Energy Efficient Large Scale Antenna

Systems for 5G Communications and

Beyond

A thesis submitted for the degree of Doctor of Philosophy

(PhD)

by

Pierluigi Vito Amadori

Communications and Information Systems Research Group

Department of Electronic and Electrical Engineering

University College London

January 2017
 Statement of Originality

I, Pierluigi Vito Amadori confirm that the work presented in this thesis is my own.

Where information has been derived from other sources, I confirm that this has been

indicated in the thesis.

Signed:

Date:

i
ii
 Abstract

The increasing popularity of mobile devices has fueled an exponential growth in data traffic.
This phenomenon has led to the development of systems that achieve higher spectral efficiencies,
at the cost of higher power consumptions. Consequently, the investigation on solutions that
allow to increase the maximum throughput together with the energy efficiency becomes crucial
for modern wireless systems. This thesis aims to improve the trade-off between performances
and power consumption with special focus toward multiuser multiple-antenna communications,
due to their promising benefits in terms of spectral efficiency.

Research envisaged massive Multi-Input-Multi-Output (MIMO) systems as the main tech-

nology to meet these data traffic demands, as very large arrays lead to unprecedented data
throughputs and beamforming gains. However, larger arrays lead to increased power consump-
tion and hardware complexity, as each radiating element requires a radio frequency chain, which
is accountable for the highest percentage of the total power consumption. Nonetheless, the avail-
ability of a large number of antennas unveils the possibility to wisely select a subset of radiating
elements. This thesis shows that multiuser interference can be exploited to increase the received
power, with significant circuit power savings at the base station.

Similarly, millimeter-wave communications experienced raising interest among the scientific

community because of their multi-GHz bandwidth and their ability to place large arrays in lim-
ited physical spaces. Millimeter-wave systems inherit same benefits and weaknesses of massive
MIMO communications. However, antenna selection is not viable in millimeter-wave communi-
cations because they rely on high beamforming gains. Therefore, this thesis proposes a scheme
that is able to reduce the number of radio frequency chains required, while achieving close-to-
optimal performances.

Analytical and numerical results show that the proposed techniques are able to improve the
overall energy efficiency with respect to the state-of-the-art, hence proving to be valid candidates
for practical implementations of modern communication systems.

iii
iv
 Acknowledgements

When I look back at the years that led to this thesis, there are a few people that I believe
need to be properly mentioned and thanked.

First and foremost, I would like to thank my supervisor Dr. Christos Masouros for believing
in me and for giving me such a great opportunity. I truly appreciate his never-ending patience
during these years, especially at the first stages of my PhD. After reading my first reports and
each chapter of this thesis, I feel like I cannot thank him enough for his constructive comments
and guidance. Words cannot express how thankful I am for trusting me when I asked to work
in remote. I would also like to thank Prof. Izzat Darwazeh for his help and support towards my
professional development.

I would like to thank my colleagues at the Communications and Information Systems group
in UCL, for their help since my arrival and for creating such a friendly environment to work in.
I would like to dedicate a special thanks to Dr. Adrian Garcia-Rodriguez, for always lending an
ear and a hand in those moments when both were greatly needed and for our daily discussions
on the way home. I would also like to thank Dr. Francesco Renna, Dr. Joao De Castro Mota,
Dr. Ka Lung Law, Mr. Jure Sokolic and Mr. Ang Li.

Additionally, I would like to thank my family. Grazie a Mamma, Papà e Marina, per
esserci stati in tutti questi anni, per avermi ascoltato quando rimpiangevo ogni mia decisione
e per avermi rincuorato quando mi preoccupavo di ogni cosa. Vorrei ringraziarvi per essere
come siete, senza mai chiedere nulla in cambio e per volermi bene, anche quando penso di non
meritarlo. Non c’é giorno che non vi pensi e che non speri di rendervi tanto orgogliosi di me,
quanto io lo sono di ognuno di voi. I ringraziamenti alla famiglia non possono finire senza un
pensiero speciale per i miei nonni. Vorrei ringraziare Nonna Savina e Nonno Pierino, per avermi
trasmesso il valore e la importanza di studio e impegno, specialmente quando da piccolo avrei
preferito passare il tempo facendo altro.

Finalment, desitjo dedicar les últimes paraules d’agraı̈ment per la Laura. Voldria dir-te
gràcies per la teva paciència, quan he començat a aprendre català, per la teva ajuda i força, en
tots els moments del meu doctorat, per les nostres trucades diàries a la nit, quan estàvem vivint
separats, i per haver decidit de venir a Londres per viure amb mi, en aquella trucada del 18 de
desembre. Gràcies per tot, per la felicitat que em dones cada dia i per haver canviat la meva
vida en el millor que hagués mai pogut desitjar.

v
vi
Contents

List of Figures xi

List of Tables xv

List of Abbreviations xvii

List of Symbols and Operators xxi

1 Introduction 1
1.1 Aim and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Energy Efficiency in Multiple Antenna Wireless Communications

Systems 9
2.1 MIMO Communications . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 MIMO Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Linear Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Non-linear Precoding . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.3 Optimization-based Beamforming . . . . . . . . . . . . . . . . . . 15
2.3 AS Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.1 Norm-based Selection . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.2 Capacity-based Selection . . . . . . . . . . . . . . . . . . . . . . 19
2.3.3 Energy Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4 Large Scale MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1 Multiuser M-MIMO Channel Model . . . . . . . . . . . . . . . . 23
2.4.2 Capacity and Signal Processing Benefits . . . . . . . . . . . . . . 24

vii
 2.4.3 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.5 Millimeter Wave MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5.1 Propagation Characteristics . . . . . . . . . . . . . . . . . . . . . 30
2.5.2 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.5.3 Beamspace mm-wave MIMO . . . . . . . . . . . . . . . . . . . . 32
2.6 Constructive Interference . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.6.1 Constructive Interference Regions for PSK modulated transmissions 35

3 Interference Exploiting Constant Envelope Precoding in Massive

MIMO 41
3.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.1.1 Benchmark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Constructive Interference Optimization Region . . . . . . . . . . . . . . 45
3.3 Constant Envelope Precoding with Constructive Interference Optimization 47
3.3.1 A CEM Solver for Constructive Interference Optimization . . . . 48
3.3.2 Two-Step Convex CEP . . . . . . . . . . . . . . . . . . . . . . . 53
3.4 Computational Complexity . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4.1 CEO-CIO Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4.2 CEO-IR Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.5 CSI-Robust CEP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.6.1 Constellation Energy . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4 Low-Complexity Interference Exploiting Antenna Selection for Mas-

sive MIMO 71
4.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 Proposed Constructive Interference Maximization Antenna Selection (CIM) 74
4.3 System Computational Complexity . . . . . . . . . . . . . . . . . . . . . 77
4.3.1 Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.3.2 Antenna Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.3.3 Transceiver Computational Costs Analysis . . . . . . . . . . . . . 82
4.4 Performance Analysis of CIM selection with MF precoding . . . . . . . 86

viii
 4.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

5 Large Scale Antenna Selection and Precoding for Interference Ex-

ploitation 95
5.1 System Model and Benchmark Techniques . . . . . . . . . . . . . . . . . 97
5.2 Proposed joint MIP Constructive Antenna Selection and Precoding (MIP-
CASP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.3 Heuristic Approaches to Joint Antenna Selection and Precoding . . . . . 99
5.3.1 3-step Successive Optimization CBF-CAS-CBF . . . . . . . . . . 100
5.3.2 2-step Successive Optimization MFCAS-CBF . . . . . . . . . . . 101
5.3.3 1-step Successive Optimization MFCAS . . . . . . . . . . . . . . 102
5.4 Optimality Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.5 Computational Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.6 Channel Uncertainty and Robust Approach . . . . . . . . . . . . . . . . 107
5.6.1 Model and Effects . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.6.2 MIP-CASP Robust Scheme . . . . . . . . . . . . . . . . . . . . . 109
5.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.7.1 Symbol Error Rate . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.7.2 Data Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.7.3 Energy Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.7.4 Robustness to CSI . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

6 Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Sys-

tems 119
6.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.1.1 Benchmark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2 Proposed Beam Selection Techniques . . . . . . . . . . . . . . . . . . . . 124
6.2.1 SINR Maximization Beam Selection (SM-BS) . . . . . . . . . . . 125
6.2.2 Capacity Maximization Beam Selection (CM-BS) . . . . . . . . . 127
6.3 Computational Complexity Analysis . . . . . . . . . . . . . . . . . . . . 131
6.4 Performance Analysis - Capacity Loss . . . . . . . . . . . . . . . . . . . 133

ix
 6.5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.5.1 Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.5.2 Mean number of beams . . . . . . . . . . . . . . . . . . . . . . . 140
6.5.3 Energy Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

7 Conclusions 147
7.1 Summary and Conclusions of the Thesis . . . . . . . . . . . . . . . . . . 147
7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

Appendices 153
Appendix A. Proof of Theorem 4.4.1 . . . . . . . . . . . . . . . . . . . . . . . 153
Appendix A. Proof of Theorem 5.6.1 . . . . . . . . . . . . . . . . . . . . . . . 156
Appendix A. Proof of Sherman-Morrison-Woodbury Identity . . . . . . . . . 158

List of References 161

x
List of Figures

1.1 Thesis Structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.1 Block diagram of a point-to-point 6 × 6 MIMO. . . . . . . . . . . . . . . 10

2.2 Block diagram of a 6 × 6 MU-MIMO. . . . . . . . . . . . . . . . . . . . 11
2.3 Downlink Precoding and Beamforming classification and representative
techniques. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Transmitter AS for MU-MIMO . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Conventional TAS classification . . . . . . . . . . . . . . . . . . . . . . . 18
2.6 Multiuser Massive MIMO representation . . . . . . . . . . . . . . . . . . 22
2.7 TDD protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.8 Visual representation of Pilot Contamination . . . . . . . . . . . . . . . 27
2.9 Visual representation of Friis equation . . . . . . . . . . . . . . . . . . . 31
2.10 Constructive regions for different constellations . . . . . . . . . . . . . . 35
2.11 Visual representation of phase-shift in Convex metrics for 8-PSK . . . . 38

3.1 Optimization region for constructive interference exploitation, 8PSK ex-

ample: (a)Interference regions for 8-PSK symbol, (b) Interference signal
after rotation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.2 Received symbols for a noiseless scenario with N = 100 antennas for
M = 20 users when using 8-PSK. . . . . . . . . . . . . . . . . . . . . . . 53
3.3 Imperfect CSI effects on the phase-shifted interfering signal t̄m . . . . . . 59
3.4 4-PSK Symbol Error Rate when M = 12, N = 64 with perfect CSI. . . . 62
3.5 8-PSK Symbol Error Rate when M = 12, N = 64 with perfect CSI. . . . 63
3.6 Symbol Error Rate as a function of the transmitted SNR when M = 6,
N = 32 with perfect CSI. . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.7 Symbol Error Rate for 8-PSK modulation when M = 12, N = 64 with
imperfect CSI and δ 2 = 0.1. . . . . . . . . . . . . . . . . . . . . . . . . . 65

xi
3.8 Symbol Error Rate for 8-PSK modulation when M = 12, N = 64 and
SN R = 10dB with imperfect CSI. . . . . . . . . . . . . . . . . . . . . . 66
3.9 Symbol Error Rate for 8-PSK modulation when M = 6 and N = 32 and
SN R = 0dB as a function of FLOPs per-frame. . . . . . . . . . . . . . . 67
3.10 Symbol Error Rate as a function of the constellation energy E = d2m =
d2 , ∀m ∈ {1, ...., M } when M = 6 and N = 32. . . . . . . . . . . . . . . . 68
3.11 Symbol Error Rate as a function of the constellation energy E = d2m =
d2 , ∀m ∈ {1, ...., M } when M = 12 and N = 64. . . . . . . . . . . . . . . 69

4.1 Computational costs as a function of the number of antennas at the trans-

mitter N for a 2-PSK modulated system with M = 5, NRF = M . . . . . 81
4.2 Computational costs as a function of Tcohe for a 2-PSK modulated system
with M = 5, NRF = M . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.3 SER as a function of the transmitted SNR for 2-PSK modulation when
N = 100, M = 5 and NRF = 5. . . . . . . . . . . . . . . . . . . . . . . . 88
4.4 SER as a function of the transmitted SNR for 4-PSK modulation when
N = 100, M = 5 and NRF = 10. . . . . . . . . . . . . . . . . . . . . . . . 88
4.5 SER as a function of the transmitted SNR for 2-PSK modulation when
N = 100, M = 5 and NRF = 5 with imperfect CSI at the transmitter
α = 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.6 Energy efficiency over throughput as a function of the transmitted SNR
for 2-PSK modulation when N = 100, M = 5 and NRF = 5. . . . . . . . 90
4.7 Energy efficiency over throughput as a function of the transmitted sym-
bols TDL when SN R = 10dB, N = 100 and M = 5. . . . . . . . . . . . 91
4.8 Energy efficiency over throughput as a function of the subset size NRF
when SN R = 0dB, N = 100 and M = 5. . . . . . . . . . . . . . . . . . . 92
4.9 Circuit power consumption at the BS as a function of the arrays size N
for a system with M = 5 mobile stations and NRF = 5. . . . . . . . . . . 93
4.10 RF Power Savings ζ as a function of the arrays size N at the transmitter
for a system with M = 5 mobile stations and NRF = 5. . . . . . . . . . . 93

5.1 Conventional TAS block diagram. . . . . . . . . . . . . . . . . . . . . . . 96

5.2 Minimum CI when M = 5 and N = 100. . . . . . . . . . . . . . . . . . . 105
5.3 Figure of merit M when M = 5 and N = 100. . . . . . . . . . . . . . . 106

xii
5.4 Frame Running time when M = 5, NRF = 5 and TDL = 4. . . . . . . . . 107
5.5 4-PSK Symbol Error Rate when M = 5, N = 100 and NRF = 5 with
perfect CSI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.6 8-PSK Symbol Error Rate when M = 5, N = 100 and NRF = 5 with
perfect CSI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.7 System Capacity comparison when M = 5, NRF = 5 and N = 100. . . . 113
5.8 4-PSK Energy Efficiency ηT when M = 5, N = 100 with perfect CSI and
SN R = 5dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.9 8-PSK Energy Efficiency ηT when M = 5, N = 100 with perfect CSI and
SN R = 10dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.10 Transmitted power for 4-PSK transmission when N = 100, M = 5 and
NRF = 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

6.1 Block diagram of a mm-wave DLA-based transceiver scheme . . . . . . 121

6.2 Comparison analytical and simulated Ψ for a n = 81 and M = 40 system 136
6.3 Capacity as a function of SNR[dB] in a line-of-sight scenario . . . . . . . 139
6.4 Capacity as a function of SNR[dB] in a multi-path scenario . . . . . . . 139
6.5 Mean number of beams (RF chains) used for transmission as a function
of the number of users in a line-of-sight scenario . . . . . . . . . . . . . 141
6.6 Mean number of beams (RF chains) used for transmission as a function
of the number of users in a multi-path scenario with Np = 2 . . . . . . . 142
6.7 Energy efficiency as a function of M in a line-of-sight scenario with Pt =
15dBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.8 Energy efficiency as a function of M in a multi-path scenario with Np = 2
and Pt = 15dBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.9 Energy efficiency as a function of Pt . . . . . . . . . . . . . . . . . . . . 145

xiii
xiv
List of Tables

2.1 Spectrum usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.1 Computational costs in FLOPs. . . . . . . . . . . . . . . . . . . . . . . . 57

4.1 Computational costs of different schemes in FLOPs . . . . . . . . . . . . 79

5.1 Computational burdens for optimal Energy Efficiency points . . . . . . . 115

6.1 Complexity in number of operations . . . . . . . . . . . . . . . . . . . . 133

xv
xvi
List of Abbreviations

A
AS Antenna Selection

AWGN Additive White Gaussian Noise

B
B-MIMO Beamspace Multiple-Input Multiple-Output

BPSK Binary Phase Shift Keying

BS Base Station

C
CAS Constructive Antenna Selection

CBF Constructive Beamforming

CCM Convex Capacity Maximization

CECI Constant Envelope Constructive Interference

CEM Cross-Entropy Method

CEO-CIO Cross-Entropy Optimization - Constructive Interference Optimization

CEO-IR Cross-Entropy Optimization - Interference Reduction

CEP Constant Envelope Precoding

CI Constructive Interference

CIB Constructive Interference Beamforming

CIM Constructive Interference Maximization

CM Capacity Maximization

CM-BS Capacity Maximization Beam Selection

CR Correlation Rotation

CSI Channel State Information

CVX-CIO Convex - Constructive Interference Optimization

xvii
 DAC Digital-to-Analog Converter

DCM-BS Decremental Capacity Maximization Beam Selection

DLA Discrete Lens Array

DLP Dynamic Linear Precoding

DPC Dirty Paper Coding

DSP Digital Signal Processor

E
ERM Error-Rate Minimization

ESPAR Electronically Steerable Parasitic Array

F
FDD Frequency Division Duplexing

FLOP Floating-Point Operation

FPGA Field-Programmable Gate Array

G
GD Gradient Descent

I
ICM-BS Incremental Capacity Maximization Beam Selection

L
LOS Line-Of-Sight

LR Likelihood Ratio

LS-MIMO Large-Scale Multiple-Input Multiple-Output

LSA Large-Scale Array

M
M-MIMO Massive Multiple-Input Multiple-Output

MEM Minimum Eigenvalue Maximization

MF Matched Filter

MIMO Multiple-Input Multiple-Output

MIP Mixed-Integer Programming

MIP-CASP Mixed-Integer Programming Constructive Antenna Selection and Precoding

xviii
MISO Multiple-Input Single-Output

MM-BS Maximum Magnitude Beam Selection

mm-wave Millimeter Wave

MP Multi-Path

MU-MIMO Multiuser MIMO

N
NLS Nonlinear Least Squares

O
OFDM Orthogonal Frequency Division Multiplexing

P
PA Power Amplifier

PAPR Peak-to-Average Power Ratio

PGM Path Gain Maximization

PSK Phase Shift Keying

Q
QAM Quadrature Amplitude Modulation

QPSK Quadrature Phase Shift Keying

R
RF Radio Frequency

RZF Regularized Zero Forcing

S
SB SINR Balancing

SDR Semidefinite Relaxation

SER Symbol Error-Rate

SIMO Single-Input Multiple-Output

SINR Signal-to-Interference-plus-Noise Ratio

SISO Single-Input Single-Output

SM Spatial Modulation

SM-BS Signal-to-Interference and Noise Maximization Beam Selection

SM-MIMO Spatial Modulation Multiple-Input Multiple-Output

xix
SO Successive Optimization

SOCP Second-Order Cone Programming

SSK Space Shift Keying

SVD Singular Value Decomposition

T
TAS Transmit Antenna Selection

TDD Time Division Duplexing

THP Tomlinson-Harashima precoding

TPM Transmitted Power Minimization

U
UHF Ultra High Frequency

ULA Uniform Linear Array

UWB Ultra Wide Band

V
VLA Very Large Array

VP Vector Perturbation

Z
ZF Zero Forcing

xx
List of Symbols and Operators

|·| Absolute value of a complex number or determinant of a matrix

T
(·) Transpose of a vector or matrix
H
(·) Hermitian transpose of a vector or matrix
−1
(·) Inverse of a matrix
p mod L Modulo operation for a real number p with base L
diag (·) Diagonal of a matrix
rank (·) Rank of a matrix
∼ Indicates “distributed as”
CN (·, ·) Complex normal statistical distribution
E {·} Statistical expectation
b·c Floor function
d·e Ceiling function
C Set of complex numbers
R Set of real numbers
Z Set of integer numbers
B Set of binary matrices
< (·) Real component of a complex number
= (·) Imaginary component of a complex number
O (·) Complexity order
⊗ Kronecker product
k·kp `p norm of a vector or matrix
k·kF Frobenius norm of a vector or matrix
(·)! Factorial of a real number
⊆ Subset operator
⊂ Strict subset operator
A0 Positive semidefinite matrix A

xxi
∪ Union of sets
∩ Intersection of sets
∈ X ∈ Y indicates that X takes values from the set Y
∞ Infinity
det (·) Determinant of a matrix
Tr (·) Trace of a matrix
a


b Binomial coefficient indexed by a and b
, Equality by definition
min (·) Minimum entry of a vector
arg min (·) Index of the minimum entry in a vector
max (·) Maximum entry of a vector
arg max (·) Index of the maximum entry in a vector
λmax (·) Maximum singular value of a matrix
λmin (·) Minimum singular value of a matrix
σmax (·) Maximum eigenvalue value of a matrix
σmin (·) Minimum eigenvalue value of a matrix
var (·) Variance of random variables
∀ “For all”

xxii
Chapter 1

Introduction

The ever-increasing popularity of mobile devices and the success of social networking

have brought data traffic to experience an exponential growth over the last decade [1].

Recent studies predict that the global mobile data traffic is expected to reach a 66%

annual growth rate [2] in the next years. Consequently, the growing need of higher data

rates has inspired both research and academia to move towards new protocols and tech-

nologies that maximize spectral efficiency and throughput. However, research on energy

consumption and hardware complexity has been often neglected in the past, affecting

both battery lifetime for mobile devices and carbon emissions. In fact, information and

communication technology is identified as the cause for ∼ 2% of the global carbon emis-

sions, and this value is expected to increase [3] each year. The effects of data traffic

escalation are not just environmental, as the employment of data-based cellular standard

is leading to higher energy costs for the base station [4] as well. Because of this, future

5G communication systems are required to be able to provide both high data rate and

higher power consumption efficiency [5–7]. The need for a direct evaluation of the rela-

tionship between power consumption and achievable rates has brought the researchers

to define a new performance metric that combines both, called energy efficiency [5].

Energy efficiency is a quantifiable evaluation of the trade-off between the total power

consumption of a communication system and its spectral efficiency. As a consequence of

the increasing importance of energy efficiency, research has witnessed the emergence of

1
Chapter 1. Introduction

a new trend: green communications [8, 9]. Green communications concepts, which can

be summarized in the aim to maximize the energy efficiency of future wireless commu-

nication systems, have found a prolific applicability to multiple antennas technologies,

generally identified as MIMO systems [10]. In fact, while MIMO technology was first

introduced to maximize achievable throughputs, new power-saving strategies rapidly

showed that multiple-antennas systems were also able to provide energy efficient com-

munications [5]. Among these, large-scale or massive MIMO [11] has risen as one of the

most promising technologies for future wireless communications.

Massive MIMO technology can be simply described as a multiple-antenna system,

where the number of radiating elements is scaled up to the order of tens or hundreds

[11]. In fact, the employment of very large arrays at the base stations is expected

to offer unprecedented benefits, such as extremely high throughputs and beamforming

gains [12]. Additionally, the higher spatial resolutions offered by large antenna arrays

allow to greatly reduce the impact of noise and multiuser interference, while requiring

lower transmission power [11]. Even though the theoretical benefits of massive MIMO

systems are undoubtedly very appealing, their practical implementation are just as

equally challenging for both hardware requirements and signal processing. In fact, a

conventional MIMO approach to large-scale systems would require to dedicate one radio

frequency (RF) chain for each radiating element, hence leading base stations equipped

with very large arrays to require an equally large number of amplifiers, analog-to-digital

converters and mixers. The above are detailed in Section 2.4.3.3.

Because of these considerations, research has focused on technologies that can pro-

mote both practicability and energy efficiency of massive MIMO systems, such as

beamspace MIMO [13, 14], Antenna Selection (AS) [15, 16] and Constant Envelope Pre-

coding (CEP) [17, 18]. All these techniques share the aim for higher energy efficiency,

either by reducing the number of RF chains required by the transceiver, as in beamspace

MIMO and AS, or by employing highly efficient power amplifiers, as for CEP. Given the

importance and relevance of green communications for future wireless communication

systems, this thesis focuses on the design of novel energy efficient transmission schemes

and approaches.

2
 Chapter 1. Introduction

1.1 Aim and Motivation

While energy efficiency is regarded as one of the main metrics for realistic imple-

mentations of future large-scale systems, most of the energy efficient massive MIMO

technologies still present a significant number of open questions.

AS was firstly introduced as an effective tool for reducing the intrinsic hardware

complexity of MIMO [15, 16, 19, 20]. In fact, even in the early stages of MIMO com-

munications study, the need for a dedicated RF chain to each radiating element was

identified as the cause of higher, and often unsustainable, power consumptions either at

the transmitter or at the receiver side. However, the techniques that originated from

these studies have proven not to be directly scalable to massive systems, as their compu-

tational costs are deemed prohibitive [21]. As a consequence of these considerations, the

design of selection algorithms specifically tailored for very large arrays is particularly

relevant, since it allows to exploit both the higher diversity offered by a large number

of antennas and the power savings deriving from utilizing a lower number of RF chains.

Likewise, constant envelope transmissions [17,18] proved that a symbol-level precod-

ing with unit peak-to-average power ratio can facilitate the implementation of massive

MIMO systems by allowing the employment of efficient non-linear RF components.

However, their design metrics are based on the minimization of multiuser interference,

which supports the search for more efficient interference-exploiting approaches that can

further increase energy efficiency. Moreover, their performances are heavily reliant on

the availability of perfect channel-state information at the transmitter, which is non-

realistic.

While AS and precoding are both capable to deliver interesting benefits for energy-

efficient large-scale systems, they conventionally operate on a separate manner [15, 16,

19,20] as they are based on disjointed performance metrics. This justified the search for

a novel transmission scheme, where both techniques are jointly performed in order to

attain the maximum combined energy efficiency benefits that each of the two approaches

can offer.

3
Chapter 1. Introduction

Finally, the application of beamspace MIMO to millimeter wave communications

has shown that near-optimal performances can be achieved even with a strongly re-

duced number of RF chains at the transmitter, yet early studies mostly focused on

the simplistic assumption of exclusively line-of-sight scenarios [14, 22, 23], which are not

realistic according to recent measurements campaigns [24]. This motivates the develop-

ment of more complete techniques that include the effects of multi-path and interference

among mobile stations, in order to fully leverage on the diversity effects of a multiuser

scenario.

1.2 Main Contributions

This thesis aims to enhance energy efficiency and practicability of large-scale multi

antenna systems through the introduction of novel transmission schemes and the im-

provement of existing approaches from the literature. The following list highlights and

summarizes the main contributions of this thesis:

• Definition of an interference exploiting scheme for CEP and solution of the

corresponding optimization problem through a cross-entropy solver and a low-

complexity convex relaxation (Chapter 3). The performance of the proposed

schemes are analyzed through extensive numerical simulations and show that in-

terference exploitation can favorably affect reception and increase the received

power.

• Development of a computationally efficient AS algorithm that exploits multiuser

interference for increasing the received power in a massive MIMO downlink trans-

mission (Chapter 4). Analytical and numerical results prove that the proposed

metric allows to greatly reduce both hardware and computational complexity of

massive MIMO, especially when compared to large-scale systems where existing

AS schemes from the literature are employed.

• Introduction of a novel transmission scheme where AS and precoding are jointly

and simultaneously performed according to interference exploiting metrics (Chap-

4
 Chapter 1. Introduction

ter 5). The optimization problem is solved both optimally, via mixed integer pro-

gramming solver, and heuristically, via convex optimization tools. Results show

a significant increase of the performances when compared to previous interference

exploiting-based AS schemes and state-of-the-art AS schemes from the literature.

• Design of several beam selection techniques that aim to maximize the received

Signal-to-Interference-plus-Noise Ratio (SINR) and system capacity in beamspace

MIMO for millimeter wave communications (Chapter 6). The results show that

the proposed schemes are able to outperform existing techniques both in terms of

capacity and energy efficiency in more realistic scenarios, where multi-path effects

are considered and the number of available RF chains is fixed.

1.3 Thesis Organization

Subsequent to this introductory chapter, this thesis is organized following the struc-

ture illustrated in Fig. 1.1 and described in the sequel.

Chapter 2 provides a thorough review of multiple antenna communications systems,

in their small-scale and large-scale implementation. The theoretical principles behind

MIMO systems are discussed, with special focus on multiuser signal processing. Numer-

ous state-of-the-art low-complexity and energy-efficient schemes for large-scale MIMO

are examined, which include millimeter wave communications and AS schemes.

Chapter 3 introduces a precoding design for an energy efficient Constant Envelope

Precoding (CEP) scheme for multiuser scenarios. Downlink precoding design is per-

formed in order to exploit the multiuser interference at the receiver side, hence increasing

the received power. Two different CEP schemes are presented: a first technique, based

on the application of the cross-entropy solver, and a two-step approach, based on an

initial relaxation of the power constraints and a subsequent enforcement of per-antenna

power constraints.

Chapter 4 presents a low complexity AS scheme based on constructive interference.

We show that the proposed AS algorithm, combined with simple matched filter pre-

5
Chapter 1. Introduction

Energy Efficient Large-Scale MIMO Systems

Hardware
Hardware-complexity Reduction
Efficiency

Antenna Selection Constant Envelope

Beam Selection
Algorithms for Precoding for
Algorithms for
Constructive Constructive
Discrete Lens Array
Interference Interference
Systems
Exploitation Exploitation

Joint Antenna Selection and Precoding for

Constructive Interference Exploitation

Figure 1.1: Thesis Structure.

coding, outperforms more complex and computationally expensive AS techniques that

involve zero forcing linear precoding. The computational burdens of the proposed tech-

nique are discussed and its benefits are analyzed by means of an energy efficiency metric

that combines throughput performance with the circuit power required.

Chapter 5 discusses several low-complexity transmission schemes for massive MIMO

based on jointly performing transmit AS and convex optimization precoding. The pre-

sented schemes are analyzed in terms of running time and compared with state-of-the

art algorithms. Their performances are evaluated in terms of symbol error rate, capacity

and energy efficiency. Finally, the effects of imperfect channel-state information at the

transmitter are studied and a methodology for designing robust joint transmit AS and

precoding is derived.

Chapter 6 presents a new millimeter-wave (mm-wave) transmission scheme that

combines Beamspace Multiple-Input Multiple-Output (B-MIMO) communications and

beam selection techniques to achieve near-optimal performances with a low hardware-

6
 Chapter 1. Introduction

complexity transceiver. The benefits of the proposed scheme are evaluated via perfor-

mance computations, comparisons of numbers of required RF chains and a joint sum-rate

and complexity metric that shows the trade-off between spectral and energy efficiency.

Chapter 7 concludes the thesis with a summary of contents and the conclusions

derived throughout. Moreover, future research lines within the framework of this thesis

are also presented.

1.4 Publications

The aforementioned contributions have been presented in the following list of pub-

lications:

Journals.

[J1] P.V. Amadori and C. Masouros, “Interference Exploiting Antenna Selection for

Hardware-Efficient Massive MU-MIMO,” submitted to IEEE Transactions on

Communications.

[J2] P.V. Amadori and C. Masouros, “Constant Envelope Precoding by Interference

Exploitation in Phase Shift Keying-Modulated Multiuser Transmission,” Accepted

for publication in IEEE Transactions on Wireless Communications.

[J3] P.V. Amadori and C. Masouros, “Interference-Driven Antenna Selection for Mas-

sive Multiuser MIMO,” IEEE Transactions on Vehicular Technology , vol. 65, no.

8, pp. 5944 - 5958, August 2016.

[J4] P.V. Amadori and C. Masouros, “Low RF-Complexity Millimeter-Wave

Beamspace-MIMO Systems by Beam Selection,” IEEE Transactions on Commu-

nications, vol. 63, no. 6, pp. 2212 - 2223, June 2015.

Conferences.

[C1] P.V. Amadori and C. Masouros, “A Mixed-Integer Programming Approach to In-

terference Exploitation for Massive-MIMO,” 2017 IEEE International Conference

7
Chapter 1. Introduction

on Communications (ICC), Submitted.

[C2] P.V. Amadori and C. Masouros, “Constructive interference based Constant En-

velope Precoding,” 2016 IEEE 17th International Workshop on Signal Processing

Advances in Wireless Communications (SPAWC), pp. 1 - 5, 2016.

[C3] P.V. Amadori and C. Masouros, “Power efficient massive MU-MIMO via an-

tenna selection for constructive interference optimization,” 2015 IEEE Interna-

tional Conference on Communications (ICC), pp. 1607 - 1612, 2015.

[C4] P.V. Amadori and C. Masouros, “Beam selection techniques in mm-wave commu-

nications,” IET Colloquium on Antennas, Wireless and Electromagnetics, pp. 1 -

20, 2015.

[C5] P.V. Amadori and C. Masouros, “Low complexity transceivers in multiuser

millimeter-wave beamspace-MIMO systems,” 2014 IEEE 25th Annual Inter-

national Symposium on Personal, Indoor, and Mobile Radio Communication

(PIMRC), pp. 118 - 122, 2014.

8
Chapter 2

Energy Efficiency in Multiple

Antenna Wireless
Communications Systems

This chapter introduces the concepts that inspired the research synthesized in this

thesis. These preliminary sections are used to briefly address the key benefits of multiple

antenna communications and their main drawbacks. In addition, a detailed presentation

of large-scale systems is provided, with special focus on the challenges they present.

2.1 MIMO Communications

Generally, the term MIMO communications is used to identify transceiver systems

that involve multiple antennas at the transmitter and the receiver. Because of this,

MIMO can be seen as a direct evolution of antenna array communications, with diversity

being simultaneously exploited both at transmitter and receiver. The interest for MIMO

applications has strengthened over the past years, thanks to the high capacity, increased

diversity and interference suppression they are able to provide [25].

Wireless channels of point-to-point MIMO systems, see Fig. 2.1, with nt antennas at

the transmitter and nr antennas at the receiver, are generally represented by a Cnr ×nt

9
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

Wireless Channel
x H y
RF Chain RF Chain

RF Chain RF Chain

u RF Chain RF Chain ⌃
u
Pre-Processing Post-Processing
RF Chain RF Chain

RF Chain RF Chain

RF Chain RF Chain

Figure 2.1: Block diagram of a point-to-point 6 × 6 MIMO.

matrix H, where each element hi,j identifies the transfer function between the j-th

transmitter and the i-th antenna of the receiver. This notation leads to a Cnr ×1 received

signal vector y that can be defined analytically as

y = Hx + n, (2.1)

where x is the Cnt ×1 transmitted signal vector and n the Cnr ×1 additive white Gaussian

noise (AWGN), i.e., n ∼ CN (0, N0 ) with N0 being the noise variance.

If transmitter and receiver possess perfect knowledge of the channel, we can de-

rive the capacity of such a system by decomposing the channel into a set of parallel,

non-interfering, Single-Input Single-Output (SISO) channels via singular value decom-

position (SVD) [26]. Accordingly, we can compute the system capacity as

nX
min
P ∗ λ2
 
C= log 1 + i i , (2.2)
N0
i=1

 
N0
where Pi∗ = max 0, µ − λ2i
is the waterfilling power allocation for the i-th eigenmode

of the channel and µ is the waterfill level, chosen to respect the total power constraint

[10].

10
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

2.2 MIMO Precoding

The use of multiple antennas at both sides of a communication link leads to increased

signal processing burdens, either at the transmitter or the receiver. The techniques used

to separate data streams can be differentiated between detection and precoding, whether

they are applied at the receiver or at the transmitter side, respectively. In multiuser

MIMO (MU-MIMO), precoding techniques are generally preferred for downlink commu-

nications and detection techniques for uplink communications. Thanks to this choice,

the burdens of signal processing are always sustained by the Base Station (BS) instead

of the mobile users, which are generally characterized by higher power and complexity

constraints.

y1
Terminal
x
RF Chain
y2
Terminal
RF Chain
y3
u RF Chain H Terminal
Precoder y4
RF Chain
Terminal
RF Chain
y5
Terminal
RF Chain yM
Terminal

Figure 2.2: Block diagram of a 6 × 6 MU-MIMO.

In this section, we focus on a brief introduction to MU-MIMO downlink precoding,

see Fig. 2.2, where the BS employs an array with N elements and communicates with

M single-antenna users or terminals. In this scenario, the CM ×N channel matrix H can

be seen as a set of multi-input single-output (MISO) channels, where each row hTm of H

represents the C1×N channel vector for the m-th user, as

H = [h1 , ..., hM ]T , (2.3)

hence leading to the following definition of the m-th user received symbol

N
X
ym = hTm x + nm = hn,m xn + nm . (2.4)
n=1

11
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

Precoding techniques that help overcoming the interference between users have at-

tracted the attention of the scientific community over the past years with several linear

and non-linear approaches, as shown in Fig. 2.3. The distinction between linear and

non-linear precoding is based on the operations the CM ×1 desired symbols vector u expe-

riences before transmission. In general, linear precoding approaches from the literature

are characterized by low-complexity computational burdens and aim to either cancel

[25,27,28] or exploit [29,30] the interference at the transmitter side, while non-linear pre-

coding techniques require more complex signal processing but are able to achieve higher

rates [29,31–33]. Finally, we include a brief description of non-linear optimization-based

downlink beamforming, based on SINR optimization [34, 35]. While such approaches

are characterized by high and near optimal performances, they are affected by higher

computational burdens both at the transmitter and at the receiver side, where channel

equalization is necessary.

