Politecnico di Torino

Porto Institutional Repository

[Doctoral thesis] Algorithms for 5G physical layer

Original Citation:
Pawan Dhakal (2017). Algorithms for 5G physical layer. PhD thesis
Availability:
This version is available at : http://porto.polito.it/2670627/ since: May 2017
Published version:
DOI:10.6092/polito/porto/2670627
Terms of use:
This article is made available under terms and conditions applicable to Open Access Policy Arti-
cle ("Creative Commons: Attribution 3.0") , as described at http://porto.polito.it/terms_and_
conditions.html
Porto, the institutional repository of the Politecnico di Torino, is provided by the University Library
and the IT-Services. The aim is to enable open access to all the world. Please share with us how
this access benefits you. Your story matters.

(Article begins on next page)

 Doctoral Dissertation
Doctoral Program in Electronics Engineering (29th cycle)

Algorithms for 5G Physical Layer

By

Pawan Dhakal
******

Supervisor(s):
Prof. Roberto Garello, Supervisor

Doctoral Examination Committee:

Prof. Paola Pierleoni, University of Marche,
Prof. Fulvio Babich, University of Trieste,
Prof. Francesca Vatta, University of Trieste,
Prof. Stefano Mangione, University of Palermo

Politecnico di Torino
2017
 Declaration

I hereby declare that, the contents and organization of this dissertation constitute my
own original work and does not compromise in any way the rights of third parties,
including those relating to the security of personal data.

Pawan Dhakal
2017

* This dissertation is presented in partial fulfillment of the requirements for Ph.D.

degree in the Graduate School of Politecnico di Torino (ScuDo).
I would like to dedicate this thesis to my loving parents, my wife and my son.
 Acknowledgements

I take immense pleasure in thanking Department of Electronics and Telecommuni-

cation Engineering, Politecnico di Torino, Italy, for providing me an opportunity to
carry out this PhD Research during the years 2014/2016.
My deepest gratitude to my advisor, Prof. Roberto Garello, for the able guidance
and for his invaluable professional and human support in all possible ways ever
since.
Then, I am grateful to Prof. Anand Gachhadar for hosting me at Kathmandu
University, Nepal. Special thanks to all my former colleagues at Department of
Electrical and Electronics Engineering of Kathmandu University for the friendly
and welcoming atmosphere that really made me feel at home. Furthermore, I want
to sincerely thank my friend and colleague Dr. Daniel Riviello for our fruitful and
lasting collaboration, Mr. Research Dawadi for his help in the reviewing process.
Special thanks to Adriana Scavello and Franco Mantione for their moral support and
deepest concern, and last but not least, my family and everyone who supported me
during these years.
 Abstract

There is a great activity in the research community towards the investigations of the
various aspects of 5G at different protocol layers and parts of the network. Among
all, physical layer design plays a very important role to satisfy high demands in
terms of data rates, latency, reliability and number of connected devices for 5G
deployment. This thesis addresses the latest developments in the physical layer
algorithms regarding the channel coding, signal detection, frame synchronization
and multiple access technique in the light of 5G use cases. These developments are
governed by the requirements of the different use case scenarios that are envisioned
to be the driving force in 5G.
All chapters from chapter 2 to 5 are developed around the need of physical layer
algorithms dedicated to 5G use cases. In brief, this thesis focuses on design, analysis,
simulation and the advancement of physical layer aspects such as 1. Reliability based
decoding of short length Linear Block Codes (LBCs) with very good properties in
terms of minimum hamming distance for very small latency requiring applications. In
this context, we enlarge the grid of possible candidates by considering, in particular,
short length LBCs (especially extended BCH codes) with soft-decision decoding;
2. Efficient synchronization of preamble/postamble in a short bursty frame using
modified Massey correlator; 3. Detection of Primary User activity using semi-
blind spectrum sensing algorithms and analysis of such algorithms under practical
imperfections; 4. Design of optimal spreading matrix for a Low Density Spreading
(LDS) technique in the context of non-orthogonal multiple access. In such spreading
matrix, small number of elements in a spreading sequences are non zero allowing
each user to spread its data over small number of chips (tones), thus simplifying the
decoding procedure using Message Passing Algorithm (MPA).
 Publicationss

Some of the contents of this thesis have previously appeared or will appear in the
following publications:

Book Chapter

1. [49] Pawan Dhakal, Shree K. Sharma, Symeon Chatzinotas, Björn Otter-

sten and Daniel Riviello, “Effect of Primary User Traffic on Largest Eigen-
value Based Spectrum Sensing Technique”, in Cognitive Radio Oriented Wire-
less Networks, 11th International Conference, CROWNCOM 2016, Grenoble,
France, May 30 - June 1, 2016, Proceedings, Springer International Publish-
ing, pp. 67-78, 2016. ISBN: 978-3-319-40351-9, http://dx.doi.org/10.1007/
978-3-319-40352-6_6.

Journal

1. [22] M. Baldi and M. Bertinelli and F. Chiaraluce and P. Dhakal and R. Garello
and N. Maturo and E. Paolini, “A Study on State-of-the-Art Space Mission
Telecommand Receivers”, In IEEE Aerospace and Electronic Systems Maga-
zine, 2017. In Press.

Conference Proceedings

1. [48] Pawan Dhakal, Roberto Garello, Shree Krishna Sharma, Symeon Chatzino-
tas and Bjorn Ottersten, “On the Error Performance Bound of Ordered Statis-
 vii

tics Decoding of Linear Block Codes”, IEEE Intl. Conf. on Commun. 2016,
Jun 2016, http://dx.doi.org/10.1109/ICC.2016.7510798.

2. [44] Pawan Dhakal, Daniel Riviello, “Multi-Antenna Energy Detector Under

Unknown Primary User Traffic”, COCORA 2016, The Fourth Intl. Conf. on
Advances in Cognitive Radio, Feb. 2016. http://www.thinkmind.org/download.
php?articleid=cocora_2016_1_40_80032.

3. [146] Daniel Riviello, Pawan Dhakal, Roberto Garello, “Performance Anal-

ysis of Multi-Antenna Hybrid Detectors and Optimization with Noise Vari-
ance Estimation”, International Conference on Advances in Cognitive Radio
(COCORA), April 2015, http://www.thinkmind.org/download.php?articleid=
cocora_2015_1_30_80029.

4. [147] Daniel Riviello, Pawan Dhakal, Roberto Garello, “On the Use of Eigen-
vectors in Multi-Antenna Spectrum Sensing with Noise Variance Estimation”,
2nd Intel. Conf. in Signal Processing and Integrated Networks, Feb. 2015,
http://dx.doi.org/10.1109/SPIN.2015.7095339.

5. [45] Pawan Dhakal, Daniel Riviello, Roberto Garello, “SNR Wall Analysis
of Multi-Sensor Energy Detection with Noise Variance Estimation”, 11th
IEEE Intl. Symp. on Wireless Commun. Systems (ISWCS), Aug. 2014,
http://dx.doi.org/10.1109/ISWCS.2014.6933440.

6. [46] Pawan Dhakal, Roberto Garello, Federico Penna and Daniel Riviello,
“Impact of Noise Estimation on Energy Detection and Eigenvalue Based Spec-
trum Sensing Algorithms”, IEEE Intl. Conf. on Commun. 2014, Jun 2014,
http://dx.doi.org/10.1109/ICC.2014.6883512.

7. [47] Pawan Dhakal, Daniel Riviello, Roberto Garello and Federico Penna,
“Hybrid Approach Analysis of Energy Detection and Eigenvalue Based Spec-
trum Sensing Algorithms with Noise Power Estimation”, COCORA 2014,
The Fourth Intl. Conf. on Advances in Cognitive Radio, Feb. 2014, http:
//www.thinkmind.org/download.php?articleid=cocora_2014_1_40_80029.
Contents

List of Figures xii

List of Tables xvii

Acronyms xviii

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Structure and Contribution of This Thesis . . . . . . . . . . . . . . 11

2 Ordered Statistics Decoding of Linear Block Codes 13

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Channel Coding for 5G . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Ordered Statistics Decoding . . . . . . . . . . . . . . . . . . . . . 19
2.4 Overview of OSD and Conventional Reprocessing . . . . . . . . . . 21
2.5 Proposed New Statistic of Ordered Vector Components . . . . . . . 23
2.6 OSD Error Performance . . . . . . . . . . . . . . . . . . . . . . . 27
2.7 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.7.1 Pmf of EL . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.7.2 Error Performance . . . . . . . . . . . . . . . . . . . . . . 29

3 Frame Synchronization 32
Contents ix

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.1 Machine to Machine Communication . . . . . . . . . . . . 32
3.2 Introduction to Frame Synchronization Problem . . . . . . . . . . . 34
3.3 System Model and Problem discussion . . . . . . . . . . . . . . . . 35
3.4 Likelihood Ratio Test . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4.1 Equivalency of Massey Formulation and Chiani/Martini for-
mulation of LRT . . . . . . . . . . . . . . . . . . . . . . . 38
3.5 Hard and Soft Correlation . . . . . . . . . . . . . . . . . . . . . . . 39
3.6 The Massey Correlation Detector . . . . . . . . . . . . . . . . . . . 40
3.7 Analytical computation of the detector performances under H1 . . . 41
3.7.1 Miss-Detection Probability . . . . . . . . . . . . . . . . . 45
3.8 Analytical computation of the detector performances under H0 . . . 45
3.8.1 False-Alarm Probability . . . . . . . . . . . . . . . . . . . 50
3.8.2 False-alarm lower bound . . . . . . . . . . . . . . . . . . . 50
3.9 ROC curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.10 The Extended Massey correlation detector . . . . . . . . . . . . . . 54
3.10.1 Analytical computation of the Extended Massey detector
pdf under H1 . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.10.2 Analytical computation of the Extended Massey detector
pdf under H0 . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.10.3 Extended Massey detector performances . . . . . . . . . . . 59
3.11 Comparison against LRT, soft and hard correlation . . . . . . . . . 60
3.11.1 Miss-detection vs. SNR performances . . . . . . . . . . . . 63
3.12 Some Aspects for Practical Implementation . . . . . . . . . . . . . 65

4 Cognitive Radio: Spectrum Sensing 66

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.1.1 Cognitive Radio Technology . . . . . . . . . . . . . . . . . 70
x Contents

4.1.2 Spectrum Sensing . . . . . . . . . . . . . . . . . . . . . . 72

4.1.3 Spectrum sensing under practical imperfection . . . . . . . 74
4.2 Semi-blind SS techniques under noise uncertainty . . . . . . . . . . 75
4.2.1 Energy Detection and its performance under Noise Uncer-
tainty (State-of-art) . . . . . . . . . . . . . . . . . . . . . . 76
4.2.2 Eigenvalue Based Detection and its performance in Noise
Uncertainty (State-of-art) . . . . . . . . . . . . . . . . . . . 77
4.2.3 Model Scenario . . . . . . . . . . . . . . . . . . . . . . . . 78
4.2.4 Mathematical framework . . . . . . . . . . . . . . . . . . . 79
4.2.5 Detection Performance Parameters . . . . . . . . . . . . . . 81
4.2.6 Maximum Likelihood Noise Estimation Model . . . . . . . 82
4.2.7 Energy Detection . . . . . . . . . . . . . . . . . . . . . . . 85
4.2.8 Hybrid ED (HED1) . . . . . . . . . . . . . . . . . . . . . . 93
4.2.9 Hybrid ED-2(HED2) . . . . . . . . . . . . . . . . . . . . . 101
4.2.10 RLRT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.2.11 Hybrid RLRT (HRLRT1) . . . . . . . . . . . . . . . . . . . 115
4.2.12 Hybrid RLRT-2(HRLRT2) . . . . . . . . . . . . . . . . . . 120
4.2.13 SNR Wall Analysis of ED and HED1 . . . . . . . . . . . . 127
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 136
4.3.1 Characterization of Primary user traffic . . . . . . . . . . . 139
4.3.2 ED under Unknown PU Traffic . . . . . . . . . . . . . . . . 144
4.3.3 EBD under Unknown PU Traffic . . . . . . . . . . . . . . . 150

5 Low Density Spreading (LDS) Multiplexing for NOMA 159

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.2 Multiplexing with Low Density Spreading . . . . . . . . . . . . . . 162
5.3 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
5.4 Calculation of average energy per bit (Eb ) . . . . . . . . . . . . . . 166
Contents xi

5.5 Spreading Matrix Design . . . . . . . . . . . . . . . . . . . . . . . 169

5.5.1 Binary Matrix Design . . . . . . . . . . . . . . . . . . . . 169
5.5.2 Design of non-zero symbols . . . . . . . . . . . . . . . . . 170
5.5.3 Distribution of non-zero symbols in a binary matrix . . . . . 178
5.6 Optimal Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
5.7 Simulation Result . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
5.7.1 A6,4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
5.7.2 A9,6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

6 Summary and Conclusions 187

References 189
List of Figures

1.1 Evolution of current MBB services towards beyond-2020. [66] . . . 3

1.2 Enhancement of key capabilities from LTE-Advanced to IMT-2020.
[135] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Enhancement of key capabilities from LTE-Advanced to IMT-2020.
[111] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1 Probability density function of ri for Eb /N0 = 3.5 dB. . . . . . . . . 26

2.2 Probability density function of E: Eb /N0 = 3.5 dB, L = k. . . . . . 29
2.3 Code Error Rate of OSD with order−I reprocessing for (64,42,8)
Reed Muller code . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Code Error Rate of OSD with order−I reprocessing for (128,99,10)
extended BCH code . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5 Code Error Rate of OSD with order−I reprocessing for (128,64,22)
extended BCH code . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.1 The binary pattern recognition problem under study. . . . . . . . . . 36

3.2 The frame synchronization problem. . . . . . . . . . . . . . . . . . 36
3.3 Pdf of random variable y for SNR = 0 dB. The red dotted curve
fg (x) is the original Gaussian random variable centered in ci = +1. . 43
3.4 Massey correlation detector: pdfs under H1 and H0 for N = 8 at
SNR = 0 dB. Analytical (a) vs. simulated (s) curves. . . . . . . . . 44
3.5 Massey correlation detector: pdfs under H1 and H0 for N = 64 for
SNR = 0 dB. Analytical (a) vs. simulated (s) curves. . . . . . . . . 44
List of Figures xiii

3.6 Massey correlation detector: miss-detection and false-alarm per-

formances, for N = 8 bits and SNR = 0 dB. Analytical (a) and
simulated (s) curves. . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.7 Massey correlation detector: miss-detection and false-alarm per-
formances, for N = 64 bits and SNR = 0 dB. Analytical (a) and
simulated (s) curves. . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.8 Pdf of random variable z for SNR = 0 dB. The red dotted curve
fg (x) is the original Gaussian random variable centered in di = −1. . 49
3.9 Massey correlation detector: ROC curves for N=8 bits and different
SNR values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.10 Massey correlation detector: ROC curves for N=64 bits and different
SNR values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.11 Massey correlation detector: ROC curves for SNR = 0 dB and dif-
ferent pattern lengths. . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.12 Extended Massey correlation detector: Pf a and Pmd vs. threshold,
for N = 8 bits and SNR = 6 dB. . . . . . . . . . . . . . . . . . . . 61
3.13 ROC curves for N = 8 bits and SNR = 0 dB: comparison between
Extended Massey detector, Massey detector, LRT, Soft correlation,
Hard correlation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.14 ROC curves for N = 64 bits and SNR = 0 dB: comparison between
Massey detector, LRT, Soft correlation, Hard correlation. . . . . . . 62
3.15 Extended Massey detector: Pmd vs. SNR, for N = 8 bits and Pf a =
10−2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.16 Massey detector: Pmd vs. SNR, for N = 64 bits and Pf a = 10−6 . . . 64

4.1 OFCOM frequency allocation chart for UK [57] . . . . . . . . . . . 69

4.2 Probability of false alarm PFa and Probability of Miss-detection PD 82
4.3 Block diagram of Maximum Likelihood Noise Estimation . . . . . 83
4.4 Hybrid approach-1: Noise variance estimated from S auxiliary noise
only slots. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
xiv List of Figures

4.5 Hybrid approach-2: Noise variance estimated in S auxiliary slots

from received signal samples declared only noise samples by ED/RLRT 85
4.6 χN2 distribution with 5, 10, and 20 degrees of freedom. A normal
distribution is superimposed to illustrate a good approximation to
χN2 by NR (20, 40) for N large . . . . . . . . . . . . . . . . . . . . . 88
4.7 Comparison of Chi Square Distribution with Gaussian approximation
for different N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.8 MSE of the approximation considering Probability Density Function
approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.9 ROC plot of ED for Normal Approximation of the Decision Statistics
for N = 100, and SNR = −10dB for varying number of sensors K . 92
4.10 Probability of Miss vs. SNR for varying number of sensors. Parame-
ters: N = 50, PFa = 0.05 and K = 5, 10 and 15 . . . . . . . . . . . . 93
4.11 Probability of Detection vs. SNR for varying amount of noise un-
σ̂ 2
certainty measured as β = σv2 , and x = 10 log10 (β ), Parameters:
v
N = 40, PFa = 0.1 and K = 1, 5 and S . . . . . . . . . . . . . . . . 94
4.12 ROC of Multi Sensor HED1, Parameters: N = 100, M = 100, K = 5,
S = 1, 2, 5, 10, and SNR = −10 dB . . . . . . . . . . . . . . . . . . 100
4.13 Probability of Miss vs SNR, Parameters: N = 50, M = 50, K = 5,
S = 2, 8, and SNR = −10dB . . . . . . . . . . . . . . . . . . . . . 100
4.14 Comparison of simulated and analytical detection performance curves
of HED2 method for N = 10, M = 10, K = 5, S = 5 and 50 . . . . . 109
4.15 Comparison of performance curves of HED2, ED and HED1 for
N = 10, M = 10, K = 5, S = 2, 5 and 10 . . . . . . . . . . . . . . . 110
4.16 Comparison of simulated and analytical ROC performance curves
of RLRT method for N = 50, K = 8 and 10, SNR = −10 dB . . . . 113
4.17 PD vs. SNR performance of RLRT. Parameters: N = 100, K = 5, 7,
PFa = 0.05 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.18 Effect of Noise Variance fluctuation on ED and RLRT. Parameters:
N = 100, K = 4, var(σ̂v2 ) = 0.0032(−25dB) given nominal variance
σv2 = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
List of Figures xv

4.19 Comparison of simulated and analytical performance curves of

HRLRT1 method for N = 80, K = 4 and S = 2 8, SNR = −10 dB . . 119
4.20 ROC performance of RLRT and ED methods compared with their
hybrid counterparts. Parameters: N = 80, K = 4 and S = 2, 5, 10,
SNR = −10 dB . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.21 Comparison of HRLRT1 and HED1 performance curves. Parameters:
N = 80, K = 4 and S = 2 and 15, SNR = −10 dB . . . . . . . . . . 121
4.22 Comparison of simulated and analytical performance curves of
HRLRT2 method. Parameters: N = 80, K = 4 and S = 2 and 8,
SNR = −10 dB . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.23 Probability of detection vs. SNR for RLRT, HRLRT1 and HRLRT2,
Parameters: N = 80, K = 4 and S = 2 and 15, SNR = −10 to −4 dB 127
4.24 Comparison of sample complexity N to achieve given PD′ and PFa ′

for single antenna vs. multi antenna environment for different SNR
assuming perfect knowledge of noise variance. Parameters: K =
5, 10, 15 and 20, P’D = 0.9, P’Fa = 0.1 . . . . . . . . . . . . . . . 129
4.25 Variation of sample complexity N as the SNR approaches SNR wall
for ED, Parameters: K = 5, P’D = 0.9, P’Fa = 0.1 [use x = 10 log10 β
in Equation (4.174)] . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.26 Probability distribution of normalized noise variance estimate V .
Parameters: S = 2, K = 2, M = 10 . . . . . . . . . . . . . . . . . . 132
4.27 Variation of noise uncertainty level with number of slots S used for
noise variance estimation. . . . . . . . . . . . . . . . . . . . . . . . 134
4.28 Variation of SNR wall level with number of auxiliary slots S used
for noise variance estimation . . . . . . . . . . . . . . . . . . . . . 135
4.29 Variation of sample complexity N as the SNR approaches SNR
wall for varying number of auxiliary slots used for noise variance
estimation in HED1 . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.30 Primary user traffic scenario and sensing slot classification . . . . . 139
4.31 Pdfs of the ED decision statistic: Parameters: N = 50, K = 4, M f =
150, Mb = 150 and SNR = −6 dB . . . . . . . . . . . . . . . . . . 145
xvi List of Figures

4.32 ROC performance for the considered scenario. Parameters: N = 100,

K = 4 and SNR = −6 dB . . . . . . . . . . . . . . . . . . . . . . . 149
4.33 Probability of missed detection vs. the number of Antennas, Param-
eters: N = 25, M f = 62, Mb = 62 and PF = 0.1 . . . . . . . . . . . 150
4.34 Sensing performance of RLRT under unknown PU traffic. . . . . . . 156
4.35 RLRT under unknown PU traffic sensing performance comparison
for different sensing parameters (N and K). . . . . . . . . . . . . . 158

5.1 Factor Graph representation of LDS spreading . . . . . . . . . . . . 164

5.2 Regular LDS spreading system . . . . . . . . . . . . . . . . . . . . 166
5.3 Euclidean distance relationship . . . . . . . . . . . . . . . . . . . . 168
5.4 Probability of a BPSK bit error for A6,4 . . . . . . . . . . . . . . . 184
5.5 Probability of a BPSK bit error for A9,6 . . . . . . . . . . . . . . . 186
List of Tables

1.1 Features of 3GPP standards up to now . . . . . . . . . . . . . . . . 2

2.1 Turbo codes and duo-binary turbo codes in standards. . . . . . . . . 16

3.1 Massey correlation detector: false-alarm lower bound values for

different lengths N. . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4.1 Potential Spectrum-Related Implications of various 5G Require-

ments [15] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2 Receiver Parameter for 802.22 WRAN [3] . . . . . . . . . . . . . . 73

5.1 Relationship of A and B with input x and y. . . . . . . . . . . . . . 172

5.2 Set of angles satisfying (5.43), (5.44) and (5.45). . . . . . . . . . . 176
5.3 Inequality check of angle sets. . . . . . . . . . . . . . . . . . . . . 177
5.4 Set of difference angles satisfying inequalities presented in (5.55),
(5.56), (5.57) and (5.59) . . . . . . . . . . . . . . . . . . . . . . . . 178
Acronyms

Acronyms / Abbreviations

3GPP 3rd Generation Partnership Project

5GNOW 5th Generation Non-Orthogonal Waveforms for Asynchronous Signalling

AW GN Additive White Gaussian Noise

BCH Bose, Chaudhuri, and Hocquenghem

BER Bit Error Rate

BP Belief Propagation

BPSK Bipolar Shift Keying

CCSDS Consultative Committee for Space Data Systems

cd f Cumulative Distribution Function

CDMA Code Division Multiple Access

CDR Constant Detection Rate

CFAR Constant False Alarm Rate

CP Cyclic Prefix

CRSC Circular Recursive Systematic Constituent

DV B − RCS Digital Video Broadcasting - Return Channel via Satellite

DV B − T Digital Video Broadcasting - Terrestrial

Acronyms xix

DySPAN Dynamic Spectrum Access Networks

eBCH Extended BCH

EBD Eigenvalue Based Detection

ED Energy Detection

ENP Estimated Noise Power

EOF End of Frame

ERD Eigenvalue Ratio Detection

ESA European Space Agency

EU European Union

FBMC Filter Bank Multi-Carrier

FDMA Frequency Division Multiple Access

FFT Fast Fourier Transform

GF Galois Field

GFDM Generalized Frequency Division Multiple Access

GLRT Generalized Root Test

GSM Global System for Mobile Communications

ICT Information and Communication Technologies

IFFT Inverse Fast Fourier Transform

IID Independent and Identically Distributed

IMT − A International Mobile Telecommunication - advanced

IOT Internet of Things

ISM Industrial Scientific and Medical

ITU − R International Telecommunication Union - Radio Communication Sector

xx Acronyms

LBC Linear Block Codes

LDPC Low Density Parity Check Codes

LDPC −CC LDPC-Convolution Code

LDS Low Density Spreading

LRT Likelihood Ratio Test

LT E − A Long Term Evolution - Advanced

M2M Machine to Machine

MAI Multiple Access Interference

MAP Maximum a Posteriori

MAS Moving Average System

MDL Minimum Description Length

MET IS Mobile and Wireless Communications Enablers for Twenty-twenty Infor-

mation Society

MIMO Multiple Input Multiple Output

ML Maximum Likelihood

MLD Maximum Likelihood Decoding

mMTC Massive Machine Type Communication

MPA Message Passing Algorithm

MPEG Moving Picture Experts Group

MRI Most Reliable Independent

MSE Mean Square Error

MUD Multi-User Detection

NGMN Next Generation Mobile Networks

Acronyms xxi

NOMA Non-Orthogonal Multiple Access

NP Neyman Pearson

NTA Nepal Telecom Authority

OFCOM Office of communications

OFDM Orthogonal Frequency Division Multiple Access

OoB Out of Band

OSA Opportunistic Spectrum Access

OSD Ordered Statistics Decoding

pd f Probability Density Function

pm f Probability Mass Function

RAT Radio Access Technology

RLRT Roy’s Largest Root Test

RMT Random Matrix Theory

ROC Receiver Operating Characteristics

ROF Rank Order Filtering

RS Reed Solomon

RSMA Resource Spread Multiple Access

SC − LDPC Spatially Coupled LDPC

SCMA Sparse Code Multiple Access

SINR Signal to Interference Plus Noise Ratio

SNR Signal to Noise Ratio

SOF Start of Frame

SW Synchronization Word
xxii Acronyms

T DMA Time division Multiple Access

T EP Test Error Pattern

TTI Transmission Time Interval

UE User Equipment

UFMC Universal Filtered Multi-Carrier

UHF Ultra High Frequency

UMP Uniformly Most Powerful

UMT S Universal Mobile Telecommunications System

V NI Visual Network Index

WiMaX Worldwide Interoperability for Microwave Access

W LAN Wireless Local Area Network

W PAN Wireless Personal Area Network

W RAN Wireless Regional Area Network

W RC World Radio Conference

W SN Wireless Sensor Network

xMBB Extreme Mobile Broadband

Chapter 1

Introduction

1.1 Motivation

Wireless communication networks are perhaps one of the most important element
in the global Information and Communication Technologies (ICT) strategy, under-
pinning many other industries. It is one of the fastest growing and most dynamic
sectors in the world. The phenomenal success of wireless mobile communications
is mirrored by a rapid pace of technology innovation. From first generation (1G)
analog FM cellular systems in 1981 to fourth generations (4G) Long Term Evolu-
tion - Advanced in 2011, the wireless mobile network has transformed from a pure
telephony system to a network that can transport rich multimedia contents [27, 62].
Long Term Evolution (LTE) and its extension have now been deployed around the
globe to deliver high quality, low latency video and multimedia applications for
wireless devices. Table 1.1 illustrates the main features of the 3rd Generation Part-
nership Project (3GPP) standards now in the market, highlighting the trend towards
widespread use of spectrum, higher bandwidth, higher spreading efficiency and lower
latency.
In 4G systems, an advanced radio interface is used with orthogonal frequency-
division multiplexing (OFDM), Multiple-Input Multiple-Output (MIMO), and link
adaptation technologies. 4G wireless networks can support data rates of up to 1 Gb/s
for low mobility, such as nomadic/local wireless access, and up to 100 Mb/s for high
mobility, such as mobile access. However, it is widely agreed that similar to previous
technologies, the current technologies also follow a similar technology maturity
2 Introduction

Table 1.1 Features of 3GPP standards up to now

Properties GSM UMTS HSPA LTE

450, 800 850, 900 850, 900 700, 800, 850, 900
Band (MHz) 900, 1800 1700, 1900 1700, 1900 1700, 1800, 1900, 2100
1900 2100 2100 2500, 2600, 3500
Max. Band (MHz) 0.20 5 10 20
Peak Data Rate 9.6 Kbps 384 Kbps 42Mbps 326 Mbps
Round Trip Time 600 ms 75 ms 41 ms 20 ms

where only incremental improvements and small amounts of new spectrum can be
expected; where a mere evolution of the status quo cannot achieve the extreme higher
aggregate data rates and much lower latency required by next generation technology.
The development of wireless technologies has greatly improved people’s ability
to communicate and live in both business operations and social functions. Essential
services such as e-banking, e-learning, and e-health will continue to proliferate
and become more mobile. On-demand information and entertainment (e.g., in the
form of augmented reality) will progressively be delivered over mobile and wireless
communication systems. More and more people crave faster internet access on the
move, trendier mobile phones, and, in general, instant communication with others or
access to information. Every year, the demand in mobile broadband communications
is increasing dramatically as more and more users subscribe to mobile broadband
packages. In addition, smart phones, super-phones, tablets and laptops with powerful
multimedia capabilities and applications are becoming increasingly popular and
are creating new demands on mobile broadband. The possible evolution of Mobile
Broad-band (MBB) from IMT-2000 and IMT-Advanced towards 2020 and beyond is
as shown in Figure 1.1 [66], based on the assumption that the traffic requirements
would be increased by a thousand times over the next decade.
New data services and applications, for example pervasive 3D multimedia, are
making the experience of using mobile broadband better and more exciting. Further-
more, it is generally predicted that today’s dominating scenarios of human-centric
communication will, in the future, be complemented by a tremendous increase in
the numbers of communicating machines. This so-called Internet of Things will
make our everyday life more efficient, comfortable, and safe. These trends, shown in
Figure 1.2, are expected to maintain their momentum over the next decade and will
be complemented by the arrival of billions of machine devices and related machine-
1.1 Motivation 3

Fig. 1.1 Evolution of current MBB services towards beyond-2020. [66]

to-machine (M2M) applications [135]. These developments will lead to an avalanche

of mobile and wireless traffic volume, predicted to increase a thousand-fold over the
next decade [60].

Fig. 1.2 Enhancement of key capabilities from LTE-Advanced to IMT-2020. [135]

Thanks largely to the annual visual network index (VNI) reports released by
Cisco, we have quantitative evidence that the wireless data explosion is real and will
continue. Driven largely by smartphones, tablets, and video streaming, the most
recent (Feb. 2016) VNI report [39] shows, just in a decade, the amount of mobile
4 Introduction

data has increased by well over a factor of 4000-fold and almost 400-million fold
over the past 15 years. Mobile networks carried fewer than 10 gigabytes per month
in 2000, and less than 1 petabyte per month in 2005. In 2015, it was 3.7 exabytes per
month. Similarly, although 4G connections represented only 14 percent of mobile
connections in 2015, they already account for 47% of the mobile data traffic and for
the first time in 2015, 4G traffic exceeded 3G traffic. This deluge of data has been
driven chiefly by video thus far, but new unforeseen applications can reasonably be
expected to materialize by 2020. In addition to the sheer volume of data, the number
of devices and the data rates will continue to grow exponentially. Smart-phones
accounted for most of that growth. Global mobile devices and connections in 2015
grew to 7.9 billion, up from 7.3 billion in 2014.
Further, machine-type applications are becoming important in addition to the
human centric communications that have been dominating the cellular scene so far.
In fact, the number of communicating machine was some point forecasted to be
trending towards the number 50 billion by 2020. The expected uptake of machine-
type and human-type wireless communications in many economic sectors and vertical
industries will lead to a large and wide diversity of communication characteristics
imposing different requirements on mobile and wireless communication systems,
e.g., in terms of cost, complexity, energy dissipation, data rate, mobility, latency, and
reliability. On the other hand, 4G networks have just about reached the theoretical
limit on the data rate with current technologies and therefore are not sufficient to
accommodate the above challenges.
All the above issues are putting more pressure on cellular service providers, who
are facing continuously increasing demand for higher data rates, larger network
capacity, higher spectral efficiency, higher energy efficiency, and higher mobility
required by new wireless applications. In this sense, we need ground breaking
wireless technologies to solve the above problems caused by trillions of wireless
devices.
Foreseeing all these future requirements to meet, the International Mobile
Telecommunication (IMT) - 2020 was established in 2013, which is dedicated
to investigate beyond 4G (B4G) or 5G wireless techniques. The aim is to connect
the entire world, and achieve seamless and ubiquitous communications between
anybody (people to people), anything (people to machine, machine to machine),
wherever they are (anywhere), whenever they need (anytime), by whatever electronic
1.1 Motivation 5

devices/services/networks they wish (anyhow). This means that 5G networks should

also be able to support communications for some special scenarios not supported by
4G networks (e.g., for high-speed train users). 5G is considered to be the enabler of
fully connected network society. It is enabler of the world where there is unlimited
access to information and sharing of data any-time and anywhere to anyone and
anything. So it is not only about mobile broadband and mobile telephony, it is
about connectivity for any kind of application. It is not only for any kind of devices
and connectivity for people. It is about connectivity for any kind of devices and
connectivity should be available truly at any time and anywhere [177].
Amount of traffic carried by communication systems will increase tremendously.
In the future, we may see 1000 times more traffic than we see today. And we have to
ensure that wireless communication systems can really handle such enormous traffic.
In addition to this, number of devices will also increase tremendously. In the future,
we will probably see 100’s of billions of devices which will be very different from
small phones and tablets that we see today. Finally, our wireless communication
system have to handle a very large range of requirements for the future [59]. It will
not only be about high data rates; some applications may require small latency, some
applications may require ultra high reliability of the system, some applications may
require extremely low device cost and extremely low device energy consumption.
5G is the overall wireless access solution of the future and that will of-course
include revolution of currently existing Radio Access Technologies (RAT) including
LTE-A, WiFi, WiMax etc. But it will also include new complementary RAT that
will address use cases and scenarios that cannot be properly handled by current
RAT. And all these together, existing RAT in combination with complementary new
technologies that we refer to as 5G [59].
Let us look at the future wireless communication scenario. In the future, the
cellular technologies such as LTE, LTE-A will continue to be very important and
will provide mobile telephony and mobile broadband with wide area coverage. But
they will be extended by lots of new applications and lots of new use cases and new
scenarios. Some of which can be revolution of existing technologies but some of
which will need new technologies. Let us take a few examples: To provide extreme
data rates or extreme traffic capacities, for instance in indoor office environments of
in outdoor hot spot scenarios, ultra dense deployment may be needed.
6 Introduction

Machine type communications, will be more and more important in the future.
This will include communications for electronics and home appliances, communica-
tions for transports, communications for medical equipments and many more. A lot
of dense communications will probably be carried out by our currently existing wire-
less access technologies but some of these use cases will have so high requirements
that they will require new wireless access technologies.
Ultra reliable communications may be needed, for example, for critical infras-
tructure, or for industrial processes. And once again this may very well require
new radio access technologies. Car-to-car or car-to roadside communication will
be introduced for traffic safety and traffic control. There may be completely new
ways to communicate for example, communicating directly between devices. It is
all those together, the revolution of currently existing RAT and new complementary
RAT that is referred to as 5G.
The following spider diagram shown in Figure 1.3 is the best way to illustrate
a wide range and expansion of the 5G requirements in comparison to prior cellular
generations such as LTE-Advanced. Some of these applications can be supported
by today’s mobile broadband networks and their future evolution. However, some
other applications will impose additional and very diverse requirements on mobile
and wireless communication systems that the next generation (Fifth Generation 5G)
will have to support:

• Far more stringent latency and reliability requirements are expected to be

necessary to support applications related to healthcare, security, logistics,
automotive applications, and mission-critical control.

• A wide range of data rates has to be supported, up to multiple gigabits per

second, and tens of megabits per second need to be guaranteed with very high
availability and reliability.

• Network scalability and flexibility are required to support a large number

of devices with very low complexity and requirements for very long battery
lifetimes.

5G research has been quite active in the past years. Thus, several funded projects
and standardization bodies like Mobile and Wireless Communications Enablers
for Twenty-twenty Information Society -II (METIS - II) [1], 5th Generation Non-
Orthogonal Waveforms for Asynchronous Signalling (5GNOW) [4], 4G Americas
1.1 Motivation 7

Fig. 1.3 Enhancement of key capabilities from LTE-Advanced to IMT-2020. [111]

[15], Next Generation Mobile Networks (NGMN), 3GPP, International Telecom-

munication Union - Radio Communication Sector (ITU-R) [111] have attempted
to create pioneering scenarios for identifying the requirements of 5G. This process
resulted in a large number of scenarios and use cases focusing on diverse require-
ments. Given the large number of use cases, it would be unrealistic to consider all the
use cases proposed by the research community and the standardization bodies and
fora for evaluation of the 5G RAN design solutions. Thus, the most representative
scenarios are noted below.

1. Mobile broadband and media everywhere and anytime: Mobile broadband is

the key use case today and it is expected to continue to be one of the key use
cases driving the requirements for 5G. It goes far beyond basic mobile Internet
access and covers rich interactive work, media and entertainment applications
in the cloud or reality augmentations (both centralized and distributed).
Data will be one of the key drivers for 5G and in new parts of this system. We
may for the first time see no dedicated voice service - in 5G, voice is expected
to be handled as an application, simply using the data connectivity provided by
the communication system. The main drivers for the increased traffic volume
8 Introduction

are the increase in size of content and the number of applications requiring
high data rates. Factors include increase in camera resolution, the rise in
screen resolution with the recent introduction of 4K (8K is already expected
beyond 2020). Streaming services (audio & video), interactive video and
mobile Internet connectivity will continue to be used more broadly as more
devices connect to the Internet. Many of these applications require always-on
connectivity to push real time information and notifications to the users.
Cloud storage and applications are rapidly increasing for mobile communi-
cation platforms. This is applicable for both work and entertainment. Cloud
storage is one particular use case driving the growth of uplink data rates. In the
past, content was mostly downloaded. 5G will also be used for remote work in
the cloud which, when done with tactile interfaces, requires much lower end-
to-end latencies in order to maintain a good user experience. Entertainment,
for example cloud gaming and video streaming, is another key driver for the
increasing need for mobile broadband capacity.
With 5G, users will experience broadband access in crowded areas like con-
certs, sporting events and festivals, alleviating issues with capacity, interference
and reliability. 5G customers will also enjoy 4K movies downloaded in just
seconds without a WiFi connection. While live TV broadcasts and sporting
events will become immersive viewing experiences, as if you were at the
event in real-life. For customers, 5G will also maximize their experience in
both indoor and outdoor connectivity and offer high QoS broadband even in
challenging network conditions.

2. Smart vehicles, transport and infrastructure: The automotive sector is expected

to be a very important driver for 5G, with many use cases for mobile com-
munications for vehicles. For example, entertainment for passengers requires
simultaneous high capacity and high mobility mobile broadband, because
future users will expect to continue their good quality connection independent
of their location and speed. Entertainment will be very essential on smart
phones and tablets everywhere, including high mobility environments such
as trains, cars and air-planes. The key application areas could be for smart
infrastructures, connected bus-stops, connected trucks and connected cars.
Thus, these applications will be more focused on massive machine type com-
1.1 Motivation 9

munication. We can consider sensors embedded in roads, railways and airfields

to communicate with each other and/or with smart vehicles.
Other use cases for the automotive sector are augmented reality dashboards.
These display overlay information on top of what a driver is seeing through
the front window, identifying objects in the dark and telling the driver about
the distances and movements of the objects.
The previous two use cases are related to content provisioning for the car users,
but the cars themselves will also be connected. Many car manufacturers are
already adding driver assistance systems based on 3D imaging and built-in
sensors. In the future, wireless modules will enable communication between
vehicles themselves, information exchange between vehicles and supporting
infrastructure and between vehicles and other connected devices, for example,
those carried by pedestrians. Additionally, use cases for traffic safety are
now widely discussed. These include cars detecting safety critical situations,
such as black ice, accidents within reach of the car and other hazardous road
conditions. Safety systems will also guide drivers on alternative courses of
action to allow them to drive more safely and lower the risks of accidents.
The next phase will be remotely controlled or even self-driven vehicles, which
will require ultra reliable and very fast communication between different self-
driving cars and between cars and infrastructure. In a plausible future, a
self-driving car takes care of all driving activity, allowing the driver to rest and
concentrate only on traffic anomalies that the car itself cannot identify. Ideally,
reading the morning newspaper while commuting will become possible. The
technical requirements for self-driving cars call for ultra-low latencies and
ultra-high reliability, increasing traffic safety to levels humans cannot achieve.

3. Human interaction - Internet of Things (IOT): 5G means unlocking the poten-

tial for us to interact and connect with machines in unbelievable ways making
the networked society and the IOT era a reality by connecting more devices
than ever. 5G will become the backbone of IOT, connecting devices in ways
we never thought possible. Users will experience smart cars that are capable
of communicating with traffic lights, augmented reality and 360 degree im-
mersive gaming and movies, and transmitting touch and texture to realize the
tactile internet. The IOT applications that 5G will help enable is truly limitless.
10 Introduction

4. Critical control of remote devices: The high reliability and low latency of
5G unlocks the ability to control critical services and infrastructure. This
creates new opportunities for public safety, government, city management
and utility companies. 5G will be a key enabler of the future digital world.
It will bring about new service capabilities for industrial stakeholders thanks
to the unprecedented on-demand performance and real-time reactivity. For
example, energy and water utilities will be capable of connecting to millions
of networked devices, taking real-time, intelligent and autonomous decisions.
5G introduces the ability to remotely control devices and heavy machinery,
thus unlocking the potential for improved medical services and surgeries that
are otherwise unavailable to rural communities. 5G would also allow for heavy
machinery to be controlled from a distance, making hazardous situations safer
for humans.
Wireless and mobile communications are becoming increasingly important
for industrial application. Wires are expensive to install and maintain and the
possibility of replacing cables with reconfigurable wireless links is a tempting
opportunity for many industries. However, achieving this requires that the
wireless connection works with a similar delay, reliability and capacity as
cables and that its management is simplified. Low delays and very low error
probabilities are new requirements that need to be addressed with 5G.
The health sector has many applications that can benefit from mobile com-
munications. Communications systems enable tele-medicine, which provides
clinical health care at a distance. It helps eliminate distance barriers and can
improve access to medical services that would often not be consistently avail-
able in distant rural communities. It is also used to save lives in critical care
and emergency situations.

The vision of the 5G system, key design principles and use cases lead to require-
ments that the future mobile broadband system will need to meet. Considering all
these aspects, the work presented in this thesis addresses in particular the physical
layer problems for the use cases defined above. A more detailed description of these
problems and their contexts of application is provided next.
1.2 Structure and Contribution of This Thesis 11

1.2 Structure and Contribution of This Thesis

The thesis is organized as follows: all chapters from Chapter 2 to 5 are developed
around the need of physical layer algorithms dedicated to above described 5G use
cases. In particular:

• Chapter 2 considers the aspect of channel coding algorithms for the ultra-low
latency use cases of 5G. In order to achieve low latency for critical control of
remote devices use case of 5G system, channel coding and modulation plays
an important role in the 5G physical layer. In this work, we enlarge the grid of
possible candidates by considering, in particular, short length LBCs (especially
eBCH codes) with soft-decision decoding. Specifically, results include:

1. Statistical analysis of the ordered vector components which highlights

and makes evidence to Ordered Statistics Decoding (OSD) property.
2. Use of novel statistics to derive a further simplified error performance
bound.
3. Analytical study of simplified expressions for the Probability Density
Function (pdf) and Cumulative Distribution Function (cdf) of the novel
statistic.

• Chapter 3 investigates the aspect of frame synchronization which is very

important in achieving the ultra-low latency and ultra reliable communication
for smart vehicles, transport and infrastructure and human interaction - IOT
use cases of 5G where a M2M communication plays a very important role.
Major findings in this chapter include:

1. Analysis of false-alarm and miss-detection probabilities for both the

Massey detector (introduced by Massey in [122]) and the Extended
Massey detector.
2. Introduction of a new test to eliminate a lower bound limitation on the
Pf a for short pattern lengths.

• Chapter 4 investigate semi-blind detection methods, especially, Energy De-

tection (ED) and Roy’s Largest Root Test (RLRT). Original contributions are
performance analysis under practical imperfections of the aforementioned
detectors. Specifically, results include:
12 Introduction

1. Analysis of false-alarm probability of Hybrid ED and Hybrid RLRT.

2. Performance analysis (detection probability) of the Hybrid ED and Hy-
brid RLRT.
3. Performance analysis (detection probability) of ED and RLRT under
unknown primary user traffic.
4. Comparison of ED and RLRT performance, and quantification of the
impact of noise level knowledge.
5. Comparison of ED and RLRT performance, and quantification of the
impact of primary user traffic knowledge.

