NOMA in Content-Centric Mobile Networks

NON-ORTHOGONAL MULTIPLEACCESS (NOMA) IN CONTENTCENTRIC MOBILE NETWORKS

A thesis submitted to the University of Manchester

for the degree of Doctor of Philosophy

in the Faculty of Science and Engineering

2021

ByMuhammad Norfauzi Pehin Dato Haji Dani

Department of Electrical and Electronic Engineering

Contents

List of Abbreviations 9

List of Mathematical Notations 13

List of Variables 14

Abstract 17

Declaration 18

Copyright 19

Acknowledgements 20

1 Introduction 221.1 Trends in Mobile Networks . . . . . . . . . . . . . . . . . . . . . 22

1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.3 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.5 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . 29

1.6 List of Publications . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2 Background and Overview 312.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.2 Wireless Channel Models . . . . . . . . . . . . . . . . . . . . . . 31

2.2.1 Large-Scale Propagation Effects . . . . . . . . . . . . . . 32

2.2.2 Small-Scale Propagation Effects . . . . . . . . . . . . . . 35

2.3 Non-Orthogonal Multiple Access (NOMA) . . . . . . . . . . . . 38

2.3.1 Superposition Coding (SC) . . . . . . . . . . . . . . . . . 39

2

2.3.2 Successive Interference Cancellation (SIC) . . . . . . . . 40

2.3.3 NOMA in Downlink Communication Systems . . . . . . 43

2.4 Layered Video Streaming in Multicast Networks . . . . . . . . . 48

2.4.1 Layered Video Streaming . . . . . . . . . . . . . . . . . . 51

2.4.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

2.5 Delivery techniques for Wireless Caching at User Equipment . 56

2.5.1 Coded Multicasting . . . . . . . . . . . . . . . . . . . . . 57

2.5.2 CIC-based NOMA . . . . . . . . . . . . . . . . . . . . . . 59

2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3 Resource Allocation for Layered Multicast Video Streaming inNOMA Systems 623.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3.2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.3 System Model And Problem Formulation . . . . . . . . . . . . . 65

3.4 Power Allocation Schemes for Two-layer with Arbitrary Sub-grouping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3.4.1 Subgradient Method . . . . . . . . . . . . . . . . . . . . . 74

3.4.2 Multicast-based Equal RB Power Allocation (M-ERPA) . 76

3.5 Power Allocation And Subgrouping Formation Schemes forMulti-layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

3.5.1 Successive Layer-based Power Allocation (SLPA) . . . . 79

3.5.2 Subgroup Formation . . . . . . . . . . . . . . . . . . . . . 82

3.5.3 Complexity of Subgroup Formation Algorithm . . . . . 85

3.6 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . 85

3.6.1 Two-Layer Case with Arbitrary Subgrouping . . . . . . . 86

3.6.2 General Multi-Layer Case . . . . . . . . . . . . . . . . . . 89

3.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

4 Low-Complexity Beamforming Algorithms for NOMA-based Lay-ered Multicast Systems 974.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4.2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

4.3 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

4.4 Problem Formulation and Proposed Beamforming Schemes . . 102

4.4.1 Maximizing the CNRs of the Weakest Users (MCNR) . . 103

3

4.4.2 Interference Nulling and Maximizing the CNRs of Se-lected Users (N-MCNR) . . . . . . . . . . . . . . . . . . . 104

4.4.3 Multiplicative Update Algorithm (NOMA-MU) . . . . . 106

4.4.4 Computational Complexities . . . . . . . . . . . . . . . . 108

4.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . 109

4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

5 NOMA and Coded Multicasting in Cache-Aided Wireless Networks114

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

5.2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

5.3 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

5.3.1 NOMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

5.3.2 Coded Multicasting . . . . . . . . . . . . . . . . . . . . . 120

5.4 Performance Analysis of NOMA and Coded Multicasting . . . 121

5.4.1 Asymptotic Sum Rate Comparison . . . . . . . . . . . . 123

5.4.2 Outage Performance . . . . . . . . . . . . . . . . . . . . . 126

5.4.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . 128

5.5 Hybrid NOMA and Coded Multicasting Scheme . . . . . . . . . 133

5.5.1 Scheme Selection . . . . . . . . . . . . . . . . . . . . . . . 134

5.5.2 Near-Optimal Joint Power and RB Allocation or UserPairing for NOMA . . . . . . . . . . . . . . . . . . . . . . 135

5.5.3 Low-complexity Joint Power and RB Allocation or UserPairing Scheme . . . . . . . . . . . . . . . . . . . . . . . . 139

5.5.4 Computational Complexities . . . . . . . . . . . . . . . . 140

5.5.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . 141

5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

6 Conclusions and Future Work 1486.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

References 153

Word count: 34,104

4

List of Tables

2.1 Path loss exponents for various propagation environments . . . 34

2.2 The parameters for ITU Pedestrian Model . . . . . . . . . . . . . 37

3.1 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . 86



5

List of Figures

2.1 Tap delay line filter representation of frequency selective fading 36

2.2 (a) Power delay profile and (b) the channel gain of a multi-path channel with respect to frequency based on ITU Pedes-trian Channel B Model . . . . . . . . . . . . . . . . . . . . . . . . 37

2.3 An example of superposition coding . . . . . . . . . . . . . . . . 40

2.4 SIC process for two-user downlink scenario . . . . . . . . . . . . 41

2.5 An example of NOMA symbol decoding . . . . . . . . . . . . . 42

2.6 Receiver structure of user 2 employing SLIC and CWIC . . . . 42

2.7 Capacity regions of downlink NOMA and OMA . . . . . . . . 46

2.8 Illustration of a layered video coding scheme . . . . . . . . . . . 52

2.9 Resource utilization in 2-layer multicast streaming in OFDMAand NOMA systems . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.10 A general network architecture of CCMN . . . . . . . . . . . . . 57

2.11 An example of coded multicasting in cache-enabled networkfor two-user scenario . . . . . . . . . . . . . . . . . . . . . . . . . 58

2.12 An example of CIC-based NOMA in cache-enabled network fortwo-user scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

3.1 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.2 SIC in multicast users’ receivers for 4-layer multicast system . . 68

3.3 The power allocation stages in SLPA for 4-layer multicast system 80

3.4 Performance in terms of sum rate versus total transmission power 87

3.5 Convergence performance of subgradient method (Algorithm3.1) when PT = 46 dBm (µ(0) = 50, τ(0) = 0, αµ = 0.4/

√t and

ατ = 5/ (1 + t)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

3.6 Transmission rate of individual users in each subgroup versustotal transmision power . . . . . . . . . . . . . . . . . . . . . . . 88

6

3.7 Sum rate performance of different subgrouping methods usingSLPA versus total transmission power for S = 2 and S = 4 cases. 90

3.8 The effect of varying the percentage difference in CNR (in dB)of Method 3 on sum rate performance when PT = 40 dBm forthe S = 2 case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

3.9 Computational complexity of different subgrouping methodsfor S = 2 and S = 4 cases. . . . . . . . . . . . . . . . . . . . . . . 91

3.10 Sum rate performance of different power allocation schemesutilizing subgrouping Method 3 versus total transmissionpower for S = 4 case . . . . . . . . . . . . . . . . . . . . . . . . . 92

3.11 Outage probability for each layer stream s in S = 4 case againstdifferent target rate at PT = 46 dBm . . . . . . . . . . . . . . . . 93

3.12 Fairness index for S = 4 case against different number of usersat PT = 46 dBm with the target rates of Φ

(1)min = 2.5 Mbps, Φ

(2)min =

5 Mbps, Φ(3)min = 7.5 Mbps, and Φ

(4)min = 10 Mbps . . . . . . . . . . 95

4.1 (a) Sum rate and (b) individual user rate performance for K = 5

(G1 = 2 and G2 = 3 ) and M = 4 . . . . . . . . . . . . . . . . . . 110

4.2 (a) Sum rate and (b) individual user rate performance for K = 8

(G1 = G2 = 4) and M = 8 . . . . . . . . . . . . . . . . . . . . . . 111

4.3 The convergence behaviour of NOMA-MU for K = 5 (G1 = 2

and G2 = 3 ) and M = 4 . . . . . . . . . . . . . . . . . . . . . . . 112

5.1 Content placement and delivery in NOMA and coded multi-casting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

5.2 Probability that the sum rate difference between NOMA andcoded multicasting is greater than R (a) when b = 5 is pairedwith a = 1, and a = 4, and (b) when b = 3 is paired with a = 1,and a = 2, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

5.3 Probability that NOMA performs better than coded multicast-ing (i.e., R = 0) against the power allocation coefficient of theweak user αa when b = 5 is paired with a = 1, a = 3 and a = 4, 130

5.4 Outage probability of NOMA and coded multicasting as a func-tion of transmit SNR with Φ

(a)min = Φ

(b)min = 0.25BPCU when

b = 5 is paired with (a) a = 1, and (b) a = 4 . . . . . . . . . . . . 132

7

5.5 Sum rate performance of NOMA, coded multicasting and hy-brid NOMA-coded multicasting versus total transmission power 142

5.6 Sum rate performance of NOMA, coded multicasting and hy-brid NOMA-coded multicasting for RC = 200 m (σ=8 dB) andRC = 50 m (σ=4 dB) . . . . . . . . . . . . . . . . . . . . . . . . . 144

5.7 Fairness index for NOMA, coded multicasting and hybridNOMA-coded multicasting case versus total transmission power 145

5.8 Sum rate performance of different resource allocation schemesin NOMA and coded multicasting versus total transmissionpower . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

5.9 Convergence performance of Algorithm 5.2 for (a) K = 2 users,and (b) K = 4 users when PT = 20 dBm . . . . . . . . . . . . . . 146

8

List of Abbreviations

1G First Generation mobile network

2G Second Generation mobile network

3G Third Generation mobile network

3GPP Third Generation Partnership Project

4G Fourth Generation mobile network

5G Fifth Generation mobile network

5G NR Fifth Generation New Radio

AR Augmented Reality

AWGN Additive White Gaussian Noise

BLER Block Error Rate

BPCU Bit Per Channel Use

BS Base Station

CA Carrier Aggregation

CCMN Content-Centric Mobile Network

CDF Cumulative Distribution Function

CDMA Code-Division Multiple Access

CIC Cache-enabled Intereference Cancellation

CINR Channel-to-Interference-plus-Noise Ratio

CNR Channel-to-Noise Ratio

CoMP Coordinated Multipoint

C-RAN Cloud Radio Access Network

9

CSI Channel State Information

CWIC Codeword-Level Interference Cancellation

D2D Device-to-Device

DCT Discrete Cosine Transform

eMBB Enhanced Mobile Broadband

eMBMS Enhanced Multimedia Broadcast and Multicast Service

FDMA Frequency Division Multiple Access

FD MIMO Full Dimension Multiple-Input Multiple-Output

FEC Forward Error Correction

FFT Fast Fourier Transform

FGS Fine Granularity Scalable

FPA Fixed Power Allocation

F-RAN Fog Radio Access Network

FSPA Full Search Power Allocation

FTPA Fractional Transmit Power Allocation

GCC Greedy Constrained Coloring

HetNet Heterogenous Network

HEVC High Efficiency Video Coding

HSDPA High Speed Downlink Packet Access

IC Index Coding

IDT Information Decomposition Techniques

IGSMS-PA Intra-Group Scalable Multicast Scheduling Power Allocation

IoT Internet of Things

ITU International Telecommunication Union

KKT Karush-Kuhn-Tucker

LDD Lagrangian Dual Decomposition

LDS-CDMA Low-Density Spreading Code-Division Multiple Access

LLR Log-Likelihood Ratio

10

LoS Line of Sight

LTE-A Long Term Evolution-Advanced

M2M Machine-to-Machine

MC-CDMA Multi-Carrier Code-Division Multiple Access

MCNR Maximizing the CNRs of the weakest users

MCS Modulation and Coding Scheme

MDC Multiple Description Coding

MEC Mobile Edge Computing

M-ERPA Multicast-based Equal RB Power Allocation

MINLP Mixed Integer Non-Linear Programming

MISO Multiple-Input Single-Output

mMTC Massive Machine Type Communication

mmWave Millimetre Wave

MPA Message Passing Algorithm

MRC Maximal Ratio Combining

MSF Multicast Subgroup Formation

MU Multiplicative Update algorithm

MUSA Multiuser Shared Access

NALU Network Abstraction Layer Unit

N-MCNR Interference Nulling and Maximizing the CNRs of users

NOMA Non-Orthogonal Multiple Access

NOMA-MU NOMA-based Multiplicative Update algorithm

OFDMA Orthogonal Frequency Division Multiple Access

OMA Orthogonal Multiple Access

PDF Probability Density Function

PDMA Pattern Division Multiple Access

PSD Power Spectral Density

QoS Quality of Service

11

QPSK Quadrature Phase-Shift Keying

RB Resource Block

SC Superposition Coding

SC-FDMA Single Carrier Frequency Division Multiple Access

SCMA Sparse Code Multiple Access

SDR Semi-Definite Relaxation

SFN Single Frequency Network

SIC Successive Interference Cancellation

SINR Signal-to-Interference and Noise Ratio

SISO Single Input Single Output

SLA Successive Linear Algorithm

SLIC Symbol-Level Interference Cancellation

SLPA Successive Layer-based Power Allocation

SNR Signal-to-Noise Ratio

SVC Scalable Video Coding

TDMA Time-Division Multiple Access

UE User Equipment

UHD Ultra-High Definition

UMTS Universal Mobile Telecommunications System

URLLC Ultra-Reliable and Low Latency Communication

V2X Vehicle-to-Everything

VR Virtual Reality

WSRM Weighted Sum Rate Maximization

12

List of Mathematical Notations

∏Product symbol∑Summation symbol

! Factorial operator

|.| Magnitude of a complex number

‖·‖2 Euclidean norm of a vector

(.)H Hermetian (conjugate) transpose

Im m×m identity matrix

0m×n m× n zero matrix

arg max Argument of the maximum

arg min Argument of the minimum

f(.) Probability density function

F (.) Cumulative distribution function

ln(.) Natural logarithm

logx (.) Logarithmic function to base x

max Maximum of a set

min Minimum of a set

O(.) Complexity order

Pr(x) Probability of x

13

List of Variables

αk,n Power allocation coefficient of user k at RB n

B Bandwidth of each RB

BT Total system bandwidth

βk,n RB allocation indicator for user k at RB n

d Distance between transmitter (or BS) and receiver (or user)

d0 Reference distance

δ Timeslot factor

ε Tolerance

gk,n Complex coefficient for fading experienced by user k at RB n

Gr Receive antenna gain

Gt Transmit antenna gain

G(s)n Total number of users in subgroup s

G(s)n Set of users in subgroup s

γ Path loss exponent

Γ SNR or SINR

hk,n Complex channel coefficient between BS and user k at RB n

hk Complex channel vector between BS and user k

14

I Total number of iterations

K Total number of users

K(s)n Number of users receiving layer stream s at RB n

K Set of users

λ Lagrange multiplier

Λ CNR

µ Lagrange multiplier

N Total number of RBs

N0 Power spectral density of noise

Ωn Power budget coefficient for RB n

Pk,n Power allocated to user k at RB n

P(s)n Power allocated to layer stream s at RB n

Pr Received power

PRB Power allocated to each RB

Pt Transmit power

PT Total transmission power of BS

PL Path loss

PL Mean path loss

PLfs Free-space path loss

PLk Path loss effect between BS and user k

φ Lagrange multiplier

Φmin Minimum rate requirement

15

Rk,n Achievable rate of user k at RB n

R(s)n Achievable rate of layer stream s at RB n

RC Cell radius

ρ Transmit SNR

S Number of layer streams

Tmax Maximum number of iterations

τ Lagrange multiplier

vBL Beamforming weight vector for base layer

vEL Beamforming weight vector for enhancement layer

wk,n AWGN experienced by user k at RB n

xk,n Transmitted symbols of user k at RB n

x(s)n Transmitted symbols of layer stream s at RB n

x(f)i Transmitted symbols of subfile f intended for user i

x(f)a⊕b Transmitted symbols of the coded message X(f)

a⊕b

X(s)n Data stream of layer s at RB n

X(f)i Subfile f intended for user i

X(f)a⊕b XOR coded message that contains f th subfiles intended for user a and b

Xσ Gaussian distributed random variable for shadowing (in dB)

ξk log-normal shadowing factor experienced by user k

yk,n Received signal at user k at RB n

Z Accuracy parameter for Chebyshev-Gauss quadrature

16

Abstract

As the demand for rich multimedia content is expected to increase drasticallyover the years, the capacity of the current mobile network will not be ableto cope with the tremendous growing traffic. Both non-orthogonal multipleaccess (NOMA) and content centric mobile network (CCMN) have emergedas the promising technologies which potentially addresses the growing traf-fic dominated by high-quality multimedia content. In this thesis, varioustechniques are developed to improve the sum rate performance of NOMA inCCMN, particularly multicast and cache-aided networks. First, the perfor-mance of NOMA-based layered multicast system is enhanced by designing aresource allocation scheme which considers the joint optimization of powerallocation and subgrouping. In addition, the performance is further improvedby employing beamforming techniques at the expense of additional computa-tional cost, which can hinder its practical implementation. Hence, three low-complexity suboptimal beamforming algorithms are designed. Simulationresults show the effectiveness of the resource allocation scheme in enhancingthe performance. Moreover, the proposed scheme offers robust delivery ofbasic quality video to all users. Furthermore, all the proposed beamformingtechniques offer significant performance gain over single-antenna multicastsystem. This thesis also studies the performance of NOMA and coded mul-ticasting as the delivery approaches in cache-aided network with the aim ofdiscovering which technique performs better under different pairing scenar-ios. Finally, a hybrid delivery scheme along with its resource allocation is de-veloped in order to further improve the performance. Numerical results showthat NOMA outperforms coded multicasting when the paired users’ channelgains are highly distinctive, while coded multicasting is preferred when thepaired users have similar channel gains. In addition, the hybrid scheme offersimproved sum rate performance over NOMA and coded multicasting.

17

Declaration

No portion of the work referred to in the thesis hasbeen submitted in support of an application for an-other degree or qualification of this or any other uni-versity or other institute of learning.

18

Copyright

i. The author of this thesis (including any appendices and/or schedulesto this thesis) owns certain copyright or related rights in it (the “Copy-right”) and s/he has given The University of Manchester certain rightsto use such Copyright, including for administrative purposes.

ii. Copies of this thesis, either in full or in extracts and whether in hardor electronic copy, may be made only in accordance with the Copyright,Designs and Patents Act 1988 (as amended) and regulations issued un-der it or, where appropriate, in accordance with licensing agreementswhich the University has from time to time. This page must form partof any such copies made.

iii. The ownership of certain Copyright, patents, designs, trade marks andother intellectual property (the “Intellectual Property”) and any repro-ductions of copyright works in the thesis, for example graphs and tables(“Reproductions”), which may be described in this thesis, may not beowned by the author and may be owned by third parties. Such Intellec-tual Property and Reproductions cannot and must not be made availablefor use without the prior written permission of the owner(s) of the rele-vant Intellectual Property and/or Reproductions.

iv. Further information on the conditions under which disclosure, pub-lication and commercialisation of this thesis, the Copyright and anyIntellectual Property and/or Reproductions described in it may takeplace is available in the University IP Policy (see http://documents.manchester.ac.uk/DocuInfo.aspx?DocID=24420), in any rele-vant Thesis restriction declarations deposited in the University Library,The University Library’s regulations (see http://www.library.

manchester.ac.uk/about/regulations) and in The University’spolicy on presentation of Theses

19

http://documents.manchester.ac.uk/DocuInfo.aspx?DocID=24420

http://documents.manchester.ac.uk/DocuInfo.aspx?DocID=24420

http://www.library.manchester.ac.uk/about/regulations

http://www.library.manchester.ac.uk/about/regulations

Acknowledgements

Praise be to Allah who has blessed me with strength, patience and health tocomplete my thesis. I would like to express my deepest gratitude to my su-pervisor, Dr. Daniel Ka Chun So whose guidance, support and expertise wereinvaluable in accomplishing my research work. This work would not havebeen possible without his patience and dedication to high-quality research. Iwould also like to express my gratitude to the Government of Brunei Darus-salam for granting the scholarship and giving me the opportunity to pursuemy studies.

I would like to thank Universiti Teknologi Brunei (UTB) for continuouslysupporting me throughout my research, particularly Dr. Hjh Noor Maya bintiHaji Md Salleh and Dr Ang Swee Peng whose inspiring advice have upliftedmy motivation in completing the PhD. Many thanks to my colleagues inElectrical and Electronic Engineering of UTB and friends in the Universityof Manchester for their support and knowledge sharing.

I would like to express my deepest gratitude to my parents (Pehin DatoHaji Dani Haji Ibrahim and Datin Hajah Norhayati Haji Abu Bakar), myparents-in-law (Dato Haji Metussin Haji Baki and the late Datin Dr. HajahSaadiah Haji Tamit), my siblings and in-laws for their endless love, supportand prayers.

Finally, I would like to give special thanks to my wife, Dr. Halimatus-saadah Dato Haji Metussin whose love, understanding and patience has al-ways keep me up during the hard times. Not to forget my children, Ameer So-lahuddeen, Nor Areesya Syamimi, Noorussyifa Az-Zahra and Nur AmeerahAmalina who bring joy and happiness in my life.

20

In the name of Allahthe Most Beneficent, the Most Merciful

Dedicated toMy Father and Mother

21

Chapter 1

Introduction

1.1 Trends in Mobile Networks

The growth of mobile subscribers may lead to the explosively increasing mo-bile data traffic. The latest data from [1] indicates that the global mobilebroadband subscriptions have reached the 5.1 billion figures in 2018 and itis predicted to grow to 5.7 billion by 2023. Furthermore, the number ofconnected devices may surpass the global population in the future due tothe growth of machine-to-machine (M2M) communications and Internet ofThings (IoT) applications, which further contributes to global mobile trafficgrowth. In addition, the demand for rich multimedia content has increaseddrastically in recent years and hence inevitably contributes to further increasein mobile traffic. As highlighted by a study in [2], mobile video content con-tributed more than half of the global data traffic since 2012 and it is expectedto reach 78% of global traffic by 2021, prompting the need for thousand-foldincrease in capacity and wireless resources in order to accommodate the mo-bile traffic demand. Innovative solutions, including massive MIMO, networkdensification and advanced multiple access schemes, are crucial to addressthe challenges of the ever-increasing capacity requirement [3].

In the last decades, the world has witnessed the evolution of mobile com-munication technology from the voice-centric first generation (1G) to the data-centric fourth generation (4G), which is associated with enhancement in termsof data rates, capacity and wide-range applications. In particular, the analog1G system implements Frequency Division Multiple Access (FDMA) to sup-port solely voice application. With the advancement of digital communication

22

CHAPTER 1. INTRODUCTION 23

system, Time-Division Multiple Access (TDMA) becomes more practical andis deployed in the second generation (2G) system for voice and text applica-tions. Code-Division Multiple Access (CDMA) is applied in the third genera-tion (3G) system to enable voice, video and data services. The 4G Long TermEvolution-Advanced (LTE-A) employ Orthogonal Frequency Division Multi-ple Access (OFDMA) and single carrier FDMA (SC-FDMA) for downlink anduplink respectively to support ubiquitous mobile broadband services whichrequire high data rates, improved latency and high capacity. This is madepossible with advanced technologies such as carrier aggregation (CA), co-ordinated multipoint (CoMP), heterogenous network (HetNet) deployment,enhanced multimedia broadcast and multicast service (eMBMS) and full di-mension multiple-input multiple-output (FD MIMO) [4–6]. The performancerequirements for LTE-A are peak data rate of 1 Gbps and 500 Mbps for down-link and uplink respectively, latency of less than 5 ms and support mobilityfor speed up to 350 km/h [7]. Nevertheless, these performance targets willnot be able to support advanced and critical future applications such as tac-tile communication, virtual reality (VR), augmented reality (AR), high qualityvideo streaming, and real-time remote control and monitoring.

It is highly anticipated that the fifth generation (5G) technology will beable to support advanced and critical use cases or applications which arebroadly identified by International Telecommunication Union (ITU) as en-hanced mobile broadband (eMBB), massive machine type communication(mMTC), and ultra-reliable and low latency communication (URLLC). In or-der to accommodate the requirements for these use cases, the ITU has definedkey performance targets which include the peak rate of 20 Gbps and 10 Gbpsfor downlink and uplink respectively, 100 Mbps user experience data rate, la-tency of less than 1 ms, mobility up to 500 km/h and connectivity density of1 million devices per square kilometre [8]. The key enabling technologies for5G includes network slicing, mobile edge computing (MEC), millimetre wave(mmWave) communications, cloud radio access network (C-RAN) [9–11]. Re-searches on diverse aspects of 5G networks have been conducted by variousorganizations including ITU, third generation partnership project (3GPP) andMETIS [12]. The first version of 5G system based on 3GPP release 15 specifi-cations is already available for commercial deployment. The second phase ofthe enhancement to the 5G features will be introduced through 3GPP release


16 specifications by 2020. Further enhancements are expected to be made for5G in order to support more advanced services and use cases [13].

1.2 Motivation

The demand for high capacity and data rate in mobile networks increasetremendously with the growing traffic due to the shift of users’ trend frombeing connection-centric (e.g. voice calls and text messages) to content centric(e.g. multimedia downloading, video streaming and multimedia-based socialmedia applications). Moreover, the emerging demand for high quality videocontent and advanced broadband applications (e.g. VR, AR and ultra-highdefinition (UHD) video) poses greater challenges for the current mobile net-work to cope with the high data rate and latency requirements. For example,the video streaming based on 4K/8K UHD needs a data rate of around 300

Mbps [9], which burden the network resources of the current network particu-larly when many user devices are served. This motivates the establishment ofcontent-centric mobile network (CCMN) for next generation mobile networks,in which, wireless edge caching and multicast technology are exploited to of-fload the traffic caused by multimedia content delivery [14, 15].

Multicasting is a technology which offers increased efficiency in terms ofresource utilization by allowing the delivery of common multimedia contentto multiple users over shared link. Furthermore, this technique effectivelyreduces the transmission power at the base station (BS). In wireless edgecaching, users are allowed to retrieve the popular contents directly from thestorage devices equipped at the BS or user equipment (UE), instead of obtain-ing these contents from remote servers. Therefore, backhaul link capacity andwireless link resources are utilized more efficiently, and transmission latencyis potentially reduced. Wireless edge caching is considered to be a promisingcandidate technology which leads to a significant paradigm shift in mobilenetworks [10].

Non-orthogonal multiple access (NOMA) has emerged as one of thepromising radio access technology which offers improved spectral efficiency,increased transmission rate and reduced transmission latency [16–20]. In ad-dition, NOMA significantly outperforms the existing orthogonal multiple ac-cess (OMA) scheme in terms of overall throughput performance in macrocell


environment and enhanced performance gain is expected with the implemen-tation of advanced power allocation scheme [18]. With these promising per-formance potentials, NOMA is deemed as one of the candidate radio accessschemes for the future generation mobile network [3]. In 2018, the applica-tion of NOMA in 5G New Radio (5G NR) is considered as a study item in3GPP works. However, further studies are required and thus it was decidedfor NOMA to be implemented beyond 5G [21].

Resource allocation has been recognized as a key technique to maximizea variety of network performance including sum rate maximization. In par-ticular, the resource allocation in NOMA system has gained much interestin literature as significant performance gain can be achieved over OMA sys-tem [22]. However, few works have studied the resource allocation for NOMAin multicast and cache-enabled networks. This research investigates efficienttechniques, including resource allocation, that potentially enhance the perfor-mance of content-centric mobile network through the integration of NOMAinto both multicast and wireless edge caching technologies.

1.3 Objectives

The main objective of this research is to investigate and design efficient tech-niques that can enhance the performance of NOMA system that is applied tocontent-centric mobile technologies, including multicasting and wireless edgecaching. The specific objectives of the research are as follows

1. To carry out general literature review on NOMA and CCMN technolo-gies, specifically, layered multicast video streaming and wireless cachingat the UEs.

2. To carry out specific literature review on the application of NOMAin layered multicast network and the delivery techniques for cached-enabled network (particularly NOMA and coded multicasting).

3. To investigate a resource allocation technique for layered multicast videostreaming system which exploits the advantages of NOMA and layeredcoding. Moreover, another objective set for this research is to design theresource allocation scheme that offers low complexity while maintainingperformance not far from the optimal one.


4. To propose low-complexity beamforming techniques which can furtherenhance the performance of NOMA-based layered multicast network.

5. To investigate the performance of NOMA and coded multicasting asdelivery techniques for caching at the UEs through analytical and simu-lation studies. Furthermore, it is another aim of this research to proposehybrid NOMA and coded multicasting scheme in order to leverage thebenefits of both delivery techniques. The design of an efficient resourceallocation for the proposed hybrid technique is aimed to further enhancethe performance.

1.4 Contributions

The main contributions of this thesis are listed as follows

C.1 Proposed two low complexity suboptimal power allocation schemes fortwo-layered multicast video streaming in NOMA network by consid-ering arbitrary subgrouping. The solutions are derived from sum ratemaximization problem which ensures both transmission power budgetand proportional rate constraints are achieved. A closed-form solutionis derived based on the assumption that the power allocated for eachresource block (RB) is equal, and transformation of the KKT conditionrelated to proportional rate constraint into a different form. Anotherscheme is based on iterative subgradient method which is demonstratedto achieve performance that is comparable to the optimal one. De-spite the superiority of subgradient method, the closed-form solutionis more suitable for practical implementation due to its low complex-ity. It is shown that both proposed schemes outperform all existinglow-complexity schemes in terms of sum rate performance and fairnesstowards the users with poor channel gains.

C.2 Extended the investigation for a general case with multiple layers bygeneralizing the two-layer based closed form solution in C.1. In partic-ular, a low-complexity power allocation scheme for general multi-layercase is developed by successively allocating power for each layer usingthe solution in C.1. Although the solution is based on the modification


of the scheme proposed in C.1, it is demonstrated that the proposedscheme maintains its property of achieving proportional fairness, whichguarantees substantial transmission rate for the high-priority base layerstream. This feature is crucial in multicast video streaming since it en-sures robust delivery of basic quality video to all users.

C.3 Investigated low-complexity joint power allocation and subgroupingschemed for NOMA-based multi layer multicast video streaming. Thejoint optimization of power allocation and subgrouping is a mixed in-teger non-linear programming (MINLP) which is a complex problemthat requires high-complexity numerical solution. In order to solvethis problem at low complexity, the main optimization problem is splitinto two subproblems. The power allocation subproblem is solved us-ing the solution developed in C.2. This power allocation technique isthen embedded into the proposed iterative subgrouping methods to es-tablish a scheme that jointly optimizes the power allocation and sub-group formation. The subgrouping method proposed in other literatureis mainly based on high complexity exhaustive search which hinderspractical implementation. This work demonstrates that the proposedlow-complexity subgrouping methods achieve performance that is closeto exhaustive search at much lower complexity.

C.4 Proposed three low complexity beamforming schemes for two-layeredmulticast video streaming in NOMA network. The first method is de-signed to direct the layer beams towards the weakest user in each respec-tive subgroup. This approach will improve the sum rate since the trans-mission rate depends on the minimum of the users’ achievable rates,which is mainly the rate of weakest user. The second technique poten-tially enhance the performance by nullifying the interference caused bythe enhancement layer towards the users in subgroup 1 while directingthe beams towards selected users. The final beamforming method isbased on iterative algorithm which proportionally improves the achiev-able rate of each user, leading to enhanced transmission rate. Subjectto the beamforming weights obtained using the proposed methods, thepower allocation can be solved using any existing power allocation tech-niques, including the solution proposed in C.1.


C.5 Studied the performance of NOMA and coded multicasting as contentdelivery techniques in cache-aided network. This is performed by de-termining the analytical expression for the probability that the sum rateof NOMA attains a specific performance gain over coded multicasting.This expression is obtained from the probability density function (PDF)for the ratio of the channel gains of paired users. This PDF expressionwhich consider the case where users are uniformly located in circularcell and are affected by Rayleigh fading is derived in this thesis, sinceit is not available in any literature. Both analytical and simulation re-sults provide an insight on which content delivery technique performsbetter under different pairing scenarios. Furthermore, the analytical ex-pression for the outage probability is also derived in this thesis in orderto examine the performance of individual users in both delivery tech-niques.

C.6 Proposed a hybrid NOMA and coded multicasting scheme for cache-aided network in order to exploit the advantages of both delivery tech-niques under different pairing scenarios. In this hybrid method, jointpower and RB allocation or user pairing is first performed by consid-ering NOMA as a delivery technique. This optimization problem isformulated as MINLP which can then be transformed to nonlinear pro-gramming through the relaxation of the integer variables constraint.Near-optimal solutions can be obtained by using iterative subgradientmethod. In this thesis, we also propose scheme that maintain goodsum rate performance while reducing the complexity. Based on theseresource allocation techniques, the selection of either NOMA or codedmulticasting as a delivery mode is performed for each RB using a con-ditional expression which is derived based on the sum rate comparisonbetween the two delivery techniques.


1.5 Thesis Organization

The rest of this thesis is organized as follows.Chapter 2 provides the theoretical background and overview on various

areas covered in this thesis. It includes the fundamental knowledge of wire-less channel models on which the simulation works in this research are based.The basic concept of NOMA is presented particularly for downlink communi-cation network. This chapter also presents the general idea on the applicationof NOMA in multicast and cache-enabled networks. The overview of layeredmulticast network and delivery techniques for caching at the UEs are alsobriefly discussed in this chapter.

Chapter 3 focuses on resource allocation that is aimed towards maximiz-ing the overall sum rate of NOMA-based layered multicast video streamingnetwork. The proposed power allocation for two-layer case with arbitrary sub-grouping is discussed in this chapter. This chapter also presents the proposedpower allocation method that successively allocates power for each layer inmulti-layer multicast system in NOMA network. Along with this power allo-cation, subgrouping techniques are also proposed in order to jointly optimizepower allocation and subgroup formation at low complexity.

Chapter 4 proposes beamforming techniques for two-layer multicast sys-tem in NOMA networks. The three subgrouping methods proposed inthis chapter is designed to achieve significant performance gain over single-antenna NOMA network at low complexity.

Chapter 5 investigates the performance of NOMA and coded multicastingin cache-aided NOMA. This is performed by deriving the analytical expres-sions for the probability of sum rate gap between the two delivery techniquesand outage probability. In this chapter, the hybrid NOMA and coded multi-casting scheme is proposed along with its resource allocation schemes withthe aim of improving the performance of a network with cache-enable UEs.

Finally, the conclusions of this thesis and outlines of the future extensionof the work are discussed in Chapter 6. This chapter is followed by the list ofreferences.


1.6 List of Publications

P.1 Chapter 3: M. N. Dani, Z. Q. Al-Abbasi and D. K. C. So, "Power Alloca-tion for Layered Multicast Video Streaming in Non-Orthogonal MultipleAccess System," in IEEE Globecom Workshops (GC Wkshps), 2017.