Linear Precoding
• Matched Filtering (MF) – Simple but low-performance.
• Zero Forcing (ZF) – More computationally demanding because of matrix inversion.
• Regularized Zero-Forcing (RZF) – Comparable complexity to ZF, higher
performances.

Non-Linear Precoding
• Dirty Paper Coding (DPC) – Optimal, but computationally prohibitive.
• Vector Perturbation (VP) – High performances at the cost of higher complexity
than linear precoding.
• Tomlinson-Harashima Precoding (THP) – High performances at the cost of high
complexity at transmitter and receiver.

Optimization-Based Beamforming
• Transmit Power Minimization (TPM) – High performances, can be solved through
Convex Optimization.
• SINR Balancing (SB) – High performances, can be solved through algorithmic
approach.

Figure 2.3: Downlink Precoding and Beamforming classification and representa-

tive techniques.

2.2.1 Linear Precoding

Linear precoding is a simple transmission approach where the transmitted signal x

is derived as a linear combination of the data symbol u. The operation used to derive

12
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

the transmitted signal vector x can be described analytically as follows

x = Gu = γFu, (2.5)

where G = γF is the CN ×M precoding matrix and is defined as the combination of two

different elements: the scaling factor γ ∈ R, necessary to ensure that the transmitted

signal x respects power constraints, i.e. E ||x||2 = P , and the precoding matrix


without scaling F. Here, we list the fundamental linear precoding techniques for MIMO

communications:

• Matched filtering (MF) represents the simplest linear precoding technique from

the literature and is designed to maximize the received signal-to-noise ratio (SNR).

Given the scaling factor γM F , the MF precoding matrix FM F can be computed as

the Hermitian of the channel matrix H [27]

HH
GM F = γM F HH = p . (2.6)
tr[(HH H)]

• Zero Forcing (ZF) is a simple linear precoding technique that has been exten-

sively studied in the past[28, 29] and is designed in order to equalize the effects of

the channel at the receiver side. Given the scaling factor γZF , the ZF precoding

matrix FZF is defined as the Moore-Penrose pseudoinverse of the channel H [36]

HH (HHH )−1
GZF = γZF HH (HHH )−1 = q . (2.7)
H −1
tr[(HH ) ]

• Regularized Zero Forcing (RZF) identifies a modification of conventional ZF,

with the aim to reduce its high susceptibility to ill-conditioned channel matrices

[25]. Performance losses are identified by a decreasing SNR at the receiver and

are caused by the increased scaling factor γZF experienced by channel matrices

with smaller condition number. The RZF precoding matrix aims to maximize the

13
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

SINR at the receiver [30] and is defined as

GRZF = γRCI HH (HHH + αI)−1

HH (HHH +αI)−1
(2.8)
= q ,
tr[(HHH +αI)−1 HHH (HHH +αI)−1 ]

where α = N · N0 is the optimal regularization factor for a single-cell scenario [29].

2.2.2 Non-linear Precoding

Non-linear precoding embraces all the signal processing techniques devoted to the

generation of the precoded vector x by means of complex non-linear operations over

the desired symbol vector u. While the use of non-linear precoding techniques at the

transmitter offers significant sum-rates benefits, it generally comes at the expense of

very sophisticated signal processing. Due to their innate high complexity, non-linear

precoding techniques are only briefly presented in this thesis, as main focus resides on

systems which require low computational burdens.

• Dirty Paper Coding (DPC) methods are based on the concepts introduced by

the seminal work in [31], which showed that DPC is able to achieve the theoretical

channel capacity if the transmitter is aware of the interference. In addition, such

remarkable performances can be achieved without the need of additional power in

transmission nor of shared channel-state-information (CSI) with the receiver [25].

Despite offering significant benefits, DPC methods suffer from low practicability in

realistic scenarios because they require complex signal processing, such as sphere-

search algorithms and infinite length codewords.

• Vector Perturbation (VP) performs a perturbation of the user data before trans-

mission in order to reduce the scaling factor of ZF precoding. While this allows to

greatly enhance the performances of linear precoding, as for RZF, its application

is affected by a significant increase in the signal processing burdens, as it requires

to solve an integer-lattice least squares problem [29, 32].

• Tomlinson Harashima Precoding (THP) is a technique that aims to equalize

14
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

the intersymbol interference. In particular, the approach to interference followed

by THP can be seen as translation of V-BLAST [33] to the transmitter side,

where the channel matrix H is decomposed into the multiplication of a unitary

matrix feedforward matrix F and a lower triangular matrix L. The aim of the

technique is the reduction of the channel matrix at the receiver side to a simple

lower triangular matrix B = DHF with unitary diagonal elements, where the

diagonal scaling matrix D entries are the inverse of the diagonal elements of the

matrix L. The transmitted signal is computed as

x = Fe
x, (2.9)

where the i-th element of x

e is obtained through a recursive approach at the receiver

over the i-th desired symbol ui

 
i−1
X
ei = ui −
x bi,j x
ej  modL , (2.10)
j=1

where bi,j is the i-th entry of the j-th column of B and [·]modL is the L-base

modulo operation that keeps the symbol x

ei within its original Voronoi region [37],

with L being a real number that depends on the modulation used. In opposition

to the previous techniques, THP requires further signal processing at the receiver

side. In fact, the i-th receiver needs to both normalize the i-th received symbol

by the i-th diagonal entry of the matrix D and to apply an additional modulo

operation in order to estimate the received symbol.

2.2.3 Optimization-based Beamforming

In addition to linear and non-linear precoding, we list the state-of-the-art on non-

linear beamforming based on the optimization of the received SINR : transmitted power

minimization [34] and SINR balancing [35]. When using beamforming, the transmitted

signal can be defined as follows

M
X
x= pm um , (2.11)
m=1

15
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

where pm represents the CN ×1 weight or beamforming vector for the m-th user and um

identifies the data or constellation symbol for the m-th user. Given the definition of the

beamforming vectors, it is possible to evaluate the received SINR for the m-th user as

N 2
P
2 hn,m pn,m
hTm pm n=1
γm = P = (2.12)
|hTm pj |2 + N0 N
P P
2
j6=m hn,m pn,j + N0
j6=m n=1 ,

where pn,m represents the n-th element of the m-th user beamforming vector pm .

• Transmitted power minimization (TPM) [34] represents a conventional ap-

proach to downlink beamforming, where interference is regarded as a harmful

element for transmission. Because of this, the beamforming optimization problem

is designed to minimize the transmitted power while respecting predefined SINR

requirements Γm , ∀m ∈ {1, ..., M }. Analytically

M
kpm k2
P
PT P M : minimize
pm m=1
2
N
(2.13)
P
hn,m pn,m
n=1
subject to γm = 2 ≥ Γm ,
P N
P
hn,m pn,j +N0
j6=m n=1

which can be efficiently solved by means of convex optimization with a semidefinite

relaxation (SDR) approach [34].

• SINR balancing (SB), also known as Max-min beamforming, aims to identify

the beamforming vectors that maximize the minimum received SINR, while re-

specting a predefined transmitted power constraint Pt . The optimization problem

is analytically defined as

PSB : maximize Γm
pm

subject to γm ≥ Γm (2.14)
M
kpm k2 ≤ Pt .
P
m=1

Differently from PT P M , the SINR balancing problem is non-convex and its solution

16
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

requires a more complex algorithmic approach [35].

2.3 AS Techniques

In the previous section, we described how MIMO communications provide benefits in

terms of capacity and reliability for radio communication. However, the use of multiple

antennas at transmitter and receiver comes with the burden of increased costs in terms of

size, power and hardware, caused by the need of higher numbers of radio frequency (RF)

chains [15], which comprise digital-analog converter, mixer and filter. In this scenario,

AS is a possible solution for exploiting the diversity gains offered by MIMO systems,

while achieving complexity reductions. Thanks to AS, multiple antenna benefits are

partially preserved with low additional costs, mostly determined by the need of RF

switches.

~
x x

y1
Terminal

RF Chain
u H
Precoder RF Chain ym
RF Chain Terminal

yM
Terminal

Figure 2.4: Transmitter AS for MU-MIMO

The concept behind AS, at the transmitter or at the receiver, is very simple: the sys-

tem chooses, according to a specific performance metric, the best (MRF , NRF ) antennas

out of the (M, N ) available antennas at the receiver or transmitter, respectively. Thanks

to this selection, receiver/transmitter experiences a significant reduction in complexity,

since the number of needed RF chains lowers down to (MRF , NRF ).

Over the past years, AS has been intensively studied for both the transmitter, Fig.

2.4 for a MU-MIMO scenario, and the receiver side, showing different benefits. While

subset AS at the receiver side offers interesting complexity reduction, transmit selection

additionally proved to increase the capacity in comparison with an open loop MIMO

17
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

system [16].

The equation of a system that involves Transmit AS (TAS) follows the definition of

a classical MU-MIMO system, as in (2.1). In particular, we have

y = Hx̃ + n, (2.15)

where x̃ identifies the precoded vector after TAS has been performed, whose elements

are null when their index corresponds to one of the deactivated antennas, i.e., x̃n =

0, ∀n ∈
/ N with N being the subset of transmitting antennas with cardinality equal to

the number of available RF chains card (N ) = NRF .

In this section, we focus on the key TAS techniques in MU-MIMO, which have been

applied according to different criteria, such as the system capacity [38], the channel

matrix eigenvalues [39], the antenna path gain [40] and the error rate [41]. These

techniques can be separated according to two different approaches or metrics: diversity

or norm-based selection and multiplexing or capacity-based selection, as in Fig. 2.5.

Norm-Based Selection
• Path Gain Maximization (PGM)– Simple, but low-performance.
• Error Rate Minimization (ERM) – High performances, nearly prohibitive for large
systems.

Capacity-Based Selection
• Capacity Maximization (CM)– Sub-optimal, but computationally demanding, i.e.,
prohibitive with an exhaustive search approach.
• Minimum Eigenvalue Maximization (MEM) – High performances at the cost of higher
complexity, prohibitive in its exhaustive search approach

Figure 2.5: Conventional TAS classification

2.3.1 Norm-based Selection

We identify as norm-based selection, the TAS techniques that aim to capture the

diversity and improve the SNR at the receiver. More specifically, we indicate as diversity

the presence of multiple signal paths that fade independently.

18
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

• Path Gain Maximization (PGM). Path gain selection at the transmitter can be

easily performed by selecting the subset of antennas with the highest path gains.

The antenna subset can be analytically identified as follows

n o
N = arg max kh1 k2 , ..., khn k2 , ..., khN k2 , (2.16)
NRF

where the notation hn is used to identify the channel response corresponding to the

n-th antenna of the BS, i.e., the n-th column of the matrix H, and max identifies
NRF
the NRF highest values of the argument.

While path gain selection represents a very appealing technique for its reduced

complexity, performances are normally very poor when compared to more sophis-

ticated approaches from the literature. However, early works on TAS in MISO or

Single-Input Multi-Output (SIMO) systems [40,42,43] showed that PGM performs

optimally when the receiver applies maximum-ratio combining.

• Error Rate Minimization (ERM). In [41] the authors present a TAS technique

that aims to minimize the error rate at the receiver, under the assumption of

a maximum likelihood detection-based receiver. While this approach is able to

achieve impressive performances in terms of symbol error rate (SER), it rapidly

becomes computationally prohibitive for large constellations, due to the exhaus-

tive search over all the possible transmitted symbols and over all the possible

combinations of transmitting antennas.

2.3.2 Capacity-based Selection

In addition to norm-based selection, TAS has been studied also as a technique to

maximize the capacity achievable by lower dimensional systems. In fact, TAS over

capacity proved to be even able to increase the capacity for systems with low-rank

channels [44]. Capacity maximization techniques are based on the formulation of ergodic

capacity for a system with DPC and equal power transmission between the antennas

[15]

C = log2 det IM + SN R · HHH ,



(2.17)

19
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

where IM is a M -dimensional identity matrix. Accordingly, we can formulate the cor-

responding optimization problem as follows:

log2 det IM + SN R · H∆HH



PCM : maximize
∆

subject to ∆n,n ∈ {0, 1} , (2.18)

PN
n=1 ∆n,n = NRF ,

where ∆ is a N × N is a binary-valued and diagonal selection matrix, whose entries can

be either null, i.e., ∆n,n = 0 if n is a non-selected antenna, or unit, i.e., ∆n,n = 1 if n is

an antenna selected for transmission.

First approaches in capacity maximization were based on the identification of the

best antenna subset by means of an exhaustive search. While performance of such

approach proved to be optimal, its application is affected by nearly prohibitive compu-

tational costs as the number of antennas at the transmitter increases. As a consequence,

researchers focused on sub-optimal approaches that allowed to perform a selection over

the capacity, while maintaining acceptable computational burdens.

• Sub-optimal Capacity Maximization (CM). This approach aims to optimize

the capacity of a system with TAS by identifying the transmitting antenna subset

in a recursive manner. The key concept behind this approach is that the deacti-

vation of each transmitting antennas in a MU-MIMO system in a Rayleigh fading

scenario can be related to a quantitative loss in the maximum achievable capacity.

The loss δn caused by the n-th antenna deactivation is analytically described [38]

as

−1
δn = hH
n IM + SN R · HH
H
hn . (2.19)

The parameter δn is computed for each of the available antennas, allowing to

identify the index d of the antenna that contributes the least to the maxi-

mum achievable capacity. The algorithm is recursively applied in order to iden-

tify the N − NRF antennas to deactivate, which are collected in the subset
n o
D = d1 , ..., dl , ..., dN −NRF , where dl is the index of the antenna deactivated

at the l-th step of the algorithm.

20
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

• Minimum Eigenvalue Maximization (MEM). Given the key role played by

channel singular values in MIMO capacity performances, their optimization rep-

resented one of the main aims of TAS. In [39], the authors present a selection

technique based on an exhaustive search of the antenna subset that leads to the

highest minimum eigenvalue of the low dimensional channel matrix. While such

approach proved to be optimal, it rapidly becomes computationally prohibitive as

the sizes of the system increase.

2.3.3 Energy Efficiency

The trade-off between system performance and RF complexity introduced by TAS

techniques can be analyzed via a parameter called energy efficiency, which is computed

according to different metrics, such as capacity and throughput. More specifically:

• Energy efficiency over capacity [45] can be analytically defined as follows

C C
ηC = = , (2.20)
PBS Pt + NRF · PRF

where C[bits/s/Hz] represents the capacity, PBS [W ] is the power consumed at

the BS, according to the modelling from the literature [46], Pt [W ] identifies the

transmitted power of the system and PRF [W ] represents the power consumed by

each RF chain, characterized by digital-analog converter, mixer and filter.

• Energy efficiency over throughput [45] is defined as follows

T T
ηT = = , (2.21)
PBS Pamp + NRF · PRF + Nops · Pf pga

where T = (1 − BLER) · l · M is the throughput, BLER is the block error

rate, l identifies the bits per symbol, i.e., l = 1 for 2-PSK and l = 2 for 4-

PSK, M is the number of users, Pamp [W ] is the power required by the amplifier,

PRF [W ] is the power consumed by a single RF chain, Nops [KF LOP s] identifies

the complexity burden of the analyzed technique in terms of 103 floating point

operations, and Pf pga [W/KF LOP s] is the power consumption per operation of

21
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

the field-programmable gate array (FPGA). When the information regarding the

floating point operation count is not available, it is possible to evaluate ηT with a

simpler model of PBS , where FPGA power consumption is not accounted for, as

follows
T
ηT = . (2.22)
Pamp + NRF · PRF

While the shown metrics have been presented in order to highlight the performance-

complexity trade-offs which characterize AS systems, they can be effectively employed to

evaluate more generalized MIMO scenarios. For example, γC and γT could be evaluated

for two separate systems as a function of the number of radiating elements NRF required

in order to achieve a fixed performance threshold, i.e., a minimum value of achievable

capacity C or throughput T , respectively.

2.4 Large Scale MIMO

In recent years, the growing interest for MU-MIMO systems has brought to the

development of the concept of large-scale MIMO (LS-MIMO) or Massive MIMO (M-

MIMO). The pioneering work of [11] proposed the introduction of schemes that tackle

the non-scalability of classical MU-MIMO by employing a large excess of antennas at

the BS for a limited number of users, see Fig. 2.6. These schemes, namely identified as

LS-MIMO or M-MIMO, rapidly experienced an increasing interest by the scientific com-

munity because of the exciting benefits they promise to provide and the new challenges

they uncover [12].

Figure 2.6: Multiuser Massive MIMO representation

22
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

In fact, the channel matrix H experiences significant changes in its statistical prop-

erties when the number of transmitting elements grows asymptotically to infinity, e.g.,

the singular values of the channel matrix approach a deterministic function [47]. While

the benefits introduced by M-MIMO are not limited to the multiuser scenario, the char-

acteristics of point-to-point M-MIMO are not addressed in this thesis for the sake of

brevity. For the interested reader, the characteristics of M-MIMO for point-to-point

communications are fully described in [48].

The following sections briefly describe the key elements that characterize M-MIMO,

with a particular focus on the benefits and the challenges brought by this technique.

More specifically, we consider a time-division duplexing (TDD) system, as it is common

in M-MIMO, because of the simplifications it brings with regards to the CSI acquisi-

tion. In fact TDD systems are characterized by channel reciprocity, where both down-

link/uplink transmission links exactly match, leading to the possibility to achieve CSI

through the uplink. This is a particularly useful property for M-MIMO, since the time

required to transmit pilots in the uplink does not depend on the number of antennas at

the transmitter, while the time required for downlink pilots depends on the number of

antennas at the BS [48].

2.4.1 Multiuser M-MIMO Channel Model

In a single cell scenario, when the BS is equipped with N antennas and serves M

single-antenna users, the path gain between the n-th antenna and the m-th user can be

defined as
p
hn,m = tn,m βm , (2.23)

where tn,m represents the complex fast-small scale fading and βm represents the real

slow-large scale fading coefficients experienced by the m-th user.

Accordingly, the uplink channel matrix of a multiuser M-MIMO system can be repre-

sented as the combination of two matrices: a CN ×M matrix T whose entries are different

for each user and each antenna and a RM ×M diagonal matrix Dβ whose entries depend

23
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

only on the user. Analytically we have

1/2
Hu = TDβ , (2.24)

where Hu represents the channel matrix for the uplink scenario. Accordingly, the CN ×1

received signal vector at the BS yu for an uplink scenario can be defined as

√
yu = ρu Hu xu + nu , (2.25)

where xu is the CM ×1 signal vector from the users to the BS, nu is the CN ×1 additive

white Gaussian noise and ρu is the uplink transmit power. Thanks to the assumption

of TDD operations, the channel model described for the uplink can be easily translated

to the downlink scenario. In particular, the received signal vector yd in a downlink

transmission can be described analytically as

√ √
yd = ρd Hd xd + nd = ρd HH
u xd + nd , (2.26)

where Hd = HH
u is the reciprocal channel matrix for downlink transmission, xd is the

CN ×1 signal vector transmitted by the BS, nd is the CM ×1 additive white Gaussian

noise and ρd is the downlink transmit power.

2.4.2 Capacity and Signal Processing Benefits

The use of Large-Scale Arrays (LSA) at the transmitter leads to significant benefits

in terms of achievable capacity and signal processing for both uplink and downlink. In

fact, when the number of antennas at the BS N tends to infinity, the channel responses

for different users become orthogonal, if the elements tn,m of T are independent [11].

This directly affects the correlation matrix as follows

1/2 1/2 1/2 1/2

Ru = HH H
u Hu = Dβ T TDβ ≈ N D β I M Dβ = N Dβ , (2.27)

24
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

hence simplifying the computation of the capacity for the uplink to

M
X
Cu = log2 det IM + ρu HH


u H u ≈ log 2 det (IM + N ρ D
u β ) = log2 (1 + N ρu βm ) .
m=1
(2.28)

Accordingly, the uplink transmission under M-MIMO experiences both a dramatic in-

crease in the achievable rate and additional simplifications in the signal processing, as

the MF detector becomes asymptotically optimal and is sufficient to obtain the rates

shown in (2.28). Given the complex data symbols vector in uplink uu , the estimated

received symbol vector of a BS with MF detection is

H √ √
ŝ = HH H
u yu = Hu ( ρu Hu uu + nu ) ≈ N ρu Dβ uu + Hu nu . (2.29)

Clearly, MF detection in M-MIMO greatly benefits from the asymptotic orthogonality of

the channel vectors, as it becomes asymptotically optimal. In fact, multiuser interference

is nullified and signals from different users over different streams are efficiently and

perfectly separated, each transmission can be considered as a SISO channel with SN R =

N ρ2 βm and, finally, noise whiteness is preserved.

In a downlink transmission, the sum capacity of a M-MIMO system with power

allocations is defined as

Cd = max log2 det IN + ρd HH

d PHd ≈ max log2 det (IN + ρd N PDβ ) , (2.30)
P P

where P is a diagonal matrix whose real entries (p1 , ..., pM ) represent the power allo-
PM
cations for each user, bound to the power constraint pm = Pt . As shown for the
m=1
uplink transmission, a simple MF precoder can be proven to be asymptotically optimal.

In fact, given the complex data symbols vector in downlink ud , the received signal vector

at the user side is defined as

√
yd = ρd Hd HH
d P
1/2
ud + nd ≈ ρd N Dβ P1/2 ud + nd , (2.31)

which, following a similar analysis to the one for (2.29), proves the optimality of the MF

25
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

precoder in a downlink transmission with N → ∞ when power allocation is optimized.

2.4.3 Challenges

While M-MIMO is able to achieve terrific benefits, the employment of LSAs unveils

a series of challenges, which are addressed in this subsection. In particular, main focus

will reside on the channel estimation challenges, briefly introduced at the beginning of

the section, followed by a description of the phenomenon called pilot contamination and,

finally, a concise presentation of the practical limitations of M-MIMO.

2.4.3.1 Channel State Information Acquisition

In the previous section, TDD operations were regarded almost as a requirement

for realistic implementations of M-MIMO, because of the benefits channel reciprocity

introduces in terms of CSI acquisition complexity. In fact, under frequency-division

duplexing (FDD), i.e., where downlink and uplink transmissions operate at different

frequencies, channel reciprocity does not apply and downlink/uplink are characterized

by two separate channel matrices. Because of this, channel training in FDD requires

two separate steps: one for the uplink channel and one for the downlink channel. Under

these conditions, downlink training might require the whole coherence time for channel

estimation, leaving no time for data symbol transmission. The challenges and feasibility

of FDD in M-MIMO are presented and discussed in [11], which shows the significant

constraints frequency-division introduces over the number of antennas at the BS N .

However, recent works [49] are trying to justify the use of FDD in M-MIMO, depict-

ing its feasibility when specific conditions apply, such as the knowledge of the channel

covariance matrix.

Tdata

TCSI
Tsp
Tul
Tdl

Tcohe

Figure 2.7: TDD protocol

26
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

While the application of M-MIMO technology to FDD transmission represents a

very interesting approach, the CSI acquisition complexity can be readily tackled via

TDD operations [50]. With this regard, [51] introduced a simple TDD protocol where

the coherence time Tcohe is divided between the time for CSI acquisition TCSI , the time

for data transmission Tdata and the time for BS processing Tsp . As shown in Fig. 2.7,

the time for data is further divided between the time for downlink transmission Tdl and

uplink transmission Tul .

2.4.3.2 Pilot Contamination

As shown, TDD transmissions in a single-cell scenario allow to exploit channel reci-

procity to estimate the channel response in uplink and use the information during down-

link transmission. However, more realistic multi-cell M-MIMO scenarios, i.e., where

terminals are distributed among different cells as in Fig. 2.8, are known to be affected

by a phenomenon called pilot contamination.

In an ideal multi-cell scenario, the m-th terminal can transmit a pilot sequence of
n o
1 , ..., ψ τ
length τ to the l-th cell BS, i.e., ψm,l = ψm,l m,l , without interfering with other

cells and BSs. Here, non-interference between terminals is achieved by transmitting

pilot sequences that are orthogonal within the same cell and between neighboring cells,

as
H
ψm,l ψn,p = δ [m − n] δ [l − p] , (2.32)

where δ [·] represents the Dirac delta, i.e., δ [x] is unitary if x = 0 and null if x 6= 0.

Pilot  Contamina+on  
 
Correct  Pilot  

Figure 2.8: Visual representation of Pilot Contamination

27
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

However, the orthogonality for all the sequences of all terminals on different cells

requires to strongly limit the number of users that can be served by a multi-cell multiuser

M-MIMO system [11]. Accordingly, terminals are expected to use non-orthogonal pilot
H ψ
sequences, i.e., ψm,l n,p 6= 0, which causes BSs to possess a CSI that is affected by

imperfections, often referred to as pilot contamination.

Recent works are trying to mitigate the pilot contamination phenomenon through

different approaches, such as protocol-based techniques [11, 52–54], precoding methods

[55–57] and blind transmissions [58].

2.4.3.3 Practical Limitations

Because of its key elements, M-MIMO inherits and amplifies most of the practical

limitations that affect MU-MIMO systems. In particular, the use of hundreds of anten-

nas at the transmitter comes with the necessity to use hundreds of RF chains, which

corresponds to hundreds of power amplifiers, digital-to-analog converters and filters. It

was already pointed out in previous sections how these components affect the hardware,

power and complexity costs of a classic MU-MIMO system. The need for cheaper and

low power hardware becomes a key for M-MIMO, since the dimensions of the system

increase drastically [12]. As a consequence, recent works approached the implementation

of M-MIMO with more efficient hardware, at the cost of increased distortions [59, 60],

or increased signal processing [17, 18].

In addition to this, while the transmission through a large number of antennas

allows to reduce the radiated power, the deployment of such a large number of elements

will have to face new challenges. In fact, even though arrays in M-MIMO can have

many different configurations and geometries, such as cylindrical or uniform-linear arrays

(ULAs), they are expected to be characterized by small active units [12] in order to

respect more stringent cost and space constraints. Toward this end, recent works [36,

61] have investigated the possibility of exploiting transmit mutual coupling at the BS,

allowing the dimensions of antenna arrays in fixed physical spaces to further increase.

Finally, the computational costs and complexity of precoding increase together with

28
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

the number of antennas at the BS. As a consequence, the use of low complexity precoding

becomes critical in M-MIMO, since the precoding at the BS reduces the time dedicated

to data transmission, as shown in Fig. 2.7. This, together with the increasing circuit

power consumption experienced by baseband signal processing, leads to the need for

highly parallel and efficient signal processing hardware [12].

2.5 Millimeter Wave MIMO

Most of the mobile communications use the Ultra High Frequency (UHF) spectrum,

i.e., in the range between 300 M Hz − 3 GHz, see Table 2.1. Since the number of mobile

devices is experiencing an unprecedented growth, wireless service providers need to

face the forthcoming bandwidth shortage [62]. Accordingly, recent studies [62] envisage

millimeter-wave (mm-wave) spectrum, i.e. 30 − 300 GHz, as a promising approach,

because of the high bandwidth it provides and because of its license free nature.

Nevertheless, the use of higher frequencies in wireless communications comes with

the burden of less favorable propagation characteristics [63], which represent one of the

main challenges of mm-wave systems. Given the peculiar propagation characteristics

of mm-wave communications, the following sections introduce both channel models and

transceiver architectures used in the literature to study the performances of systems

that exploit the 30 − 300 GHz bandwidth.

Band Uplink Downlink

700 MHz 746-763 MHz 776-793 MHz
AWS 1710-1755 MHz 2110-2155 MHz
IMT Extension 2500-2570 MHz 2620-2690 MHz
GSM 900 880-915 MHz 925-960 MHz
UMTS Core 1920-1980 MHz 2110-2170 MHz
GSM 1800 1710-1785 MHz 1805-1880 MHz
PCS 1900 1850-1910 MHz 1930-1990 MHz
Cellular 850 824-859 MHz 869-894 MHz

Table 2.1: Spectrum usage - Spectrum usage in modern mobile communica-

tions. [64]

29
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

2.5.1 Propagation Characteristics

One of the main critiques moved toward the use of mm-wave resides in the high

values of free space loss experienced during propagation. In fact, according to Friis

equation, we have
 2
Pr f
= Gt G r , (2.33)
Pt 4πRc

where Pt and Pr are transmitted and received power, respectively, Gt and Gr are the

antenna gains at transmitter and receiver, f is the signal frequency, c is the speed of light

in the medium and R is the distance between transmitter and receiver, described visually
 2
f
in Fig. 2.9. The free space loss in (2.33) is represented by the 4πRc term, describing

how attenuation experiences a quadratic growth with the frequency. Accordingly, from

(2.33) it is clear that if we compare two separate systems, operating at two different

frequencies with same transmitter/receiver gain, the system with a higher frequency

will achieve worse performances.

Nonetheless, this is not a fair comparison, as the use of shorter wavelengths λ in

mm-wave systems, leads to a tremendous increase in the number of antenna elements

within the same aperture. Consider a simple ULA of length L, the number of elements

N that can be deployed with critical spacing λ/2 can be easily evaluated as

2·L
N= , (2.34)
λ

hence leading to higher antenna gains as the frequency grows. Consequently, high fre-

quencies/short wavelengths do not necessarily experience a significant disadvantage in

terms of free space attenuation in comparison with longer wavelengths [65].

The study of mm-wave communications propagation characteristics has experienced

a significant increase in interest over the past years, supported by numerous measure-

ment campaigns [66] that aim to a better understanding of the channel for high frequency

systems. More specifically, recent works by [24, 67] proved the feasibility of mm-wave

system in outdoor urban environments.

30
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

Gt    :  Antenna  gain   Gr    :  Antenna  gain  

Pr  :  Transmitted  Power   Pr  :  Received  Power  

R  :  Distance  

Transmitter   Receiver  

Figure 2.9: Visual representation of Friis equation

2.5.2 Channel Model

When describing the channel model for a multiuser mm-wave system (BS equipped

with an N dimensional ULA), we generally refer to multiple steering-response vectors

of the array for each user. This represents the most prominent model for mm-wave

systems, as it allows to directly discriminate between multi-path and line-of-sight prop-

agation, both of which greatly characterize systems that operate at these frequencies

[24]. Analytically, for ULAs we have [10, 13]

h i
a(θ) = e−j2πθi , (2.35)
i∈I(N )

where a(θ) is the CN ×1 steering vector, θ = d

λ sin(φ) is the spatial frequency, whose

value ranges between −0.5 ≤ θ ≤ 0.5 for a critically spaced array, φ represents the

physical directions −π/2 ≤ φ ≤ π/2 and I(N ) = {i − (N − 1)/2 : i = 0, 1, ...N − 1} is a

symmetric set of indices centered around 0.

Hence the channel model for the m-th terminal can then be defined as

Np
(m) (m)
X
h(M
m
P)
= βi a(θi ), (2.36)
i=1

(m) (m)
where Np is the total number of paths, βi and θi are respectively the gain and the
(m)
direction of the i-th path of the m-th user. In particular, θi can be evaluated in terms

of spatial frequencies.

This model however does not take into account the line-of-sight (LOS) component

31
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

of the propagation, which strongly characterizes the channel in mm-wave frequencies

communications. In fact, small wavelengths λ allow to pack hundreds of antennas at

the BS, which lead to narrow and high gain beams with reduced angular spreads. Thanks

to this, in a mm-wave frequencies scenario we can consider the presence of LOS paths

for all the users.

The channel model for the LOS path can then be defined as

(m) (m)
h(LoS)
m = β0 a(θ0 ), (2.37)

(m) (m)
where θ0 represents the direction or position of the m-th user and β0 is the complex

gain for the LOS path.

We can define the channel for the single user as a sum of the two terms. Analytically

Np
(m) (m) (m) (m)
X
hm = h(LoS)
m + h(M
m
P)
= β0 a(θ0 ) + βi a(θi ), (2.38)
i=1

(M P )
where hm represents the multi-path (MP) component of the channel vector and the
Np
ratio between β02 and βi2 is called Rice factor.
P
i=1

2.5.3 Beamspace mm-wave MIMO

Thanks to the wider bandwidths they are able to provide [68], mm-wave communica-

tions have experienced a continuous increase in relevance for short-range, high-capacity

wireless link. However, a large-scale MIMO approach in mm-wave frequencies is still

prohibitive because of the number of antennas and the high transceiver complexity. In

order to tackle transceiver complexity, the combination of beamspace MIMO (B-MIMO)

[13], where data is multiplexed onto orthogonal beams, together with hybrid transceivers

[69], where analog beamformers in the RF domain are combined with a smaller num-

ber of digital beamformers in baseband, has been identified as a promising candidate

for future mm-wave MIMO applications [70]. In fact, systems based on such combina-

tion allow to achieve near-optimal performance with low-hardware complexity, which

operates on the dimension of the communication subspace.

32
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

The study on practical implementation of the combination of hybrid transceivers

and B-MIMO has mainly focused on two aspects: precoding optimization [71–76], where

analog processing is performed through phase shifters, and antenna design [14,23,77–79],

where analog processing is performed through lenses. More specifically:

• The use of hybrid analog/digital precoding for mm-wave B-MIMO transmissions is

motivated by the fact that it allows to preserve beamforming gain and diversity

order achieved by the large number of antennas, while the number of RF chains is

lower-limited by the number of data streams to be transmitted. Several approaches

have been introduced in both the single-user [71,72] and multiuser scenario [73–75],

mostly focusing on the analog/digital precoding design and the impact of imperfect

CSI. However, these approaches generally disregard the practical implications of

signal processing in the RF domain, which have been thoroughly addressed in [76].

• With regards to antenna design, discrete lens arrays (DLA)-based hybrid

transceiver architectures have found increasing interest in the research community

[14,77]. Differently from classical MIMO, DLA-based systems are characterized by

a direct correspondence between the number of transmitting beams and RF chains,

since the array behaves as a convex lens, directing the signals towards different

points of the focal surface [22]. Accordingly, DLAs preserve narrow beam-widths

in reduced RF-chain operations, allowing to reduce both the power required per

stream and the interference between the streams. In fact, DLA-based B-MIMO

systems, where analog spatial beamforming is performed through DLAs, proved

to be able to achieve near-optimal performance when combined with beam selec-

tion concepts [23,78]. More specifically, [23] considers a line-of-sight only scenario,

while the selection algorithm presented in [78] leverages on the diversity effects in

a multiuser scenario in order to increase system energy efficiency. Additionally,

recent studies [79] have proven that DLA-based systems can benefit from energy

efficient CSI estimation in point-to-point mm-wave MIMO.

33
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

2.6 Constructive Interference

In addition to the conventional signal processing approaches from the literature that

focus on interference minimization, this section introduces a different line of research that

aims at exploiting multiuser interference as a means for increasing the received SINR. In

fact, early works on linear precoding [30, 32, 80] showed that interference minimization

does not necessarily lead to the best performances in a communication system. Since

interference is data dependent, the transmitter can predict the multiuser interference at

the receiver and use this knowledge to influence it and benefit from it.

While the seminal works in [81, 82] focused on reducing the negative effects of inter-

ference while preserving its positive components, [83] finally showed that the transmitted

signal can be precoded in order to rotate the destructive component of interference into

constructive or beneficial interference. Therefore, future research is focusing onto identi-

fying new optimization metrics that exploit CSI and data knowledge at the transmitter

side to maximize the SINR of each user by capitalizing on the power contained within

multiuser interference. More recently, works on Phase-Shift Keying (PSK) modulated

signals [84–88] have introduced different metrics that prove how known interference can

be effectively used as a source of green signal power for downlink transmissions. In fact,

the symbol error probability in a L-PSK modulation is a function of the received SINR

ξ, as expressed in the following equation [83]:

( s )
sin2 Lπ · ξ
 

L−1 l·L
P = 1−
, (2.39)
L Lπ − π 1 + ξ sin2 Lπ

where ( s )
sin2 Lπ · ξ


π π
l = + tan−1 cot
. (2.40)
2 L 1 + ξ sin2 Lπ

While in this thesis the main focus resides in the application of constructive inter-

ference to downlink precoding and transmission, the same concepts can be applied to

different scenarios. In this regard, [89] has investigated the applicability of symbol-level

precoding based on relaxed receive constellations to increase the security of MU-MIMO

communication systems by means of Directional Modulation [90].

