performanceanalysisanddesigninsights1264028/fulltext01.pdf · samarbete i två udn-scenarier:...

Ultra-Densification for Future Cellular Networks

Performance Analysis and Design Insights

YANPENG YANG

PhD Thesis in Information and Communication TechnologySchool of Electrical Engineering and Computer Science

KTH Royal Institute of TechnologyStockholm, Sweden 2018

TRITA-EECS-AVL-2018:88ISBN 978-91-7873-017-9

KTH School of Electrical Engineering andComputer ScienceSE-164 40 Kista

SWEDEN

Akademisk avhandling som med tillstånd av Kungl Tekniska högskolan framläggestill offentlig granskning för avläggande av teknologie doktorsexamen i Informations-och kommunikationsteknik torsdagen den 7 December 2018 klockan 13:00 i Sal C,Electrum, Kungl Tekniska högskolan, Kistagången 16, Kista.

© Yanpeng Yang, December 2018

Tryck: Universitetsservice US AB

Abstract

The traffic volume in wireless communication has grown dramatically in the lastdecade. Recent predictions indicate such data storm will be even more violent in theshort run. Potential solutions for accommodating the rapid traffic growth can besummed up into three categories: broadening the available bandwidth, improvingthe spectral efficiency, and densifying the infrastructure. In this thesis, we focuson the densification dimension which has been proven to be the most effectiveone in the past. The current gain of network densification mainly comes from cellsplitting, thereby serving more user equipment (UE) simultaneously. This trendwill decelerate as the base station (BS) density gets closer to or even surpassesthe UE density which forms an ultra-dense network (UDN). Thus, it is crucialto understand the behavior and design operations of ultra-densification in futurenetworks.

An important question for future system design and operating strategy is whichelement is more effective than others. To this end, we start from comparing the ef-fectiveness of densification with spectrum expansion and multi-antenna systems interms of meeting certain traffic demand. Our findings show that deploying more BSsprovides a substantial gain in sparse network but the gain decreases progressively ina UDN. Meanwhile, even with the same area throughput, different combinations ofindividual throughput and UE density lead to different requirements for resources.The diminishing gain appearing in UDNs makes us curious to know if there ex-ists a terminal on the way of densification. Such uncertainty leads to the study onthe asymptotic behavior of densification. By incorporating a sophisticated boundeddual-slope path loss model and practical UE densities in our analysis, we present theasymptotic behavior of ultra-densification: the coverage probability and area spec-tral efficiency (ASE) have non-zero convergences in asymptotic regions unless theUE density goes to infinity (full load). Our results suggest that network densifica-tion cannot always improve the UE performance or boost the network throughput.

Next, we shift our focus to the operations of UDNs. We first study BS coop-erations in two UDN scenarios: homogeneous and heterogeneous UDNs which aredistinguished by BS types. In both cases, the cooperation rules become more compli-cated than those in traditional networks. Either channel state information (CSI) orextra delay information needs to be acquired in order to obtain cooperation gains.At last, we investigate the feasibility of applying random beamforming to initialaccess in millimeter-wave (mmWave) UDNs. To our surprise, the simple methodcan provide sufficient performance in both control and data plane, comparing withthe existing schemes. Therefore, it may be unnecessary to develop complex algo-rithms for initial access in future dense mmWave networks. The findings indicatethat UDN may complicate network operations while it may also facilitate the use ofsimple schemes. Our work provides insights into the understanding of the networkdensification and thus paves the way for the operational design of future UDNs.

Sammanfattning

Trafikvolymen inom trådlös kommunikation har ökat dramatiskt det senasteårtiondet och de senaste prognoserna visar på en fortsatt kraftig ökning på kortsikt. Potentiella lösningar för att kunna tillgodose den snabba trafikökningen kansammanfattas i tre kategorier: utöka den tillgängliga bandbredden, förbättra spek-traleffektiviteten, och förtäta infrastrukturen. I denna avhandling fokuserar vi påförtätning som historiskt har visat sig vara den mest givande lösningen. Prestan-davinsten av nätverksförtätning kommer huvudsakligen från celldelning som tillåterfler användarterminaler (UE) att tjänas samtidigt. Den här trenden bromsas dockupp när basstationstätheten går mot eller till och med överstiger UE-tätheten ochformar ett ultratätt nätverk (UDN). Det är därmed viktigt att förstå beteendet avultraförtätning för framtida nätverkstillhandahållning.

En viktig fråga för framtida systemdesign och operativ strategi ligger i attbestämma vilken komponent som är den mest effektiva. För detta ändamål jämförvi verkningsfullheten av förtätning med spektrumutökning och multiantennsystem isyfte att tillgodose vissa trafikkrav. Våra resultat visar att förlägga fler basstationer(BS) ger en påtaglig förbättring i glesare nätverk som dock minskar progressivti UDN. Att inkludera fler basstationsantenner eller tillgodoräkna sig mer band-bredd blir då mer fördelaktigt än fortsatt förtätning i UDN. Inverkan av trafiktypdiskuteras också. Även med samma dataflöde per areaenhet så leder olika kombi-nationer av individuella dataflöden och användartäthet till olika resurskrav. Denminskande prestandavinsten i UDN leder till frågan huruvida det finns en ändpunktpå vägen till förtätning. Denna osäkerhet leder oss till att studera asymptotiskt be-teende av förtätning. Vi införlivar en sofistikerad tvådelad sträckdämpningsmodelloch praktiska UE-tätheter i vår analys. Med hjälp av stokastisk geometri härledervi uttryck för täckningssannolikhet för en typisk UE och areaspektraleffektiviteten(ASE) för nätverket. Därutöver presenterar vi asymptotiskt beteende av ultraförtät-ning: täckningssannolikhet och ASE har en konvergens skild från noll i det asymp-totiska fallet såvida inte UE-tätheten går mot oändligheten.

Därefter skiftar vi vårt fokus till verksamheten i ett UDN. Vi studerar BSsamarbete i två UDN-scenarier: homogen och heterogen UDN där olika typer avBS-användare distribueras. I båda fallen blir samarbetsregeln mer komplicerat itraditionella nätverk. Antingen kanalkännedom (CSI) eller extra fördröjningsin-formation måste förvärvas för att få samarbetsvinst. Slutligen undersöker vi möj-ligheten att tillämpa slumpmässig strålformning till initial åtkomst i amillimeter-våg (mmWave) UDN. Till vår förvåning kan den enkla metoden ge tillräcklig pre-standa i både kontroll- och dataplan, jämföra med befintliga system. Därför kan detvara onödigt att utveckla komplexa algoritmer för första åtkomst i framtida tätammWave-nätverk. Resultaten visar att UDN kan komplicera nätverksoperationermedan det även kan underlätta enkla system. Vårt arbete ger insikt i förståelsenav nätförtätningen och banar därmed vägen för driftdesign av framtida UDNs.

Dedicatedto

My Paternal GrandfatherSongwu Yang

and

My Maternal GrandfatherJiguang Lv

AcknowledgmentsThis thesis is the culmination of my PhD journey which was just like climbing

a high peak step by step accompanied with hardship, struggle, frustration, encour-agement, and cooperation. When I found myself at the final destination, I realizedthough only my name appears on the cover of this dissertation, a great many peo-ple have helped me on my way here and contributed to accomplish this huge task.Before naming the helpful people, I wish to thank all the hardships, for not beatingme down but making me stronger.

First and foremost, I express my sincere gratitude to Dr. Ki Won Sung and Prof.Jens Zander for offering me the opportunity to start this wonderful journey in theRS-lab. I owe my biggest thanks to my main advisor Docent Ki Won Sung, foryour patient guidance, encouragement and valuable feedbacks on my work. I alsoreally appreciate him of being a good listener and mentor when I need to pour outtroubles from both work and personal life. I am deeply grateful to my co-advisorProf. Jens Zander for broadening my view of research. His critical and far-sightedthinking has always inspired me and will continue in my future career.

During my doctoral studies, I have had the opportunities to collaborate withmany extraordinary researchers. I am thankful to all my co-authors for our discus-sions which have been a great inspiration for my work. In particular, I am gratefulto Dr. Jihong Park and Dr. Hossein Shokri-Ghadikolaei for sharing your knowledgeand encouraging my research. My special thanks to Prof. Seong-Lyun Kim fromYonsei University for accepting me as a visit PhD student in his lab.

I greatly appreciate Dr. Mikael Prytz and Prof. Ben Slimane for reviewing myPhD proposal and PhD thesis and providing valuable comments. I would like tooffer my special thanks to Prof. Kaibin Huang for accepting the role of opponentin the dissertation defense. Besides, I would like to thank Prof. Anne Håkansson,Prof. Jaap van de Beek, and Dr. Olav Queseth for acting on the grading committee.

I would like to thank all my colleagues in the RS-lab for providing a greatresearch environment and nice discussions with me. I also feel grateful to the formermembers of RS-lab for setting good examples to me on the research journey. Anothercredits should be given to the PhD office and IT-service center for their excellentadministrative support. Outside my academic life, I am indebted to all my wonderfulfriends for the great moments we shared together.

Finally, it is impossible to describe my gratefulness to my parents Bin andJingxin, for their support, encouragement, patience, and endless love. Last but byno means least, I owe a big debt of gratitude to my beloved wife Jie, for brighteningmy life, now and forever.