P.2 Chapter 3: M. N. Dani, D. K. C. So, J. Tang and Z. Ding, "ResourceAllocation for Layered Multicast Video Streaming in NOMA Systems,"IEEE Transaction on Vehicular Technology (submitted).

P.3 Chapter 4: M. N. Dani, D. K. C. So, J. Tang and Z. Ding, "Low-complexity Beamforming Algorithms for NOMA-based Layered Mul-ticast Systems," IEEE Transaction on Vehicular Technology (under prepara-tion).

P.4 Chapter 5: M. N. Dani and D. K. C. So, "On the Performance of NOMAand Coded Multicasting in Cache-Aided Wireless Networks," in IEEEInternational Conference on Communications Workshops (ICC Workshops),2019.

P.5 Chapter 5: M. N. Dani, D. K. C. So, J. Tang and Z. Ding, "NOMA andCoded Multicasting in Cache-Aided Wireless Networks," IEEE Transac-tion on Wireless Communications (under preparation).

Chapter 2

Background and Overview

2.1 Introduction

This chapter presents the background and overview related to this thesis.First, the wireless channel models applied in the simulation works of this the-sis is described. In addition, this chapter includes the basic concept of NOMA,which provides the fundamental foundation for the study of NOMA in bothlayered video multicast streaming and cache-aided networks. This chapteralso provides the overview of the layered video streaming in multicast net-works, which includes the challenges in reception reliability and resource al-location. Finally, this chapter presents the overview of the delivery techniquesin wireless caching at the UEs, which mainly focuses on coded multicastingand NOMA.

2.2 Wireless Channel Models

The performance of a wireless communication system is severed by the pres-ence of noise, interference and other channel impairments. Due to user mobil-ity and diverse building or terrain features, these impairments unpredictablyvary over time, posing greater challenges in wireless communication systemanalysis. One of the channel impairments is the received signal attenuationcaused by the effect of path loss and shadowing. Path loss is caused by energydissipation from transmit antenna as well as the effects of wave propagationthrough space which leads to the decrease in signal strength as the distance

31

CHAPTER 2. BACKGROUND AND OVERVIEW 32

between transmitter and receiver increases. Meanwhile, shadowing is a signaldegradation which occurs when the signal is obstructed by large objects suchas buildings and hills. Path loss and shadowing can be classified as large-scale propagation effects since the signal variation caused by these effectsoccurs over large distance [23–25].

In practical environment, electromagnetic waves emitted by a transmittermay propagate towards one or more receivers with terrains, buildings andother objects in between by means of any of the three basic mechanisms; re-flection, diffraction and scattering [23,24]. Obstructions from any objects mayrestrict a direct line of sight (LoS) communication between a transmitter and areceiver, and instead, the receiver may receive multiple signals resulting fromreflection, diffraction or scattering by various objects. These signals propa-gate along different paths of varying lengths and may arrive at the receiverat different time instant. These multiple arrival signal paths cause signal fluc-tuations due to constructive and destructive interferences. This phenomenonis referred to as multipath fading. Moreover, the movement of a transmitteror receiver may also lead to time variation in a channel due to Doppler fre-quency shift. This effect is sometimes known as time variant fading. Bothmultipath fading and time variant fading are categorized as small-scale prop-agation effects due to the fact that the amplitude variation occurs over verysmall distances (in the order of wavelength) [23–25].

The channel models for large-scale and small-scale propagation effects arefurther described in the subsequent subsections.

2.2.1 Large-Scale Propagation Effects

The simplest signal propagation model is the free space model, which as-sumes the absence of obstacles between transmitter and receiver. In otherwords, this model considers the signal to propagate in direct LoS path andneglect the signals which arrive from reflected objects. The free space modelis represented by the Friis equation [26] as follows

Pr(d) =PtGtGrλ

2

(4πd)2 =PtGtGr

PLfs(2.1)


where Pr(d) is the received power as a function of distance between transmit-ter and receiver, Pt is the transmit power, Gr and Gt are the gains of the receiveand transmit antenna respectively, λ is the wavelength, d is the distance be-tween transmitter and receiver, and PLfs =

(4πdλ

)2 is the path loss expressionwhich is widely referred to as free-space path loss model. From (2.1), thereceived power Pr is inversely proportional to the square of the distance d,indicating that the received signal drops rapidly in relation to the distance. Inaddition, the received power Pr is directly proportional to the square of thewavelength of the signal. This means that the received power decreases withthe increase in carrier frequency (decrease in wavelength) [23, 24].

In general, a propagation model can be expressed in dB (i.e., Pr and Pt

either in dBW or dBm) as

Pr (d) = Pt +Gt +Gr − PL(d) (2.2)

where PL(d) is the signal attenuation due to path loss expressed in dB whichcan be determined by various models, including the free-space path lossmodel which is given as

PL(d) = PLfs = 20 log

(4πd

λ

). (2.3)

Free-space path loss does not represent an accurate model since it con-siders simple propagation environments. Models based on empirical mea-surements have been developed to predict the path loss in specific wirelessenvironment including urban, rural and indoors. However, these models areapplicable over a given distance in a specified frequency range. For example,the Okumura-Hata model is only valid for the frequency range of 150-1500

MHz over a distance of more than 1 km in urban, suburban or rural environ-ment [23]. The current and future mobile communication system operates at avery wide range of frequencies including frequencies higher than 2 GHz andthe system is characterized by intensive deployment of small cell and indoorBSs. Hence, Okumura-Hata cannot be used extensively in current and futuremobile communication system design. It is difficult to formulate a path lossmodel that is applicable to a wide range of frequencies over various propaga-tion environments. Since all path lost models are approximations anyways,it is common to utilize a simple generalized model for mean path loss in a


communication system analysis. The mean path loss at distance d can beexpressed in terms of path loss exponent γ as

PL(d) = PL(d0) + 10γ log

(d

d0

)(2.4)

where d0 is the reference distance and PL(d0) is the mean path loss at d0 whichis often given as free-space path loss (2.3) at a distance of d0 when the value isnot obtained through measurements. Different propagation environments canbe considered by appropriately choosing the values for d0 and γ. For instance,the reference distance d0 for macrocellular/microcellular system with a cellradius of less than 1 km could be 100 m or 1 m [26]. Meanwhile, the rangeof values for γ in different propagation environment are listed in Table 2.1[24, 26].

Table 2.1: Path loss exponents for various propagation environments

Environment γ range

Free space 2

Urban area cellular radio 2.7 - 3.5

Shadowed urban cellular radio 3 - 5

In building line-of-sight 1.6 - 1.8

Obstructed in building 4 - 6

Obstructed in factories 2 - 3

The presence and movement of various types of objects placed in differ-ent locations as well as clutter variation within the propagation environmentcaused random signal attenuation at different locations of a given distance.This effect, which is referred to as shadowing, occurs due to the blockagefrom objects. This signal variation should also be incorporated in a large-scaleeffect model by statistical means. One such model is known as log-normalshadowing which has been empirically verified to be an accurate model forboth indoor and outdoor environments [23]. By taking into account the effect


of shadowing, the path loss at a distance d in dB is formulated in [26, 27] as

PL(d) = PL(d) +Xσ = PL(d0) + 10γ log

(d

d0

)+Xσ (2.5)

where PL(d) is the mean path loss at distance d according to (2.4) and Xσ isthe Gaussian distributed random variable (in dB) with zero mean and stan-dard deviation σ (in dB). In (2.5), the path loss is a random variable withmean PL(d) dB and standard deviation σ dB. This model is implementedin the simulation works of this thesis in order to calculate the large-scale ef-fect experienced by the randomly placed users in downlink communicationsystem.

2.2.2 Small-Scale Propagation Effects

The types of fading depend on the time dispersion and frequency disper-sion effects of the channel. Time dispersion is caused by the multipath delayspread leading to fading over a frequency domain, which is either flat or fre-quency selective fading [23, 24]. On the other hand, frequency dispersion iscaused by Doppler spread leading to time variant fading [23–25]. Flat fad-ing is associated with a channel which has relatively constant gain and linearphase response over its signal bandwidth. It occurs when the bandwidth ofthe signal is less than the coherence bandwidth. In other words, the symbolperiod should be greater than the RMS delay spread for flat fading to hap-pen. In frequency selective fading scenario, the channel gains vary over thefrequency components of the signal bandwidth. The variation occurs whenthe signal bandwidth is greater than the coherence bandwidth or the symbolperiod is smaller than the RMS delay spread. The implementation of efficientfrequency domain equalizer is one of the techniques to address the effects offrequency selective fading [28]. On the other hand, the channel is deemedto experience fast fading when the signal fluctuates rapidly over time thatthe channel response varies within the symbol period. This only occurs ifthe symbol period is greater than the coherence time. Meanwhile, slow fad-ing incurs signal variation at a slower rate relative to the symbol period. Slowfading occurs when the symbol period is much lower than the coherence time.


The most common fading channel models are Rayleigh and Rician mod-els which are based on statistical means. Rayleigh fading model assumesthe arrival of the infinite number of signals paths at the receiver at the sametime at all angles. In addition, this model considers the absence of dominantLoS path and therefore all paths have zero mean and equal variance. It isalso assumed that all path gains are statistically independent. All of theseassumptions meet the criteria for central limit theorem which stipulates thateach of the quadrature components of the channel gain obeys Gaussian dis-tribution [23]. Rayleigh fading is characterized as random complex numberswhose real and imaginary components are independently and identically dis-tributed Gaussian random variables with zero mean and equal variance [25].Meanwhile, Rician fading model considers the presence of a single dominantLoS path such that the magnitude of the signal varies according to Riciandistribution [23].

Rayleigh Fading

Simulator a0

Rayleigh Fading

Simulator a1

Rayleigh Fading

Simulator aN

. . .

τ1 τN . . .

∑

Signal s(t)

r(t)

Figure 2.1: Tap delay line filter representation of frequency selective fading

Frequency selective fading can be modelled as a tap delay line filter [26] asshown in Figure 2.1. Each tap is associated with independent Rayleigh fad-ing as well as different gain and time delay to reflect the frequency selectivefading effects. If the LoS component is present, the first tap is changed fromRayleigh to Rician fading simulator and the rest of the taps are maintained asRayleigh simulator. The channel gain fluctuations in frequency domain canbe obtained by implementing Fast Fourier Transform (FFT). The values for


the gain and relative time delay are typically based on existing models, suchas the ITU Pedestrian model which is tabulated in Table 2.2 [27, 29]. In thisthesis, the small-scale effect is modelled using the ITU Pedestrian ChannelB model, where the power delay profile and the generation of channel gainsover 10 MHz spectrum are illustrated in Figures 2.2(a) and 2.2(b) respectively.

Table 2.2: The parameters for ITU Pedestrian Model

TapChannel A Channel B Doppler

SpectrumRelativedelay (ns)

AveragePower (dB)

Relativedelay (ns)

AveragePower (dB)

1 0 0 0 0 Classic

2 110 -9.7 200 -0.9 Classic

3 190 -19.2 800 -4.9 Classic

4 410 -22.8 1200 -8.0 Classic

5 2300 -7.8 Classic

6 3700 -23.9 Classic

0 1000 2000 3000 4000

Relative delay (ns)

0

0.2

0.4

0.6

0.8

1

Ave

rage

pow

er

(a)

0 2 4 6 8 10

Frequency (MHz)

-30

-20

-10

0

10

Cha

nnel

gai

n (d

B)

(b)

Figure 2.2: (a) Power delay profile and (b) the channel gain of a multipathchannel with respect to frequency based on ITU Pedestrian Channel B Model


2.3 Non-Orthogonal Multiple Access (NOMA)

NOMA allows multiple users to be allocated in the same frequency and timeresources by exploiting another domain such as power or code, producingintentional non-orthogonality for signals among the users and thus allowingmutual interference among the users. NOMA can be further classified intotwo major categories; power domain and code domain NOMA [17,19,30–32].The multiple access schemes such as sparse code multiple access (SCMA),low-density spreading CDMA (LDS-CDMA) and multiuser shared access(MUSA) are classified as code-domain NOMA.

Based on the concept of CDMA, spreading gain and shaping gain can beachieved by code domain NOMA at the expense of increased bandwidth.SCMA is an advanced version of LDS-OFDM which exploits sparse non-orthogonal spreading method to support massive connectivity. This tech-nique enables direct mapping of bit streams to different sparse codewordsand consequently multiplex the different codewords for all users over sharedorthogonal resources. A message passing algorithm (MPA) is employed atthe receiver to allow multi-user detection [16, 19, 30, 33]. Based on the con-cept of multi-carrier CDMA (MC-CDMA), the uplink MUSA scheme spreadthe symbols by applying advanced low correlation spreading sequences andthese sequences are transmitted over shared resources. At the receiver, SIC isimplemented to extract the data for each user [16,19,30]. The user overloadingfactor of MUSA scheme is high and therefore it can achieve significant systemperformance gains [16].

Power domain NOMA improves the spectral efficiency by multiplexing theusers’ signals in power domain using superposition coding (SC) at the trans-mitter and applying successive interference cancellation (SIC) at the receiverto detect the users’ signals. In addition, power domain NOMA can achieverobust performance gain regardless of channel state information (CSI) feed-back latency and UE mobility [34, 35]. The knowledge of CSI of receiversare required by transmitter for transmit power allocation and user pairing.Meanwhile, the CSI is required at the receiver for user demultiplexing [18].

Pattern Division Multiple Access (PDMA) is another NOMA schemewhich multiplexes the users’ signals in multiple domains including power,code, spatial or any combination of these domains. Therefore, PDMA is not


categorized in either power domain or code domain NOMA. In PDMA, non-orthogonal patterns are applied at the transmitter prior to the actual multi-plexing. These patterns are designed in such a manner that the diversity ismaximized and the overlaps among multiple users are minimized. The mul-tiple domain features of PDMA offers high flexibility in the implementationof coding and decoding techniques [16, 19, 30].

For brevity, the power domain NOMA is referred to as NOMA through-out this thesis. In addition, the other NOMA schemes will not be explainedhereinafter and therefore the term ’NOMA’ will only refer to power domainNOMA scheme.

2.3.1 Superposition Coding (SC)

In NOMA, the signals of multiple users are superposed in power domain us-ing SC at the transmitters. SC is a non-orthogonal approach which combineuser’s information into a single signal source and hence able to achieve the ca-pacity on a scalar Gaussian broadcast channel [17]. This allows multiple users’information to be transmitted at the same frequency and time resources.

Consider an example of a two-user scenario in which K = 1, 2 denotesthe set of users. The data bits for the kth user are mapped into its respectivem-ary modulation constellation to form complex-valued symbols, xk. For thisexample, quadrature phase-shift keying (QPSK) is considered for both usersas illustrated in Figure 2.3. In SC, the constellation of user 1 (with highertransmit power) is superposed with that of user 2 (with lower transmit power)to form superposed constellation as shown in Figure 2.3(c). The superposedsymbol for two-user scenario can be represented as

xs =√PTα1x1 +

√PTα2x2 (2.6)

where PT is the total transmission power and αk is the fraction of PT fork = 1, 2 which is subject to α1 + α2 = 1.


𝑷𝑻𝜶𝟏

(a) Signal constellation of user 1

𝑷𝑻𝜶𝟐

(b) Signal constellation of user 2

𝑷𝑻𝜶𝟐

𝑷𝑻𝜶𝟏

𝑷𝑻

(c) Constellation of superposed signal

Figure 2.3: An example of superposition coding

2.3.2 Successive Interference Cancellation (SIC)

The superposed signal is then transmitted over the same frequency and timeresources. At the receiver, the wanted signals are detected by using SIC mul-tiuser detection technique which is enabled by ensuring large differences inpower allocation between the users at the transmitter [31, 33–35]. The signalreceived by user k is denoted by

yk = hkxs + wk = hk

(√PTα1x1 +

√PTα2x2

)+ wk (2.7)


where hk is the complex channel coefficient at user k and wk denotes theadditive white Gaussian noise (AWGN). At the receiver of user 1, the desiredsignal is decoded immediately from the received signal y1 by consideringthe signal of user 2 as interference or noise. This is possible due to the factthat the received power of user 1’s signal is higher than that of user 2 (i.e.,α1 > α2 ). However this is not the case for user 2. At the receiver of user2, the signal of user 1 is decoded first. Assuming that the signal of user 1

is decoded perfectly at the receiver of user 2, the decoded signal of user 1

is y(1)2 =

√PTα1h2x1. Then, the recovered signal of user 1 is used to cancel

itself (as interference) from the received signal for the detection of the user 2’ssignal. The resulting signal prior to the detection of the signal of user 2 is y′2 =

y2−y(1)2 =

√PTα2h2x2 +w2. The SIC process for a two-user downlink scenario

and the symbol detection are illustrated in Figures 2.4 and 2.5 respectively.

User 2

User 1

BS|h2|

2

|h1|2

|h1|2 < |h2|

2

Decoder𝒚𝟏 = 𝒉𝟏 𝑷𝑻𝜶𝟏𝒙𝟏 + 𝑷𝑻𝜶𝟐𝒙𝟐 +𝒘𝟏 𝒙𝟏

Decoder𝒙𝟏

Decoder𝒙𝟐

𝑷𝑻𝜶𝟏𝒉𝟐𝒙𝟏

−

𝑷𝑻𝜶𝟐𝒉𝟐𝒙𝟐 +𝒘𝟐

𝒚𝟐 = 𝒉𝟐 𝑷𝑻𝜶𝟏𝒙𝟏 + 𝑷𝑻𝜶𝟐𝒙𝟐 +𝒘𝟐

Figure 2.4: SIC process for two-user downlink scenario

In NOMA, the two types of interference cancellation being studied are thesymbol-level interference cancellation (SLIC) and codeword-level interferencecancellation (CWIC) [35]. Figure 2.6 illustrates the receiver structure of thestronger user 2 employing both SLIC and CWIC techniques in a SIMO-basedNOMA system with two users [35]. First, the received signals are multipliedwith the maximal ratio combining (MRC) weights in order to obtain the datasymbols for the interfering signal, i.e., signal of weaker user 1. Then, thelog-likelihood ratios (LLRs) calculations are performed for these symbols. ForSLIC receiver, the interfering symbols are detected by directly using LLRs andsymbol estimator and thus the interference cancellation is performed without


prior Turbo decoding. On the contrary, for CWIC receiver, the interferingsymbols are detected and decoded using LLRs, Turbo coding and symbolestimator. Then, for both SLIC and CWIC, the interfering symbols of user 1

are substracted from the received data symbol. Finally, the signal of strongeruser 2 is recovered by applying MRC and turbo decoding.

Decodedsymbol x1

Received symbol y1

(a) Decoding of the signal of user 1

Decoded Symbol x1

Received Symbol y2SIC

(cancels x1

from y2)

Decoded Symbol x2

1

2

3

(b) Decoding of the signal of user 2

Figure 2.5: An example of NOMA symbol decoding

MRC

MRC

CP Deletion

FFTMRC

P/S

co

nve

rter

LLR computation

Channel deinterleaver

TurboDecoder

Symbol Estimator

−

−

MRC

P/S

co

nve

rter

LLR computation

Channel deinterleaver

TurboDecoder

Channel estimator

Decoder detecting signal of user 1

Decoder detecting signal of user 2

Recovered signal

CWIC

SLIC

Figure 2.6: Receiver structure of user 2 employing SLIC and CWIC


From the link level performance investigations in [36], [37] and [38], CWICreceiver outperforms SLIC receiver and achieves almost the same block er-ror rate (BLER) performance as ideal SIC, particularly when the power al-location factor for the strong user is below 0.35 (i.e., α2 ≤ 0.35). In otherwords, the power allocation factor for the weak user must be above 0.65 (i.e.,α1 ≥ 0.65) in order to keep the performance of CWIC similar to that of idealSIC. These works demonstrated that the CWIC receiver is capable of mitigat-ing both inter-user and inter-stream interference. Furthermore, it is demon-strated in [35] that the performance difference between CWIC and ideal SICis considerably minimal and hence the error propagation in CWIC is signifi-cantly small. However, the performance of CWIC is largely affected by higherimplementation complexity and latency. A practical downlink NOMA systembased on software defined radio has been implemented in [39] using CWICreceiver in order to validate the feasibility of NOMA in practical wireless net-works.

On the other hand, SLIC receiver reduces the complexity of NOMA re-ceiver, but experiences substantial performance degradation if the power allo-cation factor for the weak user is not sufficiently high [38]. It is shown in [35]that the performance of SLIC is almost similar to CWIC when the power al-location ratio is α1 : α2 = 0.9 : 0.1. However, if the power allocation ratio isα1 : α2 = 0.7 : 0.3, the performance is degraded by 6.2 dB compared to idealSIC because the detected symbols for the interfering signal (weak user’s sig-nal) are not accurate. Therefore, it is necessary to keep the power allocationfactor of the weak user larger than a specific threshold in order to maintaingood performance.

2.3.3 NOMA in Downlink Communication Systems

Consider a single cell multicarrier downlink cellular system that consists ofa single-antenna BS and K single-antenna users who are indexed as k wherek ∈ K = 1, 2, 3, ..., K. The BS sends data at a total transmit power PTover a total system bandwidth BT which are divided into N RBs of equalbandwidth B. The bandwidth of each RB is assumed to be smaller thanthe coherent bandwidth and therefore each RB will experience flat fading.However, the whole system bandwidth experiences frequency selective fad-ing. The users on each RB n are sorted in an ascending order of their channel


gains, i.e., |h1,n|2 < |h2,n|2 < |h3,n|2...... < |hK,n|2. With the aid of SC andSIC techniques at the BS and UE respectively, NOMA allows multiple usersto be served by a single RB. On the other hand, in OMA scheme such asOFDMA, a single RB can only serve a user. Through SC at the BS, users’ dataare multiplexed in power domain by allocating a fraction αk,n (0 < αk,n < 1)of the total transmit power PT to user k. Therefore, the power allocated toeach user k at RB n is determined as Pk,n = αk,nPT where the condition∑N

n=1 (α1,n + α2,n + ...+ αK,n) = 1 must be met to ensure the total transmitpower does not exceed the maximum power budget PT . In SC, higher trans-mit power is generally allocated to the signal of weaker user and lower powerto the stronger user (i.e., P1,n > P2,n > P3,n...... > PK,n > 0) in order toguarantee user fairness or satisfy certain quality of service (QoS) require-ments [16, 18, 34]. However, in terms of information-theoretic point of view,this is not necessarily the case. According to [40], power allocation dependson the specified points on the capacity region of the paired users. Therefore,less power can be allocated to the weaker user as long as the QoS require-ments for all users are satisfied.

The superposed signal received by each user k at RB n can be depicted as

yk,n = hk,nxs + wk,n (2.8)

where the superposed signal xs =∑K

k=1

√PTαk,nxk.n and wk,n is the AWGN

with power spectral density (PSD) of N0. Since the weakest user signal isbeing allocated with the highest transmit power, the weakest user 1 imme-diately decodes the intended signal by considering the signal of other users(with higher user index, k > 1) as interference which is termed as intra-cellinterference or intra-cluster interference [17]. The second weakest user 2 firstdecodes the signal of user 1. The decoded signal of user 1 is then cancelledfrom the received signal to allow the receiver to decode the wanted signal bytaking into account the signal of users with higher user index, k > 2 as inter-ference. Therefore, each user k is able to decode the signal of the weaker usersj where j < k and subsequently the decoded signal is successively cancelledfrom the received signal in order to recover the desired signal by treating thesignal of the stronger users l where l > k as interference. This process willcontinue at the receiver of the strongest user, that is, the K th user, in which,through ideal successive cancellation of interference, the signal is decoded


without any interference. Assuming perfect SC and SIC at the BS and UErespectively, the achievable data rate for user k at each RB n is represented as

Rk,n = B log2

(1 +

Pk,n|hk,n|2

|hk,n|2∑K

i=k+1 Pi,n +BN0

). (2.9)

The achievable rate for the strongest K th user at each RB n is given by

RK,n = B log2

(1 +

PK,n|hK,n|2

BN0

). (2.10)

Consider an example of a single cell NOMA based-downlink network with2 users where user 1 and 2 are weak and strong users respectively (i.e.,|h1|2 <|h2|2). For simplicity, the subscript notation n which indicates RB index isremoved to consider only single RB. Using (2.9) and (2.10), the achievable ratefor user 1 and user 2 are denoted respectively as

R1 = log2

(1 +

P1|h1|2

P2|h1|2 +N0

)= log2

(1 +

αΓ1

(1− α)Γ1 + 1

)(2.11)

R2 = log2

(1 +

P2|h2|2

N0

)= log2 (1 + (1− α)Γ2) (2.12)

where α and 1−α are the fractions of transmit power allocated to user 1 and 2

respectively, and Γ1 = PT |h1|2N0

and Γ2 = PT |h2|2N0

are the received signal-to-noiseratio (SNR) for user 1 and 2 respectively.

For OMA system, a single cell TDMA based-downlink network with 2

users is considered for comparison. Here, the entire timeslot resource cannotbe shared among the 2 users and therefore each timeslot is segregated into thefraction δ and (1− δ) for user 1 and user 2 respectively, where 0 < δ < 1 . Theachievable rate for user 1 and user 2 in OMA two-user scenario are expressedrespectively as

R1 = δ log2

(1 +

PT |h1|2

N0

)= δ log2 (1 + Γ1) (2.13)

R2 = (1− δ) log2

(1 +

PT |h2|2

N0

)= (1− δ) log2 (1 + Γ2) . (2.14)

From (2.11), the rate of user 1 can be enhanced by allocating higher power


P1, but restricted by the signal strength of the interfering user 2 who is allo-cated with power P2. As indicated in (2.12), the rate of user 2 is solely dependson the allocation of power to user 2. Therefore, the overall transmission ratecan be optimized and user fairness can be achieved by regulating the powerallocation ratio P1/P2. In NOMA, the channel gain difference is exploited toattain multiplexing gains by pairing users with large channel gain differences.Although the users are allocated relatively less power due to power sharing,the users instead benefit from being allocated with larger bandwidth. Mean-while, in OMA, multi-user gain is achieved through time or frequency domainscheduling which exploits channel diversity. [18, 34, 35].

0 1 2 3 4 5 6 7

Rate of user 2 (bps)

0

0.2

0.4

0.6

0.8

1

Rat

e of

use

r 1

(bps

)

NOMAOMA

Figure 2.7: Capacity regions of downlink NOMA and OMA

The user rates of NOMA and OMA can be compared by evaluating themultiuser capacity regions as shown in Figure 2.7 [17]. This analysis considersthat the channel is asymmetric to reflect the difference in the received SNRfor the two users, that is, Γ2 > Γ1. For this analysis, the received SNR forthe two users are set as Γ1 = 0 dB and Γ2 = 20 dB. Figure 2.7 shows thatthe capacity upper bound of NOMA is outside that of OMA indicating thatNOMA users can achieve higher performance gain in terms of individualrate as compared to OMA. In addition, this analysis proves the information-theoretic perspective that the rate regions achieved by downlink NOMA are


bounded by the capacity region of broadcast channels [16,31,34]. Consider anexample that assumes the SNR of the two users are known as Γ1 = 0 dB andΓ2 = 20 dB. Assuming that the power fraction allocated to user 1 and user2 are α = 0.8 and 1 − α = 0.2 respectively, the transmission rate of NOMAuser 1 and user 2 are R1 = 0.74 bps and R2 = 4.39 bps respectively. ForOMA, the transmission rate of user 1 and user 2 are R1 = 0.50 bps and R2 =

3.33 bps respectively considering equal timeslot for both users (i.e.,δ = 0.5).From this simple example, it can be observed that NOMA outperforms OMAin terms of achieving both individual user throughput and sum rate. From(2.11) and (2.12), it can be noted that the rate performance gain is achieved byutilizing the entire timeslot even at the expense of degradation caused by theinterference towards user 1 (weak user) and lower power allocated to user 2

(strong user). Furthermore, the spectral efficiency of OMA is low, particularlywhen some resources (in this case timeslot) are allocated to users with poorchannel quality. Meanwhile, in NOMA, the spectral efficiency is significantlyenhanced due to the fact that the strong user is allowed to access all theresources allocated to the weaker user [16]. By exploiting the channel gaindifferences for the users to achieve higher spectral efficiency and user fairness,NOMA is deemed to be one of the most promising candidate multiple accessscheme for the next generation mobile network [17, 31, 34, 35].

It is worth mentioning that, from (2.9) and (2.10), power allocation clearlyinfluence the performance of NOMA in terms of achieving increased overallthroughput and user fairness. The impact of power allocation on fairness hasbeen investigated in [41] and it is shown that high fairness requirement isachieved with appropriate power allocation scheme. Furthermore, the studyin [42] has shown that NOMA outperforms OMA in terms of outage perfor-mance considering appropriate choice of power coefficients and target rates.Therefore, power allocation has been recognized as a key factor in enhancingthe performance of NOMA and consequently has emerged as a promising re-search area in NOMA. The Full Search Power Allocation (FSPA) scheme pro-posed in [18] exploits the full potential of NOMA by achieving high through-put performance, but at the expense of excessive computational complexitywhich hinder practical implementation. Low-complexity power allocation,


such as Fixed Power Allocation (FPA) and Fractional Transmit Power Alloca-tion (FTPA) [18] does not perform closely to the optimal scheme. This moti-vates the investigation of power allocation in NOMA, such as in [43] and [44],which is aimed towards achieving near-optimal performance while offeringlow computational complexity.

From (2.9), it can be observed that the individual user rates may beseverely affected by the interference caused by other users as the numberof users increase, particularly the weakest user. In other words, NOMA isan interference-limited system and therefore performing NOMA for all theusers may not be realistic, particularly if the number of users is very high.Furthermore, the implementation of SC and SIC for higher number of usersmay increase the complexity of both transmitter and receiver [16]. In order toaddress these challenges, user pairing or clustering is applied to NOMA, inwhich the users can be divided into multiple pairs or clusters and each pairor cluster is allocated with different RBs or subbands. Due to the heterogene-ity of users’ channel conditions, the performance of this technique heavilydepends on the selection of users in each pair or cluster. The impact of userpairing is studied in [45] and it is concluded that the performance gain ofNOMA with fixed power allocation over OMA can be further enhanced bypairing the users whose channel conditions are more distinctive.

The resource allocation in NOMA, which consider both power allocationand user pairing, requires complex scheme in order to attain optimal per-formance. Mathematical optimization theories are often applied to solve theresource allocation problem in NOMA such as in [22], [44] and [46].

2.4 Layered Video Streaming in Multicast Net-

works

The multicast features in mobile communication system have evolved sinceits first introduction in the 3G Universal Mobile Telecommunications System(UMTS). Evolved multimedia broadcast multicast service (eMBMS) has beendesigned for LTE-Advance to address the capacity requirement due to theincreasing mobile data traffic dominated by multimedia services (e.g. video


downloading, live streaming, video conferences, etc.) [47]. The multicast fea-tures incorporated in eMBMS offer high resource utilization efficiency by al-lowing the same multimedia content to be delivered simultaneously to mul-tiple users on shared allocated resources. Higher and flexible bit rates, singlefrequency network (SFN) and carrier configuration flexibility are among thekey features of eMBMS which enhance the system performance over the pre-vious standards. The description of eMBMS is given in much detail in [48].Nevertheless, the demand for high quality video (e.g. 4K/8K UHD videos)and the emergence of enhanced content application (e.g. VR and AR) posedgreater challenges for the current eMBMS system to cope with the capacityrequirements. Therefore, the multicast features in eMBMS are likely to be en-hanced in future mobile networks [14,49,50]. Studies has been made recentlyto enhance broadcast and multicast features for 5G mobile networks based on5G NR specifications [51].

In mobile communication networks, multicasting techniques are catego-rized into single-rate and multi-rate transmissions [52]. Practically, the instan-taneous channel gains of mobile users are heterogeneous and therefore eachuser may experience non-uniform capacities. In the conventional single-ratemulticast transmission, the resources are shared among all the users withina multicast group and thus the BS transmits the content to all the users atthe same transmission rate. The simple implementation and low complexityfeatures of the single-rate technique are definitely attractive. Nevertheless, inorder to ensure all users in a particular multicast group are able to decode thecontent with minimal errors, the data rate is constrained by the user with theworst channel quality. In other words, this conventional scheme is associatedwith the trade-off between transmission rate and coverage. This technique isunfair towards the users with good channel quality as they potentially achievehigher transmission rate with unicast system. Moreover, the study in [53] hasshown that this scheme limits the ergodic capacity as the number of userincreases.

This challenge is addressed by the multi-rate scheme at the expense of highcomputational complexity, coding and synchronization issues caused by therequirement to transmit to multiple subgroups [52]. In multi-rate multicasttransmission, the video content is encoded into multiple streams of different


transmission rates. Cell-edge users will only receive the high-priority basic-quality video stream while the higher channel gain users tend to receive highquality video as more enhancement layer streams (including the high-prioritystream) are successfully decoded. Consequently, minimum QoS is guaranteedfor the cell-edge users while other users enjoy higher quality video transmit-ted over higher combined transmission rates. A lot of research interest onmulti-rate scheme has been received due to its potential in achieving userrate differentiation and higher spectral efficiency [50, 52]. The two techniquesmostly discussed in literature are the Information Decomposition Techniques(IDT) and Multicast Subgroup Formation (MSF) [52]. In IDT scheme, the con-tent is encoded into multiple streams of data traffic via layered coding suchas Multiple Description Coding (MDC) [54], Fine Granularity Scalable (FGS)coding [55] and Scalable Video Coding (SVC) [56]. Each stream is transmit-ted at different rates ranging from a low rate stream which is intended to bereceived by all users to higher rate streams which are received by the usersaccording to their handling capacities. The implementation of IDT in HighSpeed Downlink Packet Access (HSDPA) has shown significant efficiency interms of the utilization of resources. Nevertheless, the high computationalcomplexity and signalling overhead are the major issues in IDT [52].

Meanwhile, the MSF technique leverages the combination of the princi-ple of single-rate scheme and IDT to offer higher transmission rate potential.It involves the segregation of users into different subgroups depending onuser’s channel qualities as well as splitting the content into multiple streamsthrough layered coding. The transmission rate of each layer stream is definedby the least channel gain user in each respective subgroup. Similar to IDT, thequality of the content improves as more layered streams are successfully de-coded by the users, particularly those who join the stronger subgroup. Basedon the survey in [52], MSF is very effective in exploiting multiuser channeldiversities in order to achieve higher overall throughput performance. Forinstance, the study in [57] has demonstrated that the application of multi-layered video technique over wireless fading channels effectively enhance thechannel utilization and the system throughput with the implementation ofappropriate rate adaptation scheme.


2.4.1 Layered Video Streaming

Video compression is an essential component in video content delivery systemwhich enables the distribution of video data over limited radio resources, par-ticularly the bandwidth. Video compression standards, such as H.264/AVC,MPEG-2 and MPEG-4, have been designed to achieve a high-quality video athigher compression ratio. However, the conventional compression techniquesfail to support the requirement of various transmission scenarios includingaddressing the issue of users’ channel diversity. In a single-rate multicastvideo streaming using traditional compression technique, all the users re-ceive the video with similar quality at a rate defined by the least channelgain user. The users may experience long delays due to the low transmissionrate, particularly when high-quality video (which is associated with large filesize) is sent. Hence, there is a trade-off between delays and video quality inthis scheme. Meanwhile, the users with high channel quality in the multicastgroup deserve to receive better quality video data at higher transmission rate.Therefore, layered video coding has been designed to address the problemson users’ channel heterogeneity, and also support the requirement of variousdevices and applications (e.g. video streaming, teleconferencing etc.). Layeredvideo coding has emerged as an active research field over the past decades.Comprehensive tutorial and literature survey on layered video streaming canbe found in [58].

