumption, and high energy efficiency,
while URLLC is mainly designed to provide high reliability and low
latency for serving mission-critical applications [117]. Evidently, wireless
communications in IoT cause severe access congestion due to insufficient
frequency, time, power, antennas, and other resources. One of the efficient
techniques is to cluster massive devices into a number of small groups, with
a part of resources being shared and reused in different clusters, referred to
as multi-cluster IoT [118]. A typical cluster-based industrial IoT (IIoT) is
depicted in Fig. 8.1. Each cluster is equipped with one or more APs to
provide high-quality wireless transmissions. A number of APs are
connected by a high-speed optical fiber to data collection module and
application server for resource coordination, authentication, and
interference management. Therefore, the wireless interface management
between APs and devices is essential in IoT, from the perspectives of radio
spectrum allocation and interference management [119, 120].

Figure 8.1 Low energy consumption, high energy efficiency, and low
complexity techniques are preferable to enable an energy-efficient multi-
cluster coordinated IIoT system.
According to the coordination level among APs, the multi-cluster
coordination techniques in IoT can be mainly categorized into two modes:
partially-coordinated beamforming (PBF) and fully-coordinated joint
transmission (FJT) [117]. For the PBF mode, channel state information
(CSI) is shared by a number of APs to design cooperative beamforming for
suppressing inter-cluster interference [121, 122]. For the FJT mode, in
addition to CSI, the intended transmission data are shared [123]. Since both
CSI and the corresponding transmission data are available at all the
coordinated APs, the FJT mode allows the coordinated APs to transmit the
same data to the corresponding devices. Therefore, the FJT mode provides
an improved degree-of-freedom (DoF) in terms of spatial diversity gain and
achieves superior system performance than the PBF mode.
Generally, since only CSI is shared among APs, the beamforming design of
the PBF system makes a trade-off between inter-cluster interference
suppression and maximizing the devices’ signal-to-interference-plus-noise
ratio (SINR) within the cluster of interest, with a moderate level of
coordination overhead. On the other hand, the FJT scheme can be regarded
as a virtual multiple-input multiple-output (MIMO) system where, all AP-
device links are utilized to convey data, at the expense of higher
coordination overhead due to the CSI and data sharing mechanism among
the APs [118].
These multi-cluster coordination techniques have distinct advantages and
disadvantages, whereas their implementation in IoT is still restrictive. It is
because:

Due to the massive connections of devices, intra-cluster (multi-device)

interference within each cluster becomes a critical limitation for
ensuring QoS target of devices. In terms of multi-device interference
management, by the existing coordination systems, wireless resources,
such as time and frequency can only be exclusively assigned to one
device on the basis of OMA [124, 125]. Nevertheless, the limitations
of the orthogonality-based methods are increasingly extrusive. Since
IoT is typically loaded with massive equipments, the conventional
orthogonal approaches inevitably require extra time and frequency
resources for providing strict orthogonality, otherwise only part of
devices can be served at the cost of high latency and poor fairness.
Apart from the orthogonal time/frequency access, another approach is
based on the interference alignment (IA) technique, which provides
 spatial orthogonality for multi-device access. Nevertheless, it requires
large-scale antennas to achieve strict spatial orthogonality by carefully
designing the beamforming vector to cancel multi-device interference,
which is cost/power consuming and thus may not be suitable for the
small-sized APs.
In IoT communications, there is an urgent demand for providing high
system throughput while limiting energy consumption. Hence, energy
efficiency (EE) defined as the ratio of throughput to total power
consumption has attracted much attention from vendors and
researchers, where the conventional throughput maximization-oriented
designs in cellular networks become incongruous.

Since NOMA is able to multiplex a large number of devices onto the same
frequency resource, it is particularly suitable for the MMTC application due
to its enhanced multi-device access capability. Its applications in IoT have
been researched in terms of the NOMA-based parallel spaceborne antenna
array calibrations and impulsive noise mitigation. Nevertheless, they merely
applied NOMA for non-coordination IoT, where the system performance
may be significantly impaired by the intra-cluster interference.
Furthermore, note that the intrinsic intra-cluster multi-device interference,
also deemed as an underlying resource, has not been fully exploited when
designing non-orthogonal transmission techniques. Considering the high
device density in IoT, if the rich multi-device interference can be utilized
rather than being mitigated, the system performance can be significantly
enhanced. This inspires the interest in treating the ubiquitous interference in
IoT as a green source with the aid of the spatially non-orthogonal design,
namely constructive interference (CI).

8.2 Multi-cluster Coordination with NOMA

As mentioned above, the conventional multi-cluster coordination methods
(such as the PBF and FJT) provide orthogonal code-, time-, frequency-, or
spatial-domain multi-device access for interference management, which
strictly limit the number of devices being served. To alleviate this, it is
imperative to collaborate NOMA with multi-cluster coordination to provide
flexible multi-device access, which also enables an improved level of DoFs
in resource allocation and interference management. Together with the low
power and high EE demands of IoT communications, the green multi-
cluster coordinated NOMA IoT system has become an essential and
practical issue. In this section, we first outline the design principles of
NOMA in PBF and FJT. Then, the device association (DA), successive
interference cancellation (SIC) design, device grouping, and power control
policies tailored for IoT are detailed.

8.2.1 Multi-cluster Coordination NOMA Design

Based on the shared information among the coordinated APs, the multi-
cluster coordinated NOMA techniques can be classified into partially-
coordinated beamforming NOMA (PBF-NOMA) and fully-coordinated
joint transmission NOMA (FJT-NOMA).

1. PBF-NOMA: In the PBF-NOMA, the devices are served by their

associated local APs, and the devices within each cluster are scheduled
into NOMA groups [126, 127]. Based on the shared devices’ CSI
among the APs, inter-cluster interference is suppressed by the PBF,
while intra-cluster interference is handled by SIC design. Compared to
the conventional PBF-OMA designs, the PBF-NOMA scheme can
simultaneously support more device connections and provide a high
level of DoFs in multi-device interference management.
2. FJT-NOMA: In the FJT-NOMA, the devices can be jointly served by
the coordinated APs, where a global manner NOMA is achieved
among all the coordinated clusters. Hence, this higher level of
coordination system is more device-centric, as the devices can be
scheduled into the same groups even if they locate into different APs’
coverage [128, 129]. In addition, compared to the PBF-NOMA scheme
where inter-cluster interference is extensively suppressed, the FJT-
NOMA takes advantage of the inter-cluster interference channel for
data transmission and benefits from the spatial diversity gain from the
distributed APs [128, 129].

For the FJT mode, each AP shares CSI and the transmitted data to other
coordinated APs. With devices in each AP’s coverage and
coordinated APs, the total overhead of sharing CSI and data is given
 , where denotes the required bits for
describing one device’s CSI while denotes the required bits for
exchanging each device’s data symbols. Generally, only a few bits are
enough to describe a multiple-input single-output (MISO) channel for each
user [130]. For the overhead on sharing data symbols, it is decided by the
downlink frame structure of the communication protocol. Hence, the
overhead on sharing data symbols is reasonable, considering the fact that
the APs have been connected to a centralized cloud-manager through high-
speed optical fibers.