Yanpeng YangStockholm, Dec. 2018

Contents

Contents xi

List of Figures xiii

List of Tables xv

List of Acronyms xvii

I Thesis Overview 1

1 Introduction 31.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2 From Sparse to Ultra-Dense Networks . . . . . . . . . . . . . . . 51.3 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . 81.4 Thesis Focus and Problem Formulation . . . . . . . . . . . . . . 101.5 Overview of Thesis Contributions . . . . . . . . . . . . . . . . . . 141.6 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Research Methodology and Models 192.1 Research Methodology . . . . . . . . . . . . . . . . . . . . . . . . 19

2.1.1 Stochastic Geometry based Network Analysis . . . . . . . 192.1.2 Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . 21

2.2 System Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.2.1 Network Model . . . . . . . . . . . . . . . . . . . . . . . . 212.2.2 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . 232.2.3 Maximal Ratio Transmission Beamforming Model . . . . 262.2.4 Analog Beamforming Model . . . . . . . . . . . . . . . . . 262.2.5 3GPP NR Frame Model . . . . . . . . . . . . . . . . . . . 28

3 Key Results 293.1 Resource Trade-offs in UDN Provisioning . . . . . . . . . . . . . 29

3.1.1 Technical Rate of Substitution . . . . . . . . . . . . . . . 29

xi

xii Contents

3.1.2 TRS of Densification . . . . . . . . . . . . . . . . . . . . . 303.2 Asymptotic Behavior of Ultra-Densification . . . . . . . . . . . . 33

3.2.1 Coverage and Area Spectral Efficiency Convergence in UDNs 333.3 BS Cooperation in UDNs . . . . . . . . . . . . . . . . . . . . . . 37

3.3.1 Homogeneous UDN scenario . . . . . . . . . . . . . . . . . 373.3.2 Heterogeneous UDN scenario . . . . . . . . . . . . . . . . 40

3.4 Initial Access in mmWave UDNs . . . . . . . . . . . . . . . . . . 433.4.1 Random Beamforming based Initial Access . . . . . . . . 43

4 Conclusion and Future Work 494.1 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 494.2 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

Bibliography 53

List of Figures

1.1 Mobile data traffic forecast . . . . . . . . . . . . . . . . . . . . . . . 41.2 Resource requirement in different traffic demands. . . . . . . . . . . 41.3 Evolution from sparse network to ultra-dense network. . . . . . . . . 61.4 Approaches in propagation and load modeling. . . . . . . . . . . . . 91.5 An illustration of JT in traditional fully-loaded network and partially-

loaded UDN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.6 Existing cell search schemes . . . . . . . . . . . . . . . . . . . . . . . 13

2.1 Network topology based on the PPP assumption . . . . . . . . . . . 202.2 A bounded dual-slope path loss model . . . . . . . . . . . . . . . . . 24

3.1 Relation between antenna number and BS density. . . . . . . . . . . 303.2 Relation between spectrum and BS density. . . . . . . . . . . . . . . 313.3 Resource requirement in different traffic demands. . . . . . . . . . . 333.4 Effect of λu on coverage probability under bounded dual-slope model. 353.5 Impacts of (a) bounded path loss, (b) dual-slope path loss, and (c)

partial load models on asymptotic SIR coverage probability as BSdensity increases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.6 ASE bounds and comparison with previous models. . . . . . . . . . 363.7 Joint transmission in a homogeneous UDN. . . . . . . . . . . . . . . 373.8 UE performance with different critical distances when path loss ex-

ponents=[2,4], λb = 4000/km2, λu = 200/km2. . . . . . . . . . . . . 393.9 Cooperation gain when Rc = 70m, N = 5. . . . . . . . . . . . . . . . 403.10 An example of PW-CRAN. . . . . . . . . . . . . . . . . . . . . . . . 403.11 Average sum SE for scheduled UEs vs. the total delay τtotal. . . . . . 413.12 Average sum SE for scheduled UEs vs. the ratio α of the wireless

link delay to the total delay for varying τtotal. . . . . . . . . . . . . . 423.13 Illustration of transmission frame under random beamforming. . . . 433.14 BS density effect on failure probability. (1) refers ε = 0, αN = ∞,

(2) refers ε = 0, αL ≤ αN < ∞, and (3) refers 0 ≤ ε ≤ 1, αN =∞. ("RB", "IS", and "ES" stand for random beamforming, iterativesearch and exhaustive search) . . . . . . . . . . . . . . . . . . . . . . 45

xiii

xiv List of Figures

3.15 BS density effect on IA delay (ULA model). . . . . . . . . . . . . . . 463.16 Total data transmission latency vs packet size. . . . . . . . . . . . . 46

List of Tables

1.1 Different types of small cells. . . . . . . . . . . . . . . . . . . . . . . 71.2 Comparison of traditional cellular and high-density system charac-

teristics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2.1 Road sign of models applied in the thesis. . . . . . . . . . . . . . . . 22

3.1 TRS between antenna number and densification (4M4λb ). . . . . . . . 303.2 TRS between spectrum and densification (4MHz

4λb ). . . . . . . . . . . 31

xv

List of Acronyms

3GPP 3rd generation partnership project4G The fourth generation of cellular mobile communications5G The fifth generation of cellular mobile communicationsAoA Angle of arrivalAoD Angle of departureASE Area spectral efficiencyBS Base stationC-RAN Cloud radio access networkCAGR Compound annual growth rateCDMA Code-division multiple accessCSI Channel state informationDAS Distributed antenna systemEE Energy efficiencyIoT Internet-of-thingsJT Joint transmissionLOS Line of sightM2M Machine-to-machineMIMO Multiple input multiple outputMISO Multiple input single outputmmWave Millimeter-waveMNO Mobile network operatorMRT Maximal ratio transmission

xvii

xviii List of Acronyms

NLOS Non line of sightNR New radioOFDM Orthogonal frequency division multiplexingPP Point processPPP Poisson point processQoS Quality of serviceRRH Remote radio unitSBP Sectorized beam patternSE Spectral efficiencySIMO Single input multiple outputSINR Signal to interference-plus-noise ratioSIR Signal to interference ratioSS Synchronization signalTDD Time division duplexingTDMA Time division multiple accessTRS Technical rate of substitutionUE User equipmentULA Uniform linear arrayUDN Ultra dense network

Part I

Thesis Overview

1

Chapter 1

Introduction

"The formulation of a problem is often more essential than its solution,which may be merely a matter of mathematical or experimental skill."

Albert Einstein

1.1 Background

The volume of mobile data traffic has undergone an unprecedented growth duringthe past decades. From 2012-2017, the overall mobile data traffic grew around eighttimes [1]. Surprisingly, it is expected to grow at a compound annual growth rate(CAGR) of 47 percent in the near future and reach 49 exabytes per month by 2021,a sevenfold increase over 2017 [2]. As shown in Fig 1.1 [2], smartphones and tabletswill continue dominating mobile traffic (above 90 percent). Over 70 percent of thetraffic from the smart devices are contributed by mobile video, which are supportedby very high data rates. Meanwhile, machine-to-machine (M2M) category will gainmore share in the market. According to [1], around 20 billion internet-of-things(IoT) devices are forecast, of which 1.8 billion will be with cellular connections.

The fourth generation (4G) networks have reached a mature level and theirlimits. To accommodate such rapid traffic growth, the fifth generation (5G) ofwireless networks are expected to improve the network capacity by 1000 timescompared with that in 2010 [3] [4], which poses a tremendous challenge to wirelessresearch community. Fundamentally, the answer to the 1000 times capacity crunchcan be categorized within the following three paradigms as shown in Fig. 1.2:

• Spectrum Expansion, to seek more spectrum such as exploring millimeterwave spectrum and/or the unlicensed bands;

• Spectral Efficiency Enhancement, to design advanced physical layer tech-niques to support high transmission rate, such as efficient modulation and cod-ing schemes, cooperative transmissions and massive multiple-input-multiple-output (MIMO);

3

4 Introduction

Figure 1.1: Mobile data traffic forecast

1x4G today

Hz

Spectrum expansionMillimeter wave, unlicenced bands...

bps/Hz/cell

1000x5G beyond 2020

Further spatial reuseDense deployment, D2D...

Cells/area

Spectral efficiency enhancementMassive MIMO, CoMP...

Figure 1.2: Resource requirement in different traffic demands.

• Spatial Spectrum Reuse, to improve the spectrum reuse per unit area bydensely deploying base stations (BSs), i.e., network densification.

Different from the other two elements implemented in the infrastructures, spec-trum is considered as a ’virtual’ resource. In Shannon’s formula, the additionalbandwidth can lead to an almost linear increase in data capacity (especially in in-terference limited systems). However, the acquisition of more bandwidth is a verycostly solution, unless devices are authorized for spectrum sharing or operating inunlicensed bands. In addition, transmitting in higher carrier frequencies suffers fromlarger attenuation as well as lower penetration, and usually requires more expensiveequipments.