• Chapter 5 investigate a signature matrix for LDS multiplexing. The small

number of elements in a spreading sequences are non zero allowing each
user to spread its data over small number of chips (tones), thus simplifying
the decoding procedure using Message Passing Algorithm (MPA). Major
contributions include:

1. Analysis of upper bound on the normalized minimum euclidean distance

of output codewords of a LDS spreading system.
2. Design of generic spreading matrices for 150% overloading factor which
meets the upper bound on the normalized minimum euclidean distance.

Finally, Chapter 6 contains summarizing conclusions and remarks.

Chapter 2

Ordered Statistics Decoding of

Linear Block Codes1

2.1 Introduction

Current generation of wireless technologies is being evolved toward a fifth generation

(5G) for better serving end users and transforming our society [42, 63, 67]. A detailed
discussion of the new use cases for 5G wireless communication standards can be
found in Chapter 1. Demands for greater capacity and higher data rates are just
two of the many factors influencing the evolution of wireless access technologies.
The impact of other aspects like energy consumption, device cost, spectrum and
latency will be fundamental to the success of future networks. The next evolution
of wireless communications is to provide large variety of services with significantly
different requirements. For example, in the European Union project Mobile and
wireless communications Enablers for the METIS [1], three main service types
that are considered for 5G, namely extreme mobile broadband (xMBB), massive
Machine-Type Communications (mMTC) and ultra-reliable MTC (uMTC) as well as
the five specific use cases for the 5G RAN design which typically represent a mixture
of services. It further describes the key requirements on the 5G RAN architecture
that have been identified and derived from the diverse service and use case needs,
and explicitly elaborates on the requirements posed by the notion of Network Slicing
in 5G. Thus, 5G should not be seen as a single solution but rather as a platform on
1 Part of the work described in this chapter has been previously published in [48].
14 Ordered Statistics Decoding of Linear Block Codes

which multiple different services can be built on using the solutions selected from
the 5G toolbox.
M2M communication will make up a large part of the new types of services
and use cases that 5G systems will address. From a communication technology
perspective, M2M can be divided into two main categories: mMTC is about con-
nectivity for large numbers of low-cost and low-energy devices in the context of the
Internet of things; mission-critical MTC is envisioned to enable real-time control
and automation of dynamic processes in various fields, such as industrial process
automation and manufacturing, energy distribution, intelligent transport systems,
traffic safety, infrastructure protection, remote surgery and requires communication
with very high reliability and availability, as well as very low end-to-end latency
going down to millisecond level. In short, the requirements of mission-critical MTC
have to be fulfilled in three dimensions: latency, reliability and availability.

a. Latency refers to the time delay between data being generated – e.g., at a
sensor – and the same data being correctly received – e.g., by the actuator.
The most stringent requirement on the end-to-end latency may be 1 ms, as
explained for example in ITU - Telecommunication Standards Sector (ITU-T)
Technology Watch Report on the Tactile Internet [5].

b. Reliability refers to the capability of guaranteeing successful message transmis-

sions within a defined latency budget – or delay. The reliability requirements
vary among different mission-critical MTC services, but may go down to one
per billion messages as shown in an ETSI TR 101 557 V1.1.1 (2012-02)[2].
As an example, in industrial automation, only one message in one billion data
transfers may be lost or delayed within the given latency budget.

c. System availability has to ensure that critical applications are not in outage
when they are needed. To equate availability of wireless and wired solutions,
an availability of 99.999% can be sought.

Mission-critical MTC, however, is still in the early-development phase, and there

are a lot of challenging research problems to solve. Emerging use cases in context to
mission-critical MTC along with the anticipated killer apps of the Tactile Internet,
like immersive virtual reality, augmented reality and tele-surgery, to name a few,
present a need for ultra-low latency mobile networks [72]. Although the current
2.2 Channel Coding for 5G 15

minimum data plane latency in 4G LTE which is in the order of 20 ms is sufficient

for most of the mobile broadband applications, it may however, not be enough for
mission critical MTC applications. To ensure support for this kind of mission-critical
machine-type connectivity, next-generation wireless access should support latency
around 1 ms.
Latency can be interpreted in terms of efficiency of very low-layer procedures
allowing for time/frequency synchronization, identification/authentication (Start
of/End of frame (SOF/EOF) detection), channel setup time, channel interleaving,
channel coding etc. Thus, to enable an end-to-end latency of 1 ms or less, the
entire system needs to be designed for low latency. All in all, if we want to address
all the potential use cases of mission-critical MTC, the radio technology that we
design should also be scalable for these stringent requirements. To achieve the lower
delays and higher reliability discussed above, we can identify a number of system
modification, for instance, reduced transmission time intervals, e.g., down to 100µs,
and short symbol duration enabling fast and efficient data transmission. In addition,
redesigning of physical channels allowing early channel estimation, fast detection of
SOF/EOF and use of fast encoding-decoding channel codes also help in minimizing
the transmission/detection time. In this chapter, we consider the aspect of channel
coding algorithms for the ultra-low latency use cases of 5G while in Chapter 3, we
will consider the case of frame synchronization for the low latency aspect of 5G use
cases.

2.2 Channel Coding for 5G

In terms of error correction coding techniques, a quantum leap from 2G to 3G was

achieved of Turbo Codes [25]. However, error correction coding schemes in 3G and
4G are almost same. Another big jump from 4G to 5G is expected because of the
several new challenges in 5G.
In order to achieve low latency in 5G system, channel coding and modulation
plays an important role in the 5G physical layer. In general, the channel decoder
must share the latency budget with many other physical layer components such
as synchronization. Owing to this, the channel decoder should target the lowest
possible processing latency. Furthermore, since the channel decoder must overcome
the uncertainty introduced by noise interference and poor signal strength, it typically
16 Ordered Statistics Decoding of Linear Block Codes

Table 2.1 Turbo codes and duo-binary turbo codes in standards.

Types of Code Rates Termination

Standards
Turbo Codes
3G 1/2, 1/3 and
(UMTS and Turbo codes 1/4 Tail Bits
CDMA 2000)
4G Duo-binary
1/2 ∼ 7/8 Circular
(WiMAX) Turbo code
4G LTE Turbo Code 1/2 ∼ 7/8 Tail Bits
Duo-binary
DVB-RCS 1/3 ∼ 6/7 Circular
Turbo code
1/2, 1/3, 1/4 Tail bits
CCSDS Turbo codes
and 1/6

has a much greater complexity than channel encoder. This is the reason why the
channel decoder is typically the main concern when designing a channel code. More
specifically, the decoding complexity plays a dominant role in defining decoding
delays. Among the most attractive options emerged till now, some of the potential
options are discussed hereby.

A. Turbo Codes and Duo-binary Turbo Codes

Turbo codes [25] achieved very low error probability which is close to the
Shannon limit. The important design criteria of turbo codes are to find suitable
component codes which maximize the effective free distance and to optimize
the weight distribution of the codewords at a low EbN0 [53, 52]. One disadvan-
tage of the original turbo codes is that error floor occurs at quite high Bit Error
Rate (BER) 10−5 due to poor minimum Hamming distance. Possible solutions
are to (a) design a good interleaver, (b) increase the number of memories of the
consecutive encoder and (c) use the non-binary turbo codes such as duo binary
turbo codes[55] or 3D turbo codes. Duo-binary turbo codes are composed of
two Circular Recursive Systematic Constituent (CRSC) codes with two inputs.
Its information bits are encoded pair wise. CRSC encoders do not need any
tail bits.

B. LDPC and Non-binary LDPC

LDPC codes [80], combined with iterative Belief Propagation (BP) decoding,
have emerged after 1993 as the most promising method of achieving the goal
2.2 Channel Coding for 5G 17

set by Shannon [153] in his landmark 1948 paper: to communicate reliably

over a noisy transmission channel at a rate approaching channel capacity.
Indeed, many applications have recently adopted LDPC codes as industry
standards, such as Wireless Local Area Networks (WLANs) (IEEE 802.11n),
WiMAX (IEEE 802.16e), Digital Video Broadcasting (DVB) - S2, 10GBase-T
Ethernet (IEEE 802.3an), and the ITU-T standard for networking over power
lines, phone lines, and coaxial cable (G.hn/G.9960). The key feature that
sets LDPC codes apart from other capacity approaching codes is that with
suboptimal iterative BP decoding, complexity grows only linearly with code
block length, resulting in practically realizable decoder implementations for
powerful (long block length) codes. (The decoding complexity of optimum
Maximum Likelihood (ML) decoding, on the other hand, grows exponentially
with block length, making it impractical for large block lengths.) LDPC
block code (LDPC-BC) designs can be classified in two types: regular and
irregular. Regular codes, as originally proposed by Gallager [80] in 1962, are
asymptotically good in the sense that their minimum distance grows linearly
with block length. This guarantees, with ML decoding, that the codes do not
suffer from the error floor phenomenon, a flattening of the BER curve that
results in poor performance at high signal-to-noise ratios (SNRs), and similar
behavior is observed with iterative BP decoding as well.
Non-binary LDPC codes using iterative BP show us good performance at
medium code lengths (500 ≤ N ≤ 3000) and high order modulation commu-
nications (greater than 16QAM). However, the decoding complexity remains a
major obstacle for their commercial applications.

C. Spatially coupled LDPC (SC-LDPC) codes

SC-LDPC codes have robustly excellent performance over a broad range of
channel conditions, including both the waterfall and error floor regions of the
BER curve. SC-LDPC codes can be viewed as a type of LDPC convolutional
code (LDPC-CC), since spatial coupling is equivalent to introducing memory
into the encoding process. In channel coding parlance, the key feature of
SC-LDPC codes that distinguishes them from standard LDPC codes is their
ability to combine the best features of regular and irregular codes in a single
design. An added feature of the SC-LDPC code design is that the resulting
Tanner graph retains the essential implementation advantages associated with
18 Ordered Statistics Decoding of Linear Block Codes

the structure of regular codes compared to typical irregular designs. The

research establishing the performance characteristics of SC-LDPC codes relies
on ensemble average asymptotic methods, that is, the capacity approaching
thresholds and asymptotically good minimum distance behavior are shown to
hold for typical members of SC-LDPC code ensembles as the block length
tends to infinity [61].

D. Other Candidate Channel Codes

Other candidate channel codes are Sparse Regression Codes and Polar Codes.
Sparse Regression Codes [24] are new class of codes for Gaussian multi-
terminal source They are designed using the statistical framework of high-
dimensional linear regression. They are rate-optimal codes with low com-
plexity for compression and transmission. They are suitable for cooperative
communications. Polar codes [17] are first constructive coding to provably
achieve channel capacity using Kronecker product. Its characteristic can be
summarized as follows: (1) the block length of a polar code is normally a
power of two but the code length can be adjusted by usual code shortening
methods, (2) the rate of a polar code can be adjusted to any number between
0 and 1 in increments of 1/N, (3) polar codes are available in systematic
form, (4) polar codes have a recursive structure which makes them suitable for
low-complexity implementations and (5) polar codes have no error floor.

While turbo and LDPC codes have brought capacity-approaching performance

within reach of implementable systems, implementable does not necessarily mean
practical. The complexity of codes that perform well under practical constraints such
as limited decoding delay and high spectral efficiency is still a major hurdle for low
power implementations in integrated circuits. Furthermore, when codes are short
as required by small latency, minimum distance is more important. Turbo codes
and LDPC codes are known to have small minimum distance for short length codes.
Instead, algebraic designed block codes (e.g., BCH code [28] and Reed-Solomon
(RS) codes [145]) are known to have higher minimum distance for short length codes
when compared to LDPC and Turbo codes. In this work, we enlarge the grid of
possible candidates by considering, in particular, properly designed short length
LBCs (especially eBCH codes) with soft-decision decoding.
2.3 Ordered Statistics Decoding 19

2.3 Ordered Statistics Decoding

Low complexity decoding of LBCs has long been investigated by many coding
theorists: a detailed bibliography of the contributions in this area can be found
in [155]. Some examples of decoding algorithms are 1) Exhaustive Decoding:
Possible only for very short length codes, 2) Algebraic Decoding: Exploiting the
property of algebraic designed block codes, 3) Viterbi decoding [41, 173]: Although
all the LBCs possess the trellis structure which is the backbone of Viterbi decoding,
the number of states (min{2k , 2n−k }) becomes too large to practically implement for
long length codes and 4) Iterative decoding: Iterative decoding of powerful classical
codes such as BCH, and RS codes are quite sub-optimal with respect to ideal
Maximum-Likelihood Decoding (MLD), due to their structures. As a consequence,
sub-optimal (near optimal) soft decision decoding based on the ordered statistic of
the received noisy codeword has been proposed [54, 99, 32, 76, 81] proving to be
efficient with considerable complexity for short length LBCs.
Since Fossorier and Lin, in their original contribution [76] presented a novel OSD
scheme for soft decision decoding of LBCs based on ordered statistics of the received
noisy samples (although an algorithm belonging to this set was first proposed by
Dorsch [54] and also used by [99] and [32]), OSD is widely being studied in the
literature [76, 81, 79, 77, 14]. Over the years, various new methods [175, 75, 165,
106, 176, 97, 101] based on reliability information and many modifications [77, 14]
on original OSD have been proposed in the literature to minimize the performance-
complexity trade-offs. In addition to wide spread application of ordered Statistic
in decoding of LBCs, the contributions in [74, 172, 168, 185] show its use also in
decoding of LDPC and convolution codes. In all these optimum and sub-optimum
decoding algorithms, a reliability measure of the received symbols has been used to
reduce the search space and find the most likely codewords.
The original concept of OSD [76] is basically implemented in two stages, a)
determining the Most Reliable Independent (MRI) bits from the Most Reliable Basis
(MRB) of the code and b) Order−I reprocessing on MRI using most likely Test
Error Patterns (TEPs). Out of these two stages, order−I reprocessing is designed to
improve the hard decision decoded codeword progressively until either practically
optimum or a desired error performance is achieved. The approach of ML resource
test based on the cost function calculated from the soft valued samples of the
permuted received sequence is introduced as a stopping criterion after each stage
20 Ordered Statistics Decoding of Linear Block Codes

j, 0 ≤ j ≤ I of order−I reprocessing. This stopping criterion indeed proved excellent

in reducing the average number of computations. An upper bound on the error
performance for order−I reprocessing OSD has been derived based on the noise
statistics after reordering in [76, 78].
However, there are two major drawbacks of the performance bound derived
in [76]. The first drawback is its complexity of evaluation (requiring (I + 1) di-
mensional integral for any order−I reprocessing). The second is the tightness of
the bound, since this bound has been derived based on the assumption that the
events associated with the reordered vector components are statistically indepen-
dent, although, this does not hold true in practical scenarios. These issues related
to computationally complex error performance bound has been revisited in [79],
where a computationally simpler and comparatively tighter upper-bound on the error
performance has been derived based on the statistical approach proposed by Agrawal
& Vardy in [9]. It has been shown that, compared to (I + 1) dimensional integral
computation in [76], the expression derived in [79] requires only a 2−dimensional
integral for any order−I reprocessing. Although, the reduction of integral dimen-
sion from (I + 1) to 2 seems quite impressive, computing 2−dimensional integral
(n − k − 1) times as referred to [79] is still computationally complex.
In this chapter, we propose a novel statistic of the ordered vector components
which highlights and makes evidence to OSD property. Furthermore, the proposed
statics can be applied to derive a further simplified error performance bound. More
importantly, simplified expressions for the pdf and cdf of the proposed statistic
are derived. Subsequently, we incorporate the properties of this statistic to derive
the simplified error performance bound for OSD with order−I reprocessing. The
computational complexity of the corresponding bound is found to be even simpler
(requiring single dimensional integral evaluation) to that of the bounds derived in [76]
and [79]. Furthermore, the error performance bound derived in this work is as tight
as the one proposed in [79] and is also derived without any assumption.
2.4 Overview of OSD and Conventional Reprocessing 21

2.4 Overview of OSD and Conventional Reprocess-

ing

Given an LBC C(n, k), with a systematic generator matrix G, at the transmitter side, a
k−bit information vector, v = (v1 , v2 , . . . , vk ), vi ∈ GF(2) = {0, 1} is mapped into a
codeword c = v · G = (c1 , c2 , . . . , cn ) where GF(2) stands for Galois field of order-2.
Under Binary Phase Shift Keying (BPSK), the codeword is mapped into a real-
valued vector as

s = (s1 , s2 , . . . , sn ) si ∈ {−1, +1} ⊆ R, (2.1)

where ci = 0 → si = −1 and ci = 1 → si = +1.

The vector s is transmitted over an Additive White Gaussian Noise (AWGN)
channel. At the receiver side, we observe the received vector,

r = (r1 , r2 , . . . , rn ) ri ∈ R, (2.2)

where ri = si + wi , wi is a white Gaussian noise sample with mean zero and variance
σ 2.
Given r, we want to perform a soft-decision decoding. As already stated in
section 2.3, for small-medium block codes (with k upto some hundreds of bits),
an effective solution is provided by OSD algorithms, like the Most Reliable Basis
(MRB) algorithm [76]. This algorithm starts by reordering the received vector
in the descending order of the absolute values. In this way, the first symbols are
characterized with a high reliability, i.e., a large probability of being correct.
Given r, by reordering its components in decreasing magnitude λi = |ri |, we
obtain a vector,
r∗ = (r1∗ , r2∗ , . . . , rn∗ ) ri∗ ∈ R, (2.3)

such that |ri∗ | > |ri+1

∗ | for 1 ≤ i ≤ n. Let us define ρ as the permutation rule applied
1
∗
on r to obtain r .
22 Ordered Statistics Decoding of Linear Block Codes

The generator matrix G is also permuted using the same permutation rule ρ1 , to
give a new permuted generating matrix

G∗ = ρ1 (G). (2.4)

The matrix G∗ in (2.4) is then processed using elementary row operations to

obtain a systematic form G′ . As well known in the OSD literature [76, 79], it may
happen that the first k columns of G∗ may not be linearly independent, i.e., the k
most reliable components do not correspond to an information set. In this case, it is
necessary to slightly change the permutation ρ1 until an information set is obtained.
This introduces a second permutation defined as ρ2 which needs to be applied both
on r∗ and G∗ to obtain r′ and G′ respectively. Thus, the final permutation relations
can be written as,

r′ = ρ2 (r∗ ) ⇔ ρ2 (ρ1 (r)), G′ = ρ2 (G∗ ) ⇔ ρ2 (ρ1 (G)). (2.5)

In the following, we consider ρ2 as an identity permutation function such that

ρ2 (x) = x where x is a arbitrary vector. Thus, we suppose r′ has exact reliability
ordering. We perform a symbol-by-symbol hard decision on r′ to obtain the binary
vector,
y′ = (y′1 , . . . , y′i , . . . , y′n ) y′i ∈ GF(2), (2.6)

where ri′ < 0 → y′i = 0 and ri′ ≥ 0 → y′i = 1.

Next, we take the first k bits of y′ to form the candidate information vector,

v′ = (v′1 , . . . , v′i , . . . , v′k ) v′i ∈ GF(2). (2.7)

Due to reordering, with high probability, v′ contains a few errors because its bits
have high reliability. One of the objectives of this paper is to provide computationally
efficient expression for calculating the number of errors contained in v′ .
Exploiting the reliability property, OSD algorithm considers a set of patterns,

S = {p = (p1 , . . . , pi , . . . , pk ) pi ∈ GF(2)}, (2.8)

with Hamming weight of the TEP wH (p), 0 ≤ wH (p) ≤ I, where I is called the order
of the algorithm. Each pattern is added to the candidate information vector given
2.5 Proposed New Statistic of Ordered Vector Components 23

by (2.7), which is then encoded by the matrix G′ to obtain a reprocessing codeword

c∗ in the following way,

∀p ∈ S : v′ → v∗ = v′ + p → c∗ = v∗ · G′ . (2.9)

When all patterns are considered, the codeword c∗ in (2.9) at minimum Euclidean
distance from the permuted received vector r′ can be chosen as the received codeword.
Obviously, if we set I = K and we test all the corresponding 2k patterns, we can
surely find the ML codeword, but the algorithm becomes equivalent to an exhaustive
decoding, which is impossible for non-trivial codes [76]. Then, a key issue for these
algorithms is the choice of the order I and the set S to optimize the complexity-
performance trade-off. It has been shown that, for small/medium size codes (e.g.,
n ≤ 150, code rate k/n > 0.5), an order I ≈ dmin /4 is able to provide nearly-optimal
decoding performance, i.e, very close to that of ideal MLD [76].
An approximation of the closed form expression for the upper-bound on the
error performance of the order−I reprocessing has been first derived in [76] which
requires the computation of an (I + 1)-dimensional integral for any reprocessing
order-I. Later, a relatively simple (requiring 2−dimensional integral evaluation) and
relatively accurate (without any assumption) upper bound on the error performance of
OSD for each stage reprocessing is derived in [79] based on the statistics introduced
in [9]. In the following sections, we present a new statistics on the error properties of
the permuted binary received vector y′ which can be applied to evaluate a simplified
expression of the OSD error performance bound.

2.5 Proposed New Statistic of Ordered Vector Com-

ponents

Given an LBC C(n, k), without loss of generality, let us consider an all-zero transmit-
ted codeword, c = (0, . . . , 0, . . . , 0), which after BPSK mapping, corresponds to the
transmitted vector, s = (−1, . . . , −1, . . . , −1). At the output of the AWGN channel,
we observe the received vector r = (r1 , . . . , ri , . . . , rn ), with:

ri = −1 + wi , (2.10)
24 Ordered Statistics Decoding of Linear Block Codes

where wi is a Gaussian random variable with zero mean and variance σ 2 . All wi
noise components are statistically independent.
Each component ri in (2.10) has a pdf given by,

1 (x+1)2
−
fr (x) = √ e 2σ 2 . (2.11)
2πσ 2

If we consider the magnitude of the components of r written as λi = |ri |, the pdf

of λi is given by,

0
 if x < 0,
fλi (x) = −
(x+1)2
−
(x−1)2 (2.12)
 e√ 2σ 2 + e√ 2σ 2

if x ≥ 0,
2
2πσ 22πσ

while its cdf is given by,


0 if x < 0,
Fλi (x) = (2.13)
x+1 x−1
1 − Q


σ −Q σ if x ≥ 0,
√
1/ 2πexp(−y2 /2)dy is the standard normal tail function.
R∞
where Q(x) ≜ x

Now, let us focus on the reordered vector r′ . A first study to estimate the behavior
of the components of r′ was done in [76], where the pdf of the reordered noise
samples conditioned upon the transmitted symbol has been presented.
Given the vector r observed by transmitting the all-zero codeword over an AWGN
channel with BPSK modulation and the ordered vector r′ obtained by ordering r
in decreasing magnitude, we can obtain the pdf of the components of r′ , and the
pdf of the magnitude of the components of r′ by using arguments from ordered
statistics [78] as follows. The pdf of the i-th component ri′ of r′ is given by,

n!
∀x : fri′ (x) = (1 − Fλ (|x|))i−1
(i − 1)!(n − i)!
· (Fλ (|x|))n−i fr (x). (2.14)
2.5 Proposed New Statistic of Ordered Vector Components 25

In the similar manner, using the pdf of λi , the pdf of the magnitude of i-th
component λi′ = |ri′ | of r′ is given by,

n!
∀x : fλi′ (x) = (1 − Fλ (x))i−1
(i − 1)!(n − i)!
· (Fλ (x))n−i fλ (x). (2.15)

In the following, we define a new random variable EL which represents the

number of errors contained in the first L positions of the permuted received vector
r′ . Furthermore, we present a computationally efficient expression for the pdf of
EL . Normally, under identity permutation rule (which basically doesn’t change
the position of the vector elements) ρ2 , order−I OSD includes only the chances
of having I or less than I errors in MRPs of the permuted received vector. This
leaves behind the probability of missing the true codeword by an order−I OSD
under ML performance. In this relation, the complementary cdf of EL at EL = I
actually provides a simplified expression to evaluate the maximum probability of
missing a true codeword by an order−I OSD, thus justifying the importance of EL in
evaluating the performance bound of the order−I OSD. Further details are discussed
in Section 2.6. We present a new Theorem which provides an exact expression for
the pmf of the random variable EL .
Theorem 1. Given a random variable EL which represents the number of errors we
expect on the first L, 1 ≤ L ≤ n, positions of the permuted received vector r′ , its pmf
is given by,

Z +∞  
l
pEL (EL = j) = p(x) j (1 − p(x))l− j fλl+1
′ (x)dx, (2.16)
0 j

where,
x+1


Q σ
p(x) = x+1

−x+1

, (2.17)
1+Q σ −Q σ

′ (x) given by (2.15) and its cdf is given by,

with fλl+1

j
FEL (EL = j) = ∑ pEL (EL = i). (2.18)
i=0
26 Ordered Statistics Decoding of Linear Block Codes

pdf of ri
−4 −3 −2 −1−x 0 +x 1 2 3 4

ri

Fig. 2.1 Probability density function of ri for Eb /N0 = 3.5 dB.

Proof. Fix a value x, and suppose the magnitude of the (l + 1)-th component of the
reordered vector r′ is λl+1′ = x. Then, the vector r contains exactly l components
with |ri | ≥ x. As can be observed in Fig. 2.1 (the shaded region corresponds to
|ri | ≥ x), for each of these components, the probability of having an error is,

p(x) = P(ri > 0||ri |≥x )

x+1


Q σ
= x+1

−x+1

. (2.19)
1+Q σ −Q σ

Since we are working with the components of r we can use their implicit sta-
tistically independent property, which instead does not hold if we try to work with
the components of r′ . As a result, the probability of having j errors among these k
components is given by,
 
′ l
P(EL = j|{λl+1 = x, L = l}) = p(x) j (1 − p(x))l− j . (2.20)
j
′
The above result is obtained under the condition λl+1 = x. By integrating over
′ (x), we obtain the final pmf of EL as in (2.16) while its cdf
all x values by using fλl+1
at some value EL = j is obtained as in (2.18) simply by summing the normalized
pmf of EL for all EL : 0 ≤ EL ≤ j.
2.6 OSD Error Performance 27

2.6 OSD Error Performance

In this section, a different look at the error performance of order−I OSD is presented
based on the distribution of E detailed in Section 2.5. Let PeOSD−I denote the code
error performance of the order−I OSD and Pe (I) denote the probability that the
correct codeword is not among the candidate codewords supported by the order−I
OSD. The upper bound on the order−I OSD performance can be written as an
inequality as

PeOSD−I ≤ PeML + Pe (I)

!
More than I errors occur
≤ PeML + P , (2.21)
in 1st k positions of r′

where PeML is the MLD code error rate.

In (2.21), the probability of having more than I errors in the first k ordered
received symbols in r′ given an identity permutation function ρ2 can be simply
evaluated from the cdf of EL as,

Pe (I)|ρ2 (x)=x = 1 − FEL (I), (2.22)

where L in this case is equal to k.

In a real scenario, the permutation function ρ2 may or may not be identity. In fact,
the second permutation ρ2 directly relates with the number of column permutation
required to obtain first k columns of G∗ to be linearly independent. Let us consider
d is the number of dependent columns before kth independent one and P(d) be the
probability associated with d. It has been shown in [76] that the maximum number
of dependent columns that can be found before kth independent one for a given
generator matrix is given by,

dmax = n − k − dHmin − 1, (2.23)

where dHmin is the minimum Hamming distance of the considered code.

28 Ordered Statistics Decoding of Linear Block Codes

Thus, Pe (I) can be expressed under all cases of ρ2 as,

!
dmax
More than I errors occur
Pe (I) = ∑P · P(d)
d=0 in 1st k + d positions of r′
dmax

= ∑ P(d) 1 − FEk+d (I) . (2.24)
d=0

where FEk+d is the cdf of EL at L = k + d. The probability P(d) can be evaluated

either from simulation or from the approximated distribution proposed in [76].
As stated earlier, the random variable E denotes the number of errors, thus,
can only take integer values. Also the reprocessing order−I is always fairly small
d min
(I = H4 ≤ 5) for LBCs with practical dHmin . Thus, (2.24) can be written in terms of
pdf of EL as,
" #
dmax I
Pe (I) = ∑ P(d) 1 − ∑ fEk+d (Ek+d = i) . (2.25)
d=0 i=0

where fEk+d (Ek+d = i) can be evaluated by a single integral given by (2.16).

The above discussion shows that, for all admissible values of order-I, the theoret-
ical upper bound can be evaluated. It is quite clear that the computational complexity
of this new upper bound requires a single dimensional integral for any order-I com-
pared to (I + 1) and 2 dimensional integral for the one presented in [76] and [79]
respectively. Furthermore, this bound is as tight as the one proposed in [79] (see
Fig. 2.3 to 2.5) and is also derived without any assumption.

2.7 Numerical Results

For the purpose of comparison and performance evaluation, we adopt the same
scenario of BPSK transmissions over an AWGN channel as described in Section 2.4
and Section 2.5. In order to accommodate the code rates of different LBCs, we
adopt the ratio of the energy per bit to noise power spectral density ratio Eb /N0 with
Eb /N0 = NE0s · nk , where Es is the signal energy and N0 is the noise power spectral
density.
2.7 Numerical Results 29

100
Theoretical curves
Simulated (n = 64, k = 42)
−1
10 Simulated (n = 128, k = 99)
Simulated (n = 128, k = 64)

10−2

pdf of EL
10−3

10−4

10−5

10−6
0 2 4 6 8 10
E

Fig. 2.2 Probability density function of E: Eb /N0 = 3.5 dB, L = k.

2.7.1 Pmf of EL

Figs. 2.2 plots the pmf of random variable EL for different combinations of n and
k. For each (n, k) combinations, the simulation results are plotted and compared
with the theoretical result obtained from (2.16). We observe a perfect matching for
all permissible values of EL which justifies the validity of the expression presented
in (2.16).

2.7.2 Error Performance

It is worthwhile to mention that the evaluation of (2.21) involves the prior evaluation
of the MLD code error rate for the code under consideration. Thus, for our simulation
purpose, we use the simplest upper bound, i.e., the union bound [112],

n
√ !
d
PeML (c) ≤ ∑ Ad Q . (2.26)
min
d=dH
σ

In (2.26), d : d ≥ dHmin represents the Hamming weight of the codeword and Ad

is the multiplicity (number of codeword with hamming weight equal to d) of the
code. For simplicity, we consider at most the first four components of (2.26) which
provide a good approximation for medium dimension codes at considerable Eb /N0 .
For higher dimensional codes, the total union bound or other tight upper bound
alternatives represent more accurate performance measure [112].
30 Ordered Statistics Decoding of Linear Block Codes

100
Optimum ML
Eq. (25)

10−1 I = 0
I = 1
I = 2

Code Error Rate

10−2

10−3

10−4

10−5

10−6
1 2 3 4 5 6 7
Eb /N0

Fig. 2.3 Code Error Rate of OSD with order−I reprocessing for (64,42,8) Reed Muller code

100
MLD Union Bound
Eq. (25)
I = 0
−1
10 I = 1
I = 2
Code Error Rate

10−2

10−3

10−4

10−5
2 3 4 5 6 7
Eb /N0

Fig. 2.4 Code Error Rate of OSD with order−I reprocessing for (128,99,10) extended BCH
code

Figs. 2.3 to 2.5 depict the error performances of the (64, 42, 8) Reed Muller code,
the (128, 99, 10) eBCH code and the (128, 64, 22) eBCH code, respectively. Each
plot includes the simulation results and the corresponding upper bounds computed
from (2.21). We observe that for all values of reprocessing order-I, the theoretical
upper bounds are tight. The bounds are as tight as those derived in [76] and [79].
However, while the error performance bound of order−I reprocessing OSD based
on [76] requires the computation of (I + 1)−dimensional integral and [79] requires
the computation of a two-dimensional integral, the new upper bounds require the
computation of a single dimensional integral for any reprocessing order-I.
2.7 Numerical Results 31

100
MLD Union Bound
Eq. (25)
I = 0
−1
10 I = 1
I = 2
Code Error Rate

I = 3

−2
10

10−3

10−4

10−5
1 2 3 4 5 6 7
Eb /N0

Fig. 2.5 Code Error Rate of OSD with order−I reprocessing for (128,64,22) extended BCH
code
Chapter 3

Frame Synchronization1

3.1 Introduction

In Chapter 2, we considered the aspect of channel coding in achieving ultra-low

latency for 5G use cases. In this chapter, we consider the aspect of frame synchro-
nization which is very important in achieving the ultra-low latency and ultra reliable
communication for M2M applications.

3.1.1 Machine to Machine Communication

M2M communication, a promising technology for smart city concept, enables ubiqui-
tous connectivity between one or more autonomous devices without or with minimal
human interaction. The most important requirements for M2M communication in
5G systems currently being studied are 1) Maximum allowable end-to-end latency,
including jitters/re-transmits less than 1ms. 2) Reliability, for example packet loss
rate 10-9 [65, 69, 116]. Similarly, mission critical communications, which can help
first responders work safer, smarter and faster in disasters and day-to-day incidents,
have similar requirements as well. Among these two requirements for M2M, virtually
zero latency, i.e., radio latency lower than 1ms, is very critical for

• achieving higher data rates while keeping equipment cost low.

1 Part of the work described in this chapter has been previously published in [22].
3.1 Introduction 33

• ensuring fast procedure response time in the system (e.g. fast wake up and
dormancy, fast scheduling, fast link configuration)

• whole new range of use cases like remote control of machines and objects in
the cloud or tactile internet etc.

In addition, for real time M2M applications in the area of intelligent transport
system, remote monitoring and health will also be requiring very low latency and are
become the main focus of many mobile and it operators and vendors as a new revenue
opportunity. M2M traffic is bursty, since M2M applicatioin have not been developed
with network load shaping in mind. For instance, M2M devices have a much larger
ratio of uplink to downlink traffic volume, their traffic typically exhibits different
diurnal patterns, they are more likely to generate synchronized traffic resulting in
bursty aggregate traffic volumes, and are less mobile compared to smartphones.
Thus, burstiness is the result of synchronization effect resulting from several M2M
devices simultaneously reporting data [108]. In such application scenario where the
data transmission is bursty, a known synchronization/preamble sequence are often
periodically or aperiodically embedded in the data in order to aid in

1. data packet location during detection or decoding process,

2. resource scheduling,

3. packet synchronization,

4. determining the boundaries of the transmitted frame.

Thus, frame synchronization is a process by which incoming frame alignment

signals are identified, permitting the data bits within the frame to be extracted for
decoding and re-transmissions. In addition to the context of M2M application, binary
pattern synchronization problem is quite popular in many other scenarios. Some
important standards which make use of binary pattern synchronization are noted
here.

1. Moving Picture Experts Group (MPEG)-4 data compression standard [100]: It

utilizes synchronization patters such as the 24-bit start-code prefix and 17-bit
re-synchronization flag.
34 Frame Synchronization

2. CCSDS (Consultative Committee for Space Data Systems) Telecommand

uplink Recommendation [73]: It is used by international space agencies like
NASA and ESA for transmission from Earth to satellites/spacecrafts. The
transmission is bursty and the frames have variable lengths: every frame is
composed by n codewords, with n ranging between 2 and 147. A tail sequence
synchronization word of N = 64 bits is used to identify the frame end.

3. DVB Terrestrial (DVB-T) standard [95]: In this application we look for the
synchronization word inside a sliding window. A N = 8-bit synchronization
pattern is used by DVB-T standard for digital television broadcasting.

4. IEEE 802.15.4 standard [8]: Physical layer frame starts with a physical syn-
chronization header (5 bytes) followed by a physical payload upto 127 bytes
such that the physical synchronization header is used to alert the receiver that
a potentially receivable signal is present.

Another scenario could be in contention free random access procedure in LTE

[184], the base station assigns distinct preamble to each UE (user equipment) and
hence the concern for collision and other collision related issues are non-existent.
Contention-free Random Access are used in areas where low latency is required,
such as handover and resumption of down-link traffic for UE.

3.2 Introduction to Frame Synchronization Problem

Binary pattern recognition in a sequence of noisy samples is an important problem

in communication and signal processing. As a classical example, Synchronization
Word detection is commonly used to determine the boundaries of a transmitted
frame for both continuous and burst communication links [23, 157, 82]. Frame
synchronization literature covers a number of topics, including optimum and sub-
optimum detectors [122, 132, 151, 110, 37], important aspects for communication
systems [133, 56, 102, 103, 34–36], and binary pattern design [23, 163].
In this scenario, we focus on the basic detection problem which consists of a
known binary pattern, antipodal signaling and Gaussian noise, where soft values
are provided to the detector. Under this hypothesis, the optimal solution is provided
by the Likelihood Ratio Test (LRT). The LRT for this problem was first derived
3.3 System Model and Problem discussion 35

by Massey in [122]. A different formulation was provided by Chiani and Martini

in [37] (In succeeding section, we present the proof). Both LRT formulations are
characterized by a quite high complexity, which may be critical for some practical
applications requiring real-time implementation, like frame synchronization at high
bit rates and low latency scenario. To overcome this problem, two simplified ap-
proaches are often used: hard and soft correlation [82]. Despite their widespread use,
they are highly sub-optimal with respect to LRT.
The Massey correlation detector was introduced by Massey in [122] as an asymp-
totic approximation of LRT which holds for large SNR values. It consists of a
modified soft correlation computed only on the components where the sign of
observed and pattern symbols differs, therefore it is characterized by a very low
complexity. Typically, analytical results on the detectors are not easy to be obtained
[132]. However, they are very useful because they highlight the detector properties,
and allow to compute the performance without resorting to extremely long simula-
tions. As an example, the LRT analytical performances were presented for the first
time by Chiani and Martini in [37] (in the same paper, the analytical performances
of hard and soft correlation were reported, too).

3.3 System Model and Problem discussion

We have a fixed, known, binary pattern or Synchronization Word (SW) c, composed

by N binary antipodal symbols:

c = (c1 , · · · , ci , · · · , cN ) with fixed ci ∈ {−1, +1}.

We observe a window which may contain c or random symbols. The observed

samples are corrupted by AWGN. We formulate the problem as the binary hypothesis
testing of Figure 3.1.
Given the vector r composed by N real-value observed samples, there are two
alternatives. Under H1 the SW is present and we observe:

H1 : r = (r1 , · · · , ri , · · · , rN )
ri = ci + ni with fixed ci ∈ {−1, +1}. (3.1)
36 Frame Synchronization

Under H0 the SW is absent and we observe random symbols:

H0 : r = (r1 , · · · , ri , · · · , rN )
ri = di + ni with i.i.d. di ∈ {−1, +1}. (3.2)

In both cases, ni are samples of a Gaussian random variable with zero mean and
variance σ 2 .
The binary pattern recognition problem is quite popular in many scenarios.
For communication systems a typical example is frame synchronization: SWs are
commonly used to determine the boundaries of a transmitted frame. The observed
window where we look for the SW can be fixed or moving [23]-[82]. For example,

Fig. 3.1 The binary pattern recognition problem under study.

Fig. 3.2 The frame synchronization problem.

3.4 Likelihood Ratio Test 37

it is fixed for the CCSDS Telecommand uplink Recommendation [73], used by

international space agencies like NASA and European Space Agency (ESA) for
transmission from Earth to satellites/spacecrafts. The transmission is bursty and
the frames have variable lengths: every frame is composed by n codewords, with n
ranging between 2 and 147. A tail sequence SW of N = 64 bits is used to identify
the frame end. As shown in Figure 3.2.a, at the end of any codeword the receiver
looks for the SW inside a fixed window of 64 symbols. If it is detected, the frame is
declared closed, otherwise a new codeword is processed. The problem is very well
matched to the considered one. The hypothesis that, under H0 , the binary symbols
look as i.i.d. random symbols is well justified by the properties of the transmitted
codewords, which typically look as random sequences (see, for example, [31]).
In other applications, we look for the SW inside a sliding window, as in Figure
3.2 (b). Typically, this is the case of continuous systems, where equidistant patterns
are introduced to allow block synchronization. The receiver slides a window of N
symbols along the received sequence, looking for the SW which starts a frame. If
the window is aligned with the SW, we have the H1 case. In all the other cases, we
have the H0 case. The hypothesis that the binary symbols look random under H0 is
certainly realistic when the window is completely outside the SW (red upper H0 in
Figure 3.2.b). When the window is partially overlapped with the SW (blue lower H0
in Figure 3.2.b), the hypothesis can still be considered true for well designed SWs
with good properties in terms of aperiodic autocorrelation function [23],[163], which
look as random symbols when out of sync [37].
In this work we will often refer to frame synchronization as a typical scenario.
Obviously, binary pattern recognition problem is much more general and all the
presented results hold for a number of different applications, too.

3.4 Likelihood Ratio Test

For the problem presented in Section 3.3, two alternative, equivalent formulations
of the optimal LRT solution were derived, respectively by Massey in [122] and by
Chiani/Martini in [37].
38 Frame Synchronization

The Massey formulation [122] is:

N N r 
i
ΓLRT1 = ∑ ri ci − σ 2 ∑ log cosh . (3.3)
i=1 i=1 σ2

The Chiani/Martini formulation [37] is:

N   
2ri ci
ΓLRT2 = ∑ log 1 + exp − 2 . (3.4)
i=1 σ

3.4.1 Equivalency of Massey Formulation and Chiani/Martini

formulation of LRT

Starting from the final expression of Massey formulation in (3.3), raising both sides
of the equation by an exponent, we get,
rc
∏Ni=1 exp( σ 2 )
i i
exp(ΓLRT1 /σ 2 ) =   (3.5)
ri
∏N
i=1 cosh σ 2

Now, we can take a reciprocal of the components on both sides of the (3.5), and
using the expansion of hyperbolic cosine function as,
h    i
ri ri
∏N
i=1 exp σ2
+ exp − σ2
exp(−ΓLRT1 /σ 2 ) = rc (3.6)
∏N
i=1 exp( σ 2 )
i i

In order to simplify, we multiply ri in the exponent at the numerator by ci which

does not change the overall value of the numerator since ci is +1 or -1. Thus, we
obtain,

 N 
2 2ri ci
exp(−ΓLRT1 /σ ) = ∏ 1 + exp − 2 (3.7)
i=1 σ

Now, taking logarithm on both sides, we obtain,

 N  
2 2ri ci
− ΓLRT1 /σ = ∑ log 1 + exp − 2 (3.8)
i=1 σ
3.5 Hard and Soft Correlation 39

Finally, (3.8) proved that the LLR expression in (3.4) is a scaled version of the
LLR expression presented by Massey in (3.3).

3.5 Hard and Soft Correlation

By looking at the formulas we observe that both formulations are characterized by

non-negligible complexity which makes their application to real-time applications
problematic, especially in the case of low latency and high bit-rates system scenario.
For this reason, simplified formulations are usually adopted in practice: hard and
soft correlation [157],[82],[37].
Given the real-value received vector r = (r1 , · · · , ri , · · · , rN ), if a hard decision is
taken on each sample by looking at the sign only, it is transformed into the binary
vector:
y = (y1 , · · · , yi , · · · , yN ) ,

where:

if ri ≥ 0 → yi = +1
if ri < 0 → yi = −1.

Given the binary vector y, it is easy to show that the optimal test for detecting c is
the hard correlation detector:
N
ΓH = ∑ yi ci , (3.9)
i=1

which is equivalent to compute the Hamming distance between the two binary
vectors, i.e., the number of different symbols.
Given the hard correlation expression (3.9), we can extend it to the real-value
domain. The expression of the soft correlation detector is given by:

N
ΓS = ∑ ri ci . (3.10)
i=1

As discussed in [37], soft correlation has no theoretical justification. In fact, there

are many set of parameters where hard correlation is better than soft. Both hard and
40 Frame Synchronization

soft correlation tests are often used in practice. Unfortunately, as shown in the next
sections, they are highly sub-optimal with respect to LRT.

