Mojtaba Vaezi · Zhiguo Ding
H. Vincent Poor Editors
Multiple Access
Techniques for
5G Wireless
Networks and
Beyond
Multiple Access Techniques for 5G Wireless
Networks and Beyond
Mojtaba Vaezi Zhiguo Ding
H. Vincent Poor
•
Editors
Multiple Access Techniques
for 5G Wireless Networks
and Beyond
123
Editors
Mojtaba Vaezi
Villanova University
Villanova, PA
USA
H. Vincent Poor
Princeton University
Princeton, NJ
USA
Zhiguo Ding
The University of Manchester
Manchester
UK
ISBN 978-3-319-92089-4
ISBN 978-3-319-92090-0
https://doi.org/10.1007/978-3-319-92090-0
(eBook)
Library of Congress Control Number: 2018941989
© Springer International Publishing AG, part of Springer Nature 2019
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
Waveform design, multiple access, and random access techniques for fifth-generation
(5G) wireless networks and beyond are cutting-edge research topics that motivate a
very wide range of research problems. Despite being different, these three areas are
intertwined and lie at the heart of wireless communication systems. They allow
multiple users to effectively share a communication medium. The previous generations of cellular networks have adopted radically different multiple access techniques
with one common theme in mind: to have orthogonal signals for different users at the
receiver side. As an example, the fourth-generation (4G) cellular networks have
adopted orthogonal frequency division multiplexing (OFDM). In view of emerging
applications such as the Internet of Things (IoT), and in order to fulfill the need for
massive numbers of connections with diverse requirements in terms of latency and
throughput, 5G and beyond cellular networks are experiencing a paradigm shift in
design philosophy: shifting from orthogonal to non-orthogonal design in waveform,
multiple access, and random access techniques.
This book provides a comprehensive and intensive examination of multiple
access, random access, and waveform design techniques for 5G and beyond systems. It contains numerous state-of-the-art techniques and experimental results to
address the challenges in building 5G and beyond wireless networks. The book will
be of interest to readers from the communications, signal processing, and information theory communities. It will serve as a reference for graduate students,
researchers, and engineers involved in the design and standardization of wireless
communication systems. It can also serve as a reference book for graduate-level
courses for students in electrical engineering.
The book is organized into four parts and twenty-one chapters, each meant to be
self-contained. The contents of different chapters in each part are chosen so that
they reinforce and complement each other. Part I is focused on waveform design for
5G and beyond and includes four chapters outlining several advanced multicarrier
waveforms designs. Parts II through IV cover several related topics in multiple
access and random access. These parts include various non-orthogonal multiple
access (NOMA) techniques in the power domain, code domain, and other domains
as well as topics on random access. Power domain NOMA is mainly covered in
v
vi
Preface
Part II. Several code domain NOMA and other NOMA techniques as well as
random access schemes are discussed in Part III. Part IV includes experimental
trials and applications of NOMA in certain fields other than cellular
communications.
Part I (Chaps. 1–4) addresses waveform design. Chapter 1 introduces the reader
to 1G to 4G cellular systems and motivates the need for new multiple access and
waveforms for 5G and beyond systems. Chapter 2 presents various waveform
designs envisioned for the 5G new radio (NR). It first outlines the shortcomings of
conventional OFDM in serving the diverse use cases envisioned for 5G systems. It
then discusses design principles for new waveforms and introduces three new
waveforms considered in the standardization process of 5G NR, all developed from
OFDM. The new variants of OFDM are focused on reducing the out-of-band
(OOB) emission of cyclic-prefix OFDM by signal processing techniques such as
time-domain windowing and subband-based filtering. Chapter 3 introduces another
advanced multicarrier waveform, namely filter bank multicarrier modulation
(FBMC). FBMC can perform much better than windowed and filtered OFDM in
reducing the OOB emissions of conventional OFDM and can be considered as a
potential waveform for next generation wireless networks. Chapter 4 studies yet
another advanced multicarrier waveform, generalized frequency-division multiplexing (GFDM). GFDM is a multicarrier waveform technique that encapsulates
windowed and filtered OFDM techniques of 5G while providing an additional
design space reserved for forward comparability beyond 5G.
Part II (Chaps. 5–11) is dedicated to NOMA relying on the power domain.
Chapter 5 studies NOMA from an information-theoretic perspective. This chapter
reviews the basic premise behind NOMA in single- and multi-cell networks both in
the downlink and uplink. It also introduces various information-theoretic channels
that can be used to model physical layer security in NOMA. Chapter 6 investigates
power allocation for downlink NOMA under different performance metrics, such as
fairness, sum rate, and energy efficiency. The design principles of multiple-antenna
NOMA systems, including user clustering, channel state information acquisition,
and transmit beamforming are studied in Chap. 7. Chapter 8 is focused on applying
NOMA to millimeter wave networks with three transmissions schemes, namely
unicast, multicast and cooperative multicast. Chapter 9 is dedicated to full-duplex
NOMA, a technology that has the potential to double the spectral efficiency via
simultaneous transmissions in the uplink and downlink. Resource allocation in
heterogeneous NOMA with energy cooperation is discussed in Chap. 10. Chapter 11
evaluates the performance of NOMA in vehicle-to-vehicle massive MIMO channels.
Part III (Chaps. 12–17) introduces several code domain NOMA schemes as well
as non-orthogonal random access. All multiple access techniques presented in this
part have a common philosophy: to exploit efficient and low-complexity multiuser
detection. In particular, Chap. 12 studies sparse code multiple access (SCMA),
a code domain NOMA scheme that exploits the sparsity of the multi-dimensional
codewords to apply the low-complexity message passing algorithm for multiuser detection. Chapter 13 discusses interleave division multiple access (IDMA).
Preface
vii
IDMA applies a low-complexity iterative technique for multiuser detection and can
achieve near-capacity sum rate with proper power allocation. Chapter 14 introduces
pattern division multiple access (PDMA) which is a NOMA technique in which a
pattern defines the mapping of transmitted data to a group of time, frequency, and
spatial resources. In Chap. 15, low-density spreading (LDS), a variant of code
division multiple access (CDMA) in which spreading sequences have low density,
is studied. Owing to this, a near optimal message passing algorithm receiver with
practically feasible complexity can be exploited. Chapter 16 discusses a grant-free
multiple access scheme that enables lower transmission latency and savings in
device energy. Chapter 17 provides a comprehensive survey of random access
schemes that are suited to support IoT. These schemes are commonly based on
time, frequency, and code division multiple access techniques and different variants
of NOMA described in the previous chapters.
Part IV (Chaps. 18–21) outlines experimental trials, challenges, and future trends
of NOMA. To evaluate the performance of NOMA using real-world hardware and
in realistic radio environments, a test-bed is described in Chap. 18. Indoor and
outdoor experimental trials are then conducted which confirm that NOMA
improves user throughput as compared to orthogonal multiple access. Applications
and extension of NOMA to visible light communication networks and
terrestrial-satellite networks are studied in Chaps. 19 and 20, respectively. Finally,
Chap. 21 provides future research directions for NOMA in 5G wireless networks
and beyond as well as other fields.
We would like to extend our thanks to the people and organizations who made
this book possible. Our sincere thanks go to the chapter authors; it has been an
honor and a privilege to work with such a dedicated and talented group of authors
and researchers. Princeton University and the universities, research laboratories,
and corporations with which the authors are affiliated deserve credit for providing
facilities and intellectual environments for this project. It is also a pleasure to
acknowledge Springer and its team: Mary James, Brian Halm, and Zoe Kennedy.
Finally, we offer our deepest appreciation and gratitude to our families for their
patience and support during the months we were immersed in this project.
Villanova, PA, USA
Manchester, UK
Princeton, NJ, USA
Mojtaba Vaezi
Zhiguo Ding
H. Vincent Poor
Contents
Part I
Orthogonal Multiple Access Techniques and Waveform
Design
1
Introduction to Cellular Mobile Communications . . . . . . . . . . . . . .
Joseph Boccuzzi
2
OFDM Enhancements for 5G Based on Filtering
and Windowing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Rana Ahmed, Frank Schaich and Thorsten Wild
3
Filter Bank Multicarrier Modulation . . . . . . . . . . . . . . . . . . . . . . .
Ronald Nissel and Markus Rupp
4
Generalized Frequency Division Multiplexing: A Flexible
Multicarrier Waveform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ahmad Nimr, Shahab Ehsanfar, Nicola Michailow,
Martin Danneberg, Dan Zhang, Henry Douglas Rodrigues,
Luciano Leonel Mendes and Gerhard Fettweis
Part II
3
39
63
93
Non-Orthogonal Multiple Access (NOMA) in the Power
Domain
5
NOMA: An Information-Theoretic Perspective . . . . . . . . . . . . . . . . 167
Mojtaba Vaezi and H. Vincent Poor
6
Optimal Power Allocation for Downlink NOMA Systems . . . . . . . . 195
Yongming Huang, Jiaheng Wang and Jianyue Zhu
7
On the Design of Multiple-Antenna Non-Orthogonal Multiple
Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Xiaoming Chen, Zhaoyang Zhang, Caijun Zhong
and Derrick Wing Kwan Ng
ix
x
Contents
8
NOMA for Millimeter Wave Networks . . . . . . . . . . . . . . . . . . . . . . 257
Zhengquan Zhang and Zheng Ma
9
Full-Duplex Non-Orthogonal Multiple Access Networks . . . . . . . . . 285
Mohammed S. Elbamby, Mehdi Bennis, Walid Saad,
Mérouane Debbah and Matti Latva-aho
10 Heterogeneous NOMA with Energy Cooperation . . . . . . . . . . . . . . 305
Bingyu Xu, Yue Chen and Yuanwei Liu
11 NOMA in Vehicular Communications . . . . . . . . . . . . . . . . . . . . . . 333
Yingyang Chen, Li Wang, Yutong Ai, Bingli Jiao and Lajos Hanzo
Part III
NOMA in Code and Other Domains
12 Sparse Code Multiple Access (SCMA) . . . . . . . . . . . . . . . . . . . . . . 369
Zheng Ma and Jinchen Bao
13 Interleave Division Multiple Access (IDMA) . . . . . . . . . . . . . . . . . . 417
Yang Hu and Li Ping
14 Pattern Division Multiple Access (PDMA) . . . . . . . . . . . . . . . . . . . 451
Shanzhi Chen, Shaohui Sun, Shaoli Kang and Bin Ren
15 Low Density Spreading Multiple Access . . . . . . . . . . . . . . . . . . . . . 493
Mohammed Al-Imari and Muhammad Ali Imran
16 Grant-Free Multiple Access Scheme . . . . . . . . . . . . . . . . . . . . . . . . 515
Liqing Zhang and Jianglei Ma
17 Random Access Versus Multiple Access . . . . . . . . . . . . . . . . . . . . . 535
Riccardo De Gaudenzi, Oscar del Río Herrero, Stefano Cioni
and Alberto Mengali
Part IV
Challenges, Solutions, and Future Trends
18 Experimental Trials on Non-Orthogonal Multiple Access . . . . . . . . 587
Anass Benjebbour, Keisuke Saito and Yoshihisa Kishiyama
19 Non-Orthogonal Multiple Access in LiFi Networks . . . . . . . . . . . . . 609
Liang Yin and Harald Haas
20 NOMA-Based Integrated Terrestrial-Satellite Networks . . . . . . . . . 639
Xiangming Zhu, Chunxiao Jiang, Linling Kuang, Ning Ge
and Jianhua Lu
21 Conclusions and Future Research Directions for NOMA . . . . . . . . 669
Zhiguo Ding, Yongxu Zhu and Yan Chen
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 679
About the Editors
Mojtaba Vaezi received the Ph.D. degree in Electrical Engineering from McGill
University in 2014. From 2015 to 2018, he was with Princeton University as a
Postdoctoral Research Fellow and Associate Research Scholar. He is currently an
Assistant Professor of ECE at Villanova University and a Visiting Research
Collaborator at Princeton University. Before joining Princeton, he was a researcher
at Ericsson Research in Montreal, Canada. His research interests include the broad
areas of information theory, wireless communications, and signal processing, with
an emphasis on physical layer security and radio access technologies. Among his
publications in these areas is the book Cloud Mobile Networks: From RAN to EPC
(Springer, 2017). He has served as the president of McGill IEEE Student Branch
during 2012–2013. He is an Associate Editor of IEEE Communications Magazine
and IEEE Communications Letters. He has co-organized four international NOMA
workshops at VTC-Spring’17, Globecom’17, ICC’18, and Globecom’18. He is a
recipient of a number of academic, leadership, and research awards, including the
McGill Engineering Doctoral Award, IEEE Larry K. Wilson Regional Student
Activities Award in 2013, the Natural Sciences and Engineering Research Council
of Canada (NSERC) Postdoctoral Fellowship in 2014, and Ministry of Science and
ICT of Korea’s best paper award in 2017.
Zhiguo Ding received his B.Eng. in Electrical Engineering from the Beijing
University of Posts and Telecommunications in 2000, and the Ph.D. degree in
Electrical Engineering from Imperial College London in 2005. From July 2005 to
April 2018, he was working in Queen’s University Belfast, Imperial College,
Newcastle University and Lancaster University. Since April 2018, he has been with
the University of Manchester as a Professor in Communications. From September
2012 to September 2018, he has also been an academic visitor in Princeton
University. His research interests are 5G networks, game theory, cooperative and
energy harvesting networks and statistical signal processing. He is serving as an
Editor for IEEE Transactions on Communications and IEEE Transactions on
Vehicular Technology. He served as an Editor for IEEE Wireless Communication
Letters, IEEE Communication Letters, and Journal of Wireless Communications
xi
xii
About the Editors
and Mobile Computing. He received the best paper award in IET International
Communication Conference on Wireless Mobile and Computing, 2009, and the
IEEE WCSP 2015, IEEE Transactions on Vehicular Technologies Top Editor 2017,
and the EU Marie Curie Fellowship 2012–2014.
H. Vincent Poor received the Ph.D. degree in EECS from Princeton University in
1977. From 1977 until 1990, he was on the faculty of the University of Illinois at
Urbana-Champaign. Since 1990, he has been on the faculty at Princeton, where he
is currently the Michael Henry Strater University Professor of Electrical
Engineering. During 2006 to 2016, he served as Dean of Princeton’s School of
Engineering and Applied Science. His research interests are in the areas of information theory and signal processing, and their applications in wireless networks,
energy systems, and related fields. He is a member of the National Academy of
Engineering and the National Academy of Sciences and is a foreign member of the
Chinese Academy of Sciences, the Royal Society, and other national and international academies. Other recognition of his work includes the 2017 IEEE Alexander
Graham Bell Medal, and honorary doctorates and professorships from a number of
universities.
Acronyms
1G
2G
3G
3GPP
4G
5G
5GC
ACEP
ACK
ACLR
ACO-OFDM
ACRDA
ADC
AF
AGC
AMC
AMF
AMI
AMPS
AN
AoA
AoD
APD
APM
APP
APSK
AR
AWGN
BC
BEP
First generation
Second generation
Third generation
3rd Generation Partnership Project
Fourth generation
Fifth generation
5G core
Average codeword error probability
Acknowledgement
Adjacent channel leakage rejection
Asymmetrically clipped optical OFDM
Asynchronous contention resolution diversity ALOHA
Analog-to-digital converter
Application function
Automatic gain control
Adaptive modulation and coding
Access and mobility management function
Averagemutual information
Advanced mobile phone services
Artificial noise
Angle of arrival
Angle of departure
Avalanche photodiode
Amplitude-phase modulation
A posteriori probability
Amplitude phase-shift keying
Augmented reality
Additive white Gaussian noise
Broadcast channel
Bit error probability
xiii
xiv
BER
BF
BICM
BICM-ID
BLER
BP
BPSK
BRAM
BRU
BS
BSA
BSC
BSS
BTS
C2X
CA
CB
CBRS
CCDF
CCI
CDD
CDF
CDMA
CEP
CFO
CIM
CIR
CLT
CM
CMT
CN
CN
CNR
CoF
CoMP
CP
CPICH
CP-OFDM
CPRI
CQI
CR
CRA
C-RAN
CRC
CRDSA
Acronyms
Bit error rate
Beamforming
Bit-interleaved coded modulation
BIMC with iterative decoding
Block error rate
Belief propagation
Binary phase-shift keying
Block RAM
Basic resource unit
Base station
Binary switching algorithm
Base station controller
Base station subsystem
Base transceiver stations
Car to anything
Carrier aggregation
Coordinated beamforming
Citizens broadband radio service
Complementary cumulative distribution function
Co-channel interference
Cyclic delay diversity
Cumulative distribution function
Code division multiple access
Channel estimation preamble
Carrier frequency offset
Color intensity modulation
Channel impulse response
Central limit theorem
Cubic metric
Cosine-modulated multitone
Core network
Channel observation node
Channel-to-noise ratio
Cycle-of-four
Coordinated multipoint
Cyclic prefix
Common pilot channel
Cyclic prefix OFDM
Common public radio interface
Channel quality indicator
Cognitive radio
Contention resolution ALOHA
Cloud radio access network
Cyclic redundancy check
Contention resolution diversity slotted ALOHA
Acronyms
CR-NOMA
CRS
CS
CS
CS
CSA
CsDMA
CSI
CSIR
CSIT
CSK
CSMA
CSMA/CA
CSMA/CD
CTS
CU
CW-SIC
D/A
D2D
DA
DAC
DACE
DC
DC
DCCC
DCI
DCO-OFDM
DCS
DD
DEC
DFE
DL
DMRS
DoF
DPC
DPD
DRAM
DS-ALOHA
DS-CDMA
DSP
DS-SS
DTFT
DU
DVB
DVB-T
xv
Cognitive radio inspired NOMA
Cell-specific reference signal
Circuit switched
Cyclic suffix
Coordinated scheduling
Coded Slotted ALOHA
Code-shift division multiple access
Channel state information
Channel state information at the receiver
Channel state information at the Transmitter
Color shift keying
Carrier sense multiple access
Carrier sense multiple access/collision avoidance
Carrier sense multiple access/collision detection
Clear to send
Centralized unit
Codeword level SIC
Digital-to-analog
Device-to-device
Deferred acceptance
Digital-to-analog converter
Data-aided channel estimation
Direct current
Difference of convex functions
Discrete codebook-constrained capacity
Downlink control information
Direct current-biased optical OFDM
Dynamic cell selection
Direct detection
Decoder
Decision feedback equalizer
Downlink
Demodulation reference signal
Degree of freedom
Dirty-paper coding
Digital predistortion
Dynamic random access memory
Diversity slotted ALOHA
Direct-sequence CDMA
Digital signal processor
Direct-sequence spread spectrum
Discrete-time Fourier transform
Distributed unit
Digital video broadcasting
Digital video broadcasting terrestrial
xvi
DZT
EDGE
EE
eMBB
eMBMS
eMTC
eNB
EP
EPC
ES
ESE
E-SSA
ETU
EV-DO
EVM
EXIT
FBMC
FD
FDD
FDE
FDM
FDMA
FEC
FER
FET
FFT
FIFO
FIR
FMT
FN
FO
f-OFDM
FOV
FPA
FPGA
FSPA
FTPA
GA
GA
GB
GEO
GEVD
GF
GFDM
GGSN
Acronyms
Discrete Zak transform
Enhanced data rates for GSM evolution
Energy efficiency
Enhanced mobile broadband
Evolved multimedia broadcast multicast service
Enhanced machine type communications
Enhanced NodeB
Expectation propagation
Equal power control
Exhaustive search
Elementary signal estimation
Enhanced spread spectrum ALOHA
Extended typical urban
Evolution-data optimized
Error vector magnitude
Extrinsic information transfer
Filter bank multicarrier modulation
Full-duplex
Frequency division duplex
Frequency domain equalization
Frequency division multiplex
Frequency division multiple access
Forward error control
Frame error rate
Field-effect transistor
Fast Fourier transform
First in first out
Finite impulse repose
Filtered multitone
Function Node
Frequency offset
Filtered OFDM
Field of view
Fixed power allocation
Field programmable gate array
Full search power allocation
Fractional transmission power allocation
Genetic algorithm
Gaussian approximation
Grant-based
Geosynchronous earth orbit
Generalized eigenvalue decomposition
Grant-free
Generalized frequency division multiplexing
Gateway GPRS support node
Acronyms
GMSC
GMSK
gNB
GOCA
GP
GPRS
GSM
GSVD
GT
HARQ
HD
HetNet
HK
HLR
HOM
HOS
HPPP
HSDPA
HSPA
HSTRN
HSUPA
i.i.d.
IA
IAI
IBFD
IBI
IC
IC
ICI
ICI
ICT
IDD
IDFT
IDMA
IFFT
IFPI
IFS
IGCH
IGMA
IM
I-MRC
INI
IoT
IR
IRSA
xvii
Gateway mobile switching center
Gaussian minimum shift keying
Next generation NodeB
Group-orthogonal coded access
Guard period
General packet radio services
Global system for mobile communications
Generalized singular value decomposition
Guard tone
Hybrid automatic repeat request
Half-duplex
Heterogeneous network
Han-Kobayashi
Home location register
Higher-order modulation
Hierarchy of orthogonal sequences
Homogeneous Poisson point process
High speed downlink packet access
High speed packet access
Hybrid satellite terrestrial relay network
High speed uplink packet access
identically independently distributed
Interference alignment
Inter-antena interference
In-band full-duplex
Inter-block interference
Interference cancellation
Interference channel
Inter-carrier interference
Inter-cell interference
Information and communication technology
Iterative detection and decoding
Inverse discrete Fourier transform
Interleave division multiple access
Inverse fast Fourier transform
Interference-free pilot insertion
Inter-frame spacing
Information-guided channel hopping
Interleave grid multiple access
Intensity modulation
Iterative maximum ratio combining
Inter-numerology interference
Internet of Things
Infrared
Irregular repetition slotted ALOHA
xviii
ISI
ISM
ITS
ITU
IUI
JP
JRA
JT
J-TACS
KKT
KPI
LDM
LDPC
LDS
LED
LEO
LiFi
LLR
LMMSE
LOS
LP
LPWAN
LS
LTE
LTE-A
M2M
MA
MAC
MAC
MAC
MACA
MAI
MAP
MARSALA
MBS
MC-CDMA
MC-LDSMA
MC-NOMA
MCS
MEC
MEO
ME-SSA
MF
MF-CRDSA
MI
Acronyms
Inter-symbol interference
Industrial scientific and medical
Intelligent transportation systems
International telecommunications union
Inter-user interference
Joint processing
Joint resource allocation
Joint transmission
Japan TACS
Karush-Kuhn-Tucker
Key performance indicator
Layer division multiplex
Low-density parity-check
Low-density spreading
Light-emitting diode
Low Earth orbit
Light fidelity
Log-likelihood ratio
Linear minimum mean squared error
Line-of-sight
Linear program
Low power wide area networks
Least squares
Long term evolution
LTE advanced
Machine to machine
Multiple access
Multiple access channel
Medium access control
Medium access layer
Multiple access collision avoidance
Multiple access interference
Maximum a posterior
Multi-replica decoding using correlation based Iocalization
Macro base station
Multicarrier CDMA
Multicarrier low density spreading multiple access
Multi-channel NOMA
Modulation coding scheme
Multi-access edge computing
Medium Earth orbit
MMSE enhanced spread spectrum ALOHA
Matched filter
Multi-frequency CRDSA
Mutual information
Acronyms
MIIT
MIMO
MINLP
MISO
ML
MM
MMF
MMSE
mMTC
mmWave
MPA
MPF
MPR
MRC
MRT
MSC
MSE
MTC
MTSO
MU
MUD
MUI
MU-MIMO
MUSA
MuSCA
MUST
NAICS
N-AMPS
NB
NB-IoT
NCMA
NEF
NFV
NLOS
NMSE
NMT
NOCA
NOMA
NP
NR
O/E
OFDM
OFDMA
OMA
OOB
xix
Ministry of industrial and information technology
Multiple-input and multiple-output
Mixed integer nonlinear programming
Multiple-input and single-output
Maximum likelihood
Metameric modulation
Maximin fairness
Minimum mean squared error
Massive machine type communication
Millimeter wave
Message passing algorithm
Marginal product of functions
Multi-packet reception
Maximum ratio combining
Maximum ratio transmission
Mobile switching center
Mean squared error
Machine type communication
Mobile telephone switching office
Mobile user
Multi-user detection
Multi-user interference
Multi-user MIMO
Multi-user shared access
Multi-slots coded ALOHA
Multi-user superposition transmission
Network-assisted interference cancellation and suppression
Narrowband AMPS
NodeB
Narrow band IoT
Non-orthogonal coded multiple access
Noise enhancement factor
Network function virtualization
Non-line-of-sight
Normalized mean-squared error
Nordic mobile telephone
Non-orthogonal coded access
Non-orthogonal multiple access
Non-deterministic polynomial-time
New radio
Optical to electrical
Orthogonal frequency division multiplexing
Orthogonal frequency division multiple access
Orthogonal multiple access
Out-of-band
xx
OOK
OQAM
O-QPSK
OSS
OSTBC
OVSF
P/S
PA
PAM
PAPR
PBS
PC
PCCC
PCM
PD
PDCCH
pdf
PDMA
PDR
PEP
PER
PHY
PIA-ASP
PIC
p-i-n
PL
PLR
PN
PPM
PPP
PRB
PS
PS
PSD
PSK
PSTN
PWM
QAM
QoS
QPSK
RA
RACH
RAN
RAR
RAT
Acronyms
On-off keying
Offset quadrature amplitude modulation
Offset quadrature phase shift keying
Operation and support subsystem
Orthogonal space-time block coding
Orthogonal variable spreading factor
Parallel-to-serial
Power amplifier
Pulse amplitude modulation
Peak-to-average power ratio
Pico BS
Power control
Parallel concatenated convolutional code
Policy control function
Photodiode
Physical downlink control channel
Probability density function
Pattern division multiple access
Packet drop rate
Pairwise error probability
Packet error rate
Physical layer
Prior-information aided adaptive subspace pursuit
Parallel interference cancellation
Positive-intrinsic-negative
Primary layer
Packet loss rate
Pseudo noise
Pulse position modulation
Poisson point process
Physical resource block
Packet switched
Phase shifter
Power spectral density
Phase-shift keying
Public switched telephone network
Pulse width modulation
Quadrature amplitude modulation
Quality of service
Quadrature phase shift keying
Random access
Random access channel
Radio access network
Random access response
Radio access technology
Acronyms
RB
RC
RDMA
RE
RF
RMS
RNC
RNTI
RPC
RPMA
RR
RRC
RRC
RRH
RS
RSMA
RSRP
RTS
RV
RV
RX
S-ALOHA
SAMA
SA-SCMA
SC
SC
SCA
SC-FDMA
SCM
SCMA
SCR
SCS
SDMA
SDN
SDR
SE
SER
SFBC
SG
SGSN
SIC
SINR
SIR
SISO
SL
xxi
Resource block
Raised-cosine
Repetition division multiple access
Resource element
Radio frequency
Root mean square
Radio network controller
Radio network temporary identifier
Randomized power control
Random phase multiple access
Round robin
Radio resource control
Root-raised cosine
Remote radio head
Rate-splitting
Resource spread multiple access
Reference signal received power
Request to send
Redundancy version
Random variable
Receiver
Slotted ALOHA
Successive interference cancellation amenable multiple access
Spread asynchronous scrambled coded multiple access
Sub-carrier
Superposition coding
Successive convex approximation
Single carrier frequency division multiple access
Superposition coded modulation
Sparse code multiple access
Signal-to-clipping-noise ratio
Sub-carrier spacing
Space division multiple access
Software defined networking
Software defined radio
Spectral efficiency
Symbol error rate
Space frequency block coding
Scheduling grant
Serving GPRS support node
Successive interference cancellation
Signal-to-interference-plus-noise ratio
Signal-to-interference ratio
Single-input and single-output
Secondary layer
xxii
SLA
SLL
SLS
SL-SIC
SM
SMF
SMS
SMT
SNR
SPS
SR
SR
SrCMA
SS-ALOHA
SSD
SSK
STC
STO
SUD
SVD
Tx
TACS
TB
TCP
TDD
TDL
TDM
TDMA
TO
TO
TR
TR-STC
TS
TTI
TX
UA
UDM
UDN
UE
UFMC
UF-OFDM
UL
UMTS
UN
UNB
Acronyms
Side lobe attenuation
Side lobe level
System level simulation
Symbol level SIC
Spatial modulation
Session management function
Short message service
Staggered multitone
Signal-to-noise ratio
Semi-persistent scheduling
Scheduling request
Sum rate
Scrambled coded multiple access
Spread-spectrum ALOHA
Signal-space diversity
Space shift keying
Space-time coding
Symbol time offset
Single-user detection
Singular value decomposition
Transmit antenna
Total access communication system
Transport block
Transmission control protocol
Time division duplex
Tapped delay line
Time division multiplexing
Time division multiple access
Transmission occasion
Time offset
Technical report
Time-reversal space-time coding
Time-sharing
Transmission time interval
Transmitter
User association
Unified data management
Ultra dense network
User equipment
Universal filtered multicarrier
Universal filtered OFDM
Uplink
Universal mobile telephone system
User node
Ultra narrow band
Acronyms
U-OFDM
UPC
UPF
URLLC
UTRAN
V2I
V2N
V2P
V2V
V2X
VANET
VBLAST
VLC
VLR
VN
VR
WAVE
WBE
WCDMA
WOLA
WSR
ZF
ZFBF
ZP
ZP-OFDM
ZT-DFT
xxiii
Unipolar OFDM
Unequal power control
User plane function
Ultra-reliable low latency communications
UMTS terrestrial radio access network
Vehicle-to-infrastructure
Vehicle-to-network
Vehicle-to-pedestrian
Vehicle-to-vehicle
Vehicle-to-everything
Vehicular ad hoc network
Vertical Bell laboratories layered space-time
Visible light communication
Visitor location register
Variable node
Virtual reality
Wireless access for vehicular environments
Welch-bound equality
Wideband CDMA
Weighted overlap and add
Weighted sum rate
Zero-forcing
Zero-forcing beamforming
Zero-padding
Zero prefix OFDM
Zero tail DFT
Part I
Orthogonal Multiple Access Techniques
and Waveform Design
Chapter 1
Introduction to Cellular Mobile
Communications
Joseph Boccuzzi
1.1 Introduction
This chapter provides an overview of the evolution of the cellular mobile
communication systems. We begin with a quote from a conversation held over a
mobile cellular network from Martin Cooper on 3 April 1973 [1].
I’m calling you from a cell phone, a real handheld portable cell phone.
The mobile device used during this conversation was a Motorola DynaTAC weighing approximately 2.5 lbs with a cost of approximately $9,000 USD. This historic
event ignited a movement which would change the lives of so many people. This
life change is much, much more than supporting mobile users, it snowballed into
creating highly complex devices (presently called smart phones) that help us remain
connected to the world. These devices not only perform our much needed voice
and data communication needs, but they also take on a very wide array of supporting applications such as keeping our friends informed on social media, competing
with online gaming, consuming and producing video content, performing medical
measurements, utilizing location-based services, etc.
As these wireless devices benefited from Moore’s law, the cellular mobile technologies were able to remain a focal point to introduce such new and exciting features,
and benefits, to the end user.
This chapter is intended to address important driving technologies behind the
5G new radio (NR) system designs, which focuses on solutions to supporting 5G
new services in uplink (UL) transmissions with requirements such as low-latency
and high-reliability, energy-saving, and small packet applications. Grant-free (GF)
resources in NR UL is termed as “a configured grant,” which means that the
J. Boccuzzi (B)
Intel Corporation, San Diego, CA, USA
e-mail: Joseph.Boccuzzi@intel.com
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_1
3
4
J. Boccuzzi
pre-configured UE-specific resources will be used for UE UL transmission without dynamic scheduling/grant. Also, the base station (BS) in 5G NR network is
referred to as “next generation NodeB” or “gNB.”
1.2 Cellular Mobile Communication: A Primer
The cellular standards use a variety of multiple access (MA) techniques, which we
highlight in Table 1.1. These techniques include frequency division multiple access
(FDMA), time division multiple access (TDMA), code division multiple access
(CDMA), and orthogonal frequency division multiple access (OFDMA). We also
describe the relevant duplex method used for two-way communication and the actual
physical resources available to be assigned to each user. The duplex methods are time
division duplex (TDD) and frequency division duplex (FDD).
All of the above multiple access techniques can be viewed as a form of “orthogonal” multiple access (OMA), where the access of users, theoretically, do not interfere
with one another as they share the wireless medium. They are, however, limited to the
number of resources available that make them orthogonal to each other. An exception
to this would be CDMA, where the transmission from the wireless device to the base
station is inherently non-orthogonal.
In FDMA, the frequency is divided into channels to be utilized by various users.
In TDMA, time is divided into time slots as a means to allow various users to access
the cellular system. In CDMA, users are separated by PN codes and transmit over
the entire frequency channel, all at the same time. In OFDMA, users are allocated
to various frequency channels (groups of sub-carriers) at different time slots. For
the next generation digital cellular system called 5G, OFDMA is still used where
the sub-carrier spacing and time slot durations are flexible and scalable to support
wide-varying requirements and use cases. It is also expected to utilize NOMA in 5G.
In Fig. 1.1, we provide an overview to showcase various multiple access techniques
that will be discussed in this section. They are compared in three dimensions or
domains: power, time, and frequency.
Table 1.1 Multiple access in different generations of cellular networks
Cellular
MA technique
Duplex method
Physical
generation
resources
1G
2G
3G
FDMA
TDMA
CDMA
FDD
FDD
FDD/TDD
4G
5G
OFDMA
OFDMA
FDD/TDD
FDD/TDD
Frequency
Time slots
Time slots/PN
Codes
Time/Frequency
Time/Frequency
Notable examples
AMPS, NMT
GSM, IS-54
WCDMA
LTE, LTE-A
5G-NR
1 Introduction to Cellular Mobile Communications
5
Fig. 1.1 Overview of various multiple access techniques: a FDMA, b TDMA, c CDMA, d
OFDMA-4G
Spectral Efficiency (bps/Hz)
80
70
60
50
GSM
WCDMA/HSPA
5G
LTE
5G
40
LTE/
LTE-A
30
20
10
0
1.E+04
WCDMA/
HSDPA
GSM/
EDGE
1.E+05
1.E+06
1.E+07
1.E+08
1.E+09
1.E+10
Data Rate (bps)
Fig. 1.2 Spectral efficiencies (bps/Hz) of the digital cellular evolution
A system performance metric that continues to be improved in every generation
is spectral efficiency, bps/Hz. Figure 1.2 shows the DL spectral efficiencies of the
2G, 3G, 4G, and 5G digital cellular standards versus the peak theoretical data rates.
Notice that with each new standard the demand for higher and higher data rates along
with an increased demand for spectral efficiency becomes more pronounced.
With every cellular generation, there is not only an expectation of increased performance, but also the addition of new features. Figure 1.3 shows how the cellular
6
J. Boccuzzi
Fig. 1.3 User capabilities as
a function of cellular
generations
User Capability
5G
4G
3G
~2020
2G
~2010
1G
~1980
~1990
~2000
Cellular GeneraƟons
user capabilities (and expected features) have increased exponentially over the evolution of the cellular generations. We started with voice only and then moved on to
voice and short message service (SMS) capabilities in 2G. The data capabilities were
improved in 3G to include packet-switched services. 4G provided mobile Internet
with expanded use cases for the Internet of things (IoT), vehicle-to-everything (V2X),
device-to-device (D2D), etc. The next generation cellular system, 5G, is expected
to only increase the use case possibilities, thus opening many doors for innovative
products to be delivered.
The DL is the communication direction from the BS to the handset or user equipment (UE). The UL is the communication direction from the UEs to the BS. The
UL also consists of random access where UEs attempt to access the communication
systems resources, either from power on state or initiating a new transaction.
The method used to separate the DL and UL communication is called duplex. For
example, this operation can be performed in the time (TDD) and/or the frequency
(FDD) domains. In TDD, certain time slots are allocated to the DL and other time slots
to the UL. In FDD, the UL and DL transmission occur simultaneously in different
frequency bands. The benefits of TDD are a single spectrum is needed and shared
(no paired spectrum is necessary), and there are symmetrical channel views (UL
measurements can be used for DL communications and vice versa). The benefit
of FDD is the need for less timing synchronization requirements; however, due to
the frequency separation between the DL and UL, the UL measurements may not
be useful for DL communications as reciprocity cannot be guaranteed. Whichever
method is used, latency (time duration to access the networks resources) is becoming
more and more critical as a system performance indicator.
1 Introduction to Cellular Mobile Communications
7
1.2.1 The Evolution of Mobile Technologies
In this section, we will introduce the mobile radio access technologies (RATs) and
comprehend their evolutionary benefits and advantages. Figure 1.1 shows the cellular
standard evolution from 1G to 4G. We notice as 2G and 3G evolved, there was an
increase in system complexity across multiple standards. This changed as the industry
converged to a single 4G standard, where now there is an increase in complexity
within a single standard.
Orthogonal Multiple Access Techniques
• FDMA (frequency division multiple access)
Difficult to assign multiple carriers in the same channel
Narrowband channels (less than the coherence bandwidth of the
wireless channel) are desirable
Guard bands in frequency domain are needed to reduce spectral
emissions into adjacent frequency bands
Finite number of orthogonal resources.
• TDMA (time division multiple access)
Inter-symbol interference compensation (equalization) is needed
Uses guard bands in time domain to allow for time delay variations of
UL transmissions
Synchronization of time slots across all uses is critical to not destroy
the OMA principle
Finite number of orthogonal resources.
• CDMA (code division multiple access)
Uses the entire bandwidth at the same time utilizing spreading codes
Finite number of orthogonal resources.
• OFDMA (orthogonal frequency division multiple access)
Assigns different sub-carriers to different users (at different time slots)
Finite number of orthogonal resources.
Spectrum is very precious to the operators and remains necessary to deliver
increased system and user throughput. There is an industry-wide movement to not
only use the traditional licensed spectrum, but also embrace the unlicensed (traditionally used by WiFi devices) and the shared spectrum whenever and wherever
possible.
8
J. Boccuzzi
1.2.2 First-Generation Cellular Systems
The first-generation (1G) mobile cellular system was created to enable voice communications and support mobile users when a voice call would “hand off” to another
base station (or cell) as the mobile user physically traversed the cellular environment.
The technology used was analog frequency modulation (FM), and the spectrum was
divided into 30 kHz segments, called channels. A single user utilized the entire channel for the duration of its call. This system is called advanced mobile phone services
(AMPS) and is referred to as 1G [2].
In order to support a wide coverage area, a frequency reuse technique was introduced. Here the same frequency channels were allowed to be reused by other users, at
the same time, as long as the distance was large enough to cause minimal interference.
This interference is called co-channel interference or inter-cell interference.
In an effort to increase overall system capacity, a new technology was introduced
called narrowband AMPS (N-AMPS). Here channel spacing was reduced to 10 kHz.
Similarly, in an effort to introduce data services (which were not supported in AMPS),
cellular digital packet data was proposed which utilized frequency channels when
voice users were not present. However, it was quickly determined that an integrated
voice and data wireless network is needed to effectively and efficiently deliver such
services. Simple and robust discriminator detectors were used which were implementable, yet susceptible to random FM and deep fades from multipath observed in
the radio environment. Forcing the mobile cellular system community to move to a
different modulation technique [3, 4].
A typical cellular network architecture for 1G is provided in Fig. 1.4. Where a
cell is denoted as a hexagonal shape. To be able to increase capacity, the cells can
be divided into smaller cells, also called sectors. The mobile telephone switching
office (MTSO) connects to base transceiver stations (BTS) and the public switched
telephone network (PSTN). It also controls handovers, call routing, registration,
authentication, etc. This was a circuit switched (CS)-based network. The network
BTS
BTS
PSTN
BTS
BTS
MTSO
BTS
BTS
BTS
Fig. 1.4 1G network architecture block diagram
1 Introduction to Cellular Mobile Communications
9
used licensed spectrum to deliver the voice services, spectrum that operators purchased from the relevant governing bodies.
The 1G analog cellular standards globally deployed are listed below. Note a
single global standard did not exist.
• Advanced mobile phone services (AMPS)—US based
Analog FM modulation, FDD duplex, FDMA-based multiple access
Supports N-AMPS for narrowband, the channel bandwidth was
decreased from 30 to 10 kHz.
• Nordic mobile telephone (NMT)—The Nordic countries
Analog FM modulation, FDD duplex, FDMA-based multiple access
Channel bandwidth was dependent on the frequency band deployed:
either 25 kHz or 12.5 kHz
Supported roaming in European countries.
• Total access communication system (TACS)—UK based
Variant for Japan available (J-TACS)
Analog FM modulation, FDD duplex, FDMA-based multiple access
Channel bandwidth 30 kHz.
1.2.3 Second-Generation Cellular Systems
The second-generation (2G) mobile cellular systems were created to expand the voice
user capacity as well as to offer an integrated data services capability. The technology
moved away from analog and toward digital modulation. This shift to digital enabled
better quality voice communications via usage of voice coders (vocoders), support of
data services, initially through short messaging services (SMS), enabled encryption
to support security, and increased system capacity.
This generation created a shift from FDMA to TDMA and CDMA. These were
very interesting times the cellular users were facing; by this we mean being exposed
to incompatible 2G cellular systems. The European community was backing global
system for mobile communications (GSM), while the USA was struggling with two
competing standards: IS-54 (later renamed IS-135) based on TDMA and IS-95 (later
renamed CDMA-One) based on CDMA. All three of these cellular standards had
technical merit.
In order to increase system capacity, not only was the frequency band divided
into channels, but also time was divided into time slots for TDMA. In the CDMA
case, each user’s information was scrambled and frequency spread by a pseudonoise (PN) sequence; all users transmitted at the same time over the entire channel.
10
J. Boccuzzi
These standards used licensed spectrum purchased by network operators from the
local spectrum governing body. Receiver complexity was growing exponentially
especially when considering data rates, modulation scheme, and number of antennas
involved have increased.
The 2G digital cellular standards globally deployed are listed below. Note a
single global cellular standard did not exist.
• GSM—single standard in Europe
TDMA based
Digital modulation (GMSK), FDD duplex
Channel bandwidth = 200 kHz
Frame duration = 4.615 ms
Time slot duration = 0.557 ms (8 slots/frame)
Data Rate = 270.833 Kbps
Evolved to general packet radio services (GPRS), also considered 2.5G
Evolved to enhanced data rates for GSM evolution (EDGE), also
considered 2.75G.
• IS-54 (also called IS-136)—standard in US
TDMA based
Digital modulation (π /4-DQPSK), FDD duplex
Channel bandwidth = 30 kHz
Frame duration = 40 ms
Time slot duration = 6.67 ms (6 slots/frame)
Data rate = 48.6 Kbps.
• IS-95 (also called CDMA-One)—standard in US and Korea
CDMA based, developed by Qualcomm
Digital modulation (QPSK, O-QPSK), FDD duplex
Frame duration = 20 ms
Data rate = 115 Kbps.
These standards were all circuit switched (CS)-based networks, which over time,
had extensions (e.g., evolving from 2G → 2.5G → 2.75G) which allowed interfacing
to packet switched (PS)-based networks. Due to economies of scale, deployment
costs, patent policies, and global backing, GSM held the largest piece of the cellular
market share. The users’ appetite increased thus forcing 2G to take incremental
evolutionary steps such as 2.5G (GPRS) and 2.75 (EDGE). Both of which were
created to increase the user data rate beyond the baseline GSM capability as well as
add packet services capability. These systems are very much in use today [5].
1 Introduction to Cellular Mobile Communications
11
OSS
Core Network
VLR
BTS
GMSC
MSC
BTS
HLR
BSC
BTS
SGSN
Voice
PSTN
AUC
GGSN
Data
PDN
BSS
Fig. 1.5 2G GSM network architecture block diagram
The GSM network architecture block diagram is shown in Fig. 1.5 and is made
up of the following network elements:
• Base station subsystem (BSS) which is composed of two parts: BTSs and
base station controller (BSC)
• Operation and support subsystem (OSS) which controls and monitors the
overall GSM network
• Mobile switching center (MSC) which provides registration, authentication,
call location, call routing, etc.
• Home/visitor location register (HLR/VLR), a database of subscriber information
• Gateway mobile switching center (GMSC) obtains subscriber information
from HLR to route calls to correct MSC
• Serving GPRS support node (SGSN) for packet routing and mobility management
• Gateway GPRS support node (GGSN) organizes the GPRS network and
external packet-switched internetworking.
1.2.4 Third-Generation Cellular Systems
This third-generation (3G) digital cellular system was created to increase system
user capacity and satisfy the increasing data rate appetite. This generation provided
users the ability to surf the Internet and have simultaneous voice and data services. It
also was the ecosystem catalyst to introduce video applications to the cellular user’s
devices. Both CS and PS services were supported from its initial definition. At this
point, in the cellular evolution, mobile access to the Internet was becoming more and
12
J. Boccuzzi
more important. The MA technique shifted from using both TDMA and CDMA to
standardizing on CDMA. CDMA-One evolved into CDMA2000, and GSM/IS-136
evolved into Wideband CDMA (WCDMA).
CDMA is a multiple access technique where multiple users are separated by PN
codes and transmit at the same time over the whole bandwidth allocated. It is well
known as more users transmit, intra-cell interference grows called multiple access
interference. A power control mechanism was used in the system to not only improve
performance in a multipath fading environment, but also control the interference
introduced by each additional user in the system. Power control was the solution to
the near-far problem, with its goal of having the UE transmission flexible so that all
users received by the NodeB would have comparable energy. This created a solution
where all users were able to equally interfere with each other.
The international telecommunications union (ITU) provided 3G goals in the form
of IMT-2000 requirements. The 3GPP standards body was formed and created specifications to support implementations which satisfied these ITU requirements. The
3G cellular system continued to use the licensed spectrum. The small cells concept
was introduced in the standard and was called HomeNB. Carrier aggregation (CA)
was a seed planted into the 3G system as a method to evolve and support higher
user data rates. This seed grew and is presently benefiting the modern 4G systems.
Multiple-input and multiple-output (MIMO) spatial multiplexing was also a seed
planted into 3G where multiple streams or layers were transmitted to the user (provided the channel matrix rank requirement was satisfied). Higher-order modulation
(HOM) was also standardized; a movement from 16-QAM to 256-QAM in a land
mobile cellular system was very new during these times.
An example of a rake receiver, designed to counter the effects of multipath fading,
used for the reception of the WCDMA downlink signal is provided in Fig. 1.6. A key
WCDMA system design parameter is to have the transmission bandwidth be larger
than the coherence bandwidth of the wireless channel so that multipath (or echos)
can be used to exploit time diversity of the channel. The rake receiver consists of N
fingers which individually track multipath and demodulate the respective waveforms.
Each finger is assumed to demodulate the common pilot channel (CPICH) to support
channel estimation [6, 7].
Receiver complexity grows linearly with data rate, modulation scheme used and
number of antennas supported. In the WCDMA standard, both FDD and TDD duplex
options were provided for paired and un-paired spectrum, respectively. To aid receiver
digital signal processing, both common and dedicated pilot symbols were inserted
into the waveforms. A complete shift from non-coherent detection to coherent detection was recognized by the cellular industry.
1 Introduction to Cellular Mobile Communications
13
Searching
P()
Candidate List
Echo Profile
Manager
1 2
AcƟve List
1
...
Finger #1
Rx
Signal
N
C
O
M
B
I
N
E
R
2
Finger #2
. . . N
RAKE
Output
Finger #N
CPICH Despreader
Finger
Input
X
S*(N)
DPCH
Despreader
De-Pilot & Chan. Est.
X
C-CPICH(N)
X
X
MA
( )*
p*
Delay
X
Finger
Output
Channel CompensaƟon
C-DPCH(N)
Fig. 1.6 3G WCDMA rake receiver block diagram
The 3G WCDMA network architecture is shown in Fig. 1.7. The NodeB replaced
the BTS functions, and the radio network controller (RNC) replaced the BSC functions. WCDMA is also called universal mobile telephone system (UMTS). The
UMTS terrestrial radio access network (UTRAN) consists of NodeB and RNC groupings [8].
14
J. Boccuzzi
UTRAN
NodeB
Core Network
NodeB
RNC
NodeB
MSC
& VLR
GMSC
Voice
PSTN
HLR & AUC
SGSN
NodeB
NodeB
NodeB
NodeB:
o Power Control
o ModulaƟon
o Spreading
o Error CorrecƟon
GGSN
Data
PDN
RNC
RNC:
o Seƫng Power Parameters
o Common Channel Scheduling
o Dedicated Channel Scheduling
o Handover Control
o QoS
o Outer Loop Power Control
o Radio Resource Control
o Sets OperaƟng Frequencies
o Admission control
Fig. 1.7 3G WCDMA network architecture block diagram
The 3G cellular standards globally deployed are listed below. Note that a single
global cellular standard did not exist.
• WCDMA (also called UMTS)
Digital modulation (QPSK, 16-QAM, 64-QAM, etc.), FDD/TDD
duplex
Channel bandwidth = 5 MHz (with a chip rate = 3.84 Mcps)
Frame duration = 10 ms
Time slot duration = 0.667 ms (15 time slots/frame)
Data rates up to 1 Mbps
Defined by the 3GPP standards body.
• CDMA2000
Digital modulation (QPSK, 16-QAM, 64-QAM), FDD duplex
Channel bandwidth = 1.25 MHz × 3
Frame duration = 10 ms
Time slot duration = 0.667 ms (15 time slots/frame)
Data rates up to 1 Mbps
Defined by 3GPP2 standards.
A WCDMA high-level functional block diagram of the downlink transmitter is
shown in Fig. 1.8. Each cell has a unique scrambling code, whereas the same spreading codes (orthogonal variable spreading factor (OVSF)) are reused in every cell.
The spreading codes were also called channelization codes. The block diagram of
the uplink transmitter is also shown in Fig. 1.9. Each cell has a unique scrambling
1 Introduction to Cellular Mobile Communications
I
P-CCPCH,
S-CCPCH,
CPICH, etc.
QPSK
15
X
OVSFk
Q
X
SCH
...
I
DPCHj
QPSK
Complex
Spreading
X
OVSFj
Q
+
X
X
LPF
LPF
+
Quad Transmit
Mod Signal
...
SDL
HS-DSCH
QPSK
16QAM
64QAM
Etc.
I
X
OVSFn
Q
X
Fig. 1.8 WCDMA downlink transmitter block diagram
DPDCH (1,3,5)
HS-DPCCH
E-DPDCH
DPDCH (2,4,6)
DPCCH
HS-DPCCH
E-DPDCH
E-DPCCH
Cd
d
X
X
Chs
hs
X
X
Ced
ed
X
X
Cd
d
X
X
Cc
c
X
X
Chs
hs
X
X
Ced
ed
X
X
S
U
M
Complex
Spreading
X
LPF
LPF
Quad Transmit
Mod Signal
SUL
S
U
M
Fig. 1.9 WCDMA uplink transmitter block diagram
code, spreading codes are also reused in every cell. The difference is the uplink also
uses quadrature multiplexing between the I and Q channels [9, 10].
The WCDMA cellular system evolved to what is called high speed packet access
(HSPA)1 which consisted of both the downlink (HSDPA) and uplink (HSUPA) components. HSPA was created because an efficient way to deliver packet services was
1 The
CDMA2000 cellular system evolved to what is called evolution-data optimized (EV-DO) to
support data only extension.
16
J. Boccuzzi
Data
CRC
Turbo
Encode
HARQ
PHY
Segment
Intlv.
ModulaƟon
AdapƟve ModulaƟon & Coding (AMC)
MulƟ-Carrier
Spreading
Tx IQ
Samples
UE Feedback
Info
Fig. 1.10 HSDPA transmitter block diagram
UTRAN
NodeB
Core Network
NodeB
RNC
NodeB
MSC
& VLR
GMSC
Voice
PSTN
HLR & AUC
SGSN
NodeB
NodeB
NodeB
NodeB:
o Power Control
o ModulaƟon
o Spreading
o Scheduling
o Dynamic Resource AllocaƟon
GGSN
Data
PDN
RNC
RNC:
o Seƫng Power Parameters
o Handover Control
o QoS
o Outer Loop Power Control
o Radio Resource Control
o Sets OperaƟng Frequencies
o Admission control
Fig. 1.11 HSDPA network architecture block diagram
needed. HSPA introduced the shared channel concept and adaptive modulation and
coding (AMC) supporting hybrid automatic repeat request (HARQ). Also, in HSPA
• Whole frequency band was used (no frequency reuse greater than 1).
• Users scheduled on time slot (referred to a transmission time interval (TTI) with
a duration of 2 ms) basis and used PN codes as physical resources.
• Network architecture flattening concept was introduced to support low-latency
communications.
An HSDPA block diagram is shown in Fig. 1.10. Each user packet is protected and
transmitted to the UE; an acknowledgement (ACK) is expected to ensure error-free
communication. In the event of negative acknowledgements (NACK), the scheduler will decide which combination of coding, modulation, and physical resources
should be used to increase the likelihood of error-free communication [11]. The 3G
HSDPA network architecture block diagram is shown in Fig. 1.11. Certain functionality (highlighted in the figure) previously performed in the RNC are now performed
closer to the edge of the access network within the NodeB—supporting the network
flattening initiative.
Evolving WCDMA further became a great concern to cellular system designers.
Every known tool was being used to increase the user data rate. HOM was used to
1 Introduction to Cellular Mobile Communications
17
increase the data rate within an allowed spectral bandwidth. MIMO, in the form of
spatial multiplexing, was used to increase the data rate within an allowed spectral
bandwidth. The spectral bandwidth was also increased in the form of aggregating carriers, to increase the data rate; however, the spectral efficiency remained unchanged.
Increasing the single-carrier bandwidth brought along increased concerns. The baseline WCDMA system used a rake receiver which performs better when the processing
gain is larger, rather than smaller. This bandwidth expansion factor (signal spreading
operation), coupled with the desired high user data rates, prohibited the conventional
WCDMA system from evolving further.
1.2.5 Fourth-Generation Cellular Systems
This fourth-generation (4G) digital cellular system was created to support the exponential system capacity and data rate appetite. Much higher data rates were required to
enable mobile Internet access and video applications. Long-term evolution (LTE) is
also known as 4G and only supports PS-based networking. The standard is also evolving to use licensed, unlicensed, and shared spectrum options—all with a common
goal of increasing the user data rate, increasing system capacity, lowering latency, and
improving the user experience. The ITU provided 4G goals in the form of IMT-2010
requirements.
At this point in the cellular evolution, the industry converged to a single standard,
LTE. The LTE cellular system is based on OFDMA where the TTI has been reduced
from 2 ms (used in the 3G cellular system) to 1 ms. This TTI reduction improved
performance by being able to more quickly react to changing channel conditions, so
more efficient scheduling algorithms can be used. The reduced TTI also provided
a reduction in the end-to-end latency. The frequency bandwidth options have also
increased: 1.4, 3, 5, 10, 15, and 20 MHz to provide flexible bandwidth deployments.
To efficiently support FDMA multiple access, OFDMA (via inverse fast Fourier
transform (iFFT) and FFT operations) was chosen which divided the frequency band
into sub-channels (or sub-carriers) of 15 kHz spacing. To keep receiver signal processing complexity to a minimum, it was desirable to have the sub-carrier spacing
less than the coherence bandwidth of the wireless channel. To deliver higher data
rate services, MIMO support is mandatory to accommodate multiple layers through
spatial multiplexing [12].
Recall that with a TDMA system, the increased data rate (or decreased symbol
time duration) caused the receiver to use an equalizer to combat inter-symbol interference (ISI). The higher data rates, higher-order modulation, and longer delay spreads
caused a significant increase in equalizer complexity. With a WCDMA system, the
increased data rate (or decreased chip time duration) forced the receiver to make use
of the time diversity of the wireless channel but required a large processing gain to
adequately combat ISI. With WCDMA, the motivation was to have the transmission
bandwidth larger than the coherence bandwidth of the wireless channel; however,
for OFDM, the opposite holds true. OFDM addresses the higher data rate demand
18
J. Boccuzzi
Fig. 1.12 Time/frequency representation of the OFDM signal for LTE standard. There are four
different symbols (QPSK) each represented by one color
by generating many narrowband channels, where each narrowband channel can be
seen to experience frequency flat disturbance. This observation coupled with the fact
the frequency domain signal processing is possible, made OFDM a very attractive
multi-carrier technique to mitigate a frequency selective fading environment.
In OFDMA, users are multiplexed in both the frequency and time domains, as
depicted in Fig. 1.12 for LTE system. On the LTE air interface, the unit of allocation
is a physical resource block (PRB). A PRB is 12 sub-carriers by 7 OFDM symbols
which is equal to 84 modulation symbols. The minimum allocation to a single UE
during a subframe (1 ms) is 2 PRBs with one PRB in each slot of the subframe.
Thus, a UE will get a total of 2 PRBs/subframe which equals to 168 modulation
symbols/subframe. Note that not all these 168 modulation symbols can be used to
transmit user information, but some of these modulation symbols are used for synchronization or as pilot for channel estimation. Each PRB contains 12 sub-carriers,
and thus have a bandwidth of 12 × 15 kHz = 180 kHz. Figure 1.12 represents 2
PRB (2 × 7 symbols × 12 sub-carrier). Assuming QPSK modulation, there are four
different symbols represented by four different colors. Each color represents one
resource element (RE) and carries two bits with QPSK modulation.
A block diagram providing an example of the OFDMA waveform generation is
provided in Fig. 1.13. We also highlight the various points on the processing chain
that can significantly impact system performance. The number of sub-carriers (SCs)
has a direct impact on the data rate and user capacity of the system. From a system perspective, this value should be as large as possible; however, the occupied
bandwidth needs to be controlled via the spectral shaping function. The OFDM
symbol’s peak to average power also impacts the occupied bandwidth and imposes
linearity requirements to be met to minimize any increased spectral growth. Lastly,
note the addition of cyclic prefix (CP) removes ISI from the wireless channel.
The CP time duration should be large enough to exceed the length of the
1 Introduction to Cellular Mobile Communications
Combat ISI
User1 Data
…
UserK Data
IFFT …
P/S
CP
19
Reduce Spectral
Emissions
Spectral
Shaping
RF
SecƟon
…
*Not Standardized
by 3GPP.
(Vendor Specific)
fs = SCS x Nŏ
# of SC
SCS & Nŏ
Spectral re-growth
due to PAPR and
Non-linearity.
Fig. 1.13 OFDMA waveform generation with K sub-carrier (SC)
wireless channel time dispersion, but also as small as possible to maximize the
user data information during the subframe duration.
There are certain disadvantages with OFDM that should be addressed in future
systems, such as:
• CP overhead: The need for adding the CP introduces redundancy to the
transmitted and thus results in a loss in spectral efficiency. This loss is larger
when long CP is used or when the sub-carrier spacing (SCS) is small.
• Sensitivity to frequency and timing offsets: In order to keep the orthogonality in OFDM, the transmitter and receiver must have exactly the same
reference frequency. Any frequency offsets will ruin the orthogonality, causing sub-carrier leakage known as inter-carrier interference (ICI).
• High out-of-band (OOB) emission: OFDM assumes rectangular pulse in
time domain which is equivalent to sinc in the frequency domain which has
infinite bandwidth theoretically and cause relatively high (OOB) emissions.
The lack of spectral shaping (either filtering or windowing) is creating large
spectral side lobes in the transmit spectrum.
• High peak-to-average power ratio (PAPR): The envelope of the OFDM
waveform has a large variation which causes problems when encountering
a nonlinear device such as a transmit power amplifier. The high PAPR in
OFDM compared to the single-carrier transmission technique is due to the
summation of the many individual sub-carriers with different phases which
can results in a high PAPR when added together.
In terms of occupied bandwidth, 3GPP did not specify any spectral shaping technique in LTE and such each equipment and device vendor implements their own
solution. OFDM sub-carriers are treated as sin(x)/x, so applying spectral shaping
will help produce a more spectrally efficient waveform with minimal or no impact to
orthogonality performance. These spectral side lobes are relatively high in power due
20
J. Boccuzzi
EUTRAN
Control Plane
EPC
MME
eNodeB
HSS
eNodeB
eNodeB
S-GW
Data Plane
eNodeB:
o Packet Scheduling
o Dynamic Resource Allocation
o Load Control
o Radio Resource Management
o Admission Control
P-GW
Voice & Data
PDN
PCRF
EPC:
o Data routing
o Mobility Handling
o Sets Config Parameters
Fig. 1.14 4G LTE network architecture block diagram
to the assumed rectangular shaping. These high side lobes require a large guard band
to reduce the out of band interference. Applying spectral shaping techniques, such
as filter bank multi-carrier, universal filtered multi-carrier, etc., which are discussed
in Chap. 2 through Chap. 4 of this book, will help reduce the side lobes. The other
component can be found when viewing adjacent transmitted OFDM symbols in the
time domain; there will be phase discontinuities that also cause spectral emissions.
The peak to average power concern of OFDM can be viewed as a weighted sum
of sinusoids, which helps explain the large PAPR of the generated OFDM symbol
(as high as 12 dB). A high PAPR can be problematic if the waveform encounters
nonlinarites. Crest factor reduction is a technique used to reduce PAPR and a technique used to compensate for nonlinear distortion is digital pre-distortion. The LTE
uplink waveform uses the single-carrier FDMA (SC-FDMA) method to reduce the
PAPR impact on portable devices.
Lastly, to minimize ISI and provide the property of cyclic convolution, a small
part of the end of each symbol is added to the beginning of each transmitted OFDM
symbol. CP size depends on delay spread and LTE uses a short and a long CP. For
LTE the short CP has a value of 4.7 µs, which is approximately 8% of the symbol
time. Generally speaking, if a large delay spread is not expected to be encountered
in a particular deployment, then a lower or shortened duration CP should be used.
The 4G LTE network architecture block diagram is shown in Fig. 1.14. Note
that we now have a single global cellular standard. The evolved packet core (EPC)
replaced the core network (CN) functions, and the eNodeB replaced the NodeB
functions. The EUTRAN consists of eNodeB and EPC groupings. The EUTRAN to
EPC connection consists of both control plane and user plane signaling. This was
the beginning of an effort to separate user and control planes to allow for different
evolution rates and network deployment scenarios/options [13].
In LTE, the SCS is set to 15 kHz which translates to an OFDM symbol duration of
66.67 µs. There are 14 data symbols per time slot (1 ms), and every OFDM symbol
1 Introduction to Cellular Mobile Communications
21
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Release 8
LTE-Advanced
Release 9
LTE
MBMS
Release 10
LTE-Advanced Pro
2xCA
Relays
MBMS
DL8x8
Release 11
Relays
CoMP
CA
MBMS
[3GPP.org]
Release 12
CA
UL8x8
MTC
MBMS
CoMP
256QAM
MU-MIMO
D2D
DC
Release 13
UL64QAM
LAA
MTC
D2D
LWA
LAA
MTC
1024QAM
CA
NB-IoT
V2X
5G Phase 1
Release 14
5G Phase 2
Release 15
Release 16
Fig. 1.15 3GPP release feature timeline
requires a CP. Including the time durations of all the CPs, results in an equivalence
of 15 symbols (data + CP) in a time slot. The largest FFT size is 2048 which creates
a sampling frequency of 30.72M samples/s. Below we list some LTE features in the
3GPP standards body release schedule.
• Data rates increased by employing HOM, MIMO, and CA
• New features are added: DC, V2X, IoT, D2D, etc.
• Advanced technology to support performance improvement: coordinated multipoint (CoMP), latency reduction, beamforming
• Spectral and RAT flexibility: licensed, shared, unlicensed and LTE-WiFi
aggregation.
As discussed earlier, certain seeds were planted in the 3G cellular system to
observe how beneficial they would become to the later generations. For example,
CA continues to be useful, MIMO has become more and more essential, and HOM
is effective. In fact, all three techniques have been successfully tested and are commercially deployed; they are required to achieve the greater than 1 Gbps data rate in
LTE [14].
Figure 1.15 also reveals a departure from the typical cellular system evolution
which has been given to increasing data rates, increasing user capacity, and lowering
latency. This new trend clearly shows the additions of new services (or features, use
cases) that the industry is recognizing are required as society demands. These new
services were not really intended to be addressed when 4G was created in 2006. The
width of new expected services is growing very rapidly (as depicted in Fig. 1.3).
The network is also experiencing its own evolutionary growth. The softwaredefined networking (SDN) and network function virtualization (NFV)-based “wave
front” is departing the data center [15], making its way through the core network and on to the wireless access network. These CN and RAN workloads have
22
J. Boccuzzi
started to be implemented on a homogeneous, general-purpose CPU-based platform
(instead of the traditional dedicated logic + digital signal processor + microcontroller
approaches). This has sparked the cloud-RAN movement utilizing the Information
and communication technology (ICT) industry benefits. 4G will deploy these technologies, and when they are successful, the expectations are that 5G will be a network
upgrade.
1.3 5G Drivers, Technologies, and Spectrum
The network architecture block diagram of fifth-generation (5G) of cellular is shown
in Fig. 1.16. The 5G core (5GC) replaced EPC; next-generation radio access network
(NG-RAN) consists of distributed unit (DU) and centralized unit (CU) grouping; the
gNodeB replaced the eNodeB. As we will discuss, providing a flexible and scalable
network architecture is essential for 5G. In this theme, the combination of DU and
CU were introduced to support various RAN split options to extract the above-said
benefits [16]. The 5GC elements consists of the following:
• Access and mobility management function (AMF): performs ciphering and
integrity protection, mobility management, authentication, and authorization, etc.
• Session management function (SMF): performs UE IP address allocation and
management, selection and control of UPF, roaming, etc.
• Unified data management (UDM): performs subscription management, user data,
registration and mobility management, etc.
• Policy control function (PCF): performs policy rules for CP functions, etc.
• User plane function (UPF): performs the external interconnect point to data network, QoS handling of UP, etc.
• Application function (AF): interacts with policy framework for policy control, etc.
5GC
NG-RAN
Control
Plane
gNodeB
DU
PCF
AF
gNodeB
CU
gNodeB
DU
UDM
AMF
SMF
gNodeB
DU
gNodeB
DU
gNodeB
CU
Voice
UPF
PDN
Data
Data
Plane
Fig. 1.16 5G network architecture block diagram
1 Introduction to Cellular Mobile Communications
23
1.3.1 5G Drivers
5G cellular systems need to make a significant jump in features and performance over
LTE as incremental improvements are not wanted and do not justify the significant
capital investments operators need to commit to deploy 5G services. It is important
to note that we have maintained a single global cellular standard. The 5G driving
factors are [17]:
•
•
•
•
•
•
•
•
Increased user data rate
Increased system capacity
Massive number of connections
Reduction in end-to-end latency
Heterogenous mix of services
Flexible bandwidth deployments
Network flexibility
Move to more energy-efficient communications.
The ITU provided 5G goals in the form of IMT-2020 requirements, they are
provided in Table 1.2. For comparison purposes, we have also included the IMTadvanced requirements.
5G NR will support both non-stand-alone and stand-alone modes of deployment.
The NSA deployment will use LTE to provide wide area coverage, control and data
planes and connection into an evolving EPC. 5G services will provide high-speed data
via a dual connectivity scenario. The stand-alone deployment will provide control
and data planes as well as a connection into a 5G CN.
The ITU published the diagram shown in Fig. 1.17 to identify 5G services. The
three significant use cases (corners of the triangle) are meant to encapsulate
the expected usages of 5G in the future:
• Enhanced mobile broadband (eMBB)
• Massive machine to machine communication (mMTC)
• Ultra-reliable low-latency communications (URLLC).
These 5G use cases range from smart home, connected drones, ehealth, connected
energy, autonomous cars, real-time virtual reality/augmented reality gaming, etc.
The introduction of low latency techniques has started in LTE to aid the transition
to transforming the network in preparations for the widely varying 5G services [17].
The 5G cellular system is expected to support these usage scenarios by employing
the following technologies:
• Flexible spectrum deployments: licensed, unlicensed and shared spectrum, larger
and contiguous bandwidth, multi-RAT, etc.
• Improved network architecture: support of the ICT industry cloud trend,
SDN/NFV, network slicing, multi-access edge computing, lower latency, etc.
24
J. Boccuzzi
Table 1.2 Comparison of IMT-2010 and IMT-2020 requirements
System metrics
IMT-2010
IMT-2020
Peak data rate
DL: 1 Gbps UL: 0.5
Gbps
DL: 20 Gbps UL: 10
Gbps
Area traffic capacity
(Mbps/m2 )
Network energy
efficiency (bit/Joule)
0.1
10
1x
100x (less)
Connection density
(devices/km2 )
104
106
Latency (ms)
10
1
Mobility (kmph)
350
500
Spectral efficiency
(bps/Hz)
1x
3x (more)
User expected data
rate (Mbps)
10
100
Fig. 1.17 IMT-2020 usage scenarios
Comments
Maximum achievable
data rate under ideal
conditions
Total traffic served per
geographic area
Quantity of info bits
per unit of energy
consumption
Total number of
connected devise per
unit area
Time from when
source sends a packet
to when the
destination receives it
(end-to-end one way)
Maximum speed a
defined QoS can be
achieved
Average data
throughput per unit of
spectrum and per cell
Achievable data rate
ubiquitously available
across the coverage
area
1 Introduction to Cellular Mobile Communications
25
• Flexible numerology: Support a wide variety of use cases and spectral deployments (below and above 6 GHz bands), flexible time slots and transmission bandwidths, etc.
• Modulation and coding: QAM modulation continues to provide a reasonable
spectral and power efficiency trade-off, polar and other forward error correcting
codes.
• Advanced techniques: NOMA, full-duplex, spectral shaping, etc.
1.3.2 5G Technologies
Expected to be commercialized around 2019/2020 time frame, 5G mobile networks
are under intense reach and development activities. Compared to the current 4G
mobile networks, 5G networks are expected to support enormous system capacity,
much less latency, and about 1000 times more devices per squared kilometer, among
other requirements. To satisfy these requirements, several new technologies have
been suggested and are being developed for 5G networks. These technologies include
but are not limited to: massive MIMO, software-defined networking, mm-Wave,
cloud radio access network (cloud-RAN), non-orthogonal multiple access, M2M
communications, mobile edge computing, wireless caching, ultra-dense networks,
and full-duplex communication. In the following, we briefly describe some of these
technologies.
1.3.2.1
Massive MIMO
In discussing massive MIMO, let us first address the term “massive.” It is used to
denote the large number of antenna elements that are used in the antenna signal
processing. The number of antennas to be considered massive should be greater than
64 elements. Massive MIMO relies on the law of large numbers to make sure that the
channel and hardware imperfections (e.g., noise, fading, and hardware) average out
when signals from a large number of antennas are combined in the air together [18].
Multiple antennas afford two options in which the antennas can be used: First is to
provide an array gain by focusing energy in desired directions and nulling in unwanted
signal directions (forming a beam). Second, is to provide spatial multiplexing gain
by sending independent data streams on each antenna. Either technique can be used
to increase the overall user or system data rate. Both options are shown in Fig. 1.18
[18].
First, consider using massive MIMO for beamforming; here the antenna arrays can
be arranged in either linear, rectangular, or circular arrays that can also be stacked.
Massive MIMO will be deployed for 4G and 5G; in fact, high-frequency bands lead
to more compact, large-scale antenna arrays due to the smaller wavelength. Massive
MIMO can be deployed in either FDD or TDD duplex methods, TDD systems allow
26
J. Boccuzzi
U
E
M
I
M
O
M
I
M
O
B
F
U
E
U
E
Fig. 1.18 Massive MIMO examples: spatial multiplexing (left) and single-/multi-user beamforming
(right)
the users to invoke the theorem of reciprocity to apply what is observed on the UL
to the DL.
Next, consider using massive MIMO for spatial multiplexing, which has been
widely used for 4G and will continue in 5G deployments. Spatial multiplexing can
be achieved provided the rank of the channel matrix between the transmit and receive
antennas is greater than 1. In fact, for a 4 × 4 MIMO system, full capacity is only
observed if the channel matrix rank is full (in this case, a value of 4).
Due to the success of spatial multiplexing in LTE, it would be logical to assume
this continues for 5G and on a larger scale. This is true; however, should come with a
warning. The larger the MIMO dimension, the less likely one would experience full
rank. This means designing a 256 × 256 MIMO array and expecting to send 256 layers
to a single user, all the time is a difficult assumption to make. This is one of the reasons
5G has limited the number of DL layers per user to 8. The implementation complexity
involved in implementing massive MIMO in the digital domain is significant. Hybrid
beamforming has been proposed to provide a compromise in performance/capability
with complexity.
This brings forth an interesting question: Assuming a maximum number of layers
of 8, what can one do with the remaining degrees of freedom? Some can be used
to create (or form) beams and some can be used to multiplex other users over the
antenna array. This last comment is known as multi-user MIMO (MU-MIMO). Here,
multiple users transmit and their collective transmissions are treated as though they
came from a single source of multiplexing. The beamforming weights can create a
beam in the azimuth and elevation directions.
When considering beamforming, array gain can be used in a variety of ways. It
can be used to the extend coverage area, reduce the transmit power of devices on
the UL, improve signal-to-interference-plus-noise ratio (SINR) resulting in high user
throughput and to reduce the transmit power on the DL thus improving overall power
efficiency.
The number of antenna elements needed depends on a few items:
•
•
•
•
Array gain (coverage area, power relief, etc.)
Multiplexing layers needed
Multi-users expected to be serviced
Frequency band used (form factor, etc.)
1 Introduction to Cellular Mobile Communications
27
• Signal processing complexity (CSI estimation, analog vs. digital domain, etc.)
• System performance gains (SINR, capacity, data rate, etc.).
One of the benefits of using multiple antenna techniques, for either transmitting or
receiving, is the significant reduction in channel variation. This behavior is essential
in combating multipath fading, and having at least 64 antennas in the antenna array
significantly reduces the channel variations. Multiple 5G deployment scenarios proposed by 3GPP have varying use cases for eMBB, URLLC, and mMTC services. In
these deployment scenarios, the maximum number of DL antennas discussed was
256 and maximum number of UL antennas discussed was 32.
1.3.2.2
Software-Defined Networking
Network functions virtualization (NFV) and software-defined networking (SDN) are
supporting the movement to a software-centric network. These capabilities offer great
technical (in the form of system performance) and financial (in the form of CAPEX
and OPEX) improvements to the network operators. This movement provides the
network operators with tremendous benefits such as: a more manageable means to
monitor the network, better support of new feature roll-outs, network relocation, etc.
However, it also opens the doors for new market players (such as Internet service
giants, cable service providers, etc.) who wish to establish wireless network presence.
The adoption has been to initially virtualize the less timing critical functions, such as
in the EPC (also called vEPC) and then transition down the protocol software stack
toward the physical layer [15].
Moving to a SDN allows network operators to become nimble in deployments of
various use cases. One benefit is called network slicing. Here the network will be
able to dynamically pull together the access and core network functions necessary to
satisfy the requirements of a specific use case (latency, bandwidth, etc.). We have seen
a trend that started in 4G where a diverse set of services have emerged, and 3GPP
is addressing this demand as part of LTE’s evolution. We expect this demand to
increase and continue to create diverse requirements. The LTE network architecture
(at its conception) has been called monolithic and needs to be more flexible and
scalable as we introduce 5G services. Network slicing is a technique proposed to
support these wide variety of use cases.
Network slicing creates virtual network architectures based on SDN and NFV
principles. These virtual networks (or slices) are created on top of a common shared
physical infrastructure and can be “optimized” to meet requirements of applications,
services, or operators. The virtual networks consist of a set of network functions
instantiated to provide a complete end-to-end logical (or virtual) network to meet the
targeted performance requirements. For example, mMTC communications rely on
user capacity and not necessarily low latency, whereas autonomous cars rely on low
latency and not necessarily the highest throughput eMBB services would require.
Figure 1.19 provides a block diagram example of how the network may be sliced to
support the various 5G services discussed above.
28
J. Boccuzzi
uR-LLC
Services
Data Center
Core Network
Access Network
Front Haul
eMBB
Services
Back Haul Network
Transport Network
mMTC
Services
Fig. 1.19 Network slicing example supporting UR-LLC, eMBB and mMTC services
1.3.2.3
Multi-Access Edge Computing
To support demands for lower latency, optimizations in the 5G air interface alone
are not sufficient, we must also optimize the network. Multi-access edge computing
(MEC) is a method of moving core network or data center centric functions closer to
the edge of the network (toward the antenna) where the data will be operated upon.
It has been shown using this Principle of Relocation; the user end-to-end latency can
be significantly reduced. Additionally, the backhaul traffic can also be reduced since
the “back-and-forth” traffic has been significantly minimized by this move [19].
MEC enables cloud computing capability to be within the access network, which
is closer to the user devices. This will also be supported by fog computing. The
edge of the network is considered to be the antenna within the remote radio heads
(RRHs) which are connected to the radio access network (RAN). There are a number
of reasons to place the computing capability at the edge of the network. The most
significant reason is to reduce latency (or delay) a mobile application encounters
when trying to connect to a server. This eliminates the time a packet needs to enter
the wireless network before being acted upon. The closer the MEC server is to
the edge, the smaller the delay the applications would encounter. Examples of the
expected delays are: latencies <1 ms are needed to support industrial robots and
autonomous driving applications, latencies <10 ms are needed to support augmented
reality applications, and latencies <100 ms are needed to support-assisted driving
applications.
Figure 1.20 shows the concept of distributing the functionality which is typically
located in the CN and data center (cloud computing) to the edge (fog computing).
Besides lower application latency, we can observe lower backhaul traffic by not
sending large packets all the way into the network to be processed and then sent all
the way back to the edge [20].
MEC will perform compute and storage functionality with some market drivers
for MEC deployments being:
• Reduce total cost of ownership (OPEX and CAPEX)
1 Introduction to Cellular Mobile Communications
C-RAN
NR
29
DistribuƟng the funcƟonality.
MEC
New
Front
Haul
Device
Back
Haul
C-RAN
5G
PC
MEC
Core Network
C-RAN
MEC
Fig. 1.20 Network diagram displaying distributing computing functionality to the edge
• Increase revenue by providing ability to create new services utilizing new technology such as artificial intelligence, content distribution network, etc.
• Natural migration as virtualization proliferates out to the access network (edge
and fog)
• Improve performance (lower latency, reduce backhaul traffic).
A point also worth discussing. Why does the network edge need to be at the antennas?
We should move away from the black and white viewpoint of network/device (also
known as a cell-centric view) toward a more colorful viewpoint (also known as a
user-centric view) where the edge is more blurred. Many reports reveal the total
wireless devices are expected to be greater than 20B devices around the 2025 time
frame. We should be cognizant that the number of devices is exceeding the number
of people in the world. Also, since the computing performance of devices (handheld,
laptop, etc.) is becoming more and more complex and capable, devices should be
considered as an extension of the network—in other words the network edge.
1.3.2.4
RAN Split
The traditional and most commonly deployed fronthaul technology is based on fiber
using the common public radio interface (CPRI) protocol. CPRI carries the IQ samples between the RAN and RRHs [21]. The CPRI capabilities are being stressed to
support the evolution of LTE, especially when CA and massive MIMO deployments
are required. This stress is due to the larger bandwidth required to transport the IQ
waveform samples to the RRH, and only becomes more problematic when 5G enters
the picture. Hence, next generation front haul technology is needed to support the
expected 5G services [16].
A few front haul options exist: One solution is to standardize on another protocol
that can use higher bandwidth technologies such as ethernet-based protocols (e.g.,
25, 100 GB) while another component of the solution is to use a different RAN
split options (with lower bandwidth requirements). A few RAN split options exist
30
J. Boccuzzi
(proposed by the 3GPP) that can reduce the front haul bandwidth requirements as
well as latency, and potentially trade-off performance [16].
One RAN split option transports modulated symbols, which is a point in the
processing chain that is prior to being converted to the time domain by the iFFT
operation on the transmit side. The frequency domain sampling rate is much lower,
thus allowing more carrier-antenna combinations to be supported. This technique
still maintains a centralized processing capability to allow for more complicated
scheduling across cells. Another RAN split option transports user data packets, for
example PDCP packets. These packets have had their headers compressed and properly ciphered and protected to address any security concerns. This results in a much
lower data rate, but loses the centralized processing ability.
In addition to splitting the RAN functions, control and user planes are migrating
to become separable to allow for separate evolution rates, lower latency, and support
new deployment scenarios. This will, for example, provide the ability to have a
control plane supplied by a wide area LTE macrocell while the user plane supplied
by a small cell 5G. The 5G cellular system will also be based on OFDMA where the
time slots have been defined to be variable to handle the widely varying requirements
across all the expected services. As noticed in 4G, spectrum is extremely important
to provide higher data rates. The OFDMA parameters (sub-carrier spacing, time slot
duration, iFFT/FFT size, etc.) have been made flexible to support various spectral
deployments.
1.3.3 5G Spectrum and mm-Wave Band
LTE has a maximum bandwidth of 20 MHz, as previously discussed user data rates
have been increasing due to the use of HOM, MIMO spatial layers, and CA techniques. While present solutions support up to 5 CA, it is worth mentioning the 3GPP
LTE specifications can support up to 32 carriers. This means if we sacrifice the complexity in supporting many carriers, there is plenty of room to further increase the
data rates. In many cases, operators need to aggregate licensed and unlicensed spectrum (via license-assisted access) to reach the Gbps data rates. In fact, band number
46 (B46), whose spectral range is 5.15–5.925 GHz. is defined for that intention [22].
5G is defined to have a maximum bandwidth of 100 MHz for frequency bands
below 6 GHz. Note that large bandwidth delivers high data rates, but lower bandwidth can also provide 5G services. This coupled with the fragmented spectral band
allocations is a reason to support the need for flexibility in the OFDMA parameters discussed above. Another option besides traditional licensed and unlicensed
(5–5.9, 64–71 GHz) spectrum usage is to use the citizens broadband radio service
(CBRS) spectrum. The CBRS spectrum range is 3.55–3.7 GHz (totaling 150 MHz
of bandwidth) and is governed by a three-tiered spectrum authorization framework
to accommodate users on a shared basis with incumbent federal and non-federal
users of this band. A summary of the items that need to be considered in using 5G
frequency bands is provided in Fig. 1.21. A point worth mentioning, in these new
1 Introduction to Cellular Mobile Communications
1 GHz
31
6 GHz
o Large Range/Coverage
o Indoor PenetraƟon
o Mobile
o IoT
100 GHz
o Medium Range/Coverage
o Mobile
o Mission CriƟcal
o IoT
o Limited Range/Coverage
o Mobile
o Mission CriƟcal
Support
A Flexible & Scalable soluƟon to handle
the wide variety of Use Cases is needed.
High
Mid
Low
Low
Mid
High
Frequency
Band
Fig. 1.21 5G frequency band considerations
frequency bands the availability of paired spectrum to support FDD is minimal forcing the industry to focus more on TDD deployments. Hence, not only do we expect
bandwidth availability to vary across the low (<1 GHz), medium (<6 GHz), and
high (>6 GHz) frequency bands, but we should also expect the duplex method to
also vary.
Some operators are focusing on fixed wireless access to deliver high-speed 5G
services (approximately 1Gbps) in place of cable/fiber deployments as initial 5G
deployments in mm-Wave bands instead of, and in addition to, supporting mobile
broadband applications. This approach will help develop a mm-wave-based ecosystem that will enable 5G technologies which need to be used for battery operated
devices.
The heterogeneous spectrum usages discussed so far assumed the licensed spectrum is always used; there is an initiative to support services which only use the
unlicensed spectrum (like WiFi today). The MulteFire alliance allows LTE technology (and 5G) to be exclusively used (in a stand-alone fashion) in shared and
unlicensed spectrum to enable private services, neutral host network architecture,
industrial networks, etc.
Spectrum for 5G service will be challenging. Some of the new frequency bands
being considered in 5G NR by region is provided in Table 1.3. Operators and equipment manufacturers are faced with various options to identify spectrum (re-farm,
acquire new, partner, etc.). We see reasonable convergence (toward Global Harmonization) around the 3–4 GHz frequency bands around the world and, at the moment,
less so in the USA.
32
J. Boccuzzi
Table 1.3 New 5G frequency bands
Region
Freq. band (<6
GHz)
Europe
China
Japan
Korea
United States
3.4–3.8
3.3–3.6
3.6–4.2
3.4–3.7
3.55–3.7
<6 GHz
bandwidth
Freq. band (>6
GHz)
>6 GHz
bandwidth
400 MHz
300 MHz
800 MHz
300 MHz
150 MHz
24.25–27.35
3.1 GHz
27.5–29.5
26.5–29.5
27.5–28.35
2 GHz
3 GHz
0.85 GHz
1.4 Waveform Design for 5G
As discussed in Sect. 1.2.5, CP-OFDM has certain limitations that makes it not
the most suitable waveform for all 5G applications. However, due to its advantages
and for backward compatibility reasons, OFDM will still be the main waveform for
5G systems. On the other hand, due to its limitations, certain modifications have
been proposed in the literature to make it suitable for 5G application. Among these
limitations, fixed SCS (in 4G LTE), CP overhead, and high OOB emission are the
most important. Before listing these new waveforms, in the following, we discuss
these limitations one by one.
Internet of Things (IoT) is a main contributor to the exponential growth of users
in 5G. IoT devices, e.g., sensors, usually send sporadic short data packets and have
a limited power. On the other hand, for eMBB a large volume of data should be
transmitted in a short amount of time. Such varying characteristics of the bursts to
be transported makes CP-OFDM with a fixed SCS an inefficient waveform. For IoT
applications, 5G waveform is required to support a transmission mode with very
low air interface latency enabled by very short frames [23]. To enable low-latency
transmissions, very short TTIs are required, for energy-efficient communications by
minimizing on times of low-cost devices. OOB emission can be reduced by applying
time domain windowing that smooths the transition from one symbol to another.
As discussed earlier, the OFDM parameters have been made flexible to support
various spectral deployments. Specifically, the SCS numerology is now 15, 30, 60,
120, 240, and 480 kHz. The maximum FFT size is now set to 4096, and the maximum
number of resource blocks (RBs) that can be transmitted was also increased to 275
(or 3300 sub-carriers). Besides spectral deployment advantages, these options allow
more spectrally efficient transmissions to occur. For example, in LTE, we utilize 18
MHz of the available 20 MHz of spectrum, with the adoption of the new numerology
we are capable to utilize up to 99 MHz of the available 100 MHz of spectrum. In
considering an example 100 MHz deployment, a set of parameters can include SCS
= 30 kHz and FFT size = 4096 thus resulting in a sampling frequency of 122.88 MHz
(which is 4 times greater than LTE while utilizing 5 times greater spectrum).
Having a flexible OFDMA system is critical to efficiently deploying the wide
variety of 5G services [24]. Based on propagation characteristics, it is expected the
1 Introduction to Cellular Mobile Communications
33
lower frequency bands will be used for large-area deployments with smaller SCS
and the associated larger subframe time durations, while higher frequency bands are
expected to be used for the dense deployments with larger SCS and their associated
smaller subframe time durations. These are examples and other others can surely
exist. As can be seen, this deployment capability can be easily derived from a flexible
numerology system.
To reduce OOB emission, various filtering and windowing-based solutions are
applied to OFDM [23]. Filtered OFDM (F-OFDM), windowed OFDM (also known as
weighted overlap and add or WOLA-OFDM), universal filtered OFDM (UF-OFDM),
filter bank multi-carrier (FBMC), and other candidates have been suggested for new
waveform in 5G and beyond. These candidates will be studied Chap. 2 through
Chap. 4.
1.5 Multiple Access Techniques in 1G to 5G
Let us recall the multiple access techniques deployed in the cellular systems so far.
In the first-generation, cellular systems employed FDMA where the frequency band
was divided up into frequency channels and users were assigned channels. In the
second-generation, TDMA and CDMA were used and in both cases the frequency
band was divided into smaller frequency channels. In TDMA, the new dimension
of time was used as a resource (time slot), and in CDMA, the new dimension in
the code domain (PN sequence) was used. TDMA receiver complexity grew exponentially as the data rate was increased, modulation order increased and number
of antennas increased. In the third-generation, CDMA was deployed which utilized
larger bandwidth and more importantly introduced the concept of a shared channel.
Here the physical resources allocated to users are: time slots and PN codes. CDMA
technology complexity increased as the data rate increased. The resulting WCDMA
spread bandwidth required a larger processing gain to have reasonable inter-path
interference suppression capabilities.
The fourth-generation of cellular systems deployed OFDMA and kept the shared
channel concept. Here the physical resources were time slots and frequency subcarriers. OFDMA technology maintained the flexibility of resources and kept the
available information bandwidth at the desired value. Due to the use of the cyclic
prefix and frequency domain signal processing, the receiver complexity is manageable. It is also a reason why the fifth-generation has decided to continue with
OFDMA.
We would like to briefly discuss the differences between the DL and UL communication links; this is shown in Fig. 1.22. The DL starts with a common signal
transmitted which consists of the aggregate sum of all UEs in that cell. Each UE is
physically located in a different cell position and thus experiences different multipath fading, denoted by h i . Each UE has its own additive noise, denoted by n i . The
UL starts with individual signal transmissions that encounter different fading, due to
the physical locations within a cell. These individual signals are summed at the base
34
J. Boccuzzi
n1
+
h1
n
h1
UE1
UE1
n2
h2
.
.
.
BS
hk
+
h2
+
UE2
.
.
.
UE2
nk
hk
BS
+
UEk
UEk
Fig. 1.22 Downlink and uplink communications
R2
User 2 rate
User 2 rate
R2
User 1 rate
R1
User 1 rate
R1
Fig. 1.23 The rate region for two-user DL and UL
station receive antenna, where the base station adds its additive noise. The rate regions
of the DL and UL multiple access are shown in Fig. 1.23, for the two-user cases.
The figures compare OMA, denoted by the solid line, against superposition coding,
denoted by the dashed line. The left curve is used to showcase the DL capacity while
the right curve is used to showcase the UL capacity [25–27].
1.6 What is Non-Orthogonal Multiple Access?
In an orthogonal multiple access (OMA) system, such as TDMA and FDMA, orthogonal resource allocation is used among users to avoid intra-cell (inter-user) interference. The number of users that can be supported is then limited by the number
of orthogonal resources available. Non-orthogonal multiple access (NOMA) allows
and utilizes intra-cell interference in the resource allocation of users. Interference
cancelation techniques, such as success interference cancelation (SIC) or multi-user
detection (MUD) are used to mitigate this interference. NOMA is a technique being
considered by 3GPP in Release 16.
1 Introduction to Cellular Mobile Communications
35
OMA
Transmit Power
Transmit Power
NOMA
Frequency
Frequency
Fig. 1.24 OMA and NOMA power and spectral allocations
NOMA refers to non-orthogonal MA that can support multiple users within
a single resource and thus can improve user and overall system throughput. It
can be realized within the power domain, code domain or other domains.
Power domain NOMA (see Part II of this book) exploits the channel strength
differences between users and is the optimal capacity-achieving multiple
access technique in a single-cell network, as shown in Figs. 1.22 and 1.23.
More details and the multi-cell cases can be found in Chap. 5 and [25–27].
The spectrum and power allocation for power domain NOMA is graphically
compared with that of OMA in Fig. 1.24. In the NOMA-based systems, two
users can share the same spectral band, where each user has a different power
allocated to it.
Code domain NOMA schemes (see Part III of this book) usually exploit
low-complexity multi-user detection schemes. Sparse code multiple access
(SCMA), interleave division multiple access (IDMA), and low-density spreading (LDS)-CDMA are notable examples of code domain NOMA.
Some of the possible benefits when using NOMA are [27]:
• Massive connectivity: While OMA is limited by the number of orthogonal
resources, NOMA is not. Theoretically, NOMA can support an unlimited
number of users.
• Lower latency: OMA waits for available resource blocks to transmit which
is accomplished by waiting for an access grant whereas NOMA can support
a flexible scheduling and grant-free transmission.
• Improved spectral efficiency (bps/Hz): Every NOMA user can utilize the
entire bandwidth, whereas an OMA user can utilize a limited amount. The
data rates of properly grouped users can be increased when compared to
OMA.
36
J. Boccuzzi
The NOMA cellular system components are
• Multi-user grouping, i.e., deciding which users should be grouped together to
deploy NOMA.
• Resource allocation (power, code, etc.), e.g., for the power domain NOMA case,
users with large power differences are favorable.
• SIC or MUD interference cancelation techniques to remove the controlled NOMA
additions.
With SIC or MUD, NOMA can support this multiple access concept. As we
have seen, we expect to support an exponential growth in system capacity and user
throughput in future systems. This amount of growth presents challenges that forces
us to investigate new solutions. The choice of radio access technology plays an
important role. NOMA is a proposed scheme to address the future system demands.
1.7 Conclusion
In this chapter, we have reviewed the evolution of 1G to 5G cellular networks. A
special emphasis has been placed on orthogonal and non-orthogonal multiple access
techniques and network architectures in different generations of cellular technologies. The IMT-2020 requirements for 5G including enhanced mobile broadband, massive machine to machine communications and ultra-reliable and low-latency communications has been discussed and possible modifications such as flexible OFDM,
required to address these requirements have been briefly reviewed. A few key technical components for 5G wireless network, including massive MIMO, cloud-RAN, and
SDN, have been addressed. Advantages and issues of CP-OFDM have been listed
and possible direction for new waveform design has been outlined.
References
1. The story behind the first cell phone call ever made (Online), https://www.bloomberg.com/
news/articles/2015-04-24/the-story-behind-the-first-cell-phone-call-ever-made
2. V.H. McDonald, The cellular concept. Bell Syst. Tech. J. (1979)
3. W. Lee, Mobile Communications Engineering (McGraw-Hill, 1982)
4. J.G. Proakis, Digital Communications (McGraw-Hill, 2001)
5. J. Romero, T. Halonen, J. Melero, GSM, BPRS and EDGE Performance: Evolution Towards
3G/UMTS (Wiley, 2002)
6. F. Adachi, M. Sawahashi, H. Suda, Wideband ds-cdma for next generation mobile communications systems. IEEE Commun. Mag. 36(9), 56–69 (1998)
7. H. Andoh, M. Sawahashi, F. Adachi, Channel estimation using time multiplexed pilot symbols
for coherent Rake combining for DS-CDMA mobile radio, in Proceedings of the 8th IEEE
International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC),
vol. 3 (1997) pp. 954–958
8. H. Holma, A. Toskala, WCDMA for UMTS Radio Access for Third Generation Mobile Communications (Wiley, 2004)
1 Introduction to Cellular Mobile Communications
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
37
The 25 series of technical specification (TS) for 3G (Online), www.3gpp.org
J. Boccuzzi, Signal Processing for Wireless Communications (McGraw-Hill, 2008)
H. Holma, A. Toskala, HSDPA/HSUPA for UMTS (Wiley, 2006)
A. Ghosh, J. Zhang, J.G. Andrews, R. Muhamed, Fundamentals of LTE (Pearson Education,
2010)
S.P.E. Dahlman, J. Skold, 4G: LTE-Advanced Pro and The Road to 5G (Elsevier, 2016)
The 36 series of technical specification (TS) for LTE (Online), www.3gpp.org
M. Vaezi, Y. Zhang, Cloud Mobile Networks: From RAN to EPC (Springer, 2017)
The 38 series of technical specification (TS) for 5G-NR (Online), www.3gpp.org
ITU-R, IMT vision–framework and overall objectives of the future development of IMT for
2020 and beyond (2015)
T.L. Marzetta, E.G. Larsson, H. Yang, H.Q. Ngo, Fundamentals of Massive MIMO (Cambridge
University Press, 2016)
H. Zhang, N. Liu, X. Chu, K. Long, A.-H. Aghvami, V.C. Leung, Network slicing based 5G
and future mobile networks: mobility, resource management, and challenges. IEEE Commun.
Mag. 55(8), 138–145 (2017)
D. Bubley, Mobile/Multi-access edge computing: how can telcos monetise this cloud?
(2017) (Online), https://stlpartners.com/research/mobilemulti-access-edge-computing-howcan-telcos-monetise-this-cloud/
M. Vaezi, Y. Zhang, Radio access network evolution, in Cloud Mobile Networks (Springer,
2017), pp. 77–86
B. Ayvazian, H. Sarkissian, Spectrum Strategies for 5G. Wireless 20/20 Report (2017) (Online),
http://www.wireless2020.com/media/articles.html
F. Schaich, T. Wild, Waveform contenders for 5GOFDM vs. FBMC vs. UFMC, in Proceedings
of 6th International Symposium on Communications, Control and Signal Processing (ISCCSP)
Y. Liu, X. Chen, Z. Zhong, B. Ai, D. Miao, Z. Zhao, J. Sun, Y. Teng, H. Guan, Waveform
design for 5g networks: analysis and comparison. IEEE Access 5, 19282–19292 (2017)
D. Tse, P. Viswanath, Fundamentals of Wireless Communication (Cambridge University Press,
2005)
A. Goldsmith, Wireless Communications (Cambridge University Press, 2005)
W. Shin, M. Vaezi, B. Lee, D.J. Love, J. Lee, H.V. Poor, Non-orthogonal multiple access
in multi-cell networks: theory, performance, and practical challenges. IEEE Commun. Mag.
55(10), 176–183 (2017)
Chapter 2
OFDM Enhancements for 5G Based
on Filtering and Windowing
Rana Ahmed, Frank Schaich and Thorsten Wild
2.1 Motivation
One of the main drivers of new radio (NR) is the huge market opportunity in the
Internet of Things (IoT) applications [1]. It is predicted that such verticals will evolve
as well as their needs. New services will emerge which cannot be efficiently served
with any of the available dedicated solutions. The killer application (service) in 2030
cannot be easily predicted 10 years in advance; hence, forward compatibility is a
design principle in NR. Such a killer application will not be most efficiently served
by a fixed design, a configurable multiservice air interface is therefore the solution.
In contrast to previous long-term evolution (LTE) releases, which mainly target
serving broadband users and to which serving verticals came only as an afterthought,
e.g., narrow band IoT (NB-IoT) in Releases 13 and 14 [2, 3], 5G NR aims at serving
verticals as a basic system capability in addition to broadband users. Consequently,
the use cases considered by 5G NR are more diverse. Beyond enhanced mobile
broadband (eMBB), massive machine communication (mMTC) and ultra-reliable
low latency communication (URLLC) have to be supported. For example, NR targets applications with limited battery capability, which demand less stringent time
synchronization requirements, and at the same time NR targets applications which
are very sensitive to time delay and thus require shorter symbol transmission time. To
make it possible, various considerations for the radio access in general, and for the
design of the waveform in particular, have to be accounted for, as will be discussed
in the upcoming sections.
R. Ahmed (B) · F. Schaich · T. Wild
Nokia Bell Labs Stuttgart, Lorenzstrasse 10, Stuttgart, Germany
e-mail: rana.ahmed_salem@nokia-bell-labs.com
F. Schaich
e-mail: frank.schaich@nokia-bell-labs.com
T. Wild
e-mail: thorsten.wild@nokia-bell-labs.com
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_2
39
40
R. Ahmed et al.
2.1.1 Multi-carrier Transmission
The selection of the applied waveform is one of the most fundamental design decisions to be taken. It defines the temporal and spectral characteristics of the transmit
signal. The resulting time domain peak-to-average power ratio will impact the power
amplifier design and hence is responsible for the energy efficiency of communication.
The resulting transmit signal spectrum impacts spectral efficiency and the coexistence with other communication systems. Waveforms will carry modulated symbols,
so their design will impact the multiple access possibilities and frame structure design
for multiplexing data symbols, pilot symbols, and blocks of control symbols.
With conventional single carrier transmission, each transmitted symbol occupies
the whole transmission bandwidth. As the transmission bandwidth increases, and
thus the symbol length Ts gets shorter, the channel delay spread τmax becomes more
significant. This leads to a distortion caused by inter-symbol interference (ISI), where
the sum of several delayed replicas of the transmitted symbol is received. To compensate for this effect, multi-tap equalizers have to be employed at the receiver side,
e.g., nonlinear decision feedback equalizer (DFE) [4].
In general, if the symbol duration Ts is much larger than the maximum delay spread
Ts ≫ τmax or alternatively the signal bandwidth is much smaller than the coherence
bandwidth Bs ≪ Bc , the channel is considered as a “flat” channel; i.e., the received
signal can be considered to be a version of the transmitted signal weighted by a
complex scalar factor. Thus very low equalization effort is required. Indeed, this
is exactly the basic idea of a multi-carrier system. The idea is to divide the total
bandwidth of the signal Bs into smaller subchannels, referred to as subcarriers. Such
that each subchannel bandwidth is equal to Bsc = BNs , referred to as subcarrier spacing
∆f . The information is transmitted in parallel over these subcarriers. If the number of
subcarriers, N , is large enough for a given overall bandwidth such that ∆f ≪ Bc , the
channel experienced over every subcarrier is “flat,” and hence, a single-tap equalizer
is sufficient to compensate the channel distortion in the frequency domain.
The idea of multi-carrier transmission has emerged a long time ago as early as
the 1966 paper of Chang [5]. However, it was only considered for practical implementation when an efficient implementation using the fast Fourier transform (FFT)
[6] was proposed.
Multi-carrier modulation is therefore a favorable choice in channels with long
delay spread, since it avoids the high computational complexity needed with the
single carrier equalization. The inherent serial-to-parallel conversion of multi-carrier
modulation naturally offers a basic delay spread protection, which can be extended
by using a cyclic prefix or zero postfix.
Furthermore, multi-carrier modulation allows frequency-selective channel access,
which exploits high gain links while avoiding fading dips. The decoupling into narrowband subchannels is very appealing for multiple input multiple output (MIMO)
antenna processing techniques. The multi-carrier flexibility in time–frequency multiplexing allows good design properties for frame structures, including the multiplexing of pilot symbols and control information. Pilot symbols design can be flexibly
2 OFDM Enhancements for 5G Based on Filtering and Windowing
41
tailored to the coherence bandwidth and coherence time of the radio propagation
channel.
All those benefits made multi-carrier modulation the technique of choice in 4G
LTE, which uses orthogonal frequency division multiplexing (OFDM) in the downlink (DL). The merits of OFDM, however, come at the price of an increased peakto-average power ratio (PAPR), compared to single carrier transmission. In fact,
as the number of subcarriers increases, the PAPR increases as well. That is why
single carrier frequency division multiple access (SC-FDMA) is used in the uplink
(UL) transmission of LTE, instead of OFDM. Note, however, that via a discrete
Fourier transform (DFT) precoding, OFDM can be transformed into SC-FDMA, so
the established multi-carrier processing techniques can be reapplied to single carrier
modulation.
To provide the required orthogonality1 between the different subcarriers:
• Every OFDM symbol is appended at the beginning by a guard interval of NG I
samples, where NG I is designed to be larger than the channel delay spread. The
guard interval contains either:
1. A duplicate of the last NG I samples of the OFDM symbol and hence is referred
to as cyclic prefix (CP).
2. Or NG I zero samples and hence is referred to as zero prefix (ZP).
The guard interval is important to avoid inter-block interference (IBI) between
successive OFDM symbols. Therefore, the symbol time (in samples) is equal
to Ns = N + Nc in case of CP-OFDM or Ns = N + Nz in case of ZP-OFDM,
where Nc and Nz are the number of samples in the cyclic prefix and the zero
prefix, respectively.
• The subcarriers are arranged in the frequency domain such that the frequency
spacing between the subsequent subcarriers is
∆f =
1
1
=
Tu
NT
(2.1)
where T is the sampling period between two successive samples in the time domain
and Tu = N T is the useful symbol period in time domain. Such an arrangement
guarantees that, at the frequency sampling point of any subcarrier q, no other
′
contribution from any other subcarrier q = q exists after removal of the CP, i.e.,
orthogonality between the different subcarriers.
Assuming an UL transmitter to which Q subcarriers are allocated, transmitting at a
symbol rate T1s , where Ts = Ns T , the output of a multi-carrier CP-OFDM transmitter
1 Orthogonality
subcarriers.
here means that no crosstalk occurs in the detection process between the different
R. Ahmed et al.
Ɵ
Ɵ
42
Fig. 2.1 Transmit–receive chain for one subband in an OFDM system
at time instant n can be written as [7] (after dropping the DFT spreading for ease of
representation)
x (n) =
Q−1
i
sq,i w (n − i Ns ) e j2π fq (n−Nc −i Ns ) ,
(2.2)
q=0
where w (n) is a rectangular window function, holds the value of 1 over the interval
[0, N + Nc ]. sq,i are the i.i.d. complex-valued symbols transmitted at subcarrier frequency f q and symbol i. Figure 2.1 shows the block diagram of the OFDM transceiver
for one subband.
In frequency domain, the window function for one subcarrier in Eq. 2.2, w (n), is
written as
sin (π f (N + Nc ))
W ( f ) = (N + Nc )e− j2π f (N +Nc )
,
(2.3)
π f (N + Nc )
m
As shown in Fig. 2.2, the nulls of W ( f ) for subcarrier q occur at f = f q + (N +N
,
c)
where m is an integer not equal to zero.
It is worth noting that in ZP-OFDM, the nulls of W ( f ) for subcarrier q occur at
f = f q + mN . Therefore, one can find that the power spectral density of ZP-OFDMbased waveforms have “true” nulls in the frequency domain, since the subcarrier
spacing of the transmitted subcarriers, according to Eq. 2.1, coincides exactly with
the nulls of W ( f ).
The first notable multi-carrier modulation technique came up even before OFDM;
filter-bank multi-carrier (FBMC) [8]. FBMC is a consequent application of Gabor
signaling [9] generated by an orthogonal train of time–frequency shifted pulses. In
2 OFDM Enhancements for 5G Based on Filtering and Windowing
Fig. 2.2 Windowing
function in frequency
domain W ( f ) with
CP-OFDM
43
1
0.8
sinc(fN s)
0.6
0.4
0.2
0
-0.2
-0.4
-3
-2
-1
0
1
2
3
fNs
order to make it spectrally efficient, offset-QAM is used [10]. The overlapping symbols generated by long filters create an impressive spectral containment of signals.
However, this comes at the price of suitability for short bursts and loss in multiplexing
flexibility [11].
2.2 5G Waveform Requirements and Scenarios
NR is targetting a diverse set of use cases [12, 13]. It is foreseen that NR will have to
support a wide range of user velocities, data rates, reliability, and power efficiency
requirements. In order to be able to configure the waveform parameters to match the
requirement of every use case, NR is supposed to support subcarrier spacing scaling
principle of ∆f = 15 × 2 S kHz, where S is an integer, unlike LTE which supports
only ∆f = 15 KHz.
The coexistence of all these services (waveform configurations) together is essential for an efficient use of resources and to be able to adapt to traffic load changes.
Since the waveform is a fundamental component in the design of the air interface,
NR waveform should be designed to facilitate that coexistence. In other words, NR
waveform should be robust enough against possible inter-carrier interference (ICI)
distortions caused by the support of different services, which will be detailed in
Sects. 2.2.1 and 2.2.2
As mentioned earlier, in CP-OFDM with perfect synchronization, only a frequency domain one tap zero-forcing (ZF) equalizer is sufficient at the receiver side
to equalize the effect of the channel. However, in reality, ICI can occur at the receiver
side due to Doppler distortions caused by temporal channel variations or synchronization errors, etc. In such case, the relatively slow decay of the sinc waveform in
Fig. 2.2 is especially problematic and the overall performance is not adequate. Therefore, in NR, it is desirable to design waveforms with faster decay rates in frequency,
i.e., better frequency localization. In addition, waveforms with higher frequency
44
R. Ahmed et al.
localization than baseline OFDM can achieve higher spectrum utilization than 90%,
which was the maximum achieved in LTE [14]. A further driver for higher spectral
confinement is the created forward compatibility; i.e., any kind of signals/waveforms
which are favorable for a future use case can be inserted in the evolution of the 5G
standard, as well-defined in-band requirements allow for “cleaning-up” the spectrum from sidelobes much better than regular OFDM could do. As will be shown in
Sects. 2.3.1–2.3.3, the design is a trade-off between time and frequency localization.
2.2.1 Mixed Numerology
In LTE, OFDM parameters, namely cyclic prefix length and subcarrier spacing,
are selected as a reasonable compromise for different transmission scenarios (e.g.,
Doppler spread vs channel delay spread). In NR, because of the extreme use cases,
more configuration options are available to serve each use case most efficiently [15].
For example, on one hand with use cases requiring URLLC, in order to save on
the latency part, one option is to reduce the symbol lengths. This corresponds to
a wider subcarrier spacing. Similarly, users who travel at a very high speed (e.g.,
high-speed trains), can benefit from having a large subcarrier spacing, reducing the
interference arising from Doppler spread. On the other hand, for low-end devices,
to enhance the coverage, or for users in a high channel spread environments, longer
symbol durations (consequently longer CP) and smaller subcarrier spacings are more
favored.
For carrier frequencies below 6 GHz, Release 15 supports subcarrier spacings of
15, 30 and 60 kHz [14]. NR supports the multiplexing of these numerologies in UL
and DL [16]. On one carrier, NR supports mixed numerologies in frequency division
multiplex (FDM) or time domain multiplex manner. As discussed in [15], FDM of
different numerologies in neighboring subbands generates ICI at the edge between
the two different numerologies, which can be explained as follows:
Assuming two neighboring allocations in FDM with two numerologies, namely
(∆f 1 , Ts1 ) and (∆f 2 , Ts2 ), such that ∆f 2 = 2∆f 1 and Ts2 = T2s1 .
1. As shown in Fig. 2.3, allocation 1 suffers from ICI because the neighboring
allocation (with larger subcarrier spacing) have now nonzeros contributions at
its own subcarrier positions.
2. The ICI on the larger numerology can be better understood in time domain,
where the symbol of allocation 2 is half the duration of the symbol of allocation
1, Ts2 = T2s1 . Therefore, the demodulator of allocation 2 collects only half the
samples of allocation 1. The effect is equivalent to multiplying with a time
domain window, which is equivalent to a convolution with the frequency response
of the window [15].
2 OFDM Enhancements for 5G Based on Filtering and Windowing
45
1
sinc(fT s), sinc(2fTs)
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-5
0
5
fT s
Fig. 2.3 ICI on smaller numerology
2.2.2 Asynchronous Uplink Transmission
The OFDM orthogonality in LTE UL requires that uplink transmissions from various
users are synchronized at the BS; i.e., the different symbols from the UEs arrive at
the BS within a certain time window which does not exceed the CP length. To
compensate for different propagation delays, a user equipment (UE) would apply a
timing advance [2] depending on how far it is from the BS, which means that UE
devices which are far from the BS send their UL signals earlier than those close to
the BS. Therefore, any device wanting to transmit a few bits of data has to enter
the network via the random access procedure, wherein Msg 2 of the random access
procedure, the BS informs the UE about its timing advance value (Fig. 2.4).
The whole random access procedure includes a closed loop timing advance control, which in some cases may not be suitable for low-end machine-type communication (MTC) devices, dealing with sporadic traffic and stringent requirements on
energy efficiency. The lack of UL synchronization, as shown in Fig. 2.5, creates ICI
between neighboring UL users as explained in [17, 18]. The effect is actually similar
to the windowing effect mentioned in Sect. 2.2.1.
2.3 Candidate 5G Waveforms
As discussed in Sects. 2.1–2.2.2, there is a strong need for a waveform design in
NR, which is robust against time–frequency misalignments. NR has selected CPOFDM/DFT-s-OFDM as the baseline waveform including the optional addition of a
windowing or a filtering functionality [16, 19]. However, such improvements should
46
R. Ahmed et al.
0.05
2 'short' symbols
0.045
CP
CP
0.04
N2
L2
N2
L2
envelope
0.035
0.03
1 'long' symbol
0.025
0.02
0.015
CP
0.01
N1
L1
0.005
0
0
100
200
300
400
500
600
700
800
900
1000
time index
Fig. 2.4 ICI on larger numerology
Fig. 2.5 Multi-user scenario
(FDMA)
be agnostic to the UE/BS in Release 15 [20], meaning that a baseline CP-OFDM/DFTs-OFDM receiver should work seamlessly without prior knowledge of the method
used at the transmitter to reduce the out-of-band leakage (OOB). To enable a seamless multiplexing of different services, the used waveform should abide by in-band
requirements, which are currently discussed in RAN4 [21]. These requirements
can be met either by using gaps between the different subband allocations (guard
2 OFDM Enhancements for 5G Based on Filtering and Windowing
47
Fig. 2.6 Options for modifying the OFDM transceiver
subcarriers) or by using a frequency localized waveform. Figure 2.6 shows the options
for modifying the baseline CP-OFDM transceiver in Fig. 2.1 to achieve this goal.
In this section, we discuss waveform examples which can be used in NR. Two main
classes appear in this context; subband filtering and windowing, where the latter can
be mapped to subcarrier filtering [7]. Subband filtering is motivated by the fact that
ICI typically occur at the edge between neighboring subband allocations (blocks of
subcarriers). For example, different uplink users having different waveform configurations/requirements. Therefore, the idea is to apply a well frequency localized filter,
the bandwidth of which is close to the subbands bandwidth. As a result, only a few
subcarriers close to the edges of the subband in frequency are affected by the filter, as
the filter suppresses their out-of-subband sidelobes. By adapting filter parameter, the
distortion on subband edges can be alleviated with little negative impact, depending
on the use case. We discuss two waveforms which belong to this category, namely
UF-OFDM (a.k.a UFMC) and f-OFDM.
Windowing, on the other hand, is applied in the time domain, by modifying the
rectangular pulse shape of CP-OFDM waveform to have smoother transitions in time
at both ends. As an example of this category, we discuss weighted overlap and add
(WOLA) waveform. Other candidate windowing techniques exist [22], but will not
be discussed here.
Other waveforms which gained a lot of attention, but were not considered for
Release 15 due to incompatibility with CP-OFDM, are FBMC, mentioned at the
beginning of this chapter, and zero tail discrete Fourier transform (ZT-DFT-s) [23].
All the methods mentioned in this chapter are compatible with DFT spreading as
shown in Fig. 2.6.
48
R. Ahmed et al.
Fig. 2.7 WOLA waveform
Fig. 2.8 WOLA transmitter operation
2.3.1 Weighted Overlap and Add (WOLA)
In CP-OFDM with WOLA, the windowing function w (n) in (2.2) is replaced by a
pulse
function with soft edges at both sides, the length of the window is extended
to − L2wt , N + L2wt [24, 25], where L wt is the length of the extension beyond the
CP-OFDM length. The soft edges at the beginning and end of the window function
result in better localization of the WOLA waveform in the frequency domain.
In [24, 25], the CP-OFDM symbol is first extended by a cyclic extension in
the time domain, and both edges are shaped by a weighting function, as shown in
Fig. 2.7. As shown in Fig. 2.8, the resulting symbol is overlapped and added to the
next symbol, and hence the overhead remains the same as in CP-OFDM.
It is worth noting that, although the WOLA symbols overlap within one burst,
when considering TDD, the tails from the end of the last WOLA symbol of one burst
and from the first WOLA symbol of the following burst would already extend into
the guard period (GP). In TDD, the GP is placed between two successive bursts when
DL/UL switching is made. Therefore, the window length should be chosen carefully
so as not to hinder TDD transmission.
At the receiver side, a WOLA receiver can optionally be applied to suppress ICI,
leaked from a neighboring allocation [24, 25]. The WOLA receiver processing is
shown in Fig. 2.9.
2 OFDM Enhancements for 5G Based on Filtering and Windowing
49
Fig. 2.9 WOLA receiver operation
The design of the soft edges of w (n), including the overlap length L wt , determines
the frequency domain behavior of the WOLA symbol. In [24, 25], a raised cosine
window design is used, but this does not exclude other possible window shapes.
In general, the design is a trade-off between time and frequency localization, i.e.,
between ISI and ICI. The longer the window length L wt , the better the ICI localization
of the WOLA waveform is, but the longer the overlap between the successive WOLA
symbols, and the less robust the WOLA symbol is to channels with long delay spread.
Figure 2.10 shows the power spectral density (PSD) of WOLA for a subband of 2
PRBs with two different window lengths L wt = 144, 72 samples. Compared to CPOFDM (L wt = 0), we can see that windowing reduced the OOB leakage of the
waveform.
2.3.2 Universal Filtered OFDM (UF-OFDM)
UF-OFDM is a 5G candidate waveform, also known as universal filtered multi-carrier
(UFMC), where blocks of subcarriers (subbands) are filtered. As shown in Fig. 2.6,
this is done by passing the OFDM signal output for user k through a subband filter
f k (n) with filter order L f . f k (n) is built by shifting a prototype filter f (n) to the
center of subband of user k. The modified subband filtered OFDM signal for user k
can be written as
50
R. Ahmed et al.
Fig. 2.10 PSD of WOLA
waveform for different
window lengths at a subband
allocation size of 2 PRBs
10
WOLA Lwt =144
WOLA Lwt =72
0
CPOFDM
PSD
-10
-20
-30
-40
-50
0
10
20
30
40
50
60
freq index
xk (n) =
i
⎡
f k (n) ∗ ⎣
Q−1
q=0
⎤
sk,q,i w (n − i Ns ) e j2π fq (n−Nc −i Ns ) ⎦.
(2.4)
The input to the UF-OFDM subband filter can be either a ZP-OFDM signal or
a CP-OFDM signal (depending on the value of Nc in Eq. 2.4). The advantage of
applying the subband filter on a ZP-OFDM signal is that the resulting overall symbol
can be limited in time domain, such that no overlapping between the successive
symbols occur in an ISI-free environment (if the subband filter order is smaller than
the guard interval L f < NG I ). The prototype filter of choice in UF-OFDM is the
Dolph–Chebychev filter, but it is not restricted to this selection.
If the subband filter is applied at the transmitter on a ZP-OFDM signal, at the
receiver side, the L f samples from the tail of the received signal are simply added
to the beginning of the symbol before applying the FFT. Hence, even if a ZP-OFDM
signal is used, no extra complexity is needed at the receiver side to demodulate the
signal as explained in [26]. In [7], it was shown that subband filtering applied on CPOFDM signal has near identical rate-versus-SNR performance to subband filtering
applied on a ZP-OFDM signal.
In [27], it was proposed to use Dolph–Chebyshev-based subband filtering in combination with variable CP/ZP. The summation of the CP plus ZP parts was assumed
to be constant, and the filter order matches the CP length. Each part (CP/ZP) can be
tuned depending on the channel environment, i.e., to optimize the trade-off between
ISI and ICI rejection. However, in Release 15, the NR waveform is based on CPOFDM [16], therefore only CP-based UF-OFDM implementation can be supported
in Release 15.
Dolph–Chebychev filters are optimal in the sense that for a given side lobe level
(SLL) the main lobe width is minimized. They are adjustable by the tuning parameter
for the side lobe attenuation (SLA) as well as by the filter length L. The optimum
filter choice of L and SLA depends on the use case. For example, on the one hand,
in high ICI use cases with asynchronous transmission, it makes sense to use filters
2 OFDM Enhancements for 5G Based on Filtering and Windowing
Fig. 2.11 Time domain
impulse response for
Dolph–Chebyshev filter with
L = 72 and S L A = 35, 60
dB
51
1
SLA=35dB
SLA=60dB
0.8
0.6
0.4
0.2
0
-3
-2
-1
0
1
2
3
t[ s]
Fig. 2.12 Frequency
domain response for
Dolph–Chebyshev filter with
L = 72 and S L A = 35, 60
dB
SLA=35dB
SLA=60dB
0
-20
-40
-60
-80
-100
-120
-1000
-500
0
500
1000
frequency [kHz]
which are longer than the guard interval L > NG I , at the price of higher vulnerability
to delay spreads [7]. On the other hand, in environments with high delay spread, a
shorter filter length is used to protect against ISI. The SLA controls the trade-off
between the main lobe width and the SLL. As shown in Figs. 2.11 and 2.12, as the
SLA increases, the main lobe width increases and the SLL decreases.
In general, for any subband filtering technique, when the allocation size is larger
than a single PRB, a subband filter which has a broader passband offers a steeper
side lobe level decay at the pass-band edge.
For equalization at the receiver side, the receiver has to be aware of the filter
coefficients. In order to avoid that, pre-equalization of a UF-OFDM signal is proposed in [28] for reference signals, but can also be applied to data symbols [26]. In
addition, equal transmit power per subcarrier prevents an increased number of error
events at the pass-band edge, as those errors from weaker edge subcarriers, would
be detrimental for bit error rate (BER) performance.
52
R. Ahmed et al.
To further reduce the effect of ICI leaked from a neighboring allocation due
to mixed numerology or asynchronous transmission, etc., windowing or matched
filtering can be used at the receiver side of the UF-OFDM [7, 18, 29].
2.3.3 Filtered OFDM (f-OFDM)
Filtered OFDM (f-OFDM) is a 5G candidate waveform based on subband filtering
of a CP-OFDM signal [30, 31]. Therefore, the f-OFDM signal can also be built
according to (2.4). Compared to UF-OFDM, the key property of f-OFDM is that
the filter length, L f , can well exceed the guard interval length, which enables it to
provide very good frequency localization. As shown in [30], soft truncation of a
prototype filter is used, which in this case is a sinc impulse response p B (n). The sinc
impulse response should have a bandwidth B in the frequency domain equal to the
subband allocation size. The subband filter is obtained by applying a time windowing
mask to p B (n)
(2.5)
f k (n) = w(n) p B (n),
where w(n) is the windowing mask with duration Tw . The windowing mask has to
have smooth transitions to zero on its both ends so that it avoids abrupt jumps at the
beginning and end of the truncated filter. An example of such a windowing function
is the Hanning window as proposed in [30] or the raised cosine window as proposed
in [32]. Tw is usually chosen as Tw = T2u .
In Figs. 2.13 and 2.14, the frequency domain response and the time domain
impulse response of the designed filter for f-OFDM with bandwidth equal to 2 PRBs
(360 kHz) and 3 PRBs (540 ‘kHz) are depicted, where the Hanning window function
is used [30]. The filter length should be long in order to achieve desirable frequency
localization. With half OFDM symbol filter length for example, one f-OFDM symbol
extends into 25% of each of the previous and following f-OFDM symbols. However,
most of the energy of the time domain impulse response, for allocation sizes greater
than 3 PRBS, is limited to the CP part which is eventually dropped at the receiver
side.
Similar to WOLA, the long tails of the f-OFDM is also a problem in TDD transmission. Especially with long filter lengths, Tw = T2u . That is why in [32], signal
burst tail reduction is proposed, where a hard truncation is suggested to reduce the
tail overhead at both ends.
In [32], one design criteria is to choose the sinc filters bandwidth B to be larger
than the subband bandwidth W by a small excess bandwidth ∂ W , called tone offset
(TO), on each side, i.e., B = W + 2 × ∂ W . The motivation behind using the TO
is to guarantee a flat passband across all used subcarriers. The TO in f-OFDM is
then analogous in use to the pre-equalization used on UF-OFDM, as mentioned
in Sect. 2.3.2. The benefits of the TO in f-OFDM and the pre-equalization in UFOFDM come at the expense of a larger mainlobe (or higher SLL), and thus, a weaker
frequency localization.
