Multiuser Full-Duplex Communication

Thesis submitted in partial fulfillmentof the requirements for the degree of

Master of Sciencein

Electronics and Communication Engineering by Research

by

Chandan Pradhan201332539

chandan.pradhan@research.iiit.ac.in

International Institute of Information Technology, Hyderabad(Deemed to be University)Hyderabad - 500 032, India

April 2016

Copyright c© Chandan, 2016

All Rights Reserved

International Institute of Information TechnologyHyderabad, India

CERTIFICATE

It is certified that the work contained in this thesis, titled “Multiuser Full-Duplex Communication” byChandan Pradhan, has been carried out under my supervision and is not submitted elsewhere for adegree.

Date Adviser: Prof. Garimella Rama Murthy

To

My Family and Friends

Acknowledgments

My journey in IIIT- Hyderabad has been a wonderful experience. It could not have been so withoutthe support of many people. As I submit my MS thesis, I wish to extend my gratitude to all those peoplewho helped me in successfully completing this journey. First of all, I want to thank my guide Prof.Garimella Rama Murthy, for accepting me as a student and constantly guiding me. His guidance hashelped me improve not only as a researcher but also as a person. His help and support during difficulttimes strengthened and motivated me to move further.

I thank my colleagues and friends in SPCRC for providing a positive work environment. Manythanks to Sumit, Kunal, Priyanka and Deepti for extensive discussions and support. Special thanks toAmrisha and Sudipto for strengthening me and for all the fun-filled moments in IIIT.

I could not have accomplished it without the support and understanding of my parents. I wish to thankmy Grandparents and my sister Chandni for being my constant support and motivation. Last, but notthe least, thanks to IIIT community for giving me an inspiring environment and loads of opportunitiesto grow.

v

Abstract

The inevitable high bandwidth requirement in the future cellular network has made researchers tocome up with revolutionary ideas in recent times. One such idea is the introduction of full-duplex (FD)communication. A FD systems make the simultaneous in-band transceiving feasible, i.e, simultaneousuplink and downlink operation using the same spectrum resources. In recent years, extensive work hasbeen done in the area of self-interference cancellation (SIC) design, including for compact devices likelaptops and smart phones, enabling FD communication for both single and multiple antenna transceiverunits. The designs aim in optimal cancellation of interference from the receiver chains introduced bythe transmitter chains of the transceiver unit. While this is far from true today for cellular networks,sufficient progress is being made in this direction to start considering the FD model and its implications,especially in case of small cells.

In this work, multiple user equipments (UEs) are considered operating in an FD mode on the same setof spectrum resources simultaneously. However, the use of the same spectrum resources (or subcarriers)for both uplink and downlink results in co-channel interference (CCI) at the downlink of a UE fromuplink signals of other co-existing UEs. Two techniques have been proposed to mitigate the CCI incase of multiuser full-duplex communication. The first technique involves deployment of the smartantenna technique at the UEs with highly spatially correlated multiple antennas. The second techniqueinvolves the use of diversity gain at the UEs, acquired by using multiple antennas at eNB and UEs ina rich scattering environment. However, the dynamical nature of operating conditions can make theCCI large enough for these solutions to tackle. Hence, a dynamic resource block allocation (DRBA)algorithm is proposed which shifts the operation of co-existing UEs to different spectrum resources. Thedeployment of multiuser FD communication is shown to have a significant gain in terms of the capacityof the communication system.

Further, an analysis is carried out for a method to mitigate the path loss through the dynamic spec-trum access method. In a small cell scenario which has an operating frequency in the sub-3 or sub-6 GHzbands, the operating environment can change rapidly with sudden degradation of operating conditionsor arrival of obstruction between transmitter and receiver, resulting in possible link failure. The methodanalyzed includes dynamically allocating spectrum at a lower frequency band for a link suffering fromhigh path loss. For the analysis, a wireless link was set up using Universal Software Radio Periph-erals (USRPs). The received power is observed to increase by dynamically decreasing the operatingfrequency from 1.9 GHz to 830 MHz.

vi

Contents

Chapter Page

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Problem Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Cognitive Radio Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42.2 Regulatory Models for Spectrum Access . . . . . . . . . . . . . . . . . . . . . . . . . 52.3 Cognitive Cycle for DSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.4 Physical Layer Functionalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.4.1 Geo-location Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.5 DoA Based Cognitive Base Station . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.5.1 DoA Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.5.1.1 Root Music Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.5.2 DoA based Cognitive Base Station (CBS) . . . . . . . . . . . . . . . . . . . . 162.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 Self-Interference Cancellation: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . 183.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183.2 Self-Interference Cancellation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.2.1 Passive Self-Interference Cancellation . . . . . . . . . . . . . . . . . . . . . . 213.2.2 Analog Self-Interference Cancellation . . . . . . . . . . . . . . . . . . . . . . 233.2.3 Digital Self-Interference Cancellation . . . . . . . . . . . . . . . . . . . . . . 28

3.3 Full-Duplex MIMO Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . 313.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4 Proposed Multiuser Full-Duplex Communication . . . . . . . . . . . . . . . . . . . . . . . 354.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364.3 Case 1: FD Communication: Smart Antennas Technique . . . . . . . . . . . . . . . . 38

4.3.1 Downlink Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384.3.2 Uplink Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434.3.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4.4 Case 2: FD Communication: Diversity Gain Technique . . . . . . . . . . . . . . . . . 484.4.1 Full-Duplex Multiuser Operation . . . . . . . . . . . . . . . . . . . . . . . . 49

vii

viii CONTENTS

4.4.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524.5 Dynamic Resource Block Allocation (DRBA) . . . . . . . . . . . . . . . . . . . . . . 54

4.5.1 DRBA Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.5.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

4.6 Improving Link Quality: Software Defined Radio . . . . . . . . . . . . . . . . . . . . 614.6.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4.6.1.1 Setup Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . 624.6.1.2 Deployment of the Setup . . . . . . . . . . . . . . . . . . . . . . . 624.6.1.3 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4.6.2 System Deployment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674.6.2.1 SDR in Future Cellular Networks . . . . . . . . . . . . . . . . . . . 674.6.2.2 Challenges in Deployment of the System . . . . . . . . . . . . . . . 67

4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684.8 Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694.9 Appendix B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745.1 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

List of Figures

Figure Page

2.1 LSA Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.2 Cognitive Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.3 Classification of Spectrum sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.4 Model for implementing and maintaining Geo-DB . . . . . . . . . . . . . . . . . . . . 112.5 ULA antenna configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.6 GUI model demonstrating the CR nature of the proposed architecture a) when PU at 20o

and SU at 40o b) when both PU and SU at 40o . . . . . . . . . . . . . . . . . . . . . . 17

3.1 To provide a sufficiently high SI cancellation capability, an FD radio must be capableof cancelling 110 dB of linear component, 80 dB of non-linear component as well as 60dB of analog cancellation [1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.2 Techniques related to SI measurement and suppression. [2] . . . . . . . . . . . . . . . 213.3 Block diagram of antenna cancellation for a wireless FD SI cancellation. The power

splitters introduce a 6 dB reduction in signal, thus power from TX1 is 6 dB lower com-pared to power from TX2, without the need for an additional attenuator to compensatefor the amplitude mismatch [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.4 Block diagram representation of analog and digital SI cancellations, in which the SIinvert is executed by employing balun circuit, followed by QHx220 based delay & at-tenuation adjustment [2]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.5 Circulator aided FD radio block diagram with analog and digital cancellation stages, inwhich Tb denotes the intended baseband signal for transmission, while T is the practi-cally transmitted RF signal. For an intended receive signal R, it will be contaminatedby the strong components partially due to the undesirable leakage of the circulator [1]. 29

3.6 Cascaded Cancellation Design: Shows a 3 antenna full-duplex MIMO radio design withcascaded filter structure for cancellation [4]. . . . . . . . . . . . . . . . . . . . . . . . 33

3.7 Spectrum plot after cancellation of various self-talk and cross-talk components for RX1of a 3× 3 full-duplex system using our design [4]. . . . . . . . . . . . . . . . . . . . 33

3.8 Structure of the dual-polarized microstrip antenna. (a) Top view. (b) Multilayer antennastack-up [44]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.9 Single-antenna FD solution with TX-to-RX isolation and electrical balance duplexeroperation principle [44]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4.1 System model for multiuser full-duplex communication . . . . . . . . . . . . . . . . . 364.2 Transceiver structure for the proposed eNB architecture . . . . . . . . . . . . . . . . . 374.3 Transceiver structure for the proposed UE architecture . . . . . . . . . . . . . . . . . 37

ix

x LIST OF FIGURES

4.4 Structure for the Beamforming unit for the proposed UE architecture and the formationof directed beam toward eNB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.5 BER performance at UE1 for FD downlink . . . . . . . . . . . . . . . . . . . . . . . 484.6 BER performance at eNB for FD uplink . . . . . . . . . . . . . . . . . . . . . . . . . 484.7 Overall Spectrum efficiency per cell in downlink for various scheduling approaches . . 484.8 Proposed UE transceiver architecture for mitigating CCI using diversity gain . . . . . . 494.9 Downlink capacity vs channel SNR of UE1 for different Nr . . . . . . . . . . . . . . 534.10 Downlink capacity vs channel SNR of UE1 for different Ne . . . . . . . . . . . . . . 534.11 Basic flowchart briefly demonstrating the RB handoff procedure . . . . . . . . . . . . 574.12 UE1 and UE2 in the coverage region of eNB . . . . . . . . . . . . . . . . . . . . . . . 594.13 BER performance at UE1 and UE2 w.r.t DoAUE2

eNB . . . . . . . . . . . . . . . . . . . 594.14 BER performance at UE1 and UE2 for conventional and dynamic approach w.r.t DoAUE2

eNB 604.15 Half-power beamwidth (HPBW) w.r.t number of antennas . . . . . . . . . . . . . . . 604.16 BER performance at UE1 and UE2 for Nr = 10 w.r.t DoAUE2

eNB . . . . . . . . . . . . . 604.17 The path loss experimental setup at SPCRC lab in IIIT Hyderabad . . . . . . . . . . . 624.18 Received signal strength w.r.t. operating frequency . . . . . . . . . . . . . . . . . . . 654.19 FFT plot for transmitted signal in baseband . . . . . . . . . . . . . . . . . . . . . . . 654.20 FFT plot for received signal strength in pass band for operating frequency of a) 1.9 GHz

b) 1.6 GHz c) 1.2 GHz d) 830 MHz . . . . . . . . . . . . . . . . . . . . . . . . . . . 664.21 FFT plot for received signal strength in pass band at 1.9 GHz for transmitter RF chain

gain of a) 13 dB b) 26 dB c) 40 dB . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664.22 Drop in received signal strength due to the obstruction between eNB and UE . . . . . . 68

List of Tables

Table Page

4.1 Received signal strength (in decibel) for four sets measured . . . . . . . . . . . . . . . 64

xi

Chapter 1

Introduction

A recent article in Times Of India (TOI) headlined: TRAI moots spectrum sharing, prices may fall.TRAI (The Telecom Regulatory Authority of India) regulates telecom services in India. The articlereflects the recent problem of spectrum scarcity which is faced by almost all the telecom regulatorybodies worldwide. In the last few years, there has been a steep increase in the number of cellular serviceproviders and its subscribers. The Wireless World Research Forum (WWRF) predicted that 7 trillionwireless devices will serve 7 billion people worldwide by 2017 [5]. The explosion in the number ofmobile users has resulted in the depletion of limited spectrum resource. With 4G getting deployed orsoon to be deployed in many countries, this problem of spectrum crisis can prove to be a bottleneck.To address this issue and facilitate higher data rate, researchers have already begun working on the 5Gcellular networks. The 5G is expected to get deployed beyond 2020.

A number of organizations in different countries and regions have launched programs for 5G suchas 5th Generation Non-Orthogonal Waveforms for Asynchronous Signaling (5GNOW) [5–8] and Mo-bile and Wireless Communications Enablers for the Twenty-twenty Information Society (METIS) [9].Key technologies for future wireless communication systems will make it possible to achieve capac-ity enhancement via increased spectral efficiency, spectrum extension and network densification usingmany small cells [10]. Furthermore, the Third Generation Partnership Project (3GPP) has proposed adraft version as a roadmap for the 5G system [11]. This proposal includes requirements such as higherspectral and energy efficiency, lower end-to-end latency, and the ability to support massive numbers ofnodes for future wireless communication systems [11]. As we turn our attention to increasing spectralefficiency of massive-scale services and applications [12, 13], it is fair to ask: What candidate tech-nologies might define future wireless networking [14]. These technologies can involve systems basedon utilizing advanced air interface transmission, such as advanced interference mitigation techniques,massive multiple-input multiple-output (MIMO) [13] and non orthogonal transmission.

Among these technologies, non-orthogonal transmission methods, such as in-band full-duplex (FD)transmission [1, 4, 15, 16] have recently garnered significant attention in academia and industry becausethey have the potential of being able to increase spectral efficiency without the need for additional spec-trum resources. FD transmission has a long history and has been used in the design of continuous

1

wave (CW) radar systems since at least the 1940s [17]. Significant interest has also recently arisenin re-architecting wireless communications systems because FD transmission can double the ergodiccapacity of a MIMO system [1, 4]. Ideally, using the FD transmission mode means wireless communi-cation systems can potentially double their spectral efficiency relative to systems with half-duplex (HD)operation. In conventional systems using the HD transmission mode, time division duplex (TDD) orfrequency division duplex (FDD) are generally utilized for orthogonal transmission. However, this or-thogonal transmission with TDD or FDD can lead to decreased spectral efficiency. On a contrary, the FDsystem simultaneously transmits and receives data signals with the same spectrum resource [1,4,15,16].

In this thesis, the focus will be on the multiuser full-duplex communication, specifically in a smallcell scenario. The small cell systems are considered suitable for deploying FD communication due tothe low transmit power, short distances and low mobility [18–20]. In [19], the gain for small cell FDtransmission compared to the conventional HD system in a small cell scenario is analyzed.

1.1 Problem Overview

In this work, we have proposed a multiuser full-duplex communication system. The system is ca-pable of using the same spectrum resources for uplink and downlink, simultaneously, by multiple UEs.This is achieved by using SVD based beamforming at the downlink for transmitting parallel streams ofdata to the multiple UEs. The channel reciprocity property of the full-duplex communication allows theeNB to estimate channel state with ease and precode the UEs’ data streams. For the uplink, a successiveinterference cancellation algorithm is used to decouple the uplink signals of the multiple UEs. Here, wehave assumed perfect deployment of full-duplex circuitry at the transceivers. The next chapter providesa brief survey of these circuitries and their integration into the wireless infrastructure including compactdevices like laptops and smart phones. However, the use of same spectrum resources by the multipleUEs for simultaneous uplink and downlink results in the co-channel interference (CCI). The CCI issuffered at the downlink of a UE from uplink signals of other co-existing UEs. The UEs can only cancelout their own transmitted signal from the received signal. They cannot cancel the signal transmitted bythe other UEs. Also, this interference cannot be corrected easily by the eNodeB (eNB) as it does nothave access to the channels between the UEs.

1.2 Contribution

In this thesis, we focus on mitigating the CCI at the downlink of a UE from uplink signals of otherUEs sharing the same spectrum resources in the FD mode. For this two techniques have been discussed:

• Smart Antennas Technique: A smart antenna based approach is deployed which uses multipleantennas at UEs to form directed beams towards eNB and nulls toward other UEs coexisting in

2

the same spectrum resources. This technique considers deploying the smart antenna technique atthe UEs with highly spatially correlated multiple antennas.

• Diversity Gain Technique: In this technique, diversity gain at the receiver (UEs) is analyzed tomitigate the effect of CCI and allowing the UEs to share the same spectrum resources. Here, thesystem is considered in a highly scattered environment. The diversity gain is shown to increasewith increase in number antennas at the transmitter (eNB) or receivers (UEs) or both.

In a practical scenario, the UEs are mobile and the operating conditions may dynamically change.For instance, in case where the smart antenna technique is deployed, the UEs sharing the same spectrumresource, may align in the same direction w.r.t the eNB, resulting in CCI. Similarly, in case of diversitygain, the UEs may increase their transmit power resulting in an increase in CCI. To tackle these dy-namic rise in CCI, we have proposed a scheduling algorithm called dynamic resource block allocation(DRBA) algorithm. This algorithm is based on the initiative of the telecom regulatory bodies to allowoperators share their licensed spectrum.

1.3 Thesis Organization

This thesis is organized in five chapters. Chapter 1 provided an overview of the general backgroundand the problem setting. Also, the motivation behind the present work and the major contributions werebriefly described. In Chapter 2, we present the state of the art in cognitive radio technology which is thebasis for the dynamic allocation of spectrum resources. This chapter describes the technical challengesof cognitive networks and existing strategies for service providers and regulatory bodies. Also, variousphysical layer aspects of cognitive radio are analyzed, including a cognitive base station model based onDirection of Arrival (DoA) estimation. The effective implementation of full-duplex communication de-pends on the optimal cancellation of self-interference from the received signal. Self-interference is theinterference in the received signal of a transceiver from its own transmitted signal when operating in thefull-duplex mode. In recent years, extensive work is carried out in designing efficient self-interferencecancellation (SIC) algorithms and circuitry. These SIC algorithms and circuitry for single and multipleantenna systems is presented in Chapter 3. Chapter 4 presents the proposed multiuser FD communi-cation which allows multiple UEs to share the same spectrum resources, simultaneously. The use ofthe same subcarriers for both uplink and downlink results in the CCI at the downlink of a UE fromuplink signals of other UEs sharing the same spectrum resources. Two techniques have been discussedto mitigate this CCI: Smart antenna technique and Diversity gain technique. In a practical scenario,the operating conditions are dynamical, which can result in the failure of the proposed method to tackleCCI. The chapter introduces the solution to the problem, the dynamic resource block allocation (DRBA)algorithm, which is used for reallocation of resources to mitigate the effects of CCI. Further, the use ofsoftware defined radio to improve the link quality of a communication system is analyzed. Finally,Chapter 5 contains the conclusions of this thesis and the future work.

3

Chapter 2

Cognitive Radio Technology

2.1 Overview

Traditional wireless networks are running with fixed spectrum assignment policies regulated by gov-ernment agencies. Spectrum is assigned to service providers on a long term basis for large geographicalregions. The spectrum is allowed to be used by licensed users. The Federal Communications Com-mission (FCC) measurements have indicated that 15-85% of the time, many licensed frequency bandsremain unused while some other bands are highly over-crowded. These overcrowded bands face theissue of spectrum scarcity. In order to better utilize the licensed spectrum, the FCC has launched asecondary market initiative, whose goal is to remove regulatory barriers and facilitate the developmentof secondary markets in spectrum usage rights among the wireless radio services. The inefficient usageof the existing spectrum can be improved through opportunistic access to the licensed bands without in-terfering with the primary users. This introduces the concept of Dynamic Spectrum Allocation (DSA),which implicitly requires the use of cognitive radios to improve spectral efficiency.

Cognitive radio is an intelligent radio that is aware of its surrounding environment and dynamicallyadapts to the transmission or reception parameters to achieve efficient communication without interfer-ing with primary users. Thus, a Cognitive Radio Network (CRN) [21] consists of primary and secondaryusers. The primary users are the licensed users and hence have exclusive rights to access the radio spec-trum, whereas the cognitive users are the unlicensed users that can opportunistically access the freespectrum bands, without causing harmful interference to primary users. This introduces cognition andadaptation to the physical layer of the network. These networks work on multi-channel environmentwith dynamic availability of channels based on primary user activity. These channels can be of differentfrequencies and/or different bandwidth. The operation of these networks is very different from simplemulti-channel environments as different frequencies have different characteristics, like transmit powerlevels, transmission distance and multipath effects.

CRNs generate new challenges to a variety of stakeholders. For RF equipment vendors, the challengeis to build equipment which can be operated on any spectrum band and capable of doing on the flychanges in the radio transmission parameters. For wireless service providers, the challenge is to use

4

their own resource efficiently and opportunistically use the other bands while maximizing their profit.As other providers can use their band opportunistically, security is also a major concern from the usersperspective. Ensuring the licensing policies which enable the secondary usage of the spectrum is achallenge for regulatory bodies. Licensing policies should provide benefit to licensed users as well ascognitive users.

To deal with the cognition in radio (i.e. physical) layer, upper layers should also be modified to adaptto the reconfiguration done at the physical layer. Thus, adaptation and cognition should also be includedat upper layers to achieve the benefits of cognitive radio technology. This makes a network as cognitivenetwork [22]. According to [23] cognitive network is defined as:

A cognitive network is a network with a cognitive process that can perceive current network condi-tions, and then plan, decide and act on those conditions. The network can learn from these adaptationsand use them to make future decisions, all while taking into account end-to-end goals.

Initially the research community focused on the issues, assuming that the Cognitive User (CU) isthe unlicensed user (ISM band user). ISM band was initially reserved internationally for industrial,scientific and medical purposes other than telecommunications. Now, ISM band is shared with license-free communication applications such as wireless sensor networks, wireless LANs, cordless phones, etc.These applications have made ISM band overcrowded and hence opportunistic access to other bandsmay solve the spectrum scarcity problem in these bands. Now research community has also startedworking on Cognitive Cellular Networks (CCN). At CCN, the secondary user is also a licensed userof some other band and opportunistically accessing PU’s band. The simplest model from the regulatorypoint of view is where cellular users are allowed to opportunistically use TV white spaces. IEEE 802.22is the standard based on CR technology for WRAN that uses UHF/VHF TV bands. The work in [24]discusses about using these TV bands for LTE network.

2.2 Regulatory Models for Spectrum Access

When we talk about dynamic spectrum access, the first question comes is What are the accessibilitymodels for spectrum?

The simplest model from the regulatory point of view is where cellular users are allowed to oppor-tunistically use TV white spaces. IEEE 802.22 is the standard based on CR technology for WRAN thatuses UHF/VHF TV bands. General accessibility models which allow the opportunistic access of anyband are discussed below.

The first model is the Collective Use of Spectrum (CUS). CUS does not provide exclusive licensesto spectrum users and hence does not provide protection from interference. CUS application mustshare the spectrum with other users. However, some technical restriction can be applied by the nationalregulatory bodies. Traditionally cellular networks are assigned exclusive licenses to mobile networkoperators (MNO) and hence guarantee QoS to licensed users. This is not possible in CUS and hence

5

MNOs are not interested in this kind of shared access of the spectrum. Radio Spectrum Policy Group(RSPG) of the European commission defines CUS as:

Collective Use of Spectrum allows an unlimited number of independent users and/or devices toaccess spectrum in the same range of designated CUS frequencies at the same time and in a particulargeographic area under a well-defined set of conditions.

Authorized Shared Access (ASA) and Licensed Shared Access (LSA) are the models that allowshared use of licensed spectrum. The idea is to share the spectrum when and where the primary licenseeis not using it. According to ASA, primary user has exclusive rights for the spectrum band and couldshare its underutilized spectrum with other users with predetermined rules and conditions. These con-ditions may be static based on location and time or may be dynamic based on frequency, time, locationand traffic of primary users. ASA provides QoS to the primary users as well as increases the spec-trum utilization with non interference bases. Nokia and Qualcomm proposed the concept of ASA. Theyclaim that this model is well suited for small cells due to their lower transmit power levels. Nokia andQualcomm are working and providing solutions for LTE ASA. They claim that ASA can be applied toexisting LTE-Advanced infrastructure and there will be no special impact on the device side beyond im-plementing the ASA frequency band. On September 2013, Nokia Solutions and Networks demonstratedthe world’s first authorized Shared access field trial with TD-LTE spectrum [25]. USA based company,Spectrum Bridge Inc, is also giving solutions for this kind of shared access.