34
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

2.6.1 Constructive Interference Regions for PSK modulated transmis-

sions

When considering PSK-modulated signals, interference can be classified as construc-

tive or destructive according to simple geometrical concepts [81–85]. In fact, interference

can be considered beneficial for system performances when it leads the received symbol

further away from the decision thresholds of the desired constellation symbol. A visual

representation of constructive and destructive interference regions is presented in Fig.

2.10. Here, the received symbol benefits from interference when it falls in the construc-

tive region (i.e., the blue shaded area) and is instead affected by its negative effects when

it lies in the destructive region (i.e., the red shaded area). As we can see, when received

symbols fall in the destructive region they reside closer to the decision thresholds, rep-

resented by the bold lines, when compared to the desired symbol. On the other hand,

all the points lying in the constructive region are characterized by a larger distance

from the decision thresholds. The analytical conditions that split interference between

Figure 2.10: Constructive regions for different constellations

constructive and destructive can be defined according to two separate approaches: low-

complexity metrics, which are specifically tailored for a modulation order, and convex

optimization-based metrics, which are generalized for any-order PSK modulations. Both

approaches, including their applications, are presented in the following.

2.6.1.1 Low-Complexity metrics and their applications

Low-complexity metrics propose to identify constructive and destructive interference

according to the cross-correlation matrix R = HHH . Under such assumption, the

35
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

received signal for the m-th user is equivalent to the one of a MF precoded multiuser

systems
X
rm = ρm,m um + ρm,k uk = ρm,m um + ICIm , (2.41)
m6=k

where ρm,k is the m-th element of the k-th column of the cross-correlation matrix R and

ICIm is the inter-channel interference experienced by the m-th user. Given the set of

constructive and destructive interferers for the m-th user C and D, the set membership

conditions for 2-PSK, 4-PSK and 8-PSK modulations are defined as follows.

• 2 -PSK or Binary Phase-Shift Keying (BPSK) conditions:

C : {k|sign(um ) = sign (< [ρm,k uk ])} , (2.42)

D : {k|sign(um ) 6= sign (< [ρm,k uk ])} , (2.43)

where sign(·) and <[·] identify the sign and the real part of the argument respec-

tively.

• 4 -PSK or Quadrature Phase-Shift Keying (QPSK) conditions:

C : {k|sign(<[um ]) = sign (< [ρm,k uk ]) ∩ sign(=[um ]) = sign (= [ρm,k uk ])} ,

(2.44)

D : {k|sign(<[um ]) 6= sign (< [ρm,k uk ]) ∪ sign(=[um ]) 6= sign (= [ρm,k uk ])} ,

(2.45)

which represent a bi-dimensional version of (2.42) and (2.43), as 4-PSK constella-

tion symbols require both real and imaginary part.

• 8 -PSK conditions:

√
          
ρm,k uk ρm,k uk ρm,k uk 1 ρm,k uk
C : k|( 2 − 1)< ≤= ∩= ≤ √ < ,
um um um ( 2 − 1) um
(2.46)
√
          
ρm,k uk ρm,k uk ρm,k uk 1 ρm,k uk
D : k|( 2 − 1)< ≥= ∪= ≥ √ < .
um um um ( 2 − 1) um
(2.47)

36
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

While these metrics were initially used to remove destructive interference compo-

nents, as in dynamic linear precoding [91], successive studies proved that they could

be efficiently applied to exploit the energy of destructive components, as in correlation

rotation linear precoding [83]. More specifically:

• Dynamic Linear Precoding (DLP) is a linear technique based on the assump-

tion that destructive ICI can be predicted at the transmitter side and its effects

nullified at the receiver side. Given the conditions over interference listed above,

the m-th element of the k-th column of the constructive only correlation matrix

Rc is analytically defined as


ρm,k , ∀ k ∈ C

ρ̇m,k = (2.48)

0,
 ∀ k ∈ D.

Once the modified correlation matrix Rc has been derived, the precoding matrix

GDLP is defined as follows

−1
H H −1

H HHH Rc
GDLP = γDLP H HH Rc = r h i. (2.49)
H −1 H
tr Rc (HH ) Rc

• Correlation Rotation (CR) is a linear precoding technique based on the as-

sumption that ICI can be predicted at the transmitter side [83] and exploited to

enhance the received SINR. More specifically, the relative phase-shift between the

desired symbol um and the interference symbol uk can be defined as

ρm,k (uk )∗
φm,k = ej (U{um }−U{ρm,k uk }) = um , (2.50)
|ρm,k |

where operator U{·} identifies the phase extraction of the argument and φm,k is

the m-th element of the k-th column of the matrix Φ.

In CR precoding, the information over the interference relative phase-shift is used

to rotate the ICI in order to have constructive interference. The phase rotation is

37
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

obtained via the precoding matrix GCR

−1
H H −1

HH HHH Rφ
GCR = γCR H HH Rφ = r h i, (2.51)
H −1 H
tr Rφ (HH ) Rφ

where the phase-correcting correlation matrix Rφ = R ◦ Φ is defined as the

Hadamard product between the correlation matrix and the relative phase-shift

matrix [83].

2.6.1.2 Convex metrics and their applications

Convex constructive interference exploitation metrics are derived for a generalized

downlink transmission where the transmitter sends a precoded vector x and conditions

are imposed directly over the received signal in a noiseless scenario r = Hx instead of

the cross-correlation matrix. Additionally, convex metrics for constructive interference

exploitation offer a particularly important property, which is modulation order indepen-

dence. In fact, these metrics operate a phase-shift on the received signal rm according

to the phase of the symbol of interest for the m-th user φm = U{um }. The phase-shift is

a fundamental operation, as it allows to isolate the received amplitude and phase-shift

over the desired symbol um caused by the interference tm .

Figure 2.11: Visual representation of phase-shift in Convex metrics for 8-PSK

In [84], constructive interference conditions are analytically expressed for the m-th

38
 Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

received signal for the case where the received signal fully aligns with the desired symbol

um :

N
!
    X
−jφm −jU{um } −jU{um }
p
< rm · e = < rm · e =< hm,n xn e ≥η N0 , (2.52)
n=1

N
!
    X
= rm · e−jφm = = rm · e−jU{um } = = hm,n xn e−jU{um } = 0, (2.53)
n=1

where η ∈ R+ is a direct proportionality coefficient used to set a threshold for construc-

tive interference over the real part of tm and determines the resulting SINR. Note that

the conditions in (2.52) and (2.53) are imposed over the phase-shifted received signal

rm · e−jU{um } , according to the phase of the symbol of interest for the m-th user U{um },

as shown in Fig. 2.11b.

The condition in (2.53) can be further relaxed, as the phase of the received symbol

rm does not need to be strictly aligned with the phase of the desired symbol um . In fact,

interference is to be considered constructive and beneficial for the transmission as long

as the received symbol rm is contained in the constructive area of the constellation, as

in the 8-PSK example of Fig. 2.11.

From basic geometry properties and from the conditions (2.52) and (2.53), the con-

structive interference region for the m-th user can be defined as

     p 
= rm · e−jU{um } ≤ < rm · e−jU{um } − η N0 tan Φ, (2.54)

where Φ is the central angle of the constructive interference sectors, which depends on

the constellation order L and can be readily computed as Φ = ±π/L.

The constructive interference constraint definition in (2.54) allows the identification

of a new precoding optimization region that exploits the interfering signal power, instead

of reducing it. In fact, as shown in Fig. 2.11, the constructive interference regions can

be defined as sectors with infinite radii whose central angle depends on the constellation

order. This definition allows to relax classical optimization metrics based on interference

minimization, as the constructive interference region is only constrained by the proximity

39
Chapter 2. Energy Efficiency in Multiple-Antenna Wireless Systems

to the decision thresholds and extends infinitely in the directions away from them.

Optimization region constraints are visually represented in Fig. 2.11 by the dashed

lines.

Convex metrics for interference exploitation led to the formulation of a general-

ized beamforming, called Constructive Interference Beamforming (CIB). As described

in [84], CIB is a convex optimization-based beamforming scheme, where the downlink

beamforming problem is formulated in order to exploit the multiuser interference expe-

rienced at the receiver side as

PCIB : minimize kxk2

x (2.55)
  h   i
subject to = ḣTm x ≤ < ḣTm x − Γm tan Φ, ∀m,

where ḣm = hm e−jU{um } and the SINR requirements are identified as Γm , ∀m. In [84] it

was shown that PCIB is a Second-Order Cone Programming (SOCP) problem and can

be efficiently solved by means of standard convex optimization tools.

In addition, Constructive Interference exploitation concepts can also be applied to

the SINR balancing problem in (2.14) as follows

PSCIB : maximize Γt
x
  h   √ i
subject to = ḣTm x ≤ < ḣTm x − Γm N0 tan Φ, ∀m, (2.56)

kxk2 ≤ Pt .

Similarly to its power minimization formulation, the optimization problem PSCIB is

a SOCP problem that can be solved by means of convex optimization tools.

40
Chapter 3

Interference Exploiting Constant

Envelope Precoding in Massive
MIMO

In linear precoding-based MIMO communications, the average or instantaneous total

transmitted power is generally constrained to a specific value by sum-power constraints

[28, 29]. This is mostly supported by the fact that sum-power constraints are easy to

model and study. However, in a realistic scenario, each antenna of the base station is

typically connected to its own power amplifier (PA), which has to meet specific power

constraints. This is particularly relevant in M-MIMO, because the benefits of using

a large number of antennas at the transmitter side are followed by heavy burdens in

terms of hardware costs and power consumption, which strongly affect its feasibility for

future communication systems. In fact, the role of amplifiers is particularly critical for

M-MIMO practicability, as inefficient PAs are accountable for ∼ 40 − 50% of the total

power consumption [92].

Toward this end, the employment of non-linear RF components in conjunction with

low peak-to-average power-ratio (PAPR) precoding techniques [93] are expected to posi-

tively affect the energy efficiency of M-MIMO [17,18,94]. More specifically, [94] presents

a transmission scheme for orthogonal frequency-division multiplexing (OFDM) modu-

41
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

lations based on low PAPR precoding, while [17, 18] propose a CEP technique where

the transmitted signal amplitude corresponding to each antenna is constant and inde-

pendent from the channel realization, i.e., leading to a unitary PAPR and therefore

facilitating low cost PAs. In [17] the precoding technique is designed by minimizing

the error norm function of the received signal for a single user scenario, while in [18]

the transmitted symbols vector is designed for multiuser MIMO with the aim to reduce

the interference caused by other users. CEP was further analyzed in [95], where the

precoding design for frequency-selective MIMO channels is presented. Still, the per-

formances of CEP with interference reduction are strongly affected by the number of

iterations used and by the array size at the transmitter side [18]. In addition, the study

in [96] investigated the effects of phase constraints at the transmitter, since additional

restrictions to the change in transmitted phases at different symbol times can increase

the energy efficiency of the system. Finally, the authors in [97] further improved the

performances of interference reduction CEP, by employing cross-entropy optimization

instead of gradient descent-based algorithms.

While the above approaches focus on interference minimization, several works on lin-

ear precoding [30,32,80] have proven that interference minimization does not necessarily

lead to the best performances in a communication system. In fact, since interference

is data dependent, the transmitter is able to predict the MUI at the receiver and can

use this knowledge to influence it and benefit from it. Accordingly, this chapter intro-

duces two novel CEP techniques which exploit concepts of constructive interference for

PSK-modulated signals. In the proposed techniques, conditions over interference are

relaxed, allowing the transmitter to use the interfering signal as a green source of power

to increase the SINR at the receiver side. It is important to highlight that the proposed

schemes are particularly suitable for high-interference and low-SNR scenarios, where

low order modulations such as 2-PSK and 4-PSK are often preferred to ensure reliable

communications [98]. Nevertheless, the benefits of constructive interference could also

be extended to Quadrature Amplitude Modulation (QAM) signals, e.g., over the outer

constellation points of a 16-QAM or to the whole constellation by means of adaptive

decision thresholds [80].

42
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

The main contributions presented in this chapter are synthesized in the following:

• Definition of a new optimization region for CEP, based on the concepts of con-

structive interference.

• Introduction of two different CEP approaches, where both equality and inequality

power constraints are considered.

• Study of the computational costs of the proposed techniques in comparison with

the classical CEP approach in the literature.

• Introduction of a CSI-robust precoding scheme based on a relaxation of the inter-

ference optimization region.

• Evaluation of the performances of the proposed schemes for different PSK modu-

lation orders and in scenarios where the transmitter holds perfect and imperfect

CSI.

3.1 System Model

Consider a downlink multiuser scenario where the BS employs an N dimensional an-

tenna array to communicate with a population of M single-antenna users. The received

signal y is a CM ×1 vector that collects the M user received signals ym and is analytically

defined in accordance with eq. (2.26), without the subindex {·}d to ease the notation.

While complex channel gains hm,n in M-MIMO are modeled to include both the com-

plex small scale fading gm,n between the n-th antenna and the m-th user and the real

large scale fading coefficient βm experienced by the m-th user [11], our focus resides on a

single-cell scenario where channel gains are commonly modeled as independent Rayleigh

fading [99]. Accordingly, small scale fading gm,n are zero mean i.i.d. Gaussian variables

and large scale coefficients are considered unitary βm = 1, ∀m ∈ {1, ..., M }.

Given the total transmitted power by the antenna array Pt , the n-th element of the

43
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

transmitted signal x in a CEP-based system is defined as [18]

p
xn = Pn ejθn , (3.1)

N
P
where Pn is the power transmitted from the n-th antenna, so that Pn = Pt , and θn
n=1
represents the precoding phase of the CEP signal. Similarly, the received signal at the

m-th user can be defined as

N
X p
ym = hm,n Pn ejθn + nm . (3.2)
n=1

For simplicity and to ease the notation, throughout the chapter we assume unitary

transmitted power Pt = 1 and equally distributed power among the N antennas at the

BS, i.e., Pn = 1/N, ∀n ∈ {1, ..., N }, hence leading to

N
X 1
ym = √ hm,n ejθn + nm . (3.3)
n=1
N

The first term of the received signal ym can be rearranged in order to explicitly

discriminate between the desired signal and the interference. Analytically we have

ym = um + tm + nm , (3.4)

where um = dm ejφm is the PSK desired symbol for the m-th user, with magnitude dm

and phase φm , and tm represents the interfering signal for the m-th user

N
!
X 1
tm = √ hm,n ejθn − dm ejφm (3.5)
n=1
N
.

Accordingly, we can identify the total MUI energy as

M N
! 2
X X 1
EM U I = √ hm,n ejθn − dm ejφm (3.6)
m=1 n=1
N
.

44
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

3.1.1 Benchmark

First approaches to CEP were based on the minimization of the MUI energy [18]. In

order to minimize (3.6), the base station proceeds to identify the N dimensional transmit

phase angle vector θ = [θ1 , ..., θN ] that leads to the lowest MUI energy. Accordingly,

the CEP algorithm can be formulated as follows [18, 97]

M
 N
 2
√1 hm,n ejθn ejφm
P P
PCEIR : minimize N
− dm
θ m=1 n=1 (3.7)
subject to |θn | ≤ π, ∀n ∈ {1, ..., N } ,

which represents a non-convex non-linear least squares (NLS) problem, affected by local

minima. The optimization problem (3.7) was first solved in [18] with a gradient descent

(GD) based approach, and further improved in [97] with a direct application of cross-

entropy method [100]. Once the transmit phase vector θ is computed, the system

proceeds to compute the transmitted signal according to (3.1).

3.2 Constructive Interference Optimization Region

In Chapter 2, it was shown that for PSK-modulated signals, interference can be

classified as constructive and destructive according to simple geometrical concepts. In

fact, the interference signal tm can be considered beneficial for system performances

when it leads the noise free received symbol rm = ym − nm further away from the

decision thresholds of the desired constellation symbol um .

Previously, constructive interference conditions were expressed for the received signal

in a noiseless scenario, however, they can be explicitly imposed over the interfering signal

by substituting tm into rm as

   
= tm · e−jφm ≤ < tm · e−jφm tan Φ. (3.8)

A visual representation of the distinction between constructive and destructive inter-

ference is presented in Fig. 3.1a, where the desired symbol um is considered to be the

45
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

Figure 3.1: Optimization region for constructive interference exploitation, 8PSK

example: (a)Interference regions for 8-PSK symbol, (b) Interference
signal after rotation.

√ √

1/ 2 + j1/ 2 point of the 8-PSK constellation. Here the superscripts {·}c and {·}d

are used to differentiate between two different cases, where the received symbol falls in

the constructive region (i.e., the green shaded area) or in destructive region (i.e., the

red area), respectively. As per above, we can see that when the received symbol falls

in the destructive region it resides closer to the decision thresholds, represented by the

bold lines, when compared to the desired symbol. On the other hand, when t̄m lays in

the constructive region, its distance from the decision thresholds is greater than the one

which characterizes um .

The condition (3.8) is visually described in Fig. 3.1b for the 8-PSK case, where t̄m =

tm · e−jφm represents the rotated interfering signal for the m-th user and t̄R
m = < (t̄m )

and t̄Im = = (t̄m ) identify the shift from um suffered by the received symbol by means

of interference. More specifically, t̄R

m is an analytical measure of the amplification of the

received constellation point along the axis of um , while t̄Im represents a linear measure

of the angle shift from the original constellation point with phase φm 1 . The reader

is referred to [82–84] for more details on the definition of the constructive interference

region.

1
It is important to stress that t̄R I
m and t̄m can grow infinitely, as long as they respect the condition in
(3.8).

46
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

3.3 Constant Envelope Precoding with Constructive In-

terference Optimization

Existing studies in M-MIMO systems mostly consider precoding techniques with

sum-power constraints at the transmitter side. However, this is not a realistic assump-

tion, since each transmitting antenna is typically characterized by its own amplifier and

is hence affected by specific power constraints. Moreover, the use of precoding tech-

niques where the power at each antenna is fixed also allows the employment of highly

efficient amplifiers, hence reducing the total operational power consumption of the sys-

tem. Since CEP provides a solution to the above challenges, its joint application with

interference exploitation concepts is proposed, in order to improve the performance of

classical CEP approaches.

Toward this end, two different CEP approaches are introduced, both based on con-

structive interference exploitation concepts: one with CEP equality constraints, i.e.,

|xn | = p, ∀n ∈ {1, ..., N }, and a two-stage approach where the constraints are initially

relaxed to inequality conditions, i.e., |xn | ≤ p, ∀n ∈ {1, ..., N }, to be successively reap-

plied by means of normalization in order to perform CEP.

Following the concepts of constructive interference in (3.8), it is possible to define

a new optimization metric that maximizes the interference power, while imposing con-

straints over the phase of tm . Thanks to simple analytical operations, we can rearrange

(3.8) as
   
< tm · e−jφm tan Φ − = tm · e−jφm ≥ 0. (3.9)

The difference on the left side of the inequality can be used as an indicator of how

constructive or destructive the interfering signal tm is. In fact, if (3.9) is negative, the

interfering signal lies in the destructive region of interference, while if (3.9) is positive it

implies that the interfering signal is constructive. In addition, since the real part of (3.8)

represents the power of the interfering signal, we can infer that higher and positive values

of (3.9) lead to stronger forms of constructive interference. Accordingly, the Constant

Envelope Constructive Interference (CECI) optimization problem PCECI is defined as

47
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

follows:

min < tm e−jφm tan Φ − = tm e−jφm




PCECI : maximize
θ m (3.10)
subject to |θn | ≤ π, ∀n ∈ {1, ..., N } ,

where m ∈ {1, ..., M } and the operator min {·} represents the minimum value of the
m
argument among each of the M values. In PCECI the minimum value of the constructive

interference metric is maximized. With this approach, when the minimum value of the

metric is positive, we can automatically infer that the constructive interference condition

is verified and maximized for all the M users. In cases where the solution to PCECI

leads to negative values of the minimum, instead, it implies that the precoding phases

minimize the destructive interference as its least constructive component is maximized,

as visually described for the 8-PSK case in Fig. 3.1b. The formulation in PCECI is

clearly non-convex, however it can be efficiently solved via the cross-entropy method

(CEM).

3.3.1 A CEM Solver for Constructive Interference Optimization

The cross-entropy method can be described as an adaptive algorithm that aims

to the identification of rare events by means of variance reduction. The algorithm is

characterized by an iterative approach [100], where each iteration presents two main

steps:

• Generation of random samples based on a specific distribution f (θ, u).

• Update distribution parameters u ∈ R, according to the computed values of a

chosen cost function, in order to improve the random samples generation in the

following iterations.

The use of cross-entropy method to perform combinatorial optimization can be described

as follows. Consider the maximization problem described in PCECI , the global optimum

48
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

γ ∗ is defined as
γ ∗ = min {< (t̄∗m ) tan Φ − |= (t̄∗m )|}
m
h i (3.11)
= max min {< (t̄m ) tan Φ − |= (t̄m )|} ,
θ∈Θ m

where t̄∗m represents the m-th element of the normalized interfering signal, analytically

expressed as
N
!
X 1 ∗
t̄∗m = √ hm,n ejθn − dm ejφm e−jφm , (3.12)
n=1
N

with θn∗ being the n-th element of the optimal solution θ ∗ to the optimization prob-

lem. The application of CEM to optimization problems is based on the associa-

tion of the maximization problem with the probability estimation of a rare event.

Given a performance threshold γ, we can evaluate the probability of the rare event

min {< (t̄m ) tan Φ − |= (t̄m )|} ≥ γ as

m

 
L(γ) = Pu min {< (t̄m ) tan Φ − |= (t̄m )|} ≥ γ
n mn oo
= Eu I min {< (t̄m ) tan Φ − |= (t̄m )|} ≥ γ (3.13)
Z n m o
= I min {< (t̄m ) tan Φ − |= (t̄m )|} ≥ γ f (θ, u) dθ
m

where the operator Pu (·) evaluates the probability of the event in argument, the operator

Eu {·} represents the expectation of the argument with respect to the distribution f (θ, u)

and I {·} is a Boolean indicator function that returns 1 or 0 values when its argument

it true or false, respectively. The estimation of L(γ) can be performed through Monte

Carlo simulations 2 , by drawing a set of K random states Θ1 , ..., ΘK from f (θ, u) and

by computing

K
1 X n n    o o
L (γ) =
b I min < t̄km tan Φ − = t̄km ≥γ , (3.14)
K m
k=1

2
While analytical estimations of L(γ) can also be performed, Monte-Carlo estimation represents
the standard procedure for applications of the Cross-Entropy solver, as described in [100, 101] and as
performed in [97].

49
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

where t̄km is the m-th element of the interfering signal for the k-th state Θk =
 k
Θ1 , ..., Θkn , ..., ΘkN


N
!
X 1 k
t̄km = √ hm,n ejΘn − dm ejφm e−jφm . (3.15)
n=1
N

A direct application of (3.14) becomes rapidly prohibitive when the probability of

the event is very small, i.e., on the order of ∼ 10−5 . This can be addressed by means of

importance sampling, which estimates a different probability density function g (θ) that

more frequently generates such rare events. Under importance sampling, the estimation

problem becomes

K
1 X n n    o o f (Θ , u)
k
Lb (γ) = I min < t̄km tan Φ − = t̄km ≥γ , (3.16)
K m g (Θk )
k=1

f (Θk ,u)
where g (Θk ) represents the importance sampling distribution and g(Θk ) is defined as

the likelihood ratio (LR) estimator.

The importance sampling function is commonly chosen as a probability density func-

tion from the same family of f (θ, u), as

g (θ) = f (θ, v) , (3.17)

where v ∈ R is the tilting parameters vector and is obtained by computing the func-

tion with the minimum Kullback-Leiber distance from the ideal solution g ∗ (θ) =
I{S(θ)≥γ}f (θ,u)
L(γ) , where S(θ) is a real valued function of the optimization parameter θ.

The Kullback-Leiber distance or cross-entropy between two densities s(x) and t(x) is

analytically defined as

Z Z
D (s, t) = s(x) ln s(x)dx − s(x) ln t(x)dx (3.18)

and its minimization can be achieved through the maximization of the second term in

the equation. The tilting parameters v deriving from the minimization of the Kullback-

50
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

Leiber distance between g ∗ (θ) and f (θ, u) can be obtained as

I {S (θ) ≥ γ} f (θ, u)
Z
∗
v = arg max ln f (θ, v)dθ, (3.19)
v L (γ)

which, for the proposed optimization problem, is equivalent to the maximization [100] :

n n o o
v∗ = arg maxEu I min {< (t̄m ) tan Φ − |= (t̄m )|} ≥ γ ln f (Θ, v) . (3.20)
v m

A solution to (3.20) can be numerically estimated as

K
1 X n n    o o
b∗ =
v I min < t̄km tan Φ − = t̄km ≥ γ ln f (Θk , v). (3.21)
K m
k=1

Here, f (θ, v) is considered to be a Gaussian distribution, i.e., f (θ, v) = f (θ, [µ, σ]),

which allows to analytically estimate (3.20) as

K n o
I min < t̄km tan Φ − = t̄km
P 


≥ γ Θk
k=1 m
µ
b= K n o (3.22)
I min {< (t̄km ) tan Φ − |= (t̄km )|} ≥ γ
P
k=1 m

v 
u K n o
u P k k 2
u k=1 I min {< (t̄m ) tan Φ − |= (t̄m )|} ≥ γ (Θk − µ
b) 
u
m
σ
b = u , (3.23)
u 
K n o
t k k
P 
I min {< (t̄m ) tan Φ − |= (t̄m )|} ≥ γ
k=1 m

where µ
b and σ
b respectively represent mean and standard deviation of the importance
b ∗ = [b
sampling distribution, i.e., v µ, σ
b]. This assumption is not uncommon for continuous

optimization problems [97] and leads to efficient solutions. As previously mentioned,

CEM is based on an iterative approach and requires the tilting parameters to be updated

at each iteration. However, a direct update from (3.21) is often undesirable, as it might

rapidly converge to suboptimal solutions [100]. The occurrence of these events can be

reduced by using smooth updating procedures, as follows

µ(l) = αb
µ(l) + (1 − α) µ(l−1) (3.24)

51
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

σ (l) = αb
σ (l) + (1 − α) σ (l−1) , (3.25)

where the superscript (·)(l) represents the l-th iteration of the value in argument.

An analytical description of the constructive interference optimization precoding

based on cross-entropy optimization (CEO-CIO) technique is presented in Algorithm

3.1. Here, T represents the number of iterations, K identifies the random sample size

and ρ is direct proportionality coefficient used to compute the intermediate threshold

γ (l) . More specifically, the intermediate threshold γ (l) is identified by the cost function

evaluation Ck whose index is the smallest integer to be greater or equal to ρK and is

evaluated in Algorithm 3.1 as CdρKe .

Algorithm 3.1 CEO-CIO Precoding

Input: H, u, T , L , K
Output: x

Initialize µ(0) and σ (0)

for l = 1 → T
h i 
2 
(l) (l) (l) (l)
Θ(l) = θ1 , ..., θk , ..., θK where the columns θk ∼ N µ(l−1) , σ (l−1)
for k = 1 → K
(l) (l)
xk = √1 ejθk
N
(l) (l)
t k = H · xk − u
n    o
(l) (l)
Ck = min < tm,k e−jφm tan Φ − = tm,k e−jφm
m
end
Sort C1 ≥ C2 ≥ ... ≥ CK
γ (l) = CdρKe
b(l) and σ
µ b(l) from (3.22) and (3.23)
µ(l) and σ (l) from (3.24) and (3.25)

end
(T )
Return x = x1

The application of the proposed algorithm leads to received symbols r which preva-

lently reside in the constructive interference region. To illustrate this effect, Fig. 3.2

shows the received constellation of CEP precoded signals for the example of 8-PSK

constellation in a noise free transmission over 100 different channel realizations, in a

52
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

3
Received Symbols
Desired Symbols
2

Imaginary
0

-1

-2

-3
-3 -2 -1 0 1 2 3
Real

Figure 3.2: Received symbols for a noiseless scenario with N = 100 antennas for
M = 20 users when using 8-PSK.

scenario where the BS is equipped with N = 100 antennas and communicates with

M = 20 single-antenna users.

3.3.2 Two-Step Convex CEP

In addition to the previous approach, an additional technique for constant envelope

transmissions is presented where power constraints are initially relaxed into inequality,

allowing to use standard convex optimization techniques, and subsequently enforced to

equality via normalization at a later stage (i.e., by dividing the antenna outputs that do

not respect power constraints by their absolute value). In order to relax the conditions

in PCECI , we reformulate the optimization problem in its equivalent form where the

cost function is dependent on the transmitted signal x = [x1 , ..., xn , ..., xN ]T :

min < tm e−jφm tan Φ − = tm e−jφm




PeCECI : maximize
x m
√
subject to |xn | = 1/ N , ∀n ∈ {1, ..., N } . (3.26)
N
P
tm = hm,n xn − um .
n=1

Similarly to the optimization in PCECI , the above problem is non-convex, because

53
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

of the equality constraint over a convex set. In order to tackle this, the problem can

be convexified by imposing relaxed conditions to the transmitted signal xn ∈ C, ∀n ∈

√
{1, ..., N } and its absolute value |xn | ≤ 1/ N , ∀n ∈ {1, ..., N }. Thanks to this, the
0
optimization problem PeCECI can be reformulated into its relaxation PeCECI as

0 min < tm e−jφm tan Φ − = tm e−jφm




PeCECI : maximize
0 x m
√
subject to |x0n | ≤ 1/ N , ∀n ∈ {1, ..., N } . (3.27)
N
hm,n x0n − um .
P
tm =
n=1

where the superscript {·}0 is used to identify the solution achieved through relaxation.

Different from PeCECI , the newly formulated problem is a standard second-order cone

program (SOCP) 3 and can be effectively solved by means of standard convex optimiza-

tion techniques [84]. Since the constraints over the amplitude of the precoded signal
√
|x0n | ≤ 1/ N , ∀n ∈ {1, ..., N } cannot guarantee a strict constant envelope condition,

in order to achieve a full CEP transmission for all the antennas at the BS we need

to force the equality constrained before transmission. More specifically, in the second

and final stage of the algorithm, we can proceed by normalizing the elements where
√
|x0n | =
6 1/ N , ∀n ∈ {1, ..., N } as follows

√  √
x0n / N |x0n | ∀n where |x0n | =

 6 1/ N
xn = (3.28)
√
x0n |x0n |

 ∀n where = 1/ N .

The precoding scheme, namely identified as Convex Constructive Interference Opti-

√
mization (CVX-CIO), is analytically described in Algorithm 3.2, where x0  1/ N is
√
used to represent that x0n ≤ 1/ N , ∀n ∈ {1, ..., N }.

3
More specifically, the problem can be cast as a standard SOCP [102], as its objective function is
concave [84] as it can be decomposed into the combination of a linear function < tm e−jφm and a

concave function − = tm e−jφm . In fact, in [84] it was shown that the extraction of the imaginary

and real of a linear function preserves its linearity.

54
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

Algorithm 3.2 CVX-CIO Precoding

Input: H, u
Output: x
n o
x0 = arg max√ min [< (t̄m ) tan Φ − |= (t̄m )|]
m
x0 ∈C,|x0 |1/ N
 T
x01 x0N
Return x = [x1 , ..., xN ]T = √ , ..., √
|x01 | N |x0N | N

3.4 Computational Complexity

This section computes and analyzes the complexity of the proposed CEO-CIO in

comparison with the CEO approach to interference reduction (CEO-IR) precoding from

[97] in terms of floating-point operations (FLOPs), following the operational costs listed

in the literature [102]. More specifically, addition, subtraction and multiplication be-

tween two floating-point numbers are considered as a FLOP. Since both approaches are

characterized by the same number of iterations T , the analysis focuses on the computa-

tional burdens of the two different cost functions.

In line with Chapter 2, a simple time-division duplexing (TDD) scenario [51] is

considered, as it allows to exploit the reciprocity of the channel, enabling the CSI ac-

quisition for downlink via uplink pilots. This property is fundamental in M-MIMO

systems, as the time required by CSI acquisition TCSI becomes proportional to the

number of users M instead of the number of antennas N . In our analysis, we consider

a simple TDD case where TCSI = µM , with µ ≥ 1 being the number of pilot slots. Fi-

nally, a symmetrical transmission case is assumed, where the time for data transmission

Tdata = Tcohe − TCSI is divided between downlink and uplink transmissions according to

a parameter 0 ≤ DL ≤ 1. The parameter DL explicitly represents the portion of Tdata

devoted to downlink symbol transmission. Analytically, we have

TDL = DL (Tcohe − TCSI ) = DL (Tcohe − µM ) . (3.29)

55
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

3.4.1 CEO-CIO Costs

As previously mentioned, main costs of the proposed CEO-CIO algorithm reside in

the need to compute the cost function for each of the randomly generated samples. Cost

function computation can be synthesized in the following main operations:

• Computation of the received vector in a noise free scenario r = Hx,

• Identification of the interfering signal vector t = r − u,

• Projection of the interfering signal t̄ = t ◦ u∗ ,

• Identification of min {<(t̄) tan Φ − |=(t̄)|},

where ◦ represents the Hadamard product.

From the literature [102], the costs of each of the aforementioned operations is known:

the multiplication between a M × N matrix and an N × 1 vector requires M (2N − 1)

FLOPs, while the computation of the interfering signal and its rotation can be performed

with M FLOPs each, since they can be achieved by M subtractions and multiplications,

respectively. Finally, we can compute the costs of the identification of the minimum as

a search through an M -sized vector, hence leading to M FLOPs. It follows that the

proposed approach is characterized by a total FLOP count of M (2N − 1) + 4M FLOPs,

which includes the cost of the separation between the real and imaginary part of the

rotated interfering signal. Computational costs for the derivation and transmission of a

CEO-CIO signal are listed in Table 3.1.

3.4.2 CEO-IR Costs

The application of the conventional CEO-IR follows a similar pattern to CEO-CIO,

due to the fact that they both require the computation of the interfering signal for all

the randomly generated samples. More specifically, the computational costs of CEO-IR

can be highlighted in the following operations:

• Computation of the received vector in a noise free scenario r = Hx,

56
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

• Identification of the interfering signal vector t = r − u,

M
|tm |2 ,
P
• Computation of the interference energy
1

Following a similar approach to the previous section, we identify the multiplication costs

in M (2N − 1) FLOPs and the computation of the interfering signal as M FLOPs. Since

the interfering energy can be computed as the inner product of two M -sized vectors, i.e.,

by a cost of 2M −1 FLOPs, the total cost of the CEO-IR algorithm is M (2N −1)+3M −1

FLOPs.

As we can see, the computational costs of the proposed technique CEO-CIO are

comparable to the ones of the CEO-IR approach from the literature, as the FLOP

count difference is almost negligible. The total costs of the application of the precoding

techniques in a coherence time are listed in Table 3.1, which includes the effects deriving

by both the number of iterations T and the sample size K.

Table 3.1: Computational costs in FLOPs.