In layered video coding, a video content is encoded into multiple layers.The high-priority layer consists of basic quality video data and is termed asthe base layer. The remaining layers, which are referred to as enhancementlayers, progressively improves the video quality at the receiver. It is sufficientfor the users to encode only the base layer in order to extract the video con-tent. However, if the base layer is corrupted or lost during transmission, theenhancement layers are dropped and hence the user fails to obtain the video.A layered video coding technique is illustrated in Figure 2.8. The most ro-bust modulation and coding scheme (MCS) is often applied to the base layerto provide higher transmission resiliency and hence base layer data is trans-mitted at a very low rate. This will ensure that all the users in the multicastgroup are able to decode the base layer stream and at least satisfied withthe reception of basic quality video. Depending on the channel quality, theusers may extract the enhancement layers to improve the quality of the video.


In particular, the users with better channel conditions will be able to receivemore enhancement layers in order to retrieve better quality video at a highercombined rate. This significantly improves the sum rate performance over theconventional single-rate multicast system.

UE 2

BL

Base

EL1

Layered Video

Encoder

Video Source

BL

EL1

BL – Base layer EL1 – Enhancement layer 1 EL2 – Enhancement layer 2

EL2

Transmitter Receiver

Data Decoder

Layered Video

Decoder

Data Decoder

Layered Video

Decoder

Data Decoder

Layered Video

Decoder

Basic quality

Improved quality

High quality

128 kbps

384 kbps

896 kbps

UE 1 UE 2

BL

UE3

EL1 BL

EL2

Base station

Figure 2.8: Illustration of a layered video coding scheme

There are a number of layered video coding techniques which are alreadystandardized and studied in literature. The most common schemes are brieflyexplained as follows:

1. Fine Granularity Scalable (FGS) coding: The FGS coding technique isdefined in MPEG-4 standard. A non-scalable coding is utilized to gen-erate the base layer in order to attain the lower bound of the bit-raterange. Meanwhile, the enhancement layer data is generated by usingbit-plane coding of the Discrete Cosine Transform (DCT) coefficientswhich encode the difference between the original data and the recon-structed data. Depending on the transmission scenario, the receiver will


be able to reconstruct high quality video from the base layer and theenhancement layer streams. The details on FGS can be found in [55].

2. Scalable Video Coding (SVC): In the SVC extension of the H.264/AVCstandard, a video data is encoded into a base layer and a number ofenhancement layers. Each stream consists of a sequence of network ab-straction layer units (NALUs). The header of the NALU contains theinformation of the type of NALU and its associated function in videoreconstruction. Specifically, it describes the scalability property of thecompressed bitstream which include spatial, temporal and video qual-ity. For instance, the adaptation of video properties can be based onspatial resolution. Here, the video is encoded into multiple enhancedsubstreams known as dependency layers. The spatial resolution im-proves with successful detection of the spatial enhancement layers whichheavily dependent on the decoding of the base layer. Meanwhile, thepayload of the NALU contains the compressed video data frame. SVCis described in much detail in [56]. The SVC extension of the latesthigh efficiency video coding (HEVC) standard has been introduced re-cently [59].

3. Multiple Description Coding (MDC): It is a coding technique which splitthe video content into multiple streams known as descriptions. Basicquality content can be guaranteed with the successful decoding of eachdescription alone. The content can be further improved with the re-ception of every additional description. Unlike in other layered codingschemes, all streams in MDC have equal priority and does not dependon the successful decoding of another stream. The substream inde-pendence property of MDC has attracted much attention in research asany combinations of substream can be detected independently withoutany performance degradation. The principle idea of MDC is describedin [54].

All of the above layered video coding schemes can be applied in NOMA-basedlayered multicast video streaming. Nevertheless, the application of SVC inNOMA is mostly investigated in literature such as [60–63]. Meanwhile, theprinciples of MDC and NOMA were combined in [64] in order to furtherenhance the throughput and robustness for image transmission.


2.4.2 Challenges

Apart from the design issues in layered video coding, there are other chal-lenges in the implementation of layered multicast streaming which are de-scribed as follows:

1. Reception Reliability

One major problem in layered video streaming is that the video cannot beextracted when the base layer is not successfully decoded although the en-hancement layers are perfectly detected. In this case, the enhancement layerstreams must be discarded and therefore the radio resources assigned to theselayer streams are wasted. In order to address this issue, stringent QoS require-ments are implemented for the base layer stream in order to ensure successfulretrieval of basic quality video.

Another mitigation technique is the implementation of forward error cor-rection (FEC), which improves the resilience of the data stream at the expenseof reduced spectral efficiency. The inter-layer forward error FEC is a techniqueproposed in [65] that guarantees better reception quality. In this technique,the FEC information of the high-priority base layer is incorporated into thefirst enhancement layer data. Thus, if the first enhancement layer is success-fully decoded, the implanted FEC information intended for the base layer canbe utilized to enhance the error-resilience of the base layer. Similarly, the FECinformation of the first enhancement layer can be embedded into the secondone and so on. This technique has emerged as a promising technique for lay-ered video streaming and hence is studied intensively as highlighted in [58].

Nevertheless, investigating the technique for reliable and robust transmis-sion of the layered streams is beyond the scope of the work in this thesis.In this thesis, QoS-based target rate requirements are considered to ensurereliable and robust delivery of the video content while maximizing the per-formance through resource allocation.

2. Resource Allocation

Significant amount of literature exist on the resource allocation schemesfor OFDMA-based multicast systems as highlighted in [52]. In particular,the multi-rate or multi-layer technique has gained much interest in literature


due to its potential in achieving user rate differentiation and higher spec-tral efficiency [50, 52]. Meanwhile, the conventional multicast system is oftenassociated with fairness among all users at the expense of lower single trans-mission rate as it is restricted by the least channel gain user in the multicastgroup. However, for multi-rate or multi-layer scheme, additional orthogonalresources are required to deliver the multiple layer streams in an OFDMA-based network. This issue can be addressed by applying NOMA, which al-lows the utilization of whole bandwidth allocated to a single multicast groupby multiplexing all the layer streams in power domain. Therefore, NOMA canachieve superior performance in terms of resource utilization as compared toOFDMA as illustrated in Figure 2.9. Moreover, the resource allocation inOFDMA-based multi-layer system involves the assignment of each subcarrieror RB to the each layer stream which is not required in NOMA. Therefore,in multi-layer multicast network, the resource optimization for NOMA is lesscomplex compared to that of OFDMA.

NOMA OFDMA

Figure 2.9: Resource utilization in 2-layer multicast streaming in OFDMA andNOMA systems

Applying NOMA for video multicasting with layered coding, the weakestusers’ signal can be the base layer stream (high-priority basic-quality videodata) while stronger users are sent the enhancement layer streams (additionaldata to improve video quality). Instead of discarding the weaker user’s sig-nals, the stronger users will utilize them to enhance the overall sum rate. Mo-tivated by the benefits of exploiting the nature of NOMA and layered coding,the resource allocation for multi-layer multicast streaming in NOMA networkis investigated in Chapter 3. Optimizing the resource allocation in NOMA-based multi-layer multicast streaming is crucial to further enhance the system


performance, particularly the overall throughput. Nevertheless, the computa-tional complexity is another key factor in determining the effectiveness of theresource allocation scheme.

2.5 Delivery techniques for Wireless Caching at

User Equipment

Most researchers attempt to strengthen next generation mobile network withhigher data rates and capacity. However, this effort is constrained by the factthat radio spectrum resources are scarce. Another key technique is the densi-fication of the deployment of BSs. This will improve the spectral efficiency perunit area. Nevertheless, this technique coupled with the centralized nature ofmobile networks is associated with two main technical issues; interference,and traffic congestion in both core network and backhaul [15, 66]. Recently,the idea of caching has gained much interest as it potentially alleviates trafficcongestion in core network, backhaul and wireless links, particularly due todownloads of rich multimedia data (e.g. high-quality video streaming andvideo on demand) [15, 67]. The use of caching technique in mobile networksis often referred to as wireless edge caching, which is considered as one of thepotential key technology for the next generation mobile networks. In wirelessedge caching, caches are installed in key component of the network such asrouters, BSs, and UEs as illustrated in Figure 2.10. Generally, contents areallowed to be stored at these caches during low peak data traffic and thisprocess is known as content placement. During high peak data traffic, con-tent delivery process allows the users to retrieve the content from the nearestcaches, which improves network performance by easing the network trafficload and reducing the latency [14, 15, 66, 68].

Caching at the UEs are considered to be one of the most promising tech-nology for next generation cellular network as it ptentially relieves the trafficin both backhaul and wireless links [67]. However, the storage size of thecaches in the UEs is generally small compared to that of caches in the corenetworks or BSs [69]. Therefore, not all popular contents can be stored incache of each UE. Designing the content placement and delivery strategies


are very challenging problem for caching at the UEs. The most common con-tent delivery technique is the coded caching or coded multicasting [67,68,70],which is described in the next subsection.

Backhaul link Base Station level

o Alleviate traffic in backhaul link

o Reduce latency o Does not address

wireless traffic

User Equipment level

o Alleviate wireless traffic o Reduce latency o Energy consumption at

user end

Core Network

Internet Core Network level

o Alleviate traffic at CN-internet gateway

o Reduce latency o Does not address

backhaul traffic Content server

Figure 2.10: A general network architecture of CCMN

2.5.1 Coded Multicasting

In coded multicasting, each content in the remote server is partitioned intoa number of portions with equal size. Each portion is referred to as a sub-file. Depending on content popularity or request prediction of each user,only selected subfiles of each content are stored in users’ caches during con-tent placement phase. At the beginning of delivery phase (i.e., after contentrequests), the missing subfiles of the requested contents are encoded into asingle coded stream using XOR coding. This stream is delivered to multipleusers over shared link through multicasting instead of transmitting individualsubfiles over different time-frequency resources. Therefore, coded multicast-ing improves both spectral and energy efficiency. However, the transmissionrate of coded multicasting is restricted by the user with the weakest channelgain to ensure the subfiles are successfully decoded by all users.


Request: 𝐴

Request: 𝐵

Base Station

Content Server

power

UE 1

UE 2

Receiver at user 𝟏

Cache

𝑨𝟏 𝑨𝟐

𝑩𝟏 𝑩𝟐

Frequency/time

XOR Decoder

𝑨𝟏

𝑩𝟐

𝑨𝟐⨁𝑩𝟐 𝑨𝟐⨁𝑩𝟐

𝑩𝟐

𝑨𝟐

𝑨𝟐⨁𝑩𝟐

𝑨𝟏

𝑨𝟏

𝑨𝟐

Receiver at user 𝟐

Cache

XOR Decoder 𝑨𝟐⨁𝑩𝟐 𝑨𝟐⨁𝑩𝟐

𝑩𝟐

𝑨𝟐

𝑩𝟏

𝑨𝟐

𝑩𝟏

𝑩𝟏

𝑩𝟐

Figure 2.11: An example of coded multicasting in cache-enabled network fortwo-user scenario

A simple example of two-user coded multicasting system is illustrated inFigure 2.11. In this example, there are two content files and the set of files isdenoted as F = A,B. Each content is partitioned into two subfiles of equalsize and the set of subfiles is denoted as Fp = A1, A2, B1, B2. It is assumedthat the network applies off-line caching, that is, the content placement phasetakes place during off-peak times. In addition, the BS possesses the informa-tion on the request prediction of each user. According to this information, it ispredicted that user 1 and 2 will request content A and B respectively. Duringcontent placement phase, each user stored the first subfile of the content tobe requested (i.e., user 1 and 2 stored A1 and B1 respectively). In addition,the second subfile of the unwanted content is stored by each user (i.e., user 1

and 2 stored B2 and A2 respectively). This content placement configurationwill ensure coded multicasting is feasible. In coded multicasting, a user canonly decode the missing subfile of the requested content from the XOR codedstream (which consists of mixture of combined missing subfiles) if and onlyif the requesting user has the knowledge of all other unwanted subfiles inthe XOR coded stream. The problem is identified as index coding (IC) and


each problem can be described by the conflict graph Q which consists of thefollowing components

• Vertices of Q: The vertices corresponds to the requested subfiles. Eachvertex v has three fields in its label where ρ (v) is the subfile identity,µ (v) is the index of the user who is requesting for the subfile, and η (v)

is the index of the user who possess the subfile in the cache.

• Edges of Q: Two vertices v1 and v2 are connected by an edge if the twofollowing conditions are satisfied; (1) ρ (v1) 6= ρ (v2), and (2) µ (v1) /∈η (v2) & µ (v2) /∈ η (v1).

If two vertices are connected by an edge, the two subfiles represented bythese vertices must be transmitted over different orthogonal resources. Onthe other hand, if these vertices are not connected by an edge, the two sub-files can be encoded into a single packet by using XOR operation and hencetransmitted as a single coded packet. The conflict graph problem is difficult toapproximate in general and a polynomial-time approximation of this problemis proposed in [70] which is known as greedy constrained coloring (GCC). Inparticular, the computational complexity of IC in coded multicasting increaseswith larger number of users and subfiles [67].

2.5.2 CIC-based NOMA

NOMA is also considered as an alternative content delivery technique forcaching at the UEs [71, 72]. Similar to conventional NOMA, the users’ signalsare superposed in power domain using SC which are then transmitted by theBS. In conventional NOMA, SIC is performed at each user’s receiver by de-tecting the interfering signal first which is then used to cancel the interference.However, in cache-aided NOMA, the interfering signal is directly cancelled byexploiting the subfiles available in the cache without the need for the detec-tion of interfering signal. This means both strong and weak users are able tocancel interference provided that the interfering data or subfiles are availablein their caches. This technique, which is known as cache-enabled interferencecancellation (CIC), enhance the individual transmission rate particularly forthe weak user. However, the sharing of power among the users may affect thesum rate performance.


Request: 𝐴

Request: 𝐵

Base Station

Content Server

power

UE 1

UE 2

Receiver at user 𝟏

Cache

𝑨𝟏 𝑨𝟐

𝑩𝟏 𝑩𝟐

Frequency/time

Decoder

𝑨𝟏

𝑩𝟐

𝑨𝟐

𝑨𝟏

𝑨𝟏

𝑨𝟐

Receiver at user 𝟐

𝑨𝟐

𝑩𝟐

𝑨𝟐

𝑩𝟐 −

𝑩𝟐

𝑨𝟐

+ noise

+ noise

Cache

Decoder

𝑨𝟐

𝑩𝟐 −

𝑩𝟐

𝑨𝟐

+ noise

+ noise

𝑨𝟐 𝑩𝟏 𝑩𝟏

𝑩𝟐 𝑩𝟐

𝑩𝟏

Figure 2.12: An example of CIC-based NOMA in cache-enabled network fortwo-user scenario

An example of two-user CIC-based NOMA system is illustrated in Figure2.12. Similar content placement configuration as coded multicasting is em-ployed to ensure CIC-based NOMA is feasible. Note that, in Figure 2.12, bothusers can perform CIC to cancel the interference and detect the desired sig-nals from the residual received signal. Thus, CIC-based NOMA enhances theoverall sum rate since the individual transmission rate is not affected by in-terference. In conventional NOMA, SIC can only be performed by the stronguser as described in Section 2.3. The weak user in conventional NOMA de-tects the desired signal by treating the interference from strong user’s signalas noise. Therefore, the transmission rate of weak user is interference-limited.Practically, the conventional SIC method can be jointly implemented to cancelthe interference caused by uncached files.


2.6 Summary

In this chapter, relevant background theory required for this thesis is pre-sented. The wireless channel models, which consider both large-scale andsmall-scale propagation effects, are presented in this chapter. These modelsare applied throughout the simulation works of this thesis. This chapter alsodiscussed on the overview of NOMA, a state-of-the art technology which sig-nificantly improves spectral efficiency, increases cell-edge transmission rateand offers low latency. The performance gain achieved by NOMA over OMAis also discussed. Furthermore, this chapter explained how layered videostreaming operates in multicast networks. Two key challenges in the imple-mentation of layered video streaming are discussed in this chapter; receptionreliability and resource allocation. Finally, coded multicasting and CIC-basedNOMA are presented as the effective delivery techniques for wireless cachingat the UEs.

Chapter 3

Resource Allocation for LayeredMulticast Video Streaming inNOMA Systems

3.1 Introduction

The multicast features in mobile networks are likely to be enhanced in thefuture due to the increasing mobile data traffic which is highly dominated bymultimedia services (e.g. video downloading, live streaming, etc). The com-bination of NOMA and layered coding potentially enhance the performanceof multicast network. This chapter investigates the joint power allocationand subgrouping for layered multicast streaming in NOMA networks, whichis crucial to further enhance the system performance, particularly the sumrate. A review of existing literature on the application of NOMA in multicastnetworks is presented in Section 3.2. Section 3.3 introduces the system modeland formulates the optimization problem as an MINLP. Since MINLP requireshighly-intensive computational solution, the problem based on 2-layer casewith arbitrary subgrouping is first solved in Section 3.4. In this section, twosuboptimal power allocation schemes are proposed for the 2-layer case: aniterative subgradient technique and a closed-form solution. In Section 3.5,the closed-form solution is then generalized to a case with multiple layers inorder to derive a low-complexity power allocation technique which succes-sively allocates power for each layer. In addition, this section also describesthe low-complexity iterative subgrouping techniques which can incorporate

62

CHAPTER 3. RA FOR LAYERED MULTICAST STREAMING IN NOMA 63

the proposed multi-layer power allocation scheme to develop a solution whichjointly optimizes the power allocation and subgroup formation. Section 3.6provides the simulation results which demonstrate the effectiveness of theproposed power allocation and subgrouping schemes in enhancing the sumrate performance.

3.2 Related Works

NOMA has recently emerged as one of the promising multiple access schemeswhich can seamlessly integrate multicast network with other different net-works or technologies. In particular, the authors in [73] proposed the applica-tion of NOMA to combine multicast and wireless content caching networksin order to perform multicasting and content pushing simultaneously on thesame time-frequency resource. In [74], multiple information stream was de-livered to different multicast groups with the aid of sophisticated beamform-ing. Meanwhile, NOMA was applied to enhance the performance of content-centric multicast delivery in C-RANs [75]. In addition, the application ofNOMA in relay assisted multicast system for vehicle-to-everything (V2X) net-works was studied in [76]. Furthermore, there are a number of works whichconsider the integration of multicast and unicast transmission such as [77–80].

Nevertheless, the performance of a multicast network alone can be furtherenhanced by exploiting the nature of NOMA and layered coding. By apply-ing NOMA to multicast video streaming with layered coding, the base andenhancement layer data are multiplexed in power domain using SC. This al-lows the utilization of the whole allocated bandwidth, unlike in OMA whichrequires additional time-frequency resources for multiple layer streams. Inaddition, the allocation of RBs to each layer stream is not required in NOMAand therefore, the complexity of resource optimization is lower than that ofOMA. In NOMA-based multilayer multicast system, the high-priority baselayer stream is treated as the weakest users’ signal in order to allow all usersto retrieve the basic quality video data. Meanwhile, the stronger users per-form SIC in order to retrieve the enhancement layer streams which containadditional data that can improve video quality. Instead of discarding the theweaker users’ signal as in conventional NOMA, the stronger users utilize all


the detected layer streams to enhance the overall performance. For exam-ple, the performance in terms of signal-to-interference and noise ratio (SINR),coverage probability, average number of served users and sum rate was in-vestigated in [81] and [82] for both cooperative millimeter-wave and heteroge-neous networks, respectively. These studies demonstrate that NOMA-basedlayered multicast system offers significant performance improvement over thetraditional OMA network. Meanwhile, the works in [83] and [84] focused ondeveloping strategies to allow the weak users to receive the enhancement lay-ers through cooperative communications. However, resource allocation wasnot investigated in these works.

Motivated by the benefits of exploiting the nature of NOMA and layeredcoding, several works on resource allocation for NOMA-based layered mul-ticast system have emerged in literature. For instance, the optimization ofthe beamforming weights for MISO case was considered in [60, 61, 85, 86]with the objective of minimizing the transmission power whilst achievingthe minimum target rates. Meanwhile, power allocation schemes were pro-posed in [62, 87–89], but none of these works consider the joint optimizationof power allocation and subgrouping. Subgrouping is a promising techniquewhich potentially enhance the performance of NOMA-based layered multi-cast system by exploiting the heterogeneity of users’ channel quality [50].Nevertheless, subgrouping leads to combinatorial optimization which is acomplicated problem and different from the pairing or clustering problemin unicast NOMA system. Exhaustive search was employed in [63] and [90]to solve the optimal solution for subgrouping. However, exhaustive search isassociated with high complexity which hinders the practical implementation.Therefore, this thesis investigates the subgrouping methods which can offersignificant performance at low complexity.

In addition, this thesis also explores low-complexity power allocationschemes which can be incorporated into the subgrouping methods in order tojointly optimize the power allocation and subgrouping. In [90], a closed formsolution is derived for a single channel and two-layer case only. However,that work did not consider multiple layers and multiple RBs case, which isgenerally difficult to solve as more variables are optimized. In addition, thatwork only consider the minimum target rate constraint for the high-priority


base layer, but not the enhancement layer. On the other hand, the power al-location for multiple layers and multiple RBs case was considered in [63]. Inthat work, the proposed power allocation strategy is performed by allocat-ing the power for the lower layers to exactly meet the target rate while theremaining power is allocated to the uppermost layer. However, this strategydoes not necessarily maximize the sum rate since it was not derived usingoptimization approaches in [91]. This thesis considers the joint optimizationof power allocation and subgrouping with the aim of maximizing the sumrate while satisfying the proportional rate constraint, which is a more difficultoptimization problem.

3.3 System Model And Problem Formulation

In this chapter, a downlink NOMA-based multicast system is considered witha single-antenna BS located at the centre of the cell and K single-antennausers who request for the same video content1. The video data is deliveredby the BS to the users with a maximum transmit power budget of PT over atotal bandwidth of BT which is divided into N RBs of equal bandwidth B.It is assumed that each individual RB will experience frequency flat fadingsince the bandwidth B is smaller than the coherence bandwidth. However,frequency selective fading is experienced across all RBs. Channel impairmentsdue to path loss and shadowing are also considered in this work. Therefore,the channel gain between the BS and k th user at the n th RB is given by |hk ,n |2 =ξk|gk,n |2

PLkwhere ξk is the log-normal shadowing factor for k th user, |gk ,n |2 is the

effect of fading, and PLk is the path loss effect experienced by user k. Theusers are sorted according to ascending order of their channel gains2 and theset of all UEs at the n th RB is denoted as Mn = 1, 2, ..., K. These users aresegregated into a number of subgroups in such a manner that the sum rate isoptimized. The system model for the NOMA-based layered multicast video

1It is assumed that there are no other users requesting for different video contents inorder to focus on power allocation and subgrouping problems. Practically, the BS can servethe users requesting different video contents by allocating different subbands or RBs. Theoptimization problem involving the allocation of subbands or RBs to different group of userscan be treated independently and this problem has been studied extensively in the literatureon [52].

2In power allocation problem for NOMA, the users are commonly arranged according totheir channel gains for each RB in order to achieve the optimal SIC decoding order [44, 63].


streaming is illustrated in Figure 3.1.

Subgroup 𝟏𝟏

Subgroup 𝟐𝟐

UE 1

UE 2

UE 3

UE 4

UE k UE k+1

UE K

Subgroup 𝑺𝑺

. . . . . .

𝒉𝒉𝟏𝟏,𝒏𝒏𝟐𝟐 < 𝒉𝒉𝟐𝟐,𝒏𝒏

𝟐𝟐 < 𝒉𝒉𝟑𝟑,𝒏𝒏𝟐𝟐 < ⋯ < 𝒉𝒉𝒌𝒌,𝒏𝒏

𝟐𝟐 < 𝒉𝒉𝒌𝒌+𝟏𝟏,𝒏𝒏𝟐𝟐 < ⋯ < 𝒉𝒉𝑲𝑲,𝒏𝒏

𝟐𝟐

Figure 3.1: System model

At the BS, layered coding is employed to split the video content streaminto S layers. The first (base) layer X(1)

n consists of the most important data forbasic quality video, and therefore must be successfully decoded by all usersin order to retrieve the desired video content. Each subsequent enhancementlayer X(s)

n (for s ∈ 2, ..., S) improves the quality of the video and is onlyintended for the users in the corresponding and higher ranked subgroups. Inother words, the data in the sth layer is for the users in subgroups s to S. Letthe set of users and the total number of users in subgroup s at the nth RBbe G(s)

n and G(s)n respectively. By applying SC, all the layer streams X(s)

n aremultiplexed in power domain. The received signal at user k at RB n can beexpressed as

yk,n = hk,n

S∑s=1

√P

(s)n x(s)

n + wk,n (3.1)

where x(s)n is the transmitted symbols of the layer stream X

(s)n , P (s)

n is thepower allocated to layer stream s at RB n and wk,n is the AWGN. From (3.1),it is noted that the superposed signal, which consists of the base layer andall the remaining enhancement layer data streams, is received by all users.However, not all the users are guaranteed to successfully decode all the lay-ers due to their channel qualities. In particular, the users in subgroup 1 di-rectly detect the base layer by treating the remaining enhancement layers asnoise. This can be performed by allocating sufficiently higher power to the


base layer in comparison to the combined power of the enhancement layerstreams. However, the users in subgroup 1 will not be able to detect the en-hancement layer streams via SIC due to their poor channel conditions andlower power allocated to the enhancement layer streams. On the contrary,the users in other subgroups (for s ∈ 2, ..., S) will be able to decode thecorresponding enhancement layer streams from the received signal throughthe cancellation of the base layer and lower-level enhancement layer streams(for s ∈ 1, ..., s− 1). Unlike the conventional unicast NOMA, weaker users’streams (for s ∈ 1, ..., s− 1) are not discarded and instead, all the detectedsignal streams are combined together to enhance the quality of the video con-tent. Hence, the sum rate performance is improved due to the combinedtransmission rates of all the layered streams. In conventional layered videostreaming, the enhancement layers should be discarded when the base layer iscorrupted or lost during transmission, and this leads to the inefficient utiliza-tion of bandwidth and power [58]. The SIC decoding order in NOMA-basedmulticast system guarantees the successful decoding of the base layer firstbefore detecting the enhancement layers. Moreover, this decoding order com-plies with optimal SIC decoding scheme in conventional NOMA since thebase layer (weak user’s signal) is being decoded first prior to the detectionof enhancement layer (strong user’s signal) through SIC. Figure 3.2 presentsthe SIC process for the users in each subgroup in a 4-layer NOMA-basedmulticast system.

Assuming that the users in the sth subgroup successfully detect the lowerlayers (for s ∈ 1, ..., s− 1), the achievable rates for the sth and Sth layers atRB n can be expressed respectively as

R(s)n = B log2

(1 +

P(s)n H(s)

n

H(s)n

∑Si=s+1 P

(i)n +BN0

)(3.2)

R(S)n = B log2

(1 +

P(S)n H(S)

n

BN0

)(3.3)

where N0 is the noise power spectral density and H(s)n is the channel gain of

the weakest user in subgroup s due to multicasting, which is represented byH(s)n = min |hk ,n |2 in which k ∈ G(s)

n . First of all, this enable the weaker usersto obtain better standard quality video content. Secondly, in this multicast


system, a higher rate for the base layer will be beneficial for all users. Fur-thermore, if the base layer is not successfully decoded, the higher layer datawill be useless due to layered coding and may not be detected anyway due toerror propagation in SIC. Therefore, unlike in conventional unicast NOMA,the power allocation solution here must allocate higher power to the baselayer (i.e., weak user’s signal). The conventional minimum rate constraintwill ensure sufficient power for the weak user to attain the minimum rate,and offer as much power as possible for the strong user to maximize the sumrate. In this work, the proportional rate constraint is employed in order toproportionally allocate power to the users after the minimum rate has beenachieved [44].

Decoder

𝒚𝒌,𝒏 = 𝒉𝒌,𝒏 𝑷𝒏𝟏𝒙𝒏𝟏+ 𝒉𝒌,𝒏

𝒔=𝟐

𝟒

𝑷𝒏𝒔𝒙𝒏𝒔+ 𝒘𝒌,𝒏

𝒙𝒏𝟏

Decoder

𝒙𝒏𝟏

Decoder𝒙𝒏𝟐

𝒉𝒌,𝒏 𝑷𝒏𝟏𝒙𝒏𝟏

−

𝒌 ∈ 𝓖𝒏𝟏

𝒌 ∈ 𝓖𝒏𝟐

𝒉𝒌,𝒏 𝑷𝒏𝟐𝒙𝒏𝟐+ 𝒉𝒌,𝒏

𝒔=𝟑

𝟒


𝒙𝒏

Subgroup 2 receiver structure

𝒙𝒏𝟏+ 𝒙𝒏

𝟐

Decoder𝒙𝒏𝟑

𝒉𝒌,𝒏

𝒔=𝟏

𝟐

𝑷𝒏𝒔𝒙𝒏𝒔

−

𝒌 ∈ 𝓖𝒏𝟑

𝒉𝒌,𝒏 𝑷𝒏𝟑𝒙𝒏𝟑+ 𝒉𝒌,𝒏 𝑷𝒏

𝟒𝒙𝒏𝟒+ 𝒘𝒌,𝒏

𝒚𝒌,𝒏 = 𝒉𝒌,𝒏

𝒔=𝟏

𝟒


𝒙𝒏

Subgroup 3 receiver structure

𝒙𝒏𝟏+ 𝒙𝒏

𝟐+ 𝒙𝒏

𝟑

Decoder𝒙𝒏𝟒

𝒉𝒌,𝒏

𝒔=𝟏

𝟑

𝑷𝒏𝒔𝒙𝒏𝒔

−

𝒌 ∈ 𝓖𝒏𝟒

𝒉𝒌,𝒏 𝑷𝒏𝟒𝒙𝒏𝟒+ 𝒘𝒌,𝒏

𝒙𝒏

Basic quality video

Improved quality video

High quality video

Good quality video

Receiver Structure of Subgroup 1




Treated as noise

Treated as noise

Treated as noise


𝒔=𝟏

𝟒

𝑷𝒏𝒔𝒙𝒏𝒔+𝒘𝒌,𝒏


𝒔=𝟏

𝟒


Figure 3.2: SIC in multicast users’ receivers for 4-layer multicast system

Considering that all multicast users K successfully detect the base layerstreams and only the members of subgroup s ∈ 2, ..., S are able to detect


the desired enhancement layer data streams, the sum rate of the NOMA-basedlayered multicast system is given by

Rsum =N∑n=1

[(S−1∑s=1

K(s)n R(s)

n

)+K(S)

n R(S)n

](3.4)

where K(s)n is the total number of users who successfully detect layer stream

s at RB n, which directly relates to the total number of users in subgroup s bythe expression K(s)

n =∑S

i=sG(i)n . It is worth noting that subgrouping influence

the value of K(s)n which affects the considered channel gain H(s)

n in (3.2) and(3.3), and hence the transmission rate R(s)

n . Consequently, the sum rate can beoptimized by appropriately selecting the value of K(s)

n through subgroupingwhile allocating power to each layer streams.

It could be noted from (3.2) that the transmission rate is restricted by thechannel condition of the weakest user in each respective subgroup and theinterference caused by the higher-level enhancement layers. Specifically, therate of the first (base) layer is limited by the weakest channel condition amongall the users as well as interference from all the enhancement layers. However,in layered multicast system, it is vital to maintain a substantially high achiev-able rate for the base layer stream in order to guarantee robust transmissionof the mandatory basic video data to all users. Therefore, the optimizationhere is focused on maximizing the sum rate performance by designing a jointpower allocation and subgrouping scheme which takes into account both to-tal transmission power and proportional rate constraints. This optimizationproblem is formulated as

maximizeP

(s)n ,K

(s)n

Rsum (3.5a)

subject toN∑n=1

S∑s=1

P (s)n ≤ PT (3.5b)

P (s)n ≥ 0, ∀n, ∀s (3.5c)N∑n=1

R(1)n : . . . :

N∑n=1

R(S)n = Φ

(1)min : . . . : Φ

(S)min (3.5d)

K(s)n ∈ 1, 2, ..., K (3.5e)

K(1)n ≥ K(2)

n ≥ · · · ≥ K(S)n (3.5f)


where constraints (3.5b) and (3.5c) are the total power budget and non-negative powers constraints respectively. Constraint (3.5d) is the proportionalrate constraint that guarantees the minimum target rates for each layer streamΦ

(s)min while maintaining a proportionality among all the layer streams. It is

worth noting that once the minimum rate requirements are met, the power re-sources are distributed among the layer streams in proportional manner andtherefore, the obtained rates are not necessarily restricted to the target ratesratio [44]. The application of this constraint in NOMA ensures sufficientlyhigher power is allocated to the weaker users’ signals to allow the success-ful detection of their own signals while treating the stronger users’ signalsas noise. In addition, adequate power is assigned to weaker users’ signals inorder to allow the stronger users to perform SIC and cancel the weaker users’signals to detect their desired information. Constraint (3.5e) assures that K(s)

n

can only take integer values. Finally, constraint (3.5f) implies that not all userscan successfully decode the enhancement layer s (for s ∈ 2, ..., S) and alsoensures the expression K

(s)n =

∑Si=sG

(i)n is satisfied.

3.4 Power Allocation Schemes for Two-layer with

Arbitrary Subgrouping

The optimization problem in (3.5) is an MINLP problem which is generallydifficult to solve and associated with computationally-intensive numerical so-lution. Therefore, the problem in (3.5) is first simplified by considering onlytwo-layer streams and the users are arbitrarily grouped based on the order oftheir channel qualities3. According to (3.2), (3.3) and (3.4), the sum rate fortwo-layer case is represented by

Rsum = K(L)

N∑n=1

R(L)n +K(H)

N∑n=1

R(H)n (3.6)

3In arbitrary subgrouping, the users are first arranged according to descending order oftheir channel gains. The number of users in each subgroup G

(s)n is arbitrarily chosen. For

example, in a 2-layer case (S = 2) where the number of users for each subgroup is arbitrarilychosen as G

(1)n = 2 and G

(2)n = 3, the users ranked 4

th and 5th are selected for subgroup 1,

and those ranked 1st to 3

rd are in subgroup 2.


where K(L) and K(H) are the numbers of users who receive the base layerand enhancement layer respectively. Note that the superscripts (L) and (H)

replace the number-based layer/subgroup index to emphasize that the baselayer (equivalent to layer 1) and enhancement layer (equivalent to layer 2)streams are always being treated as weak and strong users’ signals respec-tively. Moreover, the users are arbitrarily grouped by selecting G(L) weakestusers to join subgroup L and the remaining G(H) strong users to be in sub-group H . It is considered that all the users will successfully detect the baselayer stream and hence K(L) will always be equal to the total number of users,that is, K(L) = G(L) +G(H) = K. Meanwhile, the users in subgroup H can de-tect the enhancement layer stream and thus K(H) = G(H). The achievable ratesfor the base layer and enhancement layer at RB n are expressed respectivelyas

R(L)n = B log2

(1 +

P(L)n H(L)

n

P(H)n H(L)

n +BN0

)(3.7)

R(H)n = B log2

(1 +

P(H)n H(H)

n

BN0

)(3.8)

where H(L)n is the channel gain of the weakest user in the multicast group

which is represented as H(L)n = min |hk ,n |2 in which k ∈ Kn and H(H)

n is theweakest channel gain in subgroup H by considering H(H)

n = min |hl ,n |2 wherel ∈ G(H )

n . Since the weakest channel gain is considered, the rate in (3.7) isachievable to all the users. Therefore, SIC can be successfully applied by thestrong users in subgroup H to achieve the rate for enhancement layer in (3.8).