ASA is generalized for sharing spectrum between different types of systems in LSA. LSA basicallyfocuses on the regulatory aspects of ASA. European and USA regulators are working to develop ar-chitecture/rules for LSA. Radio Spectrum Policy Group (RSPG) of European commission defines LSAas:

A regulatory approach aiming to facilitate the introduction of radio-communication systems operatedby a limited number of licensees under an individual licensing regime in a frequency band alreadyassigned or expected to be assigned to one or more incumbent users. Under the Licensed Shared Access(LSA) approach, the additional users are authorized to use the spectrum (or part of the spectrum) inaccordance with sharing rules included in their rights of use of spectrum, thereby allowing all theauthorized users, including incumbents, to provide a certain Quality of Service (QoS).

LSA regulatory architecture is shown in fig.2.1. National Regulatory Body (NRB) provides primarylicenses and LSA licenses to primary user and cognitive user/LSA licensee respectively. LSA agreementbetween primary user and LSA licensee outlines the terms and conditions of shared use of spectrum,including details about a geographical area, technical conditions, when and how to vacate the spectrum.LSA Geo-location database provides information about the spectrum availability to LSA licensee. Us-ing secure access protocols and policies, the database ensures protection and enforcement of nationalregulatory policies and objectives, while minimizing restrictions to shared use by LSA licensees. Thisdatabase may be operated by NRBs, the primary users or a trusted third party.

6

Figure 2.1 LSA Architecture

2.3 Cognitive Cycle for DSA

Cognitive cycle for dynamic spectrum access is shown in fig.2.2. The cognitive cycle builds thespectrum opportunity map identified by the sensing block and schedules the resources dynamicallyamong the cognitive users. Furthermore, the cognitive cycle allows cognitive users to vacate the selectedchannel when a primary user becomes active on that channel.

Following are the main features of spectrum management in cognitive cycle:

• Spectrum sensing: The cognitive users are allowed to use the unused portion of the spectrum.Therefore, a cognitive user should monitor the available spectrum bands to detect spectrum oppor-tunities. Spectrum sensing can be done by the cognitive users or a trusted third party. Generally,spectrum sensing techniques can be classified into three groups, namely, non-cooperative sensing,cooperative sensing, interference-based sensing as shown in fig.2.3. For detailed discussion onvarious strategies for spectrum sensing, the readers can refer to [26, 27]. In case of third partysensing, third party shares the Geo-location database of spectrum with the cognitive users. Someefforts are already taken by companies to provide Geo-location database of spectrum. The FCChas approved Spectrum Bridge Inc.’s for providing spectrum database services for television whitespaces in USA since January 26, 2012. Google is also working with industry and regulators forthe TV whitespaces database. The Google spectrum database had been certified by the FCC andis available to wireless devices that are approved by the FCC for TV white space bands. Google’sdatabase can be accessed by researchers, device manufacturers and anyone interested throughGoogle API for identification.

7

Figure 2.2 Cognitive Cycle

Figure 2.3 Classification of Spectrum sensing

8

• Spectrum decision: Once the available spectrum bands are identified, cognitive users select themost appropriate band according to their QoS requirements.

• Spectrum handoff: If primary user presence is detected in the specific portion of the spectrumin use, cognitive user should vacate that spectrum and continue their communication in someother vacant band. The spectrum handoff process needs to communicate with a spectrum sensingprocess to find the vacant band.

• Spectrum sharing: As multiple cognitive users may be looking for the spectrum at the sametime, the transmission between cognitive users should be coordinated to avoid collisions. Thisspectrum sharing capability is taken care by spectrum sharing process. Spectrum sharing can beclassified as centralized spectrum sharing and distributed spectrum sharing. Spectrum sharingcan also be classified as cooperative spectrum sharing and non-cooperative spectrum sharing. Incooperative sharing, communication effect to other cognitive users is considered while selectingthe channel. While in non-cooperative sharing, the channel is selected based on local policies.For networks where centralized infrastructure is present, spectrum sharing is done through thebase-station and hence comes under cooperative sharing. For infrastructure-less networks, wherethe controlling infrastructure is not present, spectrum sharing is done based on local policies.

2.4 Physical Layer Functionalities

Primary user detection, spectrum handoff and designing a transmitter and a receiver for cognitiveradio are the main issues for the cognitive radio physical layer. Energy consumption in transmitting asignal depends on the transmit power level of the signal, thus power control is also required in CRNs.Selecting physical layer parameters based on the channel selected at MAC layer is also a function ofthe physical layer. In this section an emerging alternative to spectrum sensing, i.e., spectrum database/Geolocation database with its key feature is presented.

2.4.1 Geo-location Database

An important issue in spectrum sensing is the reliability of the partial measurements performed bya single user. Mobile radio systems suffer from distortions like multipath fading, shadowing, causingsevere degradation to the signal, which may result in the opportunistic secondary users not detectingthe primary activity in a certain moment or location. The SUs may suffer from the hidden node prob-lem in which the primary signal strength in the non-licensed user position is weak, but opportunistictransmission will still interfere with the licensed operation. Cooperative spectrum sensing can prove tobe the useful solution to alleviate the above mentioned problem. In cooperative sensing, the decisionon channel status is based on measurements from multiple cognitive users rather than a single user.The reliability of the decision is proportional to the number of cognitive users participating in channel

9

measurement. Cooperative sensing shortens the sensing time of the spectrum while improving over-all sensitivity. In Cognitive Wireless Sensor Networks (CWSNs), cooperative sensing is facilitated byselecting a cluster head, which combines fuses the sensing data from multiple nodes to give a reliabledecision.

In cellular networks, along with the channel status, the Geo-localization of these measurements canbe used for the development of a Geo-location Database (Geo-DB) [28]. The Geo-DB stores the spec-trum holes in a certain frequency band at a given time and location. One of the examples of spectrumholes in frequency band is that of TV White Space (or spectrum holes), denoted by TVWS. The whitespace is the result of releasing of frequency band, resulting from the analog-to-digital TV switchover,referred as Digital Dividend (DD) in literature. The Digital TV (DTV) band is a typical example ofinefficient spectrum use since, depending on the geographical location, only certain channels are occu-pied. The information about these TVWS with respect to a given geographical location and time canbe stored in a Geo-DB and used to opportunistically access them by the cellular network. The latestcellular network standard being deployed is LTE Advanced (LTE-A) where LTE stands for Long TermEvolution.

In order to maintain and update the database, cognitive users sense the channel and report mea-surements using IEEE 802.21 Multimedia Independent Handover (MIH) standard [28]. The networkretrieves and compiles the cooperative sensing measurements and their geographical location in orderto decide upon the channel vacuity in a certain location and thereby updating the Geo-DB. The Geo-location information is used to propose a powerful control mechanism for opportunist access to theprimary band, aiming to minimize the interference to licensed and other secondary users. Fig.2.4 showsthe basic architecture to implement and maintain Geo-DB. A new node in the LTE-A referred to as theCognitive Resource Manager (CRM) is introduced, which coordinates the opportunistic specificationaccess (OSA) to the unlicensed spectrum based on the notifications reporting the channel status per-ceived by the cognitive user equipments (UEs). LTE-A specification considers UE localization throughthe LTE positioning protocol (LPP) and LPP annex (LPPa) (3GPP TS 36305). Several different position-ing methods are mentioned in the standard, namely: Observed Time Difference of Arrival (OTDoA),Assisted-Global Navigation Satellite System (A-GNSS) and Enhanced-Cell ID (E-CID). All of thesepositioning methods are based on measurements collected by the UE or the eNodeB. The MobilityManagement Entity (MME) is the entity that receives the request for the localization of a UE from an-other entity such as another UE, eNodeB (connected through S1 link) or other nodes. Then, the MMEsends a location service request to the Enhanced Serving Mobile Location Center (E-SMLC), whichwill execute the positioning procedure through LPP and LPPa protocols. The SLs interface betweenE-SMLC and MME serves as a tunnel for E-SMLC to transparently carry LPP and LPPa. The CRMand E-SMLC are interconnected using the MME. CRM requests the location service to the MME, whichwill activate the E-SMLC service. The resulting location calculated by the E-SMLC is sent to the MMEthat finally forwards it to the CRM. With the obtained information, the CRM can map sensing reportsand location to build the Geo-DB.

10

Figure 2.4 Model for implementing and maintaining Geo-DB

Owing to user’s positioning capability, the Geo-DB will contain valuable information about whichfrequency bands can be used by a given eNodeB at a specific moment of time and the maximum cover-age range in order not to interfere with the primary system. The CRM collects the sensing informationfrom the UEs and the positioning information from the location service provided by the MME and up-dates the database after the cooperative decision has been taken. Once this process is finished, the CRMpossesses the location of every opportunistic UE and which spectrum bands are suitable for OSA andwhich not. The Geo-DB will contain information about the occupation of different spectrum bands inthe licensed spectrum band on a per-cell basis, including also the maximum coverage distance from theeNodeB. This way, an eNodeB, identified in the Geo-DB, is able to opportunistically use those spectrumbands with reduced transmission power in case the maximum ranged is detailed in the correspondingregister. Otherwise, the corresponding resource can be used without restrictions regarding the trans-mission power. The Geo-DB also contains the final decision concerning the different resources. Theinformation contained in the Geo-DB must be periodically updated in order to take account the po-tential utilization patterns in the licensed spectrum, especially if those changes are due to the primarysystem activity. A cooperative decision taking into account the information provided by all the users ina certain range would allow increasing the sensing accuracy. An up-to-date database will reduce the col-lision probability that could be caused by the lack of synchronization between the real state of primaryspectrum and the available information stored in the database.

The cooperative decision making mechanism is implemented in the CRM. Without loss in generality,analysis here is carried out on per spectrum band basis. The LTE-A specific per-resource block (RBs)based analysis can be studied in [28] by the interested readers. The input data considered by the decisionmechanism includes all the channel state reports from the UEs served by the same eNodeB and the

11

geographical position of the UEs obtained through the aforementioned location services. Every channelstate notification creates a new entry in the CRM containing the UE identifier, the estimated geographicallocation of the UE, the channel or resource monitored, the licensed activity state sensed in that resourceand the time in seconds when the channel state report was received. Once all this data is collected fromdifferent UEs, the CRM can make decisions about the vacuity of the monitored resources at differentlocations. The collected sensing reports are classified by their distance to the eNodeB in different ranges.For each range and spectrum band, an independent cooperative decision will be made. Multiple samplesof the sensing data from different resources and UEs are combined in order to update the Geo-DB. Forthe same UE and sensed resource, only the most recently collected data is used in the cooperativedecision calculation.

There are two hard-decision rules for cooperative decision making introduced in [28]. The conserva-tive rule declares the resource as free from licensed activity if all the UEs report such state; otherwise thechannel is considered to be occupied. On the other hand, the aggressive strategy declares the resource asidle provided a single UE senses the channel as free. However, due to the above mentioned single sensormeasurement uncertainty, the different measurements reported over time must be considered in the finaldecision. That is to say, old measurements must not have the same importance in the final decision asthe newest reports due to the fast-changing radio-channel state conditions. Soft-based cooperative deci-sion stands on this idea and performs better than the above mentioned hard-decision rules. In soft-basedcooperative decision, each reported measure has an associated weight. The CRM will combine the in-formation of the spectrum band state with the weights and will make a decision by comparing the resultwith a defined threshold. Each spectrum band state notification is weighted according to the elapsedtime between the moment the notification was received by the CRM (i.e., the ith notification is receivedat time ti) and the instant when the channel state decision is made tnow as seen in (2.1). TMAX is thetime elapsed between two consecutive measures of a specific primary spectrum band. In addition to thislinear weight equation, the quadratic and the square root version of the formula is analyzed in [28].

wi =TMAX − (tnow − ti)

TMAX(2.1)

The decision regarding the resource availability for opportunistic access will be taken according tothe value of the spectrum decision metric on a given spectrum band, defined as:

U =

N−1∑i=0

(−1)diNdiwiN2

(2.2)

where di is the state of the monitored band reported in the ith notification, equal to 0, if a spectrum bandis sensed as idle or equal to 1, if a spectrum band is sensed occupied. N is the number of notificationevents considered in the decision making mechanism, including free spectrum band notifications, N0,and occupied spectrum band notifications, N1. Ndi is the number of measurements that agree withthe state of the spectrum band notified in the ith measurement report: N0 if the spectrum band wassensed occupied, N1 otherwise. As reflected by (2.2), the spectrum decision metric U will depend on

12

the number of measurements taken into consideration, the weight of each measurement and the mostreported single-sensor measurement. If most of the measurements agree on the spectrum band vacuity,a positive value of U is obtained. On the contrary, if the majority of measurements reports that thespectrum band was occupied, a negative value of U is obtained. Note that, in order to normalize theequation, it is needed to divide it by the second power ofN , in such a way that−1 ≤ U ≤ 1. Dependingon the value of U , the considered spectrum band is stated as a candidate for OSA (H0) if that result isgreater than a certain threshold, denoted as λ. Otherwise, the resource is not available for opportunisticusage (H1). The decision threshold λ must be tuned in order to provide the largest OSA probabilitywithout exceeding the interfering limit.

Decision =

H0, if U < γ

H1, if U > γ(2.3)

2.5 DoA Based Cognitive Base Station

In this section, we will present the use of Direction of Arrival (DoA) of the user equipments (UEs)at the eNodeB (eNB) for efficient opportunistic spectrum usage.

2.5.1 DoA Estimation

Accurate estimation of a signal direction of arrival (DOA) has received considerable attention incommunication and radar systems for commercial and military applications. Radar, sonar, and mobilecommunication are a few examples. There are many different super resolution algorithms for DoAestimation including spectral estimation, model based, and eigen-analysis to name a few [29–31]. Here,we concentrate the discussion on the application of estimating the DOA of multiple signals. The focusesare on a class of Multiple Signal Classification (MUSIC) algorithms known as root-MUSIC.

In this chapter, We use an uniform linear array (ULA) with Ne elements. Fig.2.5 shows the generalconfiguration for a ULA antenna having Ne elements arranged along a straight line with the distancebetween sensor elements, be h = λ/2, where λ is the incoming signal wavelength. The angle of theincoming signal, ψ, is measured relative to the antenna bore sight.

2.5.1.1 Root Music Algorithm

The root-MUSIC method relies on the following properties of the array correlation matrix: the spacespanned by its eigenvectors may be partitioned into two orthogonal subspaces, namely the signal plusnoise subspace and the noise only subspace; the steering vectors corresponding to the directional sourcesare orthogonal to the noise subspace [32]. The Ne ×Ne correlation matrix that contains K number ofincoming signals is formed by:

13

Figure 2.5 ULA antenna configuration

R = ARyyAH + σ2I (2.4)

where σ2 is the variance of the Gaussian white noise, Ryy is the signal power matrix and A is the signaldirection matrix:

Ryy = diag[P1, P2, ..., PK ] (2.5)1 1 · · · 1

e−jβ(ψ1) e−jβ(ψ2) · · · e−jβ(ψK)

......

......

e−j(Ne−1)β(ψ1) e−j(Ne−1)β(ψ2) · · · e−j(Ne−1)β(ψK)

and the phase delay between sensor elements is:

β(ψj) =2πh

λsin(ψj) (2.6)

Let λ1 ≥ λ2 ≥ ... ≥ λNe be the eigenvalues of the correlation matrix R, and v1 ≥ v2 ≥ ... ≥ vK bethe eigenvalues for ARyyAH . Then from (2.4):

λi =

vi + σ2, i = 1, 2, ...,K

σ2, i = K + 1,K + 2, ..., Ne

(2.7)

For high signal to noise ratios (SNR), vi >> σ2, the eigenvalues can be used to determine thenumber of sources that are detected by counting the number of comparatively large eigenvalues. Alter-natively, Ref. [32] suggests a more rigorous approach to determining the number of incoming sourcesthat provides better detection performance when the incoming SNR is not as high. For the purposes

14

of this, the incoming SNR is chosen to be sufficiently high as to not be in a situation where the sourcenumber detection is ambiguous.

Let q1,q2, ...,qNe be the eigenvectors associated with the decreasing ordered eigenvalues λ1 ≥λ2 ≥ ... ≥ λNe . From (6), the first K eigenvectors will span the signal plus noise subspace and theremaining Ne − K eigenvectors will span the noise only subspace, QN. By eigen-analysis, we canrepresent the Ne −K smallest eigenvectors as:

Rqi = σ2qi, i = K + 1, ..., Ne (2.8)

Using (2.8) in (2.4), can be rewritten as:

ARyyAHqi = 0, i = K + 1, ..., Ne (2.9)

Since A is a full column rank matrix and Ryy is diagonal, (2.9) becomes:

AHqi = 0, i = K + 1, ..., Ne (2.10)

or more explicitly:

α(ψk)Hqi = 0, i = K + 1, ..., Ne, k = 1, 2, ...,K (2.11)

where the streeing vector α(ψ) is given by:

α(ψk) = [1, exp(−jβ(ψk)), ..., exp(−j(Ne − 1)β(ψk))]T (2.12)

Equation (2.11) proves the orthogonality between the signal plus noise and the noise only subspaces.This is important because it shows that the angle of the incoming signals can be found by searching forsignal direction vectors that, when projected onto the noise only subspace, give a zero result. Followingthis idea, if a polynomial, J(ψ), is constructed such that:

J(ψ) = α(ψ)HQNQHNα(ψ) = 0 (2.13)

The roots of J(ψ) contain the directional information of the incoming signals. Ideally, the rootsof J(ψ) would be on the unit circle at locations determined by the directions of the incoming signals;however, due to the presence of noise the roots may not necessarily be on the unit circle. In this case,the K closest roots to the unit circle are the roots that correspond to the K incoming signals [33].Theseselected roots, by themselves, do not directly represent the incoming angle. For each root, the incomingangle is found by solving (2.13).

ψj = arcsin[ λ

2πharg(ψj)

](2.14)

15

Obviously, when the root-MUSIC algorithm is implemented, there is no prior knowledge of the in-coming signal directions or signal powers needed to construct the correlation matrix using (2.4). There-fore the correlation matrix must be estimated using only the information available from the antennaarray. There are several methods commonly used to perform this estimation such as temporal averaging,spatial smoothing or, a hybrid combination of both temporal averaging and spatial smoothing [34]. Inthis work, we use only the temporal averaging method.

The estimated correlation matrix using the temporal averaging method with k snapshots is given as:

Θ = E[Y HY ] (2.15)

where the incoming data matrix is:x1(1) x1(2) · · · x1(k)

x2(1) x2(2) · · · x2(k)...

......

...xNr(1) xNr(2) · · · xNr(k)

with xi(k) being the ith sensor output at time k.

The estimated correlation matrix,Θ, asymptotically approaches the correlation matrix, R, as thenumber of snapshots increases. Therefore, in order to have an accurate estimation of the correlationmatrix the observation time must be sufficiently long. The long observation times are not ideal forradar signal processing applications; however, there are many applications where this does not pose aproblem. Correlation matrix estimation techniques like the spatial smoothing method are better suitedfor use in time sensitive systems.

2.5.2 DoA based Cognitive Base Station (CBS)

In this section, we discuss an application of DoA in developing a cognitive base station. We herefocus on using the DoA of the incumbent to judge the vacancy of a spectrum. The conventional spectrumsensing technique is not considered under the assumption that the system is using highly directionalantennas. Thus, the CCI depends on the proximity of the UEs. The problem of CCI arises, when themobile UEs during their course of movement attend a position where they are aligned in the samedirection w.r.t eNB. The solution, we have proposed to this problem is the spectrum handoff [35]. Thereare two cases resulting in CCI:

• When the PU is aligned in the direction of other SUs due to its mobility. In this case, PU remainsin their licensed spectrum band, whereas the interfering SUs switch to other available unusedspectrum bands. The unused spectrum bands can be found in the list of backup channel as men-tioned in the previous section.

• When two or more SUs come close enough, it causes co-channel interference. In this case, thespectrum band of one of the SU remains unchanged, but the spectrum band of other SU changes.The SU with unchanged spectrum is decided according to some priority protocol.

16

Figure 2.6 GUI model demonstrating the CR nature of the proposed architecture a) when PU at 20o andSU at 40o b) when both PU and SU at 40o

The priority protocol can be designed in a number of ways. One of the methods is to keep track ofthe time period for which the SUs are connected to the BS. Top priority is given to the SU which hasthe largest time period. Another method is to assign top priority to the SU which have lowest batterypower (lifetime). The priority can also be decided according to the type of service carried out by theSU. For example, SU involved in live streaming is given higher priority as compared to SU transmittingvoice data. The SU with top priority is the user with the unchanged frequency during spectrum handoff,provided there is no PU to compete with. PU always has the highest priority. The priority level isdecided and maintained by the eNB. Once it is decided which UE would undergo spectrum handoff.The eNB controls all the spectrum handoff activities.

The spectrum handoff successfully eliminates the CCI, thereby, improving the quality of service(QoS). In fig.2.6, a GUI demonstrates the proposed method for a case of coexistence of a PU and a SUwhere both are transmitting unmodulated sinusoidal signal of frequency 8 KHz. In the fig.2.6a, we cansee that the PU and SU coexist at same operating frequency, at 2 GHz, as they have largely separatedDoA. The PU is at 20o and SU at 40o DoA. When they both come close to each other, it is seen in theFig.2.6b that the operating frequency of SU switches to the operating frequency of 3 GHz.

2.6 Summary

In this paper, we have introduced a cognitive radio technology and described features like licensedspectrum access and Geo-location database. We also discussed the process of estimation of DoA byROOT MUSIC algorithm. Finally, a DoA based cognitive base station is presented which used the DoAof the incumbent to allow opportunistic use of the licensed spectrum by the SU.

17

Chapter 3

Self-Interference Cancellation: A Survey

3.1 Overview

The problem of spectrum scarcity has led the researchers to look for solutions to meet the extensivedemand of bandwidth in future cellular networks [5]. One of the major revolutions in the future wirelessnetworks can be the introduction of full-duplex (FD) eNodeB (eNB) and user equipment (UE). The FDoperation has the capability of cutting down the spectrum requirement by half. The main challengeencountered in implementing an FD wireless device is the large power difference between the self-interference (SI) imposed by the device’s own transmissions and the signal of interest received from aremote source.

In recent years, extensive work is being done in the area of self-interference cancellation (SIC) designfor both single and multiple antenna transceiver units [1,4,15,36–46]. Most of the recent research on FDcommunication considers an FD eNB and multiple half-duplex (HD) UEs [45–48]. The computationalcomplexity of SIC circuitry limits the use of FD UEs. However, the SIC designs proposed in [1,4,40,45],have the capability of eradicating this limitation. Using these SIC techniques, simultaneous transmissionand reception on the same spectrum resource is possible using a single antenna, as opposed to usingtwo antennas. The design works for large bandwidths and high data rates. Here, the SI cancellationtechniques are described by classifying it into three categories, namely passive suppression, analogcancellation and digital cancellation.