3.6 The Massey Correlation Detector

J.L. Massey derived the expression of the Massey correlation detector in [122]
starting from the optimal LRT given by (3.3) and using this approximation, which
holds at high SNR:
   
 r  exp + σri2 + exp − σri2 1 |ri |
 
i
cosh = ≃ exp . (3.11)
σ2 2 2 σ2

By applying it, (3.3) simplifies to:

N N
ΓLRT1 ≃ ∑ ri ci − ∑ |ri | . (3.12)
i=1 i=1

By noting that:

if sgn(ri ) = ci →|ri | = ri ci
if sgn(ri ) ̸= ci →|ri | = −ri ci ,

Massey obtained the following simplified expression:

Definition 1. The Massey Correlation Detector [122]. Given the binary antipodal
pattern c = (c1 , · · · , ci , · · · , cN ) and the observed real vector r = (r1 , · · · , ri , · · · , rN ),
the Massey correlation detector computes the statistical test:

ΓM = ∑ ri ci , (3.13)
i:
sgn(ri )̸=sgn(ci )

which is compared against a threshold t:

• If ΓM ≥ t, the detector declares the event H1 : the SW is present.

• If ΓM < t, the detector declares the event H0 : the SW is absent.

3.7 Analytical computation of the detector performances under H1 41

Massey correlation is similar to soft correlation (3.10), which is now computed

only on the received symbols with a sign different from the pattern symbols, and
is much simpler than the LRT. In the next two sections, we present one of the key
results of this work: how to analytically compute the Massey detector performances
under the two hypotheses H0 and H1 .

3.7 Analytical computation of the detector performances

under H1

By looking at (3.13) we immediately note that if all ri symbols have the same sign
of the pattern symbols then ΓM = 0. Otherwise, if at least one component has a
different sign, we have ΓM < 0, with a real value depending on the ri components.
This means that ΓM has a mixed pdf with:

• A discrete component in the origin.

• A continuous component for negative values.

In the following we will express the detector performance as a function of SNR, i.e.,
SNR = NE0s = 2σ1 2 dB.
The following theorem computes this pdf under H1 .

Theorem 2. The pdf under H1 of the Massey correlation detector is a mixed pdf,
with:
− A discrete component for x = 0 given by

P(ΓM = 0|H1 ) = (1 − p)N , (3.14)

with r
1 1
p = erfc .
2 2σ 2
− A continuous component for negative values given by

N  
N c
∀x < 0 fΓM (x|H1 ) = ∑ p (1 − p)N−c F −1 [F ( fy (x))c ] , (3.15)
c=1 c
42 Frame Synchronization

where F is the Fourier transform and y is a random variable with pdf:

1 1 (x−1)2
−
fy (x) = √ e 2σ 2 for x < 0 (3.16)
p 2πσ 2
fy (x) = 0 for x ≥ 0.

Proof. Under H1 , every component ri = ci + ni is an independent Gaussian random

variable centered around a fixed symbol ci . Let us suppose (without loss of generality)
that ci = +1, ∀i. As a consequence, the condition sgn(ri ) ̸= sgn(ci ) is equivalent to
the condition ri < 0. Then we can write:

ΓM = ∑ ri ci = ∑ ri . (3.17)
i: i:
sgn(ri )̸=sgn(ci ) ri <0

The probability that a single component ri = 1 + ni is negative is given by:

r
1 1
p = P(ri < 0) = P (1 + ni < 0) = P (ni < −1) = erfc . (3.18)
2 2σ 2

Then, the probability of having C negative components is:

 
N c
P (C = c) = p (1 − p)N−c . (3.19)
c

From this expression we can compute the discrete component of the pdf in the origin:

P(ΓM = 0|H1 ) = P(C = 0) = (1 − p)N ,

which is result (3.14). To compute the continuous component of the pdf for x < 0,
we introduce the random variable:
c
α (c) = ∑ yi (3.20)
i=1

obtained as the sum of c (with 1 ≤ c ≤ N) independent random variables yi corre-

sponding to negative symbols ri :

yi = 1 + ni with yi < 0.
3.7 Analytical computation of the detector performances under H1 43

The pdf of these random variables is depicted in Figure 3.3 and is given by (3.16).
Note that the pdf of yi is written starting from the Gaussian pdf of ri , but taking into
account the condition yi < 0. For this reason, the original pdf is divided by p, which
is exactly the factor p = P(ri < 0).
3.5
f y (x)
f g (x)
3

2.5

2
f y (x)

1.5

0.5

0
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
x

Fig. 3.3 Pdf of random variable y for SNR = 0 dB. The red dotted curve fg (x) is the original
Gaussian random variable centered in ci = +1.

To derive the pdf of α(c), we can compute the convolution of the c pdfs fy (x).
As an alternative, we can exploit the characteristic function approach by computing
the Fourier transform:
Φy ( f ) = F ( fy (x))

and applying:
fα(c) (x) = F −1 [Φy ( f )c ] . (3.21)

From (3.19) and (3.21) we can finally compute the pdf continuous component, which
is result (3.15):

∀x < 0
N   N
N c
fΓM (x|H1 ) = ∑ P(C = c) fα(c) (x) = ∑ p (1 − p)N−c F −1 [F ( fy (x))c ] .
c=1 c=1 c

Theorem 2 allows to compute the Massey detector pdf for any pattern length
value N and any SNR value ES /N0 = 1/(2σ 2 ). As an example, the pdfs (both
discrete/continuous components) under H1 at SNR = 0 dB are shown in Figure 3.4
44 Frame Synchronization

for N = 8 and in Figure 3.5 for N = 64. In the same figures we also verify the perfect
agreement with the simulated pdf curves, obtained by generating a sufficiently large
number of events H1 and H0 , applying the Massey detector, and collecting the test
results. MASSEY DETECTOR - pdf
N = 8 bits, Es /N0 = 0 dB
1
f ! (x|H1 ), x<0, a
0.9 M
f ! (x|H1 ), x=0, a
M
0.8
f ! (x|H1 ), x<0, s
M

0.7 f ! (x|H1 ), x=0, s

M
f (x|H ), x<0, a
0.6 !
M
0

f (x|H ), x=0, a
f ! (x)

! 0
0.5 M
M

f ! (x|H0 ), x<0, s
M
0.4 f ! (x|H0 ), x=0, s
M

0.3

0.2

0.1

0
-12 -10 -8 -6 -4 -2 0
x

Fig. 3.4 Massey correlation detector: pdfs under H1 and H0 for N = 8 at SNR = 0 dB.
Analytical (a) vs. simulated (s) curves.

MASSEY DETECTOR - pdf

N = 64 bits, Es /N0 = 0 dB
0.5
f ! (x|H1 ), x<0, a
0.45 M
f ! (x|H1 ), x=0, a
M
0.4
f ! (x|H1 ), x<0, s
M

0.35 f ! (x|H1 ), x=0, s

M
f ! (x|H0 ), x<0, a
0.3 M
f ! (x|H0 ), x=0, a
f ! (x)

0.25 M
M

f ! (x|H0 ), x<0, s
M
0.2 f ! (x|H0 ), x=0, s
M

0.15

0.1

0.05

0
-60 -50 -40 -30 -20 -10 0
x

Fig. 3.5 Massey correlation detector: pdfs under H1 and H0 for N = 64 for SNR = 0 dB.
Analytical (a) vs. simulated (s) curves.
3.8 Analytical computation of the detector performances under H0 45

3.7.1 Miss-Detection Probability

Given the pdf, we can compute the miss-detection probability. We are under H1 , the
SW is present, but the test is erroneously below the threshold and the SW is declared
absent. The following Corollary of Theorem 2 holds.

Corollary 1. The miss-detection probability of the Massey correlation detector is

given by:
Z t
ift < 0 : Pmd = P (ΓM < t|H1 ) = fΓM |H1 (x)dx
−∞
ift = 0 : Pmd = P (ΓM < 0|H1 ) = 1 − (1 − p)N (3.22)
ift > 0 : Pmd = P (ΓM < t|H1 ) = 1.

Proof. The first line is the definition of miss-detection probability, with fΓM |H1 (x)
given by (3.15). For t = 0 the integral covers the entire continuous component, then
we obtain the complement to one of the discrete component in the origin (3.14). For
any positive t > 0 the integral covers the entire pdf, then we obtain 1.

As an example, the miss-detection probability of the Massey detector at SNR =

0 dB is plotted in Figure 3.6 for N = 8 and in Figure 3.7 for N = 64, where they
also compared against the simulated curves. (Clearly, the analytical curves allow to
compute very low error rates, without resorting to very long simulations).

3.8 Analytical computation of the detector performances

under H0

In this section, we show how to analytically compute the test performances under
H0 . Also in this case, ΓM has a mixed pdf with a discrete component for ΓM = 0
(when all received ri have the same sign of the pattern symbols) and a continuous
component for ΓM < 0 (when at least one ri has a different sign). The following
theorem computes this pdf under H0 .

Theorem 3. The pdf under H0 of the Massey correlation detector is a mixed pdf,
with:
46 Frame Synchronization

MASSEY DETECTOR - P fa, P md vs. threshold

N = 8 bits, Es /N0 = 0 dB
10 0

10 -1

10 -2

10 -3

10 -4
P fa ,P md

10 -5

10 -6

10 -7
P fa a
10 -8 P fa s
P md a
-9
10 P s
md

-10
10
-12 -10 -8 -6 -4 -2 0
t

Fig. 3.6 Massey correlation detector: miss-detection and false-alarm performances, for N = 8
bits and SNR = 0 dB. Analytical (a) and simulated (s) curves.

MASSEY DETECTOR - P fa, P md vs. threshold

N = 64 bits, Es /N0 = 0 dB
10 0

10 -1

10 -2

10 -3

10 -4
P fa ,P md

10 -5

10 -6

10 -7
P fa a
10 -8 P fa s
P md a
10 -9 P md s

10 -10
-60 -50 -40 -30 -20 -10 0
t

Fig. 3.7 Massey correlation detector: miss-detection and false-alarm performances, for
N = 64 bits and SNR = 0 dB. Analytical (a) and simulated (s) curves.

− A discrete component for x = 0 given by

1
P(ΓM = 0|H0 ) = . (3.23)
2N
3.8 Analytical computation of the detector performances under H0 47

− A continuous component for negative values given by

N  
1 N
∀x < 0 fΓM |H0 (x) = ∑ N
i=0 2 i
i N−i
i−c N − i
   
i c ′ ′ ′
∑∑ p (1 − p) (1 − p)c pN−i−c F −1 [F ( fy (x))c F ( fz (x))c(3.24)
],
c=0 c′ =0 c c′
(c̸=0) OR(c′ ̸=0)

where y is a random variable with pdf given by (3.16), and z is a random variable
with pdf given by:

1 1 (x+1)2
−
fz (x) = √ e 2σ 2 for x < 0 (3.25)
1 − p 2πσ 2
fz (x) = 0 for x ≥ 0.

Proof. Under H0 , every component ri = di + ni is an independent Gaussian random

variable centered around a random symbol di . We suppose (without loss of generality)
that ci = +1. We can write:

ΓM = ∑ ri ci = ∑ ri = ∑ ri + ∑ ri . (3.26)
i: i:ri <0 i:ri <0 i:ri <0
sgn(ri )̸=sgn(ci ) di =+1 di =−1

Given a random vector d = (d1 , · · · , di , · · · , dN ), composed by i.i.d. symbols, we

denote by:

• N + the number of symbols with di = +1.

• N − = N − N + the number of symbols with di = −1.

We have:  
+ 1 N
P(N = i) = N . (3.27)
2 i

Given r, we denote by:

• C the number of negative symbols ri < 0 corresponding to di = +1.

• C′ the number of negative symbols ri < 0 corresponding to di = −1.

48 Frame Synchronization

We note that:

P(ri < 0|di = +1) = P(ni < −1)= p

P(ri < 0|di = −1) = P(ni < +1)= 1 − p.

Then the probability that C = c and C′ = c′ is given by:

 
+ i c
P(C = c|N = i) = p (1 − p)i−c
c
N −i
 
′ ′ + ′ ′
P(C = c |N = i) = ′
(1 − p)c pN−i−c .
c

When both c = c′ = 0, all symbols ri are positive and ΓM = 0. This allows to

compute the pdf discrete component under H0 :
N N  
1 N 1
P(ΓM = 0|H0 ) = ∑ P(N + = i)P(C = 0,C′ = 0|N + = i) = ∑ N
(1 − p)i pN−i = N ,
i=0 i=0 2 i 2

which is result (3.23). Even if this procedure is useful for the remaining part of the
proof, we note here that it is possible to compute this value in a simpler way. Let us
consider the output of a binary symmetric channel when the input vector is random.
The output vector is a random vector, too, and 1/2N is exactly the probability that all
its symbols are positive.
To compute the continuous component of the pdf, when at least one symbol ri is
negative, i.e., [(c ̸= 0) OR(c′ ̸= 0)], let us consider:

c c′
′ ′
γ(c, c ) = α(c) + β (c ) = ∑ yi + ∑ zi . (3.28)
i=0 i=0

The first term α(c) corresponds to the negative symbols ri < 0 with di = +1. It can
be studied as in the previous section and its pdf is given by (3.21).
The second term β (c′ ) is obtained as the sum of c′ independent random variables
zi corresponding to negative symbols ri with di = −1:

zi = −1 + ni with zi < 0.

The pdf of each of them is depicted in Figure 3.8 and is given by (3.25).
3.8 Analytical computation of the detector performances under H0 49

6
f z (x)
f g (x)
5

f z (x)
3

0
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
x

Fig. 3.8 Pdf of random variable z for SNR = 0 dB. The red dotted curve fg (x) is the original
Gaussian random variable centered in di = −1.

The pdf of γ(c, c′ ) can be computed by using the characteristic function approach:
h i
−1 c′
fγ(c,c′ ) (x) = F F ( fy (x)) F ( fz (x)) .
c

Finally, we can group together all these results to get the pdf continuous component:
N i N−i
∀x < 0 fΓM |H0 (x) = ∑ P(N + = i) ∑∑ P(C = c,C′ = c′ |N + = i) fγ(c,c′ ) (x) =
i=0 c=0 c′ =0
(c̸=0) OR(c′ ̸=0)
N i N−i
i−c N − i
     
1 N i c ′ ′ ′
=∑ N ∑∑ p (1 − p) ′
(1 − p)c pN−i−c F −1 [F ( fy (x))c F ( fz (x))c ],
i=0 2 i c=0 c′ =0 c c
(c̸=0) OR(c′ ̸=0)

which is result (3.24).

Theorem 3 allows to compute the detector pdf for any pattern length value N and
any SNR value. As an example the pdfs (discrete and continuous components) under
H0 at SNR = 0 dB are shown in Figure 3.4 for N = 8 and in Figure 3.5 for N = 64,
where we have also plotted the simulated curves.
50 Frame Synchronization

3.8.1 False-Alarm Probability

Given the pdf, we can compute the false-alarm probability. We are under H0 , the
SW is absent, but the test is erroneously above the threshold and the SW is declared
present. The following Corollary of Theorem 3 holds.
Corollary 2. The False-Alarm probability of the Massey detector is given by:
Z 0−
1
ift < 0 : Pf a = P (ΓM ≥ t|H0 ) = fΓM |H0 (x)dx +
t 2N
1
ift = 0 : Pf a = P (ΓM ≥ 0|H0 ) = (3.29)
2N
ift > 0 : Pf a = P (ΓM ≥ t|H0 ) = 0

Proof. The false-alarm probability is obtained by integrating the detector pdf given
by Theorem 3, which has a continuous component for negative values and a discrete
component in the origin equal to 1/2N . When t = 0, Pf a is equal to the discrete
component. For positive thresholds it goes to zero.

As an example, the false-alarm probability of the Massey detector at SNR = 0 dB

is plotted in Figure 3.6 for N = 8 and in Figure 3.7 for N = 64, where they are also
compared against the simulated pdf.

3.8.2 False-alarm lower bound

By looking at Corollary 1 and 2, we can observe that:

• For t = 0 we have Pf a = 1/2N and Pmd = 1 − (1 − p)N .

• For t > 0 we have Pf a = 0 and Pmd = 1.

This means that it is useless to set a positive threshold, because the pattern is never
detected. As a consequence, it is impossible to achieve a value Pf a < 1/2N : the false-
alarm rate is always lower bounded by the term due to the pdf discrete component
under H0 :
(LB) 1
Pf a ≥ Pf a = N . (3.30)
2

This is an important issue for practical applications, because the term 1/2N is
not negligible for small N. As an example, the lower bound values for some SW
3.9 ROC curves 51

lengths N are reported in Table 3.1, which shows that the bound is quite high for
small lengths (e.g., N ≤ 16).

Table 3.1 Massey correlation detector: false-alarm lower bound values for different lengths
N.

(LB)
N Pf a
8 3.90 ·10−3
16 1.53 ·10−5
32 2.33 ·10−10
64 5.42 ·10−20

Indeed, the lower bound impact is clearly visible in Figure 3.6 for N = 8, while
it is negligible in Figure 3.7 for N = 64. This limitation for small values of N can be
a problem for some applications which require very low false-alarm rates. This issue
will be further discussed in Section 3.10, where we present a solution to eliminate it.

3.9 ROC curves

Given the miss-detection and the false-alarm probability analytical expressions

derived in Section 3.8 and Section 3.7, we can now compute the ROC performances
of the Massey detector, i.e., the miss-detection vs. false-alarm curves. As an example,
Figure 3.9 and Figure 3.10 present the ROC for N = 8 and N = 64 for different SNR
values. As another example, the ROC curves at SNR = 0 dB for different values of
N are shown in Figure 3.11. Also in this case, the impact of the false-alarm lower
bound values is clearly visible in Figure 3.9 and Figure 3.11 for N = 8, and in Figure
3.11 for N = 16.
52 Frame Synchronization

MASSEY DETECTOR - ROC

N = 8 bits
0
10

10 -1

10 -2

10 -3
P md

10 -4

10 -5 E s /N0 = -4 dB
E s /N0 = -2 dB
-6 E s /N0 = 0 dB
10
E s /N0 = 2 dB
E s /N0 = 4 dB
10 -7
E s /N0 = 6 dB

10 -8
10 -4 10 -3 10 -2 10 -1 10 0
P fa

Fig. 3.9 Massey correlation detector: ROC curves for N=8 bits and different SNR values.

MASSEY DETECTOR - ROC

N = 64 bits
10 0
E s /N0 = -4 dB
10 -1 E s /N0 = -3 dB
E s /N0 = -2 dB
-2
10
E s /N0 = -1 dB
E s /N0 = 0 dB
10 -3

10 -4
P md

10 -5

10 -6

10 -7

10 -8

10 -9

10 -10
10 -10 10 -9 10 -8 10 -7 10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 10 0
P fa

Fig. 3.10 Massey correlation detector: ROC curves for N=64 bits and different SNR values.
3.9 ROC curves 53

MASSEY DETECTOR - ROC

N = 8, 16, 32, 64 bits, ES /N0 = 0 dB
0
10

10 -1

10 -2

10 -3

10 -4
P md

10 -5

10 -6

10 -7

10 -8 N=8
N=16
N=32
10 -9
N=64

10 -10
10 -10 10 -9 10 -8 10 -7 10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 10 0
P fa

Fig. 3.11 Massey correlation detector: ROC curves for SNR = 0 dB and different pattern
lengths.
54 Frame Synchronization

3.10 The Extended Massey correlation detector

As explained in Section 3.8.2, for the Massey detector we have a false-alarm lower
(LB)
bound Pf a = 1/2N . For applications requiring low false-alarm rates, this limitation
may be a problem. As an example, a target Pf a = 10−6 (which is a realistic values for
some applications, e.g. [73]) cannot be achieved by N < 32. Unfortunately, in some
cases the use of a synchronization word with N ≥ 32 can be a problem, because
the introduced overhead may be too large, especially for small data frames. In this
section we introduce a new solution, which allows to overcome this limitation.
The lower bound is due to the pdf discrete component in zero, which is induced
by the Massey detector definition. If we have sgn(ri ) = sgn(ci ) for any received
symbol (an event with probability 1/2N ), the test result is equal to zero. As a
consequence, there is no way to further reduce the false-alarm probability from the
lower bound: for positive thresholds we have Pf a = 0, but Pmd = 1; so the pattern
cannot be revealed and the detector is useless.
We now propose an enhanced definition of the Massey detector, that we call the
Extended Massey correlation detector, which is based on the following idea. If at
least one received symbol ri has a sign different from the corresponding ci symbol,
the Massey correlation is applied. If all the signs agree, the soft correlation (which
takes into account all the marker symbols) is applied.

Definition 2. The Extended Massey Correlation Detector. Given the binary an-
tipodal pattern c = (c1 , · · · , ci , · · · , cN ) and the observed real vector

r = (r1 , · · · , ri , · · · , rN ) (3.31)

the Extended Massey correlation detector computes the statistical test:

if ∃i : sgn(ri ) ̸= sgn(ci ) → ΓE = ∑ ri ci (3.32)

i:
sgn(ri )̸=sgn(ci )
if ∀i : sgn(ri ) = sgn(ci ) → ΓE = ∑ rici. (3.33)
i

For negative thresholds, the test is identical to the Massey detector. As we will
show afterwards:
3.10 The Extended Massey correlation detector 55

• Since we can use positive thresholds, too, the lower bound limitation is elimi-
nated.

• For positive thresholds the Extended test is better than or equal to the soft
correlation.

• Overall, it is nearly optimal for both positive and negative thresholds, very
close to the LRT.

To analyze this new test, we start from the analytical computation of its pdfs. We
note that:

• We apply (3.32) if and only if ΓE < 0.

• We apply (3.33) if and only if ΓE ≥ 0.

Then, for negative threshold values, the pdfs are equal to the continuous component
of the Massey detector pdf computed in Section 3.7 (H1 ) and Section 3.8 (H0 ). For
zero and positive threshold values we present two theorems to compute them.

3.10.1 Analytical computation of the Extended Massey detector

pdf under H1

Under H1 , the following theorem provides the pdf of the new test.
Theorem 4. The pdf under H1 of the Extended Massey correlation detector is given
by:
− For negative values:

∀x < 0 fΓE |H1 (x) = fΓM |H1 (x) given by (3.15).

− For zero and positive values:

h i
−1 N
∀x ≥ 0 : fΓE |H1 (x) = (1 − p) F
N
F ( fs (x)) . (3.34)

where s is a random variable with pdf given by

1 1 (x−1)2
−
fs (x) = √ e 2σ 2 for x ≥ 0 (3.35)
1 − p 2πσ 2
fs (x) = 0 for x < 0.
56 Frame Synchronization

Proof. We suppose, without loss of generality, that all ci = +1. For positive thresh-
olds we have:
N N
∀x ≥ 0 ΓE = ∑ ri ci = ∑ ri (iff all ri ≥ 0). (3.36)
i=1 i=1

The probability that all ri ≥ 0 is equal to (1 − p)N .

We introduce the random variable
N
δ (N) = ∑ si (3.37)
i=1

obtained as the sum of N independent random variables si corresponding to positive

ri with ci = +1, i.e.,

si = 1 + ni with si ≥ 0.

The pdf of each of them is given by (3.35). (Note that this pdf is similar to the pdf of
the random variable y depicted in Figure 3.3, only reflected on the positive axis.)
To derive the pdf of δ (N) we can apply the characteristic function approach:

fδ (N) (x) = F −1 F ( fs (x))N .

 
(3.38)

Finally, the pdf of ΓE under H1 for positive values is given by:

h i
−1 N
∀x ≥ 0 : fΓE |H1 (x) = (1 − p) F
N
F ( fs (x)) , (3.39)

which is result (3.34).

For negative values we have fΓE |H1 (x) = fΓM |H1 (x) given by (3.15) because we
apply (3.32) if and only if ΓE < 0: in this region the definitions of the Massey
detector and the Extended Massey detector coincide.
3.10 The Extended Massey correlation detector 57

Given the pdf, we can compute the miss-detection probability.

Corollary 3. The miss-detection probability of the Extended Massey detector is

given by:
Z t
ift < 0 : Pmd = P (ΓE < t|H1 ) = fΓM |H1 (x)dx
−∞
ift = 0 : Pmd = P (ΓE < 0|H1 ) = 1 − (1 − p)N (3.40)
Z t
ift > 0 : Pmd = P (ΓE < t|H1 ) = 1 − (1 − p)N + fΓE |H1 (x)dx
0

Proof. For negative thresholds, the pdf is equal to the Massey continuous pdf. For
t = 0, we obtain the integral of the Massey continuous component, which is the
complement to one of its discrete component, i.e., 1 − (1 − p)N as in Corollary 1.
For positive thresholds we sum to this term the integral of the Extended Massey pdf,
which is distributed on the entire positive axis. For t going to infinity, the probability
goes to one.

3.10.2 Analytical computation of the Extended Massey detector

pdf under H0

The following theorem provides the pdf of the new test under H0 .

Theorem 5. The pdf under H0 of the Extended Massey detector is given by:
− For negative values:

∀x < 0 fΓE |H0 (x) = fΓM |H0 (x) given by (3.24).

− For zero and positive values:

N  
1 N
∀x ≥ 0 fΓE |H0 (x) = N
2 ∑ d (1 − p)d pN−d F −1 [F ( fs (x))d F ( ft (x))N−d ], (3.41)
d=0

where s is a random variable with pdf given by (3.35) and t is a random variable
with pdf:

1 1 (x+1)2
−
ft (x) = √ e 2σ 2 for x ≥ 0 (3.42)
p 2πσ 2
ft (x) = 0 for x < 0.
58 Frame Synchronization

Proof. We suppose without loss of generality that all ci = +1. For positive thresholds
we have:
N N
∀x ≥ 0 ΓE = ∑ ri ci = ∑ ri (iff all ri ≥ 0). (3.43)
i=1 i=1

Under H0 we have ri = di + ni , where di are random symbols. All the ri are positive.
Let us denote by:

• D the number of positive symbols ri corresponding to di = +1.

• D′ = N − d the number of positive symbols ri corresponding to di = −1.

We note that:  
1 N
P(D = d) = N (1 − p)d pN−d , (3.44)
2 d
where the first two terms are the probability that an N−bit random vector contains d
symbols equal to +1 and the last two are the probability that all the symbols of the
corresponding observed vector are positive. Let us consider:

d d′
′ ′
ζ (d, d ) = δ (d) + ε(d ) = ∑ si + ∑ ti . (3.45)
i=0 i=0

The first term δ (d) corresponds to the components with ri ≥ 0 and di = +1. It is the
sum of d random variables si with pdf (3.35) introduced for studying H1 . The pdf of
δ (d) is then given by:
h i
fδ (d) (x) = F −1 F ( fs (x))d . (3.46)

The second term ε(d ′ ) corresponds to the components with ri ≥ 0 and di = −1 is

obtained as the sum of d ′ independent random variables ti ,

ti = −1 + ni with ti ≥ 0.

The pdf of each of them is given by (3.42). (Note that this pdf is similar to the pdf of
the random variable z depicted in Figure 3.8, only reflected on the positive axis.)
The pdf of ε(d ′ ) can be computed by using the characteristic function approach:
h ′
i
fε(d ′ ) (x) = F −1 F ( ft (x))d .
3.10 The Extended Massey correlation detector 59

Finally, we can group together all these results to get the pdf of the Extended detector
under H0 for x ≥ 0:
N
fΓE |H0 (x) = ∑ P(D = d) fζ (d,N−d)(x)
d=0
N  
1 N
=∑ N (1 − p)d pN−d F −1 [F ( fs (x))d F ( ft (x))N−d ],
d=0 2 d

which is result (3.41).

As for Theorem 4, for negative thresholds we have fΓE |H0 (x) = fΓM |H0 (x) given
by (3.24), because we apply (3.32) if and only if ΓE < 0, then in this region the two
definitions coincide.

Given the pdf, we can compute the false-alarm probability.

Corollary 4. The false-alarm probability of the Extended Massey detector is given

by:
Z 0−
1
ift < 0 : Pf a = P (ΓE ≥ t|H0 ) = fΓM |H0 (x)dx +
t 2N
1
ift = 0 : Pf a = P (ΓE ≥ 0|H0 ) = N
(3.47)
Z2 +∞
∀t > 0 Pf a = P (ΓE ≥ t|H0 ) = fΓE |H0 (x)dx.
t

Proof. For negative thresholds we have the same false-alarm rate of the Massey
detector. For t = 0, the integral covers all the non-negative values then it is equal to
1/2N (for the Massey detector this probability is concentrated in the origin, while for
the Extended detector it is distributed on the entire non-negative axis). For positive
thresholds the integral is applied only to the positive portion of the Extended detector
pdf.

3.10.3 Extended Massey detector performances

For negative thresholds, the Extended Massey detector has the same performances of
the Massey detector, but it overcomes the lower bound limitation on the Pf a because
the threshold can assume positive values, too. As an example, the Extended detector
performances in terms of Pf a and Pmd vs. thresholds are reported in Figure 3.12 for
60 Frame Synchronization

SNR = 6 dB and N = 8 bits. For comparison, in the same figure we have also plotted
the performances of the Massey detector and the soft correlation (3.10). By looking
at these results we observe that:

• For negative thresholds the Extended Massey performances are identical to

those of the Massey detector. In fact, in this region the detector rules are
exactly the same.

• For the Extended Massey detector, the lower bound limitation no longer
applies: for positive thresholds the Pf a becomes smaller than the lower bound
(LB)
Pf a = 1/2N = 3.90 · 10−3 where the Massey detector curve stops.

• For positive thresholds the Extended Massey performances becomes similar to

the soft correlation ones for large thresholds. In this region, the two detectors
use the same correlation rule but their pdfs are different. To explain this
behavior let us suppose again (without loss of generality) that all ci = +1. The
Extended Massey detector applies the soft correlation rule under the condition
that all ri are positive. Instead, the soft correlation detector always applies
it, for any received vector. In this second case, small positive values of the
test are generated by vectors r with both negative and positive ri , while large
positive values are generated by vectors r that, with high probability, have all
positive symbols. Then, the performances of the two tests are different for
small positive values, and become similar for large values.

3.11 Comparison against LRT, soft and hard correla-

tion

In this section we compare the performances of the Massey detector and the Extended
Massey detector against the optimal LRT, the soft correlation and the hard correlation.
As an example, the ROC curves of these detectors for SNR = 0 dB are compared in
Figure 3.13 for N = 8 bits and in Figure 3.14 for N = 64 bits. By looking at Figure
3.13 we can observe that:
(LB)
• As expected, the Massey detector curve stops at the lower bound Pf a =
1/2N = 3.90·10−3 (where also the hard correlation curve stops). The Extended
Massey detector does not suffer this limitation.
3.11 Comparison against LRT, soft and hard correlation 61

EXTENDED MASSEY DETECTOR

N = 8 bits E S /N0 = 6 dB
0
10

10 -1

10 -2
P fa ,P md

10 -3 P Extended Massey
fa
P md Extended Massey
P fa Massey
10 -4 P md Massey
P fa Soft
P Soft
md

-5
10
-15 -10 -5 0 5 10 15
t

Fig. 3.12 Extended Massey correlation detector: Pf a and Pmd vs. threshold, for N = 8 bits
and SNR = 6 dB.

• Even if the SNR is quite low, the Extended Massey detector ROC is essentially
coincident with the optimal LRT ROC curve and much better than the soft and
hard correlation curves.

These results are confirmed by Figure 3.14 for N = 64 bits. For this length, there
(LB)
is no need to use the Extended Massey detector, because the lower bound Pf a ≥
5.42 · 10−20 is negligible. Also in this case, the Massey detector performances are
very close to the optimal LRT ones and much better than the soft and hard correlation
curves.
62 Frame Synchronization

COMPARISON - ROC
N = 8 bits, ES /N0 = 0 dB
10 0

10 -1
P md

10 -2

10 -3 Hard
Soft
Massey
LRT
Extended

10 -4
10 -4 10 -3 10 -2 10 -1 10 0
P fa

Fig. 3.13 ROC curves for N = 8 bits and SNR = 0 dB: comparison between Extended
Massey detector, Massey detector, LRT, Soft correlation, Hard correlation.

COMPARISON - ROC
N = 64 bits, ES /N0 = 0 dB
10 0
Hard
10 -1 Soft
Massey
10 -2 LRT

10 -3

10 -4
P md

10 -5

10 -6

10 -7

10 -8

10 -9

10 -10
10 -10 10 -9 10 -8 10 -7 10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 10 0
P fa

Fig. 3.14 ROC curves for N = 64 bits and SNR = 0 dB: comparison between Massey detector,
LRT, Soft correlation, Hard correlation.
3.11 Comparison against LRT, soft and hard correlation 63

3.11.1 Miss-detection vs. SNR performances

To better quantify the differences between the detectors, we now consider the miss-
detection vs. SNR performances at fixed false-alarm rate. Fixed a target value for Pf a ,
by using Corollary 2 for the Massey detector or Corollary 4 for the Extended Massey
detector we can compute the minimum value of the threshold t which guarantees to
achieve it. Given t, by using Corollary 1 for the Massey detector or Corollary 3 for
the Extended Massey detector, we can compute the corresponding value of the Pmd .
In this way, we can plot the miss-detection probability as a function of the SNR, for
the chosen target false-alarm probability value.
These curves are very useful for practical applications, where the systems are
typically designed for achieving a given value Pf a (the so-called CFAR (Constant
False-Alarm Rate) property): these results show which are the corresponding Pmd
values obtained for different SNR values. Moreover, fixed the target Pf a and the
target Pmd they provide the SNR value necessary to achieve them.
As an example, Figure 3.15 and Figure 3.16 show the results for N = 8 and
N = 64, with different target false-alarm rates. In Figure 3.15, since N = 8 is a small
value, we have considered the Extended Massey detector. We can observe that it is
very close to the optimal LRT, with a penalty limited to a small fraction of dB even
when the SNR is low, which disappears at higher SNR values. The gains achieved
by the Extended Massey detectors with respect to hard and soft correlation are large.
If we fix target Pmd = Pf a = 10−2 the gain is 3 dB from hard and more than 6 dB
from soft (that in this scenario is worse than hard correlation, which is not surprising
because the soft correlation is not derived as an optimal test). The gains are larger
for lower miss-detection probabilities.
In Figure 3.16, since N = 64 is a large value and the lower bound is negligible,
we have considered the Massey detector. It shows a limited penalty from the optimal
LRT, even when the SNR is low. At target Pmd = Pf a = 10−6 the Massey detector
gains 2.5 dB with respect to the hard correlator and 3 dB from the soft. Higher gains
are achieved at lower miss-detection probabilities.
64 Frame Synchronization

COMPARISON - Pmd vs. SNR

N = 8 bits, Pfa=10 -2
0
10
Hard
Soft
Extended Massey
LRT
10 -1
P md

10 -2

10 -3

10 -4
-4 -2 0 2 4 6 8 10 12 14
E s /N0 [dB]

Fig. 3.15 Extended Massey detector: Pmd vs. SNR, for N = 8 bits and Pf a = 10−2 .

COMPARISON - Pmd vs. SNR

N = 64 bits, Pfa=10 -6
10 0
Hard
10 -1 Soft
Massey
10 -2 LRT

10 -3

10 -4
P md

10 -5

10 -6

10 -7

10 -8

10 -9

10 -10
-5 -4 -3 -2 -1 0 1 2 3 4
E s /N0 [dB]

Fig. 3.16 Massey detector: Pmd vs. SNR, for N = 64 bits and Pf a = 10−6 .
3.12 Some Aspects for Practical Implementation 65

3.12 Some Aspects for Practical Implementation

The presented results show that the Extended Massey and the Massey detector:

• are nearly optimal tests at both low and high SNR values, with penalty in the
order of a fraction of dB from the optimal LRT.

• are much better than the popular hard and soft correlation tests.

The choice between the two Massey detectors depends on the target false alarm
rate. If it is below the lower bound 1/2N we use the Extended Massey detector,
otherwise we use the Massey detector. As a consequence, the Extended Massey
detector looks more appropriate for small N (e.g., for N ≤ 16) and the Massey
detector for large N.
The complexity of the two Massey detectors is very low. Obviously it is slightly
higher than the complexity of the hard correlation which works on bits, but it
is essentially identical to soft correlation complexity and much lower than LRT
complexity, which may be a serious issue for high bit rate applications.
Moreover, if we look at the false-alarm and miss-detection vs. threshold curves
of the Massey detectors shown in Figures 3.6, 3.7, and 3.12, we can see that they
have a favorable behavior for threshold setting, because the curves slope is not too
pronounced. Having fixed the target error rate to be achieved, the threshold can be
easily obtained by inverting the analytical curves.
Chapter 4

Cognitive Radio: Spectrum Sensing1

4.1 Introduction

3G and 4G technologies have mainly focused on the mobile broadband use case,
providing enhanced system capacity and offering higher data rates. This focus will
clearly continue in the future 5G era where they will provide mobile telephony and
mobile broadband with wide area coverage. Moreover, they will be extended by lot
of new applications and lot of new use cases and new scenarios. Some of them can be
the evolution of existing technologies but some of them will need new technologies.
In addition, the upcoming ‘5G wireless networks’ is sought to offer wireless access
to anyone and anything. Thus, in the future, wireless access will go beyond humans
and expand to serve any entity and device that may benefit from being connected.
Thus, 5G networks will undoubtedly explore unlimited possibilities until an official
standard by telecommunications standardisation bodies such as ITU-R, IMT and the
3GPP.
Machine type communications will be more and more important in the future.
This will include communications for electronics and home appliances, communi-
cations for transports, communications for medical equipments, internet of things
and many more. A lot of dense communications will be probably carried by existing
wireless access technologies but some of these use cases will have so high require-
ments that they will require new wireless access technologies. For example, ultra
1 Part of the work described in this chapter has been previously published in [49, 44, 146, 147, 45–
47]
4.1 Introduction 67

reliable communications may be needed for critical infrastructure, or for industrial

processes. This may very well require new radio access technologies. Car-to-car or
car-to roadside communication will be introduced for traffic safety and traffic control.
And new ways to communicate will be exploited for example, communicating di-
rectly between devices. All these factors, including new technological advancements,
emergence of new applications and growth in user demand for wireless services and
number of devices contribute to the need for additional licensed or in some respect
repurposed spectrum to accommodate new capabilities of 5G wireless systems.
Moreover, each use cases of 5G have specific technical requirements that need
to be addressed through adequate design of the 5G radio interface(s) and access
to appropriate frequency ranges. While some of these applications, such as high
resolution video, would require ultra-fast connection speeds, others might need
very robust performance and wide reaching range. It should be noted that some
of the 5G applications will be supported by evolved 4G systems with existing
spectrum. However, 5G systems will provide additional capabilities and as a result,
consideration of required spectrum for 5G should include all applications foreseen
for future networks.
To provide an example for factors affecting the amount of spectrum, ultra-high
speed connections in the range of multi-gigabit per second could potentially be
achieved through using ultra-wide carrier bandwidths in the order of up to several
hundred MHz or more. An example could be fast downloads of 4k/8k video content,
which using wide channels and multi-gigabit speeds, would take seconds. Another
example of factors affecting the spectrum choice is the case of ultra-reliable commu-
nications for mission critical applications such as public safety, where obstacle and
ground penetration for ubiquitous coverage in critical times would require use of
lower frequencies such as those in the lower Ultra-High Frequency (UHF) band. In
order to come up with a mapping between applications and required spectrum, Table
4.1 lists potential spectrum-related implications of various high- level requirements
for future 5G systems. Licensed spectrum is currently available for mobile uses in
small chunks (contiguous 50 MHz at most), which fails to meet the 5G target data
rate. In fact, predicted spectrum demands from ITU-R suggests the extra spectrum
requirements in 2020 will be 12801720 MHz [15] to supplement the current allocated
spectrum in mobile networks. Significant portion of sub-6GHz bands (with ideal
propagation characteristics for mobile uses), is currently allocated to non-mobile
spectrum users, e.g., military, radar, etc., across the globe.
68 Cognitive Radio: Spectrum Sensing

Table 4.1 Potential Spectrum-Related Implications of various 5G Requirements [15]

High-Level Requirement Potential Spectrum-Related Implications

Ultra-high speed Ultra-wide carrier bandwidths, e.g. 500 MHz
radio links Multi-gigabit fronthaul/backhaul
High speed Wide carrier bandwidths, e.g. 100 MHz
radio links Gigabit fronthaul/backhaul
Support for low to high- Depends on the throughput
Doppler environment requirement
Ultra-low latency Short range implications
Low latency Mid-short range implications
Ultra-high reliability Severe impact of rain and other atmospheric
radio links effects on link availability in higher freq.,
e.g. mm-wave for outdoor operations
Impact of rain and other atmospheric
High reliability radio links
effects on link availability in higher freq.,
e.g. mm-wave for outdoor operations
Short range Higher frequencies, e.g. mm-wave
Long range Lower frequencies, e.g. sub-3 GHz
Ground/obstacle penetration Lower frequencies, e.g. sub-1 GHz
Operation in cluttered Diffraction dominated env. in lower freqs.
environment Reflection dominated env. in higher freqs.
Operation near fast moving
Frequency-selective fading channels
obstacles
High-speed distributed wireless
Mesh networking
backhaul operating in-band or out-of band

Wireless channels are usually characterized by a fixed spectrum assignment

policy. Electromagnetic spectrum is strictly regulated and licensed by governmental
entities, for instance, Federal Communications Commission (FCC) in US, Office of
Communications (OFCOM) in UK, Ministero dello Sviluppo Economico - Diparti-
mento per le Comunicazioni in Italy and Nepal telicom Authority (NTA) in Nepal.
The rapid development in the field of wireless communications certainly creates a
big challenge for every licensing organization to accommodate all the new applica-
tions, services and use cases of 5G with the limited electromagnetic spectrum. The
frequency allocation chart of UK [57] (see Figure 4.1), Italy [57] and US [71] shows
that large portion of the radio spectrum is already assigned to traditional services
(Mobile, Maritime Mobile, Fixed Satellite Services, Radio Navigation). Even though
there are few genuinely unlicensed bands, a good example of which is Industrial
4.1 Introduction 69

Part of the Chemring Group

Fig. 4.1 OFCOM frequency allocation chart for UK [57]

Scientific & Medical (ISM) band in 2.4 GHz, available which could be the possible
solution for accommodating new services, but there are already multiple wireless
technologies in these bands such as 802.11 WLAN, Cordless Phones, Bluetooth,
Wireless Personal Area Network (WPAN), HIPERLAN etc. Some other examples of
unlicensed frequency bands include UNII-1 and UNII-2 bands where systems such
as IEEE802.11a WLAN and IEEE802.11n WLAN operate. This rapid expansion
in the field of Wireless Communications has led to a common belief that the day
is not too far when the increasing demands of the spectrum access will lead to the
spectrum scarcity.
Frequency spectrum is a scarce resource in the field of wireless communication.
Thus, its utilization efficiency should be made as efficient as possible. In contrast,
several measurements carried out to verify the spectrum utilization in different
countries have shown significant under-utilization of the allocated frequency bands.
Noticeably, the Digital Television (DTV) band is an example of inefficient spectrum
use, since depending on the geographical location, only certain channels are always
occupied both in the temporal and spatial domains. Concerning the same, FCC
Spectral Policy Task Force reported that typical radio channel occupancy is less than
70 Cognitive Radio: Spectrum Sensing

15% while the peak occupancy is close to 85% in US [70]. Similar measurements
carried out in various countries show that most of the radio frequency spectrum
is inefficiently utilized with spectrum utilization mostly in the range of 5% to
50% [96, 166]. These thought-provoking statistics in synergy provide the drive for
main challenges facing the spectral world. Experts around the world are interested
in tackling how we can manage future spectrum more effectively to serve for 5G use
cases. A number of technologies and techniques have been identified as enablers for
the 5G wireless networks and among them is Cognitive Radio technology. Many
studies are suggesting that 5G could exploit Cognitive Radio approaches to increase
band availability, especially for very small cells.

4.1.1 Cognitive Radio Technology

Cognitive Radio was first officially proposed by Joseph Mitolla II at the Royal Insti-
tute of Technology [123, 124] but the idea of using learning and sensing machines
to improve radio spectrum was envisioned several decades earlier. Now, Cognitive
Radio is seen as an emerging solution to the problem of spectrum scarcity. It has
an implicit mechanism of spectrum sensing intelligence and decision making that
enables opportunistic access to the under-utilized frequency slots of licensed or
unlicensed frequency bands which is also evident from the definitions of Cognitive
Radio adopted by FCC: [40]

“A radio or system that senses its operational electromagnetic environ-

ment and can dynamically and autonomously adjust its radio operating
parameters to modify system operation, such as maximize throughput,
mitigate interference, facilitate interoperability, access secondary mar-
kets.”