The optimization problem is transformed into a non-linear programmingand is formulated as

maximizeP

(L)n ,P

(H)n

Rsum (3.9a)

subject toN∑n=1

P (L)n +

N∑n=1

P (H)n ≤ PT (3.9b)

P (L)n ≥ 0, P (H)

n ≥ 0, ∀n (3.9c)N∑n=1

R(L)n :

N∑n=1

R(H)n = Φ

(L)min : Φ

(H)min. (3.9d)


Note that the constraints (3.5b) - (3.5d) in the original optimization prob-lem are preserved as these constraints are directly related to power allocation,whereas, the subgrouping-related constraints (3.5e) and (3.5f) are withdrawnas arbitrary subgrouping is considered.

In unicast NOMA system, the objective function for weighted sum ratemaximization (WSRM) problem is concave under some conditions [92]. Notethat, in layered multicast system, the number of users receiving each layerstream K

(s)n is equivalent to the weights in unicast system with WSRM prob-

lem. Since the proof in [92] is based on general multiple users case, it isdifficult to calculate the Hessian matrix which must be negative semidefinitein order to prove the concavity of the objective function. Therefore, the proofin [92] is simplified by dividing the objective function into multiple subfunc-tions. The concavity of each subfunction with respect to single variable isproven instead of the whole function. Meanwhile, in this thesis, the wholeobjective function is considered albeit for two-layer case only, which is suf-ficient to demonstrate that it is possible to obtain optimal or near-optimalsolution for the two-layer case investigated in this section. On the contrary,the solution for general multilayer case proposed in Section 3.5 is subopti-mal and hence the proof for this case is not required. In the following, theconcavity of the objective function is proved specifically for the two-layer case.

Proposition 3.1: The objective function (3.6) for the two-layer case is concavewithout any conditions.

Proof: First, the concavity of objective function (3.6) is assessed for each RBand hence, the objective function for each RB can be expressed as

fn(P (L)n , P (H)

n

)=K(L)B log2

((P (L)n + P (H)

n

)Λ(L)n + 1

)+K(H)B log2

(P (H)n Λ(H)

n + 1)−K(L)B log2

(P (H)n Λ(L)

n + 1)

(3.10)where Λ

(L)n = H(L)

n

BN0and Λ

(H)n = H(H )

n

BN0.

The Jacobian of fn(P

(L)n , P

(H)n

)is calculated as

∇fn(P (L)n , P (H)

n

)=

B

ln 2

(K(L)Λ

(L)n(

P(L)n +P

(H)n

)Λ

(L)n +1

)B

ln 2

(K(L)Λ

(L)n(

P(L)n +P

(H)n

)Λ

(L)n +1

+ K(H)Λ(H)n

P(H)n Λ

(H)n +1

− K(L)Λ(L)n

P(H)n Λ

(L)n +1

) .(3.11)


The Hessian of fn(P

(L)n , P

(H)n

)is calculated as

∇2fn(P (L)n , P (H)

n

)=

−A −A

−A B −A

(3.12)

where

A = Bln 2

(K(L)

(Λ

(L)n

)2

((P

(L)n +P

(H)n

)Λ

(L)n +1

)2

)(3.13)

B = Bln 2

(−

K(H)(

Λ(H)n

)2

(P

(H)n Λ

(H)n +1

)2 +K(L)

(Λ

(L)n

)2

(P

(H)n Λ

(L)n +1

)2

). (3.14)

In order to prove that fn

(P

(L)n , P

(H)n

)is concave, the Hessian of

fn

(P

(L)n , P

(H)n

)must be negative semidefinite [91], that is

[P

(L)n P

(H)n

]∇2fn

(P (L)n , P (H)

n

) [P

(L)n P

(H)n

]T≤ 0 (3.15)

The condition (3.15) can be represented as

−K(H)

(P

(H)n Λ

(H)n

)2

(P

(H)n Λ

(H)n +1

)2

−K(L)

((P

(L)n Λ

(L)n

)(P

(H)n Λ

(L)n +2P

(H)n

(P

(L)n +P

(H)n

)(Λ

(L)n

)2+(P

(L)n +P

(H)n

)Λ

(L)n

)(P

(H)n Λ

(L)n +1

)2((P

(L)n +P

(H)n

)Λ

(L)n +1

)2

)≤ 0

(3.16)

Since all the variables K(L), K(H), Λ(L)n , Λ

(H)n , P (L)

n , and P(H)n are positive

values, the condition in (3.16) always holds. This implies that fn(P

(L)n , P

(H)n

)is concave. Note that the objective function (3.6) is the summation offn

(P

(L)n , P

(H)n

), i.e.

∑Nn=1 fn

(P

(L)n , P

(H)n

). Since the concavity is preserved

for nonnegative weighted sum of concave functions [91], the objective func-tion (3.6) is concave.


3.4.1 Subgradient Method

Although the objective function (3.6) is proven to be concave, a high-complexity numerical tool is needed to solve the optimal power allocation[91, 93]. Therefore, a near-optimal iterative subgradient method is proposedto obtain the power allocated to the base and enhancement layer in all RBswith lower computational complexity. The Lagrangian dual decomposition(LDD) technique can be used to solve an optimization problem with concaveobjective function and proportional rate constraint [93, 94]. Based on LDDapproach in [91], the Lagrangian function of the optimization problem (3.9),is represented as

L(P (L)n , P (H)

n , µ, τ) = K(L)

N∑n=1

B log2

(1 + Γ(L)

n

)+K(H)

N∑n=1

B log2

(1 + Γ(H)

n

)−µ

(N∑n=1

P (L)n +

N∑n=1

P (H)n − PT

)− τ

(∑Nn=1B log2

(1+Γ

(L)n

)Φ

(L)min

−∑Nn=1B log2

(1+Γ

(H)n

)Φ

(H)min

)(3.17)

where µ and τ are the Lagrange multipliers, and Γ(L)n and Γ

(H)n are respectively

the SINR expressions of (3.7) and (3.8) which are given by Γ(L)n = P

(L)n H(L)

n

P(H )n H(L)

n +BN0

and Γ(H)n = P

(H)n H(H)

n

BN0.

To solve the optimization problem (3.9), the Karush-Kuhn-Tucker (KKT)conditions are obtained as follows

dL(P(L)n , P

(H)n , µ, τ)

dP(L)n

=N∑n=1

(K(L) − τ

Φ(L)min

)B

ln 2

Γ(L)n

P(L)n

(1 + Γ

(L)n

)− µ = 0

(3.18)

dL(P(L)n , P

(H)n , µ, τ)

dP(H)n

=N∑n=1

(K(L) − τ

Φ(L)min

)B

ln 2

−(

Γ(L)n

)2

P(L)n

(1 + Γ

(L)n

)

+

(K(H) +

τ

Φ(H)min

)B

ln 2

Γ(H)n

P(H)n

(1 + Γ

(H)n

)− µ = 0

(3.19)µ(∑N

n=1 P(L)n +

∑Nn=1 P

(H)n − PT

)= 0 (3.20)


τ

∑Nn=1B log2

(1 + Γ

(L)n

)Φ

(L)min

−

∑Nn=1 B log2

(1 + Γ

(H)n

)Φ

(H)min

= 0. (3.21)

Solving the KKT conditions (3.18) and (3.19), the power allocated to thebase layer and enhancement layer streams are respectively expressed in termsof µ and τ as

P(L)n =

[(K(L)Φ

(L)min − τ

)( 1

µΦ(L)min ln 2

−BN0Φ

(H)min

(H(H)n −H(L)

n

)(

Φ(L)minΦ

(H)min (K(L) −K(H))− τ

(Φ

(L)min + Φ

(H)min

))H(L)n H(H)

n

+ (3.22)

P(H)n =

BN0

(K(H)Φ

(H)min + τ

)Φ

(L)min(

Φ(L)minΦ


(Φ

(L)min + Φ

(H)min

))H(L)n

−BN0

(K(L)Φ

(L)min − τ

)Φ

(H)min(

Φ(L)minΦ


(Φ

(L)min + Φ

(H)min

))H(H)n

+ (3.23)

where [x]+ = max (x, 0) which ensures constraint (3.9c) is satisfied. Then, theproblem (3.9) is transformed into a dual problem which is formulated as

maximizeµ,τ

D(µ, τ) = infP

(L)n ,P

(H)n

L(P (L)n , P (H)

n , µ, τ) (3.24a)

subject to µ ≥ 0. (3.24b)

Finally, the Lagrange multipliers (dual variables) µ and τ are solved byusing the subgradient algorithm as summarized in Algorithm 3.1. In the firstiteration (t = 0), the power allocated to both the base and enhancement layersis calculated based on defined initial values of the Lagrange multipliers µ(0)

and τ(0). In each iteration t, the dual variables are updated in steps 4 and5 respectively by taking into account the dual variables in the previous itera-tion, the positive step sizes for µ and τ (denoted as δµ and δτ respectively), andthe valid subgradients. For a convex and differentiable function, its gradientis the subgradient [95]. In this case, a subgradient/gradient is obtained for


the negative dual function −L(P(L)n , P

(H)n , µ, τ) since the objective is to maxi-

mize L(P(L)n , P

(H)n , µ, τ), which is equivalent to minimize −L(P

(L)n , P

(H)n , µ, τ).

Therefore, the valid subgradients are determined from the derivatives of thenegative dual function −L(P

(L)n , P

(H)n , µ, τ) with respect to µ and τ which are

respectively given as

∇µ =N∑n=1

P (L)n +

N∑n=1

P (H)n − PT (3.25)

∇τ =

∑Nn=1B log2

(1 + Γ

(L)n

)Φ

(L)min

−

∑Nn=1B log2

(1 + Γ

(H)n

)Φ

(H)min

. (3.26)

Algorithm 3.1 Subgradient Algorithm1: Initialization: set t = 0 and ε, initialize µ(0) and τ(0)

2: while (|µ(t− 1)− µ(t)| ≥ ε or |τ(t− 1)− τ(t)| ≥ ε) and t ≤ Tmax do3: solve P (L)

n and P(H)n using (3.22) and (3.23) respectively

4: update µ(t+ 1) = [µ(t) + δµ∇µ]+

5: update τ(t+ 1) = τ(t) + δτ∇τ6: t← t+ 1

7: end while8: output the optimal solutions P (L)∗

n and P(H)∗n

Note that the initial values and the step sizes influence the convergencetowards optimal solution. In this work, diminishing step size δ = a/

√t or

δ = a/(b+ t

)is employed to guarantee near-optimal solution [95] where

a and b are fixed non-negative values. The iteration stops when the dualvariables µ and τ converge, that is, when the difference in values betweenthe dual variables in the current iteration and that in previous iteration issufficiently small and achieve a tolerance value of ε. Otherwise, the iterationterminates when the iteration index t achieves the maximum iterations Tmax.

3.4.2 Multicast-based Equal RB Power Allocation (M-ERPA)

Although the implementation of subgradient algorithm is simple, the conver-gence towards the optimal solution may be slow and therefore may incur asignificant amount of iterations. In order to further reduce the complexity


of power allocation, a closed-form solution is proposed by assuming that thepower is equally allocated to each RB [44] as well as by transforming the KKTcondition related to the proportional rate constraint to another form. In thiswork, the total power for each RB is calculated by dividing the maximumpower budget PT by the total number of RBs N and thus, is shared betweenthe base layer and enhancement layer streams in each RB as follows

P (L)n + P (H)

n =PTN

= PRB (3.27)

where PRB is the power allocated to each RB.This closed-form solution, which is referred to as Multicast-based Equal

RB Power Allocation (M-ERPA), is derived from the Lagrangian function andKKT conditions determined in subsection 3.4.1. Solving (3.18) and (3.19) forP

(H)n eliminates the Lagrange variable µ as follows

P (H)n =

BN0

(K(H)Φ

(H)min + τ

)Φ

(L)min(

Φ(L)minΦ


(Φ

(L)min + Φ

(H)min

))H(L)n

−BN0

(K(L)Φ

(L)min − τ

)Φ

(H)min(

Φ(L)minΦ


(Φ

(L)min + Φ

(H)min

))H(H)n

.

(3.28)

A high complexity numerical tool is still required to solve P (L)n using (3.20)

and (3.28). In order to simplify this problem, the equal RB power allocationassumption in (3.27) is applied. Therefore, P (L)

n is solved by allocating theremaining power in each RB to the base layer stream. Using (3.28), P (L)

n canbe expressed as

P (L)n =PRB − P (H)

n

=PRB−

BN0

(K(H)Φ

(H)min + τ

)Φ

(L)min(

Φ(L)minΦ


(Φ

(L)min + Φ

(H)min

))H(L)n

BN0

(K(L)Φ

(L)min − τ

)Φ

(H)min(

Φ(L)minΦ


(Φ

(L)min + Φ

(H)min

))H(H)n

(3.29)


Note that (3.28) and (3.29) still contain the Lagrange variable τ . This vari-able can be removed by substituting (3.28) and (3.29) into (3.21). However,according to [93], the solution can only be obtained by using a high complex-ity iterative methods, such as Newton-Raphson method. In order to derivea closed form solution, (3.21) is transformed into a different form which isexpressed as

τ

(N∑n=1

BΦ

(H)min

Φ(L)min

+N∑n=1

B log2

(1 + Γ(L)

n

)−

N∑n=1

B log2

(1 + Γ(H)

n

))= 0 (3.30)

Note that, although (3.30) is not similar to (3.21), the expression (3.30) stilloffers proportionality feature by adding the ratio between the target rate ofenhancement layer and that of base layer. By substituting (3.28) and (3.29)into (3.30), the Lagrange multiplier τ is solved as

τ =Φ

(L)minΦ

(H)min

(2(K(L) −K(H)

)ψ1 + ψ2 +BN0ψ3

)2ψ4

(3.31)

where

η =2

(Φ

(H)min/Φ

(L)min

)

ψ1 =ηPRB

(H(L)n

)2

H(H)n

(Φ

(L)min + Φ

(H)min

)ψ2 =

√BN0

(H(L)n −H(H)

n

)(K(L)Φ

(L)min +K(H)Φ

(H)min

)×√

4ψ1(Φ

(L)min+Φ

(H)min

) +BN0

((H(L)n −H(H)

n

)2

+ 4ηH(L)n H(H)

n

)ψ3 =

(K(L)Φ

(L)min −K

(H)Φ(H)min

)(H(L)n −H(H)

n

)2

+ 2ηH(L)n H(H)

n

(Φ

(L)min + Φ

(H)min

)(K(L) −K(H)

)ψ4 =ψ1

(Φ

(L)min + Φ

(H)min

)+BN0

(Φ

(L)minΦ

(H)min

(H(L)n −H(H)

n

)2

+ ηH(L)n H(H)

n

(Φ

(L)min + Φ

(H)min

)2).

Finally, the suboptimal power for the base and enhancement layer streamscan be solve by substituting (3.31) into (3.29) and (3.28) respectively to give

P(L)n = PRB+

BN0

((H(L)n Φ

(H)min+H(H)

n Φ(L)min

)(2(K(L)−K(H))ψ1+ψ2+BN0ψ3)+2

(K(H)H(H)

n −K(L)H(L)n

)ψ4

)H(L)n H

(H)n

((Φ

(L)min+Φ

(H)min

)(2(K(L)−K(H))ψ1+ψ2+BN0ψ3)+2(K(H)−K(L))ψ4

)(3.32)


P (H)n = −

BN0

((H(L)n Φ

(H)min+H(H)

n Φ(L)min

)(2(K(L)−K(H))ψ1+ψ2+BN0ψ3)+2

(K(H)H(H)

n −K(L)H(L)n

)ψ4

)H(L)n H

(H)n

((Φ

(L)min+Φ

(H)min

)(2(K(L)−K(H))ψ1+ψ2+BN0ψ3)+2(K(H)−K(L))ψ4

) .

(3.33)

It is worth noting that although all RBs have equal total power, the powerallocation ratio P

(L)n /P

(H)n vary over all RBs depending on the considered

channel gains H(L)n and H(H)

n as observed in (3.32) and (3.33).

3.5 Power Allocation And Subgrouping Formation

Schemes for Multi-layer

The sum rate performance of the NOMA-based multi-layer multicast systemcan be further improved by optimizing the subgrouping and having multiple(more than two) layer streams in the same time-frequency resource. By ex-ploiting the heterogeneity of users’ channel conditions, user transmission ratedifferentiation will be achieved, which leads to enhanced overall throughput.However, the joint power allocation and subgroup formation of the multi-layer multicast scheme in (3.5) is a difficult problem and requires a solutionwith excessive computational complexities. Therefore, in this work, the jointproblem is divided into two subproblems. The first subproblem focuses onsolving the power allocation for the multi-layer case, which is dealt in the nextsubsection. Meanwhile, the optimization of subgrouping problem is treatedas the second subproblem and will be discussed in subsection 3.5.2.

3.5.1 Successive Layer-based Power Allocation (SLPA)

In order to solve the power allocation problem for multi-layer case at a lowcomplexity, the 2-layer based M-ERPA solution derived in subsection 3.4.2 ismodified. The solutions in (3.32) and (3.33) directly allocate the power to layer1 (base) and layer 2 (enhancement) for the 2-layer (S = 2) case. For the multi-layer case, the power for each layer is successively allocated starting with thelowest layer (layer 1) and therefore the scheme is known as the successivelayer-based power allocation (SLPA). The SLPA scheme for the 4-layer (S = 4)case is shown in Figure 3.3.


𝑠 = 1

𝑠 = 2

𝑠 = 3

𝑠 = 4

𝑃𝑛1

𝑃𝑛𝐺𝐴

Stage 1

𝑠 = 2

𝑠 = 3

𝑠 = 4

𝑃𝑛2

𝑃𝑛𝐺𝐵

Stage 2

𝑃𝑛3

𝑃𝑛4

Stage 3

𝑠 = 3

𝑠 = 4

Layer L

Layer H

Layer L

Layer H

Layer L

Layer H

Figure 3.3: The power allocation stages in SLPA for 4-layer multicast system

In the first stage, layer 1 is treated as layer L and the remaining layers (i.e.,layer 2 until layer S) are grouped together to form a single imaginary upperlayer H . Considering the power budget of PRB, the solutions in (3.32) and(3.33) are then applied to calculate the power allocated to layer 1 (base layer)P

(1)n and the imaginary upper layer PGA

n . The latter is to be shared among allthe upper layers (i.e., layer 2 until layer S)4. As such, the channel gains of thebase layer and the upper layers should not be used directly as H(L)

n and H(H)n .

This is because if the imaginary upper layer has a higher channel gain, it willbe allocated with less power, which will be insufficient to be shared among allthese layers. In order to compensate for this problem, the considered channelgains for layer L (base layer) and layer H (group of enhancement layers) arerespectively calculated as

H(L)n = min

q∈Ω(L)n

H(q)n ×W−1 (3.34)

H(H)n = min

r∈Ω(H)n

H(r)n ×W (3.35)

where Ω(L)n and Ω

(H)n are the set of layer(s) which form the lower layer and

imaginary upper layer respectively, and W is a weight which is defined as

W =

∑r∈Ω

(H)nH(r)n∑S

s=1H(s)n

. The weight W is a fraction of the combined channel gain in

4It is noted that, from (3.2), the interference power is obtained from the sum of powerallocated to all the remaining upper layers. Hence, PGAn also corresponds to the interferencetowards the base layer. Based on the allocated power P (1)

n and interference PGAn , the rate ofthe base layer is guaranteed in the first stage.


the imaginary upper layer to the total channel gain of all layers. Since W isalways less than 1, it effectively reduces the considered channel gain for theimaginary upper layer and increases that for the base layer. This means thebase layer will give up some of its power in order to compensate for the powerto be shared among the enhancement layers. At each stage, the variables K(L)

and K(H) in (3.32) and (3.33) are given by K(L) = maxq∈Ω

(L)nK(q) and K(H) =

maxr∈Ω

(H)nK(r) respectively. It is worth noting that SLPA starts allocating the

power for the lowest layer in order to guarantee substantial rate for the high-priority base layer. This is performed by determining both allocated powerP

(1)n and the interference to the base layer PGA

n in the first stage. Meanwhile,the remaining power PGA

n is shared among the upper layers and hence doesnot guarantee higher rate. Nevertheless, obtaining higher rates for the upperlayers will be useless if the base layer is in outage. Therefore, this techniqueguarantee robust delivery of basic quality video data to all users by offeringhigh transmission rate for the base layer stream. This is crucial in multilayervideo coding as a higher rate for the base layer can ensure all users have astandard-quality video, while the enhancement layers enhance the quality ofthe video for the users with better channel conditions.

In the second stage, layer 2 is being treated as the lower layer L while theremaining upper enhancement layers (i.e., layer 3 until layer S) are groupedtogether as the imaginary upper layer H . The power allocated to layer 2

P(2)n and to the newly-formed imaginary upper layer PGB

n is determined byapplying (3.32) and (3.33) respectively using the power budget allocated in theprevious stage (i.e., PRB = PGA

n ). Here, the considered channel gains are alsoobtained by using (3.34) and (3.35) in order to ensure sufficient power to beshared among the upper enhancement layers. The same process is repeatedin the subsequent stages until there are only two layers left for considerationin which (3.32) and (3.33) are directly applied to solve the power for the twouppermost layers (P (3)

n and P(4)n for the 4-layer case in Figure 3.3). Note that,

the power budget allocated in the preceding stage must be considered insolving the power in each stage.

In the next subsection, low complexity iterative subgroup formationschemes are proposed, which are then incorporate the proposed SLPA schemeto develop schemes that jointly optimize the power allocation and subgroupformation.


3.5.2 Subgroup Formation

It is defined in (3.4) that the number of users receiving layer s at RB n, K(s)n , is

an integer in which the possible values can be represented as a set K(s)n . The

layer transmission rate R(s)n seems to grow linearly withK(s)

n according to (3.4).However, choosing a higher value for K(s)

n does not necessarily optimize thesum rate performance as increasing K

(s)n implies that a user with very poor

channel condition will be included into the same subgroup. The total rate forevery user in the same subgroup is thus reduced, which could lower the sumrate. In other words, the optimization of K(s)

n also depends on the weakestuser in the subgroup which affect the transmission rate of layer stream s andhence the sum rate.

In this work, it is assumed that all users successfully obtain the video con-tent by receiving the base layer data and thus, the number of users receivinglayer 1 (K(1)

n ) always equals to the total number of users (K). Consequently,the number of users receiving the remaining layers s (for s ∈ 2, ..., S) mustbe determined to maximize the sum rate. Consider an example of a 3-layerNOMA-based multicast system with 5 users, in which the number of usersreceiving layer 1 is 5 and the possible number of users receiving layer 2 isrepresented by the set K(2)

n ∈ 1, 2, 3, 4. If user 2 and 3 are definitely in sub-group 2, then K

(2)n will be 4 as user 4 and 5 (who has better channel gains)

will also be able to receive this layer. In this case, only user 4 and 5 could bein subgroup 3 and hence the set of possible number of users in layer 3 will beK(3)n ∈ 1, 2. From this example, it is noted that the search space of the setK(s)n is finite and therefore, it is possible to optimize the subgrouping through

an exhaustive search. However, this optimal search algorithm is computation-ally intensive especially when the number of users (K), the number of RBs(N ) and the number of layer streams (S) are large. Therefore, suboptimal sub-grouping schemes are proposed in this subsection with the aim of reducingthe computational complexity.

The general idea of subgrouping Method 1, which is summarized in Algo-rithm 3.2, is to examine whether removing or adding users into an arbitrarily-formed subgroup will increase the sum rate. This algorithm starts at sub-group 2 since subgroup 1 receives the base layer stream intended for all users,i.e., K(1)

n = K. Without loss of generality, consider the subgrouping for the


sth subgroup, i.e., subgroup 2 to s − 1 has already been assigned. First, sub-group s is arbitrarily formed by setting a channel-to-noise ratio (CNR) limit,Qs. All users with a CNR between Qs−1 and Qs will be admitted into this sth

subgroup, i.e., G(s)n =

k : 1 ≤ k ≤ K, Qs−1 < |hk,n |

2/BN0 < Qs

. Let the num-

ber of users receiving layer s in this arbitrary subgroup be K(s)n , and let it be

the jth element of the set K(s)n . Then, the algorithm computes the sum rate

according to this arbitrarily-formed subgroup, denoted as Rn (j). The sumrates for removing and adding a user in this subgroup (i.e., the sum ratesRn (j − 1) and Rn (j + 1) according to the (j − 1)th and (j + 1)th elements inK(s)n respectively) are also calculated. Based on these calculated sum rates, the

trend of the objective function can be approximated in order to decide on thesearch direction. If the trend is increasing (i.e., adding a user increases thesum rate), the algorithm will keep adding more users to the subgroup untilthere is no more improvement in sum rate (i.e., a local/global maximum ismet) or the maximum value of K(s)

n is reached (i.e., the last element in K(s)n ).

On the contrary, the algorithm will keep removing users from the subgroup ifa decreasing trend (i.e., eliminating a user increases the sum rate) is observedat the start of the search. If the sum rates Rn (j − 1) and Rn (j + 1) are smallerthan Rn (j) (i.e., a concave trend), the search is immediately stopped and thecurrent arbitrarily-formed subgroup (i.e. jth element) is considered to be thesolution. Note that, in this algorithm, the sum rates are calculated by apply-ing the proposed power allocation scheme and therefore, the power allocationis sub-optimized in all sum rate evaluations.

Although the computational complexity is reduced, the solution forMethod 1 may easily be trapped in a local maximum. Therefore, this methodis improved by allowing the algorithm to search for the possibility of anotherlocal maximum. Upon reaching the first local maximum (represented as thejth element of K(s)

n ), the sum rate of the(j − 2

)th or

(j + 2

)th element is as-

sessed and if found to be larger than that of(j − 1

)th or

(j + 1

)th one, the

algorithm resumes the iteration in search of the second local maximum. Thebest solution among the two local optima is then selected. This Method 2

outperforms Method 1 as it offers a higher possibility in finding the globaloptimum, but at the expense of higher complexity.


Algorithm 3.2 Subgroup Formation Method 1

1: Initialize: the set K(s)n and K

(s)n which is the jth element of the set K(s)

n .

2: for s = 2 to S3: Determine the sum rates, Rn = Rn (j − 1) , Rn (j) , Rn (j + 1)4: if arg maxRn = j + 1

5: t = j + 1

6: while Rn(t) > Rn(t− 1) & t < K(s−1)∗n

7: t = t+ 1

8: Determine Rn(t) and update Rn9: endwhile10: update new K

(s)∗n = arg maxRn

11: elseif arg maxRn = j − 1

12: t = j − 1

13: while Rn(t) > Rn(t+ 1) & t > 1

14: t = t− 1

15: Determine Rn(t) and update Rn16: endwhile17: update new K

(s)∗n = arg maxRn

18: elseif arg maxRn = j

19: K(s)∗n = K

(s)n

20: end if21: end for22: output: the suboptimal solutions K(s)∗

n & P(s)∗n

Finally, Method 3 involves the reduction of the search space by eliminatingsome elements in K(s)

n . In a multicast system, the total sum rate will be re-duced when an additional user with a much lower channel gain is added intothe subgroup. On the other hand, if the channel gain of the additional user issimilar to others, including it in the subgroup will almost certainly increasethe sum rate. Therefore, if the channel gain for some users are the same, thereis no need to calculate the sum rate for each additional user in the groupingassignment. Hence, to reduce the computational complexity, those users withsimilar channel gains can be considered together (i.e., they will be added orexcluded together into a layer). Consider an example of 2-layer NOMA multi-cast system with 5 users. If the first two strongest users (i.e. user index 1 and2 in K(2)

n ) have similar channel quality, they can be grouped together and thusK

(2)n = 1 can be removed from K(2)

n . These two users will either be selected


together for a layer or none will be selected at all. Based on the new set K(s)n ,

the solution is obtained by using Method 1. Note that Method 3 potentiallyeliminates local optimum and therefore facilitate the search for the global op-timal. However, there is also a risk of losing the global optimal due to theremoval of similar elements from K(s)

n .

3.5.3 Complexity of Subgroup Formation Algorithm

For exhaustive search, the number of power allocation and sum rate com-putations required is N K!

(K−(S−1))![63]. Meanwhile, Method 1 and 3 require

N (S − 1) (3 + I) computations where I is the total number of iterations. Theadditional three computations are accounted for the evaluation of the trendsof the objective function at the start of the algorithm. Method 2 requiresN (S − 1) (3 + I1 + 1 + I2) computations where I1 and I2 are the total numberof iterations required for the search of the first and possibly the second localmaxima respectively. Here, one more computation is needed to examine theobjective function’s trend after the first search. The three proposed subgroup-ing methods only need small number of iterations I and hence offer muchlower complexity as demonstrated in Section 3.6.

3.6 Simulation Results

The simulation in this section consider a downlink NOMA-based multicastsystem that consists of K users uniformly distributed within a circular cellradius of 500 m with a BS located at the centre. Table 3.1 depicts the para-maters associated with the effects of path loss, shadowing effect, noise andfrequency selective fading considered throughout all the simulations in thischapter, unless state otherwise. Here, Matlab’s numerical optimization toolis used to solve the optimal power allocation scheme, which is compared tothe proposed schemes and other existing low-complexity techniques. For FPAscheme, the power allocation ratio of 0.8 : 0.2 and 0.8 : 0.16 : 0.032 : 0.008 areemployed for 2-layer and 4-layer cases respectively. Note that, for the 4-layerFPA scheme, the sum of the power allocation ratio to the three uppermostlayers is 0.2, which is equal to that of layer 2 in the 2-layer case. In otherwords, the power allocation ratio between layer 1 and the combined upper


three layers is 0.8 : 0.2. The power for each layer is successively obtained byusing this 0.8 : 0.2 ratio in a manner similar to that of the proposed SLPAscheme. The main difference is that the power allocation ratio is always fixedregardless of the channel conditions of the users and hence does not necessar-ily meet the target rate. Due to the nature of this FPA scheme, it potentiallyoffer comparable performance to SLPA.

Table 3.1: Simulation Parameters

Paramaters Values/Model

No. of RB, N 25

Total Bandwidth, BT 5 MHz

Cell radius 500 m

Minimum distance from BS 10 m

Carrier frequency, fc 2 GHz

Path loss, PL 38.46+10γ log10 (d)

Path loss exponent, γ 3

Shadowing standard deviation 8dB

Noise power spectral density, N0 -174 dBm/Hz

Frequency Selective Fading ITU Pedestrian B [27]

Φ(L)min 0.5 Mbps

Φ(H)min 1 Mbps

3.6.1 Two-Layer Case with Arbitrary Subgrouping

The simulation of K = 5 multicast users is considered in this subsectionby arbitrarily grouping the users according to their channel qualities. Inthe arbitrary subgrouping, subgroup 1 (k ∈G(1 )

n ) consists of the two weak-est users while the remaining stronger users are in subgroup 2 (k ∈G(2 )

n ) toensure that more than 50% of the users will acquire excellent quality of ser-vice. The minimum target rate for base and enhancement layer streams areset as Φ

(L)min = 0.5 Mbps and Φ

(H)min = 1.0 Mbps respectively. This will ensure

that the users in subgroup 2 will achieve a combined target rate of at least


1.5 Mbps. In Figure 3.4, the sum rate performance of the the proposed NOMApower allocation schemes (subgradient method and M-ERPA) are comparedto optimal NOMA, OFDMA and other NOMA power allocation schemes in-cluding FPA, FTPA [18] and Intra-Group Scalable Multicast Scheduling PowerAllocation (IGSMS-PA) [63]. The result in Figure 3.4 shows that the proposedsubgradient method and M-ERPA performs better than other schemes (FPA,FTPA and IGSMS-PA). This is because both proposed schemes are derivedbased on sum rate maximization problem. While the target rates requirementare met, IGSMS-PA does not guarantee that the sum rate is maximized be-cause conventional optimization approaches in [34] are not applied to derivethe solutions. Moreover, the subgradient method performs very closely to theoptimal one at the expense of higher complexity over other power allocationschemes. However, the subgradient algorithm only require approximately 10

iterations to converge to near-optimal solutions as illustrated in Figure 3.5.This indicates that the computational complexity of this method is not toohigh. Despite the assumption that the power is equally allocated to each RB,the closed-form M-ERPA method maintain a good performance at a muchlower complexity. Therefore, M-ERPA is more suitable for practical imple-mentation. It is also worth mentioning that the performance gain of NOMAover optimal OFDMA is significant, even when suboptimal power allocationschemes are applied.

34 36 38 40 42 44 46

Total Power (dBm)

150

200

250

300

Sum

Rat

e (M

bps)

OptimalSubgradient methodM-ERPAFTPAFPAIGSMS-PAOFDMA

Figure 3.4: Performance in terms of sum rate versus total transmission power


5 10 15 20 25iteration

200

220

240

260

280

Sum

Rat

e (M

bps)

Figure 3.5: Convergence performance of subgradient method (Algorithm 3.1)when PT = 46 dBm (µ(0) = 50, τ(0) = 0, αµ = 0.4/

√t and ατ = 5/ (1 + t))

34 36 38 40 42 44 46

Total Power (dBm)

0

20

40

60

80

Use

r T

rans

mis

sion

Rat

e (M

bps)

Subgradient methodM-ERPAFTPAFPAIGSMS-PA

Subgroup H

Subgroup L

Figure 3.6: Transmission rate of individual users in each subgroup versustotal transmision power

Next, the performance in terms of individual user’s rates in each subgroupis presented in Figure 3.6. For the users’ rate in subgroup 1, both subgradi-ent method and M-ERPA outperforms the other schemes. This is achieved atthe expense of only small degradation to the rate of the users in subgroup 2.More importantly, both proposed schemes offer good degree of fairness sincethe difference in individual user’s rates between subgroup 1 and subgroup 2

is relatively smaller compared to that of other schemes. This indicates that theproportional rate constraint ensures a higher rate for the base layer, which canbenefit all users in this multicast multi-layered video coding system. Overall,the results in this subsection demonstrates that both subgradient method and


M-ERPA outperforms other schemes in terms sum rate performance and fair-ness requirement.