Note: For further reading on latest of FD transceiver design, readers can refer to [2]. In [2], ananalysis is carried out on the main impairments (e.g. phase noise, power amplifier nonlinearity as wellas in-phase and quadrature-phase (I/Q) imbalance, etc.) that constitutes the SI. Also, the FD based MediaAccess Control (MAC)-layer protocol design is discussed for the sake of addressing some of the criticalissues, such as the problem of hidden terminals, the resultant end-to-end delay and the high packet lossratio (PLR) due to network congestion. The potential solutions conceived for meeting the challengesimposed by the aforementioned techniques are covered. Furthermore, a range of critical issues relatedto the implementation, performance enhancement and optimization of FD systems, including importanttopics such as hybrid FD/HD scheme, optimal relay selection and optimal power allocation, etc are

18

described. Finally, a variety of new directions and open problems associated with FD technology arepointed out in [2].

3.2 Self-Interference Cancellation

The goal of FD radio is to simultaneously transmit and receive within the same frequency band, inwhich case an FD node receives not only the signal of interest, but also the signal it is transmitting,which constitutes the SI imposed upon the receive antennas (RAs). Since the strength of the SI signalobserved in FD devices may be 50 − 100 dB higher than that of the signal of interest, the stronger SIsignal will govern the gain control settings of the AGC, which scales the input signal prior to digitizationto the normalized range of [-1, 1]. If the SI power is high, it constrains the weak signal of interest tooccupy a range much smaller than [-1, 1], hence invoking a high quantization noise on the signal ofinterest as well as a significantly eroding the SINR in the digital baseband [49].

To resolve the above-mentioned problem as well as to exploit the potential FD gains, the SI has to bereduced in strength before decoding the signal of interest [50]. For example, in a scenario relying on aFD radio having a transmit power of 0 dBm and a noise floor of approximately−90 dBm, the RAs haveto be capable of reducing the SI by nearly 95 dB so as to ensure that the FD node’s own transmissions donot unduly contaminate its reception [15]. As indicated in [41], the goal of SI cancellation is to predictand model the distortions in order to compensate for them at the RAs. However, SI cancellation is by nomeans a simple linear operation, because the conventional assumption that “the radio signal preservesits original baseband representation except for power scaling and frequency shifting” turns out to bepartially incorrect [1]. To elaborate, in practical systems, the FD radios may distort the transmitted sig-nal’s digital baseband representation. Explicitly, both linear distortions (induced by signal attenuationsand reflections from the environment, etc), as well as non-linear distortions (induced by circuit powerleakage, non-flat hardware frequency response, higher-order signal harmonics, etc), the noise 1 imposedby the imperfect transmit power amplifiers and phase noise 2 generated by local oscillators are imposed.For example, in a typical Wi-Fi radio using 80 MHz bandwidth and a receiver noise floor of -90 dBm aswell as the transmit power of 20 dBm, the SI signal comprises the following typical components [1]:

• The linear (main) component of 20 dBm strength, corresponding to 110 dB above the noise floor;

• The non-linear component of -10 dBm strength, corresponding to 80 dB above the noise floor;

• The transmitter noise of -40 dBm strength, corresponding to 50 dB above the noise floor, asgraphically illustrated in fig.3.1.

1It was experimentally observed to be around the level of -50 dBm, i.e. 40 dB higher than the receiver noise floor level of-90 dBm [51].

2It is typically of the order of -40 dBm [1].

19

Figure 3.1 To provide a sufficiently high SI cancellation capability, an FD radio must be capable ofcancelling 110 dB of linear component, 80 dB of non-linear component as well as 60 dB of analogcancellation [1].

In order to suppress the SI power to a level below the noise floor, the above mentioned distortionsmust be adequately mitigated, whilst simultaneously considering the impact both of random transmitternoise and that of the ADC resolution. Explicitly, the FD devices must be capable of providing 60 dBof analog-domain cancellation plus 50 dB of digital-domain cancellation in order to reduce the SI tothe receiver noise floor. However, if by any chance the analog- and/or digital-domain cancellationssuffer from some performance degradations due to hardware imperfections and/or other impairments,their combined cancellation may not meet the decoding requirement. To mitigate the above-mentionedrequirements as well as to mitigate the analog-/digital-domain requirements, a method referred to aspassive suppression [3] can also be invoked for reducing the SI prior to reaching the RAs by exploitingthe path-loss effect between the transmit antennas (TAs) and RAs of an FD node.

In this section, the SI cancellation techniques are classified into passive suppression, analog cancel-lation and digital cancellation, as described in fig.3.1. The family-tree of SI related techniques is seenin fig.3.2. According to the order of execution of different SI suppression/cancellation modules, thepassive suppression techniques are introduced in the next subsection, followed by analog and digitalcancellations.

20

Figure 3.2 Techniques related to SI measurement and suppression. [2]

3.2.1 Passive Self-Interference Cancellation

Passive SI suppression is defined as the attenuation of the SI signal contributed by the path-loss effectdue to the physical separation/isolation between the TAs and RAs of the same node [52]. By reducingthe electromagnetic coupling between the TAs and RAs at the FD node, the power of SI can be reducedprior to its arrival at the RAs, as illustrated in fig.3.1. Numerous methods of passive SI suppressionexist [53–55]. For example, in a multi-antenna based system, the polarization decoupling techniqueenables the TAs and RAs to operate with the aid of orthogonal horizontal and vertical polarizationsfor the sake of reducing their coupling. Furthermore, passive suppression may rely on beamforming-aided techniques for directing the lobes of TAs and RAs in different directions [54], hence resultingin improved physical separation between the TAs and RAs [53]. Additionally, by employing isolationcomponents such as circulator-like devices, the transmit and receive paths of a single FD antenna canalso be isolated, providing an equivalent SI-attenuation effect [55]. In the rest of this subsection, variouspassive suppression techniques will be surveyed.

• Antenna Separation Based Passive Suppression: The simplest method for achieving passivesuppression might be resorted to the Antenna Separation (AS) technique, because in practicalsystems, increasing the pathloss effect between the TAs and RAs constitutes an effective approachto attenuate the SI signal. Consider a system, in which each node is equipped with a TA and RA, alarger TA-RA separation implies having a higher SI suppression capability. In [56], Haneda et al.,studied outdoor-to-indoor communication system operating at the center frequency of 2.6 GHz,where compact relay antenna was developed for serving as a signal repeater between the outdoorbase stations (BSs) and indoor users. In this compact relay station the TAs and RAs are attachedto the opposite sides of the physical construction for FD operation, whilst facilitating both themeasurement and suppression of the SI. The results revealed that the isolation between the TAs

21

Figure 3.3 Block diagram of antenna cancellation for a wireless FD SI cancellation. The power splittersintroduce a 6 dB reduction in signal, thus power from TX1 is 6 dB lower compared to power from TX2,without the need for an additional attenuator to compensate for the amplitude mismatch [3].

and RAs measured in a multipath environment was 48 dB for the compact relay antenna used andthis could be improved by further separating the TAs-RAs, while simultaneously optimizing theorientation of the antenna arrays. Furthermore, it was also shown that a 70 dB isolation can beachieved for a TAs-RAs separation of 5 m, while ensuring the best possible antenna orientation.Although this isolation level may still be insufficient for practical FD operation, especially forcompact relays, the employment of an interference canceller is capable of further increasing theamount of SI cancellation.

• Antenna Cancellation Based Passive Suppression: The basic philosophy of antenna cancella-tion (AC) is to employ two TAs and a single RA, where the pair of TAs is placed at distances ofd and (d + λ/2) away from the RA, respectively, with λ representing the wavelength [3]. TheRA is positioned by satisfying that its distance from the TAs differs by an odd multiple of λ/2,which results in the transmit signals being destructively superimposed for cancelling one another,as illustrated in fig.3.3. The destructive interference becomes most effective if the signal powersimpinging at the RA from the pair of TAs are identical, thus (in theory) creating a null at the po-sition of the RA. It has been shown in [3] that antenna aided cancellation techniques are capableof achieving an SI suppression of about 30 dB, and in conjunction with existing RF interferencecancellation [57] as well as digital baseband SI cancellation [58], a cancellation capability as highas 60 dB can be achieved.

22

• Directional Passive Suppression: Directional SI suppression constitutes a technique, where themain radiation lobes of the TAs and RAs of an FD node have minimal intersection [54]. Theexperimental results of [54] demonstrated that the FD mode significantly outperforms the HDmode, when relying on passive SI suppression combined with the active SI cancellation. In ascenario, where the TA-RA distance is assumed to be 10 m and the antennas are separated by anangle of 45o or more, the FD gain 3 over the HD mode may range from 60% to 90%. This gainbecomes 50% or more for an antenna distance of 15 m and for an angle ranging from 90o to 150o.

In brief, the best antenna configuration in terms of the attainable SI suppression relies on in-stalling the TAs and RAs at the opposite sides of the device in order to create the highest possibleseparation [59]. However, optimizing the antenna configuration of compact devices remains chal-lenging. Hence, there has to be a combination of passive and active suppression/cancellationtechniques for facilitating a better SI reduction in practically FD systems.

3.2.2 Analog Self-Interference Cancellation

Based on the above-mentioned discussion, it is found that the amount of SI reduction relying onthe pure passive suppression technique is insufficient for supporting high-integrity FD reception 4. Inorder to reduce the SI below the noise level, additional active cancellation techniques have to be invokedfor further reducing the residual SI after passive suppression. Hence, the objective of the additional SIcancellation modules is to minimize the SI either within the RF [3] or in the analog/digital basebandstage.

A strong SI signal would saturate the AGC, which is hence desensitized for the reception of a weakdesired signal compressed to a range much smaller than [-1; 1]. In this case, the ADC that becomesimpact of the extremely strong SI power. More explicitly, the quantization noise contaminating thedesired signal might become excessive, hence resulting in a negative effective SINR that would becomeinadequate for recovering the desired signal in the digital baseband [15]. As indicated by [60], thelimitations of the ADC, such as its estimated dynamic range and quantization resolution constitute themain obstacle in improving the achievable SI-isolation levels by employing digital cancellation.

Therefore, it is critical to further reduce the power of the SI signal prior to the digitization 5 of thedesired received signal. Specifically, a mechanism referred to as analog cancellation has to be invokedfor mitigating the SI contaminating the analog signal before it is digitized. After performing analogcancellation, the decontaminated received digital samples will exhibit a sufficiently high resolution ofthe desired received signal, thus facilitating efficient digital SI cancellation [3], as depicted in fig.3.1. Inthis subsection, a range of beneficial analog cancellation techniques will be surveyed, followed by thefamily of digital cancellation techniques.

3The FD gain as compared to the HD mode can be evaluated in terms of data rate, capacity, BER, and outage probabilityimprovement, etc.

4The signal received at the RAs will be first amplified by an AGC and then down-converted to the baseband/intermidiatefrequency, followed by filtering and sampling before the ADC to create the digital samples.

5Before performing digitization, the AGC scales the input to the normalized range of [-1; 1].

23

• Analog Cancellation for Reducing the Linear SI Component: In this section, the focus is onthe fundamentals of analog cancellation by elaborating on the reduction of the linear SI com-ponent, which constitutes the majority of the SI power, leaving the dynamic adaptation basedsolutions guarding against the non-linear components encountered in time-variant environmentsfor further study in the next part.

The principle of analog cancellation can be simply summarized as follows: In order to sufficientlyreduce the SI power, an FD radio is required for creating a reference signal corresponding to aperfect replica of the SI signal at all instances. Combining this replica and the SI signals is intheory capable of facilitating perfect SI cancellation [15]. Basically, the analog cancellation canbe performed either at the RF or at the analog baseband stage [49]. However, most of the existinganalog cancellation (e.g. [3, 15, 53]) techniques operate at the RF. Furthermore, by identifyingwhether the perfect replica based SI canceling signal is generated by processing the SI prior toor after up-conversion, the RF-based analog cancellation arrangements may be further classifiedas pre-mixer (e.g. [53]) or post-mixer schemes (e.g. [15]). The baseband analog canceler, onthe other hand, is defined as the canceler, in which the perfect replica based canceling signal isgenerated in the baseband and the cancellation occurs in the analog baseband [49].

Based on the above-mentioned principle, the operation of analog cancellation can be realized byexecuting the following three steps, including:

– Creation of SI-Inverse Signal: Basically, SI inversion can be implemented by an FD radioupon simply inverting a signal by inverting its phase. However, this phase adjustment mayonly be feasible across a limited bandwidth, which hence limits its maximum cancellationcapability. In other words, a perfect signal inversion can be attained at the central frequency,but the inverted signals will deviate at both sides of the central frequency from 180o, hencesuffering from a significant phase-distortion. To address the above-mentioned problem, in[2], focus is on sophisticated hardware/circuit design relying on:

∗ A balanced/unbalanced (balun) transformer, which is a common component in the RF,audio and video circuits, can be utilized for perfectly (in theory) converting back andforth between an input signal and its inverse at all instances [15]. As illustrated infig.3.4, the TA is assumed to transmit the positive signal. The balun output of RF ref-erence, which is subject to an adjustment on the delay and attenuation of the referencesignal, highly matches the SI signal at the RA, thus offering a reliable SI nulling bycombining the received SI signal with its negative version.

∗ Apart from that, another method of generating the RF reference signal is to view the SIcancellation as a sampling and interpolation problem, which can be resolved relying onthe delay-line based analog circuit [1]. By picking up the phase and amplitude of theSI signal (e.g. relying on the Nyquist sampling theorem), the SI signal can always bereconstructed at any instant as a weighted linear combination of samples taken before

24

Figure 3.4 Block diagram representation of analog and digital SI cancellations, in which the SI invertis executed by employing balun circuit, followed by QHx220 based delay & attenuation adjustment [2].

and after the recreation instant, with the weights of the linear combination determinedby using the so called since interpolation algorithm. A fundamental tradeoff betweenthe hardware complexity (i.e. in terms of the number of delay lines) and the cancellationcapability must be treated.

– Delay and Attenuation Adjustment: Since the signal transmitted over the ether experiencesboth attenuation and delay in all practical scenarios, an identical attenuation and delay hasto be applied to the inverted SI. The QHx220 noise cancellation chip [3], separates the SI-inversion-based RF reference signal into its in-phase and quadrature components (i.e. gi andgq), can be invoked for imposing an adaptively controllable delay on the aggregated outputsignal by carefully controlling the attenuation of those components. It is shown in [15] thatthe balun-aided cancellation is capable of achieving an impressive SI reduction across awide bandwidth, provided that both the phase and the amplitude of the inverted SI signal areset appropriately.

– Creating an SI-Null by Combining the SI and Its Inverse: The SI-inverse signal will thenbe combined with the SI signal at the RA. Without loss of generality, the Received SignalStrength Indicator (RSSI) values can be employed for representing the residual SI energyremaining after combining, as illustrated in fig.3.4. In theory, a perfect SI-inverse signal willresult in zero SI value at the output of the RSSI. However, the realistic, practical engineering

25

imperfections of the hardware components, such as power leakage or a non-flat 6 frequencyresponse at balun, will always result in a residual SI power 7 after signal combining, whichcan be minimized by carefully adapting the attenuations (gi and gq) using self-tuning algo-rithms [15].

The existing studies have demonstrated that analog cancellation techniques are capable of reduc-ing the SI strength by dozens of dBs [1, 3, 15]. For example, in order to facilitate FD communi-cation at the transmit powers typical of Wi-Fi devices [54], it was shown experimentally [61] thatthe SI-induced contamination imposed on a modulated constellation point can be significantlyreduced by using an interference canceler accommodated within the RF stage. Furthermore, asindicated in [15], the SI-inverse technique alone is capable of reducing the SI by no less than 45dB across a 40 MHz bandwidth.

Nonetheless, subtracting the SI from the received signal by simply relying on the above-mentionedSI inversion technique remains a challenge in practical systems, because the FD radio only knowsthe “clean” digital representation of the baseband signal, rather than its processed counterparttransmitted over the air. Once the signal is converted to the analog domain and upconvertedto the carrier frequency for transmission, the transmitted signal becomes an unknown non-linearfunction of the ideal source signal contaminated by unknown distortions induced either by the im-perfections of the analog components in the radio transmit chains (e.g. the third- and higher-ordersignal components created by the analog circuits, the transmitter noise due to the non-linearityof the power amplifiers, and the inaccuracy of the oscillators, etc) or by their non-flat frequencyresponse [1]. In other words, the SI cancellation circuits that simply subtract the estimate of thetransmit signal without taking into account all the non-linear distortions fail to perfectly cancelthe SI by reducing it below the noise level. As indicated by [1], no more than 85 dB of SI powerreduction can be achieved by FD designs that fail to account for the non-linear distortions. Tomake up for the deficiencies of the above-mentioned techniques, the non-linear SI componentsinduced either by hardware imperfections or by the time-variant environment has to be carefullyconsidered in designing the analog cancellation circuits.

• Dynamic Adaptation of Analog Cancellation to Remove Non-Linear SI Components inTime-Varying Environments: While the above mentioned analog cancellation schemes are capa-ble of effectively dealing with the linear SI components, a time-varying environment encounteredin the presence of channel fading, transmit power and other parameter fluctuations may impose asignificant non-linear distortion based contamination on the cancellation [1]. More importantly,

6For example, the balun circuit is not frequency flat and inverts different parts of the bands with different amplitudes, thusapplying a single attenuation and delay factor to invert the SI signal will never achieve a perfect cancellation [15]. Furthermore,the QHx220 module may also suffer from a non-linear distortion, resulting in imperfect SI cancellation for typical wirelessinput powers (0-30 dBm) [15].

7In practical designs, the combining-output energy can be further reduced in the digital-domain relying on digital cancel-lation techniques.

26

as the environment changes, the cancellation capability may drop to an inadequate level, becausethe already optimized SI cancellation parameters based on the past environmental conditions mayno longer correctly model the current SI.

To avoid the above-mentioned imperfections and provide a satisfactory cancellation performance,the FD radio must be capable of promptly tuning the analog circuit in order to adaptively respondto time-variant environments. Specifically, an adaptive scheme acting in response to the channelfluctuations must be conceived to equip the cancellation circuits with the capability of frequentlyand promptly refreshing its parameters (e.g. phase and amplitude of the SI-inverse-based RFreference signal) [1, 15]. In practical systems, both the time and frequency domain solutions maybe invoked in order to combat the non-linear SI components in the time varying environment.

– Time-Domain Solutions: In [1, 15], a method for quickly tuning the analog circuit is pro-posed. For a given time-domain reference signal c(t), the corresponding received SI signaly(t) can be modeled as a summation of weighted reference samples at different delays, i.e.

˜y(t) =∑N

i=1 αic(t − di), where N denotes the maximum number of taps, α1, α2, ..., αN

each represents the attenuation corresponding to one delayed component, and d1, d2, ..., dN

stand for delays associated with the taps, as shown in fig.3.5. The goal of the proposed tun-ing is to adaptively change α1, α2, ..., αN such that the remaining SI power is minimized,i.e:

minimizeα1,α2,...,αN

(y(t)− y(t))2 (3.1)

The above-mentioned equation can be solved by using the so called Iterative Gradient De-scent Algorithm [15]. Although this algorithm is simple, the extremely low convergencespeed substantially constrains its practical application. It was shown in [1] that the algorithmrequires nearly 40ms to converge, which cost is pretty high for practical systems (i.e. cor-responding to a 40% overhead in practical scenarios 8 that require the analog cancellation tore-tune once every 100ms). Fortunately, this high tuning cost can be substantially reduced(i.e. to about 920µs, as experimentally shown in [1]) by executing the initial settings ofthe attenuators relying on some known sequences such as the Wi-Fi preamble, followed byfinding the optimal convergence point after running a few gradient descent iterations.

– Frequency-Domain Solutions: The frequency-domain SI signal can be modeled as a functionof the tapped signal c(t) as [1]:

Y(f) = H(f)C(f) (3.2)

8Note that the re-tune period is environment dependent. In [1], the “near field coherence time” of analog cancellation isdefined to specify the time up to which the receiver remains unsaturated from the last time tune. This time duration can beused to trigger the return of the tuning algorithm.

27

where H(f) denotes the frequency-domain SI distortion induced by various factors suchas the antenna, circulator and reflections, and C(f) represents the frequency-domain rep-resentation of the tapped signal (i.e. the discrete Fourier transform (DFT) of c(t)). Similarto the time-domain solution, the frequency-domain SI-cancellation problem can be readilyformulated as:

minimizeα1,α2,...,αN

(H(f)−

N∑i=1

Hαii (f)

)2(3.3)

where Hαii (f) denotes the frequency response for attenuations αi. In theory, an exhaustive-

search scheme can be implemented to make (3.3) converge to its optimal point. However,since achieving an optimal point is an NP hard problem, a substituted solution can be pro-posed by looking for a sub optimal point that enables the circuits to provide the requiredcancellation performance. It was shown experimentally in [1] that the convergence durationof the frequency-domain quick-tuning algorithm is no longer than 900 − 1000µs, corre-sponding to less than 1% overhead for analog cancellation that performs re-tuning onceevery 100ms.

In summary, the achievable SI cancellation capability may remain limited and in fact insufficientfor high integrity detection, when relying on stand-alone analog cancellation. To offer a sufficientlyhigh cancellation capability (i.e. to make the resultant SINR high enough for high-integrity detection),digital-domain cancellation combined with analog cancellation must be employed for further mitigatingthe residual SI in the digital baseband.

3.2.3 Digital Self-Interference Cancellation

As indicated by [15], although an industry-grade balun circuit is capable of reducing the SI by asmuch as 45 dB for a 40 MHz wide SI signal, the remaining SI power may remain by up to 45 dBhigher than the noise floor (in the absence of employing passive suppression). This still excessivelyinterferes with the desired signal, either because of the residual multipath SI echoes contaminatingthe desired signal or because of the SI leakage imposed by the imperfections of the hardware circuits.Evidently, the residual SI after analog cancellation must be further reduced in the digital domain. Digitalcancellation constitutes an active SI-mitigation mechanism that by definition operates in the digitaldomain and exploits the knowledge of the interfering signal in order to cancel it after the received signalhas been quantized by the ADC [49, 62]. To achieve this, the receiver first extracts the SI and thenremodulates it and subtracts it from the received SI contaminated signal. Coherent SI-detection canalso be employed for recovering the SI by correlating the received signal with the clean hypothesizedregenerated SI-inversion based reference signal, which is available at the output of the co-located FDtransmitter [3]. This technique, then requires the receiver to estimate both the delay and phase shift

28

Figure 3.5 Circulator aided FD radio block diagram with analog and digital cancellation stages, inwhich Tb denotes the intended baseband signal for transmission, while T is the practically transmittedRF signal. For an intended receive signal R, it will be contaminated by the strong components partiallydue to the undesirable leakage of the circulator [1].

between the transmitted and the received signals relying on techniques, such as the correlation peakbased algorithm for subtracting the SI signal.

Digital cancellation can be regarded as an excellent, safety-net solution for diverse scenarios, whenanalog cancellation achieves a poor suppression [3, 53]. However, since the transmitted packets aredifferent from the generated reference signal due to a number of factors such as the hardware limita-tions and the multipath fading, subtracting the estimated signal rather than the clean signal would becapable of substantially improving the capability of digital cancellation. In practice, digital cancellationfundamentally comprises two main components, i.e. estimating the SI channel, and using the chan-nel estimation on the known transmit signal to generate digital samples for subtracting the SI from thereceived signal [15].