8.2.2 Multi-cluster Coordinated NOMA Resource

Allocation
1. Device association: DA plays a significant role in load balancing and
performance enhancement. The existing DA designs can be generally
classified into distance-based and channel gain-based methods. In
particular, the distance-based method is sensitive to channel fading and
shadowing, while the performance of the channel gain-based method
may be significantly impaired by the inter-cluster interference. As a
result, multiple criteria, such as distance, channel gain, and inter-
cluster interference, and so on, should be jointly considered to make a
comprehensive decision for DA. Recently, artificial intelligence (AI)
techniques have been proposed to strike an attractive balance among
different criteria, enhancing the robustness against the effects of
channel fading, shadowing, and inter-cluster interference [131].
2. SIC design: SIC is employed at the receive side for superposed signal
decoding and demodulating. There are two fundamental rules: (Rule 1)
the optimal SIC decoding order of devices is based on the increasing
order of their channel gains from the APs [127], and (Rule 2) the rate
of the device with a higher decoding order (e.g., the device ) to
detect the signal of the device should be no lower than the targeted
rate of the device , so that the signal of the device can be
successfully decoded and removed at the device . For the single-
cluster NOMA IoT, Rule 1 is equivalent to Rule 2. Nevertheless, in the
multi-cluster NOMA, due to the incurred inter-cluster interference, the
rate of the device to decode the device may be lower than the
 targeted rate of the device , which results in the failure of SIC
process. Therefore, to successfully perform SIC, dedicated SIC design
considering the effect of the inter-cluster interference should be
developed to guarantee the validity of Rule 2 [128, 129].
3. Device grouping and power control: For the PBF-NOMA, the
devices within each cluster can be scheduled into a number of NOMA
groups, and power allocation policies are carefully designed by the
coordinated APs to suppress the intra-cluster inter-group interference.
On the other hand, for the FJT-NOMA scheme, both the CSI and
transmission data are available at coordinated APs. As a result, the
devices located in the different clusters could be scheduled into the
same NOMA group, thereby enabling a higher DoF in device
grouping. In addition, since the inter-cluster interference channels from
the coordinated APs to devices can be utilized to carry useful
information, only the intra-cluster inter-group interference needs to be
suppressed, which results in lower power consumption and higher EE
performance.

By applying the multi-criterion-based DA, device grouping, and power

allocation schemes, the multi-cluster coordinated IoT can provide an
energy-efficient transmission [131]. Figure 8.2(a) and (b) shows the power
consumption and EE with different coordination designs. It can be observed
that, with stringent transmission rate requirements, the multi-cluster FJT-
NOMA scheme consumes lower transmission power and achieves higher
EE performance over the non-coordinated multi-cluster NOMA design, i.e.,
EE enhancement of more than 27.4% and transmission power reduction of
around 34.6 dBm. In addition, the multi-criterion-based DA scheme
provides 16.8% EE enhancement, and 31.1 dBm transmission power
reduction than the single-criterion-based methods, enabling a greener multi-
cluster coordinated NOMA IoT system. Note that though SIC incurs
additional power at the receiver sides, typically ranging from 3 to 20
dBm, this only contributes a tiny proportion to the total power consumption.
Hence, the amount of power saving achieved by the NOMA-based multi-
cluster design overwhelms the additional power consumption incurred by
the SIC operation.
Figure 8.2 Transmission power consumption and EE performance with 5
MHz bandwidth. (a) transmission power consumption performance and (b)
EE performance with different coordination designs.

8.3 CI-aided Multi-cluster Coordination with

Interference Management
As discussed above, we have reviewed the superiority of applying NOMA
into multi-cluster coordinated IoT systems. Relying on the interference-
mitigation-based design, it is able to multiplex a massive number of devices
onto the same frequency resource while ensuring a reasonable reception
quality. Due to the non-orthogonal access of devices, the incurred intra-
cluster multi-device interference needs to be strictly suppressed by SIC,
regardless of the multi-cluster coordination levels. Nevertheless, there is
scope to exploit the multi-device interference as a beneficial element based
on the concept of CI precoding. It transmits spatially non-orthogonal signals
and makes the intrinsic spatial leakage constructive [130, 132, 133] for
further enhancing system performance. In this section, we first review the
CI technique and then elaborate on it with the cluster coordination
techniques.

8.3.1 CI Signal Design

The CI exploitation involves judiciously characterizing interference by
considering the signal constellation size. Departing from the conventional
interference suppressing techniques which constrain the received signals
within a proximity area around the modulated signal constellation point, the
CI scheme can exploit, rather than suppress, the correlation between the
transmission signal to make the multi-device interference constructive [132,
133]. Specifically, to clarify the fundamental concept mentioned above, an
elementary example of the CI exploitation with two devices is given in Fig.
8.3 (a), whose signals belong to a binary phase shift keying (BPSK)
constellation. Assume that the device 1’s desired signal equals to 1 and
interfering signal equals to , respectively. Without loss of generality,
a lossless channel is assumed from the transmitter to the device 1, and the
interfering channel is denoted as . Ignoring noise, it is easy to verify that
when the interfering channel equals to 0.5, the interfering signal is
destructive to the device 1 as it pushes the received signal closer to the
BPSK decision threshold, thereby reducing the SINR performance. On the
contrary, when the interfering channel equals to , the interfering
signal becomes constructive to the desired signal , as the interfering
signal is moved further away from the decision threshold, effectively
raising the received SINR.

Figure 8.3 (a) An elementary example of CI exploitation with BPSK

constellation; (b) A quadrature phase shift keying (QPSK) constellation
example with CI precoding exploits interference as a beneficial element; (c)
A schematic representation of - constellation points.
On the basis of the CI characterization, CI precoding can be performed to
make interference constructive for devices, where the ubiquitous
interference is judiciously transformed into a green signal source for
improving the devices’ reception performance. In other words, by
employing CI, lower transmission power is required at the devices to
achieve a targeted performance since interference contributes constructively
rather than being suppressed. The above fundamental example can be
extended to general M-order of phase shift keying (PSK) (as shown in Fig.
8.3 (b)) and quadrature amplitude modulation (QAM) (as shown in Fig. 8.3
(c)) [133]. Take 16-QAM for illustration, it can be seen that for the inner
constellation points (the group of constellation points in the box labelled 1),
interference that shifts the received inner constellation point away from one
decision threshold pushes it closer to another decision threshold. Hence, the
concept of CI does not hold. However, for the outer constellation points (the
group of constellation points in the boxes labeled 2, 3, and 4), the CI can
push the received signals fall in the detection region away from the decision
boundaries to enhance the signal reception performance.
Moreover, as for the existing zero forcing (ZF) or other optimization-based
precoders, the number of the served devices should not be larger than that
of the transmit-antennas. By contrast, the CI locates the desired signals into
constructive regions (where the received signals have increased the distance
to the detection threshold of demodulation), rather than strictly locating the
signals in the proximity region around the constellation point. With a higher
design DoF at the transmitter-side, the number of the served devices can be
larger than that of the transmit-antennas [130, 132], while achieving an
enhanced reception performance. Hence, the CI is particularly suitable for
massive downlink-connectivity in MMTC scenario.