CEO-CIO
r = Hx T · K · M (2N − 1)
t=r−u T ·K ·M
t̄ = t ◦ u∗ T ·K ·M
min {<(t̄) tan Φ − |=(t̄)|} T · K · 2M
Total TDL · T · K [M (2N − 1) + 4M ]
CEO-IR
r = Hx T · K · M (2N − 1)
t=r−u T ·K ·M
tH t T · K · 2M − 1
Total TDL · T · K [M (2N − 1) + 3M − 1]

3.5 CSI-Robust CEP

In the previous sections we assumed the transmitter to possess a perfect knowledge

over the channel, allowing the definition of the constructive and destructive regions of

interference in absence of uncertainty. When the CSI acquisition is imperfect, however,

the received signal region extends according to the CSI error. The BS is assumed to be

57
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

aware of an estimated channel matrix, defined analytically as follows [84]

H
b = H + S, (3.30)

where the error matrix S represents the CSI uncertainty at the BS, statistically inde-

pendent from H, and characterized as a constrained spherical error, i.e., each element
n o
sm,n : |sm,n |2 ≤ δm,n
2 [84]. As in [84], the following analysis considers a scenario where
2
the base station is aware of the error bounds δm,n but has no knowledge over the er-

ror matrix S. Different from classical robust precoding approaches from the literature

[84, 103, 104], where the transmitted power is increased in order to overcome the effects

of CSI estimation errors, this section introduces a worst-case approach where the op-

timization region is redefined according to the CSI uncertainty, while preserving CEP

constraints. The estimated interfering signal for the m-th user, in case of imperfect CSI,

can be defined as follows

N
 
√1 b
h ejθn ejφm
P
tm =
b
N m,n
− dm
 n=1
N

√1 ) ejθn ejφm
P
= N
(hm,n + sm,n − dm (3.31)
n=1
N
 N
sm,n jθn
√1 hm,n ejθn jφ
P P
= N
− dm e m + √ e
N
,
n=1 n=1

where sm,n represents the n-th element of the m-th row of the CSI uncertainty matrix

S. As we can see in the last step of (3.31), the estimated interference signal b
tm is

characterized by two different components: the actual interference signal tm , i.e., when
N
sm,n ejθn . It follows
P
considering perfect CSI, and the uncertainty error signal zm =
n=1
that the estimated interfering signal can be defined as the sum of the two terms

tm = tm + zm .
b (3.32)

In (3.8), the interfering signal is rotated according to the desired symbol, with the

aim to have a region definition that is independent from the specific phase of the symbol

of interest um . In a similar manner, the rotated interfering signal for the m-th user in

58
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

tˉm Φ (δ)
R tˉm
Φ
zˉm zˉm
L

Φ L

a) General case b) Worst case

Figure 3.3: Imperfect CSI effects on the phase-shifted interfering signal t̄m .

presence of CSI errors b̄tm is defined as

tm · e−jφm = t̄m + z̄m .

b̄tm = b (3.33)

The second term in (3.33) can be described as the shift from the ideal interfering

signal t̄m caused by the CSI errors and can be represented as a circular constrained

region of uncertainty, as visually presented in Fig. 3.3a. Accordingly, we can identify

the worst-case scenario in the event where the actual interfering signal t̄m is within the

constructive interference region, but the uncertainty error signal z̄m moves the estimated
b̄tm away from it, as shown in Fig. 3.3b. Given the assumption of CSI errors to be

constrained within a spherical region, it is possible to analytically derive amplitude and

phase of the worst-case scenario uncertainty error signal z̄m .

Lemma 3.5.1 The amplitude of z̄m is characterized by the following analytical upper-

bound
N
P
δm,n
n=1
|z̄m | ≤ √ (3.34)
N

59
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

Proof of Lemma 3.5.1 Following the definition of z̄m we have

N
√1 sm,n ejθn e−jφm
P
|z̄m | = N
n=1 (3.35)
N
√1 |sm,n | ej(U{sm,n }+θn −φm ) ,
P
= N
n=1

where sm,n has been represented in order to show amplitude and phase and the oper-

ator U {·} identifies the phase extraction of the argument. The absolute value of zm

is evaluated as the absolute value of the sum of complex values. According to the tri-

angle inequality (i.e., given two complex numbers a, b ∈ C they satisfy the property

|a + b| ≤ |a| + |b|) we have

N N
X 1 X 1
√ |sm,n | ej(U{sm,n }+θn −φm ) ≤ √ |sm,n | . (3.36)
n=1
N n=1
N

Given the assumption of a spherical constrained error during CSI estimation, we have

N N
X 1 X 1
√ |sm,n | ≤ √ δm,n . (3.37)
n=1
N n=1
N

which ends the proof.

Finally, the worst-case scenario phase of z̄m can be readily identified as the phase

that is orthogonal to the constructive interference threshold identified by Φ.

The knowledge of the worst-case effects of CSI errors at the transmitter can be used

to relax the optimization region, in order to include the events that would be affected by

the uncertainty error signal. Thanks to this relaxation, a CSI errors robust precoding

can be derived, without the need to increase the transmitted power.

More specifically, according to simple geometrical analysis, the phase threshold Φ is

relaxed as  
 N
P 


 δm,n 


n=1
ΦR (δm ) = ΦL + arctan √ , (3.38)


 E {|tm |} N 


 

where ΦL = π/L identifies the threshold angle for the L order PSK modulation used

60
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

in transmission. Accordingly, we can define a new optimization problem, specifically

designed for the imperfect CSI case.

PCECIR : maximize min {< (t̄m ) tan ΦR (δm ) − |= (t̄m )|}

θ m (3.39)
subject to |θn | ≤ π, ∀n ∈ {1, ..., N } ,

Without loss of generality, this study considers a case where δm,n = δ, ∀m ∈

{1, ..., M } , ∀n ∈ {1, ..., N }, which leads to a simplified definition of the robust relax-

ation 
n √ o n √ o
δ N δ N π

ΦL + arctan

E{|tm |} if arctan E{|tm |} ≤ L
ΦR (δ) = (3.40)

ΦL−1 − 
 otherwise,

where  is an arbitrarily small positive quantity, which imposes an upperbound to the

growth of ΦR for high values of δ, and L − 1 identifies the modulation order which is

immediately lower than the one used during data transmission. The defined upperbound

is particularly important, given the fact that very high values of δ could cause ambiguity

with lower modulation orders, i.e., when their values lead the robust region ΦR (δ) to

coincide with or exceed ΦL−1 .

3.6 Results

This section shows the performances of the proposed precoding techniques through

Monte Carlo simulations over 50000 channel realizations. The downlink transmission

described in the previous sections is assumed, where the BS employs N = 64 antennas to

communicate with a population of M = 12 mobile users. While the proposed techniques

can be directly applied to any modulation order, results are presented for both 4-PSK

and 8-PSK. Legends are characterized by the following notation: CEO-CIO identifies

constructive interference driven precoding based on CEM, CEO-IR is used to represent

interference minimization CEO precoding and finally, CVX-CIO represents the two-step

convex CEP approach to constructive interference optimization. Both CEO techniques

are applied while considering the same parameter settings: T = 1000, ρ = 0.05 and

61
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

100

10-1

10-2
SER

10-3

10-4
CEO - IR
-5 ZF - P
10
CEO - CIO
CVX - CIO
10-6
-25 -20 -15 -10 -5 0 5
SN R [dB]

Figure 3.4: 4-PSK Symbol Error Rate when M = 12, N = 64 with perfect CSI.

α = 0.08 [97]. CEM solver parameter values have been studied in [101], where the used

settings are recommended for achieving good performances in terms of both convergence

speed and quality of the solution. In addition to CEO-IR and in line with the literature,

the proposed techniques are compared with a CEP approach to linear zero-forcing(ZF)

precoding [18], ZF-P in the legends, which can be analytically defined as

ejU{GZF u}
xZF −P = √ , (3.41)
N

where GZF is the ZF precoding matrix, as defined in Chapter 2.

Figures 3.4 and 3.5 present the SER as a function of the transmitted SNR for 4-

PSK and 8-PSK modulation respectively when considering a BS with N = 64 and

M = 12 users. As we can see from Fig. 3.4 and Fig. 3.5, the proposed approaches

greatly outperform the classical CEO-IR and ZF-P. This is due to the fact that CEO-

CIO wisely exploits the interference signal tm , ∀m ∈ {1, ..., M } to increase the received

signal power, while CEO-IR aims to a direct minimization of the interference energy.

Regarding the ZF-P approach, we can see that a direct normalization of the precoded

signal leads to a significant decrease in performances, due to its sub-obtimal approach.

In addition, Fig. 3.6 shows the SER as a function of the transmitted SNR when a

62
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

100

10-1

10-2
SER

10-3

CEO - IR
10-4 ZF - P
CEO - CIO
CVX - CIO
10-5
-15 -10 -5 0 5
SN R [dB]

Figure 3.5: 8-PSK Symbol Error Rate when M = 12, N = 64 with perfect CSI.

different topology is considered, where N = 32 and M = 6, for both the 4-PSK and

8-PSK case. Even in this different topology, the same performance trend is preserved

for all the described techniques, with the proposed schemes outperforming the CEP

approaches from the literature. It is interesting to notice that in Fig. 3.6, CEO-CIO

is able to achieve slightly better performances than CVX-CIO, differently from what

happens in Fig.s 3.4-3.5. This is due to the fact that the second-step normalization of

CVX-CIO, enforced in order to achieve a CEP transmission, has a stronger impact over

the performances achievable by systems with a lower number of transmitting antennas

and users, such as the scenario shown in Fig. 3.6. It was empirically observed that

on average both scenarios are characterized by the same number of elements where
√
|x0n | 6= 1/ N , ∀n ∈ {1, ..., N }. This means that for scenarios with a larger number of

transmitting antennas, the percentage of elements where the normalization is necessary

is lower, hence reducing the degradation in performances caused by the second-step

normalization of CVX-CIO.

In Fig. 3.7 we further characterize the proposed schemes when the base station
2
possesses imperfect CSI in the scenario where δm,n = δ 2 = 0.1, ∀m ∈ {1, ..., M } , ∀n ∈

{1, ..., N }. Here, we set the value of  = 0.1 and consider a simplified derivation of

63
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

100 100

10-1
10-1

10-2
SER

10-2
10-3
4-PSK 8-PSK
-3
CEO-IR 10 CEO-IR
10-4 ZF-P ZF-P
CEO-CIO CEO-CIO
CVX-CIO CVX-CIO
10-5 10-4
-10 0 10 -10 0 10
SN R [dB]

Figure 3.6: Symbol Error Rate as a function of the transmitted SNR when M = 6,
N = 32 with perfect CSI.

the relaxation ΦR (δ) where E {|tm |} is unitary. While more complex derivations of

E {|tm |} are expected to give a finer evaluation of ΦR (δ), it was empirically shown that

such approximation has a negligible impact in the overall system performance. More

specifically, Fig. 3.7 shows that for a system with imperfect CSI, the performances of

classical CEO-IR are strongly affected by the errors in the channel estimation, while the

performance gap with the proposed schemes is strongly accentuated. This phenomenon

is caused by the fact that CEO-IR aims to minimize the MUI over desired symbols with

unitary energy, hence leading to received points that are more prone to noise and CSI

errors. More specifically, CEP-IR leads to received symbols that more susceptible to

the imperfect CSI shift z̄m because of their shorter distance from the decision threshold

when compared to the proposed schemes. On the other hand, it is important to notice

how the performance trend of the proposed CEO-CIO scheme follows the one of the

system where perfect CSI is available at the transmitter. This is due to the interference

energy exploitation in the CIO scheme, which allows a certain robustness against noise in

the channel estimation. In addition, we can see that the robust relaxation of CEO-CIO,

identified as CEO-CIO R in the legend, is able to increase the performances achieved by

its non-robust counterpart, due to the proposed error-based optimization region, which

64
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

100

10-1

10-2
SER

10-3
ZF - P
CEO - CIO
10-4 CEO - CIO R
CEO - IR
CVX - CIO
10-5
-10 -5 0 5 10
SN R [dB]

Figure 3.7: Symbol Error Rate for 8-PSK modulation when M = 12, N = 64
with imperfect CSI and δ 2 = 0.1.

allows to partially reduce the deteriorating effects of imperfect CSI at the transmitter

side. Finally, it can be noticed how the CVX-CIO is inherently more robust to imperfect

estimations of the CSI when compared to the approach based on the CEM-solver. This

behavior is caused by the fact that CVX-CIO received signals tend be more aligned

to the corresponding desired constellation points, hence allowing a innate robustness

against noisy channels.

Fig. 3.8 studies the behavior of the proposed robust relaxation of CEO-CIO with

increasing values of the error bound δ 2 and fixed SN R = 10dB. As we can see, all the

techniques achieve lower SER performances as the error bound δ 2 grows. However, it is

interesting to notice how the CSI error-based region relaxation allows CEO-CIO R to

outperform the non-robust approach over all the spectrum of δ 2 values. Moreover, we

can see that the performance gap between the robust and non-robust version of CEO-

CIO becomes more significant as the uncertainty over the CSI grows, while the gap

between CEO-CIO R and CVX-CIO reduces as δ 2 grows. This is explained by the fact

that, when the error bound δ 2 is very low, the deriving relaxation is less appreciable,

hence leading CEO-CIO R to achieve similar performances to its non-robust counterpart.

On the other hand, for higher values of δ 2 , the CEO-CIO R is characterized by a more

65
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

100

10-1

10-2
SER

10-3
CEO - IR
ZF - P
10-4 CEO - CIO
CEO - CIO R
CVX - CIO
0.1 0.15 0.2 0.25 0.3
δ2

Figure 3.8: Symbol Error Rate for 8-PSK modulation when M = 12, N = 64 and
SN R = 10dB with imperfect CSI.

noticeable relaxation which leads to higher benefits, when compared to CEO-CIO.

In order to better represent the trade-off between complexity and performances of-

fered by the proposed scheme CEO-CIO, the achievable SER is shown in Fig. 3.9 as a

function of the per frame FLOPs count when considering an SN R = 0dB and perfect

CSI at the transmitter. More specifically, the size of the set of random states K is con-

sidered to gradually increase, as it directly affects the FLOPs count per frame shown in

Table 3.1. When computing the FLOPs per-frame, for the sake of simplicity, we consider

the frame length to be equal to the coherence time for downlink transmission TDL = 70,

in line with the LTE standard [98]. It is extremely important to highlight that, while the

proposed scheme is required to evaluate the precoded signal at a symbol-rate, such need

is shared by all the other CEP precoding schemes from the literature, for both single

[17] and multiuser scenarios[18, 95–97]. As a consequence, we can see that for similar

complexities, the proposed CEO scheme is able to greatly outperform its interference

reduction counterpart, hence showing a very positive and interesting trade-off between

complexity and performances.

66
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

100 CEO - IR
CEO - CIO

SER

10-1

10-2
0 2 4 6 8 10 12
Per-Frame Complexity [f lops] ×1010

Figure 3.9: Symbol Error Rate for 8-PSK modulation when M = 6 and N = 32
and SN R = 0dB as a function of FLOPs per-frame.

3.6.1 Constellation Energy

In previous simulations, the desired symbols are assumed to have unitary energy

constellation, i.e., dm = d = 1, ∀m ∈ {1, ..., M }. While this assumption is not un-

common in CEP literature [17, 18, 95, 97], the constellation energy can be increased to

improve CEP-IR performances. This represents one of the key drawbacks of the CEP-

IR approach, as its performances are strongly dependent on the constellation energy

E = d2 . In fact, since the expected value of the MUI is a function of both topology

(i.e., number of antennas at the BS and number of users) and modulation used in trans-

mission [18], it is not possible to know a priori the optimal constellation amplitude d∗ .

More specifically, the identification of the optimal energy would require to dynamically

estimate the SER at the transmitter side as a function of the constellation energy E,

hence increasing the computational complexity of the system. Otherwise, the search for

a sub-optimal constellation energy for CEO-IR could be performed at the transmitter

side via an additional topology-dependent optimization problem [18]. The optimization

67
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

100 CEO - IR 100 CEO - IR

ZF - P ZF - P
CEO - CIO CEO - CIO
CVX - CIO CVX - CIO
4-PSK 8-PSK
SN R = −5dB SN R = 0dB
SER

10-1 10-1

10-2 10-2
0 5 10 0 5 10
E

Figure 3.10: Symbol Error Rate as a function of the constellation energy E =

d2m = d2 , ∀m ∈ {1, ...., M } when M = 6 and N = 32.

problem that identifies the optimal constellation amplitude d∗ is defined as follows [18]

maximize d
(
M
 N
 2)
hm,n jθn
− dm ejφm
P P
subject to E √ e
N
≤γ (3.42)
m=1 n=1

dm = d, ∀m ∈ {1, ..., M }

where γ ≥ 0 ∈ R+ is a chosen threshold parameter to the MUI energy. The optimization

problem aims to identify the maximum constellation energy that preserves the expected

MUI energy within a desired threshold.

It is important to stress that the constellation energy is critical parameter for classic

CEO-IR. These considerations are visually presented in Fig. 3.10 and Fig. 3.11, for

the M = 6, N = 32 and M = 12, N = 64 scenario respectively. Both figures consider

the perfect-CSI case, while similar results can be seen for the imperfect-CSI case. In

fact, the aforementioned figures show that the performances of CEO-IR worsen as we

incautiously increase the constellation energy d, with this effect being particularly visible

for higher modulation orders such as 8-PSK. This is due to the fact that the MUI-based

metric used for CEO-IR aims to minimize the energy of the interference signal (i.e.,

68
 Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

100 CEO - IR 100 CEO - IR

ZF - P ZF - P
CEO - CIO CEO - CIO
CVX - CIO CVX - CIO
10-1
4-PSK -1 8-PSK
SER SN R = −5dB 10 SN R = 0dB

10-2

10-2
10-3

10-4 10-3
0 5 10 0 5 10
E

Figure 3.11: Symbol Error Rate as a function of the constellation energy E =

d2m = d2 , ∀m ∈ {1, ...., M } when M = 12 and N = 64.

the distance between the received symbol and the corresponding desired constellation

point), but fails to have any control over its phase U {tm }. More specifically, CEO-IR

metric is not affected by the phase of the received signal, which is particularly important

for PSK modulated signals, as information is carried through the phases of the received

signals, while their amplitude identifies their robustness against noise. Moreover, we

can see that the optimal d∗ changes when considering different scenarios and different

modulations, supporting how it is not possible to identify d∗ before transmitting.

On the other hand, the performances of the proposed techniques are not affected

by the desired symbol energy, as they aim to maximize the constructive effects of inter-

ference over the received signal. Therefore, a critical benefit of the proposed scheme is

that the additional optimization of E can be avoided, along with the significant asso-

ciated computational costs. In fact, as shown in Fig. 3.10 and Fig. 3.11 the proposed

techniques are able to outperform the classical CEO-IR for most of the energy spec-

trum. This is supported by the fact that the performances of the proposed metric are

independent from the desired symbol energy as they allow a constrained portion of the

interference at the user side. In other words, the proposed metric adaptively increases

the received constellation in function of the current CSI, without the need to addition-

69
Chapter 3. Interference Exploiting Constant Envelope Precoding in Massive MIMO

ally identify the optimal transmitted constellation energy, hence showing a very positive

complexity-performance trade-off. Additionally, we can notice that the performance gap

between CVX-CIO and CEO-CIO is larger for Fig. 3.11. This phenomenon is caused

by the aforementioned different impact of the second-step normalization of CVX-CIO,

which causes the performance gap to be more significant and beneficial for CVX-CIO

when considering systems with larger arrays at the BS.

3.7 Conclusions

This chapter proposed two CEP schemes for downlink multiuser transmissions where

interference is effectively exploited to greatly increase the received signal to interference

and noise. The proposed techniques showed that a relaxation of the optimization region

in function of the constructive interference can be beneficial to achieve reliable commu-

nications. The computational burdens of the proposed techniques have been analyzed

in terms of FLOPs, and compared with the approaches from the literature, showing

negligible differences. In addition, a precoding approach robust to bounded channel-

state information errors that does not require to increase the transmitted power has

been analytically derived and applied to scenarios where the base station possesses im-

perfect channel-state information. Finally, performances have been shown in terms of

symbol error rate for different modulation orders, proving the benefits introduced by

the proposed schemes when compared to classical CEP approaches with interference

reduction.

70
Chapter 4

Low-Complexity Interference
Exploiting Antenna Selection for
Massive MIMO

In Massive MIMO technology, the number of radiating elements can increase up

to few hundreds, performing secure, robust and energy-efficient communications [12].

However, the use of very large arrays leads to an increased and almost prohibitive

hardware complexity in terms of RF chains, as they are accountable for 50-80% of

the total transceiving power consumption [105]. Beside the application of CEP at the

downlink transmission from the previous Chapter, antenna selection can be seen as an

interesting approach to tackle the inherent hardware complexity of M-MIMO and, at

the same time, exploit the higher degrees of freedom provided by the excess of antennas

at the base station.

As described in Chapter 2, antenna selection (AS) in conventional MIMO system

has been a key topic of research in the past years [15, 20], showing the benefits in terms

of energy efficiency and performances of the use of a subset of antennas in transmission

or reception[106]. These works, among many others, proved that AS can reduce the

RF complexity at the transmitter/receiver. However, their computational costs increase

together with the system size, limiting their direct applicability in multiuser M-MIMO

71
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

scenarios. Toward this end, recent works are studying the energy efficiency benefits

offered by AS for large scale MIMO systems [107–109]. More specifically, the study in

[110] proposes an energy efficiency approach toward AS, but the energy costs caused by

the algorithms are not included in the analysis. In addition, the authors in [111] analyze

the effects of AS with a random approach, while [112] proposes an AS algorithm based

on Convex Optimization for a massively distributed antenna system.

This chapter introduces a novel low-complexity AS algorithm, specifically designed

to operate in a multiuser M-MIMO scenario and to exploit constructive interference

[80–83, 113]. In fact, the interference between the links of a MIMO system can be

beneficial for the transmission and improve the performances in terms of signal detection

by increasing the power of the desired signal. In the proposed scheme, the transmitter is

able to predict multiuser interference and can use this knowledge to identify the subset

of transmitting antennas for which interference is most constructive. The developed

algorithm makes full use of the high antenna diversity offered by very large arrays

and selects the subset that optimizes inter-channel interference, greatly reducing the

number of RF chains required for transmission and increasing the energy efficiency

of the system with a favorable trade-off between performance and complexity. The

proposed technique is characterized by a reduced digital signal processing complexity,

having overall computational burdens that are comparable or even lower than the ones

of a full M-MIMO system with a simple MF or ZF precoding.

The contributions presented in the chapter can be listed in the following:

• Introduction of a novel and low-complexity AS algorithm for multiuser M-MIMO

scenarios based on the concept of constructive interference;

• Analytical study of the computational complexity of the proposed scheme in com-

parison with previous approaches from the literature;

• Analytical derivation of the upper bound of the received SINR for each user, for

the proposed AS algorithm;

• Performance evaluation of the proposed technique in terms of SER and an energy

72
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

efficiency metric that combines throughput with system power requirements.

4.1 System Model

In this chapter, we focus on the downlink transmission of a transmit AS-based mul-

tiuser scenario where the BS equips a very large array of N antennas with NRF ≤ N RF

chains to communicate with M single-antenna users. In accordance with the previous

chapter, the channel response is modeled by a CM ×N matrix, whose channel entries

hm,n can be simply modeled as independent Rayleigh fading for a single-cell scenario

[99]. Additionally, in an effort to achieve the highest efficiency and lowest computational

complexity, the BS is assumed to perform only linear MF and ZF. Accordingly, in line

with (2.15), the linearly precoded transmitted vector for a transmit AS-based system

can be analytically defined as

x̃ = Ge u = γe Fe u, (4.1)

where Ge , Fe and γe identify the precoding matrix, precoding matrix without scaling

and the scaling factor corresponding to the equivalent channel matrix He , which is

defined as
 
He = h1,e , ..., hn,e , ..., hN,e . (4.2)

Here, the vector hn,e identifies the equivalent channel response corresponding to the n-th

antenna and its entries can be either null, when the index n corresponds to a deactivated

antenna, or equal to the channel response hn , when corresponding to a selected antenna.

Analytically 
0M ×1

 ∀n ∈
/N
hn,e = (4.3)

hn
 ∀n ∈ N .

73
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

4.2 Proposed Constructive Interference Maximization An-

tenna Selection (CIM)

The proposed scheme combines low-complexity MF and AS to optimize the con-

structive interference at the receiver side, as defined according to the low-complexity

metrics in (2.41). In Chapter 2, constructive interference for low-complexity metrics is

defined according to the cross-correlation matrix R = HHH . More specifically, inter-

channel interference for m-th user can be expressed to explicitly differentiate between

constructive and destructive components as follows

M
X
ICI ICI
ICIm = ρm,k uk = Cm + Dm (4.4)
k=1,k6=m

ICI , D ICI identify

where ρm,k identifies the m-th element of the k-th column of R and Cm m

the constructive and destructive component of ICI.

Constructive and destructive ICI can be analytically expressed as

X
ICI
Cm = ρm,k uk (4.5)
k∈C

X
DkICI = ρm,k uk . (4.6)
k∈D

where constructive and destructive subsets, i.e., C and D, are defined in (2.42) and (2.43)

for 2-PSK, (2.44) and (2.45) for 4-PSK and (2.46) and (2.47) for 8-PSK.

With the aim to exploit constructive interference energy, the transmitter can select

the antenna subset that, within a channel realization, is characterized by the highest

value of the minimum constructive interference. A straightforward application of AS

ICI for each user and select the minimum value for each of
would require to compute Cm

all the possible combinations of a subset of size NRF . This simple approach becomes

computationally prohibitive for systems with a high number of antennas and users,

leading to NN = N !(NN−N !


RF RF
)! possible subset combinations.
RF

Accordingly, a low-complexity approach to interference-exploiting AS is proposed,

74
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

Algorithm 4.1 Constructive Interference Maximization for 2-PSK

Input: H, u, NRF
Output: He

• for n = 1 → N

– Rn = hn hH
n
– [φ1 , ..., φm , ..., φM ]T = < [(Rn − diag {Rn }) u] ◦ u
– ψn = min {φ1 , ..., φm , ..., φM }

• end

• N = arg max {ψ1 , ..., ψn , ..., ψN }

NRF
 
• He = h1,e , ..., hn,e , ..., hN,e as (4.3)

based on the property

N
X N
X
HHH = hn hH
n = Rn , (4.7)
n=1 n=1

where Rn identifies the n-th antenna cross-correlation and is computed according to

n-th antenna channel response hn .

Clearly, a symbol-by-symbol control of ICI is computationally prohibitive for highly

populated scenarios, as it would require to compute constructive interference condition

M 2 − M times. In order to preserve a low computational complexity approach, the

proposed AS scheme considers a new parameter ψn that defines the interference related

to the n-th antenna. More specifically, for a 2-PSK modulated signal

   
n o  M 
(n)
X
ψn = min φ(n)
m , ∀m = min < [um ] <  ρm,k um  , ∀m (4.8)
 
k=1,k6=m

hP i
(n) M (n)
where φm = < [um ] < k=1,k6=m ρm,k uk is the decision metric for the n-th antenna and

ρ(n) is used to identify the elements of Rn . Finally, the algorithm proceeds to compute

ψn for all the N available antennas, and selects the NRF antennas that correspond to

the highest values.

75
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

Algorithm 4.2 Constructive Interference Maximization for 4-PSK

Input: H, u, NRF
Output: He

• for n = 1 → N

– Rn = hn hH
n
– t = (Rn − diag {Rn }) u
– [φ1 , ..., φM ]T = < [u] ◦ < [t] + = [u] ◦ = [t]
– ψn = min {φ1 , ..., φm , ..., φM }

• end

• N = arg max {ψ1 , ..., ψn , ..., ψN }

NRF
 
• He = h1,e , ..., hn,e , ..., hN,e as (4.3)

More specifically, the subset of selected antennas N can be defined as follows:

n n hP i o o
M (n)
N = arg max min < [um ] < ρ u
k=1,k6=m m,k k , ∀m , ∀n
NRF
n n o o (4.9)
(n)
= arg max min φm , ∀m , ∀n .
NRF

The selection technique is described analytically in Algorithm 4.1, where ◦ is used

to identify the Hadamard product.

While (4.9) is specifically tailored for performing AS in a 2-PSK modulated scenario,

constructive interference can also be used to enhance the received SINR of higher order

PSK modulated transmissions. Considering the conditions for constructive ICI described

in (2.44), the 4-PSK selection metric is defined as

φm,4P SK = < [(Rn − diag {Rn })u] ◦ < [u] + = [(Rn − diag {Rn })u] ◦ = [u] (4.10)

where the selected antennas subset N4P SK is defined as

n n o o
(n)
N4P SK = arg max min φm,4P SK , ∀m , ∀n . (4.11)
NRF

Selection for 4-PSK signaling is described analytically in Algorithm 4.2.

76
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

Unfortunately, because of the tightening on constructive interference conditions in

(2.44), 4-PSK selection preserves part of the destructive interference components. Ac-

cordingly, this section introduces a hybrid approach that nullifies the remaining destruc-

tive interference components which could not be optimized through AS, while preserving

the constructive interference benefits introduced by the selection algorithm. This is ob-

tained via the definition of a constructive correlation matrix Rφ whose entries can be

analytically described as


ρ

n,m if ρn,m ∈ C
ρφ (n, m) = (4.12)

0
 if ρn,m ∈
/C

Hence, the equivalent precoding matrix is defined as [81]

H −1
Ge,HY = γe,HY HH
e (He He ) Rφ (4.13)

where He indicates the equivalent channel matrix after the AS and γe,HY =
r h i
1/tr RHφ (HH H )−1 R
e e φ represents the corresponding scaling factor [81].

4.3 System Computational Complexity

One of the effects brought by the use of hundreds of antennas in M-MIMO systems

is a significant increase in computational costs, even when linear precoding techniques

are involved in the transmission. This section studies the complexity of precoding and

AS techniques of all the considered transceiving configurations in terms of floating-point

operations per second (FLOPs), following the analysis in [114], based on the costs listed

in the literature [115].

The proposed approaches are compared with several schemes where linear precoding

is performed in cascade with AS algorithms from the literature. More specifically, in

line with Chapter 2, the following algorithms are considered

• the low-complexity path gain maximization [40, 116] algorithm

77
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

• the recursive capacity maximization [38] algorithm

• a reduced complexity approach to the minimum eigenvalue [39] algorithm. Un-

fortunately, MEM, as proposed in [39] and described in Chapter 2, is computa-

tionally prohibitive for M-MIMO scenarios because of its exhaustive search-based

approach. Accordingly, following the work in [117], a decremental approach to

minimum eigenvalue maximization is considered, where the algorithm iteratively

identifies the antenna whose deactivation leads to the intermediate subset with

the largest smallest eigenvalue λmin .

In particular, a thorough analysis of linear precoding techniques costs for systems

that do not involve AS is first presented, then followed by a study of the computational

burden of the schemes that involve the AS techniques presented in the chapter.

4.3.1 Precoding

The dominant costs of ZF precoding can be identified in the following steps:

• Compute the correlation matrix R = HHH

• Compute the inverse of R

• Multiply R−1 by HH

• Apply the precoding matrix GZF to the data u

The number of operations necessary for each step of the precoding procedure depends

on the matrix size [115]. Matrix inversion is particularly expensive, as its computational
8 3
complexity grows exponentially with the size 3M , but it is computed only once per

coherence time. At the same time, the precoding procedure in M-MIMO becomes signif-

icant for the complexity count. In fact, due to the size of the matrices involved, the costs

of the precoding GZF u become particularly relevant in the final computational count.

Additionally, since the precoding operation is dependent on the desired symbols vector

u, it has to be performed on a symbol-by-symbol basis and is further characterized by

a factor TDL .

78
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

Full System with ZF Full System with MF

R M (2M + 1)(4N − 1) GM F MN
R−1 8
3
M3 GM F u TDL 2N (4M − 1)
GZF 4M N (4M − 1) − −
GZF u TDL 2N (4M − 1) − −
Total 2(M 2 + M )(4N − 1) + 38 M 3 + Total M N + TDL 2N (4M − 1)
4M N (4M − 1) + TDL 2N (4M − 1)
PGM Ant. Sel. with ZF CIM Ant. Sel. with MF
HHH N (2N + 1)(4M − 1) Rn N M (2M + 1)
He N ψn 2-PSK: TDL N M (2M − 3)
R M (2M + 1)(4NRF − 1) 4-PSK: 2TDL N M (2M − 3)
R−1 8
3
M 3
H e TDL N
GZF 4M NRF (4M − 1) GM F u TDL 2NRF (4M − 1)
GZF u TDL 2NRF (4M − 1) − −
Total (4M − 1)(2N 2 + N + 4M NRF + 2NRF TDL )+ Total 2-PSK: N M (2M + 1) + TDL N M (2M − 3)+
(2M 2 + M )(4NRF − 1) + 38 M 3 + N TDL N + TDL 2NRF (4M − 1)
CM Ant. Sel. with ZF
B M (2M + 1)(4N − 1) + 2M + 83 M 3
N −NRF
ABAH
P
[4(N − l)M (4M − 1) + (N − l)[2(N − l) + 1](4M − 1)]
l=1
N −NRF
[2M (4M − 1) + (4M − 1) + 4M (4M − 1) + M 2 ]
P
B
l=1
N −NRF
P
He (N − l)
l=1
R M (2M + 1)(4NRF − 1)
R−1 8
3
M3
GZF 4M NRF (4M − 1)
GZF u TDL 2NRF (4M − 1)
N −NRF
16
M3
P
Total [4(N − l)M (4M − 1) + (N − l)[2(N − l) + 1](4M − 1)] + 3
l=1
N −NRF
[2M (4M − 1) + (4M − 1) + 4M (4M − 1) + M 2 + (N − l)] + 2M
P
l=1
2
(2M + M )(4N + 4NRF − 2) + (4M − 1)(4M NRF + 2NRF TDL )
MEM Ant. Sel. with ZF
N −NRF
P
R M (2M + 1)(4(N − l) − 1)
l=1
N −NRF
[M + (N − l)( 64 M 3 + 4M 2 )]
P
λmin 3
l=1
N −NRF
P
He M
l=1
R M (2M + 1)(4NRF − 1)
R−1 8
3
M3
GZF 4M NRF (4M − 1)
GZF u TDL 2NRF (4M − 1)
N −NRF
[M (2M + 1)(4(N − l) − 1) + 2M + (N − l)( 64 M 3 + 4M 2 )]+
P
Total 3
l=1
(2M 2 + M )(4NRF − 1) + 38 M 3 + (4M − 1)(4M NRF + 2NRF TDL )

Table 4.1: Computational costs of different schemes in FLOPs

79
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

In opposition to zero forcing, MF precoding is based only on the computation of the

Hermitian transpose of the channel matrix and its application to u. As a consequence,

main costs of MF reside in the application of precoding to the data signal vector u.

4.3.2 Antenna Selection

4.3.2.1 PGM Antenna Selection

The PGM algorithm has low complexity, since it is characterized by two operations

only: the computation of the antenna path gains from the diagonal of the matrix HH H,

and the identification of the NRF highest values.

4.3.2.2 CM Antenna Selection

The CM algorithm has a very high complexity as its key operations are especially

−1
demanding. In particular, we need to: compute the matrix B = IM + SN R · HHH ,

select the minimum value of δn and, finally, update the matrix B. Main costs reside in

the iterative nature of this approach, as it leads to the need to repeat each of these steps

N − NRF times. Since the sizes of the channel matrix at intermediate stages H̄ change

at each iteration of CM, the computational costs of this technique require the use of a

summation, whose elements are a function of the iteration number and NRF .

4.3.2.3 MEM Antenna Selection

The MEM algorithm is affected by the highest computational costs as it requires to

compute the eigenvalues of intermediate stages correlation matrices several times within

a single iteration. More specifically, MEM is characterized by N − NRF iterations, each

one characterized by the necessity to compute N − l times the eigenvalues of H̄H H̄,

where l is the iteration step. In order to evaluate the computational burdens of MEM,

a tridiagonal QR algorithm is assumed for the computation of the eigenvalues, with the

assumption that R has been previously reduced to a tridiagonal form [118].

The costs of this approach are particularly high and nearly prohibitive for M-MIMO.

80
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

CIM-MF
109 PGM-ZF
PGM-MF
CM-ZF
Computational Cost [flops] 108 CM-MF
MEM-ZF
MEM-MF
MF-No AS
107
ZF-No AS

106

105

60 80 100 120 140 160 180

N

Figure 4.1: Computational costs as a function of the number of antennas at the

transmitter N for a 2-PSK modulated system with M = 5, NRF = M .

4.3.2.4 Proposed Constructive Interference Maximization (CIM) Antenna

Selection

The computational costs for CIM can be identified mainly in the following steps:

• Compute the antenna cross-correlation matrix Rn = hn hH

n

• Compute the decisional parameter ψn

• Identify maximum values of ψn

A detailed study of the computational burden of the proposed transceiving schemes,

within a single Tcohe , is reported in Table 4.1, along with the total complexity of each

of the AS schemes.

Table 4.1 does not include the computational studies for the MF precoding over

PGM, CM and MEM selection for the sake of brevity, since they can be easily obtained

by substituting the R → GZF u steps with the GM F u step of the proposed scheme.

81
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

4.3.3 Transceiver Computational Costs Analysis

From Table 4.1 we can see that computational costs for linear precoding in a classical

MIMO system reside mostly in the identification of the precoding matrix G. In fact,

costs for the application of data precoding in classic multiuser MIMO are less relevant

because of the reduced sizes of the system. However, this is not true for M-MIMO,

where the number of antennas is much larger than the number of users, hence leading

to computational costs for precoding application that are directly proportional to N .