In order to implement the digital cancellation to eliminate the residual SI power after analog can-cellation, the SI channel components comprising both the leakage over through the analog cancellationcircuit and the delayed reflections of the SI signal from the environment must be estimated [1]. Basi-cally, the residual SI can be sub-divided into linear and non-linear components. The former constitutesthe majority of the SI power and can be estimated by existing algorithms, such as the family of least-square and MMSE [63] based techniques, while the latter is induced by the non-linear distortions of theimperfect analog cancellation circuits. For example, the QHx220 hardware [15] employed in the balunbased analog cancellation scheme may cause non-linear distortions, particularly for high input powers

29

beyond say -40 dBm. Consequently, the non-linearity of the SI leakage channel must be accuratelycharacterized for the sake of high-rejection SI cancellation in the digital domain. In practical designs,the following techniques can be used for estimating the linear and nonlinear components [1].

• Estimation of the Linear Component of the SI Leakage-Channel: By modeling the linearcomponents of the SI as a non-causal linear function of the transmitted digital signal x[n], which isknown in advance, the received sample y[n] at any instant can be modeled as a linear combinationof up to k samples of x[n] before and after the instant n, where k > 0 is a function of the SIleakage channel memory:

y[n] =

k∑z=l−k

x[n− z]h[z] + w[n] (3.4)

where h[n] and w[n] represent the SI channel attenuation and the additive noise component atinstant n, respectively. By defining y = [y[0], y[1], ..., y[n]]T , h = [h[−k], ..., h[0], ..., h[k−1]]T

and w = [w[0], w[1], ...w[n]]T , where xT denotes the transpose of vector x, the SI channel vectorh can be estimated as follows [1]:

h = (AHA)−1AHy (3.5)

where A=

x[−k] · · · x[0] · · · x[k − 1]

.... . .

.... . .

...x[n− k] · · · x[n] · · · x[n+ k − 1]

and AH denotes the Hermitian transpose

of the matrix A. Since the training matrix A can be pre-computed, the computational complexityof the above mentioned algorithm can be substantially reduced.

• Estimation of the Non-Linear Components of the SI Leakage-Channel: After estimating andeliminating the linear components of the SI signal, the residual nonlinear components can befurther reduced. As indicated by [1], the power of the residual non-linear components is about20 dB higher than the noise level. Since the exact non-linear function that an FD radio appliesto the baseband transmitted signal is hard to estimate, a general model relying on Taylor seriesexpansion can be employed for approximating the non-linear function in the digital basebanddomain [1]:

y[n] =∑

m∈oddterms,n=−k,...,kx[n](|x(n)|)m−1.hm[n] (3.6)

In this only the odd-order terms correspond to nonzero energy in the frequency band of interest,as revealed in [1]. Evidently, the first term is the linear component corresponding to the major-ity of the SI power, which can be estimated and canceled using the algorithm proposed in (3.5).

30

Furthermore, it was found in [1] that in practice the higher order terms of (3.6) constitute a cor-respondingly lower power, because those terms are created by the mixing of multiple lower-orderterms, where each mixing operation reduces the combined power. Hence, only a limited numberof terms have to be considered in implementing practical SI leakage-channel estimation.

In summary, to carefully tackle the above mentioned challenges and improve the aggregated capabil-ity of the concatenated analog-digital cancellation, several critical problems have to be addressed. Onone hand, an effective balance needs to be maintained for the tradeoff between analog and digital can-cellations so as to attain the best possible aggregated cancellation in the overall solution. On the otherhand, the cancellation capability of any individual technique (e.g. passive suppression, analog or digi-tal cancellation) should be further improved. To achieve this ambitious goal, sophisticated algorithms,such as the SI model relying on Taylor series expansion [1] for mitigating both linear and non-linear SIcomponents in the analog- and/or digital-domain, can be the solution.

3.3 Full-Duplex MIMO Communication

Several efforts are now underway to include FD technology in future cellular 5G standards, as wellas explore applications of the technology in current wireless infrastructure. However, these efforts arehampered by the fact that there aren’t viable and efficient FD designs that can work in conjunctionwith MIMO. Specifically, no current practical designs are known which can enable one to build a Mantenna FD MIMO radio that can transmit and receive from all antennas at the same time and double thethroughput. The best known prior MIMO FD system, MIDU [64] requires 4M antennas for building aFD M antenna MIMO radio, and even then fails to provide the needed self-interference cancellation forWi-Fi systems (20 MHz bandwidth) to achieve the expected doubling of throughput.

Recent work has, however demonstrated that a single antenna (SISO) FD system is practically pos-sible [1]. Specifically, it demonstrates the design and implementation of a cancellation system for aSISO system that completely cancels self-interference to the noise floor and consequently achieves thetheoretical doubling of throughput. A natural question therefore is why not just replicate the same de-sign M times to build a MIMO M full duplex radio? After all, a MIMO radio can be conceptually andphysically viewed as a collection of M single antenna SISO radios.

The challenge is cross-talk interference. When a FD MIMO radio transmits, the transmission fromany one of the M antennas (interchangeably referred to as transceiver chains) propagates to the otherantenna (chains) and causes a large amount of interference. For the sake of clarity, here the self-interference at a receive chain caused by a transmission from the TX-chain with which the receivechain shares an antenna is refereed as “self-talk”, and the interference from a neighboring TX chainstransmission as “cross-talk”. Since MIMO antennas are closely spaced due to size constraints, thiscross-talk is extremely strong, almost 75-80 dB stronger than the desired signal that is being received onthat chain. Consequently, even if the cancellation circuits and algorithms cancel every chain’s self-talk,there will be an extremely strong cross talk interference that can saturate the receive chain.

31

In [4], a design and implementation of a MIMO Wi-Fi full duplex radio is presented. TheM antennaFD MIMO radio uses each antenna for simultaneous transmit and receive, i.e., it uses the same numberof antennas as a standard half duplex M antenna MIMO radio unlike prior designs. The design usesslightly more than M cancellation circuits and DSP algorithms to cancel all the self and cross talks. Inother words, complexity scales linearly with the number of chains, which is the best performance onecould expect. Further, the performance does not degrade linearly with the number of MIMO chains,i.e., the residual interference is the same as the SISO design and does not increase linearly with thenumber of chains. The design is prototyped and integrated with the off-the-shelf WARP software radiosrunning a stock Wi-Fi baseband and demonstrated experimentally that it achieves close to the theoreticaldoubling of throughput.

The design solves the key challenge of efficiently and effectively achieving the MIMO full duplexusing two major ideas as follows.

• First, a key insight is that MIMO chains are co-located, i.e., “they share a similar environment”.Intuitively, the signals transmitted by two neighboring antennas (separated by a few cm) gothrough a similar set of reflectors and attenuation in the environment. Cancellation systems areessentially trying to model these distortions, so for modeling cross-talk, the work that has beendone for modeling the chain’s own self-talk interference can be reused. This results in a novel“cascaded” filter structure for cancellation that results in an overall design that has near-linearcomplexity scaling with the number of MIMO antennas.

• Second, the reason performance degrades linearly with the SISO replication based design is thateach of the M independent cancellation algorithms for self-talk and cross-talk at a receive chainproduce their own estimation error which add up to the linear degradation. The key insight is toleverage the fact that the M transmitters are available that can concurrently send training sym-bols. Specifically, a training preamble is designed that allows each receive chain to estimate eachof the self-talk and cross-talk channels at an error that is M times lower than the SISO designby combining information from all M training symbols. Consequently, in the design when theestimation errors add up for the self-talk and cross-talk cancellations, the overall error or residueis the same as a SISO system would have achieved, which is the best one can hope for.

The resultant is the cascaded design, shown in fig.3.6, capable of efficient FD communication. In [4],experiments demonstrate that in a 3×3 configuration, the proposed system cancels the self-interference(including the cross-talk) to the noise floor, which can be observed in fig.3.7. Also, the system achievesa 95% throughput gain over half duplex radios using a standard Wi-Fi compliant OFDM PHY of 20MHz for 802.11n for all different modulations (BPSK, QPSK, 16QAM and 64 QAM) and coding ratesof (1/2, 2/3, 3/4, 5/6), supporting three streams for 3× 3 MIMO.

Note: In the next chapter, we discuss the FD communication for multiple antenna eNB and UE andthe corresponding uplink and downlink operations. We assume perfect self-interference cancellation ateNB and UE transceiver circuits. While this is far from true today, sufficient progress is being made in

32

Figure 3.6 Cascaded Cancellation Design:Shows a 3 antenna full-duplex MIMO radiodesign with cascaded filter structure for can-cellation [4].

Figure 3.7 Spectrum plot after cancellation ofvarious self-talk and cross-talk components forRX1 of a 3 × 3 full-duplex system using ourdesign [4].

this direction to start considering this model and its implications, especially in case of small cells, wherethe transmission power varies from 17 dBm to 30 dBm. Further, in [44], two SIC designs are proposed,which allow its integration to compact radios. The first design combines a dual polarized antenna with aself-tunable cancellation circuit and targets devices like small-cell base stations and tablet computers. Inthe second design a tunable electrical balance isolator/duplexer is combined with a single-port miniatureantenna. The balance circuit can be implemented in a CMOS technology, facilitating low cost and denseintegration. Hence, this can be integrated in device like smartphones sensor nodes. The two designs areshown to provide a 75 dB and 50 dB of isolation at 2.45 GHz over a signal bandwidth of 10 MHz and6 MHz, respectively. The performance can further be increased by incorporating digital cancellation.These designs are shown in fig.3.8 and fig.3.9.

3.4 Summary

The available radio spectrum is limited, hence before new commercially implementable spectral re-sources are exploited, the ever increasing throughput requirements cannot be readily satisfied withoutincreasing the achievable spectral efficiency expressed in bits/s/Hertz. The main driving force behindFD techniques is the promise of nearly doubling the data rate in comparison to their HD counterpart,while striking an attractive trade-off among the design challenges. Those challenges specifically pertainto FD communications, potentially facilitating simultaneous transmission and reception within the same

33

Figure 3.8 Structure of the dual-polarized mi-crostrip antenna. (a) Top view. (b) Multilayerantenna stack-up [44].

Figure 3.9 Single-antenna FD solution withTX-to-RX isolation and electrical balance du-plexer operation principle [44].

frequency band. One of the most challenging factors is that the family of SI suppression/cancellation so-lutions is typically based on complex and/or costly hardware designs. Hence it is of crucial importanceto closely examine cost-efficient algorithms associated with tolerable hardware/software complexity.More importantly, the most dominant hardware imperfections, such as the phase noise, non-flat fre-quency response of the circuits, power amplifier nonlinearity and transmit I/Q imbalance, etc. may allimpose limitations on the attainable SI cancellation capability and must be carefully mitigated.

Given that FD communication has become a feasible design option, the research community is turn-ing its attention to more cost-effective system design principles, where the attainable spectral efficiencymay be significantly improved by creating high-reliability, reduced complexity, reduced-cost, power-efficient FD devices.

34

Chapter 4

Proposed Multiuser Full-Duplex Communication

4.1 Overview

One of the potential solution to the problem of spectrum crunch in future cellular network is full-duplex (FD) communication. The FD communication makes the simultaneous in-band transceivingfeasible, i.e, simultaneous uplink and downlink operation using the same spectrum resources. Thechapter presents multiuser FD communication which allows multiple UEs to share the same spectrumresources, simultaneously. The use of the same subcarriers for both uplink and downlink results in co-channel interference (CCI) at the downlink of a UE from uplink signals of other co-existing UEs. Wehave proposed two solutions to mitigate the effect of the CCI from the downlink:

• Case 1: We considered deploying the smart antenna technique at the multiple antenna UEs withhigh spatially correlated multiple antennas.

• Case 2: A highly scattered environment is considered to take advantage of the diversity gain atthe UEs to mitigate the effect of CCI, in a multiple antenna system.

In a practical scenario, the operating conditions are dynamical, resulting in possible failure of theproposed method to tackle CCI. The dynamical changes include events like alignment of UEs in thesame direction w.r.t the eNB or increase of the uplink power of the UEs. Hence, we have proposed thedynamic resource block allocation (DRBA) algorithm to mitigate the effects of CCI. Further, we analyzethe use of software defined radio to combat pathloss and improve the link quality of a communicationsystem.

Note: As discussed in the previous chapter, extensive progress is made in the field of SIC deploymentin transceivers, especially in small cells, hence, we here have assumed perfect SIC at the eNB and theUE transceiver circuits.

Notation: [.]T ,(.)H denote transpose and Hermitian respectively. ||.|| denotes the Euclidean norm.(.)d and (.)u denote downlink and uplink components, respectively.

35

Figure 4.1 System model for multiuser full-duplex communication

4.2 System Model

In this work, for facilitating FD communication, both the eNB and UEs operate in the FD mode. AnFD eNB with Ne antennas and K FD UEs with Nr antennas each is considered such that Ne ≥ KNr,shown in fig.4.1. For all the proposed transceiver architectures shown in fig.4.2 and fig.4.3 for eNB andUEs for the two scenarios respectively, the Analog and Digital SIC unit at the RF front end, includesthe SIC circuitry [16,44] enabling the FD communication. The details of SIC design are analyzed in theprevious chapter.

Let each subcarrier allocated be shared by K UEs simultaneously, where K is given by [65]:

K = min(⌊Ne

Nr

⌋,K)

(4.1)

.

Each UE is allocated M (=⌊NKK

⌋) subcarriers, where N is the total number of subcarriers available.

Keeping this in mind, a case of K = K, i.e., all the K UEs are allocated all the N subcarriers, isconsidered. However, the appropriate number of co-existing UEs depends on the CCI experiencedby the UEs in their downlink and hence can be ≤ K. The channel between each eNB antenna andUEs antenna is assumed to be frequency selective with L taps. The FD operation allows the channelreciprocity between downlink and uplink:

huj,i,k(b) = hdj,i,k(b) (4.2)

where huj,i,k(b) and hdj,i,k(b) denotes bth time domain uplink and downlink channel coefficient betweenjth antenna at the eNB and kth antenna of the ith UE, respectively, b = 0, 1, 2, ..., L−1, j = 1, 2, ..., Ne,k = 1, 2, ..., Nr and i = 1, 2, ...,K.

36

Figure 4.2 Transceiver structure for the proposed eNB architecture

Figure 4.3 Transceiver structure for the proposed UE architecture

37

In a conventional LTE system, UEs are allocated subcarriers according to channel state schedulingalgorithm [65]. For uplink and downlink, single carrier frequency division multiple access (SC-FDMA)and orthogonal frequency division multiple access (OFDMA) is used for multiple access respectively.For FD operation, same subcarriers can be allocated to UEs for uplink and downlink. Hence, we proposeusing SC-FDMA for both uplink and downlink, due to its advantage, over OFDMA, in terms of bit errorrate (BER) performance and energy efficiency, particularly at the UE [65].

In the downlink, the channel reciprocity property of FD enables the transmitter (eNB) to acquire CSIwith ease. The CSI can be used to perform efficient subcarrier allocation and precoding the UE data ateNB so as to perform SVD based beamforming. In the uplink, successive interference cancellation withoptimal ordering (SSIC-OO) algorithm is used at the eNB to segregate signals of UEs sharing the samesubcarriers. Multiple antennas at the eNB and the UE are exploited to avoid interference to the UE atthe downlink from uplink signals of other UEs, sharing same subcarriers 1.

4.3 Case 1: FD Communication: Smart Antennas Technique

First, we consider the scenario where the Nr antenna elements in the UE, unlike in case of eNB,are taken closely spaced enough to allow a high spatial correlation between them. Especially, for asmall cell deployment deployed around the lower end of the super high frequency (SHF) band (2 GHzto 7 GHz [66]), angular spread can be around 20o for outdoor and 22o − 26o for indoors [66]. Thus,keeping antenna spacing below half wavelength, results in high correlation between antenna elements[67]. Hence, UE acts as a single antenna system, i.e. Nr ≈ 1, when evaluating K. This eliminates thepossibility of diversity gain at the UE. Also, the number of data streams per UE is given by Q = 1.Thehighly correlated multiple antenna UE are used to form a directed beam 2 towards the eNB and nulls inthe direction of other UEs operating in the same subcarriers. The directed beam also helps in combatinghigh pathloss [19]. This is implemented through the beamforming unit shown in fig.4.4. To keep theanalysis simple, the frequency domain MMSE equalizer is used both in downlink and uplink.

4.3.1 Downlink Operation

In fig.4.2, let xid denote the ith UE information data block of length M in downlink (denoted by d):

xdi = [xdi (1), xdi (2), ..., xdi (M)]T , i = 1, 2, ...,K (4.3)

The output of the M -point block is given by:

xdi = FMxdi (4.4)

1In a multicell scenario, intercell interference (eNB to eNB and eNB to UE) can be mitigated by methods like interferencemanagement through cloud access network (C-RAN) architecture. Discussion on inter cell interference is out of scope of thiswork.

2Operation in the GHz range considered, ease the deployment of multiple antennas at UE. Also extensive research is underway in the area of mmWaves [19] which further facilitates deployment of multiple antennas at UE.

38

Figure 4.4 Structure for the Beamforming unit for the proposed UE architecture and the formation ofdirected beam toward eNB

where FM is the M -point DFT matrix and xdi = [xdi (1), xdi (2), ..., xdi (M)]T . This output is thenpassed through the subcarrier allocation block. Let Pi

m denotes NeX1 precoding vector for ith UE onmth subcarrier. The precoded output vector of size NeX1 for ith UE on mth subcarrier is given by:

zdi (m) = Pimx

di (m) (4.5)

where m = 1, 2, ...,M , zdi (m) = [zdi,1(m), zdi,2(m), ..., zdi,Ne(m)]T . Let Ai, represents the NXM

subcarrier allocation matrix for ith UE, i = 1, 2, ...,K [65]. The NX1 vector input to the N -pointIDFT block for jth transmit chain is given by:

edj =K∑i=1

Aizdi,j (4.6)

where j = 1, 2, ..., Ne, zdi,j = [zdi,j(1), zdi,j(2), ..., zdi,j(M)]T . The output of the N -point IDFT blockfor the jth transmit chain is given by:

sdj = FNedj (4.7)

where j = 1, 2, ..., Ne, FN denotes the N -point IDFT matrix. This signal is then transmitted on jth

antenna after addition of the cyclic prefix (CP).

In this work, the UE transceiver unit (fig.4.3) consists of a uniformly spaced linear antenna array ofNr elements with an inter element distance of h. The angle with respect to the array normal at which theplane wave impinges upon the array is represented as ψ. Let A = [α(ψ1),α(ψ2), ...,α(ψj), ...,α(ψK)]

be the NrXK steering matrix such that α(ψj) is the NrX1 vector which represents the array steering

39

vector corresponding to the direction of arrival (ψj) of either the eNB or one of the other K − 1 UEssharing the same subcarriers [68]:

α(ψj) = [α0j , α

1j , ..., α

xj , ..., α

Nr−1j ]T (4.8)

where αxj = exp(−j2πxhλ sin(ψj)) and λ is the wavelength. Algorithms like Root-MUSIC (due lesscomputational complexity) is used for estimating DoA of eNB and other UEs sharing the same spectrumresource (subcarriers). By knowing the direction of the eNB, antenna array forms directed beams in itsdirection with a constant gain and minimize sensitivity towards other UEs.

At the lth UE, the downlink received signal is obtained after SIC cancellation in receive chain. Thereceived signal at kth antenna, is given by:

ydl,k =[ Ne∑j=1

hdj,l,k ⊗ sdj

]αk−1e,l +

[ K∑q=1q 6=1

Nr∑k=1

huq,k,l,k

⊗ suq,k

]αk−1q,l + nl,k (4.9)

where l = 1, 2, ...,K, k = 1, 2, ..., Nr. ⊗ denotes circular convolution operation. ydl,k is NcpX1 vectorwhere Ncp is the receive symbol size with CP. hdj,l,k is the channel coefficient between jth antenna ofeNB and kth antenna of lth UE in downlink. αk−1

e,l is the spatial response of the kth antenna of lth UEin the DoA of ψe,l. ψe,l is the DoA of eNB w.r.t lth UE. hu

q,k,l,kis the channel coefficient between kth

antenna of qth UE and kth antenna of lth UE. suq,k

is the uplink signal from the kth antenna of qth UE.

αk−1q,l is the spatial response of the kth antenna of lth UE in the DoA of ψq,l. ψq,l is the DoA of qth UE

w.r.t lth UE. nl,k ∈ N(0, σ2nINcp) is the channel noise at kth antenna of the lth UE.

In one of our work [69], we neglected the CCI assuming a non-line of sight (NLOS) scenario betweenUEs. For this, we considered an environment where a plethora of man-made and natural obstructionsare present, like buildings and trees, between the UEs. This leads to screening of signals between theUEs taking into account the effects like reflection, diffraction, absorption and shadowing. However,in this work, we consider a scenario for small cells with possible LOS between the UEs sharing thesame subcarriers. Hence, a smart antenna based approach is deployed which uses multiple antennas atUEs to form directed beams towards eNB and nulls toward other UEs coexisting in the same spectrumresources [68]. For the purpose of forming directed beam towards eNB, we formulated a optimizationproblem, given in appendix A, which looks to minimize the received signal energy by keeping the gainin the direction of eNB constant. In literature, such a problem is termed as the Constrained Least MeanSquare algorithm (CLMS) and is preferred for its low computational simplicity. The CLMS algorithmfor determining the optimal weight vector for the look direction is derived in appendix A as:

Wl(t+ 1) = P[Wl(t)− µydl (y

dl )H]

+ F (4.10)

WHl α(ψx) ≈

1, x = j

0, x 6= j, ∀t (4.11)

40

where Wl(t) = [wl,1(t), wl,2(t), ..., wl,Nr(t)]T is the complex weight vector for the lth UE, in the tth it-eration, α(ψj) is the array response on the desired look direction at ψj . ydl = [yd1,l,y

dl,2, ...,y

dl,Nr

]T

is NrXNcp matrix and ydl is the weighted sum of the signals from the Nr antennas. The µ is apositive scalar, called gradient step size that controls the convergence rate of the algorithm. P =[I − α(ψx)((α(ψx))Hα(ψx))−1(α(ψx))H ] and F = α(ψx)((α(ψx))Hα(ψx))−1. Henceforth, for the

sake of analysis, the notion for the number of iteration (t) will be ignored. All the notations will be takento be at the tth iteration, which is large enough for the CLMS algorithm to converge. The weighted sumof the outputs of the Nr antennas is then calculated as follows:

ydl =

Nr∑k=1

w∗l,kydl,k =

Nr∑k=1

[( Ne∑j=1

hdj,l,k ⊗ sdj

)w∗l,kα

k−1e,l

+( K∑q=1q 6=1

Nr∑k=1

huq,k,l,k

⊗ suq,k

)w∗l,kα

k−1q,l + w∗l,knl,k

](4.12)

Due to the high correlation between antennas at the UEs, these can be approximated as a singleantenna system and hence (4.12) is equivalent to the following:

ydl =

Ne∑j=1

(hdj,l ⊗ sdj

)[ Nr∑k=1

w∗l,kαk−1e,l

]+

K∑q=1q 6=1

Nr∑k=1

(huq,l ⊗ su

q,k

)[ Nr∑k=1

w∗l,kαk−1q,l

]+ nl (4.13)

where, for the lthUE, hdj,l,1 ≈ hdj,l,2 ≈ ... ≈ hdj,l,Nr≈ hdj,l, and nl =

∑Nrk=1w

∗l,knl,k. For the

qth UE, huq,1,l,k ≈ huq,2,l,k ≈ ... ≈ huq,Nr,l,k≈ huq,l,k, hdq,l,1 ≈ hdq,l,2 ≈ ... ≈ hdq,l,Nr

≈ hdq,l and

suq,k

= αk−1e,q wk,qs

uq .