8.3.2 CI Design for Multi-cluster Coordination

The key challenge of integrating CI and multi-cluster coordination lies in
the inter-cluster interference presented from the adjacent APs and the multi-
device interference within each cluster, which need to be separately
managed.

1. Partially-coordinated beamforming CI (PBF-CI): In the PBF-CI

scheme, since CSI is available at the coordinated APs, the intra-cluster
multi-device interference can be predicted and characterized before
transmission, which then can be treated as a beneficial element at the
receiver side, enabled by the CI precoding. However, the intended data
is merely transmitted by the local APs, which makes it difficult to
exploit inter-cluster interference as a constructive element. To this end,
 the philosophy of the PBF-CI is to exploit intra-cluster multi-device
interference while carefully mitigating inter-cluster interference [130].
2. Fully-coordinated joint transmission CI (FJT-CI): In the FJT-CI,
both CSI and the intended transmission data of the devices are shared
among the coordinated APs. As a result, not only the inherent intra-
cluster multi-device interference but also the inter-cluster interference
can be exploited as constructive elements, which achieves a higher
level of DoFs in interference management and significantly enhances
performance, at the cost of higher coordination overhead over the
PBF-CI scheme.
Figure 8.4 The performance gain of the FJT-CI and PBF-CI designs over
the conventional coordination techniques.
The impact of the devices’ SINR requirements on the total transmission
power is shown in Fig. 8.4. As can be seen, the FJT-CI scheme consumes
the lowest power, providing 12.2 dBm power saving over the FJT, 15.8
dBm over the PBF-CI, and 24.2 dBm over the PBF scheme, respectively.
The reason lies in that the FJT-CI scheme utilizes both inter-cluster and
intra-cluster multi-device interference as constructive elements, which leads
to much lower transmission power for achieving a targeted SINR. In
addition, the FJT-CI can also benefit from the spatial diversity gain by
taking advantage of the inter-cluster interference channel, thanks to the
distributed coordinated APs. For the PBF-CI scheme, as the intended
transmission data is not shared among the APs to reduce coordination
overhead, only intra-cluster multi-device interference can be made
constructive while inter-cluster interference should be carefully mitigated
by joint precoding design. It is clear from the above results that the
employment of the CI exploitation in IoT has the potential of orders-of-
magnitude reduction in power consumption.
The CI design also has a high potential in the URLLC scenario. The symbol
error rate (SER) and execution time performance are demonstrated in Fig.
8.5. It can be concluded that the CI endorses the lowest SER at
moderate/high SINR regimes, and thus outperforms the ZF, minimum mean
square error (MMSE), and SINR balancing precoders in terms of reliability
performance. In addition, the CI design requires the same execution time as
the ZF and MMSE precoders, which are known as the most practical
precoders due to their low complexities. As a result, the CI design imposes
no additional latency on the signal transmission, and is readily compatible
with existing URLLC-dedicated techniques.

Figure 8.5 The symbol error rate and execution time of different precoders
in the FJT scenario. (a) Symbol error rate of different precoders and (b)
execution time of different precoders.

8.4 Future Works

1. Hybrid NOMA and CI transmission for a multi-cluster
coordination IoT: NOMA is able to provide massive device access
with a limited number of antennas, while its performance may be
 significantly reduced if the devices are with high-channel correlation.
To this end, hybrid NOMA and CI transmission can be exploited for
multi-cluster coordination, according to the devices’ distinct traffic
demands and channel conditions. For example, devices with an instant
access request and high channel disparity can be served by NOMA,
while devices having channel correlations and requiring high
throughput can be served by CI. Hence, how to adaptively provide the
synthetic NOMA and CI transmission while guaranteeing devices’ QoS
requirements is still an open issue.
2. Multi-cluster coordination IoT with fog network computing: Fog
computing is envisioned as a crucial technology in IoT to provide low
latency and local processing, by offloading the computing and
functionalities storing from the centralized cloud center to the edge-fog
devices. Nevertheless, how to integrate fog computing-based radio
access networks (F-RAN) with multi-cluster coordination is still
challenging. It requires a new topology design, such as the design and
allocation of fog nodes across multiple clusters. Since the fog devices
in F-RAN are supplemented with additional computation capacity,
how to jointly leverage the computation capacity of fog nodes with
low overhead is yet to be explored.
3. Multi-cluster coordination design for devices with limited
statistical or unavailable CSI: The existing coordination techniques
generally rely on the availability of CSI at the transmitter (CSIT). In
some typical applications of IoT, such as ambient or back-scatter
communications, obtaining accurate CSIT of devices may be difficult.
In this case, limited statistical or unavailable CSIT of devices is more
likely to be present. How can we exploit the advantages of the CI with
limited statistical or unavailable CSIT of the devices? The energy-
efficient multi-cluster coordination in this case is not yet well
understood.
4. Joint device detection and channel estimation in grant-free multi-
cluster coordination: Low latency is an important metric for some
IoT applications, such as early warning of malfunctioned equipment.
However, the complicated handshake mechanism that has been
extensively used for grant access in cellular networks, leads to high
access delay and outdated CSI. Considering the sporadic transmission
 at uplink, joint active device detection and channel estimation would
benefit system performance in terms of low access delay and accurate
CSI acquiring. Since in practice, the number of active devices may be
much lower than the total number of devices, it is reasonable to apply
compressed sensing theory for the device detection based on the
transmission sparsity. The integration of NOMA and grant-free has
been extensively researched, however, how to further improve the
detection performance with the multi-cluster coordination needs more
fundamental analysis.
5. AI-aided non-orthogonal transmission in a multi-cluster
coordination IoT: In IoT scenario, the actions and operations of the
devices may be programmed, and thus their positions, motions, and the
associated channel quality could be observable, predictable, and
learnable. Especially, the periodic transmission is incurred by the
devices at uplink, such as the periodic monitoring and reporting of
sensors, as well as the control feedback of robotic arms. These
regularities lay a strong foundation for the AI-aided non-orthogonal
transmission designs, in terms of cluster association, device grouping,
massive device access, coordination level selection, and resource
allocation, etc. However, the burst transmission of the devices harms
the regularities in IoT, complicates the learning process, and impairs
the testing performance. This again requires fundamental analysis and
designs for the learning-aided non-orthogonal transmission in the
multi-cluster coordination IoT.

8.5 Conclusion and Remarks

In this chapter, we introduced disruptive approaches for inter-cluster and
intra-cluster interference management for multi-cluster coordinated IoT.
Focusing on the typical MMTC scenarios in multi-cluster coordination, we
have discussed the potential of accommodating non-orthogonal multi-
device access, and even exploiting the ubiquitous interference as a green
source by CI for enhancing reception performance. These novel solutions
enable an energy-efficient transmission, and offer a new design philosophy
of interference management for the emerging multi-cluster coordinated IoT.
Challenges relating to emerging multi-cluster applications are also
envisaged, and these give promise of exciting and insightful research over
the years to come.