2 OFDM Enhancements for 5G Based on Filtering and Windowing
Fig. 2.13 Frequency
domain response of designed
filter for f-OFDM with 2
PRB and 3 PRBs
53
B=540kHz
B=360kHz
0
-20
-40
-60
-80
-100
-120
-1000
-500
0
500
1000
frequency [kHz]
Fig. 2.14 Impulse response
of designed filter for
f-OFDM with 2 PRB and 3
PRBs versus sampling time
1
B=540kHz
B=360kHz
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-20
-15
-10
-5
0
5
10
15
20
t[ s]
f-OFDM is often described in conjunction with a matched filter, which suppresses
the ICI leaked into the UE subband from neighboring UEs. In addition, it maximizes
the received signal-to-noise ratio (SNR) of the UE.
2.3.4 Comparison Between the Different Waveforms
In this section, we compare the performance of the four different waveforms discussed in this chapter, namely CP-OFDM, WOLA, UF-OFDM (CP based), and
f-OFDM. Table 2.1 depicts the waveform parameters which were assumed for the
different preprocessing approaches applied on top of the baseline CP-OFDM signal.
54
R. Ahmed et al.
Table 2.1 Waveform parameters
Waveform
Baseline CP-OFDM
WOLA
CP-UFOFDM
f-OFDM
Parameters
N G I = 72 samples, FFT size N = 1024, one
subband = 1 and 4 PRBs
L wt = 144
SLA=25dB and 75dB for subband sizes 1 PRB
and 4 PRBs, respectively, L f = 73 samples,
pre-equalization applied
L f = 513, TO=0 and 2 and 4, Hanning
window used for soft truncation of sinc impulse
response
Fig. 2.15 UL polynomial power amplifier model: gain and phase distortion
2.3.4.1
Power Amplifier Model
The power amplifier, at the RF frontend, introduces signal nonlinearities in the final
transmitted signal. Hence, the effectiveness of the OOB emission reduction by subband processing is eventually limited by the spectral regrowth due to these nonlinearities. As the PAPR increases, the spectral regrowth increases for the same
power amplifier efficiency.
Third-generation partnership project (3GPP) uses modified Rapp model and polynomial model to model the effect of the power amplifier in NR DL and UL, respectively [33]. Figure 2.15 shows the UL polynomial power amplifier model gain and
phase distortion curves. In order to operate at an output power of 22 dBm, an input
power operating point of −5.12 dBm is assumed, as shown in Fig. 2.16.
As shown in Fig. 2.17, the error vector magnitude (EVM), used to describe the
signal quality [34], reaches a minimum point at a phase compensation value of
−77.1◦ . As mentioned earlier, one of the drawbacks of multi-carrier OFDM is the
increase in the PAPR. The subband processing, based on windowing or filtering,
leads to a further increase in the PAPR of the input signal to the power amplifier,
which can also be observed in Fig. 2.17, where CP-OFDM has the lowest EVM value.
2 OFDM Enhancements for 5G Based on Filtering and Windowing
55
Fig. 2.16 Output power in dBm of amplifier versus input power in dBm for UL power amplifier
Fig. 2.17 EVM versus
phase compensation value
for all waveforms for UL
power amplifier
7
CPOFDM
UFOFDM
6.5
WOLA
FOFDM
FOFDM TO=4
Averaged EVM [%]
6
5.5
5
4.5
4
3.5
-79
-78.5
-78
-77.5
-77
-76.5
-76
-75.5
-75
phaseComp[degrees]
2.3.4.2
Performance Comparison
In Figs. 2.18 and 2.19, we can see the PSD of all waveforms before and after the
UL power amplifier. As we can see, with the given waveform configurations, subband filtering (f-OFDM followed by CP-UFOFDM) have better spectral localization compared to WOLA and is therefore more effective in dealing with frequency
misalignments. After the power amplifier, the relative performance of the different
waveforms remain the same, but the gap between them is significantly reduced. A
waveform with high spectral localization translates into higher spectral efficiency,
56
R. Ahmed et al.
Fig. 2.18 PSD with 1 PRB
allocation: solid lines are
PSD without PA, dashed
lines are PSD with PA
10
CPOFDM
CPOFDM
WOLA
WOLA
CP-UFOFDM
CP-UFOFDM
FOFDM
FOFDM
FOFDM-TO2
FOFDM-TO2
0
PSD
-10
-20
-30
-40
-50
-60
0
10
20
30
40
50
60
freq index
since with high spectral localization, relatively smaller number of in-band guards are
needed to satisfy the in-band requirements of NR.
Figures 2.18 and 2.19, however, do not show the performance in a synchronous
environment, especially with high delay spread. Results for this case in [35, 36] show
an opposite ranking for the three waveforms in a synchronous environment with high
delay spread, where WOLA shows the best performance followed by CP-UFOFDM
and then f-OFDM. In [37], it is shown that by taking into consideration, the reduced
guard band overhead required in case of well spectrally localized f-OFDM, f-OFDM
without using a matched filter can have a higher spectral efficiency than than of
WOLA. No comparison was made in [37] against UF-OFDM. In [7], it is shown
that both windowing and filtering techniques are comparably effective in combating
channel time–frequency misalignments depending on the selected waveform configuration, with filtering techniques showing a slightly superior performance.
2.3.4.3
Implementation Aspects
WOLA requires an additional complexity over baseline CP-OFDM depending on
the window length, which is L 2wt multiplications, as shown in Fig. 2.7. Similarly, if a
WOLA receiver L 2wr is applied to suppress ICI from a neighboring allocation, extra
L 2wr multiplications are needed on top [36]. Therefore, the total added complexity
is that of the overlap operation and L 2wt +L 2wr .
For subband filtering, higher computational complexity is required, and therefore,
a range of low-complex solutions exists in the literature. Two main directions that
will be discussed here are:
• Multi-window approximation discussed in [38, 39]
• Fast convolution discussed in [40, 41]
2 OFDM Enhancements for 5G Based on Filtering and Windowing
57
10
0
PSD
-10
CPOFDM
CPOFDM
WOLA
WOLA
CP-UFOFDM
CP-UFOFDM
fOFDM
fOFDM
fOFDM-TO4
fOFDM-TO4
-20
-30
-40
-50
40
42
44
46
48
50
52
54
56
58
60
freq index
Fig. 2.19 PSD with 4 PRB allocation: solid lines are PSD without PA, dashed lines are PSD with
PA
Recalling Eq. 2.4, the overall signal x (n) composed of all subbands can be written
as
x (n) =
K
−1
xk (n) e
j2πk Qn
N
.
(2.6)
k=0
The complexity of generating the overall signal (e.g., in the DL case) for one symbol
scales linearly with the number of subcarriers per subband Q (to generate one subband
k) and with the number of allocated subbands K .
The multi-window approximation is based on the observation that subband filtering is equivalent to subcarrier windowing, with the result that Eq. 2.4 can be rewritten,
for one symbol duration, as [38]
xk (n) =
Q−1
f q (n)sk,q w (n)
(2.7)
q=0
where f q (n) is the effective filter that is used to modulate the qth subcarrier in
each subband. One can observe that for close subcarriers, the value of f q (n) is
not so different, depending mainly on the bandwidth for which the prototype filter
f (n) was designed. Therefore, the idea is to divide the subband allocation into
multiple subcarrier groups, every subcarrier group g consists of Q g subcarriers and
is windowed using one effective filter f g (n), with the result that, the complexity
scales only with the number of subcarrier groups G [39]. The overall signal for one
symbol duration can be approximated as
58
R. Ahmed et al.
Fig. 2.20 Overlap-save processing flow used for fast convolution
x (n) ≈
G−1
g=0
f g (n)
Q−1
K
−1
sk,q w (n) e
j2πk Qn+qn
N
,
(2.8)
k=0 q=0
where the strength of the approximation thus depends on the number of subcarrier
groups G. In the extreme case of using G = 1, this approximation has the same complexity as a windowing operation. With G = 3, the spectral localization is improved,
coming at the price of roughly 3 times baseline OFDM complexity [38].
The main idea behind fast convolution algorithms is that filtering through a higherorder impulse response can be implemented effectively through multiplication in
FFT-domain. This is done by taking the DFT of the input sequence as well as the
DFT of the filter impulse response. The time domain output signal is finally obtained
by IDFT [40]. As shown in Fig. 2.20, overlap-save processing is applied for long
sequences to combine all processed blocks into the output signal.
2.4 Summary
In this chapter, we have discussed the requirements on the waveform design in 5G
NR is driven by the new and challenging use cases in NR. To enable a seamless multiplexing of different services and to fulfill the forward comparability vision for NR,
it is highly desirable to have a frequency localized waveform design in NR without
sacrificing too much time localization. CP-OFDM has been proven as a powerful
and flexible waveform already for 4G, and hence in 5G, it is the dominant candidate
2 OFDM Enhancements for 5G Based on Filtering and Windowing
59
solution, while dealing with its weak points such as frequency localization by some
modifications To this end, three candidate waveform preprocessing techniques for
the 5G NR are discussed, namely WOLA, UF-OFDM, and f-OFDM, where each is
based on either time domain windowing or subband filtering. The design principle
and implementation aspects of each waveform are outlined.
The design parameters of each waveform can be tuned to fit the target use case,
e.g., the window length or the window shape in WOLA and the subband filter length
in subband filtering techniques, mainly optimizing a trade-off between time domain
localization and frequency domain localization. Depending on the choice of these
parameters, the performance of windowing and subband filtering techniques is found
to be comparable, with subband filtering offering a slightly better performance, at
the price of higher complexity.
Several techniques can be used to reduce the implementation complexity of the
subband filtering technique. One interesting technique, besides frequency domainbased fast convolution, is to approximate the subband filtering by a multi-windowing
operation, which in essence yields a hybrid between a subband filtering technique
and a windowing technique, with a trade-off between frequency localization (and
hence performance) versus implementation complexity cost.
References
1. Machina research, Technical report, Aug 2016
2. 3rd Generation Partnership Project; TS 36.211, E-UTRA, Physical Channels and Modulation
(Release 13) (2016)
3. 3rd Generation Partnership Project; TS 36.211, E-UTRA, Physical Channels and Modulation
(Release 14) (2017)
4. J. Proakis, Digital Communications. 4th edn. (Mc Graw-Hill Book Company, 2001)
5. R.W. Chang, High-speed multichannel data transmission with bandlimited orthogonal signals.
Bell Sys. Tech. 45, 1775–96 (1966)
6. S.B. Weinstein, The history of orthogonal frequency-division multiplexing [History of Communications]. Commun. Mag. IEEE 47(11), 26–35 (2009)
7. S. Venkatesan, R.A, Valenzuela, OFDM for 5G: cyclic prefix versus zero postfix, and filtering
versus windowing, in 2016 IEEE International Conference on Communications (ICC), pp. 1–5,
May 2016
8. B. Farhang-Boroujeny, OFDM versus filter bank multicarrier. IEEE Signal Process. Mag. 28(3),
92–112 (2011)
9. G. Wunder, P. Jung, M. Kasparick, T. Wild, F. Schaich, Yejian Chen, S. Brink, I. Gaspar, N.
Michailow, A. Festag, L. Mendes, N. Cassiau, D. Ktenas, M. Dryjanski, S. Pietrzyk, B. Eged,
P. Vago, F, Wiedmann, 5GNOW: non-orthogonal, asynchronous waveforms for future mobile
applications. IEEE Commun. Mag. 52(2), 97–105 (2014)
10. M. Bellanger, FBMC Physical Layer: A Primer (2010)
11. F. Schaich, T. Wild, Y. Chen, Waveform Contenders for 5G—suitability for short packet and
low latency transmissions, in 2014 IEEE 79th Vehicular Technology Conference (VTC Spring),
pp. 1–5, May 2014
12. NGMN Alliance, NGMN 5G White Paper (2015), http://www.ngmn.org/5g-white-paper.html
60
R. Ahmed et al.
13. A. Osseiran, V. Braun, T. Hidekazu, P. Marsch, H. Schotten, H. Tullberg, M.A. Uusitalo,
M. Schellman, The foundation of the mobile and wireless communications system for 2020
and beyond: challenges, enablers and technology solutions, in 2013 IEEE 77th Vehicular Technology Conference (VTC Spring), pp. 1–5, June 2013
14. 3GPP TR 38.912 Study on New Radio (NR) access technology (Release 14) (2017)
15. F. Schaich, T. Wild, Subcarrier spacing–a neglected degree of freedom? in 2015 IEEE 16th
International Workshop on Signal Processing Advances in Wireless Communications (SPAWC),
pp. 56–60, June 2015
16. R1-167963 Way forward on waveform RAN1#86 (2016)
17. Y. Chen, F. Schaich, T. Wild, Multiple access and waveforms for 5G: IDMA and universal
filtered multi-carrier, in 2014 IEEE 79th Vehicular Technology Conference (VTC Spring), pp.
1–5, May 2014
18. F. Schaich, T. Wild, Relaxed synchronization support of universal filtered multi-carrier including autonomous timing advance, in 2014 11th International Symposium on Wireless Communications Systems (ISWCS), pp. 203–208, Aug 2014
19. R1-163615 WF on overview of NR RAN1#84bis (2016)
20. RAN1#86bis chairmans notes
21. R4-1610921 Way forward on in-band requirements for NR RAN4#81, Nov 2016
22. Ericsson. R1-163224 Waveform candidates RAN1#84bis (2016)
23. G. Berardinelli, F.M.L. Tavares, T.B. Strensen, P. Mogensen, K. Pajukoski, Zero-tail DFTspread-OFDM signals, in 2013 IEEE Globecom Workshops (GC Wkshps), pp. 229–234, Dec
2013
24. Qualcomm, 5G Waveform & Multiple Access Techniques, https://www.qualcomm.com/
documents/5g-research-waveform-and-multiple-access-techniques (2015)
25. Qualcomm, R1-162199 Feasibility of Mixing Numerology in an OFDM System RAN1#84bis
(2016)
26. Nokia, R1-165014 Subband-wise filtered OFDM for New Radio below 6 GHz RAN1#85
(2016)
27. LG. R1-162516 Flexible CP-OFDM with variable ZP RAN1#84bis, Apr 2016
28. X. Wang, T. Wild, F. Schaich, S. ten Brink, Pilot-aided channel estimation for universal filtered
multi-carrier, in 2015 IEEE 82nd Vehicular Technology Conference (VTC2015-Fall), pp. 1–5,
Sept 2015
29. R. Ahmed, T. Wild, F. Schaich, Coexistence of UF-OFDM and CP-OFDM, in 2016 IEEE 83rd
Vehicular Technology Conference (VTC Spring), pp. 1–5, May 2016
30. J. Abdoli, M. Jia, J. Ma, Filtered OFDM: a new waveform for future wireless systems, in 2015
IEEE 16th International Workshop on Signal Processing Advances in Wireless Communications
(SPAWC), pp. 66–70, June 2015
31. X. Zhang, M. Jia, L. Chen, J. Ma, J. Qiu, Filtered-OFDM—enabler for flexible waveform
in the 5th generation cellular networks, in 2015 IEEE Global Communications Conference
(GLOBECOM), pp. 1–6, Dec 2015
32. Huawei, R1-164033 f-OFDM scheme and filter design RAN1#85 (2016)
33. R1-166004, R4-164542, Response LS on realistic power amplifier model for NR waveform
evaluation, May 2016
34. E. Dahlman, S. Parkvall, J. Skold, 4G: LTE/LTE-Advanced for Mobile Broadband. Academic
Press, 1st ed. (2011)
35. R1-1609564 Implementation-specific UF-OFDM for New Radio (2016)
36. R1-164685, OFDM based waveform single user evaluation RAN1#85 (2016)
37. R1-166093, Waveform evaluation updates for case 1a and case 1b, Aug 2016
38. M. Matthe, D. Zhang, F. Schaich, T. Wild, R. Ahmed, G. Fettweis, A reduced complexity
time-domain transmitter for UF-OFDM, in 2016 IEEE 83rd Vehicular Technology Conference
(VTC Spring), pp. 1–5, May 2016
2 OFDM Enhancements for 5G Based on Filtering and Windowing
61
39. Samsung, R1-166746 Discussion on multi-window OFDM for NR waveform RAN1#86 (2016)
40. M. Renfors, J. Yli-Kaakinen, T. Levanen, M. Valkama, T. Ihalainen, J. Vihriala, Efficient fastconvolution implementation of filtered CP-OFDM waveform processing for 5G, in 2015 IEEE
Globecom Workshops (GC Wkshps), pp. 1–7, Dec 2015
41. J. Yli-Kaakinen, T. Levanen, S. Valkonen, K. Pajukoski, J. Pirskanen, M. Renfors, M.
Valkama, Efficient fast-convolution-based waveform processing for 5G Physical Layer. CoRR,
abs/1706.02853 (2017)
Chapter 3
Filter Bank Multicarrier Modulation
Ronald Nissel and Markus Rupp
3.1 Why FBMC?
Future mobile systems will be characterized by a large range of different use cases,
ranging from enhanced mobile broadband (eMBB) over enhanced machine type communications (eMTC) to ultra-reliable low-latency communications (URLLC) [2, 41,
53, 63]. To efficiently support such diverse use cases, a flexible time–frequency allocation becomes necessary. In particular, the out-of-band (OOB) emissions must be
sufficiently low in order to efficiently support different use cases within the same
band. Furthermore, low OOB emissions reduce the synchronization requirements.
Conventional orthogonal frequency-division multiplexing (OFDM) with cyclic prefix (CP) performs poorly in this context because of the underlying rectangular prototype filter, which causes large OOB emissions. To improve spectral properties in
OFDM, the 3rd Generation Partnership Project (3GPP) is therefore considering windowing and filtering [45, 52, 63]. The windowed OFDM scheme is called OFDM with
weighted overlap and add (WOLA) and the filter-based methods are called universal filtered OFDM (UF-OFDM) and filtered OFDM (f-OFDM). While windowing
and filtering can indeed reduce the OOB emissions of conventional OFDM, filter
bank multicarrier modulation (FBMC) with offset quadrature amplitude modulation
(OQAM) [41] still performs much better, as shown in Fig. 3.1. Additionally, FBMCOQAM has a maximum symbol density, that is, a time–frequency spacing of TF
= 1 for complex-valued symbols. In OFDM-based schemes, on the other hand, the
symbol density is lower, as indicated by T F > 1, additionally worsening the spectral
efficiency. Besides the better support of different use cases, FBMC also increases
the throughput of legacy Long Term Evolution (LTE) transmissions because fewer
R. Nissel (B) · M. Rupp
TU Wien, Gusshausstraße 25, 1040 Vienna, Austria
e-mail: rnissel@nt.tuwien.ac.at
M. Rupp
e-mail: mrupp@nt.tuwien.ac.at
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_3
63
R. Nissel and M. Rupp
Power Spectral Density [dB]
64
0
−20
CP-OFDM
−40
WOLA
LTE like:
T F=1.07
24 subcarriers
T F=1.09
−60 UF-OFDM
f-OFDM
−80
FBMC
OQAM
T F=1
(complex)
−100
−50 −40 −30 −20 −10
0
10
20
30
40
50
Normalized Frequency, f /F
Fig. 3.1 FBMC has much better spectral properties compared with CP-OFDM. Windowing
(WOLA) and filtering (UF-OFDM, f-OFDM) can improve the spectral properties of CP-OFDM.
However, FBMC still performs much better and has the additional advantage of a maximum symbol
c
density, T F = 1 (complex). 2017
IEEE, [41]
FBMC-OQAM,
T F = 1 (complex)
mean
+30%
87 Subcarriers
95 % confidence
interval obtained
by bootstrapping
1.4 MHz
LTE Like
OFDM,
T F = 1.07
72 Subcarriers
6
Throughput [Mbit/s]
Fig. 3.2 Real-world testbed
measurements at 2.5 GHz
show that FBMC has a
higher throughput than
OFDM (1.4 MHz LTE
resembling SISO signal)
because of a higher available
bandwidth and no CP
overhead [38, 41]. The
channel estimation in FBMC
is based on the data
spreading approach [21, 35]
4
2
0
−5
0
5
10
15
20
25
Signal-to-Noise Ratio for OFDM [dB]
guard subcarriers are required and no CP is needed. Figure 3.2 shows real-world
testbed measurements and compares FBMC with an 1.4 MHz LTE single-input and
single-output (SISO) signal (including pilots but ignoring signaling overhead). For
high signal-to-noise ratio (SNR) values, the throughput of FBMC is approximately
30% higher than for OFDM. Even compared with f-OFDM, FBMC would still be
approximately 20% better, as indicated by the time–frequency efficiency calculations
in [41]. However, one has to keep in mind that the potential improvement strongly
depends on the number of subcarriers and the required guard band. In particular,
once the number of subcarriers is very high, windowed OFDM and filtered OFDM
will perform close to FBMC.
Unfortunately, all the nice features of FBMC-OQAM come at a price, namely,
the complex orthogonality condition is replaced by the less strict real orthogonality
condition. While this limitation has in many cases either no or only a minor influence
3 Filter Bank Multicarrier Modulation
65
on the performance, some important methods, such as channel estimation and some
multiple-input and multiple-output (MIMO) techniques, become more challenging.
There exist different variants of FBMC, but we will mainly focus on OQAMbased schemes because they provide the highest spectral efficiency. Different names
are used to describe OQAM, such as OFDM/OQAM [8], fbmc-pulse-amplitude
modulation (PAM) [26], staggered multitone (SMT) or Cosine Modulated Multitone
(CMT) [12], which, however, are essentially all the same. One can easily transform
one of those schemes into another by appropriately tuning the underlying parameters.
For example, FBMC-PAM is a conventional FBMC-OQAM scheme for which the
subcarrier spacing is reduced by a factor of two, the number of subcarriers is increased
by two, and the offset is applied in the frequency domain instead of the time domain.
In general, all those “different” FBMC schemes are characterized by:
• A prototype filter which is localized in time and frequency.
• Only real-valued information symbols can be transmitted at a given time–frequency
position.
• A time–frequency spacing of T F = 0.5 for real-valued symbols (equivalent to
T F = 1 for complex-valued symbols).
• Intrinsic imaginary interference.
Although FBMC has been considered as a strong contender for replacing OFDM in
the fifth-generation (5G) of wireless systems [4, 5, 59], 3GPP eventually decided that
they will stick to OFDM [3]. While such decision makes sense in terms of backward
compatibility to fourth-generation (4G) wireless systems, it is not the most efficient
technique for all possible use cases, especially if the number of subcarriers is low.
Thus, if the envisioned concept of different use cases within the same band turns out
to be successful in 5G, we expect that FBMC will again gain momentum for beyond
5G communications.
3.2 Multicarrier Modulation
Multicarrier modulation has a long-standing history in wireless communications [9,
51, 58]; however, widespread practical applications have only been realized in the
latest versions of wireless systems in the form of OFDM, enabled by advances in the
field of integrated circuits. Current applications of OFDM include LTE, WiFi and
digital video broadcasting-terrestrial (DVB-T).
In multicarrier systems [50], information is commonly transmitted over orthogonal pulses which overlap in time and frequency. The big advantage is that these
pulses usually occupy only a small bandwidth, so that frequency-selective broadband channels transform into multiple, virtually frequency flat, sub-channels (subcarriers). Mathematically, the transmitted signal, s(t), of a multicarrier system in the
time domain can be expressed as
66
R. Nissel and M. Rupp
s(t) =
K
−1
L−1
gl,k (t) xl,k ,
(3.1)
k=0 l=0
where xl,k denotes the transmitted symbol at subcarrier position l and time position k,
and is chosen from a symbol alphabet, usually a QAM or a PAM signal constellation.
The total number of subcarriers is denoted by L and the total number of symbols in
time by K . The basis pulse gl,k (t) in (3.1) is defined by
gl,k (t) = p(t − kT ) ej2π l F (t−kT ) e jθl,k ,
(3.2)
and is, essentially, a time and frequency shifted version of prototype filter p(t), with
T denoting the time spacing and F the frequency spacing (subcarrier spacing). The
choice of phase shift θl,k becomes relevant later in the context of FBMC-OQAM.
After transmission over a channel, the received symbols are decoded by projecting
the received signal, r (t), onto the basis pulses, gl,k (t), that is,
yl,k = r (t), gl,k (t) =
∞
∗
(t) dt.
r (t) gl,k
(3.3)
−∞
In (3.3), we implicitly apply a matched filter if the channel perturbation is additive
white Gaussian noise (AWGN), maximizing the SNR. In a doubly selective channel,
it might be better to choose the transmit and receive prototype filters slightly different, as, for example, suggested in pulse-shaped multicarrier transmissions [27] or
in the practically more relevant case of CP-OFDM. However, employing an AWGN
matched filter is usually close to the optimum because the channel induced interference is often dominated by noise, see Sect. 3.5. One of the biggest advantages of
orthogonal multicarrier systems is that the transmission in (3.3) can be modeled by
a one-tap channel, that is,
yl,k = H (kT, l F) xl,k + n l,k + zl,k ,
(3.4)
where H (t, f ) denotes the time-variant transfer function and represents the one-tap
channel. The noise in (3.4) is described by n l,k and the channel induced interference
by zl,k . Often, the delay spread and the Doppler spread are low enough so that the
channel induced interference is dominated by noise, that is, E{|n l,k |2 } ≫ E{|zl,k |2 }.
Thus, the channel induced interference can be neglected and the employment of
low-complexity one-tap equalizers correspond to the maximum likelihood symbol
detection in case of Gaussian noise.
Multicarrier systems are mainly characterized by prototype filter p(t) as well as
time spacing T and frequency spacing F, so that the ambiguity function [12, 50],
3 Filter Bank Multicarrier Modulation
A(τ, ν) =
∞
67
τ j2πνt
τ ∗
p t+
e
p t−
dt,
2
2
−∞
(3.5)
captures the main properties of a multicarrier system in a compact way. The projection
of the transmitted basis pulses gl1 ,k1 (t) onto the received basis pulses gl2 ,k2 (t) can
then be expressed by the ambiguity function according to
−k2 ) j(θl1 ,k1 −θl2 ,k2 )
A( T (k1 − k2 ), F(l1 − l2 ) ) .
e
gl1 ,k1 (t), gl2 ,k2 (t) = e−jπ T F(l1 +l2 )(k1
only a phase shift
ambiguity function
(3.6)
There exist some fundamental limitations of multicarrier systems, as formulated
by the Balian–Low theorem [13], which states that it is mathematically impossible
that the following four desired properties are fulfilled at the same time:
1. Maximum symbol density,
T F = 1,
(3.7)
2. Time localization,
∞
(t − t¯)2 | p(t)|2 dt < ∞,
(3.8)
( f − f¯)2 |P( f )|2 d f < ∞,
(3.9)
gl1 ,k1 (t), gl2 ,k2 (t) = δ(l1 −l2 ),(k1 −k2 )
(3.10)
σt =
−∞
3. Frequency localization,
σf =
∞
−∞
4. Orthogonality,
A(T (k1 − k2 ), F(l1 − l2 )) = δ(l1 −l2 ),(k1 −k2 ) ,
(3.11)
∞
with δ denoting the Kronecker delta function. The pulse P( f ) = −∞ p(t) e−j2π f t dt
∞
in (3.9) represents the Fourier transform of p(t) while t¯ = −∞ t | p(t)|2 dt corre∞
sponds to the mean time and f¯ = −∞ f |P( f )|2 d f the mean frequency of the
pulse. Furthermore, we assume that p(t) is normalized to preserve unit energy. The
localization measures in (3.8) and (3.9) can be interpreted as standard deviation, with
| p(t)|2 and |P( f )|2 representing the probability density function (pdf). This relates
the Balian-Low condition to the Heisenberg uncertainty relationship [57, Chap. 7].
The Balian-Low theorem implies that at least one of the four desired properties has
to be sacrificed when designing multicarrier waveforms. For example, CP-OFDM
sacrifices frequency localization while FBMC-OQAM the complex orthogonality
condition.
68
R. Nissel and M. Rupp
3.2.1 CP-OFDM
CP-OFDM is the most prominent multicarrier technique and is applied, for example,
in Wireless LAN and LTE [24, 29]. CP-OFDM employs a rectangular transmit and
receive pulse, which greatly reduces the computational complexity. Furthermore, the
CP guarantees orthogonality in frequency-selective channels. The transmitter (TX)
and receiver (RX) prototype filter can be expressed by
pTX (t) =
pRX (t) =
√1
T0
0
if −
T0
2
+ TCP ≤ t <
T0
2
(3.12)
otherwise
T0
2
√1
T0
if
0
otherwise
≤t <
T0
2
(3.13)
for which
(Bi)-Orthogonal : T = T0 + TCP ; F = 1/T0 → T F = 1 +
√ CP ;
Localization :
σt = T02+T
σf = ∞
3
TCP
T0
,
(3.14)
with T0 representing a time-scaling parameter which depends on the desired subcarrier spacing (or time spacing). Note that, in contrast to FBMC, the prototype filter is
differently at the TX and RX side.
Figure 3.3 shows the ambiguity function, see (3.5), for CP-OFDM (without CP,
that is, TCP = 0). Orthogonality is guaranteed for a time spacing of T = T0 and a
frequency spacing of F = 1/T0 , leading to T F = 1. This is also indicated by the
rectangular grid (the small circles) inside of Fig. 3.3. The ambiguity function decays
very slowly in frequency because of the underlying rectangular pulse. Additionally,
the CP simplifies equalization in frequency-selective channels but also reduces the
spectral efficiency. In order to reduce the OOB emissions in OFDM, 3GPP is currently
considering windowing [45] and filtering [52, 63]. As already shown in Fig. 3.1, such
|A( , )|2
Time Shift, /T0
Fig. 3.3 The ambiguity
function, 10 log10 |A(τ, ν)|2 ,
for OFDM (without CP)
shows a good localization in
the time domain, but a poor
localization in the frequency
domain. The orthogonal
time–frequency spacing is
T F = 1, indicated by the
small circles in the figure
4
3
2
1
0
−1
−2
−3
−4
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 [dB]
Frequency Shift, T0
0
−20
−40
−60
−80
−100
3 Filter Bank Multicarrier Modulation
69
methods can reduce the OOB emissions in OFDM, but still do not perform as good
as FBMC and have the additional disadvantage of a reduced symbol density, that is,
T F > 1, even in an AWGN channel. Note that for all the windowed- and filteredbased OFDM techniques, receive windowing and filtering is of utmost importance
[41]. Very often, people forget this crucial aspect and only focus on reducing the
OOB emissions at the transmitter.
3.3 FBMC-OQAM
FBMC-OQAM replaces the complex orthogonality condition with the less strict
real orthogonality condition, ℜ{gl1 ,k1 (t), gl2 ,k2 (t)} = δ(l2 −l1 ),(k2 −k1 ) , and works, in
principle, as follows:
1. Design a prototype filter with p(t) = p(−t) which is orthogonal for a time spacing
of T = T0 and a frequency spacing of F = 2/T0 , leading to T F = 2.
2. Reduce the time–frequency spacing by a factor of two each, that is, T = T0 /2
and F = 1/T0 , leading to T F = 0.5.
3. The so induced interference is shifted to the purely imaginary domain by the
phase shift θl,k = π2 (l + k) in (3.2).
Let us take a closer look at the intrinsic interference. With θl,k =
0.5, the inner product in (3.6) transforms to
π
π
gl+∆l,k+∆k (t), gl,k (t) = e j 2 (∆l+∆k) e−j 2 ∆k(2l+∆l)
purely imaginary for odd ∆k, ∆l
π
(l
2
+ k) and T F =
A(∆k T, ∆l F).
0 if both ∆k =0, ∆l =0 are even
(3.15)
The ambiguity function in (3.15) approaches zero if ∆k and ∆l are even because the
prototype filter is designed to be orthogonal for those cases. If, on the other hand,
either ∆k or ∆l is odd, A(·) no longer approaches zero, leading to interference.
The main idea of FBMC is to shift this interference to the imaginary domain. To
be specific, the exponential function in (3.15) becomes purely imaginary valued if
either ∆k or ∆l is odd. Furthermore, the ambiguity function is always real-valued
because of p(t) = p(−t), so that it has no influence on the imaginary part. Note that
we consider a phase shift of θl,k = π2 (l + k), but other phase shifts are also possible,
for example, θl,k = j π2 (l + k) − jπlk.
Similar as in orthogonal multicarrier systems, FBMC also allows the employment of low-complexity one-tap equalizers. To be specific, the transmission can be
modeled by a one-tap channel H (kT, l F), similar as in (3.4), according to
yl,k = H (kT, l F) (xl,k + j vl,k ) + n l,k + zl,k .
(3.16)
Compared with orthogonal multicarrier systems, the data symbols xl,k are real-valued
and there exists an imaginary interference term, described by jvl,k , which depends
70
R. Nissel and M. Rupp
on the adjacent symbols. The big advantage of FBMC compared with other nonorthogonal schemes is that the imaginary interference can easily be canceled, simply
by taking the real part after equalization.1 Thus, computational demanding equalization and cancellation methods are not necessary. Note that the imaginary interference
does not carry any useful information for L → ∞ and K → ∞, so that, by taking
the real part, we do not lose any useful information; see Sect. 3.6 for more details.
Although the time–frequency spacing in FBMC-OQAM is equal to T F = 0.5,
only real-valued information symbols can be transmitted in such as system, leading
to an equivalent time–frequency spacing of T F = 1 for complex-valued symbols.
Very often, the real part of a complex-valued symbol is mapped to the first time slot
and the imaginary part to the second time slot, thus the name offset-QAM. However,
such self-limitation is not necessary. We can equivalently perform this mapping over
subcarriers or directly consider PAM symbols instead of “staggered” QAM symbols.
As already mentioned in the beginning of this section, the prototype filter has
to be an even function and orthogonal for a time–frequency spacing of T F = 2.
All prototype filters which satisfy this condition can be utilized in FBMC-OQAM.
Let us discuss two prominent prototype filters, namely the Hermite filter and the
PHYDYAS filter. The Hermite prototype filter is based on Hermite polynomials
Hn (·), as proposed in [16], and can be expressed as
1 −2π
p(t) = √ e
T0
t
T0
2
√ t
ai Hi 2 π
,
T0
i={0,4,8,
(3.17)
12,16,20}
for which the coefficients can be found to be [35]
a0 =
1.412692577
a12 = −2.2611 × 10−9
a4 = −3.0145 × 10−3 a16 = −4.4570 × 10−15
a8 = −8.8041 × 10
−6
a20 =
1.8633 × 10
−16
(3.18)
,
leading to the following properties of (3.17),
Orthogonal : T = T0 ;
F = 2/T0
→ TF =2
Localization : σt = 0.2015 T0 ; σ f = 0.403/T0 .
(3.19)
Note that in practical systems, the pulse in (3.17) will be truncated so that it fits
within the time interval − O2T0 ≤ t < O2T0 , with O denoting the overlapping factor.
Figure 3.4 shows the ambiguity function for the Hermite prototype filter. The
filter has the same shape in time and frequency, allowing us to exploit symmetries.
Furthermore, the filter is based on a Gaussian function and therefore has a good joint
time–frequency localization of σt σ f = 1.02 × 1/4π , almost as good as the bound of
σt σ f ≥ 1/4π ≈ 0.08 (attained by the Gaussian pulse). In Fig. 3.4, we also observe
1 The
channel must be known at the receiver. Channel estimation itself becomes more challenging
in FBMC, as we will discuss later in this section.
3 Filter Bank Multicarrier Modulation
71
Time Shift, /T0
|A( , )|2
4
3
2
1
0
−1
−2
−3
−4
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 [dB]
0
−20
−40
−60
−80
−100
Frequency Shift, T0
Fig. 3.4 The ambiguity function for the Hermite prototype filter shows good localization in both,
time and frequency. The orthogonal time–frequency spacing is T F = 2. To improve the spectral
efficiency, the time–frequency spacing is reduced to T F = 0.5, indicated by the black markers. The
so induced interference is shifted to the purely imaginary domain
that the pulse is orthogonal for a time spacing of T = T0 and a frequency spacing
of F = 2/T0 , indicated by the small circles. In FBMC-OQAM, the time–frequency
spacing is reduced to T = T0 /2 and F = 1/T0 . This causes interference, indicated
by the black markers in Fig. 3.4, which, however, is purely imaginary valued, see
(3.15).
Another prominent filter is the PHYDYAS prototype filter [6, 7], constructed by:
p(t) =
⎧ O−1
⎨ 1+2 i=1 bi cos O2πtT0
√
O T0
⎩0
if − O2T0 ≤ t <
O T0
2
.
(3.20)
otherwise
The coefficients bi were calculated in [28] and depend on overlapping factor O. For
example, for an overlapping factor of O = 4, the coefficients become,
b1 = 0.97195983; b2 =
√
2/2; b3 = 0.23514695,
(3.21)
and lead to the following properties of (3.20)
F = 2/T0
→ TF =2
Orthogonal : T = T0 ;
Localization : σt = 0.2745 T0 ; σ f = 0.328/T0 .
(3.22)
Figure 3.5 shows the ambiguity function for the PHYDYAS prototype filter
(O = 4). Compared to the Hermite prototype filter, the PHYDYAS filter has a better frequency localization but a worse time localization. The joint time–frequency
localization of σt σ f = 1.13 × 1/4π is also worse. Note that the PHYDYAS filter is
not perfectly orthogonal, as shown in Fig. 3.5 for time position ±3 and frequency
position {−2, 0, 2}, However, this interference is very low and thus be neglected for
typical working points of wireless communications.
72
|A( , )|2
Time Shift, /T0
Fig. 3.5 The PHYDYAS
prototype filter shows good
localization in time and
frequency. Compared with
the Hermite prototype filter,
the time localization is worse
but the frequency
localization is better. The
small circles indicate the
orthogonal time–frequency
spacing, while the black
markers correspond to the
imaginary interference
R. Nissel and M. Rupp
4
3
2
1
0
−1
−2
−3
−4
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 [dB]
0
−20
−40
−60
−80
−100
Frequency Shift, T0
3.3.1 Latency
For our latency considerations, we focus on the underlying pulse duration but ignore
other sources of delays, such as channel delays and processing delays. The transmission time of FBMC-OQAM then depends on subcarrier spacing F, the number of
symbols in time K and overlapping factor O, according to,
Tblock =
0.5
O
+
(K − 1).
F
F
(3.23)
In particular, the subcarrier spacing can be utilized to reduce the latency. This, however, comes at the expense of an increased sensitivity to frequency-selective channels
(but improves the robustness in time-variant channels). When compared to OFDM,
the overlapping factor plays an important role in FBMC. A common value is O = 4,
leading to very low OOB emissions for the PHYDYAS prototype filter, see Fig. 3.1.
However, FBMC allows more flexibility than that. For example, the Hermite prototype filter can be used with a lower overlapping factor. One can also design a
prototype filter that is specifically tailored for low-latency scenarios, such as a time
domain root-raised-cosine pulse with overlapping factor O = 1.
An LTE subframe requires a transmission time of 1 ms and consists of K = 14
1
and F = 15 kHz). Because each FBMC symbol only
OFDM symbols (TCP = 14F
carries half the information of that of an OFDM symbol (same number of subcarriers),
we need in total K = 28 FBMC symbols for a fair comparison to LTE. For an
overlapping factor of O = 1, this implies that FBMC has a transmission time of
Tblock ≈ 0.97 ms and is therefore faster than LTE. For an overlapping factor of O =
1.5, the transmission time is exactly 1 ms, same as in LTE. An overlapping factor of
O = 4, on the other hand, performs relatively poor and requires 1.2 ms, 20% longer
than LTE.
Note that the ramp-up and ramp-down period in FBMC increases the latency, but
not necessarily the sum throughput of the whole system, because different blocks
can overlap in time. This works as long as the required phase pattern, which shifts the
3 Filter Bank Multicarrier Modulation
73
intrinsic interference to the purely imaginary domain, is fulfilled, as typically the case
in downlink transmissions. However, in multi-user uplink transmissions, different
users experience different phase shifts. Thus, the necessary phase pattern is violated,
leading to interference, even for a perfectly time and frequency synchronized system.
One then has to include a guard time in order to avoid interference. In such cases,
the ramp-up and ramp-down period not only increases the latency, but also reduces
the sum throughput.
3.3.2 Channel Estimation
The main idea of FBMC-OQAM is to equalize the phase, followed by taking the real
part in order to get rid of the imaginary interference. This, however, only works once
the phase is known, thus only after channel estimation. The channel estimation itself
has to be performed in the complex domain, affected by the imaginary interference,
and one observes an SIR of 0 dB. Thus, additional processing becomes necessary.
Preamble-based channel estimation was, for example, discussed in [20]. However,
LTE employs pilot-aided channel estimation because it has a low overhead and allows
a simple tracking of the channel in time. A straightforward method for pilot-aided
channel estimation in FBMC was proposed in [19], where one data symbol per pilot,
the so-called auxiliary symbol, is sacrificed to cancel the imaginary interference at
the pilot position. The big disadvantage of such method is the high power of the
auxiliary symbols, worsening the PAPR and wasting signal power. Subsequently,
different methods have been proposed to mitigate these harmful effects [10, 21, 35,
60]. In particular, the data spreading approach of [21] is promising because no energy
is wasted, there is no noise enhancement, and the performance is close to OFDM [33].
The idea of [21] is to spread data symbols over several time–frequency positions,
close to the pilot symbol, in such a way, that the imaginary interference at the pilot
position is canceled. The drawback is a slightly higher computational complexity
[35]. To reduce the computational complexity and to improve the applicability in
doubly selective channels, one can combine the data spreading approach with the
auxiliary symbol method, as proposed in [11]. Note, however, that also the classical
spreading approach performs well in doubly selective channels [43].
3.4 Discrete-Time System Model
The continuous-time representation, discussed so far, provides analytical insights
and gives physical meaning to multicarrier systems. However, such representation
becomes analytically hard to track in doubly selective channels because double integrals have to be solved. Furthermore, in practice, the signal is generated in the
discrete-time domain. Thus, we will switch from the continuous-time domain to the
discrete-time domain. In contrast to many other authors, we employ a matrix-based
74
R. Nissel and M. Rupp
system model instead of a discrete-time filter representation because it simplifies analytical investigations and provides a more compact description. If one is interested
in the conventional discrete-time filter representation, we refer to [47, 55].
In our matrix-based system model, the basis pulses in (3.2) are sampled at rate
f s = 1/∆t = F NFFT and stacked in a basis pulse vector gl,k ∈ C N ×1 according to
[gl,k ]n =
√
∆t gl,k (t)
t=n∆t−O T
,
(3.24)
for n = 0, 1, . . . , N − 1, where the total number of samples is given by N =
O NFFT + NFFT
(K − 1). The interpretation of overlapping factor O and fast Fourier
2
transform (FFT) size NFFT ≥ L becomes more clear later in this section, when we
discuss an efficient FFT implementation. Practical systems will never operate at a
critically sampling rate (NFFT = L) because this would lead to large OOB emissions, caused by the repetition of the spectrum in the frequency domain. We strongly
advise to never use a critically sampled system, which is only useful for emulating
the asymptotic case of infinitely many subcarriers, L → ∞. Unfortunately, many
authors ignore this important aspect.
By stacking all basis pulse vectors from (3.24) in a large transmit matrix G ∈
C N ×L K ,
G = g0,0 · · · g L−1,0 g0,1 · · · g L−1,K −1 ,
(3.25)
and all data symbols in a large transmit symbol vector x ∈ C L K ×1 ,
⎧⎡
⎤⎫
⎪
⎨ x0,0 · · · x0,K −1 ⎪
⎬
⎢
⎥
..
x = vec ⎣ ... . . .
⎦
.
⎪
⎪
⎩
⎭
x L−1,0 · · · x L−1,K −1
T
= x0,0 · · · x L−1,0 x0,1 · · · x L−1,K −1 ,
(3.26)
(3.27)
we can express the sampled transmit signal s ∈ C N ×1 in (3.1) by:
s = Gx.
(3.28)
Because of linearity, matrix G can easily be found even if the underlying modulation
format is not known in detail. For that, all transmitted symbols have to be set to zero,
except xl,k = 1. Vector s then provides immediately the l + Lk-th column vector of
G. Repeating this step for each time–frequency position delivers transmit matrix G.
At the receiver, we perform matched filtering by GH , so that the whole transmission
system simplifies to
y = GH H G x + GH n
≈ diag{h} GH G x + GH n,
(3.29)
(3.30)
3 Filter Bank Multicarrier Modulation
75
with y ∈ C L K ×1 denoting the received symbols, H ∈ C N ×N the banded time-variant
convolution matrix ([H]i, j = h conv. [i, i− j] with time-variant impulse response
h conv. [i, m τ ]) and n ∼ C N (0, Pn I N ) the additive white Gaussian noise in the time
domain with zero mean and variance Pn . In most practical scenarios, the delay spread
and the Doppler spread are low enough so that the channel induced interference can
be neglected [36, 41]. This allows us to factor out the channel in (3.29) according
to (3.30), for which h ∈ C L K ×1 describes the one-tap channels (frequency domain),
that is, the diagonal elements of GH H G. In particular, the l + Lk-th element of h is
given by
H
H gl,k ≈ H (kT, l F),
(3.31)
h l,k = gl,k
and represents the one-tap channel at subcarrier position l and time position k. FBMC
experiences imaginary interference, described by the off-diagonal elements of GH G
and only the real orthogonality condition holds true, that is, ℜ{GH G} = I L K .
3.4.1 IFFT Implementation
Practical systems must be much more efficient than the matrix multiplication in
(3.28). It was, for example, shown in [55] that FBMC-OQAM can be efficiently
implemented by an inverse FFT (IFFT) together with a polyphase network. Unfortunately, the authors of [55] rely on a filter bank representation which is very different
to the conventional OFDM description. We therefore consider an alternative interpretation, more closely related to conventional OFDM systems. A similar representation
was, for example, suggested in [27] for pulse-shaping multicarrier systems.
To simplify the exposition and without losing generality, we consider only time
position k = 0. Any other time position can easily be obtained by time shifting this
. The main idea is to factor out the prototype
special case by T = T20 , respectively NFFT
2
filter p(t) from (3.1), so that the sampled transmit signal can be expressed by
s0 (n ∆t) = p(n ∆t)
L−1
e
j2π l N n
FFT
e jθl,0 xl,0 ,
(3.32)
l=0
for n = − O N2FFT , . . . , O N2FFT − 1. The summation in (3.32) corresponds to an
NFFT point IFFT with the input arguments {e jθ0,0 x0,0 , e jθ1,0 x1,0 , . . . , e jθL−1,0 x L−1,0 , 0,
0, . . .}. Furthermore, because l is an integer, the summation in (3.32) is NFFT periodic with respect to n. Thus, the IFFT has to be calculated only for NFFT samples.
Those samples can then be copied O-times, followed by an element-wise multiplication with prototype filter p(n ∆t). By stacking the transmitted samples in a vector
s0 ∈ C O NFFT ×1 , we can therefore express (3.32) by
76
R. Nissel and M. Rupp
⎜
⎜
⎜
s0 = p ◦ ⎜
⎜ 1 O×1
⎜
⎝
⎤⎞
e jθ0,0 x0,0
⎥⎟
⎢
..
⎥⎟
⎢
.
⎢
⎥⎟
H
jθ L−1,0
⎢
⎟
x L−1,0 ⎥
⊗ W NFFT ⎢e
⎥ ⎟,
⎢
⎟
⎥
0
⎣
⎦⎠
..
.
IFFT
⎡
⎛
(3.33)
repeat O -times
element-wise multiplication
where ◦ denotes the element-wise Hadamard product, ⊗ the Kronecker product
and W NFFT ∈ C NFFT ×NFFT a Discrete Fourier Transform (DFT) matrix. Note that, by
circular shifting the IFFT input in (3.33), we can shift the signal in frequency by
multiples of F. The sampled prototype filter p ∈ C O NFFT ×1 in (3.33) is given by,
√
[p]n = ∆t p(t)
for n = 0, 1, . . . , O NFFT − 1.
t=n ∆t−O T
(3.34)
Figure 3.6 illustrates the efficient FBMC-OQAM implementation and compares it
to windowed OFDM. Both modulation schemes apply the same basic operations, that
is, IFFT, repetition and element-wise multiplications. However, windowed OFDM
has overall a lower complexity because the element-wise multiplication is limited to a
window of size 2 TW and time symbols are further apart, that is, T = TW + TCP + T0
in windowed OFDM versus T = T0 /2 in FBMC-OQAM. Thus, FBMC needs to
apply the IFFT more than two times (exactly two times if TW = TCP = 0). Of course,
the overhead TW + TCP in windowed OFDM reduces the spectral efficiency. Because
Windowed CP-OFDM
FBMC-OQAM
IFFT
IFFT
IFFT
IFFT
IFFT
IFFT
×
|p(t)|2
+
IFFT
IFFT
p(t)
t
t
+
t
T0
TW
IFFT
×
t
T0
2
T0
2
TCP
Fig. 3.6 From a conceptional point of view, the signal generation in windowed OFDM and FBMCOQAM requires the same basic operations, namely, an IFFT, copying the IFFT output, element-wise
c
multiplication with the prototype filter and, finally, overlapping. 2017
IEEE, [41]
3 Filter Bank Multicarrier Modulation
77
the signal generation for both modulation formats is very similar, FBMC-OQAM can
utilize the same hardware components as windowed OFDM.
The receiver works in a similar way, but in reversed order, that is, element-wise
multiplication, reshaping the received symbol vector to NFFT × O followed by a
row-wise summation and, finally, an FFT. In matrix notation, this can be expressed
by
(
)
y0 = diag e−jθ0,0 . . . e−jθL−1,0 0 L×(NFFT −L) W NFFT (11×O ⊗ I NFFT ) (p ◦ r0 ),
(3.35)
where r0 ∈ C O NFFT ×1 represents the received samples and y0 ∈ C L×1 the received
symbols, both at time position k = 0. WOLA requires at the receiver the same basic
operations as FBMC. However, in contrast to FBMC, WOLA employs a different
transmit and receive prototype filter.
3.5 One-Tap Equalizers in Doubly Selective Channels
The biggest advantage of multicarrier systems is that the transmission over a doubly
selective channel can be approximated by one-tap channels. In this section, we calculate the approximation error by considering the signal-to-interference ratio (SIR).
In orthogonal multicarrier systems, the SIR can be calculated by
SIRQAM =
E{|h l,k xl,k |2 }
E{|zl,k |2 }
(3.36)
with interference zl,k given by,
H
H G x − h l,k xl,k .
zl,k = gl,k
(3.37)
In FBMC-OQAM, on the other hand, the SIR cannot be calculated as easily as in
(3.36) because directly applying (3.36) leads to an SIR of approximately 0 dB due
to the inherent self-interference. We have to equalize the phase followed by taking
the real part, before calculating the SIR. Thus, the SIR can be expressed by
SIROQAM =
E{|ℜ{e−jϕl,k h l,k xl,k }|2 }
,
E{|ℜ{e−jϕl,k zl,k }|2 }
(3.38)
where e jϕl,k = hl,k/|hl,k | represents the phase of the one-tap channel. The SIR is very
helpful because it determines the point at which interference starts to dominate the
noise, that is, SNR > SIR, leading to a saturation of the bit error rate (BER), see
[36]. As long as the SNR is approximately 10 dB lower than the SIR, interference is
completely dominated by noise and can thus be neglected. Once the SNR approaches
78
Optimal Subcarrier Spacing
SIR
Fig. 3.7 The SIR depends
on the subcarrier spacing.
For a fair comparison of
different modulation
schemes, we consider an
optimal subcarrier spacing
R. Nissel and M. Rupp
Limiting factor:
Doppler spread
(time-variant channel)
Limiting factor:
delay spread
(multipath delays)
Subcarrier Spacing
the SIR, that is, SNR = SIR, one observes a performance degeneration equivalent
to an SNR shift of approximately 3 dB. The matrix representation in Sect. 3.4 can
be utilized to calculate the SIR in (3.36). For example, the channel power can be
T
H
T
H H
⊗ gl,k
)Rvec{H} (gl,k
⊗ gl,k
) , with channel correlacalculated by E{|h l,k |2 } = (gl,k
2
2
H
N ×N
tion matrix Rvec{H} = E{vec{H}vec{H} } ∈ C
. For OQAM, on the other hand,
additional processing becomes necessary, see [41] for more details. Note that the SIR
can also be calculated with the ambiguity function, as, for example, demonstrated in
[15, 50].
In [36], we showed that FBMC (Hermite prototype filter) outperforms CP-OFDM
in high-velocity scenarios. This, however, was only true because interference from
the Doppler spread dominated interference from the delay spread. By increasing the
subcarrier spacing, the overall SIR could be improved, as illustrated in Fig. 3.7. The
big question is then, does FBMC still outperform CP-OFDM if both modulation
schemes apply an optimal subcarrier spacing? Because 5G will include a flexible
subcarrier spacing [3], our considerations here are also relevant for future wireless
systems. As a rule of thumb, the subcarrier spacing should be chosen so that [12]
τrms
σt
≈
,
σf
νrms
(3.39)
where time localization σt and frequency localization σ f are given by (3.19) for the
Hermite pulse and by (3.22) for the PHYDYAS pulse. For FBMC-OQAM, this leads
to the following optimal subcarrier spacings:
Fopt,PHYDYAS
*
νrms
,
τrms
*
νrms
≈ 0.91 ×
.
τrms
Fopt,Hermite ≈ 0.71 ×
(3.40)
(3.41)
For a Jakes Doppler spectrum, the root mean square (RMS) Doppler spread is given
by νrms = √12 vc f c , with v denoting the velocity, c the speed of light and f c the carrier
frequency. Note that (3.39) represents only an approximation. The exact relationship
can be calculated, as, for example, done in [17] for the Gaussian pulse, and depends
on the underlying channel model and the prototype filter. However, for our chosen
numerical parameters, the differences between the optimal SIR (exhaustive search)
3 Filter Bank Multicarrier Modulation
79
and the SIR obtained by applying the rule in (3.39) is less than 0.1 dB for FBMCOQAM and less than 1 dB for FBMC-QAM. For the rest of this section, we always
find the optimal subcarrier spacing in FBMC through an exhaustive search.
As a reference, we also consider an optimal subcarrier spacing in CP-OFDM.
The rule in (3.39), however, cannot be applied because the underlying rectangular
pulse is not localized in frequency. Instead, we assume, for a fixed CP overhead of
κ = TTCP0 = TCP F = T F − 1, that the subcarrier spacing is chosen as high as possible,
while still satisfying the condition of no Inter Symbol Interference (ISI), that is,
κ
TCP = τmax . This leads to F = τmax
. For a Jakes Doppler spectrum, the SIR can be
expressed by a generalized hypergeometric function 1 F2 (·) [48], which, together
with an optimal subcarrier spacing, leads to [41],
CP−OFDM
SIRopt.,
noISI =
1 F2
1 − 1 F2
1 3
; , 2; −
2 2
τmax
π νTmaxF−1
1 3
; , 2; −
2 2
2
τmax
π νTmaxF−1
2 .
(3.42)
For our numerical example, we consider a TDL-B channel model, as proposed by
3GPP [1, Sect. 7.7.3], and a carrier frequency of 2.5 GHz. Furthermore, we assume a
long delay spread of 300 ns. We expect that in future wireless systems, the “typical”
delay spread will be much lower than 300 ns [41]. Nonetheless, such a long delay
spread allows robustness considerations. The optimal subcarrier spacing for a velocity of zero approaches F → 0 Hz, not feasible in practice. We therefore assume that
the subcarrier spacing is lower bounded by F ≥ 15 kHz. For our channel parameters,
the SIR is illustrated in Fig. 3.8 and allows the following conclusions:
1. The SIR in FBMC is high enough, so that the channel induced self-interference
is usually dominated by noise. Thus, self-interference can be neglected.2
2. For an optimal subcarrier spacing, the Hermite prototype filter outperforms the
PHYDYAS prototype filter, but only by approximately 0.6 dB.
3. For low velocities, on the other hand, the PHYDYAS prototype filter becomes
better than the Hermite filter because of a better frequency localization in combination with a fixed subcarrier spacing (lower bound).
4. For a maximum symbol density of T F = 1 (complex), FBMC performs much
better than OFDM without CP, especially for low velocities.
5. CP-OFDM, see (3.42), performs best, but also has a lower data rate than FBMC
because of the CP overhead.
κ
and TCP = 0) has a lower SIR than CP-OFDM
6. WOLA (TW,TX = TW,RX = 2F
and performs close to FBMC in high-velocity scenarios.
7. In general, OFDM-based schemes perform good in low-velocity scenarios, while
in high-velocity scenarios, they lose most of their advantages when compared
with FBMC.
2 The SNR is often below 20 dB. Wireless systems are interference limited and what we call “noise”
is in practice often interference from other users and real-world hardware effects.
80
R. Nissel and M. Rupp
Channel Model: TDL-B, Long delay spread:
50
rms =
300 ns, fc = 2.5 GHz
Signal-to-Interference Ratio [dB]
T F = 1.07
CP-OFDM, T F = 1.07, F=51 kHz, no ISI
40
WOLA, T F = 1.07
30
FBMC-OQAM
(Hermite)
20
TF = 1
(complex)
FBMC-OQAM
(PHYDYAS)
OFDM (no CP)
CP-OFDM: F=51kHz
WOLA: F=44kHz
Hermite: F=37kHz
PHDYAS: F=51kHz
OFDM (no CP): F=22kHz
10
Subcarrier Spacing, F=15kHz
Subcarrier Spacing, F>15kHz
0
0
50
100
150
200
250
300
350
400
450
500
Velocity [km/h]
Fig. 3.8 The SIR is high enough so that the channel induced interference can usually be neglected
in FBMC because it is dominated by noise
Note that Fig. 3.8 represents a typical behavior. Different channel models and RMS
delay spread values only shift the curves to some other point [41]. The main conclusion here is that one-tap equalizers are in most practical cases sufficient. For all of
our testbed measurements in [34, 38, 42], low-complexity one-tap equalizers were
sufficient. Feedback delays and repeated handovers are usually more problematic
than a small, channel induced, interference.
For the sake of completeness, we also consider a highly double-selective channel, where we assume an RMS delay spread of 720 ns and a carrier frequency of
60 GHz, as suggested in [1]. However, we want to emphasize that such extreme
channel condition will rarely happen in practice and a system should therefore not
be optimized for it, but it should be able to cope with those scenarios, at least to some
extent. There exist several ways of dealing with such harsh channel environments:
Firstly, the employment of computationally demanding equalizers, as, for example, proposed in [40, 49]. Secondly, we can treat interference as noise and accept a
(small) throughput loss, see Fig. 3.2. Thirdly, spectral efficiency can be sacrificed
in order to gain robustness. Let us discuss the last method in more detail, where
we utilize the underlying orthogonality of the prototype filter in FBMC to transmit
complex-valued symbols. This is enabled by setting some symbols, corresponding
to the black markers in Figs. 3.4 and 3.5, to zero, so that a time–frequency spacing
of T F = 2 is achieved. We call this transmission scheme FBMC-QAM but there
does not exist a unique definition and different authors use this term differently. In
FBMC-QAM, complex-valued symbols are transmitted and no intrinsic imaginary
interference appears, allowing us to straightforwardly apply all known methods of
3 Filter Bank Multicarrier Modulation
81
Channel Model: TDL-B, Very long delay spread
50
rms =
Hermite: F=204kHz
PHDYAS: F=212kHz
WOLA: F=234kHz
CP-OFDM: F=292kHz
Signal-to-Interference Ratio [dB]
CP-OFDM, T F = 2, F=292 kHz, no ISI
WOLA, T F = 2
40
720 ns, fc = 60 GHz
FBMC-QAM (Hermite)
30
TF = 2
FBMC-QAM
(PHYDYAS)
20
FBMC-OQAM
(PHYDYAS)
10
TF = 1
(complex)
Subcarrier Spacing, F=60kHz
Subcarrier Spacing, F>60kHz
0
0
50
100
150
200
250
300
350
400
450
500
Velocity [km/h]
Fig. 3.9 In the rare case of a highly double-selective channel, it is possible to sacrifice spectral
efficiency (T F = 2) in order to improve robustness. FBMC-QAM then even outperforms CPOFDM
OFDM. Figure 3.9 shows the SIR for a highly double-selective channel. The following important observations can be identified.
1. The Hermite prototype filter performs much better than the PHYDYAS prototype
filter thanks to a better joint time–frequency localization. In FBMC-OQAM, this
effect is largely lost because of the time–frequency squeezing.
2. FBMC with a Hermite prototype filter outperforms even CP-OFDM for velocities
larger than 60 km/h.
3.6 Block Spread FBMC: Enabling All MIMO Methods
The loss of complex orthogonality is the main obstacle in FBMC-OQAM and seriously hampers some important transmission techniques, such as channel estimation
[35], Alamouti’s space-time block code [23] or maximum likelihood MIMO detection [62]. In particular, the limited MIMO compatibility3 is a major issue, preventing
a widespread application of FBMC.
In this section, we investigate a method which restores complex orthogonality in
FBMC-OQAM, so that all known techniques for OFDM can be straightforwardly
3 Only
some specific MIMO techniques become more challenging in FBMC. Many other MIMO
methods, such as receive diversity or spatial multiplexing based on Zero-Forcing (ZF) or MMSE
equalization, can be straightforwardly employed in FBMC.
82
R. Nissel and M. Rupp
employed in FBMC. This is enabled by adding an additional code dimension (besides
time and frequency). In contrast to conventional FBMC, the data symbols no longer
belong to a certain time–frequency position, but are rather spread over several time
or frequency positions. Such spreading typically increases the sensitivity to doubly
selective channels. However, if the delay spread and the Doppler spread are sufficiently low, the channel induced interference can still be neglected.
Spreading is very beneficial in FBMC because it can solve the underlying problem
of Alamouti’s space-time block code and maximum likelihood (ML) MIMO detection in FBMC. For example, authors in [46] proposed a block-Alamouti scheme (over
time) which can be seen as a special kind of spreading (distributing symbols in time).
The same method was recently applied by [30] in the frequency domain. However,
Walsh–Hadamard spreading [23, 34] offers more flexibility because it restores complex orthogonality, so that it not only works for Alamouti transmissions (as in [30,
46]), but additionally allows to straightforwardly employ all other methods known in
OFDM, such as channel estimation, other space-time block codes or low-complexity
maximum likelihood symbol detection. Similar to Walsh–Hadamard spreading,
authors in [62] propose FFT spreading in time to restore (quasi)-orthogonality.
Let us describe the spreading approach in more detail. In a first step, we assume
an AWGN channel, that is, H = I N , so that (3.30) transforms to
y = GH G x + GH n.
(3.43)
Note that (3.43) describes a block transmission of L subcarriers and K symbols in
time. Several of those blocks must be concatenated in time and frequency to achieve
a desired bandwidth and transmission time.
LK
In spread FBMC, complex-valued data symbols x̃ ∈ C 2 ×1 are precoded by a
LK
coding/spreading matrix C ∈ C L K × 2 , so that the transmitted symbols x ∈ C L K ×1
can be expressed by
x = C x̃.
(3.44)
A priori, the size of C and x̃ is unknown. We will explain later in this section why
LK
the size was chosen that way. The received data symbols ỹ ∈ C 2 ×1 are obtained by
decoding of the received symbols according to
ỹ = CH y.
(3.45)
To restore complex orthogonality, the coding matrix must be chosen so that the
following condition is fulfilled,
CH GH G C = IL K/2 .
(3.46)
A straightforward way to find coding matrix C is based on an eigenvalue decompoLK
sition of GH G = UΛUH , so that coding matrix C ∈ C L K × 2 becomes,
3 Filter Bank Multicarrier Modulation
83
⎡ −1/2
Λ1
⎢
⎢ 0
⎢
⎢
C = U⎢ 0
⎢ 0
⎢
⎣ :
0
0
..
.
0
⎤
⎥
0 ⎥
⎥
1
− /2
0 ΛL K/2 ⎥
⎥,
... 0 ⎥
⎥
:
: ⎦
... 0
(3.47)
Fig. 3.10 The eigenvalues
of GH G for an
FBMC-OQAM system.
Similar as for the derivation
of the MIMO channel
capacity, the eigenvalues in
combination with
eigenvector precoding can be
utilized to determine the
optimal precoding matrix
Eigenvalues of GH G
where Λi represents the i-th eigenvalue (sorted) of GH G and U the unitary eigenvector matrix. Figure 3.10 shows the eigenvalues of a typical FBMC-OQAM transmission matrix GH G. For the limit case of K → ∞ and L → ∞, the eigenvalues
are Λ1 = Λ2 = · · · = ΛL K/2 = 2 and Λi = 0 for i > L2K . Thus, (3.47) implicitly
applies water filling [56], so that C becomes the optimal spreading matrix in terms
of maximizing the information rate. In particular, it shows that the optimal size of
LK
the spreading matrix is L K × L2K and that any matrix, C ∈ C L K × 2 , which satisfies (3.46), is optimal for K → ∞ and L → ∞ (the SNR is always the same).
Moreover, it also shows that the intrinsic imaginary interference does not consist
of any useful information and can thus be canceled by taking the real part. For
a limited number of subcarriers and time symbols, the spreading matrix in (3.47)
no longer corresponds to the optimal solution. Instead, water filling could improve
the performance, where the column size of matrix C will usually be larger than
LK
. However, a spreading matrix of size L K × L2K , which satisfies (3.46), still
2
performs close to the optimum, as indicated by the eigenvalues in Fig. 3.10. For
example, for L = 36 and K = 30, the suboptimal spreading matrix performs only
3.6% worse in terms of achievable rate than the optimal spreading matrix (water
filling) for SNR values smaller than 20 dB. Furthermore, the optimal spreading
matrix requires different code rates for layers close to the eigenvalue of ΛL K/2 . This
increases the overall complexity, while the possible improvement is rather low, so
that, employing a slightly suboptimal spreading matrix makes sense in practical systems.(Note that precoding
power
by a factor of two, that
)
( reduces the
(
)
) average transmit
is, tr GCCH GH = tr CH GH GC = L2K = 21 tr GH G . Thus, for the same SNR
as without precoding, the data symbol power has to be increased by a factor of two.
It is also interesting that the noise after despreading is white, ñ ∼ C N (0, Pn IL K/2 ),
2
L = 12, K = 10
1.5
L = 36, K = 30
1
Cyclic repetition,
corresponds to
L
,K
0.5
0
0
0.2
0.4
0.6
i-th position / LK
0.8
1
84
R. Nissel and M. Rupp
Code
Time
Frequency
Fig. 3.11 In conventional FBMC-OQAM, real-valued symbols are transmitted over a rectangular
time–frequency grid (T F = 0.5). Two real-valued symbols are required to transmit one complexvalued symbol. Thus, the name “offset”-QAM, where we apply the offset not in time (as often in
c
literature) but in frequency. 2017
IEEE, [39]
even though the spreading matrix itself is not necessarily semi-unitary. The reason behind this is again the orthogonalization condition in (3.46), implying that
Rñ = CH GH Rn GC = Pn IL K/2 .
While the spreading matrix in (3.47) provides analytical insight, it is not very
practical because of a high computational complexity and the fact that the spreading is performed in both, time and frequency, which only works for a doubly flat
channel. Walsh–Hadamard spreading [22, 23, 34, 39], on the other hand, is a much
more practical solution because it requires almost no additional complexity and the
spreading is performed in only one dimension, either in time or in frequency. In conventional FBMC-OQAM, see Fig. 3.11, each time–frequency position can only carry
real-valued symbols, so that two time–frequency positions are required to transmit
one complex-valued data symbol, indicated by the color in Fig. 3.11. In block spread
FBMC-OQAM, on the other hand, data symbols no longer belong to a specific time–
frequency position, but are spread over several subcarriers, see Fig. 3.12. To keep the
spectral efficiency the same as in FBMC-OQAM (ignoring possible guard symbols),
several data symbols are transmitted over the same time–frequency resources, but differentiated by their spreading/coding sequence. To be specific, L/2 complex-valued
data symbols are spread over L subcarriers. This leads to the same information rate as
in conventional FBMC-OQAM (again, ignoring possible guard symbols). Although
complex orthogonality can be perfectly restored within one block, there still exists
interference between different blocks. Thus, if we spread in frequency, a guard subcarrier might be necessary. If we spread in time, on the other hand, a guard symbol in
time might be required. Such guard symbols typically reduce the spectral efficiency
by a few percent, see [34, 39].
Spreading in frequency can be described by frequency spreading matrix C f ∈
L× L2
R
for which we take every second column out of a sequency-ordered [25] Walsh–
Hadamard matrix H ∈ R L×L , that is,
3 Filter Bank Multicarrier Modulation
85
Code
Time
Guard subcarrier
Channel is approximately
frequency-flat
Frequency
Fig. 3.12 In block spread FBMC-OQAM, complex-valued symbols are spread over several subcarriers (or time positions), allowing to restore complex orthogonality within one block. To improve
c
the SIR between different blocks, a guard symbol might be necessary. 2017
IEEE, [39]
[C f ]l,m = [H]l,2m−1
; for l = 1, 2, . . . , L; m = 1, 2, . . . ,
L
.
2
(3.48)
Note that matrix C f in (3.48) could equivalently be defined by [H]l,2m . Utilizing
the underlying structure of our matrix notation (vectorization) and the fact that we
LK
spread in frequency only, allows us to express overall spreading matrix C ∈ R L K × 2
by,
C = IK ⊗ C f ,
(3.49)
where Kronecker product ⊗ together with identity matrix I K map coding matrix C f
to the correct time slots.
Spreading in time can be described in a similar way except that we have to alternate
K
K
between spreading with Ct ′ ∈ R K × 2 and spreading with Ct ′′ ∈ R K × 2 for adjacent
subcarriers. The spreading matrices itself are again found by taking every second
column out of a sequency-ordered [25] Walsh–Hadamard matrix H ∈ R K ×K , that
is,
[Ct ′ ]k,m = [H]k,2m−1
[Ct ′′ ]k,m = [H]k,2m
; for k = 1, 2, . . . , K ; m = 1, 2, . . . ,
K
.
2
(3.50)
LK
To find the overall spreading matrix C ∈ R L K × 2 , we have to map the individual
spreading matrices Ct ′ and Ct ′′ to the correct subcarrier positions. For the vectorized
system model in (3.26), this implies that,
+ ,
,
00
10
C = Ct ′ ⊗ IL/2 ⊗
,
+ Ct ′′ ⊗ IL/2 ⊗
01
00
+
(3.51)
where the matrices 01 00 and 00 01 are necessary to alternate between spreading with
Ct ′ and Ct ′′ for adjacent subcarriers.
86
R. Nissel and M. Rupp
It can easily be checked by numerical evaluations that (3.49) and (3.51) satisfy
the complex orthogonalization condition in (3.46). For a formal proof that Walsh–
Hadamard spreading restores complex orthogonality in FBMC, we refer to [22].
Authors of [22] left the question open whether it is possible to find a spreading
matrix that has more than L2K columns while still satisfying (3.46). Our investigations in (3.47) show that this is not possible (ignoring any edge effects which become
negligible for a large number of K and L). A small disadvantage of Walsh–Hadamard
spreading is the fact that the spreading length has to be a power of two. This makes
the integration into existing systems problematic, but has almost no impact if a system is designed from scratch. The big advantage of Walsh–Hadamard spreading, on
the other hand, is that only additions, but no multiplications are needed. Thus, the
additional computational complexity is very low. Moreover, a fast Walsh–Hadamard
transformation can be used, further reducing the computational complexity. For
example, spreading in time only requires log2 (K ) − 1 extra additions/subtractions
per data symbol at the transmitter and log2 (K ) extra additions/subtractions per data
symbol at the receiver. For spreading in frequency, it is log2 (L) − 1, respectively
log2 (L).
Similar as in (3.29), it is possible to include a doubly selective channel into
our transmission model. The input output relationship between the transmitted data
symbols x̃ and the received data symbols ỹ can then be modeled by
ỹ = CH GH H GC x̃ + CH GH n
H
H
≈ diag{h̃} x̃ + C G n.
(3.52)
(3.53)
Similar as in (3.29), if the delay spread and the Doppler spread are sufficiently low,
the transmission can be approximated by a one-tap channel, see (3.53). If we spread
in time, block spread FBMC becomes more sensitive to time-variant channels. At
the same time, it becomes slightly more robust to multipath delays. For spreading
in frequency, the opposite holds true. In [34, 39], we discuss the effect of doubly
selective channels on block spread FBMC.
Precoding by C can also be interpreted
as transforming the underlying basis pulses
according to G̃ = GC = g̃1 · · · g̃L K/2 . Thus, instead of modulating data symbols
with gl,k (t), as in (3.1), we modulate them with g̃i (t). In contrast to conventional multicarrier systems, however, the transformed basis pulses g̃i (t) no longer all employ
the same underlying prototype filter p(t). Instead, many basis pulses have their own,
unique, prototype filter pi (t). Thus, we cannot directly implement G̃ in an efficient
way. However, by interpreting G̃ as a precoded FBMC system, the advantage of an
efficient signal generation are preserved. Moreover, such interpretation offers overall
a high flexibility.
We have validated the block spread FBMC approach by real-world testbed measurements in [34] for outdoor-to-indoor scenarios (150 m link distance, 2.5 GHz
carrier frequency) and in [42] for indoor-to-indoor scenarios (5 m link distance,
60 GHz carrier frequency). For both measurements, the assumption of a low delay
spread and a low Doppler spread was fulfilled, so that (3.53) accurately described
3 Filter Bank Multicarrier Modulation
87
2.5 GHz, 16-QAM, F=15 kHz
Bit Error Ratio
10
−1
2×2 MIMO,
ML detection
60 GHz, 4-QAM,
F=500 kHz,
10
−2
2×1
Alamouti
2×1
Alamouti
CP−OFDM
Spread FBMC
10
−3
−5
0
5
10
15
20
25
Signal-to-Noise Ratio for OFDM [dB]
Fig. 3.13 Real-world testbed measurements [34, 42] show that MIMO works in FBMC once
symbols are spread in time. The spreading process itself has a low computational complexity
because of a fast Walsh–Hadamard transformation. FBMC and OFDM experience both the same
c
BER, but FBMC has lower OOB emissions. 2017
IEEE, [41]
the true physical behavior. Because of a time-invariant channel, we spread in time
instead of frequency, see (3.50) and (3.51). For the 60 GHz measurement setup, we
employed a high subcarrier spacing of F = 500 kHz, as often considered in millimeter wave transmissions [44]. This implicitly reduces the latency, so that even
though we spread symbols in time, the overall transmission time was less than 40 µs,
satisfying the low-latency condition of 100 µs [14].
Figure 3.13 shows the measured BER over SNR. Alamouti’s space-time block
code and low-complexity ML MIMO detection performs in FBMC as good as in
OFDM, but FBMC has the advantage of much lower OOB emissions.
3.7 Pruned DFT-Spread FBMC-OQAM: Reducing the
PAPR
Besides the intrinsic imaginary interference, nonlinearities, such as a limited Digitalto-Analog Converter (DAC) resolution or a nonlinear power amplifier, are even more
problematic in practical systems because they destroy the superior spectral properties
of FBMC [41, 50]. Thus, the concept of sharp digital filters to enable a flexible time–
frequency allocation, as discussed in [41], only works as long as FBMC operates in
the linear regime. In multicarrier systems, this is, in general, hard to achieve because
of the poor peak-to-average power ratio (PAPR). To reduce the PAPR in practical
systems, LTE employs single carrier frequency division multiple access (SC-FDMA)
in the uplink [54], a DFT precoded OFDM system. The same technique will also
be used in 5G [3]. Unfortunately, simply combining FBMC and a DFT, as done
in SC-FDMA for OFDM, performs poorly in FBMC [18, 31, 61]. This motivated
us to develop pruned DFT-spread FBMC [32, 37], a novel transmission technique
R. Nissel and M. Rupp
Fig. 3.14 Precoding matrix
C shapes the transmitted
signal in such a way, that the
average transmit power
(diagonal elements of
E{s sH } = GC CH GH )
shows an almost perfect
rectangular shape, with many
beneficial properties. For the
illustration we consider only
one FBMC symbol, that is,
K =1
Transmit Power, E{|s(t)|2 }
88
1
0.8
diag{GC CH GH }
Pruned DFT-s
FBMC
0.6
0.4
Conventional
FBMC-OQAM
diag{GGH }
0.2
0
−1
−0.5
0
0.5
1
Normalized Time, t/T0
with the remarkable properties of a low PAPR, low-latency transmissions and a
high spectral efficiency. The idea of pruned DFT-spread FBMC is closely related
to Walsh–Hadamard spreading, see Sect. 3.6. In particular, we again spread L2 data
L
symbols over L subcarriers, described by frequency spreading matrix C f ∈ C L× 2 ,
C f = W L×L/2 diag{b},
L
(3.54)
where W L×L/2 ∈ C L× 2 describes a pruned DFT matrix, that is, a conventional DFT
L
matrix for which L2 column vectors are canceled. Scaling vector b ∈ R 2 ×1 , on the
other hand, guarantees that the diagonal elements of CH GH G C are exactly one, with
C = IK ⊗ C f .
Figure 3.14 shows the expected transmit power over time for one FBMC symbol.
In conventional FBMC, there exists a large overlapping of symbols in time and the
transmission requires a long ramp-up and ramp-down period. In pruned DFT-spread
FBMC, on the other hand, precoding by C f shapes the transmitted signal in such a
way, that the overlapping in time is very low and the ramp-up and ramp-down period
dramatically reduced. This reduces the overall latency. The pruned DFT matrix in
(3.54) is found by canceling those column vectors of a conventional DFT matrix,
so that the main energy is concentrated within the time interval − T40 ≤ t ≤ T40 , see
Fig. 3.14. The OOB emissions of pruned DFT-spread FBMC are comparable to
conventional FBMC transmissions, leading to a high spectral efficiency.
From a conceptional point of view, the key difference between pruned DFTspread FBMC and block spread FBMC, discussed in Sect. 3.6, is that pruned
DFT-spread FBMC spreads the data symbols over the whole bandwidth, while for
block spread FBMC the bandwidth is split into smaller chunks. Those small chunks
can then be equalized by a simple one-tap equalizer, so that Alamouti’s space-time
block code and ML MIMO detection become feasible. In pruned DFT-spread FBMC,
on the other hand, low-complexity ML detection is often not possible and one has to
rely on minimum mean squared error (MMSE) equalization before despreading by
CHf , same as in SC-FDMA. Another small drawback of pruned DFT-spread FBMC is
3 Filter Bank Multicarrier Modulation
10
4QAM, L = 128 Subcarriers
0
CP-OFDM
10
−1
FBMC
CCDF
Fig. 3.15 Pruned
DFT-spread FBMC [32, 37]
has the same PAPR as
SC-FDMA, but the
additional advantages of a
higher spectral efficiency.
Note that a simple
DFT-spread FBMC
transmission scheme
performs relatively poor [18,
31]
89
10
10
−2
−3
SC-FDMA
Simple
DFT-s
FBMC
Pruned DFT-s FBMC
5
6
7
8
9
10
11
Peak-to-Average Power Ratio (PAPR) [dB]
that orthogonality is only approximately restored, that is, CH GH G C ≈ IL K/2 , leading
to some (small) interference. By restricting the time domain of p(t) to approximately
− 43 T0 ≤ t ≤ 43 T0 , the interference can be reduced, so that it becomes neglectable in
most cases. Moreover, a frequency CP can further reduce the interference [32, 37],
if necessary.
Figure 3.15 shows the complementary cumulative distribution function (CCDF)
of the PAPR for a 4-QAM signal constellation and L = 128 subcarriers. Conventional FBMC has the same poor PAPR as OFDM. A simple DFT-spread FBMC
transmission scheme, as proposed in [18], only slightly improves the PAPR. Even an
π
π
optimal phase condition [31], that is, e j 2 (l+k) → e j 2 (l+k) e−jπlk , hardly reduces the
PAPR. In pruned DFT-spread FBMC, on the other hand, the PAPR is as good as in
SC-FDMA and approximately 3 dB better than in OFDM and FBMC.
3.8 Summary
FBMC has the best spectral properties among all 5G waveform candidates. This is
especially useful if the number of subcarriers is low, for example, in eMTC, and to
support different use cases within the same band. To efficiently generate an FBMC
signal, many hardware components from windowed OFDM can be reused. The main
drawback of FBMC is that orthogonality only holds in the real domain, which makes
some techniques, such as channel estimation or some MIMO transmission methods, more challenging. However, there exist several solutions to overcome those
limitations.
Future wireless systems will be characterized by a relatively low delay spread,
so that the channel induced self-interference in FBMC is very low and can often be
neglected. This is even more true if an optimal subcarrier spacing is employed. To
restore complex orthogonality in FBMC, one can spread data symbols over several
time or frequency positions. In this context, we presented block spread FBMC which
90
R. Nissel and M. Rupp
allows to straightforwardly employ all known methods from OFDM in FBMC (if
the delay spread and the Doppler spread are sufficiently low). If, on the other hand,
the focus lies on reducing the PAPR, pruned DFT-spread FBMC is a better option.
References
1. 3GPP, TR 38.900: study on channel model for frequency spectrum above 6GHz (release 14)
(2016a), http://www.3gpp.org/DynaReport/38900.htm
2. 3GPP, TR 38.913: study on scenarios and requirements for next generation access technologies
(release 14) (2016b), http://www.3gpp.org/DynaReport/38913.htm
3. 3GPP, TR 38.802: study on new radio access technology; physical layer aspects (release 14)
(2017), http://www.3gpp.org/DynaReport/38802.htm
4. J.G. Andrews, S. Buzzi, W. Choi, S.V. Hanly, A. Lozano, A.C. Soong, J.C. Zhang, What will
5G be? IEEE J. Sel. Areas Commun. 32(6), 1065–1082 (2014)
5. P. Banelli, S. Buzzi, G. Colavolpe, A. Modenini, F. Rusek, A. Ugolini, Modulation formats
and waveforms for 5G networks: who will be the heir of OFDM? an overview of alternative
modulation schemes for improved spectral efficiency. IEEE Signal Process. Mag. 31(6), 80–93
(2014)
6. M. Bellanger, D. Le Ruyet, D. Roviras, M. Terré, J. Nossek, L. Baltar, Q. Bai, D. Waldhauser,
M. Renfors, T. Ihalainen et al., FBMC Physical Layer: A Primer. PHYDYAS (2010)
7. M.G. Bellanger, Specification and design of a prototype filter for filter bank based multicarrier
transmission, in IEEE International Conference on Acoustics, Speech, and Signal Processing,
vol. 4, (2001), pp. 2417–2420
8. H. Bölcskei, Orthogonal frequency division multiplexing based on offset QAM, in Advances
in Gabor analysis (Springer, 2003), pp. 321–352
9. R.W. Chang, Synthesis of band-limited orthogonal signals for multichannel data transmission.
Bell Syst. Tech. J. 45(10), 1775–1796 (1966)
10. J.M. Choi, Y. Oh, H. Lee, J.S. Seo, Pilot-aided channel estimation utilizing intrinsic interference
for FBMC/OQAM systems. IEEE Trans. Broadcast. 63(4), 644–655 (2017)
11. W. Cui, D. Qu, T. Jiang, B. Farhang-Boroujeny, Coded auxiliary pilots for channel estimation
in FBMC-OQAM systems. IEEE Trans. Veh. Technol. 65(5), 2936–2946 (2016)
12. B. Farhang-Boroujeny, OFDM versus filter bank multicarrier. IEEE Signal Process. Mag. 28(3),
92–112 (2011)
13. H.G. Feichtinger, T. Strohmer, Gabor Analysis and Algorithms: Theory and Applications
(Springer Science & Business Media, 2012)
14. G.P. Fettweis, The tactile internet: applications and challenges. IEEE Veh. Technol. Mag. 9(1),
64–70 (2014)
15. M. Fuhrwerk, J. Peissig, M. Schellmann, Channel adaptive pulse shaping for OQAM-OFDM
systems, in IEEE European Signal Processing Conference (EUSIPCO) (2014), pp. 181–185
16. R. Haas, J.C. Belfiore, A time-frequency well-localized pulse for multiple carrier transmission.
Wirel. Pers. Commun. 5(1), 1–18 (1997)
17. F.M. Han, X.D. Zhang, Wireless multicarrier digital transmission via Weyl-Heisenberg frames
over time-frequency dispersive channels. IEEE Trans. Commun. 57(6) (2009)
18. T. Ihalainen, A. Viholainen, T.H. Stitz, M. Renfors, M. Bellanger, Filter bank based multi-mode
multiple access scheme for wireless uplink, in IEEE European Signal Processing Conference
(EUSIPCO) (2009), pp. 1354–1358
19. J.P. Javaudin, D. Lacroix, A. Rouxel, Pilot-aided channel estimation for OFDM/OQAM, in
IEEE Vehicular Technology Conference (VTC), vol. 3 (2003), pp. 1581–1585
20. E. Kofidis, D. Katselis, A. Rontogiannis, S. Theodoridis, Preamble-based channel estimation
in OFDM/OQAM systems: a review. Signal Proces. 93(7), 2038–2054 (2013)
3 Filter Bank Multicarrier Modulation
91
21. C. Lélé, R. Legouable, P. Siohan, Channel estimation with scattered pilots in OFDM/OQAM, in
IEEE Workshop on Signal Processing Advances in Wireless Communications (SPAWC) (2008),
pp. 286–290
22. C. Lélé, P. Siohan, R. Legouable, M. Bellanger, CDMA transmission with complex
OFDM/OQAM. EURASIP J. Wirel. Commun. Netw. Article ID 748063, 1–12 (2008)
23. C. Lélé, P. Siohan, R. Legouable, The Alamouti scheme with CDMA-OFDM/OQAM.
EURASIP J. Adv. Signal Process. Article ID 703513, 1–13 (2010)
24. Y.G. Li, G.L. Stuber, Orthogonal Frequency Division Multiplexing for Wireless Communications (Springer Science & Business Media, 2006)
25. J. Manz, A sequency-ordered fast Walsh transform. IEEE Trans. Audio Electroacoustic 20(3),
204–205 (1972)
26. D. Mattera, M. Tanda, M. Bellanger, Filter bank multicarrier with PAM modulation for future
wireless systems. Signal Process. 120, 594–606 (2016)
27. G. Matz, D. Schafhuber, K. Grochenig, M. Hartmann, F. Hlawatsch, Analysis, optimization,
and implementation of low-interference wireless multicarrier systems. IEEE Trans. Wirel.
Commun. 6(5), 1921–1931 (2007)
28. S. Mirabbasi, K. Martin, Design of prototype filter for near-perfect-reconstruction overlapped
complex-modulated transmultiplexers, in IEEE International Symposium on Circuits and Systems (2002)
29. A.F. Molisch, Wireless Communications, vol. 34 (Wiley, 2012)
30. D. Na, K. Choi, Intrinsic ICI-free Alamouti coded FBMC. IEEE Commun. Lett. 20(10), 1971–
1974 (2016)
31. D. Na, K. Choi, Low PAPR FBMC. IEEE Trans. Wirel. Commun. 17(1), 182–193 (2018)
32. R. Nissel, Filter bank multicarrier modulation for future wireless systems. Dissertation, TU
Wien, 2017
33. R. Nissel, M. Rupp, Bit error probability for pilot-symbol aided channel estimation in FBMCOQAM, in IEEE International Conference on Communications (ICC) (2016a)
34. R. Nissel, M. Rupp, Enabling low-complexity MIMO in FBMC-OQAM, in IEEE Globecom
Workshops (GC Wkshps) (2016b)
35. R. Nissel, M. Rupp, On pilot-symbol aided channel estimation in FBMC-OQAM, in IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2016c), pp.
3681–3685
36. R. Nissel, M. Rupp, OFDM and FBMC-OQAM in doubly-selective channels: Calculating the
bit error probability. IEEE Commun. Lett. 21(6), 1297–1300 (2017)
37. R. Nissel, M. Rupp, Pruned DFT spread FBMC-OQAM: low-PAPR, low latency, high spectral
efficiency. IEEE Trans. Commun. (2018)
38. R. Nissel, S. Caban, M. Rupp, Experimental evaluation of FBMC-OQAM channel estimation
based on multiple auxiliary symbols, in IEEE Sensor Array and Multichannel Signal Processing
Workshop (SAM) (2016)
39. R. Nissel, J. Blumenstein, M. Rupp, Block frequency spreading: a method for low-complexity
MIMO in FBMC-OQAM, in IEEE Workshop on Signal Processing Advances in Wireless
Communications (SPAWC) (2017a)
40. R. Nissel, M. Rupp, R. Marsalek, FBMC-OQAM in doubly-selective channels: a new perspective on MMSE equalization, in IEEE Workshop on Signal Processing Advances in Wireless
Communications (SPAWC) (2017b)
41. R. Nissel, S. Schwarz, M. Rupp, Filter bank multicarrier modulation schemes for future mobile
communications. IEEE J. Sel. Areas Commun. 35(8), 1768–1782 (2017c)
42. R. Nissel, E. Zöchmann, M. Lerch, S. Caban, M. Rupp, Low-latency MISO FBMC-OQAM:
it works for millimeter waves! in IEEE International Microwave Symposium (2017d)
43. R. Nissel, E. Zöchmann, M. Rupp, On the influence of doubly-selectivity in pilot-aided channel
estimation for FBMC-OQAM, in IEEE Vehicular Technology Conference (VTC Spring) (2017e)
44. Z. Pi, F. Khan, An introduction to millimeter-wave mobile broadband systems. IEEE Commun.
Mag. 49(6), 101–107 (2011)
92
R. Nissel and M. Rupp
45. Qualcomm Incorporated, Waveform candidates, in 3GPP TSG-RAN WG1 84b (Busan, Korea,
2016)
46. M. Renfors, T. Ihalainen, T. H. Stitz, A block-Alamouti scheme for filter bank based multicarrier
transmission, in European Wireless Conference (EW) (2010)
47. M. Renfors, X. Mestre, E. Kofidis, F. Bader, Orthogonal Waveforms and Filter Banks for Future
Communication Systems (Academic Press, 2017)
48. P. Robertson, S. Kaiser, The effects of Doppler spreads in OFDM(A) mobile radio systems, in
IEEE Vehicular Technology Conference (VTC Fall) (1999), pp. 329–333
49. F. Rottenberg, X. Mestre, D. Petrov, F. Horlin, J. Louveaux, Parallel equalization structure
for MIMO FBMC-OQAM systems under strong time and frequency selectivity. IEEE Trans.
Signal Process. 65(17), 4454–4467 (2017)
50. A. Sahin, I. Guvenc, H. Arslan, A survey on multicarrier communications: prototype filters,
lattice structures, and implementation aspects. IEEE Commun. Surv. Tutor. 16(3), 1312–1338
(2012)
51. B. Saltzberg, Performance of an efficient parallel data transmission system. IEEE Trans. Commun. Technol. 15(6), 805–811 (1967)
52. F. Schaich, T. Wild, Y. Chen, Waveform contenders for 5G-suitability for short packet and low
latency transmissions, in IEEE Vehicular Technology Conference (VTC Spring) (2014), pp. 1–5
53. F. Schaich, T. Wild, R. Ahmed, Subcarrier spacing-how to make use of this degree of freedom,
in IEEE Vehicular Technology Conference (VTC Spring) (2016), pp. 1–6
54. S. Sesia, M. Baker, I. Toufik, LTE-the UMTS Long Term Evolution: from Theory to Practice
(Wiley, 2011)
55. P. Siohan, C. Siclet, N. Lacaille, Analysis and design of OFDM/OQAM systems based on
filterbank theory. IEEE Trans. Signal Process. 50(5), 1170–1183 (2002)
56. E. Telatar, Capacity of multi-antenna gaussian channels. Trans. Emerg. Telecommun. Technol.
10(6), 585–595 (1999)
57. M. Vetterli, J. Kovačević, V.K Goyal, Foundations of Signal Processing (Cambridge University
Press, 2014)
58. S. Weinstein, P. Ebert, Data transmission by frequency-division multiplexing using the discrete
fourier transform. IEEE Trans. Commun. Technol. 19(5), 628–634 (1971)
59. G. Wunder, P. Jung, M. Kasparick, T. Wild, F. Schaich, Y. Chen, S. Brink, I. Gaspar, N.
Michailow, A. Festag et al., 5GNOW: non-orthogonal, asynchronous waveforms for future
mobile applications. IEEE Commun. Mag. 52(2), 97–105 (2014)
60. B. Yu, S. Hu, P. Sun, S. Chai, C. Qian, C. Sun, Channel estimation using dual-dependent pilots
in FBMC/OQAM systems. IEEE Commun. Lett. 20(11), 2157–2160 (2016)
61. C.H. Yuen, P. Amini, B. Farhang-Boroujeny, Single carrier frequency division multiple access
(SC-FDMA) for filter bank multicarrier communication systems, in IEEE International Conference on Cognitive Radio Oriented Wireless Networks (CROWNCOM) (2010), pp. 1–5
62. R. Zakaria, D. Le Ruyet, A novel filter-bank multicarrier scheme to mitigate the intrinsic
interference: application to MIMO systems. IEEE Trans. Wirel. Commun. 11(3), 1112–1123
(2012)
63. X. Zhang, M. Jia, L. Chen, J. Ma, J. Qiu, Filtered-OFDM-enabler for flexible waveform in the
5th generation cellular networks, in IEEE Global Communications Conference (GLOBECOM)
(2015), pp. 1–6
Chapter 4
Generalized Frequency Division
Multiplexing: A Flexible Multicarrier
Waveform
Ahmad Nimr, Shahab Ehsanfar, Nicola Michailow, Martin Danneberg,
Dan Zhang, Henry Douglas Rodrigues, Luciano Leonel Mendes
and Gerhard Fettweis
4.1 Introduction to GFDM Modulator
In this section, we give an overview of generalized frequency division multiplexing
(GFDM) waveform. We start from the continuous model representation, from which
an analytical expression of power spectral density (PSD) is derived. Based on that,
a discrete signal model representation is derived and expressed in terms of matrix
model. We show the structure of the modulation matrix by mean of decomposition.
A. Nimr (B) · S. Ehsanfar · M. Danneberg · D. Zhang · G. Fettweis
Vodafone Chair Mobile Communication Systems, Technische Universität Dresden,
Dresden, Germany
e-mail: ahmad.nimr@ifn.et.tu-dresden.de
S. Ehsanfar
e-mail: shahab.ehsanfar@ifn.et.tu-dresden.de
M. Danneberg
e-mail: martin.danneberg@ifn.et.tu-dresden.de
D. Zhang
e-mail: dan.zhang@ifn.et.tu-dresden.de
G. Fettweis
e-mail: gerhard.fettweis@ifn.et.tu-dresden.de
N. Michailow
National Instruments Corp., 11500, Mopac Expwy, Austin, TX 78759, USA
e-mail: nicola.michailow@ni.com
H. D. Rodrigues · L. L. Mendes
Intituto Nacional de Telecomunicações (Inatel), Santa Rita do Sapucai, MG, Brazil
e-mail: henry@inatel.br
L. L. Mendes
e-mail: lucianol@inatel.br
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_4
93
94
A. Nimr et al.
From this model, the design requirements and performance indicators are derived.
Finally, we show the ability of GFDM representation to generate other state-of-the-art
waveforms.
4.1.1 Continuous Signal Model
GFDM is a block-based multicarrier modulation technique that employs circular
filtering [1]. For a better understanding of the GFDM structure, we represent the
modulation technique using the continuous time model. Consider a time-frequency
resource block defined by time duration T and frequency bandwidth B. The target
is to use this resource to convey a data message of maximum length of N data
symbols. For this purpose, the available bandwidth is divided into K equally spaced
subcarriers with subcarrier spacing Δf = KB , and the available time is divided into
M subsymbols with subsymbol spacing Tsub = MT . The subcarrier spacing is related
, where
to the subsymbol spacing with the relation ΔfTsub = 1. Hence, T = NB = Δf
M
N = KM . Each pair (k, m)-(subcarrier, subsymbol) can be used to transmit one data
symbol dk,m modulated by a pulse shape gk,m (t) given by
gk,m (t) = wT (t)gT (t − mTsub )ej2πkΔft ,
(4.1)
where wT (t) is a rectangular window of duration T , namely wT (t) = 1, t ∈ [0, T ]
and 0 elsewhere. gT (t) is a prototype periodic pulse shape of period T , which can be
expressed using Fourier series. This means that the pulse shapes are generated by time
and frequency shifts of a periodic prototype pulse shape in addition to multiplication
with a finite time window to form the GFDM block. In conventional GFDM, the
number of the frequency components of gT is limited to 2M such that,
gT (t) =
M
−1
q=−M
q
g̃T [q]ej2π M Δft ,
(4.2)
where g̃T [q] are the nonzero coefficients of the Fourier series. This allows each
subcarrier to span at maximum two subcarrier spacing. The frequency response of
the prototype pules shape is given by
G T (f ) =
M
−1
q=−M
g̃T [q]δ(f −
q
M
Δf ).
(4.3)
Here, δ(·) is the Dirac pulse. GFDM commonly adopts a cyclic prefix (CP) of duration
Tcp ≥ τmax to tackle the impact of fading channel with maximum excess delay spread
τmax . In addition, a cyclic suffix (CS) with duration Tcs may be added to the end of
the block. The CP and CS can be simply introduced by extending the window to
4 Generalized Frequency Division Multiplexing …
95
Fig. 4.1 Pulse shape generation using periodic Raised-Cosine prototype pulse shape, with roll-off
factor α = 0, M = 4 for k = 0, m = 0, . . . , 3
wTs (t) where Ts = T + Tcp + Tcs , as depicted in Fig. 4.1. Mathematically, it can be
written as
(t)
gk,m
(t) = wTs (t − Tcp )gT (t − mTsub )ej2πkΔft .
(4.4)
Therefore, the frequency domain representation is given by
−j2πTcp (f −kΔf )
∗ G T (f )e−j2πmTsub f ,
G (t)
k,m (f ) = WTs (f − kΔf )e
(4.5)
where ∗ denotes the convolution operator. Replacing G T (f ) from (4.3), we get
(t)
G k,m (f ) = e−j2π Tcp (f −kΔf )
M
−1
q=−M
q
q
Δf
g̃T [q]WTs (f − (kM + q) M )e−j2π m M e+j2π Tcp Δf M .
(4.6)
In practice, not all subcarriers and subsymbols are used, thus, we define Kon and Mon
as the sets of active subcarriers and subsymbols, respectively. Therefore, the signal
corresponding to the i-th GFDM block that modulates the data symbols {dk,m,i } is
generated as
(t)
dk,m,i gk,m
(t).
(4.7)
xi (t) =
m∈M on k∈K on
Moreover, the signal of a frame that contains Ns blocks can be expressed as
x(t) =
N
s −1
i=0
xi (t − iTs ) =
N
s −1
i=0 m∈M on k∈K on
(t)
dk,m,i gk,m
(t − iTs ).
(4.8)
From (4.8) and (4.6), the PSD
of the GFDM signal assuming uncorrelated data symbols with unit power, i.e., E dk∗1 ,m1 ,i1 dk2 ,m2 ,i2 = δ(k1 − k2 )δ(m1 − m2 )δ(i1 − i2 ), can
be computed as,
96
A. Nimr et al.
Fig. 4.2 PSD of subsymbols using a periodic raised-cosine (RC) prototype pulse with roll-off
factor α = 0, M = 4 for k = 0, m = 0, . . . , 3
Sx (f ) =
1
Ts
1
=
Ts
m∈M
on
2
(t)
G k,m (f )
k∈K
on
−1
q
M
−j2π m M +j2π Tcp Δf
g̃T [q]WTs (f − (kM + q) Δf
e
M )e
m∈M on k∈K on q=−M
q
M
2 (4.9)
.
According to (4.9), the PSD is obviously influenced by the prototype pulse shape
coefficients {g̃T [q]}, the window WTs (t) and the number of subsymbols M . In addition, the active subsymbol set Mon plays an important role throughout the phase term
q
e−j2πm M , which depends on the subsymbol index m. To clarify that, Fig. 4.2 shows
the individual PSD of each subsymbol using a rectangular window and no CP nor
CS. As can be seen, the first subsymbol, namely m = 0, is the source of high out-ofband (OOB) emission for the selected prototype pulse shape. This can be intuitively
understood via the discontinuity between successive blocks, which happens when
dk,0,i and dk,0,i+1 are not identical.
4.1.2 Discrete Signal Model
The discrete time signal representation can be derived from the sampling of the
analog signal with frequency Fs = B. With that, we get K, N , Lcp , and Lcs samples
per subsymbol, symbol, CP and CS, respectively. The discrete prototype pulse shape
results from (4.2) with
4 Generalized Frequency Division Multiplexing …
g[n] =
M
−1
q=−M
97
qn
g̃T [q]ej2π N , n = 0, . . . , N − 1.
(4.10)
Let g̃ = N-DFT {g} be the N -point finite discrete Fourier transform (DFT) of g such
that
N −1
qn
1
g[n] =
g̃[q]ej2π N ,
(4.11)
N q=0
then, the relation between the Fourier series coefficient g̃T [q] of the continuous model
and the frequency bins {g̃[q]} of the discrete model can be expressed as,
g̃[q] =
1
g̃T [< q >N ],
N
(4.12)
where < · >N is the modulo-N operator. Assuming a rectangular window as in (4.1),
we get
k
(4.13)
gk,m [n] = g[< n − mK >N ]ej2π K n .
Thus, the subcarrier–subsymbol pulse shapes are generated from the circular shift of
the prototype pulse shape in the time and frequency domains. In fact, the circularity
in time results from the design with periodic pulse shape and in frequency from the
sampling. Consequently, the design of the prototype pulse shape g[n] can be carried
out in the frequency domain such that only the first and last M samples of g̃ have
nonzero values. This ensures limited inter-carrier interference (ICI) to only adjacent
subcarriers under the assumption of perfect synchronization. However, in the case of
asynchronous subcarriers, we resort to the continuous discrete-time Fourier transform
(DTFT) G k,m (ν) = DFT(gk,m [n]), which takes into account the frequency response
of the window as well. Actually, g̃k,m [q] = G k,m (ν = Nq ), as illustrated in Fig. 4.3.
Fig. 4.3 ICI of adjacent subcarriers using periodic RC prototype pulse with roll-off factor α = 0.5,
M = 16 for k = −1, 0, 1, m = 0. The sampling points stand for g̃k,m [q], and the solid line represents
G k,m (ν)
98
A. Nimr et al.
Next, we focus on the matrix representation of the GFDM block showing that both
modulation and demodulation matrices follow the same structure. This structure is
determined with the decomposition of the GFDM matrix.
4.1.2.1
Modulation Matrix Model
One GFDM block can be represented in a vector x ∈ CN ×1 such that,
[x](n) =
m∈M on k∈K on
k
dk,m g[< n − mK >N ]ej2π K n .
(4.14)
In the frequency domain, x̃ = N-DFT{x} can be written as
m
x̃ (n) =
dk,m g̃[< n − kM >N ]e−j2π M n .
(4.15)
m∈M on k∈K on
Adding CP and CS is done by copying the last Lcp to the beginning and the first
Lcs samples to the end of x. Let D ∈ CK×M be the matrix representing the data
symbols, with [D](k,m) = dk,m and for (k, m) ∈
/ Kon × Mon , dk,m = 0. We define the
data vector d = vec {D}, namely, [d](k+mK) = [D](k,m) . In addition, the modulation
matrix A ∈ CN ×N is defined by
k
[A](n,k+mK) = g[< n − mK >N ]ej2π K n .
(4.16)
Thereby,
[x](n) =
M
−1 K−1
[A](n,k+mK) [d](k+mK) ,
(4.17)
m=0 k=0
and thus,
x = Ad.
(4.18)
Taking into account the active sets of subcarriers and subsymbols, we can define a
compact representation for the active resource set
Non = {n = k + mK, (k, m) ∈ Kon × Mon }
(4.19)
x = A(on) d(on) ,
(4.20)
as
where A(on) = [A](:,N on ) stands for the active modulation matrix and d(on) = [d](N on )
is the vector of the active data symbols.
4 Generalized Frequency Division Multiplexing …
4.1.2.2
99
Demodulation Matrix Model
An equalized received signal y[n], represented by the vector y ∈ CN ×1 , is to be
demodulated using a receive prototype filter γ [n]. The estimated data symbols can
be expressed as
k
d̂k,m = γ ∗ [−n] ⊛ y[p]e−j2π K n |n=mK
d̂
(k+mK)
=
N
−1
n=0
k
y[n]γ ∗ [< n − mK >N ]e−j2π K n .
(4.21)
Then,
d̂ = BH y,
where
(4.22)
H
k
B (k+mK,n) = γ ∗ [< n − mK >N ]e−j2π K n , so that
k
[B](n,k+mK) = γ [< n − mK >N ]ej2π K n .
(4.23)
Comparing with (4.16), we conclude that the demodulation matrix has the same
structure as the modulation matrix.
4.1.3 GFDM Matrix Decomposition
As the GFDM matrix is generated by circular shift of a prototype pulse shape in the
time and frequency domains, it has a well-defined structure and its properties can be
derived from the prototype pulse shape. To investigate the structure, first, we define
several auxiliary matrices to facilitate the derivation. For a vector a ∈ CPQ×1 , we
define the following matrices
T
(a)
= unvecQ×P {a} ,
VP,Q
(a)
(a)
ZP,Q
= FP VP,Q
=
(a)
ṼP,Q ,
(4.24)
(4.25)
where unvecQ×P {a} denotes the inverse of vectorization operation. Thus,
(a)
VP,Q
(p,q)
= [a](q+pQ) .
(4.26)
ij
FP is the P-point DFT matrix, where [FP ](i,j) = e−j2π P . The matrix V(a)
P,Q represents
(a)
the polyphase components generated by the sampling of a with factor Q, while ZP,Q
is known as discrete Zak transform (DZT) transform [2]. Figure 4.4 visualizes these
matrices by mean of an example.
100
A. Nimr et al.
Fig. 4.4 DZT example. a of size 6 × 1, P = 3, Q = 2
Recall the GFDM block Eq. (4.14) with the consideration of full allocation and
defining n = q + pK, q = 0, . . . , K − 1 and p = 0, . . . , M − 1, then
[x](q+pK) =
M
−1 K−1
m=0 k=0
k
dk,m g[< q + pK − mK >N ]ej2π K q .
(g)
Based on (4.26), [x](q+pK) = V(x)
M ,K
(p,q)
and VM ,K
(p,q)
= g[< q + pK >N ]. Fol-
lowing this, it is easy to show that g[< q + pK − mK >N ] = V(g)
M ,K
Putting all notations together we get
V(x)
M ,K
(p,q)
=
=
M
−1
(g)
VM ,K
m=0
M
−1
(g)
VM ,K
m=0
(<p−m>M ,q)
(<p−m>M ,q)
K−1
(4.27)
(<p−m>M ,q)
.
k
dk,m ej2π K q
k=0
(4.28)
T H
D FK (m,q) .
(g)
The second line represents circular convolution between the q-th column of VM ,K
and the q-th column of DT FH
K . By representing the circular decomposition in the
frequency domain we get,
V(x)
M ,K
(:,q)
=
1 H
F diag
M M
(g)
ṼM ,K
(:,q)
FM DT FH
K (:,q) .
(4.29)
Finally, by stacking the columns according to the q index and using (4.25), we get
V(x)
M ,K =
1 H (g)
FM ZM ,K ⊙ FM DT FH
.
K
M
(4.30)
4 Generalized Frequency Division Multiplexing …
101
Here, ⊙ denotes the element-wise multiplication operator. Thus, x = vec
Following the same approach on x̃ defined in (4.15), we get
V(x̃)
K,M
(q,p)
=
K−1
(g̃)
VK,M
k=0
(<q−k>K ,p)
M
−1
m
dk,m e−j2π M p .
(x)
VM ,K
T
.
(4.31)
m=0
Thereby,
V(x̃)
K,M =
Similarly, x̃ = vec
V(x̃)
K,M
1 H (g̃)
F ZK,M ⊙ [FK DFM ] .
K K
(4.32)
T
. Finally, using the vectorized output of (4.30) and
(4.32) with respect to (4.14), we can express the modulation matrix A as
1 T
P
UH L(g) UK,M PM ,K UH
M ,K ,
M M ,K K,M
1 1 H
(g̃)
T
=
F PM ,K UH
M ,K L UM ,K PM ,K UK,M PM ,K .
NK N
A=
(4.33)
(4.34)
Here, PP,Q ∈ ℜPQ×PQ is the permutation matrix that fulfills for any Q × P matrix X
vec XT = PP,Q vec {X} ,
(4.35)
UP,Q = IP ⊗ FQ ,
(4.36)
where ⊗ is the Kronecker product, and
(g)
L(g) = diag vec ZM ,K ,
(g̃)
L(g̃) = diag vec ZM ,K .
(4.37)
(4.38)
Either of these diagonal matrices can be used to analyze the GFDM performance
(g)
(g̃)
indicators and derive the demodulator matrix. In other words, ZM ,K or ZM ,K is the
key of GFDM design and analysis.
4.1.3.1
Demodulation Prototype Filter
While the GFDM-based demodulator matrix has the same structure as the modulation matrix, by using the decomposition (4.33) or (4.34), the prototype filter of the
demodulator can be easily computed from the inverse of DZT as
γ =
1
(γ )
vec FMH ZM ,K
M
T
.
(4.39)
102
A. Nimr et al.
First, we need to find B. Using the representation (4.33), the product BH A can be
calculated as
1
(γ )H (g)
L UK,M PM ,K UH
UK,M PTM ,K UH
K,M L
M ,K ,
M
= KUH L(γ )H L(g) U,
BH A =
where
1
U = √ UK,M PM ,K UH
M ,K ,
N
(4.40)
(4.41)
is a unitary matrix. Thus, the diagonal elements of KL(γ )H L(g) represent the eigenvalues of BH A. Obviously, the matched filter demodulator uses γ [n] = g[n]. On
the other hand, the zero forcing (ZF) is determined when KL(γ )H L(g) = IN . Conse(γ )
quently, γzf can be computed form the DZT inverse of ZMzf,K , defined by
(γ )
ZMzf,K
(m,k)
=
1
K
(g) ∗
ZM ,K
(4.42)
.
(k,m)
The minimum mean square error (MMSE) demodulator, assuming additive white
Gaussian noise (AWGN) of noise power PN and uncorrelated data symbols with
power PD , can be computed via the matrix
BMMSE = AAH +
=
PN
I
PD N
−1
A
1 T
(g) (g)H
PM ,K UH
+
K,M KL L
M
PN
PD
−1
(4.43)
L(g) UK,M PM ,K UH
M ,K .
Comparing with (4.33), we find that
(γ
)
ZMMMSE
,K
(m,k)
=
(g)
ZM ,K
(k,m)
(g)
K ZM ,K
2
+
(k,m)
PN
PD
−1
.
(4.44)
Then, γMMSE can be found form the DZT inverse using (4.39). It is worth noting that
in the case of non-full allocation, an MMSE or least squares (LS) receiver can be
derived based on the compact model x = A(on) d(on) . However, the obtained matrix
does not necessary have a GFDM structure and may complicate the implementation.
4.1.4 Performance Indicators
In order to evaluate the design with a given prototype pulse shape, we study three performance indicators of the modulation matrix A under the full allocation assumption.
4 Generalized Frequency Division Multiplexing …
103
(g̃)
All these indicators can be computed from ZK,M using the decomposition represented
in (4.34), which can be reformulated as
1
A = √ WH L(g̃) V,
K
(4.45)
with
WH = √
1
H
FH
N PM ,K UM ,K ,
NK
1
V = √ UM ,K PTM ,K UK,M PM ,K ,
N
(4.46)
(4.47)
are unitary matrices.
Conditional Number
2
, then {σk,m
=
|zk,m |2
} correspond
K
maxk,m {σk,m }
maxk,m {|zk,m |}
=
.
mink,m {σk,m }
mink,m {|zk,m |}
(4.48)
(g̃)
Defining the short-hand notation zk,m = ZK,M
(k,m)
to the singular values of A. The conditional number of A is given by
cond (A) =
When cond (A) = 1, i.e., |zk,m | = 1, ∀(k, m), then A is orthogonal, and when there
is at least (k0 , m0 ) such that zm0 ,k0 = 0, A becomes singular. The rank of A is reduced
(g̃)
by the number of zero elements in ZK,M . The conditional number is important
in all receiver processing steps that require the computation of the inverse of A.
Thus, a well-conditioned modulation matrix with smaller conditional number is preferred. Although we can always design GFDM with an orthogonal matrix, some
other requirements cannot be achieved. More details on that are introduced in Sect.
4.1.5.
Noise Enhancement Factor
Considering the received signal in AWGN channel,
y = Ad + w,
(4.49)
with E ddH = PD IN and E wwH = PN IN , the noise enhancement factor (NEF)
is defined by the ratio of the average signal-to-noise ratio (SNR) before and after
applying the ZF demodulator;
−1
trace E AddH AH
trace E ddH
ξ=
trace E vvH
trace E A−1 vvH A−1H
1
= 2 trace AAH trace A−1 A−1H
N
104
A. Nimr et al.
(g̃) 2 (g̃)−1 2
L L
F
⎛
⎞
⎞⎛ F
1
1 ⎝
⎠.
= 2
|zk,m |2 ⎠ ⎝
N
|zk,m |2
=
1
N2
k,m
(4.50)
k,m
Noting that after applying the ZF demodulator, we get A−1 y = d + A−1 w and
because the rows of A−1 are generated from circular shift of the prototype filter
γ zf , the noise enhancement is equal on each of the elements of d in the case of white
noise. Contrariwise, it depends on the subcarrier–subsymbol index for colored noise.
Self-interference Ratio
Because the columns of A are generated from the prototype pulse shape g[n] by
circular shift in time and frequency, then
and hence,
AH A (n,n) = g 2 ,
(4.51)
1
|zk,m |2 = N g 2 .
trace AH A =
K
(4.52)
k,m
After matched filter we get
AH y = g 2 d + AH A − g 2 IN d + AH w
= g 2 d + VH K1 L(g̃)H L(g̃) − g 2 IN Vd + AH w.
(4.53)
The signal-to-interference ratio (SIR) is defined by the ratio of the signal power to
the self-interface
SIR =
1
k,m
=
K
N g
|zk,m |2 − g
N
|zk,m |2
k,m
⎡
4
K g
2
−1
2 2
2
⎛
(4.54)
2
⎞2 ⎤−1
⎢ 1 ⎜ |zk,m |
⎟ ⎥
=⎣
− 1⎠ ⎦
⎝1
2
N
|z
|
k,m
N
k,m
.
k,m
The computed self-interference is averaged over all data symbols. Nevertheless,
when full allocation is considered and all data symbols
have the same power, the SIR
is identical for each data symbol. This is because [V](i,j) = √1N , ∀i, k ∈ {0, . . . ,
N − 1}.
4 Generalized Frequency Division Multiplexing …
105
The later discussion shows the dependency of the modulation on the prototype
pulse shape, where all indicator can be expressed in terms of its DZT. Although we
use DZT of g̃, the same results hold with respect to the DZT of g. In the next section,
we introduce a method to design the prototype filter g̃.
4.1.5 GFDM Pulse Shaping Filter Design
Starting from the DTFT of a preselected basis filter h[n] of practical interests, e.g.,
RC or root-raised cosine (RRC), which is denoted as H (ν). Here, ν is the normalized
frequency and thus the period of H (ν) is equal to 1. Then, we compute g̃[n] = H ( Nn ).
With such design, it has been shown in [3] that A becomes singular for even M , K and
(g̃)
= 0. The requirement
a real symmetric filter h[n]. This is caused by ZK,M
(K/2,M /2)
of odd M or K impedes an efficient implementation in terms of low-complexity
radix-2 FFT operations. In [4], we present a design approach that overcomes this
restriction for any basis filter h[n] fulfilling the following conditions,
1. h[n] is real-valued, i.e., H (ν) = H ∗ (1 − ν) = H ∗ (−ν).
2. H (ν) spans two subcarriers within each period, namely H (ν) = 0, ∀ν ∈ [ K1 , 21 ].
3. |H (ν)| is decreasing from 1 to 0 for ν ∈ [0, K1 ].
The idea is to introduce a fractional shift λ ∈ [0, 1] when sampling H (ν), as shown
in Fig. 4.5. Accordingly, the samples of g̃ are defined by
n+λ
⎧
⎫
0≤n<M −λ
⎨ H N ,
⎬
[g̃]n (λ) = H ∗ N −n−λ , N − M − λ < n ≤ N − 1 .
N
⎭
⎩
0,
otherwise
(4.55)
Moreover, g̃ can be reshaped as in (4.24) to
(g̃)
VK,M (λ)
(k,m)
m+λ
⎧
⎫
k=0 ⎬
⎨ H N ,
= H ∗ M −m−λ , k = K − 1 .
N
⎭
⎩
0,
elsewhere
(4.56)
With this design restriction, the same frequency taps can be used with different values
of K ≥ 2. Then, we compute the DZT from (4.25), so that
zk,m (λ) = H
(
m+λ
N
)
+ H∗
(
)
M − m − λ j2π k
e K.
N
(4.57)
Due to the symmetry of H (ν), we have
k
∗
j2π K
,
zk,m (1 − λ) = zk,M
−1−m (λ)e
|zk,m (1 − λ)|2 = |zk,M −1−m (λ)|2 .
(4.58)
106
A. Nimr et al.
Fig. 4.5 Filter design with sampling. Basis filter is Raised-Cosine with roll-off factor α = 0.5,
M = 8. The samples are shown for 0 ≤ n < M − λ
(g̃)
Fig. 4.6 Filter design with ISI free with and without matched filtering. Here, VK,M (λ) is shown
Thereby, all results regarding conditional number, NEF and SIR are symmetric
around λ = 0.5. Moreover, Eq. (4.57) shows that zk,m (1 + λ) = zk,m+1 (λ). Therefore, it suffices to study the range 0 ≤ λ ≤ 0.5.
We subsequently focus on particular design cases for K = 2x , x > 1 and different
values of M . Namely, design with the families of H (ν) that fulfill the inter-symbolinterference (ISI)-free criterion without or with matched filtering, Fig. 4.6.
4.1.5.1
ISI-Free Without Matched Filter
In this case, H (ν) additionally satisfies the condition
K−1
k=0
)
(
k
= 1.
H ν−
K
(4.59)
4 Generalized Frequency Division Multiplexing …
107
From the symmetry and limited band of H (ν), it follows that
H (ν) + H
∗
(
1
−ν
K
)
+
*
1
.
= 1, ∀ν ∈ 0,
K
(4.60)
+ H ∗ M −m−λ
= 1. From that, there exists a function f (ν) =
As a result, H m+λ
N
N
jφ(ν)
∗ 1
r(ν)e
with f (ν) = −f K − ν and
H (ν) =
+
*
1
1
.
(1 + f (ν)) , ∀ν ∈ 0,
2
K
(4.61)
Let us assume a real-valued f (ν), i.e., φ(ν) = 0 and f (ν) = r(ν). A complex-valued
f (ν) as in Xia filters is treated in the following section. Due to the constraint of the
decreasing amplitude of H (ν), r(ν) must be decreasing from 1 to −1 for ν ∈ [0, K1 ].
From (4.60) and (4.61) we get
|zAk,m (λ)|2 =
)
(
(1 + fm2 (λ) (1 − fm2 (λ))
k
,
+
cos 2π
2
2
K
(4.62)
where fm (λ) = f ( m+λ
− 1. The singular values are symmetric with
) = 2H m+λ
N
N
respect to k, and decreasing with k = 0, . . . , K2 . Therefore, |zA0,m (λ)|2 = 1 and
|zA K ,m (λ)|2 = fm2 (λ) are the maximum and minimum singular values with respect
2
to k, respectively. Therefore, |zAmax (λ)|2 = 1, because fm2 (λ) ≤ 1, and |zAmin (λ)|2 is
1
,
obtained from minm {fm (λ)2 }. Since f (ν) is decreasing and antisymmetric around 2K
1
1
1
2
f (ν) is decreasing ∀ν ∈ [0, 2K ] and increasing ∀ν ∈ [ 2K , K ]. As a result, when M
is even and 0 ≤ λ ≤ 0.5, |zAmin |2 is obtained at m = M /2, and when M is odd, it is
obtained at m = (M − 1)/2. Consequently,
2
|zAmin | (λ) = f
where S(λ) =
)
S(λ)
1
.
+
2K
2N
(4.63)
2λ, M is even
.
1 − 2λ, M is odd
(4.64)
2
(
From the increasing/decreasing intervals of f 2 (ν), |zAmin |2 (λ) increases with 0 ≤ λ ≤
0.5 for even M and decreases when M is odd. Hence, the condition number can be
expressed as
1
.
(4.65)
cond(AA )(λ) =
1
+ S(λ)
f 2K
2N
Similarly, cond(AA )(λ) is decreasing for even M and increasing for odd M . Hence,
the best condition of A is attained at λ = 0.5 for even M and λ = 0 for odd M .
108
4.1.5.2
A. Nimr et al.
ISI-Free After Matched Filtering
A filter H (ν) is ISI-free after matched filtering if
K−1
k=0
(
)2
H ν − k = 1.
K
(4.66)
By exploiting the symmetry and band limit, we get
(
+
)2
*
∗ 1
1
|H (ν)| + H
,
− ν = 1, ∀ν ∈ 0,
K
K
2
(4.67)
2 ∗ M −m−λ 2
+ H
= 1. Furthermore, there exists a realand hence, H m+λ
N
1
N
valued function f (ν) = −f K − ν , which is decreasing from 1 to −1 in the interval
ν ∈ [0, K1 ] with
+
*
1
1
.
(4.68)
|H (ν)|2 = (1 + f (ν)), ∀ν ∈ 0,
2
K
Adding an (arbitrary) phase φ(ν) yields the original H (ν) by
H (ν) = e
jφ(ν)
,
+
*
1
1
.
(1 + f (ν)), ∀ν ∈ 0,
2
K
(4.69)
Using (4.67) and (4.69),
H
H
∗
(
(
m+λ
N
M −m−λ
N
)
)
=e
jφa,m (λ)
=e
jφb,m (λ)
,
,
1
(1 + fm (λ)),
2
(4.70)
1
(1 − fm (λ)),
2
b
where φma (λ) = φ m+λ
, φm (λ) = −φ M −m−λ
, and fm (λ) = f m+λ
. As special
N
N
N
1
cases, we study the phase in the form φ(ν) = −φ K − ν + β π2 , β = 0, 1, 2, 3.
Then, ejφa,m (λ) = jβ ejφb,m (λ) . The case of no ISI with and without (MF), as the Xia
filters [5] provide, is obtained with f (ν) = cos(2φ(ν)) and β = 2 or, equivalently,
φ(ν) = 21 acos(f (ν). From (4.57), we get
σB2k,m (λ) = 1 +
-
k − β K4
1 − fm2 (λ) cos 2π
K
.
(4.71)
The maximum singular value with respect to k is located at kmax = β K4 and the
minimum one at kmin =< β + 2 >4 K4 . This requires that K is a multiple of 4 for
β = 1, 3.
4 Generalized Frequency Division Multiplexing …
109
1 − fm2 (λ),
|zBkmin ,m (λ)|2 = 1 − 1 − fm2 (λ).
|zBkmax ,m (λ)|2 = 1 +
(4.72)
Following the same arguments as in the previous subsection and based on the properties of f (ν), both |zBmin (λ)|2 and |zBmax (λ)|2 are obtained at m = M /2 for even M
and m = M 2−1 for odd M . Accordingly,
.
)
S(λ)
1
,
+
=1+ 1−f
2K
2N
.
(
)
1
S(λ)
2
2
+
,
|zBmin (λ)| = 1 − 1 − f 2
2K
2N
(4.73)
and the conditional number can then be written as
1
+ S(λ)
f 2K
2N
cond(AB )(λ) =
.