Also, from (4.11), received signal in (4.13) after removing of CP can now be given by:

ydl =

Ne∑j=1

hdj,l ⊗ sdj + nl (4.14)

where l = 1, 2, ...,K, ydl is of size NX1, hdj,l = [hdj,l(0), hdj,l(1), ..., hdj,l(L− 1), (N − L)zeros]T andnl ∈ N(0, ||Wl||22N0IN ) is the additive noise vector. This signal is then converted to the frequencydomain, which is given by:

ydl = FNydl (4.15)

where FN is the N -point DFT matrix.

ydl =

Ne∑j=1

Hdj,le

dj + nl

=

Ne∑j=1

Hdj,l

K∑i=1

Aizdi,j + nl

(4.16)

41

where Hdj,l = diag(FNhdj,l) is the NXN diagonal matrix whose diagonal elements are frequency

domain coefficients between jth transmit antenna at eNB and lth UE. Let Ai be theMXN deallocationmatrix where Ai = (Ai)T . The M -point received signal for lth UE after sub-carrier deallocation isgiven by:

ydl = Aiydl

=

Ne∑j=1

K∑i=1

AiHdj,lA

izdi,j + nl

(4.17)

Now, the received signal for the lth UE on the mth subcarrier is given by:

ydl (m) = Hdl (m)

K∑i=1

Pimx

di (m) + nl(m) (4.18)

where Hdl (m) is the 1XNe frequency domain channel coefficient vector of lth UE on themth subcarrier.

(1, j)th entry is the mth diagonal element of matrix Hdj,l. nl(m) is channel noise for the lth UE on the

mth subcarrier. The SVD decomposition of channel matrix is given by [65]:

Hdl (m) = Udm,lE

dm,l(V

dm,l)

H (4.19)

where for a single antenna UE, Udm,l is a scalar such that (Udm,l)

2 = 1, Edm,l is a scalar equal to(λdm,l)

1/2 where λdm,l is the eigenvalue of Hdl (m)(Hd

l (m))H and Vdm,l is a NeX1 matrix containing the

eigenvector corresponding to non-zero eigenvalue of (Hdl (m))HHd

l (m), which is equal to λdm,l. Thereceived signal vector on mth subcarrier due to all UEs sharing the subcarriers is hence can be given by:

yd(m) = UdmEd

m(Vdm)HPmxd(m) + n(m) (4.20)

where yd(m) = [ydl (m), yd2(m), ..., ydK(m)]T , Udm = diag(Udm,1, U

dm,2, ..., U

dm,K), Vd

m = [Vdm,1,

Vdm,2, ...,V

dm,K ], Ed

m = diag(Edm,1, Edm,2, ..., E

dm,K), Pm = [P1

m,P2m, ...,P

Km] , xd(m) = [xd1(m),

xd2(m), ..., xdK(m)]T and n(m) = [n1(m), n2(m), ..., nk(m)]T . The interference from the downlink ofother UEs on the lth UE can be completely eliminated by choosing the precoding matrix as:

Pm = [(Vdm)H ]+ (4.21)

where [(Vdm)H ]+ is the pseudo inverse of (Vd

m)H . The equation (4.20) can be represented as:

yd(m) = UdmEd

mβmxd(m) + n(m) (4.22)

The received signal on mth subcarrier for the lth UE is given by:

ydl (m) = Udm,lEdm,lx

dl (m) + nl(m) (4.23)

42

In the post-processing unit, for the lth UE, the received signal is multiplied with (Udm,l)H :

ydl (m) = (Udm,l)H ydl (m)

= Edm,lxdl (m) + wl(m)

(4.24)

where for a single antenna UE, (Udm,l)H = Udm,l.

Using the above definition, the received signal vector for lth UE on the M allocated subcarriers canbe expressed by:

ydl = Edl x

dl + wl (4.25)

where ydl = [ydl (1), ydl (2), ..., ydl (M)]T , Edl = diag(Ed1,l, E

d2,l, ..., E

dM,l) , xdl = [xdl , x

dl (2), ..., xdl (M)]T

and wl = [wl(1), wl(2), ..., wl(M)]T .This is then subjected to frequency domain MMSE equalization. The received signal vector at the

output of the MMSE equalizer on the M allocated subcarriers is:

ˆxdl = ((Ed

l )H(Ed

l ) +N0Im)−1(Edl )H ydl (4.26)

This signal for the lth UE is then converted to the time domain by an M -point IDFT operation givenby:

xdl = FM ˆxdl (4.27)

where l = 1, 2, ...,K and FM is M -point inverse IDFT matrix. This is used for decoding of the signalfor the lth UE.

4.3.2 Uplink Operation

In fig.4.3, let xui denotes the information data block of length M for ith UE in uplink (denoted by u):

xui = [xui (1), xui (2), ..., xui (M)]T , i = 1, 2, ...,K (4.28)

The output of the M -point DFT block is given by:

xui = FMxui (4.29)

where xui = [xui , xui (2), ..., xui (M)]T . As discussed, Ai represents the NXM subcarrier allocation

matrix for ith UE, i = 1, 2, ,K. Due to channel reciprocity, the subcarrier allocation matrix for a UEin both uplink and downlink is equal. The NX1 vector input to the N -point IDFT block for ith UE isgiven by:

dui = Aixul (4.30)

43

The output of N -point IDFT block for the ith UE is given by:

sui = FNdui (4.31)

This signal is then multiplexed into Nr copies after addition of the cyclic prefix (CP). The signal,after multiplying with the complex weights,wi,k, is transmitted in the direction of eNB w.r.t ith UE,.i.e.,ψe,i, from the Nr antennas with spatial response αk−1

e,i for the kth antenna. This steers the beam inthe direction of eNB according to CLMS algorithm described in the previous section. The transmittedsignal from the kth antenna of ith UE is represented as follows:

sui,k = αk−1e,i wi,ks

ui , k = 1, 2, ..., Nr (4.32)

At the eNB (fig.4.2), the received signal is obtained after SIC cancellation in the Ne receive chains.The received signal vector of size NX1 at the jth receive antenna due to the K UEs, after removing theCP, is given by:

yuj =K∑i=1

Nr∑k=1

huj,i,k ⊗ sui,k + nj

=K∑i=1

Nr∑k=1

huj,i,k ⊗ [αk−1e,i wi,ks

ui ] + nj

(4.33)

where j = 1, 2, ..., Ne, huj,i,k is the channel coefficient between jth antenna of eNB and kth antenna ofith UE in uplink. nj is the channel noise introduced at the antenna of eNB. As we have assumed highlycorrelated antennas at the UEs, we take huj,i,1 ≈ huj,i,2 ≈ ... ≈ huj,i,Nr

≈ huj,i. Hence, similar to theanalysis in (4.12-4.13) for the case of downlink and using (4.11), the received signal in (4.33) can nowbe represented as:

yuj =K∑i=1

huj,i ⊗ sui + nj (4.34)

where j = 1, 2, ..., Ne, huj,i = [huj,i(0), huj,i(1), ..., huj,i(L−1), (N −L)zeros]T and nj ∈ cN(0, N0IN )

is the additive noise vector, which due to channel reciprocity, is equal for each pair of antenna in theeNB and the UE in both uplink and downlink . The output of the jth antenna received signal is thenconverted to the frequency domain by taking the N -point DFT, which is given by:

yuj = FNyuj

=K∑i=1

Huj,id

ui + nj

(4.35)

where j = 1, 2, ..., Ne and Huj,i = diag(FNh

uj,i) is theNXN diagonal matrix whose diagonal elements

are the frequency domain channel coefficients between antenna of ith UE and jth receive antenna at

44

eNB. As discussed, Ai is the MXN deallocation matrix where Ai = (Ai)T . The M -point receivedsignal on the jth antenna after sub-carrier deallocation is given by:

yuj = Aiyuj (4.36)

For the mth subcarrier, .i.e., the mth element of yuj , the received signal on the jth antenna is givenby:

yuj (m) =K∑i=1

Huj,i(m)xui (m) + nj(m) (4.37)

The signal received by the eNB on all the Ne antennas for the mth subcarrier is given by:

yu =

K∑i=1

Hui (m)xui (m) + n(m) (4.38)

where i = 1, 2, ...,K, yu(m) = [yul (m), yu2 (m), ..., yuNe(m)]T , Hu

i (m) = [Hu1,i(m), Hu

2,i(m), ...,HuNe,i

(M)]T

and n(m) = [n1, n2, ..., nNe(m)]T . For decoding of lth UE signal, the signal received given by (4.38)can be represented as:

yu = Hul (m)xul (m) +

K∑i=1i 6=l

Hui (m)xui (m) + n(m) (4.39)

where the first term represents the desired UE signal, the second term represents the interference fromthe uplink signals of other UEs and the last term n(m) is the noise term.

The received signal is subsequently passed through the frequency domain MMSE equalizer. Theestimated signal for the lth UE on the mth subcarrier is given by:

ˆxul (m) = [σ−1xul (m) + (Hu

l (m))HRuqq−1(m)Hu

l (m)]−1(Hul (m))HRu

qq−1(m)yu(m) (4.40)

where qu(m) =∑K

i=1i 6=l

Hul (m)xui (m) + n(m) is the interference and noise factor for the lth UE.

Ruqq(m) =

∑Ki=1i 6=l

Hul (m)(Hu

l (m))Hσ2xui (m) + N0INe is the covariance of qu(m). σ2

xui (m) is the sig-

nal power for the lth UE and is normalized such that σ2xui (m) = 1, hence the estimated signal term for

the lth UE on the mth subcarrier, ignoring the scaling term, is given by:

ˆxul (m) = (Hul (m))HRu

qq−1(m)yu(m) (4.41)

The estimated signal of the UEs can be estimated by the successive interference cancellation withoptimal ordering (SSIC-OO) procedure defined in the algorithm(1) below.

45

Algorithm 1 SSIC-OO for estimating UE signal1: Let c = K

2: while c > 1 do

3: Calculate the received powers (PW ) for all the K UEs

4: PWi = ||Hui (m)||2, i = 1, 2, ..., c where σ2

xul (m) = 1

5: Let l = argmaxi

(PWi)

6: Estimate ˆxul (m)

7: yu(m) = yu(m)−Hul (m)ˆxul (m)

8: c = c− 1

9: end while

10: Now, let l represent index for the last remaining unestimated UE

11: ˆxul (m) = (Hul (m))H yul (m), the system, ignoring the scaling term Ru

qq(m)−1 = [N0INe ]−1, rep-

resents a single UE and multiple receive antennas at eNB with maximal ratio combining (MRC) of

UE symbols

Let the signal for the lth UE on all the subcarriers is given by:

ˆxul = [ˆxul (1), ˆxul (2), ..., ˆxul (M)]T (4.42)

This signal for the lth UE is then converted to time domain by an M -point IDFT operation given by:

xul = FM ˆxul , l = 1, 2, ...,K (4.43)

This is used for decoding of the signal for the lth UE.

4.3.3 Simulation Results

The advantage of using SC-FDMA for downlink instead of OFDMA in terms of BER performanceis analyzed in [65]. To validate the inclusion of SIC design for FD system in our architecture, we havecarried out MATLAB simulations for BER performance in downlink and uplink. For simulation, wehave considered an FD eNB with four antennas (Ne = 4) and two FD UEs (K = 2) with four antennaseach (Nr = 4) and PW2 > PW1. The DoA of eNB w.r.t two stationary UEs, i.e, UE1 and UE2 is 10o

and 60o respectively 3. The issues due to mobility of UEs are considered later in the chapter. For smartantenna beamforming at UEs, the CLMS algorithm with root music algorithm is used. The GhorbaniModel and thermal noise (Noise temperature = 290K) is used for modeling the non-linearity which isintroduced to the complex baseband SC-OFDM symbols. The channel between each antenna of eNB

3Such that the UEs are at the null of each others beampattern.

46

and UEs’ is taken as frequency selective with L = 10 and uniform power delay profile (UPDP). Themodulation scheme used is 16-QAM (no coding). The bandwidth allocated to the UEs is taken to be 3MHz which is split into 256 subcarriers, out of which 180 subcarriers are occupied by the UEs. A cyclicprefix of duration 4.69µs is used.

For the downlink, the performance at UE1 is considered for received complex SC-FDMA symbolsfrom eNB. In fig.4.5, the effect of SIC on the receiver performance at UE1 is shown. It is observed thatwithout SIC the receiver has zero throughput. A similar result is obtained when SIC is attempted withouttaking into consideration the non-linearity (NL) components. No improvement in receiver performanceis observed by including the NL distortion components in SIC without the cross-talk cancellation (CTC).The receiver performance is equivalent to half-duplex (HD) performance when we considered both selfand cross talk along with NL distortion components for SIC.

For the uplink, the performance at multiple antenna eNB is considered for received complex SC-FDMA symbols from UE1. In fig.4.6, the effect of SIC on the receiver performance of the eNB isshown. The SIC analysis is similar to downlink, but there is an additional diversity gain introduceddue to multiple antennas in eNB according to the algorithm(2). Comparing the BER performance at UE(fig.4.5) and eNB (fig.4.6), it is observed that due to the additional diversity gain, there has been a nearly18 dB gain for the eNB over the BER performance of UE at BER 10−2. In case of uplink, the effect ofCCI has negligible impact as eNB employs SSIC-OO to cancel out interference of signal of one UE onthe signal of other UEs. Similar analysis can be done by taking UE2 into consideration.

In terms of the overall spectral efficiency (SE) per cell , the proposed FD eNB and FD UE transceiveralong with the smart antenna technique, helps in achieving higher performance4 as compared to HDtime division duplexing (TDD) system and scheduling algorithm proposed in [46]. The SE per cell indownlink for two UEs, assuming total SIC, is shown in fig.4.7. For the HD TDD system, only one UE isscheduled for uplink and downlink in alternate time slots. For scheduling algorithm [46], two HD UEs(with FD eNB) is considered such that, for a time slot, these are scheduled simultaneously in reversedirections (one in uplink and other in downlik) with their direction interchanging in every consecutivetime slot.

4The smart antenna technique helps more UEs to be scheduled simultaneously using the same spectrum resource in down-link and uplink.

47

0 5 10 15 20 25 30 35 4010

−2

10−1

100

Es/No, dB

BE

R

28 30 32

10−0.12

10−0.01

Half−Duplex system

With self and cross interference

Cancellation without NL consideration

Cancellation with NL consideration but no CTC

Complete SIC cancellation

Figure 4.5 BER performance at UE1 for FD

downlink

0 5 10 15 20 25 30 35 4010

−2

10−1

100

Es/No, dB

BE

R

Half−Duplex system

With self and cross interfernce

Cancellation without NL consideration

Cancellation with NL consideration but no CTC

Complete SIC cancellation

Figure 4.6 BER performance at eNB for FD

uplink

0 5 10 15 20 25 30 35 400

2

4

6

8

10

12

Es/No, dB

Ca

pa

city (

bits/s

/Hz)

Ergodic channel capacity

Smart antenna approach

Scheduling Algorithm [46]

Half−Duplex system

Figure 4.7 Overall Spectrum efficiency per cell in downlink for various scheduling approaches

4.4 Case 2: FD Communication: Diversity Gain Technique

Next, we consider the scenario where the environment is richly scattering, which allows to take theadvantage of diversity gain unlike in case of the smart antenna approach discussed above. Hence, in this

48

Figure 4.8 Proposed UE transceiver architecture for mitigating CCI using diversity gain

case, to mitigate CCI in the downlink, we use the diversity gain to maximize the signal-to-interference-plus-noise ratio (SINR) at the receiver,.i.e, UE. The deployment of a sufficient number of transmit andreceive antennas at the eNB and UEs, respectively, results in significant improvement in performancein the presence of CCI. Also, the proposed architecture for the UE has less computational complexitythan the architecture deploying smart antenna technology discussed in the previous section. To maintainbrevity, we focus only on downlink operation. The uplink operation is similar to the analysis carried outfor uplink in the previous section with minor modifications.

4.4.1 Full-Duplex Multiuser Operation

In the downlink, let xdi denote the ith UE information data block of length M , shown in fig.4.2. Theoutput of the M -point DFT block is given by:

xdi = FMxdi (4.44)

where i = 1, 2, ...,K, FM is the M -point DFT matrix and xdi = [xdi (1), xdi (2), ..., xdi (m), ..., xdi (M)]T .xdi (m) is the mth data symbol of ith user and unlike in [65] same M symbols are transmitted in allthe Q data streams. This output is then passed through the pre-processing / subcarrier allocation block.Similar to previous section, Pi

m denotes NeXQ precoding vector for the ith UE on the mth subcarrier.The precoded output vector of size NeX1 for the ith UE on the mth subcarrier is given by:

49

zdi (m) = Pimxdi (m) (4.45)

where m = 1, 2, ...,M and zdi (m) = [zdi,1(m), zdi,2(m), ..., zdi,Ne(m)]T . xdi (m) = (xdi (m)J1,Q)T where

xdi (m) is the mth element of xdi and J1,Q is the 1XQ unit matrix. As described in the previous section,Ai, represents the NXM subcarrier allocation matrix for the ith UE [65]. The transmit signal vector,at the jth antenna of the eNB, after subcarrier allocation and IDFT operation is given by:

sdj = FNedj (4.46)

where j = 1, 2, ..., Ne, FN denotes the N -point IDFT matrix, edj =∑K

i=1 Aizdi,j and zdi,j =

[zdi,j(1), zdi,j(2), ..., zdi,j(M)]T . This signal is then transmitted after the addition of the cyclic prefix(CP).

In this work, the UE transceiver unit (fig.4.8) consists of a uniformly spaced linear antenna array ofNr elements with an inter element distance of h. The angle with respect to the array normal at whichthe plane wave impinges upon the array is represented as ψ. Each kth antenna introduces some phasedelay αk−1

x = exp(−j2π(k−1)hλ sin(ψx)) to the received signal, where λ is the wavelength. The indexx = e represents direction of arrival (DoA) of the eNB (ψe) w.r.t lth UE or x = q represents DoA of theqth UE (q 6= l) w.r.t lth UE whose uplink signal results in CCI at downlink of the lth UE. Algorithmslike Root-MUSIC (due less computational complexity) can be used for estimating the DoAs. However,the proposed method is independent of the availability of information about DoA of interfering UEs.

At the lth UE, the downlink received signal at kth antenna is obtained after multiplying the factor4k

followed by SIC cancellation in receive chain, where 4k = (αk−1e )∗. This is required for co-phasing

of the downlink signal from the eNB by removing the phase (αk−1e ) introduced by the kth antenna [70].

Hence, the received signal at kth antenna, is given by:

ydl,k =

Ne∑j=1

hdj,l,k ⊗ sdj + nl,k +[ K∑q=lq 6=l

Nr∑k=1

huq,k,l,k

⊗ suq,k

]αk−1q 4k (4.47)

where l = 1, 2, ...,K, k = 1, 2, ..., Nr. ⊗ denotes the circular convolution operation. ydl,k is NcpX1

vector whereNcp is the receive symbol size with CP. hdj,l,k = [hdj,l,k(0), hdj,l,k(1), ..., hdj,l,k(L−1), (Ncp−L)zeros]T is the complex i.i.d Rayleigh fading channel coefficient between jth antenna of the eNB andkth antenna of the lth UE in the downlink. nl,k ∈ N(0, σ2

nINcp) is the channel noise at kth antennaof the lth UE. hu

q,k,l,k= [hu

q,k,l,k(0), hu

q,k,l,k(1), ..., hu

q,k,l,k(L − 1)), (Ncp − L)zeros]T is the complex

i.i.d Rayleigh fading channel coefficient between kth antenna of the qth UE and kth antenna of the lth

UE. suq,k

is the uplink signal from kth antenna of the qth UE. αk−1q is the spatial response (or the phase

delay) of kth antenna of the lth UE in the DoA of ψq. This signal can be represented by:

ydl,k =

Ne∑j=1

hdj,l,k ⊗ sdj + Il,k (4.48)

50

where Il,k = nl,k +[∑K

q=lq 6=l

∑Nr

k=1huq,k,l,k

⊗ suq,k

]αk−1q 4k is the noise plus interference suffered by the

lth UE at the kth antenna. After the removal of CP, the signal is converted to the frequency domain:

ydl,k = FNydl,k =

Ne∑j=1

Hdj,l,ke

dj + Il,k (4.49)

where FN is the N -point DFT matrix. Hdj,l,k = diag(FNhdj,l,k) is the NXN diagonal matrix whose

diagonal elements are frequency domain coefficients between jth transmit antenna of the eNB and kth

antenna of the lth UE.

The signal is then subjected to the subcarrier deallocation and after simplifications [65], we get:

ydl (m) = Hdl (m)

K∑i=1

Pimxdi (m) + Il(m) (4.50)

where yl(m) = [yl,1(m), yl,2(m), ..., yl,Nr(m)]T , yl,k(m) is the mth element of yl,k = Aiydl,k. Ai isthe MXN deallocation matrix. Hd

l (m) is the NrXNe frequency domain channel coefficient matrix ofthe lth UE on the mth subcarrier. (k, j)th entry in the Hd

l (m) is the mth diagonal element of matrixHdj,l,k. Il(m) is channel noise and interference for the lth UE on the mth subcarrier.

The SVD decomposition of channel matrix is given by Hdl (m) = Ud

m,lEdm,l(V

dm,l)

H . Udm,l is a

NrXQ unitary matrix containing the eigenvectors corresponding to non-zero eigenvalues of Hdl (m)(Hd

l (m))H ,Edm,l is a QXQ diagonal matrix containing the non-zero eigenvalues (λdm,l) of Hd

l (m)(Hdl (m))H such

that Edm,l = diag((λdm,l,1)1/2, (λdm,l,2)1/2, ..., (λdm,l,Q)1/2) and Vd

m,l is a NeXQ unitary matrix contain-ing the eigenvectors corresponding to non-zero eigenvalues of (Hd

l (m))HHdl (m). The received signal

vector on the mth subcarrier due to all UEs sharing the subcarriers is hence can be given by:

yd(m) = UdmEd

m(Vdm)HPmxd(m) + I(m) (4.51)

where yd(m) = [(ydl (m))T , (yd2(m))T , ..., (ydK(m))T ]T , Udm = diag(Ud

m,1,Udm,2, ...,U

dm,K), Vd

m =

[Vdm,1,V

dm,2, ...,V

dm,K ], Ed

m = diag(Edm,1,E

dm,2, ...,E

dm,K), Pm = [P1

m,P2m, ...,P

Km], xd(m) =

[(xd1(m))T , (xd2(m))T , ..., (xdK(m))T ]T and I(m) = [(I1(m))T , (I2(m))T , ..., (IK(m))T ]T .

The interference from the downlink of other UEs on the lth UE is eliminated by choosing the pre-coding matrix as Pm = [(Vd

m)H ]+. At the lth UE, the received signal on the mth subcarrier afterpost-processing, .i.e, multiplying with (Ud

m,l)H is given by [65]:

ydl (m) = Edm,lx

dl (m) + Il(m) (4.52)

Now, defining Edm,l = [Edm,l(1, 1), Edm,l(2, 2), ..., Edm,l(j, j), ..., E

dm,l(Q,Q)]T where Edm,l(j, j) is

the jth diagonal element of Edm,l. As xdl (m) = (xdl (m)J1,Q)T , above equation can be rearranged to:

ydl (m) = Edm,lx

dl (m) + Il(m) (4.53)

51

Now, assuming the rich scattering environment and equal channel conditions between the lth UE andK − 1 interfering UEs (especially in a small cell scenario), the received signal for the lth user on themth subcarrier can be obtained by MRC 5 [70] as:

ˆxdl (m) = ((Edl (m))H(Ed

l (m))−1(Edl (m))H ydl (m) (4.54)

This signal for the lth UE is then converted to the time domain by an M -point IDFT operation givenby:

xdl = FM ˆxdl (4.55)

where l = 1, 2, ...,K and FM is M -point inverse IDFT matrix and ˆxdl = [ˆxdl (1), ˆxdl (2), ..., ˆxdl (M)]T .