This uncovers additional benefits brought by AS, as costs for precoding are strongly

reduced. At the same time, it is important to highlight that the proposed AS is affected

by the necessity to repeat the algorithm for each TDL because of its data dependent

nature. Nonetheless, the higher costs of data dependent AS over classical M-MIMO

are mitigated by the higher TCSI that characterizes such systems. It is important to

stress that the values shown in Table 4.1 are computed for a single coherence time, while

considering the renew frequency of data dependent operations. In fact, since precoding

and the proposed AS have to be repeated at a symbol rate, they are characterized by

a TDL factor, typically on the order of 4 OFDM symbols for a fast-fading M-MIMO

scenario [11]. On the other hand, the costs of classical AS algorithms are considered

only once per coherence time, as they compute costly metrics that are dependent on the

channel realization.

The selection metric of the proposed algorithm changes according to the constellation

order used at the transmitter, leading to different complexities, as presented in Table

4.1. Nevertheless, as shown in Algorithm 4.2, we can note that the difference between

2-PSK and 4-PSK metric is identified only in the need to compute the interference

metric for both real and imaginary part. In addition, the algebraic property in (4.7) is

independent from the constellation used and its cost represents an important component

of the global burden of the algorithm, significantly reducing the differences between the

two cases.

The effects described can be observed in Fig. 4.1, which shows the computational

costs in FLOPs as a function of the number of antennas at the BS N , when the number

82
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

of users is fixed to a specific value of M . In particular, Fig. 4.1 represents a fast fading

scenario, i.e. Tcohe is shorter than a frame, with M = 5 users, subset size equal to the

number of users NRF = 5 and TDD is characterized by the parameters: Tcohe = 100,

µ = 2 and ηDL = 50%. The values used correspond to a coherence time tcohe ≈ 7ms

when considering current LTE standards for frame time tf = 10ms and symbol time ts =

71.4µs [11] with a single carrier transmission scheme. This assumption is not uncommon

in the study of energy efficient systems, as recent works over large-scale MIMO systems

[119] showed the energy efficiency benefits of single-carrier transmission schemes. More

specifically, [119] shows that multi-carrier OFDM modulation has a very high PAPR,

which requires the RF power amplifiers to work within an operating regime where they

have low efficiency. Toward this end, the use of CEP [18] for massive MIMO system

has further shown the energy efficiency benefits of single carrier communications. In

fact, as shown in Chapter 3, single carrier communications with CEP at the transmitter

allow the use of energy efficient/non-linear RF components. From Fig. 4.1, we can see

that previous selection techniques are characterized by high costs, due to the size of the

system, leading to near overlapping curves for ZF and MF. It is interesting to notice

that the proposed technique has always lower costs than all the other approaches and

that the difference in costs increases as the number of antennas at the BS grows.

Note that the computational costs presented in Fig. 4.1 represent the overall FLOP

count required by the systems described in the legend, including both the precoding

costs and the AS costs, where applicable. Simple massive MIMO approaches, ZF -

No AS and MF - No AS in the legend, are characterized only by zero forcing and

matched filter precoding costs respectively, as they do not involve AS algorithms. On

the other hand, the computational burden for AS systems include both precoding and

selection algorithm. In fact, AS systems are identified in the legend according to the

following notation: the first acronym for the AS algorithm, while the latter represents

the precoding technique considered.

Given the data dependent/interference based nature of CIM selection, its perfor-

mances in terms of computational costs are affected by the length of Tcohe . As we would

expect, computational costs increase as the number of transmitted symbols per coher-

83
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

×105
10
ZF-No AS
9 CIM-MF
Computational Cost [flops] Tcohe = 300
8

4 Tcohe = 200

1 Tcohe = 100

0
60 80 100 120 140 160 180
N

Figure 4.2: Computational costs as a function of Tcohe for a 2-PSK modulated

system with M = 5, NRF = M .

ence time becomes larger, but with a lower ratio than the classical MIMO approach

with ZF. This effect is well described in Fig. 4.2, which shows the computational costs

for different values of Tcohe for both approaches. In this figure, only the costs of CIM

selection and the full system with ZF are considered, as results in Fig. 4.1 show that

AS algorithms of the literature experience computational burdens that are one or more

orders of magnitude higher than the proposed technique.

Note that the proposed AS algorithm requires a fast, symbol rate, RF switching. Due

to the criticality of this element, it is important for the RF switching to be performed

with low insertion losses. Toward this end, recent developments in hardware design

show that GaN MMIC based switches by TriQuint [120] can achieve switching speeds

on the order of ns, while offering promising performances in terms of insertion losses. In

a similar manner, solid-state RF switches represent a widely used technology in modern

communication systems and are able to provide switching speed inferior to 1µs [121].

In addition, it is important to state that fast symbol-rate RF switching schemes re-

quire time and bandwidth limited pulses [122] to tackle possible spectral regrowth. This

cannot be realized through conventional shaping filters, such as the raised-cosine, as they

84
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

are bandwidth limited and time unlimited. The design of time-limited orthogonal shap-

ing filters was first introduced for Ultra Wide Band (UWB) systems[123–125], showing

that it is possible to achieve pulses which are limited in time and in frequency. It is im-

portant to stress that the algorithms proposed in [123–125] are not UWB-dependent, as

they can be tuned to respect desired time and spectral constraints, as for the proposed

scheme. More specifically, the authors in [123] present a pulse shaping methodology

based on the Hermite functions, while [124] presents an algorithm based on the numer-

ical solution of the convolution between pulse and filter responses. Finally, the study in

[125] presents a convex optimization metric for a DSP based pulse shaping. In addition

to these works, the recent study in [122] presents a thorough analysis of the performances

of different time-limited shaping filters and applies the design to a multi-antenna system

that employs symbol rate RF switching.

These critical advances have fuelled the interest over single RF-chains techniques

[126], which require symbol rate switching, as for Spatial Modulation MIMO (SM-

MIMO)1 [127], Space Shift Keying (SSK)2 [122] and electronically steerable parasitic

array (ESPAR)3 communications [128]. These techniques have been successfully im-

plemented in real systems, in [129] for SM-MIMO and [130] for ESPAR, proving how

transmission schemes with similar requirements can offer increased values of energy

efficiency in modern communication systems.

1
In Spatial Modulation MIMO, a single RF chain and an antenna array are used to simultaneously
transmit multiple symbols. The RF chain switches among the antennas at symbol rate, in order to mod-
ulate the information symbols over a PSK/QAM symbol and over the antenna chosen for transmission.
2
In SSK transmission, the information bits are mapped over the index of a single radiating transmit
antenna. The system switches at a symbol rate between the available antennas according to the data to
be transmitted, while all the other antennas radiate no power.
3
Over ESPAR MIMO communications, it is possible to transmit multiple streams over a single RF
chain by adaptively exploiting the beam pattern characteristics of the arrays involved at the transmitter
side. In fact, while the ESPAR antenna explicitly transmits a PSK/QAM symbol, additional symbols
are analogically modulated by the antenna pattern, which is modified by exploiting the mutual coupling.

85
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

4.4 Performance Analysis of CIM selection with MF pre-

coding

In order to study the performances achieved by the proposed technique, the upper

bound of the average SINR for a single user is derived. Consider the received signal for

the m-th user when the base station employs CIM

ym = γe,M F hTm gm,e um + nm (4.14)

where gm,e identifies m-th column of the equivalent CN ×M precoding Ge and nm is the

m-th entry of the noise vector n.

Theorem 4.4.1 The received SINR of a downlink multiuser transmission under CIM

selection with MF precoding is characterized by the following upperbound

√
M (M + 1) + M (M − 1)(1 + M π + (M − 2)π/4)
ξf
m = (4.15)
NRF M N0

Proof of Theorem 4.4.1 The proof can be found in Appendix A.

The analysis of the received SINR can be used to derive a lower bound for the symbol

error probability. This can be easily computed for the proposed scheme by substituting

in (2.39) both the order modulation L and the final received SINR derived in (4.15).

4.5 Results

The performances obtained by the proposed AS technique are evaluated through

Monte Carlo simulations over 50000 channel realizations. A single-cell downlink scenario

with perfect CSI at the transmitter side is considered, where the BS is equipped with

an antenna array of N = 100 elements and communicates with M = 5 single-antenna

mobile users. The BS is assumed to possess a fixed number of RF chains NRF , equal

to the number of users M . In the simulations, both 2-PSK and 4-PSK modulations are

employed.

86
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

All the schemes described in the figures are characterized by a cascade of AS and

precoding at the transmitter, with the exception of simple massive MIMO approaches

here used as performance references. Legends have been conventionally defined to first

declare the AS performed: CIM to identify the proposed selection technique, PGM when

the subset is defined according to the path gain, CM for the capacity maximization and

MEM for the minimum eigenvalue maximization selection. Finally, precoding techniques

are identified as follows: ZF and MF for zero forcing and matched filter precoding,

respectively, and HY to identify the hybrid approach for 4-PSK transmissions. Precoding

schemes followed by No AS identify classical massive MIMO approaches where the BS

uses all the available antennas.

Fig. 4.3 and Fig. 4.4 show the SER performances as a function of the transmitted

SN R = 1/N0 for all the configurations described, with 2-PSK and 4-PSK signaling,

respectively. Performances in terms of SER for both cases are higher when no AS at the

transmitter is involved, but they are achieved thanks to a higher hardware complexity,

as previously shown. All previous AS techniques with MF are characterized by strong

losses in performances when NRF equals the number of users.

It is interesting to notice how in Fig. 4.3, PGM selection is characterized by strong

losses even with ZF precoding at the transmitter. Always in Fig. 4.3, we can see that

MEM and CM approaches with ZF achieve good performances as the SNR grows, but

at the expenses of high computational costs, which are nearly prohibitive for practical

systems. On the other hand, the proposed scheme, CIM with MF, shows only minor

losses in performances when compared to the full system with both ZF and MF, while

using only M of the N antennas available at the transmitter.

Fig. 4.4 shows the performances of the proposed algorithm in a 4-PSK modulated

scenario, when the antenna subset NRF = 10 and M = 5. The performances achieved

by the proposed scheme for 4-PSK are identified by the curve that corresponds to CIM-

HY in the legend. As we can see, the proposed scheme follows similar performances to

the previous and prohibitive AS techniques of the literature, showing a positive trade-

off between complexity and performance. It is pivotal to highlight that the shown

87
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

100

10-1

CIM-MF
PGM-ZF
SER

10-2 PGM-MF
CM-ZF
CM-MF
MEM-ZF
MEM-MF
10-3
MF-No AS
ZF-No AS
CIM-MF Analytical

10-4
-20 -10 0 10 20 30
SNR [dB]

Figure 4.3: SER as a function of the transmitted SNR for 2-PSK modulation
when N = 100, M = 5 and NRF = 5.

100

10-1

CIM-HY
SER

PGM-ZF
10-2
PGM-MF
CM-ZF
CM-MF
MEM-ZF
-3
10 MEM-MF
MF-No AS
ZF-No AS
-20 -10 0 10 20 30
SNR [dB]

Figure 4.4: SER as a function of the transmitted SNR for 4-PSK modulation
when N = 100, M = 5 and NRF = 10.

88
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

100

10-1

CIM-MF
SER

PGM-ZF
10-2 PGM-MF
CM-ZF
CM-MF
MEM-ZF
10-3 MEM-MF
MF-No AS
ZF-No AS
-20 -10 0 10 20 30
SNR [dB]

Figure 4.5: SER as a function of the transmitted SNR for 2-PSK modulation
when N = 100, M = 5 and NRF = 5 with imperfect CSI at the
transmitter α = 10.

performances for CIM-HY are achieved with significantly lower computational costs

than CM or MEM, for both MF and ZF cases. In addition, we can see that previous

AS systems that employ MF are affected by error floors. This behavior, distinctive for

MF precoding and here kept for the sake of completeness, is caused by the inability of

previous AS techniques to optimize the destructive effect of interference.

In Fig. 4.5 the proposed scheme is further characterized for the case where imperfect

CSI is considered at the base station. During AS and precoding, the BS is assumed to

be aware of an estimated channel matrix, defined analytically as follows [131]

H
b =H+E (4.16)

where E ∼ CN (0, β) is the error matrix, statistically independent from H, and β =

α
SN R = α · N0 is the variance of the estimation error for a TDD system, with α being

an inverse proportionality coefficient [131]. Fig. 4.5 shows that for a system with

α = 10, the performances of the proposed technique are affected by the errors in the

channel estimation. However, it is fundamental to highlight how the performance trend

89
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

1.4
CIM-MF
PGM-ZF
1.2 PGM-MF
CM-ZF
CM-MF
ηT [bit/channel use/W]

1
MEM-ZF
MEM-MF
0.8 MF-No AS
ZF-No AS
0.6

0.4

0.2

0
-20 -10 0 10 20 30
SNR [dB]

Figure 4.6: Energy efficiency over throughput as a function of the transmitted

SNR for 2-PSK modulation when N = 100, M = 5 and NRF = 5.

of the proposed scheme follows the one of a system when perfect CSI is available at the

transmitter.

In order to better illustrate the benefits brought by the proposed scheme and

the performance-complexity trade-off, we evaluate the achieved energy efficiency over

throughput ηT , according to (2.21). When computing (2.21), we consider complexity

values reported in Table 4.1 and we model hardware consumption from practical sys-

tems [45, 132], where Pamp = Pt /ν is defined as the power required by an amplifier with

ν = 0.35 efficiency in order to have a transmitted power Pt = 30dBm, PRF = 65.9mW

and Pf pga = 5.76mW/KF LOP s.

Fig. 4.6 describes the results in terms of energy efficiency over throughput for a

2-PSK scenario, showing that the proposed approach is characterized by higher values

of efficiency than all the other techniques, including the system without selection for

both ZF and MF precoding. The higher energy efficiency of CIM-MF is supported by

the lower hardware complexity it requires, which leads to reduced values of power at the

denominator in (2.21), while achieving similar performances in terms of throughput. In

addition to the computational savings, the proposed approach with CIM AS is able to

90
 Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

1.6
CIM-MF
PGM-ZF
1.4
PGM-MF
CM-ZF

ηT [bit/channel use/W]
1.2
CM-MF
MEM-ZF
1
MEM-MF
MF-No AS
0.8
ZF-No AS
0.6

0.4

0.2

0
0 200 400 600 800 1000
TDL

Figure 4.7: Energy efficiency over throughput as a function of the transmitted

symbols TDL when SN R = 10dB, N = 100 and M = 5.

achieve the shown performances with only 5% of the RF power required by full system

BS, i.e., when no AS is involved. On the other hand, we can see that previous AS algo-

rithms are characterized by very low energy efficiency because of the high computational

burdens they are affected by. In fact, the increased power consumed by the FPGA is

high enough to overcome the RF power savings.

Figure 4.7 presents the energy efficiency as a function of the number of single carrier

symbols during the downlink TDL when SN R = 10dB. As we can see from Fig. 4.7, the

proposed technique CIM-MF maintains higher performances than the other approaches

for increasing values of TDL . The performance gap between CIM-MF and the systems

without AS reduces as TDL grows, however it is important to highlight that the proposed

scheme keeps outperforming the classical MIMO approach until TDL ≈ 1000. These

values of TDL correspond to an OFDM modulated scenario with coherence time of 10

OFDM symbols with 256 sub-carriers. In fact, if we consider a Tcohe = 10 OFDM

symbols and TCSI = 2 OFDM symbols, we would have a TDL = 4 OFDM symbols

(TDL = 4 ∗ 256 sub-carriers roughly corresponds to TDL = 1000 symbols). These values

are often used when analyzing massive MIMO systems, as in the work by [51] where a

91
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

1.4

ηT [bit/channel use/W] 1.2

1
CIM-MF
0.8 PGM-ZF
PGM-MF
CM-ZF
0.6
CM-MF
MEM-ZF
0.4 MEM-MF
MF-No AS
0.2 ZF-No AS

0
5 10 15 20
Ns

Figure 4.8: Energy efficiency over throughput as a function of the subset size NRF
when SN R = 0dB, N = 100 and M = 5.

coherence time with Tcohe = 7 OFDM symbols is considered.

In order to identify the subset size that optimizes the trade-off between complexity

and performances, the energy efficiency is studied as a function of NRF . As shown in

Fig. 4.8, NRF = M proves to be a near optimal choice, as the CIM-MF curve presents

a peak around NRF = 6/7. Given the negligible efficiency gap between NRF = M = 5

and NRF = 7, the simple case where one transmit antenna is assigned to each user

is considered. In addition, it is interesting to notice how the full system outperforms

all the previous AS algorithms, with the exception of PGM, independently from the

choice of NRF . This is justified by the high complexity of MEM and CM, whose power

consumptions overcome the savings introduced by the use of a subset of RF chains. To

this end, Fig. 4.9 shows the power consumption of all the transmission schemes presented

in the chapter. Circuit power values required by CM and MEM are characterized by

very high consumptions, rapidly increasing towards the KW scale. On the other hand,

the proposed technique shows low power consumption values which lie below the full

system approaches, requiring ∼ 6.2W less than the MF without selection for the case

studied by our simulations, i.e. when N = 100, M = 5 and NRF = 5.

92
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

22
8000
CM-ZF
20 6000
CM-MF

Circuit Power Consumption [W]

18 4000 MEM-ZF
16 MEM-MF
2000

14 0
60 80 100 120 140 160 180

12
10
8
CIM-MF
6 PGM-ZF
4 PGM-MF
MF-No AS
2 ZF-No AS
0
60 80 100 120 140 160 180
N

Figure 4.9: Circuit power consumption at the BS as a function of the arrays size
N for a system with M = 5 mobile stations and NRF = 5.

1
ZF-No AS
0.9 CIM-MF
PGM-ZF
0.8
PGM-MF
0.7
ζ

0.6

0.5

0.4

0.3

0.2
60 80 100 120 140 160 180
N

Figure 4.10: RF Power Savings ζ as a function of the arrays size N at the trans-
mitter for a system with M = 5 mobile stations and NRF = 5.

93
Chapter 4. Low-Complexity Interference Exploiting Antenna Selection for Massive MIMO

Fig. 4.10 is presented to clarify the benefits introduced by the proposed technique

by showing the power savings ζ = P/PZF over a M-MIMO system that involves ZF,

where P represents the power consumption of the studied technique and PZF is used to

identify the power required by the full system with ZF. For the sake of simplicity the

same scenario is considered, where M = 5, NRF = 5 and the RF chains power values are

modeled as in (2.21). From Fig. 4.10 it is clear that the proposed technique CIM-MF

is less affected by the increased number of antennas at the transmitter because of the

low complexity of the selection technique, since ζ decreases significantly as N increases.

In particular, we can notice a power saving of ζ = 0.38 for the scenario considered,

where N = 100, M = 5 and NRF = 5 with 2-PSK transmission, meaning that the

performances are achieved with ∼ 62% less power. At the same time, we can see how

the other low-complexity approach PGM is characterized by increasing levels of power

consumption at the base station, since the curves for both MF and ZF present values

that grow towards the equal consumption ζ = 1 threshold as the array size at the BS

N increases.

4.6 Conclusions

This chapter proved that antenna selection and constructive inter-channel interfer-

ence concepts can be jointly used to improve energy efficiency performances of M-MIMO.

It was shown through analytical and numerical studies that constructive interference at

the receiver side can be optimized by efficiently identifying a subset of antennas at

the BS. Performances of the proposed algorithm have been evaluated in terms of sym-

bol error rate and an energy efficiency metric that combines throughput and power

requirements to analyze the trade-offs introduced. The presented system was further

characterized by confirming the numerical results through the derivation of a closed

form expression of the SER, when received SINR is considered equal to its analytical

upper-bound.

94
Chapter 5

Large Scale Antenna Selection

and Precoding for Interference
Exploitation

The previous chapter showed that the hardware complexity deriving from the use of

very large antenna arrays in M-MIMO can be tackled by means of TAS. Since a direct

application of MU-MIMO AS techniques to massive systems can be impractical in terms

of computational costs, see Chapter 4, recent works focused on AS algorithms precisely

designed for M-MIMO systems [107, 108, 133, 134]. More specifically, the work in [110]

presents an AS study under the perspective of energy efficiency, while the authors in

[111] study a random selection approach. Finally, [112] and [133] proposed the use of

Convex Optimization for M-MIMO AS systems, for a massively distributed antenna

system and for channel capacity optimization, respectively.

Conventionally, TAS-based MU-MIMO systems approach selection and downlink

precoding as two disjointed optimization problems. In fact, TAS systems are character-

ized by the definition of the antenna subset to be used for transmission and followed by

either linear or nonlinear precoding [15,20,38,41,116,134], as described in Chapter 4 and

shown in Fig. 5.1. This is mostly caused by the fact that conventional TAS algorithms

and precoding designs are often based on different and disjointed metrics, hence leading

95
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

Figure 5.1: Conventional TAS block diagram.

to two separated optimization problems. In contrast to this, we propose a novel trans-

mission approach where both TAS and precoding are jointly performed according to the

same performance metric, so to achieve the highest benefits from both techniques. This

key metric is based on the concept of interference exploitation [83,84]. The deriving op-

timization is a Mixed-Integer Programming (MIP) problem and can be efficiently solved

by commercial optimization solvers. In addition to the proposed MIP-based approach,

three different heuristic solutions to the optimization problem are introduced and their

performances are analyzed.

The algorithms here presented introduce a novel approach to M-MIMO TAS sys-

tems and are designed to fully exploit both constructive interference (CI) and the high

diversity offered by very large arrays (VLAs) by jointly selecting a small subset of trans-

mitting antennas and defining the precoded signal. The joint optimization enables to

fully exploit the benefits of both TAS and precoding. In fact, the proposed joint TAS-

precoding allows to greatly reduce the number of RF chains at the BS, hence achieving

both a significant mitigation of hardware complexity and power consumption at the

transmitter side, and an important reduction of the signal processing required.

The following list synthesizes the contributions presented in the chapter:

• Introduction of a novel transmission scheme for multiuser M-MIMO scenarios

based on concepts of CI exploitation that jointly performs TAS and precoding;

• Definition of a MIP-based and three low-complexity heuristic approaches to effi-

96
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

ciently solve the proposed optimization problem. The optimality of these heuristic

approaches is further discussed and analyzed;

• Study of the effects of imperfect CSI over the presented metrics and derivation of

a CSI-robust approach for the proposed techniques;

• Evaluation of the performances obtained by the proposed schemes in terms of

SER, Capacity and energy efficiency over throughput.

5.1 System Model and Benchmark Techniques

Similarly to previous chapters, we consider the application of the proposed schemes

to the downlink transmission in a single-cell multiuser scenario, where the BS adopts a

very large N -sized array to communicate with a population of M single-antenna users.

Accordingly, the channel response is modeled as a CM ×N matrix, whose entries are

modeled as independent Rayleigh fading [99]. In line with the previous chapter on TAS,

the BS is characterized by a total number of available RF chains card (N ) = NRF ≤ N .

Throughout the chapter, the proposed scheme is compared with classical TAS sys-

tem approaches from the literature. More specifically, we consider a capacity-based TAS

technique for M-MIMO where the selection is performed through convex optimization

for multiuser scenarios [133] and PGM. With regards to the downlink beamforming

techniques, we consider two main approaches based on SINR ratio metrics: TPM beam-

forming from [34], as described in (2.13), and the algorithmic SB beamforming from

[35], as described in (2.14).

A convex approach to capacity-based TAS for M-MIMO has been recently proposed

by [133], based on a relaxation of the binary constraints imposed over the diagonal of

the selection matrix in (2.18). Accordingly, the new relaxed optimization problem for

97
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

Convex Capacity Maximization (CCM) can be defined as

log2 det IM + ρH∆HH



PCCM : maximize
∆

subject to ∆n,n ∈ [0, 1] , (5.1)

PN
n=1 ∆n,n = NRF ,

which leads to the TAS subset definition

N = arg max {∆1,1 , ..., ∆n,n , ..., ∆N,N } . (5.2)

Nt

5.2 Proposed joint MIP Constructive Antenna Selection

and Precoding (MIP-CASP)

In line with Chapter 3, CI conditions are imposed over the interfering signal for the

m-th user tm . Following the same notation, we defined the conditions for CI as

   
ωm , < tm · e−jφm tan Φ − = tm · e−jφm ≥ 0, (5.3)

Given the conditions for constructive interference in (5.3), it is possible to define a

novel optimization problem that exploits the beneficial components of MUI by jointly

performing TAS and precoding as

PCASP : maximize min {< (t ◦ u∗ ) tan Φ − |= (t ◦ u∗ )|}

a,x̃

subject to t = Hx̃ − u,

kx̃k2 ≤ 1,

|x̃|  a,
N
P
an = NRF ,
n=1

an ∈ {0, 1} ,
(5.4)

where the operator ◦ identifies the Hadamard product, b  c represents that inequality

has to be respected for each element of both vectors (i.e.,bi ≤ ci , ∀i) and a represents

98
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

the selection vector, whose entries are either one, when the corresponding antenna is

to be connected to the RF chain for transmission, or null, when the corresponding

antenna needs to be deactivated. Clearly, (5.4) jointly optimizes the precoded symbols

through x̃ and the TAS through a, subject to power constraints in kx̃k2 ≤ 1 (without

loss of generality, the total power budget is assumed unitary for simplicity), and the

typical antenna-number constraint also found in (5.1). Given its binary constraint, the

optimization problem (5.4) is clearly non-convex, however it can be efficiently solved

by means of commercial optimization tools such as MoSek. Still, it is important to

highlight that its objective function is concave [84], since it can be deconstructed into

the combination of two functions: a linear function < (t ◦ u∗ ) and a concave function

− |= (t ◦ u∗ )|, as the extraction of the imaginary and real of a linear function preserves

its linearity [84].

As we can see, PCASP is designed in order to jointly perform the TAS (i.e., iden-

tifying the subset of transmitting antennas a) and design the precoded signal x̃. The

joint optimization allows to fully exploit the beneficial components of MUI, achieving

significant transmission benefits and a particularly interesting trade-off between system

complexity and performances.

5.3 Heuristic Approaches to Joint Antenna Selection and

Precoding

While the joint MIP-CASP approach effectively reduces the RF chains at the trans-

mitter, the joint optimization of a and x̃ involved introduces a significant computational

burden. Accordingly, this section proposes three heuristic Successive Optimization (SO)

approaches based on the decomposition of PCASP into three different optimization prob-

lems. That is, PCASP can be decomposed into the succession of three convex optimiza-

tion problems, as follows:

• a full-system preliminary precoding, where the precoded signal x of the system

with no TAS is derived via constructive beamforming (CBF)

99
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

• TAS, where the sub-set of transmitting antennas a is identified via CI AS (CAS)

• subset precoding, where the transmitted signal for the chosen NRF transmitting

antennas x̃ is re-computed.

Towards reducing the involved computational complexity, the solution to the suc-

cession of these problems can be achieved through different approaches with decreasing

computational complexities, here introduced and discussed. More specifically, the three

following approaches with reducing computational complexity are presented:

• 3-step SO approach, namely CBF-CAS-CBF, involving CBF for the original pre-

coding in the first step, CAS in the TAS step, and CBF in the final precoding

step, where each one of the three aforementioned optimization problems is solved

through convex optimization tools

• 2-step SO approach, namely MFCAS-CBF, where the first step is circumvented

by employing a closed form MF precoder, while the remaining two problems are

solved by means of convex optimization techniques

• 1-step SO approach, namely MFCAS, where the first and last steps are circum-

vented by employing the MF precoder and only the antenna optimization problem

is solved by convex optimization.

The proposed schemes are described in details in the following subsections.

5.3.1 3-step Successive Optimization CBF-CAS-CBF

We identify as 3-step CBF-CAS-CBF the scheme based on the decomposition

of PCASP into three different convex problems to be solved in a sequential man-

ner. In 3-step CBF-CAS-CBF, we first derive the precoded vector for the full-system

x = [x1 , ..., xN ]T by solving the CBF optimization problem

PCASP 3a : maximize min {< (t ◦ u∗ ) tan Φ − |= (t ◦ u∗ )|}

x

subject to t = Hx − u, (5.5)

kxk2 ≤ 1.

100
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

The optimization problem PCASP 3a can be cast as a Second-Order-Cone-

Programming (SOCP) [102], since its objective function is concave [84]. Once the opti-

mal precoded vector for the full N -antenna system x is achieved, the system proceeds

to identify the antenna subset for transmission based on a. The constructive TAS is

performed according to the following CAS optimization problem

PCASP 3b : maximize min {< (t ◦ u∗ ) tan Φ − |= (t ◦ u∗ )|}

a

subject to t = Hx − u,

|x|  a, (5.6)
an ∈ [0, 1] , ∀n ∈ {1, ..., N } ,
PN
an = NRF ,
n=1

where a represents the selection vector, following the same notation as for PCASP . The

solution to PCASP 3b yields a vector with non-binary values of a, which are achieved by

selecting the NRF largest elements with their indices representing the selected antennas.

Finally, in order to achieve the final transmitted signal, the solution to PCASP 3b is used

to identify the precoded vector x̃ for the transmitting antennas subset N in the following

PCASP 3c : maximize min {< (t ◦ u∗ ) tan Φ − |= (t ◦ u∗ )|}

x̃

subject to t = Hx̃ − u,
(5.7)
kx̃k2 ≤ 1,

|x̃|  a.

5.3.2 2-step Successive Optimization MFCAS-CBF

While the previous approach is able to achieve near optimal performances, it is based

on the derivation of the precoding vector for the full-size system x, which is a computa-

tionally demanding step. Because of this, in order to further reduce the computational

complexity of the signal processing at the BS, we propose an additional approach, called

2-step MFCAS-CBF, which leverages on the known property of asymptotic optimality

for linear precoding in massive MIMO systems [11].

101
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

Thanks to this property, the computational burdens required by the convex precod-

ing in PCASP 3a are greatly reduced, as they are replaced by a simple closed-form linear

precoding approach. As it follows, the 2-step MFCAS-CBF approach can be synthesized

in the following algorithm, where we first identify the subset of transmitting antennas

based on the assumption of MF precoding

PCASP 2a : maximize min {< (c ◦ u∗ ) tan Φ − |= (c ◦ u∗ )|}

∆

subject to c = H∆HH u − u,
PN (5.8)
∆n,n = NRF ,
n=1

∆n,n ∈ [0, 1] ,

where ∆ is the selection diagonal matrix, as in PCCM , and c identifies MUI interference

under MF assumption c = H∆HH u − u, i.e., x = HH u. After identifying the antenna

subset a = diag(∆), we then proceed to derive the precoding vector x̃ as a solution to

the CBF problem

PCASP 2b : maximize min {< (t ◦ u∗ ) tan Φ − |= (t ◦ u∗ )|}

x̃

subject to t = Hx̃ − u,
(5.9)
2
kx̃k ≤ 1,

|x̃|  a,

which can be efficiently solved by standard convex optimization techniques, as in (5.7).

5.3.3 1-step Successive Optimization MFCAS

In addition to the previous schemes, we propose a final approach to MUI exploiting

AS-precoding where the computational burden is further reduced. Here, the antenna

subset selection is the only problem that requires convex optimization in order to be

solved, while precoding is performed by assuming only MF at the transmitter side.

102
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

Accordingly, we can define a new single-step optimization problem as follows

PCASP 1 : maximize min {< (c ◦ u∗ ) tan Φ − |= (c ◦ u∗ )|}

∆

subject to c = H∆HH u − u,
PN (5.10)
∆n,n = NRF ,
n=1

∆n,n ∈ [0, 1] ,

After the transmitting subset N has been identified, we proceed to compute the

transmitted signal x̃, based on MF

M

P
 1/ξn
 hm,n um , ∀ n ∈ N ,
x̃n = m=1 (5.11)

 0 ∀ n∈
/ N,

N
|xn |2 = 1.
P
where ξn is a scaling factor, which guarantees a unitary transmitted power
n=1

5.4 Optimality Evaluation

This section provides a further characterization of the proposed heuristic approaches

by studying the impact that successive optimization and closed form approximations

have over the achievable cost function values. In line with the literature [135], we define

the figure of merit M as

f
M= ? (5.12)
fM IP ,

where f defines the objective function for which we want to measure the optimality.

Clearly here

f = min {< (t ◦ u∗ ) tan Φ − |= (t ◦ u∗ )|} , (5.13)

represents the cost function of the optimization problem PCASP when the heuristic

solutions are considered, i.e., the minimum CI achieved by successive optimization tech-
?
niques, and fM IP identifies the cost function evaluation when the optimal MIP-CASP

solution is considered, i.e., when f is computed considering the x̃ solution of PCASP .

The defined metric represents a direct evaluation of the optimality of the proposed

103
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

?
heuristic approaches, as fM IP represents the optimal and maximum value of minimum

constructive interference achievable from a system. Clearly M = 1 signifies that a

MIP-equivalent solution is obtained.

Fig. 5.2 collects the cost function evaluation f for all the proposed approached with

several modulation orders when considering increasing sub-set array sizes at the base

station, i.e., increasing the number of transmitting antennas at the BS NRF . Interest-

ingly, we can notice that both the 3-step and the 2-step approaches are able to achieve
?
near optimal solutions when compared to the optimal fM IP for all the modulation or-

ders. This result is particularly important, as it proves that SO-based approaches are

able to efficiently approximate and solve the MIP equivalent formulation. On the other

hand, we can see that the closed form single-step approach, 1-step MFCAS, is char-

acterized by lower values of f when we increase the modulation order, because of its

suboptimal approach when solving the precoding problem (i.e., MF linear precoding).

Nevertheless, Fig. 5.2 shows that such approach can still represent an interesting alter-

native for low-order modulation and low-energy scenarios, as it is characterized by very

low complexity and is still able to achieve acceptable performances for the 4-PSK and

8-PSK scenario.

These results are confirmed in Fig. 5.3, where the figure of merit M curves are

presented for three different modulation orders. As we can see, these results confirm

that both the 2-step and 3-step approaches are characterized by near optimal perfor-

mances, as they rapidly and closely approach the optimality line, represented by the

unitary value. On the other hand, the 1-step MFCAS approach proves to be a valuable

alternative for low-power and low-modulation scenarios, thanks to its favorable trade-

off between complexity and performances. Towards quantifying this trade-off, below we

evaluate the computational complexity of each of the proposed schemes.

5.5 Computational Evaluation

This section analyzes the computational costs of the proposed schemes in terms

of running time for different antenna array sizes at the transmitter side. In order to

104
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

101

Minimum Constructive Interference

4-PSK

8-PSK

100

16-PSK

10-1

1-step MFCAS
2-step MFCAS-CBF
3-step CBF-CAS-CBF
MIP-CASP
10-2
5 10 15 20
NRF

Figure 5.2: Minimum CI when M = 5 and N = 100.

perform a fair evaluation, we consider the running times in [s] within a coherence time,

i.e., for the number of frames where the CSI is assumed constant. This is due to the fact

that the proposed schemes require a symbol-rate evaluation of TAS/precoding, while the

conventional CCM scheme needs to be performed on a coherence time basis. With this

regard, we consider a TDD scenario [51] with realistic values for a fast-fading scenario

(i.e., where the proposed schemes are mostly suited for) where Tcohe = 10 symbols, in

line with the work in [51]. More specifically, in reference to the notation used in (3.29),

we consider a CSI acquisition time of TCSI = 5 symbols (i.e., µ = 1) in a DL dominant

scenario with TDL = 4 symbols dedicated to downlink transmission (i.e., ηDL = 0.8,

with TU L = 1).

As we can see in Fig. 5.4, the proposed schemes are overall affected by longer com-

putational times over the length of the coherence time. This is due to the fact that the

proposed schemes require a symbol-by-symbol update, in contrast with conventional

TAS schemes from the literature. Nevertheless, it is interesting to notice that the pro-

posed 1-step MFCAS scheme is characterized by running times that can be compared to

the ones of the benchmark scheme. This strongly reaffirms that such approach represents

a particularly appealing scheme for low-modulation scenarios, as it is able to achieve

105
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

M 4-PSK
0.9

0.8
5 10 15 20
1
M 8-PSK

0.8

0.6

5 10 15 20
1
M 16-PSK

1-step MFCAS
0.5 2-step MFCAS-CBF
3-step CBF-CAS-CBF
MIP-CASP
5 10 15 20
NRF

Figure 5.3: Figure of merit M when M = 5 and N = 100.

interesting performances with non-significant additional computational costs. On the

other hand, it is important to notice how the proposed 2-step and 3-step schemes, are

almost unaffected by the increase in array sizes, while the CCM-SB is instead char-

acterized by increasing computational times as N grows. Accordingly, for very large

systems, the proposed schemes are expected to be characterized by similar complexity,

when compared to existing TAS schemes. On the other hand, we can see that the MIP-

CASP approach is characterized by higher computational times, because of its trellis

search-based solution.