,
1
+ S(λ)
1 − 1 − f 2 2K
2N
(4.74)
|zB2 max (λ)|2
2
(
It can be seen that cond(AB )(λ) is decreasing for even M and increasing for odd M
with λ ∈ [0, 0.5]. When using the same function f (ν) in cases A and B, it is notable
that that |zBmax (λ)|2 ≥ 1 = |zAmax |2 and |zBmin |2 ≤ fm2max (λ) = |zAmin |2 , and hence,
cond(AA )(λ) ≤ cond(AB )(λ),
(4.75)
proving that the condition number is smaller when using an ISI-free filter, compared
to using its square root.
4.1.5.3
Numerical Example
In this section, we study the family of prototype filters with roll-off factor α, being
obtained with the generator function Fig. 4.7,
⎫
⎧
1,
0 ≤ ν ≤ 1−α
⎪
⎪
⎪
⎪
2K
⎪
⎪
⎬
⎨ 2K
1−α
1
1+α
a
f
[ν
−
]
,
<
ν
≤
.
(4.76)
f (ν) =
α
2K
2K
2K
⎪
⎪
⎪
1 ⎪
1+α
⎪
⎪
<ν≤
−1,
⎭
⎩
2K
K
Here, f a is a real-valued antisymmetric (f a (x) = f a (−x)), and decreasing from 1 to
−1 for x ∈ [−1, 1], which produces f (ν) = −f ( K1 − ν). Hence, f (ν) can be used to
construct pulse shapes that provide ISI free without matched filter used to generate AA
110
A. Nimr et al.
Fig. 4.7 Generator function of filters with roll-off factor α
(4.61) or ISI free with matched filter used to generate AB (4.69). From (4.74), (4.65),
and (4.76), we get for M α ≤ S(λ), cond(AA ) = cond(AB ) = 1. For S(λ) ≤ M α,
1
,
cond(AA )(λ) =
a S(λ)
f
αM
a S(λ)
f
αM
cond(AB )(λ) =
.
,
S(λ)
a2
1− 1−f
αM
(4.77)
The condition number is independent of K and, based on the properties of f a ,
increases with αM . As a particular example, RC and RRC use the function f a (x) =
− sin( π2 x). Replacing in (4.77) we get,
))
( (
π S(λ) −1
cond(ARC )(λ) = sin
,
2 αM
))
( (
π S(λ) −1
.
cond(ARRC )(λ) = tan
4 αM
(4.78)
Figure 4.8 illustrates the condition number of A for different sampling shift λ
and validates the closed-form expressions (4.78) numerically. As shown, λ = 0 is
optimal for odd M and λ = 21 for even M when K is also even. In addition, as proven
in (4.78), using RC yields a better conditioned A than RRC. Furthermore, numerically
obtained values for the NEF as shown in Fig. 4.9 behave similarly as the condition
number. This can be explained by the influence of the smaller singular value on
the noise enhancement. In both
number as well as the smallest
cases, the condition
π S(λ)
=
sin
. Considering the optimum λ,
singular value depends on f a S(λ)
αM
2 αM
Fig. 4.10 shows the NEF and SIR with different M . The proper choice of λ with
4 Generalized Frequency Division Multiplexing …
111
= 0.5, K = 64
RC-Sim
RC-Closed-form
RRC-Sim
RRC-Closed-form
Conditional number
14
12
10
8
M=9
M=9
6
M=8
M=9
4
M=9
M=8
2
0
0.2
0.4
0.6
0.8
1
Sampling offset
Fig. 4.8 Conditional number
= 0.5, K = 64
8
RC
RRC
Noise enhancment [dB]
7
6
5
4
3
2
M=9
M=9
1
0
M=9
M=9
M=8
M=8
0.2
0.4
0.6
0.8
1
Sampling offset
Fig. 4.9 NEF versus λ
respect to M preserves the trend of NEF which increases with M . On the other hand,
the SIR is independent of M when M is big enough. In fact, the SIR approaches
01
the interference value that can be directly obtained from SIR = 2 21 |H (ν)|2 d ν,
2K
which is independent of λ and K but depends on α. Finally, Fig. 4.11 depicts the
dependency of the NEF on K for the case of design with ISI free without matched
filter. Although the condition number is independent of the even values of K, the
112
A. Nimr et al.
= 0.5, K = 64
NEF
NEF/SIR [dB]
0
RC
RRC
-5
SIR
-10
-15
5
10
15
20
25
30
35
40
M
Fig. 4.10 NEF and SIR for optimal λ
M = 16
8
RC( = 0.9)
RC( = 0.6)
RC( = 0.3)
NEF [dB]
6
4
2
0
5
10
15
20
K
Fig. 4.11 NEF versus even values of K
NEF is higher for smaller values, especially when α is small, and converges to fixed
value when K gets bigger. As a result, K should be bigger than the product of αM .
4 Generalized Frequency Division Multiplexing …
113
4.1.6 Multicarrier Waveforms Generator
In this section, we provide a general representation of linear multicarrier modulation
techniques inspired from the GFDM model. Then, we show how different stateof-the-art waveforms can be generated from the GFDM block modulator, block
multiplexing, windowing, and filtering. This enables the development of a universal
waveform generator.
4.1.6.1
General Linear Modulation
In general multicarrier modulation, a stream of data symbols can be split into timefrequency substreams {dk,m,i }, with k is the index in the frequency domain denoted as
subcarrier, m the index in the time domain as subsymbol, and i stands for the block
t
[n] with finite
index. Each stream is modulated by a transmitter pulse shape gk,m
/ {0, . . . , Lt − 1}, gk,m [n] = 0. The discrete
length Lt , specifically, ∀(k, m) and ∀n ∈
transmitted signal can be written as
xt [n] =
=
∞
i=0 k∈K on m∈M on
∞
i=0
xit [n
t
dk,m,i gk,m
[n − iLs ]
(4.79)
− iLs ],
where Ls is the block spacing, Kon and Mon are the sets of active subcarriers and
subsymbols, respectively, and xit [n] is the i-th multicarrier block, which is given by
xit [n] =
k∈K on m∈M on
t
dk,m,i gk,m
[n].
(4.80)
Accordingly, xit [n] has a length of Lt samples. The number of available resources
per block is denoted as N = MK. The difference between the block length and the
block spacing Lo = Lt − Ls determines the overlapping between successive blocks
as illustrated in Fig. 4.12. When Lo > 0, the blocks overlap, which means that the
last Lo samples of the previous block are added to the first Lo samples of the current
block prior to transmission. On the other hand, for Lo ≤ 0, there is a guard interval of
Lo zero padding (ZP) samples between successive blocks. Therefore, the multicarrier
waveform can be defined by knowing the set of pulse shapes {gk,m }, the set of active
resources Kon × Mon , the resource dimensions K, M and the overlapping length Lo .
114
A. Nimr et al.
Fig. 4.12 Different cases of block multiplexing
4.1.6.2
Generic Waveforms Generator
t
[n] can be generated from a prototype
In the common modulation techniques, gk,m
pulse shape with shift in time and frequency. Moreover, CP and CS can be added afterward. In addition, windowing and subband filtering may be applied. Some waveforms
involve more than one prototype pulse shape, which can be seen as superposition
of different waveforms. As an example, we list the overall procedure for generating
waveforms employing one prototype pulse shape g[n] of length N samples.
• Shifting in time and frequency
k
gk,m [n] = g[< n − mK >N ]ej2π K n , n = 0, . . . , N − 1.
(4.81)
• CP and CS insertion
cp
gk,m [n] = gm,k [< n − mK − Lcp >N ], n = 0, . . . , N + Lcp + Lcs − 1.
(4.82)
• Time domain windowing using a window function w[n] of length N + Lcp + Lcs
samples
cp,w
cp
gk,m [n] = w[n] · gm,k [n], n = 0, . . . , N + Lcp + Lcs − 1.
(4.83)
• Filtering using a filter f [n] with Lf samples
cp,w,f
cp,w
gk,m [n] = f [n] ∗ gk,m [n], n = 0, . . . , N + Lcp + Lcs + Lf − 2.
(4.84)
t
t
gk,m
[n] results from one or more of that steps. In the most complicated case gk,m
[n] =
cp,w,f
gk,m [n]. Therefore, (4.80) becomes
4 Generalized Frequency Division Multiplexing …
115
Fig. 4.13 Multicarrier waveform generator stages
⎛
xit [n] = f [n] ∗ ⎝w[n] ·
where
k∈K on m∈M on
⎞
dk,m,i gm,k [< n − mK − Lcp >N ]⎠
(4.85)
= f [n] ∗ w[n] · xi [< n − Lcp >N ] ,
xi [n] =
=
k∈K on m∈M on
k∈K on m∈M on
dk,m,i gk,m [n], n = 0, . . . , N − 1
k
dk,m,i g[< n − mK >N ]ej2π K n .
(4.86)
This means that the transmitted block can be obtained starting from a core block of
length N samples, which can be generated using the GFDM modulator. Then, the
CP and CS can be added to the core block followed by windowing and filtering.
Finally, the blocks are multiplexed in the time domain to generate the waveform.
This procedure is depicted in Fig. 4.13.
4.1.6.3
Example of the State-of-the-Art Waveforms
As discussed, the core block is the essential part of the waveform, and this can be
generated with the GFDM modulator under proper setting of the related parameters.
OFDM Variants
Obviously, orthogonal frequency division multiplexing (OFDM) is a special case
of GFDM when M = 1 and g[n] is a rectangular pulse. In the simplest form of
OFDM, only a CP is added. Thus Lt = K + Lcp and no overlapping, i.e., Ls = Lt .
In the windowed version, a window is applied after adding the CP and CS. Then,
Lt = K + Lcp + Lcs , with Lcp > Lcs . In order to reduce the overhead, the blocks are
overlapped with Lo = Lcs , so that Ls = K + Lcp , Fig. 4.12. In the filtered variants,
either the ones based on subband filtering for multiple users or the others that apply
filtering to the whole signal, a filter is applied on the generated CP symbols with
the corresponding active subcarriers. Hence, Lt = K + Lcp + Lf − 1, Ls = K + Lcp ,
and due to the filtering Lo = Lf − 1.
116
A. Nimr et al.
DFT-spread OFDM
In this modulation, [6] a set of M data symbols are transferred into the frequency
domain using FM and then allocated to the subcarrier set of N subcarriers using
1 H
F . Thus, the modulation matrix of this waveform is given by
N N
A=
1 H
F UK,M .
N N
(4.87)
with the GFDM modulaComparing with (4.34), this waveform can be
generated
tor where the input vector is given by vec DT and using a pulse shape g̃ with
(g̃)
ZK,M
(k,m)
= 1, ∀(k, m). This actually corresponds to the Dirichlet pulse given by
g[n] = ejπn
M −1
N
sin(π Kn )
.
sin(π Nn )
(4.88)
Filtered Multitone (FMT)
This waveform does not originally consider subsymbols representation [7]. However,
it is easy to transform it to fit in this framework. The prototype pulse gFMT [n] is
assumed to have a length of KMo , where Mo is denoted as the overlapping factor.
t
xFMT
[n] =
=
=
∞ K−1
p=0 k=0
k
dk,p gFMT [n − pK]ej2π K n
∞ K−1
d −1
M
i=0 k=0 m=0
∞
i=0
k
dk,m+iMd gFMT [n − (m + iMd )K]ej2π K n
(4.89)
xi [n − iMd K],
where
xi [n] =
K−1
d −1
M
k=0 m=0
k
dk,m+iMd gFMT [n − mK]ej2π K n .
(4.90)
With respect to GFDM notations, we define the number of subsymbols
M = Md + Mo − 1, where Md ≥ 1 is the number of active subsymbols, the data
symbol dk,m,i = dk,m+iMd and the pulse shape g[n] as
g[n] = gFMT [< n +
KMo
2
>N ], n = 0, . . . , N − 1.
(4.91)
Therefore, we get the GFDM block
xi [n] =
K−1
o /2−1
Md +M
k=0
m=Mo /2
k
dk,m,i g[< n − mK >N ]ej2π K n .
(4.92)
4 Generalized Frequency Division Multiplexing …
117
Fig. 4.14 FMT pulse shape to GFDM
Noting that Mon = {Mo /2, . . . , Md + Mo /2 − 1}, in other words, there are Mo − 1
subsymbols turned off, Fig. 4.14. However, while Ls = Md K and Lt = MK, there is
an overlapping of Lo = (Mo − 1)K samples, which compensates the overhead results
from the non-active subsymbols.
4.1.6.4
Further Degrees of Freedom
Recalling the compact equation
1 H (g)
F ZM ,K ⊙ FM DT FH
.
K
M M
The GFDM core block is generated as x = vec XT and then further processing
steps take place. Inspired from that, it is possible to generate three types of other
signals;
X = V(x)
M ,K =
1. By constructing the core block according to the columns, i.e., x1 = vec {X}. This
can be seen as a block of stacked K precoded OFDM symbols with M subcarriers.
2. Adding a CP to each column of X to get X(cp) , then x2 = vec X(cp) . This signal
can be seen as transmitting precoded OFDM symbols successively.
(cp)
(cp)
3. Adding a CP to each column of XT to get X3 , then x3 = vec X3 . This signal
can be seen as inserting CPs within the GFDM symbol.
(g)
Moreover, the matrix
ZM ,K can be populated with unit amplitude and variant phase
(g̃)
= 1, so that the overall modulation matrix is orthogelements, namely, ZK,M
(k,m)
n2
(g̃)
onal. For example, with M = 1 and ZK,M
= ej2πc N = L(c) (k,k) , we get
(k,1)
118
A. Nimr et al.
x = L(c) FH
K d,
(4.93)
which represents a chirp-based waveform [8].
In conclusion, the GFDM-inspired waveform generator is a very powerful tool for
unified implementation of various standard waveforms, which makes it appropriate
for mixed numerology approach.
4.1.7 Channel Estimation for GFDM Detection
Consider a multiple-input multiple-output (MIMO) transceiver with Nt transmit and
Nr receive antennas where the complex-valued data symbols are spatially multiplexed. Combining the MIMO system with a GFDM-based modulation, we encounter
a three-dimensional interference situation due to ISI, ICI as well as inter-antennainterference (IAI). Such an interference-limited scenario challenges the MIMOGFDM receiver design in different stages. A critical functional unit at the receiver
side is the channel estimation. Due to broad subcarrier spacing in GFDM, the individual subcarriers become frequency selective and correct detection of the data symbols
requires a reliable estimate of the wireless channel transfer function.
In the following, we focus on pilot-aided channel estimation techniques, where
some reference signals (also referred as pilots), which are known to both transmitter
and receiver, are multiplexed with the data symbols within the same time-frequency
resource block. Given the knowledge of the pilots at the receiver, the frequency
selective channel transfer function can be estimated and utilized for coherent detection of the data symbols. Here, we also assume that the interference autocorrelation
matrix of the MIMO-GFDM transceiver is known at the receiver side, and based on
such knowledge we derive the two well-known estimation techniques, namely least
squares (LS) and linear minimum mean squared error (LMMSE). Later on, we also
discuss an alternative approach for the pilot insertion of a MIMO-GFDM system in
order to achieve interference-free channel estimation performance.
4.1.7.1
MIMO Wireless Channel
We assume an urban scenario with multipath Rayleigh fading MIMO channels. Further, we assume that the individual channels between each antenna pairs are independent and they are block fading, namely the channel impulse response (CIR) does
not vary significantly within one GFDM symbol duration and therefore, it can be
considered as constant. Hence, we model the CIR between the it -th and ir -th antennas
as a linear finite-impulse-response filter given by
hit ir [n] =
L−1
ℓ=0
hℓ,it ir δ[n − ñℓ ],
(4.94)
4 Generalized Frequency Division Multiplexing …
119
where ñℓ is the discrete-time-delay of the ℓ-th path, hℓ,it ir is the complex-valued gain
of the ℓ-th path, and it is independent and identically distributed (i.i.d.) zero mean
Gaussian process parameterized by the power-delay profile (PDP) E |hℓ,it ir |2 =
Pit ir [ℓ]. Collecting all the L paths in a vector, we have
hit ir =
diag pit ir qit ir ,
(4.95)
where, qit ir ∼ C(0, IL ) and pit ir ∈ RL is the vector of normalized PDP.
Thanks to the utilization of CP in GFDM, the individual channels between each
antenna pair are diagonal in the frequency domain. Thus, the receive (Rx) signal in
DFT domain at the Rx antenna ir is characterized by the linear expression
ỹir =
Nt
(x̃)
(D(x̃)
p,it + Dd ,it )FN ,L hit ir + w̃ir ,
(4.96)
it =1
where w̃ir is the frequency counterpart of AWGN samples with variance σw2 , FN ,L ⊆
FN is the N -DFT matrix with the L columns associated to the discrete path delays ñℓ .
(x̃)
Moreover, Ds,i
= diag(x̃s,it ), s ∈ {p, d } and x̃s,it = FN xs,it , is the diagonal frequency
t
domain transmitted signal associated with either pilots or data. For the time domain
signal, we have xs,it = Ads,it where dp,it and dd ,it are the N -dimensional vectors of
the pilots and data symbols at data symbols at transmit (Tx) antenna it , respectively.
In addition, the multiplexing of the pilots and data symbols satisfies dp,it ◦ dd ,it = 0.
If the number of pilot subcarriers is smaller than K, i.e., the spacing between the
pilot subcarriers Δk > 1, only a subset of the observations in the frequency domain
with Np = ⌊N /Δk⌋ samples, that contain the knowledge of pilots, are of interest.
Therefore, the observed signal ỹ¯ ir of size Np × 1 at the pilot-bearing frequency bins1
follows
Nt
¯
¯
x̃)
¯i ,
ỹ¯ ir =
(D(p,i
+ D(dx̃),it )F′N ,L hit ir + w̃
(4.97)
r
t
it =1
¯
where D(s,ix̃)t = diag(x̃¯ s,it ), s ∈ {p, d }, x̃¯ s,it = F′N xs,it , F′N ,L ⊆ FN ,L of size Np × N , and
F′N ⊆ FN of size Np × N includes the rows of FN ,L and respectively FN in which
their inner product with the time domain signal xp,it is nonzero.
We rearrange the expression (4.97) into a matrix form as
Y = (Xp + Xd )F′Nt H + W,
(4.98)
herein, each of the above parameters is defined as F′Nt = INt ⊗ F′N ,L , Y = (ỹ¯ 1 , . . . ,
¯
¯
¯ N ) ∈ CNp ×Nr , Xs = (D(x̃)
¯ 1 , . . . , w̃
, . . . , D(x̃) ) ∈ CNp ×Np Nt
ỹ¯ N ) ∈ CNp ×Nr , W = (w̃
r
1 We
r
s,1
s,Nt
refer to each inner product of the rows of the DFT matrix with the time domain signal as a
frequency bin.
120
A. Nimr et al.
Fig. 4.15 Overview of the matrix structure for a 2 × 2 MIMO channel
with s ∈ {p, d }. The matrix H ∈ CLNt ×Nr is the Nt × Nr block-matrix of channel
impulse responses hit ir . An example of the matrix structures for a 2 × 2 MIMO
channel is depicted in Fig. 4.15. Note that due to the structure of H, its vectorization h = vec {H} consists of Nt Nr independent column vectors of channel impulse
responses. Thus, by assuming Rayleigh fading channels with
no spatial correlation,
the covariance matrix of all channel impulse responses E hhH becomes diagonal.
Resorting to the matrix property vec {ABC} = (CT ⊗ A)vec {B}, the associated
vectorization of the observed matrix Y yields the expression
y = vec {Y} = X̃p h + X̃d h + w,
(4.99)
where X̃s = (INr ⊗ Xs F′Nt ) ∈ CNp Nr ×LNt Nr and w = vec {W}.
4.1.7.2
Least Squares Estimation
The LS estimator of H minimizes Y − Xp H
2
with respect to H. This yields [9]
−1
ĤLS = QLS Y = (Xp F′Nt )H (Xp F′Nt )
(Xp F′Nt )H Y,
(4.100)
which subjects to estimation errors due to the enhancement of noise as well as
interference. Subsequently, the mean squared error (MSE) of the LS estimation is
given by
MSELS = E
=
H − ĤLS
2
(4.101)
σw2
1
trace INr ⊗ (QH
trace INr ⊗ (QH
Q
)
R
+
LS
Ψ
Ψ
LS
LS QLS )
Δk
Δk
where we calculate the interference covariance matrix as
H
RΨ Ψ = E vec Xd F′Nt H vec Xd F′Nt H
= EXd (INr ⊗ Xd F′Nt )E hhH |Xd (INr ⊗ Xd F′Nt )H
= EXd (INr ⊗ Xd F′Nt )Rhh (INr ⊗ Xd F′Nt )H .
(4.102)
4 Generalized Frequency Division Multiplexing …
121
In the above expression, Rhh is diagonal because independent Rayleigh fading has
been considered for the MIMO channels. Thus, the covariance matrix is given by
Rhh = diag [pT11 , . . . , pTNt 1 , . . . , pT(Nt −1)Nr , pTNt Nr ]T .
(4.103)
RΨ Ψ = RΨ Ψ (1) ⊕ · · · ⊕ RΨ Ψ (ir ) ⊕ · · · ⊕ RΨ Ψ (Nr ) ,
(4.104)
Further, due to the block-diagonal structure of (INr ⊗ Xd F′Nt ) in (4.102), the resulting
interference covariance matrix RΨ Ψ follows the form
where ⊕ is the direct sum of matrices. Moreover, we calculate the interference
covariance matrix RΨ Ψ (ir ) at Rx antenna ir as
RΨ Ψ (ir ) =
Nt
RΨ Ψ (it ,ir ) ,
(4.105)
it =1
wherein for each antenna pair it -ir , we have
′H H
RΨ Ψ (it ,ir ) = EXd ,it Xd ,it F′N ,L Eh hit ir hH
it ir FN ,L Xd ,it
= Rhhf (it ir ) ◦ RXd Xd ,it ,
(4.106)
where Rhhf (it ir ) = F′N ,L diag pit ir F′H
N ,L is the channel covariance in the frequency
domain. Furthermore, the covariance RXd Xd ,it is given by
RXd Xd ,it = F′N ARdd ,it AH F′H
N ,
(4.107)
herein, Rdd ,it = E dd ,it dH
d ,it is the covariance matrix of the data symbols transmitted
on antenna it . Assuming the data symbols are i.i.d. with unit variance, Rdd ,it becomes
diagonal with zero entries at the pilot positions.
4.1.7.3
Linear Minimum Mean Squared Error Estimation
The LMMSE estimation calculates the coefficients of a linear filter aiming at minimizing the MSE. In accordance with (4.98) and the corresponding vectorization in
(4.99), we have
H
H
ĥLMMSE = Rhh X̃p (X̃p Rhh X̃p + RΨ Ψ + σw2 INp Nr )−1 y.
23
4
1 23 4 1
Rhy
(4.108)
Ryy
Here, ĥLMMSE ∈ CLNt Nr is a column vector that contains Nt Nr individual columns of
size L associated to the LMMSE estimate of each channel impulse response.
The resulting MSE performance of the LMMSE estimation follows
122
with
A. Nimr et al.
MSELMMSE = trace RHH − RĤĤ ,
RHH = (INr ⊗ F′Nt )Rhh (INr ⊗ F′Nt )H ,
RĤĤ = (INr ⊗
4.1.7.4
H
F′Nt )Rhy R−1
yy R hy (INr
⊗
(4.109)
(4.110)
F′Nt )H .
(4.111)
Interference-Free Pilot Insertion
In this section, we slightly modify the GFDM modulation at the pilot subcarriers2 in
order to insert orthogonal pilots. The low complexity frequency domain processing
of the GFDM modulation can be written as in [10].
x = FH
N
K−1
P(k) Gδ T(δ) FM d,
(4.112)
k=0
where T(δ) is δ-fold repetition matrix which concatenates δ identity matrices IM of
size M , i.e., T(δ) = (IM IM . . .)T . The value of δ is based on the number of nonzero
values in the filter frequency response, e.g., if a filter spans over two subcarriers δ
is typically selected
as δ =2. In (4.112), due to the circular filtering, the subcarrier
(δ)
filter G(δ) = diag FδM g(δ) is diagonal
infrequency domain. The circulant filter g
is the down-sampled version of g = g[n] n=0,...,N −1 by factor K/δ. The permutation
matrix P(k) shifts the DC signals to their corresponding subcarriers (i.e., k) and is
given by
*
+T
I
0
0
P(k) = Cn′ M δ/2 M δ/2 M δ/2×(N −δM ) ,
(4.113)
IM δ/2 0M δ/2 0M δ/2×(N −δM )
where n′ = kM − M δ/2. The circulant matrix Cn′ follows
Cn′ = circ([0Tn′
T
mod N , 1, 0(N −n′ −1)modN ]),
(4.114)
here, circ(·) returns a circulant matrix associated to its input row vector.
Note that in (4.112) the M -point DFT matrix FM can be considered as a special
form of precoding. Hence, by slightly modifying such precoder we can reserve some
frequency bins specifically for the pilots without any influence from the data symbols
[11]. Thus, at the pilot subcarriers k ∈ Kp , we modify the expression (4.112) by
replacing FM with CP
x K p = FH
N
2 Pilot
P(k) G(δ) T(δ) CP dk ,
(4.115)
k∈K p
subcarrier is referred to a subcarrier in which Mp pilots are multiplexed with Md data subsymbols while M = Mp + Md .
4 Generalized Frequency Division Multiplexing …
123
Null bin
Data
Cy
clic
Cy
clic
Data
Pilot bin
Pre
fix
f
Pre
fix
f
Pilot bin
Null bin
t
t
(a) Antenna 1
(b) Antenna 2
Fig. 4.16 Pilots and data subsymbols in time-frequency resources
1
0.8
Magnitude
Fig. 4.17 DFT domain of
the signal for M = 15,
K = 4, P′ = IM
0.6
0.4
Pilot on Antenna 1
Pilot on Antenna 2
Pilot subcarrier
Data Subcarrier
0.2
0
0
1
2
3
normalized frequency f /M
4
where CP = P′ (λINt ⊕ F(M −Nt ) ). Here, λ is a scaling factor that normalizes the pilots
energy to one. P′ can be any permutation matrix of compatible size which allocates
the pilots to any frequency bin within the pilot subcarriers. For instance, the choice
P′ = IM allocates the pilots on the center frequency bins of the pilot subcarriers.
Furthermore, INt in CP ensures that the first Nt subsymbols of the pilot subcarriers (i.e.,
dk [0], dk [1], . . . , dk [Nt − 1] for k ∈ Kp which are filled with pilots) are processed
directly in the frequency domain being orthogonal to the rest of subsymbols (i.e.,
dk [Nt ], . . . , dk [M − 1] for k ∈ Kp ). Nevertheless, such orthogonality holds if and
only if the pilots are located at the frequency bins where no inter-carrier interference
is present. Moreover, reserving each orthogonal subsymbol for a specific Tx antenna,
the Nt × Nr MIMO channel can be processed in terms of Nt Nr single-input singleoutput (SISO) channels for channel estimation. The approach can be considered as
a variation of cell-specific reference signal mapping in Long-Term Evolution (LTE)
[12].
Figure 4.16 shows an example how the pilot subsymbols in the GFDM data block
are mapped into the time-frequency grid of the resources for a 2 × 2 MIMO channel.
Here, two frequency bins of the pilot subcarriers are reserved only for the pilots
while at each Tx antenna only one pilot is being transmitted. Thus, the pilot is
being transmitted during the whole GFDM symbol, while also the energies of the
data subsymbols are no longer concentrated at equally spaced M peaks. Figure 4.17
124
A. Nimr et al.
shows an example of the signal filtering in frequency domain, where the pilots at
different antennas are orthogonal to one another as well as to the data bins within
the pilot subcarrier.
4.1.7.5
Simulation Results
In this section, we verify the validity of the closed-form expressions of the channel
estimation MSE by simulation and numerical results while we also compare them
with the channel estimation performance of OFDM. Later on, we evaluate the performance of MIMO-GFDM with an interference-free pilot design, where we adopt
2 × 2 MIMO block fading multipath channel with Rayleigh distribution. Since the
interference-free pilot insertion (IFPI) in the GFDM block might modify the original
signal characteristics, we analyze the Tx signal in terms of peak-to-average power
ratio (PAPR) and OOB emission via Monte Carlo simulations.
Consider a sequence of 16-QAM symbols with energy per symbol Es being
transmitted through a multipath MIMO channel with noise energy N0 and with
Nt = {2, 3, 4} and Nr = {2, 3, . . . , 8} antennas. A single block of GFDM contains
M = 7 subsymbols, and it is filtered by an RC pulse with roll-off factor α = 0.3. For
comparison purpose, we configure OFDM to have K ′ = MK subcarriers. Assuming
both signals have an identical bandwidth, the subcarrier spacing of GFDM becomes
M times broader with respect to the OFDM one and therefore, each GFDM subcarrier
consists of M bins while OFDM has a single frequency bin per subcarrier.
Figure 4.18 illustrates the MSE evaluations for theoretical analysis as well as
simulation results. Here, each channel is chosen to have L = 9 taps with exponential
PDP. As expected from the theoretical expressions, the channel estimation for GFDM
contains an error floor due to the interference from data symbols while for OFDM,
the MSE decreases linearly with the increase of the SNR. Moreover, comparing the
100
OFDM, 2 × 2, LS, Sim.
OFDM, 2 × 2, LS, Theo.
GFDM, 2 × 2, LS, Sim.
GFDM, 2 × 2, LS, Theo.
GFDM, 2 × 8, LS, Sim.
GFDM, 2 × 8, LS, Theo.
GFDM, 4 × 4, LS, Sim.
GFDM, 4 × 4, LS, Theo.
GFDM, 4 × 8, LS, Sim.
GFDM, 4 × 8, LS, Theo.
MSE
10−1
10−2
10−3
10−4
0
5
10
15
20
25
30
OFDM, 2 × 2, LMMSE, Sim.
OFDM, 2 × 2, LMMSE, Theo.
GFDM, 2 × 2, LMMSE, Sim.
GFDM, 2 × 2, LMMSE, Theo.
GFDM, 2 × 8, LMMSE, Sim.
GFDM, 2 × 8, LMMSE, Theo.
GFDM, 4 × 4, LMMSE, Sim.
GFDM, 4 × 4, LMMSE, Theo.
GFDM, 4 × 8, LMMSE, Sim.
GFDM, 4 × 8, LMMSE, Theo.
35
Es /N0 (dB)
Fig. 4.18 MSE results of channel estimation versus Es /N0 for simulation and theoretical calculations in Rayleigh fading MIMO channel with a pilot spacing of Δk = 2 and K = 128 subcarriers
4 Generalized Frequency Division Multiplexing …
125
10 0
−20
10 −2
−40
PSD [dB]
CCDF
0
10 −1
10 −3
OFDM
IFPI GFDM
Basic GFDM
10 −4
10 −5
4
6
−60
OFDM
IFPI GFDM
Basic GFDM
W-IFPI GFDM
W-GFDM
−80
−100
8
10
12
PAPR [dB]
(a) Signal PAPR for M = 21 , K = 4 vs. its
CCDF
−120
−40
−20
0
20
40
subcarrier index f /F
(b) OOB for K = 96, M = 15 and the total
number of frequency samples F.
Fig. 4.19 Transmitted signal characteristics
GFDM channel estimation results for various number of Tx and Rx antennas, we
notice that the error does not directly depend on the number of receive antennas, e.g.,
the MSE curves for 2 × 2 versus 2 × 8 antennas are overlapped (as well as 4 × 4 vs.
4 × 8). This is due to the fact that, by linearly increasing the number of Rx antennas
we increase the number of observations while the number of estimation parameters
(i.e., channel taps) also increases linearly, e.g., doubling the number of Rx antennas,
we also double the number of channel taps while their ratio remains identical. As a
consequence, no analytical difference should be expected in this case. On the other
hand, as we increase the number of Tx-Rx antennas, the estimation performance
for both LS and LMMSE estimators degrades, because by linearly increasing the
number of Tx-Rx antennas, the number of channel taps increases quadratically and
thus, the estimation performance degrades.
The PAPR of the IFPI GFDM is compared to the original GFDM beside OFDM
in Fig. 4.19a. One can see that due to orthogonal pilot insertion, the PAPR of IFPI
GFDM increases with respect to the basic GFDM. However, it still has more than
one dB difference with the PAPR of an OFDM signal. On the other hand, comparing the power spectral densities of the signals in Fig. 4.19b, we observe that, despite
IFPI GFDM has slightly larger OOB compared to the original GFDM signal, the windowed case achieves almost the same OOB radiation as in original windowed GFDM
(W-GFDM). The window function is configured in form of an RC window with a
ramp length of a quarter subsymbol. For further details regarding the windowing
process, we refer the interested readers to [1].
The coded performances of the three receiver types are provided in Fig. 4.20.
Here, a GFDM block has M = 7 subsymbols and K = 96 subcarriers. The received
signal constellations are detected via GFDM zero forcing demodulation and MMSE
frequency domain channel equalization. The channel codes are chosen as parallel
concatenated convolutional codes (PCCCs) (1, 15/13), and they provide a gain in
spectral efficiency leading to energy per bit Eb /N0 = Es /N0 − 10 log10 (μr) where
μ and r denote the modulation order and the code rate, respectively. The detected data
126
A. Nimr et al.
IFPI GFDM, Genie-aided
OFDM, Genie-aided
Basic GFDM, Genie-aided
IFPI GFDM, LS
OFDM, LS
Basic GFDM, LS
IFPI GFDM, LMMSE
OFDM, LMMSE
Basic GFDM, LMMSE
100
BER
10−1
10−2
10−3
10−4
10−5
−2
0
2
4
Eb /N0 [dB]
(a) 1/3 code-rate
6
85
10
15
20
25
30
Eb /N0 [dB]
(b) 5/6 code-rate
Fig. 4.20 Bit error rate performance with 5% pilots overhead over 2 × 2 MIMO channel
(M = 7, K = 96)
symbols are transferred into maximum likelihood (ML) symbol log-likelihoods and
they are inserted into the soft demapper with 8 turbo decoder iterations. In Fig. 4.20,
the channel estimation performance for IFPI GFDM and OFDM is identical, although
the basic pilot insertion for GFDM has an error floor at high SNR regions as was
shown before.
Note that in Fig. 4.20, employing a robust code rate of 1/3, OFDM, IFPI GFDM,
and basic GFDM receivers obtain almost similar BER, though, basic GFDM with
LS estimation has 2–3 dB worse BER performance than the rest of receivers with
imperfect channel knowledge. Here, due to around 1 dB gap of the genie-aided
receivers of OFDM and IFPI GFDM, the latter receiver stays around 0.5 dB behind
OFDM when the channel is estimated through pilot transmission. On the other hand,
Fig. 4.20b shows that the basic GFDM channel estimation with non-orthogonal pilots
has appreciable performance loss for a high code rate of 5/6 which is due to its
large error floor in the channel estimation. Comparing OFDM and IFPI GFDM, we
observe that the performance loss in GFDM which is a non-orthogonal waveform is
not significant compared to OFDM. Furthermore, in Fig. 4.20b the BER for LS and
LMMSE estimations in OFDM as well as IFPI GFDM are identical due to identical
channel estimation performances at high SNR regions.
In short, if the SNR region is low and a robust code rate is utilized, with normal
pilot insertion in GFDM the performance has slight degradation in comparison with
OFDM. Although for higher SNR regions when faster code rates are in favor, it is
necessary to insert interference-free pilots in order to fully exploit the capacity of the
GFDM system while the advantages of its signal characteristics are also preserved.
4 Generalized Frequency Division Multiplexing …
127
4.1.8 Transmission Diversity for GFDM
Robustness against time-variant and frequency-selective channel is an important
feature for the fifth-generation (5G) networks. Transmit diversity can be exploited to
improve coverage, reduce the need for retransmissions, and improve the reliability
of the system. Although the space-time coding (STC) as proposed in [13] can be
applied to GFDM with the help of wide linear equalization, the complexity of the
receiver can hinder its application in some 5G scenarios, such as Internet of Things
(IoT) and machine-type communication (MTC).
A simple and elegant solution for this issue is based on employing the time-reversal
space-time coding (TR-STC) [14] applied to GFDM in order to achieve maximum
diversity gain and low implementation complexity without any performance loss. In
this case, two antennas are used to transmit two subsequent GFDM blocks xi [n] and
xi+1 [n], building the STC codeword as
Block i
Block i + 1
Antenna 1
Antenna 2
∗
xi [n] −xi+1
[− < n >N ] .
xi+1 [n]
xi∗ [− < n >N ]
(4.116)
After CP removal, the received signals at the j-th receiving antenna, in the frequency
domain, for the i-th and (i + 1)-th time instants are given by
ỹi,ir = D(irh̃),1 x̃i − Di(rh̃),2 x̃∗i+1 + w̃i,ir
ỹi+1,ir = D(irh̃),1 x̃i+1 + D(irh̃),2 x̃∗i + w̃i+1,ir ,
(4.117)
where D(irh̃),it = diag(h̃ir ,it ) with h̃ir ,it = FN hir ,it , x̃i = FN xi and w̃i,ir is the noise vector
on the i-th time instant and ir -th receive antenna. Assuming that Nr receiving antennas
are employed by the receiver, the received signals can be combined in the frequency
domain as
Nr
−1 (h̃) ∗
Dir ,1 ỹi,ir + D(irh̃),2 ỹ∗i+1,ir
x̃ˆ i = D(eqh̃)
ir =1
x̃ˆ i+1 =
D(eqh̃)
−1
Nr
(4.118)
D(irh̃),1
ir =1
where
D(eqh̃)
=
Nr
2
∗
ỹi+1,ir −
D(irh̃),2 ỹ∗i,ir ,
∗
Di(rh̃),it D(irh̃),it .
(4.119)
it =1 ir =1
The combined GFDM blocks in the time domain are given by
x̂i =
1 Hˆ
F x̃i ,
N N
(4.120)
128
A. Nimr et al.
which can be used to recover the data symbols, using the demodulation matrix B, as
presented in Sect. 4.1.
TR-STC can achieve full diversity gain of order 2J, which means that the approximated expression for the symbol error rate derived for the maximum-ratio combiner
(MRC) can be adapted for the TR-STC-GFDM. Assuming that a V -QAM constellation is used to map the data bits into each subcarrier and a non-orthogonal prototype
pulse that leads to an NEF of ξ = N1 tr(BH B) is employed, the symbol error probability for the TR-STC-GFDM is approximately given by
pe ≈ 4μ
2N
r −1 (
i=0
2Nr − 1 + i
i
)(
1+η
2
)i
,
(4.121)
where
√
(
)
1 − η 2J
V −1
μ=
and
√
2
V
5
6 3σe2 Es
6 V −1 ξ N
0
η=7
,
3σe2 Es
2 + V −1
ξ N0
(4.122)
(4.123)
with σe2 = n E[|hn |2 ], ES and N0 denote the average symbol energy and the noise
power, respectively.
Figure 4.21 shows the TR-STC-GFDM symbol error rate (SER) performance
assuming the parameters presented in Table 4.1 and the average channel impulse
response based on the Extended Pedestrian A model from LTE, which is described
Fig. 4.21 TR-STC-GFDM SER performance under time-variant and frequency-selective channel
4 Generalized Frequency Division Multiplexing …
Table 4.1 Simulation parameters
Parameter
Symbol
Mapper
Transmit filter
# subcarriers
# subsymbols
CP length [samples]
CS length [samples]
Detector
Noise enhancement factor
# receiving antennas
V -QAM
g[n]
K
M
Lcp
Lcs
–
ξ
J
129
GFDM
OFDM
16-QAM
RC, α = 0.25
128
7
9
3
ZF
1.02
1
16-QAM
Rect
128
1
9
3
ZF
1
1
Table 4.2 Channel power-delay profile used in the simulations
Tap (nth sample)
0
1
2
3
4
Tap gain hn (dB)
0
−1
−2
−3
−8
5
−17.2
6
−20.8
by Table 4.2. The taps of the channel model are multiplied by i.i.d. complex Gaussian variable with zero mean and unitary variance, resulting in independent Rayleigh
fading channels between each transmit and receive antennas. The results presented
in Fig. 4.21 assumes two different situations. The first one considers SISO, where
just one transmit antenna is employed to send data to the receiver. In the second
scenario, both transmit antennas are active, providing full diversity gain. It is also
assumed that the receiver knows the state information of all channels. We can see
from Fig. 4.21 that the approximation presented in (4.121) can be used to predict the
TR-STC-GFDM under time-variant frequency-selective channels. Also, it is possible
to conclude that the TR-STC is able to provide full diversity gain to GFDM, introducing a considerable SER performance gain when compared to the SISO-GFDM.
Therefore, the high SER performance gain introduced by the TR-STC comes with
a very small complexity increment on the receiver side, once the received signals
can be easily combined in the frequency domain.
4.2 Link-Level Waveform Comparison
In this section, we compare the link-level performance of advanced multicarrier waveforms under MIMO wireless communication channels. The baseline waveform is CPOFDM. In the last decade, it has evolved as a popular multicarrier scheme in different
standards, including 3GPP LTE and Wi-Fi families. However, with new and even
more stringent requirements in 5G and beyond, OFDM faces its limitations, such as
sensitivity to time-frequency misalignments, high OOB emission, limited flexibility,
130
A. Nimr et al.
and high PAPR [15, 16]. To overcome these limitations, advanced alternatives have
been intensively investigated in recent years. Two groups will be examined in this
section.
One group of waveforms attempts to improve OFDM while mostly keeping its
orthogonality. Filtered OFDM (F-OFDM) [17] linearly filters a set of contiguous
subcarriers that form a subband. It is evident that filtering is effective to limit the
OOB emission. On the other hand, the presence of inter-block-interference (IBI) is a
new issue as the linear filter tail will spread outside the duration of each OFDM block.
One pragmatic solution is to insert one or several guard tones (GTs) between adjacent
subbands. A widened subband helps reducing the filter length, thereby alleviating
IBI. Another approach is to use ZP instead of CP as ZP has zero energy and minimizes
IBI. This yields the second waveform in this group, namely universal-filtered OFDM
(UF-OFDM) [18].
The other group of waveforms, consisting of filter bank multicarrier (FBMC) [19]
and GFDM [1], completely discards the orthogonality requirement of OFDM to
achieve better temporal and spectral characteristics. Comparing with OFDM and its
variants in the first group, FBMC and GFDM have three core different features that are
in common. They are: (1) filtering on a subcarrier basis, (2) permission of more than
one data symbol per subcarrier, and (3) being subject to ISI and ICI arising from their
non-orthogonality. Between FBMC and GFDM, they both also have several distinct
features.3 FBMC adopts linear filtering to achieve ultra-low OOB emission [19, 20].
On the other hand, the long filter length makes it more suitable for continuous rather
than burst transmission, considering the usage efficiency of time resources. FBMC
does not use a CP and relies on the soft transition of its filter tail to combat multipath
fading. In GFDM, a unique feature initially adopted by it is circular filtering. This
ensures a block-based waveform with no filter tails, but at the cost of an increased
OOB emission. CP has been suggested as a default setting for GFDM, but a single
one can protect multiple data symbols for the sake of temporal efficiency.
The self-introduced interference of non-orthogonal waveforms often results in
much-increased receiver complexity, e.g., [21–23] and references therein. In MIMO
communications, IAI will additionally takes place. To tackle such three-dimensional
interference, we contributed an innovative way in [24] to perform MMSE equalization
such that they can be jointly resolved with complexity in the same polynomial order
as that of the (quasi-)orthogonal waveforms in the first group.
The focus of this section is to study the link-level performance of the abovementioned waveforms, including OOB emission, PAPR and coded frame error rate
(FER) achieved by the MMSE receiver mentioned before. Challenging channel conditions in terms of large delay spread and time-varying fading with imperfect synchronization and channel estimation are under the consideration.
3 Here we consider the features that were suggested when FBMC and GFDM were invented. As the
recent progress in waveform engineering, these features start becoming mutually usable. Therefore,
they may no longer be regarded as distinct.
4 Generalized Frequency Division Multiplexing …
131
4.2.1 System Configurations
Taking OFDM as the baseline, we use K to denote the total number of subcarriers,
where the indices of the active ones are recorded in the set Kon . The symbol duration
without CP and ZP is denoted as T . At the sampling rate K/T, the CP and ZP
respectively contain Lcp and Lzp samples. Here we assume each frame carries a
codeword, producing Ns OFDM blocks per frame. For the linear filter adopted by
F-OFDM and UF-OFDM, we use Lf to denote its length that is normalized by the
sampling period T /K. Since both GFDM and FBMC permit multiple data symbols
per subcarrier, let M denote this number and T be the time spacing between two
consecutive data symbols.
Table 4.3 summarizes the representative configurations of the waveform candidates, deriving from the baseline CP-OFDM. Their configurations attempt to fulfill
the constraint that each frame uses the same bandwidth and carries the same number
of data symbols.4
Exploiting the additional degree of freedom in the time domain, we particularly
examine two configuration types of GFDM and FBMC. The first type sets the time
spacing to be the same as the duration of the OFDM symbol, i.e., yielding identical
subcarrier spacing 1/T. Each frame consists of a single block to optimize the temporal
efficiency. On contrary, the second type makes the subcarrier spacing of GFDM and
FBMC M times wider than that of OFDM, i.e., M times shorter time spacing between
two consecutive data symbols. In doing so GFDM-II has the same block length as
OFDM, thereby yielding the same number of blocks per frame. For the sake of
temporal efficiency, FBMC-II still has one block per framework, resulting M times
more data symbols than FBMC-I per subcarrier.5
Now, let us set the parameters of the baseline CP-OFDM as: K = 1536, |Kon | =
36, T = 66.67 µs, and Ns = 7. With the subcarrier spacing T −1 = 15 kHz, the occupied subband out of the total 23.04 MHz band has about 0.5 MHz bandwidth. In
accordance with Table 4.3, the corresponding configurations for the other waveforms can be readily computed. For the type-II configuration of GFDM and FBMC
in Table 4.3, we additionally set M = 12. Furthermore, following the suggestions in
the literature, the filters adopted by the waveform candidates except CP-OFDM are
set as follows. UF-OFDM adopts the Dolph–Chebyshev filter with length Lf = 74
and the side-lobe attenuation −51 dB [25]. The filter used by F-OFDM is a Hanning
windowed sinc-function with length Lf = K/2 + 1 [26]. The PHYDYAS filter of
FBMC has the longest filter length equal to 4 K [19]. GFDM adopts a periodic RC
function with the roll-off factor α = 1.
4 Due
to different choices of the filter, it is difficult to achieve the same frame duration without
violating the bandwidth constraint. Considering the strict regulation on the spectrum, identical
bandwidth is our primary constraint.
5 For FBMC, the guard time interval to accommodate the filter tail between blocks can be too large.
Therefore, in both configuration types, we only consider one block per framework to ensure a good
temporal efficiency.
132
Table 4.3 Frame parameterization and modulation complexity
Waveform
Nr. dat.
symbs per
subcar.
Nr. blks per Nr. subcar.
frame
per blk.
Nr. act.
subcar.
Subcar.
spacing.
Sampling
rate
CP-OFDM
−
Ns
|Kon |
1/T
K/T
(F/UF)OFDM
−
Ns
K
|Kon |
1/T
K/T
GFDM-I
Ns
1
K
|Kon |
1/T
K/T
|Kon |
1/T
FBMC-I
Ns
1
K
K
K/T
GFDM-II
M
Ns
K/M
|Kon |/M
M/T
K/T
FBMC-II
Ns M
1
K/M
|Kon |/M
M/T
K/T
For UF-OFDM, its Lzp takes on the same value as Lcp
O(1) represents the arithmetic complexity per multiplication
Frame length
(K+Lcp )Ns
T
K
(K+Lcp )Ns +Lf −1
T
K
Lcp
(Ns + K )T
(Ns + 27 )T
(K+Lcp )Ns
T
K
7
(Ns + 2M )T
Arithmetic complexity per data
symb.
O
O
O
O
O
O
K
|K on |
K
|K on |
K
|K on |
K
|K on |
K
|K on |
K
|K on |
log K
log K +
log K +
log K +
KLf
|K on |
KM
|K on |
4K
|K on |
log(K/M ) +
log(K/M ) +
KM
|K on |
4K
|K on |
A. Nimr et al.
4 Generalized Frequency Division Multiplexing …
133
Fig. 4.22 Power spectral densities (PSDs) of the waveform candidates in their baseband signal
form (per transmit antenna)
4.2.2 OOB Emission
Figure 4.22 depicts the PSD of the waveform candidates in their baseband signal
form. Note that the impairment from the RF front-end is not considered here. As
expected, the OOB emission of CP-OFDM, namely outside the allocated subband
ranging from 5.5 to 6 MHz, is very high, because of the disruptive change from
one OFDM block to another in the time domain. UF-OFDM, F-OFDM, and FBMC
rely on linear filtering to smoothen the transition between blocks, thereby achieving
lower OOB emission. The longer the filter is, the lower is the achieved OOB emission.
For GFDM, circular filtering however keeps the disruptive change between blocks.
GFDM-I achieves slightly improved OOB emission performance by having longer
block duration and reducing the number of blocks per frame. On the other hand,
GFDM-II with the same number of blocks per frame as OFDM achieves nearly
identical OOB emission performance. Note that the soft transition of its PSD on the
shoulder of the occupied spectrum is due to (1) its wider subcarrier spacing than
CP-OFDM; and (2) the large roll-off factor. Such soft transition also appears with
FBMC-II for the same reasons.
Compared to linear filtering, time domain windowing can be an attractive solution
to reduce the OOB emission of block-based waveforms as well. GFDM-I uses the
minimum time resource among all evaluated waveforms, even 6Lcp shorter than the
baseline CP-OFDM. Targeting the same temporal efficiency, we can extend the CP
of GFDM-I by 3Lcp samples and add a CS with identical length to window each
GFDM block without impairing the data transmission plus one CP to combat the
multipath fading channel. Here, the time domain window takes the frequency domain
expression of an RC function, whose ramp up and down are contained within the
extended part of the CP and CS. Figure 4.22 shows that such RC windowing is
very efficient in reducing the OOB emission of GFDM-I, making it competent with
134
A. Nimr et al.
Pr(PAPR > PAPR0 )
100
CP-OFDM
DFT-s-OFDM (spreading factor 12)
UF-OFDM
F-OFDM
GFDM-I
FBMC-I
FBMC-I w./o. ramp up and down
GFDM-II
FBMC-II w./o. ramp up and down
GFDM-II (OQAM)
FBMC-II (OQAM) w./o. ramp up and down
10−1
10−2
10−3
10−4
6
7
8
9
10
11
PAPR0 (dB)
12
13
Fig. 4.23 PAPRs of the waveform candidates in their baseband signal form (per transmit antenna),
where the CCDF is empirically constructed from 106 frames and the oversampling factor is 4. The
default modulation scheme is 16-QAM (gray). The explicit label OQAM represents offset-16-QAM
(gray)
the outer waveforms using linear filters. Besides filtering and windowing, it is also
possible to reduce the OOB emission via preprocessing the transmitted data symbols,
e.g., [27–29], but it is beyond the scope of our discussion here.
4.2.3 PAPR
Figure 4.23 depicts the PAPRs achieved by the waveforms. For those using linear
filter, we note that the ramp up and down of the filtered signal reduce the average
power without affecting the peak power. This causes the PAPR increment of UFOFDM, F-OFDM, and FBMC-I compared to CP-OFDM. It is more noticeable when
the filter length is longer. Note that such PAPR increment will not impose additional
challenge on designing the power amplifier. So, it is not a concern. Since FBMC has
the longest linear filter length among all waveforms, we compute its PAPR of FBMC
excluding its ramp up and down phase on purpose.
For GFDM and FBMC, their additional degree of freedom in the time domain
can be leveraged to improve the PAPR performance. By having fewer subcarriers
and more data symbols per subcarrier, the configuration type-II achieves much lower
PAPRs than the type-I. Figure 4.23 also shows that offset quadrature amplitude modulation (OQAM) is beneficial to further reduce the PAPR of GFDM-II and FBMC-II.
If necessary, other PAPR reduction techniques, such as tone reservation and active
constellation extension, are applicable on top of these waveforms.
4 Generalized Frequency Division Multiplexing …
135
4.2.4 FER Under a Doubly Dispersive Channel
In this following, the waveforms are evaluated in a 4 × 4 spatial multiplexing MIMO
system with spatially uncorrelated multipath Rayleigh fading channels. On top of the
waveforms, two modulation coding schemes (MCSs) are applied. Namely, the turbo
code with the generator polynomial {1, 15/13}o can operate at rate 1/2 and 3/4,6
which are respectively modulated with 16- and 64-QAM (gray). Unless otherwise
stated, QAM is the default choice. Its comparison with OQAM is always under the
same modulation order. The metric Es /N0 denotes the energy per data symbol to
noise ratio.
As for the channel model, we choose the extended typical urban (ETU) model
specified by 3GPP and with the total power of the path gains normalized to one. Given
its large delay spread, we accordingly choose the long CP mode in LTE, specifically,
Lcp = 16.67 µs. Following the Jakes’ model, the maximum Doppler frequency fd
reflects the channel varying rate. For coherently equalizing each block, the used CIR
is obtained by averaging the continuously time-variant CIR over each block duration.
Additionally, we assume perfect synchronization.
Generally speaking, the time-varying channel can affect the FER performance
from two conflicting aspects. First, a continuous time-varying channel introduces
ICI that increases along with the maximum Doppler frequency. Second, the time
selectivity across the blocks is desirable for the decoder to exploit the code diversity
in order to improve the decoding performance. When the additive noise dominates
over ICI, the second aspect plays the determinant role. Therefore, lower FERs are
attained at a higher maximum Doppler frequency, e.g., CP-OFDM, UF-OFDM, FOFDM, and GFDM-II in Fig. 4.24a, c. However, when the ICI becomes the dominant
factor, we observe higher FERs as the maximum Doppler frequency increases, e.g.,
in Fig. 4.24b, d.
For a similar reason, the IBI introduced by linear filtering of UF-OFDM and
F-OFDM becomes particularly harmful at the higher MCS, which requires higher
operating Es /N0 for a satisfactory FER performance. UF-OFDM benefits from the
use of ZP and a shorter filter to suffer from less IBI than F-OFDM. As mentioned at
the beginning of this section, we can alleviate the IBI issue encountered by F-OFDM
through GT insertion. Figure 4.24b, d depict the performance achieved by having
one GT on each side of the subband, which however costs spectral efficiency, i.e.,
(2/38) ≈ 5% loss.
For FBMC and GFDM-I, they suffer from the need of a long block length. With
long blocks, the time-variant CIR cannot be well approximated by its average value
and the resulting mismatched channel knowledge can severely degrade the performance of equalization and subsequent decoding. Therefore, they perform poorly
with fd = 300 Hz. Even at a lower maximum Doppler frequency fd = 70 Hz, such
impairment is non-negligible at high SNRs, i.e., Fig. 4.24b.
6 For every six information bits input to the turbo code, we keep all information bits plus two parity
bits respectively generated by the two identical component convolutional codes.
136
A. Nimr et al.
FER (after 10 turbo iter.)
CP-OFDM
GFDM-I
FBMC-II
UF-OFDM
GFDM-II
FBMC-I (OQAM)
100
100
10−1
10−1
10−2
10−2
F-OFDM
FBMC-I
FBMC-II (OQAM)
F-OFDM with one GT
10−3
5
6
7
8
9
Es /N0 (dB)
10
11
12
10−3
16
FER (after 10 turbo iter.)
(a) Code Rate 1/2, 16 QAM, fd = 70 Hz
18
20
22
24
Es /N0 (dB)
26
28
30
(b) Code Rate 3/4, 64 QAM, fd = 70 Hz
100
100
10−1
10−1
10−2
10−2
F-OFDM with one GT
10−3
5
6
7
8
9
Es /N0 (dB)
10
11
(c) Code Rate 1/2, 16 QAM, fd = 300 Hz
12
10−3
16
18
20
22
24
Es /N0 (dB)
26
28
30
(d) Code Rate 3/4, 64 QAM, fd = 300 Hz
Fig. 4.24 FERs achieved by the waveforms under perfect synchronization and channel knowledge
Only in Fig. 4.24a, the benefit of FBMC becomes appreciable. Non-orthogonality
not only introduces interference, but also spreads the information of each data symbol
over more than one channel observations in the frequency domain. Particularly, the
type-II has larger subcarrier spacing than the type-I to ensure the necessary frequency
selectivity among the channel observations. Analogous to FBMC-II, GFDM-II is
also equipped with such feature. Between them, FBMC-II in this case has more
channel observations per data symbol to achieve a lower FER. Between GFDM-II
and FBMC-II, the former permits short block lengths without considerably reducing
the temporal efficiency. Therefore, it outperforms the latter in higher mobility case,
e.g., Fig. 4.24c, d.
Last but not least, as shown in our work [24], QAM can more efficiently exploit
the frequency selectivity of the channel than OQAM, therefore outperforming in
Fig. 4.24.
4 Generalized Frequency Division Multiplexing …
137
4.2.5 FER with Imperfect Synchronization and Channel
Estimation
This part investigates the impact of synchronization and channel estimation error
on the FER performance of the waveforms. The channel is generated by following
another 3GPP channel model termed extended vehicular A model (EVA) with the
maximum Doppler frequency equal to 30 Hz and with the sum of the average path
gains normalized to one. Due to the reduced maximum delay spread, the CP length
accordingly decreases to 4.69 µs, namely the normal mode in LTE.
For data-aided channel estimation, we insert a preamble consisting of one baseline
CP-OFDM block before each payload frame. It consists of Nt orthogonal pilot vectors that are periodically modulated onto the subcarriers belonging to the occupied
subband plus Nt subcarriers on each side for achieving sufficient channel estimation
quality also on the edge of the subband. Using such preamble, the receiver performs
LMMSE channel estimation by assuming a uniform power-delay profile with maximum delay length equal to Lcp [30]. The obtained channel estimates will be used as
the true one for coherently equalizing the whole frame. One important reason behind
this setup is to assure that the MMSE equalizer of each waveform works with the
same quality of channel knowledge.
Figure 4.25a shows that FBMC-II and GFDM-II both are robust against the channel estimation error, achieving up to 5 dB gain in comparison with CP-OFDM and
its variants. This observation indicates that it is possible to harness the benefits of
the waveform-induced interference instead of only suffering from it. To this end, we
CP-OFDM
GFDM-II
UF-OFDM
FBMC-I
GFDM-I
10−1
100
FER (after 10 turbo iter.)
F-OFDM
FBMC-II
10−2
10−1
10−3
−10 −8 −6 −4 −2
100
10−2
0
2
4
6
8
10
STO (sample)
10−1
10−2
10−3
10
12
14
16
18
Es /N0 (dB)
20
22
24
(a) Perf. Sync. and Imperf. Channel Knowledge
10−3
0
10
20
30
40
50
abs. CFO (ppm) at Carrier Freq. 2.6 GHz
(b) Sensitivity to STO and CFO at Es /N0 = 18 dB
and with Imperf. Channel Knowledge
Fig. 4.25 FERs achieved by the waveforms with the maximum Doppler frequency 30 Hz and
relying on imperfect synchronization and channel knowledge, where the 16 QAM and code rate
1/2 are the default MCS and the contrastive OQAM has the same modulation order
138
A. Nimr et al.
need to exploit the data information conveyed by the interference rather than treating
it as a part of the background white noise.
We further investigate the sensitivity of the waveforms against symbol time offset
(STO) and carrier frequency offset (CFO), respectively. Specifically, the use of guard
intervals, no matter in the form of CP, ZP with overlap-and-add or soft termination
of the filter, provides protection against negative STO estimates, i.e., the estimated
frame arrival being earlier than the true one. The performance degradation appears
once we have a positive STO estimate. How severe the degradation is depends on
the power of the first non-negligible paths of the channel in its discrete-time model.7
From Fig. 4.25b, we can infer that the initial 4 paths of the channel are insignificant.
According to Fig. 4.25c, the waveforms can work under the CFO up to ±20 ppm. For
FBMC-II and GFDM-II, even with ±30 ppm CFO, their performances are similar to
that of OFDM working with ±20 ppm CFO.
4.2.6 Section Summary
In this section, we have analyzed the link-level performance of advanced waveforms
that are being intensively researched as alternatives to CP-OFDM for future systems.
There is no single waveform that can outperform the others in all examined aspects,
i.e., OOB emission, PAPR and FER under different channel conditions.
By observing the 3GPP RAN1 discussion through publicly available materials, the
waveforms targeted by this section have all appeared in the proposals from different
organizations. At this moment, the OFDM-based waveforms are more interested
and supported by the main industrial players to ensure a good backward capability.
Nevertheless, non-orthogonal waveforms, i.e., GFDM and FBMC, are definitely
worth investigation. We believe their benefits can be exploited with a complexity that
is affordable by today’s hardware. Furthermore, non-orthogonality is not necessary
to be a curse in the system design. Further research on non-orthogonal waveforms
is no doubt valuable for a wide range of communications systems using multicarrier
waveforms, including but not limited to mobile systems.
4.3 Multiple Access with GFDM
In multicarrier-based multiple access, the time and frequency resources are distributed among users. The basic resource element corresponds to transmitting one
t
[n]. However, in practice, the
data symbol dk,m per block using the pulse shape gk,m
7 Given
the power-delay profile, the discrete-time channel model is obtained by sampling the lowpass filtered CIR, where the bandwidth equals the sampling rate. The discrete-time model very
often have more resolvable paths than the power-delay profile. This is because the delays specified
by the power-delay profile are not integer multiples of the sampling period.
4 Generalized Frequency Division Multiplexing …
139
Fig. 4.26 GFDM resource allocation. In this example, M = 5, K = 8, Mu = 4 and Ku = 2
smallest physical resource block (PRB) consists of a number of basic resources and
several blocks. For example, in LTE standard, which uses OFDM in the physical
layer (PHY), the smallest PRB consists of 12 subcarriers and 7 or 6 OFDM symbols,
for long and short CP, respectively. In addition, the scheduling, where the resources
can be reassigned to different users, takes place after certain time, for example, in
LTE after 12 or 14 symbols [31].
In GFDM, the resource allocation is done by setting the entries of the data matrix
D corresponding to the allocated {(k, m)} pairs and the other entries are set to zero.
(u)
⊂ Kon × Mon be the set of allocated {(k, m)} to the u-th user.
Let Kon(u) × Mon
(u)
Practically, we fix Mon
= Mon and vary Kon(u) . Therefore, we define a PRB that
have Mu = |Mon | subsymbols and Ku subcarriers. The scheduling can take place
after certain number of GFDM blocks. Figure 4.26 shows an example of GFDM
resource allocation.
4.3.1 Signal Model
We consider a network that consists of a base station (BS) and U users. In the downlink (DL), as seen in Fig. 4.27, the BS multiplexes the data symbols of all users in
a data vector d = vec {D} using the set of indexes Non(u) = {n = k + mK, (k, m) ∈
(u)
Kon(u) × Mon
}, with d(u) = [d]N (u) , and generates the modulated signal as dison
cussed in Sect. 4.1.6.2. Each user receives the multiplexed signal given by
y(u) [n] = ej2πfu n h(u) [n] ∗ x(u) [n] + v[n],
(4.124)
where h(u) [n] is the fading channel and fu represents the CFO between the u-th user
and the BS. The STO is implicitly included in h(u) [n]. Under perfect synchronization,
i.e., fu is perfectly estimated and the first channel tap is detected. By employing a
sufficiently long CP, we get the following signal model after block demultiplexing,
140
A. Nimr et al.
Fig. 4.27 DL signal model
Fig. 4.28 UL signal model
y(u) = H(u) Ad + v,
(4.125)
where H(u) is the circular channel matrix. This model can be used directly in a simple
receiver, where the u-th user demodulates y(u) to get d̂ and then applies resource
demapping to extract its allocated data symbols. More advanced receiver can work
on the model,
y
(u)
(u)
(u) (u)
=H A d
+H
U
(u)
v=1,v=u
A(v) d(v) + v.
(4.126)
Here, the second term corresponds to the inter-user-interference (IUI). Actually, the
interference in this case is inherited from the self-interference of the matrix A. Thus,
the study of the DL performance is similar to that of point-to-point link.
In the uplink (UL), Fig. 4.28, each user generates its modulated signal using the
truncated modulation matrix A(u) = [A]:,N (u) , as discussed in Sect. 4.1.2.1. The
on
BS receives a superposition of the signals from all users, which is expressed as
y[n] =
U
u=1
ej2πfu n h(u) [n] ∗ x(u) [n] + v[n].
(4.127)
In the UL scenario, we distinguish two multiple access (MA) schemes
• Synchronous MA: the users are strictly synchronized with the BS, i.e., fu = 0 and
no time offset, which can be achieved via closed-loop synchronization, then
4 Generalized Frequency Division Multiplexing …
y = H(u) A(u) d(u) +
U
v=1,v=u
141
H(v) A(v) d(v) + v.
(4.128)
In this case, the IUI is inherited from the self-interference properties of the GFDM
waveform. For example, when the self-interference is limited to adjacent subcarriers, the IUI can be null if a sufficiently large guard subcarrier is used between
the adjacent users.
• Asynchronous MA: the synchronization is coarse that there is a remaining CFO
modeled by fu and remaining STO which leads to the increase of the channel
delay spread [32]. Nevertheless, the CP can be extended to take into account
the maximum possible STO in additional to the channel excess delay. With that,
and assuming the BS is able to perfectly estimate fu and H(u) , we get the signal
corresponding to each user after the CFO compensation as
H
y(u) = diag φ (u) y
= H(u) A(u) d(u) +
U
v=1,v=u
diag φ (v,u) H(v) A(v) d(v) + v,
(4.129)
where φ (u) (n) = ej2π(fu n+cu ) and φ (v,u) (n) = ej2π(Δfv,u n+cv,u ) are constant phase
and Δfv,u = fv − fu is the relative CFO. In this case, the IUI arises from the relative
CFO among users.
In the next subsections, we study the IUI using the frequency domain processing.
4.3.2 Frequency Domain Processing
Consider GFDM with prototype pulse shape that has a maximum discrete frequency
response within two subcarrier spacing. Without loss of generality, we let
⎡
g̃T1
⎤
(g̃)
VK,M = ⎣ 0K−2,M ⎦ ,
g̃T2
(4.130)
where g̃1 = g̃ (0:M −1) ∈ CM ×1 and g̃1 = g̃∗ (N −M :N −1) ∈ CM ×1 . Recalling (4.31),
V(x̃)
K,M
(q,p)
=
K−1
m=0
(g̃)
VK,M
(<q−k>K ,p)
M
−1
m=0
m
dk,m e−j2π M p ,
142
A. Nimr et al.
Fig. 4.29 Frequency domain signal processing model
we get
V(x̃)
K,M
V(x̃)
K,M
(q,:)
(q+1,:)
=
K−1
(g̃)
VK,M
k=0
(<q−k>K ,:)
= g̃1 ⊙ D̃
(q,:)
= g̃1 ⊙ D̃
(q+1,:)
⊙ [D](k,:) FM
+ g̃2 ⊙ D̃
(<q+1>K ,:)
+ g̃2 ⊙ D̃
,
(<q>K ,:)
where
D̃ = DFM
(4.131)
represents the M -DFT-spread data matrix. Thus, due to the M -DFT spreading, any
allocated data symbol in the k-th subcarrier produces samples in the k-th and (k + 1)th subband, as demonstrated in Fig. 4.29. For the set of allocated subcarriers Kon(u) ,
the set of occupied frequency subbands can be defined by
K
(u)
= K1(u) ∪ K2(u) ,
(4.132)
with K1(u) = Kon(u) and K2(u) = < Kon(u) − 1 >K . In addition, let M1 and M2 be
the indexes of the nonzero elements of g̃1 and g̃2 , respectively, we define the set of
occupied frequency indexes as
S (u) = {(k, m) ∈ Ki × Mi , i = 1, 2}.
Accordingly, we adapt a masking matrix U(u) of size |K
(u)
(4.133)
| × M defined as
(u)
U (k,m) = 1, (k, m) ∈ S (u) , and 0 elsewhere.
(4.134)
4 Generalized Frequency Division Multiplexing …
143
This matrix can
be used
as q frequency domain window at the receiver. Recalling
(x̃) T
that, x̃ = vec VK,M , this window can be applied to the received vector using the
form
(u)T
w(u)
.
(4.135)
F = vec U
so that
w(u)
F
(N
(u)
)
= 1, and 0, elsewhere,
(4.136)
and
N
(u)
= {n = m + kM : (k, m) ∈ S (u) }.
(4.137)
Moreover, by taking the N -DFT of (4.129) and applying the frequency domain
windowing we get
*
H +
ỹ(u) = FN diag φ (u) y ⊙ w(u)
F
⎡
=D
h̃([)
=D
h̃([)
u]Ã
u]Ã
(u) (u)
d
(u) (u)
d
(u)
+ wF
+
⊙ ⎣F N
U
v=1,v=u
U
v=1,v=u
diag φ
(v,u)
(v)
H A d
(u)
Z(v→u) d(v) + diag wF ṽ.
⎤
(v) (v) ⎦
(u)
+ wF ⊙ ṽ
(4.138)
where
Z̃
(v→u)
1
(v,u) H h̃(v) (v)
j2πcv,u
e
F
diag
φ
F
D
= diag w(u)
à .
N
N
F
N
(4.139)
h̃(x)
= diag h̃(x) . Equation (4.139) defines the signal model after synchroand D
nization, so we can compute the SIR assuming uncorrelated data as
SIR(u) =
h̃(u) (u) 2
Ã
Pu D
F
2 ,
U
(v→u)
Pv Z̃
v=1,v=u
(4.140)
F
where · F is the Frobenius norm, and Px = E |d (x) |2 is the x-th user power per
symbol. The SIR depends on the modulation matrix, resource and power allocation,
CFO and the CIR.
The SIR metric is useful to evaluate the IUI; however, it is also important to
examine the overall performance after channel equalization and demodulation. Let
(u)
B̃ be the receiver matrix involving channel equalization and demodulation, then
144
A. Nimr et al.
d̂
(u)
(u)H (u)
= B̃
(u)H
= B̃
ỹ
D
h̃(u)
(u)
(u)H
à d(u) + B̃
U
(v→u) (v)
Z̃
d
v=1,v=u
(u)
ṽ.
+ B̃ diag w(u)
F
(4.141)
The average signal-to-interference-plus-noise ratio (SINR) is considered as a performance metric. It is given by
(u)
SINRB =
2
(u)H h̃(u) (u)
D
à − I
Pu
+
B̃
F
Actually, SINR(u)
B
symbols.
−1
U
(u)
Pu Non
v=1,v =u
.
(u)
(u) (v→u) 2
(u) 2
Pv B̃ Z̃
+ N0 B̃ diag wF
F
F
(4.142)
is the normalized mean squared error (NMSE) of the data
4.3.3 Asynchronous MA Evaluation
For the purpose of comparing the asynchronous MA with different waveform parameters, we consider two users, i.e., U = 2, with identical power allocation P1 = P2
and evaluate the SIR (4.140) and the SINR (4.142) in different scenarios.
4.3.3.1
AWGN Channel
This evaluation is useful to comprehend the interference due to the CFO. Let
C(1,2) =
then
SIR(u) =
1
FN diag φ (1,2) FH
N,
N
g̃k,m 2
2
k∈K on(1) m∈M on
2 .
(1)
(1,2)
diag
w
C
g̃
k,m
F
(4.144)
2
k∈K on(2) m∈M on
Noting that
(4.143)
(1,2)
g̃k,m (q) = G k,m (ν =
C
q
N
− Δf1,2 ),
(4.145)
where G k,m (ν) is the DTFT of gk,m [n], then
2
(1)
G k,m ( q − Δf1,2 )2 .
diag wF C(1,2) g̃k,m =
N
2
q∈N
(1)
(4.146)
4 Generalized Frequency Division Multiplexing …
Therefore,
145
(2→1) 2
S2 ( Nq − Δf1,2 ),
Z̃
=N
F
q∈N
(4.147)
(1)
where Su (ν) is the PSD of the signal of the u-th user without CP expressed as
Sx (ν) =
2
1
G k,m (ν) .
N
(x)
k∈K on
As a result, we can write the SIR in the form
SIR(1) =
q∈N
q∈N
(1)
(4.148)
m∈M on
(1)
S1 ( Nq )
S2 ( Nq − Δf1,2 )
.
(4.149)
This form is intuitively comprehensive as N (1) represents the frequency band of
the first user. Additionally, the interference of the second user is the integral over
the leakage within the first user’s band, as illustrated in Fig. 4.30. From this, it can
be shown how important to study the OOB of the waveform, which in the case of
GFDM depends on the prototype filter. This equation can be simply extended to
U > 2. Figure 4.31 justifies the closed-form solution with the numerical simulation.
In addition, it can be seen that for the same allocated bandwidth and the same guard
band, GFDM is significantly less sensitive to CFO compared to OFDM. It is important
to highlight that the first subsymbol is turned off. Additionally, the worst situation
for the used prototype filter happens when the CFO is half the frequency sample,
which is exactly half the subcarrier spacing in OFDM. On the other hand, a shift by
complete sample retains the orthogonality if a sufficient guard band is used.
Fig. 4.30 IUI due to CFO using. In the synchronized case no interference as the sampling points
are at zero crossing. With CFO, the samples are not equal to zeros
146
A. Nimr et al.
Fig. 4.31 SIR(1) simulation versus closed-form for GFDM and OFDM
4.3.3.2
Fading Channel
In the existence of fading channel, we define the modified pulse shape as
1
(u)
FN (h̃ ⊙ g̃m,k ).
N
gk,m|h(u) =
(4.150)
Following the same discussion, the PSD of the u-th user can be written as
Su|h(u) (ν) =
2
1
G k,m|h(u) (ν) ,
N
(u)
k∈K on
and then
SIR(1) =
q∈N
q∈N
(1)
(4.151)
m∈M on
(1)
S1|h(1) ( Nq )
S2|h(2) ( Nq − Δf1,2 )
.
(4.152)
By averaging over different channel realization, we achieve the SIR for certain PDP.
Figure 4.32 shows the effect of the fading channel considering the TGn channel
model in comparison with AWGN channel for different roll-off factors of RC filter.
As expected, while larger roll-off factor produces well-localized filter and reduces the
side lobes, the SIR is higher for larger roll-off. However, this gain may be reduced
after the receive filter. For example, if ZF receiver is used, then the gain will be
reduced by the NEF, which is higher for larger roll-off. On the other hand, with
fading channel the average SIR increases, that is because the channel gain may
attenuate the interfering signal. This can be beneficial if a proper equalization such
as MMSE is used.
4 Generalized Frequency Division Multiplexing …
147
Fig. 4.32 SIR(1) for different roll-off in AWGN and TGn fading channel. K = 64, M = 16, Mon =
{1, . . . , M − 1}
4.3.4 Mixed-Numerology with GFDM
The supported services in the 5G are diverse, and each service has different requirements [33]. For example, to support the low latency of tactile Internet, shorter symbol
duration needs to be used. To tackle the effect of Doppler shift in high speed vehicular
communications, larger subcarrier spacing is required. The massive machine-type
communications (mMTCs) employ fewer subcarriers to ensure lower PAPR. In order
to multiplex all these services at one BS, the need of mixed-numerology arises. As
discussed in Sect. 4.1.6, GFDM can be seen as a universal multicarrier waveform
generator, with several reconfigurable parameters that can be altered on the fly to
generate the corresponding signal. This feature promotes the GFDM framework to
be a candidate for mixed-numerology.
The main issue arises here is the inter-numerology-interference (INI), which can
be seen as IUI if we consider different users per service. In all cases, we need a model
to evaluate the interference in order to optimize the design and the allocation of the
services in the time and frequency domains.
4.3.4.1
General Inter-user-Interference Model
Without loss of generality, we consider two users u = 1, 2 transmitting their signals
with GFDM modulation using Ku subcarriers, Mu subsymbols and a pulse shape
(u)
g (u) [n] with subcarrier allocation indexes Kon,u
, whose entries are subset of Kon,u
(u)
and Mon,u = Mon,u . Note that, the sub-index indicates the numerology index, while
the sup-index represents the user index in that numerology. Let y[n] be the discrete
received signal
(4.153)
r[n] = r (1) [n] + r (2) [n] + v[n],
148
A. Nimr et al.
where r (u) [n] is the signal contribution from the u-th user and v[n] is the additive
noise. In order to decode the u-th signal, the receiver performs its operation based
on r (u) [n] configurations, while the signal of the other user is the interference. After
removing the CP, we get the block of the u-th user as
zu,i [n] = r[n + Lcp,u + iLs,u ], n = 0, . . . , Nu − 1,
(4.154)
where Lcp,u and Ls,u are the CP size and block spacing used by the u-th user, respectively. Afterward, Nu -DFT is computed, where = Ku Mu . Therefore, we get
(2→1)
+ ṽ1,i ∈ CN1 ×1 ,
z̃1,i = z̃(1)
i + z̃i
(4.155)
z̃2,i =
(4.156)
z̃(2)
i
+
z̃(1→2)
i
N2 ×1
+ ṽ2,i ∈ C
.
Here, z̃i(u1 ) is the i-th block of the u-th user containing all samples, z̃i(u1 →u2 ) is the
interfering block from u1 to u2 and ṽu1 ,i the additive noise. First, we consider the
case of AWGN. Using the multicarrier representation, we have
1)
z̃(u
=
i
(u )
k∈K on,u11
(u1 ) (u1 )
g̃k,m
dk,m,i ,
(4.157)
m∈M on,u1
(u1 )
where g̃k,m
is the Nu1 -DFT of the pulse shapes of the u1 -th user as originally generated
by the waveform. Moreover, the interfering signal can be expressed as
z̃i(u2 →u1 ) =
(u2 →u1 ) (u2 )
g̃k,m,j
dk,m,I j (i) .
(4.158)
j∈J i k∈K (u2 ) m∈M on,u2
on,u2
(u2 →u1 )
In this notation, g̃k,m,j
is the Nu1 -DFT of the pulse shapes seen by the receiver
of u1 . This pulse shape can be computed from the original pulse shape g (u2 ) and
the multiplexing parameters, as seen in Fig. 4.33. The set Ji represents a set of
indexes depending on the index i and Ij (i) is the block index of u2 that contributes
to the interference. Note that Ij1 (i1 ) = Ij2 (i2 ), i1 = i2 and j1 = j2 . This is the key
difference in mixed-numerology. If the same numerology is used, then Ji disappears
and Ij (i) = i. But in mixed-numerologies, the interference may depend on the block
index. However, by a proper design we can control the pattern, such that JPi+p =
Jp , in order to have P interference patterns applied to different block indexes, such
that
(u2 →u1 ) (u2 )
(u2 →u1 )
g̃k,m,j
dk,m,I j (i) .
(4.159)
=
z̃Pi+p
j∈J p k∈K (u2 ) m∈M on,u2
on,u2
Following the same approach in (4.149), we get the SIR for each pattern
4 Generalized Frequency Division Multiplexing …
149
Fig. 4.33 INI example. N2 = 2N1 , Lcp,2 = 2Lcp,1 . The i-th block of u = 2 interferes with the 2i-th
and (2i + 1)-th blocks of u = 1 in two different patterns. Each pattern has one pulse shape. On the
other side, two blocks of u = 1 interfere with one block of u = 2. In the latter case, there is only
one pattern but with two pulse shapes
SIRp(u1 ) =
(u )
q∈N u1 1
(u )
q∈N u1 1
S1 ( Nq )
Sp,u2 →u1 ( Nq )
.
(4.160)
Here, Nu1(u1 ) is the set of nonzero frequency samples of u1 in its numerology, and
Sp,u2 →u1 (ν) =
1
Nu1
(u →u ) 2
2
1
(ν) .
G k,m,j
(4.161)
j∈J p k∈K (u2 ) m∈M on,u2
on,u2
The average SIR can be computed by averaging the SIR overall patterns. This can
be extended in a similar way to the case of fading channel as in (4.152).
150
A. Nimr et al.
Fig. 4.34 INI for OFDM with K1 = 32, Lcp,1 = 8, K2 = 64, Lcp,2 = 8, Δf = Δf1 = 2Δf2 . The
users are allocated the same bandwidth, so that, Ku,1 = 1, Ku,2 = 2
4.3.4.2
Numerical Examples
To verify the closed-form solution in (4.160), consider OFDM systems with different
subcarrier spacing. The first system uses K1 subcarriers and the second system uses
K2 = 2K1 , in other word, the second system employs half the subcarrier spacing. The
CP length in both cases is 1/4 of the symbol length, and both systems transmit data
with unit power. Under perfect synchronization, we get the frame structure presented
in Fig. 4.33. As illustrated in Fig. 4.34, we compare the SIR of one allocated subcarrier
of the first system with two subcarriers from the other system, so we get the same
bandwidth. The guard band is an integer number of subcarriers which is normalized
to the largest subcarrier spacing Δf . As expected, the interference from user 2 to user
1 decreases with larger guard band. In addition, the guard band can be controlled by
the smaller subcarrier spacing of system 2, where Δf2 = Δf /2. Nevertheless, it can
be observed that the behavior of the interference from user 1 to the user 2 fluctuates
and is influenced by the guard interval.
Next, we compare GFDM with different settings. In this example, the target is to
have short and long blocks as in Fig. 4.33. This can be attained by changing either
the subcarrier spacing or the subsymbol spacing. The configurations and results are
shown in Fig. 4.35, where the pair (U1, U2) corresponds to changing the subcarrier
spacing, while the pair (U3, U2) realizes the different block lengths via different
subsymbol spacing. It can be seen that with the used RC pulse shape, the interference
in both directions follows the same behavior. Interestingly, it can be shown also that
both settings achieve similar performance with slight gain when different subsymbol
spacing is applied. Another advantage is that keeping smaller subcarrier spacing
allows fine control of the guard band. Finally, GFDM outperforms OFDM by a gain
up to 10 dB in the achieved SNR for the same subcarrier spacing.
4 Generalized Frequency Division Multiplexing …
151
Fig. 4.35 INI for GFDM with different settings for user 1 and user 2
4.4 GFDM Implementation
The path to meeting all requirements of 5G is a challenging endeavor. The demand
to achieve higher data rates for the enhanced media broadband (eMBB) scenario and
novel use cases like ultra-reliable and low-latency communication (URLLC) and
mMTC drive researchers and engineers to consider new concepts and technologies
for future wireless communication systems. The goal is to identify promising candidate technologies among a vast number of new ideas and decide which are suitable
to be implemented in future products. Figure 4.36 gives a rough overview of the
development process.
New ideas and concepts typically first undergo extensive software simulations,
which allow to make early predictions on the expected performance. After selection
of the best candidates, individual aspects of the envisioned system can be implemented on a hardware-accelerated platform, e.g., software-defined radio (SDR), in
order to learn about real-time behavior and over-the-air performance with real radio
frequency (RF) components. Technologies that prove promising at this stage can be
further evaluated in test beds, where the focus shifts toward the interaction of different
technology building blocks and the realization of complete end-to-end applications.
New concepts and technologies that have been proven to exhibit improved performance in practically relevant environments will ultimately find their way into new
standards and lastly, industry will adopt them in future products.
152
A. Nimr et al.
New ideas
and concepts
Simulations
Predict expected performance
based on suitable models and
reasonable assumptions.
Prototypes
Find reasonable simplifactions, algorithm partitioning and estimate
implementation complexity.
Testbeds
Standards
Demonstrate the interworking of different components under real-world
conditions and in real-time.
Decide which technologies should be included in future products.
Products
Fig. 4.36 From theory to practice
This section presents a field-programmable gate array (FPGA)-based, real-time
implementation of a modulator/demodulator for multicarrier waveforms [34]. This
proof-of-concept design provides a large number of degrees of freedom to the user
and offers the flexibility for practical evaluation of new algorithms that aim to address
various 5G aspects. Example applications include experiments with flexible numerology, which is a key differentiator of 5G new radio (NR) compared to fourth-generation
(4G) LTE, as well as the design and implementation of a corresponding scheduler
that utilizes the additional flexibility. The presented platform is a building block for
test beds that will assist the design of 5G radio interface and network architecture
[35, 36].
4.4.1 Modem Implementation
For the sake of simplicity, consider a communication system that consists of two
nodes. Each node is realized by a control PC that is connected to a software-defined
radio platform. This hardware setup is depicted in Fig. 4.37. Note that the design that
is described in this section is tailored for the USRP-RIO hardware [37]. However,
the basic principles of the implementation are valid for other platforms in general.
The block diagram in Fig. 4.38 shows how the components of the overall system are
mapped to the hardware platform.
4 Generalized Frequency Division Multiplexing …
153
Fig. 4.37 Hardware setup
Fig. 4.38 Block diagram of the transceiver
In order to be able to support the various requirements of 5G wireless systems,
the individual blocks need to be implemented with flexibility in mind. This section
will focus on the PHY aspects of a 5G NR transceiver. The corresponding signal
processing algorithms have to be implemented on FPGA, in order to be able to meet
throughput and latency requirements. As medium access layer and higher layers have
more relaxed requirements w.r.t. timing, when operating in sub-6 GHz bands, those
components rarely require specialized hardware acceleration. Hence, the assumption
is made that they are implemented using software running on standard PC hardware.
4.4.1.1
Baseband Modem
The baseband signal processing is performed by the resource mapper and modulator
blocks on the transmit path, and demodulator and resource demapper on the receive
path, respectively. Note that the block encoder, QAM mapper, decoder, and QAM
demapper are standard implementations taken from a 4G LTE library [38], and hence
will not be discussed here.
As seen in Fig. 4.39, the resource mapper takes complex symbols from various
input sources, e.g., payload data, control channel data, and reference signals, and
154
A. Nimr et al.
Fig. 4.39 Block diagram of the resource mapper and the modulator
maps them to a two-dimensional time-frequency resource grid. The mapping pattern,
i.e., the resource map, is fully programmable. This mechanism allows to support any
user-defined resource grid that can be adapted during the run-time. The resource
demapper performs the inverse operation.
The modulation operation can be separated into two main functional blocks. The
first block is an inverse discrete Fourier transform (IDFT) that transforms the data
from frequency domain into time domain. The second block, which will be called
core modem, applies a pulse shaping filter to each subcarrier of the transmit signal.
The demodulation operation consists of the same processing blocks in reversed order,
where the only differences are the direction of the Fourier transformation and the
filter coefficients that are used in the core modem.
The first task of the core modem is to split the incoming IDFT output stream
δm into M subsymbols with K samples in each, which are stored in block RAM
(BRAM) 0, . . . , M − 1. Each individual subsymbol has to be repeated M times,
such that the filter gm with N samples can be applied. Therefore, the K samples
of each subsymbol are stored inside an independent subsymbol memory bank. All
M memories are read sample-by-sample in parallel. The filter gm is also stored in
M parallel subsymbol filter banks. Each of the filter memory banks contains K
coefficients that represent a different part of the pulse shaping filter in the time
domain. The filter needs to be applied to the data in a circularly shifted way. This
is implemented with a circular pattern that dynamically selects which filter BRAM
is connected to which subsymbol BRAM. The last step in the core modem is to
accumulate the contributions from all M parallel branches to get the transmit signal.
4.4.1.2
Post Modem Processing
After the signal is modulated, CP, CS, time-windowing, and preamble are added.
Figure 4.40 depicts the general frame format. The CP and CS can be applied on both
preamble and data block. NP,CP defines the length of the CP for the preamble, NCP
for the data block, NP,CS defines the length of the CS for the preamble, and NCS for
4 Generalized Frequency Division Multiplexing …
NP,CP
NP
NP,CS
NP,W
155
NCP
N
NCS
NP,W NW
CP Preamble CS
NW
CP
GFDM block
CS
Fig. 4.40 Supported frame structure with one preamble and one data block
Fig. 4.41 CP, CS, windowing and preamble insertion
the data block. It is assumed that the window is symmetric, thus, the length of one
half is given by NP,W and NW . In addition, NP denotes the length of the preamble.
The preamble is calculated in advanced on the host computer and written to the
BRAM memory during the configuration phase of the transceiver. The CP and CS of
the modulated GFDM data block are added via first in, first out (FIFO) memory as
depicted in Fig. 4.41, which can be seen as a variable delay to shift the data samples
into the correct output order.
The windowing unit follows, where only the rising half is stored inside a memory.
An integrated counter in the control logic counts up until NW is reached to trigger the
memory for the appropriate samples. During the main data block, the unit is disabled.
Finally, the same counter is decreased to create the falling flank.
Whenever the controller has finished reading in the first data block of a frame into
the data-FIFO, the preamble insertion unit is triggered to push the preamble samples
to the digital-to-analog (D/A) converter.
4.4.2 Complete Transceiver Chain and Extension for MIMO
A complete transceiver was implemented in FPGA as a proof of concept with its
block diagram shown in Fig. 4.42. It includes all required processing functions for a
real-world wireless communication system.
156
A. Nimr et al.
PA
Rate
Adaption
Polar
Encoder
CP/CS &
Windowing
TR-STC
Encoder
Frame
Multiplex
Digital PreDistortion
TX
Front-End
PA
Synch.
Preamble
Transmit
Message
Resource
Mapper
GFDM
Modulator
Transmitter
Receiver
RX
Front-End
Imparments
Correction
RX
Front-End
Channel
Estimation
Preamble
Automatic
Gain
Control
Synchronization
Channel
Estimation
STC
Combiner
Polar
Decoder
Stuffing
Removal
Synch.
Preamble
Channel
Estimation
Preamble
GFDM Demodulator
Resource
De-mapper
Received
Message
Fig. 4.42 Transceiver block diagram
A flexible frame structure was designed to cope with many design aspects such as
synchronization, channel estimation, codeword length, resource allocation, MCS,
channel multipath protection, OOB emission, MIMO operation, etc. Figure 4.43
depicts the adopted frame structure. The number of GFDM symbols within a frame,
NG , is chosen such that there is an integer number of codewords carried by an integer
number of GFDM symbols. Therefore, the receiver can synchronously start decoding
a codeword at the beginning of the frame. Synchronization and channel estimation
were implemented using a preamble-based scheme, i.e., the transmitter inserts waveforms known by the receiver, multiplexed with the GFDM symbols. The periodicity
of the synchronization preambles is proportional to frequency precision and stability difference between transmitter and receiver time base, likewise, the channel
estimation preambles periodicity is inversely proportional to the mobile velocity.
The BSs operate in continuous transmission mode, where the carrier is always
on the air independent if useful data is available for transmission. Considering the
modulator is a wireless pipeline, it requires a constant input data rate. As the useful
transmit message rate varies over time, a rate adaption scheme is required. The
simplest solution is to fill dummy data in order to maintain a constant rate, and
remove this stuffing data at the receiver.
Polar code was derived from the channel polarization theory and introduced in
[39]. The code presents explicit construction, hardware efficient coding and decoding
algorithms, and high flexibility, namely the code rate can continuously vary from N1C
to NNC −1
, where NC is the codeword length. It makes it a strong candidate for the future
C
wireless networks, such as 5G. The implemented encoding algorithm is systematic
[40] with semi-parallel architecture [41] for improved throughput. The implemented
decoding algorithm is based on successive cancellation in the logarithm domain [42].
4 Generalized Frequency Division Multiplexing …
(a)
157
(b)
Fig. 4.43 Frame structure for a data encapsulation and; b waveforms multiplexing
R = 1/2 FP
R = 3/4 FP
R = 7/8 FP
R = 1/2 HW
R = 3/4 HW
R = 7/8 HW
10−1
Bit error rate
10−2
10−3
10−4
10−5
10−6
10−7
1
2
3
4
5
6
Eb /N0 [dB]
Fig. 4.44 Polar code bit error rate for different code rates in floating-point simulation (FP) and
hardware fixed-point implementation (HW)
A comparison between a floating-point simulation and fixed-point implementation
for the polar code is presented in Fig. 4.44.
The signal to be transmitted needs to be amplified by a power amplifier (PA).
The majority of today’s BSs employ the high electrical efficient Doherty topology [43], which is intrinsically nonlinear due to its class-C branch responsible for
amplifying signal peaks. The nonlinearity distorts the amplified signal generating
spectral regrowth, also known as intermodulation, presented at both in-band and
OOB frequencies. The in-band intermodulation affects the signal quality, and, therefore, the receiving threshold. The OOB intermodulation results in OOB emissions,
interfering in adjacent channels, and losing the benefit of using a waveform with
158
A. Nimr et al.
low-OOB emissions such as GFDM. Linearization using digital predistortion (DPD)
has been widely employed to mitigate the intermodulation [44]. The DPD system
design procedure is subdivided into three distinct tasks: (1) to choose a behavioral
model equation which is able to represent the PA characteristics, e.g., nonlinearity and memory effects, with the minimum number of coefficients; (2) to design a
real-time DPD block which is able to generate the distortion according to the model
equation and its coefficients; and (3) to select an algorithm which identifies the optimum values for the model coefficients in order to compensate for the PA distortion.
Our chosen behavioral model equation is based on the memoryless orthogonal baseband polynomial for Gaussian distributed signals [45] and modified to include the
memory as
y(n)=
M
KD
k
D −1
m=0 k=1 l=1
√ (
)
hm,k k k − 1
|x[n − m]|2(l−1) x[n − m],
l!(−1)l−k l − 1
(4.162)
where x[n] and y[n] are the input and output model signals, respectively, KD and MD
are the polynomial order and memory length, respectively, and {hm,k } are the DPD
model coefficients at the m-th tap and k-th order.
The DPD performance was tested with a gallium nitride device operating at 3 W
average, amplifying a 20 MHz wide GFDM signal centered at 723 MHz. The PA
linearization results are shown in Fig. 4.45. Table 4.4 shows the DPD performance
in terms of adjacent channel leakage rejection (ACLR).
There are impairments on the received signal caused by imperfections in the
receiver front-end analog components. It causes interference and need to be digitally
compensated in the real and imaginary parts of the received signal in three steps: (1)
remove the average value, μ; (2) remove the average correlation between parts, β;
Fig. 4.45 DPD
intermodulation reduction
performance measured with
a spectrum analyzer
−30
Power [dBm]
−40
Without DPD
With DPD
−50
−60
−70
−80
−90
673
693
713
733
Frequency [MHz]
753
773
4 Generalized Frequency Division Multiplexing …
Table 4.4 DPD performance in terms of ACLR
Lower ACLR (dB)
Without DPD
With DPD
Improvement
−36.6
−53.2
16.6
159
Upper ACLR (dB)
−37.9
−53.1
15.2
Fig. 4.46 Receiver impairments correction
and (3) equalize the power (or variance σ 2 ) difference between parts. This process
is shown in Fig. 4.46.
At the receiver side, the signal peak-to-peak voltage needs to comply with the
analog-to-digital converter (ADC) input range. If the voltage swing is too low, the
SNR is compromised by the quantization error. If the voltage swing is too high,
clipping effects will also affect the SNR. Assuming a Gaussian distribution and an
ADC with a given number of bits, it is possible to find the optimum signal amplitude
which maximizes the SNR [46]. The automatic gain control (AGC) goal is to keep
the voltage level at this optimum point as shown in Fig. 4.47.
The existing preamble-based receiver synchronization techniques for OFDM are
applicable for GFDM. For instance, a preamble with two repeated halves may be
used, where the autocorrelation
ρ[n] =
n+N
P −1
k=n
r[k]∗ r[k + NP /2],
(4.163)
can be calculated between the halves, where NP is the synchronization preamble
length. Normalizing the autocorrelation by the signal energy leads to
2 |ρ[n]|2
μS [n] = n+NP −1
.
|r[k]|2
k=n
(4.164)
160
A. Nimr et al.
Fig. 4.47 SNR due to
quantization and clipping,
and optimal operating point
120
Quantization SNR (14 bits)
Clipping SNR
Total SNR
SNR (dB)
100
80
(11.8 dB → 74 dB)
60
40
8
9
10
11
12
13
14
Clipping limit (above average power) [dB]
The presence of CP and CS produces the plateau effect which can be mitigated by
integrating (4.164) and resulting in
n
1
μM [n] =
μS [n],
L+1
(4.165)
k=n−L
where L = NCP + NCS .
The STO can also be estimated through the cross-correlation, which is given by
ρC [n] =
NP −1
1
r ′ [n + k]px∗ [k],
NP
(4.166)
k=0
where px [n] is the known preamble waveform. Finally, (4.165) and (4.166) can combined for an enhanced performance given by
μA [n] = |ρC [n]| · μM [n],
(4.167)
where the multiplication suppresses the cross-correlation side peaks, which appear
due to the repeated halves. All discussed synchronization metrics are depicted in
Fig. 4.48. The STO is estimated as the sample time index where the metric peak is.
The preamble-based channel estimation scheme is accomplished in the frequency
domain. It is straightforward since the channel estimation preamble (CEP) in the
frequency domain is known to the receiver. Considering the CEP length is usually
shorter than the GFDM symbol, the estimated channel needs to be interpolated.
However, when some subcarriers are muted, the IFFT/FFT interpolation method
4 Generalized Frequency Division Multiplexing …
161
μ S [n]
μ M [n]
| ρ C [n] |
μ A [n]= | ρ C [n] | · μ M [n]
1
Magnitude
0.8
0.6
0.4
0.2
0
50
100
200
150
250
300
Samples [n]
Fig. 4.48 Synchronization metrics
FFT
×
IFFT
Zero
Padding
Windowing
N-DFT
Smoothing
1
FFT{CEP}
Estimation
Interpolation
Fig. 4.49 Preamble-based channel estimation block diagram
fails because all the frequency components are required to calculate the interpolated
response. In order to solve this problem, a time-windowing is done on the channel
impulse response being interpolated. The detailed channel estimation block diagram
is shown in Fig. 4.49.
References
1. N. Michailow, Gaspar I. Matthé et al., Generalized frequency division multiplexing for 5th
generation cellular networks. IEEE Trans. Commun. 62, 3045–3061 (2014)
2. H. Bölcskei, F. Hlawatsch, Discrete Zak transforms, polyphase transforms and applications.
IEEE Trans. Signal Process. (1997), https://doi.org/10.1109/78.564174
3. M. Matthé et al., Generalized frequency division multiplexing in a Gabor transform setting.
IEEE Commun. Lett. (2014), https://doi.org/10.1109/LCOMM.2014.2332155
4. A. Nimr et al., Optimal Radix-2 FFT compatible filters for GFDM. IEEE Commun. Lett.
(2017), https://doi.org/10.1109/LCOMM.2017.2687926
5. X.-G. Xia, A family of pulse-shaping filters with ISI-free matched and unmatched filter properties. IEEE Trans. Commun. (1997), https://doi.org/10.1109/26.634674
162
A. Nimr et al.
6. G. Berardinelli, On the sensitivity of Zero-Tail DFT-spread-OFDM to small bandwidth allocations. IEEE Wirel. Commun. Lett. (2017), https://doi.org/10.1109/LWC.2017.2784825
7. G. Cherubini et al., Filtered multitone modulation for very high-speed digital subscriber lines.
IEEE J. Sel. Areas Commun. (2002), https://doi.org/10.1109/JSAC.2002.1007382
8. K. Jung-Hyo et al., A novel OFDM chirp waveform scheme for use of multiple transmitters in
SAR. IEEE Geosci. Remote Sens. Lett. (2013), https://doi.org/10.1109/LGRS.2012.2213577
9. S. Ehsanfar, M. Matthé, D. Zhang et al., A study of pilot-aided channel estimation in MIMOGFDM systems, in ITG/IEEE Workshop on Smart Antennas (WSA’16) (2016)
10. I. Gaspar, N. Michailow, A. Navarro et al., Low complexity GFDM receiver based on sparse
frequency domain processing, in IEEE Vehicular Technology Conference (VTC) (2013)
11. S. Ehsanfar, M. Matthé, D. Zhang et al., Interference-free pilot insertion for MIMO-GFDM
channel estimation, in IEEE Wireless Communication and Networking Conference (WCNC’17)
(2017)
12. E. U. T. R. Access, Physical channels and modulation, 3GPP TS, Vol. 36 (2009), p. V8
13. S. Alamouti, A simple transmit diversity technique for wireless communications. IEEE J. Sel.
Areas Commun. 16, 1451–1458 (1998)
14. M. Matthé, L. Mendes, I. Gaspar et al., Multi-user time-reversal STC-GFDMA for future
wireless networks. EURASIP J. Wirel. Commun. Netw. 132, 1–8 (2015)
15. L. Le, V. Lau, E. Jorswieck et al., Enabling 5G mobile wireless technologies. EURASIP J.
Wirel. Commun. Netw. (2015), https://doi.org/10.1186/s13638-015-0452-9
16. P. Banelli, S. Buzzi, G. Colavolpe et al., Modulation formats and waveforms for 5G networks:
who will be the heir of OFDM?: an overview of alternative modulation schemes for improved
spectral efficiency. IEEE Signal Process. Mag. 31, 80–93 (2014)
17. X. Zhang, M. Jia, L. Chen et al., Filtered-OFDM—enabler for flexible waveform in the 5th
generation cellular networks, in IEEE Global Communications Conference (GLOBECOM)
(2015), https://doi.org/10.1109/GLOCOM.2015.7417854
18. T. Wild, F. Schaich, Chen, 5G air interface design based on universal filtered (UF-)OFDM, in
International Conference on Digital Signal Processing (2014), https://doi.org/10.1109/ICDSP.
2014.6900754
19. M. Bellanger et al., FBMC physical layer: a primer, in The PHYDYAS project (2010), http://
ict-phydyas.org
20. A. Aminjavaheri, A. Farhang, A. Rezazadeh Reyhani et al., Impact of timing and frequency offsets on multicarrier waveform candidates for 5G, in IEEE Signal Processing and Signal Processing Education Workshop (SP/SPE) (2015), https://doi.org/10.1109/DSP-SPE.2015.7369549
21. M. Matthé, I. Gaspar, D. Zhang et al., Reduced complexity calculation of LMMSE filter coefficients for GFDM, in IEEE 82nd Vehicular Technology Conference (VTC2015-Fall) (2015),
https://doi.org/10.1109/VTCFall.2015.7391113
22. A. Farhang, N. Marchetti, L. Doyle, Low complexity GFDM receiver design: a new approach,
in IEEE International Conference on Communications (ICC) (2015), https://doi.org/10.1109/
ICC.2015.7249078
23. T. Ihalainen, A. Ikhlef, J. Louveaux et al., Channel equalization for multi-antenna
FBMC/OQAM receivers. IEEE Trans. Veh. Technol. 60, 2070–2085 (2011)
24. D. Zhang, M. Matthé, L. Mendes et al., A study on the link level performance of advanced
multicarrier waveforms under MIMO wireless communication channels. IEEE Trans. Wirel.
Commun. 16, 2350–2365 (2017)
25. M. Matthé, D. Zhang, F. Schaich et al., A reduced complexity time-domain transmitter for
UF-OFDM, in IEEE 83rd Vehicular Technology Conference (VTC Spring) (2016), https://doi.
org/10.1109/VTCSpring.2016.7504101
26. J. Abdoli, M. Jia, J. Ma, Filtered OFDM: a new waveform for future wireless systems, in IEEE
16th International Workshop on Signal Processing Advances in Wireless Communications
(SPAWC) (2015), https://doi.org/10.1109/SPAWC.2015.7227001
27. J. van de Beek, F. Berggren, N-continuous OFDM. IEEE Commun. Lett. 13, 1–3 (2009)
28. Z. Yuan, A. Wyglinski, On sidelobe suppression for multicarrier-based transmission in dynamic
spectrum access networks. IEEE Trans. Veh. Technol. 59, 1998–2006 (2010)
4 Generalized Frequency Division Multiplexing …
163
29. P. Kryszkiewicz, H. Bogucka, Out-of-band power reduction in NCOFDM with optimized
cancellation carriers selection. IEEE Commun. Lett. 17, 1901–1904 (2013)
30. Y. Li, J. Cimini, N. Sollenberger, Robust channel estimation for OFDM systems with rapid
dispersive fading channels. IEEE Trans. Commun. 46, 902–915 (1998)
31. G. Ku, J.M. Walsh, Gross, Resource allocation and link adaptation in LTE and LTE advanced:
a tutorial. IEEE Commun. Surv. Tutor. (2014), https://doi.org/10.1109/COMST.2014.2383691
32. M. Matthé et al., Asynchronous multi-user uplink transmission with generalized frequency
division multiplexing, in Communication Workshop (ICCW) (2015), https://doi.org/10.1109/
ICCW.2015.7247519
33. A.A. Zaidi, Gross et al., Waveform and numerology to support 5G services and requirements.
IEEE Commun. Mag. (2016), https://doi.org/10.1109/MCOM.2016.1600336CM
34. GFDM on GitHub, Jan 2018, https://github.com/ewine-project/Flexible-GFDM-PHY
35. The ORCA Project, Jan 2018, https://www.orca-project.eu/
36. The eWINE Project, Jan 2018, https://ewine-project.eu/
37. National Instruments reconfigurable software defined radio (USRP-RIO), Jan 2018, http://sine.
ni.com/nips/cds/view/p/lang/en/nid/212174
38. LabVIEW Communications LTE Application Framework, Jan 2018, http://sine.ni.com/nips/
cds/view/p/lang/en/nid/213083
39. E. Arikan, Channel polarization: a method for constructing capacity-achieving codes for symmetric binary-input memoryless channels. IEEE Trans. Inf. Theory (2009), https://doi.org/10.
1109/TIT.2009.2021379
40. E. Arikan, Systematic polar coding. IEEE Commun. Lett. (2011), https://doi.org/10.1109/
LCOMM.2011.061611.110862
41. G. Sarkis, I. Tal, P. Giard, A. Vardy, C. Thibeault, W.J. Gross, Flexible and low-complexity
encoding and decoding of systematic polar codes. IEEE Trans. Commun. (2016), https://doi.
org/10.1109/TCOMM.2016.2574996
42. C. Leroux, I. Tal, A. Vardy, W.J. Gross, Hardware architectures for successive cancellation
decoding of polar codes, in IEEE International Conference on Acoustics, Speech and Signal
Processing (2011), https://doi.org/10.1109/ICASSP.2011.5946819
43. R. Darraji, P. Mousavi, F.M. Ghannouchi, Doherty goes digital: digitally enhanced Doherty
power amplifiers. IEEE Microw. Mag. (2016), https://doi.org/10.1109/MMM.2016.2561478
44. J. Wood, System-level design considerations for digital pre-distortion of wireless base station
transmitters. IEEE Trans. Microw. Theory Tech. (2017), https://doi.org/10.1109/TMTT.2017.
2659738
45. R. Raich, G.T. Zhou, Orthogonal polynomials for complex Gaussian processes. IEEE Trans.
Signal Process. (2004), https://doi.org/10.1109/TSP.2004.834400
46. C. Berger, Y. Benlachtar, R. Killey, Optimum clipping for optical OFDM with limited resolution
DAC/ADC. Advanced Photonics, OSA Technical Digest, paper SPMB5 (2011)
Part II
Non-Orthogonal Multiple Access (NOMA)
in the Power Domain
Chapter 5
NOMA: An Information-Theoretic
Perspective
Mojtaba Vaezi and H. Vincent Poor
5.1 What Is Non-Orthogonal Multiple Access (NOMA)?
Multiple access lies at the heart of cellular communication systems. It refers to a
technique that allows multiple users to share a communication channel. The firstgeneration (1G) to the fourth-generation (4G) of cellular networks have adopted
radically different multiple access schemes with one common theme in mind—to
have orthogonal signals for different users at the receiver side [1]. In particular, 1G
to 4G cellular networks has adopted one or more of the following multiple access
methods:
•
•
•
•
•
Frequency division multiple access (FDMA)
Time division multiple access (TDMA)
Code division multiple access (CDMA)
Orthogonal frequency division multiple access (OFDMA)
Space division multiple access (SDMA)
For example, in OFDMA which has widely been used in 4G systems, different
users’ signals are orthogonal in the frequency and/or time domains. In other words,
one orthogonal frequency division multiplexing (OFDM) resource block (180 kHz)
cannot be allocated to more that one user.
Non-orthogonal multiple access (NOMA) [2], in contrast, allows multiple users
to share the same resource elements, be it in the time, frequency, space, or code
domain. NOMA is currently a hot research topic for 5G and beyond systems, both in
academia and industry. While it is concerned with “non-orthogonality” of multiple
access, it appears that the research community is perceiving this term in somewhat
M. Vaezi (B)
Villanova University, Villanova, PA, USA
e-mail: mvaezi@villanova.edu
H. Vincent Poor
Princeton University, Princeton, NJ, USA
e-mail: poor@princeton.edu
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_5
167
168
M. Vaezi and H. Vincent Poor
different ways. Due to the different interpretations, there is not a consensus about
applying this term to some well-known existing techniques such a CDMA. While
the majority of recent works see CDMA as an orthogonal multiple access (OMA)
technique, there are a group of other papers that categorize it as a NOMA technique.
In the following, we present and discuss different viewpoints used to define nonorthogonality in NOMA.1
NOMA Definitions and Viewpoints
• Superposition Coding with Successive Interference Cancellation: A
large body of papers consider NOMA to be equivalent to superposition coding and successive interference cancellation (SC-SIC), respectively, at the
transmitter and receivers. This is partly because the first paper using the term
NOMA considered the problem of downlink transmission with SIC [2] and
partly due to the fact that SC-SIC is the capacity-achieving technique for the
downlink channel in single-cell single-input single-output (SISO) transmission, as we discuss later in this chapter. In fact, the key term is SIC as it also
appears in the uplink transmission. With this in mind, SC-SIC is also applied
to several different cases, such as multiple-input multiple-output (MIMO)
networks and multi-cell networks (SISO or MIMO), in which SC-SIC is
suboptimal. This method is also known as power domain NOMA.
• Overloading: A second important view is to distinguish NOMA and OMA
based on the system loading. In this setting, NOMA refers to overloaded
systems and “overloading” means to have more than one user per available
resource element in the time, frequency, code, or space domain. This point
of view is rooted in CDMA systems. With this definition, a CDMA system
will be seen as a NOMA scheme if it is overloaded (i.e. when there are more
users than the number of codes) and will be considered as an OMA scheme
if there are more codes than the number of users in the systems. Examples
of the NOMA schemes developed with this view are low-density spreading (LDS) CDMA, LDS-OFDM, sparse code multiple access (SCMA), and
multi-user shared access (MUSA), which are collectively known as code
domain NOMA. This definition (overloading) can be applied to other settings such as SDMA systems. That is, similar to the CDMA case, a multi-user
MIMO (MU-MIMO) system can be seen as a NOMA or OMA scheme. The
former happens when the number of transmit antennas n t is less than the
number of users K , whereas in the latter case n t ≥ K .
• Linear Transform Decoding: Some even define NOMA based on the complexity of multi-user detection. With this point of view, in an OMA scheme,
the signals of different users can be separated in orthogonal subspaces using
a linear transform. Then, any scheme that does meet this definition can be
1 We should highlight that in this chapter we are discussing non-orthogonal multiple access methods.
This should not be confused with non-orthogonal random access which is another related topic and
will be discussed in Chap. 17 of this book.
5 NOMA: An Information-Theoretic Perspective
169
categorized as NOMA. For example, CDMA systems theoretically can be
constructed using independent random coding for different users. Such
CDMA systems are naturally non-orthogonal. Since the 1990s, there has
been a vast amount of research on the capacity of multiple access systems based on CDMA (involving power control, serial interference cancellation, dirty-paper coding, etc.)2 However, historically, CDMA implies directsequence CDMA (DS-CDMA) operation as used in 3G wideband CDMA
(WCDMA) systems. Since the 2000s, other types of CDMA schemes have
been developed, such as interleave division multiple access (IDMA) [3]. In
[3], NOMA is used to cover both DS-CDMA and IDMA. In this respect,
NOMA has a broader meaning than DS-CDMA. Therefore, when comparing CDMA and NOMA, we need to distinguish between CDMA in the
general sense, i.e. CDMA defined by information theory, and CDMA in a
narrow sense such as DS-CDMA.
• Information-Theoretic View: Looking from an information-theoretic perspective, NOMA may refer to any technique in which concurrent transmission is allowed over the same resources in time/frequency/code/space. This
achieves a better rate region when compared to orthogonalization of one or
some of the resources and includes SC-SIC but is not limited to that. In this
context, other techniques such as rate-splitting (RS) and dirty-paper coding
(DPC) can be seen as NOMA. On the other hand, CDMA techniques that
are based on orthogonality in the code domain will be seen as OMA techniques. This definition is very broad and includes many existing techniques
as a subset, and in general, its concern is to promote optimal strategies in
various uplink/downlink communication strategies.
In this chapter, NOMA refers to the last sense.3 We should highlight that the
theory of NOMA has been around for many years. In effect, the basic premise behind
NOMA is to reap the benefits promised by information theory for the downlink and
uplink transmission of wireless systems, modelled by the broadcast channel (BC) and
multiple access channel (MAC), respectively. Capacity regions of the BC and MAC
have been established several decades ago [1, 4, 5], and concurrent non-orthogonal
transmission is the optimal transmission strategy in both cases. That is, in general, to
achieve the capacity region, the users must transmit at the same time and frequency.
In particular, the capacity region of the BC is achieved using superposition coding at
the base station (BS). For decoding, the user with the stronger channel gain (usually
the one closer to the BS) uses successive interference cancellation (SIC) to decode
2 For
example, CDMA with power control is capacity achieving for the single-antenna broadcast
channel.
3 However, in the SDMA case, we assume that the system is overloaded. This is relevant in the
MIMO (and MU-MIMO) case where “space” comes in as a new resource. We note that when the
number of transmit antennas n t is greater than the number of users K , different users can be served
in different (orthogonal) spaces, reducing the underlying system to an OMA one.
170
M. Vaezi and H. Vincent Poor
User 2 signal
detecƟon
SIC of User 2
signal
Fig. 5.1
users
User 1 signal
detecƟon
Illustration of single-cell downlink NOMA using power domain multiplexing for two
its signal free of interference, while the user with the weaker channel gain treats
the signal of the stronger users as noise. This is illustrated in Fig. 5.1. Similarly, to
get the highest achievable region in multi-cell systems, concurrent non-orthogonal
transmission is still required [5–7], and orthogonal transmission is suboptimal.
5.2 What Drives NOMA?
The next generation of wireless networks must support very high throughput, low
latency, and massive connectivity. According to the international telecommunication
union (ITU) [8], 5G networks must fulfill several requirements including:
(1) a minimum peak data rate of 10 Gbps (100 times more than that in the 3rd
Generation Partnership Project (3GPP) Long-Term Evolution (LTE))
(2) a latency of 1ms (ten times lower than that in 4G networks), and
(3) a connection density of 1,000,000 devices per km2 (100 times more than 4G
networks).
Effective multiple access is a key enabler in achieving these requirements. As
noted above, the first to the fourth-generations of cellular networks have adopted different multiple access schemes with one common theme in mind—to have orthogonal
signals for different users at the receiver side. By allowing multiple users to share all
domains (e.g. time, frequency, space), NOMA can address the above challenges of the
next generation of wireless networks more efficiently than the conventional orthogonal multiple access schemes. NOMA can increase spectral efficiency and userfairness by exploiting a capacity-achieving scheme in the downlink. It can support
5 NOMA: An Information-Theoretic Perspective
171
more connections in the uplink by letting multiple users simultaneously access the
same wireless resources, which, in turn, can reduce latency [9–12].
Code domain NOMA4 uses user-specific sequences for sharing the entire radio
resource. On the other hand, power domain NOMA exploits the channel gain differences between the users for multiplexing via power allocation. In this chapter, we
study NOMA in the power domain, a technique that can improve wireless communication through the following benefits:
• Massive Connectivity: There appears to be a reasonable consensus that NOMA
is essential for massive connectivity. This is because the number of served users
in all OMA techniques is inherently limited by the number of resources (e.g. the
number of codes in CDMA and the number of resource blocks in OFDMA). In
contrast, by superimposing all users’ signals, NOMA theoretically can serve an
arbitrary number of users even within one resource block. In this sense, NOMA
can be tailored to Internet of Things (IoT) applications where a large number of
devices sporadically transmit a small number of packets. In fact, allocating an
entire resource block (180 kHz in LTE) to one device, as is done in OMA, is
extremely inefficient.
• Low Latency: Latency requirements for 5G applications are rather diverse and
very stringent in some cases. For example, ITU requires a user plan latency of
4ms and 1ms for enhanced mobile broadband (eMBB) and ultra-reliable and lowlatency communications (URLLC), respectively [8]. With OMA, it is very difficult
to guarantee such stringent delay requirements because no matter how many bits a
device wants to transmit the device must wait until an unoccupied resource block
becomes available. In contrast, NOMA supports flexible scheduling since it can
accommodate a variable number of devices depending on the application that is
being used and the perceived quality of service (QoS) of the device.
• High Spectral Efficiency: According to ITU requirements for IMT-2020, downlink peak spectral efficiency should be 30 bits/s/Hz. NOMA offers a higher spectral efficiency and a better user-fairness compared to OMA. As will be seen in
Sect. 5.3, NOMA is the theoretically optimal way of using the spectrum for both
uplink and downlink communications in a single-cell network. Such better performance is achieved due to the fact that every NOMA user can enjoy the whole
bandwidth, whereas OMA users are limited to a smaller fraction of the spectrum
which is inversely proportional to the number of users. NOMA can also be combined with other emerging technologies, such as massive MIMO and millimeter
wave (mmWave) technologies, to further improve spectral efficiency and support
higher throughput.
4 Sparse
code multiple access (SCMA) is an important variant of NOMA in the code domain.
172
M. Vaezi and H. Vincent Poor
In consideration of the above benefits, NOMA has drawn significant attention from
both academia and industry during past a few years. However, as we briefly mentioned
earlier and will see in detail in this chapter, from an information-theoretic perspective
the theory behind NOMA has been established for several decades. Nevertheless,
despite this insight from information theory, orthogonal multiple access techniques
have been used in the cellular networks from 1G to 4G. This was mainly to avoid interuser interference cancellation which would have resulted in unacceptably complex
receivers.
Today, with the advances in processing power, the implementation of interference cancellation at user equipment has been made practical. For example, a
category of relatively complex user terminals, known as network-assisted interference cancellation and suppression (NAICS) terminals, has recently been
adopted in 3GPP LTE-A. Such technological advances, in conjunction with
the need to support exponentially increasing numbers of devices and better
spectral efficiency, have motivated a new wave of research on NOMA. Saito et
al. first showed that NOMA can improve system throughput and user-fairness
over a SISO channel using OFDMA [2]. The spectral efficiency of NOMAbased systems is further boosted when combined with MIMO communication
[13–15]. Successful operation of this technique, however, depends on knowledge of the channel state information (CSI) between the BS and the end-users.
More practical solutions, e.g. those with limited and delayed CSI, are crucial in
making NOMA workable. Today, a variation of NOMA, known as multi-user
superposition transmission [12], is considered for the 3GPP LTE-A systems.
While recent advances in processor capabilities have made SIC, and consequently
NOMA, feasible, significant research challenges remain to be addressed before
NOMA can be deployed. In addition to the above practical issues, NOMA-based
transmission introduces new security and privacy challenges. This is because in
NOMA-based transmission a user with a better channel is able to decode the other
user’s signal. Even, a user with a weaker channel can also partly decode the stronger
user’s signal. On top of that, wireless transmission is naturally vulnerable to external
eavesdroppers. Although upper-layer security approaches (e.g. cryptography) can be
used to secure transmissions, there are numerous risks in cryptographic methods due
to the rapid advancement of computing technologies. Also, cryptographic approaches
require a key management infrastructure which should be secured, in turn. Moreover, traditional key agreement algorithms are not suitable for many existing and
emerging wireless networks, such as ad hoc networks and IoT, since they consume
scarce resources such as bandwidth and battery power.
5 NOMA: An Information-Theoretic Perspective
173
5.3 Theory Behind NOMA
Analysis of cellular communication can generally be classified as either downlink or
uplink. In the downlink channel, the BS simultaneously transmits signals to multiple
users, whereas in the uplink channel multiple users transmit data to the same BS.
As noted above, from an information-theoretic perspective, the downlink and
uplink are modelled by the broadcast channel and multiple access channel, respectively. The basic premise behind single-cell NOMA in the power domain is to reap
the benefits promised by the theory of multi-user channels [1, 5]. As such, we review
what information theory promises for these channels, both in the single-cell and
multi-cell settings. In particular, we seek to answer the following two questions in
this section: (1) What are the highest achievable throughputs for these multi-user
channels? and (2) how can a system achieve such rates?
5.3.1 Single-Cell NOMA
For simplicity of illustration, we first consider a network consisting of a single cell.
In addition, we assume that there are only two users in that cell. Later, we discuss
the general case with multiple cells each consisting of multiple users. For OMA, we
consider a TDMA technique5 where a fraction α of the time (0 ≤ α ≤ 1) is dedicated
to user 1 and a fraction ᾱ 1 − α of the time is dedicated to user 2.
The capacity regions of the two-user MAC and BC are achieved via NOMA,
where both users’ signals are transmitted at the same time and in the same frequency
band [5]. To gain more insight, we describe how these regions are obtained.
Throughout the chapter, we use
C (x)
1
log2 (1 + x),
2
(5.1)
and γi = |h i |2 P is the received signal-to-noise ratio (SNR) for user i, where h i is
the channel gain, P is the transmitter power, and the noise power is normalized to
unity.
5.3.1.1
Two-User MAC (Uplink)
The capacity regions of the two-user MAC is achieved via non-orthogonal transmission in which both users’ signals are transmitted at the same time and in the same
frequency band [5]. The curves labelled NOMA in Fig. 5.2 represent the capacity
regions of the MAC for different values of P1 and P2 , which are the respective transmit powers of the two users. From these figures, it can be seen that except for a
few points OMA is strictly suboptimal. One of these points is the sum capacity of
5 FDMA
has exactly the same performance [5].
174
M. Vaezi and H. Vincent Poor
NOMA
OMA−PC
OMA
R
2
R2
NOMA
OMA−PC
OMA
R
R
1
1
(a) P1 = 5, P2 = 10
(b) P1 = 10, P2 = 10
Fig. 5.2 Best achievable regions by OMA and NOMA in the two-user MAC (uplink) for different
values of the two users’ transmit powers P1 and P2
the MAC. This means that both OMA and NOMA can achieve the sum capacity of
MAC.
To gain more insight, we describe how the above regions are obtained. Using
OMA, each user sees a single-user channel in its dedicated fraction of the time. As
a result
R1 = αC (γ1 ),
R2 = ᾱC (γ2 ),
(5.2a)
(5.2b)
are achievable. If power control (PC) is applied, these rates can be increased to
R1 = αC
γ
1
,
γα
2
R2 = ᾱC
.
ᾱ
(5.3a)
(5.3b)
This region is labelled NOMA-PC in Fig. 5.2a, b.
In the case of NOMA, both users concurrently transmit, and their signals interfere
with each other at the BS. The BS can use SIC to achieve any point in the NOMA
region, which is the capacity region of this channel [1].
Theorem 1 The capacity region of the two-user MAC is the set of nonnegative
(R1 , R2 ) such that
R1 ≤ C (γ1 ),
R2 ≤ C (γ2 ),
(5.4a)
(5.4b)
R1 + R2 ≤ C (γ1 + γ2 ).