This is used for decoding of signal for the lth UE. For the analysis of system performance, we considerthe effective SINR (γeffml ) 6, derived in appendix B, for the lth UE on the mth subcarrier from theequation (4.54):

γeffml = [σ2nl

+K∑q=lq 6=l

σ2qβ

qm]−1||Ed

m,l||2(βlmQ−1) (4.56)

where σ2nl

= E[|nl,k(m)|2], ∀m, l, k such that nl,k(m) is the frequency domain i.i.d noise for kth

antenna of the lth UE on the mth subcarrier.∑K

q=lq 6=l

σ2qβ

qm is the total interference power, such that

σ2q = |Hu

q,k,l,k(m)|2, ∀m, q, k, l, k where Hu

q,k,l,k(m) is the frequency domain i.i.d channel coeffi-

cient between kth antenna of the qth UE and kth antenna of the lth UE on the mth subcarrier andβqm = |xuq (m)|2, ∀m is the total uplink power allocated to the signal of the qth UE on themth subcarrier.βlm = |xdl (m)|2,∀m, l is the total power allocated to downlink signal of the lth UE on the mth subcar-rier. Now, considering βlm ≥ βqm, ∀m, q, to simplify the equation (4.56), we take βqm = βum, ∀m, q,βlm = Qβum and σ2

q = σ2u, ∀q. Hence, the equation (4.56) can be expressed as:

γeffml = [σ2nl

+ (K − 1)σ2uβ

um]−1||Ed

m,l||2βum (4.57)

It is important to observe that the effective SINR can be controlled by the gain factor, .i.e, ||Edm,l||2.

This is discussed through simulation results in the next section.

4.4.2 Simulation Results

Here, an FD eNB and three spatially uncorrelated FD UEs (say UE1, UE2 and UE3) sharing the samespectrum resources at both the downlink and uplink, are considered. We analyze capacity 7 (bits/s/Hz)

5MMSE equalizer requires knowledge of covariance of the interference at the receiver. As this information is not availableat the UE in downlink, zero forcing equalizer (which in this work is equivalent to MRC) is deployed which requires only thedownlink channel estimate.

6We have considered equal power allocation across all the Q streams at the downlink.7The capacity here is defined as number of correct bits received per second per Hertz

52

−10 −5 0 5 101

2

3

Es/No, dB

Ca

pa

city (

bits/s

/Hz)

Ne=20, Nr=6, K=3

Ne=20, Nr=6, K=1

Ne=20, Nr=4, K=3

Ne=20, Nr=4, K=1

Ne=20, Nr=1, K=3

Ne=20, Nr=1, K=1

Figure 4.9 Downlink capacity vs channel SNRof UE1 for different Nr

−10 −5 0 5 101

2

3

Es/No, dB

Ca

pa

city (

bits/s

/Hz)

Ne=40, Nr=4, K=3

Ne=40, Nr=4, K=1

Ne=20, Nr=4, K=3

Ne=20, Nr=4, K=1

Ne=40, Nr=1, K=3

Ne=40, Nr=1, K=1

Figure 4.10 Downlink capacity vs channelSNR of UE1 for different Ne

at the downlink of UE1 with two other UEs acting as the interference source. In this work, we assumeperfect cancellation of self-interference for the FD operation. The channel between each antenna of eNBand UEs’ is taken as frequency selective with L = 10. The modulation scheme used is 16-QAM (nocoding). The bandwidth allocated to the UEs is taken to be 3 MHz, which is split into 256 subcarriers,out of which 180 subcarriers are occupied by the UEs. A cyclic prefix of duration 4.69µs is used. Thetransmit power of the UEs are normalized to unity, .i.e, βum = 1, ∀m

Fig.4.9 and 4.10 show the downlink capacity of UE1 vs. the channel SNR for the increasing numberof the receive antennas (Nr) and transmit antennas (Ne) at the UEs and the eNB, respectively. In case ofthe increasingNr at the UEs, the number of transmit antennas at the eNB is kept constant atNe = 20. Itcan be seen from fig.4.9 that with the increase inNr, there is an improvement in the downlink capacity ofthe UE. This improvement is due to the increase in the diversity order which results in higher magnitudeof ||Ed

m,l||2, ∀m, l improving the γeffml ,∀m. Similarly, for the increasing (Ne) at the eNB, the numberof receive antennas at the UE is kept constant at Nr = 4. An improvement in the downlink capacity ofthe UE can be observed (fig.4.10) with the increase in Ne. This is due to increase in magnitude of eacheigenvalue, .i.e, Edm,l(j, j), j = 1, 2, ..., Q, ∀m, l resulting in the higher ||Ed

m,l||2 and hence improvingthe γeffml . Also, from fig.4.9 and fig.4.10, it can be observed that at the higher channel SNR region, thedownlink system performance is the only interference limited.

In all the above simulations, with the increase in Nr and Ne, the downlink capacity approaches theideal value where there is no CCI, .i.e, K = 1. The results are also compared with the conventionalscenario of no diversity gain at the UE, .i.e, Nr = 1. It can be observed that there is a significantimprovement in the downlink capacity with the diversity gain. Hence, increasing both Nr and Ne

53

improves the system performance at the downlink in the presence of CCI. However, increasing Nr atthe UEs, increases the computational complexity at the UEs including the power consumption. Withthe power consumption not a constraint at the eNB, comparatively, and profound research on massiveMIMO (deploying large Ne) in recent years [71], the prospect of using large antenna arrays at eNBto tackle the CCI in the case of the FD downlink seems so be encouraging. Moreover, the channelreciprocity property of the FD communication will aid the CSI actualization at the eNB.

4.5 Dynamic Resource Block Allocation (DRBA)

In conventional a LTE system, UEs are allocated spectrum resources according to channel statescheduling algorithm [65, 72]. Recently, much research has been done to develop spectrum and powerallocation algorithms for FD systems [45–48, 73]. In these works, the eNB has the capability to operatein FD mode. In [46], a scheduler is introduced which aims to maximize the logarithmic sum of theaverage rates of all the HD UEs in both uplink and downlink. The UEs either operate in HD mode or inpairs, one each in uplink and downlink directions, for FD mode depending on the operating conditions.In [47], pairs of HD UEs are formed for FD operation and spectrum resources are allocated so asto maximize the sum-rate of the system. Each subcarrier is allocated to just one transceiver. Theproblem is solved using matching theorem. In [48], the authors address joint subcarrier and powerallocation to the set of HD uplink and downlink UEs using a suboptimal iterative algorithm based onthe Frank-Wolf approach. In [73], resource allocation is considered for a system consisting of FDeNB communicating with FD UEs. The UEs are allocated exclusive subcarriers used for downlink anduplink communications to avoid the inter-user interference. The joint radio resource allocation problemfor uplink and downlink channels is addressed with the objective of sum-rate maximization. For this,an iterative algorithm is proposed based on game theory by modeling the problem as a non-cooperativegame between the uplink and downlink channels.

In this work, the K FD UEs are taken such that they share the same set of RBs (or subcarriers) inboth uplink and downlink. The use of the same RBs for both the directions results in interference atthe downlink of a UE from uplink signals of other UEs operating in the same RBs. As described inthe previous sections, to tackle this interference, either the smart antenna technique or diversity gain areused depending on the operating environment. However, in a practical scenario, the UEs are mobileand the operating conditions may also change dynamically. For example, in case where the smartantenna technique is deployed, the UEs sharing the same spectrum recourse, may align in the samedirection w.r.t the eNB, resulting in CCI. Similarly, in case of diversity gain, the UEs may increase theirtransmit power resulting in an increase in CCI. These scenarios may make the CCI high enough forthe proposed methods to handle and hence we have proposed the dynamic resource block allocation(DRBA) algorithm.

54

4.5.1 DRBA Algorithm

The DRBA algorithm is used to mitigate CCI caused by uplink signals of a group of two or moreUEs, sharing the same RBs, at each others’ downlink. This algorithm initiates when BER of anyone ofthe UEs crosses a threshold. The frame structure taken is that of LTE FDD system and the BER of all theUEs is tracked by eNB every transmit time interval (TTI) of 1ms (one sub-frame). In the algorithm, allthe UEs, except one having the highest priority, moves to different available resource block accordingto the Resource Block (RB) handoff technique [35]. RB handoff is a technique, based on cognitiveradio (CR) technology, which determines the unused RBs where operation of the interfering UEs canbe shifted. The driving force behind the technique is the willingness of the telecom regulatory bodies tolet operators share their licensed spectrum bands.

For the RB handoff technique, the eNB keeps track of the unused RBs of every operator in thatregion (cell) and for every TTI. Consider a UE which is selected according to the DRBA algorithmfor undergoing a shift in operating spectrum resources. Currently, it is allocated RBs belonging to itsserving operator. Let this operator be termed as primary operator (PO). The steps for the proposed RBhandoff technique are as follows:

• Step 1: Select RBs belonging to the PO in the current cell (Intracellular). This is termed as Intra-band / Intra-operator RB handoff if RBs are selected from the same frequency band or Interband/ Intra-Operator RB handoff if RBs are selected from different frequency band, provided the POhas unused RBs in that particular LTE TTI in the current cell.

• Step 2: If Step 1 fails, select RBs belonging to a different operator present in the current cell(Intracellular). Operator in this case is called a secondary operator (SO). This is called Intraband/ Inter-Operator RB handoff if the SO operates in the same band or Interband / Inter-Operator RBhandoff if the SO operates in different bands, provided the SO has unused RBs in that particularLTE TTI in the current cell.

• Step 3: If there is still no unused RBs available, the network looks for a conventional handoffprocedure to a neighboring cell. This is called Intercellular / Intra-Operator RB handoff if theoperator in the target cell is PO or Intercellular / Inter-Operator RB handoff if the operator intarget cell is a SO, provided the neighboring cell has unused RBs in that particular LTE TTI. Thetarget frequency band can be same as the source frequency band (Intraband) or can be different(Interband). The selection of proper target eNB becomes crucial. This is handled by the cellselection procedure in the LTE network [35]. The Interband and Intercellular handoffs require theRF components to operate in a considerably broader bandwidth.

• Step 4: Finally, if there is still no unused RBs available, the RB handoff procedure and operationof the concerned UE are terminated. The RB procedure is briefly demonstrated in the flowchartgiven in fig.4.11.

55

Among a group of UEs, the UE which undergo RB handoff is decided according to a priority pro-tocol. The UE with lowest priority undergoes the handoff procedure. The priority protocol for UEs isdecided by the network operators and depends on a number of factors, as discussed in Chapter 2. Incase, the UE is operating under a SO, the UEs primarily belonging to the SO are given higher priority.The network operators can coordinate among themselves to come up with a consolidated scheme toallocate priority level to the UEs. The eNB keeps track of the priority level of all UEs operating underit.

There can be two modes of operation for RB handoff:

1. Full Mode: In this mode, the UE vacates all the overlapping RBs.

2. Semi Mode: Here the UE vacates only a portion of overlapping RBs due to unavailability ofsubstitute for all the RBs.

Along with the above proposed procedure to share the unused resources by the operators, extensiveresearch is underway to make use of unused spectrum band, freed by digitization of TV transmission,for providing an additional resource for the future generation networks [5, 28]. This unused spectrumresource is called TV white space (TVWS) and its information is maintained in a database called Geo-location database. In [28], a detailed procedure for creating and continuously updating the Geo-locationdatabase is analyzed. Also the work includes various technical aspects of using this database for Op-portunistic Spectrum Access (OSA) in LTE-A. However, in this work we are not considering utilizationof TVWS. Coexistence of cellular network with wireless technologies like Wi-Fi and ZigBee in theunlicensed spectrum is discussed in [74].

For K UEs sharing the same RBs, let Ul represent the lth UE and Ulm be the UE with the lowestpriority among the UEs. A threshold λUl

is set for all the UEs which is the maximum level of BERtolerable to lth UE to maintain the corresponding QoS (Quality of Service). λUl

can depend on theapplication the lth UE is running and services it acquires from the network operators. Thus, the thresholdis generally set by the network operator. The DRBA algorithm to mitigate CCI is described below.

Algorithm 2 DRBA algorithm to mitigate CCI

1: Initialize l=set of all K UEs2: for every TTI do3: while ∃b ∈ l : BERUb

> λUbdo

4: Execute RB handoff for Ulm5: Remove Ulm from set l6: c = length (l)7: if c > 1 then8: lm = arg min

b∈l{priority(Ub)}

9: end if10: end while11: end for

56

Figure 4.11 Basic flowchart briefly demonstrating the RB handoff procedure

Note: In the present work, we have considered DRBA algorithm employing full mode RB handoff.

To illustrate the DRBA algorithm, we consider a simple scenario with an eNB at the center of a cell.In the geographical area, let the AWS band is the only frequency band being used. Let operator I isassigned Block A and B (AWS is divided into 6 blocks A-F) together constituting 20 MHz and operatorII is assigned Block C of 10 MHz. Now taking two UEs (UE1 and UE2) belonging to the operator I,such that the eNB is at 10o and 60o w.r.t UE1 and UE2 respectively and share all the RBs of Block A.Block B and C are assumed to be unused. Also, UE1 has higher priority than UE2. At this point, UE2starts moving towards UE1. Depending on the beamwidth, the more the UE2 moves closer to UE1, thehigher is the chance they are invading each others’ coverage area. This results in CCI in their downlinkfrom each others’ uplink signal. The eNB continuously monitors the BER of the both UEs every TTIor 1ms. When BER of anyone of the UE crosses the threshold, UE2 can move its operation either toBlock B (Intraband / Intra-operator RB handoff) or Block C (Intraband / Inter-Operator RB handoff)from the next TTI.

4.5.2 Simulation Results

In this section, we demonstrate the utility of DRBA in case of smart antenna deployment to mitigatethe CCI. Similar analysis can be carried out for the method using diversity gain to mitigate the CCI andis not included in this work to avoid redundancy. For simulation, we have considered an FD eNB withfour antennas (Ne = 4) and two FD UEs (K = 2) with four antennas each (Nr = 4), with inter-antenna

57

distance (h) of 0.45λ. The UE1 is considered to have higher priority than UE2. The DoA of eNB w.r.ttwo stationary UEs (UE1 and UE2), i.e, DoAUE1

eNB and DoAUE2eNB is 10o and 60o respectively, such that

the UEs are at the null of each others’ beampattern. For smart antenna beamforming at UEs, the CLMSalgorithm with root music algorithm is used. The Ghorbani Model and thermal noise (Noise temperature= 290K) is used for modeling the non-linearity which is introduced to the complex baseband SC-OFDMsymbols [1,4,40,75]. The channel between each antenna of eNB and UEs is taken as frequency selectivewith L = 10 and have a uniform power delay profile (UPDP). The modulation scheme used is 16-QAM(no coding). The bandwidth allocated to the UEs is taken to be 3 MHz which is split into 256 subcarriers,out of which 180 subcarriers (15 RBs) are occupied by the UEs. A cyclic prefix of duration 4.69µs isused. The FD UEs share all the available RBs.

Here, we consider a scenario where UE1 and UE2 are at the same distance from the eNB. Nextthe UE2 starts moving towards UE1 in a circular path, such that the distance between UE2 and eNBremains unchanged, as shown in the fig.4.12. The closer the UE2 moves nearer to UE1, more is thechance that they come in each others’ coverage range, depending on their beamwidth. Hence, if theyare close enough, there is the chance of CCI in their downlink as explained in the previous sections.

The solution to the problem of CCI proposed here, is the dynamic resource block algorithm (DRBA)described in Section 3. To demonstrate the DRBA algorithm, we analyze the BER performance of theUEs in the downlink as UE2 moves closer to UE1. We set the BER threshold (λUl

) of both UE1 andUE2 to 10−1 . For the analysis henceforth, the channel condition between the eNB and UEs is usedbetween mobile UEs, with the expectation that in a small cell, eNB and UE are comparable. Withoutloss of generality, the downlink channel SNR (ES/NO) value of UE1 and UE2 is kept constant at 35 dBand 40 dB respectively. The channel SNR between UEs is kept constant at 40 dB. The downlink signalpower at the UEs are normalized to analyze the BER performance. From fig.4.13, it can be observed thatas UE2 moves closer to UE1, the BER of both the UEs increases due to interference in their downlinkfrom each others’ uplink signal. When the BER of UE1 reaches the BER threshold (λUl

), DRBA isinitiated and UE2 moves completely to a different available RBs (full-mode RB handoff) according tothe RB handoff technique described in the previous section.

Practically, the CCI does not cancel out completely using the smart antenna technique. Hence, it isobserved that the final BER of each UE is less than the initial BER due to the complete absence of CCIin the final stage, as can be observed from the fig.4.13. Initially, when DoAUE2

eNB is 60o, there is someCCI at the downlink of each UE due uplink signal of other UE. The scenario, where there is a largenumber of UEs at the initial stage, results in higher CCI at the downlink of each UE. The UEs co-exist,if their BER is below λUl

. Otherwise, they are allocated different spectrum resources. The differencebetween the initial and final BER value is reduced by improving the smart antenna implementation. Oneway to do so is by increasing the number of antennas in UEs, which is discussed later in this section.

In the analysis so far, the UE with lowest priority vacated the entire RBs when required to mitigateCCI. Now, a more dynamic approach is introduced which provides more flexibility to the RB sharingof the UEs. The approach defines a semi threshold level, λsUl

, for UEs except the UE with the lowest

58

Figure 4.12 UE1 and UE2 in the coverage re-gion of eNB

101520253035404550556010

−5

10−4

10−3

10−2

10−1

100

DoA in degrees

BE

R

BER of UE1

BER of UE2

Figure 4.13 BER performance at UE1 and UE2w.r.t DoAUE2

eNB

priority (Ulm). When any one of the UEs crosses λsUl, Ulm vacates certain fraction of RBs rather than

the entire RBs (semi-mode RB handoff). This prevents the BER of UEs from reaching the conventionalthreshold λUl

. Hence, the semi-mode of handoff facilitates the co-existence of UEs as the BER is keptbelow λUl

. Though this approach increases the computational complexity, it is particularly useful wheneNB finds network too congested to allocate Ulm a complete set of available RBs in case a handoff isinitiated.

To understand the concept, consider a scenario (shown in fig.4.14) where, enough unused RBs areavailable to allow UE2 to shift 40% of its RBs to them. In the conventional approach described in above,the UE2 terminates its call when BER of UE1 exceeds λUl

around 40o due to unavailability of enoughRBs (indicated by the absence of BER performance for UE2 in fig.4.14). Now, let us look at the dynamicapproach. For this, the semi threshold level for UE1 is set to λsUl

= 5X10−2 . Around 44o DoAUE2eNB , the

BER of UE1 reaches λsUland UE2 vacates 40% of RBs and thereby keeping BER below λUl

. Comparingwith the conventional approach, the modified approach prevents the UE2 from terminating its call. Otherthan the computational complexity, there are many challenges, like deciding upon the optimal value forλsUl

and sensing possible congestion in the network, that needs to be considered for implementing thisapproach. This detailed study of the approach is out of the scope of the work and is left for future work.

Finally, we show the performance of the DRBA algorithm by increasing the number of antennas inthe UEs. This scenario is studied keeping in mind the use of mmW communication in future cellularnetworks. The use of mmW will allow deployment of larger number of antennas in the UEs and hencemaking the beam sharper. The fact is demonstrated in fig.4.15, which shows the half-power beamwidth(HPBW) of a linear antenna array w.r.t to the number of antennas (Nr). It can be seen that for a single

59

010203040506010

−5

10−4

10−3

10−2

10−1

100

DoA in degrees

BE

R

BER of UE1 with conventional approach

BER of UE2 with conventional approach

BER of UE1 with dynamic approach

BER of UE2 with dymamic approach

Figure 4.14 BER performance at UE1 and UE2 for conventional and dynamic approach w.r.t DoAUE2eNB

antenna system, the HPBW is infinity which indicates an omnidirectional beam pattern. In fig.4.16, theBER of UE1 and UE2 w.r.t are considered by deploying 10 antennas at the UEs. It is observed thatwith the increase in the number of antennas at the UEs, the handoff takes place when UEs are furthercloser to each other as compared to when the lesser number of antennas were deployed. Also, it can beobserved that the difference between the initial BER and the final BER is decreased, which is the resultof a decrease in initial CCI due to the sharper beamwidth.

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

90

100

Number of antennas

HP

BW

in d

egre

es

Figure 4.15 Half-power beamwidth (HPBW)

w.r.t number of antennas

010203040506010

−5

10−4

10−3

10−2

10−1

100

DoA in degrees

BE

R

BER of UE1

BER of UE2

Figure 4.16 BER performance at UE1 and UE2 for

Nr = 10 w.r.t DoAUE2eNB

60

4.6 Improving Link Quality: Software Defined Radio

In the discussion so far, we have studied FD communication in a scenario where operating frequencyis in sub-3 or sub-6 GHz bands. From the concepts of electromagnetic wave propagation, higher theoperating frequency, higher is the path loss. The path loss (PL) is the reduction in power density ofa radio signal as it propagates in the space [76]. The main factors contributing to path loss includepropagation losses caused by the natural expansion of the radio wave front in free space, absorptionlosses caused when the energy of the radio signal is absorbed by the obstacles in the path and diffractionlosses when part of the radio wave front is obstructed by an opaque obstacle. The PL adversely affectthe QoS of a wireless link resulting in possible link failure and hence combating the PL is a crucial taskfor any wireless system.

In LTE, link adaptation and power control (PC) are two tools used to mitigate the effect of the PL. Thelink adaptation refers to dynamical allocation of the modulation and coding (MCS) for communication.This is controlled by the eNodeB (eNB) and help tackling path loss in the downlink. In the downlink,eNB usually transmits with the maximum power [77, 78]. In the uplink, PC provides the way to handlethe path loss. The PC in uplink includes determining PL by user equipment (UE) through the receivedsymbol received power (RSRP) from the eNB. The UE then compensates for the PL by changing itstransmit power accordingly (open-loop PC). The transmit power of UE also includes an MCS dependentoffset determined by the eNB (closed-loop PC) [77].

With the next generation cellular network, 5G, looking to exploit the higher frequency (or mmW), theproblem of path loss can prove to be a bottleneck [19]. The operating environment is highly dynamic.It can encounter sudden degradation in the operating conditions or the arrival of obstruction betweentransmitter and receiver, resulting in link failure, especially at high operating frequencies. We herepresent an analysis of a specific scenario where the already established method to tackle PL proves tobe insufficient.