On the other hand, recent studies show that point-to-point link throughput isclose to the theoretical limits [5]. With this regard, multi-antenna or even massiveMIMO system is considered to provide extra capacity. By adding multiple antennas,

1.2. From Sparse to Ultra-Dense Networks 5

it offers a spatial dimension in addition to time and frequency for communicationand yields a degree of freedom gain to accommodate more information data. Withmassive MIMO, a performance improvement can be obtained with regard to relia-bility, spectral efficiency, and energy efficiency [6] [7]. Nevertheless, the performanceof massive MIMO may be limited by the finite and potentially correlated scatteringgiven the space constraints and pilot contamination issues [7]. Meanwhile, chan-nel feedback/estimation for massive antennas is an inherent challenge. As a result,massive MIMO is more favorable in time division duplex (TDD) systems.

Densely deployed small cells are motivated by increasing the reuse of spectrumin the network area. It can decrease the number of user equipment (UE) at eachBS, thereby reducing the competition for resources. Compared with the other twoelements, network densification is considered more practical from regulatory andtechnical perspective for the following reasons. First, the cost of small cell deploy-ment is much lower than the other two options. Both capital expenditure (CAPEX)and operational expenditure (OPEX) of deploying small cells are considerably lowcompared to multi-antenna macro-BS deployment [8], where a significant portion ofthe recurring cost comes from wired backhauling, power usage, and real estate [9].Furthermore, small cells are energy efficient as they can be utilized intelligentlyand opportunistically. Depending on the traffic demand, small cells can turn intosleep mode so that the energy consumption and interference can be reduced. Atlast, by deploying small cells over the macro-cell umbrella heterogeneously, cell edgemay become vague or even vanished. The ’poor cell edge UEs’ will experience goodthroughput over the network.

According to Martin Cooper, the pioneer and visionary, the wireless capacityhas doubled about every two years and increased around 1,000,000 fold from 1950to 2000, also known as ‘Cooper’s Law’. The 1 million gains are composed by 25ximprovement from wider spectrum, 25x improvement from advanced modulationand coding schemes, and an astonishing 1600x improvement brought by deployingmore BSs and reducing cell size [10]. In this regard, network densification has madesignificant contributions to the capacity improvement in the past and is anticipatedto play an important role in the future.

1.2 From Sparse to Ultra-Dense Networks

Network densification has been employed over several cellular generations [11],which is roughly depicted in Figure 1.3. The first generation (1G) mobile net-works in the 1980s, had cell radius on the order of tens of kilometers. Since then,the cell sizes have been progressively shrinking. Later in the 1990s, an early type ofpico-cells with cell radius ranging from tens to around one hundred meters began toemerge [12]. Such kinds of small cells were essentially a smaller version of the macroBS, and required comparable planning, management and network interfaces. Theywere mainly used as a complementary for offloading traffic from busy macro-cellsand providing connections to the coverage holes.

6 Introduction

(a) Traditional sparse network

(b) Small cell network

(c) Ultra dense network

Figure 1.3: Evolution from sparse network to ultra-dense network.

Since entering the 21st century, the demand for higher data rates in wirelessnetworks is unrelenting, which has triggered the design of new cellular architecture.The concept of femto-cell represents a new thinking on the deployment and con-figuration of cellular systems which considers flexibility, energy and cost aspects.Different from the traditional small cells, a femto-cell covers even smaller area butis more autonomous and self-adaptive. Nowadays, networks are rapidly evolvingto include nested small cells such as pico-cells and femto-cells, as well as remoteunit head (RRH) and distributed antenna systems (DAS) [13–15]. The features ofdifferent small cell types are summarized in Table 1.1, which is revised from [16].Such small cells are equipped with agile antennas in replacing traditional bulkytransceiver units and can be installed arbitrarily for various scenarios [17].

Definition of Ultra-dense Network

To reach the high capacity demand for 5G, interest has lately been turning towardultra-dense BS deployment, an evolution of current small cell networks. So, whatis an ultra-dense network (UDN)? Qualitatively, UDN is a network with much

1.2. From Sparse to Ultra-Dense Networks 7

Table 1.1: Different types of small cells.

Type of small cell Deployment Scenario Coverage

Pico-cells (fully functioning) indoor/outdoor (planned) up to 100 meters

Femto-cells (fully functioning) indoor/outdoor (planned/un-planned)

10 - 30 meters

Delays (macro-extension) indoor/outdoor (planned) up to 100 meters

RRHs (macro-extension) indoor/outdoor (planned) up to 100 meters

DAS (macro-extension) indoor/outdoor (planned) up to 100 meters

higher density of radio resources than that in current networks, i.e., much densersmall cell network in terms of either relative density or absolute density of theBSs. Quantitatively, the definitions of UDN vary among the literature, which arecharacterized from different angles. In [18], the definition is given from an absoluteBS density perspective:

• The 5G systems should target the small cell networks with λ > λ1 as ultradense networks where λ1 appears to be around several 103 BSs/km2.

This definition is interchangeable with the one proposed by the industrial commu-nity [19, 20], in which the interpretation depends on inter-site distance:

• UDNs are characterized as networks with very short inter-site distances rang-ing from a few meters in indoor deployments up to roughly 50 m in outdoordeployments (every lamp post).

As densification goes further, some cells may have no UE inside and become idlewhich is illustrated in Fig. 1.3. When the BS density is sufficiently large, there willbe a body of void cells in the network. The third definition of UDN comes fromsuch scenario, which describes the relative density of BS, as in [21]:

• A cellular network with BS density λb and active UE density λu is called aUDN where ρ , λb

λu 1. The notation f g implies f is sufficiently large so

that O(g/f) is approximated by 0.

In general, the third definition should also be based on high BS density (shortinter-site distance) regime since we can barely call a network with 10 BSs and 1 UEper km2 a UDN, though the density ratio is high enough. Therefore in this thesis,we do not stick on a single UDN definition, but flexibly apply the aforementionedconcepts depending on different situations.

The differences between (ultra) dense networks and traditional networks is high-lighted in Table 1.2, which is revised from [22]. By the comparison, it should beclear that the definition of UDN and traditional cellular network differ significantlyin both engineering and business aspects. Thus, it is not likely that the successfulschemes for the traditional cellular paradigm will be successful on the UDN.

8 Introduction

Table 1.2: Comparison of traditional cellular and high-density system characteristics.

Traditional cellular Ultra-dense network

Characteristic Wide-area paradigm High-density, short range

Propagation Distance loss, shadowing, rich multi-path

Mostly LOS, body shadowing

Interference Interference sum of many compo-nents

Extremely varying interference

Engineering limita-tions

Range, interference, energy Interference

Peak rate limitationset by

Noise and interference Equipment (very high SNR)

Cost limitations Sites: acquisition, antennas, equip-ment, deployment, backhaul, spec-trum licenses

Backhaul, deployment

Active UEs per BS 1-100 0.01-1

Available bandwidth <0.5 GHz, licensed >5 GHz, secondary sharing

Design paradigm Industrial grade, centralized control,"mandatory complexity"

Consumer grade, distributed control,plug-and-play

Maintenance model Single point of failure; 24/7 monitor-ing

Graceful degradation; replace whentime available

1.3 Literature Review

Based on the given definitions, lots of existing models and operations applied tocurrent macro cellular networks or even small cells may not hold in UDN scenariosbecause the transmission distance is short and not all the BSs are transmittingsimultaneously. In this section, we give an overview of related prior work modelingand evaluating the UDNs.

Modeling and Analysis of UDNIn order to understand the impact of network densification, quantitative as wellas scalable research is indispensable. A glance at the approaches in previous worksis illustrated in Fig. 1.4, in which the assumptions of propagation environmentand network load condition are the key distinctions. Starting from step (a), [23]provides a comprehensive understanding on the impact of BS density in fully loadeddownlink cellular networks with a simple single-slope path loss model. It concludesthat BS densification does not change the coverage probability of individual UE(in interference-limited networks), but linearly improves the area spectral efficiency(ASE) defined as the sum average spectral efficiency of all active BSs achievingthe target threshold in a unit area sum rate. However, this conclusion may becomeinvalid in a UDN due to its over simplified signal propagation and load models.

To examine the impact of ultra densification, it is important to incorporate thepropagation characteristics of UDN properly. It is well known that the path loss

1.3. Literature Review 9

Unbounded

Single-Slope

Full Load

Unbounded

Single-Slope

Partial Load

Unbounded

Multi-Slope

Full Load

Bounded

Multi-Slope

Partial Load

(a)

(b)

(d)

(e)

Bounded

Single-Slope

Full Load

(c)

Figure 1.4: Approaches in propagation and load modeling.

exponent is small (or large) when radio signals experience less (or more) absorp-tion and diffraction losses. In a UDN environment, the probability for the desiredand interfering signals of experiencing different loss levels is quite high, resulting indissimilar path loss exponents. Furthermore, the probability of a link within an an-alytical reference distance, e.g. one meter, becomes high, and thus this phenomenoncannot be neglected in the analysis. Consequently, a widely accepted unboundedsingle-slope path loss model, i.e., G(d) = d−α, becomes dubious in a UDN. Withthis regard, research in bounded as well as multi-slope path loss models becomesnecessary in such scenarios. Moreover, a non-negligible portion of BSs become in-active without interfering other cells in a UDN (Fig. 1.3(c)). In this case, a full loadmodel where BSs always transmit signals will overestimate the interference level.