3.6.2 General Multi-Layer Case

For joint optimization of power allocation and subgrouping in general multi-layer case, a substantial performance gain can only be achieved with largernumber of users due to the heterogeneity of users’ channel qualities. Here, thenumber of users is set as K = 20. Based on SLPA, the sum rate performance ofthe three proposed subgrouping methods, the optimal exhaustive search so-lution and fixed arbitrary subgrouping are compared in Figure 3.7 for 2-layer(S = 2) and 4-layer (S = 4) cases. For the arbitrary subgrouping, the subgroup-ing configurations for 2-layer and 4-layer are set as K1,n, K2,n = 20, 10, ∀nand K1,n, K2,n, K3,n, K4,n = 20, 15, 10, 5, ∀n respectively. From Figure 3.7,all the proposed subgrouping methods offer superior performance gain overfixed arbitrary subgrouping for for both 2-layer and 4-layer cases. More im-portantly, the performances of all the proposed subgrouping methods arealmost comparable to that of exhaustive search solution for both 2-layer and4-layer cases. Method 1 offers the worst among all the proposed techniquesbecause it is easily trapped in local optima. Better performances are achievedby both Method 2 and 3 because of their additional features of potentiallyfinding the global optimum through the search for second local maximumand the elimination of local optima respectively.

Note that, for Method 3, the users are selected together when the differ-ence in CNR in dB is less than 5%, which is discovered to be a acceptablethreshold for all maximum total power PT scenarios. Figure 3.8 demonstratesthe influence of selecting various CNR percentage differences on the perfor-mance of Method 3. It can be observed that best sum rate performance canbe achieved when the CNR percentage differences is increased to 5-6% asmore local optima are possibly eliminated from the search space. However,when the CNR percentage differences is increased to more than 6%, the searchspace is reduced much further since many users are grouped together. Withreduced search space, global optimum and better local optima are possiblyeliminated and hence the sum rate performance drops. On the other hand, ifthe CNR percentage differences is lower than 5%, it is not sufficient to elim-inate some of the local optima and therefore significant performance gain


cannot be achieved because Method 3 may be trapped in local optima. No-tably, the sum rate at 5% difference is very close to the exhaustive searchresult, which verify the choice of the threshold that ensure the effectivenessof Method 3.

34 36 38 40 42 44 46

Total Power (dBm)

400

500

600

700

800

900

1000

Sum

Rat

e (M

bps)

Method 1Method 2Method 3 (5%)Exhaustive SearchArbitrary

41.5 42 42.5

760

770

780

790

800

41.5 42 42.5650

660

670

680

690

700

S = 4

S = 2

Figure 3.7: Sum rate performance of different subgrouping methods usingSLPA versus total transmission power for S = 2 and S = 4 cases.

0 1 2 3 4 5 6 7 8 9 10

% difference in terms of CNR in dB

600

602

604

606

608

610

Sum

Rat

e (M

bps)

Exhaustive SearchMethod 3

Figure 3.8: The effect of varying the percentage difference in CNR (in dB) ofMethod 3 on sum rate performance when PT = 40 dBm for the S = 2 case


Next, the complexity of the proposed subgrouping schemes, which is mea-sured by the average number of power allocation and sum rate computations,is presented in Figure 3.9 for both the 2-layer (S = 2) and 4-layer (S = 4) cases.Both Method 1 and 3 offer much lower complexity since Method 1 is easilyterminated at a local maximum and the search space in Method 3 is signifi-cantly reduced. Despite the good sum rate performance offered by Method 2,the computational complexity is higher due to its effort in finding the secondlocal maximum. It can be observed from Figure 3.9 that the complexity of the4-layer case is around three times higher than that of the 2-layer case. This isbecause, in the 4-layer case, the algorithm needs to optimize the number ofusers receiving the remaining three layers K(s)

n (for s ∈ 2, 3, 4) while only thenumber of users receiving the single enhancement layer K(2)

n is determined inthe 2-layer case. Note also that the order of complexity for the proposed meth-ods is almost consistent for all total transmission power PT . The above resultsdemonstrate that Method 3 is the best among all the proposed subgroupingmethods in terms of both sum rate and complexity performances. Therefore,Method 3 is employed thereafter to assess the performance of different powerallocation schemes. Moreover, in order to show results for the more generalcase, only 4-layer case is considered.

34 36 38 40 42 44 46

Total Power (dBm)

0

100

200

300

400

500

600

700

No.

of C

ompu

tatio

ns

2 Layer: Method 12 Layer: Method 22 Layer: Method 34 Layer: Method 14 Layer: Method 24 Layer: Method 3

Figure 3.9: Computational complexity of different subgrouping methods forS = 2 and S = 4 cases.

Figure 3.10 presents the sum rate performance of the proposed SLPA, FPA,


IGSMS-PA and the numerically optimized solution using Method 3 as a sub-grouping technique for the 4-layer case. From this figure, both SLPA andFPA perform better than IGSMS-PA. SLPA performs closely to FPA when thetransmission power is below 42 dBm. However, at higher transmission power,SLPA offers a degree of performance gain over FPA and is much closer tothe optimal solution. This is because, in the multi-layer case, the allocation ofhigher power to the uppermost enhancement layer will cause a higher level ofinterference to the lower layers and therefore does not necessarily increase theoverall rates. It is vital to ensure higher rates for the lower layers, particularlythe base layer which is intended for all users. Higher rate for the uppermostenhancement layer does not guarantee enhanced sum rates since there areonly small number of users receiving this layer. Meanwhile, the proposedSLPA guarantees substantial rates for the lower layers through achieving theminimum target rates and proportional fairness, and therefore potentially en-hances the sum rate performance. In contrast, FPA scheme does not complyto any minimum rate requirement and thus does not guarantee higher ratesfor the lower layers, particularly for the base layer which is mandatory for allusers to ensure the successful retrieval of video data. Meanwhile, IGSMS-PAallocates the power sufficiently to satisfy the exact minimum target rate andthen the remaining power to the uppermost layer, but does not necessarilymaximize the sum rate.

34 36 38 40 42 44 46

Total Power (dBm)

400

500

600

700

800

900

Sum

Rat

e (M

bps)

OptimalSLPAFPAIGSMS-PA

Figure 3.10: Sum rate performance of different power allocation schemes uti-lizing subgrouping Method 3 versus total transmission power for S = 4 case


0 10 20 30 40Minimum Rate (Mbps)

0

0.2

0.4

0.6

0.8

1O

utag

e pr

obab

ility

SLPAFPAIGSMS-PA

(a) Layer 1/Base layer (s = 1)


0

0.2

0.4

0.6

0.8

1

Out

age

prob

abili

tySLPAFPAIGSMS-PA

(b) Layer 2 (s = 2)


0

0.2

0.4

0.6

0.8

1

Out

age

prob

abili

ty

SLPAFPAIGSMS-PA

(c) Layer 3 (s = 3)


0

0.2

0.4

0.6

0.8

1

Out

age

prob

abili

ty

SLPAFPAIGSMS-PA

(d) Layer 4 (s = 4)

Figure 3.11: Outage probability for each layer stream s in S = 4 case againstdifferent target rate at PT = 46 dBm


In multicast video streaming, it is crucial to offer robust delivery of basicquality video to all the users to satisfy the targeted user experience. In orderto assess the robustness of the delivery of each layer stream, it is necessary toplot the performance of SLPA, FPA and IGSMS-PA schemes in terms of outageprobability as presented in Figure 3.11. Here, the total transmission power isfixed at PT = 46 dBm. From Figure 3.11(a), SLPA outperforms both FPA andIGSMS-PA in terms of the outage probability for the delivery of base layer(layer 1) stream. This demonstrates that SLPA can guarantee the delivery ofbase layer stream at higher target rates with lower outage, which is necessaryfor enhanced basic quality video. In addition, SLPA also offer better outageperformance for layer 2 and 3 as compared to FPA and IGSMS-PA as illus-trated in Figures 3.11(b) and 3.11(c). On the contrary, Figure 3.11(d) showsthat, for the uppermost layer, the outage performance of SLPA is the worstamong all the schemes. Nevertheless, it is worth pointing out that if the baselayer is in outage, receiving upper enhancement layers at high rates are use-less since these layers must be discarded in layered video coding. AlthoughFPA offers almost similar sum rate performance as SLPA, the users are oftenin outage because it does not always guarantee the successful delivery of thebase layer at a specific target rate. On the other hand, IGSMS-PA ensuresthe successful delivery of the lower layer streams by satisfying the minimumtarget rates, although the sum rate is not maximized. SLPA offers better sumrate performance while ensuring the robust delivery of lower layers, particu-larly for the base layer which is essential for the successful detection of videodata.

Finally, the fairness performance of FPA, SLPA and IGSMS-PA for the4-layer case is shown in Figure 3.12 when the transmission power is atPT = 46 dBm. The fairness is measured using Jain’s index which is defined

in [96] as J =(∑Kk=1Rk)

2

K∑Kk=1R

2k

, where Rk is the transmission rate of each user k andthe fairness index must be within the range 0 ≤ J ≤ 1. In this simulation, thelayer streams are discarded if the target rate requirement of Φ

(1)min = 2.5 Mbps,

Φ(2)min = 5 Mbps, Φ

(3)min = 7.5 Mbps, and Φ

(4)min = 10 Mbps are not achieved. In

layered coding technique, the decoding of a layer stream is highly depen-dent on the successful retrieval of lower layers. Therefore, if a layer is inoutage, the layer and its upper layers are all discarded. From Figure 3.12, itis demonstrated that SLPA achieves higher fairness compared to FPA due to


the proportional allocation of resources to all the layers. However, IGSMS-PAachieves the highest user fairness at the expense of much lower sum rate. Thisis because the rates of the lower layers are fixed at minimum target rates.

5 10 15 20 25 30

Number of Users

0.88

0.89

0.9

0.91

0.92

Fai

rnes

s In

dex

SLPAFPAIGSMS-PA

Figure 3.12: Fairness index for S = 4 case against different number of usersat PT = 46 dBm with the target rates of Φ

(1)min = 2.5 Mbps, Φ

(2)min = 5 Mbps,

Φ(3)min = 7.5 Mbps, and Φ

(4)min = 10 Mbps

3.7 Summary

In this chapter, a resource allocation scheme is developed for multi-layer mul-ticast video streaming in NOMA networks. The optimization problem isformulated with the objective of maximizing the sum rate while achievingthe transmission power budget and proportional rate constraints. First, twosuboptimal power allocation schemes for the 2-layer case considering arbi-trary subgrouping has been proposed: an iterative subgradient method and aclosed-form M-ERPA technique. Based on the 2-layer based M-ERPA solution,the work extends to a general multi-layer case by proposing a low-complexitySLPA scheme which successively allocates power to each layer. In order tojointly optimize the power allocation and subgrouping, the SLPA scheme isincorporated into three proposed low-complexity subgrouping methods. Forthe 2-layer case with arbitrary subgrouping, simulation results demonstrates


the superiority of both the subgradient method and M-ERPA over other low-complexity schemes in terms of sum rate and user fairness. Simulation resultsfor the multi-layer case have shown that the proposed subgrouping meth-ods perform close to the optimal exhaustive search technique but with lowercomplexity. In particular, Method 3 offers the best performance in terms ofachieving higher sum rates at a reduced computational complexity. Mean-while, the proposed SLPA scheme performs better than FPA and IGSMS-PAwhile being close to the optimal solution particularly at higher transmissionpower. Moreover, SLPA ensures substantial rates for the base layer streamwhich guarantees robust delivery of mandatory video data to all users.

Chapter 4

Low-Complexity BeamformingAlgorithms for NOMA-basedLayered Multicast Systems

4.1 Introduction

It was demonstrated in Chapter 3 that multicast network employing bothNOMA and layered coding offer superior performance in terms of sum rateand resource utilization when compared to OMA-based layered multicastnetwork. Similar to conventional unicast NOMA system, the performanceof NOMA-based layered multicast system can be further enhanced by im-plementing multiple-input single-output (MISO) system with sophisticatedbeamforming techniques. However, the complexity of beamforming meth-ods may hinder its practical implementation. Therefore, this chapter focuseson developing low-complexity beamforming schemes which can improve theperformance of NOMA-based layered multicast system. This chapter be-gins with the review of existing works on beamforming techniques in mul-ticast system, which is described in Section 4.2. Next, the system model forthe MISO-based layered multicast in NOMA system is presented in Section4.3. Section 4.4 formulates the sum rate maximization problem while ensur-ing transmission power budget and proportional rate constraints are satis-fied. With predetermined beamforming weights, the power allocation can besolved by using the method proposed in Chapter 3. Therefore, this chap-ter only focuses on designing three beamforming methods which can further

97

CHAPTER 4. BEAMFORMING FOR LAYERED MULTICAST IN NOMA 98

improve the sum rate performance. The first technique is designed to en-hance the CNR of the weakest user in each subgroup. Meanwhile, the sec-ond scheme focuses on nullifying the interference caused by the enhancementlayer for the users in the weaker subgroup 1 while improving the CNRs of se-lected users. The final method is an iterative algorithm which proportionallyenhances the CNRs of the users and therefore improves the worst-case CNR.Finally, the effectiveness of the proposed beamforming schemes are demon-strated through simulation results presented in Section 4.5.

4.2 Related Works

The performance of multicast networks can be enhanced by employing MISOsystem, in which the desired video signal is steered towards multicast usersthrough beamforming technique. In multicast system, the transmission rate isrestricted by the minimum of the individual achievable rates or equivalentlythe received SNRs among the users. In order to enhance the overall through-put performance, it is common to optimize the beamforming weight basedon max-min problem, which maximizes the minimum received SNR whileachieving a maximum power budget. However, it is difficult to obtain theoptimal solution for max-min problem since it is an NP-hard problem [97].Suboptimal solutions can be determined using several state-of-the-are tech-niques such as semi-definite relaxation (SDR) with randomization [97] andsuccessive linear algorithm (SLA) [98,99]. Nevertheless, these methods are as-sociated with high computational complexity which can increase the latencyof video delivery due to increased computational burdens [49]. This motivatesthe studies of low-complexity beamforming methods which can practically re-duce the computational burdens of base stations. One of the low-complexityscheme is designed by [100] with the aim of maximizing the average SNRover all users. However, it does not necessarily maximize the minimum SNRsamong the users and therefore desired performance may not be achieved atthe expense of low complexity. On the other hand, [101] proposed an itera-tive algorithm which can improve the performance by enhancing the SNR ofthe weakest user in each iteration. However, this method only focuses on theweakest user. In [99], beamforming algorithms were designed to offer muchimproved performance by taking into account the SNRs of all users.


The performance of the NOMA-based layered multicast system describedin Chapter 3 can be further improved by employing beamforming techniques.In [85], the beamforming and power allocation problem are solved based onpower minimization problem. However, the iterative beamforming algorithmproposed in that work does not guarantee optimal solution. Moreover, thisscheme can only be applied for two-user case. Multiple user case was stud-ied in [60] and [61] with the objective of minimizing the total transmissionpower minimization and maximizing total utility function respectively. Inthese works, suboptimal beamforming solutions are obtained by implement-ing high-complexity SDR-based algorithm which may hinder the practicalimplementation of NOMA-based layered multicast system. Therefore, thischapter investigates low-complexity beamforming schemes which can offersignificant performance gain over multicasting with single input single output(SISO). In this chapter, joint power allocation and beamforming is consideredwith the objective of maximizing the sum rate while satisfying the total trans-mission power and proportional rate constraint, which is a complex optimiza-tion problem. Since power allocation problem with pre-determined beam-forming weights is a convex optimization problem according to [102], it isnecessary to solve the beamforming weights first. The beamforming weightsare determined with the aim of improving the worst-case CNR which con-sequently enhance the sum rate. This chapter focuses on developing beam-forming schemes which enhances the minimum of the CNRs among all usersat a low complexity. Then, the power allocation problem can be solved usingany existing power allocation schemes. In this work, the power allocation de-veloped in Chapter 3 is employed in order to specifically solve the sum ratemaximization problem with transmission power budget and proportional rateconstraints.

Notations: Throughout this chapter, lowercase and uppercase boldface let-ters are used for vectors and matrices respectively. The Hermitian (conju-gate) transpose is denoted as (·)H denotes. The symbol Cn represents then-dimensional complex space. Meanwhile, Im and 0m×n denote an m × m

identity matrix and an m×n zero matrix respectively. The Euclidean norm ofa vector is represented as ‖·‖2.


4.3 System Model

This chapter considers a downlink NOMA-based layered multicast systemthat contains a BS equipped with M antennas and K single-antenna usersrequesting the same video content1. This video data is transmitted by theBS with a maximum transmission power of PT over a bandwidth of B2. TheBS is assumed to possess the full knowledge of the CSI for all users. Notethat the the complex channel vector between the BS and kth user hk ∈ CM

takes into account the effect of path loss, shadowing and Rayleigh fading.Let K = 1, 2, . . . , K be the set of all users who are arranged in ascendingorder of their channel gains i.e., ‖h1‖2 ≤ ‖h2‖2 ≤ . . . ≤ ‖hK‖2. This channelordering is generally valid since the effect of path loss and shadowing is moredominant. The users are split into two subgroups and the set of users insubgroup 1 and 2 are denoted as G1 and G2 respectively. These users aregrouped based on their order of channel gains ‖hk‖2 such that G1 numberof weakest users are selected for subgroup 1 and the remaining G2 strongerusers are in subgroup 2.

At the BS, the video content is split into two layers of data streams us-ing layered coding. The base layer stream XBL must be successfully detectedby all users since it contains the mandatory data elements for basic qualityvideo. Meanwhile, the enhancement layer stream XEL improves the qual-ity of the video and is only intended for the users in subgroup 2 who havebetter channel conditions. Both base and enhancement layer streams are mul-tiplexed in power domain using SC and hence the received signal at user k isrepresented as

yk = hHk

(√PBLvBLxBL +

√PELvELxEL

)+ wk (4.1)

where xBL and xEL are the transmitted symbols of the base and enhancementlayer streams respectively, PBL and PEL are the power allocated to the base

1Similar to the works in Chapter 3, it is assumed that there are no other users in the cellrequesting for other video contents in order to focus on beamforming problem. In practice,these users can be served in different subbands or RBs.

2Note that most works in beamforming for multicast system, such as [97] and [61], onlyconsider single subband or RB due to the complexity of beamforming problem. Furthermore,the constraint related to unit norm beamforming vector applies to each RB. Therefore, a singleRB case is considered in this work. Practically, video data is delivered over multiple RBs inorder to address frequency selective fading.


and enhancement layers respectively, vBL ∈ CM and vEL ∈ CM are the beam-forming weights for the base and enhancement layers respectively, and wk isthe AWGN.

Similar to the system model described in Chapter 3, all the users can de-code the base layer stream (being the weak users’ signal) by treating the en-hancement layer stream as noise. Meanwhile, the users in subgroup 2 canperform SIC in order to retrieve the enhancement layer streams. This SIC de-coding order is commonly employed in NOMA-based multicast system sinceit guarantees the decoding of the high-priority base layer first before detectingthe enhancement layers [63, 90].

The transmission rate of the base layer stream depends on the minimumachievable rates attained by all users in order to guarantee the retrieval ofbasic quality video for all users. Hence, the achievable rate of the base layerstream is given by

RBL = mink∈K

B log2

(1 +

PBLΛ(BL)k

PELΛ(EL)k + 1

)(4.2)

where Λ(BL)k =

|hHk vBL|2BN0

and Λ(EL)k =

|hHk vEL|2BN0

are the CNRs of each user kreceiving the base and enhancement layers respectively, and N0 is the noisepower spectral density. The achievable rate of the enhancement layer streamis restricted by the minimum achievable rates of the users in subgroup 2 andthus, is expressed as

REL = mink∈G2

B log2

(1 + PELΛ

(EL)k

). (4.3)

The sum rate of the layered multicast system employing in NOMA andMISO techniques is represented as

Rsum = K RBL +G2REL. (4.4)


4.4 Problem Formulation and Proposed Beamform-

ing Schemes

First, the optimization problem is formulated with the aim of maximizing thesum rate while satisfying both transmission power budget and proportionalrate constraints as follows

maximizePBL,PEL,vBL,vEL

Rsum (4.5a)

subject to PBL + PEL ≤ PT (4.5b)

PBL ≥ 0, PEL ≥ 0 (4.5c)

RBL : REL = Φ(BL)min : Φ

(EL)min (4.5d)

‖vBL‖22 = 1, ‖vEL‖2

2 = 1 (4.5e)

where constraints (4.5b) and (4.5c) are the power budget and non-negativepower constraints respectively. Constraint (4.5d) is the proportional rate con-straint which ensures the minimum target rates of Φ

(BL)min and Φ

(EL)min are achieved

by base and enhancement layer streams while maintaining a proportionalityamong both layer streams. As mentioned in Chapter 3, this constraint guar-antees that sufficiently higher power is allocated to the base layer stream inorder to allow its successful detection while treating the enhancement layerstream as noise. This will guarantee all the users to receive the basic qualityvideo. Furthermore, enough power is assigned to the base layer stream in or-der to allow users with better effective channel conditions to perform SIC andthen detect the enhancement layer data to improve the quality of the video.This SIC decoding order is commonly applied in MISO-based layered multi-cast in NOMA system such as in [60,61]. Finally, constraint (4.5e) ensures thebeamforming weights are unit norm vectors.

The problem which involves optimizing beamforming weights is gener-ally not convex [60, 61, 85] and hence it is difficult to solve the optimizationproblem (4.5). Nevertheless, sum rate maximization which optimizes onlypower allocation under fixed beamforming weights is a convex optimizationproblem [102]. Therefore, in order to solve (4.5), the beamforming weights aredetermined first by considering only the channel conditions of each user hkwhile achieving constraint (4.5e). The aim of the beamformer is to enhance


the transmission rate of each layer streams by improving the minimum of theCNRs among all users.

Once the beamforming weights vBL and vEL are obtained, the individualCNRs are known and hence, the power allocated to both base and enhance-ment layer streams (i.e., PBL and PEL ) can be solved. Here, the power allo-cation problem is solved using equal RB power allocation for multicast (M-ERPA) as proposed in Chapter 3 in order to address the remaining constraints(4.5b), (4.5c) and (4.5d). Therefore, this chapter only focuses on solving thebeamforming weights. The subsequent subsections presented the three pro-posed beamforming schemes which can be incorporated with any power al-location techniques, including M-ERPA, in order to obtain low-complexitysuboptimal solutions to problem (4.5).

4.4.1 Maximizing the CNRs of the Weakest Users (MCNR)

In this subsection, the beamforming weights for both layer streams are de-signed by directing each layer beam towards the weakest user in each re-spective subgroup. Let ABL and AEL be the matrices in which each columnvector represents the complex channel vector of the user who is being directedwith base layer and enhancement layer signal beams respectively. From (4.3),the transmission rate of the enhancement layer stream can be enhanced bydesigning the weight vEL that is aimed towards maximizing the CNRs ofthe users receiving enhancement layer in subgroup 2, i.e., Λ

(EL)k for k ∈ G2.

In order to maximize the average CNR over all users, the method proposedin [100] determine the weight vector by solving the dominant eigenvector ofa matrix which consists of the complex channel vectors of all users. How-ever, in that approach, the improvement in the CNRs of the users are uneven.Specifically, the CNR of the users with better channel conditions are signif-icantly improved while sacrifying that of the weaker users [103]. Since theachievable rate in (4.3) is restricted by the user with least CNR in subgroup2, higher transmission rate will not be achieved for the enhancement layerstream. Therefore, this work focuses on maximizing the CNR of the user withthe weakest channel condition. For this case, the considered matrix AEL onlyincludes the complex channel vector of the weakest user in subgroup 2 i.e.,AEL = [hq] for q = arg mink∈G2 ‖hk‖2. Therefore, the weight for the enhance-ment layer vEL is determined by solving the dominant unit-norm eigenvector


of AELAHEL.

Meanwhile, the transmission rate of the base layer stream according to(4.2) is restricted by the user with the worst ratio between the received baselayer signal and the interference from the enhancement layer. Since the en-hancement layer beam is directed towards the weakest user in subgroup 2, theinterference towards the base layer will inevitably be higher for that user. Inorder to address this interference, the base layer beam must also be directedtowards the weakest user in subgroup 2 and thus the complex channel vectorof this user must be included in ABL. Furthermore, the beam must also besteered towards the user with the least channel quality in order to improvethe minimum of the CNRs among all users which influence the achievablerate of the base layer. Therefore, the matrix ABL must include the complexchannel vectors of the weakest users in both subgroup i.e., ABL = [hp hq]

for p = arg mink∈K ‖hk‖2. The beamforming weight for the base layer vBL isobtained by solving the dominant unit-norm eigenvector of ABLA

HBL.

4.4.2 Interference Nulling and Maximizing the CNRs of Se-

lected Users (N-MCNR)

The users in subgroup 1 only need the base layer stream to obtain the basicquality video and thus, it is necessary for the users in subgroup 1 to eliminatethe enhancement layer from the received signal. For this reason, the weightvEL can be designed to nullify the the interference caused by the enhance-ment layer stream towards the users in subgroup 1. On the other hand, it isalso crucial to steer the enhancement layer beam towards the weakest user insubgroup 2 in order to improve the worst case CNR. However, this approachdoes not necessarily improve the sum rate because the interference nullingand steering the beam only to the weakest user may cause uneven CNR per-formance to other users. Therefore, an algorithm is proposed to examinewhether directing the beam to multiple users will improve the enhancementlayer rate. While the rate of the enhancement layer is improved, the rate of thebase layer is restricted by the interference caused by the enhancement layer.Therefore, the weight vBL is designed to encounter this interference. Sincethe interference effect experienced by the users in subgroup 1 has been nulli-fied, the users in subgroup 1 will have good achievable rates despite having


weaker channel conditions. Therefore, it is not necessary to direct the baselayer beam towards these users. The proposed algorithm will initially steerthe beam towards the weakest user in subgroup 2 and then assess whetherinluding other users will improve the sum rate. The proposed algorithm ispresented in Algorithm 4.1.

Algorithm 4.1 Nulling and maximizing individual CNR Algorithm1: Initialize: ABL and AEL , AG1 = [hk]k∈G1

, and i = 1

2: Obtain V G1 by solving AHG1V G1 = 0

3: for iEL = 1 to G2

4: Obtain vEL by solving the dominant eigenvector of

V HG1AELA

HELV G1

5: for iBL = 1 to G2

6: Calculate vEL(i) = V G1 vEL

7: Obtain vBL(i) by solving the dominant eigenvector of

ABLAHBL

8: Calculate the sum rate Rsum(i)

9: Calculate Γ(BL)k for each user in subgroup 2

10: Update ABL based on minimum Γ(BL)k

11: i← i+ 1

12: end for13: Calculate Λ

(EL)k for each user in subgroup 2

14: Update AEL based on minimum Λ(EL)k

15: end for16: Obtain i∗ = arg maxRsum(i)

17: output: the solutions w∗BL = wBL(i∗) and w∗EL = wEL(i∗)

Let AG1 = [hk]k∈G1be a matrix in which each column vector represents the

complex channel vector of each user in subgroup 1 who does not require theenhancement layer. The singular value decomposition (SVD) of this matrix isAG1 = UG1ΣG1

[VG1VG1

], where VG1 ∈ CM×(M−G1) consists of the last M − G1

right-singular column vectors with V HG1VG1 = IM−G1 . The column vectors in

VG1 are the basis for the right null space of AG1 i.e., AG1VG1 = 0G1×(M−G1).The weight vector vEL, which nullifies the enhancement layer from the signalreceived by the users in subgroup 1 is determined according to [104] as

vEL = VG1vEL (4.6)


where vEL ∈ C(M−G1)×1 is a unit norm complex vector which is designedto steer the beam towards the selected users in subgroup 2. The vectorvEL is obtained by solving the dominant unit-norm eigenvector of the ma-trix V H

G1AELA

HELV G1 . At the beginning of the algorithm, the enhancement

layer beam is steered towards the weakest user in subgroup 2 only i.e., AEL =

[hq] for q = arg mink∈G2 ‖hk‖2. In order to address the interference caused byenhancement layer, the base layer beam is also directed towards the weakestuser in subgroup 2 (i.e., ABL = [hq]).

The outer loop (step 3-4 and 13-15 of Algorithm 4.1) determines theweight vector vEL based on the updated AEL. In each iteration of theouter loop, the complex channel vector of a user from subgroup 2 withthe lowest CNR is added into the matrix AEL (i.e., AEL = [AEL hs] for

s = arg mink∈G2 Λ(EL)k =

|hHk vEL(i)|2BN0

). While the direction of the enhancementlayer beam is fixed towards specific users according to updated AEL, the in-ner loop (step 5-12 of Algorithm 4.1) calculates the sum rate Rsum(i) basedon the updated weight vectors vBL(i) and vEL(i) which are determined usingthe dominant unit norm eigenvector of the matrix ABLA

HBL and (4.6) respec-

tively. Note that the sum rate Rsum(i) is calculated based on a given powerallocation technique such as M-ERPA. Then, the matrix ABL is updated forthe next iteration by including a user from subgroup 2 who has the low-est channel-to-interference-plus-noise ratio (CINR), i.e., ABL = [ABL hr] for

r = arg mink∈G2 Γ(BL)k =

|hHk vBL(i)|2

|hHk vEL(i)|2+BN0

. At the end of the algorithm, the sub-

optimal solutions for the weight vectors are obtained as v∗BL = vBL(i∗) andv∗EL = vEL(i∗) considering that the sum rate at the i∗th iteration is the highestone.

4.4.3 Multiplicative Update Algorithm (NOMA-MU)

The beamforming techniques proposed in subsections 4.4.1 and 4.4.2 onlymaximize the CNR of selected users by directing the beams towards theseusers while ignoring other users. These beamformers may lower the CNRof the neglected users which affect the sum rate performance. Therefore,it is crucial to design a beamforming scheme which enhance the CNR ofthe weaker users by not sacrifying the users with better channel condition.This can be addressed by employing a max-min beamformer which maximize


the minimum CNR among the users. Nevertheless, max-min approach is anNP-hard problem which can only be solved using high-complexity numericaltools. According to [99], low-complexity suboptimal schemes can be derivedfrom the approximate solutions of a proportionally fair multicast beamform-ing problem that is close to the conventional max-min approach. Multiplica-tive Update (MU) algorithm is one of the techniques proposed in [99] whichis shown to offer significant performance at low complexity and guaranteedconvergence. However, that work only considered conventional multicast sys-tem, in which a single beamforming vector is obtained for the transmissionof a common signal to all users. In NOMA-based layered multicast system,beamforming weights for both base and enhancement layers can be designedby employing the MU algorithm in two stages as summarized in Algorithm4.2. This proposed technique is referred to as NOMA-based MultiplicativeUpdate (NOMA-MU).

Algorithm 4.2 NOMA-based Multiplicative Update AlgorithmFirst stage1: Initialize: vEL(1), CEL

k =hkh

Hk

BN0, t = 1, Tmax and ε

2: while |Λ(EL)min (t− 1)− Λ

(EL)min (t)| ≥ ε and t ≤ Tmax do

3: Update the weight vEL(t+ 1) according to (4.7) and (4.8)

4: Find Λ(EL)min (t+ 1) = mink∈G2

10 log10

(|hHk vEL(t+1)|2

BN0

)5: t← t+ 1

6: end while7: output: the suboptimal solution v∗ELSecond stage8: Initialize: vBL(1), CBL

k =hkh

Hk

|hHk v∗EL|2+BN0

, t = 1, Tmax and ε

9: while |Γ(BL)min (t− 1)− Γ

(BL)min (t)| ≥ ε and t ≤ Tmax do

10: Update the weight vBL(t+ 1) according to (4.9) and (4.10)

11: Find Γ(BL)min (t+ 1) = mink∈K

10 log10

(|hHk vBL(t+1)|2

|hHk v∗EL|2+BN0

)12: t← t+ 1

13: end while14: output: the suboptimal solution v∗BL

In the first stage, the weight for the enhancement layer v∗EL is solved it-eratively using the MU algorithm. The MU updating equations, which are


derived according to [99], are given by

vEL(t+ 1) =

(∑k∈G2

CELk vEL(t)

vHEL(t)CELk vEL(t)

)(4.7)

vEL(t+ 1) =vEL(t+ 1)

‖vEL(t+ 1)‖(4.8)

where CELk =

hkhHk

BN0is the CNR covariance matrix of each user who receives

the enhancement layer data. Note that the updating equations only involveCELk for the users in subgroup 2. This is because, from (4.3), the rate of en-

hancement layers only depends on the individual performance of the users insubgroup 2. It is worth mentioning that, from (4.7), the weight vEL is updatedby not only favouring the weakest user, but also other users in subgroup 2.

The weight vector for the base layer v∗BL is determined in the second stageusing the MU algorithm. However, the interference caused by the enhance-ment layer based on v∗EL solved in the first stage is taken into account inthe CINR covariance matrix of each user receiving the base layer stream, i.e.,CBLk =

hkhHk

|hHk v∗EL|2+BN0

. Here, CBLk for all users are considered since base layer

data are required by all users. The updating equations in the second stage areexpressed as

vBL(t+ 1) =

(∑k∈K

CBLk vBL(t)

vHBL(t)CBLk vBL(t)

)(4.9)

vBL(t+ 1) =vBL(t+ 1)

‖vBL(t+ 1)‖. (4.10)

In both stages, the iteration stops when the minimum CNR/CINR expres-sions (Λ(EL)

min and Γ(BL)min ) converge (i.e., when the difference between minimum

CNR/CINR in the current iteration and in the previous iteration is less thana tolerance value ε) or when the iteration index t reaches the maximum itera-tions Tmax.

4.4.4 Computational Complexities

The first method (MCNR) only needs 2 computations of dominant eigenvec-tor. On the other hand, N-MCNR algorithm requires G2 × G2 iterations andeach iteration involves the calculation of both eigenvector and sum rates. In


addition, the basis for the right null space is computed at the start of this algo-rithm. Meanhwhile, the complexity of the iterative NOMA-MU depends onthe strictness of the stopping criteria which, in this case, is the tolerance valueε. Hence, the complexity of the two-stage NOMA-MU algorithm is O

(2ε2

).

4.5 Simulation Results

In this section, the simulation considers a downlink NOMA-based multicastsystem, which consists of a BS with M antennas located at the centre and K

single-antenna users that are uniformly distributed within a circular cell ra-dius of 500 m. Table 4.1 summarizes the parameters associated with variouswireless channel impairments including path loss, shadowing effect, noiseand Rayleigh fading. In order to determine a low-complexity suboptimalsolution for the problem (4.5), M-ERPA technique proposed in Chapter 3 isapplied after the beamforming weights are solved using the techniques pro-posed in this chapter. Meanwhile, for MISO-OMA system, MU algorithm isapplied to solve the beamforming weights and then the subgradient methodis used to solve power allocation problem considering proportional rate con-straint.