There are some more definitions of Cognitive Radio which are similar in theme.
Cognitive Radio defined by Haykin [87] is:

“Cognitive radio is an intelligent wireless communication system that is

aware of its surrounding environment (i.e., outside world), and uses the
methodology of understanding by building to learn from the environment
and adapt its internal states to statistical variations in the incoming RF
4.1 Introduction 71

stimuli by making corresponding changes in certain operating parame-

ters (e.g., transmit-power, carrier-frequency, and modulation strategy)
in real-time, with two primary objectives in mind: highly reliable com-
munication whenever and wherever needed: efficient utilization of the
radio spectrum.”

Being a multidisciplinary nature of subject, working on Cognitive Radio is more

challenging and therefore exciting to most of the researchers. However, Cognitive
Radio Technology is still in its infancy, better to say in a research phase, despite
few attempts are done in developing wireless communication standards based on
the concept of Cognitive Radio. IEEE 802.22 Wireless Regional Area Network
(WRAN) [3] is the first international standard exploiting TV white spaces using Cog-
nitive Radio Technology. WRAN is envisioned to operate in Very High Frequency
(VHF) and Ultra High frequency (UHF) (54 − 862 MHz) bands currently licensed
for analog and digital television broadcasting and wireless microphones. ECMA
TC48-TG1 [64], IEEE 802.11 [94], Dynamic Spectrum Access Networks Standards
Committee (DySPAN - SC) [156], IEEE 802.16 [7], and IEEE 802.19 [6] are some
other emerging standardization initiatives based on the platform of Cognitive Radio
concept.
The Cognitive Radio Network paradigm defines two user environments coexisting
in a common spatial location exploiting the same radio environment in a quasi-
cooperative manner. In Cognitive Radio terminology, a Primary User (PU) refers
to incumbent licensees, which have a legal priority to use the allocated spectrum
whereas a Secondary User (SU) refers to unlicensed users who seek opportunity
to use the spectrum when primary users are idle. The user territory of primary
users and secondary user corresponds to primary user environment and secondary
user environment respectively. Some of the common technologies operated using
legacy rights are Global System for Mobile (GSM), Worldwide Interoperability for
Microwave Access (WiMAX) and LTE.
The concept behind Cognitive Radio (CR) is to exploit the underutilized spectral
resources by reusing them in an opportunistic manner. SU adjust its carrier frequency,
transmit power, modulation, coding etc to make best use of the available spectrum
and to achieve the desired quality of service. Even though PU no doubt can use the
spectrum as per their requirements, SU can access spectral resources of PU, when
PU is not using it and vacate the frequency band as soon as the PU becomes active
72 Cognitive Radio: Spectrum Sensing

without causing any interference. It is required that the unlicensed secondary users
should not cause harmful interference to the licensed primary users. Therefore, SU
have to make sure that the licensed spectrum bands are not utilized by the PU before
they access these bands, which makes spectrum sensing one of the key technologies
for the implementation of CR.

4.1.2 Spectrum Sensing

Among the functionalities provided by Cognitive Radio, Opportunistic Spectrum

Access (OSA) is devised as a dynamic method to increase the overall spectrum
efficiency by allowing SUs to utilize unused licensed spectrum. For this purpose, a
correct identification of channel vacuity conditions by means of spectrum awareness
techniques becomes fundamental. Spectrum sensing is defined as “The task of
finding underutilized bands both in temporal and spatial domain, by sensing the
radio spectrum in the local neighborhood of the Cognitive Radio receiver in an
unsupervised manner” [88]. Literature referring to these underutilized portions of
spectrum as spectrum holes or white spaces.
Specifically, the task of spectrum sensing may involve the following sub-tasks: [88]

• Detection of spectrum holes;

• Spectral Resolution of each spectrum hole;

• Estimation of spatial directions of incoming interferes;

• Signal classification;

In essence, spectrum hole detection is a very critical component of Cognitive

Radio Concept. Table 4.2 shows the currently understood requirements about the
secondary devices (CPEs) sensitivity for three signal types. According to 802.22
Working Group [3], for a receiver noise figure of 11 dB, the resulting required
SNR for the secondary receiver is listed, where noise power is calculated over a
bandwidth of 6 MHz for a TV signals and over a bandwidth of 200 KHz for wireless
microphones. It is also clear from above Table 4.2 that each SU sensors is required
to operate under very low SNR values. In general, such low SNR values must be
expected in all deployment scenarios of CR to protect the primary users from undue
4.1 Introduction 73

Table 4.2 Receiver Parameter for 802.22 WRAN [3]

Wireless
Parameter Analog TV Digital TV
Microphone
Probability of
0.9 0.9 0.9
Detection
Probability of
0.1 0.1 0.1
false alarm
Channel Detection
≤ 2s ≤ 2s ≤ 2s
Time
Incumbent Detection
-94dBm -116dBm -107dBm
Threshold
SNR 1dB -21dB -12dB

interference. Thus, the goal is to design detection algorithm that meet the given
sensing constraint at very low SNR.
Spectrum sensing is the most challenging task which should be taken with prime
importance for the Dynamic spectrum Access (DSA) and autonomous exploitation of
locally unused spectrum, to provide new paths for spectrum access. It is the enabling
unit of SU for the accurate identification and exploitation of the PU spectrum.
Moreover, the detection performance of the spectrum sensing device quantifies the
interference level caused to the PU transmission in the spectrum band.
Several spectrum sensing methods have been considered for Cognitive Radio
applications including Energy Detection (ED) [164], Match Filtering [29], Feature
Detection Algorithms [33], proposed for individual SU and their cooperative counter-
part for multiple SU sensing. A survey of different spectrum sensing methodologies
for Cognitive Radio [18, 16] shows that a remarkable spectrum sensing performance
can be attained with feature detection (e.g., Cyclic Prefix (CP) based detector and
Pilot Based Detector) that exploit some known characteristic of PU signals at the
expense of long observation time. Even more, Matched Filtering is assumed to
perform best with high processing gain at the constraint of knowledge of PU signal
properties [141]. ED is considered to be the simplest detector which does not make
any assumption on the PU signal statistics. Feature detection lies in the middle of
these two extremes and only makes certain assumptions on the statistical properties
of the PU signal while designing the PU signal. Even though Match filtering and
feature detection algorithms are known to outperform ED algorithm, the requirement
74 Cognitive Radio: Spectrum Sensing

of knowledge of the PU signal characteristics at the expense of long observation time

make them less suitable for spectrum sensing in Cognitive Radio Network, where
knowledge of primary user signal is usually unavailable.
In recent years, sensing techniques based on the eigenvalues of the received
signal co-variance matrix evolved as a promising solution for spectrum sensing [139].
EBD schemes can be categorized as “semi-blind EBD schemes”: which assume
knowledge of noise level and “blind EBD schemes”: which do not assume noise
level knowledge. Methods belonging to the first class provide better performance
when the noise variance is known exactly, whereas blind methods are more robust to
uncertain or varying noise level.
Semi-blind detection algorithms like ED and EBD achieve very good perfor-
mance in a known noise power level scenario. Thus, knowledge of noise power
level plays prime role for the better performance ED and EBD spectrum sensing
algorithms. Unfortunately, in real practice, the noise uncertainty is very often present.

4.1.3 Spectrum sensing under practical imperfection

The spectrum awareness problem for a CR system is different from the legacy
wireless communication ststems, for these reseasons: (i) no prior knowledge of
the PU signal structure, channel and noise variance is available (ii) the information
on spectrum occupancy, SNR, etc. must be acquired at the shortest possible time,
(iii) primary SNR may be very low due to fading and multi-path phenomenon, and
(iv) noise/interference power may vary in time giving rise to noise/interference
power uncertainty. Thus, despite the significant volume of available literature on
spectrum sensing under ideal scenarios, investigation under practical constraints and
imperfections are still lacking [154].
In this direction, the main focus of this chapter is to provide a comprehensive
analysis of the popular Semi-blind’ spectrum techniques (ED and EBD) under two
of the most important practical imperfections.

1. Noise uncertainty

2. Uncertainty on PU traffic
4.2 Semi-blind SS techniques under noise uncertainty 75

4.2 Semi-blind SS techniques under noise uncertainty

In real world scenario, the information about the PU signal is generally not available
and, even if available, Match Filtering and Feature Detection algorithms needs
sensing receivers for all types of signals. Thus, for CR application the most popular
sensing algorithms are semi-blind spectrum sensing techniques which requires the
knowledge of noise floor. Performance of these algorithms over AWGN and different
fading channels has been studied in many works including [164, 180, 50, 51, 105,
162, 161]. These works assumed a perfect knowledge of the noise power at the
receiver, which allows for a perfect threshold design. However, in real systems
the detector does not have a prior knowledge of the noise level. In recent years,
variation and unpredictability of the precise noise level at the sensing device have
been identified as a critical issue, which is also known as noise uncertainty.
Noise is an unwanted random process which may arise from several sources in
the external environment as well as from every components of a receiver chain. In
real world scenarios, noise distribution is not known to infinite precision and the
noise is neither perfectly Gaussian, perfectly white, nor perfectly stationary [159].
Therefore, the noise variance in practice has to be estimated by using a proper noise
calibration method. The noise calibration can be done either during the sensing
process or by carrying out on-site Out of Bands (OoB) measurements. Another
option for noise calibration is to use inband measurements at the frequencies where
the pilot is absent so that the noise statistics can be calibrated at the pilot frequencies
[158].
The noise estimation can be perfect in the ideal case, however, in practice, accu-
rate estimation of the noise variance is not possible, thus limiting the performance
of the semi blind sensing techniques like E and EBD, at low SNRs. For ED, the
noise variance uncertainty may lead to SNR wall phenomenon due to which it is
not possible to achieve the robust detection performance beyond a certain SNR
value even by increasing the sensing duration [158]. In this work, we present two
alternative for the estimation of the noise variance:

1. Offline noise estimation: In this approach, a maximum likelihood noise vari-

ance is used under the assumption that pure noise samples are available in
prior. The pure noise samples can be obtained from inband measurement or
from out of band measurements as stated above.
76 Cognitive Radio: Spectrum Sensing

2. Online noise estimation: In this approach, noise variance is estimated implicitly

from auxiliary received signal slots which are declared noise only samples by
the detection algorithms.

Thus, in this section we introduce the idea of auxiliary noise variance estimation
for noise uncertainty. Later Hybrid approach of ED and EBD are presented and the
impact assessment of noise power estimation is then analysed.

4.2.1 Energy Detection and its performance under Noise Uncer-

tainty (State-of-art)

For CR application, the most popular sensing algorithm is the simple ED that
compares the energy of the received signal with the noise variance σv2 . ED requires
the knowledge of σv2 value. Performance of ED in AWGN and different fading
channels has been studied in many works including [164, 50, 51, 105]. These works
assumed a perfect knowledge of the noise power at the receiver, which allows for
perfect threshold design. In that case ED can work with arbitrarily small value of
false alarm probability PFa and Miss-detection Probability PMd , by using sufficiently
large observation time, even in low SNR environment [159].
With the scope of reducing the impact of noise uncertainty on the signal detection
performance of ED, several researches have been proposed including [158, 159, 118,
119]. Hybrid Spectrum Sensing Algorithms based on the combination of ED and
Feature Detection techniques have been proposed for the reduction of the effect of
noise variance uncertainty [125]. Similar Hybrid Spectrum Sensing approach was
discussed in [19] utilizing the positive points of ED and Covariance Absolute Value
detection methods while Sequeira et al. [152] used Akaike Information Criteria (AIC),
Minimum Description Length (MDL) and Rank Order Filtering (ROF) methods for
noise power estimation in presence of signal for energy based sensing. In [159] the
fundamental bounds of signal detection in presence of noise uncertainty are analyzed.
This study showed that there are some SNR thresholds under noise uncertainty
known as SNR Wall, that prevents achieving the desired performance even if the
detection interval is made infinitely large. It concludes that the robustness of any
detector can be quantified in terms of the SNR Wall giving the threshold below which
weak signals cannot be detected reliably no matter how many samples are taken.
In [119] authors performed the asymptotic analysis of the Estimated Noise Power
4.2 Semi-blind SS techniques under noise uncertainty 77

(ENP) to derive the condition of SNR Wall phenomenon for ED. They suggested
that the SNR Wall can be avoided if the variance of the noise power estimator can
be reduced while the observation time increases. Yonghong Zeng [182] proposed
a uniform noise power distribution model for the noise uncertainty study of ED
in low SNR regime. Similarly, [109] proposed a discrete-continuous model of the
noise power uncertainty for the performance analysis of the ED in presence of noise
uncertainty. Performance of ED using Bartlett’s estimate is studied in [85].

4.2.2 Eigenvalue Based Detection and its performance in Noise

Uncertainty (State-of-art)

Indeed, in recent years various new algorithms able to outperform ED have been
applied to CR, mostly based on Random Matrix Theory (RMT) and information
theoretic criteria. Two thorough reviews have been presented in [183] and [180].
Different diversity enhancing techniques such as multiple antenna, cooperative and
oversampling techniques have been introduced in the literature to enhance the spec-
trum sensing efficiency in the wireless fading channels [143, 162, 181, 117, 169].
Most of these methods use the properties of the eigenvalues of the received signal’s
co-variance matrix and use results from advanced RMT. In particular, sensing tech-
niques based on the eigenvalues of the received signals co-variance matrix recently
emerges as a promising solution, as they also do not require a prior assumption on
the signal to be detected, and typically outperform the popular ED method when
multiple sensors are available. The various EBD algorithms can be divided in two
classes:

A. Those that require the prior knowledge of the noise variance σv2 (called“semi-
blind” in [180]). This class includes channel independent tests [183], and
Roy’s Largest Root Test (RLRT) [148] which shows the best performance in
this class (see Sub-Section 4.2). [This class also includes classical ED]

B. Those that do not require knowledge of σv2 (called “blind” in [180]). This
class includes the Eigenvalue Ratio Detection (ERD) [179], channel and noise-
independent tests [183], information theoretic criteria detectors like the AIC
and the MDL [169], and the Generalized Root Test (GLRT) [26], which shows
the best performance in this class.
78 Cognitive Radio: Spectrum Sensing

Methods belonging to the first class provide better performance when the noise
variance is known exactly, whereas blind methods are more robust to uncertain or
varying noise level. Recently, few research works have studied the impact of noise
variance uncertainty and their effect in semi-blind EBD including [129, 139, 104].
In [129] author showed the importance of the accurate noise estimation for the EBD
algorithms performance.

4.2.3 Model Scenario

This chapter focuses on the scenario described below:

• Single, unknown, primary signal. Its samples are modelled as Gaussian and
independent.

• Flat-fading channel, constant over the entire sample window.

• Additive Gaussian white noise.

The problem is formulated as a simple binary test between the mutually exclusive
hypotheses:

H0 (single primary signal absent) and H1 (single primary signal present).

This model is mostly used in detection theory because it allows a clear analytical
approach and represents a benchmark for the case of non-parametric analysis of a
single unknown primary signal. In particular relating to CR applications, where
it is very popular and it was adopted by many relevant papers (including work by
Debbah [26], and many others). The reason is that, despite of its simplicity, it is well
matched to many practical situations:

1. Single primary user. In many CR scenarios, the primary signal of interest for
a secondary opportunistic CR network is unique. This is the case, for example,
in reuse of TV signal bands [3] or the coexistence between a Wi-Fi access
point and a Wireless Sensor Network (WSN) or Bluetooth in the 2.4 GHz band,
two typical CR applications that can already be realized in practice. Moreover,
when multiple sources are present, the performance essentially depends on the
component with the highest received power.
4.2 Semi-blind SS techniques under noise uncertainty 79

2. Gaussian primary signal. Most CR sensing algorithms working on the time-

axis use non-parametric detection that does not exploit the (complete or lim-
ited) knowledge of the signal shape. This is a realistic assumption for several
CR wireless applications: even if we know that the primary signal has a
PSK/QAM/OFDM format, the secondary network is not synchronized, neither
in carrier nor in time (this would require a great amount in complexity, not
available for most applications). Then, the I/Q samples does not corresponds
to the constellation signals and do not have special properties (e.g. constant
envelope for PSK signals). Under these conditions, the Gaussian approxima-
tion for the signal amplitude turns out to be appropriate for many practical
situations.

3. Uncorrelated signal samples. In practical CR sensing, some correlation be-

tween adjacent samples may be present , but (a) it strongly depends on the
shape of the transmitting and the receiver filters and (b) the sampling frequency
is completely asynchronous with respect to the received signal. For this reason,
it is difficult to be modelled. Furthermore, including it into the framework is
expected to have a negligible impact on the detector performance.

4. Channel and noise. The flat fading channel assumption is rather realistic when
the sampling window time is relatively short and the system mobility is limited,
which is the typical scenario for current CR applications. Finally, the Gaussian
model for the noise is appropriate in general. (Impulsive noise is usually of
secondary importance for CR wireless applications.)

4.2.4 Mathematical framework

The detector builds its test statistic from K sensors (receivers, nodes or antennas)
and N time samples for each of them. Let,

y(n) = [y1 (n) · · · yk (n) · · · yK (n)]T , (4.1)

be the K × 1 received vector at time n, where the element yk (n) is the discrete
baseband complex (I/Q) sample at receiver k and time n. Under H0 , the received
vector contains only noise and consists of K independent complex Gaussian noise
80 Cognitive Radio: Spectrum Sensing

samples with zero mean and variance σv2 :

y(n)|H0 = v(n), (4.2)

where v(n) ∼ NC (0K×1 , σv2 IK×K ).

Under H1 , the received vector contains signal plus noise:

y(n) = x(n) + v(n),

= hs(n) + v(n), (4.3)

where,

• s(n) is the transmitted signal sample at time n, modeled as Gaussian with zero
mean and variance σs2 : s(n) ∼ NC (0, σs2 )

• h is the channel complex vector h = [h1 · · · hK ]T ; assumed to be constant and

memoryless during the sampling window.

Under H1 , the average SNR at the receiver is defined as

E∥x(n)∥2 σs2 ∥h∥2

ρ≜ = (4.4)
E∥v(n)∥2 Kσv2

where, ∥.∥2 denotes the Frobenius norm and E[.] is the mean operator. The received
samples are stored by the detector in the K × N received matrix

Y ≜ [y(1) · · · y(n) · · · y(N)] (4.5)

Let us introduce the 1 × N signal matrix S ≜ [s(1) · · · s(n) · · · s(N)] and the K × N
noise matrix V ≜ [v(1) · · · v(n) · · · v(N)]. All the signals samples of S and the noise
samples of V are assumed statistically independent. The detector must distinguish
between

Y|H0 = V and Y|H1 = hS + V.

The sample covariance matrix is given by

1
R≜ YY H , (4.6)
N
4.2 Semi-blind SS techniques under noise uncertainty 81

and we will denote by λ1 ≥ · · · ≥ λK its eigenvalues sorted in decreasing order.

4.2.5 Detection Performance Parameters

Let T be the test statistic employed by a detector to distinguish between H0 and

H1 . To make the decision, the detector compares T against a pre-defined threshold
t: if T > t it decides for H1 , otherwise for H0 . As a consequence, the probability
of false alarm is defined as PFa = Pr(T > t|H0 ) and the probability of detection as
PD = Pr(T > t|H1 ). (see Figure 4.2).
Both PFa and PD are key quantities for practical CRN: PFa must be low to
maximize the spectrum exploitation by the secondary user and PD must be high
to minimize the interference caused by the opportunistic user to the primary one.
As an example, the WRAN standard [3] imposes stringent requirements on both
of them: PFa < 0.1 and PD > 0.9. In practical applications, the decision threshold
t is typically computed as a function of the target PFa : this ensures the so-called
Constant False Alarm Rate (CFAR) detection.

Receiver Operating Characteristics

In order to compare the performances for different threshold values ROC curves can
be used. The ROC curve is obtained by plotting the probability of correct detection
versus the probability of false alram. ROC curves allow us to explore the relationship
between the sensitivity (Probability of detection) and specificity (Probability of false
alarm) of a sensing method for a variety of different threshold, thus allowing the
determination of an optimal threshold.

The Neyman-Pearson test

The usual criterion for comparing two tests is to fix the false alarm rate PFa and look
for the test achieving the higher PD . The Neyman-Pearson (NP) lemma [130] is
known to provide the Uniformly Most Powerful (UMP) test, achieving the maximum
possible PD for any given value of PFa . The NP criterion is applicable only when
both H0 and H1 are simple hypotheses. In our setting this is the case when both the
82 Cognitive Radio: Spectrum Sensing

Fig. 4.2 Probability of false alarm PFa and Probability of Miss-detection PD

noise level σv2 and the channel vector h are a prior known. The NP test is given by
the following likelihood ratio:

p1 (Y; h, σs2 , σv2 )

TNP = . (4.7)
p0 (Y; σv2 )

The NP test provides the best possible performance, but requires exact knowledge of
both h and σv2 . For most practical applications, the knowledge of h is questionable.
The noise variance is somewhat easier to know: since we only consider thermal
noise, if the temperature is constant some applications may possess an accurate
estimation of it.

4.2.6 Maximum Likelihood Noise Estimation Model

Maximum Likelihood Estimation (MLE) is a method of estimating the parameters

of a statistical model. When applied to a given statistical model of received noise
samples, maximum-likelihood provides estimates for the variance of a data set.
Maximum-likelihood estimation gives a unified approach to estimation, which is
well-defined in the case of the normal distribution and many other problems.
Consider Independent and Identically Distributed (IID) Gaussian noise samples
with mean zero and unknown variance. For N time samples of noise v(n) taken
by the nodes in a given detection interval, the MLE of the variance is given by the
formula [92].
2 1 K M
σ̂v = ∑ ∑ |vk (m)|2 . (4.8)
KM k=1 m=1
4.2 Semi-blind SS techniques under noise uncertainty 83

Fig. 4.3 Block diagram of Maximum Likelihood Noise Estimation

From N noise samples in one detection interval (one slot), one MLE noise estimate
is generated. To make the estimate more accurate, S such estimates are averaged
after getting new estimate of the noise variance forming a Moving Average System
(MAS) as shown in Figure 4.3. Thus, the final expression for the noise variance
estimation is given by,

1 S K M
σ̂v2 (S) = ∑ ∑ ∑ |vk (m)|2 . (4.9)
KSM s=1 k=1 m=1

Hybrid approach of noise estimation

Basically, there are two approaches of noise variance estimation depending on the
availability of the noise signal samples. Based on the noise variance estimation ap-
proach, two types of hybrid techniques of ED/RLRT are proposed in this dissertation,
namely,

A. Offline Estimation (Hybrid Approach-1)

B. Online Estimation (Hybrid Approach-2)

These methods are described in the following subsections.

84 Cognitive Radio: Spectrum Sensing

Fig. 4.4 Hybrid approach-1: Noise variance estimated from S auxiliary noise only slots.

A Offline Estimation
In this approach, noise variance is estimated explicitly assuming pure noise
samples are available. As can be seen from the block diagram shown in
Figure 4.4, noise variance is estimated from explicitly available noise only
samples taking N samples at a time. Accurate estimate of the noise variance is
obtained by averaging S variances from S noise-only slots using the MAS. The
average noise variance estimate is then used in detecting the received primary
signals using ED/RLRT detectors. After each detection interval, the average
noise variance estimate is updated using one new noise variance estimate and
the process is repeated. For ED and RLRT algorithm using this approach of
noise estimation, Hybrid Energy Detection (HED1) and Hybrid Roy’s Largest
Root Test (HRLRT1) is named respectively.

B Online Estimation
In this approach, noise variance is estimated implicitly from S auxiliary re-
4.2 Semi-blind SS techniques under noise uncertainty 85

Fig. 4.5 Hybrid approach-2: Noise variance estimated in S auxiliary slots from received
signal samples declared only noise samples by ED/RLRT

ceived signal slots which are declared only noise samples by the same algo-
rithm. As can be seen from the block diagram shown in Figure 4.5, noise
variance is calculated from signal samples implicitly declared as noise only
samples by ED/RLRT taking M samples at a time. Accurate estimate of the
noise variance is obtained by averaging S variances using a Moving Average
taking S slots into consideration. The average noise variance estimate is then
used in detecting the received primary signals using ED/RLRT detectors in
the next detection interval. Thus, after the encounter of the noise only slots,
the system updates the noise variance estimate and the process is repeated.
For ED and RLRT algorithms using this approach of noise estimation, Hybrid
Energy Detection-2 (HED2) and Hybrid Roy’s Largest Root Test-2 (HRLRT2)
are named respectively.

4.2.7 Energy Detection

Objective of every spectrum sensing scheme is to find out the detection statistic
which can be used in the decision making by properly comparing against a threshold
 86 Cognitive Radio: Spectrum Sensing

value. ED is a spectrum sensing techniques that evaluates the signal energy over
a certain time interval and compare it with the threshold to decide whether the
spectrum is in use or not. The presence of noise in the signal may affect the decision
of energy detector thus causing false alarm or even miss detection.

Formulation of the Decision Statistic

Using the information of the received signal matrix Y to develop a test statistic TED ,
which is the measure of the average energy of the received signal over a sensing
interval N, the detector compares TED against a predefined threshold t. If TED < t
then it decides in favour of Null Hypothesis H0 otherwise in favor of Alternate
Hypothesis H1 . The average energy of the received signal vector Y normalized by
the noise variance σv2 can be represented as,

K N
1
TED = ∑ ∑ |yk (n)|2.
KNσv2 k=1
(4.10)
n=1

Case 1: Null Hypothesis

For Null Hypothesis, rearranging (4.10) using yk (n) = vk (n),

K N
1
TED |H0 = ∑ ∑ |vk (n)|2
KNσv2 k=1
(4.11)
n=1
 2
1 K N  vRk (n) vCk (n) 
= ∑ ∑  √σv + j √σv  (4.12)
2KN k=1 n=1 2 2
K N  2
1  R

= β + jβ C  (4.13)

∑ ∑
2KN k=1 n=1
1 K N 2
= ∑ ∑ βR + βC2,
2KN k=1
(4.14)
n=1

where vRk (n) and vCk (n) are real and imaginary part of the noise signal vk (n)
√ √
respectively. βR = 2vRk (n)/σv and βC = 2vCk (n)/σv . As vk (n) is a zero
mean and σv2 variance complex valued Gaussian Random Variable, βR and βC
are Standard Normal Random Variables with mean zero and unity variance.
The numerator of TED in (4.14) is sum of square of 2KN Standard Normal
Random Variable with mean zero and variance 1, thus, the decision statistic
TED |H0 follows the Chi Square Distribution with 2KN degrees of freedom
 4.2 Semi-blind SS techniques under noise uncertainty 87

scaled by the factor (1/2KN). Thus, (4.14) can be written as,

1 2
TED |H0 = χ . (4.15)
2KN 2KN

Case 2: Alternate Hypothesis

For Alternate Hypothesis, rearranging the (4.10) using yk (n) = hk sk (n)+vk (n),

K N
1
TED |H1 = ∑ ∑ |hk sk (n) + vk (n)|2
KNσv2 k=1
(4.16)
n=1
 2
K 2
σtk N
 
 hk sk (n) + vk (n) 
=∑ 2 ∑

σtk
 (4.17)
k=1 KNσ v n=1  √

2

K σt2k N
= ∑ 2KNσv2 ∑ |α|2, (4.18)
k=1 n=1

where α = hk skσ(n)+v /
√ k (n) . As hk is assumed to be constant for the sensing interval
2
tk
and both the signal and noise are Complex Valued Gaussian Signals with
variances σv2 and σs2 respectively and both are independent signals, hk sk (n) +
vk (n) is also Complex Valued Gaussian Signal with mean zero and variance
σt2k . It is clear that α is also a Complex Valued Standard Normal Random
Variable with mean zero and unity variance. So the sum ∑N 2
n=1 |α| in (4.18)
follows the Chi Square Distribution with 2N degrees of freedom. And we
have:
K 
|hk |2 σs2 + σv2

2
TED |H1 = ∑ 2
χ2N (4.19)
k=1 2KNσ v
K  K
|htk |2 σs2

2 1 2
=∑ 2
χ2N + ∑ χ2N (4.20)
k=1 2KNσv k=1 2KN
2
Kρ χ2N χ2
TED |H1 = + 2KN . (4.21)
2KN 2KN

Normal Approximation of ED Decision Statistic

According to the Central Limit Theorem, when N is sufficiently large, the Chi
Squared Distributed Random Variable in (4.21) converges to a Gaussian Distribution.
88 Cognitive Radio: Spectrum Sensing

Chi-square pdf
0.25
N=3
PDF of Chi Squared Distributed Random Variable

N=5
N = 10
0.2 N = 15
N = 20
Normal(20,40)

0.15

0.1

5 · 10−2

0
0 5 10 15 20 25 30 35 40
y

Fig. 4.6 χN2 distribution with 5, 10, and 20 degrees of freedom. A normal distribution is
superimposed to illustrate a good approximation to χN2 by NR (20, 40) for N large

Figure 4.6 shows the pdf plot of the Chi Squared Distribution for 3, 5, 10, and 20
degrees of freedom superimposing NC (10, 20) to illustrate a good approximation of
χN2 by a Gaussian NC (N/2, 2N) ∼= NR (N, 2N) for large N. It is clear fron Figure 4.6
that Chi Square Distribution Variable with Degrees of Freedom 20 (N = 20) shows
a similarity with the Normal Distribution with mean and variance 20 and 40 (N and
2N).
To get an insight of accuracy of the approximation, simulation of Gaussian
approximation of Chi Square Distribution for different values of N is carried out.
Figure 4.7 shows the pdf of Chi Square Random Variable and its Gaussian Approxi-
mation counterpart for N taking values 5, 10, 20 and 40. Also Figure 4.8 shows the
plot of the Mean Square Error (MSE) of the approximation considering Probability
Density Functions of distribution functions as a comparison criteria for varying N.
The approximation shows perfect matching for N greater than 40 and it is clear
that the MSE is nearly zero, i.e. the Gaussian approximation is very accurate for
N > 40, which indicated that a degree of freedom greater than 40 is sufficient for
4.2 Semi-blind SS techniques under noise uncertainty 89

For N = 5 For N = 10
0.2 0.1
Chi−Square pdf Chi−Square pdf
Density Function

Density Function
Gaussian pdf 0.08 Gaussian pdf
0.15
0.06
0.1
0.04
0.05
0.02

0 0
0 20 40 60 80 0 20 40 60 80
x x
For N = 20 For N = 40
0.08 0.06
Chi−Square pdf Chi−Square pdf
Density Function

Density Function
0.06 Gaussian pdf Gaussian pdf
0.04
0.04
0.02
0.02

0 0
0 20 40 60 80 0 20 40 60 80
x x

Fig. 4.7 Comparison of Chi Square Distribution with Gaussian approximation for different N

approximating Chi Squared Distribution of the detection statistic to a Gaussian

distribution.
For good approximation, different models have been developed such as Edell’s
Model [58], Torrieri’s Model [161] and Berkeley Model [30] which have analyzed
the accuracy of different models in approximating the exact solution of the TED and
concluded that these models almost have the same performance for such scenario.
Thus, for the result in (4.21) and (4.15), using Berkeley Model [30] the Chi Squared
Distribution Function can be approximated to a Normal Distribution Function as,

 H0 , .
1
(

NR 1, KN
TED = 2 +2ρ+1 (4.22)
NR (ρ + 1), Kρ KN H1
90 Cognitive Radio: Spectrum Sensing

·10−4
6

4
Mean square error

0
0 10 20 30 40 50 60 70 80 90 100
Degree of Freedom

Fig. 4.8 MSE of the approximation considering Probability Density Function approximation

Detection and False Alarm Probabilities

A numerical study shows that ED ROC curves based on chi-squared distributions may
be accurately represented by binomial ROC curves. This allows the detector accuracy
and the ROC shape to be expressed simply in terms of distribution parameters [86].
For any description model, it would be useful to have a concise description of the
ROC curves that is meaningful in terms of underlying signal and noise distributions.
Such a description would facilitate the comparison between model and experiments,
and help in studying the effect of changing model parameter.
Hypothesis test is a procedure which divides the space of observations into
2 regions, Rejection Region (R) and Acceptance Region (A). The two important
characteristics of a test are called significance and power, referring to errors of type I
and II in hypothesis testing which relates to Probability of false alarm and probability
of detection respectively. The probabilities of false alarm PFa and probability of
4.2 Semi-blind SS techniques under noise uncertainty 91

detection PD for a given decision statistic referring to ED test is given by:

PFa = Prob{TED > t|H0 } (4.23)

PD = Prob{TED > t|H1 }. (4.24)

Based on the statistics of TED shown in (4.22), PFa can be evaluated as,
Z ∞
PFa = TED |H0 dt (4.25)
t
t−µ
  
1
= 1 − φ (t) ≡ 1 − 1 + er f √ (4.26)
2 2σ 2
t−µ
  
1
= 1 − er f √ (4.27)
2 2σ 2
t−µ
 
1
= er f c √ (4.28)
2 2σ 2
t−µ
 
PFa = Q √ , (4.29)
2σ 2

where φ (t) is the cdf of Normal Distribution, er f () is the error function, er f c()
is the complementary error function and Q() is the complementary cdf of Normal
Random Variable. Now putting the value of mean and variance for H0 from (4.22),
We have h √ i
PFa = Q (t − 1) KN . (4.30)

Following the same line of reasoning, the expression of PD is given by,

" √ #
(t − 1 − ρ) KN
PD = Q p . (4.31)
Kρ 2 + 2ρ + 1

Simulation Result

This Sub-Section discusses the simulation of ED in single sensor and multi-sensor

environment. Signal, channel and noise environment are set to matche the scenario
explained in Sub-Section 4.2.3.
Figure 4.9 illustrates the ROC plot of ED with its detection statistic following Nor-
mal distribution. Analytical result of ROC for ED computed using (4.30) and (4.31)
92 Cognitive Radio: Spectrum Sensing

1
ED Theory (K = 1)
ED Numerical (K = 1)
ED Theory (K = 2)
0.8 ED Numerical (K = 2)
ED Theory (K = 5)
Detection probability PD

ED Numerical (K = 5)
0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
False alarm Probability PF a

Fig. 4.9 ROC plot of ED for Normal Approximation of the Decision Statistics for N = 100,
and SNR = −10dB for varying number of sensors K

for SNR = −10 dB, N = 100, and number of sensors K = 1 (Single Sensor), 5 and
10 are compared with the numerical simulation for the same parameters. The plots
overlap with each other and model proves to be very accurate. With the increase in
number of sensors from 1 to 10, the performance of ED is greatly enhanced. The
graph converges to PD = 1 with steep slope for K = 10 if compared to the case
of K = 1 which shows that co-operative(“soft combining”) spectrum sensing can
greatly enhance the performance.
SNR performance of ED for a given PFa is illustrated in Figure 4.10. The
accuracy of the the closed-form expression derived in (4.30) and (4.31) is confirmed
by the results presented in Figure 4.10, where the theoretical formula is compared
against simulated detection performance for different SNR values. The considered
parameters are (K = 5, 10, 15, N = 50), at a false alarm rate of 5%. As shown in the
figure, the theoretical expressions are quite accurate even for a small number K of
sensors.
Figure 4.11 shows the variation of Probability of Detection vs. SNR performance
for different uncertainty in noise variance described as as the ratio between estimated
4.2 Semi-blind SS techniques under noise uncertainty 93

0.8
Numerical (K = 5)
0.7 Theoretical (K = 5)
Numerical (K = 10)
Theoretical (K = 10)
0.6 Numerical (K = 15)
Probability of Miss PM

Theoretical (K = 15)
0.5

0.4

0.3

0.2

0.1

0
−13 −12 −11 −10 −9 −8 −7 −6 −5 −4
SNR in dB

Fig. 4.10 Probability of Miss vs. SNR for varying number of sensors. Parameters: N =
50, PFa = 0.05 and K = 5, 10 and 15

noise variance and true variance level. Even though the increase in number of sensors
from 1 to 5 increases the performance of ED significantly, there is still a big gap
between known and unknown noise variance curves. This result makes it clear that,
the knowledge of exact noise variance is the crucial part for ED.

4.2.8 Hybrid ED (HED1)

From the results of previous section, it is evident that the knowledge of the noise
power (noise variance for zero mean Gaussian Noise) is imperative for achieving the
optimum ED performance. Unfortunately, the variation and the unpredictability of
noise power is unavoidable. Thus, the knowledge of the noise power is one of the
critical limitations of ED for its operation in low SNR. The only option is to estimate
the noise power. This chapter deals with the study of detection performance of the
94 Cognitive Radio: Spectrum Sensing

0.8
Probability of Detection PD

0.6

0.4 x = 0.1 dB, K = 1

x = 0.01 dB, K = 1
Known Variance, K = 1
x = 0.1 dB, K = 5
0.2 x = 0.01 dB, K = 5
Known Variance, K = 5

−10 −8 −6 −4 −2 0 2 4
SNR in dB

Fig. 4.11 Probability of Detection vs. SNR for varying amount of noise uncertainty measured
2
as β = σ̂σv2 , and x = 10 log10 (β ), Parameters: N = 40, PFa = 0.1 and K = 1, 5 and S
v

ED algorithm using estimated noise variance. Noise variance is estimated from S

auxiliary noise-only slots where we are sure that the primary signal is absent. For
ED with noise power estimation from noise only slots, which we call HED1 model,
the analytical expression of performance parameters PD and PFa in a multi-sensor2
environment is derived and simulated.

Formulation of the Decision Statistic

In this chapter, a concept of noise variance estimation from noise only samples
is proposed so as to fulfill the gap of unavailability of the noise variance. In this
technique called HED1, noise variance is estimated on S auxiliary noise only slots,
under the hypothesis that noise variance is constant over the adjacent slots.
Consider a sampling window of length M before and adjacent to the detection
window which is containing noise only samples for sure. Then the estimated noise
variance from these noise-only samples using a Maximum Likelihood noise power
2 For single sensor performance evaluation, set the number of sensors K to 1.
4.2 Semi-blind SS techniques under noise uncertainty 95

estimate can be written as,

1 K M
σ̂v2 = ∑ ∑ |vk (m)|2. (4.32)
KM k=1 m=1

If the noise variance is constant, the estimation can be averaged over S successive
noise-only slots. And (4.32) becomes:

1 S K M
σ̂v2 (S) = ∑ ∑ ∑ |vk (m)|2 . (4.33)
KSM s=1 k=1 m=1

Now, the ED Test Statistic in (4.10) can be modified to HED1 test statistic
using (4.33) as,
K N
1
THED1 = 2 ∑ ∑ |yk (n)|2 . (4.34)
KN σ̂v (S) k=1 n=1
Here, THED1 denotes the detection statistics for the HED1 scheme and the statistical
distribution of THED1 depends upon the particular estimation technique considered.
Moreover, (4.34) can be considered as the parametric likelihood ratio test when
the signal to be detected is assumed to be Gaussian with zero mean and variance σs2 .

Proposition 1. The false alarm probability PFa of HED1 for number of sensors K,
number of samples N, number of auxiliary slots S and threshold t is given by,
 
t−1
PFa = Q  q . (4.35)
MS+Nt2
KMNS

Similarly, the Probability of detection PD of HED1 for number of sensors K,

number of samples N, number of auxiliary slots S and threshold t is given by,
 
(t − 1 − ρ)
PD = Q  q . (4.36)
t2 Kρ 2 +2ρ+1
KMS + KN

The analytical proof of Proposition 1 is given below.

 96 Cognitive Radio: Spectrum Sensing

Case 1: Null Hypothesis H0

For Null Hypothesis, rearranging (4.34) using yk (n) = vk (n), we have:

K N
1
THED1 |H0 = 2 ∑ ∑ |vk (n)|2 (4.37)
KN σ̂v k=1 n=1
  2 
K
 R
N  vk (n) vCk (n) 
∑k=1 ∑n=1  √σv + j √σv  
2KMS  2 2
= (4.38)
 
2
2KN 
  
S M  vk (m)
√
 
∑s=1 ∑m=1  σ / 2 
v

∑K N
k=1 ∑n=1 |β1 |
2
 
2KMS
= , (4.39)
2KN ∑Ss=1 ∑K ∑N
k=1 n=1 |β 2 |2

where vRk (n) and vCk (n) are real and imaginary part of the noise signal vk (n)
√ √
respectively, β1 = 2vk (n)/σv and β2 = 2vk (m)/σv .
As vk (n) is a zero mean and σv2 variance complex valued Gaussian Random
variable, β1 and β2 are Standard Normal Random Variables with zero mean
and unity variance. Both numerator and denominator of THED1 in (4.39) are
the sum of square 2KMS and 2KN Standard Normal Random variables with
zero mean and variance 1. Thus, the numerator and the denominator of the
decision statistic THED1 |H0 follows the Chi Square Distribution with 2KMS
and 2KN degrees of freedom respectively, scaled by the factor 2KMS
2KN .

Finally, (4.39) can be modified as,

2
2KMS χ2KN
THED1 |H0 = 2
(4.40)
2KN χ2KMS

Case 2: Alternate Hypothesis H1

4.2 Semi-blind SS techniques under noise uncertainty 97

For Alternate Hypothesis, rearranging (4.34) using yk (n) = hk sk (n) + vk (n),

we have:
K N
1
THED1 |H1 = 2 ∑ ∑ |hk sk (n) + vk (n)|2 (4.41)
KN σ̂v k=1 n=1
   2
K KMSσt2 N  hk sk (n)+v
 k (n) 
∑k=1 KNσ 2 ∑n=1  σ /√2 
k
v tk
=  2 (4.42)
S K M  vk (m) √

∑s=1 ∑k=1 ∑m=1  σ / 2 
v
 
KMSσt2
∑k=1 KNσ 2 ∑N
K k
n=1 |α|
2
v
= , (4.43)
∑Ss=1 ∑K M
k=1 ∑m=1 |β2 |
2

hk sk (n)+v √
where α = √ k (n) and β2 = 2vk (m)/σv .
σtk / 2
The channel coefficient hk is assumed to be constant for the sensing interval
and both signal and noise are independent Complex Valued Gaussian variable
with variances σv2 and σs2 respectively. Then, hk sk (n) + vk (n) is a Complex
Valued Gaussian variable with zero mean and variance σt2k . It is clear that α
and β2 are also Complex Valued Standard Normal Random Variables with
zero mean and unity variance. So the sum ∑N 2 S K M
n=1 |α| and ∑s=1 ∑k=1 ∑m=1 |β |
2

in an expression of (4.43) follows the Chi Square Distribution with 2N and

2SKM degrees of freedom respectively. Thus, (4.43) can be re-written as,
 2 2 2
|hk | σs +σv
∑K
k=1 2KNσv2
2
χ2N
THED1 |H1 = 2KMS 2
(4.44)
χ2KMS
 
K |htk |2 σs2 2 + 2 1 2
∑k=1 2KNσ 2 χ2N ∑k=1 2KN χ2N
v
= 2KMS 2
(4.45)
χ2KMS
 2 + χ2 
KMS Kρ χ2N 2KN
THED1 |H1 = 2
. (4.46)
KN χ2KMS

Normal Approximation of HED1 Decision Statistic

For large values of N and M, the Chi Squared Random Variables in (4.40) and (4.46)
can be replaced by their normal approximates with χN2 ∼ NR (N, 2N). Thus, decision
98 Cognitive Radio: Spectrum Sensing

statistics TED |H0 and TED |H1 in (4.40) and (4.46) can be re-written as,
 
KMS NR (2KN, 4KN)
THED1 |H0 = (4.47)
KN NR (2KMS, 4KMS)
Kρ 2 +2ρ+1
 
NR ρ + 1, KN
THED1 |H1 = 1

. (4.48)
NR 1, KMS

Now, the ratio of two normal random variables in (4.47) and (4.48) can be trans-
formed to a single normal random variables as shown below under the assumption
that the denominator is always positive [121].
  