A multi-slope path loss model is applied in [24] (step (d)), which shows that theUE coverage probability is a decreasing function of BS density under such model-ing. Meanwhile, [25] investigates the effect of bounded path loss model when BSdensity is extremely large (step (c)). Both [24] and [25] show that the signal-to-interference-ratio (SIR) vanishes as BS density grows asymptotically. On the con-trary, the conclusions of [26] and [27] indicates that the SIR keeps increasing as BSdensity goes to infinity (step (b)). Such contradiction comes from the correspondingmodels in the references. The bounded and multi-slope path loss models help toprevent overestimating the signal power and underestimating the interference re-spectively, which decrease the SIR. Conversely, the partial load model can preventoverestimating the interference density in the whole network, which increases theSIR. Therefore, more consideration is required for the modeling of UDN.

10 Introduction

Operations of UDNThe BSs in UDN environments covers a small area and the inter-site distance wouldbe in the range of meters or tens of meters. This will create a more unpredictable andless manageable interference environment. On one hand, each cell is surrounded bymany neighbors in its close proximity. On the other hand, short-range transmissionshave different features than long-range transmissions in traditional sparse wirelessnetworks. For instance, the line-of-sight (LOS) component of signals appears anddominates, resulting in low signal power attenuates with distance. Therefore, drasticinterference will be generated from neighboring cells and become a limiting factor inthe system performance. Hence, interference management is unavoidable in a UDN.The interference mitigation methods including interference avoidance, cancellationand coordination schemes are investigated in [28–32].

Another major challenge is backhauling design because of the complex topologyin a UDN. Fiber-optic based backhauling is almost impossible for every small cell ina UDN due to cost prohibitive and deployment difficulty. An easy and cost efficientsolution is needed to bridge the pervasive small cells in a UDN and the fiber-basedcore network. Wireless backhauling are improving by leaps and emerging as a viablesolution. In wireless backhauling especially for high frequencies, LOS is a strict re-quirement which could be less challenging since the endpoints of the link are fixed.The solution is termed as in-band or self-backhauling if it reuses the radio accessfrequency. The advantages of such approach is increasing the resource utilizationand reducing CAPEX due to the reuse of hardware. However, it may create ex-tra interference between backhauling and access links. The existing backhaulingschemes for dense networks are discussed in [33–37] and the references therein.

Since the coverage area of an individual UDN cell is small, a UE may crossmore cells along its trajectory path and stays for a short time inside each cell.Handover executions thus occur more frequently in a UDN compared to conven-tional networks. In this case, the handover overhead increases significantly and theUEs need to allocate more resources for the handover instead of data transmissionwhich substantially degrades their quality of service (QoS). Consequently, mobilitymanagement in a UDN requires more intelligent handover schemes to overcomethese limitations. In [38], the authors study the handover process between a macroand a small cell in heterogeneous networks. A handover scheme utilizing coopera-tion based cell clustering is proposed in [39]. The authors in [40] investigates themobility management of UDNs from an energy-efficiency (EE) point of view. Ref-erence [41] and [42] introduced a handover skipping scheme in order to reduce thehandover rate.

1.4 Thesis Focus and Problem Formulation

Based on the literature review, we see that modeling and operations of a UDN arefundamentally different from those of conventional cellular networks. This thesisintends to focus on filling some research gaps in the aforementioned two aspects of

1.4. Thesis Focus and Problem Formulation 11

UDN. In what follows, the specific research questions addressed in this dissertationare defined.

Performance analysisAccording to the background, there is a general agreement on the effectiveness andpowerfulness of network densification. It has been recognized as the single most ef-fective way to increase network capacity in the past [43]. The glorious history owesto the traditional cellular scenario in which the network capacity scales linearlywith the BS density because power-law path loss models (single-slope models) holdand every BS serves at least one single UE [44]. According to the ‘theory’, keepdensifying endlessly, problem solved! Such ’myth’ has been dispelled by the newrecognition of further densification with more sophisticated models. However, theconflicted conclusions in the literature implied that the real value of ultra densifi-cation is still not crystal clear to us. This rises the first high-level research question(HQ) that we try to answer in this thesis:

• HQ 1: How much benefit can we get from ultra network densifica-tion?

Despite bringing the huge capacity increment in the past, densification is notthe only weapon when facing the 5G challenge, as always. The toolbox in Fig. 1.2provides different choices and flexibilities to the mobile network operators (MNO).Given the three elements therein, the MNOs need to make a strategic decision onwhich direction to head to, for the purpose of efficiently deploying a network andmeeting a certain traffic demand. Different MNOs with different existing networkassets, customer market segmentation, financial means and technical expertise mayrequire different solutions. However, in spite of abundant research in each of thethree ingredients, there has been little knowledge on the relationship among them.To this end, we will focus on the competence of network densification towards theother two elements (spectrum and multi-antennas) in answering HQ1. Thus, thefollowing research question (RQ) should be addressed:

• RQ 1: What is the trade-off between densification and other re-sources in order to fulfill certain UE requirement?

Recently, the well-known Moore’s Law has hit a plateau, resulting in wide rang-ing ramifications beyond the semiconductor industry. The analogous Cooper’s Lawmay face the similar problem sooner or later. As shown in previous works, theconclusion of linearly increased capacity are drawn under very simple load andpropagation assumptions. A simple single-slope path loss model cannot capture thedistinction of LOS and non line of sight (NLOS) paths. Furthermore, a full loadmodel forces BSs to always transmit signals despite non-negligible portion of UEvoid BSs, overestimating the interference. It is thus important to incorporate suchpropagation and load characteristics in detail so as to examine the impact of ultra

12 Introduction

0 50 100 1500

50

100

150

0 50 100 1500

50

100

150

near-field Active UE

Inactive UE

Active BS

Cooperating BS

Inactive BS

critical distance

bounded region radiusRb

Rc

(a) Traditional joint transmission0 50 100 150

0

50

100

150

0 50 100 1500

50

100

150

near-field Active UE

Inactive UE

Active BS

Cooperating BS

Inactive BS

critical distance Rc

bounded region radiusRb

(b) UDN joint transmission

Figure 1.5: An illustration of JT in traditional fully-loaded network and partially-loadedUDN.

densification. In this case, the ultimate gain of network densification is still unclear.Therefore, a clear understanding of the asymptotic behavior of ultra-densificationincorporating more realistic load and propagation models is needed. With this re-gard, we aim to answer the following research question:

• RQ 2: What is the asymptotic behavior of ultra-densification?

Operational guidelinesUDN is not a simple downscaling of traditional macro-networks. The densely de-ployment affects not only the system performance but also the operational strate-gies. Thus, operations and schemes fitting for macro-networks may not be applicablein a UDN. We have shown in the literature that most of the interference manage-ment, backhauling design and handover schemes are reconsidered or adapted forUDN regimes. This triggers new thoughts on the design of operation guidelines ofa UDN, which coincides with our second HQ in the thesis:

• HQ 2: How should certain operations be adjusted in a UDN sce-nario?

The existing UDN studies mainly evaluate the performance of UEs under singleBS association without any coordination. According to the third definition of UDN,there might be more BSs than UEs in a UDN scenario and the BSs without any UEin their coverage area are considered stay in idle mode, in order to save energy andreduce interference. These dormant BSs have the potential to provide extra capacityfor the active UEs through proper cooperative schemes. Thus, it is interesting toinvestigate cooperation schemes fit for UDN scenarios.

Joint transmission (JT) is a potential solution since it allows multiple BSs tojointly serve one user. In traditional fully loaded cellular networks, JT turns domi-

1.4. Thesis Focus and Problem Formulation 13

Figure 1.6: Existing cell search schemes

nant interferers into useful signals while the other interferers remain the same. Thus,the desired signal strength increases and the interference decreases simultaneously,at the cost of reduced scheduling probability. It is known that JT enhances theperformance of cell-edge users in macro cellular networks [5] (Fig 1.5(a)). However,interference nature is completely different in a UDN because turning on dormantBSs will generate extra interference and energy consumption. In this case, if UEsget assistance from nearby sleeping BSs, the interference will grow rapidly as wellas the desired signal power as shown in Fig. 1.5(b). Therefore, this thesis aims toanswer the following question:

• RQ 3: How to perform BS cooperations in UDN scenarios?

Millimeter wave (mmWave) networks rely on directional transmissions to over-come severe path loss in both control plane and data plane. Nevertheless, the use ofnarrow beams complicates the initial access procedure where we lack sufficient infor-mation for beamforming. Current initial access schemes, summarized in Fig. 1.6 [45]mainly focused on finding ’the best’ link quality. In [46], the authors analyzed var-ious design options given different scanning and signaling procedures. The syn-chronization signals and the detection approach is such that time-frequency-spatialsynchronization occurs jointly. An iterative beam searching scheme, namely a two-phase hierarchical procedure with multi-resolution codebook is discussed in [47].As the networks are deployed denser progressively, the potential data rate of di-rectional mmWave communications is rather high even when connecting with a BSother than the ’best’ choice. Therefore, finding ’an acceptable’ link especially in adense network may be more preferable, as it may substantially reduce the data-plane establishment latency while satisfying the rate demands of the UEs. Fastinitial access becomes even more important for wireless sensor networks comprisingmany wake-up radios in which the synchronization may take longer than the actualdata transmission, leading to a poor latency performance. Despite plenty of data

14 Introduction

plane analysis on mmWave dense networks, the impact of network densification oncontrol plane design has not been discussed, which prompts the following researchquestion:

• RQ 4: How does network densification affect the initial access inmmWave networks?