Bandwidth, B 200 kHz

Cell radius 500 m







Φ(BL)min 0.5 Mbps

Φ(EL)min 0.5 Mbps


20 22 24 26 28 30 32 34 36Total Power (dBm)

5

6

7

8

9

10

11

12

Sum

Rat

e (M

bps)

SISO-NOMAMISO-OMA: MUMISO-NOMA: MCNRMISO-NOMA: N-MCNRMISO-NOMA: MU

(a)

20 22 24 26 28 30 32 34 36Total Power (dBm)

0.5

1

1.5

2

2.5

3

Indi

vidu

al U

ser

Rat

e (M

bps)


Subgroup 1

Subgroup 2

(b)

Figure 4.1: (a) Sum rate and (b) individual user rate performance for K = 5(G1 = 2 and G2 = 3 ) and M = 4

First, the performances of the proposed beamforming schemes in termsof sum rate and individual user rate are presented for the case of K = 5

users and M = 4 transmit antennas at the BS in Figure 4.1. The two weakestusers are arbitrarily selected in subgroup 1 (i.e., G1 = 2) and the remainingstronger users are in subgroup 2 (i.e., G2 = 3) to ensure that more than 50%of the users achieve excellent quality of service. Figure 4.1(a) demonstratesthe effectiveness of MISO system in improving the performance of NOMA-based layered multicast system. More importantly, all proposed beamformingschemes offer enhanced sum rate performance over both SISO-based NOMA


and MISO-based OMA systems. NOMA-MU offers the best performance at ahigher complexity because the beamforming weights are iteratively updatedby taking into account the CNR of all users. The sum rate performance ofMCNR is close to that of N-MCNR although the complexity is much lower.From Figure 4.1(b), similar trends can also be observed for the individual userrate performance. Note that the gap between the rate of the users in subgroup1 and those in subgroup 2 is not large, which shows the effectiveness of theproportional rate feature in M-ERPA technique.

20 22 24 26 28 30 32 34 36Total Power (dBm)

6

8

10

12

14

16

18

20

Sum

Rat

e (M

bps)


(a)

20 22 24 26 28 30 32 34 36Total Power (dBm)

0.5

1

1.5

2

2.5

3

Indi

vidu

al U

ser

Rat

e (M

bps)


Subgroup 1

Subgroup 2

(b)

Figure 4.2: (a) Sum rate and (b) individual user rate performance for K = 8(G1 = G2 = 4) and M = 8


Figure 4.2 shows the sum rate and individual user rate performances ofthe proposed schemes when the number of users is increased to K = 8 andthe number of antennas at the BS is M = 8. Here, the number of users for bothsubgroups is set as four (i.e., G1 = G2 = 4). From Figure 4.2, it can be observedthat both NOMA-MU and N-MCNR techniques can maintain substantial per-formance over both SISO-based NOMA and MISO-based OMA systems evenwhen the number of users is increased. Note that the comparison betweenMISO-NOMA and MISO-OMA using MU algorithm demonstrates the supe-riority of NOMA over OMA system. However, the the performance of MCNRis degraded and become closer to that of SISO-based NOMA system. This isbecause MCNR only focuses on maximizing the CNR of the weakest user ineach respective subgroup. There is more likelihood that the second weakestuser has relatively the same channel quality as that of the weakest one, par-ticularly when the number of users is increased. Since the CNR of the secondweakest user is not maximized and the rate of each layer stream may thendepends on this user, the sum rate will not be enhanced especially when thenumber of users is larger.

1 2 3 4 5 6 7 8 9 10 11iteration

33

34

35

36

37

38

Min

imum

CN

R (

dB)

(a) First stage: weight for enhancement layer

1 2 3 4 5 6 7 8 9 10 11 12 13iteration

-13

-11

-9

-7

-5

-3

-1

Min

imum

CIN

R (

dB)

(b) Second stage: weight for base layer

Figure 4.3: The convergence behaviour of NOMA-MU for K = 5 (G1 = 2 andG2 = 3 ) and M = 4


Finally, the convergence rate of the the NOMA-MU algorithm is demon-strated in Figure 4.3 for the case of K = 5 (G1 = 2 and G2 = 3 ) and M = 4.It can be shown from Figure 4.3 that NOMA-MU only needs around 5-6 iter-ations to converge to a solution in each stage, which indicates that the com-plexity of this method is not too high. Therefore, NOMA-MU is suitable forpractical implementation since it can guarantee significant performance at alow complexity.

4.6 Summary

This chapter focused on investigating low-complexity beamforming schemesfor layered multicast in NOMA system. In this work, beamforming problemis solved by using three low-complexity suboptimal beamforming techniqueswhich can incorporated with any existing power allocation schemes. M-ERPAscheme proposed in Chapter 3 is applied in this work in order to maximizethe sum rate while satisfying the total transmission power and proportionalrate constraints. Simulation results demonstrate the superiority of the pro-posed beamforming techniques over SISO-based multicast system. Moreover,it was shown that NOMA maintains superior performance over OMA whensimilar beamforming scheme is applied. NOMA-MU algorithm offer betterperformance compared to other proposed schemes at the expense of highercomplexity. Nevertheless, it was shown that NOMA-MU converges quicklywhich indicates that the complexity of this scheme is still considered low.Both MCNR and N-MCNR offers good performance at a much lower com-plexity. However, the performance of MCNR is affected by larger number ofusers.

Chapter 5

NOMA and Coded Multicasting inCache-Aided Wireless Networks

5.1 Introduction

It is highly anticipated that wireless caching technology will be adopted in thefuture mobile networks in order to alleviate the traffic caused by rich multime-dia contents [15, 67]. The technology allows the retrieval of popular contentsdirectly from the caches which are installed in close proximity to the users(e.g. routers, BSs, UEs etc.). This will reduce the utilization of network trafficresources in the backhaul. The application of NOMA as a content placementor delivery technique will further ensure the efficient utilization of over-the-air traffic resources. This chapter begins with the review of existing worksrelated to the application of NOMA as a transmission technique for cachingat the BSs and UEs in Section 5.2. Nevertheless, this chapter focuses on thestudy of the delivery schemes for caching at the UEs, particularly NOMAand coded multicasting. The system model for cache-aided network utilizingboth NOMA and coded multicasting is presented in Section 5.3. In Section5.4, the performance of NOMA and coded multicasting for two-user pairingare investigated in terms of probability of sum rate comparison and outageprobability using both analytical and simulation results. In order to exploitthe benefits of NOMA and coded multicasting, a hybrid delivery scheme isproposed in Section 5.5. The performance of the hybrid scheme is furtherenhanced by jointly optimizing the power allocation and user pairing or RBallocation based on sum rate maximization problem. Both near-optimal and

114

CHAPTER 5. NOMA AND CM IN CACHE-AIDED NETWORKS 115

low-complexity schemes are proposed to solve the optimization problem. Fi-nally, the effectiveness of the proposed hybrid transmission techniques in im-proving the sum rate performance is validated through simulation results.

5.2 Related Works

There are several studies in literature that consider NOMA as an effectivedelivery scheme for caching at the BSs or dedicated content server. For in-stance, in [105] and [106], the BS is allowed to frequently update the contentin multiple cached-enabled local servers through NOMA transmission. More-over, this process can either be performed during off-peak period (i.e., contentplacement phase) or jointly with content delivery during on-peak hours. Ac-cording to that studies, this strategy significantly enhances the performancein terms of hit probability and delivery outage probability. Furthermore, thestudy in [107] have shown that NOMA outperforms OMA in terms of cover-age performance. In [108], the performance is further improved by consider-ing dynamic power control with the objective of minimizing the transmissiondelay. Meanwhile, the resource allocation work in [73] considered applyingNOMA to simultaneously perform both multicasting and content pushing.On the other hand, in [109] and [110], NOMA is considered as both contentplacement and delivery technique for cache-aided C-RANs and fog radio ac-cess networks (F-RANs) respectively.

Caching at the UEs is considered the most preferable technique due to itspotential in relieving the traffic for both backhaul and over-the-air links. Nev-ertheless, the limited cache storage in UEs poses greater challenges to the de-sign of content placement and delivery. Coded multicasting has been widelyrecognized as an effective content delivery for caching at the UEs [67,70,111].In coded multicasting, each content in the remote server is divided into anumber of subfiles. Depending on the prediction of users’ request, the usersonly store selected subfiles of each content during the content placementphase. At the beginning of the delivery phase, each user demands for themissing subfiles of the requested content, which are then encoded into a sin-gle coded stream through XOR coding. The BS delivers the coded stream tomultiple users via multicast transmission, which is associated with enhanced


spectral and energy efficiency. However, the transmission rate of coded mul-ticasting is restricted by the least channel gain user.

NOMA has recently emerged as an alternative delivery scheme. The worksin [71,72] introduced the idea of CIC technique in detecting the desired signalwith the aid of the information available in the cache. In particular, a user firstcancels the interfering signal by exploiting the cached data. This enable theuser to detect the desired signal directly from the residual signal. Differentfrom conventional NOMA, interference cancellation can also be performed byweak user and thus leading to improved interference-free transmission rate.This will further enhance the spectral efficiency over conventional NOMA.It is demonstrated in [71, 72] that CIC-based NOMA significantly improvesthe achievable downlink rate region for all the user. Due to the promisingperformance, CIC-based NOMA have recently attracted interest in literature.For instance, the works in [71, 72] studied the power allocation for CIC-basedNOMA based on delivery time minimization problem. Meanwhile, the workin [112] employs divide-and-conquer-based and deep reinforcement learningmethods to optimize the power allocation, which is aimed towards maxi-mizing the probability that all the users can detect the requested contents.High-complexity numerical techniques are implemented in [113] in order tooptimize the power allocation for the application of cache-aided NOMA invehicular network. The outage performance for CIC-based NOMA was in-vestigated for multi-relay network in [114], but did not consider power allo-cation problem. Nevertheless, none of the abovementioned works comparethe performance of NOMA and coded multicasting.

Although NOMA and coded multicasting offer enhanced spectral effi-ciency, both techniques possess some limitations. While the individual rate incoded multicasting is limited by the user with the weakest channel quality, thesum rate performance of NOMA may be affected due to the sharing of poweramong users. Moreover, it is not practical to serve all the users in the sametime-frequency resource due to the system complexity of both NOMA andcoded multicasting. This problem is addressed by user pairing or clusteringconcept, which is commonly used in conventional NOMA. In this approach,the users are divided into pairs or clusters and each allocated with differenttime-frequency resource. The impact of user pairing on conventional NOMAis studied in [45] and it was shown that sum rate performance gain can be


achieved over OMA when the paired users’ channel gains are highly distinc-tive. However, for CIC-based NOMA and coded multicasting, it is not knownwhich delivery technique performs better under different pairing scenarios(i.e., selection of users based on their channel gains). In this thesis, the per-formance of NOMA and coded multicasting is investigated in terms of theprobability that the sum rate of NOMA attains a specific performance gainover coded multicasting. In addition, the performance of individual usersin both delivery techniques is examined by analyzing the outage probability.These analytical studies provide an understanding on which delivery tech-nique performs better under different user pairing scenarios.

Since the performance of NOMA and coded multicasting depends on thechannel gains of the paired users, it is beneficial to incorporate both deliv-ery techniques as a hybrid scheme. The hybrid scheme was studied in [115]and [116] by considering power minimization problem. In [115], the jointpower allocation and user pairing problem is decoupled into two subprob-lems. Given the user pairing solutions obtained from suboptimal algorithm,the power allocation in that work is solved by an iterative method. Mean-while, the work in [116] only consider user pairing which is solved basedon minimum weight perfect matching problem. In these works, it is demon-strated that the hybrid scheme offers improved power efficiency. Neverthe-less, the sum rate performance was not investigated in these works. In thisthesis, the sum rate performance of the hybrid scheme is compared with CIC-based NOMA and coded multicasting. For the hybrid scheme, the deliveryscheme selection, and the joint optimization of power allocation and RB allo-cation or user pairing are solved separately. The joint optimization problem isformulated with the objective of maximizing the sum rate while satisfying thetotal transmission power and QoS-based constraints. This is an MINLP prob-lem which is generally difficult to solve and hence requires high-complexitynumerical solution. Therefore, this problem must be first transformed into anon-linear programming by relaxing integer variable constraint before it canbe solved using a proposed iterative method. Meanwhile, the delivery schemeselection is performed based on a derived conditional expression.


5.3 System Model

In this chapter, a downlink wireless communication system is considered witha BS located at the centre and K cache-enabled users who are uniformly dis-tributed within a circular cell of radius RC. In addition, each user is assumedto be capable of detecting both NOMA and coded multicasting signals. TheBS delivers the content to the users with a maximum transmit power budgetof PT over a total bandwidth of BT which is divided into N RBs of equal band-width B. Each individual RB is assumed to experience frequency flat fadingsince the bandwidth B is smaller than the coherence bandwidth. However,the whole bandwidth BT is affected by frequency selective fading. Channelimpairments due to path loss and shadowing are also considered and thus,the channel gain for the k th user at the n th RB is expressed as |hk ,n |2 =

ξk|gk,n |2

PLk

where ξk is the log-normal shadowing factor for user k, |gk ,n |2 is the effectof fading, and PLk is the path loss effect experienced by user k. The setof all users at the n th RB is denoted as Kn = 1, 2, ..., K in which the theusers are arranged according to ascending order of their channel gains, i.e.,|h1 ,n |2 ≤ |h2 ,n |2 ≤ · · · ≤ |hK ,n |2.

For both NOMA and coded multicasting, it is not practical to serve allthe users in the same time-frequency resource due to the system complex-ity caused by the implementation of superposition coding and SIC/CIC inNOMA [16], and IC in coded multicasting [67]. In this work, only two usersare allowed to access each RB n since two-user pairing is the most effectiveconfiguration in terms of implementation complexity and performance gain.Here, the users who are ranked ath and bth in the set Kn (i.e., a, b ∈ Kn) arepaired together in the same RB n where a < b (i.e., |ha,n |2 ≤ |hb,n |2). In ad-dition, each content in the remote server is split into two subfiles of equalsizes. Considering that the BS has the prior knowledge of users’ requests andwithout loss of generality, it is assumed that each paired user i ∈ a, b storesthe first subfile of their requested content X(1)

i during the content placementphase. In order to fully retrieve the requested content, each user requests forthe second subfile X(2)

i during the delivery phase. As in conventional codedmulticasting, it is further assumed that each user also stores the second subfileof the unwanted content X(2)

j for j 6= i during the content placement phasein order to make both transmission techniques feasible for the detection ofthe requested subfiles. In CIC-based NOMA, the requested second subfile


X(2)i is detected by first cancelling the interference without the need for prior

detection of the interfering signal caused by the unwanted subfile X(2)j since

it is available in the cache. Meanwhile, in coded multicasting, a user can de-code the requested second subfile X(2)

i from the coded stream (which consistsof a mixture of requested and unwanted subfiles) if and only if the user pos-sesses the unwanted subfile X(2)

j in the cache. The content delivery employingNOMA and coded multicasting techniques are further explained in the fol-lowing subsections. Figure 5.1 presents the system model which include thecontent placement and delivery strategy1.

UE a

UE b

Request: 𝑋𝑎2

Request: 𝑋𝑏2

Base Station

Content Server

NOMA

power

nth RB

power

Coded Multicasting

𝒉𝟏,𝒏𝟐≤ ⋯ ≤ 𝒉𝒂,𝒏

𝟐≤ ⋯

≤ 𝒉𝒃,𝒏𝟐≤ ⋯ ≤ 𝒉𝑲,𝒏

𝟐

UE 1

UE K

. . .

. .

.

(𝑎 < 𝑏)

. . .

Paired users

𝑥𝑎2

𝑥𝑏2

𝑥𝑎⨁𝑏2

. . .

. . .

Decoder𝒙𝒊𝟐

𝒉𝒊,𝒏 𝜶𝒋≠𝒊,𝒏𝑷𝑻 𝒙𝒋≠𝒊𝟐

−

𝒙𝒊𝟏

𝒙𝒊

𝒚𝒊,𝒏 = 𝒉𝒊,𝒏 𝜶𝒊,𝒏𝑷𝑻 𝒙𝒊𝟐+ 𝜶𝒋≠𝒊,𝒏𝑷𝑻 𝒙𝒋≠𝒊

𝟐+𝒘𝒊,𝒏

𝒚𝒊,𝒏 = 𝒉𝒊,𝒏 𝜶𝒊,𝒏𝑷𝑻 𝒙𝒊𝟐+𝒘𝒊,𝒏

Receiver at user 𝒊 ∈ 𝒂, 𝒃

NOMA

Coded Multicasting

XOR Decoder

𝒙𝒊𝟏

𝒙𝒊𝟐

𝒚𝒊,𝒏 = 𝒉𝒊,𝒏 𝛀𝒏 𝑷𝑻 𝒙𝒊⨁𝒋≠𝒊𝟐

+𝒘𝒊,𝒏

𝒙𝒋≠𝒊𝟐

𝒙𝒊

𝑿𝒊𝟏𝑿𝒋≠𝒊

𝟐𝑿𝟏

𝟐 𝑿𝑲𝟐... ...

Unwanted files

𝑿𝟏𝟏

𝑿𝟏𝟐

...

...

𝑿𝒂𝟏

𝑿𝒂𝟐

...

...

𝑿𝒃𝟏

𝑿𝒃𝟐

𝑿𝑲𝟏

𝑿𝑲𝟐

...

...

. . .

. . .

nth RB

Files cached during content placementContent delivery

Figure 5.1: Content placement and delivery in NOMA and coded multicasting

5.3.1 NOMA

At the BS, superposition coding is implemented to multiplex the subfiles re-quested by the paired users (X(2)

a and X(2)b ) in the power domain. The received

signal of user i ∈ a, b at the nth RB is represented as

yi,n = hi,n

(√αa,nPTx

(2)a +

√αb,nPTx

(2)b

)+ wi,n (5.1)

1In practice, the users, who have such cache arrangements and content requests, can begrouped together to perform either CIC-based NOMA or coded multicasting.


where x(2)a and x

(2)b are the transmitted symbols of the second subfiles in-

tended for user a and b respectively, αa,n and αb,n are the power alloca-tion coefficients for user a and b respectively at RB n which must complies∑N

n=1 (αa,n + αb,n) = 1, and wi,n is the AWGN. At the receiver of each user, theinterfering signal caused by the unwanted subfile is first cancelled throughCIC by exploiting the cached data. In particular, at the receiver of user a(b), the interference from the unwanted subfile X(2)

b (X(2)a ) is cancelled using

the information available in the cache. Therefore, the residual received signalafter CIC for user a and b are respectively given by

ya,n = ha,n√αa,nPTx

(2)a + wa,n (5.2)

yb,n = hb,n√αb,nPTx

(2)b + wb,n. (5.3)

Finally, both users can directly decode the requested second subfiles fromthe residual received signal. In conventional NOMA, the weaker user cannotperform SIC and thus, the requested data is affected by interference from thestronger user. Unlike in conventional NOMA, the weaker user in CIC-basedNOMA can detect the subfile without any inter-cell interference since CIC canbe employed. For CIC-based NOMA, the achievable rates for the paired usersa and b are expressed respectively as

RNOMAa,n = B log2

(1 + ραa,n|ha,n |2

)(5.4)

RNOMAb,n = B log2

(1 + ραb,n|hb,n |2

)(5.5)

where ρ = PTBN0

is the transmit SNR and N0 is the noise spectral power density.

5.3.2 Coded Multicasting

First, the BS encode the subfiles requested by the paired users (X(2)a and X

(2)b )

into a single coded message X(2)a⊕b through XOR operation (i.e., X(2)

a ⊕X(2)b ). For

example, if user 1 is paired with user 2, then the requested second subfiles(X(2)

1 and X(2)2 ) are combined through XOR coding to generate the coded

message X(2)1⊕2. The coded message X(2)

a⊕b is then delivered through multicasttransmission to both users and thus, the received signal at each user i ∈ a, bat nth RB is given by

yi,n = hi,n√

ΩnPTx(2)a⊕b + wi,n (5.6)


where x(2)a⊕b is the transmitted symbols of the coded message X(2)

a⊕b and Ωn isthe power budget coefficient for each RB n which must complies

∑Nn=1 Ωn = 1.

Since the information of the unwanted subfile X(2)b (X(2)

a ) is available in thecache of user a (b), the requested subfile X

(2)a (X(2)

b ) can be successfully de-coded through XOR decoding. Using the earlier example of the pairing be-tween user 1 and user 2, user 1 can retrieve X(2)

1 from X(2)1⊕2 due to availability

of X(2)2 in the cache. Meanwhile, user 2 can retrieve X(2)

2 from X(2)1⊕2 since X(2)

1

is already in the cache. Due to multicasting, the individual achievable rate islimited by the channel condition of the weaker user and thus, is expressed as

RCMi,n = min

i∈a,b

B log2

(1 + ρΩn|hi ,n |2

)= B log2

(1 + ρΩn|ha,n |2

)(5.7)

since |ha,n |2 ≤ |hb,n |2.It is worth pointing that, if both CIC-based NOMA and coded multicas-

ting are not feasible for the paired users because the required subfiles arenot available in the caches (i.e., cache miss), then conventional transmissionscheme is applied to deliver the requested subfiles. Conventional transmis-sion techniques are associated with lower sum rates, for example, the sumrate of conventional NOMA is affected by the intra-cell interference. Hence,it is vital to design an effective content placement strategy which optimizethe cache hit probability (i.e., the probability that the required contents arecached) in order to exploit the sum rate performance gain offered by CIC-based NOMA or coded multicasting.

5.4 Performance Analysis of NOMA and Coded

Multicasting

The performance analysis in this section only consider the channel impair-ments caused by path loss and Rayleigh fading Therefore, the channel gainbetween the BS and the kth user is modelled as |hk |2 = |gk |2

PLkwhere gk is the

Rayleigh fading channel gain and PLk is the path loss effect for user k. Inthis work, the path loss is specifically modelled as PLk = 1 + dγk where γ isthe path loss exponent, and dk is the distance from the BS to the user k. Here,the channel gains |hk |2 are random variables which are sorted in ascendingorder i.e., |h1 |2 ≤ |h2 |2 ≤ · · · ≤ |hK |2. Therefore, it is important to obtain the


cumulative distribution function (CDF) of the ordered channel gain in orderto analyze the performance in terms the sum rates comparison and outageprobability. This CDF can be derived from the CDF of an unordered channelgain |h|2 using order statistics [117]. Nevertheless, it is difficult to determinean exact CDF expression particularly for the scenario where the users are ran-domly located in a circular cell, and the channels are affected by path losswith γ > 2 and Rayleigh fading [42]. Therefore, an approximate CDF ofthe unordered channel gain |h|2, which is derived using Gaussian Chebyshevquadrature [118], is commonly used when it is very challenging to obtain ex-act closed-form analytical expressions, such as in [42], [119] and [120]. TheCDF of the unordered channel gain |h|2 is approximated as

F|h|2 (y) ≈ 1

RC

Z∑z=1

π

Zv (θz) (5.8)

where v (θz) =√

1− θ2z (1− e−czy)

(RC

2θz + RC

2

), Z is a parameter that influ-

ence the accuracy-complexity tradeoff, cz = 1 +(RC

2θz + RC

2

)γ, and θz =

cos(

2z−1Zπ).

Based on (5.8), the approximate PDF of the unordered channel gain |h|2 isobtained in [42] as

f|h|2 (y) ≈ 1

RC

Z∑z=1

ψze−czy (5.9)

where ψz = πZ

√1− θ2

z

(RC

2θz + RC

2

)cz.

Without loss of generality, the analysis in this section focuses on a singleRB scenario (i.e., Ωn = αa,n + αb,n = 1) with unity bandwidth. Therefore, theindex for RB n and the variable for bandwidth B are removed from (5.4), (5.5)and (5.7). The sum rates of NOMA and coded multicasting can be simplifiedrespectively as

RNOMAsum = log2

(1 + ραa|ha |2

)+ log2

(1 + ραb|hb|2

)(5.10)

RCMsum = 2 log2

(1 + ρ|ha |2

). (5.11)

It is noted that, in NOMA, the total power is shared among the pairedusers according to the specified power allocation coefficients. Higher sumrates can be attained by allocating more power to user b due to the higher


channel condition. However, higher power must also be allocated to the weakuser a in order to achieve high degree of fairness. This will also ensure the rateof the weak user is comparable to the individual user rate in coded multicas-ting, at the expense of lower rate for the strong user. In coded multicasting,the entire power budget is utilized to transmit the coded message and thesum rate improves with higher total transmission power. In addition, codedmulticasting is associated with good level of fairness since the coded stream isdelivered at a single transmission rate. Nevertheless, the sum rate is restrictedby the channel condition of the weak user a.

5.4.1 Asymptotic Sum Rate Comparison

First, the performance of NOMA and coded multicasting are compared interms of the sum rates given in (5.10) and (5.11). This performance analysisinvolves the derivation of the probability that the sum rate of NOMA achievesa specific performance gain over that of coded multicasting in different userpairing scenarios, i.e.,

Pr(RNOMAsum −RCM

sum > R)

= 1− P(RNOMAsum −RCM

sum < R)

(5.12)

where R is the target sum rate performance gain. Note that, the probabil-ity that NOMA achieves better performance than coded multicasting can bedetermined by setting R = 0. It is difficult to determine an exact expres-sion of this probability and thus, asymptotic result is considered in analyzingthe sum rate gap between NOMA and coded multicasting. The asymptoticexpression of the sum rate comparison can be derived based on high SNRapproximation (ρ→∞) as

RNOMAsum −RCM

sum = log2

(1 + ραa|ha |2

)+ log2

(1 + ραb|hb|2

)− 2 log2

(1 + ρ|ha |2

)≈ log2

(ραa|ha |2

)+ log2

(ραb|hb|2

)− 2 log2

(ρ|ha |2

)= log2

(αaαb|hb|2

|ha |2

).

(5.13)


Therefore, the asymptotic probability is expressed as

Pr(RNOMAsum −RCM

sum < R)

= Pr

(log2

(αaαb|hb|2

|ha |2

)< R

)= Pr

(|ha |2

|hb|2> 2−Rαaαb

).

(5.14)The PDF for the ratio of the channel gains of two users for the case of

Rayleigh distributed channel gains can be easily found in literature, specif-ically in [45], but not for the case where users are uniformly positioned incircular cell and are affected by Rayleigh fading. Hence, the PDF for the ratioof the channel gains of two users for the latter case is derived in this chapter.

First, the joint PDF of |ha |2 and |hb|2 is determined from (5.8) and (5.9) ac-cording to [117] and then simplified using multinomial theorem. By applyingthe Mellin transforms [121] to the joint PDF, the PDF for the ratio of two orderstatistics is derived as

f |ha |2|hb |2

(y) = ωa−1∑j1=0

b−a−1∑j2=0

a− 1

j1

b− a− 1

j2

(−1)j1+j2

(1

RC

)K−a+1+j1

∑p1+···+pZ=j1−j2+b−a−1

∑q1+···+qZ=j2+K−b

∑r1+···+rZ=1

∑s1+···+sZ=1

j1 − j2 + b− a− 1

p1, · · · , pZ

j2 +K − b

q1, · · · , qZ

( Z∏z=1

(ψz)pz+qz+rz+sz

(cz)pz+qz

)(u2 + u1y)−2

(5.15)where ω = K!

(a−1)!(b−a−1)!(K−b)! , u1 =∑Z

z=1 czpz +∑Z

z=1 czrz, and u2 =∑Z

z=1 czqz +∑Zz=1 czsz. By solving the integral of (5.15), the probability that NOMA attains

a performance gain over coded multicasting is given by


Pr(RNOMAsum −RCM

sum > R)

= 1− Pr(RNOMAsum −RCM

sum < R)

= 1−ˆ 1

2−Rαaαb

f |ha |2|hb |2

(y) dy

= 1−

ω a−1∑j1=0

b−a−1∑j2=0

a− 1

j1

b− a− 1

j2

(−1)j1+j2

(1

RC

)K−a+1+j1

∑p1+···+pZ=j1−j2+b−a−1

∑q1+···+qZ=j2+K−b

∑r1+···+rZ=1

∑s1+···+sZ=1

j1 − j2 + b− a− 1

p1, · · · , pZ

j2 +K − b

q1, · · · , qZ

( Z∏z=1

(ψz)pz+qz+rz+sz

(cz)pz+qz

)× 1

u1

(1

u2 + (2−Rαaαb)u1

− 1

u2 + u1

) .(5.16)

The expression in (5.16) gives a rough idea on the sum rate performance at-tained by NOMA over coded multicasting based on different pairings of usera and b for a, b ∈ Kn and a < b. Therefore, the analytical result, which willbe compared with simulation result in subsection 5.4.3, provides an under-standing on which delivery technique performs better under different pairingscenarios. In subsection 5.4.3, it will be proved that NOMA performs betterthan coded multicasting when the paired users’ channel qualities are highlydistinctive i.e., b a. However, coded multicasting is favoured when thepaired users’ channel conditions are relatively similar.

It is noted that (5.16) is not a function of transmit SNR ρ and thus onlygives asymptotic result. Moreover, (5.16) is expected to be accurate only atlarger SNR due to high SNR approximation in its derivation. The simulationresults in subsection 5.4.3 will verify the accuracy of (5.16). Nevertheless, theexpression is accurate at all SNR values for a special case when R = 0 and thepower allocation coefficient for the weaker user a is very high (i.e., αa → 1).This can be proved mathematically as follows.

Proof: For this special case (i.e., R = 0 and αa → 1), the expression of thesum rate comparison can be simplified as


RNOMAsum −RCM

sum = log2

(1 + ραa|ha |2

)+ log2

(1 + ραb|hb|2

)− 2 log2

(1 + ρ|ha |2

)≈ log2

(1 + ρ|ha |2

)+ log2

(1 + ραb|hb|2

)− 2 log2

(1 + ρ|ha |2

)= log2

((1 + ραb|hb|2

)(1 + ρ|ha |2

) ) .(5.17)

Without high SNR approximation, the exact probability is represented as

Pr(RNOMAsum −RCM

sum < 0)

= Pr

(log2

((1 + ραb|hb|2

)(1 + ρ|ha |2

) ) < 0

)= Pr

(|ha |2

|hb|2> αb

).

(5.18)It is noted that (5.18) is in a similar form to (5.14) when R = 0 and αa → 1,

i.e., 2−Rαaαb ≈ αb. Therefore, (5.18) can be directly solved by applying the theanalytical expression in (5.16). Since the derivation of (5.18) does not involvehigh SNR approximation (i.e., (5.18) is an exact probability), it is proved thatthe analytical expression (5.16) is accurate for all SNRs for the special case ofR = 0 and αa → 1. The accuracy of (5.16) for the special case (i.e., R = 0 andαa → 1) will be verified by comparing both analytical and simulation resultsin subsection 5.4.3.

5.4.2 Outage Performance

For NOMA, the outage probability for each user i ∈ a, b is obtained byusing order statistics as

PrNOMAi = Pr

(RNOMAi < Ri

)= Pr

(|hi |2 <

2Φ(i)min − 1

ραi

)

=

2Φ

(i)min−1ραiˆ

0

K!(F|h|2 (x)

)i−1 (1− F|h|2 (x)

)K−if|h|2 (x)

(i− 1)!(K − i)!dx

(5.19)

where Φ(i)min is the minimum rate requirement for each user i ∈ a, b. It is diffi-

cult to solve the integral in (5.19) based on the CDF and PDF of the unorderedchannel gains in (5.8) and (5.9) respectively. Therefore, the CDF and PDF isfurther simplified using Taylor’s approximation as F|h|2 (y) ≈ 1

RC

∑Zz=1 ψzy


and f|h|2 (y) ≈ 1RC

∑Zz=1 ψz (1− czy) respectively. Following the steps in [42],

the outage probability for each user i ∈ a, b is derived based on high SNRapproximation as

PrNOMAi ≈ ζi

2Φ

(i)min−1ραiˆ

0

(ηx)i−1 (1− ηx)K−i1

RC

Z∑z=1

ψz (1− czy) dx

≈ ζiiηi

(2Φ

(i)min − 1

ραi

)i

(5.20)

where η = 1RC

∑Zz=1 ψz and ζi = K!

(i−1)!(K−i)! . From (5.20), the outage probabilitydecreases with increasing transmit SNR ρ. In addition, the channel ranks(user index) of the paired users also influence the outage performance. Inparticular, the higher ranked strong user b tends to outperform weak usera (i.e., a < b) due to better channel condition (i.e., |ha |2 ≤ |hb|2). However,the outage performance also depends on the power allocation coefficient αi.The performance of the weaker user a can be improved with higher powerallocation coefficient and thus enhancing the user fairness.

Using similar steps as above, the outage probability for both users in codedmulticasting is estimated as

PrCMi ≈ ζaaηa

(2Φ

(a)min − 1

ρ

)a

. (5.21)

It is noted from (5.20) and (5.21) that the users in coded multicasting al-ways outperforms the weaker user a in NOMA since αa ≤ 1. The outageperformance for user a is comparable to that of the users in coded multicast-ing only when the allocated power coefficient αa is closer to 1. On the otherhand, the stronger user b in NOMA may achieve significant performance gainover the the users in coded multicasting particularly when the index of theuser b is much greater than a i.e. b a. This indicates that NOMA favoursthe situation when the channel gains between the paired users are highlydistinctive as will be shown in the next subsection.


5.4.3 Numerical Results

The simulation results in this subsection are demonstrated to verify the accu-racy of the analytical expressions derived in subsections 5.4.1 and 5.4.2, andto compare the performance of NOMA and coded multicasting under differ-ent pairing scenarios. Here, the simulation setup considers a total of K = 5

users uniformly distributed in a cell of radius of RC = 100m. Furthermore,the path loss effect with exponent of γ = 3 and Rayleigh fading are consid-ered. For NOMA, the power allocation coefficients of αa = 0.8 and αb = 0.2

are applied for weak and strong users respectively. The accuracy parameterfor Chebyshev-Gauss quadrature Z is set as 10.

Figure 5.2 presents the probability that the sum rate of NOMA is betterthan that coded multicasting by a gain of R bit per channel use (BPCU) con-sidering various pairing scenarios. First, the simulation results proved theaccuracy of the analytical expression for sum rate comparison derived in sub-section 5.4.1. In particular, the analytical results are accurate for the specialcase when R = 0 and the power allocation factor for the weak user is closeto 1 (i.e., αa → 1) as mathematically proven in subsection 5.4.1. However, theanalytical results for the non-zero R cases are only accurate at higher transmitSNR because the derivation of (5.16) involves high SNR approximation.

The probability that NOMA performs better than coded multicasting canbe assessed by setting R = 0. NOMA tends to outperform coded multicastingwhen the channel gain difference between the paired users is large. This canbe observed in the case when the strongest user user b = 5 is paired withthe weakest user a = 1 in Figure 5.2(a). In coded multicasting, the individualrates of the paired users in coded multicasting are restricted by the user withthe weakest channel condition and hence does not significantly improve theoverall sum rate. In NOMA, the weak user a can achieves slightly lower thanthe individual rate of coded multicasting by allocating most of the powerto this user (i.e., αa = 0.8). Therefore, the sum rate performance heavilydepends on the rate of the strong user b who possess excellent channel quality.However, for the case of user b = 3 paired with user a = 1 in Figure 5.2(b),it is shown that NOMA may not perform better than coded multicasting dueto the channel condition of the strong user b. As the rank between the pairedusers is closer to each other (i.e., the channel quality of these users becomerelatively similar), coded multicasting tends to perform better than NOMA,


for example, when user b = 5 is paired with user a = 4 as presented in Figure5.2(a). Coded multicasting achieves higher performance gain over NOMAbecause the whole total transmission power is utilized while the channel gainof the weak user a is comparable to that of the strong user b. Meanwhile, inNOMA, the total power is shared among both users who have similar channelgains and hence does not achieve higher performance gain.