 NR 1, 1 + t2 H0 ,
THED1 =  KN KMS 2 2
 . (4.49)
 NR (ρ + 1), t + Kρ +2ρ+1 H 1
KMS KN

Detection and False Alarm Probabilities

A numerical study shows that ED ROC curves based on chi-squared distributions may
be accurately represented by binomial ROC curves. This allows the detector accuracy
and the ROC shape to be expressed simply in terms of distribution parameters [86].
For any description model, it would be useful to have a concise description of the
ROC curves that is meaningful in terms of underlying signal and noise distributions.
Such a description would facilitate the comparison between model and experiments,
and help in studying the assessment of effect of changing model parameter.
Hypothesis test is a procedure which divides the space of observations into
2 regions, Rejection Region (R) and Acceptance Region (A). The two important
characteristics of a test are called significance and power, referring to errors of type I
and II in hypothesis testing which relates to probability of false alarm and probability
of detection respectively. The probabilities of false alarm PFa and probability of
detection PD for a given decision statistic referring to HED1 test is given by:

PFa = Prob{THED1 > t|H0 } (4.50)

PD = Prob{THED1 > t|H1 }. (4.51)
4.2 Semi-blind SS techniques under noise uncertainty 99

Based on the statistics of THED1 shown in (4.49), PFa can be evaluated as,
Z ∞
PFa = TED |H0 dt (4.52)
t
t−µ
  
1
= 1 − φ (t) ≡ 1 − 1 + er f √ (4.53)
2 2σ 2
t−µ
  
1
= 1 − er f √ (4.54)
2 2σ 2
t−µ
 
1
= er f c √ (4.55)
2 2σ 2
t−µ
 
PFa = Q √ , (4.56)
2σ 2

where φ (t) is the cdf of Normal Distribution, er f (.) is the error function er f c() is
the complementary error function and Q(.) is the complementary cdf of Normal
Random Variable.
Now putting the value of mean and variance for H0 from (4.49), we have:
 
t−1
PFa = Q  q . (4.57)
MS+Nt2
KMNS

Similarly, for the detection probability:

 
(t − 1 − ρ)
PD = Q  q . (4.58)
t2 Kρ 2 +2ρ+1
KMS + KN

Simulation Result

This Sub-Section shows the comparison between simulated ROC curves and analyti-
cal curves based on (4.35) and (4.36). The accuracy of the closed-form expression
derived in Proposition 1 is confirmed by the results presented in Figure 4.12, where
the theoretical formulas (4.35) & (4.36) are compared against simulated perfor-
mance over S auxiliary noise only slots (S ranges from 1 to 10). It can be realized
from Figure 4.12 that the analytical and the numerical curves are perfectly matched
100 Cognitive Radio: Spectrum Sensing

0.8

0.6
PD

0.4 ED Known Variance

HED1 Theoretical (S = 2)
HED1 Numerical (S = 2)
HED1 Theoretical (S = 5)
0.2
HED1 Numerical (S = 5)
HED1 Theoretical (S = 10)
HED1 Numerical (S = 10)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
PF a

Fig. 4.12 ROC of Multi Sensor HED1, Parameters: N = 100, M = 100, K = 5, S = 1, 2, 5, 10,
and SNR = −10 dB

0.8
HED1 Numerical (S = 2)
0.7 HED1 Theoretical (S = 2)
HED1 Numerical (S = 8)
HED1 Theoretical (S = 8)
0.6 ED (Known Variance)
Probabiliry of Miss PM

0.5

0.4

0.3

0.2

0.1

0
−12 −11 −10 −9 −8 −7 −6 −5 −4
SNR in dB

Fig. 4.13 Probability of Miss vs SNR, Parameters: N = 50, M = 50, K = 5, S = 2, 8, and

SNR = −10dB
4.2 Semi-blind SS techniques under noise uncertainty 101

and this validates the analytical expressions. Also, it can be noted that the increase
in number of slots for noise variance estimation correspondingly increases the per-
formance of HED1 and approaches closer to the optimal one (ED with known noise
variance). Under the considered scenario with M = N = 100 which is the number
of noise samples in each slot, just S = 10 ( i.e., 1000 samples) provides very near
convergence to ideal performance.
Fixed the probability of false alarm PFa = 0.05, the simulated and analytical
performance of HED1 in terms of probability of miss-detection is plotted against
different values of SNR in Figure 4.13. The considered parameters are (K = 5,
N = 50, M = 50). As shown in the figure, the theoretical expressions is quite accurate
even for a small number K of sensors. The figure clearly shows the performance
gap between ED (known variance) and HED1 with estimated noise variance from S
auxiliary noise only slots.

4.2.9 Hybrid ED-2(HED2)

As discussed, optimum performance of ED can be achieved even in the absence of

noise power knowledge in prior by estimating it from sufficient number of noise
samples which is also clear from the result of Sub-Section 4.2.8. But in reality, it is
difficult to guarantee the availability of signal free samples so as to estimate the noise
variance. Some literature analyzed the performance of ED using estimated noise
variance setting aside a separate frequency channel for the measurement of the noise
power [178]. Even though setting a separate frequency channel for the measurement
of the noise power is good enough for accurate noise estimation, it’s also a waste of
frequency resource. Wireless channels may show variation in the amount of noise on
different frequency slots. Since SU needs to scan and make decision on a large band
of frequencies about the presence or absence of PU, it is not always safe to assume
uniformly distributed noise in all the frequency band of concern. Similarly, Mariani
et al. [120] put forwarded the two-step sensing scenario where basic ED is used for
small sensing period whose decision promotes the periodic estimation of the noise
variance from the noise only samples decided by ED, and some other sophisticated
techniques can be used for fine sensing using long observation time.
HED2 is somewhat similar to the method of Mariani et al. [120]: HED2 does not
resort on the existence of auxiliary noise only slots, but estimates the noise variance
102 Cognitive Radio: Spectrum Sensing

information from the previous slots declared as H0 by the algorithm. Now, the noise
variance estimated from S previous slots which are declared H0 is used in current
detection interval to get the decision about the presence or absence of the primary
signal. The derivation of the analytical expression of performance parameters PD
and PFa in a multi-sensor environment for HED2 is now presented using different
probability distribution models.

Formulation of Decision Statistic

The information on the received signal matrix Y is used to develop a test statistic
THED2 , which measures the average energy of the received signal over a sensing in-
terval N and compares it against an estimated noise variance. The detector compares
THED2 against a predefined threshold t. If THED2 < t then it decides in favor of Null
Hypothesis H0 otherwise in favor of Alternate Hypothesis H1 . The average energy
of the received signal vector Y normalized by the estimated noise variance σ̂v2 can
be represented as,
K N
1
THED2 = ∑ ∑ |yk (n)|2.
KN σ̂v2 k=1
(4.59)
n=1

Proposition 2. Let, PS be the probability of receiving primary signal plus noise,

PDED the probability of detection of first stage ED, and S the number of slots. The
Maximum Likelihood noise variance estimate σ̂v2 using M received signal samples
declared noise samples by ED from K receivers each is given by,
" #
Ss K M SN K M
∑ ∑ ∑ |hk s(m) + v(m)|2 + ∑ ∑ ∑ |vk (m)|2
s=1 k=1 m=1 s=1 k=1 m=1
σ̂v2 (S) = , (4.60)
KMS

where, SS = SPS 1 − PDED is the number of primary signal slots missed by ED and

SN = S − SS is the number of noise samples successfully detected.

Proof. Consider a window of length S containing slots declared H0 by the detector.

In case of error, the estimated noise variance from the noise only samples using a
Maximum Likelihood Noise Power Estimate can be written as,
" #
K M
1
σ̂v2 = ∑ ∑ |hk s(m) + v(m)|2 . (4.61)
KM k=1 m=1
4.2 Semi-blind SS techniques under noise uncertainty 103

If a slot is slot are declared noise only by ED, there is a chance with probability
(1 − PDED ) that ED decides in favor of H0 even if it was H1 . If the estimated variance
is the mean of S successive noise variances calculated from the noise only slots
decided by ED, then out of S noise only slots decided by ED, PS (1 − PDED )S noise
variance estimates are calculated from H1 slots. Thus, the average noise variance can
now be written as the mean of the noise variance estimated from H0 slots declared
H0 by ED and the H1 slots declared H0 by ED. Finally, (4.61) can be modified and
rewritten as,
" #
Ss K M SN K M
2
∑ ∑ ∑ |hk s(m) + v(m)| +∑ ∑ ∑ |vk (m)|2
s=1 k=1 m=1 s=1 k=1 m=1
σ̂v2 (S) = . (4.62)
KMS

Proposition 3. The false alarm probability PFa of HED2 for number of sensors K,
number of samples N, number of auxiliary slots S for noise estimation using (4.62)
and threshold t is given by,
 
S
 t− 
 S + ρSS 
PFa = Q  s
 . (4.63)
2 2 2

 t N(SS Kρ + ρSS + S) + MS 
KMN(S + ρSS )2

Similarly, the Probability of Detection PD of HED2 for number of sensors K, number

of samples N, number of auxiliary slots S for noise estimation using (4.62) and
threshold t is given by,
 
S(ρ + 1)
 t− 
 S + ρS S

PD = Q  s
 2
.
 (4.64)
2 2 2
 t N(SS Kρ + ρSS + S) + MS (Kρ + 2ρ + 1) 
KMN(S + ρSS )2

The analytical proof of Proposition 3 is given below.

 104 Cognitive Radio: Spectrum Sensing

Case 1: Null Hypothesis

For Null Hypothesis, rearranging (4.59) using yk (n) = vk (n), we have:

K N
1
THED2 |H0 = 2 ∑ ∑ |vk (n)|2 (4.65)
KN σ̂v k=1 n=1
!
1 K N |vk (n)|2
 
1
= ∑ ∑ σv2 (4.66)
σ̂v2 /σv2 KN k=1 n=1
B
= , (4.67)
A
where,
A = σ̂v2 /σv2 and,
1 K N |vk (n)|2
B= ∑ ∑ σv2 .
KN k=1 n=1
A can be simplified using the expression of σ̂v2 in (4.62) as,

σ̂v2
A= 2 (4.68)
σv
Ss K M SN K M
∑∑ ∑ |hk s(m) + v(m)|2 + ∑ ∑ ∑ |vk (m)|2
s=1 k=1 m=1 s=1 k=1 m=1
= . (4.69)
MKSσv2

" ! #
1 SS K σt2k 2 2
A= ∑ ∑ 2 χ2M + χ2KMSN
, (4.70)
2MKS s=1 k=1 σv

where, σt2k is the total variance of hk sk (n) + vk (n).

We have:
" ! #
SS K
1 |hk |2 σs2 + σv2 2 2
A= ∑∑ 2
χ2M + χ2KMS N
(4.71)
2MKS s=1 k=1 σ v
" ! #
SS
1 2 2

2
= ∑ Kρ χ2M + χ2KM + χ2KMS N
(4.72)
2MKS s=1
1  2 2 2

= ρ χ2MSS + χ2MKS S
+ χ2MKS N
(4.73)
2MKS
1  2 2 2

= ρ χ2MSS + χ2(K−1)MSS + χ2MS S
+ χ2MS N
(4.74)
2MKS
1  2 2 2

= (Kρ + 1)χ2MS S
+ χ2KMS − χ2MS S
. (4.75)
2MKS
 4.2 Semi-blind SS techniques under noise uncertainty 105

After approximating each Chi Squared Random Variables to Normal Random

Variable using χN2 ∼ NR (N, 2N) transformation for 2MSS ≥ 50, (4.75) above
can be approximated to Normal Random Variable as,
2 2
 
NR KρSS + SS + KS − SS , SS K ρ +KρSMS +SS +KS−SS
A= (4.76)
KS
S + ρSS SS Kρ 2 + ρSS + S
 
= NR , . (4.77)
S KMS2

Similarly B can be simplified as,

1 K N |vk (n)|2
B= ∑ ∑ σv2 (4.78)
KN k=1 n=1
1 K N 2
= ∑ ∑ β1 , (4.79)
2KN k=1 n=1

where β1 = σvk/(n
√ is a Standard Normal Random Variable with zero mean and
2
v
unitary variance.
Now, the sum in (4.79) is the sum of square of KN Complex Standard Normal
Random variables, thus the sum follows a Chi Square Distribution of 2KN
degrees of freedom, i.e.
1 2
B= χ . (4.80)
2KN 2KN
Again, using an Normal Approximation to a Chi Square Random variable, (4.80)
can be written in the form
 
1
B = NR 1, . (4.81)
KN

Finally, using (4.77) and (4.81) in (4.66), the expression of THED2 |H0 can be
written as the ratio of two Normal Random variables as,
1


NR 1, KN
THED2 |H0 =  . (4.82)
S+ρSS SS Kρ 2 +ρSS +S
NR S , KMS2

Case 2: Alternative Hypothesis

106 Cognitive Radio: Spectrum Sensing

For Alternate Hypothesis, rearranging (4.59) using yk (n) = hk s(n) + vk (n), we

have:
K N
1
THED2 |H1 = 2 ∑ ∑ |hk s(n) + vk (n)|2 (4.83)
KN σ̂v k=1 n=1
!
1 K N |hk s(n) + vk (n)|2
 
1
= ∑∑ (4.84)
σ̂v2 /σv2 KN k=1 n=1 σv2
C
= , (4.85)
A
where,
1 K N |hk s(n) + vk (n)|2
C= ∑∑ and,
KN k=1 n=1 σv2
A = σ̂v2 /σv2 .
C can be simplified as,

1 K N |hk sk (n) + vk (n)|2

C= ∑∑ (4.86)
KN k=1 n=1 σv2
!  2
K σt2k N 
 hk sk (n) + vk (n) 

=∑ 2 ∑  σ / 2  √ (4.87)
k=1 2KNσv n=1 tk
!
K σt2k N
=∑ 2 ∑ |α|2 , (4.88)
k=1 2KNσ v n=1

hk sk (n)+v
√ k (n) .
where α = σtk / 2
As hk is assumed to be constant for the sensing interval and both the signal and
noise are independent Complex Valued Gaussian samples with variances σv2
and σs2 respectively, hk sk (n) + vk (n) is also Complex Valued Gaussian samples
with zero mean and variance σt2k . It is clear that α is a Complex Standard
Normal Random Variable with zero mean and unitary variance. So the sum
∑N 2
n=1 |α| in (4.88) follows the Chi Square Distribution with 2N and 2SKM
4.2 Semi-blind SS techniques under noise uncertainty 107

degrees of freedom respectively. Thus, (4.88) can be re-written as,

K
|hk |2 σs2 + σv2
 
2
C= ∑ 2
χ2N (4.89)
k=1 2KNσ v
K  2 2 2
|htk | σs

2 1 2
=∑ 2
χ 2N + ∑ χ2N (4.90)
k=1 2KNσv k=1 2KN
1  2 2

= Kρ χ2N + χ2KN . (4.91)
2KN
Simplifying (4.91) to a Normal Distributed Random Variable, we obtain

Kρ 2 + 2ρ + 1
 
C = NR ρ + 1, . (4.92)
KN

Using expression of A from (4.77) and C from (4.92) in (4.85), THED2 |H1 can
be written as the ratio of two Normal Random Variables as,

Kρ 2 + 2ρ + 1
 
NR ρ + 1,
KN
THED2 |H1 = (4.93)
S + ρSS SS Kρ 2 + ρSS + S
 .
NR ,
S KMS2

Noting the result,

  
1

 NR 1,
KN

 H0 ,



S + ρSS SS Kρ 2 + ρSS + S

 

 NR ,


S KMS2 
THED2 =  2 . (4.94)
Kρ + 2ρ + 1
N + 1,


 R ρ
 KN
H1



 
S + ρS S Kρ 2 + ρS + S 
S S S

,

 NR

S KMS2

This shows that, for both H0 and H1 , the decision statistic is the ratio of two
normal random variables which can be simplified to single Normal Random
108 Cognitive Radio: Spectrum Sensing

Variable as,
2 N(S Kρ 2 + ρS + S) + MS2 
 
S t S S
 NR , H0 ,


 S + ρSS 2 KMN(S + ρSS )2
THED2 = .
S(ρ + 1) t N(SS Kρ 2 + ρSS + S) + MS2 (Kρ 2 + 2ρ + 1)

, H1

 NR

S + ρSS KMN(S + ρSS )2
(4.95)

Formulation of Detection and False Alarm Probabilities

Based on the statistics of THED2 given by (4.95), probability of false alarm can be
written in terms of Q-function as,
 
S
 t− 
 S + ρS S

PFa = Q  s
 2
. (4.96)
2 2

 t N(SS Kρ + ρSS + S) + MS 
KMN(S + ρSS )2

Similarly, based on the statistics of THED2 given by (4.95), probability of detection

can be written in terms of Q-function as,
 
S(ρ + 1)
 t− 
 S + ρS S

PD = Q  2s .
 (4.97)
2 2 2
 t N(SS Kρ + ρSS + S) + MS (Kρ + 2ρ + 1) 
KMN(S + ρSS )2

Simulation Result

This Sub-Section shows the comparison between HED2 simulated curves and an-
alytical ROC curves based on (4.63) and (4.64). Figure 4.14 shows perfect match
between theoretical and numerical curves validating the considered model. Fig-
ure 4.15 illustrates the comparison of ED, HED1 and HED2 performance in varying
SNR. Performance of HED1 and HED2 varies typically around 0 dB SNR but no
big difference can be noted in extreme higher or lower SNR values. Since, there is
a chance of mis-interpretation of noise plus primary signal as only noise signal in
case of HED2, performance of HED2 is slightly lower than HED1 near 0 dB SNR.
With the increase in the number of slots used for the estimation of the noise variance,
4.2 Semi-blind SS techniques under noise uncertainty 109

1
HED2 Theory (S = 5)
HED2 Numerical (S = 5)
HED2 Theory (S = 50)
0.8 HED2 Simulated (S = 50)
Probability of Miss PM

0.6

0.4

0.2

0
−10 −8 −6 −4 −2 0 2 4 6 8 10
SNR in dB

Fig. 4.14 Comparison of simulated and analytical detection performance curves of HED2
method for N = 10, M = 10, K = 5, S = 5 and 50

the probability of miss of HED1 and HED2 is decreases and goes near to that of
ED-known variance.
110 Cognitive Radio: Spectrum Sensing

1
ED
HED2 (S = 2)
HED1 (S = 2)
0.8
HED2 (S = 5)
Mis-Detection Probability PM

HED1 (S = 5)
HED2 (S = 10)
0.6 HED1 (S = 10)

0.4

0.2

0
−10 −8 −6 −4 −2 0 2 4 6 8 10
SNR in dB

Fig. 4.15 Comparison of performance curves of HED2, ED and HED1 for N = 10, M = 10,
K = 5, S = 2, 5 and 10

4.2.10 RLRT

RLRT is a “semi-blind” eigenvalue based detection scheme originally proposed

in [148] and introduced in CR by [107]. It is a multi-sensor detection algorithm
which requires the prior knowledge of noise variance for optimum performance.

Formulation of Decision Statistic

Using the information on the received signal matrix Y and assuming a perfect
knowledge of noise variance σv2 and the channel parameter h, a test statistic T is
developed. The detector compares T against a predefined threshold t. If T < t it
decides in favor of Null Hypothesis H0 otherwise in favor of Alternate Hypothesis
H1 . In context of perfect knowledge of noise and channel parameters, the most
accurate and uniformly most powerful test of the above hypothesis is the Neyman-
Pearson Test (NP Test) [131]. Using NP test, the expression of the test statistic can
be written as
p (Y; h, σs2 , σv2 )
TNP = 1 . (4.98)
p0 (Y; σv2 )
4.2 Semi-blind SS techniques under noise uncertainty 111

Using (4.98) probability of false alarm PFa and probability of detection PD for a
given decision statistic are given by,

PFa = Prob{T > t|H0 } (4.99)

PD = Prob{T > t|H1 }, (4.100)

where t is the detection threshold. Since (4.98) in asymptotic regime N → ∞ con-

verges to depend on the largest Eigenvalue [128, 107] λ1 , resulting test is also known
as RLRT [148]. Thus, the test statistic of the RLRT test is given by,

λ1
TRLRT = . (4.101)
σv2

Formulation of Detection and False alarm Probabilities

1) False alarm probability: Concerning the Null hypothesis where Y = V, the

sample covariance matrix R follows a Wishart Distribution of degree N. From (4.99)
and (4.101), it is clear that, the false alarm rate depends upon the distribution of the
largest eigenvalue λ1 of the sample covariance matrix R. According to recent results
involving RMT, the detection statistic TRLRT under Null Hypothesis for sufficiently
large N and K follows a Tracy Widom Distribution of order 2 [98]. Thus,

TRLRT |H0 − µ
 
Prob < t → FTW 2 (t), (4.102)
ξ

where FTW 2 (t) is the Cumulative Distribution Function of the Tracy Widom Dis-
tribution of order 2 with suitably chosen centering and scaling parameters given
by:
"  1 #2
K 2
µ= +1 (4.103)
N
"  1 # "  1 #1/3
−2
K 2 K
ξ = N −2/3 +1 +1 . (4.104)
N N
112 Cognitive Radio: Spectrum Sensing

For (4.99), the probability of false alarm can be written as

PFa = Prob (PRLRT > t|H0 ) (4.105)

TRLRT |H0 − µ t−µ
 
= Prob > (4.106)
ξ ξ
t−µ
 
PFa = 1 − FTW 2 . (4.107)
ξ

2) Detection probability: Under alternate hypothesis H1 , the asymptotic distribution

of λ1 in the joint limit N, K → ∞ is characterized by a phase transition phenomenon
for smaller SNR [21]. For single signal detection, a critical detection threshold was
1
identified [139] in terms of SNR as, ρCric = √KN . This suggests that when SNR is
lower than the critical value, the limiting distribution of the detection statistics TRLRT
is the same of the largest noise eigenvalue, thus nullifying the statistical power of
a largest eigenvalue test. For ρ > ρCric the distribution of TRLRT was found to be
asymptotically Gaussian [107, 21] as shown below,

λ1
∼ NR (µ1 , σ12 ), (4.108)
σv2

where,

K −1
 
µ1 = (1 + Kρ) 1 + (4.109)
NKρ
K −1
 
2 1 2
σ1 = (Kρ + 1) 1 − . (4.110)
N NK 2 ρ 2

Thus, using the above normal approximation in (4.100), the probability of detec-
tion of RLRT can be written as

PD = Prob [TRLRT |H1 < t] , (4.111)

t − µ1
 
PD = Q . (4.112)
σ1
4.2 Semi-blind SS techniques under noise uncertainty 113

0.8
Probability of Detection (PD )

0.6

RLRT Numerical (K = 8)
0.4 RLRT Analytical (K = 8)
RLRT Numerical (K = 10)
RLRT Analytical (K = 10)
ED Numerical (K = 8)
0.2
ED Analytical (K =8)
ED Numerical (K = 10)
ED Analytical (K = 10)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of False Alarm (PF a )

Fig. 4.16 Comparison of simulated and analytical ROC performance curves of RLRT method
for N = 50, K = 8 and 10, SNR = −10 dB

0.9
Probability of Detection PD

0.8

0.7

0.6

0.5
RLRT (K = 5)
RLRT (K = 7)
0.4
ED (K = 5)
ED (K = 7)
0.3
−12 −11.5 −11 −10.5 −10 −9.5 −9 −8.5 −8 −7.5 −7 −6.5 −6
SNR in dB

Fig. 4.17 PD vs. SNR performance of RLRT. Parameters: N = 100, K = 5, 7, PFa = 0.05
114 Cognitive Radio: Spectrum Sensing

0.8
Probability of Detection (PD )

0.6

0.4

0.2 ED with known variance

ED with wrong variance
RLRT with known variance
RLRT with wrong variance
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of False Alarm (PF a )

Fig. 4.18 Effect of Noise Variance fluctuation on ED and RLRT. Parameters: N = 100, K =
4, var(σ̂v2 ) = 0.0032(−25dB) given nominal variance σv2 = 1

Simulation Result

The ROC performance of RLRT is shown in Figure 4.16 along with its comparison
with ED for the same simulation parameters. The accuracy of the closed-form
expression derived in (4.107) and (4.112) is confirmed by the results presented in
Figure 4.16, which shows a perfect match between analytical and ùàsimulated curves.
The considered parameters are (K = 8, N = 50) at −10dB SNR. As shown in the
figure, the theoretical expressions is quite accurate even for small number of sensors
K. The figure also shows the performance gap between ED and RLRT.
For given probability of false alarm PFa = 0.05, the performance of RLRT in
terms of probability of miss-detection is plotted against different values of SNR in
Figure 4.17. The considered parameters are (K = 5, 7 and N = 50). A comparison
between ED and RLRT can also be realized from Figure 4.17. It can be noted that
the detection performance of RLRT is far more better than that of ED for same
simulation parameters.
4.2 Semi-blind SS techniques under noise uncertainty 115

The effect of the noise variance fluctuation on ED and RLRT algorithms is

considered in Figure 4.18. Assuming Gaussian Distribution of the noise variance
estimate with mean equal to nominal value, ROC for ED and RLRT is being plotted,
setting var(σ̂v2 ) = 0.0032(−25dB) for N = 100 and K = 4. The result showed that,
for the same fluctuation on the noise variance estimate, the performance gap between
the ideal curve and the curve with wrong variance is larger for ED as compared to
RLRT. Thus, it could be figured out that RLRT is more robust to noise variance
fluctuation as compared to ED algorithm.

4.2.11 Hybrid RLRT (HRLRT1)

HRLRT1 addresses the problem of a priori knowledge of the noise power in RLRT.
HRLRT1 is a hybrid approach of RLRT similar to HED1 in which estimated noise
variance is used instead of nominal noise variance so as to normalize the received
signal in order to detect the primary signal.

Formulation of Decision Statistic

HRLRT1 explores the concept of noise variance estimation from noise only samples
so as to fulfill the gap of unavailability of the noise variance. Assuming the noise
variance is constant over the adjacent slots, independent noise estimation in auxiliary
“noise only” slots is performed.
Consider a sampling window of M samples before and adjacent to the detection
window which is containing only noise samples for sure. The estimated noise
variance from the noise only samples using a Maximum Likelihood noise power
estimation can be written as,

1 K M
σ̂v2 = ∑ ∑ |vk (m)|2. (4.113)
KM k=1 m=1

If the noise variance is constant over each slots, the estimation can be averaged over
S successive noise-only slots. Thus, above (4.113) can be modified by averaging
over S successive noise-only slots as,

1 S K M
σ̂v2 = ∑ ∑ ∑ |vk (m)|2 . (4.114)
KSM s=1 k=1 m=1
116 Cognitive Radio: Spectrum Sensing

Now, the HRLRT1 Test Statistic can be derived from (4.101) by replacing the exact
noise variance with the estimated noise variance using (4.114), i.e.,

λ1
THRLRT 1 = . (4.115)
σ̂v2

Equation (4.115) can be considered as the parametric likelihood ratio test when
the signal to be detected is assumed to be Gaussian with zero mean and variance σs2 .

Proposition 4. Consider the HRLRT1 test working with K sensors, N samples, SNR
value ρ and threshold t. Asymptotically in N and K, the false alarm and detection
probability of HRLRT1 are given by,

PFa = 1 − F0 (t) (4.116)

 
t−1 
PD = Q  q , (4.117)
2t2 2
D + σ1

where F0 (t), is the cdf of the pdf shown below:

r
tx − µ
   
1 D D
Z ∞
2
f0 (t) = |x|fTW 2 exp − (x − 1) dx. (4.118)
2ξ π −∞ ξ 4

The proof of the above Proposition 4 can be found in following Sub-Section.

Formulation of Detection and False Alarm Probability

For the ease of simplification, consider the test statistic of the HRLRT1 as shown
below,
λ1 |H0 1
THRLRT 1 |H0 = 2
× 2 2. (4.119)
σv σ̂v /σv
1) False alarm probability: Under Null hypothesis H0 where Y = V, the sample
covariance matrix R follows a Wishart Distribution of degree N. From (4.115)
and (4.119) it is clear that the false alarm rate depends upon the distribution of the
largest eigenvalue λ1 of the sample covariance matrix R.
4.2 Semi-blind SS techniques under noise uncertainty 117

λ |H
Now, for sufficiently large N and K, 1σ 2 0 follows a Tracy Widom Distribution of
v
order 2 [98] with suitably chosen centering and scaling parameters as shown below:
"  1 #2
K 2
µ= +1 (4.120)
N
"  1 # "  1 #1/3
−2
K 2 K
ξ = N −2/3 +1 +1 . (4.121)
N N

Let us denote the Probability Density Function of the Tracy Widom Distribution of
order 2 as,
z−µ z−µ
    
1
pdf FTW 2 = fTW 2 . (4.122)
ξ ξ ξ
Similarly, for sufficiently large value of D = 2KSM, σ̂v2 /σv2 follows a Normal
distribution NR 1, D2 with Probability Density Function,

r
D(x − 1)2
 
1 D
f1,2/D (x) = exp − . (4.123)
2 π 4

Thus, using above distribution, the test statistic THRLRT 1 |H0 in (4.119) can be re-
written as, λ | 
1 H0
−µ
σv2
FTW 2  ξ


THRLRT 1 |H0 = . (4.124)

NR 1, D2

Since the two random variables in (4.124) are statistically independent, the probabil-
ity density functions of the detection statistic THRLRT 1 |H0 can be written as,
r
tx − µ
   
1 D D
Z ∞
2
f0 (t) = |x|fTW 2 exp − (x − 1) dx. (4.125)
2ξ π −∞ ξ 4

Now, using the result in (4.125), the probability of false alarm can be written as,

PFa = 1 − F0 (t), (4.126)

where F0 (t), is the Cumulative Distribution Function of the probability distribution

function shown in (4.125).
118 Cognitive Radio: Spectrum Sensing

2) Detection probability: Under alternate hypothesis H1 , the asymptotic dis-

tribution of λ1 in the joint limit N, K → ∞ is characterized by a phase transition
phenomenon for small SNR [21]. For single signal detection, the critical detection
threshold can be expressed in terms of SNR as [139],

1
ρCric = √ , (4.127)
KN

which in fact suggests that when SNR is lower than the critical value, the limiting
λ |
distribution of 1σ H2 1 is the same as that of the largest noise eigenvalue, thus, nullifying
v
the statistical power of a largest eigenvalue test.
λ1 |H1
For ρ > ρCric the distribution of σv2
was found to be asymptotically Gaus-
sian [107, 21] as shown below,

λ1 |H1
∼ NR (µ1 , σ12 ), (4.128)
σv2

where,

K −1
 
µ1 = (1 + Kρ) 1 + (4.129)
NKρ
K −1
 
2 1 2
σ1 = (Kρ + 1) 1 − . (4.130)
N NK 2 ρ 2

Thus, the expression in (4.119) can be written as,

NR (µ1 , σ12 )
THRLRT 1 |H1 =
. (4.131)
NR 1, D2

The above ratio of normal distributed random variable can be approximated with a
single Normal Random Variable as,

2t2
 
2
THRLRT 1 |H1 = NR µ1 , + σ1 . (4.132)
D
4.2 Semi-blind SS techniques under noise uncertainty 119

0.8
Probability of Detection PD

0.6

0.4

HRLRT1 Numerical (S = 2)
0.2
HRLRT1 Theory (S = 2)
HRLRT1 Numerical (S = 8)
HRLRT1 Theory (S = 8)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of False Alarm PF a

Fig. 4.19 Comparison of simulated and analytical performance curves of HRLRT1 method
for N = 80, K = 4 and S = 2 8, SNR = −10 dB

Thus, using the above normal approximation, the probability of detection of

HRLRT1 can be computed,

PD = Prob [THRLRT 1 |H1 > t] (4.133)

 
t − µ1 
= Qq . (4.134)
2t2 2
D + σ1

Simulation Result

The accuracy of the closed-form expression derived in Proposition 4 is confirmed

by the results presented in Figure 4.19, where the analytical curves obtained by
applying theoretical formula in (4.126) and (4.134) are compared against simulated
detection performance. The considered parameters are (K = 4, N = 80, S = 2, 8 and
SNR = −10 dB). As shown in the figure, the theoretical expressions is quite accurate
even for a small number of sensors.
The performance of HED1 and HRLRT1 is compared in Figure 4.20 and Fig-
ure 4.21. The noise variance is estimated by (4.114) in S auxiliary pure noise slots
120 Cognitive Radio: Spectrum Sensing

0.8
Probability of Detection PD

0.6

ED
0.4 HED1 (S = 2)
HED1 (S = 5)
HED1 (S = 10)
RLRT
0.2
HRLRT1 (S = 10)
HRLRT1 (S = 5)
HRLRT1 (S = 2)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of False Alarm PF a

Fig. 4.20 ROC performance of RLRT and ED methods compared with their hybrid counter-
parts. Parameters: N = 80, K = 4 and S = 2, 5, 10, SNR = −10 dB

and the value is then used within the two tests for all the other slots. The curves
approach the ideal ED and RLRT curves by increasing the number of auxiliary slots
S, but the rate of convergence of HED1 is slower.

4.2.12 Hybrid RLRT-2(HRLRT2)

It is evident that the knowledge of the noise power is imperative for achieving RLRT
the optimum performance. Unfortunately, the variation and the unpredictability
of precise noise power is a critical issue for using RLRT in sensing signals with
low SNR. In Sub-Section 4.2.11, we discussed the detection performance of RLRT
with estimated noise variance, where noise variance was estimated from the noise
only slots. In reality, within the same channel, there is a serious problem in finding
noise only samples so as to estimate the variance of the noise samples. As an
alternative, HRLRT2 does not resort on the existence of auxiliary noise only slots,
4.2 Semi-blind SS techniques under noise uncertainty 121

0.8
Probability of Detection PD

0.6

0.4
HRLRT1 (S = 2)
HRLRT1 (S = 15)
RLRT
0.2
HED1 (S = 15)
HED1 (S = 2)
ED
0
−16 −15 −14 −13 −12 −11 −10 −9 −8 −7 −6 −5 −4
SNR in dB

Fig. 4.21 Comparison of HRLRT1 and HED1 performance curves. Parameters: N = 80, K =
4 and S = 2 and 15, SNR = −10 dB

but estimated the noise variance information from the previous slots declared as H0
by the algorithm. Now, for the RLRT scheme using noise variance estimated from
S preceding slots which are declared H0 by same algorithm, which we call Hybrid
RLRT-2 (HRLRT2) model, we derive the analytical expression of performance
parameters PD and PFa in a multi-sensor environment.

Formulation of the Decision Statistic

Using the information on the received signal matrix Y to develop a test statistic
THRLRT 2 which is a modification of the RLRT test statistic considering noise variance
is estimated in S slots declared H0 by RLRT in the first stage. The detector compares
THRLRT 2 against a predefined threshold t. If THRLRT 2 < t then it decides in favor of
Null Hypothesis H0 otherwise in favor of Alternate Hypothesis H1 . The ratio of
largest eigenvalue of the covariance matrix R and the estimated noise variance is
given by,
λ1
THRLRT 2 = 2 . (4.135)
σ̂v
122 Cognitive Radio: Spectrum Sensing

Proposition 5. Let us denote by PS the probability of receiving primary signal plus

noise, PDRLRT the probability of detection of first stage RLRT, and S the number of
slots. The Maximum Likelihood noise variance estimate σ̂v2 using M received signal
samples declared noise samples by RLRT from K receivers is given by,
" #
Ss K M SN K M
∑ ∑ ∑ |hk s(m) + v(m)|2 + ∑ ∑ ∑ |vk (m)|2
s=1 k=1 m=1 s=1 k=1 m=1
σ̂v2 (S) = , (4.136)
KMS

where, SS = SPS 1 − PDRLRT is the number of H1 primary signal slots missed by

RLRT and SN = S − SS is the number of H0 noise slots successfully detected.

Proof. Consider a sampling window of length M before and adjacent to the detection
window which contains slota declared H0 by first stage RLRT. But instead are H1
slots for them the estimated noise variance from the noise only samples using a
Maximum Likelihood Noise Power Estimate can be written as,
" #
K M
1
σ̂v2 = ∑ ∑ |hk s(m) + v(m)|2 . (4.137)
KM k=1 m=1

If the noise variance is constant, the estimation can be averaged over S successive
noise-only slots. Noting that since slots are declared noise only slot by RLRT, there
is a chance with probability (1 − PDRLRT ) that RLRT decides in favor of H0 even
if the received signal is the sum of noise and the primary signal. If the estimated
variance is the mean of S successive noise variances calculated from the noise only
slots decided by ED, then out of S noise variance estimates from S noise only slots
decided by RLRT, PS (1 − PDRLRT )S noise variance estimates are calculated from H1
slots containing noise and the primary signal. Thus, the estimated noise variance
can be written as the mean of the noise variance estimated from noise signal slots
declared H0 by RLRT given H0 and the noise signal slots declared H0 by RLRT
given H1 . Finally, (4.137) could be modified and rewritten as,
" #
Ss K M SN K M
∑ ∑ ∑ |hk s(m) + v(m)|2 + ∑ ∑ ∑ |vk (m)|2
s=1 k=1 m=1 s=1 k=1 m=1
σ̂v2 (S) = . (4.138)
KMS
4.2 Semi-blind SS techniques under noise uncertainty 123

Proposition 6. Consider the HRLRT2 test working with K sensors, N samples, SNR
value ρ, number of auxiliary slots S for noise estimation using (4.138) and threshold
t. Asymptotically in N and K, the false alarm and detection probability of HRLRT2
are respectively given by,

PFa = 1 − F0 (t) (4.139)

 
 t − µx /µ1 
PD = Q 
r
, (4.140)
t2 σ12 +σx2

µ12

where F0 (t), is the cumulative distribution function of the probability distribution

function shown below:

tx − µ
   
1 1
Z ∞
2
f0 (t) = √ |x|fTW 2 exp − 2 (x − µ1 ) dx. (4.141)
ξ σ12 2π −∞ ξ 2σ1

The proof of the above Proposition 6 could be found in following Sub-Section.

Formulation of Detection and False Alarm Probability

The test statistic of HRLRT2 can be written as

λ1 |H0 1
THRLRT 2 |H0 = 2
× 2 2. (4.142)
σv σ̂v /σv

1) Probability of false alarm: Cosider the Null hypothesis H0 where Y = V and

the sample covariance matrix R follows a Wishart Distribution of degree N. For
λ |H
sufficiently large N and K, 1σ 2 0 follows a Tracy Widom Distribution of order 2 [98]
v
with suitably chosen centering and scaling parameters given by.
"  1 #2
K 2
µ= +1 (4.143)
N
"  1 # "  1 #1/3
−2
K 2 K
ξ = N −2/3 +1 +1 . (4.144)
N N
124 Cognitive Radio: Spectrum Sensing

Let us denote the pdf of the Tracy Widom Distribution of order 2 as

z−µ z−µ
    
1
pdf FTW 2 = fTW 2 . (4.145)
ξ ξ ξ

For sufficiently large value of 2KSS M, σ̂v2 /σv2 follows a Normal distribution given
by(See (4.77)),
NR µ1 , σ12 ,


(4.146)

where,

S + SS
µ1 = (4.147)
S
S + 2ρSS + ρ 2 KSS
σ12 = . (4.148)
KMS2

Thus, following above distributions, the test statistic THRLRT 2 |H0 in (4.135) can
be re-written as,  λ | 
FTW 2 ξ1 1σ 2 0 − µ
H
v
THRLRT 2 |H0 = . (4.149)
NR µ1 , σ12

Since the two random variables in (4.149) are statistically independent, the probabil-
ity density functions of the detection statistic THRLRT 2 |H0 can be written as,

tx − µ
   
1 1
Z ∞
2
f0 (t) = √ |x|fTW 2 exp − 2 (x − µ1 ) dx. (4.150)
ξ σ12 2π −∞ ξ 2σ1

Finally, the probability of false alarm can be written as,

PFa = 1 − F0 (t), (4.151)

where F0 (t), is the Cumulative Distribution Function of the Probability Distribution

Function shown in (4.150).
2) Probability of detection: Under alternate hypothesis, the asymptotic dis-
tribution of λ1 in the joint limit N, K → ∞ is characterized by a phase transition
phenomenon for small SNR [21]. For single signal detection, the critical detection
threshold was expressed in terms of SNR as [139],

1
ρCric = √ , (4.152)
KN
4.2 Semi-blind SS techniques under noise uncertainty 125

This suggests that when SNR is lower than the critical value, the limiting distribution
λ |
of 1σ H2 1 is the same as that of the largest noise eigenvalue, thus nullifying the
v
λ1 |H1
statistical power of a largest eigenvalue test. For ρ > ρCric the distribution of σv2
was found to be asymptotically Gaussian [107, 21] with,

λ1 |H0
∼ NR (µx , σx2 ), (4.153)
σv2

where,

K −1
 
µx = (1 + Kρ) 1 + (4.154)
NKρ
K −1
 
2 1 2
σx = (Kρ + 1) 1 − . (4.155)
N NK 2 ρ 2

Thus, the expression in (4.135) can be written as,

NR (µx , σx2 )
THRLRT 2 |H1 =
. (4.156)
NR µ1 , σ12

The above ratio of normal distributed random variable can be approximated with a
single Normal Random Variable as,

µx t2 σ12 + σx2
 
THRLRT 2 |H1 = NR , . (4.157)
µ1 µ12

Thus, using the above normal approximation, the probability of detection of HRLRT2
can be written as
 

2
 t − µx /µ1 
PHRLRT
D = Q
r
. (4.158)
t2 σ12 +σx2

µ12

Simulation Result

Figure 4.22 shows the comparison of simulated and analytical performance curves
of HRLRT2 algorithm. Analytical evaluation of HRLRT2 ROC curves is carried
out in two phases. In first phase, the pdf in (4.150) is evaluated numerically. In
second phase, for a given probability of false alarm, corresponding threshold is
126 Cognitive Radio: Spectrum Sensing

0.8
Probability of Detection PD

0.6

0.4

HRLRT2 Theoretical (S = 2)
0.2
HRLRT2 Theoretical (S = 8)
HRLRT2 Numerical (S = 2)
HRLRT2 Numerical (S = 8)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of False Alarm PD

Fig. 4.22 Comparison of simulated and analytical performance curves of HRLRT2 method.
Parameters: N = 80, K = 4 and S = 2 and 8, SNR = −10 dB

evaluated. Finally, the probability of detection is calculated for each threshold value
using (4.158). The simulated ROC curves of HRLRT2 are superimposed to the
analytically generated ROC to measure the accuracy of the closed form expression
presented by Proposition 6.
Convergence of the Hybrid approach of RLRT detection to ideal RLRT (with
known variance) is illustrated in Figure 4.23. With the increase in number of auxiliary
slots used for noise variance estimation, the performance of HRLRT1 and HRLRT2
converge to the ideal RLRT performance.
4.2 Semi-blind SS techniques under noise uncertainty 127

Probability of Detection Pd

0.8

0.6

HRLRT2 (S = 2)
0.4 HRLRT2 (S = 15)
HRLRT1 (S = 15)
HRLRT1 (S = 2)
RLRT (Known variance)
0.2
−12 −11 −10 −9 −8 −7 −6 −5 −4
SNR in dB

Fig. 4.23 Probability of detection vs. SNR for RLRT, HRLRT1 and HRLRT2, Parameters:
N = 80, K = 4 and S = 2 and 15, SNR = −10 to −4 dB

4.2.13 SNR Wall Analysis of ED and HED1

For CR application the most popular sensing algorithm is the simpler ED that
compares the energy of the received signal against the noise variance σv2 as discussed
before. ED requires the perfect knowledge of the noise power at the receiver [164,
50, 51, 105], however, in real systems the detector does not have a prior knowledge
of the noise level. It is known from [159] that there is a certain SNR threshold in
case of noise uncertainty known as SNR wall, which prevents ED from achieving
the desired performance even if the detection interval tends to infinity.
In this work, starting from the contributions of [159, 118], we extend the con-
dition of SNR Wall [159] to multi-sensor hybrid ED with auxiliary noise variance
estimation (offline method) described in Sub-Section 4.2.6. For auxiliary noise vari-
ance estimation of White Gaussian Noise samples, the distribution of the estimated
variance is studied and linked to the uncertainty bound referred to [159]. The SNR
Wall expression is derived for multisensor ED and proved to be independent of the
number of sensors. It is concluded that the noise uncertainty can be reduced by
increasing the number of samples used for noise variance estimation, but the number
of samples/slots used for noise estimation exponentially increases as the SNR Wall
condition becomes more stringent.
128 Cognitive Radio: Spectrum Sensing

With reference to the system model of Sub-Section 4.2.4 we briefly summarize

the main results of ED presented in Sub-Section 4.2.7. Using the information on
the received signal matrix Y , ED test statistic TED is employed by a detector to
distinguish between H0 and H1 . TED is the measure of the average energy of the
received signal over a sensing interval N from K receivers which can be represented
as,
K N
1
TED = ∑ ∑ |yk (n)|2.
KNσv2 k=1
(4.159)
n=1
Here, the total energy estimated by TED is the normalized average energy over the
sensing interval. In order to make a decision, the detector compares TED against
a predefined threshold t: if TED > t it decides for H1 , otherwise for H0 . As a
consequence, the false alarm and detection probabilities are defined as,

PFa = Pr(T > t|H0 ), PD = Pr(T > t|H1 ). (4.160)

In practical applications, the decision threshold t is typically computed as a

function of the target PFa : this ensures CFAR detection.
If the noise variance is assumed to be perfectly known, then TED in (4.159) can
be simplified using a Gaussian approximation as,

 H0 , .
1
(

NR 1, KN
TED = 2 +2ρ+1 (4.161)
NR (ρ + 1), Kρ KN H1

Based on the above approximation, PFa and PD of ED can be easily simplified

as,
h √ i
PFa = Q (t − 1) KN (4.162)
" √ #
(t − 1 − ρ) KN
PD = Q p , (4.163)
Kρ 2 + 2ρ + 1

where Q(.) is the standard normal complementary cdf.