OceanofPDF.com
9
Conclusion of the Book
Internet of Things (IoT) provides enormous connections for massive
devices and sensors, which has been considered as an enabling technology
to build smart homes, smart transportations, smart cities, and smart planets.
It is expected that by 2030 over 500 billion IoT devices will be in use with
functions of sensing, computing, and communications, in which a surging
amount of data and information exchanges are carried out among various
IoT devices. Since IoT devices have sparse activity and low signaling
overhead, random access (RA) is an efficient means to relieve the burden of
machine-type communication (MTC). Considering that efficient signal
detection methods are crucial to IoT RA systems, in this book, we focus on
the signal detection for IoT, which is shown as follows.
In Chapter 3, in order to decide the activity variables of aggregated IoT
signals, some low-complexity approaches, e.g., variational inference (VI)
algorithm, are introduced. Since that the signal vector contains a number of
zero elements, the information vector being transmitted becomes sparse. In
this case, the compressive sensing (CS) reconstruction algorithms, e.g.,
block compressed sensing (BCS) and structured subspace pursuit (SSP)
algorithm, can be applied to detect sparse signals.
However, when massive IoT devices access the network, the preamble
collision occurs frequently and data packets from the device experiencing
collision may not be successfully decoded at the receiver. Thus, in Chapter
4, some approaches are described to detect the collided signals, e.g.,
automatic modulation classification (AMC) approach.
Meanwhile, the technique of non-orthogonal random access (NORA)
utilizes the different time of arrival to identify multiple devices with the
identical preamble to improve the throughput, the detection of the round-
trip delay (RTD) in NORA becomes significant. In Chapter 5, some
estimation methods of RTD of multiple signals in NORA are presented,
e.g., maximum likelihood (ML) algorithm, VI algorithm, and Markov Chain
Monte Carlo (MCMC) algorithm, respectively.
Besides, since backscatter communications are beneficial for power-limited
or battery-free IoT devices to transmit their data, where the signal detection
at the receiver becomes crucial. In Chapter 6, signal detection approaches
for backscatter devices are proposed to detect the presence of parasite
devices and separate parasite signals from host signals, e.g., Central Limit
Theorem (CLT) criterion.
Finally, wireless communications in IoT cause severe access congestion due
to insufficient frequency, time, power, antennas, and other resources,
evidently. In Chapter 7, we analyze the throughput and power consumption
of non-orthogonal multiple access (NOMA) systems and investigate the
energy efficiency maximization problem of a full-duplex cooperative
NOMA system. Furthermore, one of the efficient techniques is to cluster
massive devices into a number of small groups, with some resources being
shared and reused in different clusters, referred to as a multicluster IoT.
Therefore, in Chapter 8, signal design approaches for multicluster
coordination are proposed, e.g., successive interference cancellation (SIC)
design and constructive interference signal design.