25 30 35 40 45 50 55 60 65 70 75Transmit SNR, ρ (dB)

0

0.2

0.4

0.6

0.8

1

P(R

sum

NO

MA -

Rsu

mC

M >

R) Analytical, R = 0

Simulation, R = 0Analytical, R = 0.25 BPCUSimulation, R = 0.25 BPCUAnalytical, R = 0.5 BPCUSimulation, R = 0.5 BPCU

a = 1

a = 4

(a)

25 30 35 40 45 50 55 60 65 70 75Transmit SNR, ρ (dB)

0

0.2

0.4

0.6

0.8

1

P(R

sum

NO

MA -

Rsu

mC

M >

R)

Analytical, R = 0Simulation, R = 0Analytical, R = 0.25 BPCUSimulation, R = 0.25 BPCUAnalytical, R = 0.5 BPCUSimulation, R = 0.5 BPCU

a = 1

a = 2

(b)

Figure 5.2: Probability that the sum rate difference between NOMA and codedmulticasting is greater than R (a) when b = 5 is paired with a = 1, and a = 4,and (b) when b = 3 is paired with a = 1, and a = 2,


Figure 5.2 also demonstrate the probability that NOMA achieves a perfor-mance gain of R = 0.25BPCU and R = 0.5BPCU . From this figure, it canbe observed that the probability of NOMA outperforming coded multicast-ing drops with higher R because it is difficult for NOMA to achieve morestringent performance target. At lower SNR, the probability drops substan-tially since it becomes more difficult for NOMA to attain performance gainof 0.25BPCU and 0.5BPCU over coded multicasting with low transmissionpower. Nevertheless, NOMA can only achieve significant performance gainover coded multicasting when transmission power is high since the proba-bility is similar to that of R = 0 case at high SNR. The performance gain ofNOMA is dominated by the rate of the strong user which improve substan-tially with higher SNR ρ and excellent channel quality |hb|2 according to (5.5).

0.5 0.6 0.7 0.8 0.9 1α

a

0

0.2

0.4

0.6

0.8

1

P(R

sum

NO

MA >

Rsu

mC

M )

AnalyticalSimulation, ρ = 40 dBSimulation, ρ = 60 dBSimulation, ρ = 75 dB

a = 3

a = 4

a = 1

Figure 5.3: Probability that NOMA performs better than coded multicasting(i.e., R = 0) against the power allocation coefficient of the weak user αa whenb = 5 is paired with a = 1, a = 3 and a = 4,

The impact of the power allocation coefficient on the performance ofNOMA over coded multicasting is presented in Figure 5.3. From this fig-ure, it is shown that the analytical results become more accurate as the powercoefficient for the weak user αa is higher than 0.8 proving the accuracy in thecases with the common power allocation coefficients for NOMA at all SNRs.


In terms of sum rate performance, NOMA tends to offer significant improve-ment when more power is allocated to the strong user. This is because thestrong user rate dominates the sum rate performance of NOMA. However,the user fairness performance will be reduced since the rate for the the weakuser is lower. Furthermore, the results demonstrate the superiority of NOMAover coded multicasting when the strongest user b = 5 paired is paired withthe weakest user a = 1, but otherwise when paired with the next strongestuser a = 4. This shows that the performance gain of NOMA can be enhancedwhen the index of user b is greater than a (i.e., b a) which implies channelgains of the paired users are highly distinctive

Figure 5.4 presents the outage performance of each user in NOMA andcoded multicasting for the two extreme cases where the minimum individualtarget rate Φ

(i)min for each user i ∈ a, b is set as 0.25BPCU . Here, only two

extremes user pairing cases are discussed. The first case is the situation inwhich the users with the best channel condition (b = K = 5) and with theworst channel (a = 1) are paired together. The second one is the scenario inwhich a user (b = 5) is paired with the nearest ordered user (a = 4). The out-age probability results for the first and second extreme cases are separatelyshown in Figures 5.4(a) and 5.4(b) respectively. First, the results show that theanalytical expressions in (5.20) and (5.21) are closer to the simulated results athigh SNR because these expressions are derived based on high SNR approxi-mation. For the first case presented in Figure 5.4(a), the strong user (b = 5) inNOMA achieves superior outage performance due to excellent channel qual-ity. Meanwhile the performance of the weak user (a = 1) is close to that ofthe users in coded multicasting since most of the power is allocated to theweak user. However, the users in coded multicasting outperform both usersin NOMA for the second case as shown in Figure 5.4(b). Hence, coded multi-casting is preferred when the users with relatively similar channel conditionsare paired together. Note that, in Figure 5.4(b), the weak user performs betterthan the strong user for NOMA because higher power is allocated to weakuser while having almost similar channel quality to the strong user.


25 30 35 40 45 50 55 60 65 70 75Transmit SNR, ρ (dB)

10-4

10-3

10-2

10-1

100

Out

age

perf

orm

ance

NOMA - user a = 1 (Analytical)NOMA - user a = 1 (Simulation)NOMA - user b = 5 (Analytical)NOMA - user b = 5 (Simulation)Coded Multicasting (Analytical)Coded Multicasting (Simulation)

(a)

25 30 35 40 45 50 55 60 65 70 75Transmit SNR, ρ (dB)

10-4

10-3

10-2

10-1

100

Out

age

perf

orm

ance

NOMA - user a = 4 (Analytical)NOMA - user a = 4 (Simulation)NOMA - user b = 5 (Analytical)NOMA - user b = 5 (Simulation)Coded Multicasting (Analytical)Coded Multicasting (Simulation)

(b)

Figure 5.4: Outage probability of NOMA and coded multicasting as a functionof transmit SNR with Φ

(a)min = Φ

(b)min = 0.25BPCU when b = 5 is paired with

(a) a = 1, and (b) a = 4


5.5 Hybrid NOMA and Coded Multicasting

Scheme

In cache-aided UE systems, both NOMA and coded multicasting are consid-ered as effective techniques in delivering contents to multiple users in thesame time/code/frequency resource and hence offer enhanced spectral effi-ciency. However, the performance of NOMA is affected due to the sharingof power among the paired users. Significant performance gain can only beachieved in NOMA when the channel gains of the paired users are highly dis-tinctive. Meanwhile, the performance of coded multicasting is restricted bythe weakest user. Hence, in coded multicasting, it is only beneficial to pair theusers whose channel gain are similar. In order to exploit the benefits of bothNOMA and coded multicasting, a hybrid scheme which incorporate both de-livery techniques is proposed. This hybrid scheme selects either NOMA orcoded multicasting as a transmission mode for each RB depending on thechannel gains of the paired users. The hybrid NOMA and coded multicas-ting technique is expected to offer improved performance particularly if thechannel gain gaps between the paired users varies across the whole systembandwidth.

The hybrid strategy consists of two stages which are NOMA-based jointpower and RB allocation or user pairing, and scheme selection. In the firststage, joint power and RB allocation or user pairing for NOMA is performedwith the aim of maximizing the sum rate while satisfying the maximumpower budget and QoS constraints. The optimization problem is an MINLPproblem which can be transformed to nonlinear programming by relaxing theinteger variables constraint. Note that, for coded multicasting, the allocatedpower for each RB is obtained as the sum of paired users’ power calculatedfrom the NOMA-based power allocation. Therefore, a specific power alloca-tion scheme for coded multicasting is not needed. The second stage involvesthe selection of either NOMA or coded multicasting for each RB. This stageis described first in the next subsection by assuming that the joint power andRB allocation has already been solved.


5.5.1 Scheme Selection

The selection of transmission mode (NOMA or coded multicasting) is per-formed by comparing the achievable sum rates between NOMA and codedmulticasting for each RB using (5.4), (5.5) and (5.7). The condition in whichthe sum rate of NOMA is better than coded multicasting in each RB n isrepresented as

B log2

(1 + ραa,n|ha,n |2

)+B log2

(1 + ραb,n|hb,n |2

)≥ 2B log2

(1 + ρΩn|ha,n |2

)(5.22)

where user a and b are paired together at RB n based on the joint powerand RB allocation or user pairing proposed in subsection 5.5.2, and the RBpower budget coefficient should satisfy Ωn = αa,n + αb,n. Rearranging (5.22),the ratio of the channel gains of the paired users must satisfy the followingcondition in order to ensure the sum rate of NOMA is better than that ofcoded multicasting at RB n

|hb,n |2

|ha,n |2≥ 2Ωn + ρΩ2

n|ha,n |2 − αa,n − ραa,nαb,n|hb,n |2

αb,n. (5.23)

Algorithm 5.1 Scheme Selection Algorithm1: Initialization: Joint power RB allocation as described in subsection 5.5.2

2: for n = 1 to N3: if condition (5.23) is satisfied

4: NOMA is selected for RB n

5: Apply αa,n and αb,n initialized in step (1)

6: else7: Coded multicasting is selected for RB n

8: Calculate Ωn = αa,n + αb,n

9: end if10: end for

Note that the condition in (5.23) is in terms of the power coefficients ofthe paired users αa,n and αb,n. Therefore, the power coefficients must be op-timized first prior to transmission scheme selection. The scheme selectionprocess for the proposed hybrid technique is summarized in Algorithm 5.1.For each RB n, condition (5.23) is applied in order to select transmission mode


that can offer improved sum rate. If condition (5.23) is satisfied at RB n, theBS will select NOMA as the transmission mode with pre-determined powercoefficients αa,n and αb,n. Otherwise, coded multicasting is selected as trans-mission mode in RB n and the power budget coefficient is calculated as thesum of the power coefficients for NOMA in that RB (i.e., Ωn = αa,n + αb,n).

5.5.2 Near-Optimal Joint Power and RB Allocation or User

Pairing for NOMA

In order to solve the joint problem in NOMA, a variable βk,n is introduced torepresent the RB allocation indicator, which is set to 1 if RB n is allocated touser k and 0 otherwise. Therefore, in this work, the sum rate is maximized byjointly optimizing the power coefficient αk,n and the RB allocation indicatorβk,n. For resource allocation, the sum rate for NOMA is represented as

RNOMAsum =

K∑k=1

N∑n=1

βk,nB log2

(1 + ραk,n|hk ,n |2

). (5.24)

The joint power and RB allocation or user pairing optimization problem,which is aimed towards maximizing the sum rate while achieving the totalpower budget and minimum rate constraints, is formulated as follows

maximizeαk,n,βk,n

RNOMAsum =

K∑k=1

N∑n=1

βk,nB log2

(1 + ραk,n|hk ,n |2

)(5.25a)

subject to αk,n ≥ 0, ∀k, ∀n (5.25b)K∑k=1

N∑n=1

αk,n ≤ 1 (5.25c)

N∑n=1

βk,nB log2

(1 + ραk,n|hk ,n |2

)≥ Φ

(k)min, ∀k (5.25d)

βk,n ∈ 0, 1 , ∀k, ∀n (5.25e)K∑k=1

βk,n ≤ 2, ∀n (5.25f)

N∑n=1

βk,n ≥ Q, ∀k (5.25g)


where constraints (5.25b) and (5.25c) are non-negative power coefficients andtotal power coefficient constraint respectively. Constraint (5.25d) ensures thatthe minimum target rate Φ

(k)min is achieved by each user k. Constraint (5.25e)

restrict βk,n to integer values of either 0 or 1. Meanwhile, constraint (5.25f)ensures that each RB can only be accessed by a maximum of two users andconstraint (5.25g) guarantees that each user is being allocated with at least Qnumber of RBs.

Problem (5.25) is an MINLP problem which are generally solved by numer-ical tool with high complexity. Nevertheless, this problem can be transformedinto non-linear programming problem by relaxing the integer variables inconstraint (5.25e) into continuous variables, i.e., βk,n ∈ [0, 1] [122]. Accordingto [93], a problem which is similar to the objective function (5.25a) with con-straints (5.25b) and (5.25c) is a convex optimization problem. Furthermore,the nonlinear inequality constraint (5.25d) makes the feasible set convex sincethere are no intra-cell interference terms, and the remaining constraints areobviously affine. Thus, the problem (5.25) with relaxed constraint (5.25e) is aconvex optimization problem. This problem can be solved by employing theLDD approach in [91]. The Lagrangian function of the relaxed optimizationproblem (5.25), is given by

L(αk,n, βk,n, µ, τk, λk,n, φk) =K∑k=1

N∑n=1

βk,nB log2

(1 + ραk,n|hk ,n |2

)− µ

(K∑k=1

N∑n=1

αk,n − 1

)−

K∑k=1

τk

[Φ

(k)min −

N∑n=1

βk,nB log2

(1 + ραk,n|hk ,n |2

)]

−K∑k=1

N∑n=1

λk,n

(βk,n − 1

)−

K∑k=1

φk

(Q− βk,n

)(5.26)

where µ, τk, and φk are the Lagrange multipliers for constraints (5.25c), (5.25d)and (5.25g) respectively, while λk,n is the multiplier for relaxed constraint(5.25e). The relaxed problem (5.25) is then transformed into a dual problemwhich is expressed as

maximizeµ,τk,λk,n,φk

D(µ, τk, λk,n, φk) = infαa,n,βk,n

L(αa,n, βk,n, µ, τk, λk,n, φk) (5.27a)

subject to µ, τk, λk,n, φk ≥ 0. (5.27b)


The first order derivatives of (5.26) w.r.t. αk,n and βk,n are determinedrespectively as

dL(αk,n, βk,n, µ, τk, λk,n, φk)

dαk,n=

B

ln 2

[(1 + τk) βk,nρ|hk ,n |2

1 + ραk,n|hk ,n |2

]− µ (5.28)

dL(αk,n, βk,n, µ, τk, λk,n, φk)

dβk,n= (1 + τk)B log2

(1 + ραk,n|hk ,n |2

)− λk,n + φk.

(5.29)By setting (5.28) equals to zero, the power allocation coefficient for user k

at RB n is derived as

αk,n =

[(1 + τk) βk,nB

µ ln 2− 1

ρ|hk ,n |2

]+

(5.30)

where [x]+ = max (x, 0) which ensures constraint (5.25b) is satisfied. Similarly,when (5.29) equals to zero, λk,n is represented as

λk,n = (1 + τk)B log2

(1 + ραk,n|hk ,n |2

)+ φk. (5.31)

Since only two users are allowed to access the same RB according to con-straint (5.25f), the two users who gives the two highest values of λk,n at eachRB n are chosen as a pair. Hence, at each RB n, the RB allocation indicatorβk,n is set to 1 for the two users with the highest λk,n and 0 for the remainingusers.

It is noted from (5.31) that λk,n increases with higher channel gain |hk ,n |2.Hence, selecting users with the best channel gains will guarantee higher sumrate. However, the users with poor channel conditions may not be selectedin any RB. Therefore, constraint (5.25g) is necessary to ensure all users areallocated at least a number of RBs particularly those with poor channel gains.Note that the values of λk,n are also influenced by the multiplier related to theconstraint (5.25g) (i.e., φk) which can be increased to ensure the weak usersare served by a number of RBs.

The dual variables µ, τk, and φk in (5.30) and (5.31) are then solved byusing the iterative subgradient method. The subgradient updating equations,


which are derived according to [95], are given by

µ(t+ 1) =

[µ(t) + δµ

(K∑k=1

N∑n=1

αk,n − 1

)]+

(5.32)

τk(t+ 1) =

[τk(t) + δτk

(Φ

(k)min −

N∑n=1

βk,nB log2

(1 + ραk,n|hk ,n |2

))]+

(5.33)

φk(t+ 1) =[φk(t) + δφk

(Q− βk,n

)]+

. (5.34)

The joint power and RB allocation or user pairing algorithm is summa-rized in Algorithm 5.2. The iteration terminates when the dual variables µ, τk,and φk converge (i.e., when the difference between the dual variables in thecurrent iteration and in the previous iteration is less than a tolerance value ε)or when the iteration index t reaches the maximum iterations Tmax. In thisalgorithm, the diminishing step size δ = a/

√t or δ = a/

(b+ t

)is imple-

mented to guarantee faster convergence and near-optimal solutions, where aand b are fixed non-negative values [95]. It will be shown in subsection 5.5.5that this algorithm is near-optimal by comparing the simulation result withthe optimal solution obtained using numerical optimization tool.

Algorithm 5.2 Joint Power and RB Allocation or User Pairing Algorithm1: Initialization: set t = 0, ε and Tmax, initialize µ(0), τk(0) and φk(0), ∀k2: while (|µ(t− 1)− µ(t)| ≥ ε or |τk(t− 1)− τk(t)| ≥ ε or |φk(t− 1)− φk(t)| ≥ ε)

and t ≤ Tmax do3: Find αk,n using (5.30) assuming βk,n = 1, ∀k, ∀n4: Solve λk,n using (5.31)

5: Obtain βk,n =

1, if k corresponds to the highest two λk,n atRB n

0, otherwise

6: Solve αk,n using (5.30) based on RB allocation in Step 5

7: Update the Lagrange multipliers µ, τk, and φk according to (5.32), (5.33)

and (5.34) respectively

8: t← t+ 1

9: end while10: output the optimal solutions αk,n and βk,n

Note that, for the two-user case (i.e., K = 2), RB allocation or user pairing


is not needed since two users are allowed to access the same RB. Therefore,for this case, step 3-5 in Algorithm 5.2 and updating equation (5.34) can beneglected to solve the power allocation problem.

5.5.3 Low-complexity Joint Power and RB Allocation or User

Pairing Scheme

The joint power and RB allocation or user pairing scheme proposed in sub-section 5.5.2 iteratively calculates several variables, such as λk,n, and updatingequations for Lagrange multipliers, such as φk, in order to obtain near-optimalsolutions. The computational complexity can be reduced by decoupling theproblem into RB allocation or user pairing and power allocation subproblems.First, a low-complexity RB allocation or user pairing is investigated. Based onthe near-optimal solution in (5.31), the users with the best channel gains areselected as pairs if the Lagrange multipliers φk and τk is neglected. Withoutapplying the near-optimal scheme, two users with the best channel gains ineach RB can be chosen to maximize the overall sum rate. However, users withworse channel conditions may not be served by sufficient number of RBs toachieve the minimum target rate. Therefore, a 2-stage low complexity RB al-location algorithm is proposed to maximize the sum rate while ensuring allusers are served by a number of RBs. The proposed low-complexity schemeis presented in Algorithm 5.3.

In the first stage, two users with the best channel gains in each RB arepaired together. The second stage ensures all users are served by allowing theusers with the better channel gains to iteratively give up some RBs to otherweaker users until constraint (5.25g) is satisfied for all the users. First, eachuser k identifies the potential RBs to be given up in a set denoted as Bk =n : 1 ≤ n ≤ N, βk,n = 1,

∑Nn=1 βk,n > Q

. In particular, the users with Q RBs

or less will not give up any RBs in each iteration and hence the set of RBs to begiven up Bk will be null for these users. Let the set of RBs to be given up by allusers be B = B1, . . . ,BK. Then, an RB with the least channel gain in the setB is selected to be given up and identified as RB n′ which currently belongs touser k′. A set of potential users who require RB n′ is then formed and denotedas Kn′ =

k : 1 ≤ k ≤ K, k 6= k′, βk,n′ = 0,

∑Nn=1 βk,n < Q

. Specifically, the

set Kn′ consists of the users who do not possess sufficient number of RBs


according to constraint (5.25g). Hence, the users who met this consttraint willnot be included in the set Kn′ . The user with the largest channel gain in theset Kn′ will then be allocated with RB n′. Let this user be identified as user k∗.The RB allocation indicator is updated as βk∗,n′ = 1 and βk′,n′ = 0. Both userk∗ and k′ will no longer involve in the swapping of RB n′ and hence excludedfrom the sets B and Kn′ respectively in following iterations. Note that, in thefirst stage, two users with the highest channel gains are paired together in RBn′ including user k′. Another paired user in RB n′ (not user k′) will not beexcluded from the set B in the current iteration and therefore may give upthe RB to another user in the subsequent iterations.

After solving the RB allocation subproblem, the power allocation can beobtained using the subgradient method in Algorithm 5.2 at a much lowercomplexity since step 3-5 and updating equation (5.34) can be removed.

Algorithm 5.3 Low-complexity RB Allocation or User Pairing Algorithm

1: Initialize βk,n =

1, if k corresponds to the highest two |hk ,n |2 atRB n

0, otherwise

2: while∑N

n=1 βk,n < Q, ∀k do3: Construct/update a set of RBs to be given up, B = B1, . . . ,BK,

where Bk =n : 1 ≤ n ≤ N, βk,n = 1,

∑Nn=1 βk,n > Q

.

4: Select an RB with the lowest |hk ,n |2 in B and let this be RB n′ belongs to user k′.

5: Construct/update a set of users who need RB n′,

Kn′ =k : 1 ≤ k ≤ K, k 6= k′, βk,n′ = 0,

∑Nn=1 βk,n < Q

.

6: Select a user with the highest |hk ,n |2 in Kn′ and let this be user k∗ at RB n′.

7: RB n′ is allocated to user k∗and updates βk∗,n′ = 1.

8: User k′ has given up RB n′ and updates βk′,n′ = 0.

9: Exclude users k∗ and k′ at RB n′ from B and Kn′ respectively in the next

iterations.

10: end while11: output the solution βk,n

5.5.4 Computational Complexities

The computational complexity of the subgradient method depends on thenumber of optimizing variables and the strictness of the stopping criteriawhich is defined by a tolerance value ε. The complexity of the near-optimal


joint power and RB allocation problem is O(

2NKε2

)since both αk,n and βk,n

are optimized. In the low-complexity scheme, only αk,n is solved using thesubgradient method. The optimizing variable βk,n is solved by examiningthe channel gains without any computations. Hence, the complexity of thelow-complexity scheme is O

(NKε2

). On the other hand, the scheme selection

algorithm needs N times of computations for (5.23).

5.5.5 Simulation Results

The simulation results in this subsection are presented to compare the per-formance of the proposed hybrid scheme with CIC-based NOMA and codedmulticasting when joint power and RB allocation or user pairing optimizationis performed. The simulation considers a downlink system that consists ofK cache-enabled users uniformly distributed within a circular cell of radiusRC = 200 m with a BS located at the centre. Unless stated otherwise, the pa-rameters associated with the effects of path loss, shadowing effect, noise andfading considered in this subsection are summarized in Table 5.1. For theNOMA-based FPA scheme, the power allocation of 0.8 : 0.2 is employed.



No. of RB, N 10

Total Bandwidth, BT 2 MHz

Cell radius 200 m







Frequency Selective Fading ITU Pedestrian B [27]


10 12 14 16 18 20 22 24 26 28 30Total Power (dBm)

20

25

30

35

40

45

50

Sum

Rat

e (M

bps)

Coded MulticastingNOMA: FPANOMA: Minimum rate constraint (Optimal)NOMA: Minimum rate constraint (Near-optimal)NOMA: Without QoS constraintHybrid: FPAHybrid: Minimum rate constraint (Optimal)Hybrid: Minimum rate constraint (Near-optimal)Hybrid: Without QoS constraint

17.8 18 18.2

29.5

30

Figure 5.5: Sum rate performance of NOMA, coded multicasting and hybridNOMA-coded multicasting versus total transmission power

First, the performance of the proposed hybrid technique, CIC-basedNOMA and coded multicasting are investigated based on K = 2 users case inwhich RB allocation or user pairing is not required. For coded multicasting,conventional water-filling approach is applied to optimize the power allocatedfor each RB. Meanwhile, for NOMA and hybrid scheme, the performance arealso compared based on various power allocation schemes including FPA, op-timal QoS-based scheme (i.e., with minimum rate constraint) using Matlab’snumerical optimization tool, near-optimal QoS-based scheme in subsection5.5.2 and the one without any QoS constraint. For the latter scheme, the al-gorithm in subsection 5.5.2 is applied without considering the QoS-relatedsubgradient updating equations. In this simulation, the minimum target ratesfor the weak and strong users are set as 2.5Mbps and 5Mbps respectively. Thesum rate performance of the proposed hybrid scheme is compared to NOMAand coded multicasting against the total transmission power as presentedin Figure 5.5. From this figure, coded multicasting is being outperformed byNOMA for all considered power allocation schemes. This is because the chan-nel gains of the paired users are more likely to be distinctive due to the effect


of path loss in large cell size (i.e., RC = 200 m), shadowing and frequency se-lective fading. Nevertheless, there are also cases that favours coded multicast-ing, particularly when the channel gains of the paired users are almost similar.Therefore, it is beneficial to incorporate both NOMA and coded multicastingas a hybrid scheme in order to improve the sum rate performance as shown inFigure 5.5. In terms of the performance based on power allocation schemes,the one without QoS constraint achieve the best performance, but does notnecessarily satisfy the minimum rate requirement. In order to maximize thesum rate, more power is allocated to the strong user and therefore the targetrate of the weak users may not always be achieved. On the other hand, theQoS-based scheme ensures the minimum target rate is met, particularly forthe weak user, while experiencing only slight performance degradation. Notethat the near-optimal QoS-based scheme performs very closely to the optimalone which demonstrate that the scheme proposed in subsection 5.5.2 is near-optimal. Therefore, only the near-optimal scheme is considered thereafter dueto the high computational complexity of optimal scheme. Overall, the hybridscheme still offer improved performance gain over NOMA in all power alloca-tion schemes despite that the paired users’ channel gains are more likely to bedistinctive due to the large cell size (RC = 200 m). The performance gain canbe further improved when considering smaller cell size and lower shadowingeffect, which increase the likelihood of the paired users having similar chan-nel gains. This can be demonstrated in Figure 5.6 when the cell size is reducedfrom 200 m to 50 m and standard deviation for shadowing is decreased from8 dB to 4 dB. The increased implementation of coded multicasting in someRBs enhances the sum rate performance gain of the hybrid scheme. Note alsothat the performance of coded multicasting is much closer to NOMA whensmaller cell size and lower shadowing effect are considered. This indicatesthat the situations that favour coded multicasting have increased.


17 18 19 20 21 22 23Total Power (dBm)

25

30

35

40

45

50

55

Sum

Rat

e (M

bps)

Coded MulticastingNOMAHybrid

RC

= 50 m, = 4 dB

RC

= 200 m, = 8 dB

Figure 5.6: Sum rate performance of NOMA, coded multicasting and hybridNOMA-coded multicasting forRC = 200 m (σ=8 dB) andRC = 50 m (σ=4 dB)

Figure 5.7 presents the fairness performance of NOMA, coded multicast-

ing and hybrid scheme in terms of Jain’s index J =(∑Kk=1Rk)

2

K∑Kk=1 R

2k

, where Rk isthe transmission rate of each user k and the fairness index must be within therange 0 ≤ J ≤ 1 [96]. Coded multicasting achieves the highest level of userfairness as shown in Figure 5.7 because the coded stream is transmitted to allusers at a single rate. Meanwhile, hybrid scheme offers improved fairness ascompared to NOMA since coded multicasting can be implemented in someRBs. The result in Figure 5.7 generally demonstrates the trade-off betweensum rate performance and user fairness requirement. Despite the significantsum rate performance offered by the power allocation schemes with mini-mum rate and without the QoS constraint, both schemes do not guaranteegood level of fairness. Note that, the QoS-based scheme offer better fairnessthan the one without QoS at lower transmission power. This is because, inQoS-based scheme, higher power is allocated to the weak user to satisfy theminimum target rate. On the other hand, better fairness is achieved by FPAsince more power is allocated to the weak user (i.e., the power allocation factorof 0.8) as compared to the strong user.


10 12 14 16 18 20 22 24 26 28 30Total Power (dBm)

0.85

0.9

0.95

1

Fai

rnes

s In

dex

Coded MulticastingNOMA: FPANOMA: Minimum rate constraintNOMA: Without QoS constraintHybrid: Minimum rate constraintHybrid: Without QoS constraint

Figure 5.7: Fairness index for NOMA, coded multicasting and hybrid NOMA-coded multicasting case versus total transmission power

10 12 14 16 18 20 22 24 26 28 30Total Power (dBm)

20

25

30

35

40

45

50

Sum

Rat

e (M

bps)

NOMA: Near-optimal schemeNOMA: Low complexity schemeNOMA: Random RB allocationHybrid: Near-optimal schemeHybrid: Low complexity schemeHybrid: Random RB allocation

Figure 5.8: Sum rate performance of different resource allocation schemes inNOMA and coded multicasting versus total transmission power

Next, the performance of various RB allocation or user pairing schemes areassessed by considering the simulation for K = 4 users. Since all the 4 usersneed to share only 10 RBs, the minimum target rate is reduced to 1Mbps for allusers. In addition, each user must be served by at least Q = 4 RBs in order toguarantee that the minimum target rate is satisfied. Figure 5.8 demonstrates


the sum rate performance of the near-optimal and low-complexity schemesproposed in subsections 5.5.2 and 5.5.3 respectively which are compared torandom RB allocation or user pairing scheme. The latter scheme allocates theRB randomly to each user while ensuring constraint (5.25f) is satisfied. Notethat the QoS-based power allocation is also employed to both low-complexityand random schemes. From Figure 5.8, it is shown that both near-optimaland low-complexity schemes offer significant performance gain over the ran-dom scheme. The low-complexity scheme performs very closely to the near-optimal scheme at a much reduced complexity.

2 4 6 8 10 12 14 16 18iteration

21

22

23

24

Sum

Rat

e (M

bps)

(a)

2 4 6 8 10 12 14 16 18 20 22iteration

26

27

28

Sum

Rat

e (M

bps)

(b)

Figure 5.9: Convergence performance of Algorithm 5.2 for (a) K = 2 users,and (b) K = 4 users when PT = 20 dBm

Finally, the convergence behaviour of the proposed near-optimal scheme,which is summarized in Algorithm 5.2, is presented in Figure 5.9 for 2-userand 4-user cases. It can be observed from this figure that Algorithm 5.2 con-verges quickly for both cases, which shows the efficiency of this iterative


scheme. Note that, in Figure 5.9(b), the algorithm obtains solutions whichgive higher sum rate in specific iteration indexes before it finally settles witha lower sum rate. This is because, in earlier iterations, the algorithm mayobtain solutions that maximize the sum rate but do not fulfill the constraint(5.25g) which guarantees all users are being served by at least a number ofRBs, or the minimum rate constraint (5.25d). Therefore, users with betterchannel conditions may give up some RBs to the weaker users who do notsatisfy these constraints, and hence lowering the sum rate.

5.6 Summary

In this chapter, analytical expressions for sum rate comparison and outageprobability were derived in order to compare the performance of CIC-basedNOMA and coded multicasting. Based on analytical and simulation results, itwas demonstrated that NOMA offer significant performance gain over codedmulticasting when pairing the users whose channel gains are highly distinc-tive. Meanwhile, coded multicasting is favoured when the channel gains ofthe paired users are relatively similar. Therefore, both NOMA and codedmulticasting have the potential to enhance the performance of a cache-aidedsystem subject to the channel conditions of the paired users. By exploiting thebenefits of both delivery techniques, a hybrid delivery scheme was proposedto further improve the sum rate performance, and a near-optimal iterativealgorithm was developed to solve the the joint power and RB allocation oruser pairing for NOMA. Furthermore, a suboptimal joint power and RB al-location or user pairing scheme was proposed in this chapter with the aimof maintaining good performance at a much reduced complexity. Simulationresults proved that the proposed hybrid scheme offer improved sum rate per-formance gain over NOMA and coded multicasting while maintaining a gooddegree of user fairness. The results also showed that the performance of thelow-complexity scheme degrades, but not far from the near-optimal one.

Chapter 6

Conclusions and Future Work

6.1 Conclusions

The main aim of this research is to enhance the performance of NOMA systemin content-centric mobile networks, particularly multicast and cache-aidednetworks, by designing efficient techniques such as resource allocation andbeamforming. These techniques are mainly designed with the objective ofenhancing the sum rate performance while taking into account the radio re-source constraints and QoS requirements. In this work, various performancemetrics, such as sum rate, outage probability and computational complexity,are implemented to investigate the effectiveness of the proposed techniques.

In chapter 3, a resource allocation technique is specifically designed forlayered multicast video streaming in NOMA system. A joint optimization ofpower allocation and subgrouping is formulated based on sum rate maxim-ization problem considering total transmission power and proportional rateconstraints. Due to the complexity of this optimization problem, the 2-layercase with arbitrary subgrouping was first investigated which lead to the de-rivation of two suboptimal power allocation methods. The first method isbased on the iterative subgradient method which is obtained using LDD ap-proach. The second one is a closed form M-ERPA scheme which is derivedbased on equal RB power allocation assumption and transformation of KKTcondition for the proportional rate constraint. In order to solve the powerallocation problem for the general multiple layer case, the M-ERPA solutionwas then modified to develop a low-complexity suboptimal SLPA schemewhich successively allocates power to each layer. Finally, the SLPA scheme

148

CHAPTER 6. CONCLUSIONS AND FUTURE WORK 149

can be incorporated into any of the three proposed low-complexity iterativesubgrouping methods in order to jointly optimized the power allocation andsubgrouping for NOMA-based multi-layer multicast system. Numerical res-ults for the 2-layer case with arbitrary subgrouping show that NOMA offersuperior performance over OFDMA, even with the implementation of subop-timal power allocation schemes. Moreover, both proposed subgradient andM-ERPA techniques offer better sum rate performance than the other low-complexity schemes, including FPA, FTPA and IGSMS-PA, while achievinghigher level of fairness towards the users with poor channel conditions. Al-though subgradient method performs better than M-ERPA, M-ERPA is moresuitable for practical implementation because it offers much lower complexity.For the general multi-layer case, the simulation results demonstrate the effect-iveness of the proposed low-complexity subgrouping methods in achievingthe sum rate performance closer to the optimal exhaustive search method.Among all the proposed subgrouping methods, Method 3 offers the bestsum rate performance while having lower computational complexity. Forthe power allocation schemes in a general multilayer case, the proposed SLPAscheme performs better than FPA and IGSMS-PA, particularly at higher trans-mission power. Furthermore, SLPA guarantees robust delivery of basic qual-ity video to all users by offering superior outage probability performance forthe high-priority base layer stream.