The method analyzed here includes using software defined radio (SDR) with multi-band reconfig-urable antennas (automated by the field programmable gate array of the SDR [79]) at the RF front end ofthe communication systems. The SDR system is capable of dynamic spectrum allocation (DSA) for thecommunication. This is used to dynamically lower the operating frequency of a communication systemoperating at high frequency, in case of degradation of the link quality. In the conventional communica-tion system, it is not possible to lower the frequency on the fly as it will require complete reconfigurationof the architecture of the communication system. SDR provides the freedom of dynamically changingthe parameters of the components at the RF front end so as to operate in various frequency bands. Fur-ther, the use of SDR for DSA reinforces the deployment of DRBA algorithm, discussed in the previoussection. This system is analyzed by creating a wireless link using universal software radio peripherals(USRP).

Note: The experimental analysis carried out aims in highlighting the use of the SDR in PL mitigationand DSA in a generic communication system which can include FD communication.

61

4.6.1 Experimental Setup

4.6.1.1 Setup Requirements

Universal Software Radio Peripherals (USRPs) and a 64 bit PC constitute the hardware requirementsfor the experimental setup. In our SPCRC lab, Ettus USRP N210 was used. The N210 hardware isideally suited for applications requiring high RF performance and great bandwidth. Its architectureincludes a Xilinx Spartan 3A-DSP 3400 FPGA, 100 MS/s dual ADC, 400 MS/s dual DAC and GigabitEthernet connectivity to stream data to and from host processors. The RF front end of the USRP consistsof a daughter board which act as a transceiver and antennas used to receive the RF signals. Here, WBXdaughtercard and VERT900 antennas were used. The WBX provides 40 MHz of bandwidth capabilityand is ideal for applications requiring access to a number of different bands within its range : 50 MHz to2.2 GHz. VERT900 antennas operate at 824-960 MHz and 1710-1990 MHz bands. The PC used in theexperiment was Lenovo ThinkCare with 64 bit processor and 2 GB RAM. The software requirementsfor the setup include GNU Radio which is an open source software.

Note: Though the VERT900 antenna is said to operate in 824-960 MHz and 1710-1990 MHz bands,yet during the experiment, reliable data transmission was observed using the VERT900 antenna for thecomplete band of 824-1990 MHZ.

Figure 4.17 The path loss experimental setup at SPCRC lab in IIIT Hyderabad

4.6.1.2 Deployment of the Setup

For the deployment of hardware, as shown in fig.4.17, two USRP N210 were connected to the PCthrough a LAN switch. The LAN switch was connected to the gigabit Ethernet port of the PC. All theconnections were made using the Cat 5e cables. The WBX daughter board was used with the USRPs.VERT900 antennas were connected to the RF1 and RF2 of the USRPs. One of the USRP was used as

62

transmitter and the other as a receiver. Both the USRPs were at a distance of about 2 m from each other.The transmitter was placed 1.5 m above the floor and the receiver was placed 1 m above the floor. Therewas no line of sight (LOS) between the transmitter and receiver and they both were separated by twowooden slabs in the lab. The USRPs were interfaced with the PC using the GNU Radio software.

There are two basic experiments conducted to show the utility of SDR for combating the PL. Inthe first experiment, the effect of PL on the received power or received signal strength (RSS) withrespect to operating frequency was studied. For the experiment, an unmodulated sinusoidal signal offrequency 1 KHz was transmitted from transmitter to receiver. The RSS was measured for four operatingfrequencies: 830 MHz, 1.2 GHz, 1.6 GHz and 1.9 GHz for a transmit power of -10 dB. The ITU indoorpropagation model was used to analyze the path loss and interpolate values of RSS for other intermediateoperating frequencies. The ITU indoor model is given by [80]:

PL = 20 log f +N log d+ Pf (n)− 28 (4.58)

where PL is the path loss in decibel (dB), f is the operating frequency in MHz, N is the distancepower loss coefficient, d is the distance in meters (m), n is the number of floors between transmitter andreceiver and Pf (n) is the floor loss penetration factor. As the transmitter and receiver are on the samefloor, Pf (n) is neglected. Also PL = Pt − Pr , where Pt and Pr are the signal strength at receiver andtransmitter, respectively. Hence, the equation (4.58) can be written as:

Pr = Pt − 20 log f −N log d+ 28 (4.59)

In the experiment, the transmitted power and the distance were kept constant and therefore equation(4.59) can be represented as:

Pr = α− 20 log f (4.60)

where α = Pt − N log d + 28. Though every factor is kept constant, there can be slight variation invalue of the PL with time due to changing environmental conditions which result in minor variation inRSS. Hence equation (4.60) can be represented as a function of time:

Pr(t) = α(t)− 20 log f (4.61)

where α(t) = Pt − N log d + 28 − β(t) and β(t) is a time varying factor dependent on operatingenvironment. Four different sets of RSS were measured at time intervals of 30 minutes for four operatingfrequencies: 830 MHz, 1.2 GHz, 1.6 GHz and 1.9 GHz.

In the second experiment, the dynamic variation of RSS of an OFDM system with changing operatingfrequency is analyzed. The analysis is carried out according to the LTE standards. For transmission,random bits were generated and were modulated using 16-QAM. The modulated bits were then mappedon to the OFDM symbols. An FFT size of 512 was used for OFDM, out of which 200 tones wereoccupied. A cyclic prefix of size 128 was used. The OFDM symbols were then transmitted through

63

the USRP. At the receiver, the reverse procedure was employed to recover the transmitted bits. In theexperiment, the operating frequency was varied on-fly between 830 MHz and 1.9 GHz using the variableslider. Also, the transmitter RF chain gain was set as a variable which can be varied from 10 dB to 40dB using a slider. The receiver RF chain gain was constant at 10 dB.

4.6.1.3 Observations

The RSS at frequencies of 830 MHz, 1.2 GHz, 1.6 GHz and 1.9 GHz for four sets measured atregular interval of 30 minutes is tabulated in table.4.1. For each set, in order to compensate for possiblevariation due to a changing environment, the αavg is obtained by taking the average value of α obtainedfrom the four measured RSS using (4.61).

Table 4.1 Received signal strength (in decibel) for four sets measured

Frequency 830MHz 1.2GHz 1.6GHz 1.9GHz αavg

Set1 -43.09 -53.53 -60.85 -63.15 126.59

Set2 -43.09 -55.92 -61.84 -62.50 126.20

Set3 -41.44 -56.57 -62.14 -62.82 126.03

Set4 -42.70 -56.57 -61.84 -62.10 126.23

For each set, RSS for the frequency range of 830 MHz to 1.9 GHz is calculated through interpolationusing measured RSS at the four operating frequencies and plotted using MATLAB. The plot is shownin fig.4.18. It can be observed that higher the operating frequency, more is the attenuation suffered bythe transmitted signal and hence lesser is the RSS.

Next, we look at the second experiment which helps in understanding the utility of SDR in real timeapplication. The FFT of the trasnsmitted signal, in baseband, can be observed in the fig.4.19. All theFFT plots henceforth is shown in passband to help observe the plots w.r.t the operating frequency. Thetransmitter RF chain gain was kept constant at 0 dB. The FFT plot of received signal can be seen infig.4.20. The received signal was initially received at 1.9 GHz. The transmission frequency was thenchanged from 1.9 GHz to 1.6 GHz to 1.2 GHz and finally to 830 MHz. It is observed as the operatingfrequency is decreased the RSS is increased, resulting in the improved link quality.

In the second part of the experiment, the transmitter RF chain gain is varied on-fly. The receiverRF chain gain was kept constant at 10 dB. Initially, the received signal was received at 1.9 GHz andtransmitter RF chain gain was at 0 dB. The transmitter RF chain gain was then increased on-fly from 0dB to 13 dB to 26 dB and finally to 40 dB. From fig.4.21, it can be seen that as the gain is increased, theRSS also increases. The RSS at 1.9 GHz and transmitter RF chain gain of 40 dB is comparable to theRSS at 830 GHz and transmitter RF gain of 0 dB.

64

0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8

x 109

−60

−59

−58

−57

−56

−55

−54

−53

−52

−51

Frequency (GHz)

Re

ce

ive

d P

ow

er

(dB

)

Received Power plot−1

Received Power plot−2

Received Power plot−3

Received Power plot−4

Figure 4.18 Received signal strength w.r.t. operating frequency

Figure 4.19 FFT plot for transmitted signal in baseband

65

Figure 4.20 FFT plot for received signal strength in pass band for operating frequency of a) 1.9 GHz b)1.6 GHz c) 1.2 GHz d) 830 MHz

Figure 4.21 FFT plot for received signal strength in pass band at 1.9 GHz for transmitter RF chain gainof a) 13 dB b) 26 dB c) 40 dB

66

4.6.2 System Deployment

4.6.2.1 SDR in Future Cellular Networks

Extensive research is being done in small cells deploying millimeter-wave (mmW) for communi-cation in future cellular networks. The small cell systems are considered suitable due to low transmitpower, short distances and low mobility [13, 19]. The use of large chunks of underutilized spectrum inthe mmW bands has gained significant interest in realizing the aforementioned 5G vision and require-ments. Specifically [19] considers the 28- and 38-GHz bands to be initial frequencies where mmWcellular systems could operate. In many dense urban areas, cell sizes are now often less than 100-200 min radius, possibly within the range of mmW signals based on measurements in [19].

For mmW, path loss plays a significant role in degrading the link quality due to higher attenuation. Ina scenario, where in the small cell the operating environment is changing dynamically, sudden degrada-tion of operating conditions or arrival of obstruction between transmitter and receiver may result in linkfailure (fig.4.22). In such a scenario, one solution could be dynamically allocating spectrum resource ata lower operating frequency for a link suffering from high path loss. As observed from the experimentcarried out in the previous section, as the operating frequency is decreased, the RSS increases. This isquite helpful when the RSS is below the noise floor due to high attenuation at higher operating frequency.One technology which will complement the deployment of SDR in future cellular network is full-duplexcommunication [1]. The full-duplex communication will allow same spectrum resource for both uplinkand downlink operation. This will facilitates simultaneous change of the operating frequency at boththe uplink and downlink.

The change in the operating frequency of the communication system is dependent upon metricswhich can be decided by the network operators. For example, the increase in the block error rate(BLER) above the tolerable level or the decrease of RSS below the noise floor can trigger the change inthe operating frequency to lower values. Also, as discussed in the previous section, the RF chain gainof both transmitter and receiver can also be dynamically changed to maintain the RSS.

Note: These methods include some challenges and are considered only when current methods ofadaptive MCS and PC fails to tackle the problem of path loss at higher operating frequencies.

4.6.2.2 Challenges in Deployment of the System

There are many challenges when deploying SDR in cellular networks. Some of the important chal-lenges include: 1) As the operating frequency is decreased to lower range, there will be increasedrequirement of bandwidth at lower frequencies. Hence it should be ensured that the operating frequencyis decreased only when there is availability of enough spectrum resource at lower frequencies. 2) Thedeployment of SDR requires the use of programmable architecture at the eNB and UE front end alongwith multi-band re-configurable antennas. This will allow a dynamic shift in operating frequency. 3)Use of full-duplex in cellular network will increase the computational complexity of the communicationsystem [1].

67

Figure 4.22 Drop in received signal strength due to the obstruction between eNB and UE

4.7 Summary

In this chapter, we discussed a full-duplex (FD) communication scenario, where multiple FD UEsshare same spectrum resource (or resource blocks) simultaneously. The sharing of same resource blocksresults in CCI at the downlink of a UE from uplink signals of other UEs. To mitigate the interference,two possible solutions are proposed: 1) A smart antenna approach which includes using multiple an-tennas at UEs to form directed beams towards eNodeB and nulls towards other UEs and 2) A Diversitygain method where diversity gain at the receivers (UEs) is exploited to maximize SINR and thus mit-igate CCI. However, the approaches can fail due to dynamic nature of the operating environment andhence, we proposed a dynamic resource block allocation (DRBA) algorithm for avoiding CCI by properscheduling of the UEs. Finally, we demonstrate use of the SDR in improving the quality the communi-cation links by enhancing the received signal strength by dynamic spectrum access and increasing RFgains.

68

4.8 Appendix A

The weighted sum of the outputs of Nr antennas is given by:

ydl = WHl ydl

=

Nr∑k=1

w∗l,kydl,k

(4.62)

where Wl = [wl,1, wl,2, ..., wl,Nr ]T is the Nr × 1 beamforming vector for the lth UE. The expectedoutput power of the beamformer can be given by:

E[|ydl |2] = E[WHl ydl (y

dl )HWl]

= WHl RlWl

(4.63)

In order to minimize the interference from the other UEs at the downlink of lth UE, we formulatethe following minimization problem:

minimizeWl

WHl RlWl

subject to WHl α(ψx) = 1

(4.64)

where Rl = E[ydl (ydl )H ]. The optimal beamfromer weight vector WH

l obtained aims in decreasing theoverall output power of the receiver, keeping a constant channel response in the direction of eNB. Thisis the classical Frost adaptive beamformer algorithm [81].

The optimal weight vector can be found out by the method of Lagrange multipliers. The constraintfunction is adjoined to the cost function by an undetermined Lagrange multiplier λ. The Lagrangianfunction is given by:

L = WHl RlWl + λ

(WH

l α(ψx)− 1)

(4.65)

Taking the gradient of (4.65),with respect to WHl , we get:

∂L

∂WHl

= RlWl + λα(ψx) (4.66)

The first term in (4.66) is a vector proportional to the gradient of the cost function (4.64), and thesecond term is a vector normal to the constraint plane defined by WH

l α(ψx) − 1 = 0 . For optimalitythese vectors must be anti-parallel, which is achieved by setting (4.66) equal to zero. Hence the optimalweight vector is given by:

Wlopt = −λR−1l α(ψx) (4.67)

69

where R−1l is a positive definite matrix assuming the outputs fromNr elements of a UE are independent.

The Lagrange multiplier can be found out by substituting (4.67) in (4.64) and is given by:

λ∗ = −[(α(ψx))HR−1

l α(ψx)]−1

(4.68)

Taking the conjugate for both the sides and observing that the right hand side of the equation is a realquantity, we have:

λ = −[(α(ψx))HR−1

l α(ψx)]−1

(4.69)

where the existence of [((α(ψx))HR−1l α(ψx)]−1 follows from the facts that Rl is positive definite and

α(ψx) has full rank. From (4.67) and (4.68) the optimum weight vector is given by:

Wlopt =[(α(ψx))HR−1

l α(ψx)]−1

R−1l α(ψx) (4.70)

The correlation matrix Rl is usually unknown a priori and has to be learned using adaptive tech-niques. For this, a gradient descent algorithm is used to calculate an estimate of the optimum weightvector at regular period. For the sake of the algorithm derivation, let us assume temporarily that thecorrelation matrix Rl is known. In gradient-descent method, the weight vector is initialized at a vectorsatisfying the constraint given in (4.64), .i.e, Wl(0) = (1/Nr)α(ψx). At each iteration the weight vec-tor is moved in the negative direction of the constrained gradient (4.66). The rate of the convergence ofthe weight vector is proportional to the magnitude of the constrained gradient and is scaled by a constantµ. After the tth iteration the next weight vector is given by:

Wl(t+ 1) = Wl(t)− µ5WHlL (4.71)

Substituting (4.66) into (4.71), we get:

Wl(t+ 1) = Wl(t)− µ[RlWl(t) + λ(t)α(ψx)

](4.72)

From the constraint (4.64), we have:

WHl (t+ 1)α(ψx) =

[WH

l (t)− µWHl (t)Rl − µλ∗(t)(α(ψx))H

]α(ψx)

= 1(4.73)

Rearranging (4.73), λ(t) at the tth iteration is given by:

λ∗(t) = −(

(α(ψx))Hα(ψx))−1[

µ−1(1−WH

l (t)α(ψx)− µWHl (t)Rlα(ψx)

)](4.74)

Taking conjugate of both side of the equation we have:

70

λ(t) = −(

(α(ψx))Hα(ψx))−1[

µ−1(1−WT

l (t)(α(ψx))∗ − µWTl (t)Rl(α(ψx))∗

)](4.75)

Substituting (4.75) in (4.72), we have:

Wl(t+ 1) = Wl(t)− µ

[RlWl(t)− α(ψx)

((α(ψx))Hα(ψx)

)−1

[µ−1

(1−WT

l (t)(α(ψx))∗ − µWTl (t)Rl(α(ψx))∗

)]](4.76)

The weight vector in (4.76) can be rearranged as following:

Wl(t+ 1) = Wl(t)− µRlWl(t) + α(ψx)(

(α(ψx))Hα(ψx))−1(

1− (α(ψx))HWl(t))

− µα(ψx)(

(α(ψx))Hα(ψx))−1

(α(ψx))HRlWl(t) (4.77)

Wl(t+ 1) = Wl(t)− µ[I− α(ψx)

((α(ψx))Hα(ψx)

)−1]RlWl(t)

+ α(ψx)(

(α(ψx))Hα(ψx))−1[

1− (α(ψx))HWl(t)]

(4.78)

Let P =[I − α(ψx)

((α(ψx))Hα(ψx)

)−1(α(ψx))H

]and F = α(ψx)

((α(ψx))Hα(ψx)

)−1.

Hence, (4.78) can be simplified to the following:

Wl(t+ 1) = P[Wl(t)− µRWl(t)

]+ F (4.79)

However, the (4.80) is the gradient descent algorithm requiring the correlation matrix which is usu-ally unavailable a priori. An available and simple approximation for Rl at the tth iteration is the outerproduct ydl (y

dl )H . Substituting this approximation in (4.79), we get:

Wl(t+ 1) = P[Wl(t)− µydl (y

dl )H]

+ F (4.80)

For optimal convergence of the algorithm, the µ is chosen as [81]:

0 < µ <2

3tr[ydl (ydl )H ]

(4.81)

71

4.9 Appendix B

The noise and interference in the equation (4.48) are given by:

Il,k = nl,k +[ K∑q=lq 6=l

Nr∑k=1

huq,k,l,k

⊗ suq,k

]αk−1q 4k (4.82)

After removing the CP and converting it to the frequency domain, we have:

Il,k = FNIl,k

= nl,k +[ K∑q=lq 6=l

Nr∑k=1

Huq,k,l,k

xuq,k

]αk−1q 4k

(4.83)

where Huq,k,l,k

= diag(FNhuq,k,l,k

) is the NXN diagonal matrix whose diagonal elements are fre-

quency domain coefficients between kth transmit antenna of the qth UE and kth receive antenna of thelth UE. xu

q,kis the transmit signal from k antenna of the qth UE. This signal is then subjected to the

subcarrier deallocation and after simplifications, similar to equation (4.50), we get:

Il(m) = nl(m) +K∑q=lq 6=l

αq4[Huq,l(m)xuq (m)

](4.84)

where Il(m) = [Il,1(m), Il,2(m), ..., Il,Nr(m)]T , nl(m) = [nl,1(m), nl,2(m), ..., nl,Nr(m)]T , xuq (m) =

[xuq,1(m), xuq,2(m), ..., xuq,Nr(m)]T , αq = diag(α0

q , α1q , ...., α

Nr−1q ) and 4 = diag(41,42, ...,4Nr).

Huq,l(m) is theNrXNr frequency domain channel coefficient vector between the lth UE and the qth UE

on the mth subcarrier. This signal for kth antenna of the lth UE is given by:

Il,k(m) = nl,k(m) +[ K∑q=lq 6=l

Nr∑k=1

Huq,k,l,k

(m)xuq,k

(m)]αk−1q 4k (4.85)

Now, considering the fact that the nl,k(m) and Huq,k,l,k

(m) are i.i.d, the power in the above signal isgiven by:

E[|Il,k(m)|2] = σ2nl(m) +

K∑q=lq 6=l

Nr∑k=1

σ2q (β

qmN

−1r )

= σ2nl(m) +

K∑q=lq 6=l

σ2qβ

qm, ∀l, k

(4.86)

72

The signal is now subjected to post-processing from the equation (4.52), .i.e, multiplying with(Ud

m,l)H :

Il(m) = (Udm,l)

H Il(m) (4.87)

where Il(m) = [Il,1(m), Il,2(m), ..., Il,Q(m)]T . As Q = Nr and Udm,l is an unitary matrix, we have

|Il(m)|2 = |Il(m)|2. The power of the signal at the jth data stream of the lth UE is given by:

E[|Il,j(m)|2] = E[|Il,k(m)|2]

= σ2nl(m) +

K∑q=lq 6=l

σ2qβ

qm,∀j

(4.88)

The received signal for the lth user at the mth subcarrier obtained by the MRC is given in equation(4.54) as:

ˆxdl (m) = ((Edl (m))H(Ed

l (m))−1(Edl (m))H ydl (m)

= xdl (m) + ((Edl (m))H(Ed

l (m))−1(Edl (m))H Il(m)

(4.89)

From this signal, the interference and noise term is represented by:

In = ((Edl (m))H(Ed

l (m))−1(Edl (m))H Il(m) (4.90)

As the elements of Il(m) are i.i.d, the power in the above signal is given by:

E[|In|2] = [||Edm,l||2]−1|Il,k(m)|2

= [||Edm,l||2]−1(σ2

nl(m) +K∑q=lq 6=l

σ2qβ

qm) (4.91)

Hence, the effective SINR (γeffml ) for the lth UE on the mth subcarrier can now be given by:

γeffml = [σ2nl

+

K∑q=lq 6=l

σ2qβ

qm]−1||Ed

m,l||2|xdl (m)|2

= [σ2nl

+

K∑q=lq 6=l

σ2qβ

qm]−1||Ed

m,l||2(βlmQ−1)

(4.92)

73

Chapter 5

Conclusions and Future Work

5.1 Summary and Conclusion

Growing interest in very high data rate and high-quality services on one hand and scarcity of availablefrequency resources on the other hand urge for exploiting new methods that can facilitate significantimprovement in spectral efficiencies. Full-duplex communication is considered as a potential solutionfor increasing the throughput of the future generation of wireless networks, by allowing a transceiver totransmit and receive simultaneously in the same frequency band. Recent advances in self-interferencecancellation have enabled the commercial use of FD transceivers. In this thesis, we have proposedtransceiver architecture facilitating multiuser FD communication, where multiple UEs share the samespectrum resources. The main challenge of deploying multiuser FD communication is the co-channelinterference (CCI) at the downlink of a UE from the uplink signals of other co-existing UEs. To tacklethis interference, we have proposed two different techniques: (i) Use of the smart antenna technologyand (ii) Use of the diversity gain at the receiver. For the conventional multiuser downlink and uplinkoperations, the eNB uses SVD based beamforming to transmit parallel data-streams to the multipleco-exiting UEs in the downlink. The successive interference cancellation is used in the uplink fordecoupling the signals of the UEs.

Besides this, a literature survey is carried out for the techniques facilitating the implementation ofmultiuser FD communication. These techniques include dynamic spectrum access based on cognitiveradio technology, DoA estimation using MUSIC algorithm and self-interference cancellation for bothsingle and multiple antenna systems.

5.2 Future Work

So far we have considered the deployment of the multiuser full-duplex communication system ina sub-6 GHz band. The work can be extended to millimeter-wave frequencies (30 GHz - 300 GHz)which promise further improvement in the spectral efficiency. Communication over millimeter-wave(mmWave) frequencies is defining a new era of wireless communication and most recently cellular

74

systems. Recent studies show that the combination of high-bandwidth channels, network densification,and large antenna arrays at both the base station and mobile users can provide coverage comparableto conventional lower-frequency networks but with much higher data rates. Reaping the gains offeredby mmWave requires multiple-input multiple-output (MIMO) signal processing, which leverages thehigher aperture created by the antenna arrays in a way that respects the hardware design challenges inmmWave circuits.