Remark. It is important to highlight the fact that CCM-SB beamforming requires

to equalize the received signal at the receiver side. This means that the BS is required

to feed-forward the m-th mobile user with the product of the channel with the m-th

precoding vector, i.e., hTm pm , ∀m, in order to recover the data. For the CI precoding, as

the received symbol resides in the constructive area of the constellation, there is no need

to equalize the composite channel. Accordingly, such feedback is not required by the

CI-exploiting approaches, where all complexity resides at the BS, which also makes them

robust to the estimation and quantization errors that are involved in the feed-forwarding

process for conventional beamformers.

106
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

4 1-step MFCAS 4-PSK

1-step MFCAS 8-PSK
3.5 2-step MFCAS-CBF 4-PSK
2-step MFCAS-CBF 8-PSK
3-step CBF-CAS-CBF 4-PSK
3 3-step CBF-CAS-CBF 8-PSK
MIP-CASP 4-PSK
Running Time [s] MIP-CASP 8-PSK
2.5 CCM-SB

1.5

0.5

0
50 100 150
N

Figure 5.4: Frame Running time when M = 5, NRF = 5 and TDL = 4.

5.6 Channel Uncertainty and Robust Approach

In this section, the effects of imperfect CSI acquisition at the transmitter are analyzed

and discussed. More specifically, we define the channel uncertainty model and derive a

robust precoding technique to counteract the errors caused by imperfect CSI.

5.6.1 Model and Effects

In the following study, imperfect CSI at the transmitter side is modeled by adding

a complex random component to the channel matrix H. Without loss of generality, we

consider the case where channel uncertainty amplitude is upper bounded by a specific

value, i.e., CSI at the BS is affected by spherical noise [84].

Accordingly, the estimated channel gain between the n-th antenna and the m-th user

is analytically defined as

ĥm,n = hm,n + em,n , ∀n ∈ {1, ..., N } , ∀m ∈ {1, ..., M } , (5.14)

where ĥm,n represents the channel gain estimation available at the BS and em,n repre-

107
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

PN 2 2 with δ
sents the channel uncertainty, i.e., n=1 |em,n | ≤ δm m being the uncertainty

upper bound over the channel estimation for the m-th user.

Clearly, the presence of uncertainty over the available CSI at the BS has negative

effects over the performances of a system. In fact, if the BS possesses imperfect CSI as

modeled, the received signal in a noise free scenario becomes

N
X N
X N
X
r̂m = ĥm,n xn = hm,n xn + em,n xn , (5.15)
n=1 n=1 n=1

where the second term of the last equation explicitly represents effects of imperfect CSI

at the transmitter side during signal processing.

In line with the standard approach [84,136], the BS is assumed to have no knowledge

over the channel uncertainty em,n , ∀n ∈ {1, ..., N } , m ∈ {1, ..., M } beside the upper

bound δm related to the m-th user channel. This is a common assumption in the

literature [84], and allows to derive a robust precoding design, which guarantees the

downlink transmission to be resistant against all possible channel uncertainties within

2 .
the upper bound δm

In the following, we derive a CSI-robust TAS-precoding technique based on a com-

mon approach in robust precoding design where the aim is to minimize the overall

transmitted power Pt required to respect the constraints imposed by the specific opti-

mization.

108
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

5.6.2 MIP-CASP Robust Scheme

Given the MIP-based optimization in (5.4), we can identify a worst-case design for

imperfect CSI scenarios as

PCASP R : minimize kx̃k2

a,x̃

< t̂ ◦ u∗ tan Φ − = t̂ ◦ u∗



subject to min  0,
kem k2 ≤δm
2 ,∀m
 
t̂ = Ĥx̃ − u ,

kx̃k2 ≤ 1, (5.16)

|x̃|  a,
N
P
an = NRF ,
n=1

an ∈ {0, 1} .

Because of the infinite number of possible error values em,n , the first constraint in

PCASP R is intractable. However, by employing a worst-case approach, it is possible to

derive a MIP-CASP robust design optimization for TAS-precoding with CI exploitation.

In order to do so, we need to identify the equivalent constraint for a worst-case scenario,

where the largest error is considered, as in Theorem 5.6.1.

Theorem 5.6.1 The worst-case equivalent to the optimization problem PCASP R can be

defined as
NRF
∗ kw1 k2
P
PCASP R : minimize
a,w1 ,w2 n

subject to Constraints (A.27) and (A.28)

n o
(1:N ) (N +1:2N )
w1 , w1  a,
n o (5.17)
(1:N ) (N +1:2N )
w2 , w2  a,

w1 = Πw2 ,
N
P
an = NRF , an ∈ {0, 1} ,
n=1

Proof of Theorem 5.6.1 The proof can be found in Appendix B.

109
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

5.7 Results

In this section the performances of the proposed transmission schemes are presented

and discussed. The shown results are evaluated through Monte Carlo simulations over

50000 channel realizations. In order to study the performances of the proposed schemes,

we evaluate the SER at the receiver side, the achievable capacity and the energy efficiency

of the system. More specifically, results are presented for both 4-PSK and 8-PSK, as the

proposed transmission schemes can be directly applied to any PSK modulation order.

Legends are characterized by the following notation:

• MIP-CASP identifies the CI exploitation transmission scheme based on MIP,

• 3-step CBF-CAS-CBF is used to represent the CI transmission based on the solu-

tion of PCASP 3 ,

• 2-step MFCAS-CBF represents the 2-step TAS-precoding heuristic scheme,

• 1-step MFCAS is used to classify the single-step approach,

• CCM-SB stands for the literature approach where TAS is performed by CCM and

precoding is performed through SINR balancing.

Moreover, the proposed schemes are compared with two low-complexity additional ap-

proaches from the literature: PGM-ZF, where ZF linear precoding is considered and TAS

is performed via PGM, and CIM-HY from the previous Chapter for the 4-PSK scenarios,

where hybrid linear precoding is considered (HY) and TAS is performed according to

CIM. In the simulations, a single-cell downlink M-MIMO scenario is considered, where

the BS possesses perfect CSI, employs a VLA of N = 100 antennas and communicates

with M = 5 single-antenna mobile users, unless differently specified.

5.7.1 Symbol Error Rate

Fig. 5.5 collects the SER of the proposed and conventional approaches for the case

of 4-PSK modulation. We can see that the proposed schemes greatly outperform all

110
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

100

10-1

10-2
SER

10-3
1-step MFCAS
10
-4 2-step MFCAS-CBF
3-step CBF-CAS-CBF
MIP-CASP
10-5 CCM-SB
PGM-ZF
CIM-HY NRF = 10
10-6
-10 -5 0 5 10 15 20
SN R [dB]

Figure 5.5: 4-PSK Symbol Error Rate when M = 5, N = 100 and NRF = 5 with
perfect CSI.

the benchmark techniques, including the CCM-SB scheme, which is characterized by a

combination of CCM TAS [133] and SINR-balancing beamforming at the transmitter

side [35]. At the same time, it is interesting to notice how both 2-step MFCAS-CBF

and 3-step CBF-CAS-CBF are able to achieve near optimal performances when com-

pared to the MIP-CASP approach. This is supported by the previous results in terms of

M, which showed how the two heuristic approaches were able to achieve similar perfor-

mances to the MIP-based scheme. On the other hand, we can see that 1-step MFCAS

obtains reasonable performance in the relatively low-to-mid SNR range, as the error-

floor of the MF is reached when SER is lower than 10−4 . This is due to 4-PSK wider

constructive interference regions, which allow a relative robustness against the inability

of MF precoding to efficiently separate the stream between the users. However, such

inability becomes the main cause for errors at high SNR and leads to the typical error-

floor. This confirms our previous considerations regarding 1-step MFCAS as a valuable

approach for the low-complexity and low-power scenarios.

In Fig. 5.6, the same set-up is explored for 8-PSK modulation. The performance

trends for the proposed techniques are preserved. In fact, all the schemes based on

111
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

100

10-1

10-2
SER

10-3
1-step MFCAS
2-step MFCAS-CBF
3-step CBF-CAS-CBF
10-4
MIP-CASP
CCM-SB
PGM-ZF
10-5
-10 -5 0 5 10 15 20
SN R [dB]

Figure 5.6: 8-PSK Symbol Error Rate when M = 5, N = 100 and NRF = 5 with
perfect CSI.

CI exploitation are able to outperform both the CCM-SB and the PGM-ZF schemes.

At the same time, it is important to highlight how the error floor for the 1-step MF-

CAS approach is higher than the one achieved in the 4-PSK case. This is due to the

fact the final closed-form MF precoding is not able to correctly separate the different

data-streams for different users, hence leading to uncontrolled inter-channel interference,

whose effects are more visible in higher modulation orders.

5.7.2 Data Rate

As the conventional CCM approach is designed for capacity maximization, it is

important to compare the rate performance of the proposed and conventional schemes.

Accordingly, the throughput of the MIP-CASP scheme is compared with the capacity

achievable when considering the CCM TAS scheme from the literature. The use of
P
throughput instead of the ergodic capacity, i.e., m log2 (1 + γm ), as a performance

metric for the proposed MIP-CASP is justified by the fact that its assumption of a

specific modulation, i.e., any PSK modulation order, does not allow to support the

assumption of Gaussian signals. The throughput is defined as [83] T = (1−BLER)·l·M ,

112
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

where BLER is the block error rate, l = log2 (L) is the bit information per symbol and

M is the number of users in the chosen scenario.

Performances are presented in Fig. 5.7, where the throughput of the proposed MIP-

CASP approach for increasing modulation order is compared with the capacity of the

full system and that of the CCM selection. The solid line with circular markers in the

figure represents the peak-throughput trend for the proposed approach. We can see

that the achieved throughput of the proposed scheme with increasing modulation orders

outperforms the CCM selection from the literature. It is important to notice that the

proposed scheme is able to achieve performances that are comparable to the ones of a

full-system for low-to-mid SNR scenarios, where the gap with the CCM scheme from

the literature is more pronounced.

45
CCM Selection
40 Full System
MIP-CASP Trend
Sum-Capacity [bit/channel use]

35 MIP-CASP Throughput

30
32-PSK
25
16-PSK
20

15
8-PSK
10
4-PSK
5
2-PSK

0
-20 -10 0 10 20
SN R [dB]

Figure 5.7: System Capacity comparison when M = 5, NRF = 5 and N = 100.

5.7.3 Energy Efficiency

In order to better highlight benefits and trade-offs brought by the proposed schemes,

we analyze the energy efficiency over throughput ηT , as defined in (2.22). In line with

Chapter 4, we consider realistic power values from practical systems [132], where Pamp =

113
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

30
1-step MFCAS
2-step MFCAS-CBF
25 3-step CBF-CAS-CBF
MIP-CASP
ηT [bit/channel use/W ]

CCM-SB
20 CIM-HY
PGM-ZF

15

10

0
5 10 15 20
Nt

Figure 5.8: 4-PSK Energy Efficiency ηT when M = 5, N = 100 with perfect CSI
and SN R = 5dB.

Pt /ν is defined as the power required by an amplifier with ν = 0.35 efficiency and

transmitted power Pt = 30dBm and PRF = 65.9mW . Performances are presented in

Fig. 5.8 and Fig. 5.9 as a function of NRF with SN R = 5dB and SN R = 10dB,

respectively. The proposed metric allows to better characterize the trade-off between

power consumption at the transmitter and achieved throughput as a function of the

variation over the subset size NRF .

As already observed in the SER results, we can see that performance trends for both

4-PSK and 8-PSK are preserved. More specifically, we can see that the proposed schemes

are all able to greatly outperform schemes from the literature for all the spectrum of

NRF values. At the same time, it is interesting to notice that the proposed schemes

achieve their maximum energy efficiency between NRF = 6 and NRF = 8 for both 4-PSK

and 8-PSK. This shows that systems with low numbers of active antennas can provide

reasonable performance with a very positive trade-off between hardware complexity

and power consumptions (i.e., when compared to the simplified chosen scenario where

NRF = M ). For a direct performance-complexity comparison between the schemes,

Table 5.1 collects the computational burdens required per frame to achieve the optimal

114
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

45
1-step MFCAS
40 2-step MFCAS-CBF
3-step CBF-CAS-CBF
35 MIP-CASP

ηT [bit/channel use/W ]
PGM-ZF
30 CCM-SB

25

20

15

10

0
5 10 15 20
Nt

Figure 5.9: 8-PSK Energy Efficiency ηT when M = 5, N = 100 with perfect CSI
and SN R = 10dB.

Table 5.1: Computational burdens for optimal Energy Efficiency points

Name Max ηT [bit/c.u./W ] Time [s] NRF

1-step MFCAS 24.78 0.44 7
2-step MFCAS-CBF 26.83 0.85 6
3-step CBF-CAS-CBF 27.59 1.31 6
MIP-CASP 28.38 2 5
CCM-SB 18.4 0.55 10

value of energy efficiency shown in Fig. 5.8. There are evident complexity savings

achieved by the heuristic schemes compared to the MIP approach, with little loss on

the maximum energy efficiency. On the other hand their complexity is comparable to

conventional CCM, with a more than 50% energy efficiency improvement and a ∼94%

reduction in the RF chains required to achieve maximum PE.

5.7.4 Robustness to CSI

In order to characterize the performances of the proposed CSI-robust scheme, we in-

troduce a conventional robust scheme from the literature, which will be used as a bench-

mark technique. In line with the previous approaches, when considering the benchmark

scheme we assume a cascade of TAS, based on capacity, followed by a SINR metric-

115
Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

2.4
MIP-CASPR
2.2 MIP-CASP
CCM-TPMR
2 CCM-TPM
Transmitted Power

1.8

1.6

1.4

1.2

0.8

0.6
10-2 10-1
2
δ

Figure 5.10: Transmitted power for 4-PSK transmission when N = 100, M = 5

and NRF = 5.

based precoding. More specifically, when considering robust precoding, it is common

to proceed to identify an optimization problem that aims to minimize the transmit-

ted power required to overcome the worst-case scenario. Accordingly, Transmit Power

Minimization Robust (TPMR) precoding can be defined as follows [136]

M
P
PT P M R : minimize tr [Pm ]
Pm ≥0,sm ≥0 m=1
 
Dm ĥTm Qm
subject to ≥0
 

Qm ĥ∗m Qm + sm IN (5.18)

Dm = ĥTm Qm ĥ∗m − γm N0 − sm δm
2

M
Qm = pm pH pi pH
P
m − γm i
i=1,i6=m

where the notation A ≥ 0 is used to impose that the matrix A is semidefinite positive.

In Fig. 5.10 we can see a comparison between the two robust schemes in terms of

transmitted power as a function of the error upper-bound δm = δ, ∀m. In addition to

the robust schemes, the minimum transmitted power for the non-robust approaches is

also presented. As shown, the proposed scheme is characterized by significantly lower

116
 Chapter 5. Large Scale Antenna Selection and Precoding for Interference Exploitation

requirements in terms of transmitted power, when compared to both the robust and

non-robust approach from the literature. Additionally, we can see that the CCM-TPMR

scheme from the literature is affected by a faster growth rate as the channel uncertainty

increases when compared to MIP-CASPR.

5.8 Conclusions

This chapter proved that antenna selection and precoding based on constructive

multiuser interference concepts can be jointly used to greatly improve the energy ef-

ficiency of future M-MIMO systems. Analytical and numerical studies showed that

constructive interference at the receiver side can be optimized by simultaneously iden-

tifying a subset of transmitting antennas and the precoded signal at the base station.

The presented schemes have been characterized by analyzing the computational costs

in terms of running time in comparison with state-of-the art algorithms. Performances

have been evaluated in terms of symbol error rate, sum rate and energy efficiency to

analyze the performance-complexity trade-offs introduced by the proposed scheme. Pro-

vided analyses and results have shown that the proposed approaches offer a favorable

performance-complexity trade-off compared to conventional schemes, with a close-to-

optimal performance.

117
118
Chapter 6

Beam Selection schemes for

Millimeter-Wave
Beamspace-MIMO Systems

Despite the great benefits introduced by the exploitation of mm-wave spectrum,

energy efficient transceiver design for such frequencies still represents a key barrier to

its implementation. In fact, a direct large or massive MIMO approach is considered

prohibitive because of the high transceiver complexity [23, 137] deriving from the use of

an extremely large number of antennas and RF chains in reduced physical spaces. In fact,

recent studies [138] proved that systems with massive antenna arrays are particularly

prone to RF chain imperfections, which lead to additional degradations in performance.

Moreover it is understood that RF components may consume up to 70% of the total

transceiver power consumption [139].

In order to exploit the favorable characteristics of mm-wave frequency communica-

tions, research is focusing on the development of new techniques that aim to reduce

the hardware complexity of very high dimensional MIMO systems. Previous works on

small-scale MIMO tackled the hardware complexity with antenna selection [15, 16, 20],

amongst many others, but showed high degradation in performances compared to the

full system. Nevertheless, since the high beamforming gains introduced by the use of

119
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

large number of antennas are fundamental to overcome the higher free space loss ex-

perienced in the mm-wave bands, such approaches are not viable at these frequencies

[127].

As a consequence, research is regarding hybrid analog-digital transceivers, where

analog beamformers in the RF domain are combined with a smaller number of digital

beamformers in baseband, as a promising candidate for future mm-wave MIMO appli-

cations. This chapter focuses on a scheme that combines B-MIMO concepts [13] with

DLA-based systems [14], where RF complexity reduction is approached by selecting a

subset of transmitting beams, instead of antennas.

The following list summarizes the contributions presented in the chapter:

1. Introduction of a mm-wave transmission scheme based on beam selection for B-

MIMO, able to achieve near-optimal performances with a reduced RF complexity

transceiver, and presentation of 3 associated beam selection schemes,

2. Analytical evaluation of the computational complexity of the proposed beam se-

lection schemes, with regards to conventional B-MIMO,

3. Analytical derivation of the capacity losses caused by beam selection, identifying

an upper bound for the proposed techniques,

4. Analysis of the performances achieved by the proposed transmission schemes in

terms of sum-rate and energy efficiency.

6.1 System Model

In this chapter, we consider a single-cell downlink scenario where the BS is equipped

with a DLA and a simple linear ZF precoder to communicate with M single-antenna

users [23], as in Fig. 6.1. DLAs can be analytically modeled as a critically sampled,
λ
i.e. d = 2 spaced, uniform linear array (ULA) of length L leading to a signal space

120
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

Figure 6.1: Block diagram of a mm-wave DLA-based transceiver scheme

dimension n that can be defined analytically as

2L
n= . (6.1)
λ

In DLA-based systems, the parameter n represents the maximum number of spatial

modes that are supported in transmission/reception, i.e. the total number of orthogonal

beams that are supported by the system[14].

Accordingly, the received symbol vector of a DLA-based linearly precoded transmis-

sion in the spatial domain can be expressed as

y = Hx + n = HGu + n, (6.2)

where H = [h1 , ..., hM ]T is a CM ×n matrix that collects the Cn×1 channel response

vectors h1 , .., hM of all the users, x is the Cn×1 transmitted signal, G is the Cn×M

linear precoding matrix, u is the CM ×1 vector that contains all the data symbols that

have to be transmitted and n is the CM ×1 additive white Gaussian noise (AWGN)

vector.

It is pivotal to highlight that while the spatial domain channel model H for DLA

can be analytically modeled and studied as a n sized ULA, the hardware complexity

required by these approaches are profoundly different. In fact, as shown in Fig. 6.1,

121
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

a BS with DLA requires only a reduced set of nRF RF chains and a beam selector to

support n narrow beams. A classical MIMO approach instead, where the BS is equipped

an equivalent n-dimensional linear array, requires one RF chain for each of the radiating

elements, regardless of the number of streams to be transmitted.

In line with Chapter 2, the channel response for the m-th user in the spatial dimen-

sion for a mm-wave communications MIMO system can be modeled according to Rician

fading, here briefly reported for the sake of ease of reading, as [10, 13]

Np
(m) (m) (m) (m)
X
hm = h(LoS)
m + h(M
m
P)
= β0 a(θ0 ) + βi a(θi ), (6.3)
i=1

where a(θ) is the Cn×1 steering vector for the spatial direction θ.

Note that the spatial domain model described here implicitly includes the effects of

transmit correlation. In fact, since the spatial domain for DLA is modeled as a critically

sampled ULA, different degrees of correlation can be achieved by varying the angle

spread [36].

However, a DLA-based system operates in the beamspace domain, defined by the

beamforming matrix U which represents the operation of a perfectly designed DLA

[14]. The beamforming matrix is obtained by computing the steering vectors for n fixed

spatial frequencies with uniform spacing [13, 140]. The beamforming matrix is defined

analytically as follows
1
U = √ [a (∆θ0 i)]i∈I(n) , (6.4)
n

where I(n) = {i − (n − 1)/2 : i = 0, 1, ...n − 1} is a symmetric set of indices centered

1
around 0, hence leading to a Cn×n matrix where ∆θ0 = n is the uniform spacing used.

We can define the equation (6.2) in the beamspace domain as follows

y = Hx + n (6.5)

where the channel matrix in the beamspace domain H and the transmitted signal in

122
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

the beamspace domain x can be computed as

H = HU H , x = U Gu (6.6)

It follows that the multiplication for the beamforming matrix U represents a mapping

of the signals for each mobile station in a new domain of orthogonal beams. In the

angular or beamspace domain, each column of the channel model H represents one

of the n beams supported by the DLA. Such relationship between the channel in the

angular-beamspace domain H and the channel in the spatial domain H is well known.

In fact, since the elements of U are in the form √1 e−j2πml/n , H represents the inverse
n

discrete Fourier transform of the channel matrix in the spatial domain H [10, 13].

Note that, given its analytical definition, the beamforming matrix defined in (6.4)

is unitary, i.e., U H U = U U H = I, hence leading to the following relationships between

the spatial and the beamspace domain

x = U H x,
(6.7)
H = HU .

6.1.1 Benchmark

As a consequence, the equivalent channel experienced by a DLA-based mm-wave

transceiver has a very sparse nature, where few elements of the matrix have dominant

values near the LoS direction of the users. Because of this, early works on DLA-based

beamspace mm-wave MIMO proposed a simple beam selection scheme called Maximum

Magnitude Beam Selection (MM-BS) [23], which proved to be able to take advantage of

channel sparsity. This technique, here briefly reviewed for the sake of completeness, is

based on the definition of a set of beam indices called sparsity masks.

Sparsity masks are used by the BS to identify the dominant beams to be selected

for the transmission and are defined as follows

 
(m) 2 (m) 2
M = i ∈ I(n) : |hm,i | ≥ ξ max |hm,i | (6.8)
i

123
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

[
M= M(m) (6.9)
m=1,..,M

where hm,i is the i-th element of the m-th user channel response, M(m) is the sparsity

mask for the m-th user and ξ (m) ∈ [0, 1] is the threshold used to define it. We can see

that in order to obtain a minimum number of beams for each user, the threshold ξ (m)

is chosen independently for each user.

After the sparsity mask, we can define the equivalent channel after a subset of beams

has been deactivated as

He = [hl ]l∈M (6.10)

where the sizes nd ×M of the new channel matrix He depend on the number of dominant

beams nd = |M| identified in the sparsity mask. From (6.9) we can see that the MM-BS

algorithm leads to values of nd which change according to the channel realization. In

fact, the user-wise selection implemented by MM-BS often leads to multiple selections

of the same beam for different users and therefore to a varying number of required RF

chains for different channel realizations and user topologies. As a consequence, a direct

application of MM-BS in practical systems, where the number of RF chains is fixed, is

not viable. Note that, while MM-BS leads to a variable number of dominant beams nd ,

it is still required to be lower or equal to the total number of available RF chains at the

transmitter side, i.e., nd ≤ nRF .

Since MM-BS selects the strongest channel paths, it can be seen that it is suboptimal

in terms of achievable capacity in a multi-path scenario. In fact, MM-BS performance

are strongly dependent on the assumption of a highly sparse channel, which is a valid

assumption for channels where the multi-path component of (6.3) is negligible, but

becomes questionable as we introduce additional paths to the model.

6.2 Proposed Beam Selection Techniques

With the aim to overcome the key drawbacks of MM-BS, i.e., the impossibility to

have a fixed value of nd and the strong dependence on the channel sparsity, this section

124
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

presents three different beam selection criteria. The use of DLA at the transmitter

allows to apply the selection algorithm directly over the channel matrix in the beamspace

domain, hence without affecting the beam-width nor the gain of the antenna pattern.

As shown in Chapter 2, selection can be performed according to different parameters,

such as path magnitude [23], SINR at the receiver [141], system capacity [21, 142] and

minimum error rate [142]. Main focus resides on the application of selection criteria

based on the first three metrics, since the proposed analysis mostly focuses on the

capacity of the system.

6.2.1 SINR Maximization Beam Selection (SM-BS)

In the proposed technique, beams are chosen to maximize the SINR at the user

side; this selection criterion is defined as SINR maximization beam selection (SM-BS).

In order to identify the subset of beams used during data transmission, we need to

define the SINR metric for the chosen model. The SINR for each user depends on

the precoder used at the transmitter, identified by the precoding matrix in the angular

domain G = γF . More specifically, the received SINR of the m-th user is defined as [23]

ρ|γ|2 T 2
M |hm fm |
SIN Rm (ρ, G|H) = ρ|γ|2
(6.11)
|hTk fm |2 +
P
M N0
k6=m

where hTm is the transpose of the m-th user channel response, fm is the m-th column of

F , ρ is the transmitted power and N0 is the noise power.

In this study, main focus resides on a practical case where the BS is equipped with

a low-complexity ZF linear precoder, hence characterized by two important properties:

null interference, i.e., k6=m |hTk fm |2 = 0 and unitary gain, i.e., |hTm fm |2 = 1. Thanks to
P

the properties of ZF precoding, the received SINR equation can be simplified to [36]

ν|γ|2
SIN Rm,ZF (ν, G|H) = (6.12)
M

where ν = ρ/N0 is the SNR. Accordingly, under the assumption of a ZF precoded BS, the

SINR maximization selection at the transmitter can be directly achieved by maximizing

125
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

the scaling factor γ.

A direct application of SM-BS can be achieved by performing an exhaustive search

of the SINR for all the possible combination of beam subsets and then choose the subset

that leads to the highest value. While such approach leads to an optimal selection,
n


it rapidly becomes computationally prohibitive because of its nRF possible combina-

tions1 , where nRF is the subset size. Accordingly, we derive a suboptimal decremental

beam selection that identifies the beam subset with the minimum loss in terms of SINR,

shown in Algorithm 6.1. Using (6.12), the SINR for the reduced system after the elimi-

nation of the l-th beam can be computed as

(l)
ν|γ (l) |2

SIN Rm,ZF ν, G|H(l) = (6.13)
M

with
s
(l) ρ
γ = (6.14)
tr(F (l) F H
(l) )

where H(l) represents the channel matrix whose l-th beam has been eliminated, F (l) is

the precoding matrix obtained with the lower-dimensional channel matrix H(l) and γ (l)

is the corresponding scaling factor2 . Hence, we can identify the index of the beam to be

disabled via the following maximization criterion

( )
ν|γ (l) |2
δ = arg max (6.15)
l M

where δ is an element of the subset of disabled beams D. Since ρ, M and N0 are channel

independent, the maximization criterion can be simplified to

δ = arg max(|γ (l) |2 ). (6.16)

l

1
A scenario where n = 81 and nRF = 40 leads to 81 ≈ 2 · 1023 possible subsets, which is computa-


40
tionally prohibitive for a simulation evaluated study. However, previous works on antenna selection for
low dimensional systems [117] showed that the performances of decremental approaches are close to the
ones achieved by exhaustive search methods.
2
Note that SM-BS does not affect the transmitted power constraint E[xxH ] = 1. In fact, the system
deriving from the selection employs a ZF precoder, which is computed according to the low dimensional
channel matrix He obtained through SM-BS and uses a scaling factor to constrain the average transmit
power.

126
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

Algorithm 6.1 SM-BS

Input: H
Output: He

• C := H

• F := C(CH C)−1

• for j = 1 → n − nRF

– for l = 1 → n − j
∗ F (l) = C(l) (CH −1
(l) C(l) )
q
∗ γ (l) = ρ/tr(F (l) F H (l) )

– end
– δj = arg max |γ (l) |2

l
– D = {δ1 , ..., δj }
– C = [hl ]l∈D
/

• end

• He = [hl ]l∈D
/

While the selection metric for SM-BS derived in (6.16) is obtained by exploiting the

orthogonal properties of ZF precoding, the presented technique can be applied indepen-

dently from the precoding involved at the BS. In fact, following the notation used in

(6.11) and under a generic precoding assumption G, the SM-BS algorithm proceeds to

maximize the SINR for the reduced system

(l)


SIN Rm = SIN Rm ρ, G (l) |H(l) (6.17)

where G (l) represents the precoding matrix that corresponds to the reduced channel

model H(l) .

6.2.2 Capacity Maximization Beam Selection (CM-BS)

We define as Capacity Maximization Beam Selection (CM-BS) the algorithms that

aim to identify the beam subset with the minimum capacity loss from the full system

[21]. The CM-BS can be performed with two separate approaches:

127
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

• Decremental, where the algorithm recursively chooses the beams not to be used in

transmission,

• Incremental, where the algorithm recursively chooses the beams to be used in

transmission.

It is immediate to see that the difference between the two algorithms resides in the

computational costs [21]. In fact, the incremental selection is faster when the number

of beams to be included in the transmitting subset is lower than n/2, i.e., nRF ≤ n/2,

while the decremental is to be preferred when the number of beams to be included in

the transmitting subset is higher than n/2, i.e., nRF ≥ n/2.

6.2.2.1 Decremental CM-BS (DCM-BS)

The algorithm selects the beams whose elimination causes the minimum loss in terms

of capacity. Given the full system capacity

C(H) = log2 det(I + νHHH ). (6.18)

the capacity after the l-th beam has been disabled can be computed as [143]

C(H(l) ) = log2 det(I + νH(l) HH

(l) ) (6.19)

where the channel H(l) is related to the full system matrix according to the following

equation

H(l) HH H H
(l) = HH − hl hl . (6.20)

Substituting (6.20) in (6.19), we can show the relationship in terms of capacity

between the two channels

C(H(l) ) = log2 det(I + νHHH − νhl hH

l ) (6.21)

128
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

which can be rearranged to

1 1
C(Hl ) = log2 det(I+νHHH )+log2 det(I−(I+νHHH )− 2 νhl hH H −2
l (I+νHH ) ) (6.22)

hence leading to [21]

H −1
C(H(l) ) = C(H) + log2 1 − νhH
 
l (I + νHH ) hl . (6.23)

In particular (6.23) shows the relationship in terms of capacity between the full

system and the system where a beam has been disabled. Here, it is clear that the

second term on the right-hand side of the equation represents an analytical evaluation

of the capacity loss caused by the l-th beam deactivation. Accordingly, the selection

criterion for capacity maximization can be performed by minimizing such parameter

and is described analytically as

H −1
δ = arg min hH

l (I + νHH ) hl . (6.24)
l

Under the assumption of a fixed number of beams selected, the algorithm has to

compute all the others n − nRF different beams to eliminate. The above selection is

implemented using Algorithm 6.2.

6.2.2.2 Incremental CM-BS (ICM-BS)

The algorithm incrementally selects the beams with the highest capacity contribu-

tion. Using a similar notation as the previous one for the ICM-BS, it is possible to show

how the capacity is affected when a new beam is added to a low-dimensional channel

matrix as [21]
h  i
H
C H̄, hl = log2 det I+ν H̄H̄ + hl hH


l (6.25)

where H̄ represents the channel matrix formed by previously chosen beams and hl is the

newly added beam. The equation (6.25) can be expressed as a function of the system

129
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

Algorithm 6.2 Decremental CM-BS

Input: H , ν
Output: He

• K := H

−1
• B := I + νKKH

• for j = 1 → n − nRF

– for l = 1 → n − j
∗ Ω(l) = kH
l Bkl
– end
– δj = arg min{Ω(l) }
l
– D = {δ1 , δ2 , ..., δj }
 −1
– B := B + Bkδj ν −1 − kH δj Bk δj kH
δj B

– K := [hl ]l∈D
/

• end

• He = [hl ]l∈D
/

channel H̄ via the same procedure used for the DCM-BS, leading to

 
H −1
   
H
νhH


C H̄, hl = log2 det I + ν H̄H̄ + log2 1 + l I + ν H̄H̄ hl (6.26)

where the second term on the right-hand side of the equation represents the capacity

contribution of the l-th beam and needs to be maximized with an exhaustive search

through all the available beams. Accordingly, the selection criterion can be analytically

expressed as
  
H −1

H −1
 = arg max hl ν I + H̄H̄ hl (6.27)
l∈E
/

where E represents the subset of enabled beams . This selection technique is presented

analytically in Algorithm 6.3, where it uses a recursive update on the matrix inversion,

based on the Sherman-Morrison-Woodbury Identity [144], see Appendix C.

130
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

Algorithm 6.3 Incremental CM-BS

Input: H , ν
Output: He

• K := H

• A := νI

• 1 := arg max kkl k2

l

• for j = 1 → N − 1
 −1
– A := A − Akj 1 + kH
j Akj kH
j A

– for l = 1 → n − j
∗ Ω(l) = kH
l Akl
– end
– j+1 = arg max{Ω(l) }
l
– E = {1 , 2 , ...., j , j+1 }

• end

• He = [hl ]l∈E

6.3 Computational Complexity Analysis

For the sake of a complete and fair comparison, this section evaluates and studies

the computational complexity of each of the proposed algorithms. The computational

complexity counts are listed in the Table 6.1 for all the algorithms, where the orders of

magnitude of each operation are evaluated.

It is important to emphasize the distinction between digital signal processing (DSP)

complexity, which is the focus of this section, and RF chain complexity. In fact DSP

complexity involves the processor at the transmitter and its impact in power consump-

tion is of the order of 5.76mW/KOps − 22.88mW/KOps as for the Virtex family from

Xilinx [145], where values are expressed in watts per 103 operations. RF complexity,

instead, derives by the number of chains used in the transmission. Each chain is charac-

terized by a high number of elements, such as mixer, digital-analogic converter (DAC)

and filters, whose values of power consumption are particularly significant. Typical val-

ues of power consumption for a single RF chain are of the order of ∼ 30mW as in [139],

131
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

leading to power consumptions in the order of watts, when the amplifier is included in

the model.

The first column represents the MM-BS criterion as a reference, while the other

columns collect the analysis of SM-BS, DCM-BS and ICM-BS respectively. In particular,

nb identifies the number of beams chosen per user by MM-BS and ndel = n − nRF

represents the number of beams to be deactivated in decremental selections. We focus

our analysis on the application of the algorithms within a channel realization and on

the operations that dominate the complexity. Complexity order for each operation are

considered in line with the literature [10, 146].

The table shows that the DSP complexity for the MM-BS is lower than the other

algorithms, as the selection is based only on the amplitude of the paths. As a conse-

quence, the beam selection algorithms presented in the other columns are affected by

higher computational complexity. The higher costs are due to the necessity to compute

additional elements, such as γ for the SM-BS or A and B for the DCM-BS and ICM-BS

respectively. In order to highlight the differences between the DCM-BS and ICM-BS,

the constant terms in the complexity computations are preserved in the computational

count. Thanks to this notation, it is possible to confirm that DCM-BS is more efficient

when ndel ≤ n/2 while ICM-BS has to be chosen when nRF ≤ n/2. Since differences in

performances are negligible, the results obtained by these techniques will be addressed

as CM-BS from now on, without differentiating between incremental or decremental

approach.

Finally, it is worth noticing that, even though the MM-BS has a lower computational

time, it is affected by strong losses in performances in a realistic MP environment, as

shown in the results that follow. This consideration makes the presented algorithms rele-

vant in realistic applications, thanks to their appealing trade-off between computational

costs and performances.

132
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

MM-BS No. SM-BS No.

H ◦ H∗ O(M n) F = CH (CCH )−1 O(M n2 )
find O(nb M n) γ (ndel ) ndel O(n3 )
find (ndel ) ndel O(n)
F (ndel ) ndel O(M n2 )
Total O(M n)+ Total ndel O(n3 )+
nb O(M n) (1 + ndel )O(M n2 )
+ndel O(n)
DCM-BS No. ICM-BS No.
B O(n3 ) find O(n)
kHl Bkl (ndel ) ndel O(n2 ) A O(n3 )
find (ndel ) ndel O(n) A(nRF ) 3nRF O(n2 )
B (ndel ) 3ndel O(n2 ) kHl Akl (nRF ) nRF O(n2 )
find (nRF ) nRF O(n)
Total O(n3 )+ Total O(n3 )+
4ndel O(n2 ) 4nRF O(n2 )+
+ndel O(n) (nRF + 1)O(n)

Table 6.1: Complexity in number of operations

6.4 Performance Analysis - Capacity Loss

This section is dedicated to the analysis of the capacity losses caused by the selection

of a beam subset over the full system, providing an analytical study of the performances

achieved by the proposed algorithms. The search of the best trade-off is a critical

element of the system design, as the selection of a beam subset benefits from a hardware

complexity simplification while suffering a degradation in performances.