If we eliminate the threshold t by using a similar approach as in [159] from (4.162)
and (4.163) to calculate the number of samples in a sensing slot to achieve the re-
4.2 Semi-blind SS techniques under noise uncertainty 129

107
K=1
K=5
106 K=10
K=15

Number of samples N
K20
105

104

103

102

101
−30 −28 −26 −24 −22 −20 −18 −16 −14 −12 −10
SNR in dB

Fig. 4.24 Comparison of sample complexity N to achieve given PD′ and PFa
′ for single antenna

vs. multi antenna environment for different SNR assuming perfect knowledge of noise
variance. Parameters: K = 5, 10, 15 and 20, P’D = 0.9, P’Fa = 0.1

quired PD and PFa , we get:

hp i2
(Kρ 2 + 2ρ + 1)Q−1 (PD ) − Q−1 (PFa )
N= . (4.164)
Kρ 2

The expression of N in (4.164) exactly matches the expression of N derived

in [159] when K = 1(single antenna) and with real signals. Figure 4.24 plots the
sample complexity N for different SNR in single and multi antenna scenario. The
plot clearly unveil that multi-sensor ED requires relatively less number of samples
to achieve the given PD′ and PFa′ . It is also clear that, given a priori information on

the noise variance, received signal can be detected for any SNR maintaining the
detection interval accordingly.
Now, let us consider the case in which the noise variance ish not precisely
i known
1 2
but its deviation is known to be bounded in the interval [159] β σv , β σv2 , i.e.,
 
1 2
σ̂v2 ∈ 2
σ , β σv , (4.165)
β v

where σv2 is the nominal (true) noise power and β > 1 is the parameter that quantifies
the level of uncertainty. It is evident that the knowledge of the noise variance is
imperative for the optimum performance of ED. In practice it is not possible to know
130 Cognitive Radio: Spectrum Sensing

the exact value of the noise variance, so the only option is to estimate it from the
noise samples. The performance of ED with auxiliary noise variance estimation
known as Hybrid ED (HED1) using Maximum Likelihood Estimation (MLE) was
studied in previous sections. If we denote the estimate of the noise variance with σ̂v2
and consider the ED detection statistic as,

1 K N
T̃ED = ∑ ∑ |yk (n)|2,
KN k=1
(4.166)
n=1

the Gaussian approximation of the detection statistic in (4.166) can be written as,

  4

 NR σ̂v2 , σ̂v
KN H0 ,
T̃ED =  2 4 2 2 4
 (4.167)
 NR ∥h∥σ 2 + σ̂ 2 , K∥h∥ σs +2∥h∥σs σ̂v +σ̂v H1 ,
s v KN

whose performance parameters can be written as,

√
(t − σ̂v2 ) KN
 
PFa = Q (4.168)
σ̂v2
" √ #
(t − σ̂v2 − ∥h∥σs2 ) KN
PD = Q p . (4.169)
K∥h∥2 σs4 + 2∥h∥σs2 σ̂v2 + σ̂v4

Now, the worst case scenario of PD and PFa , represented by the notations P’D
and P’Fa , can be analyzed as,

√
(t − σ̂v2 ) KN
 
P’Fa = hmax iQ (4.170)
σ̂v2 ∈ β1 σv2 ,β σv2 σ̂v2
 
t
 √ 
σv2
−β N
P’Fa = Q   (4.171)
β
√ !
(t − σ̂v2 − ∥h∥σs2 ) KN
P’D = h min iQ p (4.172)
K∥h∥2 σs4 + 2∥h∥σs2 σ̂v2 + σ̂v4
σ̂v2 ∈ β1 σv2 ,β σv2
  √ 
t
σv2
− ρ β1 KN
P’D = Q  q . (4.173)
2 1 1
Kρ + 2ρ β + β 2
4.2 Semi-blind SS techniques under noise uncertainty 131

10
K = 1 and x = 0.1
K = 10 and x = 0.1
K = 1 and x = 0.01
8 K = 10 and x = 0.01

6
log10 N

0
−24 −22 −20 −18 −16 −14 −12 −10 −8 −6 −4 −2 0
SNR in dB

Fig. 4.25 Variation of sample complexity N as the SNR approaches SNR wall for ED,
Parameters: K = 5, P’D = 0.9, P’Fa = 0.1 [use x = 10 log10 β in Equation (4.174)]

By solving (4.171) and (4.173) and eliminating t, we get,

h q i2
β Q−1 (P’Fa ) − Q−1 (P’D ) Kρ 2 + 2ρ β1 + β12
N= h  i2 . (4.174)
K ρ − β − β1

Equation (4.174) gives the expression of the number of samples in a sensing slot
for given SNR condition to achieve the required performance in terms of PFa and PD ,
which is also an extension of the result in [159] for a multi-sensor ED. The number
of samples N goes to infinity as SNR approaches β − β1 . This condition is known as
SNR Wall condition, which means that under this SNR wall value we cannot achieve
the required performance in a given level of noise uncertainty even if the sample
number is made sufficiently large. Thus, SNR Wall is given by the expression as
shown below,  
ED 1
SNRwall = β − . (4.175)
β

SNR wall makes clear that the Energy Detector cannot robustly detect the signal
if the signal power is less than the uncertainty of the noise power, i.e.,
 
1
σs2 < β− σv2 . (4.176)
β
132 Cognitive Radio: Spectrum Sensing

·10−2
6
Emperical pdf of V
Analytical pdf of V
5 p-value

Probability 4

3 β = 1.4 for α = 0.01

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
V

Fig. 4.26 Probability distribution of normalized noise variance estimate V . Parameters:

S = 2, K = 2, M = 10

Our study shows that there is no difference in the SNR wall expression by using
a multi antenna or a single antenna in ED. Figure 4.25 illustrates the SNR wall
condition for single/multi antenna case and the variation of the sample complexity
N, as SNR approaches SNR wall for different levels of noise uncertainty.

Noise Uncertainty Distribution and Formulation of Uncertainty Bound

Let us suppose that noise variance is estimated in S auxiliary slots with M noise only
samples in each slot. Then the ML noise variance estimate using SKM noise-only
samples obtained from K receivers with S slots each can be written as,

1 S K M
σ̂v2 (S) = ∑ ∑ ∑ |vk (m)|2. (4.177)
KSM s=1 k=1 m=1

If we focus on noise variance distribution under the same scenario, let us consider a
random variable V (normalized
 2 noise variance estimate), which associates a unique
σ̂v
numerical value from σ 2 for each noise variance estimation, where σ̂v2 corresponds
v
to (4.177).
4.2 Semi-blind SS techniques under noise uncertainty 133

Now, V can be written as

σ̂v2
 
V= (4.178)
σ2
 v  2 
S K M
1  vk (m)  
 
= ∑ ∑ ∑  σ /2  .
p (4.179)
2KMS s=1

k=1 m=1 v

It can be noted that the expression in (4.179) is the sum of squared standard
normal noise samples, thus it has a Chi-Square Distribution with 2KMS degrees of
freedom. Then we have
 2 
χ2KMS
V= , (4.180)
2KMS

2
where χ2KMS represents a Chi-Squared random variable with 2KMS degrees of
freedom.
Figure 4.26 shows the pdf of the random variable V according to (4.180) com-
pared to its empirical pd f . Now, the uncertainty bound β can easily related with V
using (4.165) and (4.180) as,

β = max (V ) (4.181)
F −1 (1 − α, 2KMS)
= Chi , (4.182)
2MKS
−1
where FChi () is the Inverse Cumulative Distribution Function of a Chi Square Dis-
tributed Random Variable and α is the significance level of a Chi Square Distribution.
As we know, a Chi Square Distributed Random variable can take any values in the
range (−∞, +∞), which leads (4.182) to infinity. In (4.182), (1 − α) is the signif-
icance level, which gives an insight of percentage coverage of all possible values
smaller than the value given by the inverse cdf. Thus,

−1
FChi (1 − α, 2KMS)
β= . (4.183)
2MKS

Finally, β can be expressed in terms of the total number of noise samples (i.e.
KMS) considered for variance estimation as shown in (4.183) and can be easily
134 Cognitive Radio: Spectrum Sensing

1.4
α = 0.001
α = 0.001
1.2 α = 0.0001

Noise Uncertainty x (in dB)

1

0.8

0.6

0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
Number of slot S used for noise variance estimation

Fig. 4.27 Variation of noise uncertainty level with number of slots S used for noise variance
estimation.

evaluated. For example, for α = 0.01, β is found to be 1.4 as shown in Figure 4.26.
It is clear from (4.183) that the level of noise uncertainty decreases with the increase
of the number of samples and the number of slots averaging the estimation of the
noise variance.

Simulation Result

The variation of the noise uncertainty bound β with number of slots S used in noise
variance estimation for different significance level, is presented in Figure 4.27. The
plot seems to have a steep slope (∼ - 0.08dB/slot) for small values of S, but for larger
number of estimation slots S (S > 30) the slope starts to flatten suggesting large
changes in S are required for small gain in noise uncertainty bound. Figure 4.28
illustrates the variation of SNR wall as a function of the number of slots used for
noise variance estimation. Curves are plotted for different significance parameter α.
As expected, the level of SNR Wall decreases with the increase in the number of slots
used in noise estimation but curve saturates for low values of SNR Wall (< −12dB).
As a matter of fact, for given parameters (M = 100, K = 5 and α = 0.01), the number
of slots required to overcome the SNR Wall condition for SNR ≤ to −12dB is S > 80.
Similarly, Figure 4.29 plots the SNR wall condition and the sample complexity
as a function of the number of auxiliary slots S. It shows that the minimum attainable
SNR wall value of −12 dB obtained for S = 100 and α = 10−5 . Moreover, for
4.2 Semi-blind SS techniques under noise uncertainty 135

−2
α = 0.01
α = 0.001
−4 α = 0.0001

−6
SNR wall (in dB)

−8

−10

−12

−14
0 10 20 30 40 50 60 70 80 90 100
Number of slots S used for noise variance estimation

Fig. 4.28 Variation of SNR wall level with number of auxiliary slots S used for noise variance
estimation

asymptotic number of slots S (then asymptotic of total number of samples KMS), the
noise uncertainty x (in dB) decreases to zero with σ̂v2 = σv2 resulting no SNR wall,
proving the finding of [119].
136 Cognitive Radio: Spectrum Sensing

9
S=1
S = 10
8

Number of samples (log10N) required

S = 50

to achieve PD = 0.9 and PFa = 0.1

S = 100
7

1
−15 −10 −5 0
SNR in dB

Fig. 4.29 Variation of sample complexity N as the SNR approaches SNR wall for varying
number of auxiliary slots used for noise variance estimation in HED1

4.3 Semi-blind sensing techniques under unknown Pri-

mary User Traffic

Among many practical imperfections and constraints for spectrum sensing in CR

scenarios mentioned in the literature, the unknown PU traffic is one of the most
important constraints which significantly limits the sensing performance of the
secondary user. In the existing literature on spectrum sensing, the SUs are assumed
to have a perfect knowledge of the exact time slot structure of PU transmissions
providing a solid basis for guaranteeing that PU traffic transitions occur only at the
beginning of the SU sensing slots. However, in practice, the SU may not have the
knowledge of exact time slot structure of PU transmissions. Moreover, it is also
possible that the communications among PUs are not based on synchronous schemes
at all [73, 186]. It means, under practical scenarios that the primary traffic transition
may occur during the sensing period, especially when a long sensing period is used
to achieve a good sensing performance, or when spectrum sensing is performed
for a network with high traffic load. Thus, it is necessary to analyze the sensing
performance of existing spectrum sensing techniques under unknown PU traffic.
Among a limited number of literature including[138, 113, 170, 140, 174, 160]
that deal with unknown primary traffic scenario, [138] was the first one to study
the performance of well known semi-blind spectrum sensing algorithms including
ED and Roy’s Largest Root Test (RLRT) under bursty primary traffic, in which
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 137

the burst interval is comparable to or smaller than the spectrum sensing interval.
The traffic model used is limited to constant length bursts of the PU data, whose
length is smaller than the SU sensing duration. However, the burst length of the
PU may be varying with time following some stochastic models [137, 84]. A more
general scenario, in which the PUs traffic transition is completely random, may
affect the spectrum sensing performance. The analysis of the spectrum sensing
performance has been presented in [170, 140, 174, 160] by modeling the PU traffic
as an independent and identically distributed two state Markov’s model. Using
this primary traffic model, authors in [170, 174, 160] analyzed the effect of PU
traffic on the sensing performance and the sensing-throughput trade-off considering
ED as a sensing technique under the half duplex scenario. Moreover, the effect of
multiple PUs traffic on the sensing-throughput trade-off of the secondary system has
been studied in [140]. Although all the aforementioned contributions recognized
the fact that the PU traffic might affect the sensing performance including sensing-
throughput trade-off, none of them considered the realistic scenario of multi-antenna
spectrum sensing in a complex signal sample domain and the sensing performance
of other spectrum sensing techniques including Eigenvalue Based Detection (EBD)
techniques under unknown primary user traffic.
In this section we study the effect of PU traffic on the performance of multi-
antenna spectrum sensing especially ED and RLRT under the complex domain of PU
signal, noise and channel. In contrast to commonly used continuous time Markov
model in the existing literature, a novel technique of modeling PU traffic is proposed
which is only based on discrete time distribution of PU free and busy periods.
The proposed model is more realistic and simple compared to the continuous time
Markov model proposed in the previous literature [170, 140, 174, 160]. Moreover,
an analytical performance evaluation of the decision statistic in terms of ROC under
the considered scenario is carried out.

Mathematical Framework

We consider a single source scenario (single primary transmitter) whereas multiple

antennas are employed by an SU. Suppose the SU has K antennas and each antenna
receives N samples in each sensing slot. In a given sensing frame, the detector
calculates its decision statistic TD by collecting N samples from each one of the K
antennas. Subsequently, the received samples are collected by the detector in the
138 Cognitive Radio: Spectrum Sensing

form of a K × N matrix Y. When the primary transmissions are not based on some
synchronous schemes or the sensing unit at the SU does not have any information
about the primary traffic pattern, the received vector at the sensing unit may consist
in part by the samples from one PU state and the remaining part by alternate PU state.
To simplify the scenario, we begin with the following classification of the sensing
slots based on the PU traffic status, which is also illustrated in Fig. 4.30.

1. Steady State (SS) sensing slot: In such type of sensing slot, all the received
samples in one sensing slot are obtained from the same PU state.

2. Transient State (TS) sensing slot: In such type of sensing slot, a part of the
received samples within the sensing slot are obtained from one PU state and
the remaining from the another PU state.

In general, the probabilities of observing SS and TS sensing slots are dependent on

the PUs traffic model. In contrast to the commonly used hypothesis definition in
spectrum sensing literature, we define two hypotheses in the following way:

H0 : the channel is going to be free,

H1 : the channel is going to be busy.

This hypothesis formulation implies that the decision is based on the PU status at the
end of the sensing interval. Thus, in a TS sensing slot, a transition from the PU busy
state to the PU free state is considered H0 , while a transition from the PU free state
to the PU busy state is considered H1 . In the considered scenario, in an SS sensing
interval, the generic received signal matrix under each hypothesis can be written as,

(
V [K,N] (H0 ),
Y SS = , (4.184)
S[K,N] (H1 ),

where V [K,N] ≜ [v(1) · · · v(n) · · · v(N)] is the K × N noise matrix ,S[K,N] = h[K,1] s[1,N] +
V[K,N] is the K × N received noisy signal matrix when PU signal is present, h[K,1] =
[h1 · · · hK ]T is the channel vector and s[1,N] ≜ [s(1) · · · s(n) · · · s(N)] is a 1 × N PU
signal vector. In the TS sensing interval, the generic received signal matrix under
each hypothesis can be written as,
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 139

Fig. 4.30 Primary user traffic scenario and sensing slot classification

(
[S[K,N−D0 ] |V[K,D0 ] ] (H0 ),
Y TS = . (4.185)
[V[K,N−D1 ] |S[K,D1 ] ] (H1 ),

where D0 represents the number of pure noise samples in TS sensing slot under
H0 , D1 represents the number of noise plus PU signal samples in TS sensing slot
under H1 , S[K,N−D0 ] = h[K,1] s[1,N−D0 ] +V[K,N−D0 ] is the (K ×N −D0 ) received noisy
signal matrix when PU signal is present only for (N − D0 ) sample periods. Similarly,
S[K,D1 ] = h[K,1] s[1,D1 ] + V[K,D1 ] is the K × D1 received noisy signal matrix when
PU signal is present only for D1 sample periods. In each of them, the unknown
primary transmitted signal s(n) at time instant n is modelled as independent and
identically distributed (IID) complex Gaussian with zero mean and variance σs2 :
s(n) ∼ NC (0, σs2 ). The noise sample vk (n) at the kth antenna of the SU at the
time instant n is also modelled as complex Gaussian with mean zero and variance
σv2 : vk (n) ∼ NC (0, σv2 ). The channel coefficient hk of the kth antenna is assumed to
be constant and memory-less during the sensing interval. The average SNR at the
σ 2 ∥h∥2
receiver is defined as, ρ = sKσ 2 , where ||.|| denotes the Euclidean norm.
v

4.3.1 Characterization of Primary user traffic

In this sub-section, we characterize the mathematical model of PU traffic. Based on

the proposed stochastic PU traffic model, we build the PU’s probability transition
140 Cognitive Radio: Spectrum Sensing

matrix, which leads to analytical formulation of the SU’s probability of receiving SS

sensing frame and TS sensing frame under null and alternate hypothesis.
In this work, we model the PU traffic as a two state Markov process (On-Off
process: PU ‘On’ representing busy state and PU ‘Off’ representing free state).
The length of free as well as busy period are independent geometrically distributed
random variables with parameters α and β , respectively. Essentially, the parameters
α and β represent the state transition probabilities in single sample duration. The
mean length of free period M f and busy period Mb of PU traffic can be related to
parameters α and β as, M f = α1 and Mb = β1 , respectively.
Mf
At any time instant, the PU is in free state with probability Pf = Mb +M f and,
similarly, in the busy state with probability, Pb = MbM+M
b
f
. We further assume that
the parameters (α and β ) of geometrically distributed length of PU free and busy
periods are constant over time. Thus, the corresponding two-state Markovs process
can be considered homogeneous in nature. Using this homogeneity property and the
Chapman-Kolmogorov equation, the PU n-step transition probability matrix is given
by:

" #
pn00 pn01
Pn =
pn10 pn11
" #
1 β + α(1 − α − β )n α − α(1 − α − β )n
= , (4.186)
α +β β − β (1 − α − β )n α + β (1 − α − β )n

which reduces to (4.187) for single step transition matrix:

" # " #
p00 p01 1−α α
P= = . (4.187)
p10 p11 β 1−β

As already mentioned earlier in Sub-Section 4.3, the stochastic nature of the PU

state transition gives a mixed nature of received signals in a TS sensing slot resulting
in random variables (RVs) D0 and D1 . Thus, in each PU state transition from Busy
to Free State, the sensing unit has to decide based on D0 pure noise samples and
(N − D0 ) noise plus primary signal samples which actually affects the overall sensing
performance. Thus, with the support of above analysis and also keeping (4.184) and
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 141

(4.185) in reference, it is clear that to find the distribution of the decision statistic
under different hypotheses, the prior deduction of the probability of occurrence of
SS sensing slot, TS sensing slot, probability mass function (pmf ) of D0 and the pmf
of D1 are inevitable.
The following Lemmas compute the pmf s of D0 and D1 based on the two state
PU traffic model described above.

Lemma 1. Given the N number of samples in a sensing slot, the probability transition
matrix P as in (4.187) with comparable mean parameters M f and Mb ,

1. The probability of having D0 noise only (PU signal free) samples in a TS

sensing slot under H0 reduces to,

1
PD0 (d0 )|H0 = . (4.188)
N −1

2. The probability of having D1 noise-plus-PU-signal samples in a TS sensing

slot under H1 reduces to,

1
PD1 (d1 )|H1 = . (4.189)
N −1

Proof. As mentioned earlier during binary hypothesis formulation, the PU state

transition from Busy State to Free State corresponds to H0 sensing slot and viceversa
for H1 . We consider thus, without loss of generality, while dealing with a H0
sensing slot, that the TS sensing slot occurs at the beginning of the PU Free State.
Thus, given that the PU is initially in Busy State, the probability of having a PU state
transition after N − D0 samples, leading to D0 pure noise samples in a TS sensing
slot under H0 is given by,

0 −1
P(T S, D0 , H0 ) = pb · pN−D
11 p10 pD
00 .
0
(4.190)

The probability in (4.190) is the PU state transition probability after N − D0 sample

instances from a busy PU state to a free PU one. This probability can be normalized
142 Cognitive Radio: Spectrum Sensing

by the sum of the transition probabilities for all range of D0 and we obtain
N−d0 −1
pb · p11 p10 pd000
PD0 (d0 )|H0 = N−1 N−a−1
(4.191)
∑a=1 pb · p11 p10 pa00
(1 − β )N−d0 −1 (1 − α)d0
= N−1
. (4.192)
∑a=1 (1 − β )N−a−1 (1 − α)a

When the mean parameters M f and Mb are close (4.192) can be approximated to:

(1 − α)N−d0 −1 (1 − α)d0
PD0 (d0 )|H0 = (4.193)
∑N−1
a=1 (1 − α)
N−a−1 (1 − α)a

1
= . (4.194)
N −1
This proves the first claim. By using the same line of reasoning, the proof of the
second claim is straightforward.

The following Lemmas compute the probability of occurrence of SS sensing slot

pSS |H0 under H0 and the probability of occurrence of TS sensing slot, which is the
complementary, i.e., pT S |H0 = 1 − pSS |H0 .

Lemma 2. Given the number of samples in a sensing slotN and the probability
transition matrix P as in (4.187),

1. The probability of receiving SS sensing slot under H0 is given by:

1
PSS |H0 = . (4.195)
1 + α ∑N
d0 =1 (1 − β )
N−d0 −1 (1 − α)d0 −N

2. The probability of receiving SS sensing slot under H1 is given by:

1
PSS |H1 = . (4.196)
1+β ∑N
d1 =1 (1 − α)
N−d1 −1 (1 − β )d1 −N

Proof. First of all, the probability of having no PU state transition under H0 is given
by:

P(SS, H0 ) = p f · pN
00 . (4.197)
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 143

Similarly, the probability of having a PU state transition from Busy State to Free
State is given by,
! !
PU is in PU transits from Busy to Free
P(T S, H0 ) = P ·P . (4.198)
Busy State State during sensing interval

Essentially, the PU state transition may occur at any time during the sensing
duration,
N−1
0 −1
P(T S, H0 ) = pb · ∑ pN−d
11 p10 pd000 . (4.199)
d0

Thus, by using (4.187) and (4.199), the probability of having a SS sensing slot
belonging to H0 is given by,

P(SS, H0 )
PSS |H0 = (4.200)
P(SS, H0 ) + P(T S, H0 )
p f · pN
00
= N−1 N−d0 −1
. (4.201)
p f · pN
00 + pb · ∑d0 p11 p10 pd000

By replacing p f , pb and the elements of the probability transition matrix by their

respective expressions in terms of α and β , further simplification yields (4.195) and
proves claim 1 of Lemma 2.
Through the same line of reasoning, the proof of claim 2 of Lemma 1 is straight-
forward.

Corollary 5. When the mean parameters M f and Mb are comparable, the probabili-
ties in (4.195) and (4.196) reduce to simple expressions given by,

1
PSS |H0 = (4.202)
1 + (N − 1)α
1
PSS |H1 = . (4.203)
1 + (N − 1)β

In Figure 4.31, the pdf of the decision statistic under ideal PU-SU sensing slot
synchronization is compared with the pdf of the decision statistic under unknown
PU traffic considering both hypotheses. In addition, the accuracy of the derived
144 Cognitive Radio: Spectrum Sensing

analytical pdf expressions is confirmed by the results presented in Figure 4.31b,

where the theoretical formulas are compared against the numerical results obtained
by Monte-Carlo simulation. The perfect match of the theoretical and the numerical
pdfs validates the derived analytical expressions.

4.3.2 ED under Unknown PU Traffic

ED computes the average energy of the received signal matrix Y normalized by the
noise variance σv2 and compares it against a predefined threshold TED :

1 K N
TED = ∑ ∑ |yk (n)|2.
σv2 k=1
(4.204)
n=1

To analyze ED performance, it is necessary to express the pdf of the decision statistic

in case of unknown primary traffic. The following theorem computes the pdf of the
ED decision statistic under both H0 and H1 , by using the PU traffic characterization
presented Sub-Section 4.3.1.

Theorem 6. Given a multi-antenna sensing unit with K receiving antennas, N

received samples in each slot and a random PU traffic with geometrically distributed
free state duration, the pdf of the ED decision statistic under H0 and H1 is given
by (4.205) and (4.206), respectively, where fG (x, α, β ) is the pdf of a Gamma
distribution with shape parameter α and rate parameter β and fN (x, µ, σ 2 ) is the
pdf of Gaussian distribution with mean µ and variance σ 2 .

N−1
fTED |H0 (x) = pSS |H0 fG (x, KN, 1) + pT S |H0 ∑ PD (d0 ) [ fG (x, 2Kd0 , 1) + fG (x, N − d0 , Kρ)
0
d0 =1
+ fG (x, K(N − d0 ), 1) + fN (x, 0, 2ρK(N − d0 ))] , (4.205)
fTED |H1 (x) = pSS |H1 ( fG (x, N, Kρ) + fG (x, KN, 1) + fN (x, 0, 2ρKN))
N−1
+ pT S |H1 ∑ PD (d1 ) [ fG (x, d1 , Kρ) + fG (x, Kd1 , 1)
1
d1 =1
+ fN (x, 0, 2ρKd1 ) + fG (x, K(N − d1 ), 1)] . (4.206)

Proof. As noted from Sub-Section 4.3, the term within the summation in (4.204)
is different for the SS sensing slot and TS sensing slot. Under the null hypothesis
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 145

·10−4
3
H0
H1
2
pdf

0
0 200 400 600
TED
(a) Known PU Traffic (synchronized PU-SU
sensing slots)
·10−4
Numerical
H0 (Theoretical)
2 H1 (Theoretical)
pdf

0
0 200 400 600
TED
(b) Unknown PU Traffic

Fig. 4.31 Pdfs of the ED decision statistic: Parameters: N = 50, K = 4, M f = 150, Mb = 150
and SNR = −6 dB
146 Cognitive Radio: Spectrum Sensing

H0 , the ED decision statistic in (4.204) can be decomposed as a probabilistic sum

SS | TS
of TED H0 and TED |H0 .

"
K  vk (n) 2 pT S |H0 K D0  vk (n) 2
N
   
pSS |H0
TED |H0 =
2 ∑ ∑  √  + 2 ∑ ∑  σ /√2 
k=1 n=1 σv / 2 k=1 n=1 v
K N
 2 #
 hk s(n) + vk (n) 
+∑ ∑  σ /√2  . (4.207)
k=1 n=N−D0 +1 v

Next, the distribution of each sum in (4.207) can be derived as [46],

N−1
pSS |H0 2 pT S |H0  2
TED |H0 = χ2KN + ∑ PD (d0 )
0 χ2Kd0
2 2 d0 =1
i
2
+ Kρ χ2(N−d0)
+ χ 2
2K(N−d0 ) + N (0, 2ρ(N − d0 )K) , (4.208)

where χA2 represents a Chi-squared random variable with A degrees of freedom. and
N (µ, σ 2 ) represents the Normal random variable with mean µ and variance σ 2 .
In fact, the product of a Chi-squared RV with a constant is a Gamma RV, thus,
with this replacement we obtain,

N−1
TED |H0 = pSS |H0 G (KN, 1) + pT S |H0 ∑ PD (d0 ) [G (Kd0 , 1)
0
d0 =1
+ G (N − d0 , Kρ) + G (K(N − d0 ), 1) + N (0, 2ρ(N − d0 )K)] . (4.209)

In addition, G (α, β ) represents a Gamma random variable with a shape parameter

α and a rate parameter β . Since the goal is to find the pdf of the sum in (4.204)
under H0 , we replace the random variables in (4.209) with their respective pdfs to
obtain (4.205).
Similarly, under the alternate hypothesis H1 , the ED decision statistic in (4.204)
SS | TS
can be decomposed as a probabilistic sum of TED H1 and TED |H1 .
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 147

pSS |H1 K N  hk s(n) + vk (n) 2 pT S |H1

 
TED |H1 =
2 k=1 ∑ ∑  σ /√2  + 2
n=1 v
" 2 K  #
K D1  N  vk (n) 2
 
hk s(n) + v k (n) 
· ∑ ∑  √  +∑ ∑  √  . (4.210)
k=1 n=1 σv / 2 
k=1 n=N−D1 +1 σv / 2

Using the fact that D1 is a random variable,

pSS |H1 K N  hk s(n) + vk (n) 2 pT S |H1 N−1

 
TED |H1 =
2 k=1 ∑ ∑  σ /√2  + 2 ∑ PD1 (d1 )
n=1 v d1 =1
" #
K d1 
 2 K N
 2
hk s(n) + v k (n)   vk (n) 
· ∑ ∑  √  +∑ ∑  √  . (4.211)
k=1 n=1 σv / 2 
k=1 n=N−d1 +1 σv / 2

Deriving the distribution of each sum in (4.211) using [46], we obtain:

TED |H1 = pSS |H1 (G (N, Kρ) + G (KN, 1) + N (0, 2ρKN))

N−1
+ pT S |H1 ∑ PD (d1 ) [G (d1 , Kρ) + G (Kd1 , 1)
1
d1 =1
+ N (0, 2ρKd1 ) + G (K(N − d1 ), 1)] . (4.212)

Finally, we replace the random variables in (4.212) with their respective pdfs to
obtain (4.206).

In essence, the pdfs in (4.205) and (4.206) consist of the sum of independent
random variables. From a statistical point of view, the pdf of the sum of two
independent pdfs can be realized as the convolution of these pdfs [92]. Thus, the
sum of pdfs can be computed using convolution as an alternative, we can exploit
the characteristic function approach by computing Fourier transform. In conclusion,
(4.205) and (4.206) can be easily evaluated by using standard Fast Fourier Transform
(FFT) techniques.
A. Probability of False Alarm: Given the pdf of the decision statistic in (4.205),
we can compute the false alarm probability. Under H0 , the PU is in free state at
148 Cognitive Radio: Spectrum Sensing

the end of the sensing interval, but the decision statistic is erroneously above the
threshold τ and the PU signal is declared present. For false alarm PFa , the following
Corollary of Theorem 6 holds.

Corollary 6. The false-alarm probability of the ED test under unknown PU traffic

and complex signal space scenario is:

Z +∞
PFa = P(TED |H0 ≥ τ) ≡ fTED |H0 (x)dx. (4.213)
τ

B. Probability of Detection: Given the pdf of the decision statistic in (4.206), we

can compute the detection probability. Under H1 , i.e., the PU is in busy state at the
end of the sensing interval. Under this scenario, if the decision statistic is above the
threshold, the PU signal is declared present. The following Corollary of Theorem 6
holds for defines the probability of detection PD .

Corollary 7. The detection probability of the ED test under unknown PU traffic and
the complex signal space scenario is given by:

Z +∞
PD = P(TED |H1 ≥ τ) ≡ fTED |H1 (x)dx. (4.214)
τ

Simulation Results

In this Sub-Section, the effect of PU traffic on the multi-antenna ED is analyzed based

on the the traffic model developed in Sub-Section 4.3.1. The analytical expressions
derived are validated via numerical simulation. The length of the free and busy
periods of the PU traffic are measured in terms of the discrete number of samples
where each of them has Geometric distribution with probability of success parameters
p f and pb , respectively. Here, we use mean and busy period denoting M f = p1f and
1
Mb = pb , respectively. Under multiple antenna sensing scenario, the average SNR at
σs2 ∥h∥2
the receiver is defined as, ρ = Kσv2
, where ||.|| denotes the Euclidean norm.
Figure 4.32 illustrates the ED ROC performance for different values of the
mean free and busy period of the PU traffic. It shows that as the mean free and
busy periods of the primary traffic increases, the detection performance of SU also
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 149

0.8
Mf , Mb ↑
0.6
PD

0.4 Classical ED
Mf = Mb = 550
0.2 Mf = Mb = 250
Mf = Mb = 100
Mf = Mb = 75
0
0 0.2 0.4 0.6 0.8 1
PF
Fig. 4.32 ROC performance for the considered scenario. Parameters: N = 100, K = 4 and
SNR = −6 dB

increases. Clearly, the conventional model with perfect synchronization of the

PU-SU sensing slots performs better than the one with unknown PU traffic.
The behaviour variation of the sensing performance for different number of
receiving antennas is plotted in Figure 4.33. It can be observed that unlike the
rapid increase in sensing performance with increasing number of receiving antennas
under synchronized PU-SU sensing slot scenario (rapid decrease in missed-detection
probability with increasing number of receiving antennas), the sensing performance
is almost constant even if we increase the number of antennas under unknown
PU traffic. During a TS sensing slot, from each receiving antenna, the received
signal samples are a mixture of pure noise samples and the samples with both noise
and PU signal. Thus, even if we use multiple antennas, the nature of the received
signal doesn’t change much which is the reason why the sensing performance
improvement is suppressed by the unknown PU traffic (more specifically, the TS
sensing performance) when the length of the free and busy periods of PU traffic are
quite small (a few multiples of the sensing window length).
150 Cognitive Radio: Spectrum Sensing

100

10−1
PM d

10−2

10−3
Unknown PU traffic (SNR = −5 dB)
Unknown PU traffic (SNR = −3 dB)
Known PU traffic (SNR = −5 dB)
Known PU traffic (SNR = −3 dB)

−4
10
2 3 4 5 6 7 8 9 10
K
Fig. 4.33 Probability of missed detection vs. the number of Antennas, Parameters: N = 25,
M f = 62, Mb = 62 and PF = 0.1

4.3.3 EBD under Unknown PU Traffic

In this Sub-Section, we consider the performance under unknown PU traffic of an

another important class of detection techniques designed for multi-sensor detectors,
based on the eigenvalues of the received signal covariance matrix. Receive diversity
can be achieved either by multiple users (cooperative detection) or by multiple
antennas. Given a K × N received signal matrix Y, the sample covariance matrix is
defined as R ≜ N1 YYH and λ1 ≥ · · · ≥ λk are its eigenvalues sorted in the decreasing
order.
Eigenvalue based detection techniques infer the presence of signal from eigenval-
ues λi . In particular, as already mentioned the detection technique which considers
the largest one (λ1 ) and compare it against the noise variance is known as RLRT
[149] and its test statistics is,

λ1
TRLRT ≜ . (4.215)
σv2
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 151

RLRT is a “semi-blind” algorithm, as it requires the exact knowledge of noise

variance and is considered to be asymptotically optimum test in this setting[46].
Here, we analyze in detail the RLRT method. However, the results can be
extended to the other eigenvalue methods as well. To analyze the RLRT performance,
it is necessary to express the test statistics pdf for the case of unknown PU traffic.
The following theorem computes the RLRT decision statistic pdf under both the
hypotheses H0 and H1 using the PU traffic characterization presented in Sub-Section
4.3.1.

Theorem 7. Given a multi-antenna sensing unit with K receive antennas, N received

samples in each slot and a random PU traffic with geometrically distributed free and
busy state duration, let c = K/N , Ns a independent parameter and define:

K −1 K −1
    
Ns 2 Ns
µ1 (Ns ) = Kρ + 1 1+ , σ1 (Ns ) = 2 (Kρ + 1) 1 − (4.216)
N Ns Kρ N Ns K 2 ρ 2
√ 2 √  1 1/3
 
−2/3
 
µN,K = 1 + c , σN,K = N 1+ c 1+ √ . (4.217)
c

Then, the pdfs of RLRT decision statistic under H0 and H1 are given by (4.218) and
(4.219) respectively:

N−1
x − µN,K
 
fTRLRT |H0 (x) = PSS |H0 fTW 2 + PT S |H0 ∑ PD (d0 ) fD (x, N − d0 ),
0 (4.218)
σN,K d0 =1
N−1
fTRLRT |H1 (x) = PSS |H1 fD (x, N) + PT S |H1 ∑ PD (d1 ) fD (x, d1 ).
1 (4.219)
d1 =1

where,
(
K−1
fN (µ1 (d), σ12 (d)) if, d > K2ρ 2
,
fD (x, d) = . (4.220)
fTW 2 (µN,K , σN,K ) otherwise.

In (4.220), fN (µ1 (d), σ12 (d)) denotea a Gaussian pdf with mean µ1 (Ns ) and vari-
ance σ12 (Ns ) provided in (4.216) at Ns = d. Next, fTW 2 (µN,K , σN,K ) is the pdf of
Tracy-Widom distribution of order 2 with parameters µN,K and σN,K provided in
(4.217).

Proof. As noted from Sub-Section 4.3, the nature of the received signal matrix is
different for the SS sensing slot and TS sensing slot. Under null hypothesis H0 ,
152 Cognitive Radio: Spectrum Sensing

received sample covariance matrix R|H0 can be decomposed as a probabilistic sum

of RSS |H0 and RT S |H0 in the following way,

R|H0 = PSS |H0 RSS |H0 + PT S |H0 RT S |H0 . (4.221)

Since sensing slots are independent from each other, we treat each covariance matrix
in (4.221) independently. Given an SS sensing slot under null hypothesis, all the
received samples yk (n) are homogeneous in nature comprising the i.i.d. Gaussian
noise samples with mean zero and variance σv2 . Thus, the sample covariance matrix
RSS |H0 follows a Wishart distribution whose largest eigenvalue normalized by noise
variance can be expressed by a Tracy-Widom distribution of second order [46, 20].

λ1SS |H0 x − µN,K

 
= fTW 2 , (4.222)
σv2 σN,K

where µN,K and σN,K are given in (4.217).

Next, given a TS sensing slot under null hypothesis, all the received samples
yk (n) are not homogeneous in nature. To provide a better understanding, we express
the covariance matrix in a TS sensing slot under H0 as,

RT S |H0 = RS (N − D0 ) + RN (D0 ), (4.223)

where,

1
RS (N − D0 ) ≜ S SH , (4.224)
N − D0 [K,N−D0 ] [K,N−D0 ]
1
RN (D0 ) ≜ V VH , (4.225)
D0 [K,D0 ] [K,D0 ]

are the partial covariance matrices built respectively from signal-plus-noise and
only-noise samples. RS (N − D0 ) is a standard spiked population covariance matrix
of rank-1 and RN (D0 ) is Wishart matrix. The largest eigenvalue of RN (D0 ) is negli-
gible compared to the largest eigenvalue of RS (N − D0 ) given a signal identifiability
condition is met [138]. It is known that the fluctuation of the largest eigenvalue of
a rank-1 spiked population matrix normalized by the noise variance are asymptoti-
cally Gaussian [46, 139] if the signal identifiability condition is met, otherwise its
distribution is again a Tracy-Widom of order 2.
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 153

λ1T S |H0
= fD (x, (N − D0 )) (4.226)
σv2

Using the results from (4.222) and (4.226), the RLRT decision statistic under null
hypothesis can be written as:
λ1 |H0
TRLRT |H0 = (4.227)
σv2
λ1SS |H0 λ1T S |H0
= pSS |H0 + pT S |H (4.228)
σv2 0
σv2
t − µN,K
 
= pSS |H0 fTW + pT S |H0 fD (x, N − D0 ). (4.229)
σN,K

Using the fact that D0 is a random variable distributed as in (4.188), we obtain

the final distribution of the decision statistic of RLRT test under null hypothesis as
in (4.218).
We consider now the case when the PU signal is present (hypothesis H1 ). In this
case, an error is made if the presence of PU signal is not detected. Under alternate
hypothesis H1 , the received sample covariance matrix R|H1 can be decomposed as
the probabilistic sum of RSS |H1 and RT S |H1 .

R|H1 = pSS |H1 RSS |H1 + pT S |H1 RT S |H1 . (4.230)

Since RSS |H1 is a standard spiked population covariance matrix of rank-1, the
distribution of the largest eigenvalue normalized by the noise variance in a SS
sensing slot under H1 can be approximated as [46, 139],

λ1SS |H1
= fD (x, N). (4.231)
σv2

Using the same line of reasoning as in H0 , we get,

λ1T S |H1
= fD (x, D1 ). (4.232)
σv2
154 Cognitive Radio: Spectrum Sensing

Using (4.231) and (4.232), the distribution of the RLRT decision statistic under
alternate hypothesis can be written as,

λ1 |H1
TRLRT |H1 = (4.233)
σv2
λ1SS |H1 λ1T S |H1
= pSS |H1 + pT S |H (4.234)
σv2 1
σv2
= pSS |H1 fD (x, N) + pT S |H1 fD (x, D1 ). (4.235)

Incorporating the pmf of D1 (derived in (4.189)) in (4.235) yields (4.219).

A. Probability of False Alarm: Given the pdf of the decision statistic in (4.218),
we can now compute the false alarm probability. Under H0 , the PU is in free state
at the end of the sensing interval, but the decision statistic is erroneously above the
threshold τ and the PU signal is declared present. For the probability of false-alarm
PFa , the following Corollary of Theorem 7 holds.

Corollary 8. The false-alarm probability of the RLRT test under unknown PU traffic
and complex signal space scenario is given by:

Z +∞
PFa = P(TRLRT |H0 ≥ τ) ≡ fTRLRT |H0 (x)dx. (4.236)
τ

B. Probability of Detection: Given the pdf of the decision statistic in (4.219), we

can now compute the detection probability. Under H1 , i.e., the PU is in busy state at
the end of the sensing interval. Under this scenario, if the decision statistic is above
the threshold, the PU signal is declared present. The following Corollary of Theorem
7 defines the probability of detection PD .