MIMO precoding/combining in mmWave systems is generally different than precoding at lowerfrequencies used in current cellular systems. One reason is that hardware constraints are different:while the small wavelength of mmWave signals allows a large number of antennas to be packed into asmall form factor, the high cost and power consumption of some mixed signal components, like high-resolution analog-to-digital converters (ADCs), makes it difficult to dedicate a separate complete radiofrequency (RF) chain with these components for each antenna. This makes the conventional architec-ture in current cellular systems, where precoding and combining are performed entirely in the digitalbaseband, infeasible. A second difference is that MIMO systems in mmWave will make use of a largenumber of antennas. This impacts the complexity of key signal processing functions like channel es-timation, precoding, combining, and equalization. Moreover, mmWave propagation characteristics aredifferent, so that the MIMO channel is not as “rich” at mmWave. For example, measurements show thatthe mmWave channel is sparse in the angular domain, which can be leveraged to realize efficient pre-coding/combining algorithms. Finally, mmWave communication channels will use a large bandwidth,meaning that broadband channel equalization will still be required. Because of the hardware constraints,the large number of antennas, the different channel conditions, and the larger channel bandwidth, newMIMO transceiver architectures are needed for mmWave systems.

In [82], two potential mmWave MIMO transceiver architectures is presented inspired by the hardwareconstraints while still providing high data rates. The first solution is a hybrid analog/digital precoding(and combining) in which the required precoding and beamforming are divided between the analog anddigital domains. The second solution is the use of low-resolution ADCs to reduce power consumptionat the receiver.

It will be challenging research problem to develop a self-interference architecture, keeping in mindthe challenges of mmWave communication and using the system for multiuser full-duplex communica-tion.

75

Related Publications

• Pradhan, C., Sankhe, K., Kumar, S., & Rama Murthy, G., “Cognitive base station design forefficient spectrum utilization in cellular network.”, in Wireless and Optical CommunicationsNetworks (WOCN), 2014 Eleventh International Conference on (pp. 1-7). IEEE.

• Sankhe, Kunal, Chandan Pradhan, Sumit Kumar, and Garimella Rama Murthy. “Cost effectiverestoration of wireless connectivity in disaster hit areas using OpenBTS.” In India Conference(INDICON), 2014 Annual IEEE, pp. 1-6. IEEE, 2014.

• Pradhan, Chandan, Kunal Sankhe, Sumit Kumar, and Garimella Rama Murthy. “Revamp ofeNodeB for 5G networks: Detracting spectrum scarcity.” In Consumer Communications andNetworking Conference (CCNC), 2015 12th Annual IEEE, pp. 862-868. IEEE, 2015.

• Pradhan, Chandan, Kunal Sankhe, Sumit Kumar, and Garimella Rama Murthy. “Full-Duplex eN-odeB and UE Design for 5G Networks.” , accepted in Wireless Telecommunication Symposium(WTS 2015), New York.

• Sankhe, Kunal, Chandan Pradhan, Sumit Kumar, and Garimella Ramamurthy. “Machine Learn-ing Based Cooperative Relay Selection in Virtual MIMO.”, accepted in Wireless Telecommu-nication Symposium (WTS 2015), New York.

• Pradhan, Chandan, and Garimella Rama Murthy. “Full-Duplex Transceiver for Future Cel-lular Network: A Smart Antenna Approach.”, accepted in IEEE International Conference onAdvanced Networks and Telecommunications Systems (ANTS 2015), ISI Kolkata.

• Pradhan, Chandan, and Garimella Rama Murthy. “Analysis of Path Loss mitigation throughDynamic Spectrum Access: Software Defined Radio.”, accepted in International Conferenceon Microwave, Optical and Communication Engineering (ICMOCE-2015), IIT Bhubaneswar.

• Pradhan, Chandan, and Garimella Rama Murthy. “Full-Duplex Communication for FutureWireless Networks: Dynamic Resource Block Allocation Approach.”, accepted in PhysicalCommunication, Elsevier.

• Pradhan, Chandan, and Garimella Rama Murthy. “Analysis of Full-Duplex Downlink UsingDiversity Gain.”, submitted in Physical Communication, Elsevier.

76

Bibliography

[1] D. Bharadia, E. McMilin, and S. Katti, “Full duplex radios,” in ACM SIGCOMM Computer Com-munication Review, vol. 43, pp. 375–386, ACM, 2013.

[2] Z. Zhang, K. Long, A. V. Vasilakos, and L. Hanzo, “Full-duplex wireless communications: Chal-lenges, solutions and future research directions,” Proceeding of the IEEE, pp. 1–45, 2015.

[3] J. I. Choi, M. Jain, K. Srinivasan, P. Levis, and S. Katti, “Achieving single channel, full duplexwireless communication,” in Proceedings of the sixteenth annual international conference on Mo-bile computing and networking, pp. 1–12, ACM, 2010.

[4] D. Bharadia and S. Katti, “Full duplex mimo radios,” in 11th USENIX Symposium on NetworkedSystems Design and Implementation (NSDI 14), pp. 359–372, 2014.

[5] C.-X. Wang, F. Haider, X. Gao, X.-H. You, Y. Yang, D. Yuan, H. Aggoune, H. Haas, S. Fletcher,and E. Hepsaydir, “Cellular architecture and key technologies for 5G wireless communicationnetworks,” Communications Magazine, IEEE, vol. 52, no. 2, pp. 122–130, 2014.

[6] A. Gohil, H. Modi, and S. K. Patel, “5G technology of mobile communication: A survey,” inIntelligent Systems and Signal Processing (ISSP), 2013 International Conference on, pp. 288–292,IEEE, 2013.

[7] G. Wunder, M. Kasparick, S. ten Brink, F. Schaich, T. Wild, I. Gaspar, E. Ohlmer, S. Krone,N. Michailow, A. Navarro, et al., “5GNOW: Challenging the LTE design paradigms of orthog-onality and synchronicity,” in Vehicular Technology Conference (VTC Spring), 2013 IEEE 77th,pp. 1–5, IEEE, 2013.

[8] G. Wunder, P. Jung, M. Kasparick, T. Wild, F. Schaich, Y. Chen, S. Brink, I. Gaspar, N. Michailow,A. Festag, et al., “5GNOW: Non-orthogonal, asynchronous waveforms for future mobile applica-tions,” Communications Magazine, IEEE, vol. 52, no. 2, pp. 97–105, 2014.

[9] A. Osseiran, F. Boccardi, V. Braun, K. Kusume, P. Marsch, M. Maternia, O. Queseth, M. Schell-mann, H. Schotten, H. Taoka, et al., “Scenarios for 5G mobile and wireless communications: Thevision of the METIS project,” Communications Magazine, IEEE, vol. 52, no. 5, pp. 26–35, 2014.

77

[10] Y. Kishiyama, A. Benjebbour, T. Nakamura, and H. Ishii, “Future steps of LTE-A: Evolutiontoward integration of local area and wide area systems,” Wireless Communications, IEEE, vol. 20,no. 1, pp. 12–18, 2013.

[11] K. Mallinson, “The 2020 vision for LTE,” Available: http://www.fiercewireless.com/europe/story/mallinson-2020-vision-lte/ 2012-06-20#ixzz1yVRoLFcK.

[12] M. N. Tehrani, M. Uysal, and H. Yanikomeroglu, “Device-to-device communication in 5G cellularnetworks: Challenges, solutions, and future directions,” Communications Magazine, IEEE, vol. 52,no. 5, pp. 86–92, 2014.

[13] V. Jungnickel, K. Manolakis, W. Zirwas, B. Panzner, V. Braun, M. Lossow, M. Sternad, R. Apel-frojd, and T. Svensson, “The role of small cells, coordinated multipoint, and massive MIMO in5G,” Communications Magazine, IEEE, vol. 52, no. 5, pp. 44–51, 2014.

[14] T. Irnich, J. Kronander, Y. Selen, and G. Li, “Spectrum sharing scenarios and resulting technicalrequirements for 5G systems,” in Personal, Indoor and Mobile Radio Communications (PIMRCWorkshops), 2013 IEEE 24th International Symposium on, pp. 127–132, IEEE, 2013.

[15] M. Jain, J. I. Choi, T. Kim, D. Bharadia, S. Seth, K. Srinivasan, P. Levis, S. Katti, and P. Sinha,“Practical, real-time, full duplex wireless,” in Proceedings of the 17th annual international con-ference on Mobile computing and networking, pp. 301–312, ACM, 2011.

[16] A. Sabharwal, P. Schniter, D. Guo, D. W. Bliss, S. Rangarajan, and R. Wichman, “In-band full-duplex wireless: Challenges and opportunities,” Selected Areas in Communications, IEEE Journalon, vol. 32, no. 9, pp. 1637–1652, 2014.

[17] F. OHara and G. Moore, “A high performance CW receiver using feedthru nulling,” MicrowaveJournal, vol. 6, no. 9, pp. 63–71, 1963.

[18] T. S. Rappaport, F. Gutierrez, E. Ben-Dor, J. N. Murdock, Y. Qiao, J. Tamir, et al., “Broadbandmillimeter-wave propagation measurements and models using adaptive-beam antennas for outdoorurban cellular communications,” Antennas and Propagation, IEEE Transactions on, vol. 61, no. 4,pp. 1850–1859, 2013.

[19] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz,M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5G cellular: It willwork!,” Access, IEEE, vol. 1, pp. 335–349, 2013.

[20] W. Roh, J.-Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, and F. Aryanfar, “Millimeter-wave beamforming as an enabling technology for 5G cellular communications: Theoretical feasi-bility and prototype results,” Communications Magazine, IEEE, vol. 52, no. 2, pp. 106–113, 2014.

78

[21] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “Next generation/dynamic spectrum ac-cess/cognitive radio wireless networks: A survey,” Computer networks, vol. 50, no. 13, pp. 2127–2159, 2006.

[22] R. W. Thomas, D. H. Friend, L. A. Dasilva, and A. B. Mackenzie, “Cognitive networks: Adaptationand learning to achieve end-to-end performance objectives,” Communications Magazine, IEEE,vol. 44, no. 12, pp. 51–57, 2006.

[23] R. W. Thomas, D. H. Friend, L. A. DaSilva, and A. B. MacKenzie, Cognitive networks. Springer,2007.

[24] J. Xiao, R. Hu, Y. Qian, L. Gong, and B. Wang, “Expanding LTE network spectrum with cognitiveradios: From concept to implementation,” Wireless Communications, IEEE, vol. 20, no. 2, pp. 12–19, 2013.

[25] M. Matinmikko, M. Palola, H. Saarnisaari, M. Heikkila, J. Prokkola, T. Kippola, T. Hanninen,M. Jokinen, and S. Yrjola, “Cognitive radio trial environment: First live authorized shared access-based spectrum-sharing demonstration,” Vehicular Technology Magazine, IEEE, vol. 8, no. 3,pp. 30–37, 2013.

[26] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms for cognitive radio applica-tions,” Communications Surveys & Tutorials, IEEE, vol. 11, no. 1, pp. 116–130, 2009.

[27] G. P. Joshi, S. Y. Nam, and S. W. Kim, “Cognitive radio wireless sensor networks: Applications,challenges and research trends,” Sensors, vol. 13, no. 9, pp. 11196–11228, 2013.

[28] V. Osa, C. Herranz, J. F. Monserrat, and X. Gelabert, “Implementing opportunistic spectrum accessin LTE-Advanced,” EURASIP Journal on Wireless Communications and Networking, vol. 2012,no. 1, pp. 1–17, 2012.

[29] A. Lee, L.-n. Chen, A. Song, J. Wei, and H. Hwang, “Simulation study of wideband interferencerejection using adaptive array antenna,” in Aerospace Conference, 2005 IEEE, pp. 1–6, IEEE,2005.

[30] M. Grice, J. Rodenkirch, A. Yakovlev, H. Hwang, Z. Aliyazicioglu, and A. Lee, “Direction ofarrival estimation using advanced signal processing,” in Recent Advances in Space Technologies,2007. RAST’07. 3rd International Conference on, pp. 515–522, IEEE, 2007.

[31] M. I. Skolnik, “Introduction to radar,” Radar Handbook, vol. 2, 1962.

[32] L. C. Godara, Smart antennas. CRC press, 2004.

[33] D. FINDING, “Utilizing waveform features for adaptive beamforming and direction finding withnarrowband signals,” Lincoln Laboratory Journal, vol. 10, no. 2, 1997.

79

[34] T. E. Tuncer and B. Friedlander, Classical and modern direction-of-arrival estimation. AcademicPress, 2009.

[35] W.-Y. Lee and I. F. Akyildiz, “Spectrum-aware mobility management in cognitive radio cellularnetworks,” Mobile Computing, IEEE Transactions on, vol. 11, no. 4, pp. 529–542, 2012.

[36] S. Han, L. Dai, Q. Sun, Z. Xu, et al., “Full duplex networking: Mission impossible?,” arXivpreprint arXiv:1410.5326, 2014.

[37] D. Nguyen, L.-N. Tran, P. Pirinen, and M. Latva-aho, “On the spectral efficiency of full-duplexsmall cell wireless systems,” Wireless Communications, IEEE Transactions on, vol. 13, no. 9,pp. 4896–4910, 2014.

[38] A. Tang and X. Wang, “Balanced RF-circuit based self-interference cancellation for full duplexcommunications,” Ad Hoc Networks, vol. 24, pp. 214–227, 2015.

[39] S. S. Hong, J. Mehlman, and S. Katti, “Picasso: Flexible RF and spectrum slicing,” ACM SIG-COMM Computer Communication Review, vol. 42, no. 4, pp. 37–48, 2012.

[40] D. Bharadia, K. Joshi, and S. Katti, “Robust full duplex radio link,” in Proceedings of the 2014ACM conference on SIGCOMM, pp. 147–148, ACM, 2014.

[41] S.-K. Hong, J. Brand, J. Choi, M. Jain, J. Mehlman, S. Katti, and P. Levis, “Applications of self-interference cancellation in 5G and beyond,” Communications Magazine, IEEE, vol. 52, no. 2,pp. 114–121, 2014.

[42] X. Wang, A. Tang, and P. Huang, “Full duplex random access for multi-user OFDMA communi-cation systems,” Ad Hoc Networks, vol. 24, pp. 200–213, 2015.

[43] S. Huberman and T. Le-Ngoc, “MIMO full-duplex precoding: A joint beamforming and self-interference cancellation structure,” Wireless Communications, IEEE Transactions on, vol. 14,no. 4, pp. 2205–2217, 2015.

[44] B. Debaillie, D.-J. van den Broek, C. Lavin, B. van Liempd, E. Klumperink, C. Palacios, J. Cran-inckx, B. Nauta, A. Parssinen, et al., “Analog/RF solutions enabling compact full-duplex radios,”Selected Areas in Communications, IEEE Journal on, vol. 32, no. 9, pp. 1662–1673, 2014.

[45] B. Yin, M. Wu, C. Studer, J. R. Cavallaro, and J. Lilleberg, “Full-duplex in large-scale wirelesssystems,” in Signals, Systems and Computers, 2013 Asilomar Conference on, pp. 1623–1627,IEEE, 2013.

[46] S. Goyal, P. Liu, S. Panwar, R. DiFazio, R. Yang, J. Li, E. Bala, et al., “Improving small cell capac-ity with common-carrier full duplex radios,” in Communications (ICC), 2014 IEEE InternationalConference on, pp. 4987–4993, IEEE, 2014.

80

[47] L. Song, Y. Li, and Z. Han, “Resource allocation in full-duplex communications for future wirelessnetworks,” arXiv preprint arXiv:1505.02911, 2015.

[48] M. Al-Imari, M. Ghoraishi, and P. Xiao, “Radio resource allocation for full-duplex multicarrierwireless systems,” 2015.

[49] A. Sahai, G. Patel, C. Dick, and A. Sabharwal, “On the impact of phase noise on active cancellationin wireless full-duplex,” Vehicular Technology, IEEE Transactions on, vol. 62, no. 9, pp. 4494–4510, 2013.

[50] T. Riihonen, S. Werner, R. Wichman, and J. Hamalainen, “Outage probabilities in infrastructure-based single-frequency relay links,” in Wireless Communications and Networking Conference,2009. WCNC 2009. IEEE, pp. 1–6, IEEE, 2009.

[51] T. H. Lee, The design of CMOS radio-frequency integrated circuits. Cambridge university press,2003.

[52] M. Duarte, C. Dick, and A. Sabharwal, “Experiment-driven characterization of full-duplex wire-less systems,” Wireless Communications, IEEE Transactions on, vol. 11, no. 12, pp. 4296–4307,2012.

[53] M. Duarte and A. Sabharwal, “Full-duplex wireless communications using off-the-shelf radios:Feasibility and first results,” in Signals, Systems and Computers (ASILOMAR), 2010 ConferenceRecord of the Forty Fourth Asilomar Conference on, pp. 1558–1562, IEEE, 2010.

[54] E. Everett, M. Duarte, C. Dick, and A. Sabharwal, “Empowering full-duplex wireless communi-cation by exploiting directional diversity,” in Signals, Systems and Computers (ASILOMAR), 2011Conference Record of the Forty Fifth Asilomar Conference on, pp. 2002–2006, IEEE, 2011.

[55] M. E. Knox, “Single antenna full duplex communications using a common carrier,” in Wirelessand Microwave Technology Conference (WAMICON), 2012 IEEE 13th Annual, pp. 1–6, IEEE,2012.

[56] K. Haneda, E. Kahra, S. Wyne, C. Icheln, and P. Vainikainen, “Measurement of loop-back in-terference channels for outdoor-to-indoor full-duplex radio relays,” in Antennas and Propagation(EuCAP), 2010 Proceedings of the Fourth European Conference on, pp. 1–5, IEEE, 2010.

[57] B. Radunovic, D. Gunawardena, P. Key, A. P. N. Singh, V. Balan, and G. Dejean, “Rethinkingindoor wireless: Low power, low frequency,” tech. rep., full duplex. Technical report, MicrosoftResearch, 2009.

[58] S. Gollakota and D. Katabi, Zigzag decoding: Combating hidden terminals in wireless networks,vol. 38. ACM, 2008.

81

[59] A. Sahai, G. Patel, and A. Sabharwal, “Pushing the limits of full-duplex: Design and real-timeimplementation,” arXiv preprint arXiv:1107.0607, 2011.

[60] M. A. Khojastepour and S. Rangarajan, “Wideband digital cancellation for full-duplex communi-cations,” in Signals, Systems and Computers (ASILOMAR), 2012 Conference Record of the FortySixth Asilomar Conference on, pp. 1300–1304, IEEE, 2012.

[61] C. R. Anderson, S. Krishnamoorthy, C. G. Ranson, T. J. Lemon, W. G. Newhall, T. Kummetz, andJ. H. Reed, “Antenna isolation, wideband multipath propagation measurements, and interferencemitigation for on-frequency repeaters,” in SoutheastCon, 2004. Proceedings. IEEE, pp. 110–114,IEEE, 2004.

[62] T. Riihonen and R. Wichman, “Analog and digital self-interference cancellation in full-duplexMIMO-OFDM transceivers with limited resolution in A/D conversion,” in Signals, Systems andComputers (ASILOMAR), 2012 Conference Record of the Forty Sixth Asilomar Conference on,pp. 45–49, IEEE, 2012.

[63] L. S. Addison, “Statistical signal processing-detection, estimation and time series analysis,” 1991.

[64] E. Aryafar, M. A. Khojastepour, K. Sundaresan, S. Rangarajan, and M. Chiang, “MIDU: Enablingmimo full duplex,” in Proceedings of the 18th annual international conference on Mobile comput-ing and networking, pp. 257–268, ACM, 2012.

[65] H. S. Eshwaraiah and A. Chockalingam, “SC-FDMA for multiuser communication on the down-link,” in Communication Systems and Networks (COMSNETS), 2013 Fifth International Confer-ence on, pp. 1–7, IEEE, 2013.

[66] Y. S. Cho, J. Kim, W. Y. Yang, and C. G. Kang, MIMO-OFDM wireless communications withMATLAB. John Wiley & Sons, 2010.

[67] J. Salz and J. H. Winters, “Effect of fading correlation on adaptive arrays in digital mobile radio,”Vehicular Technology, IEEE Transactions on, vol. 43, no. 4, pp. 1049–1057, 1994.

[68] M. Dosaranian-Moghadam and H. Bakhshi, “Tracking mobile user through adaptive beamform-ing for wireless cellular networks in a 2D urban environment,” Indian Journal of Science andTechnology, vol. 5, no. 4, pp. 2569–2577, 2012.

[69] C. Pradhan, K. Sankhe, S. Kumar, and G. R. Murthy, “Full-duplex enodeB and UE design for 5Gnetworks,” arXiv preprint arXiv:1506.02132, 2015.

[70] A. Goldsmith, Wireless communications. Cambridge university press, 2005.

[71] K. Zheng, L. Zhao, J. Mei, B. Shao, W. Xiang, and L. Hanzo, “Survey of large-scale MIMOsystems,” Communications Surveys & Tutorials, IEEE, vol. 17, no. 3, pp. 1738–1760, 2015.

82

[72] L. A. M. R. De Temino, G. Berardinelli, S. Frattasi, and P. Mogensen, “Channel-aware schedulingalgorithms for SC-FDMA in LTE uplink,” in Personal, Indoor and Mobile Radio Communications,2008. PIMRC 2008. IEEE 19th International Symposium on, pp. 1–6, IEEE, 2008.

[73] M. Al-Imari, M. Ghoraishi, and P. Xiao, “Radio resource allocation and system-level evaluationfor full-duplex systems,” 2015.

[74] H. Zhang, X. Chu, W. Guo, and S. Wang, “Coexistence of Wi-Fi and heterogeneous small cellnetworks sharing unlicensed spectrum,” Communications Magazine, IEEE, vol. 53, no. 3, pp. 158–164, 2015.

[75] D. Falconer, T. Kolze, and J. L. YigalLeiba, “System impairment model,” Presentation to IEEE,vol. 802, 2000.

[76] T. S. Rappaport et al., Wireless communications: principles and practice, vol. 2. Prentice HallPTR New Jersey, 1996.

[77] A. Simonsson and A. Furuskar, “Uplink power control in LTE-overview and performance, subtitle:Principles and benefits of utilizing rather than compensating for SINR variations,” in VehicularTechnology Conference, 2008. VTC 2008-Fall. IEEE 68th, pp. 1–5, IEEE, 2008.

[78] X. Xu, G. Kutrolli, and R. Mathar, “Dynamic downlink power control strategies for LTE femto-cells,” in 2013 Seventh International Conference on Next Generation Mobile Apps, Services andTechnologies, 2013.

[79] C. G. Christodoulou, Y. Tawk, S. A. Lane, and S. R. Erwin, “Reconfigurable antennas for wirelessand space applications,” Proceedings of the IEEE, vol. 100, no. 7, pp. 2250–2261, 2012.

[80] P. Series, “Propagation data and prediction methods for the planning of indoor radiocommunica-tion systems and radio local area networks in the frequency range 900 MHz to 100 GHz,” 2012.

[81] O. L. Frost III, “An algorithm for linearly constrained adaptive array processing,” Proceedings ofthe IEEE, vol. 60, no. 8, pp. 926–935, 1972.

[82] A. Alkhateeb, J. Mo, N. Gonzalez-Prelcic, and R. W. Heath, “MIMO precoding and combin-ing solutions for millimeter-wave systems,” Communications Magazine, IEEE, vol. 52, no. 12,pp. 122–131, 2014.

83