OceanofPDF.com
References
1 H. Zeng, J. Wang, Z. Wei, et al. Multicluster-coordination Industrial
Internet of Things: The era of nonorthogonal transmission. IEEE
Vehicular Technology Magazine, 17(3):84–93, 2022.
2 K. Ashton. That ‘Internet of Things’. RFID Journal, 22:97–114, 2009.
3 L. Atzori, A. Iera, and G. Morabito. Understanding the Internet of Things:
Definition, potentials, and societal role of a fast evolving paradigm. Ad
Hoc Networks, 56:122–140, 2017.
4 G. Singh, J. Singh, D. Mitra, et al. A roadmap toward prospects for IoT
enabled 5G networks. In 2023 7th International Conference on
Computing Methodologies and Communication (ICCMC), pages 1405–
1410, 2023.
5 China Academy of Information and Communications Technology.The
white paper on the IoT (2011). Public Security in China (Comprehensive
Edition), 7(3):148–152, 2012.
6 K. Katsalis, N. Nikaein, E. Schiller, et al. Network slices toward 5G
communications: Slicing the LTE network. IEEE Communications
Magazine, 55(8):146–154, 2017.
7 J. Ordonez-Lucena, P. Ameigeiras, D. Lopez, et al. Network slicing for
5G with SDN/NFV: Concepts, architectures and challenges. IEEE
Communications Magazine, 55(5):80–87, 2017.
8 L. Chettri and R. Bera. A comprehensive survey on Internet of Things
(IoT) toward 5G wireless systems. IEEE Internet of Things Journal,
7(1):16–32, 2020.
9 M. Yang, S. Lim, S.-M. Oh, et al. An uplink transmission scheme for TSN
service in 5G industrial IoT. In 2020 International Conference on
Information and Communication Technology Convergence (ICTC),
pages 902–904, 2020.
10 P. A. Hafeez, G. Singh, J. Singh, et al. IoT in agriculture and healthcare:
Applications and challenges. In 2022 3rd International Conference on
Smart Electronics and Communication (ICOSEC), pages 446–450, 2022.
11 S. Kar, P. Mishra, and K.-C. Wang. 5G-IoT architecture for next
generation smart systems. In 2021 IEEE 4th 5G World Forum (5GWF),
pages 241–246, 2021.
12 Y. P. E. Wang, X. Lin, A. Adhikary, et al. A primer on 3GPP narrowband
Internet of Things (NB-IoT). IEEE Communications Magazine,
55(3):117–123, 2016.
13 R. Han, M. Liu, J. Wang, et al. Anti-jamming strategy for satellite
Internet of Things: Beam switching and optimization. IEEE Internet of
Things Journal, 10(23):20254–20263, 2023.
14 Y. Meng, W. Zhang, H. Zhu, et al. Securing consumer IoT in the smart
home: Architecture, challenges, and countermeasures. IEEE Wireless
Communications, 25(6):53–59, 2019.
15 D.-D. Tran, V. N. Ha, and S. Chatzinotas. Novel reinforcement learning
based power control and subchannel selection mechanism for grant-free
NOMA URLLC-enabled systems. In 2022 IEEE 95th Vehicular
Technology Conference: (VTC2022-Spring), pages 1–5, 2022.
16 H. Xiao, B. Ai, and W. Chen. Grant-free massive machine-type
communications with backward activity detection. In 2019 IEEE Global
Communications Conference (GLOBECOM), pages 1–6, 2019.
17 L. Liu, E. G. Larsson, W. Yu, et al. Sparse signal processing for grant-
free massive connectivity: A future paradigm for random access
protocols in the Internet of Things. IEEE Signal Processing Magazine,
35(5):88–99, 2018.
18 Y. D. Beyene, R. Jäntti, and K. Ruttik. Random access scheme for
sporadic users in 5G. IEEE Transactions on Wireless Communications,
16(3):1823–1833, 2017.
19 C. Bockelmann, N. K. Pratas, G. Wunder, et al. Towards massive
connectivity support for scalable mMTC communications in 5G
 networks. IEEE Access, 6:28969–28992, 2018.
20 R. Ghods, C. Jeon, A. Maleki, et al. Optimal large-MIMO data detection
with transmit impairments. In 2015 53rd Annual Allerton Conference on
Communication, Control, and Computing (Allerton), pages 1211–1218,
2015.
21 E. Biglieri, G. Taricco, and A. Tulino. Performance of space-time codes
for a large number of antennas. IEEE Transactions on Information
Theory, 48(7):1794–1803, 2002.
22 J. Zheng and Y. Sun. Generalized space-time shift keying: Randomized
bitwise Markov chain Monte Carlo detection and code design. In 2013
IEEE International Conference on Signal Processing, Communication
and Computing (ICSPCC 2013), pages 1–4, 2013.
23 H. Zhu, B. Farhang-Boroujeny, and R.-R. Chen. On performance of
sphere decoding and Markov chain Monte Carlo detection methods.
IEEE Signal Processing Letters, 12(10):669–672, 2005.
24 W. Kim and K. M. Lee. Continuous Markov random field optimization
using fusion move driven Markov chain Monte Carlo technique. In 2010
20th International Conference on Pattern Recognition (ICPR 2010),
pages 1364–1367, 2010.
25 D. M. Blei, A. Kucukelbir, and J. D. McAuliffe. Variational inference: A
review for statisticians. Journal of the American Statistical Association,
112(518):859–877, 2017.
26 C. M. Bishop and N. M. Nasrabadi. Pattern Recognition and Machine
Learning, volume 4. Springer, 2006.
27 D. L. Donoho. Compressed sensing. IEEE Transactions on Information
Theory, 52(4):1289–1306, 2006.
28 N. Serafimovski, A. Younis, R. Mesleh, et al. Practical implementation
of spatial modulation. IEEE Transactions on Vehicular Technology,
62(9):4511–4523, 2013.
29 A. Garcia-Rodriguez and C. Masouros. Low-complexity compressive
sensing detection for spatial modulation in large-scale multiple access
 channels. IEEE Transactions on Communications, 63(7):2565–2579,
2015.
30 R. Han, L. Bai, W. Zhang, et al. Variational inference based sparse signal
detection for next generation multiple access. IEEE Journal on Selected
Areas in Communications, 40(4):1114–1127, 2022.
31 D. Tse and P. Viswanath. Fundamentals of Wireless Communication.
Cambridge University Press, 2005.
32 J. Choi. Optimal Combining and Detection. Cambridge University Press,
2010.
33 R. Y. Mesleh, H. Haas, S. Sinanovic, et al. Spatial modulation. IEEE
Transactions on Vehicular Technology, 57(4):2228–2241, 2008.
34 A. Younis, N. Serafimovski, R. Mesleh, et al. Generalised spatial
modulation. In 2010 Conference Record of the Forty Fourth Asilomar
Conference on Signals, Systems and Computers, pages 1498–1502,
2010.
35 L. Applebaum, W. U. Bajwa, M. F. Duarte, et al. Asynchronous code-
division random access using convex optimization. Physical
Communication, 5(2):129–147, 2012.
36 G. Wunder, P. Jung, and W. Chen. Compressive random access for post-
LTE systems. In 2014 IEEE International Conference on
Communications Workshops (ICC), pages 539–544, 2014.
37 J. Choi. Stability and throughput of random access with CS-based MUD
for MTC. IEEE Transactions on Vehicular Technology, 67(3):2607–
2616, 2018.
38 L. Xiao, P. Yang, Y. Xiao, et al. Efficient compressive sensing detectors
for generalized spatial modulation systems. IEEE Transactions on
Vehicular Technology, 66(2):1284–1298, 2017.
39 H. Zhu and G. B. Giannakis. Exploiting sparse user activity in multiuser
detection. IEEE Transactions on Communications, 59(2):454–465, 2011.
40 Y. C. Eldar and G. Kutyniok. Compressed Sensing: Theory and
Applications. Cambridge University Press, 2012.
41 M. Opper and C. Archambeau. The variational Gaussian approximation
revisited. Neural Computation, 21(3):786–792, 2009.
42 L. L. Scharf. Statistical Signal Processing: Detection, Estimation, and
Time Series Analysis. Addison-Wesley Publishing Company, 1991.
43 C. M. Biship. Pattern Recognition and Machine Learning (Information
Science and Statistics). Springer, New York, 2007.
44 D. A. Harville. Matrix Algebra from a Statistician’s Perspective. Taylor
& Francis, 1998.
45 B. L. Sturm and M. Christensen. Comparison of orthogonal matching
pursuit implementations. In 2012 Proceedings of the 20th European
Signal Processing Conference (EUSIPCO), pages 220–224, 2012.
46 J. Choi, K. Lee, and N. Y. Yu. Compressive random access using
multiple resource blocks for MTC. In 2016 IEEE Globecom Workshops
(GC Wkshps), pages 1–5, 2016.
47 N. Jiang, Y. Deng, A. Nallanathan, et al. Analyzing random access