Corollary 9. The detection probability of the RLRT test under unknown PU traffic
and the complex signal space scenario is given by:

Z +∞
PD = P(TRLRT |H1 ≥ τ) ≡ fTRLRT |H1 (x)dx. (4.237)
τ
4.3 Semi-blind sensing techniques under unknown Primary User Traffic 155

Simulation Results

In this Sub-Section, the effect of PU traffic on the RLRT detection method is analyzed
based on the the traffic model developed in Sub-Section 4.3.1. The length of the free
and busy periods of the PU traffic are measured in terms of the discrete number of
samples where each of them has Geometric distribution with mean parameters M f
and Mb , respectively. Under multiple antenna sensing scenario, the average SNR at
σ 2 ∥h∥2
the receiver is defined as, ρ = sKσ 2 , where ||.|| denotes the Euclidean norm. The
v
analytical expressions derived are validated via numerical simulations.
In Fig. 4.34, the sensing performance of RLRT under unknown PU traffic is
compared with the ideal RLRT performance. It can be well understood that the con-
ventional model with perfect synchronization of the PU-SU sensing slots performs
better than the one with unknown PU traffic. The perfect match of the theoretical
and Monte-Carlo numerical curves validates the derived analytical expressions of
PFa and PD . The ROC performance of RLRT in the considered PU traffic model for
different PU traffic parameters is presented in Figure 4.34a. The sensing performance
degrades significantly when the mean lengths of busy and free periods are compa-
rable with the length of the sensing interval or in a few multiples of it. However,
an improvement in the sensing performance can be seen if the length of the mean
parameters M f and Mb is increased. In Figure 4.34b we present the missed detection
probability (PMd ) as a function of SNR. From this figure, we can see that, for a given
PU traffic parameters, increasing the SNR improves the sensing performance for
certain lower range of SNR. However, in contrast to RLRT sensing performance
under known PU traffic , the RLRT sensing performance under unknown PU traffic
levels to some point (1 > PMd >> 0) above certain SNR. This is due to the effect of
the PU signal-plus-noise and only-noise samples mixing in the TS sensing slot.
In Fig. 4.35, the RLRT sensing performance is plotted as a function of sensing
parameters N and K. The variation of the sensing performance of RLRT detector
for different number of receiving antennas (K) is plotted in Figure 4.35a. It can
be observed that, unlike the rapid increase in sensing performance (decrease in
missed-detection probability) with the increasing number of receiving antennas under
synchronized PU-SU sensing slot scenario, the RLRT sensing performance under
unknown PU traffic is almost constant even if we increase the number of antennas.
During a TS sensing slot, from each receiving antenna the received signal samples are
the mixture of pure noise samples and samples with both noise and PU signal. Thus,
156 Cognitive Radio: Spectrum Sensing

1
Mb , Mf ↑

0.8
Known PU Traffic
Numerical
0.6
PD

Unknown PU Traffic
Theoretical
0.4 Mf = Mb = 800
Mf = Mb = 300
Mf = Mb = 100
0.2

0
0 0.2 0.4 0.6 0.8 1
PF

(a) ROC performance. Parameters: N = 50,

K = 4 and SNR = −6 dB
100

10−1

Known PU Traffic
PM d

10−2 N = 50
N = 100

Unknown PU Traffic
10−3 Theoretical
N = 200
N = 600
N = 1000
10−4
−10 −5 0 5 10
SNR (dB)

(b) Probability of Missed detection vs. SNR,

Parameters: K = 8, PFa = 0.01 and M f =
Mb = 3000.

Fig. 4.34 Sensing performance of RLRT under unknown PU traffic.

4.3 Semi-blind sensing techniques under unknown Primary User Traffic 157

even if we use multiple antennas, the nature of the received signal doesn’t change
much. This is the reason why the sensing performance improvement is suppressed
by the unknown PU traffic (more specifically, the TS sensing performance) when
the length of the free and busy periods of PU traffic are quite small (a few multiples
of the length of the sensing window). Furthermore, we present in Figure 4.35b, the
numerical simulation of detection probability (PD ) as a function of sensing window
(N). Note that, unlike RLRT detection probability under known PU traffic which
monotonically increases indefinitely until ‘PD = 1’ with increasing length of sensing
window, the detection probability of RLRT under unknown PU traffic do not have a
monotonic property as a function of the sensing window length.
158 Cognitive Radio: Spectrum Sensing

100

10−1

Known PU Traffic
PM d

−2 SNR = −8 dB
10
SNR = −10 dB

Unknown PU Traffic
10−3 Theoretical
SNR = −12 dB
SNR = −8 dB
SNR = −4 dB
10−4
4 6 8 10 12
K

(a) Prob. of Missed Detection vs. the number

of receiving antennas (K). Parameters: N =
100, PF = 0.01 and M f = Mb = 1500
1
Mb = Mf = 1500
Mb = Mf = 5000

0.8
PD

0.6

0.4

0 500 1,000 1,500 2,000 2,500 3,000

N

(b) Detection Probability for increasing N. Pa-

rameters: SNR = −10 dB, K = 8 and PF =
0.01

Fig. 4.35 RLRT under unknown PU traffic sensing performance comparison for different
sensing parameters (N and K).
Chapter 5

Low Density Spreading (LDS)

Multiplexing for NOMA

Low Density Spreading (LDS) Multiplexing for Non-orthogonal Multiple Access

5.1 Introduction

5G wireless networks are expected to support very diverse services - from very low
latency to very high delay tolerant, and from very small to very large packets in
different applications. An important 5G requirement is will be to support massive
connectivity with a large number of devices such as smart phones, tablets and
machines. The current LTE [43] system is not able to efficiently support massive
connectivity, especially in uplink. While fundamental research for 5G is now well
under way [1, 10], the question what actually makes a 5G system and what are the
drivers is still open and part of intensive discussions.
A new flexible and adaptable interface is proposed to meet the complex and
diverse application requirements of air interface technology. New air interface con-
sists of building blocks and configuration mechanisms such as adaptive waveforms,
adaptive protocols, adaptive frame structures, adaptive coding and modulation fam-
ily, and adaptive multiple access schemes. With these blocks and mechanisms, the
air interface will be able to accommodate the future wide variety of user services,
spectrum bands and traffic levels. The quest of improving the spectral efficiency has
160 Low Density Spreading (LDS) Multiplexing for NOMA

been regarded as the most important but yet challenging task in the design of future
wireless communication system due to the fact that the rapid growth of multimedia
services, such as interactive game and television applications, cannot be coped with
the scarce radio frequency (RF) spectrum resources. As a result, multiple access
schemes will play a critical role in providing the increasing demand in services for
future terminals and applications.
Multiple access (MA) technique is a major building block of cellular systems.
Through the MA technique, the users can simultaneously access the physical medium
and share the finite resources of the system, such as spectrum, time and power. Due to
the rapid growth in data demand in mobile communications, there has been extensive
research to improve the efficiency of cellular systems. A significant part of this effort
focuses on developing and optimizing the MA techniques. As a result, many MA
techniques have been proposed systematically over the years, and some of these MA
techniques are already been adopted in the cellular system standards such as Time
Division Multiple Access (TDMA), Frequency Division Multiple Access (FDMA),
Orthogonal Frequency Division Multiple Access (OFDMA) and Code Division
Multiple Access (CDMA). There are many factors that determine the efficiency of
MA techniques, including spectral efficiency, low complexity implementation as
well low envelope fluctuations.
OFDM and OFDMA are the modulation technique and the multiple access strat-
egy adopted in LTE fourth genereation (4G) cellular network standards, respectively
[83]. OFDM and OFDMA succeeded CDMA, employed in third generation (3G)
networks for several reasons, such as the ease of implementation of both transmitter
and receiver thanks to the use of FFT and inverse FFT (IFFT) blocks; the ability to
counteract multipath distortion, the orthogonality of subcarriers which eliminates
intercell interference; the possibility of adapting the transmitted power and the mod-
ulation cardinality; and the ease of integration with multi-antenna hardware, both at
the transmitter and receiver. Nonetheless, despite such a pool of positive properties,
OFDM/ OFDMA are not exempt of defects, and their adoption in the forthcoming
generation of wireless networks is not taken for granted. Indeed, the spectral effi-
ciency of OFDM is limited by the need of a CP and by its large sidelobes (which
require some null guard tones at the spectrum edges). OFDM signals may exhibit
large peak-to-average-power ratio values [136] and the impossibility of having strict
frequency synchronization among subcarriers makes OFDM and OFDMA not really
orthogonal techniques. In particular, synchronization is a key issue in the uplink of a
5.1 Introduction 161

cellular network wherein different mobile terminals transmit separately [127], and,
also, in the downlink when base station coordination is used [93, 68]. For instance,
with regard to the spectral efficiency loss of side lobes and the CP, in an LTE system
operating at 10 MHz bandwidth, only 9 MHz of the band is used. In addition, the
loss of the CP is around 7%, so the accumulated loss totals at 16%. These drawbacks,
which invalidate many of the above-mentioned OFDM/OFDMA advantages, form
the basis of an open and intense debate on what the modulation format and multiple
access strategy should be in next-generation cellular networks.
The limitations of OFDM-based waveforms were identified as research topics
for future 5G waveforms. One aspect is the requirement for much shorter latency
to enable new services and applications like autonomous driving, that demands an
ultra-low latency and a highly resilient communication link. Another approach is
to make the cyclic prefix optional and work with shorter symbol durations. All this
has led to several candidate waveforms, such as Generalized frequency division
multiplexing (GFDM), Filter bank multi-carrier (FBMC), Universal filtered multi-
carrier (UFMC), Filtered OFDM (f-OFDM), etc [142]. The performance of these
waveform candidates is currently being analyzed and evaluated. At the same time
new multiple access schemes are also being researched, including Sparse Code
Multiple Access (SCMA) [134], Non-Orthogonal Multiple Access (NOMA) [150]
and Resource Spread Multiple Access (RSMA) [142].
However, it has not been decided yet which of these waveform candidates or
multiple access schemes will be utilized in a future 5G system. It is up to the
standardization committees to evaluate different proposal to the IMT-2020 group
within the ITU.
Moreover, 3GPP has already started the first study items for the choice of the
new Radio Access Technology (RAT) for next generation 5G mobile networks.
Currently, there is a lot of interest in Non-Orthogonal Multiple Access (NOMA)
schemes, which allow to distribute resources among users without a strict orthog-
onal division. For non-orthogonal MA techniques, all the users share the entire
signal dimension, and there is a Multiple Access Interference (MAI). Thus, for
non-orthogonal transmission, more complicated receivers are required to deal with
the MAI comparing to orthogonal transmission. NOMA is more practical in the
uplink scenario because the base station can afford the Multi-user Detection (MUD)
complexity. On the other hand, for downlink, orthogonal MA is more suitable due
162 Low Density Spreading (LDS) Multiplexing for NOMA

to the limited processing power at the user equipment. Many non-orthogonal MA

techniques have been overlooked due to the implementation complexity. Clearly,
the recent advancements in signal processing have opened up new possibilities for
developing more sophisticated and efficient MA techniques. Thus, more advanced
MA techniques have been proposed lately. However, in order to adopt these new MA
techniques in the mobile communication systems, many challenges and opportunities
need to be studied.
To better understand the problem, let us consider 4G LTE, which uses an OFDM
(Orthogonal Frequency Division Multiplexing) radio access. In LTE, a resource
block is made of 12 OFDM carriers (corresponding to 180 kHz of band), which are
allocated for a TTI (Transmission Time Interval) 1ms long. In LTE, different users
receive disjoints sets of resource blocks, so they transmit on disjoint bands. Instead,
in a non-orthogonal approach, the same resource block may be jointly assigned to
different users. Therefore, two or more users may transmit simultaneously on the
same band.
Examples of non-orthogonal schemes under discussion for 5G access schemes
are super-position modulation [89], Low Density Signature-OFDM [91] and Sparse
Code Multiple Access [134]. In this chapter, we focus on Low Density Signature
schemes, which look very promising thanks to their simplicity.

5.2 Multiplexing with Low Density Spreading

Low Density Spreading is one of the promising multiplexing scheme for NOMA
techniques proposed for 5G multiple access [171, 11, 12]. In LDS-Multiplexing, low
density spreading codes are used for transmission, where each user spreads its data
on a small set of sub-carriers. Due to the low density structure, every data symbol
will be spread over a small subset of sub-carriers (effective processing gain) and also
every sub-carrier will only be used by small subset of data symbols that could belong
to different users. There is no exclusivity in the sub-carriers allocation and more than
one user can share each sub-carrier. Therefore, each user transmission on a given
sub-carrier will only be interfered by a small number of other user’s transmission.
The LDS structure can be captured by a low density graph thus similar to application
of LDS for CDMA system, thus the detection of LDS symbols could be based on
the MPA presented in [90]. The LDS system can be understood as a system which
5.2 Multiplexing with Low Density Spreading 163

applies LDS as multiplexing technique together with any orthogonal multi-carrier

modulation.
The main features of the LDS scheme can be summarized as follow [171]

• At each sub-carrier, a user will have relativity small number of interferers

comparing to the total number of users. Consequently, the search space will
be smaller and more complex MUD techniques can be implemented.

• Higher Signal to Interference plus Noise Ratio (SINR) can be achieved at each
sub-carrier, which results in reliable detection process.

• Each user will experience interference from different users at different sub-
carriers, which results in interference diversity by avoiding strong interferers
to destroy the signal of a user on all the sub-carriers.

• Belief propagation based MUD can be implemented with linear complexity in

the number of sub-carriers [126].

Spreading sequences of the low-density type (containing many zeros) were con-
ceptually introduced in [38] for the purpose of allowing low-complexity multiuser
detection, while for the multicarrier system they were introduced in [90]. These
sequences, when designed in accordance with appropriate rules [126], allow appli-
cation of a belief propagation algorithm at the receiver. This was shown to yield
promising performance for overloaded systems - systems with more spreading se-
quences than chips for BPSK modulated data. Later, the concept of LDS-OFDM was
introduced as an uplink multi-carrier multiple access scheme and a close-to-optimum
MUD receiver performance was analyzed using EXIT charts [91]. The Performance
comparison with other existing multiple access techniques like OFDMA and SC-
FDMA were investigated together with their suitability for next generation cellular
system in [144, 13].
The focus of above literature was on the conceptual multiuser receiver structure
rather than on the particular choice of the set of spreading sequences. Bypassing de-
tails of spreading sequence synthesis, low-density sequences were found by trial and
error. Moreover, user-specific pseudo-random phases were applied in the evaluation
of the receiver structure [90].
In this chapter we address the design of the user’s spreading sequences. We
refer to them as signatures sets assuming they have the low-density structure of [90].
164 Low Density Spreading (LDS) Multiplexing for NOMA

Instead of relying on trial-and-error searches and on pseudo-random design steps,

we take a structured approach towards the design of spreading sequences. We focus
on their performance on the AWGN channel which is a good model for downlink
multiple access systems operating in a flat, slowly fading radio environment. In the
downlink scenario, a single physical transmitter can simultaneously transmit multiple
information streams, each information stream intended to a separate user, i.e., to a
separate receiver. Thus, from a single user reception performance perspective, the
situation is equivalent to have multiple users transmitting simultaneously.
We investigate here a signature matrix of the low-density type, which has non-
zero entries only in positions indicated by a low-density spreading binary-matrix
[90]. The design of this kind of matrices has been subject of extensive research over
the past years in a completely different context (the design of LDPC codes [114])
than the one addressed here (multiple-access signatures). Binary matrices designed
for good error-correction performance usually have a large girth-value (defined as
the length of the smallest cycle in the factor graph associated with the binary matrix).
We anticipate that matrices designed according to these principles also are beneficial
when used for signature design and we therefore adopt these binary matrices as
the basis for further signature design. We assume perfect synchronization of the
transmitted signatures, typically accomplished when any orthogonal modulation
scheme is the carrier transmission technology for the transmitted chip sequences.
Demodulation is done by the near-optimally belief-propagation structure of [90]. We
evaluate the performance of our sequence constructions and we show that they have
beneficial properties in terms of their distance spectrum.

5.3 System Model

We consider a NOMA Scheme in order to allow N simultaneously users to share K

orthogonal carriers where N > K. We define the ratio γ = NK as the overload factor
of the scheme. A simple factor graph representation of the LDS spreading system is
shown in Figure 5.1.
Let us denote by x, a complex symbol vector of length
N generated by N users,
x1 ··· ··· xN
x = (x1 , · · · , xn , · · · , xN ) xn ∈ C, (5.1)

+ + + +
s1 ··· ··· sK

Fig. 5.1 Factor Graph repre-

sentation of LDS spreading
5.3 System Model 165

and K complex symbols transmitted on the K tones by s,

s = (s1 , · · · , sk , · · · , sK ) sk ∈ C. (5.2)

The problem of distributing N > K symbols can be ap-

proached by transmitting on each tone a linear combina-
tion of the N user symbols.

N
si = ∑ xnαnk xk ∈ C , (5.3)
n=1

where αnk is a linear combination coefficient. This equation can be written by using
vector representation as,

s = xA. (5.4)

where the N × K matrix A known as Spreading Matrix, contains the linear com-
bination coefficients as its columns, known as spreading sequence, i.e., A =
[a1 · · · ak · · · aK ] and ak = [α1k · · · αnk · · · αNk ]T . The spreading sequences are assigned
uniquely to each user and are completely shared among the transmitter/receiver.
In LDS multiplexing, a small number of elements of a spreading sequences are
non zero allowing each user to spread its data over small number of tones. The
relationship between the LDS matrix A and the user spreading can be better explained
using following elements:

• Binary Matrix: A binary matrix B is given by B = [b1 , · · · , bK ] where b ∈ BK

2
is a binary vector for each tone. The nth element of b is denoted by bnk and is
given by, (
0, if αnk = 0
bnk = ; n = 1, . . . , N. (5.5)
1, if αnk ̸= 0

The set of positions of 1’s in the nth row denotes the set of output tones over
which user n spread its data, while the set of positions of 1’s in the kth column
represents the set of users that can contribute with their data at the kth tone.
Note that the 1′ s components of the binary matrix B represents the non zero
components of the matrix A and the non zero components of matrix A can
take any complex values.
166 Low Density Spreading (LDS) Multiplexing for NOMA

• Row Weight and Column Weight: Let us define the cardinality of the set of
tones over which any user spreads its data as wr , known as row weight, and the
cardinality of the set of users that can contribute their data at any tone as wc ,
known as column weight. If the parameters wr and wc are constant for all users
and tones respectively, then the resulting Spreading Matrix is known as regular
spreading matrix. In the following, we discuss the problem of constructing
optimum performing Spreading Matrix A. For simplicity, focus on a regular
Spreading Matrix for the rest of the discussion.

• Column Correlation Factor: Let us define a variable w f , which denotes the

maximum correlation among the non-zero positions in two columns of the
binary matrix and is given by,
n
w f = max ∑ bnk1 bnk2 . (5.6)
k1 ,k2 n=1
k1 ̸=k2

• Non-zero elements: We consider non-zero elements of the spreading matrix

taken from a small and finite complex-valued constellation v = {vk } referred
to as signature constellation. Signature constellation are taken from the ring
of constant magnitude in a 2-D complex state space. A constant amplitude is
considered so as to keep the power of each input symbols constant during a
linear combination. Each output symbol is the linear combination of few input
symbols for each entry αkn in A,

αkn ∈ v ∪ {0}. (5.7)

5.4 Calculation of average energy per bit (Eb)

It is clear that the output symbol vector is the

complex linear combination of the input sym-
bols. The output energy per bit varies depending x1 x2 x3 xN
···
on the overloading factor, the number of input
wr
symbols linearly combined to form an output and
energy per bit at the input modulated symbol. In wc
order to derive the upper bound on minimum + + ··· +
s1 s2 sK

Fig. 5.2 Regular LDS spreading sys-

tem
5.4 Calculation of average energy per bit (Eb ) 167

Euclidean distance achieved by these schemes,

we first derive the expression for the output av-
erage energy per bit for a regular LDS spreading
system employing BPSK as input modulation
scheme.

Proposition 7. For an LDS system with system parameters (N, K, wr , wc ) employing

BPSK as an input modulation scheme, the output average energy is given by,

Kwc
Eb = Ebi ⇒ wr × Ebi . (5.8)
N

Proof. Consider a LDS system with system parameters (N, K, wr , wc ) employing

√ √
BPSK as an input modulation scheme, xn ∈ {+ Ebi , − Ebi } with input energy per
bit Ebi .
The average energy at one of the output symbol can be calculated as the sum
of the energy of every possible output symbol normalized by their number. If wc
different inputs are linearly combined to form an output symbol sk (see Figure 5.2),
wc
it means there are 2wc different possible sk : sk ∈ {s1k , · · · , s2k }. If wc inputs linearly
sums to an output sk , the remaining N − wc inputs can give 2N−wc possible input
combinations that can give the same output sk , thus the average energy of the output
symbol is given by:

wc
2N−wc ∑2i=1 |sik |2
EO (k) = . (5.9)
2N
Since there are K output symbols and N inputs, the output average energy per bit
can be written as

1 K
Eb = ∑ EO(k). (5.10)
N k=1

Since wc is constant for all the output symbols sk , the average output energy per
symbol is constant, thus:
168 Low Density Spreading (LDS) Multiplexing for NOMA

wc
K2N−wc 2
Eb = ∑ |sik |2.
N2N i=1
(5.11)

wc
It can easily be found that the sum ∑2i=1 |sik |2 is equal to wc × Ebi × 2wc .
This leads to a simplified expression of the Average Output energy per bit as in
(5.8).

Next, the following proposition derives the upper bound on

minimum euclidean distance using the above derived average
output energy per bit. xn
wr
Proposition 8. For an LDS system with system parameters
(N, K, wr , wc ) employing BPSK as an input modulation scheme,
the upper bound on minimum Euclidean distance is given by, + +
sk1 sk2

2
dmin
|max | = 4. (5.12) Fig. 5.3 Euclidean
Eb distance relation-
ship
Proof. In a considered LDS system with parameters (N, K, wc , wr ),
the least possible Hamming distance among two different binary
input sequence is 1. Since each input is connected to wr number
of output tones for a regular LDS spreading system (See Figure 5.3), the difference in
√
one bit corresponds to an Euclidean distance of at most 2 × Ebi in each connected
output tones. Thus, the maximum squared Euclidean distance we can achieve is,

2
dmin |max = wr × 4Ebi . (5.13)

Thus, the upper bound on the normalized minimum Euclidean distance expressed
in (5.12) can be easily obtained using (5.13) and (5.8).
5.5 Spreading Matrix Design 169

5.5 Spreading Matrix Design

The performance of the LDS system directly depends on the Euclidean distance
profile of the considered spreading matrix A for the given LDS system parameters. In
essence, the BER performance depends mostly on the minimum Euclidean distance
provided by the spreading matrix and partly on the multiplicity of other distances
(distance distribution).
The design of a spreading matrix A with parameters (N, K, wc , wr ) involves the
separate design of three different components, which are:

• Binary Matrix

• Non-zero elements

• Distribution of non-zero elements in a binary matrix

5.5.1 Binary Matrix Design

Binary matrix serves as a way to relate the input symbol sets with output tones. The
LDS binary matrix B must be designed to meet the system parameters, i.e., N, K, wr
and wc . Like for LDPC matrices, a factor graph can be associated to a binary matrix
B by constructing the edges connecting factor (i.e. column) nodes and variable (i.e.
row) nodes. In a factor graph a loop (or cycle) is a closed path with no repeated
nodes, and must therefore be of even length. There is at most one edge between
any two nodes, and so the shortest length a loop can have is 4, the next largest is 6,
and so on. We refer to such loops as 4-loops, 6-loops, etc., and in general a loop of
length n is an n−loop. The girth of a graph is the length of the shortest loop. The
main problem with a loop is that the value of an incorrect bit is propagated back to
itself, effectively reinforcing itself and resisting the efforts of the algorithm to correct
it. Thus, an important consideration in designing a binary matrix is to avoid small
loops. In this work, we introduce a novel design concept for the construction of a
binary matrix, which completely avoids 4-loops.
Given a binary matrix B, it is very simple to check the presence of 4-loops by
using the column correlation factor w f defined in Section 5.3. As we know that the
column position of a binary matrix corresponds to a factor node in a factor graph, this
170 Low Density Spreading (LDS) Multiplexing for NOMA

means that the correlation in the non-zero position of a two column pair represents
the edge of a factor graph. Thus, considering any two columns of a binary matrix,
a column correlation factor greater than one confirms the presence of 4-loops in
the binary matrix. It is important to note that it may not be possible to construct
the binary matrix B for all the matrix size and overloading factor with a column
correlation factor w f = 1. The following relation, which is derived based on the
t-design [115] concept, helps us to check the possibility of constructing the binary
matrix with w f = 1 for given LDS system parameters.
j k
For an LDS system with system parameters (N, K, wr , wc ), if wK−1r −1
× wKr ≥ N,
there exists at least one binary matrix design which has a column correlation factor
w f = 1. Design guidelines are provided for every overloading factor in next sections.

5.5.2 Design of non-zero symbols

The concept behind Low Density Spreading is to uniquely distribute the spreading
codes which contains less number of non-zero elements allowing each user to spread
its data over a small number of chips (tones, carriers, channels). In order to uniquely
decode the linearly combined output symbols, the one-to-one mapping relationship
between input and output is a must. This necessitates a careful design of non-zero
symbols of a spreading matrix. In addition, the choice of non-zero symbols should
not change the overall magnitude of the input symbols while spreading. Thus, it is
obvious that the non-zero symbols must be chosen from the unit circle of a complex
2-dimensional space which can be written in a complex exponential form as,

vk = exp(θk i). (5.14)

We distribute these symbols in the position of 1’s of each column in the binary
matrix. Let us consider the set of unique elements which are to be distributed as
signature constellation set. One of the easiest way to ensure a one to one relationship
between input and output of a LDS system is to ensure one-to-one relationship
among the inputs corresponding to single channel outputs [167].
As presented before, the objective of this work is to find a design rule for the
construction of a complete spreading matrix which ensures the achievement of upper
bound on the minimum normalized Euclidean distance. In the following we present
5.5 Spreading Matrix Design 171

a theorem that proves that for wc > 2, there is no such unitary magnitude signature
constellation for a single channel output that gives a one to one mapping matching
the upper bound on the minimum normalized Euclidean distance.

Lemma 3. Given an LDS binary matrix with parameters N, K = 1, wc ,

- there exists a signature constellation for wc = 2 to obtain the upper bound

on the normalized Euclidean distance, and the angle separation among the
signature constellation points should fall in the following range.

− 2π/3 ≤ θ1 − θ2 ≤ −π/3, −5π/3 ≤ θ1 − θ2 ≤ −4π/3,

π/3 ≤ θ1 − θ2 ≤ 2π/3, 4π/3 ≤ θ1 − θ2 ≤ 5π/3, (5.15)

where θ1 and θ2 are the angles of the signature constellation points.

- no signature constellation exists that gives an upper bound on the normalized

minimum Euclidean distance for wc ≥ 3

Proof. Let (x1 , · · · , xwc ) and (y1 , · · · , ywc ) be two input sets that correspond to non-
zero positions of the considered column of a binary matrix. Single channel outputs
corresponding to these input sets can be written as,

O1 = v1 x1 + v2 x2 + · · · + vwc xwc (5.16)

O2 = v1 y1 + v2 y2 + · · · + vwc ywc . (5.17)

A first and a simple case would be to consider wc = 2, thus,

O1 = v1 x1 + v2 x2 (5.18)
O2 = v1 y1 + v2 y2 . (5.19)

Now, the squared Euclidean distance among these outputs can be written as,

(O1 − O2 )2 = |v1 (x1 − y1 ) + v2 (x2 − y2 )|2 . (5.20)

 172 Low Density Spreading (LDS) Multiplexing for NOMA

Table 5.1 Relationship of A and B with input x and y.

x1 x2 y1 y2 A B
-1 -1 -1 -1 0 0
-1 +1 -1 +1 -2Ebi -2Ebi
+1 -1 +1 -1 2Ebi 2Ebi
+1 +1 +1 +1 0 0

Substituting the values of v1 and v2 with their exponential form as in (5.14), we have,

(O1 − O2 )2 = |(cos θ1 + i sin θ1 )(x1 − y1 ) + (cos θ2 + i sin θ2 )(x2 − y2 )|2 . (5.21)

Let A = x1 − y1 and B = x2 − y2 , then simplifying we get,

(O1 − O2 )2 = A2 + B2 + 2AB cos(θ1 − θ2 ). (5.22)

For BPSK modulated input symbols with bit energy of Ebi , Table 5.1 shows
the different values of A and B for different combinations of x1 , x2 and y1 , y2 . As
explained above, the squared Euclidean distance should be greater or equal to 4Ebi 2

for two distinct output tones:

A2 + B2 + 2AB cos(θ1 − θ2 ) ≥ 4Ebi

2
. (5.23)

Now, we define different cases based on the combination of these input sequences.

Case 1: x1 = x2 and y1 = y2 ( corresponds to the first and the last rows of the above
table)
When the inputs are same then, A and B becomes zero, thus (O1 − O2 ) = 0
which is a obvious result.

Case 2: If one of the input is equal and the other is different, x1 = x2 and y1 = −y2 OR
x1 = −x2 and y1 = y2
A = 0 and B = ±2Ebi OR A = ±2Ebi and B = 0, thus the inequality turns to

2 2
4Ebi = 4Ebi , (5.24)

which is also an obvious result.

 5.5 Spreading Matrix Design 173

Case 3: When both the inputs are different, x1 = −x2 and y1 = −y2 OR x1 = −x2 and
y1 = y2 , i.e., A = ±2Ebi and B = ±2Ebi thus the inequality turns to

2 2 2
4Ebi + 4Ebi + 2(±2Ebi )(±2Ebi ) cos(θ1 − θ2 ) ≥ 4Ebi . (5.25)

Here are two sub-cases,

Sub-case 1: When both A and B are positive or both negative, the product AB
is always positive, thus,

2 2 2
8Ebi + 8Ebi cos(θ1 − θ2 ) ≥ 4Ebi (5.26)
cos(θ1 − θ2 ) ≥ −0.5. (5.27)

Sub-case 2: When A and B are of opposite sign, the product AB is always

negative, thus,

2 2 2
8Ebi − 8Ebi cos(θ1 − θ2 ) ≥ 4Ebi (5.28)
− cos(θ1 − θ2 ) ≥ −0.5 (5.29)
cos(θ1 − θ2 ) ≤ 0.5. (5.30)

From (5.27) and (5.30),

− 0.5 ≤ cos(θ1 − θ2 ) ≤ 0.5 (5.31)

− 2π/3 ≤ θ1 − θ2 ≤ −π/3, −5π/3 ≤ θ1 − θ2 ≤ −4π/3,
π/3 ≤ θ1 − θ2 ≤ 2π/3, 4π/3 ≤ θ1 − θ2 ≤ 5π/3. (5.32)

Thus, if (θ1 − θ2 ) falls in the above range, then the squared Euclidean distance among
two different outputs is always greater or equal to 4Ebi 2 . This proves the first claim.

Next, we consider wc = 3.

O1 = v1 x1 + v2 x2 + v3 x3 (5.33)
O2 = v1 y1 + v2 y2 + v3 y3 . (5.34)
 174 Low Density Spreading (LDS) Multiplexing for NOMA

The squared Euclidean distance between O1 and O2 is,

(O1 − O2 )2 = |v1 (x1 − y1 ) + v2 (x2 − y2 ) + v3 (x3 − y3 )|2 . (5.35)

Substituting the values of v1 , v2 and v3 with their exponential form as in (5.14), we

have,

(O1 − O2 )2 = |(cos θ1 + i sin θ1 )(x1 − y1 ) + (cos θ2 + i sin θ2 )(x2 − y2 )

+ (cos θ3 + i sin θ3 )(x3 − y3 )|2 . (5.36)

Let A = x1 − y1 , B = x2 − y2 and C = x3 − y3 , then simplifying we get,

(O1 −O2 )2 = A2 +B2 +C2 +2AB cos(θ1 −θ2 )+2BC cos(θ2 −θ3 )+2AC cos(θ3 −θ1 ).
(5.37)
Our objective is to obtain the angle parameters relationship in order to obtain the
maximum possible normalized euclidean distance. The inequality to satisfy can be
written as
(O1 − O2 )2 ≥ 4Ebi
2
. (5.38)

We consider the values of A, B and C based on the above table. Based on the input
situation here we define three cases:

Case 1: When all the inputs are equal giving A = B = C = 0, then (O1 − O2 )2 = 0
which is obvious.

Case 2: When one of them is different and two of them are equal, i.e., A = ±2Ebi and
B = C = 0 OR B = ±2Ebi and A = C = 0 OR C = ±2Ebi B = A = 0, In each
case,
(O1 − O2 )2 = 4Ebi
2
, (5.39)

which satisfies (5.38)

Case 3: When two of them are different and one is equal, i.e., A = B = ±2Ebi and
C = 0 OR A = C = ±2Ebi and B = 0 OR B = C = ±2Ebi A = 0, In each case,

(O1 − O2 )2 = 8Ebi
2
+ 2(±2Ebi )(±2Ebi ) cos(θ1 − θ2 ), (5.40)
(O1 − O2 )2 = 8Ebi
2
+ 2(±2Ebi )(±2Ebi ) cos(θ1 − θ3 ), (5.41)
(O1 − O2 )2 = 8Ebi
2
+ 2(±2Ebi )(±2Ebi ) cos(θ2 − θ3 ). (5.42)
 5.5 Spreading Matrix Design 175

When conditioned with (S1 − S2 )2 ≥ 4Ebi

2 , we obtain

− 2π/3 ≤ θ1 − θ2 ≤ −π/3, −5π/3 ≤ θ1 − θ2 ≤ −4π/3,

π/3 ≤ θ1 − θ2 ≤ 2π/3, 4π/3 ≤ θ1 − θ2 ≤ 5π/3. (5.43)
− 2π/3 ≤ θ2 − θ3 ≤ −π/3, −5π/3 ≤ θ2 − θ3 ≤ −4π/3,
π/3 ≤ θ2 − θ3 ≤ 2π/3, 4π/3 ≤ θ2 − θ3 ≤ 5π/3. (5.44)
− 2π/3 ≤ θ3 − θ1 ≤ −π/3, −5π/3 ≤ θ3 − θ1 ≤ −4π/3,
π/3 ≤ θ3 − θ1 ≤ 2π/3, 4π/3 ≤ θ3 − θ1 ≤ 5π/3. respectively. (5.45)

We know that the complex angle at the exponent is symmetric over a multiples
of 2π. Thus, to simplify the above range in (5.43), (5.44) and (5.45), let us
consider,
π π
θ1 − θ2 = and θ2 − θ3 = + ε. (5.46)
3 3

Since θ3 − θ1 = (−(θ2 − θ3 ) − (θ2 − θ3 ), we have,

θ3 − θ1 = −π/3 − π/3 − ε (5.47)

θ3 − θ1 = −2π/3 − ε. (5.48)

It can easily be shown that the only values of ε that satisfies above inequalities
are ε = 0, −π. It means that the values of θ1 − θ2 , θ2 − θ3 and θ3 − θ1 are
fixed and are given Table 5.2,

Case 4: When all the inputs are different, i.e., A = B = C = ±2Ebi .

Sub-Case 1: One is positive and two are negative, i.e., A = −2Ebi , B = −2Ebi and
C = +2Ebi OR A = −2Ebi , B = +2Ebi and C = −2Ebi OR A = +2Ebi ,
B = −2Ebi and C = −2Ebi .
Sub-Case 2: One is negative and two are positive, i.e., A = +2Ebi , B = +2Ebi and
C = −2Ebi OR A = +2Ebi , B = −2Ebi and C = +2Ebi OR A = −2Ebi ,
B = +2Ebi and C = +2Ebi
176 Low Density Spreading (LDS) Multiplexing for NOMA

Table 5.2 Set of angles satisfying (5.43), (5.44) and (5.45).

Sets θ1 − θ2 θ2 − θ3 θ3 − θ1
−π −π 2π
Set1 3 3 3
−π 2π −π
Set2 3 3 3
2π −π −π
Set3 3 3 3
π π −2π
Set4 3 3 3
π −2π π
Set5 3 3 3
−2π π π
Set6 3 3 3
−4π 2π 2π
Set7 3 3 3
2π −4π 2π
Set8 3 3 3
2π 2π −4π
Set9 3 3 3
−2π −2π 4π
Set10 3 3 3
−2π 4π −2π
Set11 3 3 3
4π −2π −2π
Set12 3 3 3

For both of these sub-cases we have,

(O1 − O2 )2 = 12Ebi
2 2
+ 8Ebi 2
cos(θ1 − θ2 ) − 8Ebi cos(θ2 − θ3 )
2
− 8Ebi cos(θ3 − θ1 ) (5.49)
(O1 − O2 )2 = 12Ebi
2 2
− 8Ebi 2
cos(θ1 − θ2 ) − 8Ebi cos(θ2 − θ3 )
2
+ 8Ebi cos(θ3 − θ1 ) (5.50)
2 2 2 2
(O1 − O2 ) = 12Ebi − 8Ebi cos(θ1 − θ2 ) + 8Ebi cos(θ2 − θ3 )
2
− 8Ebi cos(θ3 − θ1 ) respectively. (5.51)

After simplification we get,

(O1 − O2 )2 θ2 − θ3 θ3 − θ1 θ1 − θ2
     
2
= 4 − 32 sin sin cos OR
Ebi 2 2 2
(5.52)
θ2 − θ3 θ3 − θ1 θ1 − θ2
     
= 4 − 32 sin cos sin OR
2 2 2
(5.53)
θ2 − θ3 θ3 − θ1 θ1 − θ2
     
= 4 − 32 cos sin sin
2 2 2
(5.54)
 5.5 Spreading Matrix Design 177

Table 5.3 Inequality check of angle sets.

Sets θ1 − θ2 θ2 − θ3 θ3 − θ1 (5.55) (5.56) (5.57)

−π −π 2π √ √
Set1 ×
3
−π
3
2π
3
−π √ √
Set2 ×
3
2π
3
−π
3
−π √ √
Set3 ×
3 3 3
−2π √ √
Set4 π π
×
3 3
−2π
3 √ √
Set5 π π
×
3
−2π
3 3 √ √
Set6 π π
×
3
−4π
3
2π
3
2π √ √ √
Set7 3
2π
3
−4π
3
2π √ √ √
Set8 3
2π
3
2π
3
−4π √ √ √
Set9 3
−2π
3
−2π
3
4π √ √ √
Set10 3
−2π
3
4π
3
−2π √ √ √
Set11 3
4π
3
−2π
3
−2π √ √ √
Set12 3 3 3

Considering the expressions in (5.52), (5.53) and (5.54) which when

conditioned with (S1 − S2 )2 ≥ 4Ebi
2 , we obtain,

θ2 − θ3 θ3 − θ1 θ1 − θ2
     
sin sin cos ≤0 (5.55)
2 2 2
θ2 − θ3 θ3 − θ1 θ1 − θ2
     
sin cos sin ≤0 (5.56)
2 2 2
θ2 − θ3 θ3 − θ1 θ1 − θ2
     
cos sin sin ≤ 0. (5.57)
2 2 2

We can check easily if the above explained sets satisfy these inequalities
in Table 5.3,
Sub-case 3: All values are positive OR all are negative

(S1 − S2 )2 θ2 − θ3 θ3 − θ1 θ1 − θ2
     
2
= 4 + 32 cos cos cos .
Ebi 2 2 2
(5.58)

When (5.58) is conditioned with (S1 − S2 )2 ≥ 4Ebi

2 , we obtain

θ2 − θ3 θ3 − θ1 θ1 − θ2
     
cos cos cos ≥ 0. (5.59)
2 2 2
178 Low Density Spreading (LDS) Multiplexing for NOMA

Table 5.4 Set of difference angles satisfying inequalities presented in (5.55), (5.56), (5.57)
and (5.59)

Sets θ1 − θ2 θ2 − θ3 θ3 − θ1 (5.55) (5.56) (5.57) (5.59)

−π −π 2π √ √ √
Set1 ×
−π
3 3
2π
3
−π √ √ √
Set2 ×
3
2π −π
3 3
−π √ √ √
Set3 ×
3 3 3
−2π √ √ √
Set4 π π
×
3 3
−2π
3 √ √ √
Set5 π π
×
3
−2π
3 3 √ √ √
Set6 π π
×
3
−4π
3
2π
3
2π √ √ √
Set7 ×
3
2π
3
−4π
3
2π √ √ √
Set8 ×
3
2π
3
2π
3
−4π √ √ √
Set9 ×
3
−2π
3
−2π
3
4π √ √ √
Set10 ×
3
−2π
3
4π
3
−2π √ √ √
Set11 ×
3
4π
3
−2π
3
−2π √ √ √
Set12 3 3 3 ×

Again checking the inequality for the above explained sets in Table 5.3,
we get,

None of the sets satisfy all the inequalities thus proving that there is no such set
of angles that can achieve the upper bound on the normalized Euclidean distance.

5.5.3 Distribution of non-zero symbols in a binary matrix

It is clearly evident that an optimal signature constellation doesn’t exist for a single
channel output this means that a smart distribution rule should be designed in order
to obtain an optimal overall Euclidean distance for a given spreading matrix. This
distribution rule depends both on the arrangement of the binary matrix and the
signature constellation points, thus in next section we present in detail the unique
distribution rule for each considered spreading matrices type in next section.
5.6 Optimal Design 179

5.6 Optimal Design

In the previous section, we have shown that a unitary signature constellation doesn’t
exist for a single channel output. In this section we show that by carefully selecting
the signature constellation distribution rule, we can still achieve the upper bound on
normalized minimum Euclidean distance for the signature constellation set having
one of the angle separation among the list of sets presented in Table 5.4. Next, we
provide a complete guide to the design of spreading matrix for an LDS system to
achieve the overloading capacity of 150%.
To design a spreading matrix to achieve the upper bound on the normalized
minimum Euclidean distance, we start from the signature constellation. This gives
some hint for the design of binary matrix and a signature constellation distribution
rule.
We have wc = 3, for which we have already proved in previous section that an
optimal signature constellation doesn’t exist for single channel output. However, it
can be observed from Table 5.4 that there are some sets for which the number of
output overlapping is limited to a pair. Thus, in the following design, we choose a
signature constellation whose angle separations fall under one of the sets provided in
the Table 5.4. In the following we consider a signature constellation,

v1 = exp(iθ1 ), v2 = exp(iθ2 ) and v3 = exp(iθ3 ), (5.60)

such that
π π π
θ1 − θ2 = − , θ2 − θ3 = and θ3 − θ1 = − . (5.61)
3 3 3
For these angle parameters, next we discuss the nature of the single channel
output and the input conditions which lead to non-optimal Euclidean distances.
With these angle parameters, the squared Euclidean distance among single
channel outputs from two different inputs can be written as

(O1 −O2 )2 = A2 +B2 +C2 +2AB cos(θ1 −θ2 )+2BC cos(θ2 −θ3 )+2AB cos(θ3 −θ1 ).
(5.62)
By replacing the values of the angle separation parameters from (5.62) using (5.62),
we obtain
180 Low Density Spreading (LDS) Multiplexing for NOMA

(O1 − O2 )2 = A2 + B2 +C2 + AB + BC − AB. (5.63)

We know that when all the inputs are different, the magnitude of A, B and C are
equal. Thus, when the last three products in (5.63) provide a negative contribution to
the sum, then (O1 − O2)2 may end up with zero. This happens when,

A is positive, B is negative and C is positive OR

A is negative, B is positive and C is negative . (5.64)

Thus, for the above angle parameters, the single channel output is not one to one
since for the input conditions given by (5.64), the squared Euclidean distance is zero.
Next, we look forward to design a binary matrix which gives a one-to-one
I/O mapping for the above set of signature constellation. By looking at the input
conditions that gives a squared euclidean distance zero, we can rewrite the conditions
as:

• A (corresponding to θ1 ) and C (corresponding to θ3 ) should be non-zero and

have same sign.

• B (corresponding to θ2 ) should be non-zero and having a sign different from A

and C

With above criterion in hand, we choose to consider w f = 1. While keeping w f = 1,

a single column is connected with other three columns in only one position, in
this way all the columns are interlinked by w f = 1 position. Thus, in a worst case
scenario, when one of the column is giving a zero Euclidean distance, the overall
Euclidean distance is zero only when all the input symbols are different and satisfy
the condition (5.64) for all column outputs.
Thus, when the Euclidean distances from two separate output channels are
zero at the same time and if we can avoid the Euclidean distance from the third
output channel to be zero, then the final spreading matrix will have a one-to-one I/O
relationship. For this, we have to identify three columns such that the two non-zero
5.6 Optimal Design 181

positions are linked to other two columns at one position each. For example,
 
1 1 0 ···
0 0 1 · · ·
 
 
0
 1 1 · · ·

A = 1
 0 0 · · ·
 (5.65)
· · ·
 
1 0 1
 
0
 1 0 · · ·

.. .. .. ...
. . .