The capacity achievable by a multiuser system can be defined as

M
X
C= log2 (1 + SIN Rm ). (6.28)
m=1

As shown, under the assumption of ZF linear precoding, the received SINR in a

multiuser scenario depends only on the scaling factor γ, and can be computed as [36]

νρ
SIN Rm,ZF = . (6.29)
M · tr[(HH H)−1 ]

Clearly, in this specific scenario, (6.29) leads to the same value of SINR for all the

133
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

users, hence allowing to simplify the capacity evaluation to

C = M log2 (1 + SIN RZF ). (6.30)

Capacity losses caused by the elimination of one beam can be defined as the difference

between the performances achieved by the full system and by the system where one beam

is eliminated as

(l)
Ψ(l) , M log2 (1 + SIN RZF ) − M log2 (1 + SIN RZF )
    (6.31)
νρ/M νρ/M
, M log2 1 + tr[(HH
H)−1 ]
− M log 2 1 + tr[(HH H )−1 ] l l

(l)
where SIN RZF represents the SINR for the system without the l-th beam.

The equation (6.31) is particularly useful to study SM-BS optimality. In fact, the

best trade-off between performances and hardware complexity corresponds to the case

where capacity losses caused by the selection are minimized. Here, the first term in

(6.31) does not depend on the selection because it represents the full system, while the

second term depends on the criterion used to identify the l-th beam. Accordingly, the

minimum loss Ψ(l) is obtained when the second term is maximized.

In particular, the second term of the equation can be rearranged by using the matrix

properties showed in (6.20) and (A.32) as

 
ν/M
M log2 1 + (6.32)
tr[R + R hl (1 − hH
−1 −1 −1 −1 H −1
l R hl ) hl R ]

where R = HH H.

Hence, thanks to the properties of logarithms, equation (6.31) can be rearranged as

   
ν/M
1+ tr(R−1 )
Ψ(l) = M log2    . (6.33)
ν/M
1+ tr[R−1 +R−1 hl (1−hH −1 h )−1 hH R−1 ]
l R l l

134
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

With some straightforward algebra, equation (6.33) can be simplified to the form

Ψ(l) = M log2 (1 + ι) (6.34)

where the parameter ι in the argument of the logarithm of (6.34) is

ν
M tr(Rhl (1 − hH H
l Shl )hl S)
ι= (6.35)
tr(S)2 + ν
M tr(S) + tr(S)tr(Shl (1 − hH −1 H
l Shl ) hl S)

where S = R−1 .

Results obtained in (6.34) (6.35) can be generalized to identify the global loss caused

by the selection of a subset of beams as

!
ν
M tr(TD )
Ψ = M log2 1+ (6.36)
tr(R−1 2
E ) +
ν −1 −1
M tr(RE ) + tr(RE )tr(TD )

−1 H −1 −1
where RE = HH H
E HE , TD = RE HD (I − HD RE HD ) HD RE and the subindices E

and D represent the enabled and disabled subset of beams respectively. In particular

HD = [hk ]k∈D and HE = [hk ]k∈E .

Hence the loss Ψ is a function of ν and approaches an upper bound [117] as ν → ∞,

defined in the following equation

!
tr(TD )
Ψ ≤ M log2 1+ . (6.37)
tr(R−1
E )

The analytical results of the loss are confirmed by simulations (n = 81 and M = 40

system), as shown in Fig. 6.2 for a MP scenario. In particular, the upperbounds derived

through (6.37) are indicated in the legend as CM-BS nRF = 40 Analytical and SM-BS

nRF = 40 Analytical for the selection over capacity and SINR respectively, while the

numerical values are addressed as Simulated for all the selection techniques.

The MM-BS criterion is characterized by fluctuations in the size of the beam subset,

leading to losses that are not limited by the upperbound defined in (6.37).

135
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

70

Performance loss Ψ [bits/s/Hz] 60

50

40

30

20 SM-BS nRF = 40 Analytical

CM-BS nRF = 40 Analytical
10 SM-BS nRF = 40 Simulated
CM-BS nRF = 40 Simulated
MM-BS 2-beam Simulated
0
-10 0 10 20 30 40 50
SN R [dB]

Figure 6.2: Comparison analytical and simulated Ψ for a n = 81 and M = 40

system

6.5 Numerical Results

This section presents the numerical results obtained through Monte Carlo simu-

lations over 10000 channel realizations. In the simulations, the BS is assumed to be

equipped with an DLA with n = 81 available beams and to communicate with M = 40

single-antenna users, in line with [23].

Two different channel scenarios are considered, both where perfect channel state

information is available at the transmitter3 : one with only the LOS component (2.37)

in accordance to [23] and one with the additional MP components as in (2.38) with

NP = 2. Complex path gains are defined as

βi = |βi | e−jψi (6.38)

3
Perfect CSI is a common assumption in the literature [15] for systems that involve AS at the
transmitter. Due to the sizes of the systems involved in mm-wave communications, the acquisition
of channel state information represents a critical step. Recent works on M-MIMO approached the
problem, with the aim to reduce the signal processing complexity [147] or the time [148] required for
CSI acquisition. In the chosen scenario, to retain the benefit of reduced RF-chain operation, a trivial
approach for the CSI provisioning would be by scheduled acquisition and RF switching, although more
sophisticated and efficient approaches can be found in the literature.

136
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

where the MP component (i 6= 0) is set |βi |2 = −10dB with ψi being uniformly

distributed between 0 and 2π and the LoS component (i = 0) is characterized by

|β0 |2 = 0dB and ψ0 = 0. In order to have a simplified definition for the angles of arrival

in the MP component, we consider a scenario where the distance between BS and users

is wide enough so that |θm | is uniformly distributed between ∆θ − ∆θ ∆θ

 
4 , ∆θ + 4 and

sign (θ) is chosen randomly.

The algorithms are applied with two different approaches: one where the beam

subset size is fixed to nRF and one where beam selection is recursively applied in order

to capture a certain percentage η of the total channel power σc2 . Accordingly, in the

figures, the following notation is used: Full System to denote the performances obtained

by the scheme without beam selection, MM-BS 2-beam to identify the performances

obtained by magnitude selection with nb = 2 in accordance to [23], η = 95% to classify

the approach where the percentage of channel power captured by the subset of beams

is fixed, nRF = 40 and nRF = M to address the approach where the maximum number

of beams is fixed at 40 or at the number of users M , respectively.

It is worth to notice that in practical systems the number of RF chains is generally

fixed, making the MM-BS of [23] inapplicable and the nRF = {40, M } approaches shown

here particularly relevant in realistic scenarios.

For the η = 95% approach, given the total channel power σc2 , we can compute the

captured channel power ratio η as

tr He HH tr He HH



e e
η= =
. (6.39)
σc2 tr HHH

6.5.1 Capacity

This section evaluates the capacity achieved by the proposed selection algorithms

when the transmitter is equipped with a ZF precoder. The proposed schemes are com-

pared to the full system, i.e., where no selection is performed and nRF = n, and to the

MM-BS criterion from the literature [23].

137
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

Using the previous definition of SINR in (6.11), the capacity can be computed as

M
X
C(SIN R, G|H) = log2 (1 + SIN Ri (γ, G|H)) (6.40)
i=1

which considers the full channel model H, but it can be directly applied to the low-

dimensional equivalent channel by replacing H with He .

In Fig. 6.3 and Fig. 6.4 we can see the capacity as a function of the SNR for both

fixed and variable beam subset size, i.e., nRF = 40 and η = 95% in the legend, and both

proposed algorithms,i.e., SM-BS and CM-BS in the legend. Since both incremental and

decremental CM-BS achieve the same performances, DCM-BS has been used to obtain

the shown results. When comparing the performances of CM-BS and SM-BS, we can

see a gap in the low SNR region. This gap, beneficial for SM-BS, is justified by the

different metrics used by the two algorithms. In fact, whilst CM-BS does not consider

the precoding involved at the BS, the SM-BS algorithm maximizes the SINR at the

receiver for the particular ZF precoding used, by maximizing the scaling factor γ. The

impact of this difference over the received SINR is described analytically in (6.11), where

the noise component of the denominator is inversely proportional to the scaling factor4 .

Beam selection losses in pure line-of-sight scenarios, i.e., Np = 0, are almost neg-

ligible, as we can see in Fig. 6.3. In particular the figure shows that the nRF = 40

approach is characterized by higher performance degradations in the low-SNR region,

but it rapidly achieves similar or better performances than the MM-BS for SN R ≥ 10dB

for both CM-BS and SM-BS. On the other hand, beam selection algorithms with a chan-

nel power-based approach show a significant performance increase for both CM-BS and

SM-BS, with SM-BS nearly achieving full system optimal performances.

Fig. 6.4 shows the spectral efficiencies obtained in the MP environment. For the

nRF = 40 case we see that the CM-BS algorithm performs well as the SNR grows, while

the SM-BS algorithm outperforms the previous MM-BS approach even in low SNR re-

gions. When the beam subset size is defined according to the η = 95% approach instead,

4
Hence, the gap between the two techniques is wider in the low-SNR region, since the dependence of
the SINR over the scaling factor is more visible at high values of noise power.

138
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

102
Spectral efficiency [bits/s/Hz]

101
Full System
MM-BS 2-beam
CM-BS nRF = 40
SM-BS nRF = 40
CM-BS η = 95%
SM-BS η = 95%
100
-5 0 5 10 15 20 25 30
SN R [dB]

Figure 6.3: Capacity as a function of SNR[dB] in a line-of-sight scenario

Spectral efficiency [bits/s/Hz]

102

101 Full System

MM-BS 2-beam
CM-BS nRF = 40
SM-BS nRF = 40
CM-BS η = 95%
SM-BS η = 95%
0
10
-10 0 10 20 30
SN R [dB]

Figure 6.4: Capacity as a function of SNR[dB] in a multi-path scenario

139
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

we can see that both algorithms are able to achieve near optimal performances. In par-

ticular the SM-BS algorithm performs very closely to the full-system with a negligible

degradation.

It is important to notice that the improvement in performances achieved by the

η = 95% approach over the nRF = 40 scheme resides in a larger beam subset size, as

shown in the following section, which focuses on this aspect.

6.5.2 Mean number of beams

In this section we study the mean number of selected beams, i.e., the mean beam

subset size, for both CM-BS and SM-BS as a function of the number of users M in

the system. Beam usage represents a fundamental parameter for DLA-based schemes

study, as it provides a direct evaluation of the RF complexity reduction achieved by

beam selection at the transmitter. Clearly, the nRF = M scenario is characterized by

a number of beams that is a linear function of the number of users M . This holds for

both CM-BS and SM-BS and leads to matching results, presented in Fig. 6.5 and Fig.

6.6 with CM-BS nRF = M only.

In Fig. 6.5 we can see that MM-BS 2-beam selects a number of beams that grows

constantly and rapidly with the number of users in the scenario. In particular, it shows

that the nRF = M approach achieves higher RF-complexity reductions than the other

approaches in the region 20 ≤ M ≤ 46. Additionally, for the η = 95% approach, CM-BS

is characterized by a number of beams which is always lower than the MM-BS, except

for M = 20, while SM-BS is affected by larger beam subset sizes than MM-BS but still

provides an interesting complexity reduction.

In a more realistic multi-path scenario, see Fig. 6.6, the MM-BS technique uses a

higher number of beams than the nRF = M approach for both CM-BS and SM-BS until

the number of users M ≤ 50. In particular, it is interesting to notice how this further

complexity reduction is accomplished, while still providing higher spectral efficiencies.

When CM-BS is applied with the power-based η = 95% approach, gains in terms of

beam subset size are more visible in highly populated scenarios, i.e., where M ≥ 50,

140
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

90

80

Mean number of selected beams

70

60

50

40

30 Full System
MM-BS 2-beam
20
CM-BS nRF = M
10 CM-BS η = 95%
SM-BS η = 95%
0
10 20 30 40 50 60
M

Figure 6.5: Mean number of beams (RF chains) used for transmission as a func-
tion of the number of users in a line-of-sight scenario

while the SM-BS is always characterized by larger subsets than MM-BS. For the case

with M = 40 users, we can see that CM-BS selects in average only ∼ 2 more beams

than MM-BS and yet provides considerable benefits in terms of capacity, leading to a

very advantageous trade-off. Finally, while power-based SM-BS is affected by a larger

subset size than MM-BS or nRF = M approaches, it is able to provide near-optimal

performances and significant complexity reduction when compared to the full system.

While both Fig. 6.5 and Fig. 6.6 are achieved for a SN R = 15dB scenario, it is

important to highlight that the effects of this assumption over the number of selected

beams are negligible. In fact, selections with a fixed number of beams, MM-BS and

power-based SM-BS are independent from the SNR, while CM-BS with the channel

power approach showed imperceptible differences in a low SNR scenario.

6.5.3 Energy Efficiency

In order to better characterize the trade-off introduced by the proposed schemes,

we compare the achievable energy efficiency of each scheme, according to the definition

in (2.20). Practical values are assumed for PRF = 34.4mW [45], accounting for mixer,

141
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

90

80
Mean number of selected beams
70

60

50

40

30
Full System
20 MM-BS 2-beam
CM-BS nRF = M
10 CM-BS η = 95%
SM-BS η = 95%
0
10 20 30 40 50 60
M

Figure 6.6: Mean number of beams (RF chains) used for transmission as a func-
tion of the number of users in a multi-path scenario with Np = 2

DAC and filters, and Pt = 15dBm [149], to model a small cell transmission. This metric

is particularly useful for a direct evaluation of the joint effects caused by beam selection

over the power needed by the system PBS , and over the average capacity of the system

C.

Fig. 6.7 illustrates how the nRF = M approach, with both the CM-BS and SM-

BS, outperforms greatly all the others until the number of users M ≤ 31. This is

due to the fact that the number of beams, i.e., RF chains, is much smaller than other

approaches, yielding a great reduction in power consumption. Power-based η = 95%

approaches are characterized by lower energy efficiencies in low populated scenarios, due

to the independence of the selection criterion from the number of users. However, as

the number of users grows to M ≥ 35, power-based CM-BS is able to outperform all the

other techniques. In addition, CM-BS η = 95% shows an interesting behavior in highly

populated scenarios, i.e. M > 50, where the values of energy efficiency start to increase.

This behavior is caused by a phenomenon visible in Fig. 6.5, where the beam subset size

of CM-BS η = 95% is lower than all the other approaches when M > 50, even lower than

the number of users in the scenario. Consequently, the increasing capacity, combined

142
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

180
Full System 20

160 MM-BS 2-beam

SM-BS nRF = M 15

Energy Efficiency [bits/s/Hz/W]

140 CM-BS nRF = M
CM-BS η = 95% 10
120 SM-BS η = 95%
5
100

80 0
35 40 45 50
60

40

20

0
10 20 30 40 50 60
M

Figure 6.7: Energy efficiency as a function of M in a line-of-sight scenario with

Pt = 15dBm

with lower power requirements, leads to an increase in terms of energy efficiency. Finally,

for SM-BS η = 95%, we can see that such approach is characterized by higher energy

efficiency values than the MM-BS algorithm in M ≥ 36 scenarios. This is due to the

fact that the effects of a larger beam subset sizes are mitigated by near optimal capacity

values.

In Fig. 6.8 we can see that the nRF = 40 approaches are still preferable for systems

with reduced populations. In particular, we can see that the both algorithms lead to

similar energy efficiency values, with CM-BS being gradually outperformed by SM-BS

as the number of users grows. This is due to the fact that the interference among

users, which is optimized by SM-BS, becomes more relevant as the scenario gets more

populated. When considering the η = 95% approach, we can see that CM-BS starts

outperforming all other schemes as the number of users increases to M ≈ 35. This

is caused by the high capacity values CM-BS η = 95% leads to. Finally, it is worth

noticing how the values of energy efficiency obtained by the MM-BS criterion rapidly

decrease as M grows, because of the lower capacities obtained with such approach when

the effects of scattering and multi-path increase.

143
Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

180
Full System 40

160 MM-BS 2-beam

Energy Efficiency [bits/s/Hz/W] SM-BS nRF = M 30
140 CM-BS nRF = M
CM-BS η = 95% 20
120 SM-BS η = 95%
10
100

80 0
35 40 45 50

60

40

20

0
10 20 30 40 50 60
M

Figure 6.8: Energy efficiency as a function of M in a multi-path scenario with

Np = 2 and Pt = 15dBm

While Fig. 6.7 and Fig. 6.8 collect energy efficiency performances for the high SNR

regime, i.e., where SN R = 15dB, we can infer that the benefits of the proposed schemes

extend to low SNR scenarios. In fact, the capacity performance gap with the full system

is narrow and hardware complexity savings are not affected by the SN R. Moreover,

as shown in Fig. 6.2, the losses of the proposed techniques in comparison with the

full system decrease together with the SNR. As a consequence, this suggests that the

performance trends in terms of energy efficiency are not affected by the SNR.

In Fig. 6.9, energy efficiency is computed as a function of the transmitted power.

In particular, we consider a scenario where M = 40 and SN R = 15dB. As in the

previous figures, we can see that the CM-BS η = 95% is able to outperform all other

schemes. Clearly, energy efficiency performances decrease as transmitted power grows,

since the RF-chain term in (2.20) becomes less relevant. Additionally, we can see that

MM-BS performs poorly in comparison with the proposed schemes because of the low

values of capacity and larger beam subset size. The nRF = 40 strategy provides better

performances than the full system in the region where Pt ≤ 26dBm for CM-BS, and in

the region where Pt ≤ 34dBm for SM-BS.

144
 Chapter 6. Beam Selection schemes for Millimeter-Wave Beamspace-MIMO Systems

25

Energy Efficiency [bits/s/Hz/W]

20

15

10
Full System
MM-BS 2-beam
CM-BS nRF = M
5
SM-BS nRF = M
CM-BS η = 95%
SM-BS η = 95%
0
5 10 15 20 25 30 35 40
Pt [dBm]

Figure 6.9: Energy efficiency as a function of Pt

6.6 Conclusions

This chapter introduced several beam selection techniques for DLA-based BS-MIMO,

which allow to reduce the RF complexity of mm-wave transmitters while achieving

near-optimal capacity performances, in both line-of-sight and multi-path environments.

Shown results prove that the proposed transmission schemes are particularly promising

for realistic implementations of mm-wave systems, where RF chain costs and power

consumption are of crucial importance.

In particular, analytical and numerical results show that beam selection algo-

rithms allow to achieve higher energy efficiencies than a full system, while reducing

the transceiver RF complexity according to the number of users. It also demonstrated

that beam selection algorithms with a channel power approach can lead to near-optimal

performances in both line-of-sight and multi-path scenarios, while still achieving signif-

icant RF complexity reductions.

145
146
Chapter 7

Conclusions

This thesis showed how energy efficiency represents a fundamental evaluation metric

for realistic implementation of future large-scale wireless communications systems. Since

the employment of VLAs at the BS is envisaged as the necessary approach to satisfy

the ever-growing spectral efficiency requirements, in this thesis we have discussed and

proposed several transceiving schemes that aim at increasing the energy efficiency of M-

MIMO in the downlink multiuser scenario for both the mm-wave and in the Ultra-High

Frequency spectrum.

7.1 Summary and Conclusions of the Thesis

In this thesis, Chapter 2 presented a general overview of energy-efficient multiple

antenna systems. Conventional precoding and detection schemes were first introduced

for small-scale multiuser MIMO, followed by a description of general strategies for re-

ducing the hardware complexity of multiple antenna systems. After highlighting the

main shortcomings of these techniques, we proceeded to describe the main research

contributions included in this thesis. More specifically:

• In Chapter 3, we aimed to increase the energy efficiency of large-scale MIMO by

employing Constant Envelope Precoding in the multiuser downlink transmission.

The combination of Constant Envelope Precoding with Constructive Interference

147
Chapter 7. Conclusions

concepts was proven to be able to greatly outperform classical interference reduc-

tion approaches in terms of symbol error rate with gains in the order of 5 dB, when

the desired constellation energy is considered unitary. In addition, a relaxation of

the optimization region for the case of imperfect Channel-State Information was

analytically derived, showing that a robust formulation of the proposed problem

allows to increase the performances of its non-robust counterpart without the need

to increase the transmitted power. The key contributions that can be identified in

this chapter are:

C3.1. Constructive Interference can now be applied to Constant Envelope Pre-

coding systems to further increase the power efficiency in large-scale MIMO

communication systems. This allows to seamlessly optimize the transmitted

constellation energy, without the need to solve additional optimization prob-

lems. The newly formulated non-convex optimization problem was solved

via Cross-Entropy Optimization and a low-complexity Convex Optimization

approach, based on an initial relaxation of the power constraints.

C3.2. The performance enhancements shown by the robust formulation of the

proposed scheme proved that a relaxation on constructive interference con-

straints allows to achieve robustness against channel uncertainties, especially

in the low-SNR region. However, excessive relaxation can lead to perfor-

mance worsening in the high-SNR region.

• Low-Complexity Constructive Interference metrics were exploited in Chapter 4 for

defining a highly power-efficient Transmit Antenna Selection algorithm in the mas-

sive multiuser downlink. The proposed algorithm demonstrated that modulation-

specific metrics can be efficiently employed to achieve significant benefits while

greatly reducing the hardware complexity of massive MIMO schemes. The pro-

posed schemes employ 95% less RF chains than massive MIMO systems where no

antenna selection is performed, yet they show reduced performance losses. The

gains of the proposed schemes directly translate into energy efficiency improve-

ments. These considerations suggest that the employment of very-large arrays at

148
 Chapter 7. Conclusions

the base station does not automatically imply the requirement for an equally large

number of RF chains, as long as the subset of transmitting antennas is properly

identified. The main observations included in this chapter are:

C4.1. Tailoring antenna selection algorithms for M-MIMO systems with Construc-

tive Interference can be greatly beneficial for energy efficiency with 30%

gains in the considered scenarios. Moreover, it was shown that state-of-the-

art antenna selection for small-scale MIMO cannot be directly applied to

large-scale systems, as their complexity is too high to attain any benefits.

C4.2. The proposed scheme is most useful in the low to mid SNR range for low-

order modulations, where the proposed schemes are able to nearly triplicate

the power-efficiency of large-scale MIMO system without antenna selection.

However, at higher-order modulations, the proposed metrics appear to be

less efficient in identifying the best transmitting antenna subset for exploit-

ing interference.

• Considerations and results described in Chapters 3 and 4 motivated the definition

of a novel transmission scheme in Chapter 5. Here, transmit antenna selection

and downlink precoding are jointly performed to achieve the maximum benefits

from both techniques. The use of a unified metric for both problems leads to the

formulation of a Mixed-Integer Programming problem, hence non-convex. While

a solution to the MIP formulation can be achieved through commercial solvers,

three separate heuristic approaches are proposed and their optimality is studied.

The main contributions of the chapter can be highlighted in the following list:

C5.1. Jointly performing antenna selection and precoding under a unified metric

allows to greatly enhance the performance from both techniques, allowing to

reduce both hardware complexity and computational costs of systems where

the two problems are solved separately. This is visible in the shown Symbol-

Error Rate performances, where the proposed schemes are characterized by

gains on the order of tens of dB when compared to state-of-the-art transmit

antenna selection-based schemes.

149
Chapter 7. Conclusions

C5.2. While the solution of the MIP-based approach is characterized by the high-

est performances, the optimality analysis of three heuristic schemes proved

that the problem can be efficiently and near-optimally solved via successive

optimization. More specifically, it was shown that the proposed successive

optimization approaches are able to achieve increasing close-to-optimal per-

formances as their complexity grows. Accordingly, the proposed scheme is

able to offer a very high scalability with an always-positive adaptive trade-

off between complexity and performances, which can be selected according

to the specific requirements of the considered scenario.

• Chapter 6 proposed several beam selection schemes for DLA-based multiuser

beamspace mm-wave MIMO systems. The proposed schemes showed that more so-

phisticated beam selection algorithms, based on antenna selection concepts, allow

to jointly achieve a reduced hardware complexity and near-optimal performances

in a multi-path scenario. However, these improvements come at the cost of an

increased signal processing complexity. Main remarks from this chapter can be

synthesized in the following:

C6.1. The proposed algorithms allow to identify a fixed subset-size of transmitting

beams, i.e., generally set equal to the number of users nRF = M , differently

from the conventional approaches from the literature.

C6.2. When considering multi-path scenarios, the proposed schemes show higher

robustness to additional paths with gains on the order of 5 dB in the high

SNR scenarios. This justifies the employment of the proposed transceiving

schemes in more realistic communication channels.

7.2 Future Work

The studies described in this thesis represent the foundation for future works which

could focus on further developing novel techniques that aim at increasing the energetic

efficiency of large-scale MIMO systems. We identify and discuss possible future lines of

150
 Chapter 7. Conclusions

work in this section. More specifically:

• Characterization of the doughnut channel for multiuser large-scale Con-

stant Envelope Precoding: The results and conclusions derived in Chapter 3

motivate further research with the objective of identifying the bounds of the so-

called doughnut channel for multiuser large-scale MIMO. As shown in [17], CEP in

single-user scenarios leads to a doughnut shaped channel, where the minimum and

maximum received constellation amplitude can be analytically derived. However,

such derivations are not available for the multiuser scenario, hence not allowing

to identify the optimal desired constellation amplitude for the considered channel.

Therefore, the identification of the optimal desired constellation amplitude via a

combination of constructive interference and doughnut channel concepts would

allow to extend the benefits of interference exploiting CEP to other modulations,

such as Quadrature Amplitude Modulation.

• Low-Complexity Antenna selection for Quadrature Amplitude Modu-

lation: Chapter 4 showed that low-complexity constructive interference metrics

can be applied to antenna selection for massive MIMO system under PSK trans-

mission. More specifically, these results proved that such metrics allow to both

reduce the computational complexity and the hardware requirements of large-scale

systems. Nonetheless, these metrics are most efficient in lower modulations and

for the specific case of PSK signaling. As a consequence, the possibility to ex-

tend these concepts to different modulations, such as Amplitude and Phase Shift

Keying described in [150], definitely represents an interesting direction for future

research developments.

• Limited Connectivity in joint antenna selection and precoding: The main

shortcomings of the application of constructive interference concepts to antenna

selection reside in the need to operate switching at a symbol-by-symbol rate. These

issues, briefly addressed in Chapter 4, directly translate to the transceiving scheme

presented in Chapter 5. Recent works [151] considered a switching network where

each RF chain can be connected to a predefined subset of antennas in order to

151
Chapter 7. Conclusions

reduce switching complexity in the VLAs. Therefore, in an effort to increase both

energy efficiency and practicability of such schemes, the study of a combination of

limited connectivity switching matrices with symbol level antenna selection metrics

surely poses an appealing challenge for future research.

• Employing the concepts of beamspace mm-wave MIMO onto spatial

modulation: The beam selection-based scheme developed in Chapter 6 allows

to greatly reduce the hardware complexity of mm-wave systems, while preserving

the high antenna gains required by higher frequencies. Accordingly, Discrete Lens

Arrays (DLAs) could allow Spatial Modulation (SM) [127] to be applied in the

mm-wave spectrum. In its simpler formulation, SM allows to convey two separate

data streams while using only one transmitting antenna, hence one RF chain. This

is performed by selecting a single transmitting antenna among a set of available

antennas, each corresponding to a specific symbol or bit. Here, while assuming

that the receiver is aware of the channel, it is possible to decode two separate data

streams: one corresponding to the transmitted stream and one corresponding to

the antenna used for transmission. As it follows, a direct and straightforward

application of SM concepts to mm-wave frequencies is not viable because of their

characteristic higher free loss attenuations. However, DLA-based schemes might

suit the task, with future works aiming to characterize a beam modulation tech-

nique, where the beams can be selected according to the symbols to be transmitted.

Taking everything into consideration, this thesis presented and analyzed several

energy-efficient designs for large-scale communication systems. The observations and

conclusions derived from this thesis will be helpful for a more profound understanding

of the challenges and practical issues involved in the employment of very large arrays in

future wireless communication systems.

152
Appendices

Appendix A. Proof of Theorem 4.4.1

Generalize equation (4.14) to explicitly represent the effects of interference as

X X
ym = γM F |ρm,m |um + γM F |ρm,k |uk + γM F |ρm,k |uk + nm . (A.1)
k∈C k∈D

Accordingly, we can define the received SINR for the m-th user as [81]

 2
γ2 |2 γ2
P
M F,e
|ρm,m + M F,e
|ρm,k |
k∈C
ξm =  2 . (A.2)
2
P
N0 + γM F,e
|ρm,k |
k∈D

where the constructive interference that contributes to the signal power appears at the

numerator and the destructive component of the interference is added to N0 at the

denominator, since it can be interpreted as an additional source of noise.

ICI = 0 with an
The received SINR is upper bounded by the condition where Dm

optimal CIM antenna selection at the transmitter. Accordingly, a SINR upper bound

can be derived as  2
γ2 |2 γ2
P
M F,e
|ρm,m + M F,e
|ρm,k |
k∈C
ξm = (A.3)
N0

which can be seen as a generalized form of SNR, as the interference is a constructive

parameter. Since we are interested in the average value of the SINR, we can apply the

153
APPENDICES

expectation over equation (A.3), leading to

( M 2 ) n o
(
M
2 )
γ2 2
P P
E M F,e
|ρm,k | E γM F,e
E |ρm,k |
k=1 k=1
ξem = = . (A.4)
N0 N0

In this equation we consider γM F,e to be data independent, even though the con-

ditions used to perform the antenna selection do not support this assumption. This

simplification is often performed and necessary to derive a closed form definition of ξf

m.

In order to derive the expected value of the received SINR ξf

m , we need to identify

the statistical properties of the correlation matrix R = HHH and its entries ρm,k .

In an independent Rayleigh fading scenario where the entries of H are modelled as

i.i.d Gaussian variables, the correlation matrix R is known to be a Wishart matrix,

characterized by the following distribution function [47]

π −M (M −1)/2 det An−M −T r[Σ−1 A]

fR (A) = M
e (A.5)
n
Q
det Σ (n − k)!
k=1

where the matrix Σ is the covariance matrix of the correlation matrix R.

We can define the absolute value of the entries ρm,k of the correlation matrix, related

to the m-th user, as [152]

v 2  2
u N NRF
u X RF X
hR R I I  + hIm,j hR R I  ,
u
|ρm,k | = t m,j hk,j + hm,j hk,j k,j − hm,j hk,j (A.6)
j=1 j=1

where hR and hI are used to identify respectively the real and imaginary part of h.

Thanks to the assumption of independent Rayleigh fading propagation, for hR

m,j ∼

CN (0, 1/2) we have

 
NX
RF 
E hR R I I
m,j hk,j + hm,j hk,j = 0, ∀j ∈ {1, .., NRF } (A.7)
 
j=1

154
 APPENDICES

 
NX
RF  1
var hR R I I
m,j hk,j + hm,j hk,j = , ∀j ∈ {1, .., NRF } (A.8)
  2
j=1

where var {·} is used to identify the variance of the argument. These equations can
n o n o
be derived from E hR R
m,j hk,j = 0 and var hR R
m,j hk,j = 1/4, thanks to the linearity

of E {·} and the uncorrelation of the entries of H. Hence the variables |ρm,k | can be

distinguished between [152]:

√
Mπ
and E |ρm,k |2 = M

• Rayleigh variables when k 6= i with E {|ρm,k |} = 2

• χ-squared variables when m = k with E {|ρm,k |} = M and E |ρm,k |2 = M (M +



1).

In order to complete the study of the upper bound received SINR for the proposed
n o
2
technique we need to identify the expected value of the scaling factor E γM . Fol-
F,e

lowing (2.6), we have

H −1
n o n   o
2
E γM F,e
= E tr H H
e e . (A.9)

As previously stated, the statistical properties of the matrix He HH

e lead to [47]

n
2
o 1
E γM F,e
= . (A.10)
NRF M

Hence, we can evaluate the upper bound of the received SINR for the m-th user as

 M  ( !)
|ρm,k |2
P P
E +E |ρm,k ||ρm,j |
k=1 j6=k
ξem = (A.11)
NRF M N0

The first and second term on the numerator of (A.11) can be rearranged in order to

exploit the statistical properties listed above. In particular, for the first term we have:
 
 X 
E |ρm,m |2 + |ρm,k |2 = M (M + 1) + (M − 1)M (A.12)
 
k6=m

155
APPENDICES

and for the second term

√
 
X X  Mπ Mπ
E |ρm,m ||ρm,j | + |ρm,k ||ρm,j | = 2(M − 1)M + (M − 2)(M − 1)
  2 4
j6=m j6=k,k6=m
(A.13)

which is derived thanks to the independence of the random variables.

Hence, the equation (A.11) can be evaluated analytically as

√
M (M + 1) + M (M − 1)(1 + M π + (M − 2)π/4)
ξem = (A.14)
NRF M N0

which ends the proof.

Appendix B. Proof of Theorem 5.6.1

Consider an imperfect CSI and noiseless scenario, received symbols can be decom-

posed in order to explicitly show their real and imaginary part as follows

N
P
ŷm = ĥm,n xn
n=1 (A.15)
N
(ĥR R I I R I I R
P
= m,n xn − ĥm,n xn ) + j(ĥm,n xn + ĥm,n xn ),
n=1

where real and imaginary part can be rearranged in order to explicitly identify the effects

of imperfect CSI over the received symbols as

N
X
= {ŷm } = (hR R I I R R I I
m,n xn − hm,n xn ) + (em,n xn − em,n xn ) (A.16)
n=1

N
X
< {ŷm } = (hR I I R R I I R
m,n xn + hm,n xn ) + (em,n xn + em,n xn ). (A.17)
n=1

Both (A.16) and (A.17) can be presented in a more compact manner by exploiting

auxiliary vectors, which lead to the following set of equations

T T
= {ŷm } = f̂m w1 = fm w1 + ēTm w1 (A.18)

156
 APPENDICES

T T
< {ŷm } = f̂m w2 = fm w2 + ēTm w2 , (A.19)

where the vectors fm = hR I T and ē = eR , eI T represent the real-valued chan-

   
m , hm m m m

nel response for the m-th user and the corresponding channel estimation error vector re-
T T
spectively, with f̂m = f +ēm . In a similar manner, w1 = xI , xR and w2 = xR , −xI
 

are two auxiliary real-valued representations of the precoded vector.

Likewise, we can rewrite the first constraint of PCASP R with the same notation

˙
 
= t̂m e−jφm = f̂ T w1 = ḟm
T
w1 + ē˙ T w1 (A.20)

˙
 
< t̂m e−jφm = f̂ T w2 = ḟm
T
w2 + ē˙ T w2 − 1, (A.21)

where
T
ḟm = (hIm uR R I R R I I

m − hm um ), (hm um + hm um ) (A.22)

T
ē˙ m = (ēIm uR R I R R I I

m − ēm um ), (ēm um + ēm um ) (A.23)

represent the real-valued channel vector and CSI error vector for the m-th user, whose

representations have been modified in order to include the phase shift. Without loss of

generality, for the sake of simplicity of notation, from now on we consider ē˙ m = ēm .