(5.4c)
5 NOMA: An Information-Theoretic Perspective
175
Particularly, to achieve the right corner point of the capacity region (in which
R1 = C (γ1 )), the BS first decodes user 2’s signal, treating the other signal as noise.
2
). The BS then removes user 2’s signal and decodes
This results in R2 = C ( γ1γ+1
user 1’s signal free of interference; i.e. R1 = C (γ1 ). As a result, the sum capacity
R1 + R2 = C (γ1 + γ2 ) is achievable. To achieve the other corner point, the order
of decoding needs to be changed. From Fig. 5.2, it is seen that the gap between the
NOMA and OMA regions becomes larger if power control is not used in OMA.
5.3.1.2
Two-User BC (Downlink)
Similar to the two-user MAC, the capacity region of the two-user BC is known
and is achieved via non-orthogonal transmission in which both users’ signals are
transmitted at the same time and in the same frequency band [5]. The curves in
Fig. 5.3a–c represent the capacity regions of the BC as well as the best achievable
regions obtained by OMA for different values of channel gains. From these figures, it
can be seen that except for a few points, OMA is strictly suboptimal in the downlink.
In fact, OMA can only achieve
(5.5a)
(5.5b)
R1 = αC (γ1 ),
R2 = ᾱC (γ2 ).
However, making use of a NOMA scheme can strictly increase this rate region
as shown in Fig. 5.3. In particular, the capacity region of this channel is known and
can be achieved using superposition coding at the BS and successive interference
cancellation at the receiver. For decoding, the user with the stronger channel uses
SIC to decode its signal free of interference at a rate of R1 = C (βγ1 ) while the other
NOMA
OMA
R1
(a) P1 = 1, P2 = 10
2
NOMA
OMA
R
R
2
R2
NOMA
OMA
R1
(b) P1 = 5, P2 = 10
R
1
(c) P1 = 9, P2 = 10
Fig. 5.3 Best achievable regions by OMA and NOMA in the BC (download) for different values
of P1 and P2
176
M. Vaezi and H. Vincent Poor
2
user is capable of decoding at a rate of R2 = C ( βγβ̄γ2 +1
) where β is the fraction of the
BS power allocated to user 1’s data and β̄ = 1 − β. By varying β from 0 to 1, any
rate pair (R1 , R2 ) on the boundary of the capacity region of the BC (NOMA region)
can be achieved. That is,
Theorem 2 The capacity region of the two-user BC is the set of nonnegative (R1 , R2 )
such that
R1 ≤ C (βγ1 ),
β̄γ2
,
R2 ≤ C
βγ2 + 1
(5.6a)
(5.6b)
in which β ∈ [0 1] and β̄ = 1 − β.
The fact that the capacity region of downlink NOMA is known enables us to find
the optimum power allocation corresponding to any point (R1 , R2 ) on the boundary
of the capacity region. In fact, all we need to know to achieve such a rate pair is to
find what fraction of the BS power should be allocated to each user. Corresponding
to each (R1 , R2 ), there is a 0 ≤ β ≤ 1 such that β P and β̄ P are the optimal powers
for user 1 and user 2, respectively, where P is the BS power. Conversely, every β
generates a point on the boundary of the capacity region.
The above argument implies that NOMA can improve user-fairness smoothly
and in an optimal way by flexible power allocation. Suppose that a user has a poor
channel or it has not been served for a long time (in OMA). To boost this user’s rate
and improve user-fairness, the BS can simply increase the fraction of power allocated
to this user. We can look at this problem from yet another perspective. To increase
the rate of such a user, we can maximize the weighted sum-rate µR1 + R2 where
a high weight (µ) is given to such a user. This is because, to maximize µR1 + R2
for any µ ≥ 0, there exists an optimal power allocation strategy, determined by β.
Seeing that µ > 1 (µ < 1) corresponds to the case where user 1 has higher (lower)
weight than user 2, to improve user-fairness, we can assign an appropriate weight to
the important user and find the corresponding β.
5.3.1.3
K -User Uplink/Downlink
In the above, we described coding strategies for the two-user uplink/downlink channels. Interestingly, very similar coding schemes are still capacity achieving for the
K -user MAC and BC, as described in the following.
K -User MAC: To achieve the capacity region of the K -user MAC, the users
transmit their signals concurrently and the BS applies SIC decoding. Specifically,
we have [5].
Theorem 3 The capacity region of the K -user MAC is the set of nonnegative
(R1 , . . . , R K ) such that
5 NOMA: An Information-Theoretic Perspective
j∈S
Rj ≤ C
γj
j∈S
177
for every S ⊆ [1 : K ].
(5.7)
For example, for K = 3, the capacity region is the set of nonnegative (R1 , R2 , R3 )
such that
R1 ≤ C (γ1 ),
(5.8a)
R2 ≤ C (γ2 ),
R3 ≤ C (γ3 ),
R1 + R2 ≤ C (γ1 + γ2 ),
(5.8b)
(5.8c)
(5.8d)
R1 + R3 ≤ C (γ1 + γ3 ),
R2 + R3 ≤ C (γ2 + γ3 ),
(5.8e)
(5.8f)
R1 + R2 + R3 ≤ C (γ1 + γ2 + γ3 ).
(5.8g)
The capacity-achieving scheme is based on non-orthogonal transmission that allows
multiple users to transmit at the same time and frequency. To be specific, the capacity
region is achieved by point-to-point codes, successive cancellation decoding, and
time-sharing. Again, OMA is strictly suboptimal [1].
K -User BC: To achieve the capacity region of the K -user BC, the users’ signals
are superimposed at the BS and transmitted altogether. Without loss of generality
assume that γ1 ≥ γ2 ≥ · · · ≥ γ K . At the receiver sides, receiver k ∈ [1, . . . , K ] first
decodes the signal of users k + 1, . . . , K and removes them from the received signals.
It (receiver k) then decodes its own signal treating users’ 1, . . . , k − 1 signals as noise.
As a result, we obtain
Theorem 4 The capacity region of the K -user BC is the set of nonnegative
(R1 , . . . , R K ) such that
Rk ≤ C
1+
in which k ∈ [1, . . . , K ], β j ≥ 0 ∀ j and
βk γk
k−1
j=1
K
j=1
β j γk
,
(5.9)
β j = 1.
Remark 1 To achieve the above rates, it is important to note that the single-antenna
K -user Gaussian BC is a set of degraded channels. This implies that the users can
be ordered based on their channel strengths (for example, in Theorem 4, we have
assumed that γ1 ≥ γ2 ≥ · · · ≥ γ K ).
We note that due to the SIC decoding, the user with the strongest channel (user 1) is
able to decode its own signal free of interference, whereas the user with the weakest
channel (user K ) has to treat all other users’ signals as noise. For example, for K = 3
the capacity region is the set of nonnegative (R1 , R2 , R3 ) such that
178
M. Vaezi and H. Vincent Poor
R1 ≤ C (β1 γ1 ),
β2 γ2
R2 ≤ C
,
1 + β1 γ2
β3 γ3
R3 ≤ C
.
1 + (β1 + β2 )γ3
(5.10a)
(5.10b)
(5.10c)
So far, we have assumed a network with a single cell. In fact, most of the work on
NOMA is limited to single-cell analysis, where there is no co-channel interference
caused by an adjacent BS. However, to verify the benefits of NOMA in a more
realistic setting, it is necessary to consider a multi-cell network. Specifically, as
wireless networks get denser and denser, inter-cell interference (ICI) becomes a
major obstacle to achieving the benefits of NOMA. In the next section, we discuss
the theory behind NOMA in multi-cell networks.
5.3.2 Multi-Cell NOMA
In a multi-cell setting, finding the best achievable region is much more involved than
in the single-cell case, and simple channel models are insufficient. A few models,
including the interference channel, interfering MAC, and interfering BC can be used
to model multi-cell networks. Unfortunately, the capacities of these channels are
unknown in general. However, the known achievable rate regions for these channels
indicate the superiority of NOMA to OMA.
5.3.2.1
Interference Channel (IC)
The capacity region of the two-user IC is not known in general; however, it is known
that OMA is strictly sub-optimal. Han and Kobayashi introduced an achievable region
in 1981 [6], which is still the best known inner bound for the general interference
channel. In the Han-Kobayashi (HK) scheme, each user can split its message to be
sent into two submessages of smaller rates and power. These are known as private
and common messages; the former is intended to be decoded only at the respective
receiver, whereas the latter can be decoded at both receivers. The rationale behind
this coding scheme is to decode part of the interference (the common message) and
treat the rest as noise. The optimal input distributions are not known for the HK
region. As such, commonly a subset of the HK region with Gaussian codebooks is
used to represent the HK region for the Gaussian channel; see, e.g. [16–19].
The basic HK scheme employs rate-splitting (RS) and superposition coding (SC)
at each transmitter and SIC at the receiver. Since both transmitters send their signals
concurrently at the same frequency, the HK scheme is a NOMA scheme. Flexibility in splitting each user’s transmission power into the common/private portions
of information and time-sharing between them make the HK scheme very strong, but
5 NOMA: An Information-Theoretic Perspective
179
complicated. Not surprisingly, though, the optimal HK strategy is not well-understood,
in general. Nevertheless, the HK scheme is known to be within 21 bit of the capacity
region of the two-user Gaussian interference channel [16]. Although this gap could
be due to a suboptimal scheme, loose outer bounds, or both, the outer bounds seem
to be the most crucial.
In general, the HK scheme applies time-sharing to improve the basic HK region
and can be seen as a combination of NOMA and OMA [7]. Although the basic HK
scheme is optimal for the strong and very strong interference regimes, when one
interference link is strong and the other one is weak, time-sharing can enlarge the
basic HK region in general. Specifically, we have [7, 20]
Lemma 1 The HK achievable region for the one-sided IC is the set of rate pairs
(R1 , R2 ) satisfying
R1 ≤ λ1 R11 ,
R2 ≤ λ1 R21 + λ2 R22 ,
(5.11a)
(5.11b)
in which
P1
λ1
R11 ≤ γ
1 + aβ1 P21
R21 ≤ γ
P1
λ1
a β̄1 P21
1+
(5.12a)
,
+ aβ1 P21
+ γ (β1 P21 ),
R22 ≤ γ (P22 ),
(5.12b)
(5.12c)
where λ1 + λ2 = 1, λ1 P21 + λ2 P22 = P2 , 0 ≤ β1 ≤ 1, and β̄1 = 1 − β1 .
It is worth mentioning that Lemma 1 characterizes a strictly better region when
compared with the HK region without time-sharing. In particular, for λ1 = 1, this
lemma reduces to the HK region without time-sharing (i.e., the basic HK region).
These regions are compared in Fig. 5.4. Also, substituting P21 = 0, Lemma 1 reduces
to
P1
,
(5.13a)
R 1 ≤ λ1 γ
λ1
P2
,
(5.13b)
R2 ≤ (1 − λ1 )γ
1 − λ1
for 0 ≤ λ1 ≤ 1. This region is known as the TDMA (or FDMA) region. The main
difference between the TDMA and time-sharing regions is in the fact that in the
TDMA approach only one user is allowed to transmit during each subband, while in
time-sharing method, both users can transmit in the same subband (e.g. during λ1 in
the region defined by Lemma 1).
M. Vaezi and H. Vincent Poor
R
R
2
2
180
TS
NOMA
OMA
R1
(a) P1 = 1, P2 = 3
TS
NOMA
OMA
R1
(b) P1 = 3, P2 = 3
Fig. 5.4 Best achievable regions by OMA, NOMA, TS (applying time-sharing to OMA and NOMA
in different time slots) in the IC (multi-cell network) for different values of P1 and P2
In other words, the HK scheme that combines NOMA and OMA gives the largest
rate region [7], as shown in Fig. 5.4 for different values of P1 and P2 . In these plots,
OMA refers to TDMA, whereas NOMA refers to the basic HK scheme in which
time-sharing is not applied. In this case, both transmitters send their messages using
the HK strategy but in one time-slot only. The third curve, labelled TS, is based on the
HK scheme with time-sharing (TS) in which two time slots are used: in one time-slot,
both users are active, while in the other time-slot, only one of them is transmitting. As
can be seen from this figure, both NOMA and OMA are suboptimal when compared
with the case where NOMA and OMA are combined. We should highlight that the
rate region obtained via the HK scheme without time-sharing (NOMA) is however
very close to that with time-sharing (OMA + NOMA). That is, NOMA yields a good
approximation of the best achievable region for this channel, for the two-user IC.
5.3.2.2
Interfering MAC and BC
Consider a mutually interfering two-cell network in the uplink, where each cell
includes one MAC. Assume that only one of the transmitters of each MAC (typically
the closest one to the cell-edge) is interfering with the BS of the other MAC. In this
network, the interfering transmitters can employ HK coding, similar to that used
in the IC, while the non-interfering transmitters in each MAC employ single-user
coding. This NOMA-based transmission results in an inner bound which is within a
one-bit gap of the capacity region [21]. Likewise, one can use the interfering BC to
model a mutually interfering two-cell downlink network.
5 NOMA: An Information-Theoretic Perspective
181
Despite years of intensive research in information theory, finding optimal
uplink and downlink transmit/receive strategies for multi-cell networks remains
rather elusive. In fact, as discussed earlier, even for a much simpler case of the
two-user IC, the optimal coding strategy is still unknown. Nonetheless, fundamental results from information theory as a whole suggest that NOMA-based
techniques result in a superior rate region when compared with OMA.
It should be highlighted that, despite the above insight from information theory,
OMA techniques have been used in the cellular networks from 1G to 4G, mainly to
avoid interference and due to its simplicity. In addition, the lack of understanding
of optimal strategies for multi-cell networks has motivated pragmatic approaches in
which interference is simply treated as noise.
5.3.3 NOMA in MIMO Networks
With the rapid advancement of multi-antenna techniques, today wireless nodes (particularly BSs) are often equipped with more than one antenna. Multi-antenna systems
create spatial dimension which, in turns, opens the door to SDMA. In SDMA, multiple users can communicate at the same time-frequency but are distinguished in space.
That is, each user has a different beam. Then, in the downlink and uplink of singlecell networks, we will have the MIMO-BC and MIMO MAC. In addition, similar
to the multi-cell SISO networks, we can use MIMO IC to model a multi-antenna
multi-cell network. Although in some cases the capacity region of these channels is
unknown in general, similar to the SISO case, it is known from information theory
that TDMA (OMA) is suboptimal and concurrent transmission (NOMA) can result
in a better rate region. We discuss this in the following sections.
5.3.3.1
MIMO BC
Unlike the Gaussian SISO BC, the Gaussian MIMO BC is non-degraded in general. That is, in general, the MIMO BC users cannot be ordered based on their
channel strengths. This is because the users’ channels are matrices (or vectors), and
matrices cannot be ordered in general, unlike the scalars in the SISO BC. This has
been the main difficulty in proving the capacity of the MIMO BC. However, the
capacity region of the MIMO BC has recently been established in [22] in which
it is proved that the capacity region can be achieved by DPC [23]. Because of
the non-degradedness of the MIMO BC, SC-SIC which is the capacity-achieving
technique for the SISO BC is not capacity achieving for the MIMO BC. While
in SC-SIC interference cancellation is performed at the receiver side, DPC constitutes interference cancellation at the transmitter side and consequently the transmitter requires knowledge of CSI. The same strategy is also the only technique that
182
M. Vaezi and H. Vincent Poor
achieves the capacity region of the multiple-input single-output (MISO) BC. DPC is a
non-orthogonal transmission (NOMA) strategy, and it is well-understood that TDMA
(OMA) is suboptimal for the MIMO BC and the MISO BC.
DPC is, however, prohibitively complex for practical systems. Practical MIMO
systems usually use linear precoding to simplify the transmitter design. This technique creates different beams for different users (or sets of users) and allocates a
fraction of the total transmit power to each beam. At the receiver side, the interference from other users is treated as noise. Due to its simplicity, SDMA is usually
implemented using linear precoding. This access technique is the basic principle
behind several well-established techniques in 4G and upcoming 5G networks. The
examples include but are not limited to multi-user MIMO, network MIMO, coordinated multipoint (CoMP), millimeter-wave MIMO, and massive MIMO [24].
Another line of research seeks to extend the SISO NOMA principles to MIMO
NOMA transmission. In fact, the performance of NOMA can be further boosted in
multi-antenna networks. MIMO NOMA solutions exploit multiplexing and diversity
gains to improve outage probability and throughput, by converting the MIMO channel
into multiple parallel channels. One approach is to directly apply SISO NOMA
methods by making the MIMO NOMA networks degraded. That is to order the users
based on an effective scalar channel and decode their messages using SIC. A second
approach combines SDMA with superposition coding at the receiver and SIC at the
transmitter. Such solutions try to allocate different beams to each group of users and
then to use SISO NOMA within each group. By allocating different beams to each
cluster (group), the interference between the clusters can be managed and removed.
Within each cluster, the SISO NOMA solutions are then applied.
SDMA and MIMO NOMA can both be viewed as superpositions in the power
domain with different approaches at the receiver side. In the former, the users are
separated spatial beamformers, whereas in the latter case SIC is used at the receivers
to separate the users [14]. In other words, SDMA fully treats the interference as
noise whereas NOMA with SIC fully decodes interference. Recalling the flexible
rate-splitting in the HK scheme for the IC, in which a user’s transmission power is
split between the common/private portions of information and part of the interference
(the common message) is decoded while treating the remainder as noise, makes the
achievable region larger, researchers have applied the same concept to the MIMO
BC channel. RS obviously can bridge between the SDMA and NOMA with SIC by
enabling a receiver to decode part of the interference and treat the remaining part
of that as noise. In this sense, SDMA and NOMA with SIC can be seen as two
extreme cases of RS. All of these approaches suggested by information theory are
based on the concurrent transmission of multiple users signal without making them
orthogonal and thus can be seen as different variants of NOMA. To summarize, from
information-theoretic results, it is clear that OMA is suboptimal in MIMO networks
too. This is valid both for single-cell and multi-cell (network MIMO) cases. The best
techniques are either based on NOMA or a combination of NOMA and OMA using
time-sharing. Several other techniques such as superposition coding, SIC, RS, and
time-sharing fall within these NOMA schemes.
5 NOMA: An Information-Theoretic Perspective
5.3.3.2
183
MIMO IC
The capacity region of the MIMO IC, similar to many other multi-user networks,
is unknown. Finding the exact capacity region has been quite challenging for those
channels. Because of this, approximations are widely used to get insight into the
behavior of these channels. One commonly used approximation metric is degrees of
freedom (DoF), or multiplexing gain. The DoF gives the pre-log of the capacity of a
given channel in the high SNR regime; i.e.
DoF =
lim
S N R→∞
C (S N R)
.
log(S N R)
(5.14)
For example, the DoF of the MIMO channel with M and N antennas at the
transmitter and receiver is min(M, N ) which is its multiplexing gain. The DoF of
the MIMO broadcast channel with M antennas at the transmitter and N1 and N2
antennas at the receivers is min(M, N1 + N2 ). Interference alignment (IA) is the
main technique used to achieve the degrees of freedom of interferences channels.
IA was introduced by Maddah-Ali et al. in the context of the MIMO X channel
[25, 26], where an iterative achievable scheme for this channel built upon dirtypaper coding and successive decoding was introduced. IA refers to a mechanism for
aligning (overlapping) interference spaces. It was then applied to the K -user SISO IC
in [27] leading to the surprising conclusion that wireless networks are not essentially
interference limited.
The DoF of the K -user SISO IC is equal to K2 [27], meaning that each user can
enjoy half of the spectrum in the high SNR regime. The DoF of the K -user user
MIMO Gaussian IC with M antennas at each transmitter and N antennas at each
receiver is also known [28, 29] and is obtained using IA. All these results show the
suboptimality of orthogonalization. For example, in the three-user MIMO IC with
two antennas at each node, IA allows a total of three DoF, whereas the schemes
based on orthogonalization can allow a maximum of two DoF per channel use. In the
orthogonalized solutions, e.g. TDMA, only one user will be active at a time and can
send two symbols on its 2 × 2 MIMO channel, whereas IA allows each user to send
only one information symbol, but all the users are active at all times. These results
show that the DoF of the K -user MIMO IC is achieved when all the users are active
at all times and orthogonalization is suboptimal. In other words, NOMA performs
better than OMA.
5.4 Moving from Theory to Practice
As explained in Sect. 5.3, the basic theory behind NOMA has been around for many
years. Specifically, the capacity region of SISO NOMA in the single-cell setting
has been known for several decades. Moreover, it is known that to get the highest
achievable region in multi-cell systems concurrent non-orthogonal transmission is
184
M. Vaezi and H. Vincent Poor
required [5–7], and orthogonal transmission is suboptimal. This conclusion is based
on the exact and approximated (DoF) analysis, as discussed in the previous section.
Furthermore, the theoretical results for MIMO networks (MIMO BC, MIMO MAC,
and MIMO IC) on the whole indicate that OMA is suboptimal and NOMA, with
sophisticated schemes such as DPC and IA, can achieve better rates.
Despite these insights from information theory, orthogonal multiple access
techniques have been used in the cellular networks from 1G to 4G. This has
mainly been to avoid inter-user interference cancellation which would have
resulted in unacceptably complex receivers. In fact, schemes such as SIC and
DPC used for interference cancellation are very complex to implement in user
equipment. Today, with the advances in processing power driven by Moore’s
law,6 the implementation of interference cancellation at user equipment has
been made practical, as discussed in Sect. 5.4.1. Such technological advances,
in conjunction with the need to support exponentially increasing numbers of
devices and better spectral efficiency, have motivated a new wave of research
on NOMA.
5.4.1 SIC in 4G Networks
Owing to advances in processing power, interference cancellation on user equipment
has become more practical and is realized in LTE-A networks in a few different
settings. Some of these are listed below.
• Network-Assisted Interference Cancellation and Suppression (NAICS): NAICS
refers to a category of relatively complex user terminals that has recently been
adopted by 3GPP LTE-A. Network-assisted interference cancellation/suppression
can enable more effective interference cancellation/suppression at the user-side
with possible network coordination [30]. In developing NAICS, an extensive study
was done on advanced receivers with various capabilities of interference cancellation/suppression. As an example, single-user MIMO with minimum mean square
error successive interference cancellation (MMSE-SIC) has been designed in LTE
Release 8 [31].
• Multi-User Superposition Transmission (MUST): MUST is a recent proposal
for 3GPP for downlink mobile broadband (MBB) services [12]. MUST has different categories corresponding to different transmitting schemes [32]. Although
the interference scenarios are not the same in NAICS and MUST, many of the
receivers proposed for NAICS can also be used for MUST.
6 Moore’s
Law, hypothesized by Intel founder Gordon Moore in 1965, states that the number of
transistors in a dense integrated circuit will double approximately every two years. This enables
a larger number of transistors to be concentrated in a given area which, in turn, results in a faster
processor that can operate at lower power.
5 NOMA: An Information-Theoretic Perspective
185
Saito et al. first showed that NOMA can improve system throughput and user-fairness
over a SISO channel using OFDMA [2]. The spectral efficiency of NOMA-based
systems is further boosted when combined with MIMO communication [13–15].
Successful operation of this technique, however, depends on knowledge of the CSI
between the BS and the end-users. More practical solutions, e.g. those with limited
and delayed CSI, are crucial to making NOMA workable. Today, a variation of
NOMA, known as MUST [12], is being considered for use in 3GPP LTE-A systems,
as noted above.
5.4.2 Multi-Cell NOMA Solutions
Inter-cell interference reduces a cell-edge user’s performance and is the main issue
in multi-cell networks. ICI management approaches are used to improve cell-edge
users’ performance. Depending on the availability of the data messages desired at the
users among multiple BSs, multi-cell techniques can be categorized into coordinated
scheduling/beamforming (CS/CB) and joint processing (JP) [33]. Specifically, in
CS/CB, data for a user is only available at and transmitted from a single BS, whereas
in JP, the data is shared among multiple BSs. ICI management approaches can be
combined with NOMA resulting in multi-cell NOMA solutions, e.g. NOMA-JP and
NOMA-CS/CB.
In NOMA-JP, the users’ data symbols are available at more than one BS. NOMAJP has two main categories: NOMA-joint transmission (JT) and NOMA-dynamic
cell selection (DCS). In the former, multiple BSs are active and simultaneously
serve a cell-edge user using a shared wireless resource. NOMA-JP can significantly
improve the quality of the signals received by cell-edge users as the two BSs cooperate
instead of interfering with each other. In NOMA-DCS, although the user’s data is
shared among multiple BSs, this data is transmitted only from one BS at a time. The
transmitting BS can be dynamically changed over time.
In NOMA CS/CB, the users’ data is not shared among two or more cooperating
BSs. Specifically, in NOMA CB, the beamforming decision is made with coordination of other BSs. As an example, IA-based CB is applied in [34] in which two
BSs jointly optimize their beamforming vectors in order to improve the data rates
of cell-edge users by removing ICI. On the other hand, in NOMA-CS multiple BSs
coordinate scheduling to serve NOMA users with less ICI. The cooperating BSs in
NOMA CS/CB need to exchange global CSI and cooperative scheduling information
via a standardized interface named X2 which may result in considerable overhead
especially when the users are highly mobile. A review of different multi-cell NOMA
techniques can be found in [11].
While recent advances in processing capabilities have paved the way for SIC, and
consequently NOMA, significant research challenges remain to be addressed before
NOMA can be deployed. In addition to the above practical issues, NOMA-based
transmission may introduce new security and privacy challenges. We introduce the
channel modes relevant to the physical layer security of NOMA in the next section.
186
M. Vaezi and H. Vincent Poor
5.5 Physical Layer Security in NOMA
In a NOMA-based transmission, a user with a better channel is capable of decoding
the other user’s signal. Even, a user with a weaker channel can also partly decode
the stronger user’s signal. This may introduce new security and privacy challenges.
Moreover, wireless transmission is naturally vulnerable to external eavesdroppers.
Although upper-layer security approaches (e.g. cryptography) are still relevant since
only the legitimate user has a key to decode its message, there are numerous risks
in cryptographic methods due to the rapid advancement of computing technologies.
Besides, cryptographic approaches require a key management infrastructure which
should be secured, in turn. Moreover, traditional key agreement algorithms are not
suitable for many existing and emerging wireless networks, such as ad hoc networks
and IoT, since they consume scarce resources such as bandwidth and battery power.
Considering these challenges, physical layer security schemes are of interest.
5.5.1 Description of the Channel Models
The goal of this section is to identify and leverage information-theoretic channel
models for securing communication in the context of NOMA. Three basic channel
models are considered as depicted in Figs. 5.5, 5.6 and 5.7. In these figures, n t , n 1 ,
n 2 , and n e are the number of antennas at the transmitter (Tx), legitimate receivers
(Rx1 and Rx2), and eavesdropper (Eve), respectively. For the purpose of illustration,
we only consider the case of two legitimate receivers. These models can be used for
NOMA with two users. A more general setting is the case with K (K ≥ 2) legitimate
receivers. It should be highlighted that these channels also can be seen as multi-user
MIMO BCs.
Strictly speaking, with these channel models, NOMA is relevant only in cases in
which the number of transmit antennas (n t ) is less than the number of users (K ), i.e.
when the system is overloaded. This is because for n t < K , there is more than one
user per resource block (time/frequency/space), while for n t ≥ K , this figure is less
than or equal to one, implying that each user’s signal can be transmitted in a dedicated
resource block orthogonal to the other users’ resources. In the latter case, there is
at least one spatial degree of freedom per user which is the meaningful operating
regime for SDMA.
It should be highlighted that regardless of the relation between n t and K , optimal
beamforming provides interesting open problems. In the conventional multi-user
MIMO BC, n t users are selected out of K to satisfy the constraint n t ≤ K . However,
in 5G networks, due to the explosion of connected devices, K is usually very large. In
addition, many of these connections are from low-rate IoT devices, and a dedicated
resource block may allocate more resources than needed.7 In view of these factors,
7 In
LTE one resource block is 180 kHz.
5 NOMA: An Information-Theoretic Perspective
187
1
W1
1
Rx1
H1
n1
W2
W1, W2
nt
H2
1
W2
Rx2
n2
W1
Fig. 5.5 MIMO BC with confidential messages
NOMA is a useful technology for accommodating a higher number of connections
and improving spectral efficiency.
5.5.1.1
MIMO BC with Confidential Messages
A MIMO BC, similar to any other BC, has an inherent vulnerability in terms of
security and privacy, even if there is no external eavesdropper. This is due to the
fact that each legitimate user can, partly or wholly, decode the other legitimate user’s
message. To study this issue, the MIMO BC with confidential messages (see Fig. 5.5)
has been introduced, in which independent messages W1 and W2 are intended for
their respective receivers but need to be kept secret from the other receiver. This
model is important in NOMA transmission as each legitimate user is a potential
threat to security or privacy of its counterpart.
5.5.1.2
MIMO BC with an External Eavesdropper
In the above channel model, there is no external eavesdropper, but each legitimate
user is seen as a potential adversary (eavesdropper) to the other legitimate user. In
Fig. 5.6, another channel that models a class of BCs with an external eavesdropper
is depicted. In this model, the messages are to be secured from the eavesdropper but
not necessarily from other legitimate users. This channel is also known as the MIMO
multi-receiver wiretap channel whose capacity was established in [35], under a matrix
power constraint. This channel model is relevant for secure NOMA transmission
when external eavesdropping is the only security issue but the legitimate users are not
security threats to each other. It can be used, for example, for the case where the users
are resource-limited single-antenna devices such as sensors and, thus, are not capable
of or interested in decoding the other users’ data. Therefore, the confidentiality of
188
M. Vaezi and H. Vincent Poor
1
W1
1
Rx1
H1
n1
W1, W2
H2
1
W2
nt
Rx2
G
n2
1
W2
Eve
ne
W1
Fig. 5.6 MIMO BC with an external eavesdropper
a message from other legitimate users is not an issue. Obviously, in such cases, the
channel model reduces to a MISO BC.
5.5.1.3
MIMO BC with Confidential Messages and an External
Eavesdropper
A third and more general channel modeling secure communication in a NOMA-based
transmission is the MIMO BC with confidential messages and an external eavesdropper, depicted in Fig. 5.7. This channel consists of a transmitter who wishes to communicate two messages to their respective receivers, each needing to be kept secret
from the other legitimate receiver and a third unauthorized receiver (eavesdropper).
This configuration models NOMA with two users and one external eavesdropper.
While the capacity region of the channel in Fig. 5.7 is unknown, the capacity
regions of the channel models in Fig. 5.5 and Fig. 5.6 are established in [35] and [36],
respectively. In [36], using a combination of DPC and stochastic encoding, known
as secret DPC, it is shown that both messages can be simultaneously transmitted at
their respective maximal secrecy rates under a matrix power constraint. Similar to
the channel model in Fig. 5.5, secret DPC achieves the capacity region of the channel
in Fig. 5.6, under a matrix power constraint. However, under the more practical total
average power constraint, a computable secrecy capacity expression is not known
for any of those channels, in general. Because of this and also due to the fact that the
secret DPC is prohibitively complex for practical implementations, it is important to
find simple solutions, e.g. based on linear precoding, that maximize the achievable
secret rate and/or can achieve the secret capacity region in practice.
5 NOMA: An Information-Theoretic Perspective
189
1
W1
1
Rx1
H1
n1
W1, W2
H2
W2
1
W2
nt
Rx2
G
n2
W1
1
W2
Eve
ne
W1
Fig. 5.7 MIMO BC with confidential messages and an external eavesdropper
5.5.2 Physical Layer Security via Beamforming
Developing practical physical layer security techniques that address the security
issues of NOMA in the downlink is an important research topic. In particular, securing
communication for the three basic channel models illustrated in Figs. 5.5, 5.6 and
5.7 and their extension to K -user cases leads to practically important open problems.
Specifically, secure transmission strategies based on linear beamforming techniques
are of interest. In the following, we summarize different types of beamforming that
have been proposed for physical layer security.
Beamforming is one of the most widely studied approaches to physical layer
security for multi-antenna systems. Beamforming approaches can be categorized as
follows:
• Zero-forcing beamforming: One intuitive approach to secure transmission is to
send the signal as orthogonal to the eavesdropper’s channel as possible. When the
transmitter has a larger number of antennas than the eavesdropper, it is possible
to transmit the information in the null-space of the eavesdropper’s channel. This
approach, which is called zero-forcing beamforming, is a simple but suboptimal
beamforming scheme for multi-user MIMO systems [37].
• GSVD-based beamforming: Generalized singular value decomposition (GSVD)based precoding has been widely used for security and confidentiality purposes
in the MIMO wiretap and BC channels [38, 39] as well as in multicasting [40].
Simplicity is the main advantage of GSVD-based transmission as it decouples
the MIMO channel between the transmitter and the receivers into several parallel
subchannels which can be selected independently and then be encoded separately.
However, even in the case of the MIMO wiretap channel, GSVD-based precoding is
neither capacity achieving nor very close to capacity in general [41]. Nevertheless,
190
•
•
•
•
M. Vaezi and H. Vincent Poor
this approach is useful in that it can usually result in reasonably high achievable
rates with low complexity.
GEVD-based transmission: Linear beamforming based on the generalized eigenvalue decomposition (GEVD) is known to be optimal for both the MISO wiretap
channel [39] and MISO BC with confidential messages [42]. Despite being theoretically appealing for its optimality, in practice, there is no guarantee to have
single-antenna eavesdroppers. It is worth noting that under a matrix power constraint, GEVD-based precoding also achieves the secrecy capacity of the MIMO
wiretap channel [43].
Trigonometric approach: This is another simple linear beamforming approach
which was recently introduced in the context of the MIMO wiretap channel in [41,
44] and shown to be optimal for any numbers of antennas at the eavesdropper and
the legitimate receiver when the transmitter has two antennas.
Convex-optimization-based precoding: Numerical solutions based on convex optimization have also been proposed to compute a transmit covariance matrix for the
secrecy capacity maximization in the MISO and MIMO channels [45, 46]. These
methods solve the underlying non-convex optimization problem in an iterative
way and thus are very complex because the objectives are matrices.
Artificial Noise (AN)-aided transmission: When the eavesdropper’s CSI is not
available at the transmitter, an AN-aided transmission is useful for providing security at the physical layer. In this method, multiple antennas at the transmitter and
legitimate receiver are used to inject artificial noise into directions orthogonal to
those of the main channel [47]. Due to its simplicity and practicality against passive
eavesdroppers, AN-based beamforming is widely used for secure transmissions
[48, 49], particularly when the eavesdropper’s CSI is not known. In some cases,
e.g. for the fading MISO wiretap channel with no eavesdropper’s CSI, it is shown
that the optimal solution converges to transmitting AN in all null-space dimensions
of the main channel [50].
5.5.3 Research Directions
It is of considerable interest to develop physical layer security for NOMA transmission with both in-network and external eavesdroppers. The former is specifically
important in a NOMA-based transmission since a user with a stronger channel is
able to decode a weaker user’s signal, at the physical layer, and compromise their
privacy. Therefore, security with respect to external eavesdroppers and confidentiality (privacy) with respect to legitimate users are both critical in NOMA. The case
with multiple-antenna transmitters should be the main focus of such a study, as those
are now ubiquitous in many wireless networks, including cellular networks. In this
case, space-time signal processing such as linear beamforming with/without artificial
noise can be used to put physical layer security into practice and to improve secure
transmission rates. The key is to use signal processing techniques to increase the
signal strength difference between the legitimate user and eavesdropper.
5 NOMA: An Information-Theoretic Perspective
191
One set of problems in this setting is to identify linear beamforming techniques
that maximize the secrecy rate region of the models shown in Figs. 5.5, 5.6 and 5.7
assuming the availability of the legitimate users’ instantaneous perfect CSI at the
transmitter (CSIT). Another set of problems may consider the above security problems with more realistic assumptions, i.e. developing linear beamforming techniques
that maximize physical layer security with imperfect CSIT.
References
1. D. Tse, P. Viswanath, Fundamentals of Wireless Communication (Cambridge university press,
2005)
2. Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, K. Higuchi, Non-orthogonal
multiple access (NOMA) for cellular future radio access, in Proceedings of the 77th IEEE
Vehicular Technology Conference (VTC Spring) (2013), pp. 1–5
3. P. Wang, J. Xiao, P. Li, Comparison of orthogonal and non-orthogonal approaches to future
wireless cellular systems. IEEE Veh. Technol. Mag. 1(3), 4–11 (2006)
4. T.M. Cover, J.A. Thomas, Elements of Information Theory (Wiley, New York, 2006)
5. A. El Gamal, Y.H. Kim, Network Information Theory (Cambridge University Press, 2011)
6. T. Han, K. Kobayashi, A new achievable rate region for the interference channel. IEEE Trans.
Inf. Theory 27, 49–60 (1981)
7. M. Vaezi, H.V. Poor, Simplified Han-Kobayashi region for one-sided and mixed Gaussian interference channels, in Proceedings of the IEEE International Conference on Communications
(ICC) (2016), pp. 1–6
8. ITU, Minimum requirements related to technical performance for IMT-2020 radio interface(s),
February 2017, https://www.itu.int/md/R15-SG05-C-0040/en
9. S.R. Islam, N. Avazov, O.A. Dobre, K.-S. Kwak, Power-domain non-orthogonal multiple access
(NOMA) in 5G systems: potentials and challenges. IEEE Commun. Surv. Tutor. (2017)
10. Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, I. Chih-Lin, H.V. Poor, Application of nonorthogonal multiple access in LTE and 5G networks. IEEE Commun. Mag. 55(2), 185–191
(2017)
11. W. Shin, M. Vaezi, B. Lee, D.J. Love, J. Lee, H.V. Poor, Non-orthogonal multiple access
in multi-cell networks: theory, performance, and practical challenges. IEEE Commun. Mag.
55(10), 176–183 (2017)
12. 3GPP TD RP-150496, Study on Downlink Multiuser Superposition Transmission, Mar 2015
13. Q. Sun, S. Han, C.-L. I, Z. Pan, On the ergodic capacity of MIMO NOMA systems. IEEE
Wirel. Commun. Lett. 4(4), 405–408 (2015)
14. Z. Ding, F. Adachi, H.V. Poor, The application of MIMO to non-orthogonal multiple access.
IEEE Trans. Wirel. Commun. 15(1), 537–552 (2016)
15. Z. Ding, R. Schober, H.V. Poor, A general MIMO framework for NOMA downlink and uplink
transmission based on signal alignment. IEEE Trans. Wirel. Commun. 15(6), 4438–4454 (2016)
16. R.H. Etkin, D.N. Tse, H. Wang, Gaussian interference channel capacity to within one bit. IEEE
Trans. Inf. Theory 54(12), 5534–5562 (2008)
17. A.S. Motahari, A.K. Khandani, Capacity bounds for the Gaussian interference channel. IEEE
Trans. Inf. Theory 55(2), 620–643 (2009)
18. X. Shang, G. Kramer, B. Chen, A new outer bound and the noisy-interference sum-rate capacity
for Gaussian interference channels. IEEE Trans. Inf. Theory 55(2), 689–699 (2009)
19. V.S. Annapureddy, V.V. Veeravalli, Gaussian interference networks: sum capacity in the lowinterference regime and new outer bounds on the capacity region. IEEE Trans. Inf. Theory
55(7), 3032–3050 (2009)
192
M. Vaezi and H. Vincent Poor
20. M. Vaezi, H.V. Poor, On limiting expressions for the capacity regions of Gaussian interference
channels, in Proceedings of the 49th Asilomar Conference on Signals, Systems and Computers,
Nov (2015), pp. 1292–1298
21. Y. Pang, M. Varanasi, Approximate capacity region of the MAC-IC-MAC (2016),
arXiv:1604.02234
22. H. Weingarten, Y. Steinberg, S.S. Shamai, The capacity region of the Gaussian multiple-input
multiple-output broadcast channel. IEEE Trans. Inf. Theory 52(9), 3936–3964 (2006)
23. M. Costa, Writing on dirty paper. IEEE Trans. Inf. Theory 29(3), 439–441 (1983)
24. V.O.L. Yijie Mao, B. Clerckx, Rate-splitting multiple access for downlink communication
systems: bridging, generalizing and outperforming SDMA and NOMA, arXiv:1710.11018v3
25. M.A. Maddah-Ali, A.S. Motahari, A.K. Khandani, Signaling over MIMO multi-base systems:
combination of multi-access and broadcast schemes, in Proceedings of the IEEE International
Symposium on Information Theory (2006), pp. 2104–2108
26. M.A. Maddah-Ali, A.S. Motahari, A.K. Khandani, Communication over MIMO X channels:
interference alignment, decomposition, and performance analysis. IEEE Trans. Inf. Theory
54(8), 3457–3470 (2008)
27. V.R. Cadambe, S.A. Jafar, Interference alignment and degrees of freedom of the K -user interference channel. IEEE Trans. Inf. Theory 54(8), 3425–3441 (2008)
28. C.M. Yetis, T. Gou, S.A. Jafar, A.H. Kayran, On feasibility of interference alignment in MIMO
interference networks. IEEE Trans. Signal Process. 58(9), 4771–4782 (2010)
29. T. Gou, S.A. Jafar, Degrees of freedom of the K user M × N MIMO interference channel.
IEEE Trans. Inf. Theory 56(12), 6040–6057 (2010)
30. 3GPP TR 36.866 V12.0.1, Study on Network-Assisted Interference Cancellation and Suppression (NAIC) for LTE (Release 12), Mar 2014
31. Q. Li, G. Li, W. Lee, M.-i. Lee, D. Mazzarese, B. Clerckx, Z. Li, MIMO techniques in WiMAX
and LTE: a feature overview. IEEE Commun. Mag. 48(5), (2010)
32. Y. Yuan, Z. Yuan, G. Yu, C.-H. Hwang, P.-K. Liao, A. Li, K. Takeda, Non-orthogonal transmission technology in LTE evolution. IEEE Commun. Mag. 54(7), 68–74 (2016)
33. D. Lee, H. Seo, B. Clerckx, E. Hardouin, D. Mazzarese, S. Nagata, K. Sayana, Coordinated
multipoint transmission and reception in LTE-advanced: deployment scenarios and operational
challenges. IEEE Commun. Mag. 50(2), 148–155 (2012)
34. W. Shin, M. Vaezi, B. Lee, D.J. Love, J. Lee, H.V. Poor, Coordinated beamforming for multi-cell
MIMO-NOMA. IEEE Commun. Lett. 21(1), 84–87 (2017)
35. E. Ekrem, S. Ulukus, The secrecy capacity region of the Gaussian MIMO multi-receiver wiretap
channel. IEEE Trans. Inf. Theory 57(4), 2083–2114 (2011)
36. R. Liu, T. Liu, H.V. Poor, S. Shamai, Multiple-input multiple-output Gaussian broadcast channels with confidential messages. IEEE Trans. Inf. Theory 56(9), 4215–4227 (2010)
37. K. Wang, X. Wang, X. Zhang, SLNR-based transmit beamforming for MIMO wiretap channel.
Wirel. Pers. Commun. 71(1), 109–121 (2013)
38. S.A.A. Fakoorian, A.L. Swindlehurst, Optimal power allocation for GSVD-based beamforming
in the MIMO Gaussian wiretap channel, in Proceedings of the IEEE International Symposium
on Information Theory (2012), pp. 2321–2325
39. A. Khisti, G.W. Wornell, Secure transmission with multiple antennas I: the MISOME wiretap
channel. IEEE Trans. Inf. Theory 56(7), 3088–3104 (2010)
40. W. Mei, Z. Chen, J. Fang, GSVD-based precoding in MIMO systems with integrated services.
IEEE Signal Process. Lett. 23(11), 1528–1532 (2016)
41. M. Vaezi, W. Shin, V. Poor, Optimal beamforming for Gaussian MIMO wiretap channels with
two transmit antennas. IEEE Trans. Wirel. Commun. 16, 6726–6735 (2017)
42. R. Liu, H.V. Poor, Secrecy capacity region of a multiple-antenna Gaussian broadcast channel
with confidential messages. IEEE Trans. Inf. Theory 55(3), 1235–1249 (2009)
43. R. Bustin, R. Liu, H. V. Poor, S. Shamai, An MMSE approach to the secrecy capacity of the
MIMO Gaussian wiretap channel. EURASIP J. Wirel. Commun. Netw. (1) (2009)
44. M. Vaezi, W. Shin, H. V. Poor, J. Lee, MIMO Gaussian wiretap channels with two transmit
antennas: optimal precoding and power allocation, in Proceedings of the IEEE International
Symposium on Information Theory (2017), pp. 1708–1712
5 NOMA: An Information-Theoretic Perspective
193
45. Q. Li, W.-K. Ma, Optimal and robust transmit designs for MISO channel secrecy by semidefinite
programming. IEEE Trans. Signal Process. 59(8), 3799–3812 (2011)
46. Q. Li, M. Hong, H.-T. Wai, Y.-F. Liu, W.-K. Ma, Z.-Q. Luo, Transmit solutions for MIMO
wiretap channels using alternating optimization. IEEE J. Sel. Areas Commun. 31(9), 1714–
1727 (2013)
47. S. Goel, R. Negi, Guaranteeing secrecy using artificial noise. IEEE Trans. Wirel. Commun.
7(6), 2180–2189 (2008)
48. Z. Li, P. Mu, B. Wang, X. Hu, Optimal semiadaptive transmission with artificial-noise-aided
beamforming in MISO wiretap channels. IEEE Trans. Veh. Technol. 65(9), 7021–7035 (2016)
49. T.-X. Zheng, H.-M. Wang, J. Yuan, D. Towsley, M.H. Lee, Multi-antenna transmission with
artificial noise against randomly distributed eavesdroppers. IEEE Trans. Commun. 63(11),
4347–4362 (2015)
50. S. Gerbracht, C. Scheunert, E.A. Jorswieck, Secrecy outage in MISO systems with partial
channel information. IEEE Trans. Inf. Forensics Secur. 7(2), 704–716 (2012)
Chapter 6
Optimal Power Allocation for Downlink
NOMA Systems
Yongming Huang, Jiaheng Wang and Jianyue Zhu
6.1 Introduction
With the popularity of smartphones and Internet of Things, there is an explosive
demand for new services and data traffic for wireless communications. The capacity
of the fourth-generation (4G) mobile communication system is insufficient to satisfy
such a demand in the near future. The development of the fifth-generation (5G) mobile
communication system has been placed on the agenda with higher requirements in
data rates, latency, and connectivity [1]. In order to meet the new standards, some
potential technologies, such as massive multiple-input–multiple-output (MIMO) [2],
millimeter wave [3], and ultra densification [4, 5], will be introduced into 5G. Meanwhile, new multiple access technologies, which are flexible, reliable, and efficient in
terms of energy and spectrum, are also considered for 5G communication.
Conventionally, cellular systems have adopted orthogonal multiple access (OMA)
approaches, in which wireless resources are allocated orthogonally to multiple users.
The common OMA techniques include frequency-division multiple access (FDMA),
time division multiple access (TDMA), code-division multiple access (CDMA), and
orthogonal frequency-division multiple access (OFDMA). Ideally, in OMA, the intracell interference does not exist as result of dedicated resource allocation. Also, for
this reason, the information of multiple users can be retrieved at a low complexity.
Nonetheless, the number of served users is limited by the number of orthogonal
resources, which is generally small in practice. Consequently, it is difficult for OMA
systems to support a massive connectivity.
Recently, non-orthogonal multiple access (NOMA) technologies are developed
and proposed for 5G, which will contribute to disruptive design changes on radio
access and alleviate the scarcity of suitable spectra. By using superposition coding
Y. Huang (B) · J. Wang · J. Zhu
Southeast University, Nanjing 210096, China
e-mail: huangym@seu.edu.cn
J. Wang
e-mail: jhwang@seu.edu.cn
J. Zhu
e-mail: zhujy@seu.edu.cn
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_6
195
196
Y. Huang et al.
at the transmitter with successive interference cancellation (SIC) at the receiver,
downlink NOMA allows one (frequency, time, code, or spatial) channel to be shared
by multiple users simultaneously [6, 7], thus leading to better performance in terms of
spectral efficiency, fairness, or energy efficiency [8]. Therefore, NOMA has received
much attention recently. Its combinations with MIMO and multi-cell technologies
were studied in [9, 10] and [11], respectively. NOMA was also considered to be used
in, e.g., visible light communication [12] and millimeter wave communication [13].
The principle of NOMA is to implement multiple access in the power domain
[14]. Hence, allocation is critical for NOMA systems. In the literature, there are a
number of works on power allocation for NOMA. In particular [15, 16] focused
on power allocation in a two-user NOMA system and [17–21] investigated power
allocation for multiple users (more than two) sharing one channel, which is referred
as multi-user NOMA (MU-NOMA). There were also some works, e.g., [12, 22–28],
studying the resource allocation problems in multi-channel NOMA (MC-NONA)
systems, where multiple channels are available for NOMA. Different criteria, such
as maximin fairness [18–20, 22], sum rate [15–17, 22, 28–30], and energy efficiency
[21–23, 26], were considered.
This chapter focuses on power allocation for downlink NOMA. We first briefly
review the basic concepts of downlink NOMA transmission and introduce the twouser NOMA, MU-NOMA, and MC-NOMA schemes. Then, we investigate the optimal power allocation strategies for these NOMA schemes under different performance criteria such as the maximin fairness, sum rate, and energy efficiency along
with user weights and quality-of-service (QoS) constraints. We show that the optimal
NOMA power allocation can be analytically characterized in most cases, otherwise
it can be numerically computed via convex optimization methods.
This chapter is organized as follows. Section 6.2 introduces the fundamentals of
downlink NOMA and the two-user NOMA, MU-NOMA, and MC-NOMA schemes.
In Sects. 6.3–6.5, we investigate the optimal power allocation for two-user NOMA,
MU-NOMA, and MC-NOMA schemes, respectively. The performance of the NOMA
power allocation strategies is evaluated in Sect. 6.6 via simulations, and the conclusion is drawn in Sect. 6.7.
6.2 Fundamentals of Downlink NOMA
In this section, we review the basic concepts of downlink NOMA transmission in a
single-cell network.1 To begin with, we start from the simplest two-user case, where
a base station (BS) serves two users, namely UE1 and UE2 , on the same frequency
band with bandwidth B. The BS transmits a signal sn for user n (UEn , n = 1, 2) with
transmission power pn . The total power budget of the BS is P, i.e., p1 + p2 ≤ P.
Such a simple downlink NOMA system is displayed in Fig. 6.1 [31].
1 For
multi-cell NOMA, the reader is referred to [11].
6 Optimal Power Allocation for Downlink NOMA Systems
197
Fig. 6.1 A downlink NOMA system with one base station and two users
Superposition Coding: According to the NOMA principle, the BS exploits the
superposition coding and broadcasts the signal
x=
√
p 1 s1 +
√
(6.1)
p 2 s2
to both users. The received signal at UEn is
yn = h n
−ρ
√
p 1 s1 +
√
p 2 s2 + z n ,
(6.2)
where h n = gn dn is the channel coefficient from the BS to UEn , gn follows a
Rayleigh distribution, dn is the distance between the BS and UEn , ρ is the path loss
exponent, and z n is the additive
white Gaussian noise with zero mean and variance
σn2 , i.e., z n ∼ C N 0, σn2 .
Successive Interference Cancellation (SIC): In NOMA systems, each user exploits
SIC at its receiver. Let Γn = |h n |2 /σn2 be the channel-to-noise ratio (CNR) of UEn .
Assume without loss of generality (w.l.o.g.) that the users are ordered by their normalized channel gains as Γ1 ≥ Γ2 , i.e., UE1 and UE2 are regarded as the strong and
weak users, respectively. It is expected that more power is allocated to the weak user
UE2 and less power is allocated to the strong user UE1 , i.e., p1 ≤ p2 [14, 25]. Then,
UE1 first decodes the message of UE2 and removes it from its received signal, while
UE2 treats the signal of UE1 as interference and decodes its own message.
Achievable Rate: Suppose that the channel coding is ideal and UE1 is able to
decode the message of UE2 successfully. Then, the achievable rates of UE1 and UE2
are given respectively by
R1 = B log (1 + p1 Γ1 ) , R2 = B log 1 +
p 2 Γ2
1 + p 1 Γ2
(6.3)
198
Y. Huang et al.
which are often used as the design objectives of NOMA systems.
Multi-User NOMA (MU-NOMA): Consider a more general case where a BS serves
N ≥ 2 users on the same spectrum, which are indexed by n = 1, . . ., N. The broadcast
signal by the BS is then given by
x=
N
√
pi si
(6.4)
i=1
and then the received signal at each UEn is given by
yn = h n
N
√
i=1
pi si + z n .
(6.5)
Similarly, suppose that the users are ordered by their normalized channel gains as
Γ1 ≥ Γ 2 ≥ · · · ≥ Γ N
(6.6)
and the NOMA protocol allocates higher powers to the users with lower CNRs,
leading to p1 ≤ p2 ≤ · · · ≤ p N . Therefore, UEn is able to decode the message of
UEl for l > n and remove it from the received signal so that UEn is only interfered
by UE j for j < n. Therefore, after SIC, the achievable rate of UEn is
p n Γn
Rn = log 1 + n−1
j=1 p j Γn + 1
(6.7)
for n = 1, . . . , N .
Multi-Channel NOMA (MC-NOMA): The frequency band shared by the users
could be viewed as a channel, which may also be a time slot, spread code, or resource
block. In cellular systems, there are often multiple channels available, which leads to
a more general NOMA scheme called multi-channel NOMA (MC-NOMA), where
multiple users share multiple channels. Specifically, in a downlink MC-NOMA network, the BS serves N users through M channels and the total bandwidth B is
equally divided to M channels so the bandwidth of each channel is Bc = B/M. Let
Nm ∈ {N1 , N2 , ..., N M } be the number of users using channel m for m = 1, 2, . . ., M
and UEn,m denotes user n on channel m for n = 1, 2, . . . , Nm . The signal transmitted
by the BS on each channel m can be expressed as
xm =
Nm
√
pn,m sn ,
(6.8)
n=1
where sn is the symbol of UEn,m and pn,m is the power allocated to UEn,m . The
received signal at UEn,m is
6 Optimal Power Allocation for Downlink NOMA Systems
yn,m =
Nm
√
i=1
199
pi,m h n,m si + z n,m .
(6.9)
It is easily seen that on each channel m is an MU-NOMA scheme. Similarly, assume
w.l.o.g. that the CNRs of the users on channel m are ordered as
Γ1,m ≥ · · · ≥ Γn,m ≥ · · · ≥ Γ Nm ,m ,
(6.10)
which will lead to p1,m ≤ · · · ≤ pn,m ≤ · · · ≤ p Nm ,m . Then, the achievable rate of
UEn,m using SIC is
Rn,m = Bc log 1 +
pn,m Γn,m
n−1
1 + i=1 pi,m Γn,m
.
(6.11)
The basic idea of NOMA is to implement multiple access in the power domain
[14]. Hence, power allocation is the key to achieve the full benefit of NOMA transmission. In the following parts, we will investigate the optimal power allocation strategies
for different NOMA schemes, including the simplest two-user case, the MU-NOMA
scheme, and the MC-NOMA scheme, under different performance measures.
6.3 Two-User NOMA
In this section, we investigate the optimal power allocation for the two-user NOMA
scheme. Although the two-user scheme is the simplest case of NOMA, the results
and insights obtained in this case will serve the more complicated MU-NOMA and
MC-NOMA schemes.
6.3.1 Optimal Power Allocation for MMF
The NOMA scheme enables a flexible management of users’ achievable rates and
provides an efficient way to enhance user fairness. A widely used fairness metric
is the maximin fairness (MMF), which is achieved by maximizing the worst (i.e.,
minimum) user rate. According to (6.3), the power allocation to achieve the MMF is
given by the solution to the following optimization problem:
T U MMF :
max min {R1 ( p1 , p2 ), R2 ( p1 , p2 )}
p1 , p2
s.t. 0 ≤ p1 ≤ p2 , p1 + p2 ≤ P
This problem admits a closed-form solution as follows.
200
Y. Huang et al.
Proposition 1 Suppose that Γ1 ≥ Γ2 . Then, the optimal solution to T U MMF is given
by p1⋆ = Λ and p2⋆ = P − p1⋆ , where Γl |h l |2 /σl2 and
Λ
− (Γ1 + Γ2 ) +
(Γ1 + Γ2 ) 2 + 4Γ1 Γ22 P
2Γ1 Γ2
.
(6.12)
Proof Please refer to the proof of Proposition 1 in [22].
Remark 1 It can be verified that at the optimal point R1 ( p1⋆ , p2⋆ ) = R2 ( p1⋆ , p2⋆ ), i.e.,
UE1 and UE2 achieve the same rate. This indicates that, under the MMF criterion,
the NOMA system will provide absolute fairness for two users on one channel.
To elaborate another important insight, we introduce the following definition.
Definition 1 A NOMA system is called SIC stable if the optimal power allocation
satisfies p1 < p2 on one channel.
Remark 2 In NOMA systems, SIC is performed according to the order of the CNRs
of the users on one channel [14, 25], which is guaranteed by imposing an inverse order
of the powers allocated to the users, i.e., p1 ≤ p2 . Specifically, UE1 (the stronger
user) first decodes the signal of UE2 (the weaker user) and then subtracts it from the
superposed signal. Therefore, from the SIC perspective, a large difference between
the signal strengths of UE2 and UE1 is preferred [32]. However, even with the power
order constraint p1 ≤ p2 , the power optimization may lead to p1 = p2 ; i.e., UE1
and UE2 have the same signal strength, which is the worst situation for SIC. In this
case, SIC may fail or has a large error propagation and thus is unstable. Indeed, the
authors in [33] pointed out that the power of the weak user must be strictly larger than
that of the strong user, otherwise the users’ outage probabilities will always be one.
Definition 1 explicitly concretizes such a practical requirement in NOMA systems.
Lemma 1 The NOMA system is SIC stable for T U MMF .
Proof Please refer to the proof of Lemma 1 in [22].
Indicated by Lemma 1, the two-user NOMA system is always SIC stable under the
MMF criterion, as in this case the optimal power allocation always satisfies p1⋆ < p2⋆ .
On the other hand, in the subsequent subsections, we will show that a NOMA system
may not always be SIC stable under different criteria.
6.3.2 Optimal Power Allocation for SR Maximization
In this subsection, we seek the optimal power allocation for maximizing the sum rate
(SR). In SR maximization, to take user priority or fairness into account, user weights
or quality-of-service (QoS) constraints are often adopted.
6 Optimal Power Allocation for Downlink NOMA Systems
6.3.2.1
201
Weighted SR Maximization (SR1)
According to (6.3), the problem of maximizing the weighted SR (WSR) is given by
T U SR1 :
max W1 R1 ( p1 , p2 ) + W2 R2 ( p1 , p2 )
p1 , p2
s.t. 0 ≤ p1 ≤ p2 , p1 + p2 ≤ P
where Wi denotes the weight of UEi for i = 1, 2. Note that T U SR1 is a nonconvex
problem due to the interference between UE1 and UE2 . Nevertheless, its optimal
solution can be found as follows.
Proposition 2 Suppose that Γ1 ≥ Γ2 , W1 < W2 and P > 2Ω, with
Ω
W 2 Γ2 − W 1 Γ1
.
Γ1 Γ2 (W1 − W2 )
(6.13)
Then, the optimal solution to T U SR1 is given by p1⋆ = Ω and p2⋆ = P − p1⋆ .
Proof Please refer to the proof of Proposition 2 in [22].
Remark 3 In Proposition 2, the conditions W1 < W2 and P > 2Ω are both to avoid
a failure of SIC. Indeed, if W1 ≥ W2 , the solution to T U SR1 is p1⋆ = p2⋆ = P/2 ;
i.e., the NOMA system is unstable according to Definition 1. SIC may also fail if
P ≤ 2Ω, which will lead to p1⋆ = p2⋆ = P/2 as well. Therefore, the two-user NOMA
system is SIC stable for the WSR maximization if and only if W1 < W2 and P > 2Ω.
6.3.2.2
SR Maximization with QoS (SR2)
Now, we consider maximizing the SR with QoS constraints. In this case, the power
allocation problem is given by
max R1 ( p1 , p2 ) + R2 ( p1 , p2 )
p1 , p2
T U SR2 : s.t. 0 ≤ p1 ≤ p2 , p1 + p2 ≤ P
Ri ≥ Rimin , i = 1, 2.
where Rimin is the QoS threshold of UEi . The optimal solution to T U SR2 is provided
in the following result.
Proposition 3 Suppose that Γ1 ≥ Γ2 , A2 ≥ 2, and P ≥ Υ , with
min
Al = 2 Rl , Υ
A2 − 1
A2 (A1 − 1)
+
,
Γ1
Γ2
Then, the optimal solution to T U SR2 is given by p1⋆ =
Γ2 P − A 2 + 1
.
A 2 Γ2
and p2⋆ = P − p1⋆ .
(6.14)
202
Y. Huang et al.
Proof Please refer to the proof of Proposition 3 in [22].
Remark 4 Similarly, in Proposition 3, the conditions A2 ≥ 2 and P ≥ Υ are to
guarantee the SIC stability. Indeed, if A2 < 2, then > P/2 and the optimal solution
will be p1⋆ = p2⋆ = P/2, which may lead a failure of SIC. At the same time, SIC may
also fail if P < Υ , which will lead to p1⋆ = p2⋆ = P/2 as well. Therefore, the NOMA
system is SIC stable in this case if and only if A2 ≥ 2 and P ≥ Υ .
According to Proposition 3, if the NOMA system is SIC stable, the optimal solution will be p1⋆ = and p2⋆ = P − p1⋆ . Hence, we have R2 ( p1⋆ , p2⋆ ) = R2min , implying
that the user with a lower CNR (i.e., UE2 ) receives the power to meet its QoS requirement exactly, while the remaining power is used to maximize the rate of the user
with a higher CNR (i.e., UE1 ).
6.3.3 Optimal Power Allocation for EE Maximization
In this subsection, we investigate the optimal power allocation for maximizing the
energy efficiency (EE), which is defined as the ratio between the rate and the consumed power. Similarly, user weights and QoS constraints are considered.
6.3.3.1
Weighted EE Maximization (EE1)
The problem of maximizing the weighted EE is formulated as follows:
T UaEE1 :
max η = W1 R1 ( p1 , p2 )+W2 R2 ( p1 , p2 )
p1 , p2
PT + p1 + p2
s.t.0 ≤ p1 ≤ p2 , p1 + p2 ≤ P
where PT is the power consumption of the circuits. Given the fraction form of the
objective, T UaEE1 is more complicated than T U SR1 . In the following, we show that
this problem can also be optimally solved.
We introduce an auxiliary variable q with p1 + p2 = q. Then, T UaEE1 can be
equivalently written into
T UbEE1 :
max η = W1 R1 ( p1 )+W2 R2 ( p1 ,q)
p1 ,q
PT +q
s.t. q ≥ 2 p1 , q ≤ P
where the rate of UE2 can be expressed as
R2 ( p1 , q) = B log
1 + qΓ2
1 + p 1 Γ2
.
(6.15)
6 Optimal Power Allocation for Downlink NOMA Systems
203
To deal with the fractional form, let us introduce the following objective function:
H ( p1 , q, α) W1 R1 ( p1 ) + W2 R2 ( p1 , q) − α (PT + q) ,
(6.16)
where α is a positive parameter. Then, we consider the following problem for given
α:
max H ( p1 , q, α)
T UcEE1 : p1 ,q
s.t. q ≥ 2 p1 , q ≤ P.
The relation between T UcEE1 and T UbEE1 is stated in the following result.
Lemma 2 ([34, pp. 493–494]) Let H ⋆ (α) be the optimal objective value of T UcEE1
and p⋆ (α) be the optimal solution of T UcEE1 . Then, p⋆ (α) is the optimal solution to
T UbEE1 if and only if H ⋆ (α) = 0.
According to Lemma 2, the optimal solution to T UbEE1 can be found by solving
T UcEE1 parameterized by α such that H ⋆ (α) = 0. Since H ⋆ (α) is monotonic in α,
one can use any line search method, e.g., the bisection method, to find α such that
H ⋆ (α) = 0. Then, the left question is how to solve T UcEE1 with given α.
Theorem 1 Suppose that Γ1 ≥ Γ2 . Then, T UcEE1 is a convex problem if one of the
following conditions hold
C1 : W1 ≥ W2 ;
(Γ1 +Γ1 Γ2 P)2
2
C2 : 1 < W
.
≤ (Γ
W1
+Γ Γ P)2
2
1 2
Proof See Appendix A.
Theorem 1 reveals that T UcEE1 is in fact a convex problem if condition C1 or C2 holds.
Consequently, T UcEE1 can be efficiently solved via convex optimization methods. The
optimal solution to T UcEE1 can further be analytically characterized.
Proposition 4 Suppose that Γ1 ≥ Γ2 , C2 holds, and P ≥ 2Ω with
Ω=
W 2 Γ2 − W 1 Γ1
.
Γ1 Γ2 (W1 − W2 )
(6.17)
Then, the optimal solution to T UcEE1 is p1⋆ = Ω and
q⋆ =
1
W2 B
−
α ln 2 Γ2
Proof See Appendix B.
P
2Ω
= max 2Ω, min
1
W2 B
− ,P
α ln 2 Γ2
.
(6.18)
204
Y. Huang et al.
Remark 5 Similarly, in Proposition 4, the conditions C2 and P > 2Ω are to avoid
a failure of SIC. Although T UcEE1 is convex under C1, condition C1 will lead to
p1⋆ = p2⋆ , which, according to Definition 1, is SIC-unstable.
6.3.3.2
EE Maximization with QoS Constraints (EE2)
Then, we consider maximizing the EE with QoS constraints. In this case, the power
allocation problem is given by
max
p1 , p2
R1 ( p1 , p2 )+R2 ( p1 , p2 )
PT + p1 + p2
T UaEE2 : s.t. 0 ≤ p1 ≤ p2 , p1 + p2 ≤ P
Ri ≥ Rimin , i = 1, 2.
This problem can be optimally solved following the similar steps as in the previous
subsection.
Specifically, using p1 + p2 = q, T UaEE2 can be equivalently transformed into
max
p1 ,q
R1 ( p1 )+R2 ( p1 ,q)
PT +q
T UbEE2 : s.t. q ≥ 2 p1 , q ≤ P
Ri ≥ Rimin , i = 1, 2.
Then, we consider the following problem with given α:
T UcEE2
max Q ( p1 , q, α)
p1 ,q
: s.t. q ≥ 2 p1 , q ≤ P
Ri ≥ Rimin , i = 1, 2
where
Q ( p1 , q, α) R1 ( p1 ) + R2 ( p1 , q) − α (PT + q) .
(6.19)
According to Lemma 2, the optimal solution to T UbEE2 can be found by solving
T UcEE2 for a given α and updating α until the optimal objective value of T UcEE2 ,
denoted by Q ⋆ (α), satisfies Q ⋆ (α) = 0.
From Theorem 1, under condition C1, the objective of T UcEE2 is concave and
T UcEE2 is a convex problem too. Therefore, T UcEE2 can be efficiently solved. In fact,
the optimal solution to T UcEE2 can also be analytically characterized.
Proposition 5 Suppose that Γ1 ≥ Γ2 , A2 ≥ 2, and P ≥ Υ , with
Al = 2
Rlmin
B
, Υ
A2 − 1
A2 (A1 − 1)
+
.
Γ1
Γ2
(6.20)
6 Optimal Power Allocation for Downlink NOMA Systems
205
Then, the optimal solution to T UcEE2 is
1 + q ⋆ Γ2 − A 2
,
A 2 Γ2
A2
A2 − 1
1
−
+
q⋆ =
α
Γ1
Γ2
p1⋆ =
(6.21)
P
= max Υ, min
Υ
A2
A2 − 1
1
−
+
,P
α
Γ1
Γ2
.
(6.22)
Proof See Appendix C.
Similarly, it can be verified that the power allocation obtained from T UcEE2 (or
T UaEE2 ) is SIC stable if and only if P ≥ Υ and A2 ≥ 2.
6.4 MU-NOMA
In this section, we consider the more general MU-NOMA scheme, where a BS
serves N ≥ 2 users on the same channel. Similarly, the optimal MU-NOMA power
allocation is investigated under the MMF, SR, and EE criteria with user weights or
QoS constraints.
6.4.1 Optimal Power Allocation for MMF
According to (6.7), the power allocation problem under the MMF criterion is formulated as
min {Ri }
max i=1,...,N
MMF
p
N
:
MUa
s.t. 0 < p1 < · · · < p N ,
pi ≤ P
i=1
N
where p = { pi }i=1
denotes the powers allocated to the users. It has been shown in
[20] that, though nonconvex, MUaMMF is a quasi-convex problem. Thus, the optimal
solution to MUaMMF can be found by solving a sequence of convex problems.
Specifically, MUaMMF is equivalent to
max t
p,t
MUbMMF
N
: s.t. 0 < p1 < . . . < p N , pi ≤ P
i=1
Ri ≥ t, i = 1, . . . , N .
The constraint Ri ≥ t can be rewritten into
206
Y. Huang et al.
⎞
⎛
i−1
2 t − 1 ⎝
p j Γi + 1⎠ .
pi ≥
Γi
j=1
(6.23)
Hence, for fixed t, MUbMMF is a linear program (LP) and can be efficiently solved
by a number of LP solvers. Then, one can exploit the bisection method to search the
optimal t. Note that the optimal solution to MUbMMF for fixed t can be analytically
characterized if there is no power order constraint.
Proposition 6 In the absence of power order constraint, the solution to MUbMMF is
given by
⎞
⎛
i−1
2 t − 1 ⎝
pi =
(6.24)
p j Γi + 1⎠ , i = 1, · · · , N .
Γi
j=1
Proof Please refer to the proof of Theorem 1 in [20].
The solution in (6.24) implies that all users achieve the same data rate equal to t.
Hence, in this case, the NOMA system will provide absolute fairness for all users.
Note that, however, the solution in (6.24) is obtained without the power order constraint. One may wonder if this solution is still optimal if the power order constraint
is not omitted. The following result provides a sufficient condition to characterize
the optimality of (6.24).
Theorem 2 The solution in (6.24) is optimal for MUbMMF if P ≥ χ , where χ =
N 2 N −i
i=1 Γi .
Proof See Appendix E.
Theorem 2 indicates that the power order constraints can be omitted under some
conditions. In this case, the solution in (6.24) is indeed optimal for MUbMMF . On
the other hand, it is unknown if the solution in (6.24) is optimal if the condition in
Theorem 2 is not satisfied. Nevertheless, in this case, one can always numerically
solve the linear problem MUbMMF for fixed t.
6.4.2 Optimal Power Allocation for SR Maximization
In this subsection, we investigate the SR maximization problems in MU-NOMA
systems with user weights or QoS constraints.
6.4.2.1
Weighted SR Maximization (SR1)
The weighted SR maximization for MU-NOMA is formulated as
6 Optimal Power Allocation for Downlink NOMA Systems
max Rsum =
p
MUaSR1
:
s.t.
N
i=1
N
207
Wi Ri
i=1
pi ≤ P
p1 ≤ p2 ≤ · · · ≤ p N .
Unlike MUaMMF for MMF, MUaSR1 in its original form is neither a convex nor quasiconvex problem, making it difficult to solve it. Nevertheless, we show that MUaSR1 can
be transformed into a convex problem via a linear transformation of the optimization
variables.
Introduce the following variable transformation: qi = ij=1 p j for i = 1, 2, . . .,
N; and conversely pi = qi − qi−1 for i = 2, . . ., N and p1 = q1 . In this way, we have
R1 = log (q1 Γ1 + 1) and
⎛
⎞
i
qi Γi + 1
j=1 p j Γi + 1 ⎠
= log (qi Γi + 1) − log qi−1 Γi + 1
Ri = log ⎝ i−1
= log
qi−1 Γi + 1
j=1 p j Γi + 1
(6.25)
for i = 2, . . ., N. Therefore, the weighted sum rate can be expressed as
N
i=1
Wi Ri = W1 log (q1 Γ1 + 1) +
N
i=2
N
Wi log (qi Γi + 1) − log qi−1 Γi + 1 =
f i (qi ) ,
i=1
(6.26)
where
f i (qi ) = Wi log (qi Γi + 1) − Wi+1 log (qi Γi+1 + 1)
(6.27)
for
N i = 1, . . ., N − 1 and f N (q N ) = W N log (q N Γ N + 1). The power constraint
i=1 pi ≤ P is equal to q N ≤ P. The power order constraint p1 ≤ p2 ≤ · · · ≤ p N
is equal to q1 ≤ q2 − q1 ≤ · · · ≤ q N − q N −1 . Consequently, problem MUaSR1 can be
equivalently transformed into the following problem:
max
N
i=1 f i (qi )
MUbSR1 : s.t. q N ≤ P
0 ≤ q1 ≤ q2 − q1 ≤ · · · ≤ q N − q N −1
q
Then, the following result identifies the convexity of MUaSR1 (or MUbSR1 ).
Theorem 3 MUaSR1 (or MUbSR1 ) is a convex problem if one of the following conditions hold for i = 1, . . ., N − 1:
T1 : Wi ≥ Wi+1 ;
2
i Γi+1 P)
.
≤ (Γ(Γi +Γ
T2 : 1 < WWi+1
+Γ Γ P)2
i
i+1
i
i+1
Proof Please refer to the proof of Theorem 1 in [17].
Remark 6 Theorem 3 indicates that MUaSR1 (or MUbSR1 ) is a convex problem under
some conditions of the user weights. From T1, if the user weights are in the same
208
Y. Huang et al.
order as the channel gains, i.e., W1 ≥ W2 ≥ · · · ≥ W N , then the objective function
is concave and the problem is convex. Note that this situation includes the most
common sum rate as a special case. On the other hand, the user weights can also be
in the inverse order of the channel gains, i.e., W1 ≤ W2 ≤ · · · ≤ W N , but in this case
the ratio between Wk+1 and Wk cannot be too large according to T2. Consequently,
one can find the optimal power allocation via standard convex optimization methods,
e.g., the interior point method.
6.4.2.2
SR Maximization with QoS (SR2)
Then, we consider the SR maximization problem with QoS constraints for MUNOMA, which is given by
max
p
MUaSR2 :
s.t.
N
i=1
N
i=1
Ri
pi ≤ P, p1 ≤ p2 ≤ · · · ≤ p N
Ri ≥ Rimin , i = 1, . . . , N
Similarly, although MUaSR2 is nonconvex in its original formulation, it can be transformed into a convex problem.
In particular, we exploit the same variable transformation: qi = ij=1 p j for i =
1, 2, . . ., N. Then, MUaSR2 is transformed into
max
q
N
k=1
gi (qi )
MUbSR2 : s.t. q N ≤ P
(a1 − 1) /Γ1 ≤ q1 ≤ q2 − q1 ≤ · · · ≤ q N − q N −1
qi−1 ≤ ai qi − εi , i = 2, . . . , N
where
gi (qi ) = log (qi Γi + 1) − log (qi Γi+1 + 1)
(6.28)
for i = 1, . . ., N − 1 and g N (q N ) = log (q N Γ N + 1), ai = 2−Ri and εi =
(1 − ai ) /Γi . According to condition T1 in Theorem 3, the objective in MUbSR2 is
concave and thus MUbSR2 is a convex problem. Therefore, MUbSR2 can also be efficiently solved via convex optimization methods. Moreover, we show that if the power
order constraint p1 ≤ p2 ≤ · · · ≤ p N is absent, the optimal solution to MUbSR2 can
be analytically characterized.
N
ϕi , where ςi = 2 Ri − 1/Γi ,
Proposition 7 Suppose that P ≥ i=1
ϕi =
⎧
⎪
⎨ς1 ,
⎪
⎩max ϕi−1 , ςi 1 + Γi
i−1
j=1
ϕj
i =1
, i = 2, . . . , N
(6.29)
6 Optimal Power Allocation for Downlink NOMA Systems
209
and the power order constraint is absent in MUaSR2 and MUbSR2 . Then, the solution
to MUbSR2 is
ai+1 q̃i+1 − εi+1 , k = 1, . . . , N − 1
q˜i =
(6.30)
P,
k=N
and the solution to problem MUaSR2 is
q˜1 ,
k=1
p̃i =
(1 − ai ) q̃i + εi , k = 2, . . . , N .
(6.31)
Proof Please refer to the proof of Proposition 3 and Lemma 2 in [17].
Then, a natural question is when the solution in Proposition 7 is indeed optimal with
the power order constraint. The answer is given below.
Theorem 4 The solution in (6.31) is optimal for problem MUaSR2 with the power
order constraint if T3: R2min ≥ 1 and
min
Rimin ≥ log 2 − 2−Ri+1 , i = 3, . . . , N .
(6.32)
Proof Please refer to the proof of Theorem 2 in [17].
Corollary 1 Condition T3 in Theorem 4 holds if Rimin ≥ 1 for i = 2, . . ., N.
Theorem 4 indicates that the power order constraint can be omitted without loss of
optimality if the QoS thresholds of the last N − 1 users are not small. Corollary
1 specifies that the QoS thresholds are only required to be no less than 1bps/Hz,
which is usually satisfied in practice. Therefore, for the SR maximization with QoS
constraints, the optimal power allocation is given by Proposition 7 in practical MUNOMA systems.
6.4.3 Optimal Power Allocation for EE Maximization
In this subsection, we investigate the EE maximization for MU-NOMA systems.
6.4.3.1
Weighted EE Maximization (EE1)
The EE maximization with user weights in an MU-NOMA system is formulated as
max η =
p
MUaEE1 :
s.t.
N
i=1 Wi Ri
N
pi
PT + i=1
N
i=1
pi ≤ P
p1 ≤ p2 ≤ · · · ≤ p N .
210
Y. Huang et al.
To address this problem, we follow the similar steps for MUaSR1 to simplify MUaEE1 .
Specifically, using the variable transformation: qi = ij=1 p j for i = 1, 2, . . ., N,
MUaEE1 can be reformulated as
MUbEE1 :
max η =
q
N
k=1 f i (qi )
PT +q N
qN ≤ P
0 ≤ q1 ≤ q2 − q1 ≤ · · · ≤ q N − q N −1
s.t.
where f i (qi ) is defined in (6.27).
Then, we introduce the following objective function:
H (q, α)
N
i=1
f i (qi ) − α PT +
N
pi
(6.33)
i=1
and consider the following problem parameterized by α:
max H (q, α)
MUcEE1
q
: s.t. q N ≤ P
0 ≤ q1 ≤ q2 − q1 ≤ · · · ≤ q N − q N −1 .
According to Lemma 2, the optimal solution to MUbEE1 can be found by solving
MUcEE1 with α chosen such that H ⋆ (α) = 0, where H ⋆ (α) is the optimal objective
value of MUcEE1 . The desirable α can be found via a line search method by exploring
the monotonicity of H ⋆ (α). To solve MUcEE1 , we provide the following result.
Theorem 5 Given Γ1 ≥ Γ2 ≥ · · · ≥ Γ N , MUcEE1 is a convex problem if T1 or T2 in
Theorem 3 holds for i = 1, . . ., N − 1.
Proof Please refer to the proof of Theorem 1 in [17].
Theorem 5 indicates that, under the same condition in Theorem 3, MUcEE1 is a convex
problem. Therefore, one can efficiently compute its optimal solution via optimization
tools, e.g., the interior method.
6.4.3.2
EE Maximization with QoS Constraints (EE2)
Then, we focus on maximizing EE with QoS constraints for MU-NOMA and the
corresponding optimization problem is given by
max η =
p
MUaEE2 : s.t.
N
i=1
N
i=1 Ri
N
PT + i=1
pi
pi ≤ P
p1 ≤ p2 ≤ · · · ≤ p N
Ri ≥ Rimin , i = 1, . . . , N .
6 Optimal Power Allocation for Downlink NOMA Systems
By using the same variable transformation: qi =
MUaEE2 can be transformed into
max η =
q
211
i
j=1
p j for i = 1, 2, . . ., N,
N
k=1 gi (qi )
PT +q N
MUbEE2 : s.t. q N ≤ P
(a1 − 1) /Γ1 ≤ q1 ≤ q2 − q1 ≤ · · · ≤ q N − q N −1
qi−1 ≤ ai qi − εi , i = 2, . . . , N
where gi (qi ) is given in (6.28). Similarly, we introduce the following objective function
N
N
(6.34)
pi ,
gi (qi ) − α PT +
Q (q, α)
i=1
k=1
and consider the problem parameterized by α:
max Q (q, α)
q
MUcEE2 : s.t. q N ≤ P
(a1 − 1) /Γ1 ≤ q1 ≤ q2 − q1 ≤ · · · ≤ q N − q N −1
qi−1 ≤ ai qi − εi , i = 2, . . . , N .
Similarly, to obtain the optimal solution to MUbEE2 , one can solve MUcEE2 for given
α and search α such that the optimal objective value of MUcEE2 satisfies Q ⋆ (α) = 0,
for which we refer the reader to the previous subsection. To solve MUcEE2 , we provide
the following result.
N
ϕi , where ςi = 2 Ri − 1/Γi and
Proposition 8 Suppose that P ≥ i=1
ϕk =
⎧
⎪
⎨ς1 ,
⎪
⎩max ϕi−1 , ςi 1 + Γi
i−1
j=1
ϕj
i =1
, i = 2, . . . , N
,
(6.35)
then MUcEE2 is feasible and convex.
Proof Please refer to the proof of Theorem 1 and Proposition 3 in [17].
Proposition 8 indicates that if the power budget of BS is not too small, MUcEE2 can
be solved by convex optimization methods, e.g., the interior point method.
6.5 MC-NOMA
In this section, we consider the MC-NOMA scheme, where multiple users share
multiple channels. In this case, the resource optimization includes power allocation and channel assignment. However, the joint optimization results in a mixed
212
Y. Huang et al.
integer problem and finding its solution requires exhaustive search [35], which causes
prohibitive computational complexity. Therefore, in practice, power allocation and
channel assignment are often separately and alternatively optimized [26, 28, 35].
In this section, we focus on seeking the optimal power allocation for given channel
assignment.
Note that using SIC at each user’s receiver causes additional complexity, which is
proportional to the number of users on the same channel. Thus, in the multi-channel
case, each channel is often restricted to be shared by two users [25, 26, 36], which is
also beneficial to reduce the error propagation of SIC. In this section, we would also
like to focus on this typical situation. In this case, suppose w.l.o.g. that the CNRs of
UE1,m and UE2,m are ordered as Γ1,m ≥ Γ2,m . Then, the rates of UE1,m and UE2,m
on channel m are given, respectively, by
R1,m = Bc log 1 + p1,m Γ1,m , R2,m = Bc log 1 +
p2,m Γ2,m
.
p1,m Γ2,m + 1
(6.36)
6.5.1 Optimal Power Allocation for MMF
In MC-NOMA systems, the power allocation problem under the MMF criterion is
given by
MCaMMF :
max
p1 ,p2
min
m=1,...,M
R1,m ( p1,m , p2,m ), R2,m ( p1,m , p2,m )
s.t. 0 ≤ p1,m ≤ p2,m , m = 1, . . . , M,
M
m=1
p1,m + p2,m ≤ P
M
M
and p2 = { p2,m }m=1
. Note that MCaMMF is a nonconvex probwhere p1 = { p1,m }m=1
M
, where qm
lem. To address it, we first introduce auxiliary variables q = {qm }m=1
represents the power budget for channel m with p1,m + p2,m = qm . Suppose that
M
the channel power budgets {qm }m=1
are given. Then, MCaMMF is decomposed into a
group of subproblems and each subproblem is same with T U MMF in the two-user
case with P replaced by qm .
With given channel power budget qm , the optimal power allocation for the two
users on channel m has been provided in Proposition 1. We can use this result to
further optimize the power budgets {qm }. According to MCaMMF and T U MMF , the
corresponding power budget optimization problem is
max
MCbMMF :
q
s.t.
min
m=1,...,M
M
m=1
f mMMF⋆ (qm )
qm ≤ P, q ≥ 0
where f mMMF⋆ (qm ) is the optimal objective value of T U MMF for each channel m.
Using Proposition 1, we obtain
6 Optimal Power Allocation for Downlink NOMA Systems
⎛
f mMMF⋆ Bc log ⎝
Γ2,m − Γ1,m +
213
⎞
2
Γ1,m + Γ2,m 2 + 4Γ1,m Γ2,m
qm
⎠.
2Γ2,m
(6.37)
Then, we show that MCbMMF has a closed-form solution.
Theorem 6 The optimal solution to MCbMMF is given by
qm⋆
Z (λ) Γ2,m + Γ1,m (Z (λ) − 1)
, ∀m,
=
Γ1,m Γ2,m
(6.38)
where
Z (λ) X +
X2 +
Bc
2λ
M
m=1
and λ is chosen such that
M
m=1
1/Γ1,m
, X
M
m=1
qm⋆ = P.
Γ2,m − Γ1,m / Γ1,m Γ2,m
,
M
4 m=1
1/Γ1,m
(6.39)
Proof Please refer to the proof of Theorem 1 in [22].
Consequently, the optimal MC-NOMA power allocation under the MMF criterion is
fully characterized by Theorem 6 and Proposition 1. It follows from (6.38) that qm⋆
M
is monotonically decreasing in λ, so the optimal λ satisfying m=1
qm⋆ = P can be
efficiently found via a simple bisection method.
6.5.2 Optimal Power Allocation for SR Maximization
In this subsection, we investigate the SR maximization problem with weights or QoS
constraints in MC-NOMA systems.
6.5.2.1
Weighted SR Maximization (SR1)
With given channel assignment, the problem of maximizing the weighted sum rate
is formulated as the following power allocation problem:
max
MCaSR1 :
M
p1 ,p2 m=1
W1,m R1,m ( p1,m , p2,m ) + W2,m R2,m ( p1,m , p2,m )
s.t. 0 ≤ p1,m ≤ p2,m , m = 1, . . . , M,
M
m=1
p1,m + p2,m ≤ P
M
that represent the
To solve it, similarly we introduce auxiliary variables q = {qm }m=1
power budgets on each channel m with p1,m + p2,m = qm . Then, MCaSR1 is decom-
214
Y. Huang et al.
posed into a group of subproblems, where each subproblem is the same with T U SR1
except P replaced by qm and its solution has been provided in Proposition 2.
Next, we further optimize the power budget qm for each channel m. According
to Remark 3, to guarantee that the NOMA
Msystem is SIC stable, it is reasonable to
Θm for some positive Θm . Then, from
assume that qm ≥ Θm > 2Ωm and P ≥ m=1
MCaSR1 and T U SR1 , the corresponding power budget optimization problem is given
by
M
f SR1⋆ (qm )
max
q m=1 m
SR1
MCb :
M
s.t.
qm ≤ P, qm ≥ Θm , ∀m
m=1
where f mSR1⋆ (qm ) is the optimal objective value of each subproblem. Using Proposition 2, we obtain
qm Γ2,m + 1
.
(6.40)
f mSR1⋆ (qm ) = W1,m log 1 + Ωm Γ1,m + W2,m log
Ωm Γ2,m + 1
It is easily seen that f mSR1⋆ (qm ) is a concave function, so MCbSR1 is a convex problem,
whose solution is provided in the following result.
Theorem 7 The optimal solution to MCbSR1 is given by
qm⋆ =
where λ is chosen such that
M
m=1
W2,m Bc
1
−
λ
Γ2,m
∞
(6.41)
,
Θm
qm⋆ = P.
Proof The solution to MCbSR1 is given by the well-known waterfilling form.
Consequently, the optimal power allocation for the weighted sum rate maximization
in MC-NOMA systems is jointly characterized by Theorem 7 and Proposition 2
under the SIC stability.
6.5.2.2
SR Maximization with QoS (SR2)
Now, we consider maximizing the SR with QoS constraints. In this case, the power
allocation problem is given by
max
MCaSR2 :
M
p1 ,p2 m=1
R1,m ( p1,m , p2,m ) + R2,m ( p1,m , p2,m )
s.t. 0 ≤ p1,m ≤ p2,m , m = 1, . . . , M,
M
m=1
( p1,m + p2,m ) ≤ P
min
Rn,m ≥ Rn,m
, n = 1, 2, m = 1, . . . , M.
6 Optimal Power Allocation for Downlink NOMA Systems
215
We use the similar method to address MCaSR2 . By introducing the power budget qm
on each channel m, MCaSR2 decomposes into several subproblems and each of them
has the same structure as T U SR2 . Thus, the optimal solution to each subproblem is
given in Proposition 3 with P replaced by qm .
Then, we focus on optimizing the power budget qm for each channel. Similarly,
is SIC stable, we assume that
according to Remark
4,Mto guarantee the NOMA system
qm ≥ Υm and P ≥ m=1
Υm . According to MCaSR2 and T U SR2 , the corresponding
power budget optimization problem is as follows
MCbSR2
:
max
q
s.t.
M
m=1
M
m=1
f mSR2⋆ (qm )
qm ≤ P, qm ≥ Υm , ∀m
where f mSR2⋆ (qm ) is the optimal objective value of each subproblem and given by
f mSR2⋆ (qm ) = Bc log
A2,m Γ2,m − A2,m Γ1,m + Γ1,m Γ2,m qm + Γ1,m
A2,m Γ2,m
min
+ R2,m
.
(6.42)
Since f mSR2⋆ (qm ) is a concave function, MCbSR2 is a convex problem, whose solution
is also given in a waterfilling form.
Theorem 8 The optimal solution to MCbSR2 is given by
qm⋆ =
where λ is chosen such that
A2,m
Bc
A2,m
1
−
+
−
λ
Γ1,m
Γ2,m
Γ2,m
M
m=1
∞
,
(6.43)
Υm
qm⋆ = P.
Proof The proof is simple and thus omitted.
Therefore, the optimal power allocation for the SR maximization with QoS constraints in MC-NOMA systems is jointly characterized by Proposition 3 and
Theorem 8.
6.5.3 Optimal Power Allocation for EE Maximization
In this subsection, we investigate the optimal power allocation for maximizing the
EE with weights or QoS constraints in MC-NOMA systems.
216
Y. Huang et al.
6.5.3.1
EE Maximization with Weights (EE1)
With given channel assignment, the problem of maximizing the weighted EE is
formulated as the following power allocation problem:
MCaEE1
: max
p1 ,p2
M
m=1
W1,m R1,m ( p1,m , p2,m ) + W2,m R2,m ( p1,m , p2,m )
M
PT + m=1
p1,m + p2,m
s.t. 0 ≤ p1,m ≤ p2,m , m = 1, . . . , M,
M
m=1
p1,m + p2,m ≤ P.
The difficulties in solving MCaEE1 lie in its nonconvex and fractional objective. In
the following, we will show that this problem can also be optimally solved.
We use the similar trick to address this problem, i.e., introducing the auxiliary variM
with p1,m + p2,m = qm for each channel m. Then, MCaEE1 is decomables {qm }m=1
posed into a group of subproblems. Each subproblem is the same with T U SR1 except
P replaced by qm , and thus, its solution is provided in Proposition 1.
Then, we concentrate on searching the optimal power budget qm for each channel.
Similarly, to guarantee
NOMA system is SIC stable, it is assumed that qm ≥
the
M
Θm > 2Ωm and P ≥ m=1
Θm for some positive Θm . According to Proposition 1
and MCaEE1 , the power budget optimization problem is formulated as
MCbEE1 : max η(q)
q
s.t.
M
m=1
M
f mSR1⋆ (qm )
M
PT + m=1 qm
m=1
(6.44)
qm ≤ P, qm ≥ Θm , ∀m
where f mSR1⋆ (qm ) is given in (6.40). Although f mSR1⋆ (qm ) is a concave function,
MCbEE1 is nonconvex due to the fraction form. To solve it, we introduce the following objective function:
H (q, α)
M
m=1
=
f mSR1⋆ (qm ) − α PT +
M
m=1
R̃1,m
M
qm
m=1
qm Γ2,m + 1
+ W2,m log
Ωm Γ2,m + 1
− α PT +
M
qm , (6.45)
m=1
where R̃1,m W1,m log 1 + Ωm Γ1,m and α is a positive parameter. Then, we consider the following convex problem with given α:
6 Optimal Power Allocation for Downlink NOMA Systems
217
max H (q, α)
MCcEE1
q
:
s.t.
M
m=1
qm ≤ P, qm ≥ Θm , ∀m.
According to Lemma 2, the optimal solution to MCbEE1 can be found by solving
MCcEE1 with given α and then updating α until H ⋆ (α) = 0. Hence, we first solve
MCcEE1 with given α, whose solution is provided in the following result.
Theorem 9 The optimal solution to MCcEE1 is
qm⋆ =
where λ is chosen such that
M
m=1
W2,m Bc
1
−
α+λ
Γ2,m
∞
(6.46)
,
Θm
qm⋆ = P.
Proof The solution is obtained by exploiting the KKT conditions of MCcEE1 .
After the optimal solution to MCcEE1 is obtained, we shall find an α such that H ⋆ (α) =
0. Since H ⋆ (α) is monotonic in α, one can use the bisection method to find α.
Thereby, the optimal power allocation for the weighted EE maximization in MCNOMA systems is provided Proposition 2 and Theorem 9.
6.5.3.2
EE Maximization with QoS (EE2)
In this part, we consider maximizing the EE with QoS constraints. The corresponding
power allocation problem is given by
max
p1 ,p2
MCaEE2 :
M
m=1
R1,m p1,m , p2,m + R2,m p1,m , p2,m
M
PT + m=1
p1,m + p2,m
s.t. 0 ≤ p1,m ≤ p2,m , m = 1, . . . , M,
M
m=1
( p1,m + p2,m ) ≤ P
min
, l = 1, 2, m = 1, . . . , M.
Rl,m ≥ Rl,m
M
We can use the similar method to solve MCaEE2 . Briefly, we also adopt {qm }m=1
EE2
with p1,m + p2,m = qm and decompose MCa into a group of subproblems, whose
solution is coincided with T U SR2 and provided in Proposition 3.
Next, we optimize the
channel power budget qm for each channel. First, we assume
M
that qm ≥ Υm and P ≥ m=1
Υm to guarantee the SIC stability. Then, according to
EE2
Proposition 3 and MCa , the power budget optimization problem is given by
218
Y. Huang et al.
MCbEE2
: max η(q)
q
s.t.
M
m=1
M
f mSR2⋆ (qm )
M
PT + m=1 qm
m=1
qm ≤ P, qm ≥ ϒm , ∀m
where f mSR2⋆ (qm ) is given in (6.42). To solve MCbEE2 , we introduce the objective
function parameterized by α:
Q(q, α)
M
f mSR2⋆ (qm )
m=1
=
M
W1,m log
m=1
−α PT +
− α PT +
M
M
qm
m=1
A2,m Γ2,m − A2,m Γ1,m + Γ1,m Γ2,m qm + Γ1,m
A2,m Γ2,m
min
+ R2,m
(6.47)
qm ,
m=1
and formulate the following problem with given α:
max Q (q, α)
MCcEE2
q
:
s.t.
M
m=1
qm ≤ P, qm ≥ ϒm , ∀m.
Then, from Lemma 2, we shall solve MCcEE2 , which is a convex problem since
Q(q, α) is concave in q. The optimal solution to MCcEE2 is provided below.
Theorem 10 The optimal solution to MCcEE2 is
qm⋆ =
A2,m
A2,m
1
W1,m Bc
−
+
−
λ+α
Γ1,m
Γ2,m
Γ2,m
where λ is chosen such that
M
m=1
∞
,
(6.48)
ϒm
qm⋆ = P .
Proof The solution is obtained by exploiting the KKT conditions of MCcEE2 .
Then, we can exploit the bisection method to find an α such that the optimal objective
value of MCcEE2 satisfies Q ⋆ (α) = 0. Consequently, the optimal power allocation for
the EE maximization with QoS constraints in MC-NOMA systems is obtained by
using Theorem 10 and Proposition 3.
6 Optimal Power Allocation for Downlink NOMA Systems
219
6.6 Numerical Results
This section evaluates the performance of the optimal power allocation investigated
in this chapter. In simulations, the BS is located in the cell center and the users
are randomly distributed in a circular range with a radius of 500 m. The minimum
distance between users is set to be 40 m, and the minimum distance between the users
and the BS is 50 m. Each channel coefficient follows an i.i.d. Gaussian distribution
as g ∼ CN (0, 1) and the path loss exponent is ρ = 2. The total power budget of the
BS is P = 41 dBm and the circuit power consumption is PT =30 dBm. The noise
power is σ 2 = B N0 /M, where the bandwidth is B = 5 MHz and the noise power
spectral density is N0 = −174 dBm.
First, we evaluate the performance of the proposed optimal power solutions for
two-user NOMA (N = 2) and MU-NOMA (N = 6) systems. The user weights satisfy Wi+1 /Wi = 0.5 for i = 1, . . ., N − 1 and the QoS thresholds to be Rimin = 2
bps/Hz for i = 1, . . ., N. In addition, we compare the NOMA schemes with OFDMA
and the DC (difference of two convex functions) approach in [26], where the power
allocation is optimized via waterfilling and via DC programming, respectively.
Figure 6.2 shows the minimum user rates of the two-user NOMA and MU-NOMA
schemes using the optimal power allocation under the MMF criterion and the minimum user rate of the OFDMA scheme for different total power budgets and user
numbers. The minimum user rate in the NOMA system is higher than that in the
OFDMA system especially in the two-user case, implying that NOMA provides
better fairness than OFDMA.
106
2
1.8
NOMA N=2
NOMA N=6
OFDMA N=2
OFDMA N=6
Minimum User Rate (Mbps)
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
2
3
4
5
6
7
8
9
10
Power of BS (Watt)
Fig. 6.2 Minimum user rate for different number of users versus BS power
11
12
220
Y. Huang et al.
106
16
NOMA N=6
NOMA N=2
NOMA DC N=2
OFDMA N=2
1.8
1.6
12
Sum Rate (Mbits/Joule)
Weighted Sum Rate (Mbits/Joule)
14
107
2
10
8
6
1.4
NOMA N=6
NOMA N=2
NOMA DC N=2
OFDMA N=2
1.2
1
0.8
0.6
4
2
0.4
0
5
10
Power of BS (Watt)
15
0.2
0
5
10
15
Power of BS (Watt)
Fig. 6.3 Sum rate versus BS power
In Fig. 6.3 the left subfigure shows the weighted sum rate and the right subfigure
shows the sum rate with QoS constraints. Here, in each subfigure, we compare the
proposed methods with the OFDMA scheme and the NOMA scheme using DC programming in the two-user case. While NOMA outperforms OFDMA, NOMA with
the optimal power allocation also achieves a higher sum rate than the DC approach,
as the DC approach generally leads to a suboptimal power allocation. Meanwhile,
as expected, the (weighted) sum rate increases with the user number, implying the
potential of NOMA. In Fig. 6.4, the similar phenomenon can be observed, i.e., NOMA
using the optimal power allocation outperforms OFDMA as well as the (suboptimal)
DC approach in terms of energy efficiency.
Then, we show the performance of the optimal power allocation in MC-NOMA
systems. The user weights are set to be W1,m = 0.9 and W2,m = 1.1 for ∀m and
min
the QoS thresholds are set to be Rl,m
= 2 bps/Hz for l = 1, 2, ∀m. In Fig. 6.5, we
compare the joint resource allocation (JRA) method, which uses the optimal power
allocation and the matching algorithm [22, 26] for channel assignment, with the
exhaustive search (ES), which provides the jointly optimal solution but has high
complexity. We set the number of users N = 6 and the power budget of the BS
ranges from 2 to 12 W. From Fig. 6.5, the performance of JRA is very close to the
6 Optimal Power Allocation for Downlink NOMA Systems
106
NOMA N=6
NOMA N=2
NOMA DC N=2
OFDMA N=2
3
NOMA N=6
NOMA N=2
NOMA DC N=2
OFDMA N=2
3.5
2.5
2
1.5
1
3
2.5
2
1.5
1
0.5
0
106
4
Energy Efficiency (Mbits/Joule)
Weighted Energy Efficiency (Mbits/Joule)
3.5
221
0.5
0
5
10
0
15
0
Power of BS (Watt)
5
10
15
Power of BS (Watt)
140
Sum rate (Mbps)
20
15
10
MMF ES
MMF JRA
5
2
4
6
8
10
120
100
SR1 ES
SR1 JRA
SR2 ES
SR2 JRA
80
60
12
2
4
Power of BS (Watt)
50
EE1 ES
EE1 JRA
40
30
20
10
2
4
6
8
10
6
8
10
12
Power of BS (Watt)
12
Energy efficiency (Mbps/Joule)
Weighted Energy efficiency (Mbps/Joule)
Minimum User Rate (Mbps)
Fig. 6.4 Energy efficiency versus BS power
Power of BS (Watt)
Fig. 6.5 Comparison with the exhaustive search (ES)
50
EE2 ES
EE2 JRA
40
30
20
10
2
4
6
8
10
Power of BS (Watt)
12
222
Y. Huang et al.
globally optimal value and the maximum gap is less than 5%. Therefore, the optimal
power allocation method along with efficient (suboptimal) matching algorithm is
able to achieve near-optimal performance with low complexity.
6.7 Conclusion
In this chapter, we discussed a promising multiple access technology, i.e., NOMA,
for 5G networks and focused on the key problem of power allocation in NOMA
systems. We have investigated the optimal power allocation for different NOMA
schemes, including the two-user MU-NOMA, and MC-NOMA schemes. The optimal
power allocation was derived under different performance measures, including the
maximin fairness, weighted sum rate, and energy efficiency, wherein user weights
or QoS constraints were also considered. We showed that in most cases the optimal
NOMA power allocation admits an analytical solution, while in other cases it can be
numerically computed via convex optimization methods.
Appendix
A. Proof of Theorem 1
Since the constraints in T UcEE1 are all linear, it suffices to investigate the concavity
of H ( p1 , q, α). The second-order derivative of H ( p1 , q, α) with respect to p1 is
1
∂2 H
=
ln 2
∂ p12
ΥΘ
( p1 Γ1 + 1)2 ( p1 Γ2 + 1)2
(6.49)
,
√
√
√
where√ Υ = W√
+ 1) + √W1 BΓ
2 BΓ2 ( p1 Γ1√
1 ( p1 Γ2 + 1) and Θ = W2 B
Γ2 − W1 BΓ1 + BΓ1 Γ2 p1 W2 − W1 . Given Γ1 ≥ Γ2 , if W1 ≥ W2 , then
∂2 H
≤ 0. On the other hand, with q ≤ P and if C2 holds, we have
∂ p2
1
W2 BΓ2 −
W1 BΓ1 +
√
BΓ1 Γ2 p1
W2 −
W1
W2 BΓ2 −
W1 BΓ1 +
W2 −
W1
√
BΓ1 Γ2 P ≤ 0,
(6.50)
2
2
also implying ∂∂ pH2 ≤ 0. Following the similar manner, it can be verified that ∂∂qH2 ≤ 0,
∂2 H
∂q∂ p1
= 0 and
1
2
∂ H
∂ p1 ∂q
= 0. Therefore, the Hessian matrix
∂2 H
∂2 H
∂ p12 ∂q∂ p1
∂2 H
∂2 H
∂ p1 ∂q ∂q 2
is a negative semidefinite matrix, indicating that H ( p1 , q, α) is a concave in ( p1 , q).
6 Optimal Power Allocation for Downlink NOMA Systems
223
B. Proof of Proposition 4
The Lagrange of T UcEE1 is given by
L = W1 R1 ( p1 ) + W2 R2 ( p1 , q) − α (PT + q) + μ (q − 2 p1 ) − λ (q − P) (6.51)
with Lagrange multipliers μ and λ ≥ 0. According to Theorem 1, T UcEE1 is a convex
problem under condition C1 or C2. Therefore, its optimal solution is characterized
by the following Karush–Kuhn–Tucker (KKT) conditions:
W2 BΓ2
∂L
W1 BΓ1
=
−
− 2μ = 0,
∂ p1
ln 2 (1 + p1 Γ1 ) ln 2 (1 + p1 Γ2 )
(6.52)
W2 BΓ2
∂L
=
− α + μ − λ = 0,
∂q
ln 2 (1 + qΓ2 )
(6.53)
μ (q − 2 p1 ) = 0,
(6.54)
λ (q − P) = 0.
(6.55)
According to Definition 1, if p1 = q/2, then the NOMA system is SIC-unstable.
Therefore, from (6.54), considering the SIC stability, we have μ = 0. Hence, from
(6.52) we obtain the optimal p1⋆ = Ω. It follows from (6.55) that if q < P, then
λ = 0. Then, from (6.53) we obtain
2Ω ≤ q =
1
W2 B
−
< P.
α ln 2 Γ2
(6.56)
On the other hand, if q = P, then from (6.53) we have
λ=
which leads to
W2 BΓ2
− α ≥ 0,
ln 2 (1 + PΓ2 )
1
W2 B
−
≥ P.
α ln 2 Γ2
Therefore, the optimal q is given by q ⋆ =
W2 B
α ln 2
−
(6.57)
(6.58)
P
1
.
Γ2 2Ω
224
Y. Huang et al.
C. Proof of Proposition 5
The Lagrange of T UcEE2 is given by
L = R1 ( p1 ) + R2 ( p1 , q) − α (PT + q) + μ (q − 2 p1 ) − λ (q − P)
A1 − 1
+ σ2 (1 + qΓ2 − A2 − A2 p1 Γ2 ) ,
+ σ1 p1 −
Γ1
(6.59)
where μ, λ, σ1 , and σ2 are the Lagrange multipliers. The optimal solution is characterized by the following KKT conditions:
BΓ1
BΓ2
∂L
−
− 2μ + σ1 − σ2 A2 Γ2 = 0,
=
∂ p1
ln 2 (1 + p1 Γ1 ) ln 2 (1 + p1 Γ2 )
(6.60)
BΓ2
∂L
=
− α + μ − λ + σ2 Γ2 = 0,
∂q
ln 2 (1 + qΓ2 )
(6.61)
μ (q − 2 p1 ) = 0,
(6.62)
λ (q − P) = 0,
(6.63)
(6.64)
σ1
A1 − 1
p1 −
Γ1
= 0,
σ2 (1 + qΓ2 − A2 − A2 p1 Γ2 ) = 0.
(6.65)
In (6.62), considering the SIC stability, we have q > 2 p1 and hence μ = 0. Note
that σ2 = 0. To see this, if σ2 = 0, according to (6.60), we have
BΓ1
BΓ2
−
+ σ1 = 0
ln 2 (1 + p1 Γ1 ) ln 2 (1 + p1 Γ2 )
(6.66)
BΓ2
BΓ1
− ln 2(1+
+ σ1 > 0 with Γ1 ≥ Γ2 .
which, however, does not hold since ln 2(1+
p1 Γ1 )
p1 Γ2 )
We consider two cases: (1) σ1 = 0, σ2 = 0; and (2) σ1 = 0, σ2 = 0. First, if σ1 =
0, σ2 = 0, the optimal solution can be easily obtained as
p1⋆ =
1 + q ⋆ Γ2 − A 2 ⋆
, q =Υ
A 2 Γ2
(6.67)
from (6.64) and (6.65). Then, if σ1 = 0, σ2 = 0, according to (6.60) and (6.61), we
have
1
1
A 2 Γ2
= (α + λ) A2 .
+
(6.68)
−
1/Γ1 + p1
1/Γ2 + p1
(1 + qΓ2 )
6 Optimal Power Allocation for Downlink NOMA Systems
From (6.65), we obtain p1 =
1+qΓ2 −A2
,
A 2 Γ2
q⋆ =
225
which along with (6.68) leads to
A2
A2 − 1
1
−
+
.
α+λ
Γ1
Γ2
(6.69)
It follows from (6.63) that if q < P, then λ = 0. From (6.69), we obtain
ϒ ≤q=
1
A2
A2 − 1
−
+
< P.
α
Γ1
Γ2
(6.70)
On the other hand, if q = P, then from (6.53) we have
λ=
Γ1 Γ2
− α ≥ 0,
A2 Γ2 − (A2 − 1) Γ1 + PΓ2 Γ1
which leads to
A2
A2 − 1
1
−
+
≥ P.
α
Γ1
Γ2
Therefore, optimal q is given by q ⋆ =
W2 B
α ln 2
−
(6.71)
(6.72)
P
1
.
Γ2 ϒ
D. Proof of Theorem 2
i−1
p
Γ
+
1
can be transj=1 j i
j=1
t
t
N (2 −1)2(N −i)t
formed into qi = qi−1 2t + 2 Γ−1
. Thus, we obtain P = q N = i=1
≥ χ,
Γi
i
implying t ≥ 1 and pi ≥ pi−1 for i = 2, . . . , N . Therefore, this solution satisfies the
power order constraint.
Let qi =
i
p j , then q N = P and pi =
2t −1
Γi
References
1. J.G. Andrews, S. Buzzi, W. Choi, S.V. Hanly, A. Lozano, A.C. Soong, J.C. Zhang, What will
5G be? IEEE J. Sel. Areas Commun. 32(6), 1065–1082 (2014)
2. V. Jungnickel, K. Manolakis, W. Zirwas, B. Panzner, V. Braun, M. Lossow, M. Sternad, R. Apelfrojd, T. Svensson, The role of small cells, coordinated multipoint, and massive MIMO in 5G.
IEEE Commun. Mag. 52(5), 44–51 (2014)
3. M. Xiao, S. Mumtaz, Y. Huang, L. Dai, Y. Li, M. Matthaiou, G.K. Karagiannidis, E. Bjrnson,
K. Yang, I. Chih-Lin, A. Ghosh, Millimeter wave communications for future mobile networks.
IEEE J. Sel. Areas Commun. 35(9), 1909–1935 (2017)
4. J.G. Andrews, H. Claussen, M. Dohler, S. Rangan, M.C. Reed, Femtocells: past, present, and
future. IEEE J. Sel. Areas Commun. 30(3), 497–508 (2012)
5. J. Wang, W. Guan, Y. Huang, R. Schober, X. You, Distributed optimization of hierarchical
small cell networks: a GNEP framework. IEEE J. Sel. Areas Commun. 35(2), 249–264 (2017)
226
Y. Huang et al.
6. Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, K. Higuchi, Non-orthogonal
multiple access (NOMA) for cellular future radio access, in Proceeding of IEEE Vehicular
Technology Conference (VTC Spring) Dresden, Germany, June 2013, pp. 1–5
7. Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, I. Chih-Lin, H.V. Poor, Application of nonorthogonal multiple access in LTE and 5G networks, IEEE Commun. Mag. 55(2), 185–191
(2017)
8. L. Dai, B. Wang, Y. Yuan, S. Han, I. Chih-Lin, Z. Wang, Non-orthogonal multiple access
for 5G: solutions, challenges, opportunities, and future research trends. IEEE Commun. Mag.
53(9), 74–81 (2015)
9. Z. Ding, R. Schober, H.V. Poor, A general MIMO framework for NOMA downlink and uplink
transmission based on signal alignment. IEEE Trans. Wirel. Commun. 15(6), 4438–4454 (2016)
10. Z. Ding, F. Adachi, H.V. Poor, The application of MIMO to non-orthogonal multiple access.
IEEE Trans. Wirel. Commun. 15(1), 537–552 (2016)
11. W. Shin, M. Vaezi, B. Lee, D.J. Love, J. Lee, H.V. Poor, Non-orthogonal multiple access
in multi-cell networks: theory, performance, and practical challenges. IEEE Commun. Mag.
55(10), 176–183 (2017)
12. X. Zhang, Q. Gao, C. Gong, Z. Xu, User grouping and power allocation for NOMA visible
light communication multi-cell networks. IEEE Commun. Lett. 21(4), 777–780 (2017)
13. Y. Huang, C. Zhang, J. Wang, Y. Jing, L. Yang, X. You, Signal processing for MIMO-NOMA:
present and future challenges. IEEE Wirel. Commun. 25(2), 32–38 (2018)
14. L. Zhang, M. Xiao, G. Wu, M. Alam, Y.C. Liang, S. Li, A survey of advanced techniques for
spectrum sharing in 5G networks. IEEE Wirel. Commun. 24(5), 44–51 (2017)
15. C.-L. Wang, J.-Y. Chen, Y.-J. Chen, Power allocation for a downlink non-orthogonal multiple
access system. IEEE Wirel. Commun. Lett. 5(5), 532–535 (2016)
16. Z. Yang, Z. Ding, P. Fan, N. Al-Dhahir, A general power allocation scheme to guarantee quality
of service in downlink and uplink NOMA systems. IEEE Trans. Wirel. Commun. 15(11), 7244–
7257 (2016)
17. J. Wang, Q. Peng, Y. Huang, H.M. Wang, X. You, Convexity of weighted sum rate maximization
in NOMA systems. IEEE Signal Process. Lett. 24(9), 1323–1327 (2017)
18. J. Choi, Power allocation for max-sum rate and max-min rate proportional fairness in NOMA.
IEEE Commun. Lett. 20(10), 2055–2058 (2016)
19. J. Cui, Z. Ding, P. Fan, A novel power allocation scheme under outage constraints in NOMA
systems. IEEE Signal Process. Lett. 23(9), 1226–1230 (2016)
20. S. Timotheou, I. Krikidis, Fairness for non-orthogonal multiple access in 5G systems. IEEE
Signal Process. Lett. 22(10), 1647–1651 (2015)
21. Y. Zhang, H.M. Wang, T.X. Zheng, Q. Yang, Energy-efficient transmission design in nonorthogonal multiple access. IEEE Trans. Veh. Technol. 66(3), 2852–2857 (2017)
22. J. Zhu, J. Wang, Y. Huang, S. He, X. You, L. Yang, On optimal power allocation for downlink
non-orthogonal multiple access systems. IEEE J. Sel. Areas Commun. 35(12), 2744–2757
(2017)
23. F. Fang, H. Zhang, J. Cheng, S. Roy, V.C.M. Leung, Joint user scheduling and power allocation
optimization for energy efficient NOMA systems with imperfect CSI. IEEE J. Sel. Areas
Commun. 35(12), 2874–2885 (2017)
24. B. Di, S. Bayat, L. Song, Y. Li, Radio resource allocation for downlink non-orthogonal multiple
access (NOMA) networks using matching theory, in Proc. of IEEE Global Communication
Conference (GLOBECOM), pp. 1–6
25. Z. Ding, M. Peng, H.V. Poor, Cooperative non-orthogonal multiple access in 5G systems. IEEE
Commun. Lett. 19(8), 1462–1465 (2015)
26. F. Fang, H. Zhang, J. Cheng, V.C. Leung, Energy-efficient resource allocation for downlink
non-orthogonal multiple access network. IEEE Trans. Commun. 64(9), 3722–3732 (2016)
27. L. Lei, D. Yuan, C.K. Ho, S. Sun, Joint optimization of power and channel allocation with
non-orthogonal multiple access for 5G cellular systems. in Proceeding of IEEE Global Communication Conference (GLOBECOM), San Diego, CA, Dec 2015, pp. 1–6
6 Optimal Power Allocation for Downlink NOMA Systems
227
28. P. Parida, S.S. Das, Power allocation in OFDM based NOMA systems: a DC programming
approach, in Proceeding of IEEE Globecom Workshops, Dec 2014, pp. 1026–1031
29. Y. Sun, D.W.K. Ng, Z. Ding, R. Schober, Optimal joint power and subcarrier allocation for
full-duplex multicarrier non-orthogonal multiple access systems. IEEE Trans. Commun. 65(3),
1077–1091 (2017)
30. M.R. Hojeij, J.Farah, C.A. Nour, C. Douillard, Resource allocation in downlink non-orthogonal
multiple access (NOMA) for future radio access, in Proceeding of IEEE Vehicular Technology
Conference (VTC Spring), Dresden, Germany, May 2015, pp. 1–6
31. Z. Wei, J. Yuan, D.W.K. Ng, M. Elkashlan, Z. Ding, A Survey of Downlink Non-orthogonal
Multiple Access for 5G Wireless Communication Networks. CoRR. arXiv: 1609.01856, https://
dblp.org/rec/bib/journals/corr/WeiYNED16 (2016)
32. M.S. Ali, H. Tabassum, E. Hossain, Dynamic user clustering and power allocation for uplink
and downlink non-orthogonal multiple access (NOMA) systems. IEEE Access 4, 6325–6343
(2016)
33. Z. Ding, Z. Yang, P. Fan, H.V. Poor, On the performance of non-orthogonal multiple access
in 5G systems with randomly deployed users. IEEE Signal Process. Lett. 21(12), 1501–1505
(2014)
34. W. Dinkelbach, On nonlinear fractional programming. Manag. Sci. 13(7), 492–498 (1967)
35. S. Zhang, B. Di, L. Song, Y. Li, Radio resource allocation for non-orthogonal multiple access
(NOMA) relay network using matching game,” in Proceeding of IEEE International Conference Communication (ICC), Kuala Lumpur, Malaysia, May 2016, pp. 1–6
36. Z. Ding, P. Fan, H.V. Poor, Impact of user pairing on 5G nonorthogonal multiple-access downlink transmissions. IEEE Trans. Veh. Technol. 65(8), 6010–6023 (2016)
Chapter 7
On the Design of Multiple-Antenna
Non-Orthogonal Multiple Access
Xiaoming Chen, Zhaoyang Zhang, Caijun Zhong
and Derrick Wing Kwan Ng
7.1 Introduction
Current wireless communications in general adopt various types of orthogonal multiple access (OMA) technologies for serving multiple users, such as time division
multiple access (TDMA), frequency division multiple access (FDMA), and code
division multiple access (CDMA), where one resource block is exclusively allocated
to one mobile user (MU) to avoid possible multiuser interference. In practice, the
OMA technologies are relatively easy to implement, albeit at the cost of low spectral
efficiency. Recently, with the rapid development of mobile Internet and proliferation of mobile devices, it is expected that future wireless communication systems
should be able to support massive connectivity, which is an extremely challenging
task for the OMA technologies with limited radio resources. Responding to this,
non-orthogonal multiple access (NOMA) has been recently proposed as a promising
access technology for the fifth-generation (5G) mobile communication systems, due
to its potential in achieving high spectral efficiency and supporting massive access
[1–4].
X. Chen (B) · Z. Zhang · C. Zhong
College of Information Science and Electronic Engineering, Zhejiang University,
Hangzhou, China
e-mail: chen_xiaoming@zju.edu.cn
Z. Zhang
e-mail: ning_ming@zju.edu.cn
C. Zhong
e-mail: caijunzhong@zju.edu.cn
D. W. K. Ng
School of Electrical Engineering and Telecommunications,
The University of New South Wales, Sydney, NSW, Australia
e-mail: w.k.ng@unsw.edu.au
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_7
229
230
X. Chen et al.
The principle of NOMA is to exploit the power domain to simultaneously serve
multiple MUs utilizing the same radio resources [5–7], with the aid of sophisticated
successive interference cancellation (SIC) receivers [8, 9]. Despite the adoption of
SIC, inter-user interference still exists except for the MU with the strongest channel gain, which limits the overall system performance [10]. To address this issue,
power allocation has been considered as an effective method to harness multiuser
interference [11, 12]. Since the overall performance is limited by the MUs with weak
channel conditions, it is intuitive to allocate more power to the weak MUs and less
power to the strong MU in order to enhance the effective channel gain and minimize
the interference to the weak MUs [13]. For the specific two-user case, the optimal
power allocation scheme was studied in [14], and [15] proposed two sub-optimal
power allocation schemes exploiting the Karush–Kuhn–Tucker (KKT) conditions,
while the issue of quality of service (QoS) requirements of NONA systems was
investigated in [16]. For the case with arbitrary number of users, the computational
complexity of performing SIC increases substantially and the design of the optimal
power allocation becomes intractable. To facilitate an effective system design, clustering and user pairing have been proposed [17, 18]. Generally speaking, multiple
MUs with distinctive channel gains are selected to form a cluster, in which SIC is
conducted to mitigate the interference [19, 20]. In general, a small cluster consisting
a small number of MUs implies low complexity of SIC, but leads to high intercluster interference. Thus, it makes sense to dynamically adjust the size of a cluster
according to performance requirements and system parameters, so as to achieve a
balance between implementation complexity and interference mitigation [21]. However, dynamic clustering is not able to reduce the inter-cluster interference, indicating
the necessity of carrying out dynamic clustering in combination with efficient interference mitigation schemes.
It is well known that the multiple-antenna technology is a powerful interference
mitigation scheme [22–25], hence, can be naturally applied to NOMA systems [26,
27]. In [28], the authors proposed a beamforming scheme for combating inter-cluster
and intra-cluster interference in a NOMA downlink, where the base station (BS)
was equipped with multiple antennas and the MUs have a single antenna each.
A more general setup was considered in [29], where both the BS and the MUs
are multiple-antenna devices. By exploiting multiple antennas at the BS and the
MUs, a signal alignment scheme was proposed to mitigate both the intra-cluster
and inter-cluster interference. It is worth pointing out that the implementation of the
two above schemes requires full channel state information (CSI) at the BS, which is
usually difficult and costly in practice. To circumvent the difficulty in CSI acquisition,
random beamforming was adopted in [30], which inevitably leads to performance
loss. Alternatively, the work in [31] suggested to employ zero-forcing (ZF) detection
at the multiple-antenna MUs for inter-cluster interference cancelation. However, the
ZF scheme requires that the number of antennas at each MU is greater than the
number of antennas at the BS, which is in general impractical.
To effectively realize the potential benefits of multiple-antenna techniques, the
amount and quality of CSI available at the BS play a key role [32, 33]. In practice,
the CSI can be obtained in several different ways. For instance, in time duplex
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
231
division (TDD) systems, the BS can obtain the downlink CSI through estimating the
CSI of uplink by leveraging the channel reciprocity [34]. While in frequency duplex
division (FDD) systems, the downlink CSI is usually first estimated and quantized
at the MUs, and then is conveyed back to the BS via a feedback link [35]. For
both practical TDD and FDD systems, the BS has access to only partial CSI. As a
result, there will be residual inter-cluster and intra-cluster interference. To the best of
the authors’ knowledge, previous works only consider two extreme cases with full
CSI or no CSI, the design, analysis and optimization of multiple-antenna NOMA
systems with partial CSI remains an uncharted area. Motivated by this, we present
a comprehensive study on the impact of partial CSI on the design, analysis, and
optimization of multiple-antenna NOMA downlink communication systems.
The rest of this chapter is organized as follows: Sect. 7.2 gives a brief introduction
of the considered NOMA downlink communication system and designs the corresponding multiple-antenna transmission framework. Section 7.3 first analyzes the
average transmission rates in presence of imperfect CSI and then proposes three performance optimization schemes. Section 7.4 derives the average transmission rates
in two extreme cases through asymptotic analysis and presents some system design
guidelines. Section 7.5 provides simulation results to validate the effectiveness of
the proposed schemes. Finally, Sect. 7.6 concludes this chapter.
7.2 System Model and Framework Design
Consider a downlink communication scenario in a single-cell system, where a base
station (BS) broadcasts messages to multiple MUs, cf. Fig. 7.1. Note that the BS is
equipped with M antennas, while the MUs have a single antenna each due to the size
limitation.
7.2.1 User Clustering
To strike a balance between the system performance and computational complexity in NOMA systems, it is necessary to carry out user clustering. In particular,
user clustering can be designed from different perspectives. For instance, a signalto-interference-plus-noise ratio (SINR) maximization user clustering scheme was
adopted in [36] and quasi-orthogonal MUs were selected to form a cluster in [37].
Intuitively, these schemes perform user clustering by the exhaustive search method,
resulting in high implementation complexity. In this chapter, we design a simple
user clustering scheme based on the information of spatial direction.1 Specifically,
the MUs in the same direction but with distinctive propagation distances are arranged
1 The
spatial direction of users can be found via various methods/technologies such as GPS or user
location tracking algorithms.