Now we distribute the signature constellation points as shown below,

 
v1 v2 0 ···
 0 0 v3 · · ·
 
 
 0 v3 v2
 · · ·
A = v3 0 0
 · · · (5.66)
· · ·
 
v2 0 v1
 
0 v 0
 1 · · ·
.. .. .. ..
. . . .

Next we check when these Euclidean distances of each column are zero.

• An Euclidean distance from first column is zero when A, B and C correspond-

ing to input positions I1 , I5 and I4 are with A = positive, B = negative and C =
positive.

• For second column, from first column it is fixed that B is positive corresponding
to I1 and v2 . Thus, an Euclidean distance from second column is zero when
A and C corresponding to input positions I3 and I6 respectively are with A =
negative and C = negative.

• For third columns the input positions corresponding to I3 and I5 are already
negative. Thus, if we replace these positions in the third column of the binary
matrix with v1 and v2 , when they are of the same sign they will never give a
zero Euclidean distance.
182 Low Density Spreading (LDS) Multiplexing for NOMA

 
I1 v1 (+) v2 (+) 0 ···
I2 0 0 v3 · · ·
 
 
I3
 0 v3 (−) v2 · · ·
I v (+)
A= 3 3 0 0 · · · (5.67)
· · ·
 
I4 v2 (−) 0 v1
 
I
5 0 v1 (−) 0 · · ·
.. .. .. ..
I. . . . .
It is also important to note that when all A, B and C have non-zero magnitude and
the single channel squared Euclidean distance is non-zero, it is always greater than
2.
or equal to 8Ebi

5.7 Simulation Result

In this section, we present the simulation result for the LDS spreading system. We
consider the 150%-load signature matrices A6,4 and A9,6 which were designed based
on the criterion presented Section 5.6. Each of these matrices provide a normalized
minimum Euclidean distance of 2 with the smallest loop girth of 6. We choose the
signature constellation with a guidelines presented as in Section 5.5, i.e.,

v1 = exp(iθ1 ), v2 = exp(iθ2 ) and v3 = exp(iθ3 ) (5.68)

with θ1 = 0, θ2 = π/3 and θ3 = 2π/3.

We now compare the simulation results against the union bound, iterative decod-
ing and an exhaustive method analytical curves.
5.7 Simulation Result 183

5.7.1 A6,4

A binary matrix of size 6 × 4 is shown below, where wr = 2, wc = 3, w f = 1.

 
1 1 0 0
0 0 1 1
 
 
0 1 1 0
A=
1
 (5.69)
 0 0 1

 
1 0 1 0
0 1 0 1

Distributing the signature constellation points in a way to achieve a one to one

mapping using the above described method, we obtain a spreading matrix shown
below.

 
v3 v2 0 0
 0 0 v2 v1 
 
 
 0 v3 v1 0 
A=
  (5.70)
v2 0 0 v3 

 
v1 0 v3 0 
0 v1 0 v2

Figure 5.4 shows the bit error probability of a BPSK modulated LDS spreading
system with an overloading factor of 150%. An iterative decoding performance is
compared with union bound performance and exhaustive decoding. In addition, we
find out the normalized minimum Euclidean distance for the considered spreading
matrix A6,4 designed based on the procedure presented in Section 5.5. We find that
the normalized minimum Euclidean distance to be 2 which is equal to the upper
bound on the normalized euclidean distance that can be calculated using Proposition
8. The obtained result was very significant in achieving the best possible bit error
rate for the overloading factor of 150%. Since the matrix are scalable by changing N
and K keeping the ratio constant, by using the design principle presented in section
5.5, we can obtain a spreading matrix of any arbitrary size with minimum density
of non-zero symbols with overloading factor 1.5 without changing the minimum
Euclidean distance. Another example of the matrix is presented next.
184 Low Density Spreading (LDS) Multiplexing for NOMA

100

10-1

10-2

10-3

10-4

10-5

10-6
0 1 2 3 4 5 6 7 8 9 10

Fig. 5.4 Probability of a BPSK bit error for A6,4

5.7.2 A9,6

A binary matrix of size 9 × 6 is shown below, where wr = 2, wc = 3, w f = 1.

 
1 1 0 0 0 0
 
0 0 1 1 0 0
 
0 0 0 0 1 1
 
 
1 0 0 1 0 0
 
A=
0 0 1 0 1 0
 (5.71)
0 1 0 0 0 1
 
 
0 0 1 0 0 1
 
1 0 0 0 1 0
 
0 1 0 1 0 0

Distributing the signature constellation points in a way to achieve one to one

mapping using the above described method, we obtain a spreading matrix shown
below.
5.7 Simulation Result 185

 
v1 v2 0 0 0 0
 
 0 0 v1 v3 0 0 
 
 0 0 0 0 v1 v2
 
 
v2 0 0 v1 0 0 
 
A=
 0 0 v2 0 v3 0 
 (5.72)
 0 v1 0 0 0 v3
 
 
 0 0 v3 0 0 v2
 
v3 0 0 0 v2 0 
 
0 v3 0 v2 0 0

In Figure 5.5 we present the bit error probability of a BPSK modulated LDS
spreading system with an overloading factor of 150% with these parameters. Iterative
decoding performance is compared with an union bound performance and exhaustive
decoding. In addition, we find out the normalized minimum Euclidean distance for
the considered spreading matrix A9,6 designed based on the procedure presented
in Section 5.5. We find that the normalized minimum Euclidean distance to be
2 which is equal to the upper bound on the normalized Euclidean distance that
can be calculated using Proposition 8. The obtained result is very significant in
achieving the best possible bit error rate for the overloading factor of 150%. Since
the matrices are scalable by changing N and K keeping the ratio constant, by using
the design principle presented in section 5.5, we can obtain a spreading matrix of any
arbitrary size with minimum density of non-zero symbols with overloading factor 1.5
without changing the minimum Euclidean distance. Another example of the matrix
is presented next.
186 Low Density Spreading (LDS) Multiplexing for NOMA

100

10-1

10-2

10-3

10-4

10-5

10-6
0 1 2 3 4 5 6 7 8 9 10

Fig. 5.5 Probability of a BPSK bit error for A9,6

Chapter 6

Summary and Conclusions

In this thesis, we have presented various physical layer algorithms in context to 5G

wireless communications. The focus has been on low latency and high reliability
use cases, spectrum scarcity and spectrum efficiency problems and efficient Non-
orthogonal Multiplexing Access. Different specific problems – reliability based
decoding of Linear Block Codes, binary pattern synchronization, signal detection
under practical imperfection and non-orthogonal multiplexing technique - have been
formulated under a 5G physical layer framework.
In section 2, a simplified statistical approach has been investigated to analyse the
properties of ordered vector components of Linear Block Codes. In this case, we
have presented a novel statistic which computes the number of errors contained in
the ordered received noisy codeword, which highlights and makes evidence to OSD
property. This is a key parameter that can be applied to derive a further simplified
error performance bound. Moreover, simplified expressions for the pdf and cdf of
the proposed statistic have also been derived. The main application of these results
is in highly efficient decoding of short-length algebraic Linear Block Codes which
can be used for low latency use cases of 5G.
In section 3, Massey detector has been revisited in the context of efficient frame
synchronization for low latency, high reliability use cases of 5G. We have presented a
complete set of new analytical results for the pdfs and the false-alarm/miss-detection
probabilities for both of these detectors. A new Extended Massey Detector has been
introduced to eliminate a lower bound limitation on the Pf a for short pattern lengths.
We have shown that the two Massey detectors achieve nearly-optimal performances
188 Summary and Conclusions

both at low and high SNR values. In comparison to optimal Likelihood Ratio test
they have very limited penalty (small fraction of dB), but much less complexity.
When compared to the classical soft and hard correlation tests, the Massey detectors
achieve significant gains of several dBs. All the results have suggested that the
Extended Massey detector (for small pattern length N) and the Massey detector
(for large N) can be considered the best solution for many demanding applications
requiring less complex and short binary pattern identification, like high bit rate, low
latency and high reliability frame synchronization.
In section 4, examples of semi-blind signal detection algorithms under practical
imperfections have been considered having in mind their possible application to
Cognitive Radio - like spectrum sharingapproach for 5G. Both theoretical and
numerical results have been presented, showing that a substantial performance
loss (in terms of detection and false alarm probabilities) can be experienced under
practical imperfections like noise uncertainty and unknown primary user traffic. By
adopting approaches based on ML estimation of noise variances, the performance
reduction under noise uncertainty can be improved. Also, specific expressions of
signal detection performance have been derived for the considered scenarios, taking
into account unknown noise variance and unknown primary user traffic.
In section 5, motivated by the the demand of efficient non-orthogonal multiplex-
ing system for 5G networks, a structured approach to design low-density signatures
providing 150% overloading for non-orthogonal LDS multiplexing system over the
AWGN channel has been presented. Moreover, we have also provided a theoretical
proof on the upper bound on the minimum Euclidean distance for the LDS multi-
plexing system with regular spreading matrix. On one hand, our sequences are of
low density type a property which highly simplifies receiver operation. On the other
hand, their minimum Euclidean distance properties ensure good performance on
AWGN channel. Instead of relying on trial-and-error searches and on pseudo random
design steps, we provided a a theoretical, structured approach towards the design of
optimal spreading sequences. We have focused on their performance in the AWGN
channel which is a good model for downlink multiple access systems operating in a
flat, slowly fading radio environment. We evaluate the performance of our sequence
constructions and we showed that they have beneficial properties in terms of their
distance spectrum.
References

[1] Fp7 european project 317669 metis (mobile and wireless communications
enablers for the twenty-twenty information society) 2012. URL https://www.
metis2020.com.
[2] Etsi tr 101 557 v1.1.1 (2012-02). URL http://www.etsi.org/deliver/etsi_tr/
101500_101599/101557/01.01.01_60/tr_101557v010101p.pdf.
[3] Ieee standard for information technology- local and metropolitan area
networks- specific requirements- part 22: Cognitive wireless ran medium
access control (mac) and physical layer (phy) specifications: Policies and
procedures for operation in the tv bands. Ieee std 802.22-2011, July 2011.
[4] Fp7 european project 318555 5g now (5th generation nonorthogonal wave-
forms for asynchronous signalling), 2012. URL http://www.5gnow.eu.
[5] The tactile internet itu-t technology watch report august 2014. Techni-
cal report, August 2014. URL https://www.itu.int/dms_pub/itu-t/oth/23/01/
T23010000230001PDFE.pdf.
[6] Ieee 802.19 wireless coexistence working group (wg). Technical report, July,
2013. URL http://ieee802.org/19/.
[7] Ieee 802.16 working group on broadband wireless access standards. Technical
report, July, 2013. URL http://ieee802.org/16/.
[8] J. Adams. An introduction to ieee std 802.15.4. In IEEE Aerospace Confer-
ence, 2006.
[9] D. Agrawal and A. Vardy. Generalized minimum distance decoding in
euclidean-space: Performance analysis. IEEE Trans. Inform. Theory, 46:
60–83, January 2000.
[10] G. Wunderet al. 5gnow: Challenging the lte design paradigms of orthogonality
and synchronicity. In Mobile and Wireless Communication Systems for 2020
and beyond, Workshop at 77th IEEE Vehicular Technology Spring Conference,
June 2013.
[11] M. Al-Imari, M. A. Imran, and R. Tafazolli. Low density spreading for next
generation multicarrier cellular systems. In International Conference on
Future Communication Networks (ICFCN, 2012.
190 References

[12] M. Al-Imari, M. A. Imran, R. Tafazolli, and D. Chen. Performance eval-

uation of low density spreading multiple access. In International Wireless
Communications and Mobile Computing Conference (IWCMC), 2012.
[13] M. AL-Imari, M. A. Imran, R. Tafazolli, and D. Chen. Performance evaluation
of low density spreading multiple access. In International Wireless Communi-
cations and Mobile Computing Conference (IWCMC) 2012, Limassol, 2012.
[14] Saif E. A. Alnawayseh and P. Loskot. Complexity reduction of ordered
statistic decoding using side information. IEEE Comm. Letters, 16(2):249–
251, November 2012.
[15] 4G Americas. 5g spectrum recommendation. White paper, Aug.
2015. URL http://www.4gamericas.org/files/514/3930/9262/4G_Americas_
5G_Spectrum_Recommendations_White_Paper.pdf.
[16] D. D. Ariananda, M. K. Lakshmanan, and H. Nikookar. A survey on spectrum
sensing techniques for cognitive radio. In Second International Workshop on
Cognitive Radio and Advanced Spectrum Management, 2009. CogART 2009.,
May 2009. doi: 10.1109/COGART.2009.5167237.
[17] E. Arikan. Channel polarization: A method for constructing capacity achiev-
ing codes for symmetric binary-input memoryless channels. IEEE Trans. Inf.
Theory, 55:3051–3073, July 2009.
[18] E. Axell, G. Leus, E. G. Larsson, and H. V.: Poor. Spectrum sensing for
cognitive radio:state-of-the-art and recent advances. IEEE Signal Process.
Mag., 29(3):101–116, May 2012.
[19] S. Badrinath and V. U. Reddy. A hybrid energy detection approach to spectrum
sensing. In First UK-India International Workshop on Cognitive Wireless
Systems (UKIWCWS), 2009, volume 1, pages 10–12, December 2009. doi:
10.1109/UKIWCWS.2009.5749385.
[20] Z. D. Bai. Methodologies in spectral analysis of large-dimensional random
matrices, a review. Statistica Sinica, 9:611–677, 1999.
[21] Jinho Baik and Jack W. Silverstein. Eigenvalues of large sample covariance
matrices of spiked population models. Journal of Multivariate Analysis, 97.6:
1382–1408, 2006.
[22] M. Baldi, M. Bertinelli, F. Chiaraluce, P. Dhakal, R. Garello, N. Maturo, and
E. Paolini. A study on state-of-the-art space mission telecommand receivers.
IEEE Aerospace and Electronic Systems Magazine, 2017. In Press.
[23] R. H. Barker. Group synchronization of binary digital systems. Communica-
tion Theory, London UK., 1953.
[24] A. Barron and A. Joseph. Toward fast reliable communication at rates near
capacity with Gaussian noise. In Proc. of IEEE ISIT 2010, 2010.
References 191

[25] C. Berrou, A. Glavieux, and P. Thitimajshima. Near shannon limit error

correcting coding and decoding: Turbo codes. In in IEEE Proc. Int. Conf.
Communications, Geneva, Switzerland., May 1993.
[26] P. Bianchi, M. Debbah, M. Maida, and J. Najim. Performance of statistical
tests for single-source detection using random matrix theory. Information
Theory, IEEE Transactions on, 57:2400–2419, April 2011. doi: 10.1109/TIT.
2011.2111710.
[27] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski. Five
disruptive technology directions for 5g. in IEEE Communications Magazine,
52(2):74–80, February 2014. doi: 10.1109/MCOM.2014.6736746.
[28] R. C. Bose and D. K. Ray-Chadhuri. On a class of error correcting binary
group codes. Inform. Control, 3:68–79, March 1960.
[29] D. Cabric, A. Tkachenko, and R. W. Brodersen. Spectrum sensing measure-
ments of pilot, energy, and collaborative detection. In Military Communica-
tions Conference, October 2006. doi: 10.1109/MILCOM.2006.301994.
[30] Danijela Cabric, Artem Tkachenko, and Robert W. Brodersen. Experimental
study of spectrum sensing based on energy detection and network coopera-
tion. In First international workshop on Technology and policy for accessing
spectrum (TAPAS ’06)., 2006.
[31] G. P. Calzolari, F. Chiaraluce, R. Garello, and E. Vassallo. Symbol synchro-
nization properties of ccsds turbo codes. Intl. Journal of Satellite Commun.,
20:379–390, 2002.
[32] D. Chase. A class of algorithms for decoding block codes with channel
measurement information. IEEE Trans. Inform. Theory, 8:170–182, January
1972.
[33] Jian Chen, A. Gibson, and J. Zafar. Cyclostationary spectrum detection in
cognitive radios. In IET Seminar on Cognitive Radio and Software Defined
Radios: Technologies and Techniques, 2008, September 2008.
[34] M. Chiani and M. G. Martini. Optimum synchronization of frames with
unknown, variable lengths on Gaussian channels. In Proc. IEEE Global
Telecomm. Conf., Dallas, US., November 2004.
[35] M. Chiani and M. G. Martini. Practical frame synchronization for data with
unknown distribution on awgn channels. IEEE Commun. Lett, 9(5):456–458,
May 2005.
[36] M. Chiani and M. G. Martini. On sequential frame synchronization in awgn
channels. IEEE Trans. Commun, 54(2), February 2006.
[37] M. Chiani and M. G. Martini. Analysis of optimum frame synchronization
based on periodically embedded sync words. IEEE Trans. Commun, 55(11):
2056–2060, November 2007.
192 References

[38] J. Choi. Low density spreading for multicarrier systems. In Proceedings

of the Eighth International Symposium on Spread-Spectrum Techniques and
Applications, 2004.
[39] Cisco. Visual networking index. White paper at cisco.com, 2016.
[40] Federal Communication Commission. Notice of proposed rule making and
order: Facilitating opportunities for flexible, efficient, and reliable spectrum
use employing cognitive radio technology. Feb 2005.
[41] J. H. Conway and N. J. A. Sloane. Soft decoding techniques for codes and
lattices, including the golay code and the leech lattice. IEEE Trans. Inform.
Theory, 32:41–50, January 1986.
[42] E. Dahlman, G. Mildh, S. Parkvall, J. Peisa, J. Sachs, Y. Selén, and J. Sköld.
5g wireless access: requirements and realization. IEEE Commun. Mag, 52
(12):42–47, December 2014.
[43] Erik Dahlman, Stefan Parkvall, and Johan Skold. 4G: LTE/LTEAdvanced for
Mobile Broadband, 2nd ed. Elsevier, Waltham, MA, USA, 2014.
[44] P. Dhakal and D. Riviello. Multi-antenna energy detector under unknown
primary user traffic. In International Conference on Advances in Cognitive
Radio (COCORA), Feb. 2016.
[45] P. Dhakal, D. Riviello, and R. Garello. Snr wall analysis of multi-sensor
energy detection with noise variance estimation. In IEEE ISWCS, Aug. 2014.
[46] P. Dhakal, D. Riviello, F. Penna, and R. Garello. Impact of noise estimation
on energy detection and eigenvalue based spectrum sensing algorithms. In
IEEE International Conference on Communications (ICC), June 2014.
[47] P. Dhakal, D. Riviello, F. Penna, and R. Garello. Hybrid approach analysis
of energy detection and eigenvalue based spectrum sensing algorithms with
noise power estimation. In The Fourth Int. Conf. on Advances in Cognitive
Radio, 2014.
[48] P. Dhakal, R. garello, S. K. Sharma, S. Chatzinotas, and B. Otternsten. On the
error performance bound of ordered statistics decoding of linear block codes.
In IEEE ICC 2016, 2016.
[49] P. Dhakal, S. K. Sharma, S. Chatzinotas, B. Otternsten, and D. Riviello.
Effect of largest eigenvalue based spectrum sensing technique. In Springer
International Publishing, 2016.
[50] F. F. Digham, M.-s. Alouini, and Marvin K. Simon. On the energy detection
of unknown signals over fading channels. In IEEE International Conference
on Communications, 2003., May 2003. doi: 10.1109/ICC.2003.1204119.
References 193

[51] F. F. Digham, M. S. Alouini, and Marvin K. Simon. On the energy detection of

unknown signals over fading channels. Communications, IEEE Transactions
on, 55:21–24, January 2007. doi: 10.1109/TCOMM.2006.887483.
[52] D. Divsalar and F. Pollara. One the design of turbo codes. Tda progress report
42-123, November 1995.
[53] D. Divslar and R. J. McEliece. Effective freee distance of turbo codes. IEEE
Electronic Letters, 32(5):445–446, 1996.
[54] B. G. Dorsch. A decoding algorithm for binary block codes and j-ary output
channels. IEEE Trans. Inform. Theory, 20:391–394, May 1974.
[55] C. Douillard and C. Berrou. Turbo codes with rate-m/(m+1) constituent
convolutional codes. IEEE Transactions On Communications, 53:1630–1638,
2005.
[56] P. F. Driessen. Performance of frame synchronization in packet transmission
using bit erasure information. IEEE Trans. Commun., 39(4):567–573, April
1991.
[57] ECC. Spectrum allocation in europe. Technical report, 2013. URL http:
//www.erodocdb.dk/Docs/doc98/official/pdf/ercrep025.pdf.
[58] John D. Edell. Wideband, noncoherent, frequency-hopped waveforms and
their hybrids in low-probability-of-intercept communications. No. NRL-8025.
NAVAL RESEARCH LAB WASHINGTON DC, 1976.
[59] Ericsson. More than 50 billion connected devices. White paper, 2011.
[60] A. Osseiran et al. Scenarios for 5g mobile and wireless communications: the
vision of the metis project. in IEEE Communications Magazine, 52(5):26–35,
May 2014. doi: 10.1109/MCOM.2014.6815890.
[61] D. J. Costello et al. Spatially coupled sparse codes on graphs: Theory and
practice. IEEE Commun. Mag, 52(7):168–179.
[62] J. G. Andrews et al. What will 5g be? in IEEE Journal on Selected Areas
in Communications, 32(6):1065–1082, June 2014. doi: 10.1109/JSAC.2014.
2328098.
[63] J. Thompson et al. 5g wireless communication systems: prospects and
challenges. IEEE Commun. Mag, pages 62–64, February 2014.
[64] Jianfeng Wang et al. First cognitive radio networking standard for per-
sonal/portable devices in tv white spaces. April 2010. doi: 10.1109/DYSPAN.
2010.5457855.
[65] Popovski P. et al. Deliverable d1.1 scenarios, requirements and kpis for 5g
mobile and wireless system. Ict-317669-metis/d1.1, April 2013.
194 References

[66] Popovski P. et al. Deliverable d6.3 intermediate system evaluation results.

Ict-317669-metis/d6.3, August 2014.
[67] R. Baldemair et al. Evolving wireless communications: addressing the chal-
lenges and expectations of the future. IEEE Vehic. Tech. Mag, 8(1):24–30,
March 2013.
[68] R. Irmer et al. Coordinated multipoint: Concepts, performance, and field trial
results. IEEE Commun. Mag, 49(2):102–111, February 2011.
[69] T. R. 102 889-2 Etsi. Part 2: Technical characteristics for srd equipment for
wireless industrial applications using technologies different from ultra-wide
band (uwb). Technical report, Aug 2011.
[70] FCC. Federal communication commission spectrum policy task force report.
Technical report, Nov 2002.
[71] FCC. Spectrum allocation in us. Technical report, 2003. URL www.ntia.doc.
gov/files/ntia/publications/2003-allochrt.pdf.
[72] F. Fettweis and S. Alamouti. 5g: personal mobile internet beyond what cellular
did to telephony. IEEE Commun. Mag, 52(2):140–145, February 2014.
[73] Consultative Committee for Space Data Systems. Tc synchronization and
channel coding. Ccsds 231.0-b-2 blue book, September 2010.
[74] M. Fossorier. Iterative reliability-based decoding of low density parity check
codes. IEEE Trans. Inf. Theory, 19(5):908–916, May 2001.
[75] M. Fossorier. Reliability-based soft-decision decoding with iterative informa-
tion set reduction. IEEE Trans. Inf. Theory, 48(12):3101–3106, December
2002.
[76] M. Fossorier and S. Lin. Soft-decision decoding of linear block codes based
on ordered statistics. IEEE Trans. Inform. Theory, 41:1379–1396, September
1995.
[77] M. Fossorier and S. Lin. Computationally efficient soft decision decoding of
linear block codes based on ordered statistics. IEEE Trans. Inform. Theory,
42:738–750, May 1996.
[78] M. Fossorier and S. Lin. First order approximation of the ordered binary
symmetric channel. IEEE Trans. Inform. Theory, 42:1381–1387, September
1996.
[79] M. Fossorier and S. Lin. Error performance analysis for reliability based
decoding algorithms. IEEE Trans. Inform. Theory, 48:287–293, January
2002.
[80] R. G. Gallager. Low-density parity-check codes. IRE Trans. Inform. Theory,
8:21–28, January 1962.
References 195

[81] D. Gazelle and J. Snyders. Reliability-based code search algorithms for

maximum likelihood decoding of block codes. IEEE Trans. Inform. Theory,
43:239–249, January 1997.
[82] C. Georghiades and E. Serpedin. Communication handbook. CRC, Boca
Raton, FL, 2002.
[83] A. Ghosh, J. Zhang, J. G. Andrews, and R. Muhamed. Fundamentals of LTE.
Pearson Education, Upper Saddle River, NJ, 2010.
[84] A. Ghosh, R. Jana, V. Ramaswami, J. Rowland, and N. K. Shankaranarayanan.
Modeling and characterization of large-scale wi-fi traffic in public hot-spots.
In IEEE INFOCOM, April 2011.
[85] E. H. Gismalla and E. Alsusa. On the performance of energy detection
using bartlett’s estimate for spectrum sensing in cognitive radio systems.
Signal Processing, IEEE Transactions on, 60:3394–3404, July 2012. doi:
10.1109/TSP.2012.2191544.
[86] Glenn L. Greig. On the shape of energy-detection roc curves. Perception &
Psychophysics, 48(1):77–81, 1990. doi: 10.3758/BF03205013.
[87] Simon Haykin. Cognitive radio: brain-empowered wireless communications.
IEEE Journal on Selected Areas in Communications, 23:201–220, February
2005. doi: 10.1109/JSAC.2004.839380.
[88] Simon Haykin, D. J. Thomson, and J. H. Reed. Spectrum sensing for cognitive
radio. In Proceedings of the IEEE, May 2009.
[89] P. A. Hoeher and T. Wo. Superposition modulation: myths and facts. IEEE
Communications Magazine, 49(12):110–116, December 2011.
[90] R. Hoshyar, F. P. Wathan, and R. Tafazolli. Novel low-density signature
for synchronous cdma systems over awgn channel. IEEE Trans. on Signal
Processing, 56(4):1616–1626, April 2008.
[91] R. Hoshyar, R. Razavi, and M. Al-Imari. Lds-ofdm an efficient multiple
access technique. In IEEE 71st Vehicular Technology Conference, Taipei,
2010.
[92] Hwei P. Hsu. Theory and Problems of Probability, Random Variables, and
Random Processes. SCHAUM’S OUTLINE SERIES McGRAW-HILL, 1996.
[93] T. Hwang, C. Yang, G. Wu, S. Li, and G. Ye Li. Ofdm and its wireless
applications: A survey. IEEE Trans. Veh. Technol., 58(4):1673–1694, May
2009.
[94] IEEE. 802.11 wireless local area networks. Technical report, 2013. URL
http://ieee802.org/11/.
196 References

[95] European Telecommunications Standards Institute. Digital video broadcasting

(dvb). framing structure, channel coding and modulation for digital terrestrial
television. Etsi en 300 744, v1.6.1, January 2009.
[96] M. H. et al. Islam. Spectrum survey in singapore: Occupancy measurements
and analyses. In Cognitive Radio Oriented Wireless Networks and Communi-
cations (CROWNCOM), 2008, May 2008. doi: 10.1109/CROWNCOM.2008.
4562457.
[97] W. Jin and M. Fossorier. Reliability-based soft decision decoding with multi-
ple biases. IEEE Trans. Inf. Theory, 53(1):105–120, January 2007.
[98] Iain M. Johnstone. On the distribution of the largest eigenvalue in principal
components analysis. The Annals of statistics, 29.2:295–327, 2001.
[99] G. D. Forney Jr. Generalized minimum distance decoding. IEEE Trans.
Inform. Theory, 12:125–131, April 1966.
[100] Iso/iec Jtc1/sc29/wg11. Mpeg-4 overview. (v.18 - singapore version), March
2001.
[101] A. Kabat, F. Guilloud, and R. Pyndiah. New approach to order statistics
decoding of long linear block codes. In Proc. Globecom, pages 1467–1471,
November 2007.
[102] A. Kopansky and M. Bystrom. Detection of aperiodically embedded syn-
cronization patterns. In Proc. IEEE International Symp. Inf. Theory (ISIT
2001), June 2001.
[103] A. Kopansky and M. Bystrom. Detection of aperiodically embedded syn-
cronization patterns. IEEE Trans. Wireless Commun, 3(5):1386–1392, October
2004.
[104] A. Kortun, T. Ratnarajah, M. Sellathurai, Ying-Chang Liang, and Yonghong
Zeng. Throughput analysis using eigenvalue based spectrum sensing under
noise uncertainty. August 2012. doi: 10.1109/IWCMC.2012.6314237.
[105] V. I. Kostylev. Energy detection of a signal with random amplitude. In IEEE
International Conference on Communications, 2002., 2002. doi: 10.1109/
ICC.2002.997120.
[106] A. Kothiyal, O. Y. Takeshita, W. Jin, and M. Fossorier. Iterative reliability-
based decoding of linear block codes with adaptive belief propagation. IEEE
Commun. Lett, 9(12):1067–1069, December 2005.
[107] Shira Kritchman and Boaz Nadler. Non-parametric detection of the number
of signals: hypothesis testing and random matrix theory. Signal Processing,
IEEE Transactions on, 57.10:3930–3941, 2009.
References 197

[108] M. Laner, P. Svoboda, N. Nikaein, and M. Rupp. Traffic models for machine
type communications. In Proc. Wireless Commun. Syst. (ISWCS’13), pages
1–5, August 2013.
[109] Wei Lin and Qinyu Zhang. A design of energy detector in cognitive radio
under noise uncertainty. In 11th IEEE Singapore International Conference on
Communication Systems, 2008. ICCS 2008., November 2008. doi: 10.1109/
ICCS.2008.4737174.
[110] G. L. Lui and H. H. Tan. Frame synchronization for Gaussian channels. IEEE
Trans. Commun, COM-35(8):818–829, August 1987.
[111] ITU-R M. Imt vision - framework and overall objectives of the future devel-
opment of imt for 2020 and beyond. Itu working document 5d/temp/224-e,
July 2013.
[112] X. Ma, J. Liu, and B. Bai. New techniques for upper-bounding the ml decoding
performance of binary linear codes. IEEE Trans. Commun, 61(3):842–851,
2013.
[113] S. L. MacDonald and D. C. Popescu. Impact of primary user activity on the
performance of energy-based spectrum sensing in cognitive radio systems.
December 2013.
[114] D. J. MacKay. Good error-correcting codes based on very sparse matrices.
IEEE Trans. Inform. Theory, 45(2):399–431, March 1999.
[115] F. J. MacWilliams and N. J. A. Sloane. The Theory of Error-Correcting Codes:
Part 2. Elsevier, 1977.
[116] G. C. Madueno, C. Stefanovic, and P. Popovski. Reliable reporting for
massive m2m communications with periodic resource pooling. IEEE Wireless
Communications Letters, 3(4):429–432, August 2014.
[117] A. Mahram and M. C. Amirani. The complexity as a criterion for blind
spectrum sensing in cognitive radio. Developments in E-systems Engineering
(DeSE), 444:6–8, December 2011. doi: 10.1109/DeSE.2011.62.
[118] A. Mariani, A. Giorgetti, and M. Chiani. Effects of noise power estimation
on energy detection for cognitive radio applications. IEEE Transactions on
Communications,, 59:3410–3420, December 2011. doi: 10.1109/TCOMM.
2011.102011.100708.
[119] A. Mariani, A. Giorgetti, and M. Chiani. Snr wall for energy detec-
tion with noise power estimation. In IEEE International Conference on
Communications (ICC), 2011, volume 1, pages 5–9, June 2011. doi:
10.1109/icc.2011.5963367.
198 References

[120] A. Mariani, A. Giorgetti, and M. Chiani. Effects of noise power estimation

on energy detection for cognitive radio applications. Communications, IEEE
Transactions on, 59:3410–3420, December 2011. doi: 10.1109/TCOMM.
2011.102011.100708.
[121] George Marsaglia. Ratios of normal variables. in Journal of Statistical
Software, 16(4), May 2006.
[122] J. L. Massey. Optimum frame synchronization. IEEE Trans. Commun, 20(2):
115–119, April 1972.
[123] J. Mitola and G. Q. Maguire, Jr. Cognitive radio: making software radios
more personal. Personal Communications, IEEE, 6:13–18, 1999. doi: 10.
1109/98.788210.
[124] Joseph Mitola. Cognitive radio: An integrated agent architecture for software
defined radio. PhD thesis, Doctor of Technology, Royal Inst. Technol.(KTH),
Stockholm, Sweden, 2000.
[125] F. Moghimi, R. Schober, and R. K. Mallik. Hybrid coherent/energy detection
for cognitive radio networks. Wireless Communications, IEEE Transactions
on, 10:1594–1605, May 2011. doi: 10.1109/TWC.2011.030411.100973.
[126] A. Montanari and D. Tse. Analysis of belief propagation for non-linear
problems: The example of cdma (or: How to prove tanaka’s formula). In
Proceedings of the 2006 IEEE Information Theory Workshop, 2006.
[127] M. Morelli. Timing and frequency synchronization for the uplink of an ofdma
system. IEEE Trans. on Commun., 52(2):296–306, February 2004.
[128] Robb J. Muirhead. Latent roots and matrix variates: a review of some asymp-
totic results. The Annals of Statistics, pages 5–33, 1978.
[129] B. Nadler, F. Penna, and R. Garello. Performance of eigenvalue-based signal
detectors with known and unknown noise level. June 2011. doi: 10.1109/icc.
2011.5963473.
[130] J. Neyman and E. Pearson. On the problem of the most efficient tests of
statistical hypotheses. Philosophical Transactions of the Royal Society of
London, 231:289–337, 1933.
[131] J. Neyman and E. S. Pearson. On the problem of the most efficient tests of
statistical hypotheses. Philosophical Transactions of the Royal Society, 231:
289–337, 1933.
[132] P. T. Nielsen. Some optimum and suboptimum frame synchronizers for binary
data in Gaussian noise. IEEE Trans. Commun, 21(6):770–772, June 1973.
[133] P. T. Nielsen. On the expected duration of a search for a fixed pattern in
random data. IEEE Trans. Inf. Theory, 19(6):702–704, September 1973.
References 199

[134] H. Nikopour and H. Baligh. Sparse code multiple access. In 2013 IEEE
24th Annual International Symposium on Personal, Indoor, and Mobile Radio
Communications (PIMRC, 2013.
[135] Nokia. 2020: Beyond 4g radio evolution for the gigabit experience. White
paper, 2011. URL http://www.nokiasiemensnetworks.com.
[136] H. Ochiai and H. Imai. On the distribution of the peak-to-average power ratio
in ofdm signals. IEEE Trans. on Commun., 49(2):282–289, February 2001.
[137] R. Palit, K. Naik, and A. Singh. Anatomy of wifi access traffic of smart-
phones and implications for energy saving techniques. International Journal
of Energy, Information and Communication, 3(1):1–16, February 2012.
[138] F. Penna and R. Garello. Detection of discontinuous signals for cognitive
radio applications. IET Communications, 5(10):1453–1461, January 2011.
[139] F. Penna, R. Garello, and M. A. Spirito. Probability of missed detection
in eigenvalue ratio spectrum sensing. In IEEE conference on Wireless and
Mobile Computing, Networking and Communications, 2009., October 2009.
doi: 10.1109/WiMob.2009.29.
[140] H. Pradhan, A. S. Kalamkar, and A. Banerjee. Sensing-throughput tradeoff in
cognitive radio with random arrivals and departures of multiple primary users.
IEEE Communications Letters, 19(3):415–418, March 2015.
[141] J. G. Proakis. Digital Communications. McGraw-Hill, 4th edition, 2001.
[142] Qualcomm. 5g waveforms and multiple access techniques. Tech-
nical report. URL https://www.qualcomm.com/media/documents/files/
5g-research-on-waveform-and-multiple-access-techniques.pdf.
[143] D. Ramirez, G. Vazquez-Vilar, R. Lopez-Valcarce, J. Via, and I. Santamaria.
Multiantenna detection under noise uncertainty and primary user’s spatial
structure. May 2011. doi: 10.1109/ICASSP.2011.5946275.
[144] R. Razavi, M. AL-Imari, M. A. Imran, R. Hoshyar, and D. Chen. On receiver
design for uplink low density signature ofdm (lds-ofdm). in IEEE Transactions
on Communications, 60(11):3499–3508, November 2012.
[145] L. S. Reed and G. Solomon. Polynomial codes over certain finite fields. SIAM
J. Appli. Math, 8:300–304, 1960.
[146] D. Riviello, P. Dhakal, and R. Garello. Performance analysis of multi-antenna
hybrid energy detectors and optimization with noise variance estimation. In
International Conference on Advances in Cognitive Radio (COCORA), April
2015.
[147] D. Riviello, P. Dhakal, and R. Garello. On the use of eigen vectors in multi-
antenna spectrum sensing with noise variance estimation. In International
Conference in Signal Processing and Integrated Networks, Feb. 2015.
200 References

[148] S. N. Roy. On a heuristic method of test construction and its use in multivariate
analysis. The Annals of Mathematical Statistics, 24:220–238, 1953.
[149] S. N. Roy. On a heuristic method of test construction and its use in multivariate
analysis. Ann. Math Stat., pages 220–238, 1953.
[150] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and K. Higuchi.
Non-orthogonal multiple access (noma) for cellular future radio access. In
IEEE 77th Vehicular Technology Conference (VTC Spring), Dresden, 2013.
[151] R. A. Scholtz. Frame synchronization techniques. IEEE Trans. Commun,
COM-28:1204–1213, August 1980.
[152] S. Sequeira, R. R. Mahajan, and P. Spasojevic. On the noise power estimation
in the presence of the signal for energy-based sensing. In 35th IEEE Sarnoff
Symposium (SARNOFF), 2012, May 2012. doi: 10.1109/SARNOF.2012.
6222753.
[153] C. E. Shannon. A mathematical theory of communication. Bell Sys. Tech., 27:
379–423, October 1948.
[154] S. K. Sharma, T. E. Bogale, S. Chatzinotas, B. Ottersten, L. B. Le, and
X. Wang. Cognitive radio techniques under practical imperfections: A survey.
IEEE Commun. Surveys Tutorials, 2015.
[155] B. Sklar and F. J. Harris. The abcs of linear block codes. IEEE Signal Process.
Mag, 21:14–35, July 2004.
[156] DySPAN standards committe. Technical report. URL http://grouper.ieee.org/
groups/dyspan/.
[157] J. J. Stiffler. Theory of synchronous communications. Prentice-Hall, Engle-
wood Cliffs, NJ, 1971.
[158] R. Tandra and A. Sahai. Noise calibration, delay coherence and snr walls
for signal detection. In 3rd IEEE Symposium on New Frontiers in Dynamic
Spectrum Access Networks, 2008. DySPAN 2008., volume 1, pages 14–17,
October 2008. doi: 10.1109/DYSPAN.2008.42.
[159] R. Tandra and A. Sahai. Snr walls for signal detection. IEEE Journal of
Selected Topics in Signal Processing,, 2:4–17, February 2008. doi: 10.1109/
JSTSP.2007.914879.
[160] L. Tang, Y. Chen, E. L. Hines, and M. S. Alouini. Effect of primary user
traffic on sensing-throughput tradeoff for cognitive radios. IEEE Transactions
on Wireless Communications, pages 1063–1068, Apr 2011.
[161] D. J. Torrieri. Principles of Secure Communication Systems. Artech House,
Boston, US, 1992.
References 201

[162] J. K. Tugnait. On multiple antenna spectrum sensing under noise variance

uncertainty and flat fading. Signal Processing, IEEE Transactions on, 60(4):
1823–1832, April 2012. doi: 10.1109/TSP.2011.2180721.
[163] R. Turyn. Sequences with small correlation. Wiley, New York, 1968.
[164] Harry Urkowitz. Energy detection of unknown deterministic signals. In
Proceedings of the IEEE, April 1967.
[165] A. Valembois and M. Fossorier. Box and match techniques applied to soft
decision decoding. IEEE Trans. Inf. Theory, 50(5):796–810, May 2004.
[166] V. Valenta, R. Marsalek, G. Baudoin, M. Villegas, M. Suarez, and F. Robert.
Survey on spectrum utilization in europe: Measurements, analyses and obser-
vations. In Cognitive Radio Oriented Wireless Networks and Communications
(CROWNCOM), 2010, June 2010.
[167] J. van de Beek and B. M. Popovic. Multiple access with low-density signatures.
In Proc. of the IEEE Global Telecommunications Conference GLOBECOM,
Honolulu, HI, 2009.
[168] H. G. Vu, H. H. Nguyen, and D. E. Dodds. Iterative ordered-statistic decoding
for ldpc coded modulation systems. In Proc. CCECE, pages 196–199, May
2005.
[169] Rui Wang and Meixia Tao. Blind spectrum sensing by information theoretic
criteria for cognitive radios. Vehicular Technology, IEEE Transactions on, 59:
3806–3817, October 2010. doi: 10.1109/TVT.2010.2065250.
[170] T. Wang, Y. Chen, E. Hines, and B. Zhao. Analysis of effect of primary user
traffic on spectrum sensing performance. In Fourth International Conference
on Communications and Networking in China. ChinaCOM, 2009.
[171] F. Wathan. Low-density signature communications in overloaded CDMA
systems. 2007.
[172] A. R. Williamson, M. J. Marshall, and R. D. Wesel. Reliability-output decod-
ing of tail-biting convolution codes. IEEE Trans. Commun, 62(6):1768–1778,
April 2014.
[173] J. K. Wolf. Efficient maximum likelihood decoding of linear block codes
using a trellis. IEEE Trans. Inform. Theory, 24:76–80, January 1978.
[174] J. Y. Wu, P. H. Huang, T. Y. Wang, and V. W. S. Wong. Energy detection
based spectrum sensing with random arrival and departure of primary user’s
signal. December 2013.
[175] Y. Wu and C. Hadjicostis. Soft-decision decoding of linear block codes using
efficient iterative g-space encodings. In Proc. Globecom, pages 921–925,
November 2001.
202 References

[176] Y. Wu and C. N. Hadjicostis. Soft-decision decoding using ordered recodings

on the most reliable basis. IEEE Trans. Inf. Theory, 53(2):829–836, February
2007.
[177] L. Xichun, A. Gani, R. Salleh, and O. Zakaria. The future of mobile wireless
communication networks. In Proc. Int. Conf. Commun. Softw. Netw, February
2009.
[178] Zhuan Ye, G. Memik, and J. Grosspietsch. Energy detection using estimated
noise variance for spectrum sensing in cognitive radio networks. March 2008.
doi: 10.1109/WCNC.2008.131.
[179] Yonghong Zeng and Ying-Chang Liang. Maximum-minimum eigenvalue
detection for cognitive radio. September 2007. doi: 10.1109/PIMRC.2007.
4394211.
[180] Yonghong Zeng and Ying-Chang Liang. Spectrum-sensing algorithms for cog-
nitive radio based on statistical covariances. IEEE Transactions on Vehicular
Technology,, 58:1804–1815, May 2009. doi: 10.1109/TVT.2008.2005267.
[181] Yonghong Zeng and Ying-Chang Liang. Eigenvalue-based spectrum sensing
algorithms for cognitive radio. Communications, IEEE Transactions on, 57:6,
June 2009. doi: 10.1109/TCOMM.2009.06.070402.
[182] Yonghong Zeng, Ying-Chang Liang, Anh Tuan Hoang, and E.C.Y Peh. Re-
liability of spectrum sensing under noise and interference uncertainty. In
IEEE International Conference on Communications, Workshops, 2009. ICC
Workshops 2009., June 2009. doi: 10.1109/ICCW.2009.5208033.
[183] Q. T. Zhang. Advanced detection techniques for cognitive radio. June 2009.
doi: 10.1109/ICC.2009.5198712.
[184] Wei Zhao and Tobias Tynderfeldt. Contention-free random access procedure
in wireless networks. U.S. Patent US, 9301(US9301323 B2), March 2016.
[185] D. Zijian, L. Fei, and S. Yaolian. Improved syndrome-based ordered statistic
decoding algorithm for ldpc codes. In Proc. ICSPS, pages 295–297, Jul 2010.
[186] M. Zorzi, A. Gluhak, S. Lange, and A. Bassi. From today’s intranet of things
to a future internet of things: a wireless- and mobility-related view. IEEE
Wireless Communications, 17(6):44–51, December 2010.