Accordingly, we can rewrite the first constraint of PCASP R , as

n o
T
min (ḟm w2 + ēTm w2 − 1) tan Φ − ḟm
T
w1 + ēTm w1 ≥ 0. (A.24)
kem k2 ≤δm
2

which can be equivalently decomposed into two different constraints

n o
T
min (ḟm w2 + ēTm w2 − 1) tan Φ − (ḟm
T
w1 + ēTm w1 ) ≥ 0, (A.25)
kem k2 ≤δm
2

n o
T
min (ḟm w2 + ēTm w2 − 1) tan Φ + (ḟm
T
w1 + ēTm w1 ) ≥ 0. (A.26)
kem k2 ≤δm
2

The assumption of a spherical error over the CSI acquisition allows us to derive a

robust formulation for (A.25) and (A.26). In fact, the worst-case scenario is characterized

by the channel errors to be kem k2 = δm

2 , ∀m, hence causing the constraints to be lower-

157
APPENDICES

bounded by the following equations

h i
T T
(ḟm w2 − 1) tan Φ − ḟm w1 − δ kw1 − w2 tan Φk ≥ 0, (A.27)

h i
T T
(ḟm w2 − 1) tan Φ + ḟm w1 − δ kw1 + w2 tan Φ)k ≥ 0. (A.28)

Thanks to the new robust formulation for the constraints of PCASP R , we can derive

a MIP representation of the worst-case design for imperfect CSI scenarios that can be

efficiently solved by means of optimization tools as its non-robust counterpart. More

specifically, the optimization problem PCASP R in its MIP representation becomes

NRF
∗ kw1 k2
P
PCASP R : minimize
a,w1 ,w2 n

subject to Constraints (A.27) and (A.28)

n o
(1:N ) (N +1:2N )
w1 , w1  a,
n o (A.29)
(1:N ) (N +1:2N )
w2 , w2  a,

w1 = Πw2 ,
N
P
an = NRF , an ∈ {0, 1} ,
n=1

where Π = [0N , −IN ; IN , 0N ], a(1:N ) notation is used to identify the new vector b =

[a1 , ..., aN ] and {a, b}  c is used to impose the inequality to both vectors (i.e., a  c

and b  c).

Appendix C. Proof of Sherman-Morrison-Woodbury Iden-

tity

Sherman-Morrison-Woodbury Identity states that for an invertible matrix A and

two or more non-invertible matrices B,C

−1
(A + BC)−1 = A(I + A−1 BC)


−1 −1 (A.30)
(A + BC)−1 = I + A−1 BC A

158
 APPENDICES

which, thanks to the identity (I + P)−1 = I − (I + P)−1 P, can be modified to

h
−1 −1 i
(A + BC)−1 = I − I + A−1 BC A BC A−1 (A.31)

and rearranged with the identity P + PQP = P (I + QP)−1 to

(A + BC)−1 = A−1 − A−1 B(1 + CA−1 B)−1 CA−1 . (A.32)

159
160
List of References

[1] S. Sesia, I. Toufik, and M. Baker, LTE-the UMTS long term evolution. Wiley
Online Library, 2015.

[2] “Cisco Visual Networking Index : Global mobile data traffic forecast update,”
White Paper, pp. 2012–2017, 2013.

[3] Z. Hasan, H. Boostanimehr, and V. K. Bhargava, “Green cellular networks: A

survey, some research issues and challenges,” IEEE Communications Surveys and
Tutorials, vol. 13, no. 4, pp. 524–540, April 2011.

[4] Deloitte, “Technology, media and telecommunications predictions 2012,” Media,

pp. 1–53, 2012.

[5] G. Y. Li, Z. Xu, C. Xiong, C. Yang, S. Zhang, Y. Chen, and S. Xu, “Energy-
efficient wireless communications: tutorial, survey, and open issues,” IEEE Wire-
less Communications, vol. 18, no. 6, pp. 28–35, December 2011.

[6] G. Fettweis and E. Zimmermann, “ICT energy consumption-trends and chal-

lenges,” in Proceedings of the 11th International Symposium on Wireless Personal
Multimedia Communications, vol. 2, no. 4. Lapland, 2008, p. 6.

[7] E. Björnson, L. Sanguinetti, J. Hoydis, and M. Debbah, “Optimal design of energy-

efficient multi-user MIMO systems: Is massive MIMO the answer?” IEEE Trans-
actions on Wireless Communications, vol. 14, no. 6, pp. 3059–3075, June 2015.

[8] Y. Chen, S. Zhang, S. Xu, and G. Y. Li, “Fundamental trade-offs on green wireless
networks,” IEEE Communications Magazine, vol. 49, no. 6, pp. 30–37, 2011.

[9] J. Wu, S. Rangan, and H. Zhang, Green communications: theoretical fundamen-

tals, algorithms and applications. CRC Press, 2012.

[10] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge

University Press, 2005.

[11] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base

station antennas,” IEEE Transactions on Wireless Communications, vol. 9, no. 11,
pp. 3590–3600, November 2010.

161
LIST OF REFERENCES

[12] E. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive MIMO for next
generation wireless systems,” IEEE Communications Magazine, vol. 52, no. 2, pp.
186–195, February 2014.

[13] A. Sayeed, “Deconstructing multiantenna fading channels,” IEEE Transactions

on Signal Processing, vol. 50, no. 10, pp. 2563–2579, 2002.

[14] J. Brady, N. Behdad, and A. M. Sayeed, “Beamspace MIMO for millimeter-wave

communications: System architecture, modeling, analysis, and measurements,”
IEEE Transactions on Antennas and Propagation, vol. 61, no. 7, pp. 3814–3827,
July 2013.

[15] S. Sanayei and A. Nosratinia, “Antenna selection in mimo systems,” IEEE Com-
munications Magazine, vol. 42, no. 10, pp. 68–73, October 2004.

[16] ——, “Capacity of MIMO channels with antenna selection,” IEEE Transactions
on Information Theory, vol. 53, no. 11, pp. 4356–4362, November 2007.

[17] S. Mohammed and E. Larsson, “Single-User beamforming in Large-Scale MISO

systems with Per-Antenna Constant-Envelope constraints: The doughnut chan-
nel,” IEEE Transactions on Wireless Communications, vol. 11, no. 11, pp. 3992–
4005, November 2012.

[18] ——, “Per-Antenna constant envelope precoding for large Multi-User MIMO sys-
tems,” IEEE Transactions on Communications, vol. 61, no. 3, pp. 1059–1071,
March 2013.

[19] A. Molisch, M. Win, and J. Winters, “Capacity of mimo systems with antenna
selection,” IEEE International Conference on Communications, vol. 2, pp. 570–
574, 2001.

[20] A. Molisch and M. Win, “MIMO Systems with antenna selection,” IEEE Mi-
crowave Magazine, vol. 5, pp. 46–56, March 2004.

[21] A. Gorokhov, D. Gore, and A. Paulraj, “Receive antenna selection for MIMO
flat-fading channels: theory and algorithms,” IEEE Transactions on Information
Theory, vol. 49, no. 10, pp. 203–206, October 2003.

[22] A. Sayeed and N. Behdad, “Continuous Aperture Phased MIMO: A new architec-
ture for optimum line-of-sight links,” IEEE APSURSI, pp. 293–296, July 2011.

[23] A. M. Sayeed and J. Brady, “Beamspace MIMO for high-dimensional multiuser

communication at millimeter-wave frequencies,” in GLOBECOM 2013, 2013, pp.
3784–3789.

162
 LIST OF REFERENCES

[24] T. S. Rappaport, F. Gutierrez, E. Ben-Dor, J. N. Murdock, Y. Qiao, and J. I.

Tamir, “Broadband millimeter-wave propagation measurements and models us-
ing adaptive-beam antennas for outdoor urban cellular communications,” IEEE
transactions on antennas and propagation, vol. 61, no. 4, pp. 1850–1859, April
2013.

[25] Q. Spencer, C. Peel, A. Swindlehurst, and M. Haardt, “An introduction to the

multi-user MIMO downlink,” IEEE Communications Magazine, vol. 42, no. 10,
pp. 60–67, October 2004.

[26] E. Telatar, “Capacity of Multi-antenna Gaussian Channels,” European transac-

tions on telecommunications, vol. 10, no. 6, pp. 585–595, November 1999.

[27] A. H. Mehana and A. Nosratinia, “Diversity of MIMO linear precoding,” IEEE

Transactions on Information Theory, vol. 60, no. 2, pp. 1019–1038, February 2014.

[28] M. Joham, W. Utschick, and J. Nossek, “Linear transmit processing in MIMO

communications systems,” IEEE Transactions on Signal Processing, vol. 53, no. 8,
pp. 2700–2712, August 2005.

[29] C. Peel, B. Hochwald, and A. Swindlehurst, “A vector-perturbation technique for

near-capacity multiantenna multiuser communication - Part I: Channel inversion
and regularization,” IEEE Transactions on Communications, vol. 53, no. 1, pp.
195–202, January 2005.

[30] V. Nguyen and J. Evans, “Multiuser transmit beamforming via regularized channel
inversion: A large system analysis,” IEEE GLOBECOM, pp. 1–4, 2008.

[31] M. Costa, “Writing on dirty paper,” IEEE Transactions on Information Theory,

vol. 29, no. 3, pp. 439–441, May 1983.

[32] B. M. Hochwald, C. B. Peel, and A. L. Swindlehurst, “A Vector-Perturbation

technique for Near-Capacity multiantenna multiuser communication - Part II: Per-
turbation,” IEEE Transactions on Communications, vol. 53, no. 3, pp. 537–544,
March 2005.

[33] C. Windpassinger, R. F. H. Fischer, T. Vencel, and J. B. Huber, “Precoding in

multiantenna and multiuser communications,” IEEE Transactions on Wireless
Communications, vol. 3, no. 4, pp. 1305–1316, July 2004.

[34] A. B. Gershman, N. D. Sidiropoulos, S. Shahbazpanahi, M. Bengtsson, and B. Ot-

tersten, “Convex optimization-based beamforming,” IEEE Signal Processing Mag-
azine, vol. 27, no. 3, pp. 62–75, May 2010.

163
LIST OF REFERENCES

[35] M. Schubert and H. Boche, “Solution of the multiuser downlink beamforming prob-
lem with individual SINR constraints,” IEEE Transactions on Vehicular Technol-
ogy, vol. 53, no. 1, pp. 18–28, January 2004.

[36] C. Masouros, M. Sellathurai, and T. Ratnarajah, “Large-Scale MIMO transmitters

in fixed physical spaces: The effect of transmit correlation and mutual coupling,”
IEEE Transactions on Communications, vol. 61, no. 7, pp. 2794–2804, July 2013.

[37] ——, “Interference optimization for transmit power reduction in Tomlinson-

Harashima precoded MIMO downlinks,” IEEE Transactions on Signal Processing,
vol. 60, no. 5, pp. 2470–2841, May 2012.

[38] A. Gorokhov, D. Gore, and A. Paulraj, “Receive antenna selection for MIMO
spatial multiplexing: Theory and algorithms,” IEEE Transactions on Signal Pro-
cessing, vol. 51, no. 11, pp. 2796–2807, November 2003.

[39] R. Chen, R. W. Heath, and J. G. Andrews, “Transmit selection diversity for

unitary precoded multiuser spatial multiplexing systems with linear receivers,”
IEEE Transactions on Signal Processing, vol. 55, no. 3, pp. 1159–1171, March
2007.

[40] Z. Chen, J. Yuan, and B. Vucetic, “Analysis of transmit antenna

selection/maximal-ratio combining in Rayleigh fading channels,” IEEE Transac-
tions on Vehicular Technology, vol. 54, no. 4, pp. 1312–1321, July 2005.

[41] R. Heath and A. Paulraj, “Antenna selection for spatial multiplexing systems
based on minimum error rate,” IEEE International Conference on Communica-
tions (ICC), vol. 7, pp. 2276–2280, 2001.

[42] T. Eng, N. Kong, and L. B. Milstein, “Comparison of diversity combining tech-

niques for Rayleigh-fading channels,” IEEE Transactions on Communications,
vol. 44, no. 9, pp. 1117–1129, September 1996.

[43] M. Z. Win and J. H. Winters, “Analysis of hybrid selection/maximal-ratio com-

bining in Rayleigh fading,” in Communications, 1999. ICC’99. 1999 IEEE Inter-
national Conference on, vol. 1. IEEE, 1999, pp. 6–10.

[44] D. A. Gore, R. U. Nabar, and A. Paulraj, “Selecting an optimal set of transmit

antennas for a low rank matrix channel,” in Acoustics, Speech, and Signal Pro-
cessing, 2000. ICASSP’00. Proceedings. 2000 IEEE International Conference on,
vol. 5. IEEE, 2000, pp. 2785–2788.

[45] C. Masouros, M. Sellathurai, and T. Ratnarajah, “Computationally efficient vector

perturbation using thresholded optimization,” IEEE Transactions on Communi-
cations, vol. 61, no. 5, pp. 1880–1890, May 2013.

164
 LIST OF REFERENCES

[46] S. Cui, A. J. Goldsmith, and A. Bahai, “Energy-constrained modulation opti-

mization,” IEEE Transactions on Wireless Communications, vol. 4, no. 5, pp.
2349–2360, September 2005.

[47] A. M. Tulino and S. Verdu, “Random matrix theory and wireless communications,”
Foundations and Trends in Communications and Information Theory, vol. 1, no. 1,
pp. 1–182, June 2004.

[48] F. Rusek, D. Persson, B. K. Lau, E. Larsson, T. Marzetta, O. Edfors, and

F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with Very Large
Arrays,” IEEE Signal Processing Magazine, vol. 30, no. 1, pp. 40–60, January
2013.

[49] A. Adhikary, J. Nam, J.-Y. Ahn, and G. Caire, “Joint spatial division and multi-
plexing - The large-scale array regime,” IEEE transactions on information theory,
vol. 59, no. 10, pp. 6441–6463, October 2013.

[50] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL of cellular
networks: How many antennas do we need?” IEEE Journal on Selected Areas in
Communications, vol. 31, no. 2, pp. 160–171, February 2013.

[51] T. L. Marzetta, “How much training is required for multiuser MIMO?” Asilomar
Conference on Signals, Systems and Computers (ACSSC), pp. 359–363, October
2006.

[52] Y. Li, Y.-H. Nam, B. L. Ng, and J. Zhang, “A non-asymptotic throughput for
massive MIMO cellular uplink with pilot reuse,” in Global Communications Con-
ference (GLOBECOM), 2012 IEEE. IEEE, 2012, pp. 4500–4504.

[53] K. Appaiah, A. Ashikhmin, and T. L. Marzetta, “Pilot contamination reduction

in multi-user TDD systems,” in Communications (ICC), 2010 IEEE International
Conference on. IEEE, 2010, pp. 1–5.

[54] F. Fernandes, A. Ashikhmin, and T. L. Marzetta, “Inter-cell interference in non-

cooperative tdd large scale antenna systems,” IEEE Journal on Selected Areas in
Communications, vol. 31, no. 2, pp. 192–201, February 2013.

[55] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot contamina-

tion and precoding in multi-cell TDD systems,” IEEE Transactions on Wireless
Communications, vol. 10, no. 8, pp. 2640–2651, August 2011.

[56] H. Huh, S.-H. Moon, Y.-T. Kim, I. Lee, and G. Caire, “Multi-cell MIMO downlink
with cell cooperation and fair scheduling: A large-system limit analysis,” IEEE
Transactions on Information Theory, vol. 57, no. 12, pp. 7771–7786, December
2011.

165
LIST OF REFERENCES

[57] H. Huh, A. M. Tulino, and G. Caire, “Network MIMO with linear zero-forcing
beamforming: Large system analysis, impact of channel estimation, and reduced-
complexity scheduling,” IEEE Transactions on Information Theory, vol. 58, no. 5,
pp. 2911–2934, May 2012.

[58] H. Q. Ngo and E. G. Larsson, “EVD-based channel estimation in multicell mul-

tiuser MIMO systems with very large antenna arrays,” in 2012 IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2012,
pp. 3249–3252.

[59] E. Björnson, J. Hoydis, M. Kountouris, and M. Debbah, “Massive MIMO systems

with non-ideal hardware: Energy efficiency, estimation, and capacity limits,” IEEE
Transactions on Information Theory, vol. 60, no. 11, pp. 7112–7139, November
2014.

[60] T. Schenk, RF imperfections in high-rate wireless systems: Impact and digital

compensation. Springer Science & Business Media, 2008.

[61] C. Masouros and M. Matthaiou, “Space-Constrained Massive MIMO: Hitting the

wall of favorable propagation,” IEEE Communications Letters, vol. 19, no. 5, pp.
771–774, May 2015.

[62] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadband sys-

tems,” IEEE Communications Magazine, vol. 49, no. 6, pp. 101–107, June 2011.

[63] Millimeter Wave Propagation : Spectrum Management Implications. Federal

Communications Commission. Office Of Engineering and Technology, 1997, no. 70.

[64] T. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, J. K. Wong, G.

N.and Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave mobile communi-
cations for 5G cellular: It will work!” IEEE Access, vol. 1, pp. 335–349, 2013.

[65] D. M. Pozar, Microwave engineering. John Wiley & Sons, 2009.

[66] A. V. Alejos, M. G. Sánchez, and I. Cuiñas, “Measurement and analysis of prop-

agation mechanisms at 40 GHz: Viability of site shielding forced by obstacles,”
IEEE Transactions on Vehicular Technology, vol. 57, no. 6, pp. 3369–3380, Novem-
ber 2008.

[67] M. Samimi, K. Wang, Y. Azar, G. N. Wong, R. Mayzus, H. Zhao, J. K. Schulz,

S. Sun, F. Gutierrez, and T. S. Rappaport, “28 GHz angle of arrival and angle
of departure analysis for outdoor cellular communications using steerable beam
antennas in New York City,” in Vehicular Technology Conference (VTC Spring),
2013 IEEE 77th. IEEE, 2013, pp. 1–6.

166
 LIST OF REFERENCES

[68] T. S. Rappaport, R. W. Heath Jr, R. C. Daniels, and J. N. Murdock, Millimeter

wave wireless communications. Pearson Education, 2014.

[69] X. Zhang, A. F. Molisch, and S.-Y. Kung, “Variable-phase-shift-based RF-

baseband codesign for MIMO antenna selection,” IEEE Transactions on Signal
Processing, vol. 53, no. 11, pp. 4091–4103, November 2005.

[70] R. W. Heath, N. Gonzalez-Prelcic, S. Rangan, W. Roh, and A. M. Sayeed, “An

overview of signal processing techniques for millimeter wave MIMO systems,”
IEEE Journal of Selected Topics in Signal Processing, vol. 10, no. 3, pp. 436–453,
April 2016.

[71] V. Venkateswaran and A.-J. van der Veen, “Analog beamforming in MIMO com-
munications with phase shift networks and online channel estimation,” IEEE
Transactions on Signal Processing, vol. 58, no. 8, pp. 4131–4143, August 2010.

[72] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel estimation and

hybrid precoding for millimeter wave cellular systems,” IEEE Journal of Selected
Topics in Signal Processing, vol. 8, no. 5, pp. 831–846, October 2014.

[73] A. Alkhateeb, G. Leus, and R. W. Heath, “Limited feedback hybrid precoding for
multi-user millimeter wave systems,” IEEE Transactions on Wireless Communi-
cations, vol. 14, no. 11, pp. 6481–6494, November 2015.

[74] X. Gao, L. Dai, S. Han, I. Chih-Lin, and R. W. Heath, “Energy-efficient hybrid

analog and digital precoding for mmWave MIMO systems with large antenna
arrays,” IEEE Journal on Selected Areas in Communications, vol. 34, no. 4, pp.
998–1009, April 2016.

[75] L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid precoding in massive

multiuser MIMO systems,” IEEE Wireless Communications Letters, vol. 3, no. 6,
pp. 653–656, December 2014.

[76] A. Garcia-Rodriguez and C. Masouros, “Exploiting the increasing correlation of

space constrained massive MIMO for CSI relaxation,” IEEE Transactions on Com-
munications, vol. 64, no. 4, pp. 1572–1587, April 2016.

[77] Y. Zeng and R. Zhang, “Millimeter wave MIMO with lens antenna array: A new
path division multiplexing paradigm,” IEEE Transactions on Communications,
vol. 64, no. 4, pp. 1557–1571, April 2016.

[78] X. Gao, L. Dai, Z. Chen, Z. Wang, and Z. Zhang, “Near-Optimal Beam Selec-
tion for Beamspace MmWave Massive MIMO Systems,” IEEE Communications
Letters, vol. 20, no. 5, pp. 1054–1057, May 2016.

167
LIST OF REFERENCES

[79] L. Yang, Y. Zeng, and R. Zhang, “Efficient channel estimation for millimeter
wave MIMO with limited RF chains,” in Communications (ICC), 2016 IEEE
International Conference on. IEEE, 2016, pp. 1–6.

[80] C. Masouros, T. Ratnarajah, M. Sellathurai, C. Papadias, and A. Shukla, “Known

interference in the cellular downlink: A performance limiting factor or a source
of green signal power?” IEEE Communications Magazine, vol. 51, no. 10, pp.
162–171, October 2013.

[81] C. Masouros and E. Alsusa, “Two-Stage transmitter precoding based on data-

driven code hopping and partial zero forcing beamforming for MC-DCMA com-
munications,” IEEE Transactions on Wireless Communications, vol. 8, no. 7, pp.
3634–3645, July 2009.

[82] ——, “Soft linear precoding for the downlink of DS/CDMA communication sys-
tems,” IEEE Transactions on Vehicular Technology, vol. 59, no. 1, pp. 203–215,
January 2010.

[83] C. Masouros, “Correlation rotation linear precoding for MIMO broadcast commu-
nications,” IEEE Transactions on Signal Processing, vol. 59, no. 1, pp. 252–262,
January 2011.

[84] C. Masouros and G. Zheng, “Exploiting known interference as green signal power
for downlink beamforming optimization,” IEEE Transactions on Signal Process-
ing, vol. 63, no. 14, pp. 3628–3640, July 2015.

[85] M. Alodeh, S. Chatzinotas, and B. Ottersten, “Constructive multiuser interference

in symbol level precoding for the MISO Downlink channel,” IEEE Transactions
on Signal Processing, vol. 63, no. 9, pp. 2239–2252, May 2015.

[86] ——, “Energy-efficient symbol-level precoding in multiuser miso based on relaxed

detection region,” IEEE Transactions on Wireless Communications, vol. 15, no. 5,
pp. 3755–3767, May 2016.

[87] A. Li and C. Masouros, “Exploiting constructive mutual coupling in P2P MIMO

by Analog-Digital phase alignment,” IEEE Transactions on Wireless Communi-
cations, In Press.

[88] K. L. Law, C. Masouros, K. K. Wong, and G. Zheng, “Constructive interference ex-

ploitation for Outage-Probability constrained downlink precoding optimization,”
IEEE Transactions on Signal Processing, In Press.

[89] A. Kalantari, M. Soltanalian, S. Maleki, S. Chatzinotas, and B. Ottersten, “Direc-

tional modulation via symbol-level precoding: A way to enhance security,” arXiv
preprint arXiv:1606.04488, 2016.

168
 LIST OF REFERENCES

[90] M. P. Daly and J. T. Bernhard, “Directional modulation technique for phased ar-
rays,” IEEE Transactions on Antennas and Propagation, vol. 57, no. 9, September
2009.

[91] C. Masouros and E. Alsusa, “Dynamic Linear Precoding for the Exploitation of
Known Interference in MIMO Broadcast Systems,” IEEE Transactions on Wire-
less Communications, vol. 8, no. 3, pp. 1396–1404, March 2009.

[92] V. Mancuso and S. Alouf, “Reducing costs and pollution in cellular networks,”
IEEE Communications Magazine, vol. 49, no. 8, pp. 63–71, August 2011.

[93] S. H. Han and J. H. Lee, “An overview of Peak-to-Average power ratio reduc-
tion techniques for multicarrier transmission,” IEEE Wireless Communications
Magazine, vol. 12, no. 2, pp. 56–65, April 2005.

[94] C. Studer and E. Larsson, “PAR-Aware Large-Scale Multi-User MIMO-OFDM

downlink,” IEEE Journal on Selected Areas in Communications, vol. 31, no. 2,
pp. 303–313, February 2013.

[95] S. Mohammed and E. Larsson, “Constant-Envelope Multi-User precoding for

Frequency-Selective massive MIMO systems,” IEEE Wireless Communications
Letters, vol. 2, no. 5, pp. 547–550, October 2013.

[96] S. Mukherjee and S. K. Mohammed, “Constant-envelope precoding with time-

variation constraint on the transmitted phase angles,” IEEE Wireless Communi-
cations Letters, vol. 4, no. 2, pp. 221–224, April 2015.

[97] J. C. Chen, C. K. Wen, and K. K. Wong, “Improved constant envelope multiuser

precoding for massive MIMO systems,” IEEE Communications Letters, vol. 18,
no. 8, pp. 1311–1314, August 2014.

[98] Evolved Universal Terrestrial Radio Access (E-UTRA); Physical Channels and
Modulation,. 3GPP TS 36.211, V8.2.0 (2008-03), Release 8.

[99] H. Yang and T. L. Marzetta, “Performance of conjugate and zero-forcing beam-

forming in Large-Scale Antenna systems,” IEEE Journal on Selected Areas in
Communications, vol. 31, no. 2, pp. 172–179, February 2013.

[100] P. T. De Boer, D. P. Kroese, S. Mannor, and R. Y. Rubinstein, “A tutorial on the

cross-entropy method,” Annals of operations research, vol. 134, no. 1, pp. 19–67,
2005.

[101] R. Rubinstein and D. Kroese, The Cross-Entropy Method: A Unified Approach to

Combinatorial Optimization, Monte-Carlo Simulation and Machine Learning, ser.
Information Science and Statistics. Springer New York, 2011.

169
LIST OF REFERENCES

[102] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press,

2010.

[103] N. Vucic and H. Boche, “Robust QoS-Constrained optimization of downlink mul-

tiuser MISO systems,” IEEE Transactions on Signal Processing, vol. 57, no. 2, pp.
714–725, February 2009.

[104] B. Chalise, S. Shahbaz Panahi, A. Czylwik, and A. Gershman, “Robust downlink

beamforming based on outage probability specifications,” IEEE Transactions on
Wireless Communications, vol. 6, no. 10, pp. 3498–3503, October 2007.

[105] L. Correia, D. Zeller, O. Blume, D. Ferling, Y. Jading, I. Godor, G. Auer, and

L. Van der Perre, “Challenges and enabling technologies for energy aware mobile
radio networks,” IEEE Communications Magazine, vol. 48, no. 11, pp. 66–72,
November 2010.

[106] T. Gucluoglu and T. Duman, “Performance analysis of transmit and receive an-
tenna selection over flat fading channels,” Wireless Communications, IEEE Trans-
actions on, vol. 7, no. 8, pp. 3056–3065, August 2008.

[107] K. Dong, N. Prasad, X. Wang, and S. Zhu, “Adaptive antenna selection and Tx/Rx
beamforming for large-scale MIMO systems in 60 GHz channels,” EURASIP Jour-
nal on Wireless Communications, vol. 2011, no. 59, August 2011.

[108] B. Tae-Won and J. Bang Chul, “A practical antenna selection technique in mul-
tiuser massive MIMO networks,” IEICE Transactions on Communications, vol.
E96.B, no. 11, pp. 2901–2905, 2013.

[109] X. Gao, O. Edfors, J. Liu, and F. Tufvesson, “Antenna selection in measured mas-
sive MIMO channels using convex optimization,” 2013 IEEE Globecom Workshops
(GLOBECOM), pp. 129–134, 2013.

[110] H. Li, L. Song, and M. Debbah, “Energy efficiency of large-scale multiple antenna
systems with transmit antenna selection,” IEEE Transactions on Communica-
tions, vol. 62, no. 2, pp. 638–647, February 2014.

[111] B. M. Lee, J. Choi, J. Bang, and B.-C. Kang, “An energy efficient antenna selection
for large scale green mimo systems,” IEEE International Symposium on Circuits
and Systems (ISCAS), pp. 950–953, May 2013.

[112] S. Mahboob, R. Ruby, and V. Leung, “Transmit antenna selection for downlink
transmission in a massively distributed antenna system using convex optimiza-
tion,” International Conference on Broadband, Wireless Computing, Communica-
tion and Applications (BWCCA), pp. 228–233, November 2012.

170
 LIST OF REFERENCES

[113] G. Zheng, I. Krikidis, C. Masouros, S. Timotheou, D. A. Toumpakaris, and

Z. Ding, “Rethinking the role of interference in wireless networks,” IEEE Com-
munications Magazine, vol. 52, no. 11, pp. 152–158, November 2014.

[114] A. Müller, A. Kammoun, E. Björnson, and M. Debbah, “Linear precoding based

on polynomial expansion : Reducing complexity in massive MIMO,” IEEE Trans-
actions on Information Theory, 2014, Submitted.

[115] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge University Press,

2004.

[116] Z. Chen, J. Yuan, B. Vucetic, and Z. Zhou, “Performance of alamouti scheme with
transmit antenna selection,” Electronics Letters, vol. 39, no. 23, pp. 1666–1668,
November 2003.

[117] P.-H. Lin and S.-H. Tsai, “Performance analysis and algorithm designs for transmit
antenna selection in linearly precoded multiuser MIMO systems,” IEEE Transac-
tions on Vehicular Technology, vol. 61, no. 4, pp. 1698–1708, May 2012.

[118] J. W. Demmel, Applied numerical linear algebra. Society for Industrial and
Applied Mathematics, 1997.

[119] A. Pitarokoilis, S. Mohammed, and E. Larsson, “On the optimality of single-carrier

transmission in Large-Scale Antenna systems,” IEEE Wireless Communications
Letters, vol. 1, no. 4, pp. 276–279, August 2012.

[120] C. Campbell and D. Dumka, “Wideband high power gan on sic spdt switch
mmics,” IEEE MTT-S International Microwave Symposium Digest (MTT), pp.
1–1, May 2010.

[121] “Agilent solid state switches (White paper),” Application Note, 2010.

[122] M. Di Renzo, D. De Leonardis, F. Graziosi, and H. Haas, “Space Shift Keying

(SSK) MIMO with practical channel estimates,” IEEE Transactions on Commu-
nications, vol. 60, no. 4, pp. 998–1012, April 2012.

[123] J. Ney da Silva and M. de Campos, “Spectrally efficient UWB pulse shaping with
application in orthogonal PSM,” IEEE Transactions on Communications, vol. 55,
no. 2, pp. 313–322, February 2007.

[124] B. Parr, B. Cho, K. Wallace, and Z. Ding, “A novel ultra-wideband pulse design
algorithm,” IEEE Communications Letters, vol. 7, no. 5, pp. 219–221, May 2003.

[125] X. Wu, Z. Tian, T. Davidson, and G. Giannakis, “Optimal waveform design for
UWB radios,” IEEE Transactions on Signal Processing, vol. 54, no. 6, pp. 2009–
2021, June 2006.

171
LIST OF REFERENCES

[126] A. Mohammadi and F. Ghannouchi, “Single RF front-end MIMO transceivers,”

IEEE Communications Magazine, vol. 49, no. 12, pp. 104–109, December 2011.

[127] M. Di Renzo, H. Haas, A. Ghrayeb, S. Sugiura, and L. Hanzo, “Spatial Mod-

ulation for generalized MIMO: Challenges, opportunities, and implementation,”
Proceedings of the IEEE, vol. 102, no. 1, pp. 56–103, January 2014.

[128] O. N. Alrabadi, C. B. Papadias, A. Kalis, and R. Prasad, “A universal encoding

scheme for mimo transmission using a single active element for psk modulation
schemes,” IEEE Transactions on Wireless Communications, vol. 8, no. 10, pp.
5133–5142, September 2009.

[129] N. Serafimovski, A. Younis, R. Mesleh, P. Chambers, M. Di Renzo, C.-X. Wang,

P. Grant, M. Beach, and H. Haas, “Practical implementation of spatial modula-
tion,” IEEE Transactions on Vehicular Technologies, vol. 62, no. 9, pp. 4511–4523,
November 2013.

[130] J. Perruisseau-Carrier, O. Alrabadi, and A. Kalis, “Implementation of a reconfig-

urable parasitic antenna for beam-space bpsk transmissions,” European Microwave
Conference (EuMC), pp. 644–647, September 2010.

[131] C. Masouros, M. Sellathurai, and T. Ratnarajah, “Vector perturbation based on

symbol scaling for limited feedback MISO downlinks,” IEEE Transactions on Sig-
nal Processing, vol. 62, no. 3, pp. 562–571, February 2014.

[132] D. Ha, K. Lee, and J. Kang, “Energy efficiency analysis with circuit power
consumption in massive MIMO systems,” IEEE 24th International Symposium
on Personal Indoor and Mobile Radio Communications (PIMRC), pp. 938–942,
September 2013.

[133] X. Gao, O. Edfors, F. Tufvesson, and E. G. Larsson, “Massive mimo in real prop-
agation environments: Do all antennas contribute equally?” IEEE Transactions
on Communications, vol. 63, no. 11, pp. 3917–3928, November 2015.

[134] P. V. Amadori and C. Masouros, “Interference driven antenna selection for Massive
Multi-User MIMO,” IEEE Transactions on Vehicular Technology, vol. PP, no. 99,
pp. 1–1, 2015.

[135] A. Garcia-Rodriguez, C. Masouros, and L. Hanzo, “Pre-Scaling optimization for

Space Shift Keying based on Semidefinite Relaxation,” IEEE Transactions on
Communications, vol. 63, no. 11, pp. 4231–4243, November 2015.

[136] G. Zheng, K.-K. Wong, and T.-S. Ng, “Robust linear mimo in the downlink: A
worst-case optimization with ellipsoidal uncertainty regions,” EURASIP Journal
on Advances in Signal Processing., 2008.

172
 LIST OF REFERENCES

[137] F. Rusek, D. Persson, B. K. Lau, E. Larsson, T. Marzetta, O. Edfors, and

F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with very large
arrays,” IEEE Signal Processing Magazine, vol. 30, no. 1, pp. 40–60, January 2013.

[138] A. Hakkarainen, J. Werner, K. R. Dandekar, and M. Valkama, “Widely-linear

beamforming and RF impairment suppression in massive antenna arrays,” Journal
of Communications and Networks, vol. 15, no. 4, pp. 383–397, August 2013.

[139] S. Cui, A. J. Goldsmith, and A. Bahai, “Energy-efficiency of MIMO and cooper-

ative MIMO techniques in sensor networks,” IEEE Journal on Selected Areas in
Communications, vol. 22, no. 6, pp. 1089–1098, August 2004.

[140] A. S. Y. Poon, R. W. Brodersen, and D. Tse, “Degrees of freedom in multiple-

antenna channels: A signal space approach,” IEEE Transactions on Information
Theory, vol. 51, no. 2, pp. 523–536, February 2005.

[141] M. Sadek, A. Tarighat, and A. H. Sayed, “Active antenna selection in multiuser

MIMO communications,” IEEE Transactions on Signal Processing, vol. 55, no. 4,
pp. 1498–1510, April 2007.

[142] I. Berenguer, X. Wang, and V. Krishnamurthy, “Adaptive MIMO antenna selec-

tion via discrete stochastic optimization,” IEEE Transactions on Signal Process-
ing, vol. 53, no. 11, pp. 4315–4329, November 2005.

[143] A. Gorokhov, “Antenna selection algorithms for MEA transmission systems,”

IEEE ICASSP, vol. 3, no. 2, pp. 2857–2860, May 2002.

[144] H. Henderson and S. Searle, “On deriving the inverse of a sum of matrices,” Siam
Review, vol. 23, no. 1, pp. 53–60, January 1981.

[145] D. Curd, “Power consumption in 65 nm FPGAs,” Xilinx White Paper, February

2007.

[146] R. Hunger, Floating point operations in matrix-vector calculus. Munich University

of Technology, Inst. for Circuit Theory and Signal Processing Munich, 2005.

[147] H. Prabhu, J. Rodrigues, O. Edfors, and F. Rusek, “Approximative matrix inverse

computations for very-large mimo and applications to linear pre-coding systems,”
IEEE Wireless Communications and Networking Conference (WCNC), pp. 2710–
2715, April 2013.

[148] S. L. H. Nguyen and A. Ghrayeb, “Compressive sensing-based channel estima-

tion for massive multiuser mimo systems,” IEEE Wireless Communications and
Networking Conference (WCNC), pp. 2890–2895, April 2013.

173
[149] T. S. Rappaport, J. N. Murdock, and F. Gutierrez, “State of the art in 60-ghz
integrated circuits and systems for wireless communications,” Proceedings of the
IEEE, vol. 99, no. 8, August 2011.

[150] M. Alodeh, S. Chatzinotas, and B. Ottersten, “Symbol-Level Multiuser

MISO Precoding for Multi-level Adaptive Modulation: A Multicast View,”
arXiv:1601.02788., 2016.

[151] R. Méndez-Rial, C. Rusu, N. González-Prelcic, A. Alkhateeb, and R. W. Heath,

“Hybrid MIMO architectures for millimeter wave communications: Phase shifters
or switches?” IEEE Access, vol. 4, pp. 247–267, 2016.

[152] S. M. Razavi, T. Ratnarajah, and C. Masouros, “Transmit-power efficient lin-

ear precoding utilizing known interference for the multiantenna downlink,” IEEE
Transactions on Vehicular Technology, vol. 63, no. 9, pp. 4383 – 4394, November
2014.

174