The performance of the NOMA-based layered multicast video streamingcan be further enhanced by employing beamforming technique at the expenseof additional computational complexity. Chapter 4 of this thesis presents thethree suboptimal beamforming algorithms which improve the sum rate per-formance at a low complexity. The first method is designed to maximize theCNRs of the weakest user in each subgroup. Meanwhile, the second schemefocuses on nullifying the inteference caused by the enhancement layer to theusers in subgroup 1 while enhancing the CNRs of selected users. The finalmethod is an iterative algorithm which proportionally enhances the CNRs ofthe users with the aim of improving the worst-case CNR and hence the sumrate. All the proposed beamforming algorithms can be embedded with any


existing power allocation schemes. In this work, the M-ERPA scheme pro-posed in Chapter 3 is applied in order to specifically solve sum rate maxim-ization problem with power budget and proportional rate constraints. Simu-lation results demonstrate the superior performance gain achieved by a mul-ticast system with any of the proposed beamforming techniques over SISOsystem. Furthermore, NOMA offers better performance than OMA in MISOcase considering the application of MU beamforming algorithm. Among theproposed beamforming techniques, NOMA-MU offers the best sum rate per-formance although having a higher complexity. Nevertheless, it is shown thatthe complexity of NOMA-MU is deemed low considering its fast convergence.On the other hand, MCNR has the lowest complexity but the performance de-grades when the number of users is larger.

Finally, the performance of NOMA and coded multicasting in a networkwith cache-aided UEs is compared in Chapter 5. In order to compare theperformance of both delivery techniques, analytical expressions for the prob-ability of sum rate comparison and outage probability are derived. Both ana-lytical and simulation results show that NOMA offers superior performanceover coded multicasting when the channel gains of the paired users are highlydistinctive. However, coded multicasting performs better than NOMA whenthe users having relatively similar channel gains are paired together. There-fore, both NOMA and coded multicasting potentially improves the perform-ance of cached-aided system depending on the user pairing scenario. In orderto further improve the sum rate performance, a hybrid delivery technique isdeveloped by allowing the selection of either NOMA or coded multicasting asa transmission mode for each RB. In addition, the NOMA-based joint powerallocation and RB allocation or user pairing, which is an MINLP problem,is investigated in this work. This problem is first transformed into non-linear programming by relaxing the integer variables into continuous vari-ables. Then, iterative subgradient method is applied to obtain near-optimalsolution. In addition, a low-complexity algorithm is also developed in orderto reduce the overall computational complexity. Simulation results demon-strate the effectiveness of the proposed hybrid scheme in enhancing the sumrate performance gain over NOMA and coded multicasting while offeringimproved fairness over NOMA. Meanwhile, the proposed low-complexity al-gorithm only experience slight degradation when compared to the optimal


one, despite the reduced complexity.

6.2 Future Work

This thesis highlighted several important techniques for NOMA system inmulti-layer multicast and cache aided networks. However, there are still sev-eral features and applications of NOMA in content-centric mobile networkthat are not considered in this thesis. The work in this thesis can be furtherextended in the future based on the following suggestions.

Cooperative NOMA-based layered multicast systemIn conventional multi-layer multicast system, the cell edge users will only beable to retrieve basic quality video through the successful detection of baselayer stream. In practice, there are cell edge users who require enhancementlayer stream to obtain high quality video. Through cooperative communic-ation, a user who successfully detect the enhancement layer stream can beselected as a potential relay to forward this layer stream to the requesting celledge users. However, if half-duplex cooperative system is employed, addi-tional time resources are required, resulting in low spectral efficiency. There-fore, it is necessary to implement full-duplex relaying in order to maintainhigh spectral efficiency in layered multicast system. However, self interfer-ence remains a challenging issue which needs to be addressed in order tomaintain better sum rate performance.

NOMA and coded multicasting in cache-aided NOMA considering mul-tiple user clusteringThe performance of NOMA and coded multicasting in Chapter 5 is comparedbased on two-user pairing (i.e., restricting to two users in a single RB). Thework can be extended to the case of multiple user clustering, in which a singleRB can serve multiple users. Multiple user clustering can also be applied tothe proposed hybrid scheme.

Application of joint CIC and SIC-based NOMA in the hybrid schemeThe UE cache configuration considered in Chapter 5 is specifically designedto make NOMA and coded multicasting feasible for all users. Practically, theimplementation of joint CIC and SIC decoding can accomodate different UE


cache cofigurations [71, 72]. In some cases, the individual rate is affected byintra-cell interference. It is interesting to study the impact of user pairingfor different UE cache cofigurations. In addition, the resource allocation forthe hybrid scheme considering joint CIC and SIC decoding can be furtherinvestigated.

Device-to-device (D2D) communications in cache-aided networkIn a cached-aided network, a missing subfile requested by a user may beavailable in the cache of another user. Instead of relying on BS to deliver themissing subfile, the user who possesses the requested subfile can cooperat-ively deliver this subfile to the requesting user via D2D communication. Inorder to ensure the efficient use of radio resources, the D2D communicationmay utilize the same time-frequency resource as the BS, which serves otherusers through NOMA or coded multicasting or hybrid scheme. However,the interference caused by the D2D link may result in adverse effect to theoverall performance. Therefore, investigating the resource allocation prob-lem, which deals with this interference issue, is an interesting direction forfuture research.

Resource allocation based on machine learningMost of the resource allocation schemes in this thesis are developed usingLDD approach. It is interesting to implement methods based on machinelearning to solve the resource allocation problem. The machine learning-based schemes can be compared with this work in terms of sum rate per-formance and computational complexity.

References

[1] Cisco. (2020) Cisco annual internet report (2018-2023). [Online]. Available: https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.pdf

[2] ——. (2017, Feb.) Cisco visual networking index: Global mobiledata traffic forecast update, 2016-2021. [Online]. Available: http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visual-networking-index-vni/mobile-white-paper-c11-520862.pdf

[3] P. K. Agyapong, M. Iwamura, D. Staehle, W. Kiess, and A. Benjebbour,“Design considerations for a 5G network architecture,” IEEE Communi-cations Magazine, vol. 52, no. 11, pp. 65–75, Nov. 2014.

[4] J. Lee, Y. Kim, Y. Kwak, J. Zhang, A. Papasakellariou, T. Novlan, C. Sun,and Y. Li, “LTE-advanced in 3GPP rel-13/14: an evolution toward 5G,”IEEE Communications Magazine, vol. 54, no. 3, pp. 36–42, Mar. 2016.

[5] E. Dahlman, S. Parkvall, and J. Skold, 4G, LTE-Advanced Pro and TheRoad to 5G, Third Edition, 3rd ed. USA: Academic Press, Inc., 2016.

[6] I. Selinis, K. Katsaros, M. Allayioti, S. Vahid, and R. Tafazolli, “The raceto 5G era; LTE and Wi-Fi,” IEEE Access, vol. 6, pp. 56 598–56 636, Oct.2018.

[7] 3rd Generation Partnership Project (3GPP), “Requirements for furtheradvancements for Evolved Universal Terrestrial Radio Access (E-UTRA)(LTE-Advanced)(Release 14),” Tech. Rep. TR 36.913, 2017.

153

https://www.cisco.com/c/en/us/solutions/collateral/executive-perspectives/annual-internet-report/white-paper-c11-741490.pdf



http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visual-networking-index-vni/mobile-white-paper-c11-520862.pdf



REFERENCES 154

[8] International Telecommunication Union (ITU), “Minimum require-ments related to technical performance for IMT-2020 radio interface(s),”Report ITU-R M.2410-0, 2017.

[9] W. Lei, A. C. Soong, L. Jianghua, W. Yong, B. Classon, W. Xiao, D. Maz-zarese, Z. Yang, and T. Saboorian, 5G System Design. Springer Interna-tional Publishing, 2020.

[10] I. Parvez, A. Rahmati, I. Guvenc, A. I. Sarwat, and H. Dai, “A surveyon low latency towards 5G: RAN, core network and caching solutions,”IEEE Communications Surveys Tutorials, vol. 20, no. 4, pp. 3098–3130, May2018.

[11] P. Popovski, K. F. Trillingsgaard, O. Simeone, and G. Durisi, “5G wire-less network slicing for eMBB, URLLC, and mMTC: A communication-theoretic view,” IEEE Access, vol. 6, pp. 55 765–55 779, Sep. 2018.

[12] H. Tullberg, P. Popovski, Z. Li, M. A. Uusitalo, A. Hoglund, O. Bu-lakci, M. Fallgren, and J. F. Monserrat, “The METIS 5G system concept:Meeting the 5G requirements,” IEEE Communications Magazine, vol. 54,no. 12, pp. 132–139, Dec. 2016.

[13] A. Ghosh, A. Maeder, M. Baker, and D. Chandramouli, “5G evolution:A view on 5G cellular technology beyond 3GPP release 15,” IEEE Access,vol. 7, pp. 127 639–127 651, Sep. 2019.

[14] K. Poularakis, G. Iosifidis, V. Sourlas, and L. Tassiulas, “Exploitingcaching and multicast for 5G wireless networks,” IEEE Transactions onWireless Communications, vol. 15, no. 4, pp. 2995–3007, Apr. 2016.

[15] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. C. Leung, “Cache inthe air: Exploiting content caching and delivery techniques for 5G sys-tems,” IEEE Communications Magazine, vol. 52, no. 2, pp. 131–139, Feb.2014.

[16] Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashan, C. I, and H. Poor, “Ap-plication of non-orthogonal multiple access in LTE and 5G networks,”IEEE Communications Magazine, vol. 55, no. 2, pp. 185–191, Feb. 2017.

REFERENCES 155

[17] S. M. Riazul Islam, N. Avazov, O. A. Dobre, and K. Kwak, “Power-domain non-orthogonal multiple access (NOMA) in 5G systems: Poten-tials and challenges,” IEEE Communications Surveys & Tutorials, vol. 19,no. 2, pp. 721–742, 2017.

[18] A. Benjebbour, A. Li, Y. Saito, Y. Kishiyama, A. Harada, and T. Naka-mura, “System-level performance of downlink NOMA for future LTEenhancements,” in Proc. IEEE Globecom Workshops (GC Wkshps), Dec.2013, pp. 66–70.

[19] L. Dai, B. Wang, Z. Ding, Z. Wang, S. Chen, and L. Hanzo, “A survey ofnon-orthogonal multiple access for 5G,” IEEE Communications SurveysTutorials, vol. 20, no. 3, pp. 2294–2323, May 2018.

[20] Y. Cai, Z. Qin, F. Cui, G. Y. Li, and J. A. McCann, “Modulation and mul-tiple access for 5G networks,” IEEE Communications Surveys Tutorials,vol. 20, no. 1, pp. 629–646, 2018.

[21] B. Makki, K. Chitti, A. Behravan, and M. Alouini, “A survey of NOMA:Current status and open research challenges,” IEEE Open Journal of theCommunications Society, vol. 1, pp. 179–189, Jan. 2020.

[22] S. M. R. Islam, M. Zeng, O. A. Dobre, and K. Kwak, “Resource allocationfor downlink NOMA systems: Key techniques and open issues,” IEEEWireless Communications, vol. 25, no. 2, pp. 40–47, Apr. 2018.

[23] A. Goldsmith, Wireless Communications, 1st ed. Cambridge: CambridgeUniversity Press, 2005.

[24] T. S. Rappaport, Wireless Communications: Principles and Practice, 2nd ed.Upper Saddle River, New Jersey: Prentice Hall PTR, 2002.

[25] D. Tse and P. Viswanath, Fundamentals of Wireless Communication, 1st ed.USA: Cambridge University Press, 2005.

[26] Y. S. Cho, J. Kim, W. Y. Yang, and C. G. Kang, MIMO-OFDM WirelessCommunications with MATLAB, 1st ed. Wiley Publishing, 2010.

[27] K. Wehrle, M. Gnes, and J. Gross, Modeling and Tools for Network Simula-tion, 1st ed. Springer Publishing Company, Incorporated, 2010.

REFERENCES 156

[28] H. Zarrinkoub, Understanding LTE with MATLAB: From MathematicalModeling to Simulation and Prototyping, 1st ed. Wiley Publishing, 2014.

[29] International Telecommunication Union (ITU), “Minimum require-ments related to technical performance for IMT-2020 radio interface(s),”Recommendation ITU-R M.1225, 1997.

[30] L. Dai, B. Wang, Y. Yuan, S. Han, C. L. I, and Z. Wang, “Non-orthogonalmultiple access for 5G: Solutions, challenges, opportunities, and futureresearch trends,” IEEE Communications Magazine, vol. 53, no. 9, pp. 74–81, Sep. 2015.

[31] M. Vaezi, Z. Ding, and H. V. Poor, Multiple Access Techniques for 5GWireless Networks and Beyond, 1st ed. Springer Publishing Company,Incorporated, 2018.

[32] Y. Liu, Z. Qin, and Z. Ding, Non-orthogonal multiple access for massiveconnectivity, 1st ed., ser. SpringerBriefs in Computer Science. Cham:Springer, 2020.

[33] Z. Ding, X. Lei, G. K. Karagiannidis, R. Schober, J. Yuan, and V. K. Bhar-gava, “A survey on non-orthogonal multiple access for 5G networks:Research challenges and future trends,” IEEE Journal on Selected Areas inCommunications, vol. 35, no. 10, pp. 2181–2195, Oct. 2017.

[34] F. L. Luo and C. Zhang, Signal Processing for 5G: Algorithms and Imple-mentations, 1st ed. Wiley-IEEE Press, 2016.

[35] R. Vannithamby and S. Talwar, Towards 5G: Applications, requirements &candidate technologies, 1st ed. Chichester, West Susses, United Kingdom:John Wiley & Sons Inc., 2017.

[36] K. Saito, A. Benjebbour, Y. Kishiyama, Y. Okumura, and T. Nakamura,“Performance and design of SIC receiver for downlink NOMA withopen-loop SU-MIMO,” in Proc. IEEE International Conference on Commu-nication Workshop (ICCW), June 2015, pp. 1161–1165.

[37] K. Saito, A. Benjebbour, A. Harada, Y. Kishiyama, and T. Nakamura,

REFERENCES 157

“Link-level performance of downlink NOMA with SIC receiver consid-ering error vector magnitude,” in Proc. IEEE 81st Vehicular TechnologyConference (VTC Spring), May 2015, pp. 1–5.

[38] C. Yan, A. Harada, A. Benjebbour, Y. Lan, A. Li, and H. Jiang, “Receiverdesign for downlink non-orthogonal multiple access (NOMA),” in Proc.IEEE 81st Vehicular Technology Conference (VTC Spring), May 2015, pp.1–6.

[39] X. Wei, H. Liu, Z. Geng, K. Zheng, R. Xu, Y. Liu, and P. Chen, “Soft-ware defined radio implementation of a non-orthogonal multiple accesssystem towards 5G,” IEEE Access, vol. 4, pp. 9604–9613, Dec. 2016.

[40] M. Vaezi, R. Schober, Z. Ding, and H. V. Poor, “Non-orthogonal multi-ple access: Common myths and critical questions,” IEEE Wireless Com-munications, vol. 26, no. 5, pp. 174–180, Oct. 2019.

[41] S. Timotheou and I. Krikidis, “Fairness for non-orthogonal multiple ac-cess in 5G systems,” IEEE Signal Processing Letters, vol. 22, no. 10, pp.1647–1651, Oct. 2015.

[42] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance ofnon-orthogonal multiple access in 5G systems with randomly deployedusers,” IEEE Signal Processing Letters, vol. 21, no. 12, pp. 1501–1505, Dec.2014.

[43] P. Parida and S. S. Das, “Power allocation in OFDM based NOMA sys-tems: A DC programming approach,” in Proc. IEEE Globecom Workshops(GC Wkshps), Dec. 2014, pp. 1026–1031.

[44] Z. Q. Al-Abbasi and D. K. C. So, “Resource allocation in non-orthogonaland hybrid multiple access system with proportional rate constraint,”IEEE Transactions on Wireless Communication, vol. 16, no. 10, pp. 6309–6320, Oct. 2017.

[45] Z. Ding, P. Fan, and H. V. Poor, “Impact of user pairing on 5Gnonorthogonal multiple-access downlink transmissions,” IEEE Transac-tions on Vehicular Technology, vol. 65, no. 8, pp. 6010–6023, Aug. 2016.

REFERENCES 158

[46] W. Liang, Z. Ding, Y. Li, and L. Song, “User pairing for downlink non-orthogonal multiple access networks using matching algorithm,” IEEETransactions on Communications, vol. 65, no. 12, pp. 5319–5332, Dec. 2017.

[47] M. Gruber and D. Zeller, “Multimedia broadcast multicast service: newtransmission schemes and related challenges,” IEEE CommunicationsMagazine, vol. 49, no. 12, pp. 176–181, Dec. 2011.

[48] D. Lecompte and F. Gabin, “Evolved multimedia broadcast/multicastservice (eMBMS) in LTE-Advanced: Overview and rel-11 enhance-ments,” IEEE Communications Magazine, vol. 50, no. 11, pp. 68–74, Nov.2012.

[49] G. Araniti, M. Condoluci, P. Scopelliti, A. Molinaro, and A. Iera, “Mul-ticasting over emerging 5G networks: Challenges and perspectives,”IEEE Network, vol. 31, no. 2, pp. 80–89, Mar. 2017.

[50] J. Montalban, P. Scopelliti, M. Fadda, E. Iradier, C. Desogus, P. Angueira,M. Murroni, and G. Araniti, “Multimedia multicast services in 5G net-works: Subgrouping and non-orthogonal multiple access techniques,”IEEE Communications Magazine, vol. 56, no. 3, pp. 91–95, Mar. 2018.

[51] J. J. Gimenez, J. L. Carcel, M. Fuentes, E. Garro, S. Elliott, D. Var-gas, C. Menzel, and D. Gomez-Barquero, “5G new radio for terrestrialbroadcast: A forward-looking approach for NR-MBMS,” IEEE Transac-tions on Broadcasting, vol. 65, no. 2, pp. 356–368, June 2019.

[52] R. O. Afolabi, A. Dadlani, and K. Kim, “Multicast scheduling and re-source allocation algorithms for OFDMA-based systems: A survey,”IEEE Communications Surveys & Tutorials, vol. 15, no. 1, pp. 240–254,2013.

[53] C. Suh and J. Mo, “Resource allocation for multicast services in multi-carrier wireless communications,” IEEE Transactions on Wireless Commu-nications, vol. 7, no. 1, pp. 27–31, Jan. 2008.

[54] J. Wolf, A. Wyner, and J. Ziv, “Source coding for multiple descriptions,”The Bell System Technical Journal, vol. 59, no. 8, pp. 1417–1426, Oct. 1980.

REFERENCES 159

[55] W. Li, “Overview of fine granularity scalability in MPEG-4 video stan-dard,” IEEE Transactions on Circuits and Systems for Video Technology,vol. 11, no. 3, pp. 301–317, Mar. 2001.

[56] H. Schwarz, D. Marpe, and T. Wiegand, “Overview of the scalable videocoding extension of the H.264/AVC standard,” IEEE Transactions on Cir-cuits and Systems for Video Technology, vol. 17, no. 9, pp. 1103–1120, Sep.2007.

[57] J. Xu, X. Shen, J. W. Mark, and J. Chai, “Adaptive transmission of multi-layered video over wireless fading channels,” IEEE Transactions on Wire-less Communications, vol. 6, no. 6, pp. 2305–2314, June 2007.

[58] Y. Huo, C. Hellge, T. Wiegand, and L. Hanzo, “A tutorial and review oninter-layer FEC coded layered video streaming,” IEEE CommunicationsSurveys Tutorials, vol. 17, no. 2, pp. 1166–1207, 2015.

[59] J. M. Boyce, Y. Ye, J. Chen, and A. K. Ramasubramonian, “Overviewof SHVC: Scalable extensions of the high efficiency video coding stan-dard,” IEEE Transactions on Circuits and Systems for Video Technology,vol. 26, no. 1, pp. 20–34, Jan. 2016.

[60] C. Guo, Y. Cui, D. W. K. Ng, and Z. Liu, “Power-efficient multi-qualitymulticast beamforming based on SVC and superposition coding,” inProc. IEEE Global Communications Conference (Globecom), Dec. 2017, pp.1–7.

[61] ——, “Multi-quality multicast beamforming based on scalable videocoding,” in arXiv:1610.09530, 2016.

[62] X. Jiang, H. Lu, C. W. Chen, and F. Wu, “Receiver-driven video multi-cast over NOMA systems in heterogeneous environment,” in Proc. IEEEConference on Computer Communications (INFOCOM), Apr. 2019, pp. 982–990.

[63] H. Zhu, Y. Cao, T. Jiang, and Q. Zhang, “Scalable NOMA multicast forSVC streams in cellular networks,” IEEE Transactions on Communications,vol. 66, no. 12, pp. 6339–6352, Dec. 2018.

REFERENCES 160

[64] T. Tang, A. Wang, S. Muhaidat, S. Li, M. Li, and J. Liang, “MDC-NOMA:Multiple description coding-based nonorthogonal multiple access forimage transmission,” IEEE Systems Journal, pp. 1–10, Aug. 2020.

[65] W. T. Tan and A. Zakhor, “Video multicast using layered FEC and scal-able compression,” IEEE Transactions on Circuits and Systems for VideoTechnology, vol. 11, no. 3, pp. 373–386, Mar. 2001.

[66] W. Han, A. Liu, and V. K. N. Lau, “PHY-caching in 5G wireless net-works: design and analysis,” IEEE Communications Magazine, vol. 54,no. 8, pp. 30–36, Aug. 2016.

[67] Y. Fadlallah, A. M. Tulino, D. Barone, G. Vettigli, J. Llorca, and J. Gorce,“Coding for caching in 5G networks,” IEEE Communications Magazine,vol. 55, no. 2, pp. 106–113, Feb. 2017.

[68] M. A. Maddah-Ali and U. Niesen, “Coding for caching: Fundamentallimits and practical challenges,” IEEE Communications Magazine, vol. 54,no. 8, pp. 23–29, Aug. 2016.

[69] G. Paschos, E. Bastug, I. Land, G. Caire, and M. Debbah, “Wirelesscaching: Technical misconceptions and business barriers,” IEEE Com-munications Magazine, vol. 54, no. 8, pp. 16–22, Aug. 2016.

[70] M. Ji, A. M. Tulino, J. Llorca, and G. Caire, “Order-optimal rate ofcaching and coded multicasting with random demands,” IEEE Trans-actions on Information Theory, vol. 63, no. 6, pp. 3923–3949, June 2017.

[71] L. Xiang, D. W. K. Ng, X. Ge, Z. Ding, V. W. S. Wong, and R. Schober,“Cache-aided non-orthogonal multiple access,” in Proc. IEEE Interna-tional Conference on Communications (ICC), May 2018, pp. 1–7.

[72] ——, “Cache-aided non-orthogonal multiple access: The two-usercase,” IEEE Journal of Selected Topics in Signal Processing, vol. 13, no. 3,pp. 436–451, June 2019.

[73] Z. Zhao, M. Xu, Y. Li, and M. Peng, “A non-orthogonal multiple access-based multicast scheme in wireless content caching networks,” IEEEJournal on Selected Areas in Communications, vol. 35, no. 12, pp. 2723–2735, Dec. 2017.

REFERENCES 161

[74] P. Henarejos, M. Shaat, and M. Navarro, “NOMA assisted joint broad-cast and multicast transmission in 5G networks,” in Proc. InternationalSymposium on Wireless Communication Systems (ISWCS), Aug. 2017, pp.420–425.

[75] S. R. Lee, S. H. Park, and I. Lee, “NOMA systems with content-centricmulticast transmission for C-RAN,” IEEE Wireless Communications Let-ters, pp. 828–831, Oct. 2018.

[76] G. Liu, Z. Wang, J. Hu, Z. Ding, and P. Fan, “Cooperative NOMAbroadcasting/multicasting for low-latency and high-reliability 5G cel-lular V2X communications,” IEEE Internet of Things Journal, vol. 6, no. 5,pp. 7828–7838, Oct. 2019.

[77] Z. Ding, Z. Zhao, M. Peng, and H. V. Poor, “On the spectral effi-ciency and security enhancements of NOMA assisted multicast-unicaststreaming,” IEEE Transactions on Communications, vol. 65, no. 7, pp.3151–3163, July 2017.

[78] Z. Yang, J. A. Hussein, P. Xu, Z. Ding, and Y. Wu, “Power alloca-tion study for non-orthogonal multiple access networks with multicast-unicast transmission,” IEEE Transactions on Wireless Communications,vol. 17, no. 6, pp. 3588–3599, June 2018.

[79] L. Lv, J. Chen, Q. Ni, and Z. Ding, “Design of cooperative non-orthogonal multicast cognitive multiple access for 5G systems: Userscheduling and performance analysis,” IEEE Transactions on Communi-cations, vol. 65, no. 6, pp. 2641–2656, June 2017.

[80] L. Yang, J. Chen, Q. Ni, J. Shi, and X. Xue, “NOMA-enabled cooperativeunicast-multicast: Design and outage analysis,” IEEE Transactions onWireless Communications, vol. 16, no. 12, pp. 7870–7889, Dec. 2017.

[81] Z. Zhang, Z. Ma, Y. Xiao, M. Xiao, G. K. Karagiannidis, and P. Fan,“Non-orthogonal multiple access for cooperative multicast millimeterwave wireless networks,” IEEE Journal on Selected Areas in Communica-tions, vol. 35, no. 8, pp. 1794–1808, Aug. 2017.

REFERENCES 162

[82] Z. Zhang, Z. Ma, M. Xiao, G. Liu, and P. Fan, “Modeling and analysis ofnon-orthogonal MBMS transmission in heterogeneous networks,” IEEEJournal on Selected Areas in Communications, vol. 35, no. 10, pp. 2221–2237,Oct. 2017.

[83] Y. Liu, Y. Liu, and G. Ma, “Hybrid decode-forward amplify-forwardrelaying with opportunistic layered multicast,” in Proc. 10th InternationalConference on Wireless Communications and Signal Processing (WCSP), Oct.2018, pp. 1–7.

[84] L. Yang, Q. Ni, L. Lv, J. Chen, X. Xue, H. Zhang, and H. Jiang, “Co-operative non-orthogonal layered multicast multiple access for hetero-geneous networks,” IEEE Transactions on Communications, vol. 67, no. 2,pp. 1148–1165, Feb. 2019.

[85] J. Choi, “Minimum power multicast beamforming with superpositioncoding for multiresolution broadcast and application to NOMA sys-tems,” IEEE Transactions on Communications, vol. 63, no. 3, pp. 791–800,Mar. 2015.

[86] T. Li, H. Zhang, X. Zhou, and D. Yuan, “NOMA-enabled layered videomulticast in wireless-powered relay systems,” IEEE Communications Let-ters, vol. 23, no. 11, pp. 2118–2121, Nov. 2019.

[87] R. H. Gau and H. T. Chiu, “Scalable NOMA multicast in cellular net-works,” in Proc. IEEE 27th Annual International Symposium on Personal,Indoor, and Mobile Radio Communications (PIMRC), Sep. 2016, pp. 1–6.

[88] H. Duan, Y. Zhang, and J. Song, “Block-level utility maximization forNOMA-based layered broadcasting,” IEEE Transactions on Broadcasting,vol. 66, no. 1, pp. 21–33, Mar. 2020.

[89] J. Wu, B. Tan, J. Wu, and M. Wang, “Video multicast: Integrating scal-ability of soft video delivery systems into NOMA,” IEEE Wireless Com-munications Letters, vol. 8, no. 6, pp. 1722–1726, Dec. 2019.

[90] M. Chen and S. Li, “Power allocation for NOMA based layered multicasttransmission,” in Proc. IEEE 4th International Conference on Computer andCommunications (ICCC), Dec. 2018, pp. 678–682.

REFERENCES 163

[91] S. Boyd and L. Vandenberghe, Convex Optimization, 1st ed. CambridgeUniversity Press, 2004.

[92] J. Wang, Q. Peng, Y. Huang, H. Wang, and X. You, “Convexity ofweighted sum rate maximization in NOMA systems,” IEEE Signal Pro-cessing Letters, vol. 24, no. 9, pp. 1323–1327, Sep. 2017.

[93] Z. Shen, J. G. Andrews, and B. L. Evans, “Adaptive resource allocationin multiuser OFDM systems with proportional rate constraints,” IEEETransactions on Wireless Communications, vol. 4, no. 6, pp. 2726–2737, Nov.2005.

[94] W. Yu, R. Lui, and R. Cendrillon, “Dual optimization methods for mul-tiuser orthogonal frequency division multiplex systems,” in Proc. IEEEGlobal Communications Conference (Globecom), Nov. 2004, pp. 225–229.

[95] D. Palomar and M. Chiang, “A tutorial on decomposition methods fornetwork utility maximization,” IEEE Journal on Selected Areas in Commu-nications, vol. 24, no. 8, pp. 1439–1451, Aug. 2006.

[96] H. Shi, R. V. Prasad, E. Onur, and I. G. M. M. Niemegeers, “Fairness inwireless networks: Issues, measures and challenges,” IEEE Communica-tions Surveys Tutorials, vol. 16, no. 1, pp. 5–24, 2014.

[97] N. D. Sidiropoulos, T. N. Davidson, and Zhi-Quan Luo, “Transmitbeamforming for physical-layer multicasting,” IEEE Transactions on Sig-nal Processing, vol. 54, no. 6, pp. 2239–2251, June 2006.

[98] L. Tran, M. F. Hanif, and M. Juntti, “A conic quadratic programmingapproach to physical layer multicasting for large-scale antenna arrays,”IEEE Signal Processing Letters, vol. 21, no. 1, pp. 114–117, Jan. 2014.

[99] B. Gopalakrishnan and N. D. Sidiropoulos, “High performance adaptivealgorithms for single-group multicast beamforming,” IEEE Transactionson Signal Processing, vol. 63, no. 16, pp. 4373–4384, Aug. 2015.

[100] A. Narula, M. J. Lopez, M. D. Trott, and G. W. Wornell, “Efficient useof side information in multiple-antenna data transmission over fad-ing channels,” IEEE Journal on Selected Areas in Communications, vol. 16,no. 8, pp. 1423–1436, Oct. 1998.

REFERENCES 164

[101] A. Lozano, “Long-term transmit beamforming for wireless multicast-ing,” in Proc. IEEE International Conference on Acoustics, Speech and SignalProcessing (ICASSP), Apr. 2007, pp. 417–420.

[102] D. Christopoulos, S. Chatzinotas, and B. Ottersten, “Sum rate maxi-mizing multigroup multicast beamforming under per-antenna powerconstraints,” in Proc. IEEE Global Communications Conference (Globecom),Dec. 2014, pp. 3354–3359.

[103] Y. Sun and K. J. R. Liu, “Transmit diversity techniques for multicast-ing over wireless networks,” in Proc. IEEE Wireless Communications andNetworking Conference (WCNC), Mar. 2004, pp. 593–598.

[104] M. Alageli, A. Ikhlef, and J. Chambers, “Optimization for maximizingsum secrecy rate in MU-MISO SWIPT systems,” IEEE Transactions onVehicular Technology, vol. 67, no. 1, pp. 537–553, Jan. 2018.

[105] Z. Ding, P. Fan, G. K. Karagiannidis, R. Schober, and H. V. Poor,“NOMA assisted wireless caching: Strategies and performance analy-sis,” IEEE Transactions on Communications, vol. 66, no. 10, pp. 4854–4876,Oct. 2018.

[106] ——, “On the application of NOMA to wireless caching,” in Proc. IEEEInternational Conference on Communications (ICC), May 2018, pp. 1–7.

[107] Z. Zhao, M. Xu, W. Xie, Z. Ding, and G. K. Karagiannidis, “Coverageperformance of NOMA in wireless caching networks,” IEEE Communi-cations Letters, vol. 22, no. 7, pp. 1458–1461, July 2018.

[108] Y. Fu, W. Wen, Z. Zhao, T. Q. S. Quek, S. Jin, and F. Zheng, “Dynamicpower control for NOMA transmissions in wireless caching networks,”IEEE Wireless Communications Letters, vol. 8, no. 5, pp. 1485–1488, Oct.2019.

[109] J. Zhao, Y. Liu, T. Mahmoodi, K. K. Chai, Y. Chen, and Z. Han, “Re-source allocation in cache-enabled CRAN with non-orthogonal multipleaccess,” in Proc. IEEE International Conference on Communications (ICC),May 2018, pp. 1–6.

REFERENCES 165

[110] S. Yan, L. Qi, Y. Zhou, M. Peng, and G. M. S. Rahman, “Joint user accessmode selection and content popularity prediction in non-orthogonalmultiple access-based F-RANs,” IEEE Transactions on Communications,vol. 68, no. 1, pp. 654–666, Jan. 2020.

[111] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,”IEEE Transactions on Information Theory, vol. 60, no. 5, pp. 2856–2867,May 2014.

[112] K. N. Doan, M. Vaezi, W. Shin, H. V. Poor, H. Shin, and T. Q. S. Quek,“Power allocation in cache-aided NOMA systems: Optimization anddeep reinforcement learning approaches,” IEEE Transactions on Commu-nications, vol. 68, no. 1, pp. 630–644, Jan. 2020.

[113] S. Gurugopinath, P. C. Sofotasios, Y. Al-Hammadi, and S. Muhaidat,“Cache-aided non-orthogonal multiple access for 5G-enabled vehicularnetworks,” IEEE Transactions on Vehicular Technology, vol. 68, no. 9, pp.8359–8371, Sep. 2019.

[114] H. Liu, W. Li, B. Lu, and X. Li, “Multi-relay non-orthogonal multiple ac-cess network with cache-enabled users,” IEEE Access, vol. 7, pp. 133 090–133 099, Sep. 2019.

[115] Y. Fu, Y. Liu, H. Wang, Z. Shi, and Y. Liu, “Mode selection between in-dex coding and superposition coding in cache-based NOMA networks,”IEEE Communications Letters, vol. 23, no. 3, pp. 478–481, Mar. 2019.

[116] Y. Fu, K. W. Shumt, C. W. Sung, and Y. Liu, “Optimal user pairing incache-based NOMA systems with index coding,” in Proc. IEEE Interna-tional Conference on Communications (ICC), May 2019, pp. 1–6.

[117] H. A. David and H. N. Nagaraja, Order Statistics, 3rd ed. New York,NY, USA: Wiley, 2003.

[118] E. Hildebrand, Introduction to Numerical Analysis, 1st ed. New York,NY, USA: Dover, 1987.

[119] Y. Liu, Z. Ding, M. Elkashlan, and H. V. Poor, “Cooperative non-orthogonal multiple access with simultaneous wireless information and

REFERENCES 166

power transfer,” IEEE Journal on Selected Areas in Communications, vol. 34,no. 4, pp. 938–953, Apr. 2016.

[120] H. Lei, J. Zhang, K. Park, P. Xu, I. S. Ansari, G. Pan, B. Alomair, andM. Alouini, “On secure NOMA systems with transmit antenna selectionschemes,” IEEE Access, vol. 5, pp. 17 450–17 464, Aug. 2017.

[121] K. Subranhmaniam, “On some applications of Mellin transforms tostatistics: Dependent random variables,” SIAM J. Appl. Math., vol. 19,no. 4, pp. 658–662, Dec. 1970.

[122] T. Abrão, L. D. H. Sampaio, S. Yang, K. T. K. Cheung, P. J. E. Jeszensky,and L. Hanzo, “Energy efficient OFDMA networks maintaining statis-tical QoS guarantees for delay-sensitive traffic,” IEEE Access, vol. 4, pp.774–791, Feb. 2016.

Documents

NOMA in Content-Centric Mobile Networks