232
X. Chen et al.
Fig. 7.1 A multiuser NOMA communication system with 4 clusters
into a cluster. On one hand, the same direction of the MUs in a cluster allows the
use of a single beam to nearly align all MUs in such a cluster, thereby facilitating
the mitigation of the inter-cluster interference and the enhancement of the effective
channel gain. On the other hand, a large gap of propagation distances avoids severe
inter-user interference and enables a more accurate SIC at the MUs [38–40]. If two
MUs are close to each other with almost equal channel gains, it is possible to assign
them in different clusters by improving the spatial resolution via increasing the number of spatial beams and the number of BS antennas. Without loss of generality, we
assume that the MUs are grouped into N clusters with K MUs in each cluster. To
1/2
facilitate the following presentation, we use αn,k hn,k to denote the M-dimensional
channel vector from the BS to the kth MU in the nth cluster, where αn,k is the largescale channel fading, and hn,k is the small-scale channel fading following zero mean
complex Gaussian distribution with unit variance. It is assumed that αn,k remains
constant for a relatively long period, while hn,k keeps unchanged in a time slot but
varies independently over time slots.
7.2.2 CSI Acquisition
For the TDD mode, the BS obtains the downlink CSI through uplink channel estimation. Specifically, at the beginning of each time slot, the MUs simultaneously
send pilot sequences of τ symbols to the BS, and the received pilot at the BS can be
expressed as
YP =
K
N
P
αn,k hn,k Φ n,k + N P ,
τ Pn,k
n=1 k=1
(7.1)
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
233
P
where Pn,k
is the transmit power for the pilot sequence of the kth MU in the nth cluster,
N P is an additive white Gaussian noise (AWGN) matrix with i.i.d. zero mean and
unit variance complex Gaussian distributed entries. Φ n,k ∈ C 1×τ is the pilot sequence
sent from the kth MU in the nth cluster. It is required that τ > N K , such that the
H
= 1, ∀(n, k) = (i, j), can be
pairwise orthogonality that Φ n,k Φ i,Hj = 0 and Φ n,k Φ n,k
guaranteed. By making use of the pairwise orthogonality, the received pilot can be
transformed as
H
H
P
αn,k hn,k + N P Φ n,k
.
(7.2)
= τ Pn,k
Y P Φ n,k
Then, by using minimum mean squared error (MMSE) estimation, the relation
between the actual channel gain hn,k and the estimated channel gain ĥn,k can be
expressed as
hn,k =
√
ρn,k ĥn,k + 1 − ρn,k en,k ,
(7.3)
where en,k is the channel estimation error vector with i.i.d. zero mean and unit variance
complex Gaussian distributed entries, and is independent of ĥn,k . Variable ρn,k =
P
τ Pn,k
αn,k
P
1+τ Pn,k
αn,k
= 1 − 1+τ P1P α is the correlation coefficient between hn,k and ĥn,k . A
n,k n,k
large ρn,k means a high accuracy for channel estimation. Thus, it is possible to
P
or the length τ of
improve the CSI accuracy by increasing the transmit power Pn,k
pilot sequence.
For the FDD mode, the CSI is usually conveyed from the MUs to the BS through
a feedback link. Since the feedback link is rate-constrained, CSI at the MUs should
first be quantized. Specifically, the kth MU in the nth cluster chooses an optimal
(1)
(2 Bn,k )
, . . . , h̃n,k
}
codeword from a predetermined quantization codebook Bn,k = {h̃n,k
( j)
Bn,k
of size 2 , where h̃n,k is the jth codeword of a unit norm and Bn,k is the number
of feedback bits. Mathematically, the codeword selection criterion is given by
H ( j) 2
j ⋆ = arg max hn,k
h̃n,k .
B
1≤ j≤2
(7.4)
n,k
Then, the MU conveys the index j ⋆ to the BS with Bn,k feedback bits, and the BS
( j ⋆)
recoveries the quantized CSI h̃n,k from the same codebook. In other words, the BS
only gets the phase information by using the feedback scheme based on a quantization
codebook. However, as shown in below, the phase information is sufficient for the
design of spatial beamforming. Similarly, the relation between the real CSI and the
obtained CSI in FDD mode can be approximated as [41, 42]
h̃n,k =
√
⋆
ρn,k h̃n,k
+ 1 − ρn,k ẽn,k ,
(7.5)
n,k
⋆
is the phase of the channel hn,k , h̃n,k
is the quantized phase
where h̃n,k = hhn,k
information, ẽn,k is the quantization error vector with uniform distribution, and
234
X. Chen et al.
Bn,k
ρn,k = 1 − 2− M−1 is the associated correlation coefficient or CSI accuracy. Thus,
it is possible to improve the CSI accuracy by increasing the size of quantization
codebook for a given number of antennas M at the BS.
7.2.3 Superposition Coding and Transmit Beamforming
Based on the available CSI, the BS constructs one transmit beam for each cluster, so as
to mitigate or even completely cancel the inter-cluster interference. To strike balance
between system performance and implementation complexity, we adopt zero-force
beamforming (ZFBF) at the BS. We take the design of beam wi for the ith cluster as
an example. First, we construct a complementary matrix H̄i 2 as:
H̄i = [ĥ1,1 , . . . , ĥ1,K , . . . , ĥi−1,K , ĥi+1,1 , . . . , ĥ N ,K ] H .
(7.6)
Then, we perform singular value decomposition (SVD) on H̄i and obtain its right
singular vectors ui, j , j = 1, . . . , Nu , with respect to the zero singular values, where
we can design the beam as wi =
Nu is the number of zero singular values. Finally,
Nu
Nu
θ
u
,
where
θ
>
0
is
a
weight
such
that
i, j
j=1 i, j i, j
j=1 θi, j = 1. Thus, the received
signal at the kth MU in the nth cluster is given by
yn,k =
=
√
H
αn,k hn,k
N
i=1
√
wi si + n n,k
H
αn,k hn,k
wn sn +
H
αn,k (1 − ρn,k )en,k
N
i=1,i=n
wi si + n n,k ,
(7.7)
where si = Kj=1 Pi,Sj si, j is the superposition coded signal with Pi,Sj and si, j being
transmit power and transmit signal for the jth MU in the ith cluster, and n n,k is the
AWGN with unit variance. In general, Pi,Sj should be carefully allocated to distinguish
the MUs in the power domain, which we will discuss in detail below. Note that
√
H
H
H
wi = ρn,k ĥn,k
wi + 1 − ρn,k en,k
wi =
Eq. (7.7) holds true due to the fact that hn,k
H
3
1 − ρn,k en,k wi for ZFBF in TDD mode. With perfect CSI at the BS, i.e., ρn,k = 1,
the inter-cluster interference can be completely canceled.
2 In
⋆ , . . . , h̃⋆ , . . . ,
FDD mode, the complementary matrix is given by H̄i = [h̃1,1
1,K
⋆
⋆
⋆
⋆
H
h̃i−1,1 , . . . , h̃i−1,K , h̃i+1,1 , . . . , h̃ N ,K ] .
3 In FDD mode, we have h H w = √ρ h (h̃⋆ ) H w + 1 − ρ h ẽ H w =
i
n,k
n,k
n,k
n,k n,k i
n,k
n,k i
d
d
H
H
1 − ρn,k hn,k ẽn,k wi = 1 − ρn,k en,k wi , where = denotes the equality in distribution. If
ρn,k = ρn,k , Eq. (7.7) also holds true in FDD mode. In the sequel, without loss of generality, we
no longer distinguish between TDD and FDD.
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
235
7.2.4 Successive Interference Cancellation
Although ZFBF at the BS can mitigate partial inter-cluster interference from the
other clusters, there still exists intra-cluster interference from the same cluster. In
order to improve the received signal quality, the MU conducts SIC according to the
principle of NOMA. Without loss of generality, we assume that the effective channel
gains in the ith cluster have the following order:
√
√
H
H
wi |2 ≥ · · · ≥ | αi,K hi,K
wi |2 .
| αi,1 hi,1
(7.8)
It is reasonably assumed that the BS may know MUs’ effective gains through the
channel quality indicator (CQI) messages, and then determines the user order in (7.8).
Thus, in the ith cluster, the jth MU can always successively decode the lth MU’s
signal, ∀l > j, if the lth MU can decode its own signal. As a result, the jth MU can
subtract the interference from the lth MU in the received signal before decoding its
own signal. After SIC, the signal-to-interference-plus-noise ratio (SINR) at the kth
MU in the nth cluster can be expressed as
γn,k =
H
αn,k |hn,k
wn |2
k−1
j=1
H w |2 P S
αn,k |hn,k
n
n,k
,
K
N
S
Pn, j + αn,k (1 − ρn,k )
|e H wi |2
P S + 1
i=1,i=n n,k
l=1 i,l
AWGN
Intra-cluster interference
Inter-cluster interference
(7.9)
where the first term in the denominator of (7.9) is the residual intra-cluster interference after SIC at the MU, the second one is the residual inter-cluster interference
after ZFBF at the BS, and the third one is the AWGN. For the 1st MU in each
cluster, there is no intra-cluster interference, since it can completely eliminate the
intra-cluster interference. Note that in this chapter, we assume that perfect SIC can
be performed at the MUs. In practical NOMA systems, SIC might be imperfect due
to a limited computational capability at the MUs. Thus, there exists residual intracluster interference from the weaker MUs even after SIC [43]. However, the study of
the impact of imperfect SIC on the system performance is beyond the scope of this
chapter and we would like to investigate it in the future work. Moreover, the transmit
power has a significant impact on the SIC and the performance of NOMA [44]. Thus,
we will quantitatively analyze the impact of transmit power and then aim to optimize
the transmit power for improving the performance in the following sections.
7.3 Performance Analysis and Optimization
In this section, we concentrate on performance analysis and optimization of multiantenna NOMA downlink with imperfect CSI. Specifically, we first derive closed-
236
X. Chen et al.
form expressions for the average transmission rates of the 1st MU and the other
MUs, and then propose separate and joint optimization schemes of transmit power,
feedback bits, and transmit mode, so as to maximize the average sum rate of the
system.
7.3.1 Average Transmission Rate
We start by analyzing the average transmission rate of the kth MU in the nth cluster.
First, we consider the case k > 1. According to the definition, the corresponding
average transmission rate can be computed as
Rn,k = E log2 1 + γn,k
k
N
K
S
S
H w |2
H
2
αn,k |hn,k
n
j=1 Pn, j + αn,k (1 − ρn,k )
i=1,i =n |en,k wi |
l=1 Pi,l + 1
= E log2
K
N
k−1 S
S
H
H w |2
2
αn,k |hn,k
n
l=1 Pi,l + 1
i=1,i =n |en,k wi |
j=1 Pn, j + αn,k (1 − ρn,k )
K
N
k
H
H
= E log2 αn,k |hn,k
Pi,lS + 1
|en,k
wi |2
Pn,S j + αn,k (1 − ρn,k )
wn |2
i=1,i =n
j=1
N
k−1
H
Pn,S j + αn,k (1 − ρn,k )
wn |2
−E log2 αn,k |hn,k
i=1,i =n
j=1
l=1
H
wi |2
|en,k
K
l=1
Pi,lS + 1
.
(7.10)
Note that the average transmission rate in (7.10) can be expressed as the difference of two terms, which have a similar form. Hence, we concentrate on the
derivation of the
we use
to denote the term
first term. For notational convenience,
N
W
K
S
H
H
2
wn |2 kj=1 Pn,S j + αn,k (1 − ρn,k ) i=1,i
αn,k |hn,k
=n |en,k wi |
l=1 Pi,l . To compute
the first expectation, the key is to obtain the probability density function (pdf) of W .
H
wn |2 in W , since wn of unit norm is designed
Checking the first random variable |hn,k
H
2
2
independent of hn,k , |hn,k wn | is χ distributed with 2 degrees of freedom [45].
H
Similarly, |en,k
wi |2 also has the distribution χ 2 (2). Therefore, W can be considered
as a weighted sum of N random variables with χ 2 (2) distribution. According to [46],
W is a nested finite weighted sum of N Erlang pdfs, whose pdf is given by
f W (x) =
where
q
ηn,k
=
N
i=1
q
i
N
g(x, ηn,k
),
Ξ N i, {ηn,k }q=1
⎧
k
⎪
⎪
Pq,S j
⎨ αn,k
if q = n
j=1
K
⎪
⎪
S
⎩ αn,k (1 − ρn,k )
Pq,l
if q = n
l=1
,
(7.11)
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
i
g(x, ηn,k
)
q
N
Ξ N i, {ηn,k }q=1
237
1
x
= i exp − i
,
ηn,k
ηn,k
−1
N
−1
i
(−1) N −1 ηn,k
1
1
− s+U(s−i)
,
= N l
i
ηn,k
ηn,k
l=1 ηn,k
s=1
and U(x) is the well-known unit step function defined as U(x ≥ 0) = 1 and zero othq
N
erwise. It is worth pointing out that the weights Ξ N are constant for given {ηn,k }q=1
.
Hence, the first expectation in (7.10) can be computed as
E[log2 (1 + W )] =
=
∞
0
N
i=1
=−
log2 (1 + x) f W (x)d x
1
x
log2 (1 + x) i exp − i
dx
ηn,k
ηn,k
0
1
1
q
N
exp
i, {ηn,k }q=1
Ei − i
, (7.12)
i
ηn,k
ηn,k
q
N
Ξ N i, {ηn,k }q=1
N
1
ΞN
ln(2) i=1
∞
x
where Ei (x) = −∞ exp(t)
dt is the exponential integral function. Equation (7.12)
t
H
wn |2 k−1
follows from [47, Eq. (4.3372)]. Similarly, we use V to denote αn,k |hn,k
j=1
K
N
S
H
2
Thus,
w
|
P
in
the
second
term
of
(7.10).
|e
Pn,S j + αn,k (1 − ρn,k ) i=1,i
i
t=1 i,t
=n n,k
the second expectation term can be computed as
N
1
1
1
v
N
E[log2 (1 + V )] = −
Ξ N i, {βn,k }v=1 exp
Ei − i
, (7.13)
i
ln(2)
βn,k
βn,k
i=1
where
v
βn,k
=
⎧
k−1
S
⎪
⎪
Pv, j
⎨ αn,k
if v = n
j=1
K
⎪
⎪
S
⎩ αn,k (1 − ρn,k )
Pv,l
if v = n
.
l=1
Hence, we can obtain the average transmission rate for the kth MU in the nth cluster
as follows
N
1
1
1
v
N
Rn,k =
Ξ N i, {βn,k }v=1 exp
Ei − i
i
ln(2) i=1
βn,k
βn,k
N
1
1
1
q
N
exp
Ξ N i, {ηn,k }q=1
Ei − i
.
(7.14)
−
i
ln(2) i=1
ηn,k
ηn,k
238
X. Chen et al.
Then, we consider the case k = 1. Since the first MU can decode all the other
MUs’ signals in the same cluster, there is no intra-cluster interference. In this case,
the corresponding average transmission rate reduces to
Rn,1
N −1
1
1
1
N
−1
v
=
Ξ N −1 i, {βn,1
}v=1 exp
Ei − i
i
ln(2) i=1
βn,1
βn,1
N
1
1
1
q
N
Ξ N i, {ηn,1 }q=1 exp
−
Ei − i
,
i
ln(2) i=1
ηn,1
ηn,1
where
q
ηn,1 =
and
v
βn,1
=
⎧
S
⎨ αn,1 Pq,1
⎩ αn,1 (1 − ρn,1 )
K
l=1
if q = n
S
Pq,l
if q = n
⎧
K
⎪
S
⎪
Pv,l
⎨ αn,1 (1 − ρn,1 )
⎪
⎪
⎩ αn,1 (1 − ρn,1 )
l=1
K
l=1
S
Pv+1,l
(7.15)
,
if v < n
.
if v ≥ n
Combing (7.14) and (7.15), it is easy to evaluate the performance of a multipleantenna NOMA downlink with arbitrary system parameters and channel conditions.
In particular, it is possible to reveal the impact of system parameters, i.e., transmit
power, CSI accuracy, and transmission mode.
7.3.2 Power Allocation
From (7.14) and (7.15), it is easy to observe that with imperfect CSI, transmit power
has a great impact on average transmission rates. On one hand, increasing the transmit
power can enhance the desired signal strength. On the other hand, it also increases the
interference. Thus, it is desired to distribute the transmit power according to channel
conditions.
To maximize the sum rate of the considered multiple-antenna NOMA system
subject to a total power constraint, we have the following optimization problem:
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
J1 : max
S
Pn,k
s.t. C1 :
C2
K
N
239
Rn,k
n=1 k=1
N
K
S
S
Pn,k
≤ Ptot
n=1 k=1
S
:Pn,k
> 0,
(7.16)
S
where Ptot
is the maximum total transmit power budget. It is worth pointing out that
in certain scenarios, user fairness might be of particular importance. To guarantee
user fairness, one can replace the objective function of J1 with the maximization of
a weighted sum rate, where the weights can directly affect the power allocation and
thus the MUs’ rates. Unfortunately, J1 is not a convex problem due to the complicated
expression for the objective function. Thus, it is difficult to directly provide a closedform solution for the optimal transmit power. As a compromise solution, we propose
an effective power allocation scheme based on the following important observation
of the multiple-antenna NOMA downlink system:
Lemma 1 The inter-cluster interference is dependent of power allocation between
the clusters, while the intra-cluster interference is determined by power allocation
among the MUs in the same cluster.
N
Proof A close observation of the inter-cluster interference αn,k (1 − ρn,k ) i=1,i
=n
K
K
H
Pi,lS in (7.9) indicates that l=1
wi |2 l=1
|en,k
Pi,lS is the total transmit power for
the ith cluster, which suggests that inter-cluster power allocation does not affect the
inter-cluster interference.
Inspired by Lemma 1, the power allocation scheme can be divided into two steps.
In the first step, the BS distributes the total power among the N clusters. In the
second step, each cluster individually carries out power allocation subject to the
power constraint determined by the first step. In the following, we give the details of
the two-step power allocation scheme. First, we design the power allocation between
the clusters from the perspective of minimizing inter-cluster interference. For the ith
cluster, the average aggregate interference to the other clusters is given by
⎡
Ii = E ⎣
⎛
=⎝
N
K
n=1,n=i k=1
N
K
n=1,n=i k=1
H
αn,k (1 − ρn,k )|en,k
wi |2
⎞
αn,k (1 − ρn,k )⎠ PiS ,
K
l=1
⎤
Pi,lS ⎦
(7.17)
K
where PiS = l=1
Pi,lS is the total transmit power of the ith cluster. Equation (7.17)
H
follows the fact that E[|en,k
wi |2 ] = 1. Intuitively, a large interference coefficient
K
N
k=1 αn,k (1 − ρn,k ) means a more severe inter-cluster interference caused
n=1,n=i
240
X. Chen et al.
by the ith cluster. In order to mitigate the inter-cluster interference for improving the
average sum rate, we propose to distribute the power proportionally to the reciprocal
of interference coefficient. Specifically, the transmit power for the ith cluster can be
computed as
PiS =
N
n=1,n=i
N N
K
k=1
n=1,n=l
l=1
αn,k (1 − ρn,k )
−1
K
k=1 αn,k (1 − ρn,k )
−1
S
Ptol
.
(7.18)
Then, we allocate the power in the cluster for further increasing the average
sum rate. According to the nature of NOMA techniques, the first MU not only has
the strongest effective channel gain for the desired signal, but also generates a weak
interference to the other MUs. On the contrary, the K th MU has the weakest effective
channel gain for the desired signal and also produces a strong interference to the other
MUs. Thus, from the perspective of maximizing the sum of average rate, it is better
to allocate the power based on the following criterion:
S
S
S
≥ · · · ≥ Pn,k
≥ · · · ≥ Pn,K
.
Pn,1
(7.19)
On the other hand, in order to facilitate SIC, the NOMA in general requires the
transmit powers in a cluster to follow a criterion below [31]:
S
S
S
≤ · · · ≤ Pn,k
≤ · · · ≤ Pn,K
.
Pn,1
(7.20)
Under this condition, the MU performs SIC according to the descending order of the
user index, namely the ascending order of the effective channel gain. Specifically,
the kth MU cancels the interference from the K th to the (k + 1)th MU in sequence.
Thus, the SINR for decoding each interference signal is the highest, which facilitates
SIC at MUs [44].
To simultaneously fulfill the above two criterions, we propose to equally distribute
the powers within a cluster, namely
S
= PnS /K .
Pn,k
(7.21)
Substituting (7.18) into (7.21), the transmit power for the kth MU in the nth cluster
can be computed as
S
Pn,k
=
N
i=1,i=n
K
N N
l=1
K
j=1
i=1,i=l
αi, j (1 − ρi, j )
K
−1
j=1 αi, j (1 − ρi, j )
−1
S
Ptol
.
(7.22)
Thus, we can distribute the transmit power based on (7.22) for given channel statistical
information and the CSI accuracy, which has a quite low computational complexity.
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
241
Remark 1 We note that path loss coefficient αn,k , ∀n, k, remain constant for a relatively long time, and it is easy to obtain at the BS via long-term measurement.
Hence, the proposed power allocation scheme incurs a low system overhead and can
be implemented with low complexity.
7.3.3 Feedback Distribution
For the FDD mode, the accuracy of quantized CSI relies on the size of codebook
2 Bn,k , where Bn,k is the number of feedback bits from the kth MU in the nth cluster. As
observed in (7.14) and (7.15), it is possible to decrease the interference by increasing
feedback bits. However, due to the rate constraint on the feedback link, the total
number of feedback bits is limited. Therefore, it is of great importance to optimize
the feedback bits among the MUs for performance enhancement.
According to the received signal-to-noise ratio (SNR) in (7.9), the CSI accuracy only affects the inter-cluster interference. Thus, it makes sense to optimize the
feedback bits to minimizing the average sum of inter-cluster interference given by
⎡
Iinter = E ⎣
=
N
K
n=1 k=1
K
N
αn,k (1 − ρn,k )
αn,k
N
N
i=1,i=n
H
|en,k
wi |2
Bn,k
PiS 2− M−1 .
K
l=1
⎤
Pi,lS ⎦
(7.23)
i=1,i=n
n=1 k=1
Hence, the optimization problem for feedback bits distribution can be expressed as
J2 : min
Bn,k
s.t. C3 :
N
K
n=1 k=1
N
K
n=1 k=1
αn,k
N
Bn,k
PiS 2− M−1
i=1,i=n
Bn,k ≤ Btot ,
C4 :Bn,k ≥ 0,
(7.24)
where Btot is an upper bound on the total number of feedback bits. J2 is an integer
programming problem, hence is difficult to solve. To tackle this challenge, we relax
the integer constraint on Bn,k . In this case, according to the fact that
242
X. Chen et al.
N
K
αn,k
N
i=1,i=n
n=1 k=1
⎛
B
n,k
PiS 2− M−1
≥ NK ⎝
N
K
= NK 2
αn,k
= NK 2
N K
k=1 Bn,k
− n=1 M−1
Btot
− M−1
B
i=1,i=n
n=1 k=1
N
⎛
1
NK
⎝
⎞
1
NK
n,k
PiS 2− M−1 ⎠
1
NK
N
K
⎛
⎝
N
K
αn,k
i=1,i=n
n=1 k=1
αn,k
n=1 k=1
N
N
i=1,i=n
⎞
1
NK
PiS ⎠
⎞
1
NK
PiS ⎠
,
(7.25)
Bn,k
N
S − M−1
, ∀n, k are equal. In
where the equality holds true only when αn,k i=1,i
=n Pi 2
other words, the objective function in (7.24) can be minimized while satisfying the
following condition:
αn,k
N
PiS 2−
Bn,k
M−1
i=1,i=n
Btot
= 2− M−1
1
NK
⎛
⎝
K
N
n=1 k=1
αn,k
N
i=1,i=n
⎞ N1K
PiS ⎠
.
(7.26)
Hence, based on the relaxed optimization problem, the optimal number of feedback
bits for the kth MU in the nth cluster is given by
Bn,k
⎛
⎞
⎛
⎞
N
N
N K
Btot
1
S
S
=
Pl ⎠ .
Pl ⎠ + log2 ⎝αn,k
log2 ⎝αi, j
−
NK
NK
i=1 j=1
l=1,l=i
(7.27)
l=1,l=n
Given channel statistical information and transmit power allocation, it is easy to
determine the feedback distribution according to (7.27). Note that there exists an
integer constraint on the number of feedback bits in practice, so we should utilize
the maximum integer that is not larger than Bn,k in (7.27), i.e., ⌊Bn,k ⌋, ∀n, k.
Remark 2 The number of feedback bits distributed to the kth MU in the nth cluster is
determined by the average inter-cluster interference generated by the kth MU in the
nth cluster with respect to the average inter-cluster interference of each MU. In other
words, if one MU generates more inter-cluster interference, it would be allocated
with more feedback bits, so as to facilitate a more accurate ZFBF to minimize the
total interference.
7.3.4 Mode Selection
As discussed above, the performance of the multiple-antenna NOMA system is limited by both inter-cluster and intra-cluster interference. Although ZFBF at the BS and
SIC at the MUs are jointly applied, there still exists residual interference. Intuitively,
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
243
the strength of the residual interference mainly relies on the number of clusters N
and the number of MUs in each cluster K . For instance, increasing the number of
MUs in each cluster might reduce the inter-cluster interference, but also results in
an increase in intra-cluster interference. Thus, it is desired to dynamically adjust the
transmission mode, including the number of clusters and the number of MUs in each
cluster, according to channel conditions and system parameters. For dynamic mode
selection, we have the following lemma:
Lemma 2 If the BS has no CSI about the downlink, it is optimal to set N = 1. On the
other hand, if the BS has perfect CSI about the downlink, K = 1 is the best choice.
Proof First, if there is no CSI, namely ρn,k = 0, ∀n, k, ZFBF cannot be utilized to
mitigate the inter-cluster interference. If all the MUs belong to one cluster, interference can be mitigated as much as possible by SIC. In the case of perfect CSI at the
BS, ZFBF can completely the interference. Thus, it is optimal to arrange one MU in
one cluster.
In above, we consider two extreme scenarios of no and perfect CSI at the BS,
respectively. In practice, the BS has partial CSI through channel estimation or quantization feedback. Thus, we propose to dynamically choose the transmission mode
for maximizing the sum of average transmission rate, which is equivalent to an optimization problem below:
J3 : max
N ,K
N
K
Rn,k
n=1 k=1
s.t. C5 : N K = Nu ,
C6 : N > 0,
C7 : K > 0,
(7.28)
where Nu is the number of MUs in the multiple-antenna NOMA system. J3 is also
an integer programming problem, so it is difficult to obtain the closed-form solution.
Under this condition, it is feasible to get the optimal solution by numerical search
and the search complexity is O(N K ). In order to control the complexity of SIC, the
number of MUs in one cluster is usually small, e.g., K = 2. Therefore, the complexity
of numerical search is acceptable.
7.3.5 Joint Optimization Scheme
In fact, transmit power, feedback bits and transmission mode are coupled, and determine the performance together. Therefore, it is better to jointly optimize these variables, so as to further improve the performance of the multiple-antenna NOMA
systems. For example, given a transmission mode, it is easy to first allocate transmit
244
X. Chen et al.
power according to (7.22), and then distribute feedback bits according to (7.27).
Finally, we can select an optimal transmission mode with the largest sum rate. The
complexity of the joint optimization is mainly determined by the mode selection. As
mentioned above, if the number of MUs in one cluster is small, the complex of mode
selection is acceptable.
7.4 Asymptotic Analysis
In order to provide insightful guidelines for system design, we now pursue an asymptotic analysis on the average sum rate of the system. In particular, two extreme cases
are studied, namely interference limited and noise limited.
7.4.1 Interference Limited Case
S
S
With loss of generality, we let Pn,k
= θn,k Ptot
, ∀n, k, where 0 < θn,k < 1 is a power
N
v=1,v =n
allocation factor. For instance, θn,k is equal to
K
N N
l=1
K
j=1
v=1,v =l
αv, j (1−ρv, j )
K
j=1
−1
αv, j (1−ρv, j )
−1
in
S
the proposed power allocation scheme in Sect. 7.3.2. If the total power Ptot
is large
enough, the noise term of SINR in (7.9) is negligible. In this case, with the help of
[47, Eq. (4.3311)], the average transmission rate of the kth MU (k > 1) in the nth
cluster reduces to
N
Rn,k =
1
q
N
i
ln(ηn,k
)
Ξ N i, {ηn,k }q=1
ln(2) i=1
N
−
1
i
v
N
ln(βn,k
),
Ξ N i, {βn,k
}v=1
ln(2) i=1
(7.29)
where we have also used the fact that
N
i=1
q
N
=
Ξ N i, {ηn,k }q=1
N
i=1
v
N
= 1.
Ξ N i, {βn,k
}v=1
(7.30)
Similarly, the asymptotic average transmission rate of the 1st MU in the nth MU
can be obtained as
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
245
N
Rn,1 =
1
q
i
N
ln ηn,1
Ξ N i, {ηn,1 }q=1
ln(2) i=1
−
N −1
1
N −1
v
i
ln βn,1
.
Ξ N −1 i, {βn,1
}v=1
ln(2) i=1
(7.31)
Combining (7.29) and (7.31), we have the following important result:
Theorem 1 In the region of high transmit power, the average transmission rate is
S
S
independent of Ptot
, and there exists a performance ceiling regardless of Ptot
, i.e.,
S
once Ptot is larger than a saturation point, the average transmission rate will not
increase further even the transmit power increases.
i
i
i
i
S
Proof According to the definitions, ηn,k
and βn,k
can be rewritten as ηn,k
= ωn,k
Ptot
i
i
S
and βn,k = ψn,k Ptot , where
i
ωn,k
=
and
i
ψn,k
=
⎧
k
⎪
⎪
θi, j
⎨ αn,k
if i = n
j=1
K
⎪
⎪
⎩ αn,k (1 − ρn,k ) θi,l if i = n
⎧
k−1
⎪
⎪
θi, j
⎨ αn,k
,
l=1
if i = n
j=1
K
⎪
⎪
⎩ αn,k (1 − ρn,k ) θi,l if i = n
,
l=1
q
N
N
v
S
respectively. Thus, Ξ N i, {ηn,k }q=1
and Ξ N i, {βn,k
}v=1
are independent of Ptot
.
Hence, Rn,k in (7.29) can be transformed as
N
Rn,k
1
q
S
i
N
(ln(Ptot
) + ln(ωn,k
))
Ξ N i, {ηn,k }q=1
=
ln(2) i=1
N
1
N
i
v
S
))
}v=1
) + ln(ψn,k
Ξ N i, {βn,k
(ln(Ptot
−
ln(2) i=1
N
=
N
1
1
q
i
i
N
v
N
ln(ωn,k
)−
ln(ψn,k
),
Ξ N i, {ηn,k }q=1
Ξ N i, {βn,k
}v=1
ln(2) i=1
ln(2) i=1
(7.32)
N
q
N
=
i, {ηn,k }q=1
Ξ
where Eq. (7.32) follows the fact that
N
i=1
N
N
v
i=1 Ξ N i, {βn,k }v=1 = 1. Similarly, we can rewrite Rn,1 in (7.31) as
246
X. Chen et al.
N
Rn,1 =
1
q
i
N
ln ωn,1
Ξ N i, {ηn,1 }q=1
ln(2) i=1
−
N −1
1
N −1
v
i
ln ψn,1
,
Ξ N −1 i, {βn,1
}v=1
ln(2) i=1
where
i
=
ωn,1
and
i
=
ψn,1
⎧
S
⎨ αn,1 θi,1
⎩ αn,1 (1 − ρn,1 )
K
l=1
if i = n
θi,lS if i = n
⎧
K
⎪
⎪
⎨ αn,1 (1 − ρn,1 ) θi,lS
⎪
⎪
⎩ αn,1 (1 − ρn,1 )
l=1
K
l=1
(7.33)
,
if i < n
.
S
θi+1,l
if i ≥ n
S
, which proves Theorem 1.
Note that both (7.32) and (7.33) are regardless of Ptot
Now, we investigate the relation between the performance ceiling in Theorem 1
and the CSI accuracy ρn,k . First, we consider Rn,k with k > 1. As ρn,k asymptotically
approaches 1, the inter-cluster interference is negligible. Then, Rn,k can be further
reduced as
⎡
⎛
ideal = E ⎣log ⎝α |h H w |2
Rn,k
n,k n,k n
2
⎛
⎞
k
j=1 ωn, j ⎠
= log2 ⎝ k−1
.
j=1 ψn, j
k
j=1
⎡
⎞⎤
⎛
S ⎠⎦ − E ⎣log ⎝α |h H w |2
Pn,
n,k n,k n
2
j
k−1
j=1
⎞⎤
S ⎠⎦
Pn,
j
(7.34)
It is found that even with perfect CSI, the average
transmission
rate for the (k > 1)th
MU is still upper bounded. The bound log2
k
j=1
k−1
j=1
ωn, j
ψn, j
is completely determined by
channel conditions, and thus cannot be increased via power allocation. Differently,
for the 1st MU, if the CSI at the BS is sufficiently accurate, the SINR γn,1 becomes
high. As a result, the constant term 1 in the rate expression is negligible, and thus
the average transmission rate can be approximated as
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
Rn,1
247
S
H
αn,1 |hn,1
wn |2 Pn,1
≈ E log2
N
K
H
αn,1 (1 − ρn,1 ) i=1,i=n |en,1
wi |2 l=1
Pi,lS
H
S
= E log2 αn,1 |hn,1
wn |2 Pn,1
Ideal average rate
⎡
⎛
− E ⎣log2 ⎝αn,1 (1 − ρn,1 )
N
i=1,i=n
H
|en,1
wi |2
K
l=1
Rate loss due to imperfect CSI
⎞⎤
Pi,lS ⎠⎦ .
(7.35)
In (7.35), the first term is the ideal average transmission rate with perfect CSI, and
the second one is rate loss caused by imperfect CSI. We first check the term of the
ideal average transmission rate, which is given by
ideal
S
H
Rn,1
= E log2 αn,1 Ptot
θn,1 |hn,1
wn |2
C
S
.
= log2 αn,1 Ptot
θn,1 −
ln(2)
(7.36)
Note that if there is perfect CSI at the BS, the average transmission rate of the
S
1st MU increases proportionally to log2 (Ptot
) without a bound. However, as seen
in (7.34), the (k > 1)th MU has an upper bounded rate under the same condition,
which reconfirms the claim in Lemma 2 that it is optimal to arrange one MU in each
cluster in presence of perfect CSI. Then, we investigate the rate loss due to imperfect
CSI, which can be expressed as
N
K
H
S
loss
|en,1
wi |2
θi,t
Rn,1
= E log2 αn,1 (1 − ρn,1 )Ptot
i=1,i =n
S
−
= log2 αn,1 (1 − ρn,1 )Ptot
where
μvn,1 =
1
ln(2)
t=1
N
−1
i=1
⎧ K
⎪
⎪
⎨ θv,l
l=1
K
⎪
⎪
⎩
l=1
N −1
C − ln μin,1
Ξ N −1 i, {μvn,1 }v=1
,
(7.37)
if v < n
.
θv+1,l if v ≥ n
loss
S
enlarges as the total transmit power Ptot
increases.
Given a ρn,1 , the rate loss Rn,1
ideal
In order to keep the same rate of increase to the ideal rate Rn,1 , the CSI accuracy
ρn,1 should satisfy the following theorem:
S
Theorem 2 Only when (1 − ρn,1 )Ptot
is equal to a constant ε, the average transmission rate of the 1st MU in the nth cluster with imperfect CSI remains a fixed gap
with respect to the ideal rate. Specifically, the transmit power for training sequence
248
X. Chen et al.
p
P S /ε−1
should satisfy Pn,1 = totαn,1 τ in TDD systems, while the number of feedback bits
S
/ε) in FDD systems.
should satisfy Bn,1 = (M − 1) log2 (Ptot
Proof The proof is intuitively. By substituting ρn,1 = 1 −
1
P
1+τ Pn,1
αn,1
Bn,1
M−1
into (1 −
S
S
ρn,1 )Ptot
into (1 − ρn,1 )Ptot
= ε for
= ε for TDD systems and ρn,1 = 1 − 2−
S
Ptot /ε−1
p
S
FDD systems, we can get Pn,1 = αn,1 τ and Bn,1 = (M − 1) log2 (Ptot /ε), which
proves Theorem 2.
p
Remark 3 For the CSI accuracy at the BS, Pn,1 τ (namely transmit energy for training
Bn,1
sequence) in TDD systems and M−1
(namely spatial resolution) in FDD systems are
two crucial factors. Specifically, given a requirement on CSI accuracy, it is possible
to shorten the length of training sequence by increasing the transmit power, so as to
leave more time for data transmission in a time slot. However, in order to keep the
pairwise orthogonality of training sequences, the length of training sequence τ must
be larger than the number of MUs. In other words, the minimum value of τ is N K .
Similarly, in FDD systems, it is possible to reduce the feedback bits by increasing
the number of antennas M. Yet, in order to fulfill the spatial degrees of freedom
for ZFBF at the BS, M must be not smaller than (N − 1)K + 1. This is because
the beam wi for the ith cluster should be in the null space of the channels for the
(N − 1)K MUs in the other N − 1 clusters.
Furthermore, substituting (7.36) and (7.37) into (7.35), we have
Rn,1 ≈ − log2 (1 − ρn,1 ) + log2 (θn,1 ) −
N
−1
i=1
N −1
log2 μin,1 . (7.38)
Ξ N −1 i, {μvn,1 }v=1
Given a power allocation scheme, it is interesting that the bound of Rn,1 is independent
of channel conditions. As analyzed above, it is possible to improve the average rate
by improving the CSI accuracy. Especially, for FDD systems, we have the following
lemma:
Lemma 3 At the high power region with a large number of feedback bits, the average
rate of the 1st MU increases linearly as the numbers of feedback bits increase.
Bn,1
Proof Replacing ρn,1 in (7.38) with ρn,1 = 1 − 2− M−1 , Rn,1 is transformed as
Rn,1 ≈
N −1
Bn,1
N −1
+ log2 (θn,1 ) −
Ξ N −1 i, {μvn,1 }v=1
log2 μin,1 , (7.39)
M −1
i=1
which yields Lemma 3.
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
249
7.4.2 Noise-Limited Case
If the interference term is negligible with respect to the noise term due to a low
transmit power, then the SINR γn,k , ∀n, k is reduced as
H
S
wn |2 Pn,k
,
γn,k = αn,k |hn,k
(7.40)
H
wn |2 is
which is equivalent to the interference-free case. As discussed earlier, |hn,k
χ 2 (2) distributed, then the average transmission rate can be computed as
∞
S
log2 1 + Pn,k
αn,k x exp(−x)d x
1
1
= − exp
Ei − S
.
S
Pn,k
αn,k
Pn,k αn,k
Rn,k =
0
(7.41)
Note that Eq. (7.41) is independent of the CSI accuracy, thus it is unnecessary to
carry out channel estimation or CSI feedback in this scenario. Since both intra-cluster
interference and inter-cluster interference are negligible, ZFBF at the BS and SIC at
the MUs are not required, and all optimization schemes asymptotically approach the
same performance.
7.5 Simulation Results
To evaluate the performance of the proposed multiple-antenna NOMA technology,
we present several simulation results under different scenarios. For convenience, we
set M = 6, N = 3, K = 2, Btot = 12, while αn,k and ρn,k are given in Table 7.1 for
all simulation scenarios without extra specification. In addition, we use SNR (in dB)
S
.
to represent 10 log10 Ptot
First, we verify the accuracy of the derived theoretical expressions. As seen in Fig.
7.2, the theoretical expressions for both the 1st and the 2nd MUs in the 1st cluster well
coincide with the simulation results in the whole SNR region, which confirms the
high accuracy. As the principle of NOMA implies, the 1st MU performs better than
Table 7.1 Parameter Table for (αn,k , ρn,k ), ∀n ∈ [1, 3], and k ∈ [1, 2]
n
k
1
2
1
2
3
(1.00, 0.90)
(0.95, 0.85)
(0.90, 0.80)
(0.10, 0.70)
(0.20, 0.75)
(0.15, 0.80)
250
X. Chen et al.
Average Rate (b/s/Hz)
2
1.8
Theoretical (R1,1)
1.6
Simulation (R1,1)
Theoretical (R1,2)
1.4
Simulation (R1,2)
1.2
1
0.8
0.6
0.4
0.2
0
−10
−5
0
5
10
15
20
25
30
35
SNR (dB)
Fig. 7.2 Comparison of theoretical expressions and simulation results
the second MU. At high SNR, the average rates of the both MUs are asymptotically
saturated, which proves Theorem 1 again.
Secondly, we compare the proposed power allocation scheme with the equal power
allocation scheme and the fixed power allocation scheme proposed in [5]. Note that
the fixed power allocation scheme distributes the power with a fixed ratio 1:4 between
the two MUs in a cluster so as to facilitate the SIC. It is found in Fig. 7.3 that the
proposed power allocation scheme offers an obvious performance gain over the
two baseline schemes, especially in the medium SNR region. Note that practical
communication systems, in general, operate at medium SNR, thus the proposed
scheme is able to achieve a given performance requirement with a lower SNR. As
the SNR increases, the proposed scheme and the equal allocation scheme achieve
the same saturated sum rate, but the fixed allocation scheme has a clear performance
loss.
Next, we examine the advantage of feedback allocation for the FDD-based NOMA
system with equal power allocation, cf. Fig. 7.4. As analyzed in Sect. 7.4.2, at very low
SNR, namely the noise-limited case, the average rate is independent of CSI accuracy,
and thus the two schemes asymptotically approach the same sum rate. As SNR
increases, the proposed feedback allocation scheme achieves a larger performance
gain. Similarly, at high SNR, both the two schemes are saturated, and the proposed
scheme obtains the largest performance gain. For instance, at SNR = 30 dB, there
is a gain of more than 0.5 b/s/Hz. Furthermore, we investigate the impact of the total
number of feedback bits on the average rates of different MUs at SNR = 35 dB. As
shown in Fig. 7.5, the performance of the 1st MU is clearly better than that of the
2nd MU. Moreover, the average rate of the 1st MU is nearly a linear function of the
number of feedback bits, which reconfirms the claims of Lemma 3.
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
251
6
Sum of Average Rate (b/s/Hz)
5
4
3
2
Adaptive Power Allocation
Equal Power Allocation
Fixed Power Allocation
1
0
-10
-5
0
5
10
15
20
25
30
35
SNR (dB)
Fig. 7.3 Performance comparison of different power allocation schemes
Sum of Average Rate (b/s/Hz)
3
2.5
2
1.5
1
Adaptive Feedback Allocation
Equal Feedback Allocation
0.5
0
−10
−5
0
5
10
15
20
25
30
35
SNR (dB)
Fig. 7.4 Performance comparison of different feedback allocation schemes
Then, we investigate the impact of the transmission mode on the performance of
the NOMA systems at SNR = 10 dB with equal power allocation in Fig. 7.6. To
concentrate on the impact of transmission mode, we set the same CSI accuracy of
all downlink channels as ρ. Note that we consider four fixed transmission modes
under the same channel conditions in the case of six MUs in total. Consistent with
the claims in Lemma 2, mode 4 with N = 1 and K = 6 achieves the largest sum rate
at low CSI accuracy, while mode 1 with N = 6 and K = 1 performs best at high CSI
accuracy. In addition, it is found that at medium CSI accuracy, mode 2 with N = 3 and
K = 2 is optimal, since it is capable to achieve a best balance between intra-cluster
252
X. Chen et al.
1.4
Average Rate (b/s/Hz)
1.2
1
0.8
Rideal
1,2
0.6
R1,1
R
1,2
0.4
0.2
0
0
5
10
15
20
25
30
35
Total Number of Feedback (bits)
Fig. 7.5 Asymptotic performance with a large number of feedback bits
Sum of Average Rate (b/s/Hz)
7
Mode 1 (N=6, K=1)
Mode 2 (N=3, K=2)
Mode 3 (N=2, K=3)
Mode 4 (N=1, K=6)
Dynamic Mode
6
5
4
3
2
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ρ
Fig. 7.6 Performance comparison of different transmission modes
interference and inter-cluster interference. Thus, we propose to dynamically select
the transmission mode according to channel conditions and system parameters. As
shown by the red line in Fig. 7.6, dynamic mode selection can always obtain the
maximum sum rate.
Finally, we exhibit the superiority of the proposed joint optimization scheme for
the NOMA systems at SNR = 10 dB. In addition, we take a fixed scheme based on
NOMA and a time division multiple access (TDMA) based on OMA as baseline
schemes. Specifically, the joint optimization scheme first distributes the transmit
power with equal feedback allocation, then allocates the feedback bits based on the
distributed power, finally selects the optimal transmission mode. The fixed scheme
always adopts the mode 2 (N = 3, K = 2) with equal power and feedback allocation.
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
253
The TDMA equally allocates each time slot to the six MUs and utilizes maximum
ratio transmission (MRT) based on the available CSI at the BS to maximize the
rate. For clarity of notation, we use ρ to denote the CSI accuracy based on equal
feedback allocation. In other words, the total number of feedback bits is equal to
Btot = −K ∗ N ∗ (M − 1) ∗ log2 (1 − ρ). As seen in Fig. 7.7, the fixed scheme performs better than the TDMA scheme at low and high CSI accuracy, and slightly worse
at the medium regime. However, the proposed joint optimization scheme performs
much better than the two baseline schemes. Especially at high CSI accuracy, the
performance gap becomes substantially large. For instance, there is a performance
gain of about 3 b/s/Hz at ρ = 0.8, and up to more than 5 b/s/Hz at ρ = 0.9. As
analyzed in Lemma 2 and confirmed by Fig. 7.6, when ρ is larger than 0.8, which is
a common CSI accuracy in practical systems, mode 2 is optimal for maximizing the
system performance. Thus, the joint optimization scheme is reduced to joint power
and feedback allocation, which requires only a very low complexity. Thus, the proposed NOMA scheme with joint optimization can achieve a good performance with
low complexity, and it is a promising technique for future wireless communication
systems.
7.6 Conclusion
This chapter provided a comprehensive solution for designing, analyzing, and optimizing a NOMA technology over a general multiuser multiple-antenna downlink
in both TDD and FDD modes. First, we proposed a new framework for multipleantenna NOMA. Then, we analyzed the performance and derived exactly closed-form
expressions for average transmission rates. Afterward, we optimized the three key
9
Sum of Average Rate (b/s/Hz)
8
NOMA (Joint Optimization)
NOMA (Fixed Scheme)
OMA (TDMA)
7
6
5
4
3
2
1
0
0.1
0.2
0.3
0.4
ρ
0.5
0.6
0.7
0.8
0.9
Fig. 7.7 Performance comparison of a joint optimization scheme and a fixed allocation scheme
254
X. Chen et al.
parameters of multiple-antenna NOMA, i.e., transmit power, feedback bits, and transmission mode. Finally, we conducted asymptotic performance analysis and obtained
insights on system performance and design guidelines.
References
1. V.W.S. Wong, R. Schober, D.W.K. Ng, L.-C. Wang, Key Technologies for 5G Wireless Systems
(Cambridge University Press, Cambridge, U.K., 2017)
2. Y. Saito, Y. Kishiyama, A. Benjebour, T. Nakamura, A. Li, K. Higuchi, Non-orthogonal multiple access (NOMA) for cellular future radio access, in Proceedings of the IEEE Vehicular
Technology Conference (VTC-Spring), June 2013, pp. 1–5
3. Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, C-L. I, H.V. Poor, Application of non-orthogonal
multiple access in LTE and 5G networks. IEEE Commun. Mag. 55(2), 185–191 (2017)
4. L. Dai, B. Wang, Y. Yuan, S. Han, C-L. I, Z. Wang, Non-orthogonal multiple access for 5G:
solutions, challenges, opportunities, and future research trends. IEEE Commun. Mag. 53(9),
74–81 (2015)
5. Z. Ding, Z. Yang, P. Fan, H.V. Poor, On the performance of non-orthogonal multiple access
in 5G systems with randomly deployed users. IEEE Signal Process. Lett. 21(12), 1501–1505
(2014)
6. Y. Yuan, Z. Yuan, G. Yu, C.-H. Hwang, P.-K. Liao, A. Li, K. Takeda, Non-orthogonal transmission technology in LTE evolution. IEEE Commun. Mag. 54(7), 68–74 (2016)
7. W. Shin, M. Vaezi, B. Lee, D.J. Love, J. Lee, H.V. Poor, Non-orthogonal multiple access
in multi-cell networks: theory, performance, and practical challenges. IEEE Commun. Mag.
55(10), 176–183 (2017)
8. S. Timotheou, I. Krikidis, Fairness for non-orthogonal multiple access in 5G systems. IEEE
Signal Process. Lett. 22(10), 1647–1651 (2015)
9. S.-L. Shieh, Y.-C. Huang, A simple scheme for realizing the promised gains of downlink
nonorthogonal multiple access. IEEE Trans. Commun. 64(4), 1624–1635 (2016)
10. P. Xu, Y. Yuan, Z. Ding, X. Dai, R. Schober, On the outage performance of non-orthogonal
multiple access with 1-bit feedback. IEEE Trans. Wirel. Commun. 15(10), 6716–6730 (2016)
11. Z. Yang, Z. Ding, P. Fan, Z. Ma, Outage performance for dynamic power allocation in hybrid
non-orthogonal multiple access systems. IEEE Commun. Lett. 20(8), 1695–1698 (2016)
12. F. Fang, H. Zhang, J. Cheng, V.C.M. Leung, Energy-efficient resource allocation for downlink
orthogonal multiple access (NOMA) network. IEEE Trans. Commun. 64(9), 3722–3732 (2016)
13. H. Tabassum, M.S. Ali, E. Hossain, M.J. Hossain, D.I. Kim, Non-orthogonal multiple access
(NOMA) in cellular uplink and downlink: challenges and enabling techniques, August 2016
[Online], http://128.84.21.199/abs/1608.05783
14. J. Choi, On the power allocation for a practical multiuser superposition scheme in NOMA
systems. IEEE Commun. Lett. 20(3), 438–441 (2016)
15. C.-L. Wang, J.-Y. Chen, Y.-J. Chen, Power allocation for downlink non-orthogonal multiple
access system. IEEE Wirel. Commun. Lett. 5(5), 532–535 (2016)
16. Z. Yang, Z. Ding, P. Fan, N. Al-Dhahir, A general power allocation scheme to guarantee quality
of service in downlink and uplink NOMA systems. IEEE Trans. Wireless Commun. 15(11),
7244–7257 (2016)
17. Y. Liu, M. Elkashlan, Z. Ding, G.K. Karagiannidis, Fairness of user clustering in MIMO nonorthogonal multiple access systems. IEEE Commun. Lett. 20(7), 1465–1468 (2016)
18. H. Zhang, D.-K. Zhang, W.-X. Meng, C. Li, User pairing algorithm with SIC in non-orthogonal
multiple access system, in Proceedings of the IEEE International Conference on Communication (ICC), May 2016, pp. 1–6
7 On the Design of Multiple-Antenna Non-Orthogonal Multiple Access
255
19. J. Mei, L. Yao, H. Long, K. Zheng, Joint user pairing and power allocation for downlink nonorthogonal multiple access systems, in Proceedings of the IEEE International Conference on
Communication (ICC), May 2016, pp. 1–6
20. Z.Q. Al-Abbasi, D.K.C. So, User-pairing based non-orthogonal multiple access (NOMA) system, in Proceedings of the IEEE Vehicular Technology Conference (VTC-Spring), April 2016,
pp. 1–5
21. Md.S. Ali, H. Tabassum, E. Hossain, Dynamic user clustering and power allocation for uplink
and downlink non-orthogonal multiple access (NOMA) systems. IEEE Access 4, 6325–6343
(2016)
22. H. Weingarten, Y. Steinberg, S.S. Shamai, The capacity region of the Gaussian multiple-input
multiple-output broadcast channel. IEEE Trans. Inf. Theory 52(9), 3936–3964 (2006)
23. M.A. Maddah-Ali, M.A. Sadrabadi, A.K. Khandani, Broadcast in MIMO systems based on a
generalized QR decomposition: signaling and performance analysis. IEEE Trans. Inf. Theory
54(3), 1124–1138 (2008)
24. A.D. Dabbagh, D.J. Love, Precoding for multiple antenna Gaussian broadcast channels. IEEE
Trans. Signal Process. 55(7), 3837–3850 (2007)
25. X. Chen, C. Yuen, Performance analysis and optimization for interference alignment over
MIMO interference channels with limited feedback. IEEE Trans. Signal Process. 62(7), 1785–
1795 (2014)
26. Q. Sun, S. Han, C-L. I, Z. Pan, On the ergodic capacity of MIMO NOMA systems. IEEE Wirel.
Commun. Lett. 4(4), 405–408 (2015)
27. J. Choi, On the power allocation for MIMO-NOMA systems with layered transmission. IEEE
Trans. Wirel. Commun. 15(5), 3226–3237 (2016)
28. Z. Chen, Z. Ding, X. Dai, Beamforming for combating inter-cluster and intra-cluster interference in hybrid NOMA systems. IEEE Access 4, 4452–4463 (2016)
29. Z. Ding, R. Schober, H.V. Poor, A general MIMO framework for NOMA downlink and uplink
transmission based on signal alignment. IEEE Trans. Wirel. Commun. 15(6), 4438–4454 (2016)
30. K. Higuchi, Y. Kishyama, Non-orthogonal access with random beamforming and intra-beam
SIC for cellular MIMO downlink, in Proceedings of the IEEE Vehicular Technology Conference
(VTC-Fall), September 2013, pp. 1–5
31. Z. Ding, F. Adachi, H.V. Poor, The application of MIMO to non-orthogonal multiple access.
IEEE Trans. Wirel. Commun. 15(1), 537–552 (2016)
32. X. Chen, Z. Zhang, H.-H. Chen, On distributed antenna system with limited feedback
precoding-opportunities and challenges. IEEE Wirel. Commun. 17(2), 80–88 (2010)
33. D.J. Love, R.W. Heath Jr., V.K.N. Lau, D. Gesbert, B.D. Rao, M. Andrews, An overview of
limited feedback in wireless communication systems. IEEE J. Sel. Areas Commun. 26(8),
1341–1365 (2008)
34. X. Chen, Z. Zhang, C. Zhong, R. Jia, D.W.K. Ng, Fully non-orthogonal communications for
massive access. IEEE Trans. Commun. 66(4), 1717–1731 (2018)
35. X. Chen, Z. Zhang, C. Zhong, D.W.K. Ng, Exploiting multiple-antenna techniques for nonorthogonal multiple access. IEEE J. Sel. Areas Commun. 35(10), 2207–2220 (2017)
36. N. Nonaka, Y. Kishiyama, K. Higuchi, Non-orthogonal multiple access using intra-beam superposition coding and SIC in base station cooperative MIMO cellular downlink, in Proceedings
of the IEEE Vehicular Technology Conference (VTC-Spring), September 2014, pp. 1–5
37. S. Ali, E. Hossain, D.I. Kim, Non-orthogonal multiple access (NOMA) for downlink multiuser
MIMO systems: user clustering, beamforming, and power allocation. IEEE Access 5, 565–577
(2017)
38. M.B. Shahab, M. Irfan, M.F. Kader, S.Y. Shin, User pairing schemes for capacity maximization
in nonorthogonal multiple access systems. Wirel. Commun. Mob. Comput. 16(17), 2884–2894
(2016)
39. Z. Ding, P. Fan, V. Poor, Impact of user pairing on 5G non-orthogonal multiple access downlink
transmissions. IEEE Trans. Veh. Technol. 65(8), 6010–6023 (2016)
40. M.B. Shahab, M.F. Kader, S.Y. Shin, A virtual user pairing scheme to optimally utilize the
spectrum of unpaired users in non-orthogonal multiple access. IEEE Signal Process. Lett.
23(12), 1766–1770 (2016)
256
X. Chen et al.
41. N. Jindal, MIMO broadcast channels with finite-rate feedback. IEEE Trans. Inf. Theory 52(11),
5045–5060 (2006)
42. X. Chen, Z. Zhang, C. Yuen, Adaptive mode selection in multiuser MISO cognitive networks
with limited cooperation and feedback. IEEE Trans. Veh. Technol. 63(4), 1622–1632 (2014)
43. K. Saito, A. Benjebbour, Y. Kishiyama, Y. Okumura, T. Nakamura, Performance and design
of SIC receiver for downlink NOMA with open-loop SU-MIMO, in Proceedings of the IEEE
International Conference on Communication Workshop (ICCW), June 2015, pp. 1161–1165
44. M.B. Shahab, M.F. Kader, S.Y. Shin, On the power allocation of non-orthogonal multiple
access for 5G wireless networks, in Proceedings of the International Conference on Open
Source Systems Technology (ICOSST), December 2016, pp. 89–94
45. K.K. Mukkavilli, A. Sabharwal, E. Erkip, B. Aazhang, On beamforming with finite rate feedback in multiple-antenna systems. IEEE Trans. Inf. Theory 49(10), 2562–2579 (2003)
46. G.K. Karagiannidis, N.C. Sagias, T.A. Tsiftsis, Closed-form statistics for the sum of squared
Nakagami-m variates and its application. IEEE Trans. Commun. 54(8), 1353–1359 (2006)
47. I.S. Gradshteyn, I.M. Ryzhik, Tables of Integrals, Series, and Products (Academic Press, USA,
2007)
Chapter 8
NOMA for Millimeter Wave
Networks
Zhengquan Zhang and Zheng Ma
8.1 Introduction
Millimeter wave (mmWave) communications [1–3] are one of the most important
technologies for 5G wireless networks and beyond, due to the rich spectrum resources
in the mmWave band from 30 to 300 GHz. According to Shannon’s capacity theorem,
increasing the system bandwidth is an effective way to achieve high data rate transmissions. However, compared with conventional low-frequency cellular networks
working in 900 MHz to 3.5 GHz, mmWave networks suffer from high path loss and
directional transmissions, and are sensitive to blockage. To promote the application
of mmWave communications in 5G and beyond, some efforts from both academic
and industry have been devoted.
Conventionally, it is considered that mmWave communications will be mainly
used to achieve high data rate transmissions for enhanced mobile broadband (eMBB)
scenario. Recently, some works also discussed the potential that mmWave communications are used to other application scenarios, such as Internet of Things (IoT)
cloud-enabled autonomous vehicles [4]. Therefore, to further improve spectrum efficiency and support massive connectivity in mmWave networks, the combination of
mmWave communications with other key technologies, e.g., non-orthogonal multiple access (NOMA) [5–7] and multicast transmission [8], is still very important. The
works [9–14] studied NOMA for mmWave networks. The multicast transmission for
mmWave networks was also studied in [14, 15].
Z. Zhang (B) · Z. Ma
Southwest Jiaotong University, West Section, High-tech Zone, Chengdu, Sichuan, China
e-mail: zhang.zhengquan@hotmail.com
Z. Ma
e-mail: zma@swjtu.edu.cn
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_8
257
258
Z. Zhang and Z. Ma
In this chapter, the application of NOMA to mmWave networks is studied, and
unicast, multicast, and cooperative multicast transmissions for mmWave-NOMA
networks are discussed. Their performance in terms of coverage probability, outage
probability, and sum rate is also given. In Sect. 8.2, the fundamentals of mmWave
communications including path loss model, directivity gain model, and user association are briefly discussed. Then, unicast transmissions for single-tier mmWaveNOMA networks are studied in Sect. 8.3, which enable a user pair by NOMA to be
transmitted on the same radio resources. Next, to further improve the network performance, multicast transmissions for single-tier mmWave-NOMA networks are studied
in Sect. 8.4, which distribute media data to all interested users in a non-orthogonal
manner. Furthermore, considering two-tier mmWave heterogeneous networks (HetNets), cooperative multicast transmissions for mmWave-NOMA HetNets are further
studied in Sect. 8.5, followed by summary in Sect. 8.6.
8.2 Fundamentals of mmWave Communications
8.2.1 Path Loss and Small-Scale Fading
mmWave communications exhibit obvious line-of-sight (LOS) and non-LOS (NLOS)
propagations due to the very short wavelength, which is different from conventional
low-frequency cellular networks. To simply characterize LOS and NLOS links, different path loss exponents can be used. Besides, it is assumed that the type of link
observed by users is probabilistic. This means that for a link with length d, the probability that it is a LOS link is p L (d), while it is a NLOS link with probability (w.p)
p N (d) = 1 − p L (d). Therefore, the path loss model can be given by [16, 17]
L S (d) =
C S,L d −αS,L , w.p p L (d),
C S,N d −αS,N , w.p p N (d),
(8.1)
where for s ∈ {L (LOS), N (NLOS)}, α S,s is the path loss exponent for the s link.
The intercept of path loss formula for the s link, C S,s , is a function of reference
distance and wavelength (or carrier frequency) and is 10−2 log10 (4π/λc ) for the close-in
reference distance dref = 1 m [18].
According to [16, 17], the small-scale fading of mmWave link is assumed to be
Nakagami-m fading. The Nakagami-m fading parameters, N S,L and N S,N , are used
to characterize the fading of the mmWave LOS and NLOS links, respectively. Let
h S,i be the channel coefficient of the link between the i-th mmWave small cell and
the user. Then, HS,i = |h S,i |2 follows a normalized Gamma distribution. Similar to
[16, 17], shadowing is ignored.
8 NOMA for Millimeter Wave Networks
259
8.2.2 Directivity Gain
Beamforming [2] is a key technique to overcome high path loss in mmWave communications by using antenna arrays to form directional beams. Especially, analog
beamforming is a simple way to form directional beams by using low-cost phase
shifters (PSs). The deployment of antenna arrays at both BSs and user equipments
(UEs) can be considered. According to 3GPP TR 38.913, up to 256 Tx and Rx BS
antenna elements and 32 Tx and Rx UE antenna elements are assumed. To approximate the beamforming pattern for tractable analysis, sectored antenna model is
available. In this case, the directivity gain can be given by [16, 17, 19]
G(φ) =
G M , |φ| ≤ θ,
G m , |φ| > θ,
(8.2)
where G M and G m are the main and side lobe gains, respectively, φ is a certain
angle, and θ is the beamwidth of the main lobe. Furthermore, according to [19], G M
and G m can be equal to 2π−(2π−θ)ε
and ε, respectively. At the mmWave base station
θ
side, G S,M , G S,m , and θ S denote the main lobe gain, side lobe gain, and beamwidth,
respectively, which are denoted by G U,M , G U,m , and θU at the user side. Therefore,
the directivity gain of the communication link between the user and the i-th mmWave
base station is G i = G S (φ S )G U (φU ), where φ S and φU are the angle of departure
(AoD) and the angle of arrival (AoA) of the signal, respectively. Further, the user is
always assumed to be aligned with its serving base station, B0 , such that the directivity
gain is G 0 = G U,M G S,M . According to [16, 17], the directivity gain of the i-th
interference mmWave link is assumed to be a discrete random variable (RV), whose
probability distribution is G S,i = ak with probability bk , k ∈ {1, 2, 3, 4}, where ak
and bk are constants defined in Table 8.1. āk = G U,MakG S,M is the normalized directional
gain. Note that for macro BSs, the omnidirectional antennas are considered, i.e.,
θ = 2π . As a result, there is no directivity gain.
8.2.3 User Association
Maximum average received power-based user association scheme enables users to
be associated with a base station having the minimum average path loss, which averages the effect of fading and can provide robust performance. Note that this user
Table 8.1 Probability mass function of G S,i (i ≥ 1)
k
1
2
ak
bk
3
4
G U,M G S,M
G U,M G S,m
G U,m G S,M
G U,m G S,m
θU θ S
(2π )2
θU (2π −θ S )
(2π )2
(2π −θU )θ S
(2π )2
(2π −θU )(2π −θ S )
(2π )2
260
Z. Zhang and Z. Ma
association strategy will not necessarily result in the maximum average user performance due to unbalanced load distribution. However, load balancing technique can
be available to change connected users’ distribution between adjacent base stations
by optimizing their handover parameters.
In mmWave networks, each user can associate with one LOS or NLOS mmWave
base station, which depends on the user location and the distance between user and
mmWave base station. According to [16, 17], the probability that user associates
with one LOS base station is
∞
ψ L (x)
p N (t)tdt
(8.3)
g L (x)d x,
e−2πλS 0
A L = BL
0
where the probability that the user has at least one LOS base station is
B L = 1 − e−2πλS
∞
0
r p L (r )dr
(8.4)
,
and given the user observes at least one LOS base station, the conditional probability
density function (PDF) of the distance to its nearest LOS base station is
g L (x) = 2π λ S x p L (x)e−2πλS
x
0
r p L (x)dr
(8.5)
/B L .
The association probability to a NLOS base station is A N = 1 − A L . The probability
that the user has at least one NLOS base station is
B N = 1 − e−2πλS
∞
0
r (1− p L (r ))dr
(8.6)
,
and given the user observes one or more LOS base stations, the conditional PDF of
the distance to its nearest LOS base station is
g N (x) = 2π λ S x(1 − p L (x))e−2πλS
x
0
r (1− p L (x))dr
/B N .
(8.7)
8.3 Unicast Transmissions for mmWave-NOMA Networks
In this section, we will study unicast transmissions for mmWave-NOMA networks.
Conventionally, unicasting employs point-to-point mechanism to achieve data transmissions, which has been widely used in cellular networks. However, with NOMA,
the messages of multiple users can be multiplexed in the power domain by superposition coding at the transmitter side, and then each user decodes its desired message
by successive interference cancelation (SIC). This non-orthogonality overcomes the
loss of degree-of-freedom (DoF) caused by orthogonal transmission at the cost of
increasing processing complexity.
8 NOMA for Millimeter Wave Networks
261
8.3.1 System Model
Figure 8.1 illustrates the system model of single-tier mmWave-NOMA networks with
downlink unicast transmissions. The mmWave base stations are located according to
a homogeneous Poisson point process (HPPP) Φ S with density λ S and are assumed to
have same transmit power PS . We consider an M-user NOMA scenario, which means
that a NOMA user pair with M users sorted by ascending order. The M users are
randomly distributed in an mmWave service area. After channel ordering, we have
H1 L(r1 )
I1 +σ 2
H2 L(r2 )
≤
I2 +σ 2
M
satisfies m=1
≤
··· ≤
HM L(r M )
.
I M +σ 2
The power allocated to the m-th user is Pm > 0
and
Pm = PS . According to the principle of NOMA, the weak users
are allocated to more power in order to ensure that they can decode successfully.
Therefore, we have P1 ≥ P2 ≥ · · · ≥ PM .
NOMA transmissions enable the messages of all users from a NOMA user pair
to be multiplexed in the power domain to form a superposed signal, and then this
superposed signal is transmitted on the same radio frequency. Therefore, the signal
transmitted by the base station to users is
x=
M
Pm xm ,
(8.8)
m=1
where xm is the message of the m-th user and satisfies E[|xm |] = 1. The m-th user
not only receives the signal from its serving base station, B0 , but also suffers from
co-channel interference (CCI) from neighboring base stations and can be expressed
as
Fig. 8.1 System model of single-tier mmWave-NOMA networks with downlink unicast transmissions
262
Z. Zhang and Z. Ma
ym1
=
√
h m G 0 L(rm )x
,
h m,i G i L(rm,i )xi +n m
(8.9)
Bi ∈Φ\B0
ICCI,m
where n m is the additive Gaussian noise with power σ 2 for the m-th user.
When the users receive signal from base stations, they employ successive SIC to
decode their desired messages. According to the principle of SIC, the m-th user first
decodes and cancels the messages of UEi , i = 1, . . . , m − 1 from the received sum
signal orderly. The signal-to-interference-plus-noise ratio (SINR) that UEm detects
UEi can be expressed as
SINR1m,i = M
Hm G 0 L(rm )Pi
SINR1m = M
Hm G 0 L(rm )Pm
2
j=i+1 Hm G 0 L(r m )P j + ICCI,m + σ
.
(8.10)
Then, UEm decodes its own message after SIC with detecting SINR given by
j=m+1
Hm G 0 L(rm )P j + ICCI,m + σ 2
.
(8.11)
Note that for the M-th user, its SINR after SIC is given by
SINR1M =
HM G 0 L(r M )PM
.
ICCI,M + σ 2
(8.12)
8.3.2 Performance Analysis
G 0 L(r )
In mmWave networks, for a random variable (RV), X = HICCI
, according to [16,
+σ 2
17], its complementary cumulative distribution function (CCDF) can be expressed
as
(8.13)
F̄(T ) = A L F̄L (T ) + A N F̄N (T ),
while its CDF is F(T ) = 1 − F̄(T ). For s ∈ {L , N }, F̄s (T ) is the conditional CCDF
given that the user is associated with a base station in Φs and can be written as
F̄L (T ) ≈
NL
n=1
(−1)
n+1
NL
n
0
∞ −
e
α
nη L x S,L T σ S2
C L G S,0
+Q n (T,x)+Vn (T,x)
f L (x)d x, (8.14)
8 NOMA for Millimeter Wave Networks
263
and
∞
NN
−
NN
F̄N (T ) ≈
e
(−1)n+1
n
0
n=1
where
B N g N (x) −2πλ 0(C L /C N )1/αL x α N /αL
e
AN
f N (x) =
Q n (T, x) = 2π λ S
4
Vn (T, x) = 2π λ S
bk
Wn (T, x) = 2π λ S
4
bk
8.3.2.1
bk
k=1
4
k=1
k=1
Z n (T, x) = 2π λ S
+Wn (T,x)+Z n (T,x)
f N (x)d x, (8.15)
B L g L (x) −2πλ 0(C N /C L )1/α N x αL /α N (1− pL (t))tdt
,
e
AL
f L (x) =
and
α
nη N x S,N T σ S2
C N G S,0
4
∞
F(N L ,
x
∞
∞
k=1
∞
x
nη L āk T x αS,L
) p L (t)tdt,
N L t αS,L
(8.17)
(8.18)
nC N η L āk T x αS,L
) p N (t)tdt,
C L N N t αS,N
(8.19)
F(N L ,
nC L η N āk T x αS,N
) p L (t)tdt,
C N N L t αS,L
(8.20)
F(N N ,
nη N āk T x αS,N
) p N (t)tdt.
N N t αS,N
(8.21)
ψ N (x)
bk
,
F(N N ,
ψ L (x)
p L (t)tdt
(8.16)
Coverage Probability
The coverage probability can be used to characterize the quality of network coverage,
which refers to the probability that the SINR received at user exceeds certain threshold
T . According to SIC decoding, the coverage probability of the m-th ordered user in
a NOMA user pair is defined as
1
(T ) = P[SINR1m,1 > T, . . . , SINR1m,m > T ].
Pc,m
(8.22)
Since the maximum SINR that the m-th user detects the message of the i-th user
from the superposed signal is lim Hm →∞ SINR1m,i = M Pi P , the coverage probabilj=i+1
j
ity is equal to zero when the SINR threshold T is equal or greater than this maximum
detecting SINR. When the SINR threshold T is below this maximum detecting SINR,
substituting (8.10) and (8.11) into (8.22) and according to the law of total probability,
the coverage probability can be rewritten as
264
Z. Zhang and Z. Ma
1
Pc,m
(T ) = P[X m > b1 , . . . , X m > bm ]
n
n
[1 − F̄(max(b1 , . . . , bm ))]i F̄ n−i (max(b1 , . . . , bm )),
i
i=m
(8.23)
G 0 L(rm )
1
,
b
=
,
i
=
1,
.
.
.
,
m,
and
(a)
follows
(8.13)
and
where X m = HImCCI,m
i
Pi
M
+σ 2
(a)
=1 −
T
−
j=i+1
Pj
the property of order statistics [22].
Therefore, the coverage probability of the m-th ordered user in a NOMA user pair
with M users in mmWave-NOMA networks can be finally expressed as
⎧
p1
pm
⎪
⎨0, T > Mj=2 P j or, . . . , or T > Mj=m+1 P j ,
1
n
Pc,m (T ) =
n
⎪
[1 − F̄(max(b1 , . . . , bm ))]i F̄ n−i (max(b1 , . . . , bm )), otherwise.
⎩1 −
i=m i
(8.24)
8.3.2.2
Outage Probability
The outage probability is used to characterize the probability that the user cannot
achieve a target rate τ , and is defined as Po P[R < τ ]. Further, due to R = log2 (1 +
SINR) < τ , its form related with SINR can be expressed as Po P[SINR < 2τ − 1].
We assume that the target rate for the m-th user is τm > 0 and define the outage event
E m,i = {SINRm,i < γi }, where γi = 2τi − 1. This means that the m-th user failed
to decode the message of the i-th user. Correspondingly, the complementary outage
event is Ē m,i = {SINRm,i ≥ γi }. According to the principle of NOMA, the outage
probability of the m-th user can be expressed as
1
= P[E m,1 ∪ E m,2 ∪ · · · ∪ E m,m ] = 1 − P[ Ē m,1 ∩ Ē m,2 ∩ · · · ∩ Ē m,m ].
Po,m
(8.25)
With SIC decoding, when the SINR that the m-th user detects the message of the
i-th user is smaller than its maximum SINR, SINR1m,i = M Pi P , the SIC decoding
j=i+1
failed. That is, when γ1 ≥
P
M 1
is equal to one. When γ1 <
j=2 P j
P
M 1
j=2 P j
or . . . or γm ≥
M
Pm
j=m+1
and . . . and γm <
j
, the outage probability
Pj
Pm
M
j=m+1
Pj
, substituting (8.10)
and (8.11) into (8.25), the outage probability can be written as
1
Po,m
= 1 − P[X m > c1 , . . . , X m > cm ]
n
(b)
=
i=m
n
[1 − F̄(max(c1 , . . . , cm ))]i F̄ n−i (max(c1 , . . . , cm )),
i
(8.26)
8 NOMA for Millimeter Wave Networks
where ci =
Pi γi−1 −
1
M
j=i+1
Pj
265
, and (b) follows the property of order statistics and the
complementary property of CCDF and CDF. Therefore, the outage probability of the
m-th ordered user in a NOMA user pair with M users can be finally expressed as
1
Po,m
⎧
Pm
P1
⎪
⎨1, γ1 ≥ Mj=2 P j or . . . or γm ≥ Mj=m+1 P j ,
n
=
n
⎪
[1 − F̄(max(c1 , . . . , cm ))]i F̄ n−i (max(c1 , . . . , cm )), otherwise.
⎩
i=m i
(8.27)
8.3.2.3
Sum Rate
To the m-th user’s message, xm , it should ensure that all users after it can also decode
the message xm in order to perform SIC successfully. Therefore, the data rate for the
m-th user is equal to
1
1
, Rm+1,m
, . . . , R 1M,m ,
Rm1 = min Rm,m
(8.28)
1
where Rn,m
= log2 (1 + SINR1n,m ), n = m, . . . , M is the rate that the n-th user
decodes the data xm . Further, we have
Hn G 0 L(rn )Pm
1
Rn,m
= log2 1 + M
Hn G 0 L(rn )P j + ICCI,n + σ 2
⎞
P
m
⎠.
= log2 ⎝1 + M
ICCI,n +σ 2
P
+
j=m+1 j
Hn G 0 L(rn )
j=m+1
⎛
2)
1)
≤ H2 L(r
≤ ··· ≤
Due to H1 L(r
I1
I2
Therefore, we have
HM L(r M )
,
IM
(8.29)
1
1
we have Rm,m
≤ Rm+1,m
≤ · · · ≤ R 1M,m .
1
1
1
.
, Rm+1,m
, . . . , R 1M,m = Rm,m
Rm1 = min Rm,m
(8.30)
The sum rate for a typical NOMA user pair consisting of M users is defined as
the sum of the average rate for each user and is
τt1
M
m=1
E[Rm1 ]
M
1
=
E[ln(1 + SINR1m )].
ln 2 m=1
(8.31)
266
Z. Zhang and Z. Ma
For the m-th ordered user, m = 1, . . . , M − 1, the average rate is taken over both
the spatial PPP and the fading distribution. We have
1
E[ln(1 + SINR1m )]
ln 2 ⎧
⎨ ln 1+ M Pm P
(c) 1
j=m+1 j
=
P Xm >
AL
ln 2 ⎩
r >0 0
R̄m =
+ AN
=
r >0
ln 1+ M Pm
j=m+1 P j
F̄m
0
where (c) follows E[W ] =
lim
ICCI,m →0,Hm →∞
P Xm >
0
j=i+1 P j
ln 1+ M Pm
Pm
et −1
t>0
−
1
M
j=m+1
Pm
et −1
Pm
et −1
Pj
−
−
1
M
j=m+1
1
M
j=m+1
Pj
Pj
dt f L (r )dr
dt f N (r )dr
⎫
⎬
⎭
dt,
(8.32)
P(W > t)dt for a positive RV, W [20], and
ln 1 + M
Hm G 0 L(rm )Pm
2
l=m+1 H M G 0 L(rm )P j + ICCI,m + σ
= ln 1 + M
Pm
j=m+1 P j
.
(8.33)
For the M-ordered user, the integral domain of variable t in (8.32) is (0, ∞). Therefore, the average rate can be written as
R̄ M =
∞
0
F̄M
et − 1
dt.
PM
(8.34)
Combining (8.31), (8.32), and (8.34), the sum rate for a NOMA user pair in
mmWave networks can be expressed as
M−1
ln 1+Pm ( Mj=i+1 P j )−1
1
1
τt1 =
F̄m
dt
Pm
ln 2 m=1 0
− M
j=m+1 P j
et −1
∞
et − 1
dt .
+
F̄M
PM
0
(8.35)
8.3.3 Numerical Results
To evaluate the performance of mmWave-NOMA networks, the parameters are used
as follows: the carrier frequency is 28 GHz and the system bandwidth is 100 MHz;
the base station transmit power is 30 dBm; the density of mmWave base station is
1
λ S = π200
2 ; the path loss exponents for LOS and NLOS links are set as 2 and 4,
8 NOMA for Millimeter Wave Networks
267
1
SINR Coverage Probability
0.9
0.8
0.7
0.6
0.5
0.4
0.3
Analytical Results
Simulations
mmWave−NOMA User1
mmWave−NOMA User2
mmWave−OMA User1
mmWave−OMA User2
0.2
0.1
0
−10
−8
−6
−4
−2
0
2
4
6
8
10
SINR Threshold (dB)
Fig. 8.2 SINR coverage probabilities of mmWave-NOMA networks with fixed power ratio
(0.8, 0.2)
respectively, and a NOMA user pair consists of two random users with ascending
order. To give a fair comparison to NOMA, OMA employs equal resource allocation
for each user.
Figure 8.2 shows the coverage probabilities of mmWave-NOMA networks with
fixed power ratio (0.8, 0.2). The results show that the coverage probabilities of both
NOMA user1 and NOMA user2 are lower than that of OMA. This is because interuser interference introduced by NOMA deteriorates the SINR received at users. The
results also show that the coverage probabilities of NOMA users decline to zero when
the SINR threshold exceeds a certain value (i.e., PP21 ), while OMA can still achieve
some coverage. The reason is that the maximum SINR detecting NOMA user1 is
limited by PP21 , according to SIC decoding.
Figure 8.3 shows the outage probabilities of mmWave-NOMA networks with fixed
power ratios (0.8, 0.2) and (0.9, 0.1). The results show that compared with OMA, the
outage probability of the NOMA-weak user (i.e., User1) can be improved, while the
NOMA-strong user (i.e., User2) suffers from some loss. The results also show that
with the increase of power allocated to the NOMA-weak user, the outage probability
of the NOMA-weak user can be further improved, while the NOMA-strong user’s
outage probability becomes worse.
Figure 8.4 shows the average rates for mmWave-NOMA networks with fixed
power ratios (0.8, 0.2) and (0.9, 0.1). The results show that NOMA can achieve
higher data rate than conventional OMA. This is because NOMA can enable all
users in a NOMA user pair to occupy whole system bandwidth such that there is no
loss in DoF, at the cost of the increase of processing complexity and the introduction
268
Z. Zhang and Z. Ma
(a)
0
Outage Probability
10
−1
10
Analytical Results
Simulations
mmWave−NOMA User1
mmWave−NOMA User2
mmWave−OMA User1
mmWave−OMA User2
−2
10
0
0.1
0.2
0.3
0.4
0.5
0.6
Rate Threshold (b/s/Hz)
(b)
0
Outage Probability
10
−1
10
Analytical Results
Simulations
mmWave−NOMA User1
mmWave−NOMA User2
mmWave−OMA User1
mmWave−OMA User2
−2
10
0
0.1
0.2
0.3
0.4
0.5
0.6
Rate Threshold (dB)
Fig. 8.3 Outage probabilities of mmWave-NOMA networks with fixed power ratios: a (0.8, 0.2);
b (0.9, 0.1)
of inter-user interference. However, the results also show that for a certain transmit
power range, e.g., [0, 5] dBm, NOMA with power ratio 0.9 can achieve higher data
rate for the weak user (i.e, User1), while suffers from a lower data rate for the strong
user (i.e., User2) and sum rate. This is because the fixed power ratio for NOMA
is not optimal for all transmit powers. Comparing Fig. 8.4a, b, the weak user with
power ratio 0.9 can achieve higher average rate than that of power ratio 0.8, while
the opposite trend is for the strong user, which results in a lower sum rate.
8 NOMA for Millimeter Wave Networks
269
(a) 9
mmWave−NOMA Sum Rate
mmWave−NOMA User1
mmWave−NOMA User2
mmWave−OMA Sum Rate
mmWave−OMA User1
mmWave−OMA User2
8
Rate (b/s/Hz)
7
6
5
4
3
2
1
0
0
5
10
15
20
25
30
25
30
Transmit Power (dBm)
(b)
9
mmWave−NOMA Sum Rate
mmWave−NOMA User1
mmWave−NOMA User2
mmWave−OMA Sum Rate
mmWave−OMA User1
mmWave−OMA User2
8
Rate (b/s/Hz)
7
6
5
4
3
2
1
0
0
5
10
15
20
Transmit Power (dBm)
Fig. 8.4 Average rates for mmWave-NOMA networks with fixed power ratios: a (0.8, 0.2);
b (0.9, 0.1)
8.4 Multicast Transmissions for mmWave-NOMA
Networks
Compared with unicast transmissions discussed in Sect. 8.3, multicast transmissions
employ the point-to-multipoint mechanism to distribute the same media data to multiple interested users on the same radio resources such that higher spectrum efficiency
can be achieved. However, conventional multicast transmissions just deliver single data stream with a low data rate in order to ensure that most users can decode
the media. With the help of NOMA, multicast transmissions can be enhanced by
270
Z. Zhang and Z. Ma
multiplexing multiple data streams with different data rates in the power domain. As
a result, users can decode data streams according to their channel conditions. In this
section, we will discuss multicast transmissions for mmWave-NOMA networks.
8.4.1 System Model
Figure 8.5 illustrates the system model of multicast transmissions in mmWaveNOMA networks. The mmWave base stations are located according to an HPPP
Φ S with density λ S . The mmWave base stations are assumed to have same transmit power PS . The users are located according to an HPPP ΦU with density λU .
With NOMA, N -layer superposition coded multicast transmission can be achieved
by power allocation, which consists of one primary layer and N − 1 secondary layers. The primary layer carries the basic data, while the secondary layers carry the
corresponding enhanced data. The fixed data rate for each layer is also assumed.
Generally, the scalable media can be encoded into one basic data and N − 1
enhanced data by source layered coding, where the basic data provides the basic
service quality, while the enhanced data are used to improve the service quality.
Note that the enhanced data cannot work without the basic data. With NOMAenabled multicast transmissions, the basic data and enhanced data are multiplexed
in the power domain to form a superposed signal. Then, this superposed signal is
distributed on the same radio resources to all users who desire to receive the media.
When users receive this superposed signal, they first decode the basic data to obtain
the basic service quality directly, then try to decode the enhanced data to achieve
better service quality. For the weak users, they can just decode the low-rate basic
data, while the strong users can decode both the low-rate basic data and high-rate
enhanced data. Therefore, NOMA-enabled multicast transmissions can fully utilize
the channel difference between users to improve the performance of the strong users.
This technique overcomes the shortage of conventional multicast transmissions that
Fig. 8.5 System model of single-tier mmWave-NOMA networks with multicast transmissions
8 NOMA for Millimeter Wave Networks
271
the strong users cannot fully utilize their good channel to obtain better service quality,
as the conventional one just transmits single low-rate data to ensure that all users can
decode the media successfully.
Without loss of generality, two-layer superposition coded multicast transmission
is considered as an example. The mmWave base station transmits a superposed signal
to all users within its coverage as
x=
√
αpxB + 1 − αpxE ,
(8.36)
where 0 < α p < 1 is the power allocation factor (PAF), x B and x E are the transmit
messages of the primary and secondary layers, respectively. The signal received at
the user with random distance, d0 , from its serving mmWave base station can be
expressed as
yS2 = h S,0 G S,0 PS L S (d0 )x +
Xi ∈Φ S \B0
h S,i G S,i PS L S (di )xi +n S .
(8.37)
IS
After receiving the superposed signal, the user first decodes the primary layer,
and then cancel it from the received signal before decoding the secondary layer.
Therefore, substituting (8.36) into (8.37), the SINRs of detecting the primary and
secondary layers can be written, respectively, as
SINR2S,P L =
α p HS G S,0 L S (d0 )
,
HS,i G S,i L S (di ) +σ S2
(1 − α p )HS G S,0 L S (d0 ) +
X ∈Φ \B
i S 0
IS
(8.38)
and
SINR2S,SL =
(1 − α p )HS G S,0 L S (d0 )
Xi ∈Φ S \B0
HS,i G S,i L S (di ) +σ S2
.
(8.39)
IS
Note that σ S2 is the thermal noise power, normalized by transmit power, PS .
8.4.2 Performance Analysis
8.4.2.1
Coverage Probability
The coverage probability that users can decode the primary layer relative to the SINR
threshold TP L can be written as
272
Z. Zhang and Z. Ma
2
2
Pc,P
L (T P L ) = E R [P[SINR P L > T P L | R = r ]].
(8.40)
Considering the maximum SINR for detecting the primary layer, lim Hs →∞ SINR2P L =
αp
, the integral domain is
1−α p
!
"
D = (HS ) | SINR2P L > TP L
#
(8.41)
TP L (I S + σ S2 )
αp
.
|TP L <
= (HS ) | HS >
(α p − (1 − α p )TP L )G S,0 L S (r0 )
1 − αp
Therefore, the coverage probability can be expressed as
$ $
2
Pc,S,P
P HS >
(T
,
α
)
=
E
P
L
p
R
L
TP L (I S + σ S2 )
|R=r
(α p − (1 − α p )TP L )G S,0 L S (r0 )
%%
.
(8.42)
According to [16, 17] and after some manipulations, the coverage probability of
the primary layer can be obtained as
2
Pc,S,P
L (T P L , α p )
⎧
α
nη L x S,L T P L σ S2
⎪
⎪
⎪
−
+Q n (T P L ,x)+Vn (T P L ,x)
N
∞
⎪
L
C
(α
−(1−α
)T
)G
⎪
p
p
L
PL S
⎪
⎪
(−1)n+1 NnL
e
f L (x)d x
⎪ AL
⎪
⎪
n=1
0
⎨
α S,N
2
nη N x
TP L σS
≈
+Wn (T P L ,x)+Z n (T P L ,x)
−
N
⎪
∞
N
⎪
C
(α
−(1−α
)T
)G
p
p
α
N
P
L
S
⎪
⎪
f N (x)d x, T P L < 1−αp ,
e
+A N
(−1)n+1 NnN
⎪
⎪
p
⎪
n=1
⎪
0
⎪
αp
⎪
⎩ 0,
T
≥
.
PL
1−α p
(8.43)
The coverage probability of both the primary and secondary layers relative to the
SINR thresholds TP L and TSL can be written as
2
2
2
Pc,P
SL (T P L , TSL , α p ) = E R [P[{SINR P L > T P L ∩ SINR SL > TSL } | R = r ]].
(8.44)
Its integral domain is
&
'
D = (HS ) | {SINR2P L > T P L ∩ SINR2S L > TS L }
))
(
(
TS L (I S + σ S2 )
T P L (I S + σ S2 )
.
= (HS ) | HS >
∩ HS >
(α p − (1 − α p )T P L )G S,0 L S (r0 )
(1 − α p )G S,0 L S (r0 )
Let
TP L (I S +σ S2 )
(α p −(1−α p )TP L )G S,0 L S (r0 )
=
TS L (I S +σ S2 )
,
(1−α p )G S,0 L S (r0 )
we can obtain α p =
Further, the coverage probability can be expressed as
(8.45)
TP L (1+TS L )
.
TS L +TP L (1+TS L )
8 NOMA for Millimeter Wave Networks
273
2
Pc,S,P
SL (T P L , TSL , α p )
* *
++
⎧
TP L (I S +σ S2 )
P L (1+TS L )
⎨ E R P HS >
|
R
=
r
, α p ≤ TS LT+T
,
(α p −(1−α p )TP L )G S,0 L S (r0 )
P L (1+TS L )
* *
++
=
2
T
(I
+σ
)
SL S
⎩E P H >
P L (1+TS L )
S
| R = r , α p > TS LT+T
.
S
R
(1−α p )G S,0 L S (r0 )
P L (1+TS L )
(8.46)
Let
$ $
%%
TSL (I S + σ S2 )
2
(8.47)
|R=r .
(TSL , α p ) = E R P HS >
Pc,S,SL
(1 − α p )G S,0 L S (r0 )
2
Therefore, the coverage probability, Pc,P
SL (T P L , TSL , α p ), can finally be written as
2
Pc,S,P
SL (T P L , TSL , α p ) =
(
TP L (1+TS L )
2
Pc,S,P
L (T P L , α p ), α p ≤ TS L +TP L (1+TS L ) ,
TP L (1+TS L )
2
Pc,S,SL (TSL , α p ), α p > TS L +TP L (1+TS L ) ,
(8.48)
2
2
where Pc,S,P
L (T P L , α p ) is expressed as in (8.43) and Pc,S,SL (TSL , α p ) is approximated as
2
Pc,S,SL
(TSL , α p )
∞
NL
−
NL
≈ AL
e
(−1)n+1
n
0
n=1
∞
NN
−
n+1 N N
+ AN
e
(−1)
n
0
n=1
8.4.2.2
α
nη L x S,L TS L σ S2
C L (1−α p )G S
+Q n (TS L ,x)+Vn (TS L ,x)
α
nη N x S,N TS L σ S2
C N (1−α p )G S
+Wn (TS L ,x)+Z n (TS L ,x)
f L (x)d x
f N (x)d x.
(8.49)
Average Number of Served Users
For the mmWave multicast cluster, Bo , the average number of served users by the
primary layer can be expressed as
⎡
Eo [N P2 L ] Eo ⎣
y∈ΦU,Bo
⎤
I(E 2P L (y))⎦ ,
(8.50)
where E 2P L (y) = {SINR2S,P L ≥ 2 R P L − 1}. Given the user density λU , the average
number of users covered by a beam with width θ S is λU θ S (2π λ S )−1 . Further, considering the average coverage probability, Eo [N P2 L ] can be finally expressed as
2
−1
Eo [N P2 L ] = λU Pc,S,P
L (T P L , α p )θ S (2π λ S ) ,
(8.51)
274
Z. Zhang and Z. Ma
The average number of served users, who can decode the data contained in both
the primary and secondary layers, can be written as
⎡
Eo [N P2 SL ] Eo ⎣
y∈ΦU,Bo
⎤
I(E 2P SL (y))⎦ ,
(8.52)
where E 2P SL (y) = {{SINR2S,P L ≥ 2 R P L − 1} ∩ {SINR2S,SL ≥ 2 R S L − 1}}. Similarly,
Eo [N P2 SL ] can be finally expressed as
2
−1
Eo [N P2 SL ] = λU Pc,S,P
SL (T P L , TSL , α p )θ S (2π λ S ) .
8.4.2.3
(8.53)
Sum Rate
The sum rate for NOMA multicast is defined as the mean of the sum rate of all users
in coverage of the multicast cluster, who successfully decode the primary layer with
data rate, R P L , or both the primary and secondary layers with data rate, R P L + R SL .
This is given by
2
= (Eo [N P2 L ] − Eo [N P2 SL ])R P L + Eo [N P2 SL ](R P L + R SL )
R̄sum
= Eo [N P2 L ]R P L + Eo [N P2 SL ]R SL .
(8.54)
Note that Eo [N P2 L ] − Eo [N P2 SL ] is the average number of served users by the
mmWave multicast cluster, who can only decode the primary layer.
Combining (8.43), (8.49), (8.51), (8.53), and (8.54), the sum rate for the mmWave
multicast cluster can be expressed as
2
R̄sum
⎧ (R +R )λ P 2 (T ,α )θ
PL
S L U c,S,P L
PL
p S
P L (1+TS L )
⎪
, α p ≤ TS LT+T
and TPL <
⎪
2πλ S
P L (1+TS L )
⎪
⎪
2
2
⎨ R P L Pc,S,P L (TP L ,α p )λU θS
R S L Pc,S,S L (TP L ,TS L ,α p )λU θ S
+
,
2πλ S
2πλ S
=
α
TP L (1+TS L )
⎪
α p > TS L +TP L (1+TS L ) and TPL < 1−αp p ,
⎪
⎪
⎪
αp
⎩ 0,
TP L ≥ 1−α p ,
αp
,
1−αp
(8.55)
where, TP L = 2 R P L − 1 and TSL = 2 R S L − 1.
8.4.3 Numerical Results
Figure 8.6 depicts the coverage probabilities of multicast transmissions for mmWaveNOMA networks with fixed power ratio (0.8, 0.2). It can be observed that NOMA
multicast can provide a similar primary coverage layer as the conventional one and
achieve a secondary coverage layer as well. However, when the SINR threshold
8 NOMA for Millimeter Wave Networks
275
0.8
SINR Coverage Probability
0.7
0.6
0.5
0.4
0.3
Analytical results
0.2
Simulations
Multicast
NOMA multicast (PL)
0.1
0
−10
NOMA multicast (SL)
−8
−6
−4
−2
0
2
4
6
8
10
SINR Threshold (dB)
Fig. 8.6 SINR coverage probabilities of multicast transmission for mmWave-NOMA networks
with fixed power ratio (0.8, 0.2)
αP
, that users can detect the primary layer,
T is larger than the maximum SINR, 1−α
P
the SINR coverage probabilities of the primary and secondary layers are equal to
zero. This is because if users failed to decode the primary layer, they do not further
decode the secondary layer through SIC.
Figure 8.7 depicts the sum rates for multicast transmission for mmWave-NOMA
networks with fixed power ratios (0.8, 0.2) and (0.95, 0.05) and the secondary layer
data rate, R SL = 4 b/s/Hz. The results show that NOMA multicast can achieve a
significant gain of sum rate, compared with the conventional one in the low multicast
rate region. This is because NOMA multicast can fully utilize the channel conditions
of strong users. However, the maximum SINR for detecting the primary layer is
αP
such that NOMA multicast cannot work when the multicast rate for the primary
1−α P
αP
). Comparing Fig. 8.7a, b, more power is allocated to
layer exceeds log2 (1 + 1−α
P
the primary layer, a higher multicast rate for the primary layer can be provided, yet
a lower sum multicast rate is achieved. This means that NOMA multicast gradually
degrades to the conventional one, with the increase of power allocated to the primary
layer.
Figure 8.8 depicts the sum rates for NOMA multicast with different power allocations, given 0.4, 1 b/s/Hz for the primary layer and 2, 4, 6 b/s/Hz for the secondary
layer. It is shown that for each multicast rate pair, as the power ratio grows, the sum
rate first experiences a sharp rise to the maximum, then it falls slowly in the medium
power ratio region. Finally, it declines rapidly to the lowest point in the high power
ratio region. Furthermore, with fixed multicast rate for the secondary layer, the sum
rate of the primary layer with rate 1 b/s/Hz is higher than that of the primary layer
with rate 0.4 b/s/Hz, which requires more power to be allocated to the primary layer.
The sum rate of the secondary layer with rate 6 b/s/Hz is higher than that of the
276
Z. Zhang and Z. Ma
(a)
100
90
80
Sum Rate (b/s/Hz)
70
60
50
40
Analytical results
Simulations
30
Multicast
NOMA multicast
20
10
0
0
1
2
3
4
5
6
7
8
9
10
Multicast Rate (b/s/Hz)
(b)
100
90
Sum Rate (b/s/Hz)
80
70
60
50
40
Analytical results
Simulations
Multicast
NOMA multicast
30
20
10
0
0
1
2
3
4
5
6
7
8
9
10
Multicast Rate (b/s/Hz)
Fig. 8.7 Sum rates for multicast transmission for mmWave-NOMA networks with fixed power
ratios. a (0.8, 0.2); b (0.95, 0.05)
secondary layer with rates 4 and 2 b/s/Hz, when the multicast rate for the primary
layer is fixed.
8.5 Cooperative Multicast Transmissions
for mmWave-NOMA HetNets
In this section, we will further discuss multicast transmissions in a two-tier mmWaveNOMA HetNet consisting of one low-frequency macro base station (MBS) tier and
8 NOMA for Millimeter Wave Networks
277
90
80
Sum Rate (b/s/Hz)
70
60
50
40
Analytical results
Simulations
PL0.4−SL2
PL0.4−SL4
PL0.4−SL6
PL1.0−SL2
PL1.0−SL4
PL1.0−SL6
30
20
10
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Power Ratio
Fig. 8.8 Sum rate for multicast transmissions for mmWave-NOMA networks with different power
ratios
one mmWave base station tier, which is the typical deployment for mmWave wireless
networks, and introduce a cooperative multicast scheme for mmWave-NOMA networks to further improve the system performance. The cooperative multicast scheme
can increase the success probability of decoding the primary layer with the help of
cooperation from MBSs.
8.5.1 System Model
Figure 8.9 illustrates the system model of cooperative NOMA-enabled multicast
transmissions for a two-tier mmWave HetNet, which consists of one low-frequency
MBS tier with transmit power PM and one mmWave base station tier with transmit
power PS . Cooperative NOMA-enabled multicast enables the MBS tier to cooperatively transmit the primary layer with a low data rate. With the cooperation of MBSs,
the users who failed to decode the primary layer of the superposed signal from the
mmWave base station tier will try to decode it from the macro BS tier. If the primary
layer is decoded successfully, they cancel it from the superposed signal and further
decode the secondary layer. Thus, this increases the success probability that decodes
the primary and secondary layers, such that the NOMA multicast performance can
be improved.
The signal received at the user with random distance, d0 , from its serving and
interfering MBSs can be expressed as
−α /2
3
yM
= h M,0 PM d0 M x B +
Xi ∈Φ M \B0
h M,i
−α M /2
PM di
IM
xi,B +n M ,
(8.56)
278
Z. Zhang and Z. Ma
Fig. 8.9 System model of cooperative multicast transmissions for mmWave-NOMA HetNets
where n M is the additive Gaussian noise with power σ M2 . Therefore, the SINR of
decoding the data can be written as
SINR3P L =
HM,0 PM d0−α M
Xi ∈Φ M \B0
HM,i PM di−α M +σ M2
.
(8.57)
IM
The SINR of decoding the primary layer, SINR3S,P L , from the mmWave base station
tier, can be written as in (8.38), while the SINR of the secondary layer, SINR3S,SL ,
can be expressed as in (8.39).
8.5.2 Performance Analysis
8.5.2.1
Coverage Probability
With cooperative NOMA multicast in a two-tier mmWave HetNet, when users decode
the primary layer from a macro or mmWave base station successfully, the basic data
can be recovered. Therefore, the coverage probability of the primary layer can be
expressed as
3
3
3
Pc,P
L (T P L , α p ) = P{{SINR M,P L > T P L } ∪ {SINR S,P L > T P L }}
= P{SINR3M,P L > TP L } + P{SINR3S,P L > TP L }
−
P{SINR3M,P L
>
TP L }P{SINR3S,P L
(8.58)
> TP L },
where P{SINR3S,P L > TP L } is obtained as in (8.43) due to SINR3S,P L = SINR2S,P L .
3
And according to [21], Pc,M,P
L (T P L ) can be obtained as
8 NOMA for Millimeter Wave Networks
279
3
3
Pc,M,P
L (T P L ) = P{SINR M,P L > T P L }
$
%
HM PM R −α M
P
>
T
|
R
=
r
f R (r )dr
=
PL
I M + σ M2
r >0
∞
2 −1 α M /2
= π λM
e−πλ M (1+ρ(TP L ,α M ))x−TP L (PM /σ M ) x d x,
(8.59)
0
where
2/α M
ρ(TP L , α M ) = TP L
∞
−2/α M
TP L
(1 + t α M /2 )−1 dt.
(8.60)
The coverage probability of the secondary layer can be expressed as
3
3
3
Pc,P
SL (T P L , TSL , α p ) = P{{SINR S,SL > TSL ∩ SINR S,P L > T P L }
∪ {SINR3S,SL > TSL ∩ SINR3M,P L > TP L }}.
(8.61)
After some manipulations,
3
3
3
Pc,P
S L (T P L , TS L , α p ) = P{SINR S,S L > TS L ∩ SINR S,P L > T P L }
+ P{SINR3S,S L > TS L }P{SINR3M,P L > T P L }
− P{SINR3S,S L > TS L ∩ SINR3S,P L > T P L }P{SINR3M,P L > T P L }.
(8.62)
Since SINR3S,SL = SINR2S,SL and SINR3S,P L = SINR2S,P L , P{SINR3S,SL > TSL ∩
SINR3S,P L > TP L } can be written as in (8.48), and P{SINR3S,SL > TSL } can be
expressed as in (8.49). Therefore, combining (8.48), (8.49), (8.58), and (8.62),
3
Pc,P
SL (T P L , TSL , α p ) can be finally expressed as
3
2
3
Pc,P
SL (T P L , TSL , α p ) = Pc,S,P SL (T P L , TSL , α p )(1 − Pc,M,P L (T P L ))
3
2
+ Pc,S,SL
(TSL , α p )Pc,M,P
L (T P L ).
8.5.2.2
(8.63)
Average Number of Served Users
For the mmWave multicast cluster, Bo , the average number of served users by the
primary layer can be expressed as
⎡
Eo [N P3 L ] Eo ⎣
y∈ΦU,Bo
⎤
I(E 3P L (y))⎦ ,
(8.64)
where E 3P L (y) = {SINR3M,P L ≥ 2 R P L − 1 ∪ SINR3S,P L ≥ 2 R P L − 1}. Given the user
density λU , the average number of users covered by a beam with width θ S is
280
Z. Zhang and Z. Ma
λU θ S (2π λ S )−1 . Further, considering the average coverage probability, Eo [N P3 L ] can
be finally expressed as
3
−1
Eo [N P3 L ] = λU Pc,P
L (T P L )θ S (2π λ S ) .
(8.65)
The average number of served users by the secondary layer can be expressed as
⎡
Eo [N P3 SL ] Eo ⎣
y∈ΦU,Bo
⎤
I(E 3P SL (y))⎦ ,
(8.66)
where E 3P SL (y) = {{SINR3M,P L ≥ 2 R P L − 1 ∪ SINR3S,P L ≥ 2 R P L − 1} ∩ {SINR3S,SL
≥ 2 R S L − 1}}. Similarly, Eo [N P3 SL ] can be finally expressed as
3
−1
Eo [N P3 SL ] = λU Pc,P
SL (T P L , TSL , α p )θ S (2π λ S ) .
8.5.2.3
(8.67)
Sum Rate
The sum rate for cooperative NOMA multicast is defined as the mean of the sum rate
for all users in coverage of the multicast cluster and is equal to
3
= R P L (Eo [N P3 L ] − Eo [N P3 SL ]) + (R P L + R SL )Eo [N P3 SL ]
R̄sum
= R P L Eo [N P3 L ] + R SL Eo [N P3 SL ].
(8.68)
Combining (8.65), (8.67), and (8.68), the sum rate can be finally expressed as
3
3
3
−1
R̄sum
= (R P L Pc,P
L (T P L ) + R SL Pc,P SL (T P L , TSL , α p ))λU θ S (2π λ S ) .
(8.69)
8.5.3 Numerical Results
Figure 8.10 plots the coverage probabilities of cooperative multicast transmissions
for mmWave-NOMA HetNets with fixed power ratio (0.8, 0.2). The results show that
compared with multicast and NOMA multicast, cooperative NOMA multicast can
further improve the coverage probabilities of the primary and secondary layers. More
specifically, compared with multicast, cooperative NOMA multicast can achieve
superior coverage in the low SINR threshold region, while compared with NOMA
multicast, it can achieve better primary coverage, especially in the low and high SINR
threshold regions, while it provides better secondary coverage, especially in the high
SINR threshold region. This is because cooperative NOMA multicast enables the
MBS tier to transmit a copy of the primary layer without power split as well. As a
result, users can receive two copies of the primary layer, which increases the success
8 NOMA for Millimeter Wave Networks
281
probability of decoding the primary layer. This also increases the success probability
of decoding the secondary layer, as users can try to further decode the secondary
layer when they fail to decode the primary layer from a mmWave base station but
succeed to decode the primary layer from a MBS.
Figure 8.11 plots the sum rates for cooperative multicast transmission for
mmWave-NOMA HetNets with fixed power ratios (0.8, 0.2) and (0.95, 0.05). The
results show that cooperative NOMA multicast can achieve higher sum rate than
NOMA multicast, especially in the medium multicast rate region. This is because in
cooperative NOMA multicast scheme, MBS transmits a replica of the primary layer
αP
, of decoding the primary
as well, which overcomes the maximum SINR limit, 1−α
P
layer caused by NOMA transmission. As a result, this increases the success probability of decoding the primary and secondary layers. Comparing Fig. 8.11a, b, more
power is allocated to the primary layer, a higher multicast rate can be provided, yet a
lower sum multicast rate is achieved. This means that cooperative NOMA multicast
gradually degrades to conventional multicast, with the increase of power allocated
to the primary layer.
8.6 Summary
This chapter applied NOMA to mmWave networks and discussed unicast, multicast
and cooperative multicast transmissions for mmWave-NOMA networks. An analytical framework for performance analysis of large-scale mmWave-NOMA networks
by using stochastic geometry was also given. Based on this framework, analytical
1
SINR Coverage Probability
0.9
0.8
0.7
0.6
0.5
0.4
Analytical results
Simulations
Multicast
NOMA multicast (PL)
NOMA multicast (SL)
Cooperative NOMA multicast (PL)
Cooperative NOMA multicast (SL)
0.3
0.2
0.1
0
−10
−8
−6
−4
−2
0
2
4
6
8
10
SINR Threshold (dB)
Fig. 8.10 SINR coverage probabilities of cooperative multicast transmission for mmWave-NOMA
HetNets with fixed power ratio (0.8, 0.2)
282
Z. Zhang and Z. Ma
(a)
100
90
80
Sum Rate (b/s/Hz)
70
60
50
40
Analytical results
Simulations
Multicast
NOMA multicast
Cooperative NOMA multicast
30
20
10
0
0
1
2
3
4
5
6
7
8
9
10
Multicast Rate (b/s/Hz)
(b)
100
90
Sum Rate (b/s/Hz)
80
70
60
50
Analytical results
40
Simulation results
Multicast
30
NOMA multicast
20
Cooperative NOMA multicast
10
0
0
1
2
3
4
5
6
7
8
9
10
Multicast Rate (b/s/Hz)
Fig. 8.11 Sum rates for cooperative multicast transmission for mmWave-NOMA HetNets with
fixed power ratios: a (0.8, 0.2); b (0.95, 0.05)
expressions for SINR coverage probability, outage probability, and sum rate were
provided to evaluate the performance of the presented schemes. It can be concluded
that: (1) NOMA can achieve better performance than OMA, by multiplexing multiple users in the power domain; (2) compared with multicast, NOMA multicast
multiplexes multiple data streams with different multicast rates in the power domain
such that users can decode data streams according to their channel conditions, which
significantly increases the sum rate; (3) the cooperative NOMA multicast increases
the success probability of decoding data streams from the superposed signal with the
help of cooperation of MBSs such that it further improves the performance of NOMA
8 NOMA for Millimeter Wave Networks
283
multicast. Some further research directions on NOMA for mmWave networks are
pointed out as follows:
• Power Allocation for mmWave-NOMA Networks: the power ratio is a key factor
that NOMA achieves better performance than the orthogonal one. The fixed power
ratio is a simple way for NOMA, but it cannot always achieve optimal performance.
This is because it does not utilize channel state information (CSI) in real time. The
optimization of power ratio to further improve the system performance according
to the instantaneous CSI is a great challenge, especially for multicast transmissions.
• Cooperation for mmWave-NOMA Networks: HetNet is an important deployment
form for mmWave communications. This chapter only discusses cooperative multicast for mmWave-NOMA networks that MBSs cooperate transmission for the primary layer. It is important to exploit other cooperation schemes to further improve
the performance of unicast and multicast transmissions for mmWave-NOMA networks.
References
1. T.S. Rappaport et al., Millimeter wave mobile communications for 5G cellular: it will work!
IEEE Access 1, 335–349 (2013)
2. W. Roh et al., Millimeter-wave beamforming as an enabling technology for 5G cellular communications: theoretical feasibility and prototype results. IEEE Commun. Mag. 52, 106–113
(2014)
3. M. Xiao et al., Millimeter wave communications for future mobile networks. IEEE J. Sel. Areas
Commun. 35, 1909–1935 (2017)
4. L. Kong, M.K. Khan, F. Wu, G. Chen, P. Zeng, Millimeter-wave wireless communications for
IoT-cloud supported autonomous vehicles: overview, design, and challenges. IEEE Commun.
Mag. 55, 62–68 (2017)
5. Z. Ding, Z. Yang, Z. Fan, H.V. Poor, On the performance of non-orthogonal multiple access in
5G systems with randomly deployed users. IEEE Signal Process. Lett. 21, 1501–1505 (2014)
6. Z. Ding, X. Lei, G.K. Karagiannidis, R. Schober, J. Yuan, V.K. Bhargava, A survey on nonorthogonal multiple access for 5G networks: research challenges and future trends. IEEE J.
Sel. Areas Commun. 35, 2181–2195 (2017)
7. Z. Zhang, H. Sun, R.Q. Hu, Downlink and uplink non-orthogonal multiple access in a dense
wireless network. IEEE J. Sel. Areas Commun. 35, 2771–2784 (2017)
8. G. Araniti, M. Condoluci, P. Scopelliti, A. Molinaro, A. Iera, Multicasting over emerging 5G
networks: challenges and perspectives. IEEE Netw. 31, 80–89 (2017)
9. Z. Ding, P. Fan, H.V. Poor, Random beamforming in millimeter-wave NOMA networks. IEEE
Access 5, 7667–7681 (2017)
10. D. Zhang, Z. Zhou, C. Xu, Y. Zhang, J. Rodriguez, T. Sato, Capacity analysis of NOMA with
mmWave massive MIMO systems. IEEE J. Sel. Areas Commun. 35, 1606–1618 (2017)
11. B. Wang, L. Dai, Z. Wang, N. Ge, S. Zhou, Spectrum and energy-efficient beamspace MIMONOMA for millimeter-wave communications using lens antenna array. IEEE J. Sel. Areas
Commun. 35, 2370–2382 (2017)
12. Z. Ding, L. Dai, R. Schober, H.V. Poor, NOMA meets finite resolution analog beamforming in
massive MIMO and millimeter-wave networks. IEEE Commun. Lett. 21, 1879–1882 (2017)
13. A.J. Morgado, K.M.S. Huq, J. Rodriguez, C. Politis, H. Gacanin, Hybrid resource allocation
for millimeter-wave NOMA. IEEE Wirel. Commun. 24, 23–29 (2017)
284
Z. Zhang and Z. Ma
14. Z. Zhang, Z. Ma, Y. Xiao, M. Xiao, G.K. Karagiannidis, P. Fan, Non-orthogonal multiple access
for cooperative multicast millimeter wave wireless networks. IEEE J. Sel. Areas Commun. 35,
1794–1808 (2017)
15. S. Naribole, E. Knightly, Scalable multicast in highly-directional 60-GHz WLANs. IEEE Trans.
Netw. 25, 2844–2857 (2017)
16. T. Bai, R.W. Heath, Coverage and rate analysis for millimeter-wave cellular networks. IEEE
Trans. Wirel. Commun. 14, 1100–1114 (2015)
17. J.G. Andrews, T. Bai, M. Kulkarni, A. Alkhateeb, A. Gupta, R.W. Heath, Modeling and analyzing millimeter wave cellular systems. IEEE Trans. Commun. 65, 403–430 (2017)
18. T.S. Rappaport, G.R. MacCartney, M.K. Samimi, S. Sun, Wideband millimeter-wave propagation measurements and channel models for future wireless communication system design.
IEEE Trans. Commun. 63, 3029–3056 (2015)
19. H.S. Ghadikolaei, C. Fischione, G. Fodor, P. Popovski, M. Zorzi, Millimeter wave cellular
networks: a MAC layer perspective. IEEE Trans. Commun. 63, 3437–3458 (2015)
20. J.G. Andrews, F. Baccelli, R.K. Ganti, A tractable approach to coverage and rate in cellular
networks. IEEE Trans. Commun. 59, 3122–3134 (2011)
21. Andrews, J.G., Gupta, A.K., Dhillon, H.S.: A Primer on Cellular Network Analysis Using
Stochastic Geometry, http://arxiv.org/abs/1604.03183
22. H.A. David, H.N. Nagaraja, Order Statistics, 3rd edn. (Wiley, Hoboken, New Jersey, 2003)
Chapter 9
Full-Duplex Non-Orthogonal Multiple
Access Networks
Mohammed S. Elbamby, Mehdi Bennis, Walid Saad, Mérouane Debbah
and Matti Latva-aho
9.1 Introduction
In conventional wireless networks, resources are assumed to be allocated exclusively
to users by considering half-duplex (HD) transmissions in serving uplink (UL) and
downlink (DL) requests. Operating simultaneously in UL and DL over the same
frequency band, known as in-band full-duplex (IBFD) or more commonly as fullduplex (FD), has been avoided for a long time, due to the inability to suppress
the self-interference in the full-duplex radio to a feasible operating point. Furthermore, orthogonal multiple access techniques, such as orthogonal frequency-division
multiple access (OFDMA), is used in multi-carrier settings to allocate subcarriers
to users in an exclusive manner. This restriction is imposed to avoid the multi-user
interference (MUI) resulting from scheduling multiple users/messages over the same
subcarrier. The unprecedented growth in data rate requirement and the number of connected devices mandates going beyond the traditional ways of handling the scarcity
of bandwidth in future wireless networks. A fundamental shift in the in the way
wireless resources are allocated and managed is thus necessary.
M. S. Elbamby (B) · M. Bennis · M. Latva-aho
Centre for Wireless Communications, University of Oulu, Oulu, Finland
e-mail: mohammed.elbamby@oulu.fi
M. Bennis
e-mail: mehdi.bennis@oulu.fi
M. Latva-aho
e-mail: matti.latva-aho@oulu.fi
W. Saad
Wireless@VT, Bradley Department of Electrical and Computer Engineering,
Virginia Tech, Blacksburg, VA, USA
e-mail: walids@vt.edu
M. Debbah
Mathematical and Algorithmic Sciences Lab, Huawei France, 92100 Paris, France
e-mail: merouane.debbah@huawei.com
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_9
285
286
M. S. Elbamby et al.
Full-duplex (FD) is ideally a spectrum efficiency doubler. By relaxing the constraint of orthogonal UL and DL transmissions, transceivers in the user and base
station nodes can exploit the non-orthogonality to boost the spectral efficiency [21].
However, HD was adopted as the default setting in wireless networks for multiple reasons. First, FD radios can experience self-interference leaked from the UL
transmitter to the DL receiver when operating in the same time–frequency resource.
Recent advances in self-interference cancellation techniques have challenged this
assumption. By using a combination of analogue and digital cancellation techniques,
self-interference can be reduced to a level close to the receiver noise floor. Second,
operating in FD results in increased inter-user interference due to the larger number of
transmissions per resource. In particular, DL base station to UL base station and UL
user to DL user interference will occur when operating simultaneously in UL and DL
in the same frequency resource. Furthermore, this interference can occur at the intracell level, unlike the orthogonal resources per cell associated with OFDMA-based
HD systems.
Another way of boosting the cellular network bandwidth utilization is the use of
non-orthogonal multiple access (NOMA) technique to schedule users with potentially overlapping resources [11]. NOMA has been recently discussed as a promising
way to boost the network spectral efficiency as compared to orthogonal multiple
access (OMA) techniques. NOMA provides diversity in the power domain by transmitting different messages to/from different users using the same time–frequency
resource. If the different signals are received with enough power disparity, the signals can be decoded for example using successive interference cancellation (SIC)
techniques.
Hence, incorporating both FD and NOMA into current wireless networks will
pose different challenges in terms of interference and network management. Different solutions are needed to address these challenges, ranging from smart resource
scheduling, power allocation, duplex and multiple access mode switching. This
chapter sheds light on both FD and NOMA network operation and discusses potential
future research directions for this topic. First, preliminaries on FD, self-interference,
and the interplay between FD and NOMA are discussed in Sect. 9.2. The objectives
and tools used for FD and NOMA networks are discussed and surveyed in Sect. 9.3.
Numerical results on the performance of FD-NOMA networks are presented in Sect.
9.4. Finally, Sect. 9.5 concludes the chapter and discusses some open problems.
9.2 Full-Duplex NOMA Networks
9.2.1 Preliminaries
Here, we provide the basics of FD and NOMA that can be helpful to the reader. We
particularly define the basic concepts of FD and self-interference cancellation and
then briefly introduce FD and NOMA network operation challenges.
9 Full-Duplex Non-Orthogonal Multiple Access Networks
9.2.1.1
287
Full-Duplex
Conventional wireless networks operate in HD mode, meaning that one direction of
transmission is allowed at any given time and frequency resource. Different duplexing
techniques have been considered in HD networks to duplex the UL and DL transmissions. In particular, time division duplex (TDD) and frequency division duplex (FDD)
are both used commonly in today’s networks. FDD dedicates frequency resources to
UL and DL where communication in both directions is orthogonal in the frequency
domain. TDD splits the time resources between UL and DL transmissions either in
a static manner (fixed UL and DL duty cycles on each time frame) or a dynamic
manner (where UL and DL duty cycles can change to match the network UL and DL
load conditions [14].
Alternatively, bidirectional transmissions can be used simultaneously, namely
full-duplex communication. FD can ideally double the spectrum efficiency by relaxing the constraint on UL and DL orthogonality [21]. This has been an ideal assumption for a long time. However, taking it into practice was hindered by the complexity of removing the self-interference leaked from the transmitter to the receiver of
the same device. The transmitted signal is much stronger than the received one, as
the latter is significantly weakened by the path and propagation losses. Hence, the
transmitted signal saturates the receiver radio chains and prevents signal reception.
Self-interference has been seen as the major impairment to FD operation. Ideally,
self-interference should be cancelled to the same level as the receiver noise floor
such that the received signal is decoded in the same level as in HD. Otherwise, the
residual interference is added to the received signal and hence decreases the receive
SNR and throughput.
9.2.1.2
Self-Interference Cancellation
Cancelling the self-interference from the transmitted signal of a FD radio is not
as easy as it might sound. Although the transmitter knows what it is transmitting,
this knowledge is of the signal in the baseband level [12, 20]. The baseband signal
experiences several linear and nonlinear in the analogue radio chain before it is
converted into a transmitted signal. Hence, subtracting the baseband signal does not
help in removing the self-interference.
Recently, the self-interference cancellation capability has significantly evolved.
The study in [20] has shown simultaneous transmission and reception (with a single
antenna) is achievable using analogue and digital cancellation techniques to cancel
the self-interference. The leaked self-interference is reduced to the noise floor such
that the received signal is not degraded. This breakthrough in the self-interference
cancellation capability motivated the consideration of full-duplex in future networks
for user scheduling [17] as well as relaying [38]. Furthermore, the shift from traditional macrocells to low-power small-cell networks eases the process of integrating
FD into future networks. As compared to 46 dBm transmit power of macrocells, femtocells operate in power levels as low as 20 dBm, which makes the self-interference
process feasible.
288
M. S. Elbamby et al.
Fig. 9.1 An illustration of the different types of interference in multi-cell FD operation (UDI:
UL-to-DL interference, DUI: DL-to-UL interference and SI: self-interference.)
9.2.1.3
FD Network Operation
Once the self-interference is cancelled, the link throughput of FD can potentially
double that of an HD link. However, as FD operates in a network level, another issue
arises which is the inter-link interference due to having simultaneous transmission
and reception, as illustrated in Fig. 9.1. This interference situation is similar to that
of dynamic TDD networks, where base stations operate with different TDD configurations to satisfy their individual UL and DL traffic requirements, resulting in
additional inter-link inter-cell interference. In addition to that, FD networks will suffer intra-cell interference between the user pairs operating in UL and DL. Therefore,
intelligent user pairing methods are needed to select the appropriate pairs of users to
be scheduled in each time–frequency resource such that this interference is avoided.
9.2.1.4
FD-NOMA Network Operation
FD-NOMA refers to the concept of operating in IBFD mode and using NOMA to
serve multiple requests. It is shown in the literature that NOMA can operate in both
UL and DL. To reap the benefits promised by NOMA over OMA, power disparity
has to be guaranteed. In UL, decoding the signals of multiple users sharing the time–
frequency resource will occur in the base station level. It is intuitive therefore to
select users with different receive power levels to be scheduled simultaneously. For
example, scheduling a pair of cell-centre and cell-edge users is a convenient approach.
In DL, different power levels should be allocated to the messages of different users,
according to the decoding order they are going to select. This is essential to ensure
a successful SIC process to cancel the intra-cell interference.
When FD is combined with NOMA, additional levels of interference are introduced generated from the opposite direction [3]. Different approaches can be considered to mitigate the effect of this unwanted interference. One approach [13] is
9 Full-Duplex Non-Orthogonal Multiple Access Networks
289
to select between operating in FD while using an OMA technique or operating in
HD and using NOMA. The scheduling and power allocation schemes play a key
role in finding the optimal operating methods given the network conditions and optimization objectives. Prioritizing the serving of users in certain link direction means
that NOMA should be selected to operate. On the other hand, FD can be selected to
simultaneously serve UL and DL requests. While this approach restricts operation to
FD-OMA or HD-NOMA at a certain time instant, it avoids the additional intra-cell
interference of FD on NOMA operation and hence allows for more users to be served
by NOMA in a certain link direction.
Another way to blend the use of FD and NOMA together is to allow the network
to operate in FD and NOMA in a given time instant. In this regard, the inter-link
interference can be treated as noise [35] and will affect the NOMA performance. In
particular, DL users will experience high UL-to-DL interference from multiple users
operating in UL-NOMA. This interference can be handled by limiting the maximum
number of users allowed to operate in UL-NOMA if the performance of DL users is
degraded.
9.2.2 Challenges of FD-NOMA Resource Optimization
As discussed above, resource optimization is a key role in reaping the benefits of
FD-NOMA operation. Since the challenges of implementing them in practice are
similar, there are common tools that can be used to optimize the resource and power
allocation when either or both of them are considered. In this chapter, we focus on
two main optimization problems that are fundamental for enabling NOMA, which
are user pairing/scheduling and power allocation. Furthermore, we discuss the impact
of the different objective functions on the resource optimization.