1.5 Overview of Thesis Contributions

This section summarizes the main contributions of the thesis. To put our work intothe context of current research, an overview of the thesis contribution is listed asfollows:

Resource tradeoffs in UDN provisioningIn this part, we first study the resource trade-offs in network provisioning. Weintroduce the microeconomics terminology ’technical rate of substitution’ (TRS)[48] to illustrate the value of densification in terms of spectrum and antenna numberper BS. The substitutability of densification is discussed by calculating the TRSbetween BS density and other elements in different scenarios. Furthermore, we alsodemonstrate the importance of specifying demand types in the dimensioning ofwireless networks. We show that different resources are needed in different trafficdemand types even with the same area throughput. This section is based on thefollowing papers:

• Paper I (Paper A in the appendix): Y. Yang and K. W. Sung, "Tradeoffbetween Spectrum and Densification for Achieving Target User Throughput,"2015 IEEE 81st Vehicular Technology Conference (VTC Spring), Glasgow,2015.

• Paper II (Paper B in the appendix): Y. Yang and K. W. Sung, "TechnicalRate of Substitution of Spectrum in Future Mobile Broadband Provisioning,"2015 IEEE International Symposium on Dynamic Spectrum Access Networks(DySPAN), Stockholm, 2015. (Best poster award).

• Paper III: A. Widaa, Y. Yang, K. W. Sung and J. Markendahl, "On theEngineering Value of Spectrum in Dense Mobile Network Deployment Sce-narios," 2015 IEEE International Symposium on Dynamic Spectrum AccessNetworks (DySPAN), Stockholm, 2015.

The author of this thesis acted as the main author of Paper I and Paper IIand developed the simulation methodologies. The problem formulations are jointlyproposed with Dr. Ki Won Sung. The author of the thesis acknowledges Dr. AndersFuruskär at Ericsson Research for his valuable suggestions and feedback. Paper IIIwas written in collaboration with Ashraf Widaa. The author of this thesis hascontributed to the analytical modeling and proofreading.

1.5. Overview of Thesis Contributions 15

Asymptotic behavior of UDNOn the next step, the asymptotic behavior of ultra-densification is studied. We com-bine step (b), (c) and (d) altogether, resulting in step (e) in Fig. 1.4, and investigatetheir aggregate impact on the asymptotic SIR coverage behavior. To the best ofour knowledge, this is the first UDN work combining both the effect of UE den-sity and a bounded multi-slope path loss model. We discuss the convergence of theasymptotic behaviors depending on different load situations. Numerically tractableintegral-expression as well as closed-form bounds of coverage probability and ASEare derived. Furthermore, the impact of UE density on coverage probability andASE are analyzed. This section is based on the following paper:

• Paper IV (Paper C in the appendix): Y. Yang, J. Park. and K. W.Sung, "On the asymptotic behavior of ultra-densification under a boundeddual-slope path loss model," European Wireless 2017; 23th European WirelessConference, Dresden, Germany, 2017.

The author of the thesis proposed the original problem, provided the derivationsand acted as the leading author of the paper. Dr. Jihong Park made contributionsby refining the structure and polishing the writing of the paper. Dr. Ki Won Sungprovided directions and valuable insights in the paper.

BS cooperation in UDNsThis section intends to provide insights of BS cooperation in two UDN scenarios:homogeneous UDN where all the BSs are the same type and heterogeneous UDNwhere different kinds of BSs are deployed. For the homogeneous UDN scenario, weinvestigated two BS cooperation schemes: non-coherent JT without the assistanceof instantaneous channel state information (CSI) and coherent JT with full CSIknowledge. For the heterogeneous UDN scenario, we study coordination betweenCloud-RAN (C-RAN) based BSs and independent operating BSs. In both scenarios,we studied the effect of imperfect CSI on cooperation and categorized the conditionswhich are favorable for cooperation. The findings are presented in the followingpapers:

• Paper V (Paper D in the appendix): Y. Yang, K. W. Sung, J. Park S. L.Kim and K. Kim, "Cooperative transmissions in ultra dense networks under abounded dual-slope path loss model," 2017 European Conference on Networksand Communications (EuCNC), Oulu, 2017.

• Paper VI (Paper E in the appendix): D. Kim, Y. Yang, K. W. Sung,and J. Kang, "Cooperation Strategies for Partly Wireless C-RAN," IEEECommunications Letters, vol. 22, no. 6, pp. 1248-1251, June 2018.

The author of this thesis proposed the original problem formulation, imple-mented the simulation methodologies and acted as the leading author of Paper V.

16 Introduction

Dr. Jihong Park and Dr. Ki Won Sung offered valuable feedbacks on writing andthe structure of the paper. Prof. Seong-Lyun Kim and Prof. Kwang Soon Kim pro-vided comments on the problem formulation. In Paper VI, the author of this thesisrefined the problem, developed simulation framework and provided proofreading.

Initial access design in mmWave UDNWe then turn our attention to the control plane design of mmWave UDN. In thispart, we design a low-complexity initial access scheme under random beamformingin mmWave UDNs. A stochastic geometry based analytical framework for evalua-tion is developed. We investigated the effect of BS density, environmental blockageand antenna beamwidth on initial access performance. Meanwhile, we character-ize the tradeoff between failure probability and expected latency. Furthermore, wecompared our proposed scheme with the existing techniques in both control planeand data plane performance. This section is based on the following papers:

• Paper VII (Paper F in the appendix): Y. Yang, H. S. Ghadikolaei, C. Fis-chione, M. Petrova and K. W. Sung, "Reducing Initial Cell-search Latency inmmWave Networks," IEEE INFOCOM 2018 - IEEE Conference on ComputerCommunications Workshops (INFOCOM WKSHPS), Honolulu, HI, 2018.

• Paper VIII: H.S. Ghadikolaei, Y. Yang, C. Fischione, M. Petrova, and K.W. Sung. "Fast and Reliable Initial Cell-search for mmWave Networks", InProceedings of the 2nd ACM Workshop on Millimeter-Wave Networks andSensing Systems 2018 (mmNets’18).

• Paper IX (Paper G in the appendix): Y. Yang, H.S. Ghadikolaei, C.Fischione, M. Petrova, K. W. Sung, "Fast and Reliable Initial Access withRandom Beamforming for mmWave Networks", submitted to IEEE Transac-tions on Wireless Communications.

The problems in this section were jointly proposed and formulated with Dr.Hossein Shokri. The author of this thesis provided mathematical derivations, nu-merical results and acted as the leading author of Paper VII and Paper IX. PaperVIII was drafted by Dr. Hossein Shokri while the author of this thesis contributed tothe numerical results and was involved in editing the paper. Prof. Carlo Fischione,Prof. Marina Petrova and Dr. Ki Won Sung provided comments on the problemformulation and insights in the direction of the papers.

1.6 Thesis Outline

The remainder of the thesis is organized in two parts: an overview of the thesis(Chapter 1 to 4) and verbatim copies of the papers includes in this thesis. In PartI, Chapter 2 introduces the modeling and methodology employed in this thesis. The

1.6. Thesis Outline 17

key results of the thesis are present in Chapter 3. Finally, the concluding remarksand future work are discussed in Chapter 4.

Chapter 2

Research Methodology and Models

"Essentially, all models are wrong, but some are useful."George Box

This chapter describes the research methodologies and system models we usedthroughout the thesis. We first introduce the stochastic geometry based analysisand Monte Carlo simulation as our main methodology. Then, further details ontheoretical models in this thesis are present.

2.1 Research Methodology

2.1.1 Stochastic Geometry based Network AnalysisA wireless communication network can be viewed as a collection of points, whichcan in turn represent transmitters or receivers. Accurate models are essential to thedevelopment of standards and the evaluation of system performance or potentialschemes in wireless networks. The most widely used and accepted model is thehexagonal grid model. It is frequently used in system-level simulations but is notgenerally possible for analytical study [49] [50]. A popular analytical model formulti-cell systems is the Wyner model [51] which assumes a unit gain from eachBS to its serving UE and equal channel gains to the two neighboring UEs in aone-dimensional space. Such a model is reasonable in a system where other-cellinterference is approximated to a constant, such as the code division multiple access(CDMA) uplink, but it is highly inaccurate in general and particularly for systemswith 1 or 2 dominant interferers, like an orthogonal frequency-division multiplexing(OFDM) based 4G and beyond network [52].

Recently, a prominent approach has drawn attentions in academia, which usesrandom spatial models from stochastic geometry to capture the real deployment.Stochastic geometry is the theory of random sets, which forms random processesof points, particles, fibres or surfaces in the Euclidean space [53]. Specifically, itmodels BS and UE locations as realizations of spatial point processes on the two

19

20 Research Methodology and Models

0 0.2 0.4 0.6 0.8 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

BS

User

Inactive BS

Active BS

Figure 2.1: Network topology based on the PPP assumption

dimensional plane and enables the characterization of signal-to-interference-plus-noise-ratio (SINR) distribution and mean data rate. By this means, the theoreticalexpressions can reveal the impact of system parameters on the key performancemetrics, thereby providing insights on system design and optimization. Therefore,stochastic geometry is able to provide good tractablility for the analysis of homo-geneous, heterogeneous and cognitive multi-tier cellular networks [54].