collisions in massive IoT networks. IEEE Transactions on Wireless
Communications, 17(10):6853–6870, 2018.
48 H. Shariatmadari, R. Ratasuk, S. Iraji, et al. Machine-type
communications: Current status and future perspectives toward 5G
systems. IEEE Communications Magazine, 53(9):10–17, 2015.
49 P. Patel and J. Holtzman. Analysis of a simple successive interference
cancellation scheme in a DS/CDMA system. IEEE Journal on Selected
Areas in Communications, 12(5):796–807, 1994.
50 L. Liu and W. Yu. Massive connectivity with massive MIMO—Part I:
Device activity detection and channel estimation. IEEE Transactions on
Signal Processing, 66(11):2933–2946, 2018.
51 J. G. Andrews, F. Baccelli, and R. K. Ganti. A tractable approach to
coverage and rate in cellular networks. IEEE Transactions on
 Communications, 59(11):3122–3134, 2011.
52 T. Bai and R. W. Heath. Analyzing uplink SINR and rate in massive
MIMO systems using stochastic geometry. IEEE Transactions on
Communications, 64(11):4592–4606, 2016.
53 W. Xiao, R. Ratasuk, A. Ghosh, et al. Uplink power control, interference
coordination and resource allocation for 3GPP E-UTRA. In IEEE
Vehicular Technology Conference, pages 1–5, 2006.
54 M. Kiessling and J. Speidel. Analytical performance of MIMO zero-
forcing receivers in correlated Rayleigh fading environments. In 2003
4th IEEE Workshop on Signal Processing Advances in Wireless
Communications-SPAWC 2003, pages 383–387, 2003.
55 J. Ding, D. Qu, H. Jiang, et al. Success probability of grant-free random
access with massive MIMO. IEEE Internet of Things Journal, 6(1):506–
516, 2018.
56 L. Bai and J. Choi. Low Complexity MIMO Detection. Springer, 2012.
57 L. Bai, J. Choi, and Q. Yu. Low Complexity MIMO Receivers. Springer,
2014.
58 E. Bjornson, E. De Carvalho, E. G. Larsson, et al. Random access
protocol for massive MIMO: Strongest-user collision resolution
(SUCR). In 2016 IEEE International Conference on Communications
(ICC), pages 1–6, 2016.
59 J. Hoydis, S. Ten Brink, and M. Debbah. Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need? IEEE Journal on
Selected Areas in Communications, 31(2):160–171, 2013.
60 Y. Han, G. Wei, C. Song, et al. Hierarchical digital modulation
recognition based on higher-order cumulants. In 2012 2nd International
Conference on Instrumentation, Measurement, Computer,
Communication and Control, pages 1645–1648, 2012.
61 Z. Zhang, X. Wang, Y. Zhang, et al. Grant-free rateless multiple access:
A novel massive access scheme for Internet of Things. IEEE
Communications Letters, 20(10):2019–2022, 2016.
62 F. Cong, A. K. Nandi, Z. He, et al. Fast and effective model order
selection method to determine the number of sources in a linear
transformation model. In 2012 Proceedings of the 20th European Signal
Processing Conference (EUSIPCO), pages 1870–1874, 2012.
63 L. Lu, G. Y. Li, A. L. Swindlehurst, et al. An overview of massive
MIMO: Benefits and challenges. IEEE Journal of Selected Topics in
Signal Processing, 8(5):742–758, 2014.
64 R. Abbas, M. Shirvanimoghaddam, Y. Li, et al. A novel analytical
framework for massive grant-free NOMA. IEEE Transactions on
Communications, 67(3):2436–2449, 2018.
65 J. Liu, N. Kato, H. Ujikawa, et al. Device-to-device communication for
mobile multimedia in emerging 5G networks. ACM Transactions on
Multimedia Computing, Communications, and Applications (TOMM),
12(5s):1–20, 2016.
66 J. Dai, J. Liu, Y. Shi, et al. Analytical modeling of resource allocation in
D2D overlaying multihop multichannel uplink cellular networks. IEEE
Transactions on Vehicular Technology, 66(8):6633–6644, 2017.
67 Y. Liang, X. Li, J. Zhang, et al. Non-orthogonal random access for 5G
networks. IEEE Transactions on Wireless Communications, 16(7):4817–
4831, 2017.
68 C. Stefanovic, M. Momoda, and P. Popovski. Exploiting capture effect in
frameless ALOHA for massive wireless random access. In 2014 IEEE
Wireless Communications and Networking Conference (WCNC), pages
1762–1767, 2014.
69 G. Bartoli, N. C. Beaulieu, R. Fantacci, et al. An effective multiuser
detection scheme for MPR random access networks. IEEE Transactions
on Communications, 65(3):1119–1130, 2016.
70 J. Choi. NOMA-based random access with multichannel ALOHA. IEEE
Journal on Selected Areas in Communications, 35(12):2736–2743, 2017.
71 R. Han, L. Bai, J. Yi, et al. Multiple delay estimation using genetic
algorithm-based MCMC in non-orthogonal random access. IEEE
 Wireless Communications Letters, 9(3):398–401, 2020.
72 L. Bai, R. Han, J. Liu, et al. Multiple delay estimation for collision
resolution in non-orthogonal random access. IEEE Transactions on
Vehicular Technology, 69(1):497–508, 2020.
73 3GPP. Evolved universal terrestrial radio access (E-UTRA); base station
(BS) radio transmission and reception. Technical Specification 36.104,
3rd Generation Partnership Project (3GPP), October, 2014.
74 T. Morohashi, C.-H. Liao, A. Koizuka, et al. A high-performance RACH
detection scheme for random access overload in LTE-advanced. In 2015
IEEE Conference on Standards for Communications and Networking
(CSCN), pages 1–6, 2015.
75 M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, et al. An introduction to
variational methods for graphical models. Machine Learning, 37:183–
233, 1999.
76 J. Choi. MCMC-based detection for random access with preambles in
MTC. IEEE Transactions on Communications, 67(1):835–846, 2018.
77 L. Bai, T. Li, J. Liu, et al. Large-scale MIMO detection using MCMC
approach with blockwise sampling. IEEE Transactions on
Communications, 64(9):3697–3707, 2016.
78 G. Casella and E. I. George. Explaining the Gibbs sampler. The
American Statistician, 46(3):167–174, 1992.
79 K.-S. Tang, K.-F. Man, S. Kwong, et al. Genetic algorithms and their
applications. IEEE Signal Processing Magazine, 13(6):22–37, 1996.
80 Y. Sakamoto, M. Ishiguro, and G. Kitagawa. Akaike information
criterion statistics. Dordrecht, The Netherlands: D. Reidel,
81(10.5555):26853, 1986.
81 A. A. Neath and J. E. Cavanaugh. The Bayesian information criterion:
Background, derivation, and applications. Wiley Interdisciplinary
Reviews: Computational Statistics, 4(2):199–203, 2012.
82 S. M. Kay. Fundamentals of Statistical Processing: Detection Theory,
volume 2. Pearson Education India, 2009.
83 R. Han, L. Bai, Y. Wen, et al. UAV-aided backscatter communications:
Performance analysis and trajectory optimization. IEEE Journal on
Selected Areas in Communications, 39(10):3129–3143, 2021.
84 W. Zhang, Y. Qin, W. Zhao, et al. A green paradigm for Internet of
Things: Ambient backscatter communications. China Communications,
16(7):109–119, 2019.
85 D. Dobkin. The RF in RFID: UHF RFID in Practice. Newnes, 2012.
86 R. Han, Y. Wen, L. Bai, et al. Rate splitting on mobile edge computing
for UAV-aided IoT systems. IEEE Transactions on Cognitive
Communications and Networking, 6(4):1193–1203, 2020.
87 L. Bai, R. Han, J. Liu, et al. Air-to-ground wireless links for high-speed
UAVs. IEEE Journal on Selected Areas in Communications,
38(12):2918–2930, 2020.
88 Q. Zhang, H. Guo, Y.-C. Liang, et al. Constellation learning-based signal
detection for ambient backscatter communication systems. IEEE Journal
on Selected Areas in Communications, 37(2):452–463, 2018.
89 F. Rezaei, C. Tellambura, and S. Herath. Large-scale wireless-powered
networks with backscatter communications—a comprehensive survey.
IEEE Open Journal of the Communications Society, 1:1100–1130, 2020.
90 J. D. Griffin and G. D. Durgin. Gains for RF tags using multiple
antennas. IEEE Transactions on Antennas and Propagation, 56(2):563–
570, 2008.
91 A. Varshney, O. Harms, C. Pérez-Penichet, et al. LoRea: A backscatter
architecture that achieves a long communication range. In Proceedings
of the 15th ACM Conference on Embedded Network Sensor Systems,