9.2.3 User Pairing and Power Optimization
To illustrate the problem of user scheduling FD-NOMA networks, we assume a general multi-cell wireless network that comprises a set B of B FD-capable base stations,
with self-interference cancellation capability of ζ , and a set U of U users requesting
either UL or DL service. Furthermore, we assume that both base stations and users
can operate in NOMA scheme, where the resulting multi-user interference can be
cancelled by performing SIC at the receiver side. The user scheduling parameters in
UL
(t) ∈ X UL ,
UL and DL at time instant t are defined using the binary parameters xbu
DL
UL
DL
xbu (t) ∈ X ∀b ∈ B, u ∈ U . Let f b be the allocated bandwidth, vb (t) and vDL
b (t)
be the vectors of the UL and DL successive ordering index of all the users in base
290
M. S. Elbamby et al.
(.)
station b, where vbu
(t) ∈ v(.)
b (t) is the decoding order of user u when it is scheduled
DL
UL
be the UL and DL data rates between base station
and Rbu
by base station b, and Rbu
b and user u, then for a general network with open-access policy, the UL and DL
service rates of user u can be expressed using the Shannon formula as:
ruUL (t) =
UL
UL
xbu
(t)Rbu
(t),
b∈B
=
UL
UL
xbu
(t) f b log2 (1 + Γbu
(t)),
(9.1)
b∈B
ruDL (t) =
DL
xbu
(t)RuDL (t),
b∈B
=
DL
DL
xbu
(t) f b log2 (1 + Γbu
(t)).
(9.2)
b∈B
Assuming that the information to be transmitted is encoded based on a Gaussian
distribution and a zero-mean additive white Gaussian noise with variance N0 , the
UL and DL signal-to-interference-plus-noise ratios (SINRs) between base station b
and user u at time instant t are given by:
UL
(t) =
Γbu
DL
Γbu
(t) =
puUL (t)h bu (t)
,
NOMA-UL
(t) + pbDL (t)/ζ
N0 + IbUL-UL (t) + IbDL-UL (t) + Ibu
N0 +
IuDL-DL (t)
DL
(t)h bu (t)
pbu
,
NOMA-DL
(t)
+ IuUL-DL (t) + Ibu
(9.3)
(9.4)
DL
where puUL is the UL transmit power of user u, and pbDL = u∈U pbu
is the total DL
2
transmit power of base station b, h x,y (t) = |gx,y (t)| is the channel gain between the
two nodes x and y, gx,y (t) is the propagation channel between nodes x and y, and
the interference terms in (9.3) and (9.4) are expressed as follows1 :
IbUL-UL (t) =
puUL
′ (t)h bu ′ (t),
u ′ ∈U \{u}
IbDL-UL (t) =
pbDL
′ (t)h b′ b (t),
b′ ∈B \{b}
IuDL-DL (t) =
pbDL
′ (t)h b′ u (t),
b′ ∈B \{b}
IuUL-DL (t)
=
puUL
′ (t)h u ′ ,u (t),
u ′ ∈U
1 Note
that the term IUL-DL includes the intra-cell interference, to account for the interference due
to FD operation.
9 Full-Duplex Non-Orthogonal Multiple Access Networks
NOMA-UL
(t) =
Ibu
291
puUL
′ (t)h bu ′ (t),
UL
u ′ ∈U \{u}|xbu
′ =1,
UL
UL
vbu
(t)<vbu
′ (t)
NOMA-DL
Ibu
(t) =
u
′
DL
pbu
′ (t)h bu (t),
DL
∈U \{u}|xbu
′ =1,
DL
DL
vbu
(t)<vbu
′ (t)
pbDL (t)/ζ is the leaked self-interference.
Remark 1 To guarantee a successful NOMA operation in the DL, a user u ′ should
u ′ DL
decode the data of user u with an SINR level Γbu
(t) that is at least equal to the user
DL
u received SINR Γbu
(t). Otherwise, the data rate of user u is higher than what user
u ′ DL
DL
(t) ≥ Γbu
(t) must hold, where:
u ′ can decode. Accordingly, the inequality Γbu
′
u DL
Γbu
(t) =
DL
pbu
′ (t)h bu ′ (t)
DL
(t)h bu ′ (t)
pbu
.
NOMA-DL
DL-DL
(t)
(t) + Ibu
(t) + IuUL-DL
+ N 0 + Iu ′
′
′
This condition is met if the following metric is greater or equal to zero:
′
DL
u DL
DL DL
).
− Γbu
Yuu ′ = 1vbuDL (t)<vDL′ (t) xbu
xbu ′ (Γbu
bu
Note that the above condition is satisfied by default in the UL, since all users’ signals
are decoded in the base station’s receiver.
A general FD-NOMA resource optimization problem can be defined to be optimizing user scheduling, NOMA decoding ordering, and UL and DL power allocations,
UL
DL
] and X DL = [xbu
], the decoding orderi.e., the scheduling matrices X UL = [xbu
DL
UL
ing vectors vb and vb , and the power allocation vectors, pUL = [ p1UL , ..., pUUL ]
DL
DL
and pDL
b = [ pb1 , ..., pbU ]. Note that the selection of the scheduling parameters
essentially includes the problems of user association, UL/DL mode selection, and
OMA/NOMA mode selection. Accordingly, a general optimization problem can be
cast as follows:
P1:
max
UL
DL
X ,X ,
DL
{vUL
b },{vb }
pUL ,{ pDL
b }
U {ruUL }, {ruDL }
(9.5a)
UL
subject to puUL ≤ Pmax
, ∀u ∈ U ,
puDL
DL
Pmax
,
≤
UL
xbu
+
∀b ∈ B,
DL
xbu
≤ 1, ∀u ∈ U ,
(9.5b)
(9.5c)
(9.5d)
b∈B
u∈U
UL
xbu
≤ q UL ,
DL
xbu
≤ q DL ∀b ∈ B.
(9.5e)
u∈U
Yuu ′ ≥ 0 ∀u, u ′ ∈ U .
(9.5f)
292
M. S. Elbamby et al.
Constraints (9.5b) (9.5c) limit the UL and DL transmit powers to their maximum
UL
DL
values Pmax
and Pmax
, respectively. Constraint (9.5d) restricts the connection of a
user to one base station in either UL or DL. The number of users a base station can
simultaneously serve in UL or DL-NOMA is limited to a quota of q UL and q DL users,
respectively, by constraint (9.5e).2 Constraint (9.5f) guarantees a successful SIC in
the DL.
The utility in (9.5) should be selected to reflect the objective of the network
optimization problem. The work in [35] considers a utility that maximizes the sum
rate in a multi-carrier setting. In this regard, the resource allocation includes the
subcarrier allocation. In [13], a weighted rate maximization utility was developed
from a stochastic optimization problem, where the user weights are derived from the
user traffic queue and virtual queue backlogs.3
9.2.4 Optimization Tools
Rate-based utility maximizations are often combinatorial and non-convex optimization problems that are computationally complex to solve optimally. In particular, the
user scheduling problem is a combinatorial problem whose complexity will increase
exponentially as the number of base stations and users increase.
In many cases, the optimization problem has to be solved dynamically. For example, if the weighted rate maximization based on user’s queue state is considered as
the objective, both the user pairing and power allocation will need to be dynamically adjusted. Efficient and local solutions tools are therefore needed. Matching is
a powerful tool that can solve the user scheduling problem [26]. Matching theory
is a framework that provides solutions for the combinatorial problem of matching
members of two disjoint sets of players in which each player is interested to associate with one or more player the other set. Matching is performed on the basis of
preference profiles defined by the players of each side, providing a low-complexity
stable matching. Although matching does not necessarily guarantee finding the optimal solution, its suitability for dynamic systems and local implementations made it
a popular solution to reduce the complexity of the combinatorial problems [18, 19,
31]. Matching can also be used with other game-theoretic tools, such as cooperative
game theory [28, 29], to further solve user grouping problems. Matching should be
performed assuming an initial feasible power allocation policy. Subsequently, the
power allocation is performed for the selected matching setting. The decoupling of
the user scheduling and power allocation problems simplifies both problems, since
power allocation is performed to the reduced set of scheduled users [13].
2 Note
that, in theory, the number of users that can be served simultaneously using NOMA is
unrestricted. However, we impose a quota to avoid high-complexity SIC in the receiver side if a
high number of users are scheduled.
3 Virtual queues result from applying the Lyapunov framework to convert the time-averaged constraints into virtual queues such that the constraints are met as the virtual queues are stabilized.
9 Full-Duplex Non-Orthogonal Multiple Access Networks
293
In addition to the complexity of the user scheduling problem, the power allocation
problem is non-convex, due to having interference terms in the denominator of the
SINR in the rate expression. Several approaches have been proposed in the literature
to deal with the non-convexity of the problem. In [35], the global optimal solution
of the joint problem of resource and power allocation is solved using monotonic
optimization. The solutions are, however, with high computational complexity. The
authors provide a lower complexity solution to solve the problem using successive
convex approximation (SCA), which is shown to achieve a close to global optimal
solution. SCA and similar tools to convexify the non-convex terms in the optimization
problem are used in several works [13, 15, 24, 35, 36]. In this case, the convexified
problem is solved iteratively using convex optimization tools until some convergence
criterion is met.
The NOMA decoding order significantly affects the user performance in both
UL and DL. In DL-NOMA, decoding users with the lower channel strength first is
optimal in the single-cell scenario [11]. In UL-NOMA, the opposite decoding order
based on the user channel strength is shown in [6] to result in a gain over OMA
based on users channel gain disparity. The decoding ordering in a multi-cell scenario
is a challenging task, especially in the DL where different users might experience
different inter-cell interference levels.
9.3 State of the Art in FD and NOMA Resource
Optimization
This section surveys the recent works in the problem of wireless resource optimization in FD and NOMA networks. First, with the promises of doubling the link
throughput, the study of the when and how the potential gains can be achieved in
a network level is surveyed. Following that, we overview the studies of resource
optimization in NOMA networks in UL and DL. Finally, we highlight the works that
combine both FD and NOMA schemes.
9.3.1 FD Resource Optimization
In [16], a hybrid HD/FD scheduler for a single-cell network is proposed. The scheduler assigns FD resources only when it is advantageous over HD resource assignment.
The joint problem of subcarrier and power allocation are optimized in [36] using an
SCA algorithm. An auction-based algorithm to pair UL and DL users in a single-cell
IBFD network is proposed in [33]. Subcarrier and power allocation is optimized using
a heuristic algorithm with polynomial complexity for a single-cell IBFD network in
[37]. The study also considers the case of imperfect self-interference cancellation
on the performance of FD as compared to HD. It concludes that a higher number of
294
M. S. Elbamby et al.
users can be served using FD as the cancellation capability increases. In [17], the
authors extended the work in [16] to multi-cell networks. To reduce the complexity
of a centralized solution, a distributed resource allocation scheme is developed where
each base station selects locally which user to serve, and then, it coordinates with
the neighbouring base stations to coordinate their transmission powers such that the
inter-cell interference is minimized. The study shows that FD can achieve up to double the throughput of HD in indoor scenarios and 65% throughput gain in outdoor
scenarios. The work in [34] also considered the FD resource and power allocation in
multi-cell FD networks but with frequency reuse allowed only once among different
cells. Matching theory is leveraged in [7] to develop a resource allocation algorithm.
A matching algorithm is proposed to assign subcarriers to UL and DL user pairs.
In [5], the user scheduling in FD ultra-dense-networks is optimized using different
schemes with and without power optimization. The scheduling is carried out locally
assuming no knowledge of the inter-cell interference. Table 9.1-a summarizes the
contributions on the FD resource allocation.
Remark 2 A common assumption in FD scheduling is that the base station knows
the individual channel between its own users. This assumption is necessary to select
FD pairs that do not significantly interfere on each other [16]. The assumption should
be practical as FD is expected to be feasible in low-power small-cell networks where
a low number of users are served by each base station.
9.3.2 NOMA Resource Optimization
Recently, several works have looked into the resource optimization in UL and/or DLNOMA networks. Here, we shed light on the papers focusing on the scheduling and
power optimization in NOMA networks. The authors in [8] propose a many-to-many
algorithm to assign subcarriers to users in a single DL-NOMA network. Many-tomany matching is used in [9] for a multi-cell DL-NOMA scenario. Matching is also
used for DL-NOMA resource allocation in [15] with a focus on energy efficiency.
The power allocation problems are convexified and solved using difference of convex
functions (DC) programming. DC programming is also used in [24] to optimize the
power allocation in OFDM-based DL-NOMA networks, whereas a greedy algorithm
is proposed for the user selection problem. A greedy approach is also used for the user
scheduling in a multiple-input multiple-output (MIMO) DL-NOMA network in [32],
and a minimum mean squared error (MMSE)-based power allocation is considered.
Several works have looked into NOMA in the UL direction. The performance
of NOMA in the UL is investigated in [2] using an iterative channel allocation
algorithm. In [30], the problem of user pairing in UL-NOMA for users with single
and multi-antennas is optimized. User grouping and power optimization in ULNOMA is studied in [6] where the impact of user ordering and imperfect SIC is
9 Full-Duplex Non-Orthogonal Multiple Access Networks
295
Table 9.1 Summary of existing literature in FD- and NOMA-based resource allocation problems
(a) FD
References Network
Implementation FD scheduling
Power allocation
scenario
[16]
[36]
Single-cell
Single-cell
Local
Local
[33]
Single-cell
Local
[37]
[17]
Single-cell
Multi-cell
Local
Local
[34]
[7]
Multi-cell
Central
(subcarrier is
reused once)
Multi-cell
Central
[5]
Multi-cell
Local
Link
scenario
Network
scenario
Scheduling and power allocation scheme
DL
DL
DL
DL
DL
UL
UL
UL
UL + DL
Single-cell
Single-cell
Single-cell
Single-cell
Multi-cell
Single-cell
Single-cell
Heterogeneous
Multi-cell
Matching algorithm
Subchannel assignment and power allocation
User selection and power optimization
User pairing and power allocation
Matching algorithm and power allocation
User pairing for multi-antenna systems
Iterative subcarrier and power allocation
User clustering and power allocation
Optimal power allocation for a limited
number of users
(b) NOMA
References
[8]
[15]
[24]
[32]
[9]
[30]
[2]
[6]
[4]
(c) FD-NOMA
References Network
scenario
[35]
[13]
Single-cell
Multi-cell
HD/FD mode selection
Joint subchannel and power
allocation
FD user pairing and channel
allocation
OFDMA channel allocation
Suboptimal HD/FD user
selection
Mode selection and
subcarrier allocation
×
Matching subcarriers to
user pairs
Local scheduling
Scheduling and power allocation scheme
Joint subchannel and powerallocation
Joint user scheduling and power allocation
investigated. In [4], the authors consider a multi-cell UL and DL-NOMA system
where a user grouping and power optimization scheme are proposed. The optimal
power allocation is derived for a single macro-cell and a limited number of users.
296
M. S. Elbamby et al.
9.3.3 FD-NOMA Resource Optimization
Two recent studies have looked into the incorporation of FD into NOMA networks
and the impact on the scheduling and power allocation. In [35], the authors proposed
an FD-NOMA approach in which users can be scheduled simultaneously in UL and
DL in the same time–frequency resource and NOMA can be used in both directions.
SIC is used in UL and DL to decode the messages of different users, whereas the
inter-link interference due to FD is treated as noise. The joint subcarrier and power
optimization problem is formulated, and the global optimal solution is found using
monotonic optimization. Due to the high complexity of finding the global optimal
solution, a low-complexity solution based on SCA is found. The results have shown
that FD-NOMA improves the spectral efficiency as compared to HD-NOMA. Moreover, the effect of imperfect SIC is shown to impact the performance of the FD
scheme.
In [13], FD-NOMA is investigated in a dynamic multi-cell scenario where a
stochastic optimization problem based on the Lyapunov framework is considered.
The benefits of operating in HD or FD, as well as in OMA or NOMA modes, depending on traffic conditions, network density, and self-interference cancellation capabilities are investigated. The optimization problem is decomposed into two subproblems
that are solved independently per small-cell base station. User association and mode
selection are formulated as a many-to-one matching problem. A distributed matching
algorithm aided by an inter-cell interference learning mechanism is proposed which
is shown to converge to a pairwise stable matching. The matching algorithm allows
small-cells to select between HD and FD and to operate either in OMA or NOMA
schemes to serve their users. Subsequently, the UL/DL power optimization problem
is formulated as a sequence of convex problems, and an iterative algorithm to allocate
the optimal power levels for the matched users and their base stations is proposed.
It was shown that using matching theory, the network can dynamically select when
to operate in HD or FD and when to use OMA or NOMA to serve different users,
which yields significant gains in UL and DL user throughput and packet throughput,
as compared to HD-NOMA, FD-OMA, and HD-OMA schemes.
9.4 Numerical Results
In this section, we present some numerical results to assess the performance of the
queue-aware FD and NOMA resource optimization. We consider a continuous utility
function of time-averaged UL and DL service rates. The problem can be decomposed
using the Lyapunov framework into an instantaneous weighted rate maximization
in which the user weights are their queue backlogs. The network consists of smallcell base stations with a varying self-interference cancellation capability, serving
multiple users in an open-access manner. Scheduling can be in HD or FD modes,
and in HD mode, users can be scheduled in OMA or NOMA. To cancel the resulting
9 Full-Duplex Non-Orthogonal Multiple Access Networks
297
multi-user interference, base stations or users operating in NOMA can perform SIC at
the receiver side. The decoding ordering is assumed to be done in a descending order
of channel strength in DL-NOMA, and an ascending order of channel strength in
UL-NOMA. We assume that the user’s mean packet arrival rate and mean packet size
follow Poisson and exponential distributions, respectively. To satisfy queue stability
requirements, base stations need to ensure that user’s traffic queues are mean rate
stable. Equivalently, constraints are imposed to ensure that the average service rate
is higher or equal to the average arrival rate. The resource optimization problem
consists of the scheduling problem (which includes the mode selection) and the
power allocation problem. To reduce the complexity of the combinatorial scheduling
problem, a many-to-one matching algorithm based on the deferred acceptance (DA)
matching [13] is considered to dynamically schedule one or more users to each
base station at each time instant. Preference profiles for users and base stations are
selected as to maximize their individual weighted rates. The matching algorithm can
be performed locally at each base station, which significantly reduce the amount of
signalling exchange. After the matching is performed, each base station coordinates
with its neighbours to optimize the allocated power, such that a feasible policy is
achieved and inter-cell interference is minimized. The multi-cell power optimization
is non-convex. Hence, the DC programming is used to convexify the problem, which
is guaranteed to converge to a local optimal solution. System level simulations are
carried out to show the gains brought by FD and NOMA, as well as to investigate
the impact of queue stability constraints on the network performance. Simulation
parameters are presented in Table 9.2. For the sake of comparison, the following
schemes are considered in the simulation:
1. HD-OMA scheme: users are associated to the nearest base station and are allocated
orthogonal resources for UL and DL. Requests are served using a round robin
(RR) scheduler.
2. HD-NOMA only scheme: users are associated to the nearest base station, and
RR scheduler is used to serve UL and DL queues. If there are multiple users in a
scheduling queue, they are ranked according to their channel gains and are served
using NOMA if the ratio between their channel gains is at least 2; otherwise, OMA
is used. Power is allocated to NOMA users based on their channel ranking, in
a uniform descending order for UL-NOMA and a uniform ascending order for
DL-NOMA. Base stations operate either in UL or in DL depending on the queue
length on each link direction.
3. FD-OMA scheme: users are associated to the nearest base station, and a pair of
users is served in FD mode if the channel gain between them is greater than a
certain threshold; otherwise, users are served in HD mode using RR scheduler.
4. Uncoordinated scheme: in this scheme, users can be served in either HD or FD
and in OMA or NOMA modes. The many-to-one matching algorithm is used for
mode selection and user scheduling, with the user queues taken into account in
the weighted rate maximization. Power is assumed to be fixed for OMA and is
similar to that of scheme 2 for NOMA. No inter-cell interference coordination is
considered.
298
Table 9.2 Simulation parameters
Parameter
System bandwidth
Duplex modes
Multiplexing mode
Subframe duration
Network size
Number of base stations
Avg. number of users per base station
Small-cell radius
Max. base station transmit power
Max. user transmit power
Path loss model
Shadowing standard deviation
Penetration loss
Self-interference cancellation
Packet arrival rate per user
Max. quota of NOMA users q UL , q DL
M. S. Elbamby et al.
Value
10 MHz
TDD HD/ FD
OMA/NOMA
1 ms
500 × 500 m2
10
10
40 m
22 dBm
20 dBm
Multi-cell pico scenario [1]
4 dB
0 dB
110 dB
10 packet/s
5
5. Proposed scheme: In this scheme, the many-to-one matching algorithm is used
for mode selection and user scheduling as in scheme 4. In addition, inter-cell
interference coordination is considered by optimizing the UL and DL power
allocation using DC programming.
We begin by evaluating the performance of the proposed scheme under different
traffic intensity conditions. Traffic intensity is varied by changing the mean packet
size between 50 and 400 kb. In Fig. 9.2, we show the impact of traffic intensity on the
normalized UL and DL user throughput. The normalized user throughput is defined as
the user service rate divided by its data arrival rate. Figure 9.2a shows that in the UL,
all schemes but the proposed one achieve lower UL throughput as compared to the
HD-OMA scheme. The performance drop is due to the DL-to-UL interference that
has a significant impact on the uncoordinated schemes due to the higher transmitting
power of base stations and the lower path loss between the base station and user. The
proposed scheme outperforms the different schemes by mitigating the DL-to-UL
interference through power optimization.
In Fig. 9.2b, we can see that, in the DL case, the effect of UL-to-DL interference
is less significant, as it can be avoided within each cell through the pairing process.
By leveraging both matching and the UL and DL power optimization, the proposed
scheme outperforms the baseline schemes, with UL and DL gains of 10 and 20%
over the HD-OMA scheme. The figure also shows that the coordination gain (over
the uncoordinated scheme) is even more evident in the UL as in the DL due to the
dominance of the DL-to-UL interference in the uncoordinated scenario.
9 Full-Duplex Non-Orthogonal Multiple Access Networks
(a) UL
1
Normalized DL user throughput (Mbps)
Normalized UL user throughput (Mbps)
0.98
0.96
0.94
0.92
0.9
0.88
0.86
Proposed Scheme
HD-OMA
HD-NOMA
FD-OMA
Uncoordinated
0.84
0.82
0.8
(b) DL
1
0.98
299
1
2
0.96
0.94
0.92
0.9
0.88
0.86
Proposed Scheme
HD-OMA
HD-NOMA
FD-OMA
Uncoordinated
0.84
0.82
3
0.8
4
Mean traffic intensity (Mbps)
1
2
3
4
Mean traffic intensity (Mbps)
Fig. 9.2 Normalized a UL and b DL user throughput performance for different schemes as the
user traffic intensity increases, for a network of ten base station and an average of ten users per base
station
101
Average user Queue length (Mbit)
Fig. 9.3 Average user queue
length performance for
different schemes as the user
traffic intensity increases, for
a user of ten base station and
an average of ten users per
base station
10
0
10
-1
10
-2
10
-3
10
-4
0.5
Proposed Scheme
HD-OMA
HD-NOMA
FD-OMA
Uncoordinated
1
1.5
2
2.5
3
3.5
4
Mean user traffic intensity (Mbps)
Next, we investigate the queue behaviour of the different schemes as the traffic
intensity varies. In Fig. 9.3, the average queue length is shown as function of the
traffic intensity. Figure 9.3 shows that the average queue length grows as the traffic
intensifies. In low traffic conditions, FD-OMA and HD-NOMA have smaller queue
lengths as compared to the HD-OMA scheme. Also, the uncoordinated FD-NOMA
scheme maintains a smaller queue length, since coordination is not crucial in low
traffic intensity conditions. As the traffic intensity increases, we can see that the queue
length grows significantly for the uncoordinated schemes. The proposed scheme
300
5.5
Proposed Scheme
HD-OMA
HD-NOMA
FD-OMA
Uncoordinated
5
Average cell throughput (bit/s/Hz)
Fig. 9.4 Average cell
throughput (in bit/s/Hz)
performance for different
schemes as the base station
self-interference cancellation
capability varies, for a
network of ten base stations,
an average of ten users per
base station, and a mean
traffic intensity rate of 3
Mbps per user
M. S. Elbamby et al.
4.5
4
3.5
3
2.5
2
1.5
30
40
50
60
70
80
90
100
110
Self-interference cancellation (dB)
maintains small queue lengths as it seeks to stabilize the user queues through the
weighted rate maximization.
Finally, we show the impact of the self-interference cancellation capability on the
performance of the FD schemes. Figure 9.4 compares the average cell throughput
(in bit/s/Hz) of the different schemes as the self-interference cancellation capability
varies from 30 to 110 dB, which is the highest reported value [20]. As shown in
the figure, the throughput of the FD schemes degrades with lower self-interference
cancellation levels due to the interference leakage on the UL signal. It is also shown
that only a slight degradation in the throughput of the proposed FD-NOMA is
observed. As the proposed scheme optimizes the mode selection between HD/FD
and OMA/NOMA, it can select more frequently the UL-NOMA instead of FD to
serve UL users, such that it avoids high interference from the base station’s DL. This
shows that enabling both FD and NOMA has the potential to enable higher network
spectral efficiency in different network conditions.
9.5 Conclusions and Open Problems
This chapter has provided an overview of the topic of full-duplex (FD) nonorthogonal multiple access (NOMA) from a network optimization point of view.
Different challenges on the optimization of FD-NOMA networks are discussed.
It has highlighted the particular importance of the user pairing and scheduling in
both FD and NOMA networks. Several directions are still open for research. As
was mentioned throughout the chapter, the decoding ordering is a key factor in the
performance of the NOMA systems. Some studies are carried out on the optimal
ordering of users, most of which are assuming single-cell operation and focus on
the optimal solution from the point of sum rate maximization. Finding the optimal
9 Full-Duplex Non-Orthogonal Multiple Access Networks
301
decoding ordering is a challenging task in a multi-cell scenario and is even challenging in FD networks in which both intra-cell and inter-cell interference impact the
network performance. The objective of the decoding order optimization should also
take into account the notion of fairness between the users with different channel and
queue state conditions. Moreover, enabling NOMA for emerging 5G systems, such as
vehicle-to-everything (V2X) networks [10, 25] and networks with unmanned aerial
vehicles [22, 23, 27], poses a wide range of open problems. Finally, looking into
different objectives beyond the average rate maximization problems is necessary,
especially in the context of ultra-reliable and low latency communication (URLLC),
which brings further challenges to the system design.
Acknowledgements This research was supported in part by the Academy of Finland CARMA
Project, in part by the U.S. National Science Foundation under Grant CNS-1513697, and Grant
CNS-1617896, and in part by the ERC Starting Grant MORE (Advanced Mathematical Tools for
Complex Network Engineering) under Grant 305123.
References
1. 3GPP, Further enhancements to LTE TDD for DL-UL interference management and traffic
adaptation. TR 36.828, 3rd Generation Partnership Project (3GPP) (2012), http://www.3gpp.
org/DynaReport/36828.htm
2. M. Al-Imari, P. Xiao, M.A. Imran, R. Tafazolli, Uplink non-orthogonal multiple access for
5G wireless networks, in 2014 11th International Symposium on Wireless Communications
Systems (ISWCS) (2014), pp. 781–785
3. K.S. Ali, H. Elsawy, A. Chaaban, M.S. Alouini, Non-orthogonal multiple access for large-scale
5G networks: interference aware design. IEEE Access 5, 21204–21216 (2017)
4. M.S. Ali, H. Tabassum, E. Hossain, Dynamic user clustering and power allocation for uplink
and downlink non-orthogonal multiple access (noma) systems. IEEE Access 4, 6325–6343
(2016)
5. I. Atzeni, M. Kountouris, G.C. Alexandropoulos, Performance evaluation of user scheduling
for full-duplex small cells in ultra-dense networks, in 22th European Wireless Conference on
European Wireless 2016 (2016), pp. 1–6
6. A. Celik, R.M. Radaydeh, F.S. Al-Qahtani, A.H.A. El-Malek, M.S. Alouini, Resource allocation and cluster formation for imperfect NOMA in DL/UL decoupled hetnets, in IEEE Global
Communications Conference (GLOBECOM) (2017), pp. 1–7
7. B. Di, S. Bayat, L. Song, Y. Li, Radio resource allocation for full-duplex OFDMA networks
using matching theory, in 2014 IEEE Conference on Computer Communications Workshops
(INFOCOM WKSHPS) (2014), pp. 197–198
8. Di, B., Bayat, S., Song, L., Li, Y.: Radio resource allocation for downlink non-orthogonal multiple access (NOMA) networks using matching theory. In: 2015 IEEE Global Communications
Conference (GLOBECOM), pp. 1–6 (2015)
9. B. Di, L. Song, Y. Li, Sub-channel assignment, power allocation, and user scheduling for
non-orthogonal multiple access networks. IEEE Trans. Wirel. Commun. 15(11), 7686–7698
(2016)
10. B. Di, L. Song, Y. Li, G.Y. Li, Noma-based low-latency and high-reliable broadcast communications for 5G v2x services, in 2017 IEEE Global Communications Conference on GLOBECOM,
pp. 1–6 (2017)
302
M. S. Elbamby et al.
11. Z. Ding, X. Lei, G.K. Karagiannidis, R. Schober, J. Yuan, V.K. Bhargava, A survey on nonorthogonal multiple access for 5G networks: research challenges and future trends. IEEE J.
Sel. Areas Commun. 35(10), 2181–2195 (2017)
12. M. Duarte, A. Sabharwal, Full-duplex wireless communications using off-the-shelf radios:
feasibility and first results, in 2010 Conference Record of the Forty Fourth Asilomar Conference
on Signals, Systems and Computers (2010), pp. 1558–1562
13. M.S. Elbamby, M. Bennis, W. Saad, M. Debbah, M. Latva-aho, Resource optimization and
power allocation in in-band full duplex-enabled non-orthogonal multiple access networks.
IEEE J. Sel. Areas Commun. 35(12), 2860–2873 (2017)
14. M.S. ElBamby, M. Bennis, W. Saad, M. Latva-aho, Dynamic uplink-downlink optimization in
TDD-based small cell networks, in 2014 11th International Symposium on Wireless Communications Systems (ISWCS) (2014), pp. 939–944
15. F. Fang, H. Zhang, J. Cheng, V.C.M. Leung, Energy-efficient resource allocation for downlink
non-orthogonal multiple access network. IEEE Trans. Commun. 64(9), 3722–3732 (2016)
16. Goyal, S., Liu, P., Panwar, S., DiFazio, R.A., Yang, R., Li, J., Bala, E.: Improving small cell
capacity with common-carrier full duplex radios, in 2014 IEEE International Conference on
Communications (ICC) (2014), pp. 4987–4993
17. S. Goyal, P. Liu, S.S. Panwar, User selection and power allocation in full-duplex multicell
networks. IEEE Trans. Veh. Technol. 66(3), 2408–2422 (2017)
18. Y. Gu, W. Saad, M. Bennis, M. Debbah, Z. Han, Matching theory for future wireless networks:
fundamentals and applications 53(5), 52–59 (2015)
19. Y. Gu, W. Saad, M. Bennis, M. Debbah, Z. Han, Matching theory for future wireless networks:
fundamentals and applications. IEEE Commun. Mag. 53(5), 52–59 (2015)
20. M. Jain, J.I. Choi, T. Kim, D. Bharadia, S. Seth, K. Srinivasan, P. Levis, S. Katti, P. Sinha,
Practical, real-time, full duplex wireless, in Proceedings of the 17th Annual International
Conference on Mobile Computing and Networking, MobiCom’11, New York, NY, USA (2011),
pp. 301–312, http://doi.acm.org/10.1145/2030613.2030647
21. R. Li, Y. Chen, G.Y. Li, G. Liu, Full-duplex cellular networks. IEEE Commun. Mag. 55(4),
184–191 (2017)
22. M. Mozaffari, W. Saad, M. Bennis, M. Debbah, Unmanned aerial vehicle with underlaid deviceto-device communications: performance and tradeoffs. IEEE Trans. Wirel. Commun. 15(6),
3949–3963 (2016)
23. M. Mozaffari, W. Saad, M. Bennis, M. Debbah, Wireless communication using unmanned
aerial vehicles (UAVS): optimal transport theory for hover time optimization. IEEE Trans.
Wirel. Commun. 16(12), 8052–8066 (2017)
24. P. Parida, S.S. Das, Power allocation in OFDM based NOMA systems: a DC programming
approach, in 2014 IEEE Globecom Workshops (GC Wkshps) (2014), pp. 1026–1031
25. L.P. Qian, Y. Wu, H. Zhou, X. Shen, Dynamic cell association for non-orthogonal multipleaccess v2s networks. IEEE J. Sel. Areas Commun. 35(10), 2342–2356 (2017)
26. A. Roth, M. Sotomayor, Two-Sided Matching: A Study in Game-theoretic Modeling and Analysis (Cambridge University Press, Cambridge, 1992)
27. N. Rupasinghe, Y. Yapici, I. Guvenc, Y. Kakishima, Non-Orthogonal Multiple Access for
mmWave Drones with Multi-Antenna Transmission (2017), http://arxiv.org/abs/1711.10050
28. W. Saad, Z. Han, M. Debbah, A. Hjorungnes, A distributed coalition formation framework
for fair user cooperation in wireless networks. IEEE Trans. Wirel. Commun. 8(9), 4580–4593
(2009)
29. W. Saad, Z. Han, M. Debbah, A. Hjorungnes, T. Basar, Coalitional game theory for communication networks. IEEE Signal Process. Mag. 26(5), 77–97 (2009)
30. M.A. Sedaghat, R.R. Müller, On user pairing in NOMA uplink (2017), http://arxiv.org/abs/
1707.01846
31. O. Semiari, W. Saad, M. Bennis, Joint millimeter wave and microwave resources allocation in
cellular networks with dual-mode base stations. IEEE Trans. Wirel. Commun. 16(7), 4802–
4816 (2017)
9 Full-Duplex Non-Orthogonal Multiple Access Networks
303
32. L. Shi, B. Li, H. Chen, Pairing and power allocation for downlink nonorthogonal multiple
access systems. IEEE Trans. Veh. Technol. 66(11), 10084–10091 (2017)
33. J.M.B. da Silva, Y. Xu, G. Fodor, C. Fischione, Distributed spectral efficiency maximization
in full-duplex cellular networks, in 2016 IEEE International Conference on Communications
Workshops (ICC) (2016), pp. 80–86
34. R. Sultan, L. Song, K.G. Seddik, Y. Li, Z. Han, Mode selection, user pairing, subcarrier allocation and power control in full-duplex OFDMA HetNets, in 2015 IEEE International Conference
on Communication Workshop (ICCW) (2015), pp. 210–215
35. Y. Sun, D.W.K. Ng, Z. Ding, R. Schober, Optimal joint power and subcarrier allocation for
full-duplex multicarrier non-orthogonal multiple access systems. IEEE Trans. Commun. 65(3),
1077–1091 (2017)
36. Y. Sun, D.W.K. Ng, R. Schober, Joint power and subcarrier allocation for multicarrier fullduplex systems, in 2017 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP) (2017), pp. 6548–6552
37. P. Tehrani, F. Lahouti, M. Zorzi, Resource allocation in OFDMA networks with half-duplex
and imperfect full-duplex users, in 2016 IEEE International Conference on Communications
(ICC) (2016), pp. 1–6
38. T.K. Vu, M. Bennis, S. Samarakoon, M. Debbah, M. Latva-aho, Joint load balancing and
interference mitigation in 5G heterogeneous networks. IEEE Trans. Wirel. Commun. 16(9),
6032–6046 (2017)
Chapter 10
Heterogeneous NOMA with Energy
Cooperation
Bingyu Xu, Yue Chen and Yuanwei Liu
10.1 Background
10.1.1 Resource Allocation in NOMA HetNets
Heterogeneous networks (HetNets) are one of the “big three” technologies which
can achieve the system performance requirement of fifth-generation (5G) networks.
In conventional one-tier homogeneous networks, there is only one macro-BS in
each cell. In HetNets, which are also called multi-tier networks, small BSs such as
pico BSs and micro-BSs which have lower transmit powers and smaller coverages
are deployed within the coverage of the macrocell. These smaller BSs obtain higher
spectrum efficiency and throughput through reusing the spectrum. They are typically
deployed in the hotpot zones (e.g., areas with high traffic demand) and dead zones
(e.g., cell edge and areas cannot receive signal).
Recently, NOMA-enabled HetNets have attracted significant research interests. It
is worth introducing NOMA in HetNets due to the following key advantages: (1) in
HetNets, users are closer to their associated BSs, which can reduce the interference
between users and increase the accuracy of successive interference cancelation (SIC)
in NOMA systems. (2) NOMA is capable to deal with the fairness issue among users,
which is one of the main challenges of HetNets.
Although intensive research contributions have been conducted on the design
of NOMA transmission, resource allocation in NOMA-enabled HetNets remains
an open problem. Considering the fact that user association (UA) determines that
B. Xu (B) · Y. Chen · Y. Liu
Queen Mary University of London, Mile End Road, London E1 4NS, UK
e-mail: bingyu.xu@qmul.ac.uk
Y. Chen
e-mail: yue.chen@qmul.ac.uk
Y. Liu
e-mail: yuanwei.liu@qmul.ac.uk
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_10
305
306
B. Xu et al.
users should be connected to a specified BS to form a user group for superposition
transmission [1], the number of users associated with a BS has a significant effect
on the spectral and energy efficiency in NOMA multi-cell networks [2]. In addition,
power control is of great importance in such networks, since the intra-cell interference
and inter-cell interference need to be coordinated. Otherwise, the performance of cell
edge users will be significantly degraded [3].
A cooperative NOMA scheme in HetNets was proposed in [4] where each user
was served by a macro-BS and a pico BS simultaneously. Zhao et al. [5] and [6]
focused on downlink joint spectrum and power allocation problem which aims at
maximizing the sum rate of small cell users. The macro-BS used conventional OMA
protocol while small cells use NOMA protocol. Meanwhile, many works combined
the NOMA enabled HetNets with other 5G techniques, such as massive MIMO and
cloud radio access. A user association scheme was proposed in [7] for massive MIMO
enabled NOMA HetNets to maximize the biased average received power of users.
Meanwhile, the downlink power allocation problem in NOMA HetNets with cloud
radio access network was investigated in [8]. It analyzed the energy efficiency of the
network by finding the optimal number of BSs.
10.1.2 Energy Cooperation
In 5G mobile systems, one main goal is to improve energy efficiency significantly
compared to today’s networks [9]. Indeed, such large level of connectivity will
inevitably give rise to an unprecedented surge in global energy consumption, especially in networks with HetNets, which has a vast number of small cells. The latest
analysis shows that the energy demand for information and communications technology already accounts for almost 10% of the world’s total energy consumption [10].
In addition, critical environmental issues such as high carbon emissions are a big
concern. Hence, “greener” solutions need to be developed to enhance the network
energy efficiency. Among the emerging technologies, energy harvesting is regarded
as one viable solution [11]. By allowing base stations (BSs) to harvest energy from
renewable energy sources such as solar and wind, the conventional grid energy consumption of wireless networks can be greatly reduced.
Although renewable energy harvesting is a viable solution for cutting the conventional grid energy consumption in cellular networks, there are many challenges
for integrating energy harvesting capabilities into BSs [12]. In renewable energy
harvesting-enabled networks, BSs harvest variable amounts of renewable energy, due
to the fluctuating nature of renewable energy sources. When the renewable energy
harvested by BSs is insufficient to meet their load conditions, some user equipments
(UEs) have to be offloaded to distant BSs with abundant energy and may suffer more
from signal degradation. Moreover, some BSs may always have excessive harvested
energy (e.g., because of more favorable weather conditions) that will eventually be
wasted. Since the deployment of BSs with large energy storage capabilities brings
10 Heterogeneous NOMA with Energy Cooperation
307
high expenditure of networks [13], the energy fluctuation problem cannot be solely
solved by using storage.
To improve the utilization rate of renewable energy, with the development of
smart grid, the definition of energy cooperation is proposed which allows energy
transferred through grids [14]. By this way, energy can be shared between BSs with
acceptable energy loss during the energy transmission process.
Energy cooperation in the point-to-point transmission scenario has been studied
in [14–16]. In [15], one-way energy transfer in the Gaussian two-way channel and
multiple access channel were considered respectively. This line of work was extended
to the two-way case in [16]. The implementation of energy cooperation in multiple
access channels and multiple access relay networks were studied in [17] and [18],
respectively. In [19], an energy cooperation scheme in cognitive radio networks was
proposed to improve both the spectral and energy efficiency.
Recently, the potential of energy cooperation in renewable energy-enabled cellular
networks has been explored, and various energy-cooperation optimization problems
have been studied. In [13], a joint energy and spectrum allocation problem between
two neighboring cellular systems was formulated, which aimed to minimize the cost
of energy and bandwidth, and the problem was solved by convex optimization. The
power control problem between two BSs was considered in [20] under the assumption that the harvested energy, and the energy demand of BSs were deterministic.
In [21], the energy cost of cellular networks was minimized with the assumption
that BSs traded energy via the smart grid with different prices. The work of [22]
aimed to maximize the sum rate through optimizing the transmit powers of BSs in a
coordinated multipoint cluster. In [23], the energy trading problem was formulated
to minimize the average cost of energy exchange between BSs, and a dynamic algorithm was proposed based on the Lyapunov optimization technique, which did not
require the statistical knowledge of the channel and energy.
Due to the energy-saving feature of energy cooperation, it is worth to deploy
it in NOMA networks. Meanwhile, most of existing NOMA works such as [24–
29] only consider the case consisting of one BS and a group of users. Besides, the
practical multi-cell scenario is evaluated in this chapter considering the effect of
inter-cell interference which has a substantial impact on the system performance.
In this chapter, we study the power control and UA problem in energy-cooperationenabled two-tier NOMA HetNets. The content of this chapter is mainly based on our
previous published journal [30].
10.2 Network Model and Problem Formulation
In this section, the system model for energy cooperation in two-tier NOMA HetNets is
presented, and the corresponding joint UA and power control problem is formulated.
308
B. Xu et al.
10.2.1 Downlink NOMA Transmission
As shown in Fig. 10.1, a two-tier energy-cooperation-enabled HetNet consisting of
one macro BS (MBS) and M pico BSs (PBSs) is considered, where the NOMAbased downlink transmission is utilized, and all BSs are assumed to share the same
frequency band. In such a network, BSs are powered by both the conventional power
grid and renewable energy sources, and energy can be shared between BSs through
the smart grid. Let m ∈ {1, 2, 3, ..., M + 1} be the mth BS, in which m = 1 denotes
the MBS, and the other values denote PBSs. There are N randomly located user
equipments (UEs) in this network, and each UE is associated with only one BS.
All BSs and UEs are single-antenna nodes. It is assumed that the global perfect
channel state information (CSI) is available. Let j ∈ {1, 2, 3, ..., N } index the jth
signal transmitUE. According to the NOMA scheme [31, 32], the superimposed
H
N
=1, ∀m, j, where
ted by the BS m is sm = j=1 x jm P jm s jm with E s jm s jm
x jm ∈ {0, 1} is the binary UA indicator, i.e., x jm = 1 when the jth UE is associated
with the mth BS and otherwise it is zero, s jm is the jth user-stream and P jm is the
corresponding allocated transmit power. When the jth UE is associated with the mth
BS, its received signal can be expressed as
Smart Grid
MBS
PBS 1
PBS 2
UE
Renewable energy flow
Conventional grid energy flow
Superposition coded data flow
Solar panel
Renewable energy level
Fig. 10.1 An example of an energy-cooperation enabled two-tier NOMA HetNet powered by both
solar panels and the conventional grid
10 Heterogeneous NOMA with Energy Cooperation
y jm =
h jm
P jm s jm + h jm
′
309
N
Pj ′ m s j ′ m
x j′ m
′
j =1, j = j
Intra−cell interference
+
′
M+1
′
m =1,m =m
′
⎛
h mjm ⎝
N
j ′ =1
x j ′ m′
⎞
P j ′ m ′ s j ′ m ′ ⎠ +̟m ,
(10.1)
Inter−cell interference
where x j ′ m , x j ′ m ′ ∈ {0, 1}, h jm is the channel coefficient from the associated BS m,
′
′
h mjm is the interfering channel coefficient from the BS m , and ̟m is the additive
white Gaussian noise. The power density of ̟m is σ 2 . In NOMA systems, SIC
is employed at UEs, to cancel the intra-cell interference from the stronger UEs’
data signals. Without loss of generality, assuming that there are km (km ≤ N ) UEs
constituting a group that is served by the mth BS at the same time and frequency
band, the corresponding channel to inter-cell interference plus noise ratios (CINRs)
are ordered as
h k m 2
h jm 2
|h 1m |2
m
≥ · · · ≥ (2)
,
(10.2)
≥ · · · ≥ (2)
(2)
I1m
+ σ2
I jm + σ 2
Ik m m + σ 2
(2)
where I jm
is the inter-cell interference power at the jth UE and σ 2 is the noise power.
Based on the principle of multi-cell NOMA [31], the power allocation of the users’
data signals in the mth cell needs to satisfy
0 < P1m ≤ · · · ≤ P jm ≤ · · · ≤ Pkm m ,
km
P jm = Pm ,
(10.3)
j=1
where Pm is the total transmit power of the mth BS. Such order is optimal for decoding
and guarantees the user fairness [31], namely the data signals of users with weaker
downlink channels and larger interference need to be allocated more transmit power
(2)
= 0, (10.3)
to achieve the desired QoS. For the special case of single-cell, i.e., I jm
reduces to the order based on the channel power gains, as seen in [32]. Therefore,
based on (10.1), the data rate after SIC at the jth UE is given by
τ jm = W log2 1 + γ jm ,
(10.4)
where W is the system bandwidth, and γ jm is the signal-to-interference-plus-noise
ratio (SINR) given by
310
B. Xu et al.
2
P jm |h jm |
γ jm =
M+1
2 j−1
h jm
Pj ′ m +
′
′
′
m =1,m =m
j =1
(1)
I jm
=
(2)
I jm
P jm
j−1
′
j =1
(2)
P j ′ m +(I jm
+σ 2 )/|h jm |
2
,
′ 2
m
h jm Pm ′ +σ 2
(10.5)
j ≤ km
(1)
in which I jm
is the remaining intra-cell interference after SIC, and Pm ′ =
N
′
′ ′
′ ′
j ′ =1 x j m P j m is the total transmit power of the m th BS. Although this chapter
focuses on the single-carrier system, it can be straightforwardly extended to the
multi-carrier system by letting W be the subcarrier bandwidth and τ jm multiply the
subcarrier indicator to be the data rate of a subcarrier. Thus, the optimal solution over
all subcarriers in the multi-carrier case can be iteratively obtained by following the
decomposition approach of this chapter.
10.2.2 Energy Model
Each BS is powered by both the conventional grid and renewable energy sources.
The energy drawn by the mth BS from the conventional grid is denoted as G m . The
energy harvested by the mth BS from renewable energy sources is denoted by E m .
′
The energy transferred from BS m to BS m is denoted as Emm ′ , and the energy
transfer efficiency factor between two BSs is denoted as β ∈ [0, 1]. Hence (1 − β)
specifies the level of energy loss during the energy transmission process. In addition,
it is assume that there is no battery to avoid the time-consuming and expensive energy
waste during the charging/discharging process, and the energy-cooperation problem
in each time slot is independent. The time slot length is normalized as one to simplify
the power-to-energy conversion. Therefore, the transmit energy consumption at the
mth BS should satisfy.
Pm ≤ G m + E m +
β
M+1
Em ′ m
m ′ =1,m ′ =m
Energy received from other BSs
−
′
M+1
Emm ′
, (10.6)
′
m =1,m =m
Energy transferred to other BSs
where Pm = Nj=1 x jm P jm is the total transmit power of the mth BS.
From (10.6), it is seen that in energy-cooperation-enabled networks, the grid
energy consumption of a BS depends on its harvested renewable energy, transferred
energy and transmit power. Given a BS’s transmit power, its grid energy consumption
needs to be formulated as a random variable, since the amount of harvested renewable
10 Heterogeneous NOMA with Energy Cooperation
311
energy and transferred energy is uncertain, which is different from the conventional
network without energy cooperation.
10.2.3 Problem Formulation
Our aim is to maximize the energy efficiency of such networks. The energy efficiency
(bits/Joule) is defined as the ratio of the overall network data rate to the overall grid
energy consumption; i.e., the network utility is
U (x, P, E , G) =
N
M+1
m=1 j=1
M+1
Gm .
x jm τ jm /
(10.7)
m=1
In this way, the harvested renewable energy can be maximally utilized to reduce the
grid energy consumption [11]. Therefore, our problem can be formulated as follows:
P1 :
max
x,P,E ,G
s.t. C1 :
M+1
U (x, P, E , G)
(10.8)
x jm τ jm ≥ τ min , ∀ j,
m=1
C2 :
M+1
x jm = 1, ∀ j,
m=1
C3 : Pm +
′
M+1
Emm ′ ≤ G m + E m + β
′
m =1,m =m
C4 :
N
M+1
Em ′ m , ∀m,
m ′ =1,m ′ =m
x jm P jm = Pm , ∀m,
j=1
C5 : x jm ∈ {0, 1} , ∀ j, ∀m,
C6 : G m ≥ 0, Emm ′ ≥ 0, ∀ j, ∀m,
m
C7 : 0 ≤ Pm ≤ Pmax
, P jm ≥ 0, ∀ j, ∀m,
where x = x jm , P = P jm , E = Emm ′ , G = [G m ], τ min denotes the required
m
minimum data rate for a UE, Pmax
is the maximum transmit power of the BS m.
Constraint C1 guarantees the QoS. C2 and C5 ensure that each UE cannot be associated with multiple BSs. C3 is the energy consumption constraint, and C4 is the
power allocation under NOMA principle in a cell. C6 indicates that the consumed
grid energy and transferred energy are nonnegative values, and C7 is the maximum
transmit power constraint.
312
B. Xu et al.
From the objective of P1 and its constraint C3, it can be found that when more
renewable energy is harvested and shared between BSs, the total grid energy consumption of the network can be reduced, which boosts the energy efficiency.
10.3 Proposed Resource Allocation Scheme
10.3.1 Resource Allocation Under Fixed Transmit Power
P1 is a mixed integer nonlinear programming (MINLP) problem, and constitutes a
challenging problem. In this section, it is assumed that the transmit power is fixed,
and accordingly the original problem P1 can be simplified as
P2 :
U (x, E , G)
max
x,E ,G
(10.9)
s.t. C1, C2, C3, C4, C5, C6.
The problem P2 is still a combinatorial problem due to its discrete nature. To
efficiently solve it, we adopt a decomposition approach. For a given G and E , the
above problem can be rewritten as
max U (x)
P2.1 :
(10.10)
x
s.t. C1, C2, C4, C5.
10.3.1.1
Lagrangian Dual Analysis
Based on P2.1, the Lagrangian function can be written as
L (x, λ, θ) = U (x) −
N
j=1
λ j τ min −
M+1
x jm τ jm
m=1
−
M+1
θm
m=1
N
j=1
x jm P jm − Pm , (10.11)
where λ j and θm are the nonnegative Lagrange multipliers. Then, the dual function
is given by
g(λ, θ ) =
max L(x, λ, θ )
x
s.t. C2, C5
,
(10.12)
and the dual problem of P2.1 is expressed as
min g (λ, θ ) .
λ,θ
(10.13)
10 Heterogeneous NOMA with Energy Cooperation
313
Given the dual variables λ j and θm , the optimal solution for maximizing the
Lagrangian w.r.t. x is
x ∗jm
=
1, if m = m ∗
,
0, otherwise
(10.14)
where m ∗ = argmax μ jm with
m
μ jm = τ jm /
M+1
(10.15)
G m + λ j τ jm − θm P jm .
m=1
The solution of (10.14) can be intuitively interpreted based on the fact that given the
grid energy consumption, users select BSs which provide the maximum data rates.
Since the objective of the dual problem is not differentiable,
we utilize the subgradient
method to obtain the optimal solution λ∗ , θ ∗ of the dual problem, which is given
by
λ j (t + 1) = λ j (t) − δ (t)
M+1
x jm τ jm − τ min
m=1
θm (t + 1) = θm (t) − δ (t) Pm −
N
j=1
x jm P jm
+
,
(10.16)
,
(10.17)
+
where [a]+ = max {a, 0}, t is the iteration index, and δ (t) is the step size. Note that
there exist several step size selections such as constant step size and diminishing step
size. Here, the nonsummable diminishing
step length is used [33].
After obtaining the optimal λ∗ , θ ∗ based on (10.16) and (10.17), the corresponding x is the solution of the primal problem P2.1. Therefore, based on the Lagrangian
dual analysis, UA can be determined in a centralized or distributed way. The centralized UA is intuitive and requires a central controller, which has the global CSI
and determines which user is connected to a BS in this network. In this chapter,
a distributed UA algorithm which does not require any centralized coordination is
proposed, as summarized in Algorithm 1. Since our problem satisfies the conditions of the convergence proof in [33], the convergence of the proposed algorithm
is guaranteed. The complexity of the proposed algorithm is O ((M + 1)N ) for each
iteration and the convergence is fast (less than 40
iterations in the simulation), which
is much lower than the brute force algorithm O (M + 1) N . Note that the broadcast
operations have a negligible effect on computational complexity (Table 10.1).
314
B. Xu et al.
Table 10.1 Algorithm 1 distributed user association
Step 1: At user side
1: if t = 0
2: Initialize λ j (t), ∀ j. Each UE measures its received inter-cell interference via pilot signal
from all BSs, and feedbacks the CINR values to the corresponding BSs. Meanwhile, each
UE selects the BS with the largest CINR value.
3: else
4: User j receives the values of μ jm and τ jm from BSs.
5: Determines the serving BS m according to m ∗ = argmax μ jm .
m
6: Update λ j (t) according to (10.16).
7: end if
8: t ← t + 1.
9: Each user feedbacks the UA request to the chosen BS, and broadcasts the value of λ j (t).
Step 2: At BS side
1: if t = 0
2: Initialize θm (t), ∀m.
3: else
4: Receives the updated user association matrix x.
6: Updates θm (t) according to (10.17), respectively.
7: Each BS calculates μ jm and τ jm under NOMA principle.
8: end if
9: t ← t + 1.
10: Each BS broadcasts the values of μ jm and τ jm .
10.3.1.2
Genetic Algorithm
In this subsection, a genetic algorithm (GA)-based UA is proposed to solve the
problem P2.1. Such algorithm will be compared with the proposed Algorithm 1.
GA can achieve good performance when the population of candidate solutions is
sufficient [34]. Specifically, each feasible chromosome represents a possible solution
that satisfies the constraints of problem P2.1, which is defined as
Di = {[m 1i ] , [m 2i ] , . . . , [m N i ]} , i ∈ {1, . . . , K },
(10.18)
where m ji is the gene representing the index of the BS that the jth UE is associated
with, and it has an integer value varying from 1 to M + 1, and K is the population size.
During each generation, the fitness of each chromosome is evaluated, to select high
fitness chromosomes and produce higher fitness offsprings. Based on the objective
of problem P2.1, the fitness value of the chromosome Di is calculated as
10 Heterogeneous NOMA with Energy Cooperation
315
Table 10.2 Algorithm 2 genetic algorithm-based user association algorithm
1: if t = 0
2: Initialize a set of feasible chromosomes {Di } with population size K , and the maximum
number of generations tmax .
3: else
4: Rank {Di } based on the fitness values given by (10.19).
5: Based on the selection probability ρs (r ), chromosomes are selected to produce offspring
via uniform crossover and mutation operations.
6: if exceed the maximum number of generations
! "
! "
7:
x ∗jm := Di∗ , where Di∗ is the feasible chromosome with the highest fitness value.
8:
break
9: else
10:
t ← t + 1.
11: end if
12: end if
Φi (Di ) = U (Di ) .
(10.19)
Then, all chromosomes are ranked from the best to the worst with ranking r , based
on their fitness values. The probability that a chromosome is selected as a parent
q(1−q)r −1
to produce offspring is given by ρs (r ) = 1−(1−q)
K with a predefined value q [34].
In each generation process, a uniform crossover operation with the probability ρc
is utilized to produce offspring by swapping and recombining genes based on the
parental chromosomes. In addition, a uniform mutation operation with the probability
ρm is employed. Such generation procedure is repeated until reaching the maximum
number of generations, and is summarized in Algorithm 2. Given the maximum
number of generations Ω and fixed population size K , the complexity of the proposed
algorithm is O (Ω K log(K )) [35]. The performance of the GA-based UA algorithm
heavily depends on the population size and number of generations, due to the inherent
nature of GA [34]. In the simulation results of Sect. 10.5, we will demonstrate that
overall, the proposed Algorithm 1 outperforms GA-based Algorithm 2 which is
shown in Table 10.2 when the population size of GA is not very large, and thus has
lower complexity.
The aforementioned approach
UA solutions for problem P2.1. After
provides
obtaining the UA solution x = x ∗jm , the corresponding pair (G, E ) is obtained by
solving the following simple linear programming (LP):
316
B. Xu et al.
Table 10.3 Algorithm 3 resource allocation algorithm under fixed transmit power
1: if t = 0
2: For a fixed P, initialize G m , ∀ j, m.
3: else
4: Determine x jm (t) under fixed (E , G) by selecting the user association algorithm from
Algorithm 1 or Algorithm 2.
5: Given x jm (t), update the energy allocation policy (E , G) by solving the LP P2.2 via CVX.
6: if convergence
7:
Obtain optimal resource allocation policy (x∗ , E ∗ , G∗ ).
8:
break
9: else
10:
t ← t + 1.
11: end if
12: end if
Table 10.4 Algorithm 4 one-dimensional search algorithm
1: if t = 0
l = 0, ν h = ν max , ∀m, calculate F =
2: Given χ j , initialize νm
l
m
m
Fh =
N
j=1
j-th
N
j=1
∗(l)
x jm P jm and
$
$
#
#
∗(h)
∗(h)
∗(l)
x jm P jm , where P jm and P jm are the allocated transmit powers of the
l and ν h respectively, which are calculated by using
UE’s data stream for the cases of νm
m
(10.28).
3: else
4: while Fl = ϕm and Fh = ϕm
ν l +ν h
5:
Let νm = m 2 m , and compute Fm .
6:
if Fm = νm
7:
The optimal dual variable νm∗ is obtained.
8:
break
9:
elseif Fm < ϕm
10:
νmh = νm .
11:
else Fm > ϕm
l =ν .
12:
νm
m
13:
end if
14: end while
15: end if
10 Heterogeneous NOMA with Energy Cooperation
317
Table 10.5 Algorithm 5 Joint User Association and Power Control
1: if t = 0
2: Initialize Pm , G m , E m , ∀m
3: else
4: Determine x jm (t) under (P, G, E ) by selecting the user association algorithm from
Algorithm 1 or Algorithm 2.
5: Given x jm (t) and the corresponding (G, E ), update the transmit power P based on the
following rule:
Loop:
(2)
a) Given Θ jm , loop over UE j:
! "
! "
i): Obtain νm∗ using Algorithm 4 given χ j
!
"
ii): Obtain P jm according to (10.28) with νm∗ , χ j .
! "
iii): Update χ j using subgradient method.
iv): Update P jm using (10.29).
Until convergence.
(2)
b) Update Θ jm using (10.27).
Until convergence.
6: Based on the updated P, update G m and Emm ′ by solving LP problem P2.2 via CVX.
7: if convergence
8:
Obtain optimal resource allocation policy (x∗ , P∗ , E ∗ , G∗ ).
9:
break
10: else
11:
t ← t + 1.
12: end if
13: end if
P2.2 : min
E ,G
M+1
Gm
(10.20)
m=1
s.t. C3, C6.
The problem P2.2 can be efficiently solved by using existing software, e.g. CVX [36].
grid
When no energy cooperation is allowed, i.e., Emm ′ = 0, ∀ j, ∀m, the optimal
energy consumption G of problem P2.2 under the UA solution x = x ∗jm is directly
obtained as
G ∗m = [Pm − E m ]+ ,
(10.21)
where Pm = Nj=1 x ∗jm P jm .
Based on the solutions of subproblems P2.1 and P2.2, we propose an iterative
algorithm to solve the problem P2, which is summarized in Algorithm 3 (Table 10.3).
318
B. Xu et al.
10.3.2 Resource Allocation Under Power Control
In this subsection, we consider the joint resource allocation and power control design.
Specifically, we develop an algorithm to solve the MINLP problem P1 through the
decomposition approach. As discussed in the previous section, we first determine the
UA indicators given the resource allocation policy (P, E , G), which can be obtained
problem P2.1 via Algorithm 1 or Algorithm 2. Then, under a fixed UA
"
!by solving
x jm , the problem for optimizing (P, E , G) is written as
P3 :
max
P,E ,G
U (P, E , G)
(10.22)
s.t. C1, C3, C4, C6, C7.
From the utility function, we find that the power allocation vectors P and G are
coupled within the objective of problem P3. Thus, given G and E , the above problem
can be decomposed into
P3.1 : max
P
N
M+1
(10.23)
x jm τ jm
m=1 j=1
s.t. C1, C3, C4, C7.
Problem P3.1 is non-convex. Hence, we provide a tractable suboptimal solution
based on the Karush–Kuhn–Tucker (KKT) conditions. The Lagrangian function of
problem P3.1 is
L (P, ν, χ ) =
N
M+1
x jm τ jm −
m=1 j=1
−
M+1
m=1
νm
N
+1
j=1
N
j=1
x jm P jm
M+1
χ j τ min −
x jm τ jm
m=1
− ϕm ,
(10.24)
$
#
M+1
m
′
′, P
where
ϕm = min G m + E m + β mM+1
E
′
′
′ =1,m ′ =m Em m −
mm
max
m =1,m =m
according to constraints C3 and C7, and χ j and νm are the nonnegative Lagrange
multipliers.
Without loss of generality, assuming that the jth UE is associated with the BS m,
i.e., x jm = 1, based on the KKT conditions, we have
∂L
∂ P jm
where Λ jm =
WΛ
(2)
− Θ (1)
= 1 + χ j 1+P jm jm
jm − Θ jm − νm log(2)
Λ jm
=
0,
|h jm |2
(1)
(2)
I jm
+I jm
+σ 2
(10.25)
is referred to as the channel-to-interference-plus-noise ratio
at the jth UE. Based on (10.3) and (10.5), Θ (1)
jm resulting from the intra-cell interfer-
10 Heterogeneous NOMA with Energy Cooperation
319
ence is given by
Θ (1)
jm
=
km
(1 + χℓ )
ℓ> j
W γℓm
Λlm ,
1 + γℓm
(10.26)
and Θ (2)
jm resulting from the inter-cell interference is given by
Θ (2)
jm =
′
M+1
′
m =1,m =m
2
1 + χ j ′ x j ′ m ′ W γ j ′ m ′ h mj′ m ′
.
(1)
(2)
2
′ ′
′
I
+
I
+
σ
1
+
γ
′
′
′
′
j
m
j =1
j m
j m
N
(10.27)
Based on (10.25), the transmit power allocated to the jth user-stream in the mth
cell is obtained as
∗
=
P jm
(1 + χ j )W
(2)
Θ (1)
jm + Θ jm + νm log(2)
−
1 +
.
Λ jm
(10.28)
the allocated transmit power is a monotonic function of νm . As such, given
!In (10.28),
"
χ j , we adopt a one-dimensional search over the Lagrange multipliers {νm }, which
can efficiently obtain the optimal ν ∗ that satisfies constraints C3 and C7. Accord∗
max
ing to (10.28),# we can easily find that νm∗ needs
to satisfy
$ 0 ≤ νm ≤ νm , where
(2)
∗
(1 + χ j )W Λ jm − Θ (1)
νmmax = max j
jm − Θ jm / log(2) . Here, νm = 0 represents
that there is no limitation on the transmit power of the jth user-stream and νm∗ = νmmax
corresponds to the! case
" that no transmit power is allocated to the jth user-stream.
Thus, by fixing χ j , ν ∗ can be obtained by using Algorithm 4 (Table 10.4).
For achieving a specific accuracy ς , the complexity of Algorithm 4 is O (log (1/ς )).
After obtaining ν ∗ , the Lagrange multiplier χ j can be updated by using the subgradient method, which is similar to (10.16).
To ensure the system stability, we utilize the Mann iterative method to update the
transmit power in each iteration [37], which is given by
(ℓ)
(ℓ+1)
∗
,
+ η(ℓ)P jm
= (1 − η(ℓ))P jm
P jm
(10.29)
where ℓ is the iteration index, 0 < η(ℓ) < 1 is the step size, which is usually chosen as
ℓ
. After obtaining the optimal solution of problem P3.1, the corresponding
η(ℓ) = 2ℓ+1
(G, E ) can be updated by solving the LP problem P2.2 via CVX. As such, the solution
of problem P3 can be iteratively obtained. Note that the convergence of the KKTbased algorithm is usually faster than the gradient-based designs [38].
Based on the previous analysis, the proposed joint UA and power control scheme
in energy-cooperation enabled NOMA HetNets is summarized in Algorithm 5
(Table 10.5).
320
B. Xu et al.
10.3.3 Comparison with FTPA
In 4G networks, fractional transmission power allocation (FTPA) scheme is
adopted [31]. The rule of FTPA is that the transmit power will be allocated based
on the UEs’ channel conditions, i.e., the data signals of UEs with weaker downlink
channels will own more transmit power. Based on the CINR order in (10.30), the
transmit power allocated to the jth UE’s data stream in the mth cell under FTPA
protocol is expressed as [31]
P jm
−α
−α N
h jm 2
|h lm |2
/
xlm
,
= Pm
(2)
(2)
I jm
+ σ2
Ilm
+ σ2
l=1
(10.30)
where 0 ≤ α ≤ 1 is the decay factor. Here, α = 0 represents equal power allocation.
For larger α, the transmit power allocated to the data-stream of the user with largest
CINR becomes lower, and more power will be allocated to the data-stream of the user
with the lowest CINR, in order to achieve the user fairness and the optimal decoding.
However, the detrimental effect of using such simple power allocation scheme is
that distant users may receive severer inter-cell interference without power control
among BSs, due to the fact that each BS has to assign larger transmit power to the faraway users. Therefore, compared to the single-cell NOMA case [32], the inter-cell
interference has a significant impact on the power allocation of multi-tier NOMA
HetNets.
10.3.4 Comparison with No Renewable Energy
When there is no renewable energy harvesting (i.e., E m = 0, ∀m), no renewable
′
energy can be shared between BSs (i.e., Emm ′ = Em ′ m = 0, ∀m, m ), and thus, the
required energy can only be supplied by the conventional grid. In this case, Pm =
G m , ∀m, and the original problem P1 reduces to
P4 : max
x,P
M+1 N
m=1
j=1
M+1 N
m=1
j=1
x jm τ jm
x jm P jm
(10.31)
s.t. C1, C2, C4, C5, C7.
The above problem is nonlinear fractional programming and NP-hard, which can be
solved by using the proposed Algorithm 5 with E m = 0 and Emm ′ = Em ′ m = 0.
10 Heterogeneous NOMA with Energy Cooperation
321
10.3.5 Comparison with No Energy Cooperation
In this case, the energy transfer efficiency β is set to 0, which means that the harvested renewable energy cannot be transferred between BSs. Each BS is powered by
the conventional grid and its harvested renewable energy; i.e., the transmit energy
consumption at a BS needs to satisfy Pm ≤ G m + E m , ∀m. Then, the proposed Algorithm 5 can still be applied to solve this problem, and during each iteration, the grid
energy consumption is updated as G m = [Pm − E m ]+ based on the updated Pm .
10.4 Simulation Results
In this section, we present numerical results to demonstrate the effectiveness of
the proposed algorithm compared with other schemes as well as the conventional
counterpart. Since the renewable energy arrival rate changes slowly in practice and is
stationary at each information transmission time slot [39], we consider the amounts
of harvested energy at the MBS and PBSs to be constant and each PBS has the same
level of renewable energy during each transmission time slot for the sake of simplicity.
For simplicity, the amount of harvested energy E m of BS m is modeled as a uniform
distribution Um [am , bm ], and varies across different transmission blocks, where am
and bm are the minimum and maximum harvested energy values of BS m follows
uniform distributions as shown in Table 10.6, respectively [40]. Our analysis and
proposed algorithm are independent of the specific renewable energy distribution.
For the channel h jm , we focus on the large-scale channel fading condition in low
mobility environment, due to the fact that UA is carried out in a large time scale
and the small-scale fading can be averaged out [41, 42]. In addition, PBSs and UEs
are uniformly distributed in a macrocell geographical area. The basic simulation
parameters are shown in Table 10.6.
10.4.1 User Association Under Fixed Transmit Power
In this subsection, we study different UA algorithms under fixed transmit power, i.e.,
power control is unavailable at BSs. Based on the NOMA power allocation condition
m
, and
in (10.3), we consider that the total transmit power at each BS is Pm = Pmax
adopt an arithmetic progression power allocation approach for the sake of simplicity,
2j
namely the transmit power of the jth user’s data signal is P jm = km (1+k
Pm , j ∈
m)
{1, 2, 3, ..., km } when km users are multiplexed in the power domain of the mth cell.
We also provide the comparison with the conventional reference signal received
power (RSRP)-based UA. The aim of this subsection is to show the performance of
different UA algorithms under the same fixed power allocation condition.
322
B. Xu et al.
Table 10.6 Simulation parameters
Parameter
Value
System bandwidth
Noise power density
Cell radius
Path loss of MBS
Path loss of PBS
Min harvested energy of MBS
Max harvested energy of MBS
Min harvested energy of PBS
Max harvested energy of PBS
Max transmit power of MBS
Max transmit power of PBS
Fig. 10.2 Energy efficiency
versus the number of UEs for
different UA algorithms
10 MHz
−174 dBm/Hz
500 m
128.1 + 37.6log10 d (km)
140.7 + 36.7log10 d (km)
575 W
660 W
15 W
25 W
46 dBm [43]
30 dBm [43]
5.5
×10 6
Energy Efficiency (bits/Joule)
5
NOMA, Proposed UA
NOMA, GA-based UA, Fixed Population Size=200
NOMA, RSRP-based UA
OMA, RSRP-based UA
4
3
2
10
15
20
25
30
Number of UEs
Figure 10.2 shows the energy efficiency versus the number of UEs with the number
of PBSs M = 6 and the energy transfer efficiency factor β = 0.9. We set the minimum QoS as τ min = 0.1 bits/s/Hz and the amount of energy harvested by MBS and
PBS as 37 dBm and 27 dBm, respectively.1 The maximum number of generations for
the GA-based UA is 10, q = 0.1, and ρc = ρm = 0.4. The proposed UA scheme with
NOMA achieves better energy efficiency than the other cases. The energy efficiency
1 In
real networks, the renewable energy generation rate is constant during a certain period, and
the time scale of the UA and power control process is much shorter, typically less than several
minutes [41, 42]. In addition, the amount of energy harvested by an MBS is usually larger than that
at a PBS, since MBS can fit larger solar panel [42, 44].
10 Heterogeneous NOMA with Energy Cooperation
Fig. 10.3 Energy efficiency
versus the number of PBSs
for different UA algorithms
Energy Efficiency (bits/Joule)
7
323
×10 6
6
5
4
NOMA, Proposed UA
NOMA, GA-based UA, Fixed Population Size=600
NOMA, RSRP-based UA
OMA, RSRP-based UA
3
5
10
15
20
25
Number of PBSs
increases with the number of UEs because of the multiuser diversity gain (i.e., different users experience different path loss, and more users with lower path loss help
enhance the overall energy efficiency.) [45]. The use of NOMA outperforms OMA.
By using the GA-based UA, the energy efficiency slowly increases with the number
of UEs, due to the fact that the efficiency of the GA-based algorithm depends on
the population size [34]. In other words, given the population size (e.g., K = 200 in
this figure), the GA algorithm may not obtain good solutions when the number of
UEs grows large, which indicates that larger populations of candidate solutions are
needed [34].
Figure 10.3 shows the energy efficiency versus the number of PBSs with the
number of UEs N = 40 and the energy transfer efficiency factor β = 0.9. We set the
minimum QoS as τ min = 0.1 bits/s/Hz and the amount of harvested energy at MBS
and PBS as 37 dBm and 27 dBm, respectively. The maximum number of generations
for GA is 10, q = 0.1, and ρc = ρm = 0.4. NOMA achieves higher energy efficiency
than OMA, since NOMA can achieve higher spectral efficiency. The proposed UA
algorithm outperforms the other cases and the performance gap between the proposed
UA and the conventional RSRP-based UA is larger when deploying more PBSs, due
to the fact that the proposed UA can achieve more BS densification gains [9]. For the
GA-based UA algorithm with the population size K = 600, solutions are inferior
when the number of PBSs is large, as larger populations of candidate solutions are
needed [34].
324
B. Xu et al.
10.4.2 Power Control Under Fixed User Association
In, this subsection, we consider three power allocation schemes, namely the power
control method proposed in Sect. 10.4, fractional transmission power allocation
(FTPA) and the conventional fixed transmit power, to confirm the advantages of
our proposal. We adopt the conventional RSRP-based UA in the simulation, and all
the considered cases experience the same UA condition. In addition, BSs use their
maximum transmit powers in the OMA scenario, and the total transmit power of
m
, m ∈ {1, 2, 3, ..., M + 1}.
each BS for FTPA is set as Pm = Pmax
Figure 10.4 shows the energy efficiency versus the number of PBSs with the number of UEs N = 50 and the energy transfer efficiency factor β = 0.9. We set the
minimum QoS as τ min = 1 bits/s/Hz and the amount of energy harvested by MBS
and PBS as 37 dBm and 27 dBm, respectively. We see that by using NOMA with
the proposed power control, energy efficiency rapidly increases with the number of
PBS. The proposed algorithm achieves better performance than the other cases. When
deploying more PBSs, the performance gap between the proposed solution and the
other cases is larger, which indicates that the proposed power control algorithm can
achieve more BS densification gains and efficiently coordinate the inter-cell interference. When the number of PBSs is not large, NOMA with FTPA can outperform
the conventional OMA case, since NOMA can achieve better spectral efficiency than
OMA [32]. However, when adding more PBSs, NOMA with FTPA may not provide
higher energy efficiency. The reason is that more UEs will be offloaded to picocells,
and the inter-cell interference will become severer, which means that the transmit
power of each user-stream needs to be larger to combat the inter-cell interference.
As suggested in Sect. 10.3.3, FTPA with α = 0 achieves a higher energy efficiency
of the network than the α = 0.7 case, since the data-streams for UEs with poorer
Fig. 10.4 Energy efficiency
versus the number of PBSs
for different power allocation
policies
2
×10 7
NOMA, Proposed Power Control
1.8
NOMA, FTPA with
NOMA, FTPA with
α =0
α =0.7
Energy Efficiency (bits/Joule)
OMA, Fixed Transmit Power
1.4
1
0.6
0.2
1
5
10
Number of PBSs
15
20
10 Heterogeneous NOMA with Energy Cooperation
Fig. 10.5 Energy efficiency
versus energy transfer
efficiency factor for different
power allocation policies
2.2
×10 7
NOMA, Proposed Power Control
NOMA, FTPA with α =0
NOMA, FTPA with α =0.7
OMA, Fixed Transmit Power
1.8
Energy Efficiency (bits/Joule)
325
1.4
1
0.6
0.2
0
0.2
0.4
0.6
0.8
1
Energy Transfer Efficiency Factor
channel condition (i.e., lower CINR) have to be allocated more power in the case
of FTPA with α = 0.7, which reduces the total throughput of the network under the
same energy consumption.
Figure 10.5 shows the energy efficiency versus the energy transfer efficiency factor
β with the number of PBSs M = 3 and the number of UEs N = 40. We set the
minimum QoS to τ min = 1 bits/s/Hz and the amount of harvested energy at MBS and
PBS to 40 dBm and 35 dBm, respectively. Compared to the no energy-cooperation
case (i.e., β = 0), the use of energy cooperation can enhance the energy efficiency,
particularly when the energy transfer efficiency factor is large. The implementation
of NOMA can achieve higher energy efficiency than the conventional OMA system
because of higher spectral efficiency, and the proposed power control algorithm
outperforms the other cases. Moreover, the energy efficiency grows at a much higher
speed when applying the proposed algorithm. For a specified β, FTPA with α = 0
achieves higher energy efficiency of the network than the α = 0.7 case, as suggested
in Fig. 10.4.
Figure 10.6 shows the trade-off between the energy efficiency and the minimum
QoS with the number of PBSs M = 3 and the number of UEs N = 30. We set the
energy transfer efficiency factor to β = 0.9 and the amount of energy harvested by
MBS and PBS to 37 dBm and 27 dBm, respectively. For a given minimum QoS,
the proposed power control under NOMA achieves higher energy efficiency than
conventional OMA. When better QoS is required by the UE, energy efficiency of
both NOMA and OMA cases decreases. The reason is that for the proposed solution,
more transmit power will be allocated to the UEs with lower CINRs to achieve such
minimum QoS, which results in more energy consumption; for conventional OMA,
it means that more users cannot obtain the desired QoS and have to experience an
outage. We see that energy efficiency decreases significantly in the low minimum
QoS regime, because many UEs receive low QoS and increasing the level of the
326
B. Xu et al.
Fig. 10.6 Trade-off between
the energy efficiency and the
minimum QoS for NOMA
and OMA
7
×10 6
Energy Efficiency (bits/Joule)
6
5
4
NOMA, Proposed Power Control
OMA, Fixed Transmit Power
3
2
1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Minimum QoS (bits/s/Hz)
minimum QoS means that these UEs cannot be served. In practice, the minimum
QoS can be found in an off-line manner [46].
10.4.3 Joint User Association and Power Control
In this subsection, we examine the benefits of joint UA and power control design
in energy-cooperation-enabled NOMA HetNets. We also present comparisons by
considering different power allocation schemes with the conventional RSRP-based
m
in the OMA
UA. In the OMA scenario, transmit power at the BS is set to Pm = Pmax
scenario.
Figure 10.7 shows the energy efficiency versus the number of UEs with the number of PBSs M = 5 and the energy transfer efficiency factor β = 0.9. We set the
minimum QoS as τ min = 0.5 bits/s/Hz and the amount of harvested energy at MBS
and PBS as 32 dBm and 22 dBm, respectively. We see that the proposed joint UA and
power control algorithm achieves higher energy efficiency than the other cases, and
significantly improves the performance when more UEs are served in the network.
The reason is that the proposed algorithm is capable of obtaining larger multiuser
diversity gains. The use of NOMA can obtain higher energy efficiency than the OMA
case, due to NOMA’s capability of achieving higher spectral efficiency. Additionally,
when equal power allocation is adopted in NOMA HetNets with the conventional
RSRP-based UA, energy efficiency decreases with increasing the number of UEs of
the network, which can be explained by the fact that given the total transmit power of
a BS, the transmit power allocated to the data-streams of the UEs with better channel
condition reduces when more UEs are served simultaneously.
10 Heterogeneous NOMA with Energy Cooperation
Fig. 10.7 Energy efficiency
versus the number of UEs for
different joint UA and power
allocation designs
12
327
×10 6
Energy Efficiency (bits/Joule)
10
8
6
NOMA, Proposed Joint UA and Power Control
NOMA, RSRP-based UA with Proposed Power Control
NOMA, RSRP-based UA with Equal Power Allocation
OMA, RSRP-based UA with Fixed Transmit Power
4
2
10
15
20
25
30
Number of UEs
Fig. 10.8 Energy efficiency
versus the number of PBSs
for different joint UA and
power allocation designs
2.2
×10 7
NOMA, Proposed Joint UA and Power Control
NOMA, RSRP-based UA with Proposed Power Control
NOMA, RSRP-based UA with Equal Power Allocation
OMA, RSRP-based UA with Fixed Transmit Power
Energy Efficiency (bits/Joule)
1.8
1.4
1
0.6
0.2
2
3
4
5
6
Number of PBSs
Figure 10.8 shows the energy efficiency versus the number of PBSs with the
number of UEs N = 50 and the energy transfer efficiency factor β = 0.9. We set the
minimum QoS as τ min = 0.1 bits/s/Hz and the amount of energy harvested by MBS
and PBS as 37 dBm and 27 dBm, respectively. The proposed design outperforms
the other cases. By using the proposed joint UA and power control with NOMA, the
energy efficiency significantly increases with the PBS number, since the proposed
design can obtain more BS densification gains. Again, the use of NOMA achieves
better performance than OMA. For the case of RSRP-based UA with NOMA and
equal power allocation, energy efficiency decreases with increasing the number of
328
B. Xu et al.
PBSs, because the inter-cell interference has a big adverse effect on the NOMA
transmission [3].
10.5 Conclusion and Future Work
This chapter studied UA and power control in energy-cooperation- aided two-tier HetNets with NOMA. A distributed UA algorithm was proposed based on the Lagrangian
dual analysis, which does not require a central controller. Then, we proposed a joint
UA and power control algorithm which achieves higher energy efficiency performance than the existing schemes. The proposed power control algorithm satisfies
the KKT optimality conditions. Simulation results demonstrate the effectiveness of
the proposed algorithms. The results showed that the proposed algorithm can efficiently coordinate the intra-cell and inter-cell interference and has the capability of
exploiting the multiuser diversity and BS densification. The application of NOMA
can achieve larger energy efficiency than OMA due to the higher spectral efficiency
of NOMA.
To further extend this line of work, other UA optimization designs in multi-cell
NOMA networks such as proportional fairness or max-min fairness would be of
interest, and they are not trivial extensions since the optimization problems involved
will be distinct. Moreover, imperfect CSI can have a substantial effect on outage
probability and average data rate in NOMA networks, as analyzed in [47]. One of
the challenges for optimization designs under imperfect CSI is that error propagation
occurs since intra-cell interference cannot be perfectly canceled. Therefore, robust
optimization designs need to be developed in multi-cell NOMA networks. In addition, the application of MIMO technology in NOMA networks is another important
research area, which can significantly improve the performance gain [32]. In MIMONOMA networks, inter-user pair/group interference can deteriorate the performance,
as analyzed in [32, 48]. Therefore, how to mitigate the inter-user pair/group interference is crucial. Currently, UA and power control solutions in multi-cell MIMONOMA networks are not available, and more research efforts need to be made in this
area.
References
1. D. Liu, L. Wang, Y. Chen, M. Elkashlan, K.K. Wong, R. Schober, L. Hanzo. User association in
5G networks: a survey and an outlook. IEEE Commun. Surv. Tutor. 18(2), 1018–1044 (2016)
2. Y. Liu, Z. Qin, M. Elkashlan, Z. Ding, A. Nallanathan, L. Hanzo, Non-orthogonal multiple
access for 5g and beyond. Proc. IEEE 105(12), 2347–2381 (2017)
3. W. Shin, M. Vaezi, B. Lee, D.J. Love, J. Lee, H.V. Poor, Non-orthogonal multiple access
in multi-cell networks: Theory, performance, and practical challenges. IEEE Commun. Mag.
55(10), 176–183 (2017)
10 Heterogeneous NOMA with Energy Cooperation
329
4. Y. Xu, H. Sun, R. Q. Hu, Y. Qian, Cooperative non-orthogonal multiple access in heterogeneous
networks, in Proceedings of IEEE GLOBECOM (San Diego, CA, USA, Dec 2015), pp. 1–6
5. J. Zhao, Y. Liu, K.K. Chai, A. Nallanathan, Y. Chen, Z. Han, Resource allocation for nonorthogonal multiple access in heterogeneous networks. In Proceedings of IEEE ICC (Paris,
France, May 2017), pp. 1–6
6. J. Zhao, Y. Liu, K.K. Chai, A. Nallanathan, Y. Chen, Z. Han, Spectrum allocation and power
control for non-orthogonal multiple access in hetnets. IEEE Trans. Wirel. Commun. 16(9),
5825–5837 (2017)
7. Y. Liu, Z. Qin, M. Elkashlan, A. Nallanathan, J.A. McCann, Non-orthogonal multiple access in
large-scale heterogeneous networks. IEEE J. Sel. Areas Commun. 35(12), 2667–2680 (2017)
8. Q.T. Vien, T.A. Le, B. Barn, C.V. Phan, Optimising energy efficiency of non-orthogonal multiple access for wireless backhaul in heterogeneous cloud radio access network. IET Commun.
10(18), 2516–2524 (2016)
9. J.G. Andrews, S. Buzzi, W. Choi, S.V. Hanly, A. Lozano, A.C.K. Soong, J.C. Zhang, What
will 5g be? IEEE J. Sel. Areas Commun. 32(6), 1065–1082 (2014)
10. D. Jiang, P. Zhang, Z. Lv, H. Song, Energy-efficient multi-constraint routing algorithm with
load balancing for smart city applications. IEEE Internet Things J. 3(6), 1437–1447 (2016)
11. T. Han, N. Ansari, Powering mobile networks with green energy. IEEE Wirel. Commun. 21(1),
90–96 (2014)
12. O. Ozel, K. Tutuncuoglu, J. Yang, S. Ulukus, A. Yener, Transmission with energy harvesting
nodes in fading wireless channels: optimal policies. IEEE J. Sel. Areas Commun. 29(8), 1732–
1743 (2011)
13. Y. Guo, J. Xu, L. Duan, R. Zhang, Joint energy and spectrum cooperation for cellular communication systems. IEEE Trans. Commun. 62(10), 3678–3691 (2014)
14. B. Gurakan, O. Ozel, J. Yang, S. Ulukus, Energy cooperation in energy harvesting communications. IEEE Trans. Commun. 61(12), 4884–4898 (2013)
15. B. Gurakan, O. Ozel, J. Yang, S. Ulukus, Two-way and multiple-access energy harvesting
systems with energy cooperation, in Proceedings of IEEE ASILOMAR (Pacific Grove, CA,
USA, Nov. 2012), pp. 58–62
16. B. Gurakan, O. Ozel, J. Yang, S. Ulukus, Energy cooperation in energy harvesting two-way
communications. in Proceedings of IEEE ICC (Budapest, Hungary, June 2013), pp. 3126–3130
17. K. Tutuncuoglu, A. Yener, Multiple access and two-way channels with energy harvesting and
bi-directional energy cooperation. in Proceedings of IEEE ITA (San Diego, CA, USA, Feb
2013), pp. 1–8
18. K. Tutuncuoglu, A. Yener, Cooperative energy harvesting communications with relaying and
energy sharing. in Proceedings of IEEE ITW (Sevilla, Spain, Sept 2013), pp. 1–5
19. D. Wang, P. Ren, Y. Wang, Q. Du, L. Sun, Energy cooperation for reciprocally-benefited
spectrum access in cognitive radio networks, in Proceedings of IEEE GlobalSIP (Atlanta, GA,
USA, Dec 2014), pp. 1320–1324
20. Y.K. Chia, S. Sun, R. Zhang, Energy cooperation in cellular networks with renewable powered
base stations. IEEE Trans. Wirel. Commun. 13(12), 6996–7010 (2014)
21. J. Xu, L. Duan, R. Zhang, Cost-aware green cellular networks with energy and communication
cooperation. IEEE Commun. Mag. 53(5), 257–263 (2015)
22. J. Xu, R. Zhang, Comp meets smart grid: a new communication and energy cooperation
paradigm. IEEE Trans. Veh. Technol. 64(6), 2476–2488 (2015)
23. S. Lakshminarayana, T.Q.S. Quek, H.V. Poor, Cooperation and storage tradeoffs in power grids
with renewable energy resources. IEEE J. Sel. Areas Commun. 32(7), 1386–1397 (2014)
24. Y. Zhang, H.M. Wang, T.X. Zheng, Q. Yang, Energy-efficient transmission design in nonorthogonal multiple access. IEEE Trans. Veh. Technol. 66(3), 2852–2857 (2017)
330
B. Xu et al.
25. Z. Yang, Z. Ding, P. Fan, N. Al-Dhahir, A general power allocation scheme to guarantee quality
of service in downlink and uplink NOMA systems. IEEE Trans. Wireless Commun. 15(11),
7244–7257 (2016)
26. J. Cui, Z. Ding, P. Fan, A novel power allocation scheme under outage constraints in NOMA
systems. IEEE Signal Process. Lett. 23(9), 1226–1230 (2016)
27. F. Fang, H. Zhang, J. Cheng, V.C.M. Leung, Energy-efficient resource allocation for downlink
non-orthogonal multiple access network. IEEE Trans. Commun. 64(9), 3722–3732 (2016)
28. Z. Wei, D.W.K. Ng, J. Yuan, Power-efficient resource allocation for MC-NOMA with statistical
channel state information, in Proceedings of IEEE GLOBECOM (Washington, DC, USA, Dec
2016), pp. 1–7
29. Y. Sun, D.W.K. Ng, Z. Ding, R. Schober, Optimal joint power and subcarrier allocation for MCNOMA systems, in Proceedings of IEEE GLOBECOM (Washington, DC, USA, Dec 2016),
pp. 1–6
30. B. Xu, Y. Chen, J.R. Carrión, T. Zhang, Resource allocation in energy-cooperation enabled
two-tier noma hetnets toward green 5g. IEEE J. Sel. Areas Commun. 35(12), 2758–2770
(2017)
31. Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, K. Higuchi, Non-orthogonal
multiple access (NOMA) for cellular future radio access, in Proceedings of IEEE VTC (Spring)
(Dresden, Germany, June 2013) pp. 1–5,
32. Z. Ding, F. Adachi, H.V. Poor, The application of MIMO to non-orthogonal multiple access.
IEEE Trans. Wirel. Commun. 15(1), 537–552 (2016)
33. S. Boyd, A. Mutapcic. Subgradient Methods. Stanford University (2008)
34. K. Yasuda, L. Hu, Y. Yin, A grouping genetic algorithm for the multi-objective cell formation
problem. Int. J. Prod. Res. 43(4), 829–853 (2005)
35. D.E. Goldberg, K. Deb, A Comparative Analysis of Selection Schemes Used in Genetic Algorithms (Morgan Kaufmann Publishers Inc., San Mateo, CA, 1991)
36. M. Grant, S. Boyd. Cvx: Matlab software for disciplined convex programming
37. Z. Han, D. Niyato, W. Saad, T. Baar, A. Hjrungnes, Game Theory in Wireless and Communication Networks: Theory, Models, and Applications (Cambridge University, Press, 2012)
38. M. Kobayashi, G. Caire, An iterative water-filling algorithm for maximum weighted sum-rate
of Gaussian MIMO-BC. IEEE J. Sel. Areas Commun. 24(8), 1640–1646 (2006)
39. S. Zhang, N. Zhang, S. Zhou, J. Gong, Z. Niu, X.S. Shen, Energy-aware traffic offloading for
green heterogeneous networks. IEEE J. Sel. Areas Commun. 34(5), 1116–1129 (2016)
40. M. Zheng, P. Pawelczak, S. Stanczak, H. Yu, Planning of cellular networks enhanced by energy
harvesting. IEEE Commun. Lett. 17(6), 1092–1095 (2013)
41. K. Son, H. Kim, Y. Yi, B. Krishnamachari, Base station operation and user association mechanisms for energy-delay tradeoffs in green cellular networks. IEEE J. Sel. Areas Commun.
29(8), 1525–1536 (2011)
42. D. Liu, Y. Chen, K.K. Chai, T. Zhang, M. Elkashlan, Two-dimensional optimization on user
association and green energy allocation for HetNets with hybrid energy sources. IEEE Trans.
Commun. 63(11), 4111–4124 (2015)
43. A. Ghosh, N. Mangalvedhe, R. Ratasuk, B. Mondal, M. Cudak, E. Visotsky, T.A. Thomas, J.G.
Andrews, P. Xia, H.S. Jo, H.S. Dhillon, T.D. Novlan, Heterogeneous cellular networks: from
theory to practice. IEEE Commun. Mag. 50(6), 54–64 (2012)
44. T. Han, N. Ansari, Green-energy aware and latency aware user associations in heterogeneous
cellular networks, in Proceedings of IEEE GLOBECOM (Atlanta, GA, USA, Dec. 2013), pp.
4946–4951
45. D. Tse, P. Viswanath, Fundamentals of Wireless Communication (Cambridge University Press,
Cambridge, U.K., 2005)
10 Heterogeneous NOMA with Energy Cooperation
331
46. D.W.K. Ng, E.S. Lo, R. Schober, Energy-efficient resource allocation in ofdma systems with
large numbers of base station antennas. IEEE Trans. Wirel. Commun. 11(9), 3292–3304 (2012)
47. Z. Yang, Z. Ding, P. Fan, G.K. Karagiannidis, On the performance of non-orthogonal multiple
access systems with partial channel information. IEEE Trans. Commun. 64(2), 654–667 (2016)
48. Z. Ding, R. Schober, H.V. Poor, A general MIMO framework for NOMA downlink and uplink
transmission based on signal alignment. IEEE Trans. Wirel. Commun. 15(6), 4438–4454 (2016)
Chapter 11
NOMA in Vehicular Communications
Yingyang Chen, Li Wang, Yutong Ai, Bingli Jiao and Lajos Hanzo
11.1 Background and Motivation
With the rapid development of intelligent transportation systems (ITS), the broad
objective of vehicular communications is to improve the travel experience of users.
To support a variety ITS applications, the integrated vehicular networking concept
termed as ‘vehicle-to-everything’ (V2X) has been proposed. To elaborate a little
further, this includes four main types of communications, namely vehicle-to-vehicle
(V2V), vehicle-to-pedestrian (V2P), vehicle-to-infrastructure (V2I), and vehicle-tonetwork (V2N) scenarios, where the ultimate objective is that of offering improved
road safety, traffic efficiency, and infotainment services [1].
The IEEE 802.11p standard was conceived in support of wireless access for vehicular environments (WAVE), specifically dedicated to vehicular safety applications
and to the provision of data rates ranging from 6 to 27 Mbps for short-transmission
distances [2]. However, 802.11p fails to support flexible scalability and to provide
quality of service (QoS) guarantees. Furthermore, it has a potentially unbounded
delay. These characteristics of the IEEE 802.11p standard prevent its employment in
demanding V2X services requiring low latency and high reliability [3, 4]. Moreover,
Y. Chen · B. Jiao
School of Electronics Engineering and Computer Science, Peking University,
Beijing 100871, China
e-mail: chenyingyang@pku.edu.cn
B. Jiao
e-mail: jiaobl@pku.edu.cn
L. Wang (B) · Y. Ai
School of Electronic Engineering, Beijing University of Posts and Telecommunications,
Beijing 100876, China
e-mail: liwang@bupt.edu.cn
Y. Ai
e-mail: ytailiwang@bupt.edu.cn
L. Hanzo
School of Electronics and Computer Science, University of Southampton,
Southampton SO17 1BJ, U.K.
e-mail: lh@ecs.soton.ac.uk
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_11
333
334
Y. Chen et al.
due to the asynchronous nature of the IEEE 802.11p transmission, its performance
is inevitably degraded by the packet collisions imposed by the hidden node problem
encountered in carrier sense multiple access with collision avoidance (CSMA/CA)
solutions. Finally, the current status of IEEE 802.11p does not support an evolutionary path for improving its reliability, robustness, and coverage [5].
As an alternative to the IEEE 802.11p-based vehicular ad hoc network (VANET)
concept, the long-term evolution (LTE)-based V2X solution has been actively developed by the 3rd Generation Partnership Project (3GPP), and it was redefined as
LTE V2X within the 3GPP standardization framework so as to provide a beneficial
solution for V2X communications [6]. As a benefit of the global deployment and
commercialization of LTE systems, LTE V2X serves as an integrated solution for
vehicular communications. To be more specific, given an LTE network, both V2I
and V2N services can be supported at a high data rate, whilst maintaining an excellent QoS with the aid of the so-called eNodeBs. Meanwhile, LTE can be extended
to support V2V and V2P services by invoking direct device-to-device (D2D) communications for satisfying the QoS requirements even in the case of high vehicular
densities [7]. Finally, in contrast to 802.11p, the hidden node problem can be avoided
in LTE V2X scenario due to its synchronous nature, whilst both the reliability and
latency can also be improved.
The rest of this section is organized as follows. First, we provide an overview
of existing LTE-based V2X systems. Then, we describe the applicability of a range
of popular transmission techniques to vehicular communications. A non-orthogonal
multiple access (NOMA) and spatial modulation (SM)-based transmission scheme
are proposed for supporting the high data rate and high-reliability demands of V2X
systems. Finally, we detail the outline of this chapter.
11.1.1 Overview of LTE-Based V2X
11.1.1.1
Typical V2X Services
The operational LTE system is already capable of supporting ITS applications. The
V2X services can be broadly classified into three typical types, namely road safety,
traffic efficiency, and infotainment enhancement applications [1, 8]. Road safety
enhancements aim for reducing the risk of accidents and hence have to satisfy stringent reliability and latency specifications. Basically, the road safety services are short
messages and periodically broadcast from each vehicle to its neighbours within a particular geographic region.
As the second category of vehicular applications, traffic efficiency enhancements
aim for optimizing the platooning of vehicles by reducing traffic congestion. The
vehicles are required to collect sensed data and send them to the remote management
servers for route planning. Although the reliability and delay requirements related
to traffic efficiency enhancements are less strict than those of the safety enhance-
11 NOMA in Vehicular Communications
335
ments, it is still necessary to keep the packet loss and latency low in high-velocity
environments.
In contrast to the previous two categories, infotainment services include a range
of traditional and emerging Internet applications, such as popular content download
and dissemination, social networking, and Web browsing, with the goal of providing
an improved driving experience.
11.1.1.2
LTE-Based V2X Communication Modes
To support key applications in vehicular communications, the LTE system offers a
pair of communication modes.
LTE D2D for V2V/V2P: Although the original system does not support V2V/V2P
communications natively, LTE has been extended to support V2V/V2P direct communications based on a device-to-device (D2D) sidelink design through the so-called
PC5 interface [7]. The LTE D2D mode allows the terminals in close proximity of each
other to communicate directly without involving the base stations. As a benefit, the
end-to-end latency can be reduced for satisfying the associated QoS requirements.
At the time of writing, the LTE D2D mechanism is considered as the baseline for
PC5-based V2V/V2P communications [9]. However, in high user density scenarios,
the attributes of V2X services are different from those of the legacy LTE D2D communications, since V2X services are periodic or event-triggered. Hence, efficient
resource allocation has to be conceived for dense, high-mobility scenarios.
Cellular LTE for V2I/V2N: Cellular LTE refers to the common communication
mode between vehicles and infrastructure/network units. Specifically, there are two
main cellular LTE mechanisms, namely unicast and multicast. In the case of unicast, the vehicles are addressed individually. By contrast, in the multicast case, all
vehicles in the relevant area are collectively addressed. LTE supports high-quality
multicast and transmissions through the evolved multimedia broadcast multicast service (eMBMS) capabilities in the radio access network [10]. Compared to unicast
services, multicast offers the capability of geocasting the data to a set of users more
resource efficiently, although at the cost of longer delays due to the cumbersome
eMBMS session set-up, especially in the face of a heavy traffic load.
11.1.1.3
LTE-Challenges in Vehicular Scenarios
The LTE system is capable of providing a round-trip delay below 10 ms and a radio
access latency of less than 100 ms. It is based on orthogonal frequency-division multiple access (OFDMA) in the downlink and single-carrier frequency-division multiple access (SC-FDMA) in the uplink. It exhibits flexible resource allocation and
scheduling. The LTE system also relies on multiple-input multiple-output (MIMO)
techniques for improving the diversity and/or multiplexing gain of the previous generations, making LTE attractive in dynamic vehicular wireless propagation environments.
336
Y. Chen et al.
However, the ever-growing demands for vehicular communications increase the
tele-traffic congestion. Hence, it is desirable to achieve a high bandwidth efficiency,
massive connectivity, high reliability, and low latency in V2X communications.
One of the limitations of the LTE systems arises from the fact that LTE was designed
for supporting the user terminals sharing the wireless resources using orthogonal
multiple access (OMA), which can be potentially improved by NOMA schemes in
V2X communications.
In order to support advanced V2X services, given their stringent reliability and
latency requirements, multiple-input multiple-output (MIMO) techniques may be
invoked. Traditionally, MIMO schemes have been designed either for enhancing
the diversity gain by combating the channel fading (e.g. Alamouti code), or for
spatial multiplexing (e.g. Vertical Bell Laboratories Layered Space-Time, termed
VBLAST), albeit they are amalgamated by the multi-functional MIMO concept of
[11, 12]. To accommodate the ever-increasing demands of multimedia services and
applications, the massive MIMO concept emerged [13, 14]. Theoretically, massive
MIMO is able to reap all the benefits of conventional MIMO and offers abundant
degrees of freedom (DoFs). By exploiting the knowledge of the channel state information at the transmitter (CSIT), a massive antenna array becomes capable of simultaneously serving a large number of users by sharing its multiplexing gain among
them, while providing higher data rates and transmission reliability. Furthermore,
in contrast to shirt-pocket-sized handsets, the employment of large-scale MIMO
schemes becomes realistic in V2X scenarios, since multiple antennas can be realistically accommodated [15, 16].
However, massive MIMOs suffer from various problems, including the interantenna interference (IAI) and the high complexity of the receivers. It would be a
particularly costly process to acquire CSIT in frequency-division duplexing (FDD)
systems. Moreover, the hardware cost (e.g. a dedicated radio frequency (RF) chain
associated with each antenna) becomes excessive for large antenna arrays. In vehicular wireless communications, the gravest challenge is the hostile high-Doppler propagation imposed. For example, the dominant Doppler effect aggravates the intersubcarrier interference of orthogonal frequency-division multiplexing (OFDM), and
the strong line of sight (LoS) component of V2V channels would aggravate the
spatial correlation between antennas. Therefore, the direct applications of massive
MIMO in vehicular transmissions are deemed to be problematic, and another version
of massive antenna technology is required to be fit for LTE V2X communications.
11.1.2 The Applicability of NOMA to V2X Communications
To mitigate the probability of access collision in V2X environments, a range of novel
multiple access techniques has been proposed, such as sparse code multiple access
(SCMA), pattern division multiple access (PDMA), and non-orthogonal multiple
access (NOMA) to support higher bandwidth efficiency and massive connectivity [17,
18]. Among these techniques, NOMA exhibits an appealing low receiver complexity,
11 NOMA in Vehicular Communications
337
high bandwidth efficiency, and massive connectivity by allowing multiple users to
share the same channel resource via power domain multiplexing. Thus, NOMA is
considered to be a promising candidate for future wireless access [19]. To mitigate
the multiple access interference (MAI), multi-user detection (MUD) techniques such
as successive interference cancellation (SIC) [20] can be applied to the end-user
receivers for detecting the desired signals. Through power domain multiplexing at
the transmitter and SIC at the receivers, NOMA becomes capable of fully exploiting
its capacity region hence outperforming the OMA schemes [21].
The specific design aspects of NOMA schemes in cellular environments have
been discussed in [22–24]. Explicitly, in [22], the concept of basic NOMA with
SIC was introduced and its performance was compared to that of the traditional
orthogonal frequency-division multiple access (OFDMA) scheme through a systemlevel evaluation. A beneficial power allocation scheme was designed in [23] for
striking compelling tradeoffs between the user fairness and system throughput. Lv
et al. [24] studied a new cooperative NOMA transmission scheme and derived the
outage probability associated with fixed power allocation.
In vehicular environments, NOMA provides a new dimension for V2X services to
alleviate the access collisions, thereby improving the bandwidth efficiency as well as
supporting massive connectivity. The authors of [25] proposed an contention-based
uplink NOMA solution in order to reduce the control signalling overhead. In [26],
the NOMA concept was exploited to enhance the transmission of safety information,
which required low latency and high reliability within a dense vehicular communication network. The authors of [27] invoked the NOMA principle for boosting the
bandwidth efficiency of the infotainment applications in V2X services. In conclusion, NOMA is eminently applicable for supporting V2X services with enhanced
bandwidth efficiency and QoS support.
11.1.3 The Applicability of SM to V2X Communications
In recent years, spatial modulation (SM) [28] has grown in popularity, because in
contrast to the traditional MIMO configurations, it only activates a single transmit
antenna (Tx) at every transmission instance. Hence, it only requires a single-RF chain.
As a benefit, the inter-antenna interference (IAI) can be completely eliminated. Thus,
a reduced implementational cost and complexity are achieved [29, 30].
The basic idea of SM was initially derived from Chau and Yu’s work dating
back to 2001 [31], where the receiver decodes the signals transmitted from different
antennas. Then, a compelling SM-MIMO solution was proposed by Mesleh et al. in
[32]. Since then, SM has been extensively studied in the scenario of point-to-point
communications. In [28], the authors studied the channel capacity of the SM system
under the parlance of information-guided channel hopping (IGCH). It was shown that
IGCH provides better spectral efficiency than orthogonal space-time block coding
(OSTBC). In [33], the SM concept was studied by using a low-complexity two-stage
demodulator, and the potential advantages of SM-MIMO compared to the exist-
338
Y. Chen et al.
ing spatial-multiplexing and Alamouti schemes were shown. In [34], Jeganathan et
al. developed the maximum likelihood (ML)-optimum demodulator for SM-MIMO
and a range of performance improvements was shown compared to the suboptimal
demodulator introduced in [33].
The SM philosophy is that not only the classic quadrature amplitude modulation
(QAM) symbols but also the index of the active Tx (spatial constellation) convey
information for the sake of achieving bandwidth efficiency enhancements without
sacrificing the advantages of a single-RF stage. Consequently, SM was proposed to be
combined with massive MIMOs, yielding the novel concept of massive SM-MIMO,
where each UE still has one RF chain combined with a massive Tx configuration [35,
36, 38]. Due to the single-RF structure of SM, both the cost and the design complexity
of each user terminal remain similar to those of SM-MIMOs, while the data rates
can be boosted by conveying more information bits via employing a large Tx array.
More specifically, a large-scale multi-user SM-MIMO system was proposed in [35]
along with multi-user detection (MUD) schemes. In [36], Wang et al. proposed
an uplink transceiver scheme for massive SM-MIMO within frequency-selective
fading environments. The authors of [37] investigated the achievable uplink spectral
efficiency in a multi-cell massive SM-MIMO scenario, and [38] further investigated
the optimal number of Txs at the user equipment.
Indeed, a recent survey of SM can be found in [39]. In [40–43], SM and its
extensions were considered in vehicular environments. A differential SM scheme
was proposed for vehicle communications in [40], exhibiting robustness against timeselective fading and Doppler effects. Fu et al. [41] studied the bit error rate (BER)
performance of SM under a three-dimensional V2V channel model. Peppas et al. [42]
applied space shift keying (SSK) in inter-vehicular communications and derived a
closed-form expression for the pairwise error probability. In [43], the performance of
massive SM-MIMO over a spatio-temporally correlated Rician channel was analysed
under a high-speed railway scenario. Moreover, Cui and Fang have demonstrated that
by activating a single Tx, SM is capable of alleviating the channel correlation. In
conclusion, SM has become increasingly appealing for V2V systems.
11.1.4 NOMA-SM Tailored for Vehicular Communications
Let us continue by conceiving a novel transmission scheme, termed NOMA-SM,
by intrinsically amalgamating NOMA and SM in support of vehicular communications [27]. Specifically, in synergy with the inherent requirement of high bandwidth
efficiency, NOMA is invoked for non-orthogonally accessing all the resources combined with the single-RF benefits of SM. The bandwidth efficiency of the proposed
NOMA-SM scheme is further boosted by a massive Tx configuration.
Against this background, the key points of the proposed scheme are threefold:
firstly, the novel NOMA-SM concept is proposed and its link reliability is quantified. Secondly, the capacity of NOMA-SM is derived and verified by Monte Carlo
11 NOMA in Vehicular Communications
339
simulations. Thirdly, a pair of upper bounds on the capacity of NOMA-SM is formulated in closed form and a power allocation optimization is considered.
Explicitly, instead of simply combining a pair of popular techniques, their benefits are intrinsically amalgamated. By investigating the BER performance of NOMA
in comparison to different MIMO techniques and the bandwidth efficiency of SM
combined with distinct multiple access methods, NOMA and SM are shown to cooperatively improve V2V transmissions.
11.1.5 Outline of the Chapter
The rest of this chapter is organized as follows. In Sect. 11.2, the system model of
NOMA-SM is presented, while Sect. 11.3 provides the capacity analysis and mutual
information (MI) evaluation of NOMA-SM. Our capacity upper bound derivations
and power allocation problem are considered in Sect. 11.4. Simulation results and
discussions of the BER performance are provided in Sect. 11.5, together with the
numerical capacity analysis and power allocation optimization. In the final section,
we offer the main conclusions of this chapter and discuss some open problems as
well as a range of promising potential research directions. For convenience, we list
the most important notations here.
Notation: Uppercase and lowercase bold-faced letters indicate matrices and
vectors, respectively. (·)−1 , (·) H , det (·), and [·] p,q represent inverse, conjugatetranspose, determinant, and the entry in the pth row and q-column of a matrix,
respectively. E X {·} denotes the expectation on the random variable X . A ∈ C M×N
is a complex-element matrix with dimensions M × N , and I N is an N × N identity
matrix. |·| and (·)∗ imply the absolute value and the conjugate of a complex
scalar,
while · denotes the Euclidean norm of a vector. Finally, x ∼ C N μ, σ 2 indicates that the random variable x obeys a complex Gaussian distribution with mean
μ and variance σ 2 .
11.2 System Model
We consider a generic vehicular communication system, where the vehicle-toinfrastructure (V2I), V2V, and intra-vehicle transmissions are all included. As shown
in Fig. 11.1, a base station (BS) is located at the roadside while the vehicles V1 and
V2 are in motion. There is a mobile user U in V1 who requests to download a file
locally cached at the BS. Vehicle V2 also requests to download its own intended
signal from BS. We assume that V1 has also acquired the signal of V2 , as a result of
the first transmission phase, during which the messages of V1 and V2 are transmitted
simultaneously from the BS. For example, BS employs a NOMA technique to multiplex signals of V1 and V2 in the power domain. By involving the classical SIC, V1
extracts the signal of V2 in the spirit of cooperation. Another appropriate interpre-
340
Y. Chen et al.
Phase I
Phase II (exploiting NOMA-SM)
BS
V1
V2
U
Fig. 11.1 An illustration of the considered vehicular communication system, where NOMA-SM
is applied in Phase II
tation is related to the distribution of popular multimedia contents in VANET [44],
using peer-to-peer protocols for exchanging popular packets through V2V channels.
Therefore, as shown in Fig. 11.1, cooperative inter-vehicle transmission is constructed during the second phase to enhance the reception reliability. Specifically,
V1 forwards the desired signal to V2 for cooperatively enhancing the reception at V2 .
Furthermore, the second phase scenario can be generalized to various situations. For
example, user U can be a roadside unit, aiming for exchanging information with the
onboard unit of the vehicle V1 . While U may be a vehicle which is much closer to V1
than V2 . Similar to the concept in [45], a VANET is formed among these vehicles for
exchanging safety information, or for cooperatively distributing popular multimedia
contents within a geographical area of interest. In general, our model is valid in a
wide range of vehicular scenarios.
In the light of bandwidth scarcity, cognitive radio techniques can be exploited
in the second stage to opportunistically exploit the spectrum holes in the licensed
spectrum. For example, V1 may be permitted to share the cellular uplink, for which
the data traffic is typically lighter than for the downlink, hence resulting in potential
spectrum wastage [46]. Basically, underlay cognitive transmission is feasible without
traversing through the primary network. However, the interference imposed by V1
on the BS in the second stage should be carefully managed, albeit this is beyond
the scope of this article. Our main focus is on the second stage of the cooperative
transmission in Fig. 11.1, since the performance in the first phase can be analysed
similarly. Particularly, the NOMA-SM strategy is employed in the second stage for
both V1 –V2 and V1 –U links.
The schematic diagram of NOMA-SM operated in the second stage is presented
in Fig. 11.2, where V1 assigns distinct transmit power to V2 and U . The user access
is based on NOMA, combined with SM. Although there is the literature proposing
11 NOMA in Vehicular Communications
341
yU
Signal for U
Signal for V2
Superposition
coding
l
1
Spatial
demodulation
SIC of signal
for V2
Desired signal
detection
m
U
yV
V1
Spatial
demodulation
Desired signal
detection
V2
Fig. 11.2 The schematic diagram of the proposed NOMA-SM strategy
multi-user SM schemes [47, 48], we use a classical SM designed for point-to-point
transmission [28, 49] in vehicular environments. In what follows, we first elaborate
on the principles of the proposed NOMA-SM scheme. Then, our V2V channel model
is detailed.
11.2.1 The Principles of NOMA-SM
Let us assume that Nt , Nr , and Nu omnidirectional antennas are employed at V1 , V2 ,
and U , respectively. As illustrated in Fig. 11.2, the proposed NOMA-SM strategy is
applied both for the V1 –V2 and V1 –U links. At the transmitter V1 , two independent
bit streams are prepared for transmission. The bit stream for V2 is partitioned into
two parts: the first log2 (Nt ) bits are used for Tx activation, activating a specific Tx
index n t (n t ∈ {1, . . . , Nt }). The other log2 (M) bits destined for V2 are combined
coding.
with log2 (L) bits for U , employing superposition
√
√
Subsequently, the modulated symbol αγl + 1 − αχm is radiated from the
activated Tx n t , where γl and χm are intended for the in-car user U of V1 and for V2 ,
respectively, satisfying E{|γl |2 } = E{|χm |2 } = E s , where E s is the average energy
per transmission at V1 , while α is the power allocation factor. According to the
NOMA principle [23], the transmit power of the distant user in Fig. 11.2 must be
higher than that of the close-by user, that is (1 − α) E s > α E s . With this, 0 < α < 21
should be guaranteed since the in-car user has a good channel. As a result, a block of
identifies the active Tx n t and the superimposed
log2 (Nt M L) bits√unambiguously
√
complex symbol αγl + 1 − αχm transmitted from it. Hence, a NOMA-SM super
symbol can be expressed as
x = en t
√
αγl +
√
1 − αχm ,
where en t is the n t th column of the identity matrix I Nt , indicating that the n t th Tx
of V1 is activated, while the other (Nt − 1) Txs are deactivated. Furthermore, χm is
342
Y. Chen et al.
the mth symbol in the M-ary amplitude-phase modulation (APM) used for V1 –V2
transmission, while γl is the lth symbol in the L-ary APM for V1 –U transmission.
Considering the propagation inside the vehicle V1 , we assume that the in-car
user U experiences a frequency-flat Rayleigh channel. For example, the Txs of V1
are installed on the central column of the vehicular dashboard, while the receive
antennas (RAs) of U are placed behind the passenger front seat, without LoS from
V1 . In [50], this scenario has been shown to be well suited to characterize diffuse
scattering. Thus, we let G ∈ C Nr ×Nt denote the channel matrix between V1 and U ,
and assume that all entries of G are independent identically distributed (i.i.d), obeying
the distribution C N (0, 1). The signal vector received at U and V2 can be written as
yU = gn t
yV =
√
√
p0 hn t
αγl +
√
√
1 − αχm + wU ,
αγl +
√
1 − αχm + wV ,
(11.1)
(11.2)
respectively, where p0 represents the average power drop between V1 and V2 due
to the large-scale fading. Furthermore, gn t ∈ C Nu ×1 is the n t th column of G, representing the channel vector between U and the n t th Tx of V1 , while hn t ∈ C Nr ×1
is the n t th column of the V2V channel matrix H ∈ C Nr ×Nt , indicating the complex
fading envelope between V2 and the n t th Tx of V1 . Finally, w(·) denotes a complex
additive white Gaussian noise (AWGN) vector with a power spectrum density of σ02
per entry. For the inter-vehicle channel, the path loss is considerable in (11.2), while
it is neglected between the in-car user and the antenna array of V1 .
In our system, the transmitter and both receivers are assumed to have perfect
synchronization in both time and frequency. Full channel state information is assumed
to be available at receivers (i.e. CSIR). In principle, both V2 and U first have to detect
the signal destined for V2 , i.e. the activated Tx index n̂ t and the APM symbol χm̂ at
each particular time instant. The corresponding optimum maximum likelihood (ML)
detector is invoked at U and V2 according to
2
√
n̂ t , χm̂ = arg min yU − 1 − αgn t χm ,
n t ,m
2
n̂ t , χm̂ = arg min yV − p0 (1 − α)hn t χm .
n t ,m
(11.3)
(11.4)
After eliminating the interference imposed by nˆt , χm̂ on yU , U becomes capable
of performing another ML detection to acquire the desired signal γlˆ.
11 NOMA in Vehicular Communications
343
11.2.2 V2V Massive MIMO Channel Model
In contrast to the conventional fixed-to-mobile cellular radio systems, in V2V systems, both the transmitter and receiver are in motion and both are equipped with
low-elevation antennas, which will result in quite different propagation conditions.
Hence, a non-isotropic scattering V2V stochastic model was proposed in [51] for
characterizing a wide variety of V2V scenarios by adjusting relevant model parameters. In [41], a novel three-dimensional V2V geometry-based stochastic channel
was proposed for accurately capturing the effect of vehicular traffic density on the
channel.
In this article, we consider a spatio-temporally correlated Rician channel model
for characterizing our narrowband V2V massive MIMO channel, which has also
been exploited in [43] and [52]. We describe the underlying V2V channel as a matrix
of complex fading envelopes, i.e. H ∈ C Nt ×Nr , which can be expressed as
H=
K
H̄ +
K +1
1
H̃,
K +1
where K is the Rician factor, while H̄ is the fixed part related to the LoS component.
Furthermore, H̃ represents the variable part, whose entries are correlated complex
˜ , we assume that
Gaussian variables. Given H̃
= h p,q
p,q
E h̃ Rp,q h̃ Rp̂,q̂ = E h̃ Ip,q h̃ Ip̂,q̂ ,
E h̃ Rp,q h̃ Ip̂,q̂ = E h̃ Ip,q h̃ Rp̂,q̂ = 0,
where p, p̂ ∈ {1, . . . , Nr } and q, q̂ ∈ {1, . . . , Nt }. Explicitly, for each h̃ p,q , the autocorrelations of the real and imaginary parts are identical and the cross-correlations
between real and imaginary parts are equal to zero. Hence, the correlated channel
matrix can be described by the widely used Kronecker correlation model [53], which
is expressed as
1
1
H̃ = r2 Ĥ t2 .
Here, t ∈ C Nt ×Nt and r ∈ C Nr ×Nr are the correlation matrices at V1 and V2 ,
t
{1, . . . , Nt }, and
respectively, with the elements defined as [ t ]q,q̂ = σq,
q̂ for q, q̂ ∈
r
{1, . . . , Nr }. Furthermore, Ĥ is the independent Rayleigh
[ r ] p, p̂ = σ p,
p̂ for p, p̂ ∈
channel matrix whose entries are i.i.d complex Gaussian random variables, i.e.
[Ĥ] p,q = ĥ p,q ∼ C N (0, 1). Specifically, the correlation matrices t and r can
be determined according to a concrete model. Here, the exponential model of Loyka
|q−q̂ |
t
and
[54] is adopted, and the correlation matrix entries are formed as σq,
q̂ = κt
p−
p̂
|
|
r
, where κt and κr are the adjacent antenna correlation coefficients at
σ p,
p̂ = κr
V1 and V2 , respectively.
344
Y. Chen et al.
In order to mimic the influence of the V2V channel’s time-varying effects, we
take the temporal correlation into consideration, which is defined as
δ (τ ) = E Ĥ (t) Ĥ (t + τ ) ,
where τ is the sampling time. In [43], Jakes’ model is used for describing the temporal
correlation expressed as δ (τ ) = J0 (2π f D τ ), where f D is the maximum Doppler
frequency related to both the carrier frequency and the velocity of the terminal. For
simplicity of analysis, in the following, we omit the index τ . Observe that δ = 1
indicates that the underlying V2V channel is quasi-static, while δ < 1 is related to
a time-varying channel due to mobility. Naturally, both the spatial and temporal
correlations would affect the performance of the receivers.
11.3 Capacity Analysis of the NOMA-SM System
Recall that the proposed NOMA-SM transmission scheme relies on a pair of independent√
spaces: the
√ classical signal-domain, pertaining to the radiated superimposed
symbol αγl + 1 − αχm , and the Tx-domain, pertaining to the activated Tx index
n t . More specifically, the message intended for V2 is conveyed by both of the two
streams. While the message destined for U is only mapped to the classical signaldomain, superimposed with part of V2 ’s signal in the power domain. In what follows,
we investigate the capacity of the collaboration-aided vehicle V2 and the in-car user
U . Monte Carlo estimates are also provided for MI evaluation, followed by an illustrative example to augment the theoretical analysis.
11.3.1 Capacity Analysis of the Collaboration-Aided Vehicle
In the NOMA protocol, the transmit power assigned by V1 to the distant user V2 has
to be higher than that to the close-by user U . Then, the distant user directly detects
its signal, since the interference induced by the close-by user is lower and can thus
be treated as background noise. Considering that all Txs of V1 are activated with
the same probability for NOMA-SM, the instantaneous capacity pertaining to the
classical signal-domain of V2V transmission is given by
sig
C V = max I (χ ; yV |n t )
fχ
Nt
E p h 2 +σ 2
1
log2 α Es p0 hi 2 +σ0 2 .
= Nt
i=1
s
0
i
(11.5)
0
Observe that no practical modulation constellation is assumed, when performing
these capacity calculations. Since the channel capacity relates to the highest rate in
11 NOMA in Vehicular Communications
345
bits per channel use at which information can be sent with arbitrarily low probability
of error, in (11.5), we substitute χm by χ , which denotes a random input signal
alphabet with a distribution of f χ . On the other hand, the MI conveyed by the spatialdomain Tx-constellations can be written as
I (n t ; yV ) =
Nt
Pr ( yV | hi )
1
dyV ,
Pr ( yV | hi ) log2
Nt i=1
Pr (yV )
(11.6)
where Pr (yV |hi ) denotes the probability density function (PDF) of the channel
output yV received over the ith channel vector of H, given by
Pr (yV |hi ) =
π Nr
1
exp −yVH i−1 yV ,
det ( i )
where i = σ02 I + pE s hi hi H . As a result, the instantaneous capacity of V2 in the
NOMA-SM system is formulated as
sig
C V = C V + I (n t ; yV ) .
(11.7)
sig
Remark It is worth noting that in (11.5), C V is achievable where the optimum
input distribution for χ is Gaussian. In fact, this optimum input distribution is also
regarded as the optimum input distribution for a conventional SM system. This is a
common assumption in the majority of SM capacity-related contributions [28, 55–
57], effectively simplifying the analysis. Nevertheless, a fundamental weakness of
the Gaussian input assumption is that f χ affects both I (χ ; yV |n t ) and I (n t ; yV ).
Clearly, the Gaussian input distribution maximizes I (χ ; yV |n t ), but it is unclear
whether it maximizes I (n t ; yV ). In addition, the equiprobable activation of antennas
is a widely accepted assumption for SM-enabled systems, albeit this activation regime
cannot guarantee the optimal spatial design capable of achieving the capacity in
the Tx-domain. Actually, Liu et al. in [55] studied the optimal antenna activation
required for Tx-domain capacity maximization. Moreover, Basnayaka et al. [30]
have demonstrated that the Gaussian input does not achieve the upper limit of the
MI provided by an SM-aided system. As a further insight, although the MI conveyed
by the Tx-domain cannot be formulated as an analytical expression, we are inspired
to derive the capacity upper bound and to conceive the associated power allocation
optimization schemes, which will be addressed in Sect. 11.5.