In stochastic geometry analysis, the network is abstracted to a point process(PP) to model the positions of the network entities according to the network typeand/or the MAC layer behavior. The popular PPs include Poisson point process(PPP), hard core point process (HCPP), Poisson cluster process (PCP) and etc.Among these PPs, the PPP is the most popular because the assumption of indepen-dent BS locations allows substantial tools from stochastic geometry and providestractable analysis. The accuracy of the PPP model is validated by comparing withthe grid model and an actual 4G network [23].

Stochastic geometry analysis provides a good complement to exhaustive simu-lations, since a large subset of undesirable deployment parameters can be quicklyeliminated through the analytical expressions. The major drawback of stochasticgeometry modeling is the limitation of spatial averaging. Considering only spatialaverages may hide important details that affect the network performance. It willlimit the insights that can be obtained and the effect of the design parameters dueto the spatial randomness [55]. For instance, it is hard to capture the performanceof cell edge or lower 5 percentile UEs. Accordingly, we cannot design the networkwhere at least 95% of the UEs experience an outage probability less than 5% only

2.2. System Models 21

with the spatial average of the outage probability. Thus, a combination of usingdifferent tools is necessary for different purposes of analysis.

2.1.2 Monte Carlo SimulationIn Monte Carlo simulation, the entire system is simulated a large number (e.g.,10000) of times. Each simulation is equally likely, referred to as a realization ofthe system. For each realization, all of the uncertain parameters are sampled (i.e.,a single random value is selected from the specified distribution describing eachparameter). The system is then simulated through time (given the particular set ofinput parameters) such that the performance of the system can be computed. Thesimulation provides a large number of separate and independent results, each rep-resenting a possible "future" for the system (i.e., one possible path the system mayfollow through time). The results of the independent system realizations are assem-bled into probability distributions of possible outcomes. As a result, the outputsare not single values, but probability distributions.

The main advantage of the comprehensive system level simulations is that itcan investigate any arbitrary network scenario to any detail. This is enabled byimplementing the desired network topology, wireless channel model, traffic model,network protocols, and so on. Therefore, many key performance metrics of cellularnetworks, including the SINR distribution and the data rate, can be accuratelydetermined through simulations. However, a major limitation of simulation is thateach desired scenario needs to be simulated separately using distinctive systemparameters. As the cellular network continues to evolve with more complicatedfeatures and more system parameters to optimize, simulating all the scenarios ofinterest can be very time-consuming. Due to the limited number of scenarios thatcan be simulated, the corresponding design insights can also be quite restricted,and hard to infer for the scenarios that are not simulated. Hence, for your systemto be sound, it is very necessary for a harmony between analysis and simulations.

2.2 System Models

2.2.1 Network ModelWe consider a downlink cellular network in all of our works where the BSs and UEsare modeled to be distributed according to two independent homogeneous PPPs,Φb and Φu. The densities of BS and UE are denoted as λb and λu respectively. Thecoverage area of each BS comprises a Voronoi tessellation as shown in Fig. 2.1. Inmost cases (except Chapter 3.4), we assume closest association rule, i.e., the UEsin the Voronoi cell of a BS are served by it. As the BS density increases, some BSsmay not contain any active UEs in its coverage area. In this case, the BS becomesinactive without transmitting any signal. Each active BS has at least one UE in itscell and will randomly choose one of them to serve. The probability of a BS beingactive is derived in [56], given as:


Table 2.1: Road sign of models applied in the thesis.

Model Applied in Chapter

PPP based network Chapter 3.1, 3.2, 3.3, 3.4

Single-slope path loss model Chapter 3.1

Bounded dual-slope path loss model Chapter 3.2, 3.3

Blockage based path loss model Chapter 3.4

MRT beamforming model Chapter 3.1, 3.3

Analog beamforming model Chapter 3.4

3GPP NR model Chapter 3.4

pa = 1− (1 + λu3.5λb

)−3.5. (2.1)

Without loss of generality, we conduct analysis on the performance of a typicalUE located at the origin o and randomly selected by its serving BS, which is per-missible in a homogeneous PPP according to Slivnyak’s theorem [57]. The SINR ofthe typical UE served by BS xi is given by:

SINRxi = PtGBSGUE|hxi |2`(|xi|)Sxi∑xj∈Φb\xi

PtGBSGUE|hxj |2`(|xj |)Sxj +W,

where Pt and W denote the BS transmit power and the thermal noise; GBS andGUE represent the antenna gain of the BS and UE which are set to 1 for omni-directional antenna; h and |x| denote the channel and distance between BS andUE respectively; `(|x|) is the distance path loss which will be introduced in thenext section; Si is a binary term indicating LOS condition (used in Chapter 3.4).When there are many idle BSs in the network, Φb , Φ∗b represents the set of activeBSs which is not a homogeneous PPP because the active probability depends onthe cell area. Nevertheless, we can model Φ∗b as a homogeneous PPP with densityλ∗b = λbpa, by thinning the process Φb. This approximation has been shown to beaccurate according to [58, 59].

As we are dealing with UDN scenarios where BS-UE distance is quite close inthe thesis, it is reasonable to assume an interference-limited system (except forChapter 3.4). Thus, we will neglect the noise power and examine the SIR insteadof SINR throughout the thesis. Such assumption is justified through simulationin [21, 58]. The SIR of the typical UE with omni-directional antenna and LOScondition can be expressed as:


SIRxi = |hxi |`(|xi|)∑xj∈Φb\xi

|hxj |`(|xj |). (2.2)

Given the downlink SINR of the typical UE and the target SINR level T, thecoverage probability is defined as:

Pc(λb, λu, T ) , P[SINRxi > T ]. (2.3)

The network ASE Γ is given by

Γ(λb, λu, T ) , paλbPc(λb, λu, T ) log2(1 + T ) (2.4)

where paλbPc can be interpreted as the density of the BSs that successfully transmitthe symbols to their UEs.

2.2.2 Channel ModelIn wireless communications, the channel refers to the medium between the trans-mitter and receiver. Due to the electromagnetic wave propagation mechanisms,the wireless channel modeling usually includes path loss, large-scale fading andsmall-scale fading. For large-scale fading, it has been shown in [60] that the PPPassumption can be viewed by combining BS locations and shadowing with suffi-ciently large variance. Meanwhile, it will not significantly affect the accuracy of theanalysis [23, 58] without shadow fading. Therefore, we ignore the effect of shadow-ing in this thesis similar as [40, 61]. Rayleigh fading is used as small-scale fading,with the fading coefficients h are i.i.d zero mean unit variance complex normal dis-tributed random variables. Compared to more realistic models for LOS paths suchas Nakagami fading, Rayleigh fading provides very similar design insights whileleads to more tractable results [62].

The complexity of radio propagation makes it difficult to obtain a single modelthat accurately characterizes path loss across a range of different environments.Accurate path loss models can be obtained from complex ray tracing models orempirical measurements. However, for general system analysis, it is better to applya model that is simple yet able to captures the essence of the signal propagation.Thus, the following models for path loss as a function of distance is commonly usedfor system design.

Single-slope Path Loss Model

One of the most widely used path loss models is the distance-dependent standardpath loss model which is expressed below:

`(α, d) = d−α, (2.5)


0 50 100 1500

50

100

1500 50 100 150

0

50

100

150

BS active probability

3. Partial load model

pa

Ro bounded region radius

2. Bounded path loss model

Rc near-field region radius

↵c near-field path loss exponent1. Dual-slope path loss model

near-field

Baseline. Unbounded single-slope path loss model + Full load model

Active UE

Active BSInactive BS

Figure 2.2: A bounded dual-slope path loss model

where α is the path loss exponent that can be roughly fit to the environment.The standard path loss model is the basis for most existing cellular network

theory, analysis, simulation and design. We also used it in our initial studies inChapter 3.1. Nevertheless, as discussed in the Introduction, this simple model isalso known to lead to unrealistic results in some special cases [63] and cannotaccurately capture the dependence of the path loss exponent on the link distancein UDN scenarios. Thus we turn to more sophisticated path loss models in thefollowing sections.

Bounded dual-slope Path Loss Model

In a UDN environment where cell sizes are much smaller, the single-slope standardpath loss model becomes too idealized. Therefore, it is important to capture thepropagation characteristics of a UDN for providing more accurate results. With thisregard, a bounded dual-slope path loss model is applied in our UDN analysis.

We consider that path loss attenuation from a BS to a UE is distinguishedunder three different regions, bounded region, near-field and far-field at the UEas illustrated in Fig. 2.2. Bounded region is a closed circle centered at the UE,


inside which the path loss is assumed to be constant (e.g. no loss within 1 meter).The radius of the circle is always no smaller than 1 so that the transmitted signalpower cannot be amplified at the receiver side. The near-field and far-field are theareas outside the bounded region, separated from a critical distance to the UE. Itworth noting that the definition of near-field and far-field here is different from theelectromagnetic concept. In this thesis, our definition of the two fields describes thedistance differences from the transmitters to the UE, as in [24]. Near-field denotesthe region closer to the receive antenna, wherein the signal has a higher probabilityto transmit in a LOS path. Hence, the signal power decays slowly within this region.In the far-field, the signal experiences more absorption and diffraction which makeslarger path loss exponents than in the near-field regions.