pages 1–14, 2017.
92 J. Wang, J. Zhang, R. Saha, et al. Pushing the range limits of commercial
passive RFIDs. In 16th USENIX Symposium on Networked Systems
Design and Implementation, pages 301–316, 2019.
93 M. Y. Abdelsadek, Y. Gadallah, and M. H. Ahmed. Matching-based
resource allocation for critical MTC in massive MIMO LTE networks.
IEEE Access, 7:127141–127153, 2019.
94 K. Senel and E. G. Larsson. Grant-free massive MTC-enabled massive
MIMO: A compressive sensing approach. IEEE Transactions on
Communications, 66(12):6164–6175, 2018.
95 P. Popovski, C. Stefanovic, J. J. Nielsen, et al. Wireless access in ultra-
reliable low-latency communication (URLLC). IEEE Transactions on
Communications, 67(8):5783–5801, 2019.
96 E. Björnson, E. G. Larsson, and M. Debbah. Massive MIMO for
maximal spectral efficiency: How many users and pilots should be
allocated? IEEE Transactions on Wireless Communications, 15(2):1293–
1308, 2015.
97 G. Yang, Y.-C. Liang, and Q. Zhang. Cooperative receiver for ambient
backscatter communications with multiple antennas. In 2017 IEEE
International Conference on Communications (ICC), pages 1–6, 2017.
98 W. Zhao, G. Wang, S. Atapattu, et al. Channel estimation for ambient
backscatter communication systems with massive-antenna reader. IEEE
Transactions on Vehicular Technology, 68(8):8254–8258, 2019.
99 H. Jiang, D. Qu, J. Ding, et al. Multiple preambles for high success rate
of grant-free random access with massive MIMO. IEEE Transactions on
Wireless Communications, 18(10):4779–4789, 2019.
100 H. Ding, D. B. Da Costa, and J. Ge. Outage analysis for cooperative
ambient backscatter systems. IEEE Wireless Communications Letters,
9(5):601–605, 2019.
101 J.-B. Kim and I.-H. Lee. Capacity analysis of cooperative relaying
systems using non-orthogonal multiple access. IEEE Communications
Letters, 19(11):1949–1952, 2015.
102 Y. Liu, Z. Qin, M. Elkashlan, et al. Non-orthogonal multiple access for
5G and beyond. Proceedings of the IEEE, 105(12):2347–2381, 2017.
103 Z. Ding, M. Peng, and H. V. Poor. Cooperative non-orthogonal multiple
access in 5G systems. IEEE Communications Letters, 19(8):1462–1465,
2015.
104 G. Liu, X. Chen, Z. Ding, et al. Hybrid half-duplex/full-duplex
cooperative non-orthogonal multiple access with transmit power
adaptation. IEEE Transactions on Wireless Communications, 17(1):506–
519, 2017.
105 L. Zhang, J. Liu, M. Xiao, et al. Performance analysis and optimization
in downlink NOMA systems with cooperative full-duplex relaying.
IEEE Journal on Selected Areas in Communications, 35(10):2398–2412,
2017.
106 Z. Zhang, Z. Ma, M. Xiao, et al. Full-duplex device-to-device-aided
cooperative nonorthogonal multiple access. IEEE Transactions on
Vehicular Technology, 66(5):4467–4471, 2016.
107 J. Choi. Minimum power multicast beamforming with superposition
coding for multiresolution broadcast and application to NOMA systems.
IEEE Transactions on Communications, 63(3):791–800, 2015.
108 F. Fang, H. Zhang, J. Cheng, et al. Energy-efficient resource allocation
for downlink non-orthogonal multiple access network. IEEE
Transactions on Communications, 64(9):3722–3732, 2016.
109 Q. Sun, S. Han, I. Chin-Lin, et al. Energy efficiency optimization for
fading MIMO non-orthogonal multiple access systems. In 2015 IEEE
International Conference on Communications (ICC), pages 2668–2673,
2015.
110 Z. Zhang, K. Long, A. V. Vasilakos, et al. Full-duplex wireless
communications: Challenges, solutions, and future research directions.
Proceedings of the IEEE, 104(7):1369–1409, 2016.
111 B. Klaiqi, X. Chu, and J. Zhang. Energy-efficient and low signaling
overhead cooperative relaying with proactive relay subset selection.
IEEE Transactions on Communications, 64(3):1001–1015, 2016.
112 Y. Xiao and L. J. Cimini. Impact of overhead on spectral efficiency of
cooperative relaying. IEEE Transactions on Wireless Communications,
12(5):2228–2239, 2013.
113 Z. Wei, X. Zhu, S. Sun, et al. Energy-efficiency of millimeter-wave
full-duplex relaying systems: Challenges and solutions. IEEE Access,
4:4848–4860, 2016.
114 D. Nguyen, L.-N. Tran, P. Pirinen, et al. Precoding for full duplex
multiuser MIMO systems: Spectral and energy efficiency maximization.
IEEE Transactions on Signal Processing, 61(16):4038–4050, 2013.
115 B. Li, X. Qi, K. Huang, et al. Security-reliability tradeoff analysis for
cooperative NOMA in cognitive radio networks. IEEE Transactions on
Communications, 67(1):83–96, 2018.
116 Y. Tao, L. Liu, S. Liu, et al. A survey: Several technologies of non-
orthogonal transmission for 5G. China Communications, 12(10):1–15,
2015.
117 T. Qiu, J. Chi, X. Zhou, et al. Edge computing in Industrial Internet of
Things: Architecture, advances and challenges. IEEE Communications
Surveys and Tutorials, 22(4):2462–2488, 2020.
118 K. Wang, Y. Wang, Y. Sun, et al. Green Industrial Internet of Things
architecture: An energy-efficient perspective. IEEE Communications
Magazine, 54(12):48–54, 2016.
119 J. An, K. Wang, S. Wang, et al. Antenna array calibration for IIoT
oriented satellites: From orthogonal CDMA to NOMA. IEEE Wireless
Communications, 27(6):28–36, 2020.
120 B. Selim, M. S. Alam, J. V. Evangelista, et al. NOMA-based IoT
networks: Impulsive noise effects and mitigation. IEEE Communications
Magazine, 58(11):69–75, 2020.
121 Q. Chen, K. Yang, H. Jiang, et al. Joint beamforming coordination and
user selection for CoMP-enabled NR-U networks. IEEE Internet of
Things Journal, 9(16):14530–14541, 2021.
122 P. Jia, X. Wang, and K. Zheng. Distributed clock synchronization based
on intelligent clustering in local area industrial IoT systems. IEEE
Transactions on Industrial Informatics, 16(6):3697–3707, 2019.
123 B. Cheng, M. Wang, S. Zhao, et al. Situation-aware dynamic service
coordination in an IoT environment. IEEE/ACM Transactions on
Networking, 25(4):2082–2095, 2017.
124 Y. Yuan, Z. Yuan, G. Yu, et al. Non-orthogonal transmission technology
in LTE evolution. IEEE Communications Magazine, 54(7):68–74, 2016.
125 A. I. Perez-Neira, M. Caus, and M. A. Vazquez. Non-orthogonal
transmission techniques for multibeam satellite systems. IEEE
Communications Magazine, 57(12):58–63, 2019.
126 X. Liu and X. Zhang. NOMA-based resource allocation for cluster-
based cognitive Industrial Internet of Things. IEEE Transactions on
Industrial Informatics, 16(8):5379–5388, 2019.
127 Y. Cao, N. Zhao, Y. Chen, et al. Secure transmission via beamforming
optimization for NOMA networks. IEEE Wireless Communications,
27(1):193–199, 2019.
128 M. S. Ali, E. Hossain, and D. I. Kim. Coordinated multipoint
transmission in downlink multi-cell NOMA systems: Models and
spectral efficiency performance. IEEE Wireless Communications,
25(2):24–31, 2018.
129 S. Khairy, P. Balaprakash, L. X. Cai, et al. Constrained deep
reinforcement learning for energy sustainable multi-UAV based random
access IoT networks with NOMA. IEEE Journal on Selected Areas in
Communications, 39(4):1101–1115, 2020.
130 Z. Wei, C. Masouros, K.-K. Wong, et al. Multi-cell interference
exploitation: Enhancing the power efficiency in cell coordination. IEEE
Transactions on Wireless Communications, 19(1):547–562, 2019.
131 H. Zeng, X. Zhu, Y. Jiang, et al. A green coordinated multi-cell NOMA
system with fuzzy logic based multi-criterion user mode selection and
 resource allocation. IEEE Journal of Selected Topics in Signal
Processing, 13(3):480–495, 2019.
132 Z. Wei, C. Masouros, F. Liu, et al. Energy-and cost-efficient physical
layer security in the era of IoT: The role of interference. IEEE
Communications Magazine, 58(4):81–87, 2020.
133 T. Xu, C. Masouros, and I. Darwazeh. Waveform and space precoding
for next generation downlink narrowband IoT. IEEE Internet of Things
Journal, 6(3):5097–5107, 2019.