11.3.2 Capacity Analysis of the In-Car User
In contrast to the receiver of V2 , the receiver of U can detect its own signal after
removing the interference imposed by V2 , as seen in Fig. 11.2. To demonstrate the
feasibility of this SIC procedure, we first deduce the maximum rate of which U can
346
Y. Chen et al.
detect the message of V2 . Specifically, the maximum rate for U detecting the message
related to the classical signal-domain of V2 is given by
V,sig
CU
Nt
E s gi 2 + σ02
1
=
log2
.
Nt i=1
α E s gi 2 + σ02
(11.8)
The MI associated with U detecting the information embedded in the Tx-constellation
of V2 can be written as
I (n t ; yU ) =
Nt
Pr ( yU | gi )
1
dyU ,
Pr ( yU | gi ) log2
Nt i=1
Pr (yU )
(11.9)
where Pr (yU |gi ) denotes the PDF of the channel output yU received over the ith
channel vector of G given by
Pr (yU |gi ) =
π Nu
1
exp −yUH i−1 yU ,
det (i )
where i = σ02 I + E s gi gi H . As a result, the instantaneous capacity for U detecting
the signal of V2 can be expressed as
V,sig
CUV = CU
+ I (n t ; yU ) .
(11.10)
It may be readily seen that CUV > C V is always satisfied, since gi 2 > p0 hi 2 ,
guaranteeing the success of SIC. Hence, the capacity of U detecting its own desired
signal is written as
CU = max I ( γ ; yU | n t , χ , G)
fγ
Nt
(11.11)
= N1t
log2 1 + ασE2s gi 2 ,
i=1
0
where γ denotes the random input signal variable related to the desired message
of U , with a distribution of f γ . The capacity for U detecting γ indeed becomes
achievable when the channel’s input distribution f γ is Gaussian.
11.3.3 Mutual Information
To appreciate the above theoretical analysis in terms of its relevance, next, we characterize the bandwidth efficiency of the proposed NOMA-SM. Assuming perfect
knowledge of the instantaneous channel state information at both receivers, the MI
achieved by V2 and U with the aid of practical APM constellations is evaluated
by the classical Monte Carlo method. For the collaboration-aided vehicle V2 , the MI
11 NOMA in Vehicular Communications
347
between a discrete signal input (n t , χm ) and the received signal yV can be formulated
as
yV |n t ,χm ,H)
I ( n t , χm ; yV | H) = En t ,χm ,yV log2 Pr( Pr(
yV |H)
yV |χm ,hi )
= Nt1M × Pr ( yV | χm , hi ) log2 Pr(Pr(y
dyV ,
V |H )
(11.12)
where the conditional probability Pr ( yV | χm , hi ) is expressed as
H
√
exp − yV − p0 (1 − α)hi χm
√
× i−1 yV − p0 (1 − α)hi χm ,
1
Pr( yV | χm , hi ) = π Nr det(
i)
with i = σ02 I + α p0 E s hi hiH . With regard to the in-car user U performing SIC first,
the MI between the information input (n t , χm ) and the received signal yU is given
by
I ( n t , χm ; yU | G) = Nt1M ×
(11.13)
yU |χm ,gi )
Pr ( yU | χm , gi ) log2 Pr(Pr(y
dyU ,
U |G )
where the conditional probability Pr ( yU | χm , gi ) is expressed as
1
Pr ( yU | χm , gi ) = π Nu det(
×
Hi ) −1
√
√
exp − yU − 1−αgi χm i yU − 1−αgi χm ,
with i = σ02 I + α E s gi giH .
Subsequently, the MI between the information input γl and the received signal yU
after perfect SIC is expressed as
I γl ; ỹU | gn t =
1
Nt L
Pr ( ỹU | γl , gi ) log2
1
Nt L
Pr( ỹU |γl ,gi )
dỹU ,
Pr( ỹU |γk ,g j )
(11.14)
k, j
√
where ỹU = yU − 1 − αgi χm with i ∈ {1, . . . , Nt } and m ∈ {1, . . . , M} denotes
the received vector after SIC. The conditional probability Pr ( ỹU | γl , gi ) is given by
ỹU − √αgi γl 2
Pr( ỹU | γl , gi ) =
.
N exp −
σ02
π σ02 u
1
348
Y. Chen et al.
11.3.4 An Illustration
In this part, a simulation-based study of our theoretical expressions is provided with
the aid of the MI attained by practical APM constellations. We set Nt = 64, Nr =
Nu = 2 for our MIMO configurations in conjunction with α = 0.1, E s = 1 and
p0 = 10−3 are given. The channel matrix H is generated according to Sect. 11.2.2,
where K = 0.2, κt = κr = 0.5, and δ = 1 are used. Each entry of G is identically and
independently generated according to a complex Gaussian distribution C N (0, 1).
In our Monte Carlo evaluations, the 16PSK signal constellation is chosen as the
APM for χm and γl ; hence, we have M = L = 16. The effective transmit signalto-noise ratio (SNR) at V1 is given by p0 E s /σ02 as the horizontal axis of Fig. 11.3.
Notice that the transmit-SNR cannot be readily interpreted physically, because it
relates the transmitter power to the noise power at the receiver, but its notion is
convenient to use in NOMA-aided scenarios. Given the effective transmit-SNR at
V1 as SNR = p0 E s /σ02 , the average receive-SNR at V2 can be computed as
SNRrV2 =
(1 − α) SNR
.
1 + αSNR
14
12
bps/Hz
10
I(nt; y V)
8
I(nt; y U)
C sig
V
6
C V,sig
U
4
m
; y V)
I(nt,
m
; y U)
CV
2
0
-10
I(nt,
CV
U
0
10
20
30
40
SNR [dB]
Fig. 11.3 Capacity and MI performance for Nt = 64, Nr = Nu = 2, M = L = 16, and α = 0.01.
sig
V,sig
Specifically, C V , CU , I (n t ; yV ), and I (n t ; yU ) are obtained from (11.5), (11.8), (11.6),
and (11.9), respectively. While I (n t , χm ; yV ) and I (n t , χm ; yU ) are generated from (11.12) and
(11.13), respectively, after averaging over multiple channel realizations. Finally, C V and CUV are
evaluated from (11.7) and (11.10), respectively
11 NOMA in Vehicular Communications
349
Similarly, the average receive-SNR at U for detecting the signal of V2 and that of
itself is respectively expressed as
(1−α)SNR
,
p0 +αSNR
αSNR
.
p0
SNRrU,V2 =
SNRrU =
Hence, the effective transmit-SNR at V1 is unambiguously related to the SNRs at each
receiver. Furthermore, we use SNR = p0 E s /σ02 in all of the subsequent performance
analyses.
The relevant results of Fig. 11.3 are discussed as follows.
sig
• The capacity of V2 gleaned from the signal-domain, that is C V obtained from
(11.5), increases steadily up to a saturation point as the SNR increases. By contrast,
V,sig
the capacity for U detecting the signal-domain destined for V2 , i.e. CU obtained
sig
from (11.8), is higher than C V in the low and moderate SNR domain. Clearly, a
successful detection of the signal-domain of V2 can be performed by U .
• The MI I (n t ; yV ) generated using (11.6) increases with the SNR and saturates at 6
bps/Hz, since the input entropy of the Tx-domain space is log2 (Nt ). By contrast,
the MI I (n t ; yU ) attained by (11.9) is as high as 6 bps/Hz across almost the entire
SNR range, since the channel quality of U is much higher than that of V2 , implying
that U can successfully detect the signal of V2 embedded in the Tx-domain.
• The capacity of V2 , i.e. C V , grows steadily as the SNR increases up to its saturation
at high SNRs, but it remains lower than CUV . Since C V is obtained by the summation
V,sig
sig
of C V and I (n t ; yV ), and CUV equals to the sum of CU and I (n t ; yU ). Naturally,
sig
V,sig
CUV > C V is satisfied, as CU
and I (n t ; yU ) are higher than C V and I (n t ; yV ),
respectively. Therefore, U can always perform successful SIC.
• The MI curves I (n t , χm ; yV ) and I (n t , χm ; yU ) are generated from (11.12) and
(11.13), respectively, after averaging over multiple channel realizations. It may be
observed that the simulated curve I (n t , χm ; yV ) matches the analytical capacity
C V quite closely upto an SNR of 5 dB, but beyond that I (n t , χm ; yV ) starts to
drift away from C V . By contrast, the drift of I (n t , χm ; yU ) from CUV remains
nearly unchanged. Both drifts are due to the fact that the MI attained with the aid
of practical APM modulation is upper bounded by the capacity, namely by the
maximum data rate related to the optimal input distribution.
11.4 Power Allocation Algorithms
It has been demonstrated that the MI conveyed by the Tx-domain cannot be readily
formulated as a closed-form expression, only by resorting to simulations. Thus, it
is very hard to perform an optimal power allocation for NOMA-SM. To circumvent
this problem, we first derive an upper bound of the NOMA-SM capacity. Then the
power allocation, which is capable of maximizing the capacity bound is considered,
leading to the optimal solution.
350
Y. Chen et al.
11.4.1 Problem Formulation
Theoretically, the instantaneous capacity of V2 in the NOMA-SM system can be
expressed as
C V = max I (n t , χ ; yV ) = max h (yV ) − h (yV |n t , χ ) ,
fχ
fχ
(11.15)
where h (·) denotes the differential entropy. The conditional differential entropy
h (yV |n t , χ ) in (11.15) is explicitly given by
h (yV |n t , χ ) =
Nt
1
log2 det π e p0 α E s hi hiH + σ02 I .
Nt i=1
To determine C V , we have to evaluate h (yV ), which requires the knowledge of the
distribution of yV . It may be readily seen that the MI I (n t , χ ; yV ) is maximized
if the vector variable yV has a Gaussian distribution. Thus, we assume that the
received vector yV has a Gaussian distribution, which is a zero-mean vector having
a covariance matrix presented as
E yV yVH = HEn t en t Eχ p0 (1 − α) χ χ ∗ enHt H H
+ HEn t en t Eγ p0 αγ γ ∗ enHt H H + σ02 I
Nt
1
ei e H p0 (1 − α) E s H H
=H
Nt i=1 i
Nt
1
+H
ei e H p0 α E s H H + σ02 I
Nt n =1 i
t
p0 E s
=
HH H + σ02 I.
Nt
An upper bound of h (yV ) can be formulated as
p0 E s
HH H + σ02 I
.
h (yV ) ≤ log2 det π e
Nt
Hence, we obtain an upper bound of C V which is written as
11 NOMA in Vehicular Communications
351
C V ≤ log2 det π e pN0 Et s HH H + σ02 I
Nt
log2 det π e p0 α E s hi hiH + σ02 I
− N1t
i=1
Nr
log2 pN0 Et s λ2j + σ02
=
(11.16)
j=1
− N1t
Nt
i=1
log2 p0 α E s hi 2 + σ02 = C VB1 ,
where λ j is the jth singular value of H with j ∈ {1, . . . , Nr }. Clearly, C VB1 has Nr
DoFs, and it is the same as the capacity of an (Nt × Nr )-element spatially multiplexed
MIMO system, subject to inter-user interference.
On the other hand, the MI of the Tx-domain has a natural upper bound written as
I (n t ; yV ) ≤ log2 (Nt ),
which corresponds to the maximum MI that can be conveyed by the Tx-domain of
the V2V transmission link. Now, another upper bound of C V may also be formulated
as
sig
C V ≤ C V + log2 (Nt )
Nt
(11.17)
E p h 2 +σ 2
= N1t
log2 α Es p0 hi 2 +σ0 2 + log2 (Nt ) = C VB2 .
i=1
s
0
i
0
Before proceeding, we provide a numerical illustration in order to evaluate both
of the upper bounds on the capacity of V2 . Figure 11.4 depicts C V and both upper
bounds of the NOMA-SM system in conjunction with Nt = 64, Nr = 2, M = 16, and
α = 0.1, which exhibit distinct approximations of C V within certain SNR regions.
The upper bound C VB1 gives a tight bound of C V at low SNRs, indicating that the
NOMA-SM capacity at V2 is almost the same as that of a spatially multiplexed
MIMO system of the same configuration in the presence of inter-user interference.
However, the MI embedded in the Tx-domain saturates as the SNR increases, which
is due to the fact that Nt is finite. Hence, at high SNRs, C VB2 is much tighter.
Based on the above observations, a refined upper bound on the capacity of V2 in
the NOMA-SM system is represented as
C VB = min C VB1 , C VB2 .
(11.18)
Considering the QoS of the two receivers from a practical perspective, we define
the minimum rate requirement of V2 and U as C̃ V and C̃U , respectively. The optimization problem constructed for maximizing the sum capacity with a power allocation
factor of α can be formulated as
P : max CU + C VB
α
352
Y. Chen et al.
25
CB1
V
20
CB2
V
CV
bps/Hz
15
10
5
0
-10
0
10
20
30
40
50
SNR [dB]
Fig. 11.4 Capacity and two upper bounds of the V2V transmission link with Nt = 64, Nr = 2,
M = 16, and α = 0.01. Specifically, C V , C VB1 , and C VB2 are evaluated from (11.7), (11.16), and
(11.17), respectively
⎧
⎪
⎨CU ≥ C̃U , (a)
s.t. C VB ≥ C̃ V , (b)
⎪
⎩
0 < α < 21 . (c)
(11.19)
11.4.2 The Proposed Power Allocation Algorithm
To solve the proposed optimization problem, we first express the derivatives of CU ,
C VB1 , and C VB2 with respect to α as
dCU
dα
B
dC V 1
dα
B
dC V 2
dα
=
1
Nt
Nt
i=1
= − N1t
= − N1t
E s gi 2
,
α E s gi 2 +σ02
Nt
n t =1
Nt
n t =1
E s p0 hi 2
,
α E s p0 hi 2 +σ02
(11.20)
E s p0 hi 2
,
α E s p0 hi 2 +σ02
respectively. Observe from (11.20) that CU is a monotonically increasing function of
α, given its positive derivative, while both C VB1 and C VB2 are decreasing ones. Thus,
when the constraint (c) of (11.19) is taken into account, there exist both minimum
11 NOMA in Vehicular Communications
353
and maximum capacities that V2 and U can achieve. Furthermore, to satisfy the
constraint (a) and (b), we have the following conditions for C̃U and C̃ V , respectively
1
,
0 < C̃U < CU α =
2
C VB
1
α=
< C̃ V < C VB (α = 0) .
2
Given the above conditions, we can rewrite the constraints of problem P in a compact
form as
g −1 C̃U < α < f −1 C̃ V ,
where g −1 (·) and f −1 (·) indicate the inverse function of CU and C VB , respectively. To
guarantee that the feasible set of problem P is non-empty, a further refined condition
for setting C̃ V is given by
1
< C̃ V < C VB α = g −1 C̃U .
C VB α =
2
2
2
since gn t > p0 hn t is always satisfied, the derivative of
Moreover,
CU + C VB can be guaranteed to have a positive value. Accordingly, the objective
function of problem P is a monotonically increasing function and can be maximized, when α reaches the upper bound of its feasible set. With C̃U and C̃ V being
appropriately set, we find that the upper bound of α’s feasible set is related to the
constraint (b) of (11.19), and the lower bound corresponds to the constraint (a) of
(11.19). Thus, the optimal solution of problem P is
P
= f −1 C̃ V .
αopt
(11.21)
This optimal solution implies that the amount of power allocated to V2 is ‘just’
sufficient to meet the minimum rate requirement C̃ V , while the remaining power is
used for U , aiming for maximizing its capacity. Nevertheless, we should notice that
there may exist some practical considerations, which require us to give high priority
to the V2V transmission link, such as those of safety applications, which have to be
served reliably. By contrast, the transmissions for in-car users are typically related to
infotainment applications, for example peer-to-peer video sharing and multimedia
advertisements [58]. Hence, it may be desirable to maximize the data rate of the V2V
link, while guaranteeing the minimum rate requirement of the in-car user. To this
end, we develop an alternative optimization problem formulated as
354
Y. Chen et al.
O : max C VB
α
⎧
⎪
⎨CU ≥ C̃U , (a)
s.t. C VB ≥ C̃ V , (b)
⎪
⎩
0 < α < 21 . (c)
(11.22)
Clearly, the objective function of (11.22) is a monotonically decreasing function of
α, and it is maximized, when the constraint (a) is inactive. Therefore, the optimal
solution of problem O can be written as
O
= g −1 C̃U .
αopt
(11.23)
So far, we have proposed a pair of power allocation schemes and analysed the
solvability of the optimization problems considered. Explicitly, we provided an algorithm for finding the optimal solution of each problem, which are summarized in
Table 11.1. The proposed algorithm essentially performs bounding through with the
aid of a bisection procedure, yielding globally optimal solutions at linearly increasing computational complexity [59]. In specific, the minimum rate requirements of
V2 and U are respectively set as
C̃U =
C̃ V =
C U (α =
2
C VB (α =
1
2
),
1
2
)+C VB
2
α = g −1 (C̃U )
(11.24)
,
for simplicity. Basically, both of the two power allocation optimization problems
satisfy realistic practical considerations, and the suitable one can be flexibly selected
based on the specific data priority of the distinct transmission links.
11.5 Simulations and Discussions
In this section, simulation results are provided for evaluating the performance of the
proposed NOMA-SM scheme. The system parameters are summarized as follows.
The MIMO configurations for the NOMA-SM system are set as Nt = 64, Nr =
Nu = 2. We fix p0 = 10−3 , or, equivalently, the path loss exponential is set to 3, and
the distance between V1 and V2 is assumed to be 10 m, which is typical for urban
environments, especially during rush hours.
11 NOMA in Vehicular Communications
355
Table 11.1 Power Allocation Algorithm
Power Allocation Algorithm for Problem P and Problem O
1. Initialization
Set tolerance 0 < ε ≪ 1. Calculate CU α = 21 and set C̃U = CU α = 21 /2.
2. Determine the lower bound of α and find the optimal solution of problem O
Set α L = 0 and αU = 21 .
While α L − αU > ε
U
Set α = α L +α
. Calculate CU (α).
2
If CU (α) − C̃U > 0
αU = α
Else
α L = α.
End
End
U
Set C̃ V = C VB α = 21 + C VB α = α L +α
/2.
2
O
O
U
The optimal solution to the problem O is obtained as αopt
and
. Calculate CU αopt
= α L +α
2
O
C VB αopt
.
3. Determine the upper bound of α and find the optimal solution of problem P
U
Set αmin = α L +α
and αmax = 21 .
2
While αmax − αmin > ε
max
. Calculate C VB (α).
Set α = αmin +α
2
If C VB (α) − C̃ V > 0
αmin = α
Else
αmax = α.
End
End
P
P
max
The optimal solution of the problem P is obtained as αopt
= αmin +α
. Calculate CU αopt
2
P
.
and C VB αopt
11.5.1 BER Results and Discussions
In this subsection, the BER performance of the NOMA-SM scheme is compared
to NOMA relying on the popular VBLAST technique, termed NOMA-VBLAST.
Specifically, we focus on the receiver performance of V2 . The effects of the Rician
K -factor, adjacent antenna correlation coefficient, temporal correlation, and power
allocation factor are all taken into consideration. The Rician K -factors are configured
as K = 2.186 and K = 0.2 for low and high vehicular traffic density, respectively
(see [51] for more details). More specifically, we consider a pair of references:
NOMA-VBLAST applied with 16QAM and Nt = 2, and NOMA-VBLAST adopted
QPSK and Nt = 4. The MIMO configuration of the references is the same as that
of NOMA-SM except for Nt . Besides, QPSK is applied for NOMA-SM. Thus, the
following BER comparisons are carried out for the same bandwidth efficiency of
8 bits per channel use (bpcu). The optimum ML detector described in (11.4) is
356
Y. Chen et al.
10
-1
-2
10
-3
10
-4
10
-5
10
-6
BER
10
NOMA-SM K=0.2
NOMA-SM K=2.168
NOMA-VBLAST K=0.2 16QAM
NOMA-VBLAST K=2.168 16QAM
NOMA-VBLAST K=0.2 QPSK
NOMA-VBLAST K=2.168 QPSK
25
30
35
40
45
SNR [dB]
Fig. 11.5 BER comparisons with different Rician K -factor when κt = κr = 0.2 and δ = 1 are
given, and the power allocation factor is fixed at α = 0.001, as evaluated by the Monte Carlo
simulation with 106 channel realizations
10
-1
NOMA-SM, t =0.2
NOMA-SM, t =0.8
10
NOMA-VBLAST, t =0.2,16QAM
NOMA-VBLAST, t =0.8,16QAM
-2
NOMA-VBLAST, t =0.2,QPSK
NOMA-VBLAST, t =0.8,QPSK
-3
10
-4
10
-5
10
-6
BER
10
25
30
35
40
45
SNR [dB]
Fig. 11.6 BER comparisons with different adjacent antenna correlation coefficient at V1 , i.e. κt ,
when K = 0.2, κr = 0.5, and δ = 1 are given, and the power allocation factor is fixed at α = 0.001,
as evaluated by the Monte Carlo simulation with 106 channel realizations
employed at V2 in both schemes. All simulation results of this subsection are obtained
through a Monte Carlo method.
In Fig. 11.5, we show the BER performance for different Rician K -factor. It is
observed that NOMA-SM outperforms the benchmark especially in the high SNR
regime. Additionally, the increase of K imposes a more dominant degradation on
11 NOMA in Vehicular Communications
357
10-1
BER
10-2
NOMA-SM =1
NOMA-SM =0.9
NOMA-VBLAST
NOMA-VBLAST
NOMA-VBLAST
NOMA-VBLAST
=1 16QAM
=0.9 16QAM
=1 QPSK
=0.9 QPSK
10-3
10-4
10-5
25
30
35
40
45
SNR [dB]
Fig. 11.7 BER comparisons with different temporal correlation coefficient δ when K = 0.2 and
κt = κr = 0.5 are given, and the power allocation factor is fixed at α = 0.001, as evaluated by the
Monte Carlo simulation with 106 channel realizations
-1
10
-2
10
-3
=0.01
NOMA-SM
=0.001
NOMA-SM
=0.0001
NOMA-VBLAST
=0.01 16QAM
NOMA-VBLAST
=0.001 16QAM
NOMA-VBLAST
=0.0001 16QAM
NOMA-VBLAST
=0.01 QPSK
NOMA-VBLAST
=0.001 QPSK
NOMA-VBLAST
=0.0001 QPSK
BER
10
NOMA-SM
-4
10
-5
10
-6
10
25
30
35
40
45
SNR [dB]
Fig. 11.8 BER comparisons with different power allocation factor α when K = 0.2, κt = κr = 0.5,
and δ = 1 are given, as evaluated by the Monte Carlo simulation with 106 channel realizations
358
Y. Chen et al.
both of the NOMA-VBLAST schemes, which rely more vitally on the presence of
rich non-LoS scattering. This phenomenon can be explained as follows. The higher
Rician factor K represents a stronger LoS component, which increases the spatial
correlation among the adjacent channel paths. For NOMA-VBLAST schemes, the
multiple-stream information is conveyed with the aid of multiple DoFs. By contrast,
for NOMA-SM, although the more severe spatial correlation of the LoS scenario
makes it difficult to determine the index of the activated Tx, the remaining information
related to the APM signal-domain is transmitted over a single DoF; hence, it is less
susceptible to spatial correlation.
Figure 11.6 investigates the BER results associated with different adjacent Txcorrelation coefficients at V1 . Compared to κt = 0.8, κt = 0.2 represents an insignificant spatial correlation. Again, observe from Fig. 11.6 that NOMA-SM is less susceptible to spatial correlation. This phenomenon can be interpreted similarly to the trend
of Fig. 11.5. Besides, for these two figures, we notice that NOMA-SM temporarily
loses its advantage over NOMA-VBLAST adopted QPSK and Nt = 4 within moderate SNR regime. This observation results from the superiority that QPSK brings
to NOMA-VBLAST compared to 16QAM. However, as the increment of SNR,
NOMA-SM achieves its dominance in terms of higher diversity gain.
Below, we investigate the impact of the V2V channel’s time-varying nature.
Observe from Fig. 11.7 that compared to the performance of no time-varying effect
associated with δ = 1, the BER has been substantially degraded in all schemes
for δ = 0.9. Although a perfect channel estimation procedure is assumed for the
receivers, the estimated channel coefficients used for ML detection becomes partially
outdated due to the channel’s time-varying nature, hence resulting in a degraded BER
performance. Nevertheless, the proposed NOMA-SM scheme maintains its advantage over the reference within the medium and high SNR regime, regardless of the
grade of temporal correlation.
Figure 11.8 shows the BER performance associated with different α values. For all
schemes, the lower α values exhibit a better detection performance, since less power is
allocated to U and hence V2 suffers from a lower inter-user interference. More importantly, we observe that NOMA-SM consistently outperforms the NOMA-VBLAST
scheme applied with 16PSK and Nt = 2. For the cases of α = 0.001 and α = 0.0001,
though NOMA-VBLAST adopted QPSK and Nt = 4 holds a dominance within
the moderate SNR regime, NOMA-SM keeps improving its performance with the
increase of SNR and outperform the references in terms of higher diversity gain. By
jointly considering the above observations, we conclude that NOMA-SM constitutes
a potent amalgam.
11.5.2 Capacity Results and Discussions
Below, we evaluate the capacity of the NOMA-SM system associated with different
power allocation strategies. All results presented in this subsection are obtained
by averaging the instantaneous capacities over multiple channel realizations. In
11 NOMA in Vehicular Communications
359
particular, we fix K = 0.2, κt = κr = 0.5, and δ = 1 unless otherwise stated. For
benchmarking, we use an OMA-SM system, where V1 transmits messages to V2
using SM in the first slot. Then, V1 sends messages through the previously activated
antenna to U , without activating another antenna. This OMA-SM model constitutes
a fair reference for the NOMA-SM system, since the signal intended for V2 is conveyed by both the APM signal- and Tx-domain, whereas the signal destined for U is
only embedded in the classical signal-domain. The distinctive feature of OMA-SM
is that data transmissions destined for V1 –V2 and V1 –U are operated in an orthogonal
time division way within the classical APM signal-domain. Accordingly, the capacity upper bound for V2 and the capacity for U in the OMA-SM system are expressed
as
CVB = min CVB1 , CVB2 ,
Nt
log2 1 + ασE2s gi 2 ,
CU = 2N1 t
i=1
(11.25)
0
respectively, where
CVB1 = 2N1 t
Nt
i=1
log2 1 +
CVB2 = 21 log2 det I +
(1−α)E s p0
hi 2
σ02
(1−α)E s p0
HH H
σ02 Nt
+ 21 log2 (Nt ),
(11.26)
.
Let us first check the capacity associated with a fixed power allocation, that is
α = 0.01. Figure 11.9 depicts the capacity of V2 and U , as well as the sum capacity
versus SNR for both NOMA-SM and OMA-SM. Compared to OMA-SM, NOMASM provides substantial capacity gains both for the collaboration-aided vehicle V2
and for the in-car user U and accordingly obtains a significant sum capacity enhancement. Specifically, the capacity CU has been beneficially boosted by the proposed
scheme, about twice as high as that of OMA-SM. Since the APM signal-domain of
the proposed scheme is combined with a NOMA strategy, each user accesses the
channel resources via power domain multiplexing.
Subsequently, we investigate the efficiency of the proposed power allocation optimization. Specifically, the power allocation optimization denoted by Q is considered
for OMA-SM, which is formulated as
Q : max CU + CVB
α⎧
⎨ CU ≥ C˜U ,
s.t. CVB ≥ C˜VB ,
⎩
0 < α < 1.
(11.27)
For simplicity, the minimum rate requirements of V2 and U are set to C˜U = C U (α2 = 1)
C B (α = 0)
, which respectively correspond to the lower bound and upper
and C˜B = V
V
2
360
Y. Chen et al.
35
CU+CB
in NOMA-SM
V
CB
in NOMA-SM
V
30
CU in NOMA-SM
CU+CB
in OMA-SM
V
25
CB
in OMA-SM
V
bps/Hz
CU in OMA-SM
20
15
10
5
0
0
5
10
15
20
25
30
35
40
45
SNR [dB]
Fig. 11.9 Capacity of V2 and U , or the sum capacity versus SNR for the NOMA-SM and OMA-SM
scheme with a fixed power allocation factor, i.e. α = 0.01. Specifically, C VB and CU in NOMA-SM
are evaluated from (11.18) and (11.11), while C VB and CU in OMA-SM are obtained from (11.25)
bound of α’s feasible set. Then, a full-search algorithm is applied for OMA-SM
within the feasible set.
Figure 11.10 illustrates the capacity of V2 and U for NOMA-SM with optimization
P or O, where the QoS of the collaboration-aided vehicle V2 and the in-car user U ,
i.e. C̃ VB and C̃U , are also plotted for reference. It can be observed that C VB always meets
the requirement of C̃ VB with the aid of the optimization P, and CU associated with
the optimization O exactly meets the QoS C̃U . This observation is in accordance with
the foregoing analysis, which indicates that the optimization P intends to maximize
CU , while maintaining the QoS C̃ VB for V2V transmission, whereas the optimization
O aims for maximizing C VB while guaranteeing the minimum rate requirement C̃U
for the in-car user. Thus, we find that the optimized CU of P is higher than that of
O, whereas the optimized C VB of O outperforms that of P. Accordingly, the more
appropriate optimization scheme can be readily selected based on the data priority
of distinct transmission links.
Figure 11.11 compares the results of the optimization Q to that of P and O. Let
us contrast P and Q first. Clearly, both C VB and CU in NOMA-SM with optimization
P have been remarkably improved, demonstrating that the NOMA strategy offers
a bandwidth efficiency improvement. By considering the results of O and Q in
Fig. 11.11, we find that CU of NOMA-SM associated with optimization O is tightly
lower bounded by that of OMA-SM associated with optimization Q, and C VB with
O provides a substantial gain, achieving nearly twice that of Q.
11 NOMA in Vehicular Communications
361
20
18
16
14
bps/Hz
12
10
8
CU in NOMA-SM with P
CB
in NOMA-SM with P
V
6
CU in NOMA-SM with O
CB
in NOMA-SM with O
V
4
QoS for CU
2
0
QoS for CB
V
0
5
10
15
20
25
30
35
40
45
SNR [dB]
Fig. 11.10 Capacity of V2 and U , or the respective QoS versus SNR for NOMA-SM with power
allocation optimization P or O . Specifically, C VB and CU in NOMA-SM with P or O are evaluated
with the aid of the algorithm in Table 11.1. The QoS for C VB and CU , i.e. C̃ V and C̃U , are set
according to (11.24)
20
18
16
14
bps/Hz
12
10
8
CU in NOMA-SM with P
6
CB
in NOMA-SM with P
V
4
CU in NOMA-SM with O
2
CU in OMA-SM with Q
CB
in NOMA-SM with O
V
CB
in OMA-SM with Q
V
0
0
5
10
15
20
25
30
35
40
45
SNR [dB]
Fig. 11.11 Capacity of V2 and U versus SNR for NOMA-SM with power allocation optimization
P or O , and OMA-SM with power allocation optimization Q , respectively. Specifically, C VB and
CU in NOMA-SM with P or O are evaluated with the aid of the algorithm in Table 11.1. While
C VB and CU in OMA-SM with Q are obtained from a full-search algorithm
362
Y. Chen et al.
35
CU+CB
in NOMA-SM with P
V
30
CU+CB
in NOMA-SM with O
V
CU+CB
in OMA-SM with Q
V
bps/Hz
25
20
15
10
5
0
5
10
15
20
25
30
35
40
45
SNR [dB]
Fig. 11.12 Sum capacity versus SNR for NOMA-SM with power allocation optimization P or
O , and OMA-SM with power allocation optimization Q , respectively. Specifically, C VB and CU in
NOMA-SM with P or O are evaluated with the aid of the algorithm in Table 11.1. While C VB and
CU in OMA-SM with Q are obtained from a full-search algorithm
Furthermore, it can be observed from Fig. 11.12 that the NOMA-SM systems
achieve higher sum capacity than OMA-SM. Specifically, optimization P provides
higher capacity gain than O, since P aims for maximizing the data rate of the incar user U , which experiences a much better channel than the collaboration-aided
vehicle V2 .
11.6 Chapter Summary and Future Outlook
In this chapter, we introduce NOMA and SM techniques into V2X scenarios in
order to support high bandwidth efficiency and enhanced link reliability. The BER
performance of the new NOMA-SM transmission strategy has been investigated
with the impact of the Rician K -factor, spatial correlation of antenna array, timevarying effect of the V2V channel, and the power allocation factor being discussed.
Compared to NOMA relying on VBLAST, NOMA-SM has been demonstrated to
exhibit improved robustness against the spatial and temporal effects of the V2V
channel. By analysing the capacity and deriving analytical upper bounds in closed
form, a pair of power allocation optimization schemes have been formulated for
NOMA-SM. The optimal solutions have also been shown to be achievable with the aid
of the proposed power allocation algorithm. Our numerical results have verified that
with the aid of an appropriate power allocation, NOMA-SM is capable of satisfying
the QoS support of a low priority flow, whilst maximizing the throughput of the
11 NOMA in Vehicular Communications
363
high priority flow. In summary, NOMA-SM has been demonstrated to cooperatively
improve the link reliability and bandwidth efficiency of V2V transmissions.
Nonetheless, several open issues still need to be carefully addressed before NOMA
can be practically exploited in vehicular environments. Here, we discuss two potential
research topics in this field as examples.
Parallel Interference Cancellation-Aided NOMA: There is a much broader range
of V2X applications to be considered in VANETs, especially within the automated
driving field, whose characteristics are more stringent, as captured by the ultrareliable low-latency constraints. The traditional NOMA schemes use the classic SIC
technique, where a high received signal power difference is preferred. However, this
condition cannot always be guaranteed, especially in a traffic jam, where all cars
tend to have similar channel conditions. Furthermore, SIC receivers were reported
to exhibit an error floor in high-order modulation modes due to error propagation
across the cancellation stages [60]. By contrast, parallel interference cancellation
(PIC) does not require any specific detection order and all users reconstruct the
signals of all the other users in parallel. Then, they subtract the reconstructed signals
from the composite signal. Hence, PIC outperforms SIC when the received signal
powers for all users are similar. With the advent of a PIC receiver, NOMA is expected
to possess enhanced transmission reliability as well as better applicability to highlyloaded vehicular scenarios. In a nutshell, the performance analysis of NOMA with
PIC should be addressed in our further research.
NOMA for Cognitive V2X: The large amount of data generated by the vehicles
might impose excessive traffic demands on traditional cellular traffic. Hence, cognitive NOMA principles can be conceived for delay-tolerant vehicular communications services to opportunistically access the channels originally occupied by the
cellular users. Given a dedicated spectral band, cellular users and vehicles can be
regarded as primary users and secondary users, respectively. The vehicles would
only be permitted to access the channel when the services of cellular users are not
affected. In contrast to the traditional cognitive radio scheme, both of the power control and resource allocation need to be designed elaborately, and efficient transmission
schemes accommodating both primary and secondary users should be proposed and
analysed carefully.
References
1. 3GPP TS 22.185, Service requirements for V2X services, Feb 2016
2. D. Jiang, L. Delgrossi, IEEE 802.11p: towards an international standard for wireless access
in vehicular environments, in Proceedings VTC Spring 2008—IEEE Vehicular Technology
Conference (Singapore, May 2008), pp. 2036–2040
3. S.H. Sun, J.L. Hu, Y. Peng, X.M. Pan, L. Zhao, J.Y. Fang, Support for vehicle-to-everything
services based on LTE. IEEE Wirel. Commun. 23(3), 4–8, June 2016
4. M. Amadeo, C. Campolo, A. Molinaro, Enhancing IEEE 802.11p/WAVE to provide infotainment applications in VANETs. Elsevier Ad Hoc Netw. 10(2), 253–269 (2012)
5. S. Chen et al., Vehicle-to-everything (v2x) services supported by LTE-based systems and 5G.
IEEE Commun. Stand. Mag. 1(2), 70–76 (2017)
364
Y. Chen et al.
6. S. Chen, J. Hu, Y. Shi, L. Zhao, LTE-V: a TD-LTE-based V2X solution for future vehicular
network. IEEE Internet Things J. 3(6), 997–1005 (2016)
7. 3GPP RP-152293, New WI proposal: support for V2V services based on LTE sidelink, Dec
2015
8. G. Araniti, C. Campolo, M. Condoluci, A. Iera, A. Molinaro, LTE for vehicular networking: a
survey. IEEE Commun. Mag. 51(5), 148–157 (2013)
9. 3GPP TS 23.285, v.14.1.0, Architecture enhancements for V2X services, Dec 2016
10. 3GPP, TS 36.440, General aspects and principles for interfaces supporting multimedia broadcast multicast service (MBMS) within E-UTRAN, Rel. 11, Sept 2012
11. L. Hanzo, O. Alamri, M. El-Hajjar, N. Wu, Near-capacity multi-functional MIMO systems:
sphere-packing, iterative detection and cooperation (Wiley, New York, NY, USA, 2009)
12. L. Wang, R. Li, C. Cao, G.L. Stüber, SNR analysis of time reversal signaling on target and
unintended receivers in distributed transmission. IEEE Trans. Commun. 64(5), 2176–2191
(2016)
13. E.G. Larsson, O. Edfors, F. Tufvesson, T.L. Marzetta, Massive MIMO for next generation
wireless systems. IEEE Commun. Mag. 52(2), 186–195 (2014)
14. Y. Wu, R. Schober, D.W.K. Ng, C. Xiao, G. Caire, Secure massive MIMO transmission with
an active eavesdropper. IEEE Trans. Inf. Theory 62(7), 3880–3900 (2016)
15. R. Zhang, Z. Zhong, J. Zhao, B. Li, K. Wang, Channel measurement and packet-level modeling
for V2I spatial multiplexing uplinks using massive MIMO. IEEE Trans. Veh. Technol. 65(10),
7831–7843 (2016)
16. P. Harris et al., Performance characterization of a real-time massive MIMO system with LOS
mobile channels. IEEE J. Select. Areas Commun. 35(6), 1244–1253 (2017)
17. Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, K. Higuchi, Non-orthogonal
multiple access (NOMA) for cellular future radio access, in Proceedings IEEE 77th Vehicular
Technology Conference (Dresden, Germany, June 2013), pp. 1–5
18. Y. Chen, L. Wang, B. Jiao, Cooperative multicast non-orthogonal multiple access in cognitive
radio, in Proceedings 2017 IEEE ICC (Paris, France, May 2017), pp. 1–6
19. L. Dai, B. Wang, Y. Yuan, S. Han, I. Chih-Lin, Z. Wang, Non-orthogonal multiple access
for 5G: solutions, challenges, opportunities, and future research trends. IEEE Commun. Mag.
53(9), 74–81 (2015)
20. J. Liberti, S. Moshavi, P. Zablocky, Successive interference cancellation. U.S. Patent 8670418
B2, 11 Mar 2014
21. Q. Sun, S. Han, I. Chih-Lin, Z. Pan, On the ergodic capacity of MIMO NOMA systems. IEEE
Wirel. Commun. Lett. 4(4), 405–408, Aug 2015
22. Y. Saito, A. Benjebbour, Y. Kishiyama, T. Nakamura, System-level performance evaluation of
downlink non-orthogonal multiple access (NOMA), in Proceedings IEEE 24th International
Symposium Personal Indoor and Mobile Radio Communications (PIMRC) (London, UK, Sept
2013), pp. 611–615
23. Z. Yang, Z. Ding, P. Fan, N. Al-Dhahir, A general power allocation scheme to guarantee quality
of service in downlink and uplink NOMA systems. IEEE Trans. Wirel. Commun. 15(11), 7244–
7257 (2016)
24. L. Lv, J. Chen, Q. Ni, Cooperative non-orthogonal multiple access in cognitive radio. IEEE
Commun. Lett. 20(10), 2059–2062 (2016)
25. K. Au et al., Uplink contention based SCMA for 5G radio access, in Proceedings IEEE GLOBECOM Workshop (Austin, TX, Dec, 2014), pp. 900–905
26. B. Di, L. Song, Y. Li, G.Y. Li, Non-orthogonal multiple access for high-reliable and low-latency
V2X communications in 5G systems. IEEE J. Select. Areas Commun. 35(10), 2383–2397
(2017)
27. Y. Chen, L. Wang, Y. Ai, B. Jiao, L. Hanzo, Performance analysis of NOMA-SM in vehicleto-vehicle massive MIMO. IEEE J. Select. Areas Commun. 35(12), 2653–2666 (2017)
28. Y. Yang, B. Jiao, Information-guided channel-hopping for high data rate wireless communication. IEEE Commun. Lett. 12(4), 225–227 (2008)
11 NOMA in Vehicular Communications
365
29. L. He, J. Wang, J. Song, L. Hanzo, On the multi-user, multi-cell massive spatial modulation
uplink: how many antennas for each user? IEEE Trans. Wirel. Commun. 16(3), 1437–1451
(2017)
30. D.A. Basnayaka, M. Di Renzo, H. Haas, Massive but few active MIMO. IEEE Trans. Veh.
Technol. 65(9), 6861–6877 (2016)
31. Y. Chau, S.-H. Yu, Space modulation on wireless fading channels, in Proceedings IEEE Vehicular Technology Conference—Fall, vol. 3 (Atlantic City, NJ, Oct 2001), pp. 1668–11671
32. R. Mesleh, H. Haas, C. W. Ahn, S. Yun, Spatial modulation—a new low complexity spectral efficiency enhancing technique, in Proceedings 2006 First International Conference on
Communications and Networking in China (Beijing, 2006), pp. 1–5
33. R.Y. Mesleh, H. Haas, S. Sinanovic, C.W. Ahn, S. Yun, Spatial modulation. IEEE Trans. Veh.
Technol. 57(4), 2228–2241 (2008)
34. J. Jeganathan, A. Ghrayeb, L. Szczecinski, Spatial modulation: optimal detection and performance analysis. IEEE Commun. Lett. 12(8), 545–547 (2008)
35. T.L. Narasimhan, P. Raviteja, A. Chockalingam, Large-scale multiuser SM-MIMO versus massive MIMO, in Proceedings, Information Theory and Applications Workshop (ITA) (San Diego,
CA, 2014), pp. 1–9
36. S. Wang, Y. Li, M. Zhao, J. Wang, Energy-efficient and low-complexity uplink transceiver for
massive spatial modulation MIMO. IEEE Trans. Veh. Technol. 64(10), 4617–4632 (2015)
37. P. Patcharamaneepakorn et al., Spectral, energy, and economic efficiency of 5G multicell massive MIMO systems with generalized spatial modulation. IEEE Trans. Veh. Technol. 65(12),
9715–9731 (2016)
38. L. He, J. Wang, J. Song, On massive spatial modulation MIMO: spectral efficiency analysis
and optimal system design, in Proceedings IEEE GLOBECOM (2016), pp. 1–6
39. M. Di Renzo, H. Haas, A. Ghrayeb, S. Sugiura, L. Hanzo, Spatial modulation for generalized
MIMO: challenges, opportunities, and implementation. Proc. IEEE 102(1), 56–103 (2014)
40. M. Zhang, X. Cheng, L.Q. Yang, Differential spatial modulation in V2X, in Proceedings 2015
9th European Conference on Antennas and Propagation (EuCAP) (Lisbon, Apr 2015), pp. 1–5
41. Y. Fu et al., BER performance of spatial modulation systems under 3-D V2V MIMO channel
models. IEEE Trans. Veh. Technol. 65(7), 5725–5730 (2016)
42. K.P. Peppas, P.S. Bithas, G.P. Efthymoglou, A.G. Kanatas, Space shift keying transmission for
intervehicular communications. IEEE Trans. Intell. Transp. Syst. 17(12), 3635–3640 (2016)
43. Y. Cui, X. Fang, Performance analysis of massive spatial modulation MIMO in high-speed
railway. IEEE Trans. Veh. Technol. 65(11), 8925–8932 (2016)
44. T. Wang, L. Song, Z. Han, Coalitional graph games for popular content distribution in cognitive
radio VANETs. IEEE Trans. Veh. Technol. 62(8), 4010–4019 (2013)
45. H. Ilhan, I. Altunbas, M. Uysal, Optimized amplify-and-forward relaying for vehicular ad-hoc
networks, in Proceedings IEEE Vehicular Technology Conference (Calgary, BC, Sept 2008),
pp. 1–5
46. K. Huang, V.K.N. Lau, Y. Chen, Spectrum sharing between cellular and mobile ad hoc networks:
transmission-capacity trade-off. IEEE J. Select. Areas Commun. 27(7), 1256–1267 (2009)
47. T. Lakshmi Narasimhan, P. Raviteja, A. Chockalingam, Generalized spatial modulation in
large-scale multiuser MIMO systems. IEEE Trans. Wirel. Commun. 14(7), 3764–3779 (2015)
48. Y. Chen, L. Wang, Z. Zhao, M. Ma, B. Jiao, Secure multiuser MIMO downlink transmission
via precoding-aided spatial modulation. IEEE Commun. Lett. 20(6), 1116–1119 (2016)
49. L. Wang, S. Bashar, Y. Wei, R. Li, Secrecy enhancement analysis against unknown eavesdropping in spatial modulation. IEEE Commun. Lett. 19(8), 1351–1354 (2015)
50. O. Delangre, S. Van Roy, P. De Doncker, M. Lienard, P. Degauque, Modeling in-vehicle wideband wireless channels using reverberation chamber theory, in Proceedings 2007 IEEE 66th
Vehicular Technology Conference (Baltimore, MD, Oct 2007), pp. 2149–2153
51. X. Cheng, C.X. Wang, D.I. Laurenson, S. Salous, A.V. Vasilakos, An adaptive geometry-based
stochastic model for non-isotropic MIMO mobile-to-mobile channels. IEEE Trans. Wirel.
Commun. 8(9), 4824–4835 (2009)
366
Y. Chen et al.
52. M. Koca, H. Sari, Performance analysis of spatial modulation over correlated fading channels,
in Proceedings, IEEE Vehicular Technology Conference (VTC Fall) (Quebec City, QC, Sept
2012), pp. 1–5
53. J.P. Kermoal, L. Schumacher, K.I. Pedersen, P.E. Mogensen, F. Frederiksen, A stochastic
MIMO radio channel model with experimental validation. IEEE J. Select. Areas Commun.
20(6), 1211–1226 (2002)
54. S.L. Loyka, Channel capacity of MIMO architecture using the exponential correlation matrix.
IEEE Commun. Lett. 5(9), 369–371 (2001)
55. C. Liu, M. Ma, Y. Yang, B. Jiao, Optimal spatial-domain design for spatial modulation capacity
maximization. IEEE Commun. Lett. 20(6), 1092–1095 (2016)
56. X. Guan, Y. Cai, W. Yang, On the mutual information and precoding for spatial modulation
with finite alphabet. IEEE Wirel. Commun. Lett. 2(4), 383–386 (2013)
57. Z. An, J. Wang, J. Wang, S. Huang, J. Song, Mutual information analysis on spatial modulation
multiple antenna system. IEEE Trans. Commun. 63(3), 826–843 (2015)
58. Y. Toor, P. Muhlethaler, A. Laouiti, A.D. La Fortelle, Vehicle Ad Hoc networks: applications
and related technical issues, IEEE Commun. Surv. Tutor. 10(3), 74–88, Third Quarter (2008)
59. S. Boyd, L. Vandenberghe, Convex Optimization (Cambridge Univ. Press, Cambridge, U.K.,
2004)
60. L. Hanzo, L.-L. Yang, E.-L. Kuan, K. Yen. Single-and multi-carrier DS-CDMA: multi-user
detection, space-time spreading, synchronisation, standards and networking (Wiley, 2003)
Part III
NOMA in Code and Other Domains
Chapter 12
Sparse Code Multiple Access (SCMA)
Zheng Ma and Jinchen Bao
12.1 General Description
Overloaded systems, in which the number of users is greater than the dimension of
signal-space, are of practical interest in bandwidth-efficient multi-user communications. One kind of such systems is sparse code multiple access (SCMA), which
is a promising code-domain non-orthogonal multiple access technique to address
the challenges for the fifth-generation (5G) mobile networks [1–4]. Non-orthogonal
multiple access has the potential to accommodate more users with limited resources,
which provides many advantages over orthogonal multiple access including multiuser capacity, supporting overloaded transmission, enabling reliable and low latency
grant-free transmission, enabling flexible service multiplexing, etc. Applications of
non-orthogonal signaling for multi-user communications have been investigated several years ago, significant efforts were paid to the optimal signaling design and
intensive multi-user detection techniques, to suppress the multiple access interference (MAI) for lowering probability of error or increasing capacity. Hoshyar and
Guo suggest the low-density signature (LDS)-based multiple access [5], or sparsely
spread code-division multiple access (CDMA) [6], which intentionally arranges each
user to spread its data over a fraction of the chips, instead of all chips, to reduce both
the MAI and the complexity of multi-user detection. Inspired by the overloading
capability and the low-complexity feature of LDS, SCMA is developed by inheriting
from LDS the sparse sequence structure, such that the message-passing algorithm
(MPA) is available in multi-user detection to achieve near-optimal performance. In
contrast to the LDS scheme, multi-dimensional signal constellations, instead of the
spreading, are utilized in SCMA to combat the channel fading and MAI. As a result,
Z. Ma (B) · J. Bao
Southwest Jiaotong University, West Section, High-tech Zone, Chengdu, Sichuan, China
e-mail: zma@swjtu.edu.cn
J. Bao
e-mail: jinchen_bao@my.swjtu.edu.cn
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_12
369
370
Z. Ma and J. Bao
the larger coding gain and better spectrum efficiency are achievable for SCMA due
to the improved codebooks, compared to LDS.
As one of NOMA family, SCMA is capable of supporting overloaded access over
the coding domain, hence increasing the overall rate and connectivity. By carefully
designing the codebook and multi-dimensional modulation constellations, the coding
and shaping gain can be obtained simultaneously. In an SCMA system, users occupy
the same resource blocks in a low-density way, which allows affordable low multiuser joint detection complexity at receiver. The sparsity of signal guarantees a small
collision even for a large number of concurrent users, and the spread-coding like
codes design brings good coverage and anti-interference capability due to spreading
gain as well.
12.1.1 System Model
12.1.1.1
Multiple Access Procedure
An SCMA transmission system can be simply illuminated in Fig. 12.1. Suppose that
there are J synchronous users multiplexing over K shared orthogonal resources,
e.g., K time slots or orthogonal frequency division multiplexing (OFDM) tones,
and each user employs one SCMA layer.1 The forward error control (FEC) coding
scheme can be low-density parity-check (LDPC) codes or polar codes which have
been adopted for 5G recently. Each SCMA modulator/encoder maps the coded bits
to a K -dimensional complex codeword, and the resulted J codewords constitute an
SCMA block, as is shown in Fig. 12.1 (J = 6, K = 4 in the figure). The multi-user
codewords in each SCMA block are multiplexed over the air transmissions in uplink
multiple access channel (MAC), or they are superimposed at the transmitter of the
downlink broadcast channel (BC). Since each SCMA block occupies K resources for
codeword transmitting, the resulted overloading factor is J/K . This multiple access
process is similar to that of CDMA, where the spread signals in CDMA are replaced
with the SCMA codewords. Multi-user detection is carried out at the receiver to
recover the colliding codewords.
For the uplink MAC, the received signal vector after the synchronous user multiplexing is expressed as
y=
J
j=1
diag(h j )x j + n
(12.1)
t
and h j = [h j [1], . . . , h j [K ]]t , are the
where x j = x j [1], . . . , x j [K ]
K -dimensional codeword and the corresponding channel gain for the jth user, respectively, and diag(h j ) denotes the diagonal matrix with h j [k] being the kth diagonal
1 In
practical scenarios, each user employs one or multiple layers.
12 Sparse Code Multiple Access (SCMA)
371
Fig. 12.1 The system model for SCMA
element. The K -vector n is the additive white Gaussian noise (AWGN) with zero
mean and variance N0 per dimension. It is convenient to view the MAC model as
an equivalent “MIMO” communication system, and the received vector in (12.1)
becomes
y = HX + n
(12.2)
where H = diag(h1 ), diag(h2 ), . . . , diag(h J ) , is the equivalent “MIMO” channel
t
matrix, and X = x1t , x2t , . . . , xtJ , is the combined multi-user codeword representing
an SCMA block.
For the downlink BC, the codewords from multiple users are superimposed before
the transmission, so that they experience the same fading. In the case of absence of
interference between K resources, the received signal vector is given by
y = diag(h)
J
j=1
x j + n = diag(h)X + n
(12.3)
where a single receiver is considered here for simplicity, and X = Jj=1 x j , is the
superimposed codeword of J users at the input of a BC, which also represents an
SCMA codeword block.
In the following, the upper case X always denotes the combined multi-user codeword of J users in the MAC model, or the superimposed codeword in the BC model.
12.1.1.2
SCMA Codebook Mapping
Unlike the modulation used for 3G and 4G, the modulation and codebook mapping
in SCMA are designed jointly in a multi-dimensional and sparsely spread way. An
SCMA modulator/encoder maps the input bits to a K -dimensional sparse codeword,
372
Z. Ma and J. Bao
Fig. 12.2 Illustration of SCMA codebooks and bits to codeword mapping
which is selected from a layer-specific codebook of size M. The K -dimensional
complex codewords of the codebook are sparse vectors with N < K nonzero entries,
and all the codewords contain 0 in the same dimensions. Then, the codebook is sparse,
and this is where the “sparse code multiple access” is named from.
The codebooks are constructed by a mapping from an N -dimensional complex
constellation with a mapping matrix. Denote the constellation for the jth layer/user
with C j , which contains M j constellation points of length N j . The mapping matrix
V j maps the N j -dimensional constellation points to SCMA codewords to form the
codebook X j . To simplify our illumination and analysis, we assume that all layers
have the same constellation size and length, i.e., M j = M, N j = N , ∀ j. In summary,
the resulting codebook for the jth user contains M codewords, each codeword consists of K complex values from which only N are nonzero specified by the mapping
matrix V j .
An example of the codebook mapping is shown in Fig. 12.2, where a codebook
set containing 6 codebooks for transmitting 6 SCMA layers is illustrated (J = 6).
Each codebook contains 8 four-dimensional codewords (M = 8, K = 4), and two
of the four entries in the codewords are nonzero (N = 2). Upon transmission, the
codeword of each layer is selected based on the labeling of the bit sequence.
12.1.1.3
Factor Graph Representation
The low-density structure of SCMA codewords can be efficiently characterized by a
factor graph, which is analogous to that for LDPC codes. A binary column vector f j
of length K is used to indicate the positions of zero (with digit 0) and nonzero (with
digit 1) entries of the jth codebook. Then, a K × J sparse matrix F = [f1 , . . . , f J ],
called factor graph matrix, can be used to indicate the relationships between the
layers and resources. The rows of F indicate the resources and the columns indicate
12 Sparse Code Multiple Access (SCMA)
373
Fig. 12.3 Factor graph
representation for SCMA
the layers. The (k, j)th element of F, denoted as f k, j , is 1 if the jth layer contributes
its data to the kth resource.
Correspondingly, let the J variable nodes (VNs) and K function nodes (FNs) in
the factor graph represent the layers and resources, respectively, and the jth VN is
connected to the kth FN if and only if f k, j = 1. In the following, we denote
φk = { j : 1 ≤ j ≤ J, f k, j = 1},
ϕ j = {k : 1 ≤ k ≤ K , f k, j = 1}
(12.4)
the index set of layers contributing to the kth resource, and the index set of resources
occupied by the jth layer, respectively. For a regular factor graph matrix, |φ1 | =
· · · = |φ K | and |ϕ1 | = · · · = |ϕ J |, and let d f = |φk | and dv = |ϕ j |.
Example 1 Consider a 6-user SCMA transmission system with J = 6, K = 4, such
a system permits a transmission overloading 150%, and the system model is depicted
in Fig. 12.1. If we carefully design the factor graph matrix F to allow the users to
collide over only one nonzero element, then a choice of F is given by
⎡
1
⎢0
F=⎢
⎣1
0
0
1
0
1
1
1
0
0
0
0
1
1
1
0
0
1
⎤
0
1⎥
⎥
1⎦
0
(12.5)
In the sparse matrix settings, matrix (12.5) has d f = 3 and dv = 2, which means
that each FN is connected to three VNs and each VN is connected to two FNs. The
corresponding factor graph is shown in Fig. 12.3, and an example of a codebook
(with size M = 4) is listed in Table 12.1.
In summary, the main features of SCMA lie in:
• Code-domain non-orthogonal multiplexing: SCMA allows superposition of multiple codewords from different users over several resources, which supports overloading. The superposition pattern on each resource is defined in codebooks.
• Sparse spreading: SCMA uses sparse spreading to reduce inter-layer interference,
so that more codewords collisions can be tolerated with low receiver complexity.
374
Z. Ma and J. Bao
Table 12.1 An Example of SCMA Codebook (K = M = 4, N = 2, J = 6)
SCMA
SCMA Codebook for each layer
codebook
index
⎡
⎤
0.7851
−0.2243
0.2243
−0.7851
⎢
⎥
0
0
0
0
⎢
⎥
Codebook 1 ⎢
⎥
⎣ −0.1815 − 0.1318i −0.6351 − 0.4615i 0.6351 + 0.4615i 0.1815 + 0.1318i ⎦
0
0
0
0
⎤
⎡
0
0
0
0
⎢ −0.1815 − 0.1318i −0.6351 − 0.4615i 0.6351 + 0.4615i 0.1815 + 0.1318i ⎥
⎥
⎢
Codebook 2 ⎢
⎥
⎦
⎣
0
0
0
0
0.7851
−0.2243
0.2243
−0.7851
⎡
⎤
−0.6351 + 0.4615i 0.1815 − 0.1318i −0.1815 + 0.1318i 0.6351 − 0.4615i
⎢ 0.1392 − 0.1759i 0.4873 − 0.6156i −0.4873 + 0.6156i −0.1392 + 0.1759i ⎥
⎢
⎥
Codebook 3 ⎢
⎥
⎣
⎦
0
0
0
0
0
0
0
0
⎤
⎡
0
0
0
0
⎥
⎢
0
0
0
0
⎥
⎢
Codebook 4 ⎢
⎥
⎦
⎣
0.7851
−0.2243
0.2243
−0.7851
−0.0055 − 0.2242i −0.0193 − 0.7848i 0.0193 + 0.7848 0.0055 + 0.2242i
⎡
⎤
−0.0055 − 0.2242i −0.0193 − 0.7848i 0.0193 + 0.7848 0.0055 + 0.2242i
⎢
⎥
0
0
0
0
⎢
⎥
Codebook 5 ⎢
⎥
⎣
⎦
0
0
0
0
−0.6351 + 0.4615i 0.1815 − 0.1318i −0.1815 + 0.1318i 0.6351 − 0.4615i
⎡
⎤
0
0
0
0
⎢
⎥
0.7851
−0.2243
0.2243
−0.7851
⎢
⎥
Codebook 6 ⎢
⎥
⎣ 0.1392 − 0.1759i 0.4873 − 0.6156i −0.4873 + 0.6156i −0.1392 + 0.1759i ⎦
0
0
0
0
• Multi-dimensional modulation: SCMA employs multi-dimensional constellations
for better spectral efficiency.
12.1.2 Multi-user Detection
This subsection discusses multi-user detection schemes for SCMA, including the
optimal detection, the MPA receiver and other advanced receivers.
12 Sparse Code Multiple Access (SCMA)
12.1.2.1
375
Optimal/Quasi-optimal Multi-user Detection
A. Optimal Multi-user Detection
Assume that channel state is perfectly estimated at the receiver, given the received
signal vector y, the joint optimum maximum a posteriori probability (MAP) detection, for multi-user codeword X and for the jth user’s codeword x j , can be written
as
p(X|y)
(12.6)
X̂ = arg max p(X|y), and x̂ j = arg max
x j ∈X j ,∀ j
x j ∈X j
xi ∈Xi ,∀i= j
respectively. Using Bayes’ rule
p(X|y) =
where
p(y|X) p(X)
∝ p(y|X)P(X)
p(y)
J
p(X) =
j=1
p(x j ), and p(y) =
p(y|X) p(X)
x j ∈X j ,∀ j
are the joint a prior probability2 for each user’s codeword, and the probability of the
received signal vector, respectively.
By assuming that the noise components over the K resources are identically
independently distributed (i.i.d.), it holds
K
p(y|X) =
p(y[k]|X)
k=1
and considering only d f users actually collided over the kth resource, we have
⎞
⎛
2
1
1
p(y[k]|X) =
h j [k]x j [k] ⎠
exp ⎝− y[k] −
π N0
N0
j∈φ
(12.7)
k
Thus, the MAP decision for the jth user’s codeword is given by
K
x̂ j = arg max
x j ∈X j
P(X)
xi ∈Xi ,∀i= j
k=1
p(y[k]|X), ∀ j
(12.8)
With the codeword probability for each user, it is straightforward to calculate the
log-likelihood rate (LLR) for each coded bit, so that they can serve as the input for
2 Without
feedback from the FEC decoder, p(x j ) =
1
M
for all the users.
376
Z. Ma and J. Bao
the FEC decoder. For the jth user, the LLR considering the mth bit b j,m is calculated
by
Pr{b j,m = 1|y}
Λ(b j,m ) = log
Pr{b j,m = 0|y}
K
(12.9)
x j ∈X1j,m
xi ∈Xi ,∀i= j P(X)
k=1 p(y[k]|X)
= log
K
x j ∈X0j,m
xi ∈Xi ,∀i= j P(X)
k=1 p(y[k]|X)
where X1j,m and X0j,m are subsets of X j for which the mth bit of the jth user b j,m = 1
and b j,m = 0, respectively. Note that solving (12.8) is equivalent to solve the marginal
product of functions (MPF) problem, which is of exponential complexity with bruteforce searching, and is prohibitive to employ when the number of users increases.
B. MPA Detection
As the SCMA encoding can be represented by a factor graph with sparse property,
the low-complexity MPA can be used to solve the MPF problem with near-optimum
performance.
(t)
Let Ik→
j be the extrinsic information to be passed from FN k to VN j at the tth
(t)
iteration, and I j→k
be the extrinsic information to be passed from VN j to FN k.
Given the a prior probability p(x j ), the probability that x j is transmitted by the jth
user given the channel sample is updated as
(t)
I j→k
(x j ) = p(x j )
(t)
Il→
j (x j )
l∈ϕ j \k
Then, for any x j ∈ X j , the probability of the received signal y[k] given that x j is
transmitted by the jth user, marginalized over all possible codewords of the other
users, is given by
(t)
Ik→
j (x j ) =
xi ∈Xi ,∀i∈φk \ j
(t−1)
Ii→k
(xi )
p (y[k]|X)
(12.10)
i∈φk \ j
After a number of iterations, the posterior probability of x j produced by the MPA is
proportional to
I j (x j ) = p(x j )
k∈ϕ j
(T )
Ik→
j (x j ), x j ∈ X j , j = 1, . . . , J
(12.11)
where T is the number of iterations at the termination.
Similar to that for MAP detection, the LLR of the mth bit of the jth user b j,m is
calculated by
x j ∈X1j,m I j (x j )
Λ(b j,m ) = log
(12.12)
x j ∈X0j,m I j (x j )
12 Sparse Code Multiple Access (SCMA)
377
where X1j,m and X0j,m are the same as that in (12.9).
The main complexity of MPA comes from the calculation of (12.10), the summation over xi adds up M |φk |−1 terms while M probabilities should be calculated in each
iteration, which leads to a complexity order O(T K M d f ), and is far below that of
the optimal MAP detection. In practical implementations, the exponential function
in MPA algorithm may cause large dynamic ranges and very high storage burden,
then the logarithmic domain MPA is preferred to avoid the exponential operations.
For the log-MPA operation, the information exchanged between the FNs and VNs
can be expressed as
(t)
(x j ) = log p(x j ) +
I j→k
(t)
Ik→
j (x j ) =
max
xi ∈Xi ,∀i∈φk \ j
(t)
Il→
j (x j )
l∈ϕ j \k
⎧
⎨
⎩
log p (y[k]|X) +
(t−1)
Ii→k
(xi )
i∈φk \ j
⎫
⎬
⎭
p
i
where Jacobi’s logarithm formula log
≈ maxi pi is applied for a complexity
i e
reduction to a certain degree, which results in the max-log-MPA detection. The output
LLR of the MPA detector is given by
Λ(b j,m ) = max I j (x j ) − max I j (x j )
x j ∈X1j,m
where
I j (x j ) = log p(x j ) +
12.1.2.2
x j ∈X0j,m
(T )
Ik→
j (x j )
k∈ϕ j
Other Advanced Detectors
The MPA detector still has exponential complexity with respect to the codebook size
(M) and the number of accessed users at each resource (d f ), which may become
impractical for the implementation of very large codebook size (e.g., M ≥ 64) and
very high overload (e.g., d f ≥ 8). Some other advanced detectors can harness the
potential gain of SCMA while provide sufficient flexibility for a good trade-off
between the performance and detection complexity [7, 8].
A. EPA Detector
Expectation propagation algorithm (EPA) is an approximate Bayesian inference
method in machine learning for estimating sophisticated posterior distributions with
simple distributions through distribution projection, and it turns out to be an efficient iterative multi-user detector for SCMA as well as some other multiple access
schemes [8]. It approximates the discrete message in MPA as continuous Gaussian
message using the minimum Kullback–Leibler (KL) divergence criterion, and use
the a posteriori probabilities fed back from the FEC decoder to compute the approx-
378
Z. Ma and J. Bao
imate symbol belief and the approximate message, such that the message passing
reduces to mean and variance parameters update. The detailed algorithm is given in
[8]. With EPA, the complexity order of SCMA detection is reduced to linear complexity, i.e., it only scales linearly with the codebook size M and the average degree
of the factor nodes d f , while simulation results show that the EPA detector shows
nearly the same error performance as MPA for SCMA with receiver diversity. As a
result, the computation burden of the SCMA receiver is significantly alleviated and
is no longer a problem for implementation in real systems.
B. SIC-MPA Detector
Successive interference cancelation (SIC) receiver is a kind of multi-user receiver that
treats all the other undecoded users as interference when decoding a target user, and
can be implemented as either symbol level or codeword level. It works well when
the received SNR among users are quite different from each other. However, the
detecting performance deteriorates when the SNR difference is not obvious between
users, in which case error propagation happens.
To strike a good balance between link performance and implementation complexity, it is reasonable to combine SIC with an MPA (SIC-MPA) receiver. More
specifically, MPA is applied to a limited number of users firstly, so that the number
of colliding users over each resource does not exceed a given threshold value (e.g., ds
users). Then, the successfully decoded users are removed by SIC and the procedure
continues until all users are successfully decoded or no new user gets successfully
decoded in MPA. In the case of ds = d f , full MPA is realized, and when ds = 1, it
becomes a pure SIC receiver. Due to the fact that MPA is used for a very limited
number of users instead of all the users, the decoding complexity is greatly reduced,
which is of the order O(T K M ds ).
12.2 Performance Evaluation
The error performance and capacity are excellent measures that indicating the goodness of a system, and more importantly, they serve as powerful tools for the practical
system design. For SCMA, the multi-user codebook plays a key role in the system
performance improvement, and it is necessary to establish performance criterion to
guide the codebooks design. In this section, error performance and capacity analysis
for uplink and downlink SCMA systems are provided, and independent Rayleigh
fadings are assumed.
12.2.1 Average Error Probability
The error probability, e.g., the average codeword error probability (ACEP), is one of
the most important performance criteria, since it is most revealing about the nature
12 Sparse Code Multiple Access (SCMA)
379
of a system behavior. However, it is quite difficult to evaluate the exact ACEP for
SCMA systems, since one needs to average over several fading statistics due to the
multi-channel transmissions, and the integration involves a decision cell of a multidimensional signal point. As an alternative approach, it is convenient to resort to an
upper bound or approximation on the ACEP [9]. In this subsection, we use union
bound to evaluate the error performance of uplink and downlink SCMA under joint
maximum likelihood (ML) multi-user detection.
12.2.1.1
PEP over Uplink MACs
Consider the equivalent “MIMO” channel (12.2). Under the assumption of perfect
channel estimation at the receiver, the joint ML detection of multi-user codewords
is equivalent to the joint minimum distance decoding
X̂ = arg min y − HX
X
The pairwise error probability (PEP), defined as the probability that received signal
vector y is detected into Xb given that Xa is transmitted, is given by [10]
⎡
⎛
P{Xa → Xb } = EH ⎣ Q ⎝
H(Xa − Xb )
2N0
2
⎞⎤
⎠⎦
(12.13)
∞ 2
where, Q(x) = √12π x e−t /2 dt, is the well-known Gaussian function [11], and
EH [·] denotes the mean.
Let x j,a [k] and x j,b [k] be the kth entries of the jth user’s codewords x j,a and x j,b ,
corresponding to Xa and Xb , respectively. Due to the sparseness of the codewords,
/ φk . Now we define a distance for the MAC.
x j,a [k] = x j,b [k] = 0 whenever j ∈
Definition 1 The kth dimension-wise distance, between the multi-user combined
codewords Xa and Xb , for the uplink MAC is defined as
λ2k =
J
j=1
|x j,a [k] − x j,b [k]|2 =
j∈φk
|δ j [k]|2 , ∀k
(12.14)
where δ j [k] = x j,a [k] − x j,b [k].
Assume that there are repeated values among the set {λ21 , . . . , λ2K }, such that they
can be divided into V (1 ≤ V ≤ K ) groups, and each group contains the collection
of a certain value λ̂2v . Let λ̂ = [λ̂21 , . . . , λ̂2V ]t , be the vector of V distinct elements
among {λ21 , . . . , λ2K }, and r = [r1 , . . . , r V ]t , where rv is the number of elements in
V
{λ21 , . . . , λ2K } that equals to λ̂2v , such that v=1
rv = K .
380
Z. Ma and J. Bao
Definition 2 Define the parameter
V
K
Ar,λ =
λ−2
k
k=1
=
λ̂v−2rv
(12.15)
v=1
as the reciprocal of the product of the dimension-wise distances.
Definition 3 For positive integers l, v and vectors r and λ̂, define the parameter
V
l+1
Bv,l,r,λ̂ = (−1)
η∈Ωv,l j=1, j=v
ηj + rj − 1
ηj
1
λ̂2j
−
1
−(r j +η j )
(12.16)
λ̂2v
where the vector η = [η1 , . . . , ηV ]t is created from the set Ωv,l of all nonnegative
integer partitions of l − 1 (with ηv = 0). The set Ωv,l is defined as
Ωv,l
"
!
V
t
V
η j = l − 1, ηv = 0, η j ≥ 0 ∀ j
= η = [η1 , . . . , ηV ] ∈ Z ;
j=1
Next, we provide the main result regarding the PEP.
Theorem 1 For J users using K -dimensional codebooks in the uplink SCMA systems, the PEP between Xa and Xb is given by
P{Xa → Xb } = Ar,λ̂
×
rv
V
λ̂2L
v Bv,rv −L+1,r,λ̂
v=1 L=1
1 − µv
2
L
L−1
k=0
L −1+k
k
1 + µv
2
where
µv =
λ̂2v
4N0 + λ̂2v
Proof Consider the metric in (12.13),
H(Xa − Xb )
2
=
K
k=1
h[k]† (xa [k] − xb [k])
2
k
(12.17)
12 Sparse Code Multiple Access (SCMA)
381
where xa [k] = [x1,a [k], . . . , x J,a [k]]t , is the vector of the kth component for J users,
and h[k]† = [h 1 [k], . . . , h J [k]], are the corresponding channel gains, and [·]† denotes
conjugate transpose. Using the matrix decomposition, it holds that
(xa [k] − xb [k])(xa [k] − xb [k])† = Uk Λk Uk†
where Uk is unitary and Λk is a diagonal matrix, i.e., Λk = diag(λ̃2k,1 , . . . , λ̃2k,J ), with
λ̃2k, j being the ordered singular values of the matrix (xa [k] − xb [k])(xa [k] − xb [k])† .
Note that the matrix (xa [k] − xb [k])(xa [k] − xb [k])† is of rank 1 and the unique
nonzero singular value in Λk is
λ̃2k,1 = xa [k] − xb [k]
2 (a)
=
j∈φk
|x j,a [k] − x j,b [k]|2
where (a) is due to the sparseness of the codebooks. Obviously, the nonzero eigenvalue is equal to the dimension-wise distances defined in Definition 1, namely
λ̃2k,1 = λ2k . Hence,
h[k]† (xa [k] − xb [k])
2
= h[k]† Uk Λk Uk† h[k]
= h̃[k]† Λk h̃[k] = λ2k |h̃ 1 [k]|2
where we define, h̃[k]† = h[k]† Uk = [h̃ 1 [k], . . . , h̃ J [k]]. Thus, h̃[k] has the same
distribution as h[k], since multiplying with unitary matrix Uk doesn’t change the
amplitudes. Thus, the average PEP in (12.13) is equal to
⎡
⎛
P{Xa → Xb } = Eh̃[1],...,h̃[K ] ⎣ Q ⎝
⎞⎤
λ2k |h̃[k]|2
⎠⎦
2N0
K
k=1
(12.18)
where the index 1 is dropped here for h̃ 1 [k]2 .
For i.i.d. Rayleigh fading, h̃[1], . . . , h̃[K ] are i.i.d.
Gaussian random
K complex
variables with zero mean and unit variance. Thus, k=1
λ2k |h̃[k]|2 is the sum of K
exponential random variables with different means, or a linear combination of V
independent χ 2 -distributed random variables with 2r1 , . . . , 2r V degrees of freedom,
which follows Gamma or Generalized chi-squared distribution [12], with PDF given
by
f (x; r, λ̂) = Ar,λ̂
rv
V
Bv,l,r,λ̂
v=1 l=1
(rv − l)!
x rv −l e
−
x
λ̂2
v
382
Z. Ma and J. Bao
Then, the PEP can be obtained as
P{Xa → Xb } =
#
∞
0
= Ar,λ̂
= Ar,λ̂
Q
$
x
2N0
f (x; r, λ̂)dx
#
rv
V
Bv,l,r,λ̂
v=1 l=1
rv
V
(rv − l)!
0
∞
Q
$
x
2N0
x rv −l e
rv −l+1
(1 − µv )λ̂2v
Bv,l,r,λ̂
2
v=1 l=1
rv −l
1 + µv k
rv − l + k
×
k
2
k=0
−
x
λ̂2
v
dx
(12.19)
where the last step follows from (13.4-15) in [11]. Substituting l with rv − L + 1,
(12.17) is proved. This concludes the proof.
The PEP is uniquely determined by the set of all λ2k s and the SNR, which is
valid for any multi-dimensional codebooks and an arbitrary number of users. By
reducing the number of users to 1, P{Xa → Xb } becomes the PEP between two multidimensional constellation points for a single-user transmission system. Therefore,
the PEP of a joint multi-user detector is actually identical to that of the PEP of a
single-user transmitting over a fading channel, where an equivalent K -dimensional
constellation is employed such that the dimension-wise distances between the two
constellation points are λ21 , . . . , λ2K .
12.2.1.2
PEP over Downlink BCs
Consider the received signal vector of the downlink BC in (12.3), where X =
J
j=1 x j is the superimposed codeword of multiple users at the transmitter. Obviously, the model is exactly the same with that in the single-user communications,
where X is used as the K -dimensional transmitted codeword. The ML multi-user
detection for the superimposed codeword X becomes
X̂ = arg min y − diag(h)X
X
Similar to that in uplink SCMA, we define the distances for downlink BC model.
Definition 4 Let Xa and Xb be two superimposed codewords, and x j,a [k] and x j,b [k]
are the kth entries of the jth user’s codeword corresponding to Xa and Xb , respectively. The kth dimension-wise distance between Xa and Xb , for the downlink broadcast channel, is defined as
12 Sparse Code Multiple Access (SCMA)
383
2
J
2
δ j [k] , ∀k
x j,a [k] − x j,b [k] =
τk2 =
(12.20)
j∈φk
j=1
where δ j [k] = x j,a [k] − x j,b [k].
Theorem 2 The PEP of a Rayleigh broadcast channel is the same as that in (12.17),
after the substitution of λ2k with τk2 .
Proof Similar to that of the uplink case, the average PEP between Xa and Xb is equal
to
⎞⎤
⎡ ⎛
2
diag(h) (Xa − Xb ) ⎠⎦
P{Xa → Xb } = Eh ⎣ Q ⎝
2N0
⎡ ⎛
⎞⎤
(12.21)
K
2
2
k=1 τk |h[k]| ⎠⎦
= Eh ⎣ Q ⎝
.
2N0
As h[1], . . . , h[K ] are independent Rayleigh distributed random variables, the integral has been solved in (12.18), and the PEP has the similar expression as that in the
MAC case, after the substitution of λ2k with τk2 . This completes the proof.
It should be noted that, while the PEP of a BC can be evaluated through the
same expression as that in the MAC case, τk2 is different from the dimension-wise
distance λ2k in MAC, due to the absence of cross components δ j [k] × δi [k], j = i,
between different users. This is because in the MAC, the receiver distinguishes the
multi-user signals by exploiting the differences among the channel coefficients, and
only the amplitude of δ j [k] contributes to the PEP. However, in the broadcast channel
case, since the receiver exploits the differences among the multiple users’ signals to
perform the joint detection, both the amplitude and signs of δ j [k] will influence the
result of PEP.
12.2.1.3
PEP over the AWGN Channel
For the AWGN channel, where h j [k] is a constant for all j and k (assume that
|h j [k]| = c), the expressions of the received signal vector in (12.1) for uplink channels and (12.3) for downlink channels are the same. Then, according to (12.21), it is
straightforward to derive the PEP as
⎛
P{Xa → Xb } = Q ⎝
c2
⎞
2
τ
k=1 k ⎠
2N0
K
⎛&
2 ⎞
'
' c2 K
⎜(
j∈φk δ j [k] ⎟
k=1
⎟
⎜
= Q⎝
⎠
2N0
(12.22)
384
Z. Ma and J. Bao
where τk2 is the dimension-wise distance defined in (12.20), and δ j [k] = x j,a [k] −
x j,b [k].
12.2.1.4
Upper Bounds on PEP
In the codebook design, sometimes it is sufficient and easier to optimize the performance through a bound or an approximation of PEP. The exact PEP in (12.17) is a
little complicated for large K , due to the large number of enumerations in Ωv,l , when
calculating Bv,rv −L+1,r,λ . An alternative way to evaluate (12.17) is to use an upper
bound for the Q-function as [13]
Q(x) ≤
N
2
ai e−bi x , for x > 0,
i=1
where N , ai , bi are constants. Note that the upper bound in Sect. 12.2.1.4 tends to
the exact value as N increases.
For the multiple access and broadcast channels, since X = |ĥ[k]|2 is an exponen∞
1
, for
tial random variable with unit mean, holds that E X [et X ] = 0 et x e−x d x = 1−t
t ≤ 1, and
* N
K
2
2
k=1 λk |h̃[k]|
ai exp −
P{Xa → Xb } ≤ Eh̃[1],...,h̃[K ]
2N0
i=1
*
+
N
K
bi λ2k |h̃[k]|2
ai
=
Eh̃[k] exp −
2N0
i=1
k=1
=
N
i=1
bi
+
K
ai
2N0
2
2N
0 + bi λk
k=1
By choosing N = 1, a1 = b1 = 21 , we get the Chernoff bound with a scaling factor
of 0.5 as
K
Pch {Xa → Xb } ≤
4N0
1
2 k=1 4N0 + λ2k
(12.23)
In general, the Chernoff bound may be a little loose, but this does not affect the
optimization criteria in the constellation design. It is obvious from (12.23) that a
good direction is to design multi-dimensional multi-user codebooks, such that λ2k
to span in as many dimensions as possible (maximizing the diversity) and to make
the maximum PEP or maximum of Pch {Xa → Xb } as small as possible. If for any
12 Sparse Code Multiple Access (SCMA)
385
codeword pair Xa and Xb , all the λ2k are positive, then the maximal diversity order of
K can be achieved. Due to the sparseness of the codebooks, the diversity is always
less than K . A tight and simple bound (or approximation) is to choose N = 2, a1 =
1
, a2 = 41 , b1 = 21 , b2 = 23 , which is denoted as Pub {Xa → Xb }.
12
12.2.1.5
A Universal Bound of ACEP for Joint ML Detection of
Multiple Signals
A commonly used approach for the error performance analysis is the evaluation of the
ACEP by using a union bound, assuming that the codewords are equiprobable transmitted. In general, the ACEP is dominated by the nearest neighbors of codewords,
which result in a tight upper bound. However, it is quite difficult (if not impossible)
to find the nearest neighbors in multi-user scenarios. To deal with this, we take into
account all possible codewords that contribute to the ACEP.
Let M1 , . . . , M J be the codebook size for J users, respectively. We define {X j } Jj=1
as the set of all Jj=1 M j possible combined codewords of J users, and let Xa , Xb ∈
{X j } Jj=1 be two different elements of {X j } Jj=1 . Here, the combined codeword Xa
and Xb are a J K -dimensional vector for the MAC, or the sum of J K -dimensional
codewords for the BC. Denote x j,a and x j,b the transmitted codewords of the jth user
corresponding to Xa and Xb . Then, there are M j possible values for x j,a and x j,b .
Note that x j,a and x j,b are K -dimensional vector with complex entries, i.e., SCMA
codeword. Following the approach in [14] for multiple signals and [15] for MIMO
channels, the ACEP for the jth user with joint ML detection of J users’ signals is
upper bounded by
P j (e) ≤ J
1
j=1 M j
Xa
⎛
⎝
Xb ,x j,b =x j,a
⎞
P {Xa → Xb }⎠
(12.24)
The ACEP of the system can
be obtained by taking the mean of all the single-user
ACEPs, namely P(e) = 1J Jj=1 P j (e).
On the right-hand side of (12.24), the summation over Xa will add up Jj=1 M j
terms and the summation over Xb will add up (M1 − 1) Jj=2 M j terms. Thus, there
will be up to (M12 − M1 ) Jj=2 M 2j PEPs in (12.24), which is intractable for a large
constellation size and number of users. However, we can simplify it by using the
symmetry of the dimension-wise distances. For example, consider the ACEP for
the first user P1 (e) here. The upper bound of P1 (e) can be decomposed into the
summation of two parts as
386
Z. Ma and J. Bao
1
J
j=1
+ J
Mj
Xa
Xb ,x1,b =x1,a ,
[x2,b ,...,x j,b ]=[x2,a ,...,x j,a ]
Mj
1
j=1
Xa
P {Xa → Xb }
P {Xa → Xb } .
Xb ,x1,b =x1,a ,
[x2,b ,...,x j,b ]=[x2,a ,...,x j,a ]
(12.25)
The first part in (12.25) is the union bound of the probability of the event that all
users’ signals are correctly detected except for the first user, namely the ACEP for
the first user with single-user
detection in the absence of interference. This part is
a summation of (M1 − 1) Jj=1 M j PEPs, while only 21 M1 (M1 − 1) different PEP
values should be calculated, due to the symmetry of the dimension-wise distance for
the first user. The second part is the probability of the event that the errors happen
for the first user
and for at least
one user among {2, . . . , J }, which is the summation
of (M1 − 1)( Jj=2 M j − 1) Jj=1 M j PEPs, but only one-fourth of them should be
considered. A further simplification can be achieved for the MAC by considering
more decompositions of the second part.
In general, SCMA codebooks of all users are constructed from a common mother
constellation [16], with some layer-specific operations over this constellation to get
their own layer’s codebook. These layer-specific operations do not change the fundamental properties of the mother constellation, such as the Euclidean distance. The
layer operation losses their efficiency in the uplink multiple access fading channels,
due to the distinctness of each user’s channel gain. If the factor graph matrix is regular
as that in (12.5), every user will suffer from the same interference from other users.
Then, the system results in the same performance for all users, while for other cases,
the ACEP is asymmetric for each user.
The MPA detection is believed to be an efficient approach for SCMA systems.
Theoretically, the MPA detector is asymptotically equivalent to the optimal MAP
detector [17, 18] (or ML conditioned on equal probably transmissions) for a sparsely
spread system with long signatures. The analytical bounds, proposed in this subsection, work for ML detector as well as for the MPA detector.
12.2.1.6
Numerical Results and Simulations
We consider an SCMA system illustrated in Example 1, and the four-dimensional
four-ary codebooks are listed in Table 12.1. The ACEP of SCMA over AWGN and
uplink Rayleigh fading channels for 2, 4, and 6 user cases are evaluated. For the
Rayleigh fading channel, we give analytical results of the union bound on the ACEP,
corresponding to exact PEP (denoted as Pjml ), the upper bound on PEP Pub , and the
scaled Chernov bound Pch , respectively.
Results for AWGN channels are shown in Fig. 12.4a. The analytical bound of a
joint ML detector closely coincides with the simulation curves for large SNR. The
bound is quite tight for values of ACEP below 10−3 , even for six users. Thus, this
12 Sparse Code Multiple Access (SCMA)
387
0
0
10
10
ML
MPA
Bound − Pjml
Bound − Pub
Bound − Pch
−1
10
−1
10
−2
10
−2
ACEP
ACEP
10
−3
10
−4
10
2 Users
4 Users
−3
10
−4
6 Users
10
−4
−5
10
10
ML
MPA
Bound − Pjml
−6
10
0
5
22.5
−5
10
10
15
20
0
23
23.5
5
24
10
15
SNR [dB]
(a) SCMA over AWGN channel.
0
−1
−2
10
−4
MPA
ML
Bound − Pjml
Bound − Pub
Bound − Pch
10 −1
ACEP
ACEP
30
10 0
MPA
ML
Bound − Pjml
Bound − Pub
Bound − Pch
10
−3
25
(b) SCMA with 2 users in Rayleigh fading.
10
10
20
SNR[dB]
10 −2
10 −3
10
−4
10
−4
10 −4
10
26
−5
10
5
26.5
10
27
27.5
28
15
20
25
30
SNR [dB]
(c) SCMA with 4 users in Rayleigh fading.
10 −5
26
5
26.5
10
27
27.5
28
15
20
25
30
SNR [dB]
(d) SCMA with 6 users in Rayleigh fading.
Fig. 12.4 ACEP of uplink SCMA over AWGN and Rayleigh fading channels
bound is sufficient for the analysis and design of a signal constellation in AWGN.
Surprisingly, there are bends for the ACEP curves of an ML detector and analytical
bound for the six user case. The performance turns better than expected within the
SNR region from 12 to 18 dB, which is due to the sparse codebooks. This phenomenon
happens if the distance profile of the multi-user codebooks is uneven. For example,
a quite small distance exists while the others are very large. In general, the ACEP of
a constellation in AWGN channels is proportional
to√the summation of Q-function
of the distances d and SNR, i.e., P(SNR) ∝ d Q( dSNR), where d is the set of
distances among the constellation points. If there is a large difference between two
distance components, P(SNR) is not a convex function and a bend appears in the
P(SNR) vs SNR curves in log–log scale at low SNR. The theoretical bound is still
quite close to the actual ACEP within the bend region. It can be seen from Fig. 12.4a
that there is nearly 0.4 dB gap between the performance of the MPA detector and the
388
Z. Ma and J. Bao
ML detector at the SNR of 14 dB. The performance of the MPA detector is improved
asymptotically and approaches that of ML detector at high SNRs.
Figure 12.4b–d present the performance for 2, 4, and 6 users over Rayleigh fading
channels, respectively. All the bounds are asymptotically tight as SNR increases. The
analytical bound Pjml is quite tight for values of ACEP below 10−3 for all numbers
of users, and the gap between Pjml and the exact ACEP is almost constant at high
SNRs, when the number of users increases. Moreover, the bounds become looser at
low SNRs as the number of users increases. The upper bound Pub shows superiority
over all the other bounds, since it is much easier to calculate than Pjml while it has
only a little difference. It should be noted that the scaled Pch is much looser compared
to Pub . As expected, the MPA detector shows exactly the same performance as the
ML detector for any number of users and any values of SNR over Rayleigh fading
channels.
12.2.2 Capacity and Cutoff Rate
This subsection discusses the sum rate analysis of SCMA systems. The channel
capacity characterizes the limit information rate that can be reliably transmitted over
a channel. It is well known that the sum rate of multi-channel transmissions is simply
the sum of per channel rate, and in the uplink SCMA, the communications over each
SCMA resource constitutes a multiple access process, then, the sum rate of uplink
SCMA is
⎡
⎛
⎞⎤
K
|h j [k]|2 ⎠⎦
Eh 1 [1],...,h J [K ] ⎣log2 ⎝1 + ρ
C=
j∈φk
k=1
d f d f −i
(12.26)
K e1/ρ (−1)d f − j−i ρ i+ j
= df
Γ ( j, 1/ρ)
ρ ln 2 i=1 j=0 j!(d f − i − j)!
∞
where ρ is the SNR, and Γ (a, x) = x t a−1 e−t dt, is the incomplete Gamma function. In the above sum rate evaluation, we assume that the users have the same
transmitting power, and each SCMA resource carries the same number of users, i.e.,
d f = |φk |. To achieve any point on the sum rate curve, codebooks with Gaussian
distributions and successive interference cancelation (SIC) receivers are generally
required.
In practical cases, it is more valuable to investigate the capacity restricted by
specific codebooks, i.e., the discrete codebook-constrained capacity (DCCC). Consider the equivalent linear system of uplink SCMA in (12.2). Assuming that perfect
channel knowledge is available at the receiver. The conditional probability density
function (PDF) of the received signal vector is
12 Sparse Code Multiple Access (SCMA)
389
1
y − HX
f (y|X, H) =
exp −
N0
(π N0 ) K
2
(12.27)
The mutual information I (X; y) between the discrete input X and the continuous
output y, or the DCCC, is given by [11]
,
Xb f (y|Xb , H)
I (X; y) = log2 M − Ey,Xa ,H log2
f (y|Xa , H)
⎤
⎡
#
f
(y|X
,
H)
1
b
Xb
= log2 M J − EH ⎣ J
dy⎦
f (y|Xa , H) log
M y X
f (y|Xa , H)
J
a
(12.28)
where Xa , Xb ∈ {X j } Jj=1 are two combined codewords for J users. Obviously, it
is quite difficult—if not impossible—to deal with the expression for the mutual
information, and a closed-form solution is unattainable. In the following, we resort
to the cutoff rate analysis.
12.2.2.1
Cutoff Rate of Uplink MAC
The channel cutoff rate R0 , which is a lower bound on the channel capacity, is
another commonly used metric characterizing the channel rate. The cutoff rate is
more informative than the DCCC, since it provides a good estimate of the capacity
as well as a tight upper bound on the error probability of an optimal detector.
The cutoff rate can be defined by [11]
⎡
R0 = − log2 ⎣
Xa
Xb
⎤
(12.29)
p(Xa ) p(Xb )∆Xa ,Xb ⎦
where p(Xa ) = p(Xb ) = M1J , and ∆Xa ,Xb is the Bhatacharyya bound on the PEP
between Xa and Xb , which is given by [11]
∆Xa ,Xb = EH
,#
.
,
− 1
= EH e 4N0
0 1
−
= EH e 4N0
p (y | Xa , H) p (y | Xb , H) dy
H(Xa −Xb )
2
H(Xa −Xb )
2
#
1
-
/ H(Xa +Xb ) /2
1
− N1 /y−
/
2
0
e
dy
(π N0 ) K
/
Note that ∆Xa ,Xa = 1, then the cutoff rate can be written as
/
-
(12.30)
390
Z. Ma and J. Bao
⎞
1
∆Xa ,Xb ⎠
R0 = log2 M J − log2 ⎝1 + J
M X X =X
⎛
a
b
(12.31)
a
It is observed that, the term M1J Xa Xb =Xa ∆Xa ,Xb , inside the bracket of (12.31), is
the union-Bhatacharyya bound on the joint codeword error probability for multiple
users. Therefore, optimizing the mean cutoff rate is equivalent to the optimization
of the error probability, and cutoff rate can be used as a good performance criterion
for the system design.
For the uplink MAC, according to the analysis in Theorem 1,
H(Xa − Xb )
2
=
K
k=1
λ2k |h̃[k]|2
where λ2k is the kth dimension-wise distance in the MAC defined in Definition 1, and
|h̃[1]|, . . . , |h̃[K ]| are independent Rayleigh distributed random variables. Therefore,
−1
K
K
0 1 2
1
λ2
−
λ |h̃[k]|2
=
Eh̃[k] e 4N0 k
∆Xa ,Xb =
1+ k
4N0
k=1
k=1
and thus the cutoff rate for uplink MAC is given by
⎡
⎤
−1
K
λ2k
1
⎦
R0 = log M − log ⎣1 + J
1+
M X X ,b=a k=1
4N0
J
a
(12.32)
b
The average sum rate and cutoff rate for uplink SCMA in Rayleigh fading are
depicted in Fig. 12.5, where the 4-ary codebook in Table 12.1 is adopted. The sum
rates of SCMA with 2, 4 and 6 users are represented by the uppermost curves, which
increase almost linearly with the SNR when SNR becomes very large. Due to the
discrete codebooks, the DCCC and the cutoff rate are upper bounded by KJ log2 (M).
However, significant rate improvement can be achieved by overloaded access for
moderate to large SNRs. As it is observed, the cutoff rate establish a lower bound to
the DCCC, and it asymptotically approaches the DCCC with increasing SNRs.
12.2.2.2
Cutoff Rate of Downlink BC
Consider the downlink BC model in (12.3), for the jth user, the cutoff rate corresponding to the mutual information I (x j ; y) is given by [19]
R0 = log2 M − log2
1
1+
∆x j,a ,x j,b
M x x =x
j,a
j,b
j,a
12 Sparse Code Multiple Access (SCMA)
Fig. 12.5 Capacity and
cutoff rate of uplink SCMA
in Rayleigh fading
Shannon limits
Mutual information
Cutoff rate
2 users
4 users
6 users
3
Mutual Information [bits/s/Hz]
391
2.5
2
1.5
1
0.5
0
−10
−5
0
5
10
15
20
SNR [dB]
and the Bhatacharyya parameter is
∆x j,a ,x j,b
,# 2
p y | x j,a , h p y | x j,b , h dy
= Eh
⎡
⎤
#
1
= Eh ⎣ J −1
p (y|Xb , h)dy⎦
p (y|Xa , h)
M
x ∈X ,∀i= j
x ∈X ,∀i= j
i,a
i,b
i
i
(12.33)
In the integral of (12.33) a square root of the double sum of the products
“ p (y|Xa , h) p (y|Xb , h)” is involved, which makes it excessively complex for a
large number of users. Hence, we will attempt to obtain reasonable bounds for the
cutoff rate. To deal with the expression, we first calculate ∆Xa ,Xb . According to the
PEP analysis in (12.21) for downlink SCMA, the channel-dependent metric is equal
to
K
τk2 |h[k]|2
diag(h)(Xa − Xb ) 2 =
k=1
where τk2 is the dimension-wise distance defined in Definition 2, and |h[k]| is the
Rayleigh distributed random variables. Similar to that in the uplink case, the Bhatacharyya parameter considering the superimposed codewords Xa and Xb is given
by
1
0 1
−
diag(h)(Xa −Xb ) 2
∆Xa ,Xb = Eh e 4N0
−1
K
1
0 1 K 2
τ2
τ |h[k]|2
−
1+ k
= Eh e 4N0 k=1 k
=
4N0
k=1
392
Z. Ma and J. Bao
By applying Holder’s inequality,
p (y|Xa , h)
p (y|Xb , h) ≥
xi,b ∈Xi ,∀i = j
xi,a ∈Xi ,∀i = j
(xi,a ,xi,b )∈P i ,∀i = j
.
p (y|Xa , h) p (y|Xb , h)
where Pi is the set of a point-to-point pairing of codewords (xi,a , xi,b ) for all i = j,
which contains M elements.3 Then it holds that
⎡
⎤
# .
1
∆x j,a ,x j,b ≥ Eh ⎣ J −1
p (y|Xa , h) p (y|Xb , h)dy⎦
M
(xi,a ,xi,b )∈P i ,∀i= j
=
1
M J −1
∆Xa ,Xb
(xi,a ,xi,b )∈P i ,∀i= j
Thus, an upper bound on the cutoff rate of downlink SCMA is
R0 |
upper
= log2 M − log2
,
1
1+ J
M x x =x
j,a
j,b
−1 τk2
1+
4N0
k=1
K
j,a
(xi,a ,xi,b )∈P i ,∀i= j
(12.34)
For the sake of deriving the lower bound of the cutoff rate, we may invoke the
following simple inequality
p (y|Xb , h)
xi,b ∈Xi ,∀i= j
xi,a ∈Xi ,∀i= j
≤
p (y|Xa , h)
xi,a ∈Xi ,∀i= j xi,b ∈Xi ,∀i= j
.
p (y|Xa , h) p (y|Xb , h)
we get that
∆x j,a ,x j,b ≤
1
M J −1
∆Xa ,Xb
xi,a ∈Xi ,∀i= j xi,b ∈Xi ,∀i= j
and a lower bound on the cutoff rate is obtained
,
R0 |lower = log2 M − log2 1 +
1
MJ x
K
j,a x j,b =x j,a xi,a ∈Xi ,∀i = j xi,b ∈Xi ,∀i = j k=1
,
1
= log2 M − log2 1 + J
M
K
Xa Xb ,x j,b =x j,a k=1
3 There
1+
τk2
4N0
1+
τk2
4N0
−1 -
−1 -
(12.35)
are M! possible pairing patterns for (xi,a , xi,b ), hence M! choices for Pi . The tightness of
the bound is determined by the specific selection of the pairing patterns. A detailed seek for the
appropriate pairing pattern can be found in [19].
12 Sparse Code Multiple Access (SCMA)
(a) Cutoff rate.
393
(b) ACEP.
Fig. 12.6 Cutoff rate and ACEP of the worst user for downlink SCMA in Rayleigh fading
As discussed in (12.24), the expression inside the square bracket of (12.35) is the
union-Bhatacharyya bound on the ACEP for the jth user.
With the derived upper and lower bound on R0 or ∆x j,a ,x j,b , the corresponding
bounds to the union-Bhatacharyya bound on ACEPs, i.e., M1
x j,a
x j,b =x j,a ∆x j,a ,x j,b ,
can be obtained straightforwardly.
We will now verify the cutoff rate bounds and the corresponding bounds for
ACEPs by simulations. If R0 |upper is sufficiently tight, it may be regarded as satisfactory approximation of R0 . In order to emphasize the primary common characteristic
between R0 |upper and R0 |lower , we can readily refer to R0 |upper as the approximated
Chernov bound (Approxi. CB), and R0 |lower as relaxed Chernov bounds (Relaxed
CB). The results for the cutoff rate of the worst user in downlink SCMA and the corresponding ACEPs are plotted in Fig. 12.6a and Fig. 12.6b, respectively. The curve
of “Approxi. CB” and that of “Relaxed CB” merge with each other in the high-SNR
region, while the “Approxi. CB” bound gets significantly close to the associated
practical performance. Hence we may claim that R0 |upper indeed represents a satisfactory approximation of R0 . This conclusion may be verified by the associated
simulation results shown in Fig. 12.6b, where the “Approxi. CB” over ACEP gives a
better estimation of the practical ACEP than that of “Relaxed CB”, for both full-rank
(with 4 users) and rank-deficient (with 6 users) SCMA systems.
12.3 Codebook Design
As the performance of SCMA strongly depends on the multi-dimensional codebooks, codebook design constitutes one of the most important issues for SCMA, and
it is what distinguishes SCMA from other non-orthogonal multiple access schemes.
394
Z. Ma and J. Bao
Incorporating a sophisticated codebook design into SCMA has the potential of significantly improving the spectrum efficiency, and reducing the detection complexity.
12.3.1 General Design Rules
The design of SCMA codebook is a joint optimization of the sparse mapping matrix
and the multi-dimensional constellations. Assume that all layers have the same constellation size and length. An SCMA codebook can be represented by structure
S(V , C ; J, M, N , K ), where V = {V j } Jj=1 , is the set of mapping matrices, and
C = {C j } Jj=1 , is the set of signal constellations for J layers. Thus, the SCMA codebook designing is equivalent to solve the optimization problem [1]
V + , C + = arg max Υ (S(V , C ; J, M, N , K ))
V ,C
(12.36)
where the function Υ (·) is somehow the design criterion.
Unfortunately, for a given criterion Υ (·) and such a multi-dimensional problem,
the optimum solution cannot be found. In practice, a suboptimal multi-stage optimization approach is adopted, by optimizing the mapping matrices and constellations
separately. The set of mapping matrices V is generally selected in order to meet the
maximum overloading, while the design of J multi-dimensional constellations is
simplified to the design of a mother constellation and multiple layer-specific operators.
12.3.1.1
Mapping Matrices
The set of mapping matrices V should be pre-determined before the constellation
design, since it determines the number of users/layers interfering at each resource
node and complexity of the multi-user detection. As V can be characterized and
uniquely determined by the factor graph matrix, the design of V can borrow the idea
from the design of LDPC codes. However, here we introduce general rules for the
designing:
• V j ∈ B K ×N , and Vi = V j , ∀i = j
[φ]
[φ]
• V j = I N , where V j is obtained by removing all-zero rows in V j
Thus we may insert K − N all-zero row vectors into rows of I N to obtain the unique
solution V + for problem (12.36).
If we take Example 1 as the illumination, we have following properties and relations for SCMA encoding parameters
• Choose the constellation length N = 2,
the codebook length K = 4
and
K
=6
• The maximum number of layers J =
N
12 Sparse Code Multiple Access (SCMA)
395
• The number of multiplexed layers over each resource d f = JKN = 3
• Overloading factor λ = KJ = 1.5
• max(0, 2N − K ) ≤ l ≤ N − 1, where l is the number of the overlapping elements
of any two distinct f j vectors. Thus 0 ≤ l ≤ 1 if K = 4 means that the codeword
are either orthogonal or collide at only one overlap nonzero element over any two
rows.
The resulting factor graph matrix F is the same as (12.5) and the factor graph is
shown in Fig. 12.3.
12.3.1.2
Multi-dimensional Constellations
Having the mapping set V + , the optimization problem of an SCMA is reduced to
C + = arg max Υ S(V + , C ; J, M, N , K )
C
(12.37)
which is to find J different N -dimensional complex constellations, each contains M
signal points. In general, the joint design of multiple multi-dimensional constellations
is challenging, a further simplification of (12.37) can be conducted by dividing the
problem into the design of a mother constellation and J layer-specific operators,
and optimizing them separately. Without loss of generality, define C j ≡ Θ j (C), ∀ j,
where Θ j (·) denotes a constellation operator. Thus the optimization problem in
(12.37) becomes
4J
3
C+ , Θ +j
j=1
= arg
5
8
6
7J
max Υ S(V + , C ≡ Θ j (C) j=1 ; J, M, N , K )
J
C,{Θ j } j=1
(12.38)
A. Mother Constellation
In general, a constellation with large minimum Euclidean distance achieves good performance when no collisions occur among users/layers over a tone. With increasing
number of users/layers, the collisions are unavoidable and the multi-user interference will be introduced. To mitigate such interference, it is required to induce dependency among the nonzero elements of the codewords, such that the receiver can
recover colliding codewords from other tones. In general, the mother constellation
can be any form of a multi-dimensional constellation with a maximized minimum
Euclidean distance. To control dimensional dependency and power variation without destroying the Euclidean distance profile, a unitary rotation can be applied to
the mother constellation. For transmission over fading channels, the performance is
dominated by the product distance of a constellation at high-SNR region. Thus the
goal of designing a good mother constellation for SCMA is trying to optimize both
the minimum Euclidean distance and product distance. Fortunately, the optimization
of the product distance could be realized by unitary rotation as well. Thus the two
types of distances can be optimized separately. In [20], using the Chernoff bounding
396
Z. Ma and J. Bao
technique, it is shown that for Rayleigh fading channels, the error probability of a
multi-dimensional signal set is essentially dominated by four factors. To improve
performance is necessary to
• minimize the average energy per constellation point;
• maximize the modulation or signal-space diversity;
• maximize the minimum product distance
d p,min = min
xa ,xb
xa [k]=xb [k]
|xa [k] − xb [k]|
(12.39)
between any two points xa and xb in the constellation;
• minimize the product kissing number for the minimum product distance, i.e., the
total number of points at the minimum product distance.
For low rates, constellation design can be done by brute-force searching, however,
this is not necessarily the case for higher rates and a larger number of users/layers
due to the prohibitive searching complexity. Under this circumstance, the structured
construction is required. Lattice constellation construction can be considered as a
possible way to design good mother constellations. If we construct a constellation
from the lattice Z2N with gray labeling, the construction could be done effectively by
forming orthogonal QAM constellations on different complex planes. To maximize
the minimum product distance of rotated lattice, the unitary rotations of QAM lattice
constellations might be optimized as in [20, 21].
B. Constellation Operators
After obtaining the mother constellation C+ , layer-specific operators should be
designed to guarantee the unique decodability of the multi-layer signals at the
receiver, and also lower the multi-user interference. The optimization problem for
the operators can be formulated as
{Θ +j } Jj=1 = arg max Υ S(V + , C ≡ {Θ j (C+ )} Jj=1 ; J, M, N , K )
{Θ j } Jj=1
(12.40)
Note that here, the design criterion Υ (·) are not necessarily the same as that in (12.38)
for the joint design of mother constellation and constellation operators.
The constellations for different SCMA layers might be constructed with different
operators Θ j (·), and the constellation operators generally include complex conjugate, phase rotation and dimensional permutation. Generally speaking, if the different users have different power levels, the interfering codewords would be easily
separated at receiver due to the power diversity. To do this, it is obliged to have a
diverse average power level over the constellation dimensions when designing the
mother constellations, which could be done by an appropriate rotation of the lattice
constellation as discussed in [16]. Thus the task of optimization problem can be the
permutation operators which enable the SCMA codebooks to capture as much power
diversity as possible over the interfering users. The optimization for power variation
12 Sparse Code Multiple Access (SCMA)
397
over users can be designed to permute each codebook set to avoid interfering with
the same dimensions of a mother constellation over a resource node.
As discussed in Sect. 12.2.1.2, the constellation operators is unnecessary for the
uplink SCMA in fading channels. On the one hand, in MAC, the fading itself takes the
role of constellation operations, and the receiver exploiting the differences among the
channel fadings to separate the multi-user signals. On the other hand, the constellation
operators like phase rotation and complex conjugate don’t change λ2k in Definition
1, hence don’t change the error probability. However, it is important to design layerspecific operator for downlink SCMA, because all users experience the same channel
condition and the destructive codeword collision can be avoided by careful design
of Θ j (·) in the downlink.
12.3.1.3
Constellations for Lower Receiver Complexity
This part introduces two kinds of multi-dimensional constellations for SCMA, that
allow MPA receiving with reduced complexity.
A. Shuffled Multi-dimensional Constellation
The dependency among the complex dimensions of the mother constellation guarantees an efficient detection and diversity for fading channels. It is possible to construct
a mother constellation such that the real part and imaginary part are independent with
each other, while the complex dimensions are still dependent. One kind of approach
is the shuffling [16], which enables the MPA to reduce the complexity from M d f
to M d f /2 . The shuffling method rotates two independent N -dimensional real constellations to maximize the minimum product distances, with the same or different
unitary rotations, then generates an N -dimensional complex mother constellation
by concatenation of the two N -dimensional rotated real constellations. One of the
two N -dimensional real constellations corresponds to the real part of the points of
Fig. 12.7 Shuffling construction of the mother constellation [16]
398
Z. Ma and J. Bao
Fig. 12.8 An example of shuffling construction of two-dimensional 16-ary SCMA constellation
[16]
the mother constellation, and the other one corresponds to the imaginary part. The
construction is illustrated in Fig. 12.7.
Example 2 The construction of a 16-point SCMA mother constellation applicable to
codebooks with two nonzero position
5 (N√ =8 2) by shuffling is illuminated in Fig. 12.8.
Its optimum rotation angle is tan−1 1+2 5 , which maximizes the minimum product
distance.
B. Low-Projected Multi-dimensional Constellation
A key feature of SCMA codebooks is that the multi-dimensional constellation allows
a few constellation points to collide over some of the dimensions, as they can still be
separated through other components. An example is shown in Fig. 12.9, in which the
constellation points corresponding to 01 and 10 collide over the first dimension, but
are separated over the second tone, making the number of projection points equal to
3 instead of 4. By employing this low-projected constellation, the MPA receiver is
able to reduce the number of probability calculations at the FNs during each iteration.
d
As a result, the complexity is reduced to O(M p f ), where M p ≤ M is the number of
projection points.
To do this, it is obliged to let the minimum “product distance”4 be zero during
the mother constellations design by rotation. However, the zero minimum product
4 This
is the relaxed product distance that takes the product of all the dimension-wise distance
between two points into consideration.
12 Sparse Code Multiple Access (SCMA)
399
Fig. 12.9 An example of a
4-ary constellation with 3
projections per complex
plane
distance would cause the performance degradation at high SNR, thus the tradeoff between the performance and complexity should be considered for different
scenarios.
12.3.2 Multi-user Codebooks Design for Uplink SCMA
Systems
In this subsection, we introduce a practical codebook design approach for uplink
SCMA systems over Rayleigh fading channels. Instead of optimizing the mother
constellation and constellation operators separately, we address the joint design of
multi-user constellations for small constellation size and number of users [22].
12.3.2.1
Design Criterion
To address the design of good codebooks, we need to establish appropriate performance criteria for a given system, i.e., determine the Υ (·) in (12.37). It is straightforward to use the DCCC I (X, y) in (12.28), or the ACEP in (12.24), as the criterion, for
increasing capacity or lowering probability of error. However, it is inefficient to use
the DCCC or the ACEP as the cost function directly, since the evaluation of I (X, y)
involves either Monte Carlo simulations or a large amount of numerical integration,
and the calculation of the union bound on the ACEP is a little bit complicated.
As an alternative metric, the cutoff rate also gives an approximated evaluation for
the capacity as well as the error probability and allows us to optimize the codebooks
at a target value of SNR. Therefore, we can formulate the criterion for the multi-user
codebooks design, by making the cutoff rate as large as possible, or equivalently
the union-Bhatacharyya bound on the ACEP as small as possible. According to the
cutoff rate analysis in (12.32) for MAC, maximizing R0 is equivalent to choose the
combined codewords such that
400
Z. Ma and J. Bao
−1
λ2
1+ k
4N0
,b=a k=1
K
{X+j } Jj=1 = arg min
{X j } Jj=1
E
s.t.
X
Xa Xb
2
(12.41)
= Es
where E s is the total power constraint on the multi-user codebooks. It is expected
that the criterion is optimal in the sense of designing codebooks with large DCCC
or small ACEP, since it involves all pairs of possible codewords for multiple users.
12.3.2.2
Signal-Space Diversity Scheme
Recall that the problem of designing multi-dimensional constellations for fading
channels has been solved by using signal-space diversity (SSD), which rotates the
QAM constellations with a unitary matrix, constructed either from the algebraic
number theory or by computer search [21, 23]. Therefore, here, we use the rotated
constellation to build the multi-user codebooks. In particular, we obtain the rotation
matrices through computer search over compact parameterizations of unitary matrices. Note, that the null dimensions of codebooks are discarded before the rotations.
The N × N unitary matrix can be written as [23]
R=
N −1
N
Tm,n
(12.42)
m=1 n=m+1
where Tm,n is a complex Givens matrix given by [24]
⎡
Tm,n
⎤
..
⎢ .
⎢
cos θm,n
⎢
⎢
..
=⎢
.
⎢
iηm,n
⎢
−e
sin θm,n
⎣
. . . e−iηm,n sin θm,n
..
..
.
.
...
cos θm,n
..
.
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
which changes the identity matrix by replacing its (m, m)th, (n, n)th, (m, n)th and
(n, m)th elements with cos θm,n , cos θm,n , e−iηm,n sin θm,n and −eiηm,n sin θm,n , respectively. The angles satisfy θm,n ∈ [−π, π ] and ηm,n ∈ [−π/2, π/2]. So, the search for
R is reduced to the search for a sequence of phase of
θ = {θ1,2 , . . . , θ1,N , θ2,3 , . . . , θ2,N , . . . , θ N −1,N },
η = {η1,2 , . . . , η1,N , η2,3 , . . . , η2,N , . . . , η N −1,N }.
(12.43)
It seems that optimizing the constellations using the above criterion is intractable
even for a moderate number of users and codebook size, since we need to search
12 Sparse Code Multiple Access (SCMA)
401
N (N − 1)J angles, and the summation in the right-hand side of (12.41) will add up
M J (M J − 1) terms. However, searching results for two-dimensional constellations
with a small number of users show that the rotation matrices are the same for all
codebooks, and are independent of the number of users. Therefore, we simplify
the optimization process by searching over a single rotation matrix, and reducing
the number of accessed users, even though this is suboptimal. Furthermore, we can
use the approach developed in [25], where all the entries of the rotation matrix6are7
equal in magnitude. Therefore, by expanding the product in (12.42), we get θ = π4
7
6
and θ = π4 , 0.6155, π4 for N = 2 and N = 3, respectively. Exhaustive search is
computationally feasible, provided, that each user occupies a moderate number of
resources such that N ≤ 3.
In the signal-space diversity scheme, the constellations are restricted to lattice
constellations such that the rotated QAMs are suggested. In practice, the rotation
can be done over any multi-dimensional constellations to improve the cutoff rate,
e.g., the rotated spherical codebook [26] and rotations over the product of other
low-dimensional constellations.
12.3.2.3
Simulations and Discussions
Consider the SCMA system in Example 1, which supports an overloading factor
150%. Simulation results of packet error rate (PER) for uncoded SCMA and DCCC
are provided, which are performed over i.i.d Rayleigh fading channels for 4-ary and
16-ary codebooks. Four kinds of codebooks including the codebooks through SSD
scheme discussed in this subsection (named as G4,4 /G16,16 ), the codebook from [16]
(named as C4,4 /C16,16 ), spherical codebook [26], star-QAM-based codebooks [27],
are employed, and we also provide the results of the star-QAM-based codebooks
after optimization using the criterion (12.41), for which we extend α to complex
numbers and get β = 1, α = −i and α = 0.8 − 0.8i for 4-ary and 16-ary codebooks,
respectively.5
Figure 12.10 plots the DCCC of 4-ary and 16-ary codebooks for four users,
together with the theoretical limit rates of i.i.d Gaussian inputs. As it is evident,
the SSD scheme outperforms all the other codebooks in the high rate region for both
4-ary and 16-ary codebooks, while the mutual information gain is more clear for
4-ary case. While the rate of the star-QAM scheme is small, a significant gain is
achieved after optimization with the criterion in (12.41), and it becomes as good as
the SSD scheme for 4-ary codebook.
Figure 12.11 compares the PER performance of different codebooks for uplink
SCMA with six users, where two antennas are employed for receive diversity, and
the MPA detector is used with six iterations all the time. As it is observed, the SSD
scheme has a gain about 0.8 dB over C4,4 and 0.6 dB over C16,16 , and a gain about 0.5
dB and 0.3 dB over the spherical codebook for 4-ary and 16-ary cases, respectively.
5 The
star-QAM-based codebook targets on downlink channels, while its performance deteriorates
in the uplink and for large constellation size.
402
Z. Ma and J. Bao
4
Shannon limit on 4 users
C4,4
2
Shannon limit on 4 users
C
16,16
Star−QAM
Star−QAM with optimization
Spherical codebook
SSD − G
1.8
Mutual Information [bits/s/Hz]
Mutual Information [bits/s/Hz]
2.2
4,4
1.6
1.4
1.2
1
1.7
0.8
0.6
1.62
4.8
0.4
0
2
4
Star−QAM
Star−QAM with optimization
Spherical codebook
SSD − G16,16
3.5
3
2.5
2
3
1.5
5.8
6
8
6
1
10
2
3
4
5
6
7
8
7
9
10
Eb/N0 [dB]
Eb/N0 [dB]
(a) 4-ary SCMA codebooks.
(b) 16-ary SCMA codebooks.
Fig. 12.10 Mutual information of uplink SCMA with 4 users
0
10
0
10
−1
10
−1
PER
PER
10
−2
10
−2
10
C
4,4
−3
10
10
C
Star−QAM
Star−QAM with optimization
Spherical codebook
SSD − G4,4
12
14
16,16
−3
16
18
E /N [dB]
b
Star−QAM with optimization
Spherical codebook
SSD − G16,16
0
(a) 4-ary SCMA codebooks.
20
10
10
12
14
16
18
20
E /N [dB]
b
0
(b) 16-ary SCMA codebooks.
Fig. 12.11 PER of uplink SCMA systems over Rayleigh fading channels
Without optimization, the star-QAM scheme yields the worst error performance.
However, it performs much better after optimization, which coincides with the result
of mutual information in Fig. 12.10.
12 Sparse Code Multiple Access (SCMA)
403
12.3.3 Low-Projected Multi-dimensional Constellations
Design
As is discussed above, by employing the multi-dimensional constellations with low
projections, the MPA receiver is able to utilize the constellation structure to reduce
the receiver complexity. This subsection introduces an approach of constructing
low-projected multi-dimensional constellations for uplink coded SCMA. In particular, constellation optimization for bit-interleaved convolutional coded SCMA with
iterative multi-user detection is considered.
12.3.3.1
Transfer Characteristics of Turbo-MPA Detector
Extrinsic information transfer (EXIT) characteristics are investigated to find the
effect of multi-user constellations on the performance of the MPA detector, and
give us insights on the constellation optimization criteria. For each user, the EXIT
chart analysis computes the average mutual information (AMI) between the extrinsic
LLR (L e ), or the a priori LLR (L a ), and each coded bit. Thus, the extrinsic AMI is
calculated as [28]
1
Idet,e = 1 − .
2π σe2
#
+∞
−∞
2 +
− l − σe2 /2
log2 (1 + e−l )dl
exp
2σe2
*
1
1
1 or 2 users
1 or 2 users
0.8
0.6
0.6
6 users
I
I
det,e
det,e
0.8
6 users
0.4
4 users
0.4
4 users
8−ary constellation,
Eb/N0 = 6 dB, Rc = 0.5
8−ary constellation,
Eb/N0 = 6 dB, Rc = 0.5
0.2
0.2
Labeling 1
Labeling 2
Labeling 3
Constellation 1: G
8,8
Constellation 2: LDS − 8PSK
0
0
0.2
0.4
0.6
0.8
1
Idet,a
(a) EXIT charts for different constellations.
0
0
0.2
0.4
0.6
0.8
1
Idet,a
(b) EXIT charts for different labelings.
Fig. 12.12 Impacts of constellations and labelings on the detector’s transfer characteristics
(E b /N0 = 6 dB, MPA detector with 3 iterations)
404
Z. Ma and J. Bao
where σe2 is the variance of the extrinsic LLR. It is worth noting that due to the
multi-user interference, the a priori AMI (Idet,a ) and extrinsic AMI (Idet,e ) for each
user will be influenced by the other users, and hence, a J -dimensional EXIT chart
is necessary to characterize the transfer function. Here, the AMI is averaged over all
the users such that the EXIT curves can be depicted on a one-dimensional complex
plane.
Now, we investigate the transfer characteristics of the turbo-MPA detector for
uplink SCMA over i.i.d. Rayleigh fading, and a factor graph matrix as in (12.5)
is considered. Figure 12.12 presents the detector’s transfer characteristics of twodimensional 8-ary constellations for different number of users (J ) at E b /N0 = 6 dB,
where the MPA detector performs 3 iterations. Note that when J = 2, the signals
from the two users are orthogonal with each other, so that they have the same AMI
as that in the single-user case.
The impact of different constellations on the detector’s transfer characteristics
is shown in Fig. 12.12a, where the detector’s EXIT curves of two different 8-ary
constellations6 with the same labelings are provided. Obviously, constellation 1 outperforms constellation 2, and the superiority of constellation 1 over constellation 2
is independent with the number of users. This implies that the effect of the constellation on the single-user system agrees with its multi-user counterpart, even though
the EXIT curves become steeper as the number of users increases. In Fig. 12.12b,
the results for the same 8-ary constellation with different mappings/labelings are
demonstrated. It is observed that different labelings result in transfer characteristics
curves of different slopes, for all the number of users cases, and the labeling with a
steeper EXIT curve in the single-user case shows the larger slope in its multi-user
counterpart.
The conclusion implies that the influence of constellations and labelings on the
single-user system is consistent with that on the multi-user case. More precisely, a
constellation or a labeling that is good for single-user systems will be beneficial to
the multi-user systems. Therefore, we suggest to simplify the complicated multi-user
constellations optimization in SCMA to the suboptimal single-user system design.
It is expected that the constellation designing criteria for the single-user system is
efficient for multi-user cases.
12.3.3.2
Design Criteria of Multi-dimensional Constellations
A. Links Between EXIT Charts and Constellation Design
The EXIT chart is a good tool to guide the system design. For iteratively decoded
systems, given an outer convolutional code, the constellation should be designed to
form a tunnel between the transfer curves of the detector and the decoder, and the
6 The
constellation 1 is constructed by rotation over the product of a binary phase-shift keying
(BPSK) and a quadrature phase-shift keying (QPSK) constellation with Gray labelings, using the
approach in Sect. 12.3.2.2 (G8,8 ), and constellation 2 is the repetition over an 8PSK constellation
with Gray labeling, i.e., the LDS scheme [29].
12 Sparse Code Multiple Access (SCMA)
405
starting point of the detector curve and the intersection point between the detector
curve and the decoder curve should be as high as possible, to guarantee a low threshold
as well as a low error floor.
At a given value of SNR, the transfer characteristics of the detector are affected
by both the constellation itself and the labeling, as shown in Fig. 12.12. In terms
of the constellation, it is known that the area under the detector’s EXIT curve is
approximately equal to the DCCC per number of bits of a constellation point [30].
Based on this property, once a constellation with a larger DCCC is constructed,
a larger area is obtained and then it has the potential of providing a wider EXIT
tunnel, or equivalently, it would be easier to let the detector’s curve to be above
the decoder’s curve. In terms of the labeling, for a given constellation, the detector
curves corresponding to distinct labelings are rotations with each other, since the
labeling does not change the DCCC and hence the area below the detector curve.
A good labeling rotates the detector curve such that a large AMI is produced when
Idet,a = 1, which provides an error floor that reaches the BER range of practical
interest, and at the same time, to make the tunnel between the transfer curves of the
detector and the decoder still open.
Based on these facts, we divide the constellation optimization framework into
two steps. First, try to design the multi-dimensional constellation by maximizing the
DCCC; Second, optimize the labeling by EXIT curve-fitting. In the following, we
introduce two figures of merits for the constellation and the labeling.
B. Constellation Figure of Merit
As discussed above, the cutoff rate, corresponding to the DCCC, is a good criterion
that allows us to optimize the constellation at a target value of SNR. Considering
the received signal y = diag(h)x + n, the cutoff rate constrained by an M-ary K dimensional signal set C in i.i.d. Rayleigh fading, is given by [25]
⎡
1
ΨCFM (C) = log2 M − log ⎣1 +
M x ∈C x ∈C,x =x
a
b
b
a
⎤
−1
δk2
⎦
1+
4N
0
k=1
K
(12.44)
where δk = |xa [k] − xb [k]|, is the dimension-wise distance between any two distinct K -dimensional symbols xa and xb . We take the quantity ΨCFM (C) as the SNRdependent constellation figure of merit, which is a function of SNR and the constellation C, or the set of all pairwise distances between the constellation points. It
involves all pairs of multi-dimensional symbols, and is independent of the labeling
or any channel codes.