The bounded dual-slope path loss model can be formulated in a piece-wisefunction of the transmission distance d, shown as

`(α1, α2, d) =

1, 0 ≤ d ≤ Rb;τ1d−αc , Rb < d ≤ Rc;

τ2d−α, d > Rc

(2.6)

where Rb > 0 is the radius of the bounded path loss region, i.e., the path loss inthe range of [0, Rb] is assumed constant; τ1 , Rb

αc and τ2 , Rcα−αc which keeps

the continuity of the function (we set Rb = 1 in our work thus τ1 = 1); Rc ≥ Rb isthe critical distance to divide the near- and far-field; and αc and α are the near-and far-field path loss exponents for 2 ≤ αc ≤ α, respectively.

Blockage-based Path Loss Model

A distinguishing feature of mmWave cellular communication is that mmWave sig-nals are more sensitive to blockage effects than signals in lower frequency bands.Channel measurements using directional antennas have revealed that blockagescause substantial differences between LOS and NLOS paths [64]. Thus in mmWavenetwork analysis, each BS is characterized by either LOS or NLOS propagation tothe typical UE, independently from others. Define the LOS probability functionpLOS(r) as the probability that a link of distance r is in LOS condition. We applya stochastic exponential blockage model for the function where the obstacles aremodeled with a rectangle Boolean scheme [65]: pLOS(r) = exp(−βr), where β is aparameter determined by the density and the average size of the blockages, andβ−1 represents the average length of a LOS link. For the tractability of analysis, wefurther assume independent LOS probabilities between different links which onlycause a minor loss of accuracy in the SINR evaluation [65, 66].

Given LOS probability pLOS(r), the path loss for a link with distance r is givenby:

`(‖xi‖) =

(C(‖xi‖))αL , if LOS;(C(‖xi‖))αN , if NLOS (blocked)

(2.7)


where C(‖xi‖) = c4π‖xi‖fc , c is the light speed, fc is the operating frequency, αL and

αN represent the path loss exponent for LOS and NLOS links, and ‖xi‖ representsthe distance between xi and the origin. When ignoring NLOS effect, we set αN =∞and αL = α for simplicity.

2.2.3 Maximal Ratio Transmission Beamforming Model

Maximal Ratio Transmission (MRT) is one technique of linear precoding whichmaximizes the signal gain in the absence of interference, or when interference istreated as background noise [67]. We select MRT as a representative of digitalbeamforming techniques in our multi-antenna and cooperation analysis. Other tech-niques are also valid and left as future work. We assume a multi-input-single-output(MISO) system where each BS hasM antennas and each UE is equipped with a sin-gle antenna. With MRT beamforming, the received signal for a typical user servedby BS located at xi given by

yi = |xi|−α2 h∗iwi

√Ptsi +

∑j 6=i|xj |−

α2 h∗jwj

√Ptsj +W (2.8)

where hi ∼ CN (0, I) is a M × 1 vector denoting the MISO channel, the MRTprecoder is designed as wi = h∗i

‖hi‖ . Thus, the SINR of the typical UE can beexpressed as

SINRxi = gi`(|xi|)∑xj∈Φb\xi

gj`(|xj |)(2.9)

where gi =‖ hi ‖2∼ Gamma(M, 1) and gj = ‖h∗i hj‖2

‖hj‖2 ∼ Exp(1) for j 6= i [68].

2.2.4 Analog Beamforming Model

To analyze the initial cell-search in mmWave network, we consider analog beam-forming because digital or hybrid beamforming does not suit due to the existenceof many antenna elements and lack of prior channel knowledge, translated intothe need for costly pilot transmission schemes. We first model the actual antennapatterns by a sectorized beam pattern (SBP) as in [69]. The advantage of such mod-eling is the analytical simplicity, through which we can derive tractable expressionsof detection failure probability and delay. As a complementary for SBP, we alsoconsider the uniform linear array (ULA) antenna model for the due to two reasons:1) verifying the analytical insights and performance trends, obtained by the SBPmodel, and 2) obtaining the SINR in data plane with beam refinement.


Sectorized Beam Pattern

In the SBP model, BSs and UEs are equipped with an antenna array of NBS andNUE antennas respectively to support directional communications with the corre-sponding antenna gains GBS and GUE, where NBSNUE ∈ N+. Thus, the antennabeamwidth of BSs and UEs are θBS = 2π/NBS and θUE = 2π/NUE. We assumeboth BSs and UEs have 1 RF chain, such that they can transmit to or receive fromonly one direction at a time. In an ideal sectorized antenna pattern, the antennagain Gx, x ∈ BS,UE, as a function of beamwidth θx is a constant in the mainlobe and a smaller constant in the side lobe, given by

Gx(θx) =

2π−(2π−θx)εθx

, in the main lobe,ε, in the side lobe,

(2.10)

where typically ε 1. In some cases for analytical tractability, we ignore thesidelobe gain by setting ε = 0. Similar to [46, 62], we assume BSs and UEs sweepthe entire angular space by NBS and NUE beamforming vectors defined in thecodebooks, respectively.

Uniform Linear Array

In the ULA model, the array response vector can be expressed as

a (K, θ) = 1√k

[1 ejπ sin(θ) . . . ej(k−1)π sin(θ)

]H. (2.11)

The parameters of the channel model depend both on the carrier frequency and onbeing in LOS or NLOS conditions and are given in [70, Table I]. More specifically,we define f(k,K, θ) as

f(k,K, θ) :=

[1 ejπ sin(θ) . . . ej(k−1)π sin(θ) 01×(K−k)

]H

√k

, (2.12)

for integers k and K such that 0 < k ≤ K. 0x is an all zero vector of size x. Letvcb and wc

u be the precoding vector of BS b ∈ B and the combining vector of UEu ∈ U in mini-slot c of the cell-search phase. We define

vcb = f(kcb , NBS, φcb) , (2.13a)

wcu = f(kcu, NUE, θ

cu) . (2.13b)

The BSs and UEs can control the antenna boresight by changing φcb and θcu andcontrol the antenna beamwidth by changing kcb and kcu. At each BS b, we keep alocal codebook Vcb that contains MBS precoding vectors. Each vector vcb ∈ Vcb is ofthe form (2.13a) such that the codebook collectively spans all angular space. Thecardinality of the codebook is based on the half-power beamwidth and determinesthe antenna gain and sidelobe interference caused by every beam.


2.2.5 3GPP NR Frame ModelTo verify our proposed initial-cell search scheme in a system analysis, we apply thelatest 3GPP NR framework. The 3GPP technical specification for NR introducesthe concept of synchronization signal (SS) block and burst. An SS block spansfour orthogonal frequency division multiplexing (OFDM) symbols in time and 240subcarriers in frequency, see [71] and references therein. Each SS block is mappedto a certain angular direction. An SS burst is a group of several SS blocks, and theinterval between consecutive SS bursts TSS can take from 5, 10, 20, 40, 80, 160 ms.Higher values correspond to lower synchronization overhead. Within one frame ofNR, there could be several pilots, called channel-state information reference signal(CSI-RS), to enable optimal beamforming design for the data-plane.

Chapter 3

Key Results

"Be a voice, not an echo."Albert Einstein

In this chapter, we present the key results 1 of the thesis to answering theresearch questions in the Introduction. The results were obtained by employing themethodologies and models described in Chapter 2.

3.1 Resource Trade-offs in UDN Provisioning

In order to quantify the resource requirement in UDN in terms of BS density,antenna numbers and spectrum, we need to figure out the relationships amongthem. Besides, resource requirement depends on the traffic demand. Area capacityis widely used for describing demand target in previous works. However, it hardlycaptures the individual UE requirement which is a key performance indicator in 5Gor beyond [72]. Same area capacity may result in various types of traffic demandthat need different amount of resources. Therefore specifying demand types canhelp operators allocating resources properly. In this section, we study the relationsamong the three elements under different types of traffic demand. The details ofthe analysis and more results can be found in Papers I, II and III.

3.1.1 Technical Rate of Substitution

The concept from microeconomics, TRS, is introduced to explain the substitutabil-ity of each element. The definition of TRS is given as:

• TRS measures the rate at which the firm will have to substitute one input foranother in order to keep output constant.

1Parts of the results in this chapter has been presented in the Licentiate thesis of the author.The material is reused with permission.

29

30 Key Results

Antenna Number per BS

BS

de

nsity in

lo

grith

m s

ca

le (

log

10λ

b)

2 4 6 8 10 12 14 16 18 200

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2Double Rate

Basic Service

Double Usage

(a) Sparse network

Antenna Number per BS

BS

de

nsity in

lo

grith

m s

ca

le (

log

10λ

b)

2 4 6 8 10 12 14 16 18 202

2.2

2.4

2.6

2.8

3

3.2

3.4

3.6

3.8

4Double Rate

Basic Service

Double Usage

(b) Dense network

Figure 3.1: Relation between antenna number and BS density.

Table 3.1: TRS between antenna number and densification (4M4λb

).