OceanofPDF.com
Index
Symbols

2-norm 18

a
a posteriori probability (APP) 18, 40, 45
a priori probability (APRP) 18, 39, 55
acronyms, xxvii
additive white Gaussian noise (AWGN) 34, 128, 139
asynchronous 16

b
bandwidth 90, 109, 138, 145, 155
base station (BS) 8, 11, 40, 63, 118, 137
beamforming 66, 86, 120, 148, 150, 151, 152, 158
binary phase shift keying (BPSK) 102, 156
bit error rate (BER) 21, 58, 127

c
candidate vectors 18, 21
channel estimation 25
channel gain 65, 83, 118, 120, 153, 154, 155
channel matrix 17
channel state information (CSI) 89, 130, 137, 151
Chernoff bound 20
Chi-square 50, 77
circular symmetric complex Gaussian (CSCG) 55
column vector 35, 36
compression measurement 35, 36
compressive sensing (CS) 8
computational complexity 14, 15, 16, 18, 40, 43, 45, 95
constellation points 68, 157, 158
correlation 66, 124, 129, 130, 13, 134, 156, 161
cross entropy 33

d
data rate 39
degrees of freedom 50
downlink 11, 137, 146, 153, 158

e
eigenvalue 50, 70
entry 72
error probability 18
Euclidean distance 18
evidence lower bound (ELBO) 33
exhaustive search 21, 40, 94, 113
expectation 19, 45, 72, 73, 76, 77, 95, 97, 98, 99, 125

f
false alarm (FA) 50
filter bank multicarrier (FBMC) 2
filter-based orthogonal frequency division multiplexing (F-OFDM) 2
Fourier transform basis 36
fractional power control 65 118
Frobenius norm 19
full-duplex 9, 137, 164

g
generalized likelihood ratio test (GLRT) 124
Genetic Algorithm 100
Gibbs Sampler 29
grant-based random access 11
grant-free random access 14

h
half-duplex 137
high-order cumulant 68

i
identity matrix 26, 37
information entropy 33
interfering signal 156, 157
inter-symbol interference (ISI) 89

j
J-divergence 50
Jensen’s inequality 54

k
Kullback–Leibler (KL) divergence 32

l
least absolute shrinkage and selection operator (LASSO) 42
likelihood function 18, 39, 93, 100, 102, 111, 123, 124
linear detector 22
linear transformation 22
log-likelihood ratio (LLR) 29
low-density parity check (LDPC) 2
lower bound 33, 45, 95

m
MAP detection 40
MATLAB 143
matrix inversion 46, 47, 51, 96
maximum a posteriori probability (MAP) 18, 40
maximum likelihood (ML) 92, 126, 163
mean square error (MSE) 26
minimum mean square error (MMSE) 16, 40, 68, 161
missed detection (MD) 50
modulation 9, 18, 20, 23, 26, 29, 30, 39, 63, 66, 68, 79, 120, 122, 158,
163
Monte Carlo 8, 11, 99
Markov chain 29
multiple-input multiple-output (MIMO) 15, 39, 63, 118, 151

n
noise 15, 17, 19, 20, 21, 23, 25, 26, 28
non-deterministic polynomial (NP) hard 37
non-orthogonal multiple access (NOMA) 9

o
observation matrix 36
orthogonal 52, 65, 68, 79, 89, 120, 129
orthogonal basis 37
orthogonal frequency division multiplexing (OFDM) 89
orthogonal matching pursuit (OMP) 42
orthonormal 36, 119

p
pairwise error probability (PEP) 18
physical downlink shared channel (PDSCH) 11
physical random access channel (PRACH) 11
preamble 16, 17
probability density function (pdf) 31, 65, 72, 93
pseudo-inverse 92, 93

q
Q-function 19
quadrature amplitude modulation (QAM) 20, 68, 158

r
random variable 18, 29, 31, 39, 50, 52, 53, 64, 95, 123, 125, 126
Rayleigh fading 64, 92, 119, 120, 121, 146
real-part operation 19
received signal 15
receiver 15
restricted isometry property (RIP) 15

s
sampling rate 35
sensing matrix 36
signal candidates 18
signal constellation 18
signal estimation 25
signal reconstruction 35
signal-to-interference-plus-noise ratio (SINR) 53, 71, 137
signal-to-noise ratio (SNR) 21, 55, 69, 79, 102, 128
sparse signal 15
spatial diversity gain 18, 151, 153, 159
spatial modulation (SM) 39
spectral efficiency 17, 39
spectrum 2
statistical sampling 28
statistics 15
submatrix 42
suboptimal 124
subspace 9
subvector 46
successive interference cancellation (SIC) 9
symbol energy 26, 34, 43, 55, 120
symbol error rate (SER) 159

t
throughput 138, 139, 142, 144
transmission rate 7, 154
transmitted symbol 29
transmitter 18
tree search 41, 42, 48, 61

u
underdetermined 37
uplink 12, 63, 64, 65, 79, 162
upper bound 19, 20, 77, 82, 85, 138, 143

v
variance 19, 20, 23, 24, 26, 43, 52, 55, 64, 69, 72, 92
variational distribution 31
variational inference (VI) 8
vector 141, 142, 143, 151

z
zero-forcing (ZF) 66
zero-forcing beamforming 66

OceanofPDF.com
WILEY END USER LICENSE AGREEMENT
Go to www.wiley.com/go/eula to access Wiley’s ebook EULA.

OceanofPDF.com