C. Labeling Figure of Merit
The constellation labeling is a crucial design parameter to achieve a high coding
gain over the iterations for iteratively decoded bit-interleaved coded modulation
(BICM) systems. To obtain an optimization criterion for the labelings, we resort to
the error performance of multi-dimensional constellations under ideal interleaving.
Let x̃(i) = [x̃(i) [1], . . . , x̃(i) [K ]]t , be the symbol having the same label with that of x
406
Z. Ma and J. Bao
except at the ith bit position. The effect of labeling µ on the performance of BICM
with iterative decoding (BICM-ID) systems employing multi-dimensional signal
constellation can be characterized by [31]
1
m
K
1
1 2 −1
1+
ΨLFM (µ) =
δ
m M i=1 b=0
4N0 k
b k=1
(12.45)
x∈Ci
where δk = |x[k] − x̃(i) [k]|, and Cib is the subset of C that consisting of symbols
whose label has the value b in the ith bit position. The SNR-dependent object function
ΨLFM (µ) is able to characterize the influence of both the constellation C and the
labeling µ to the bit error rate (BER) performance of BICM-ID systems. With this
criterion, one can optimize the bit labeling when fixing the signal constellation,
or optimize the constellation for a given labeling, or optimize them jointly. Since
optimizing the labeling µ by decreasing ΨLFM (µ) improves the BER performance,
we take ΨLFM1 (µ) as the labeling figure of merit to guide the labeling design for a given
multi-dimensional constellation.
12.3.3.3
Design Multi-dimensional Constellations
The multi-dimensional constellation with the same projections over each dimension
can be viewed as a multi-modulation scheme [32], where the data bits are modulated into multiple one-dimensional symbols that are chosen from a one-dimensional
complex constellation A, called subconstellations in the following. The difference
among the modulations for each dimension is that they have different labelings. This
implies that the multi-dimensional constellation can be constructed by permutations
of the one-dimensional subconstellation A, dimensionally. Therefore, the problem
is to design an M-ary subconstellation A with M p distinct signal points, and the
specific mapping or permutation for each dimension. In the following, we propose
a multi-stage optimization, and the K -dimensional constellation is constructed by
three steps:
(a) Determine the desired number of projection points M p such that M p ≤ M,
choose a one-dimensional M-ary subconstellation A with M p projections;
(b) Based on the one-dimensional subconstellation A, construct a K -dimensional
constellation C using permutations;
(c) Design a labeling for the K -dimensional constellation C.
A. Design One-Dimensional Subconstellation A
Different from the traditional constellation design, the M-ary constellation with M p
projections imply that there are M − M p signal points that overlap with others. We
first choose an M p -ary constellation A p without overlappings, then allocate the M p
signal points with M labels to obtain A. The choice of A p is various, any onedimensional complex constellation, e.g., quadrature amplitude modulation (QAM)
12 Sparse Code Multiple Access (SCMA)
407
or phase-shift keying (PSK), is available. Here, we construct A p using the amplitude
phase-shift keying (APSK) constellation, since it is able to provide good DCCC
compared to other conventional modulations [33, 34].
An M p -APSK constellation is composed of L concentric rings, each with uniformly spaced PSK points. The M p -APSK constellation can be expressed as [33]
6
7
A p = r1 e jθ1 P(m 1 ), . . . , r L e jθL P(m L )
where P(m l ) is an m l -ary PSK constellation with unit average energy, and rl , θl are
the radius and phase offset of the lth ring, respectively. Let m= [m 1 , . . . , m L ]t , be
L
m l . To guarantee
the vector of the number of points over each ring so that M p = l=1
a good distance profile, it is preferred to locate fewer constellation points on the inner
rings than that on the outer rings. Then, for a set of ordered radius r1 < · · · < r L , it
is suggested that m 1 ≤ · · · ≤ m L .
Following the general APSK design procedure proposed in [34], the M p -APSK
constellation can be constructed as
• Select the number of
rings L and the number of constellation points on each ring
L
m l , such that M p = l=1
ml ;
• Determine the radius of each ring rl ,
&
*
l−1
+
'
'
m
1
l
mi +
rl = (− ln 1 −
;
M p i=1
2
• Set θl as 0 or π/m l .
Given the designed M p -APSK constellation A p , we allocate the M-ary constellation with the M p signal points. The problem can be formulated as how to put M
numbers, 0, 1, . . . , M − 1, into M p sets, where each set represents a signal point in
A p . The allocation strategy is preferred to follow the rules:
• The numbers that are allocated to a set should be less than or equal to M p , and
greater than or equal to 1, such that the overlapped points can be separated through
other dimensions;
• The numbers in each set should be as less as possible, such that the resulted multidimensional constellation has a good distance profile;
• Symmetry of the constellation A is preferred so that it has a zero mean;
• The sets with low power levels may be allocated with more numbers, such that the
constellation has a small average energy.
Note that in some cases, the allocation yields a constellation A with nonzero mean,
then we shift A toward the origin, such that the mean of all signals is zero and
therefore more energy-efficient.
Now, we give an example to illustrate the allocations. For a given 9-APSK A p that
is constructed with 3 rings and m = [1, 3, 5]t , a 16-ary subconstellation A can be
obtained by allocating 16 numbers into 9 sets. Some possible allocations are given
408
Z. Ma and J. Bao
in Fig. 12.13. Among the four strategies A, B, C, and D, while the strategy A is
the most energy-efficient, it shows the worst performance when used to construct
multi-dimensional constellations. This is because too many points overlap with each
other, leading to a very poor distance profile for the multi-dimensional constellation.
Numerical results show that the strategies B and C are equally efficient, and the
strategy D is the best one, since the largest number of overlappings is only two.
B. Construct K -dimensional Constellation C
Given the designed M-ary subconstellation A, denote πk (A) a column vector of
the kth permutation of the signals in A, and let π1 (A) = A. The K -dimensional
constellation through the permutation construction can be expressed with a K × M
matrix as
C = [π1 (A), . . . , π K (A)]t
where each column of the matrix corresponds to a K -dimensional symbol. Then,
constructing a K -dimensional constellation requires to find K − 1 permutations
π2 , . . . , π K , such that the constellation figure of merit ΨCFM (C) in (12.44) is maximized.
We focus on two-dimensional constellations (K = 2), by maximizing the constellation figure of merit, the unique permutation function π is selected as
−1
|xa [k] − xb [k]|2
1+
4N0
k=1
K
π = arg min
π̂
xa ∈C xb ∈C,xb =xa
There are M! different choices for the permutations, for small constellation size where
M ≤ 8, the optimum solution can be solved by exhaustive search with a reasonable
complexity. However, it becomes intractable for high order constellations. Note that
this problem is similar to the labeling map of a constellation in bit-interleaved coded
modulation with iterative decoding (BICM-ID) systems, which can be efficiently
solved by using the binary switching algorithm (BSA) [35], or iteratively searching
inside a randomly selected list, and a local optimum permutation can be found for a
given cost function.
As for a larger dimension where K > 2, the search for K − 1 permutations is
challenging. A suboptimal solution can be used by successively optimizing the multidimensional constellation from lower dimensions to higher dimensions, such that
only one permutation is needed to be checked in every round.
C. Labeling the K -dimensional Constellation
When a multi-dimensional constellation is found, we should choose an appropriate
labeling for the constellation. In terms of EXIT chart, optimizing the labeling is to
adjust the slope of the detector’s curve. Our approach is to obtain a set of labelings
with various slopes in their EXIT curves, firstly. Then, to choose a labeling from
the set such that the detector EXIT curve matches with the decoder curve of a given
convolutional code.
12 Sparse Code Multiple Access (SCMA)
(a)
(b)
10
51
11
2
6
9
7
12
7
2
4
13
3
(c)
12 6
1
14 0
8 15 4
409
0
8
9 2
5
15 3
10
15 9
11
7 12
11 6
4 1
0
4
14
3
14
13 1
8
7
10
13
(d)
6
11
5
8
2
13
5
0
3
12
15 10
14
9
Fig. 12.13 Examples of 16-ary subconstellation A based on 9-APSK with m = [1, 3, 5]t
The labeling figure of merit ΨLFM1 (µ) in (12.45) represents the ultimate performance
with perfectly known a priori AMI, and in EXIT charts, it corresponds to the maximum achievable value of Idet,e with Idet,a = 1, denoted as I ∗ , and I ∗ becomes larger
with decreasing ΨLFM (µ). It is observed in Fig. 12.12a, b that for single-user systems,
the detector’s EXIT curves corresponding to distinct labelings can be approximated
to straight lines, with a common intersection around the point with Idet,a = 0.5. Then,
the slope of the EXIT curve corresponding to a labeling can thus be determined by I ∗ ,
approximately. A labeling with a larger I ∗ can have a steeper transfer curve. Therefore, we can use ΨLFM (µ) to approximately control the slope of the EXIT curves,
and the detector’s curve becomes steeper with decreasing ΨLFM (µ).
Denote the set of labelings as Ω. For constellations with small sizes (M ≤ 8),
Ω is chosen to be the set of all possible labelings with distinct ΨLFM (µ). For higher
order constellations, the BSA can be used once again to obtain Ω. Begin with a
given original labeling, by minimizing ΨLFM (µ) using the BSA, new labelings with
increasing slopes may be obtained during the search, we output these labelings and
store them into Ω. Similarly, new labelings with decreasing slopes may be obtained
by maximizing ΨLFM (µ) with the same original labeling. Then, the labelings in Ω
are sorted with increasing values of ΨLFM (µ). The set Ω can also be obtained by
iteratively searching inside a randomly selected list.
Now, we choose a labeling from Ω, with the aid of EXIT chart analysis. At an
appropriate SNR, the following two conditions have to be fulfilled for the labeling:
(a) the slope of either the single-user or multi-user detector EXIT curve should be
as steep as possible, to achieve a low BER error floor;
(b) the tunnel between the decoder and the multi-user detector curves should be
open, or the intersection point between them should be as high as possible, to
guarantee a low threshold.
As an illustrative example, Fig. 12.14 shows the choices of labelings for a twodimensional 8-ary constellation (with 3 projections) and a 16-ary constellation (with
9 projections) for SCMA. The detector curves of several labelings, with distinct
ΨLFM (µ), as well as the decoder curve are provided, where a half-rate four-state
non-recursive convolutional code with generator [5, 7]8 is employed as the outer
channel code, and Idec,a (Idec,e ) denotes the AMI between the a priori (extrinsic)
LLR and the transmitted coded bit at the input (output) of the convolutional decoder.
410
Z. Ma and J. Bao
1
1 user, Eb/N0 = 7.5 dB
0.8
1 user, Eb/N0 = 6 dB
Idet,e/Idec,a
0.6
I
/I
det,e dec,a
0.8
6 users, Eb/N0 = 6 dB
0.6
6 users, Eb/N0 = 7.5 dB
0.4
0.4
Labeling with minimum Ψ
(µ)
Labeling with minimum Ψ LFM( µ )
LFM
Labeling with maximum Ψ LFM( µ )
Labeling with maximum Ψ LFM( µ )
Proposed labeling
CC [5, 7]
0.2
0
0.2
0.4
I
0.6
Proposed labeling
0.2
0.8
1
CC [5, 7]
0
0.2
I
/I
0.6
0.8
1
/I
det,a dec,e
det,a dec,e
(a) Labeling for an 8-ary constellation
with 3 projections.
0.4
(b) Labeling for a 16-ary constellation
with 9 projections.
Fig. 12.14 Examples of labelings for two-dimensional constellations
As it is observed in Fig. 12.14, the labeling with the maximum ΨLFM (µ) yields
a relatively flat slope in the 6 users case, and the one with the minimum ΨLFM (µ)
closes the tunnel between the decoder and the multi-user detector curves. In contrast,
the proposed labeling, which shows a very steep slope in the EXIT curve while still
keeps the tunnel open, achieves a good trade-off between the threshold and the BER
performance.
With the proposed approach, it is possible to construct constellations with any
projections. Figure 12.15a and 12.15b show the examples of the designed twodimensional 8-ary constellations with various projections, and a four-dimensional
64-ary constellation with 16 projections, respectively.
12.3.3.4
Simulations and Discussions
In the following, simulations are conducted to evaluate the performance of lowprojected constellation for an uplink convolutional coded SCMA system. The
detailed simulation configuration is given in Table 12.2. The SCMA system follows a factor graph matrix as (12.5), which supports an effective system loading of
75% (J Rc /K ). For the sake of simplicity, let A M,M p denote the APSK-based M-ary
constellation with M p projections. The constellations in [16, 36, 37] are named as
C M,M p (the SSD scheme in Sect. 12.3.2 is denoted as G M,M p ), which are used as the
benchmark.
The simulated BER performances of 4-ary, 8-ary and 16-ary codebooks are
depicted in Figs. 12.16 and 12.17, respectively. It is obvious that the A M,M p code-
12 Sparse Code Multiple Access (SCMA)
3
6
4
A83
2
7
3
5
2
4
2
5 0
A85
7
2
6
7
7 0
5
3
2
1
7
A87
4
0
5
2
0
1
7
6
6
3
A88
4
2
5
7
37
46
0
50
20
9
53
42
61
6
17
39
62
4
60
36
16
1
49
48
10
22
6
0
3
0
5
54
58
20
44
5
(a) Two-dimensiona l8-ary constellations with
various projections.
0
33
41
40
7
36
31
23 28
59
42
62
24
57
32
46
56
27
5
44
58
30
51
55
22
63
14
15
7
63
23
53
25
45
26
40
12
24 50
55
54 22
62 31
1
32
34
52
33
48
13
45
12
21
19 3
42 5
31
47
13
57
19 17
31 24
3
29
43
51
41
49
8
18
19
59
11
27
28
56
9
35
30 14
59 32
27 20
26 18 0
43 35 52
60 58 23
63
34 6 21 16
51 46 24 47
11
29 7 2
29 15 53
61 50 26
56 48
2 30
21 38
7
6
8
34
1 35
54 17
3
52
4
4
2
19
14
11
39
5
3
1
6
4
5
0
18
16
25
10
2
43
47
37
3
1
4
A86
1
3
2
1
5
2
0
5
7
4
6
3
6
1
6
6
7
4
5 0
0
38
15
2
5
3
6
3
0
6
4 1
2
1
1
7
3
4
A84
4
7
0
1
411
9 1
40 13
8 0
49 21
18 9
28 16
32 15 33
29 23
51
42 20 22 13
60 43 39 57
11
27 25
34 4 61
48 63
59 58
46 12
47 36
44 38
62 50
36 8
45 39
41
61
38 28
57 37
10
44
49 12
60 33
25 4
55 17
14 1
53 26
5
37
10 3
35 30
7
56
41 2
55 52
40 6
45 54
(b) A four-dimensional 64-ary
constellation with 16 projections.
Fig. 12.15 Examples of APSK-based low-projected SCMA constellations
Table 12.2 Simulation parameters
Parameters
Channel model
Target spectral efficiency
Number of users
FEC coding
Interleaving
Codebooks
Receiver
Values
Uplink Rayleigh fading channel
1.5, 2.25, 3 bits/resource
6
1/2-rate convolutional code with generator
[5, 7]8
Random interleaver, interleave length: 1024
bits
C43 /A43 , C44 /A44 /G44 , C83 /A83 , C84 /A84 ,
C85 /A85 , A88 /G88 C169 /A169 ,
C1616 /G1616 /A1616
Turbo-MPA (3 MPA iterations + 6 BICM
iterations)
books outperform others in the large SNR region, for almost all the simulation cases.
In Fig. 12.16a, the BER floor of A4,3 is lower than C4,3 when SNR is less than 8
dB, and A4,4 shows much better performance than C4,4 , and has a gain about 0.25
dB over G4,4 . Note that A4,4 outperforms G4,4 in the whole SNR region. This is
because A4,4 has the same labeling with G4,4 but the larger DCCC. For the 8-ary
codebooks shown in Fig. 12.16b, the error floors of the A M,M p codebooks happen at
the BER level below 10−5 , which are much lower than the other codebooks. Thus,
much smaller values of SNR are required to achieve a BER value of 10−6 . Similar
results are also obtained for 16-ary codebooks, which are shown in Fig. 12.17. The
gain of the A M,M p codebooks over C M,M p is smaller than that in the 8-ary codebooks
case. The BER curve of C16,16 degrades earlier than the others, but it arrives the error
412
Z. Ma and J. Bao
0
0
10
10
C4,4/A4,4/G4,4
A8,3/C8,3
C4,3/A4,3
−1
10
A8,4/C8,4
A88
−1
10
A
/C
A
/G
8,5
A85
8,8
8,5
8,8
A83
−2
−2
10
10
A84
BER
BER
C44
−3
10
C43
−3
10
C85
−4
−4
10
C83
10
C84
A43
G88
G44
−5
10
−5
10
A44
−6
−6
10
1
2
3
4
5
6
7
8
10
9
3.5
4.5
E /N [dB]
b
5.5
6.5
7.5
8.5
9.5
E /N [dB]
0
b
(a) 4-ary codebooks.
0
(b) 8-ary codebooks.
Fig. 12.16 BER performance of uplink coded SCMA for 4-ary and 8-ary codebooks
Fig. 12.17 BER
performance of uplink coded
SCMA for 16-ary codebooks
0
10
−1
10
C1616
C169
A169
−2
10
G1616
BER
A1616
−3
10
−4
10
−5
10
A16,9/C16,9
−6
10
5
A16,16/C16,16/G16,16
6
7
8
9
10
Eb/N0 [dB]
floor quickly. The codebook G16,16 shows a very large threshold, and A16,16 achieves
a good trade-off. Even though the BER threshold of A16,9 is larger than C16,9 , A16,9
shows almost the same BER performance with C16,9 at the SNR around 9.5 dB. To
achieve the BER performance of 10−6 , equal or less SNR are required for the A M,M p
codebooks.
12 Sparse Code Multiple Access (SCMA)
413
12.4 SCMA for 5G Radio Transmission
12.4.1 Application Scenarios for 5G Networks
In addition to achieving higher transmission rates, faster access, supporting of larger
user density and better user experience in enhanced mobile broadband (eMBB),
the 5G air interface connects to new vertical industries and new devices, creating
new application scenarios such as massive machine type communications (mMTC)
and ultra reliable low latency communications (URLLC) services, by supporting
massive number of devices and enabling mission-critical transmissions with ultra
high reliability and ultra-low latency requirement, respectively. This presents new
challenges and considerations for the radio multiple access to be fully scalable to
support these diverse service requirements. The current orthogonal multiple access
might not be able to fulfill some of the requirements, such as services for dense
MTC devices deployments, and SCMA can be considered as a promising candidate
to meet the 5G performance requirements. In particular, SCMA is proposed for 5G
to achieve the following benefits:
• for eMBB: larger capacity region by non-orthogonal multiplexing; robustness to
fading and interference with code-domain design; robust link adaptation with
relaxed CSI accuracy.
• for URLLC: higher reliability through diversity gain achieved by multidimensional constellations, and robustness to collision by carefully design the
codebooks; latency reduction and more transmission opportunities by enabling
grant-free access; Non-Orthogonal Multiplexing of mixed traffic types.
• for mMTC: higher connection density with high overloading; reduction of signaling overhead and power consumption by enabling grant-free access.
Moreover, it is also possible to extend SCMA application to unlicensed spectrum
and V2X systems, since the non-orthogonal transmissions can help to increase the
system efficiency and deal with the interference.
The link-level performance evaluation for some uplink SCMA scenarios is provided in [38], which compares SCMA with orthogonal frequency division multiple access (OFDMA) in typical scenarios and investigate the robustness of SCMA
to overloading and codebook collision. Results show that SCMA achieves significant gain over orthogonal multiple access with good codebook design, and the gain
increases as the supported number of users and target spectrum efficiency increases.
Moreover, high overloading with stable performance is feasible with SCMA design,
which enables robust overloaded transmission, and the performance loss with codebook collision is negligible with SCMA design, which enables robust grant-free
transmission.
414
Z. Ma and J. Bao
12.4.2 Challenges and Future Works
While SCMA is able to greatly enhance the system capability for 5G networks, some
further issues on design and implementation of SCMA remain to be resolved, which
can be listed as follows:
• Reduced complexity receiver design: Even though MPA or EPA receiver is able
to significantly reduce the complexity of SCMA, the complexity of MPA is still
very high and iterative multi-user receiver is usually required, which brings several
challenging issues for practical implementation:
– It limits the capability of SCMA to support massive connectivity;
– The iterative multi-user detection brings a large processing delay;
– The complexity makes it difficult for SCMA to employ constellations with large
sizes, hence limits the transmission rate.
Sophisticated multi-user detection schemes should be developed to address the
high complexity.
• Theoretical analysis: Further theoretical analysis of SCMA is needed to get more
insights on the practical system design. For example, the capacity or error performance with randomly codebook allocations. Also, interference cancelation may
be incorporated into the MPA detection for lower complexity, then it is desirable
to determine the performance and capacity under practical detectors.
• Codebook design: The codebook design is complicated, especially for highdimensional codebooks and that with large size. Advanced multi-dimensional constellation construction is necessary, and the joint design of factor graph matrix and
constellations is to be developed, for further performance improvement. Moreover,
the design for the scenario that all the overloaded users have different transmission
rates (codebook sizes) is to be investigated, to improve the link adaptation.
• Other issues: System scalability of supporting various loading, SCMA in both
uplink and downlink transmissions, supporting of other techniques such as MIMO,
resource/codebook allocation, channel estimation for uplink SCMA, etc.
References
1. H. Nikopour, H. Baligh, Sparse code multiple access, in IEEE 24th International Symposium
on Personal Indoor and Mobile Radio Communications (PIMRC’13) (2013), pp. 332–336
2. S. Zhang, X. Xu, L. Lu, Y. Wu, G. He, Y. Chen, Sparse code multiple access: an energy
efficient uplink approach for 5G wireless systems, in IEEE Global Communications Conference
(GLOBECOM’14) (2014), pp. 4782–4787
3. Y. Wu, S. Zhang, Y. Chen, Iterative multiuser receiver in sparse code multiple access systems,
in Proceedings of IEEE International Conference on Communications (ICC’15) (2015), pp.
2918–2923
4. J. Zhang, L. Lu, Y. Sun et. al., PoC of SCMA-Based Uplink Grant-Free Transmission in UCNC
for 5G. IEEE J. Sel. Areas Commun. 35, 1353–1362 (2017)
12 Sparse Code Multiple Access (SCMA)
415
5. R. Hoshyar, F.P. Wathan, R. Tafazolli, Novel low-density signature for synchronous CDMA
systems over AWGN channel. IEEE Trans. Signal Process. 56, 1616–1626 (2008)
6. D. Guo, C. Wang, Multiuser detection of sparsely spread CDMA. IEEE J. Sel. Areas Commun.
26, 421–431 (2008)
7. R1-166098: Discussion on feasibility of advanced MU-detector. Huawei, HiSilicon, 3GPP
TSG RAN WG1 Meeting #86 (2016)
8. X. Meng, Y. Wu, Y. Chen, M. Cheng M, Low complexity receiver for uplink SCMA system
via expectation propagation, in Proceedings of Wireless Communications and Networking
Conference (WCNC’ 17), (2017), pp. 1–5
9. J. Bao, Z. Ma, G.K. Karagiannidis, M. Xiao, Z. Zhu, Joint multiuser detection of multidimensional constellations over fading channels. IEEE Trans. Commun. 65, 161–172 (2017)
10. D. Tse, P. Viswanath, Fundamentals of Wireless Communications (Cambridge University Press,
2005)
11. J.G. Proakis, M. Salehi, Digital Communications (McGraw-Hill, New York, 2008)
12. E. Björnson, D. Hammarwall, B. Ottersten, Exploiting quantized channel norm feedback
through conditional statistics in arbitrarily correlated MIMO systems. IEEE Trans. Signal
Process. 57, 4027–4041 (2009)
13. M. Chiani, D. Dardari, M.K. Simon, New exponential bounds and approximations for the
computation of error probability in fading channels. IEEE Trans. Wirel. Commun. 2, 840–845
(2003)
14. S.J. Grant, J.K. Cavers, Performance enhancement through joint detection of cochannel signals
using diversity arrays. IEEE Trans. Commun. 46, 1038–1049 (1998)
15. X. Zhu, R.D. Murch, Performance analysis of maximum likelihood detection in a MIMO
antenna system. IEEE Trans. Commun. 50, 187–191 (2002)
16. M. Taherzadeh, H. Nikopour, A. Bayesteh, H. Baligh, SCMA codebook design, in Proceedings
of IEEE 80th Conference on Vehicular Technology (VTC Fall’14) (2014), pp. 1–5
17. A. Montanari, D. Tse, Analysis of belief propagation for nonlinear problems: the example
of CDMA (or: How to prove Tanaka’s formula), in Proceeding IEEE Information Theory
Workshop (ITW) (2006), pp. 122–126
18. C.C. Wang, D. Guo, Belief propagation is asymptotically equivalent to MAP estimation for
sparse linear systems, in Proceedings of 44th Annual Allerton Conference on Communication,
Control, and Computing (2006), pp. 926–935
19. L. Li, Z. Ma, L. Wang, P. Fan, L. Hanzo, Cutoff rate of sparse code multiple access in downlink
broadcast channels. IEEE Trans. Commun. 65, 3328–3342 (2017)
20. J. Boutros, E. Viterbo, C. Rastello, J.C. Belfiore, Good lattice constellations for both Rayleigh
fading and Gaussian channels. IEEE Trans. Inf. Theory. 42, 502–518 (1996)
21. J. Boutros, E. Viterbo, Signal space diversity: a power- and bandwidth-efficient diversity technique for the rayleigh fading channel. IEEE Trans. Inf. Theroy 44, 1453–1467 (1998)
22. J. Bao, Z. Ma, Z. Ding, G.K. Karagiannidis, Z. Zhu, On the design of multiuser codebooks for
uplink SCMA systems. IEEE Commun. Lett. 20, 1920–1923 (2016)
23. Y. Xin, Z. Wang, G.B. Giannakis, Space-time diversity systems based on linear constellation
precoding. IEEE Trans. Wireless Commun. 2, 294–309 (2003)
24. G.H. Golub, C.F. Van Loan, Matrix Computations (Johns Hopkins University Press, 1996)
25. S.P. Herath, N.H. Tran, T. Le-Ngoc, Rotated multi-D constellations in rayleigh fading: mutual
information improvement and pragmatic approach for near-capacity performance in high-rate
regions. IEEE Trans. Commun. 60, 3694–3704 (2012)
26. J. Bao, Z. Ma, M.A. Mahamadu, Z. Zhu, D. Chen, Spherical codes for SCMA codebook, in
Proceedings of IEEE 83th Conference on Vehicular Technology (VTC Spring’16) (2016), pp.
1–5
27. L. Yu, X. Lei, P. Fan, D. Chen, An optimized design of SCMA codebook based on star-QAM
signaling constellations, in Proceedings of International Conference on Wireless Communications & Signal Processing (WCSP’15) (2015), pp. 1–5
28. S.T. Brink, Convergence behavior of iteratively decoded parallel concatenated codes. IEEE
Trans. Commun. 49, 1727–1737 (2001)
416
Z. Ma and J. Bao
29. J.V.D. Beek, B.M. Popović, Multiple access with low-density signatures, in Proceedings of
IEEE Conference on Global Communications (GLOBECOM) (2009)
30. A. Ashikhmin, G. Kramer, S.T. Brink, Extrinsic information transfer functions: model and
erasure channel properties. IEEE Trans. Inf. Theory. 50, 2657–2673 (2004)
31. N.H. Tran, H.H. Nguyen, Design and performance of BICM-ID systems with hypercube constellations. IEEE Trans. Wirel. Commun. 5, 1169–1179 (2006)
32. A. Seyedi, Multi-QAM modulation: a low-complexity full rate diversity scheme, in Proceedings
of IEEE International Conference on Communications (ICC) (2006), pp. 1470–1475
33. C.M. Thomas, M.Y. Weidner, S.H. Durrani, Digital amplitude-phase keying with M-ary alphabets. IEEE Trans. Commun. 22, 168–180 (1974)
34. Q. Xie, Z. Yang, J. Song, L. Hanzo, EXIT-chart-matching-aided near-capacity coded modulation design and a BICM-ID design example for both gaussian and rayleigh channels. IEEE
Trans. Veh. Tech. 62, 1216–1227 (2013)
35. F. Schreckenbach, N. Görtz, J. Hagenauer, G. Bauch, Optimization of symbol mappings for
bit-interleaved coded modulation with iterative decoding. IEEE Commun. Lett. 7, 593–595
(2003)
36. M.T. Boroujeni, A. Bayesteh, H. Nikopour, M. Baligh, System and method for generating
codebooks with small projections per complex dimension and utilization thereof, U.S. Patent
0,049,999, 18 Feb 2016
37. A. Bayesteh, H. Nikopour, M. Taherzadeh, H. Baligh, J. Ma, Low complexity techniques for
SCMA detection, in Proceedings of IEEE Globecom Workshops (2015), pp. 1–6
38. R1-164037: LLS results for uplink multiple access. Huawei, HiSilicon, 3GPP TSG RAN WG1
Meeting #85 (2016)
Chapter 13
Interleave Division Multiple Access
(IDMA)
Yang Hu and Li Ping
13.1 Overview
The capacity of a multiple access channel was studied in [1, 2]. It can generally
be achieved by random coding together with other techniques, e.g., power control,
linear precoding, and dirty paper coding [3–6]. Random coding does not involve
orthogonality among users so it is inherently non-orthogonal. The sub-optimality
of orthogonal multiple access (OMA) was investigated in [7, 8]. The gain of nonorthogonal multiple access (NOMA) over OMA was assessed in [9, 10] for both
single-input single-output (SISO) and multiple-input multiple-output (MIMO) systems. Recently, NOMA has been promoted for improving system fairness [11, 12].
However, many practical systems still belong to OMA category. This is mainly due
to complexity concerns. OMA can work with low-cost single-user detection (SUD),
while NOMA may require more complex multi-user detection (MUD). Thus, these
two options entail different trade-off between cost and performance.
Historically, the third-generation (3G) direct-sequence code-division multiple access (DS-CDMA) system is non-orthogonal. Normally, only SUD is used in DSCDMA to avoid high complexity, which is sub-optimal. The spreading operation in
DS-CDMA reduces rate, so DS-CDMA is not convenient for high-rate applications.
The fourth-generation (4G) orthogonal frequency division multiple access (OFDMA) system returns to OMA. OFDMA allows flexibility for resource allocation
over time and frequency, which can bring about noticeable gain.
Y. Hu (B) · L. Ping
Department of Electronic Engineering, City University of Hong Kong,
Hong Kong, SAR, China
e-mail: yhu228-c@my.cityu.edu.hk
L. Ping
e-mail: eeliping@cityu.edu.hk
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_13
417
Y. Hu and L. Ping
average sum rate (bits/channel use)
418
3
2.5
2
SNRsum = 10 dB
1.5
1
0.5
0
SNRsum= 0 dB
0
4
8
12
16
20
24
28
number of users K
Fig. 13.1 Achievable sum-rate of NOMA and OMA under equal energy constraint. Note that the
two strategies have exactly the same performance. SNRsum is the sum signal-to-noise ratio (SNR)
of all users. Complex channels with both slow- and fast-fading factors are considered. Path loss
is based on a hexagon cell with a normalized side length = 1. The minimum normalized distance
between users and the base station is 35/289, corresponding to an unnormalized distance of 35 m
for an LTE cell with radius 289 m. Path loss factor = 3.76 and lognormal fading deviation = 8 dB.
The channel samples are normalized such that the average power gain = 1
Recently, NOMA has been discussed widely for the fifth-generation (5G) [13–
16]. A natural question is whether the possible gain of NOMA can justify its higher
receiver cost. We examine this question using achievable sum-rate below.
In Fig. 13.1, we compare the achievable sum-rate of NOMA and OMA with equal
energy constraint per frame per user in a multi-user SISO system. We can see NOMA
improves sum-rate when the number of users K increases, but OMA can achieve the
same gain through resource allocation. NOMA has no advantage here since OMA is
capacity achieving in this case.
The situation is slightly different if the energy per frame per user can also be
freely optimized under the sum energy constraint. For example, consider maximizing
sum-log-rate under the proportional fairness criterion [3]. Figure 13.2 illustrates the
related numerical results for K = 2. The sum-rate curves in Fig. 13.2 show that
NOMA is only slightly better than OMA with resource allocation (about 8% gain at
SNRsum = 10 dB).
The advantage of NOMA over OMA seen in Figs. 13.1 and 13.2 is disappointing
compared with many results in the literature. This is mainly for two reasons. First,
comparisons with OMA without resource allocation are not fair as resource allocation
has already been widely used in LTE. Second, a practical signal-to-noise ratio (SNR)
range should be used for comparison. A standard way for this purpose is using the
following approximation of the signal-to-noise-plus-interference ratio (SINR) in a
cellular system [3]:
Psum
,
(13.1)
SINRsum =
βPsum + σ 2
Fig. 13.2 Achievable
sum-rate of NOMA and
OMA under sum energy
constraint with proportional
fairness, K = 2. Other
system settings are the same
as those in Fig. 13.1
average sum rate (bits/channel use)
13 Interleave Division Multiple Access (IDMA)
419
3
2.5
NOMA
2
OMA with resource
allocation
1.5
1
OMA without resource
allocation
0.5
0
0
2
4
6
8
10
12
14
16
18
20
SNRsum (dB)
where Psum is the sum received powers of all users in a cell, β a cross-cell interference
factor and σ 2 the noise power. As a rule of thumb, a typical value is β = 0.6. Then, we
have SINRsum ≤ 2.2 dB. Treating interference as noise and allowing a certain range
of β, we may consider 0–10 dB as a typical range for SNRsum , as used in Figs. 13.1
and 13.2. Clearly, the gain of NOMA is marginal over such a practical SNR range.
Furthermore, NOMA performance may deteriorate seriously if a practical forward
error control (FEC) code is used. To see this, consider a successive interference
cancelation (SIC) process with K users in a descendant order of user index k. We
employ an FEC code that can achieve (almost) error-free decoding at SNR = Γ .
Assume that, when we decode for user k, the signals of all users with indexes k ′ > k
have been successfully decoded and subtracted from the received signal. Let qk be
the received power of user k. Then, user k can achieve error-free decoding provided
that
qk
≥ Γ.
(13.2a)
SNRk =
2
′
k <k qk ′ + σ
For sum-power minimization, qk can be calculated using the following recursion
(with q0 = 0):
2
qk = Γ ×
qk ′ + σ .
(13.2b)
k ′ <k
Ideally, the minimum value for Γ can be found from Shannon capacity R =
log2 (1 + Γ ). For a practical code, however, a larger Γ is required to ensure (almost)
error-free decoding. This overhead accumulates during SIC, which can amount to a
considerable loss.
Specifically, we consider a practical rate-1/6 channel coding as an example. It
achieves bit error rate (BER) ≈ 10−5 (approximately error-free) at about SNR =
−3.5 dB with a relatively short block length. There is a 2.35 dB gap compared with
SNR = −5.85 dB calculated from the Shannon formula for Gaussian signaling. Such
420
30
required sum-SNR
(dB)
Fig. 13.3 Accumulated loss
in the SIC process
Y. Hu and L. Ping
2.35dB
25
20
15
8dB
Rate-1/2
7.4dB
10
5
Rate-1/6
0
SIC
Capacity
-5
-10
2
4
6
8
10
12
number of users K
loss accumulates in the SIC process as shown in Fig. 13.3. The accumulated loss is
roughly 7.4 dB for 12 users.
The problem is more serious for higher rate, as seen from Fig. 13.3 for rate-1/2
coding. Assume the same initial single-user gap of 2.35 dB. The accumulated gap
for six users is 8 dB.
Nevertheless, NOMA is still useful. In many situations, it is difficult to establish orthogonality due to the lack of centralized control or accurate channel state information
(CSI). Then, we may have to resort to NOMA. In particular, as we will show below,
NOMA based on interleave-division multiple access (IDMA) [17–22] offers robust
and flexible solutions in such environments. IDMA can also recover a considerable
portion of the accumulated loss suffered by SIC as shown in Fig. 13.3. We will use
numerical results to verify these claims. Some of the software used in this chapter are
available at: http://www.ee.cityu.edu.hk/%7Eliping/Research/Simulationpackage/.
13.2 Basic Principles of IDMA
Following the advent of turbo and low-density parity-check (LDPC) codes [23–
25], iterative detection techniques were developed in late 1990s for equalization in
multipath channels [26] and MUD in DS-CDMA systems [27–29]. It was shown in
[30] that two independently interleaved code sequences can be separated by iterative
detection. This inspired the IDMA scheme in which users are solely separated by
interleavers [17]. Intuitively, a randomly interleaved code results in a different code.
Thus we may also say that different users in IDMA are separated by different codes.
This follows the basic principle of CDMA, except now the set of codes are generated
by a master code followed by different interleavers. We therefore can regard IDMA
as a special case of CDMA. However, IDMA is fundamentally different from DSCDMA. The former does not rely on spreading for user separation and so can avoid
the rate loss suffered by the latter.
In the following, we will start with a graphic model originally presented in [31].
We will show that the primary motivation behind IDMA is to break short cycles,
13 Interleave Division Multiple Access (IDMA)
421
since the latter is detrimental for message-passing detection. The same principles
have been successfully used in turbo and LDPC codes.
13.2.1 IDMA Transmitter Principles
Throughout this chapter, we will assume an underlying OFDMA layer that removes
inter-symbol interference (ISI). We will focus on uplink multiple access techniques
built on this OFDMA layer.
Let K be the user number and ck = {ck (j), j = 1, 2, . . . , J } a length-J codeword
generated by users k. The transmitted symbols {xk (j), j = 1, 2, . . . , J } are generated
from {ck (j)} after certain operations, such as spreading, scrambling, interleaving, and
modulations. At the receiver, the received signals {y(j)} are given by
y(j) =
K
k=1
√
hk ek xk (j) + η(j), j = 1, 2, . . . , J ,
(13.3)
where hk is the channel coefficient for user k, ek the transmitted power of user k
and η(j) an additive white Gaussian noise (AWGN) sample with variance σ 2 per
dimension.
Figure 13.4 illustrates the factor graph representation [32] for (13.3) with K = 2,
J = 8, and the same LDPC coding for all users. A circle in Fig. 13.4 represents a
variable and a square a constraint. Three types of constraints are involved, namely
a square marked with “+” for linear additions in (13.3), a white square for LDPC
coding, and a square marked with “×” for modulation.
Optimal detection for the system in (13.3) typically requires prohibitively high
complexity. Low-cost message-passing detection, similar to that used for LDPC
codes, can be applied instead, as Fig. 13.4 is sparse when K ≪ J . However, short
cylices constitute a problem. To see this, let us call a circle involving m coded
(1)
Channel
y(2) y(3) y(4) y(5) y(6) y(7) y(8)
y(1)
Received signals
+
(2)
+
(3) (4) (5) (6) (7) (8)
+
+
+
+
+
+
h1
h2
Transmitted signals
x1(1) x1(2) x1(3) x1(4) x1(5) x1(6) x1(7) x1(8)
x2(1) x2(2) x2(3) x2(4) x2(5) x2(6) x2(7) x2(8)
c1(1) c1(2) c1(3) c1(4) c1(5) c1(6) c1(7) c1(8)
c2(1) c2(2) c2(3) c2(4) c2(5) c2(6) c2(7) c2(8)
user1
user2
Modulation &
Scrambling
LDPC coding
Fig. 13.4 Factor graph of a NOMA LDPC-coded system where K = 2 and J = 8
422
Y. Hu and L. Ping
(a)
Channel
y(2) y(3) y(4) y(5) y(6) y(7) y(8)
y(1)
Received signals
(1)
+
(2)
+
(3) (4) (5) (6) (7) (8)
+
+
+
+
+
+
h1
h2
Transmitted signals
x1(1) x1(2) x1(3) x1(4) x1(5) x1(6) x1(7) x1(8)
Modulation
x2(1) x2(2) x2(3) x2(4) x2(5) x2(6) x2(7) x2(8)
Interleaving
c1(1) c1(2) c1(3) c1(4) c1(5) c1(6) c1(7) c1(8)
c2(1) c2(2) c2(3) c2(4) c2(5) c2(6) c2(7) c2(8)
user1
user2
LDPC coding
(b)
y(1) y(2) y(3) y(4) y(5) y(6) y(7) y(8)
Received signals
(2)
(1)
+
Channel
+
(3) (4) (5) (6) (7) (8)
+
+
+
+
+
+
h1
h2
Transmitted signals
x1(1) x1(2) x1(3) x1(4) x1(5) x1(6) x1(7) x1(8)
Modulation
x2(1) x2(2) x2(3) x2(4) x2(5) x2(6) x2(7) x2(8)
c1(2) c1(4) c1(5) c1(7) c1(8) c1(1) c1(6) c1(3)
c2(1) c2(5) c2(4) c2(3) c2(6) c2(8) c2(2) c2(7)
Interleaved
LDPC coding
user1
user2
Fig. 13.5 a Factor graph of a two-user IDMA system. b An equivalence form of (a) with re-shuffled
{ck (j)}. Note that in (a) the interleavers for LDPC coding are the same for both users, while in (b)
they are different
bits as a size-m cycle. An example of a size-4 cycle is shown by the black circles
{c1 (1), c1 (3), c2 (1), c2 (3)} in Fig. 13.4. There are a large number of such size-4
cycles in Fig. 13.4. Correlation may build up along these short cycles during messagepassing detection, which is detrimental to performance [33].
Figure 13.4 can also be used to represent a DS-CDMA system. For example, a
spreading operation involving binary sequence of ±1 can be regarded as a repetition
code plus a scrambling operation of sign changes. Repetition coding can be merged
with FEC coding. Scrambling can be incorporated in function of the modulation
nodes in Fig. 13.4. Note that scrambling does not change the topology of the graph.
The problem of short cycles remains the same with or without scrambling. The same
conclusion applies to other modulation techniques.
Inspired by the success of turbo and LDPC codes [23, 25], the IDMA scheme proposed in [31] employs user-specific interleaving to reduce short cycles in a statistical
sense. This is illustrated by the shuffled edge connections between {ck (j)} and {xk (j)}
in Fig. 13.5a. For example, compared with Fig. 13.4, {c1 (1), c1 (3), c2 (1), c2 (3)} no
longer form a size-4 cycle after interleaving in Fig. 13.5a. This is beneficial for
13 Interleave Division Multiple Access (IDMA)
423
Fig. 13.6 An MA node.
Here, an inbound message
LLRDEC (xk (j)) is from an
LDPC decoder. An outbound
message LLRESE (xk (j)) is
generated from the MA node.
ESE stands for “elementary
signal estimation” and DEC
for decoder
y(j)
(j)
+
LLRESE(xk(j))
LLRDEC(xk(j))
x1(j)
...
xk(j)
...
xK(j)
message-passing detection. Similar principles can also be applied to systems involving convolutional or turbo coding.
Two interleavers are used by each user in Fig. 13.5a, one for LDPC encoding
and one for multiple access. The former is the same for all users, and the latter is
user-specific. We can combine these two interleavers by re-shuffling {ck (j)} for each
user, resulting in Fig. 13.5b. The latter interpretation of IDMA is based on [34].
The advantage of Fig. 13.5b is its simpler implementation. If an interleaver is
random, its shifted version can be approximately regarded as another independent
random interleaver. Thus, Fig. 13.5b can actually be realized by using a shifted
version of an LDPC encoder involving a common underlying interleaver. The users,
in this case, are separated by the amount of shift after the common encoder structure.
Such shifting can be realized with very little cost. We may name such a scheme as
code-shift division multiple access (CsDMA).1 This concept was first discussed in
[35].
Note that the use of interleavers in Fig. 13.5 does not incur any rate loss. This is
a noticeable advantage of IDMA for high-rate applications.
13.2.2 Operations on a Multiple Access Node
We now consider message-passing detection based on Fig. 13.5. For convenience,
we refer to a square marked with “+” in Fig. 13.5 as a multiple access (MA) node.
We will only discuss the operations for such an MA node shown in Fig. 13.6. The
operations for other nodes follow the standard treatments for an LDPC code [25].
Denote by DEC k the decoder for user k. We define two messages: an inbound message LLRDEC (xk (j)) and an outbound message LLRESE (xk (j)) that are, respectively
log-likelihood ratios (LLRs), generated by DEC k and elementary signal estimation
1 The
underlying code in CsDMA should be properly interleaved. An LDPC code naturally meets
this requirement. Without interleaving, however, the correlation among the consecutive bits in a
convolutional code may cause a problem in CsDMA. This problem can be easily avoided by shuffling
the coded sequence.
424
Y. Hu and L. Ping
(ESE) operations at the jth MA node in Fig. 13.6. The discussions for LLRDEC (xk (j))
follow the standard LDPC decoding principles [25] and so will be omitted. We will
focus on LLRESE (xk (j)) below since its computation is not part of a standard LDPC
decoder.
For simplicity, let us first assume binary phase-shift keying (BPSK) modulation
xk (j) = ±1 for all k. We define LLRESE (xk (j)) by the following LLR for xk (j):
LLRESE (xk (j)) = log
Pr (y(j)|xk (j) = +1)
,
Pr (y(j)|xk (j) = −1)
(13.4)
where y(j) and xk (j) are defined in (13.3). Assume that η(j) in (13.3) is Gaussian with
mean μ(j) = E(η(j)) and variance v = Var(η(j)). (For simplicity, we will assume
that v is not a function of j. We will explain the rationale for this assumption later.)
For a single-user system with K = 1 in (13.3), the conditional probabilities in (13.4)
are given by
√
2
y(j)−(μ(j)±h1 e1 ))
(
exp −
2v
Pr (y(j)|x1 (j) = ±1) =
,
√
2π v
so
√ y(j) − μ(j)
.
LLRESE (x1 (j)) = 2h1 e1
v
(13.5)
(13.6)
For K > 1, the problem is much more complicated. We need to consider all possible
combinations of {xk (j)}. The exact result is the maximum likelihood (ML) estimator
[36] below:
i
i
(j)
(j) Pr X∼k
Pr y(j)|xk (j) = +1, X∼k
i
LLRESE (xk (j)) = log
,
i
i Pr y(j)|xk (j) = −1, X∼k (j) Pr X∼k (j)
i
(13.7)
i
(j) is one among all 2K−1 possibilities of the set {x1 (j), x2 (j), . . . ,
where X∼k
′
xk−1 (j), xk+1 (j), . .i . , xK (j)} (since xk ′ (j) ∈ {−1, +1}, ∀k ). In i (13.7), Pr
y(j)|xk (j) = ±1, X∼k (j) can be computed similarly to (13.5) and Pr X∼k (j) can be
computed from messages {LLRDEC (xk (j))}. Following [25], we define LLRDEC (xk (j))
by an LLR:
Pr (xk (j) = +1)
.
(13.8)
LLRDEC (xk (j)) = log
Pr (xk (j) = −1)
We can obtain Pr(xk (j) = ±1) by solving (13.8) together with Pr(xk (j) = +1) +
Pr(xk (j) = −1) = 1. Then,
i
Pr X∼k
(j) =
Pr (xk ′ (j)).
k ′ =1,k ′ =k
(13.9)
13 Interleave Division Multiple Access (IDMA)
425
The complexity of ML is O(2K ) for BPSK, which increases exponentially with K.
For a higher order modulation with an M -point constellation, the complexity of ML
is O(M K ). This can be a serious problem in practice.
Gaussian approximation (GA) is a low-cost alternative. We rewrite (13.3) as
√
y(j) = hk ek xk (j) + ζk (j),
(13.10a)
where
√
ζk (j) = y(j) − hk ek xk (j) =
K
k ′ =1,k ′ =k
√
hk ′ ek ′ xk ′ (j) + η(j)
(13.10b)
is the distortion (including interference-plus-noise) with respect to user k. From
the central limit theorem, we apply GA to ζk (j) in (13.10b) and assume ζk (j) ∼
N (μk (j), Var(ζk (j))). Now we can treat (13.10a) as a single-user system. For simplicity, we assume a real channel. (We will discuss a complex channel in Sect. 13.5.2.)
Then, we have
√
2
y(j)−(μk (j)±hk ek ))
(
exp −
2Var(ζk (j))
Pr (y(j)|xk (j) = ±1) =
.
(13.11)
2π Var(ζk (j))
Substituting (13.11) into (13.4) and evaluating μk (j) via (13.10b), we have the
following ESE operations for the jth MA node in Fig. 13.6. (We will discuss the
generation of Var(ζk (j)) in (13.12c) later in (13.13).)
ESE operations
(i) E(xk (j)) = Pr(xk (j) = +1) − Pr(xk (j) = −1),
(ii) μk (j) =
K
k ′ =1
√
√
hk ′ ek ′ E(xk ′ (j)) − hk ek E(xk (j)),
√ y(j) − μk (j)
.
(iii) LLRESE (xk (j)) = 2hk ek
Var(ζk (j))
(13.12a)
(13.12b)
(13.12c)
The following are some details related to the ESE operations.
• Initially, there is no decoder feedback, and we can set E(xk (j)) = 0 in (13.12a) for
∀k, j.
• GA is approximate. However, we observed a very good performance based on GA.
• The summation in (13.12b) can be shared by all users. The cost per information
bit per user is independent of the number of users K.
• The principles of GA for higher-order modulations (such as quadrature phaseshift keying (QPSK)) and complex channels can be derived similarly. Related
discussions will be shown in Sect. 13.5.2.
426
Y. Hu and L. Ping
We now discuss the evaluation of Var(ζk (j)) involved in (13.12c). From (13.10b),
Var(ζk (j)) =
K
k ′ =1
|hk ′ |2 ek ′ Var(xk ′ (j)) − |hk |2 ek Var(xk (j)) + σ 2 .
(13.13a)
We observed that the system performance is not sensitive to Var(ζk (j)). Therefore,
we take the following approximation
Var(ζk (j)) ≈
K
k ′ =1
|hk ′ |2 ek ′ vk ′ − |hk |2 ek vk + σ 2
(13.13b)
based on the following assumption
Var(xk (j)) = 1 − (E(xk (j)))2 ≈ vk , ∀j.
(13.13c)
To evaluate vk in (13.13c), we can simply compute a few samples of Var(xk (j)) and
take their average. In practice, computation for each Var(xk (j)) can be implemented
using a look-up table. We observed that the required number of samples is small, so
the related cost is negligible.
From the above discussions, the total cost for the ESE operations is
(approximately) four additions and two multiplications per chip per iteration. (Some
√
√
operations, such as hk ek and 2hk ek /Var(ζk (j)), can be precalculated and need
not be repeated for every j. The related cost is negligible.)
13.2.3 Overall IDMA Receiver
Now return to an overall IDMA system in Fig. 13.5. We divide the receiver into an
ESE module and K DEC modules. The operations in the ESE module is based on
(13.12). Consider two types of schedules below.
• Serial schedule: In each iteration, operations are carried out as follows:
ESE for user 1, DEC for user 1, ESE for user 2, DEC for user 2, . . .
As an example, ESE for user 1 means running (13.12) for every j with k fixed to
1. The LLRs generated in (13.12) are then fed to DEC for user 1. Then, {μ1 (j)}
and v1 are updated based on the DEC outputs and the process continues to user 2.
• Parallel schedule: In each iteration, operations are carried out as follows:
ESE for all users in parallel, then
DEC for all users.
In the above, in each iteration, (13.12) is run through every pair of j and k. After
the ESE operations, the outputs are fed to the K local DECs. Then, the K DECs
are run in parallel. Afterward, {μk (j)} and vk are updated simultaneously for all k.
13 Interleave Division Multiple Access (IDMA)
Fig. 13.7 Performance of
IDMA with GA in AWGN
channels
427
0
10
-1
Scramling, 4 users,
1 inner itearion,
30 outer itearions
10
-2
single user
BER
10
-3
10
IDMA,4 users,
5 inner itearions,
20 outer itearions
-4
10
IDMA, 4 users,
1 inner itearion,
10 and 20 outer itearions
-5
10
0
0.5
1
1.5
2
2.5
3
Eb/N0 (dB)
Note that in the parallel schedule, the value of the summation in (13.12b), that is
generated using the results of the previous iteration, remains the same for all users
in one iteration. In the serial schedule, this summation is updated user-by-user. We
noticed that serial scheduling converges slightly faster than the parallel one.
A question arises whether it is helpful if multiple inner DEC iterations are carried
out between two consecutive ESE iterations. For example, with the serial schedule,
we can run multiple LDPC decoding iterations for one user before going to the
next user. We observed that such inner iterations are generally unnecessary. For
fixed overall cost, better performance is achieved without inner iterations. However,
slipping some inner or outer iterations may lead to reduced complexity.
Intuitively, we can treat an IDMA system in Fig. 13.5 as a generalized code system
on a graph (where an MA node is just for a special type of constraint). The inneriteration method means more iterations on some parts of the graph for LDPC coding
constraints. Such uneven message-passing process does not help in general.
For comparison of overall cost, let us consider a K-user TDMA system in which
LDPC decoding is individually carried out for each user. The only difference between
SUD for such a TDMA system and MUD for IDMA is the ESE operations. As we
have seen above, the cost of the extra ESE operations is quite moderate. Thus, an
IDMA receiver involving (13.12) has only moderately higher complexity than SUD
for corresponding TDMA systems with the same number of users.
Figure 13.7 shows an example of a four-user IDMA system in AWGN channels.
A rate-1/2 LTE turbo code with 1200 information bits per user is used, followed by
a rate-1/8 repetition code and QPSK modulation. We can see that iterative detection
nearly converges with about ten outer iterations between ESE and DECs with one
inner-iteration. Here, one inner-iteration means running each component decoder
once in a turbo decoder. We can also run multiple iterations in each turbo decoder
within each outer iteration. The result of five inner iterations is shown in Fig. 13.7.
We can see that multiple inner iterations can only offer marginal improvement on
performance, even though at considerably higher overall cost.
428
Y. Hu and L. Ping
Figure 13.7 also shows the result with user-specific scrambling by spreading each
coded bit with a random binary sequence of +1 and −1 before modulation. No
user-specific interleaving is used in this case. It is seen that interleaving offers better
performance than scrambling. This is due to the short cycle problem as explained
earlier.
13.2.4 Performance Evaluation Through SNR Evolution
We now outline an SNR evolution technique [17] for tracking IDMA performance.
Similar techniques have been successfully applied to turbo and LDPC codes [25]
and more recently to AMP algorithms [37]. This analysis method also provides the
basis of power allocation for IDMA performance optimization in the next section.
Figure 13.8a is a so-called protograph [38] representation of Fig. 13.5, in which
each circle represents a vector and each edge represents a vector connection. Relatively thick lines are used in Fig. 13.8a to distinguish it from Fig. 13.5. The messages
LLRDEC,k and LLRESE,k in Fig. 13.8a are LLR sequences generated by the coding
and MA constraints, respectively.
We will use the following SNR-variance relationship to characterize the behavior
√
of the system in Fig. 13.8a. Recall (13.10a): yk (j) = hk ek xk (j) + ζk (j). We define
the average SNR for user k as
√
E |hk ek xk (j)|2
|hk |2 ek
=
.
SNRk ≡
E(Var(ζk (j)))
E(Var(ζk (j)))
(13.14)
From (13.13), we have
E(Var(ζk (j))) =
k ′ =1,k ′ =k
(a)
|hk ′ |2 ek ′ vk ′ + σ 2 .
(13.15)
(b)
Received signals
Received signals
y
MA constraints
LLRESE,1
y
ESE
MA constraints
+
LLRDEC,1
Transmitted signals
K
LLRESE,2
LLRDEC,2
x1
x2
c1
c2
Modulation
ESE
+
SNR1 = 1(v2)
v1 = (SNR1)
Transmitted signals
x1
Modulation
SNR2 = 2(v1)
v2 = (SNR2)
x2
DEC
DEC
Coded bits
LDPC constraints
Coded bits
c1
c2
LDPC constraints
Fig. 13.8 a Protograph representation of Fig. 13.5. b Evolution characterization for (a)
13 Interleave Division Multiple Access (IDMA)
429
Since LLRDEC,k is the output of DEC k with input LLRESE,k characterized by SNRk ,
we can write vk as a function vk = ψ(SNRk ). Here for simplicity, we assume that
all users employ the same LDPC code and therefore have the same ψ(·). (Note that
the re-shuffling operation in Fig. 13.5b has no effect on ψ(·).) Combining this with
(13.14) and (13.15), we have the following recursion to characterize an iterative
IDMA detector:
SNR(t)
k = K
|hk |2 ek
(t)
2
2
k ′ =1,k ′ =k |hk ′ | ek ′ vk ′ + σ
vk(t)
=ψ
SNRk(t−1)
(t)
(t)
≡ φk v1(t) , . . . vk−1
, vk+1
, . . . vK(t) ,
,
(13.16a)
(13.16b)
where t is an iteration index. The initialization is vk(0) = 1, ∀k, implying no information from DECs. In general, there is no closed form expression for ψ(·), but it
can be obtained by simulating a single-user APP decoder in an AWGN channel with
specified SNRs. Using {SNRk(T ) } in the final iteration in (13.16), we can estimate the
BER by a function
(13.17)
BERk = g SNRk(T ) ,
where g(·) can be obtained through simulation of DECs [17].
13.2.5 Superposition Coded Modulation (SCM)
The above IDMA scheme involves multiple signal streams from different users. We
may simply allocate these signal streams to a single-user. Such scheme is referred to
as superposition coded modulation (SCM) [39, 40].
We define a standard QPSK constellation as SQPSK = {00 → (+1, +1), 01 →
(+1, −1), 10 → (−1, +1), 11 → (−1, −1)}. Figure 13.9a is a 16-ary scheme
formed by superimposing SQPSK and a scaled version of SQPSK with a scaling factor of
2 and a 45◦ phase shift [39]. Figure 13.9b is a 64-ary scheme formed by superimposing SQPSK and two scaled versions of SQPSK with, respectively, scaling factors of 1.18
and 1.10 plus 60◦ and 120◦ phase shifts. From the central limit theorem, the SCM
signaling is more Gaussian-like when the number of streams is large. This can offer
the so-called shaping gain as analyzed in [41, 42]. It has been proved that, among
all possible signaling methods, an SCM constellation achieves the minimum mean
squared error (MMSE) bound; that is, it minimizes the function ψ(·) in (13.16b) for
a fixed underlying binary decoder. The details are discussed in [39].
430
Y. Hu and L. Ping
1000 0000
00
(a)
10
1100 0100
1010 0010
1001 0001
00
+
11
10
01
1101 0101
1110 0110
1011 0011
01
11
1111 0111
(b)
+
+
Fig. 13.9 a A 16-ary SCM signaling by superimposing two streams of QPSK constellations. b A
64-ary SCM signaling by superimposing three streams of QPSK constellations
13.3 Power Control for IDMA
At a relatively low sum-rate, such as less than 1 in a complex channel, IDMA with
a GA receiver works well with equal received power. At a higher sum-rate, unequal
power control is required. The situation is similar to power control for SIC in (13.2),
except that iterative detection makes the problem more complicated.
13.3.1 Transmitted and Received Power Minimization
Denoting the received powers by
qk = |hk |2 ek .
(13.18)
We combine (13.16a) and (13.16b) into a compact form as
SNR(t)
k = K
k ′ =1,k ′ =k
qk
+ σ2
qk ′ ψ SNR(t−1)
k′
, ∀k.
(13.19)
Let ek and qk be, respectively, sum transmitted power and sum received power.
We can minimize them, respectively. The latter is simpler since the channel gains
{|hk |2 } are not involved. The following Remark establishes a connection between
these two problems [7, 9, 43].
13 Interleave Division Multiple Access (IDMA)
431
∗
Remark 1 Assume
for
qk . Then, {ek∗ = qk∗ /|hk |2 } is a
that {qk } is a minimizer
∗
2
minimizer for ek provided that {qk } and {|hk | } have the same order, i.e., qk∗ ≤ qk∗′
if |hk |2 ≤ |hk ′ |2 .
Based on Remark 1, we can first find the minimizer {qk∗ } for
qk . We then
re∗
2
ek
label {qk } such that it has the same order as {|hk | }. Then, the minimizer for
can be obtained as {ek∗ = qk∗ /|hk |2 }.
Incidentally, Remark 1 implies that a user with a higher channel gain should be
assigned a higher transmitted power and vice versa. Next, we focus on minimizing
{qk }.
13.3.2 Feasible Profile
We now impose an SNR requirement Γ after T iterations. This SNR requirement
can be equivalently translated into a BER requirement through (13.17). We write the
received power optimization problem as follows.
minimize
(13.20a)
qk ,
subject to SNR(t)
k = K
k ′ =1,k ′ =k
SNRk(T ) ≥ Γ, ∀k.
qk
+ σ2
qk ′ ψ SNR(t−1)
k′
, ∀k,
(13.20b)
(13.20c)
The problem in (13.20) is non-convex. We will outline two searching techniques for
this problem. For convenience, we will call {qk } a feasible profile if it ensures the
constraints in (13.20b) and (13.20c).
Incidentally, it is interesting to compare (13.2a) and (13.20b). In (13.20b),
represents the residual interference from user k ′ after soft canqk ′ ψ SNR(t−1)
k′
celation. Such terms disappear in (13.2a) for decoded users due to the error-free
assumption and hard cancelation.
13.3.3 Greedy Search
We first set T = 1, i.e., only one iteration. Assume that approximate error-free decoding can be achieved at a sufficiently large SNR in the single-user case. We can
construct an initial feasible profile Q = {qk } according to (13.2). The sum-power for
such a Q is typically large.
We next consider a general T . Starting from the above initial Q, we search for
a minimum value for each qk individually to achieve (13.20c), while keeping other
elements in Q unchanged. This involves a one-dimensional search, so its complexity
432
Y. Hu and L. Ping
is affordable. Let the search result be qk∗ . We then update qk ← qk − ǫ(qk − qk∗ ) in
Q, where ǫ is a damping factor (e.g., ǫ = 0.5). We repeat the above process for all
k iteratively. We observed reasonably good performance of this simple method for a
relatively small K.
13.3.4 Approximate Linear Programming Method
Inspired by the linear program technique for LDPC code design [25], we can use the
approximate technique below for a large K. The key idea is to transform the problem
of finding the power for each user into finding the number of users on different given
power levels, which makes the problem convex [44, 45].
Let us quantize the received power into M + 1 discrete values: {q(m), m =
0, 1, . . . , M } with q(m − 1) < q(m). The received powers of all users are selected
from {q(m)}. We partition K users into M + 1 groups according to their power levels.
Let λ(m) be the number of users assigned with power level q(m) and z(m) be the
total power of these λ(m) users. As such,
λ(m) = K,
λ(m)q(m) =
m
(13.21a)
z(m) = λ(m)q(m)
(13.21b)
and the sum received power
k
qk =
m
z(m).
(13.21c)
m
Denote by SNR(m) the SNR for the users in the mth group with power q(m). Define
I=
m
z(m)ψ(SNR(m)) + σ 2 ,
(13.22)
which is the total interference power (including noise) after soft cancelation. When
K is large, (13.20b) can be approximated as
SNR(m)(t) =
I (t)
q(m)
q(m)
≈ (t) ,
(t−1)
I
− q(m)ψ SNR(m)
(13.23)
where I (t) denotes the value of I at the tth iteration. Using (13.22) and (13.23), we
have the update rule
q(m)
I (t) =
(13.24)
z(m)ψ (t−1) + σ 2 .
I
m
13 Interleave Division Multiple Access (IDMA)
Fig. 13.10 IDMA with EPC
and UPC in AWGN
channels. Rate-1/3 turbo
coding followed by rate-1/2
repetition coding is used for
each user. Information length
of each user is 1200. QPSK
modulation. Sum-rate = 1, 2,
and 4 for K = 3, 6, and 12,
respectively
433
0
10
K=12
EPC
UPC
-1
K=6
10
-4
10
-5
10
-2
0
Capacity for K = 6
10
Capacity for K = 12
K=3
-3
Capacity for K = 3
BER
-2
10
2
4
6
8
10
K=12
12
14
16
18
SNR sum (dB)
Equation (13.24) characterizes the evolution of the total interference at each iteration.
If iterative detection converges, I (t) should be lower than I (t−1) . Equivalently, we can
write the convergence condition as
z(m)ψ
m
q(m)
I
+ σ 2 ≤ (1 − δ)I , Imin ≤ I ≤ Imax .
(13.25)
where 0 < δ < 1 is a decay factor that controls the convergence speed. Imax and Imin
specify the total interference at the beginning and end of the iterative detection. In
summary, we re-formulate the optimization problem as,
minimize
subject to
m
(13.26a)
z(m),
z(m)ψ
z(m) ≥ 0, ∀m.
q(m)
I
+ σ 2 ≤ (1 − δ)I , Imin ≤ I ≤ Imax ,
(13.26b)
(13.26c)
The above optimization problem is linear with respect to {z(m)}. Hence, it can be
resolved by linear programming. More details can be found in [44, 45].
Figure 13.10 is an example to illustrate the necessity of unequal power control
(UPC). We can see that UPC improves the system performance for all the three cases
of K = 3, 6, and 12 compared with equal power control (EPC). Particularly, when
K = 12, the IDMA system does not work at all with EPC, but works well with UPC.
The gap between IDMA with UPC and capacity is about 3.7 dB for K = 12.
Compared with the 12-user example in Fig. 13.3, we can see that IDMA recovers
about 3.7 dB relative to the loss incurred by SIC. This is impressive, but there is
definitely room for improvement. The gap towards capacity can potentially be further
narrowed using more sophisticated techniques, such as curve-matching-based degree
434
Y. Hu and L. Ping
sequence optimization [46, 47], spatial coupling [48–57], or more sophisticated
modulations [58].
A criterion called overloading, which represents the user capacity in a NOMA
system, is used to assess the system performance in the recent literature. Figure 13.10
demonstrates that IDMA can offer very high overloading with centralized power control. In the next section, we will discuss IDMA techniques to achieve high throughput
without centralized power control.
13.4 Random Access via IDMA
13.4.1 Limitations of Conventional Systems
A conventional uplink system with centralized control involves a connection setup
procedure before data transmission. The overhead incurred by this procedure is not
serious for services with long-lasting connections since it can be amortized across the
connection duration. One of the main tasks envisaged for the next 5G cellular systems
is to support machine-type communication (MTC) which is characterized by short
and sporadic data communication. In this case, the cost of establishing centralized
control can be substantial.
Random access is decentralized, in which each user makes an individual decision
to transmit data packets. This avoids the overhead of connection setup, but the packets
from different users may collide. In a conventional random-access scheme, such as
ALOHA [59], colliding packets are discarded, which reduces throughput. For this
reason, such techniques cannot satisfy the demand of high spectral efficiency in 5G
cellular systems.
In a fading channel, the received powers of different users may form a feasible
profile defined in Sect. 13.3.2, even without centralized control. This is captured in
the multi-packet reception (MPR) model [60, 61]. Most existing MPR techniques
rely on channel fading to form feasible profiles. Such a passive approach achieves
limited throughput gain. In what follows, we will discuss an active approach based on
the power control technique. The main idea is to optimize the probability of forming
a feasible profile through decentralized power control at the transmitters.
13.4.2 Random IDMA with Decentralized Power Control
13.4.2.1
Problem Formulation
Recall from Sect. 13.3.2, a feasible received power profile {qk } can be formed by
centralized control. Without centralized control, however, it is difficult to guarantee
13 Interleave Division Multiple Access (IDMA)
435
this. A randomized power control (RPC) technique [62–67] is discussed below to
handle this difficulty.
The principles of RPC are as follows. Let {Q(l) , l = 1, 2, . . . , L} (L is the maximum level index) be a set of pre-defined power levels and {P (l) , l = 1, 2, . . . , L} a
set of related probabilities. Upon a packet arrival, each user randomly draws a power
level Q(l) with probability P (l) and uses it to transmit. Different users act individually
and so their transmissions may collide. However, as long as their received powers
{qk } form a feasible profile, their signals can still be recovered. Our aim is to optimize
the probability that {qk } form a feasible profile.
Mathematically, {qk } defined above are the realizations of an underlying random
variable. The random variable is characterized by a probability mass distribution
{P (l) }. Each qk is independently drawn by a user from the support {Q(l) }. Once the
distribution is given, there is no need for centralized control.
13.4.2.2
Type-2 Collisions
The problem formulated above turns out to be difficult when K is large. So far, we
have no general solution. We will discuss a sub-optimal technique below.
We say that a collision is of type-M if it involves M active users. Let Q(0) = 0 be
an element in {Q(l) }. In RPC, a user will not transmit if its selected power is Q(0) .
The related P (0) is equivalent to the back-off probability in 802.11 Wi-fi systems.
Intuitively, collisions are dominated by type-2 ones when P (0) is sufficiently large.
Therefore, we will focus on type-2 collisions.
We consider a type-2 collision involving user i and user j. We define the union of all
possible feasible profiles of the received power pair {qi , qj } as a feasible region. The
collision is resolvable if {qi , qj } falls in this region. Figure 13.11a shows an example
of the feasible region for SIC with ideal coding and decoding. The area marked by “A”
in Fig. 13.11a is formed by all possible {qi , qj } that meets the following conditions
(see (13.2)):
qj
≥ Γ,
qi + σ 2
qi
SNRi = 2 ≥ Γ.
σ
SNRj =
(13.27a)
(13.27b)
The area marked by “B” in Fig. 13.11a is formed similarly by changing decoding
order. The value of Γ here is determined by the Shannon capacity R = log2 (1 + Γ ).
Any power pair in the feasible region is resolvable by SIC.2
Figure 13.11b is an example of an IDMA system involving two LDPC-coded
users with coding rate 0.5 per user. The receiver can achieve BER ≤ 10−5 in the
feasible region, which is regarded as approximately error-free. The border of this
feasible region is obtained using simulation.
2 Figure 13.11a is for R
< 1. If R ≥ 1, the feasible region is divided into two disjoint sub-regions A
and B symmetric to the 45◦ line qi = qj [63].
436
Y. Hu and L. Ping
(a)
(b)
qj
Q(2)
(Q(3), Q(4))
(Q(2), Q(3))
qj
...
A B
...
...
Q
(3)
B
(Q(4), Q(3))
(Q(3), Q(2))
(Q(1), Q(2))
Q
(4)
Q(2)
qj = ϕ (qi)
(Q(3), Q(4))
qi = Q(1)
Q(4)
Q(3)
qj = ϕ (qi)
qi = Q(1)
...
A
0
(Q(2), Q(1))
(Q(4), Q(3))
(Q(1), Q(2))
Q (1)
Q (2) Q (3)Q (4)
Q(1)
qj = Q(1)
...
qi
0
...
(Q(2), Q(3))
(Q(3), Q(2))
qi = ϕ (qj)
qi = ϕ (qj)
Q(1)
...
qj = Q(1)
(Q(2), Q(1))
Q (1)
Q (2)
Q (3)
Q (4)
... qi
Fig. 13.11 Feasible regions for a a two-user ideally coded system with SIC, and b a two-user
LDPC-coded IDMA system
Each of the two feasible regions in Fig. 13.11 is bounded by four curves qi =
Q(1) , qj = φ(qi ), qj = Q(1) and qi = φ(qj ). Here, the function φ(·) is determined by
(13.27a) taking mark of equality (for Fig. 13.11a) or by simulation (for Fig. 13.11b).
Q(1) is the minimum power for successful single-user transmission. We construct the
set {Q(l) } as follows:
⎧
0,
l = 0,
⎪
⎨
(1)
(l)
l = 1,
(13.28)
Q = Q ,
⎪
⎩
(l−1)
φ(Q
), l > 1.
For {qk } randomly selected from {Q(l) }, we have the following situations:
Case 1: All {qk } are zeros. In this case, throughput is zero.
Case 2: Only one element in {qk } is nonzero. The transmission of the only active
user will be successful.
Case 3: Exactly two elements in {qk }, say qi and qj , are nonzero. This is a type-2
collision. It can be shown that {qi , qj } falls in the feasible region (so collision is
resolvable) provided that qi = qj .
Case 4: More than two elements in {qk } are nonzero. For simplicity, such events
are regarded as unresolvable, which is a pessimistic assumption.
Based on the above cases, we can find an optimized probability set {P (l) } for a K
user system. For convenience, we assume that the packets of all users arrive independently, following the Bernoulli process with parameter λ. The system throughput
is then given by
(13.29)
T = T1 + T2 .
In (13.29), T1 is the throughput related to transmissions without collision:
13 Interleave Division Multiple Access (IDMA)
T1 =
K
k=1
437
CKk λk (1 − λ)(K−k) Ck1 (1 − P (0) )(P (0) )k−1
(13.30)
=Kλ(1 − P (0) )(1 − λ + λP (0) )K−1 ,
where CKk λk (1 − λ)(K−k) is the probability of k users among total K users having
packets to transmit and Ck1 (1 − P (0) )(P (0) )k−1 the probability of only one user among
these k users transmitting with nonzero power. Also in (13.29), T2 is the throughout
related to type-2 collisions. From case 3 above, a type-2 collision is unresolvable
only if the two transmitting users are using the same received powers. Therefore, we
have
K
(0) 2
(l) 2
k k
(K−k) 2
(P ) (P (0) )k−2
CK λ (1 − λ)
Ck (1 − P ) −
T2 =2
l>0
k=2
2
(0) K−2
=K(K − 1)λ (1 − λ + λP )
where (1 − P (0) )2 −
(0) 2
(1 − P ) −
l>0
(l) 2
(P )
(13.31)
,
(P (l) )2 is the probability that two active users transmit with
l>0
different received powers.
We further consider an average transmitted power constraint q̄ for each user. In
AWGN channels with unit channel power gain, the constraint is given by
l≥0
Q(l) P (l) ≤ q̄.
(13.32)
Under such power constraint, we can search for {P (l) } that maximize the throughput T
in (13.29). It can be verified that the problem is convex if P (0) is fixed. The treatments
for fading channels are somewhat more complicated. The details can be found in [63].
Figure 13.12 shows two examples, one for an ideally coded system and the other
for an LDPC-coded IDMA system [63]. Conventional ALOHA is included as a
reference. We can see that the RPC-based scheme can offer noticeably throughput
gain compared with ALOHA.
It is proved in [63] that the {Q(l) } in (13.28) forms an optimal support for the
decentralized power control when K = 2. It is sub-optimal for K > 2, but it can still
provide excellent performance gain, as seen in Fig. 13.12.
As a short summary, we can treat collisions as NOMA cases. A conventional
scheme, such as ALOHA, treats such NOMA cases as failures. The discussions in
this section aim at optimizing the probability of successful detection in such NOMA
cases.
438
Y. Hu and L. Ping
(b) 1.5
1
RPC
0.5
ALOHA
0
2
4
6
8
system throughput
system throughput
(a) 1.5
RPC
1
0.5
ALOHA
02
10 12 14 16 18 20
number of active users K
4
6
8
10 12 14 16 18 20
number of active users K
Fig. 13.12 Performance comparison of RPC and ALOHA in a an ideally coded system and b an
LDPC-coded IDMA system. Rayleigh fading channel with averaged power gain 1 for all users.
Same power constraint for RPC and ALOHA
13.5 IDMA in MIMO Systems
13.5.1 Multi-User Gain in MIMO Systems
Fig. 13.13 Achievable
sum-rate of ZF under perfect
CSI. The number of antennas
at the base station is 64.
Single antenna assumed is
for each user. Other system
parameters are the same as
those in Fig. 13.1
average sum rate (bits/channel use)
MIMO is a wireless technology employing multiple transmit and receive antennas
[68–74]. Multi-user gain refers to the advantage of allowing a large number of users
to transmit simultaneously over the same time and same frequency in MIMO [10, 75].
This is illustrated in Fig. 13.13 by the potential sum-rate capacity gain for a singlecell system. Perfect CSI is assumed in Fig. 13.13. The curves apply to both up- and
downlinks following the duality principle [3, 76, 77]. We can see that multi-user
gain is very attractive. The potential gain is in the order of tens of times. Diversifying
power over more users, i.e., increasing K, is a very effective way to increase sum-rate.
60
50
Capacity
40
OMA via ZF
30
20
10
0
0
8
16
24
32
40
number of users K
48
56
64
13 Interleave Division Multiple Access (IDMA)
439
Figure 13.13 also includes the performance of zero-forcing (ZF) with proper
power allocation. ZF is an OMA technique [75]. Different users are divided into
different orthogonal subspaces in ZF, which avoids interference among users. It is
seen from Fig. 13.13 that, with accurate CSI, ZF can offer very good multi-user
gain. Since the gap between ZF and capacity is small, any further gain by NOMA is
limited. In this case, OMA via ZF can be preferred for its low-cost SUD receiver.
However, in practice, we usually do not have reliable CSI to establish spatial orthogonality initially. ZF performance deteriorates seriously when CSI is not accurate.
In the following, we will see that NOMA via IDMA offers a solution to the problem.
13.5.2 Iterative Maximum Ratio Combining (I-MRC)
We first extend the GA-based detection technique in Sect. 13.2.2 to MIMO. Consider
a multi-user uplink system model with NBS antennas at the base station. For simplicity,
we assume a single antenna at each user. The IDMA principle discussed in Sect.
13.2.1 can be directly used here. We assume perfect CSI first and will return to the
CSI estimation problem later.
The received signal at time j is written as
y(j) =
K
k=1
√
hk ek xk (j) + η(j),
(13.33)
where y(j) is an NBS × 1 signal vector received at base station antennas, hk an NBS × 1
complex channel coefficient vector, ek the transmitted power of user k, xk (j) a symbol
transmitted from user k, and η(j) an NBS × 1 vector of complex AWGN with mean
0 and variance σ 2 = N0 /2 per dimension.
Maximum ratio combining (MRC) is a common strategy for MIMO systems. An
MRC estimator is defined in a symbol-by-symbol form as
x̂k (j) = hH
k y(j).
(13.34)
x̂k (j) = λk xk (j) + ξk (j),
(13.35a)
Substituting (13.33) into (13.34),
√
√
where λk ≡ hk 2 ek = hH
k hk ek is a scalar and
ξk (j) ≡
K
k ′ =1,k ′ =k
√
H
hH
k hk ′ ek ′ xk ′ (j) + hk η(j)
(13.35b)
440
Y. Hu and L. Ping
is an interference (from {xk ′ (j), k ′ = k} to xk (j)) plus noise term. MRC does not
involve matrix inversion and so has low cost. However, interference is a problem for
MRC, especially when K is large. Iterative GA technique can alleviate this problem.
Similar as that in Sect. 13.2.2, we approximate ξk (j) in (13.35b) by a Gaussian
random variable. We assume that the real and imaginary parts of xk (j) carry two bits
of information in the QPSK modulation. Similar to (13.12), the real part of xk (j) can
be estimated as
LLRESE (Re(xk (j))) =
2λk
Re x̂k −E ξk (j) .
Var Re ξk (j)
(13.36)
The mean and variance in (13.36) are updated as
E ξk (j) = hk H
Var Re ξk (j)
=
K
k ′ =1
K
k ′ =1
hk ′
√
√
ek ′ E(xk ′ (j)) − hk ek E(xk (j)) ,
√
Var Re hH
k hk ′ ek ′ xk ′ (j)
(13.37a)
√
− Var Re hk 2 ek xk (j) + hk 2 σ 2 .
(13.37b)
Some detailed computation techniques for (13.37) can be found in [75]. The imaginary part of xk (j) can be estimated similarly. We call the above process iterative
MRC (I-MRC).
Figure 13.14 illustrates the effectiveness of I-MRC [75]. We consider three different settings:
(i) K = 1 and sum-rate Rsum = 5 with five signal streams (each stream with rate 1)
assigned to the sole user using the SCM principle discussed in Sect. 13.2.5,
(ii) K = 8 and Rsum = 16 with two streams per user, and
(iii) K = 8 and Rsum = 24 with three streams per user.
For K = 1, all the signal streams see the same channel so there is no spatial diversity among them, which results in poor performance. Increasing K from 1 to 8 results
in drastically enhanced rate or reduced power or both in Fig. 13.14. Figure 13.14 is
a compelling evidence for multi-user gain: allowing more concurrent transmitting
users is more efficient than increasing single-user rate.
13.5.3 Data-Aided Channel Estimation (DACE)
We now consider the CSI acquisition problem. Many factors may affect CSI accuracy
in MIMO. In particular, the correlation among the pilots used by different users can
lead to the pilot contamination problem [78].
13 Interleave Division Multiple Access (IDMA)
441
0
10
-1
10
BER
-2
10
-3
10
8 users,
Rsum = 16
Single user,
Rsum = 5
8 users,
Rsum =
24
-4
10
-5
10
-4
-3
-2
-1
0
1
2
3
4
SNRsum (dB)
Fig. 13.14 Multi-user gain for K = 8 with I-MRC. Rayleigh fading. NBS = 64. Equal transmitted
power is assumed for different users. Power control is used for the streams assigned to the same
user. The power allocation levels are obtained through heuristic search. Rate-1/2 turbo coding and
information length = 1200 for each stream. QPSK modulation. A codeword is transmitted over ten
resource blocks. Each resource block contains 180 symbols experiencing the same fading conditions
IDMA with data-aided channel estimation (DACE) [17, 79–83] technique can be
used to improve CSI accuracy. The basic principle of DACE is as follows. Recall that
a key difference between pilot and data is that the former is known at the receiver,
while the latter is not. Therefore, if a data symbol is known, it can be used as a pilot.
Furthermore, partial information of a data symbol, such as its mathematical mean,
can also be used to refine the channel estimates. Such partial information is readily
available in an IDMA receiver (as given in (13.8)).
DACE can be used jointly with I-MRC, which involves iterations of the following
two operations [75]:
(a) using both pilots and partially decoded data information to refine CSI, and
(b) using improved channel estimates to refine data estimation by I-MRC.
The advantages of DACE are twofolds: (i) With DACE, the estimated data is gradually
used for channel estimation. Pilot energy can be greatly reduced since only very
coarse CSI is required initially. (ii) Data sequences are typically much longer than
pilots and correlation is low among them. Therefore, DACE is robust against the
pilot contamination problem. Such problem is typically caused by the correlation
among the pilot sequences re-used in neighboring cells. Without DACE, longer pilot
sequences will be required to reduce such correlation. Thus, DACE also reduces the
time overhead related to pilots.
I-MRC and DACE can be naturally combined in an overall iterative process.
After MRC and decoding operations in each iteration, partially detected data are
used to refine channel estimates that are in turn used for MRC and decoding in the
next iteration. This is referred to as I-MRC-DACE. Figure 13.15 compares BER
442
Y. Hu and L. Ping
(a) 100
(b) 10 0
-1
-1
ZF
10
10
-2
BER
BER
-2
10
β = 0.9
-3
10
β = 0.6
β=0
-4
10
-4
I-MRC-DACE
-3
10
β = 0 β = 0.6
β = 0.8
β = 0.8
-4
β = 0.9
10
-5
10
10
-5
-2
0
2
4
SNRsum (dB)
6
8
10
10
-4
-2
0
2
4
6
8
10
SNRsum (dB)
Fig. 13.15 Performance comparison of a ZF and b I-MRC-DACE with different β values. Rayleigh
fading. NBS = 64 and K = 16. Rate-1/3 turbo coding and information length = 1312 for each user.
QPSK modulation. Each codeword is divided into 12 sections, and each section is transmitted
over a resource block (including 16 pilot symbols). Different users in a cell are assigned different
orthogonal pilots. These pilots are repeated for users in different cells. The pilot and data symbols
have the same average power
performance for ZF and I-MRC-DACE, in which β is the cross-cell interference
factor defined in (13.1). A larger β indicates a more serious pilot contamination
problem due to more severe interference among the pilots. From Fig. 13.15, we can
see that I-MRC-DACE noticeably outperforms ZF. The difference becomes very
significant when β is large (e.g., β ≥ 0.6).
IDMA is a natural choice for I-MRC-DACE since it is beneficial for iterative
detection. Note that Fig. 13.5 can also be used to characterize IDMA in MIMO
systems, if each scalar y(j) in Fig. 13.5 is replaced by its vector counterpart y(j) in
(13.33). The discussions on short cycles in Sect. 13.2.1 are still applicable after such
replacement.
IDMA also allows a superimposed pilot scheme that can reduce the power overhead and rate loss. The related discussions can be found in [83–86].
13.6 Prospective Applications of IDMA in 5G Systems
Various approaches have been proposed recently for 5G radio link under LTE, including IDMA [87], RSMA [88], IGMA [89], PDMA [90] and SCMA [91, 92]. In
the following, we will show that these schemes all share, explicitly or implicitly, the
basic principle of IDMA.
We first represent these different schemes using a unified protograph framework.
Assume that N resource blocks (RBs) defined in LTE are available for transmission.
We label the observations from these RBs by {y(1) , y(2) , . . . , y(N ) }.
Figure 13.16 shows a scheme in which each user transmits on all available RBs
as illustrated for two system settings: (a) three users over two RBs, and (b) six users
13 Interleave Division Multiple Access (IDMA)
443
(b)
(a)
|
RB1
|
RB2
y(1)
|
|
RB2
|
RB3
|
RB4
y(3)
(2)
+
+
+
|
y(2)
(1)
(2)
(1)
RB1
y(1)
y(2)
|
y(4)
(3)
(4)
+
+
+
x1
x2
x3
x1
x2
x3
x4
x5
x6
c1
c2
c3
c1
c2
c3
c4
c5
c6
Fig. 13.16 Protograph representations of DS-CDMA and IDMA with a three users over two RBs,
and b six users over four RBs
over four RBs. This can be realized by transmitting replicas of each xk over multiple
RBs. Alternatively, we may use a low-rate code to generate each xk . Each xk can be
segmented into several blocks, with each block transmitted over an RB. The latter
approach can potentially provide higher coding gain [93]. We may also use different
modulations for the bits on different RBs, as for SCMA [92].
Incidentally, both DS-CDMA and IDMA can be represented using the protographs
in Fig. 13.16. They are distinguished by the absence or presence of user-specific
interleaving within each RB. RSMA [88] is a DS-CDMA scheme. However, userspecific interleaving is stated as an option for RSMA. If this option is used, it is
equivalent to IDMA. The advantage of this option can be seen in Fig. 13.7.
Alternatively, each user can transmit over only some of the available RBs. This
is referred to as sparse coding in [91]. Figure 13.17 shows an example for sparse
coding. IGMA, PDMA, and SCMA all involve such treatment. Note that symbol-level
interleaving as in Fig. 13.5a is not explicitly seen in Figs. 13.16 and 13.17. If such
underlying interleaving is not used, size-4 cycles can be a problem in Fig. 13.16.
Sparse coding in Fig. 13.17 avoids this problem. Clearly, sparse coding leads to
user-specific edge connections between users and RBs. It has the same effect as
symbol-level interleaving; they both reduce short cycles.
With sparse coding, each user does not fully occupy all RBs. This may cause
problem for decentralized grant-free [13] or random-access applications, where each
user determines its activity individually. In these cases, the number of active users,
denoted by Kactive , is a random variable. When Kactive is small, sparse coding may
lead to inefficient use of the available RBs and so low power efficiency. This implies
poor scalability of user numbers. On the other hand, an IDMA system in Fig. 13.16
based on symbol-level interleaving does not have this problem, since all available
RBs are fully used for any value of Kactive .
444
Y. Hu and L. Ping
Fig. 13.17 Protograph
representation of sparse
coding
|
RB1
|
y(1)
RB2
|
y(2)
(1)
RB3
(2)
RB4
|
y(4)
(3)
(4)
+
+
|
y(3)
+
+
x1
x2
x3
x4
x5
x6
c1
c2
c3
c4
c5
c6
Fig. 13.18 IDMA with GA
versus SCMA with ML. Ten
iterations for both schemes
0
10
BLER
SCMA with ML
IDMA with GA
-1
10
-2
10
-4
-2
0
2
4
6
8
single-user SNR (dB)
Also, multi-user gain in MIMO is determined by the number of users concurrently
transmitting in each RB. Therefore, sparse coding may not be an efficient option in
MIMO (especially in massive MIMO).
Figure 13.18 compares IDMA and SCMA in quasi-static Rayleigh fading channels. The channels remain unchanged within each transmission. Both schemes are
with six users, two receiver antennas, sum-rate = 3 and equal transmitted power per
user. A rate-1/2 LTE turbo code is used followed the following transmitter structures:
• IDMA is with a length-2 spreading and QPSK modulation.
• SCMA is based on Fig. 13.17 with the 16-point modulation in [92].
We can see from Fig. 13.18 that the two schemes have similar performance. However,
SCMA in Fig. 13.18 is based on ML, while IDMA based on GA. The latter has much
lower complexity.
13 Interleave Division Multiple Access (IDMA)
445
13.7 Summary
We have shown that the real attractiveness of NOMA is in systems without centralized
control or without accurate CSI. It is difficult or too costly to establish orthogonality
in such channels, so we have to resort to NOMA. Iterative processing holds the key;
interference can be gradually resolved and CSI can be gradually refined through
iterative processing. IDMA is a simple implementation technique for NOMA. The
features of IMDA can be seen from its sparse graphic representation. The interleaved
edge connections in IDMA facilitate iterative processing at the receiver. We have
demonstrated that IDMA can offer significant performance gain in random access
and MIMO systems. IDMA also offers lower detection complexity compare with
other alternatives.
References
1. R. Ahlswede, Multi-way communication channels, in Second International Symposium on
Information Theory (1971), pp. 103–135
2. H. Liao, A coding theorem for multiple access communications, in Proceedings of the International Symposium on Information Theory, Asilomar (1972)
3. D. Tse, P. Viswanath, Fundamentals of Wireless Communication (Cambridge University Press,
2005)
4. P. Viswanath, D.N.C. Tse, V. Anantharam, Asymptotically optimal water-filling in vector
multiple-access channels. IEEE Trans. Inf. Theory 47(1), 241–267 (2001)
5. W. Yu, W. Rhee, S. Boyd, J.M. Cioffi, Iterative water-filling for Gaussian vector multiple-access
channels. IEEE Trans. Inf. Theory 50(1), 145–152 (2004)
6. M. Kobayashi, G. Caire, An iterative water-filling algorithm for maximum weighted sum-rate
of Gaussian MIMO-BC. IEEE J. Sel. Areas Commun. 24(8), 1640–1646 (2006)
7. R.G. Gallager, An inequality on the capacity region of multiaccess multipath channels, in
Communications and Cryptography: Two Sides of One Tapestry (Kluwer, Boston, 1994), pp.
129–139
8. R.R. Muller, A. Lampe, J.B. Huber, Gaussian multiple-access channels with weighted energy
constraint, in IEEE Information Theory Workshop, June 1998, pp. 106–107
9. P. Wang, J. Xiao, L. Ping, Comparison of orthogonal and non-orthogonal approaches to future
wireless cellular systems. IEEE Veh. Technol. Mag. 1(3), 4–11 (2006)
10. P. Wang, L. Ping, On maximum eigenmode beamforming and multi-user gain. IEEE Trans.
Inf. Theory 57(7), 4170–4186 (2011)
11. N. Otao, Y. Kishiyama, K. Higuchi, Performance of non-orthogonal access with SIC in cellular
downlink using proportional fair-based resource allocation, in IEEE ISWCS 2012, August 2012,
pp. 476–480
12. Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, I. Chih-Lin, H.V. Poor, Application of nonorthogonal multiple access in LTE and 5G networks. IEEE Commun. Mag. 55(2), 185–191
(2017)
13. L. Dai, B. Wang, Y. Yuan, S. Han, I. Chih-Lin, Z. Wang, Non-orthogonal multiple access
for 5G: solutions, challenges, opportunities, and future research trends. IEEE Commun. Mag.
53(9), 74–81 (2015)
14. G. Wunder, P. Jung, M. Kasparick, T. Wild, F. Schaich, Y. Chen, L. Mendes, 5GNOW: nonorthogonal, asynchronous waveforms for future mobile applications. IEEE Commun. Mag.
52(2), 97–105 (2014)
446
Y. Hu and L. Ping
15. F. Schaich, B. Sayrac, S. Elayoubi, I.P. Belikaidis, M. Caretti, A. Georgakopoulos, B.
Mouhouche, FANTASTIC5G: flexible air interface for scalable service delivery within wireless
communication networks of the 5th generation. Trans. Emerg. Telecommun. Technol. 27(9),
1216–1224 (2016)
16. W. Shin, M. Vaezi, B. Lee, D.J. Love, J. Lee, H.V. Poor, Non-orthogonal multiple access
in multi-cell networks: theory, performance, and practical challenges. IEEE Commun. Mag.
55(10), 176183 (2017)
17. L. Ping, L. Liu, K. Wu, W.K. Leung, Interleave division multiple-access. IEEE Trans. Wirel.
Commun. 5(4), 938–947 (2006)
18. P.A. Hoeher, H. Schoeneich, J.C. Fricke, Multi-layer interleave-division multiple access: theory
and practice. Trans. Emerg. Telecommun. Technol. 19(5), 523–536 (2008)
19. T. Yang, J. Yuan, Z. Shi, Rate optimization for IDMA systems with iterative joint multi-user
decoding. IEEE Trans. Wirel. Commun. 8(3), 1148–1153 (2009)
20. K. Kusume, G. Bauch, W. Utschick, IDMA vs. CDMA: analysis and comparison of two multiple
access schemes. IEEE Trans. Wirel. Commun. 11(1), 78–87 (2012)
21. K. Wu, K. Anwar, T. Matsumoto, BICM-ID-based IDMA: convergence and rate region analyses. IEICE Trans. Commun. 97(7), 1483–1492 (2014)
22. G. Song, J. Cheng, Distance enumerator analysis for interleave-division multi-user codes. IEEE
Trans. Inf. Theory 62(7), 4039–4053 (2016)
23. C. Berrou, A. Glavieux, P. Thitimajshima, Near Shannon limit error-correcting coding and
decoding: turbo-codes. 1, in IEEE ICC’93, vol. 2, May 1993, pp. 1064–1070
24. S. Benedetto, G. Montorsi, Unveiling turbo codes: some results on parallel concatenated coding
schemes. IEEE Trans. Inf. Theory 42(2), 409–428 (1996)
25. T. Richardson, R. Urbanke, Modern Coding Theory (Cambridge University Press, 2008)
26. C. Douillard, M. Jzquel, C. Berrou, D. Electronique, A. Picart, P. Didier, A. Glavieux, Iterative correction of intersymbol interference: turbo-equalization. Trans. Emerg. Telecommun.
Technol. 6(5), 507–511 (1995)
27. X. Wang, H.V. Poor, Iterative (turbo) soft interference cancellation and decoding for coded
CDMA. IEEE Trans. Commun. 47(7), 1046–1061 (1999)
28. M.C. Reed, C.B. Schlegel, P.D. Alexander, J.A. Asenstorfer, Iterative multiuser detection for
CDMA with FEC: near-single-user performance. IEEE Trans. Commun. 46(12), 1693–1699
(1998)
29. F.N. Brannstrom, T.M. Aulin, L.K. Rasmussen, Iterative multi-user detection of trellis code
multiple access using a posteriori probabilities, in IEEE ICC 2001, vol. 1, June 2001, pp. 11–15
30. M. Moher, An iterative multiuser decoder for near-capacity communications. IEEE Trans.
Commun. 46(7), 870–880 (1998)
31. L. Ping, L. Liu, W.K. Leung, A simple approach to near-optimal multiuser detection: interleavedivision multiple-access, in IEEE WCNC 2003, vol. 1, March 2003, pp. 391–396
32. F.R. Kschischang, B.J. Frey, H.A. Loeliger, Factor graphs and the sum-product algorithm. IEEE
Trans. Inf. Theory 47(2), 498–519 (2001)
33. Y. Mao, A.H. Banihashemi, A heuristic search for good low-density parity-check codes at short
block lengths, in IEEE ICC 2001, vol. 1, June 2001, pp. 41–44
34. R. Zhang, L. Xu, S. Chen, L. Hanzo, Repeat accumulate code division multiple access and its
hybrid detection, in IEEE ICC’08, May 2008, pp. 4790–4794
35. Z. Chenghai, H. Jianhao, The shifting interleaver design based on PN sequence for IDMA
systems, in IEEE FGCN 2007, vol. 2, December 2007, pp. 279–284
36. M. Noemm, T. Wo, P.A. Hoeher, Multilayer APP detection for IDM. Electron. Lett. 46(1),
96–97 (2010)
37. D.L. Donoho, A. Maleki, A. Montanari, Message-passing algorithms for compressed sensing.
Proc. Natl. Acad. Sci. 106(45), 18914–18919 (2009)
38. J. Thorpe, Low-density parity-check (LDPC) codes constructed from protographs. IPN Prog.
Rep. 42(154), 42–154 (2003)
39. L. Ping, J. Tong, X. Yuan, Q. Guo, Superposition coded modulation and iterative linear MMSE
detection. IEEE J. Sel. Areas Commun. 27(6) (2009)
13 Interleave Division Multiple Access (IDMA)
447
40. P.A. Hoeher, T. Wo, Superposition modulation: myths and facts. IEEE Commun. Mag. 49(12)
(2011)
41. R. Laroia, N. Farvardin, S.A. Tretter, On optimal shaping of multidimensional constellations.
IEEE Trans. Inf. Theory 40(4), 1044–1056 (1994)
42. X. Ma, L. Ping, Coded modulation using superimposed binary codes. IEEE Trans. Inf. Theory
50(12), 3331–3343 (2004)
43. C.H.F. Fung, W. Yu, T.J. Lim, Precoding for the multiantenna downlink: multiuser SNR gap
and optimal user ordering. IEEE Trans. Commun. 55(1), 188–197 (2007)
44. L. Ping, L. Liu, Analysis and design of IDMA systems based on SNR evolution and power
allocation, in IEEE VTC2004-Fall, vol. 2, September 2004, pp. 1068–1072
45. L. Liu, J. Tong, L. Ping, Analysis and optimization of CDMA systems with chip-level interleavers. IEEE J. Sel. Areas Commun. 24(1), 141–150 (2006)
46. X. Yuan, L. Ping, C. Xu, A. Kavcic, Achievable rates of MIMO systems with linear precoding
and iterative LMMSE detection. IEEE Trans. Inf. Theory 60(11), 7073–7089 (2014)
47. Y. Zhang, K. Peng, J. Song, Enhanced IDMA with rate-compatible raptor-like quasi-cyclic
LDPC code for 5G, in IEEE GC Workshops 2017 (2017)
48. K. Takeuchi, T. Tanaka, T. Kawabata, Improvement of BP-based CDMA multiuser detection
by spatial coupling, in IEEE ISIT 2011, July 2011, pp. 1489–1493
49. D. Truhachev, C. Schlegel, Spatially coupled streaming modulation, in IEEE ICC 2013, June
2013, pp. 3418–3422
50. S. Kudekar, T. Richardson, R.L. Urbanke, Spatially coupled ensembles universally achieve
capacity under belief propagation. IEEE Trans. Inf. Theory 59(12), 7761–7813 (2013)
51. C. Schlegel, D. Truhachev, Multiple access demodulation in the lifted signal graph with spatial
coupling. IEEE Trans. Inf. Theory 59(4), 2459–2470 (2013)
52. A. Yedla, Y.Y. Jian, P.S. Nguyen, H.D. Pfister, A simple proof of Maxwell saturation for coupled
scalar recursions. IEEE Trans. Inf. Theory 60(11), 6943–6965 (2014)
53. S. Kumar, A.J. Young, N. Macris, H.D. Pfister, Threshold saturation for spatially coupled LDPC
and LDGM codes on BMS channels. IEEE Trans. Inf. Theory 60(12), 7389–7415 (2014)
54. D.J. Costello, L. Dolecek, T. Fuja, J. Kliewer, D. Mitchell, R. Smarandache, Spatially coupled
sparse codes on graphs: theory and practice. IEEE Commun. Mag. 52(7), 168–176 (2014)
55. D.G. Mitchell, M. Lentmaier, D.J. Costello, Spatially coupled LDPC codes constructed from
protographs. IEEE Trans. Inf. Theory 61(9), 4866–4889 (2015)
56. C. Liang, J. Ma, L. Ping, Towards Gaussian capacity, universality and short block length, in
IEEE ISTC 2016, September 2016, pp. 412–416
57. C. Liang, J. Ma, L. Ping, Compressed FEC codes with spatial-coupling. IEEE Commun. Lett.
21(5), 987–990 (2017)
58. H.H. Chung, Y.C. Tsai, M.C. Lin, IDMA using non-Gray labelled modulation. IEEE Trans.
Commun. 59(9), 2492–2501 (2011)
59. D.P. Bertsekas, R.G. Gallager, P. Humblet, Data Networks, vol. 2 (Prentice-Hall, Englewood
Cliffs, NJ, 1987)
60. S. Ghez, S. Verdu, S.C. Schwartz, Stability properties of slotted Aloha with multipacket reception capability. IEEE Trans. Autom. Control 33(7), 640–649 (1988)
61. M.H. Ngo, V. Krishnamurthy, L. Tong, Optimal channel-aware ALOHA protocol for random
access in WLANs with multipacket reception and decentralized channel state information.
IEEE Trans. Signal Process. 56(6), 2575–2588 (2008)
62. C. Xu, P. Wang, S. Chan, L. Ping, Decentralized power control for random access with iterative
multi-user detection, in IEEE ISTC 2012, August 2012, pp. 11–15
63. C. Xu, L. Ping, P. Wang, S. Chan, X. Lin, Decentralized power control for random access with
successive interference cancellation. IEEE J. Sel. Areas Commun. 31(11), 2387–2396 (2013)
64. H. Lin, K. Ishibashi, W.Y. Shin, T. Fujii, A simple random access scheme with multilevel power
allocation. IEEE Commun. Lett. 19(12), 2118–2121 (2015)
65. M. Zou, S. Chan, H.L. Vu, L. Ping, Throughput improvement of 802.11 networks via randomization of transmission power levels. IEEE Trans. Veh. Technol. 65(4), 2703–2714 (2016)
448
Y. Hu and L. Ping
66. C. Xu, X. Wang, L. Ping, Random access with massive-antenna arrays, in IEEE VTC2016Spring, May 2016, pp. 1–5
67. Y. Hu, C. Xu, L. Ping, NOMA and IDMA in random access, invited paper, in IEEE VTC2018Spring, June 2018, pp. 1–5
68. J. Mietzner, R. Schober, L. Lampe, W.H. Gerstacker, P.A. Hoeher, Multiple-antenna techniques
for wireless communications-a comprehensive literature survey. IEEE Commun. Surv. Tutor.
11(2) (2009)
69. G. Caire, N. Jindal, M. Kobayashi, N. Ravindran, Multiuser MIMO achievable rates with
downlink training and channel state feedback. IEEE Trans. Inf. Theory 56(6), 2845–2866
(2010)
70. D. Gesbert, S. Hanly, H. Huang, S.S. Shitz, O. Simeone, W. Yu, Multi-cell MIMO cooperative
networks: a new look at interference. IEEE J. Sel. Areas Commun. 28(9), 1380–1408 (2010)
71. F. Rusek, D. Persson, B.K. Lau, E.G. Larsson, T.L. Marzetta, O. Edfors, F. Tufvesson, Scaling
up MIMO: opportunities and challenges with very large arrays. IEEE Signal Process. Mag.
30(1), 40–60 (2013)
72. J. Hoydis, S. Ten Brink, M. Debbah, Massive MIMO in the UL/DL of cellular networks: how
many antennas do we need? IEEE J. Sel. Areas Commun. 31(2), 160–171 (2013)
73. L. Lu, G.Y. Li, A.L. Swindlehurst, A. Ashikhmin, R. Zhang, An overview of massive MIMO:
benefits and challenges. IEEE J. Sel. Top. Signal Process. 8(5), 742–758 (2014)
74. P.A. Hoeher, N. Doose, A massive MIMO terminal concept based on small-size multi-mode
antennas. Trans. Emerg. Telecommun. Technol. 28(2) (2017)
75. C. Xu, Y. Hu, C. Liang, J. Ma, L. Ping, Massive MIMO, non-orthogonal multiple access and
interleave division multiple access. IEEE Access 5, 14728–14748 (2017)
76. S. Vishwanath, N. Jindal, A. Goldsmith, Duality, achievable rates, and sum-rate capacity of
Gaussian MIMO broadcast channels. IEEE Trans. Inf. Theory 49(10), 2658–2668 (2003)
77. W. Yu, Uplink-downlink duality via minimax duality. IEEE Trans. Inf. Theory 52(2), 361–374
(2006)
78. E.G. Larsson, O. Edfors, F. Tufvesson, T.L. Marzetta, Massive MIMO for next generation
wireless systems. IEEE Commun. Mag. 52(2), 186–195 (2014)
79. H. Schoeneich, P.A. Hoeher, Iterative pilot-layer aided channel estimation with emphasis on
interleave-division multiple access systems. EURASIP J. Appl. Signal Process. 2006, 250–250
(2006)
80. M. Zhao, Z. Shi, M.C. Reed, Iterative turbo channel estimation for OFDM system over rapid
dispersive fading channel. IEEE Trans. Wirel. Commun. 7(8) (2008)
81. C. Novak, G. Matz, F. Hlawatsch, IDMA for the multiuser MIMO-OFDM uplink: a factor
graph framework for joint data detection and channel estimation. IEEE Trans. Signal Process.
61(16), 4051–4066 (2013)
82. J. Ma, L. Ping, Data-aided channel estimation in large antenna systems. IEEE Trans. Signal
Process. 62(12), 3111–3124 (2014)
83. P. Hoeher, F. Tufvesson, Channel estimation with superimposed pilot sequence, in IEEE
GLOBECOM’99, vol. 4 (1999), pp. 2162–2166
84. A. Ashikhmin, T. Marzetta, Pilot contamination precoding in multi-cell large scale antenna
systems, in IEEE ISIT 2012, July 2012, pp. 1137–1141
85. J. Ma, C. Liang, C. Xu, L. Ping, On orthogonal and superimposed pilot schemes in massive
MIMO NOMA systems. IEEE J. Sel. Areas Commun. 35(12), 2696–2707 (2017)
86. Y. Chen, Low-cost superimposed pilots based receiver for massive MIMO in multicarrier
system, in IEEE VTC2017-Spring, June 2017
87. Nokia, Alcatel-Lucent Shanghai Bell, Performance of interleave division multiple access
(IDMA) in combination with OFDM family waveforms, document R1-165021, in 3GPP TSGRAN WG1 #85 (2016)
88. Qualcomm Incorporated, RSMA, document R1-164688, in 3GPP TSG-RAN WG1 #85 (2016)
89. Samsung, Link level performance evaluation for IGMA, document R1-166750, in 3GPP TSG
RAN WG1 Meeting #86 (2016)
13 Interleave Division Multiple Access (IDMA)
449
90. S. Chen, B. Ren, Q. Gao, S. Kang, S. Sun, K. Niu, Pattern division multiple access—a novel
nonorthogonal multiple access for fifth-generation radio networks. IEEE Trans. Veh. Technol.
66(4), 3185–3196 (2017)
91. H. Nikopour, H. Baligh, Sparse code multiple access, in IEEE PIMRC’13, September 2013,
pp. 332–336
92. HiSilicon, Huawei, LLS results for uplink multiple access document R1-164037, 3GPP TSG
RAN WG1 Meeting #85, May 2016
93. L. Ping, L. Liu, K.Y. Wu, W.K. Leung, Approaching the capacity of multiple access channels
using interleaved low-rate codes. IEEE Commun. Lett. 8(1), 4–6 (2004)
Chapter 14
Pattern Division Multiple Access
(PDMA)
Shanzhi Chen, Shaohui Sun, Shaoli Kang and Bin Ren
Generated from former research achievements on successive interference cancellation amenable multiple access (SAMA) [1–5] technology, pattern division multiple
access (PDMA) [6–8] was proposed in 2014. It is a type of non-orthogonal multiple access (NOMA) technology based on the principle of the introduced reasonable
diversity between multi-user to promote the capacity, which can obtain higher multiuser multiplexing and diversity gain by designing multi-user diversity PDMA Pattern
matrix to implement non-orthogonal signals transmission in such domains as time,
frequency, code, space and power.
This chapter provides a whole picture of PDMA, including the origination and
principle, pattern design, receiver algorithms, performance evaluation, extension
design, applications, challenges and trends.
14.1 Origination and Principles of PDMA
As described above, PDMA is originated from SAMA which takes successive interference cancellation (SIC) detection in the receiver. To solve the error propagation
problem of SIC, transmitter and receiver joint design is a good solution [8]. Therefore, in this section, based on the explanation of error propagation problem in SIC
S. Chen (B) · S. Sun · S. Kang · B. Ren
China Academy of Telecommunication Technologies, No. 40,
Xueyuan Road, Haidian District, Beijing, China
e-mail: chensz@datanggroup.cn
S. Sun
e-mail: sunshaohui@catt.cn
S. Kang
e-mail: kangshaoli@catt.cn
B. Ren
e-mail: renbin@catt.cn
© Springer International Publishing AG, part of Springer Nature 2019
M. Vaezi et al. (eds.), Multiple Access Techniques for 5G Wireless Networks
and Beyond, https://doi.org/10.1007/978-3-319-92090-0_14
451
452
S. Chen et al.
and the idea of transmitter receiver joint design, PDMA definition and framework
are described, and also PDMA transmitting and receiving schemes are explained.
14.1.1 Error Propagation Problem in SIC
According to theoretical results of multi-user channel [9], superposition coding at a
transmitter and SIC at a receiver, are able to achieve capacity boundary of multiple
access channels (MAC) or degraded broadcast channels (BC) when transmitter and
receiver are working together. From theoretical perspective it is rational to use SIC
to achieve channel capacity, since the packet error rate tends to be zero with the
increased code length as long as a user’s transmission rate is below the channel
capacity. However, in a real system, detection error is inevitable due to various nonideal conditions, such as, limited code length, channel fading, and glitches, etc.
For SIC receiver, if a former user’s packet is detected erroneously, it is very
unlikely that the following user’s packet could be detected correctly. This is the
so-called error propagation problem. Since multiple users are detected one by one
in serial order, the detection order of all users is usually arranged according to their
signal strength. That is, the signal of the first detected user is the strongest, the signal of
the second detected user is weaker, and so on. For the first detected user, it is recovered
directly from the original receiving. While for the following detected users, they are
recovered respectively from related cancellation receiving which should cancel those
former detected users from the original receiving by user reconstruction. If a user is
not correctly detected, its reconstruction is impossible to be correct. In addition, the
accuracy of reconstruction also impacts on the performance of following users. For
example, based on distorted channel estimation, the reconstructed signal will also be
distorted. Even though the user’s packet is detected correctly, it still has an adverse
effect on the following users’ detection.
Error propagation is a crushing blow for multi-user detection and it will deteriorate
the performance of SIC-based multi-user system. In general, two approaches can
be considered to alleviate the error propagation problem. The first is to enhance
the reliability of those early-decoded users, the second is to adopt more advanced
and complex detection algorithm than SIC. These approaches relate to joint design
between transmitter and receiver, which is the origination of PDMA technology.
14.1.2 Transmitter and Receiver Joint Design
One approach to alleviate error propagation problem is to enhance the reliability of
those early-decoded users, either by selecting users with good channel condition or
by designing transmission parameters such that the early-decoded users have higher
reliability and better channel condition.
14 Pattern Division Multiple Access (PDMA)
453
Analytical results of multiple-input and multiple-output (MIMO) detection from
[10, 11] show that the ith detected layers of SIC receiver could achieve diversity
order
(14.1)
Ndiv (i) = N R − N T + i
where N R is the receiving antenna number, N T is number of data layers.
The diversity order increases with the detection order. PDMA design is inspired
by above result [1–8]. Multi-user channel can be viewed as a virtual MIMO channel
and the above result could be generalized to multi-user non-orthogonal transmission. For non-orthogonal transmission employing SIC receiver, diversity order of
each user varies with the order of detection. The first detected user has the lowest
diversity order, and the last detected user has the highest diversity order. In a fading channel, diversity order affects transmission reliability significantly. Increasing
the diversity order typically leads to more reliable transmission. With SIC receiver,
the first detected user actually determines the overall detection performance, but
unfortunately its diversity order is the lowest. To optimize system performance, it is
desirable to have identical pro-detection diversity order for each user.
Diversity could be obtained from transmission or reception, or from both. Assuming that transmission diversity order of the ith detected user is DT (i), the diversity
order after SIC receiver can be expressed as
Ndiv (i) = DT (i) − K + i
(14.2)
where K is the number of users. By joint design from transmitter and receiver, PDMA
deliberately selects DT (i) so that the diversity order after SIC receiver is as close as
possible.
The definition of transmission diversity means that multiple copies of a signal
are transmitted from independent resources to avoid transmission error due to deep
fading on one resource. The resources could be time, frequency, code, spatial or
power resource.
PDMA maps transmitted data onto a group of resources according to PDMA pattern to realize disparate transmission diversity order. A PDMA pattern defines the
mapping from transmitted data to a resource group. A resource group can consist of
time resource, frequency resource, code resource, spatial resource, power resource
or any combination of these resources. The number of mapped resources in a group
determines the order of transmission diversity. Data of multiple users can be multiplexed onto the same resource group with different PDMA patterns. In this way,
non-orthogonal transmission is realized. By assigning PDMA pattern with different diversity order, disparate transmission diversity orders among users could be
achieved.
Another approach to alleviate error propagation problem is to adopt more
advanced and complex detection algorithm such as maximum likelihood (ML) or
maximum a posterior (MAP). It is anticipated that PDMA with advanced detection
algorithm can alleviate error propagation effect to a substantial degree. However, ML
or MAP algorithm incur tremendous detection complexity and it is hard to implement.
454
S. Chen et al.
Fig. 14.1 PDMA pattern for 6 users on 4 REs
Fortunately, the detection complexity could be reduced significantly by making the
PDMA pattern sparse. That is, data are only mapped to a small part of the resources in
the resource group. This draws on the idea of sparse coding in low density parity check
(LDPC) coding. Sparsity makes it possible to use low complexity belief propagation
(BP) algorithm to approach the MAP detection. In addition, the convergence of BP
algorithm could be speeded up by disparate transmission diversity of PDMA.
In summary, PDMA uses PDMA pattern to define sparse mapping from data to
a group of resources. PDMA pattern could be represented by a binary vector. The
dimension of the vector equals to number of resources in a group. Each element in
the vector corresponds to a resource in a resource group. A “1” means that data shall
be mapped to the corresponding resource. Actually, number of “1” in the PDMA
pattern is defined as its transmission diversity order.
Figure 14.1 shows an example of resource mapping according to PDMA pattern.
Six users are multiplexed on four resource elements (REs). A PDMA pattern is
assigned to a user. User1’s data are mapped to all four REs in the group, and user2’s
data are mapped to the first three REs, etc. The order of transmission diversity of the
six users is 4, 3, 2, 2, 1, and 1 respectively.
14.1.3 PDMA Definition and Framework
PDMA is proposed as a novel NOMA scheme based on code pattern. Joint optimization of transmitting and receiving is considered with SIC amenable pattern design at
the transmitter side and SIC-based detection at the receiver side. PDMA pattern is
designed to offer different orders of transmission diversity, so that disparate diversity
order between multiple users could be introduced to alleviate the error propagation
problem of SIC receiver. PDMA pattern is also required to be sparse to facilitate
advanced detection algorithm such as BP. Iterations between BP and channel decoding could further boost system performance. PDMA pattern can be also extended to
include power scaling and phase shifting to harvest additional constellation shaping
gain.
PDMA can design pattern for a specific user in time, frequency and space
resources. Figure 14.2 shows the technical framework of the PDMA uplink application, and Fig. 14.3 shows that of the PDMA downlink application. As shown
14 Pattern Division Multiple Access (PDMA)
455
Fig. 14.2 The technical framework of the PDMA uplink application
Fig. 14.3 The technical framework of the PDMA downlink application
in Figs. 14.2 and 14.3, the PDMA technical framework includes two parts: the
transmitter and the receiver, which reflects that the PDMA technology considers
the joint design of the transmitter and the receiver based on the optimization point
of view for multi-user communication system. At transmitter side, users are distinguished by the non-orthogonal characteristic pattern based on the multiple signals
domain (including time, frequency, code, power and the space domain, etc.). At the
receiver side, general SIC type sub-optimal multiuser detection algorithms can be
realized based on the features of the user pattern.
456
S. Chen et al.
14.1.4 PDMA Transmitting and Receiving Scheme
With above framework of PDMA, taking orthogonal frequency division multiplexing
(OFDM) waveform as a baseline, the transmitting and receiving schemes of a PDMA
based system are further explained.
Figure 14.4 shows the uplink process of the PDMA based system. At the transmitting end, the system completes transmitting signal processing by multiple user
data forward error correction channel coding, PDMA code modulation, PDMA subcarrier resource mapping and OFDM modulation. At the receiving end, the base
station performs the opposite process, namely the system gets the transmitting data
of each terminal through the OFDM demodulation and general SIC type detection
like belief propagation iterative detection and decoding (BP-IDD). In the process of
the PDMA modulation and coding, the symbol level mapping and spread spectrum in
the frequency domain are achieved at the same time. The receiver adopts the BP-IDD
algorithm which is essentially a joint iterative processing of the Turbo decoder and
BP detecting.
Figure 14.5 shows the downlink process of the PDMA based system. At the
transmitting end of the downlink PDMA system, the base station performs data forward error correction channel coding for multiple user like PDMA code modulation,
PDMA subcarrier resource mapping and OFDM modulation. At the receiving end
of the downlink PDMA system, each user performs the opposite process, including
Fig. 14.4 Illustration of the transmitting and receiving scheme of the PDMA based uplink system
14 Pattern Division Multiple Access (PDMA)
457
Fig. 14.5 Illustration of the transmitting and receiving schemes of the PDMA based downlink
system
OFDM demodulation and general SIC type detection like BP-IDD. In this process,
the downlink PDMA coding is conducted on modulation symbol level and finishes
mapping on symbol level and realizes spread spectrum in frequency domain. The
receiver adopts the BP-IDD algorithm which is essentially the Turbo decoding and
a joint iterative processing of BP detecting.
14.2 Pattern Design of PDMA
The PDMA pattern defines the rule of mapping data to the radio resource, which
can be defined as a binary vector. Each binary vector represents the PDMA pattern
of one user equipment (UE). The dimension of the vector equals to the number of
resources in a group. Those patterns with UEs sharing the same set of resources are
arranged together to constitute the PDMA pattern matrix. Overload factor (OF) is
defined as the ratio of the number of columns to the number of rows in a PDMA
pattern matrix. It reflects the excessive number of UEs multiplexed on the same
resources of PDMA relative to that of orthogonal multiple access (OMA) scheme.
Taking resource number N = 4 and user number K = 8 as an example, the OF is
then α = K /N = 200%, which means that PDMA supports two times the number
of UEs compared with that of OMA. Properties of PDMA pattern matrix such as
dimension and level of sparsity contribute to both receiver complexity and system
performance.
458
S. Chen et al.
14.2.1 Basic Pattern Matrix
Without loss of generality, both transmitter and receiver are assumed using single
antenna, K UEs map onto N REs in the domains of time and frequency, in which
each UE has a unique PDMA pattern. The PDMA received signal on the resource
group composed by N REs at the base station (BS) is expressed as:
y=
K
diag(hk )g k xk + n = H x + n
(14.3)
k=1
where y denotes a vector composed by received signal on the N resources, with length
N ; x = [x1 , x2 , . . . , x K ]T represents modulated signal vector transmitted by K UEs,
with length K , and xk are the modulated signals of kth UE; n indicates the Gaussian
,K ]
noise vector with length N , where n ∼ C N (0, N0 I N ×N ); H = H C H • G [N
P D M A and
H C H = [h1 , h2 , . . . , h K ] are the PDMA equivalent channel response matrix and
original channel response matrix of K UEs multiplexed on N REs, respectively and
both have dimensions of N × K , h K is the uplink channel response of the kth UE
with length N, diag(hk ) represents a diagonal matrix with elements from hk , the
(n, k) elements of H C H is the channel response from the kth UE to the BS on the nth
,K ]
RE, and • denotes element-wise product of two matrices; G [N
P D M A denotes a PDMA
pattern matrix with the dimensions of N × K , where g k is the PDMA pattern used
,K ]
by the kth UE, corresponding to the kth column of G [N
P DM A.
Given a certain overload factor, there are a number of pattern matrices available,
as long as resource number N and user number K are selected properly. For example,
overload factor of 150% could be achieved by a 2 × 3 pattern matrix, i.e., 3 users
are multiplexed on 2 REs. The pattern matrix is:
110
=
G [2,3]
P DM A
101
and another design for 150% overload is 4 × 6 pattern matrix:
⎡
⎤
101110
⎢1 1 0 1 0 1⎥
⎢
⎥
G [4,6]
P DM A = ⎣ 1 1 1 0 1 0 ⎦
011001
Though both pattern matrices having the same overload factor, G [4,6]
P D M A can
achieve better performance while the cost of detection complexity is higher comparing to that of G [2,3]
P DM A.
PDMA pattern matrices with different dimensions are able to achieve a given
overload factor. With a higher dimension, detection complexity is also higher, and
better performance is expected. Given an overload factor, the dimension of pattern
14 Pattern Division Multiple Access (PDMA)
459
matrix shall be selected to reach a tradeoff between detection complexity and system
performance.
If N is the size of resource group (row number of PDMA pattern matrix), there
are 2 N − 1 possible binary vectors for a pattern matrix. Assuming K is the column
number determined based on overload factor, we can thus choose K patterns out from
2 N − 1 candidates to construct PDMA pattern matrix.
Selection of patterns also gives impacts on performance and complexity:
(1) A pattern with heavier weight (number of “1” elements in the pattern) provides
higher diversity order. More reliable data transmission can be anticipated, and detection complexity is also increased. If the system can conduct complex computation,
patterns with heavy weight will be preferable; otherwise, light weight patterns have
to be selected, aiming at sparse PDMA pattern matrix.
(2) According to the design principle of PDMA, it is desirable to have different
diversity orders in the pattern matrix to alleviate error propagation problem of SIC
receiver or fasten convergence of BP receiver. Thus the selected patterns shall have
as many different diversity orders as possible.
(3) For patterns with identical diversity order, smaller inner product between the
patterns leads to less interference against each other. Small inner product means that
the two patterns have less “1” elements in common positions. That is, the number of
REs shared by the two patterns is low. Data of two users are multiplexed on only few
REs. For example, if two patterns have inner product of 0, the two patterns actually
map data onto a different set of REs, hence there is no interference between the
two patterns. For a given diversity order, the selected patterns shall minimize the
maximum inner product between any two patterns. Of course this rule is also applied
to patterns with different diversity order.
The design of pattern matrix shall take overload factor, diversity order and detection complexity into account. A good pattern matrix can reach good trade-off among
these aspects.
14.2.2 Pattern Optimization Method
Taking PDMA pattern matrix G [2,3]
P D M A as an example, data of 3 users are mapped
onto two REs. The transmission signal on these REs can be expressed as:
v1
v2
⎡ ⎤
x
1 1 0 ⎣ 1⎦
x2
=
101
x3
(14.4)
where v j is the transmission signal on the jth RE, and xk is the modulation symbol
of the k th user. Unlike orthogo