Sparse Case M=1 M=4 M=8 M=16λu=100 0.374 2.664 5.739 23.244

Ra=15MbpsDense Case M=1 M=4 M=8 M=16λu=100 0.0026 0.022 0.075 0.154

Ra=420Mbps

Suppose that we have factors, e.g. spectrum and BS density, and are operating atsome point (W,λb) with output of data rate Ra. Consider a change in W and λbwhile Ra keeps fixed:

Ra = f(λb,W ) = f(λb +4λb,W +4W ). (3.1)

Then the TRS of λb with regard to W at point (λb, W) is defined as

TRS(λb,W ) =∣∣∣∣4W4λb

∣∣∣∣ (3.2)

which indicates the substitutability of BS density. If the TRS of λb is low, it meansBS can be easily replaced by acquiring more spectrum. Otherwise, deploying moreBSs is more effective.

3.1.2 TRS of DensificationWe will evaluate the relations among BS density λb, spectrum bandwidthW and an-tenna number per BSM . Our objective is to calculate TRS(λb,W ) and TRS(λb,M)at each operating point of basic service. Beyond that, we consider increasing thenetwork capacity from either speed or usage perspective and investigate the resourcerequirement for the two network upgrading purposes.

3.1. Resource Trade-offs in UDN Provisioning 31

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20

200

400

600

800

1000

1200

1400

BS density in logrithm (log10

λb)

Spectr

um

[M

Hz]

Densification−Spectrum Indifference Curve (Sparse)

Basic Service: Ra=15Mbps λ

u=100

Double Usage: Ra=15Mbps λ

u=200

Double Rate: Ra=30Mbps λ

u=100

(a) sparse

2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 40

50

100

150

200

250

300

350

400

450

500

BS density in logrithm (log10

λb)

Spectr

um

[M

Hz]

Densification−Spectrum Indifference Curve (Dense)

Basic Service: Ra=420Mbps λ

u=100

Double Usage: Ra=420Mbps λ

u=200

Double Rate: Ra=840Mbps λ

u=100

(b) dense

Figure 3.2: Relation between spectrum and BS density.

Table 3.2: TRS between spectrum and densification (4MHz4λb

).

Sparse Case λb=1 λb=5 λb=10λu=100 349.650 23.256 6.289

Ra=15MbpsDense Case λb=100 λb=500 λb=1000λu=100 2.398 0.080 0.030

Ra=420Mbps

MRT beamforming is chosen for multi-antenna technique and the standard dis-tance dependent path loss model is employed with exponent α = 4. BSs and UEsfollow two independent PPPs and the BSs are in idle mode if no UE in its servingarea. All the models are described in Chapter 2.2.1, 2.2.2 and 2.2.3.

We choose the typical UE data rate Ra as our performance metric, which isdefined as the average downlink ergodic rate of a typical UE in the network. Racan be expressed as a function of coverage probability Pc and is given by:

Ra = WE[log2(1 + SIR)]

= W

ln2

∫T>0

PcT + 1dT.

(3.3)

Given the SIR of the typical UE, the coverage probability is derived in [58] as:

Pc = 1pa

∥∥∥∥∥[(k0 + 1

pa

)I−QM

]−1∥∥∥∥∥

1

(3.4)

where ‖ · ‖1 is the L1 induced matrix norm, I is an M ×M identity matrix, QM is

32 Key Results

an M ×M Toeplitz matrix given by

QM =

0k1 0k2 k1 0...

. . .kM−1 kM−2 · · · k1 0

(3.5)

k0 =2αT

1− 2α

2F1(1, 1 − 2α ; 2 − 2

α ,−T ), and ki =2αT

i

i− 2α

2F1(i + 1, i − 2α ; i + 1 − 2

α ,−T )for i ≥ 1, where 2F1(·) is the Gauss hypergeometric function.

We consider two scenarios in this section: sparse network representing the cur-rent deployment and dense network for the future. Sparse and dense are interpretedby the relative BS-UE density ratio ρ = λb

λu. We choose ρ = 0.1 and ρ = 10 for

sparse and dense scenario respectively while λu is fixed for both scenarios. The re-source requirement for two purposes are investigated: increasing the individual UEdata rate or accommodating more active UEs, both of which constitute the samearea capacity increase.

Figure 3.1 depicts the relationship between BS density and antenna number. Theindifference curves represent resource combinations that can fulfill the same trafficdemand. For basic service, adding antenna number has a diminishing gain in termsof BS density which is also reflected in the growing TRS in Table 3.1. Meanwhile,multi-antenna performs better in dense networks according to the small TRS values.

The spectrum-densification trade-offs are illustrated in Fig. 3.2 and Table 3.2.In sparse region, the slope of the curve is sharp which indicates densification cansave large amount of spectrum. When it comes to dense network, only a portion ofBSs are active since BSs outnumber UEs. Therefore, a further densification will addeven more idle BSs which is not effective, though slightly helpful, for the throughputincrease. Similar with installing antennas, investing spectrum is more effective indense network as well (with smaller TRS).

In Figs. 3.1 and 3.2, we also present the resource requirement for 2 times morearea capacity but with different traffic types. In sparse network, the coincidentindifference curves demonstrate that doubling data rate and doubling usage re-quires same amount of resource. Meanwhile, BS density and spectrum are linearlyexchangeable in this region, i.e., either 2x spectrum or 2x densification can accom-plish 2x capacity requirement. Nevertheless, in dense network, twofold densificationcan still achieve double usage but is far apart from meeting double data rate re-quirement. This phenomenon is further explained in Fig. 3.3 in which five traffictypes with the same area throughput are compared. It is shown that densificationplays a more important role in accommodating massive UEs while spectrum has asignificant impact in the bursty situation.

3.2. Asymptotic Behavior of Ultra-Densification 33

25 100 4000

100

200

300

400

500

600

700

800

900

1000

BS number

Required S

pectr

um

(M

Hz)

# User=1000, Ω=10Mbps





Figure 3.3: Resource requirement in different traffic demands.

Summary

By studying the TRS, we measure the relative effectiveness of network densificationover acquiring spectrum or installing multiple antennas on BSs. The high TRSindicates that network densification is the key solution to capacity increment insparse network. Spectrum and multi-antenna2 can have additional merits especiallyin dense networks. We also demonstrate the importance of specifying user demand innetwork dimensioning by showing that the resource requirement in different trafficdemand types with the same area throughput. It is observed that increasing datarate of individual users requires more resources than accommodating more usage.

3.2 Asymptotic Behavior of Ultra-Densification

In the previous section, we have observed a diminishing gain of densification as BSdensity keeps increasing. However, the end point of densification is not foreseeablefrom there. In this section, we further extend our work to the asymptotic behaviorof ultra-densification. The details of the analysis and more results can be found inPaper IV.

3.2.1 Coverage and Area Spectral Efficiency Convergence inUDNs

By incorporating the bounded dual-slope propagation (Chapter 2.2.2) and partialload, the exact form of coverage probability along with its upper and lower boundsis derived in Proposition ?? and ?? of Paper C. Based on the bounds, we further

2Only MRT beamforming is evaluated, zero forcing or other schemes may have different results.

34 Key Results

show the convergence of the coverage probability and ASE in asymptotic regions:As λb →∞, SIR coverage probability and ASE converge to finite values as follows:

limλb→∞

Pc(λb, λu, T ) = e−λuπcT (α,αc,Rc,Rb) (3.6)

limλb→∞

Γ(λb, λu, T ) = λue−λuπcT (α,αc,Rc,Rb) log2(1 + T ) (3.7)

where cT (α, αc, Rc, x) := Rc2F(

2α ,

1T

[Rcx

]α)−Rb2 [F ( 2α ,

1T

)− T

1+T

]+

2TRbαRc2−α

αc−2 F(

1− 2αc, T[xRc

]α)and F (b, z) = 2F1(1, b, 1+b,−z) with 2F1(a, b, c, z)

being the Gauss hypergeometric function.The convergence of Pc emphasizes the importance of considering UE density in

a UDN. The converged value is decreases with λu and tends to 0 when λu → ∞,i.e., in a fully loaded network. Thus, deploying infinite number of BSs will not bringthe UE performance to a unprecedented level. Unlike the coverage probability, theasymptotic ASE will be beneficial from ultra-desification to some extent and butstill highly depends on the UE density.

Numerical Results

Monte Carlo simulations are implemented to validate our theoretical analysis. Inthe simulation, we set Rb = 1m, Rc = 70m, αc = 2.5, α = 4 and the SIR thresholdT = 10dB in all of our results. To calculate or simulate ‘fully loaded network’, weset λu = 2× 108/km2 which is a sufficiently large value so that pa ≈ 1.

The effect of UE density on coverage probability is illustrated in Fig. 3.4 whereour analysis and simulations are well-matched. Meanwhile, we find that coverageprobabilities show completely different trends among different UE densities. In fullload model, the coverage probability starts to decline when the BS density is suf-ficiently large because there are interfering BSs inside the near-field of the typicalUE. It will keep decreasing towards zero as BS density grows. When λu is high butfinite (e.g. λu = 200 or 2000), all the UEs will get served as BS density increases andthe interference becomes saturated. Thus the coverage probability will first decreaseand start increasing again after the saturation. In contrast, when UE density is low(e.g. λu = 20), the probability of no interferer inside the near-field of the typicalUE is very high. Hence, the coverage probability becomes a non-decreasing functionof λb. Meanwhile, we find that the converged value of the coverage probability willdecrease as UE density grows and decay to zero in the full load case. The reason isthat the desired signal is restricted by the bound effect in asymptotic regions andthe overall performance will depend on the interference which refers to UE density.

In Fig. 3.5, we compare the our finding with the previous models to emphasisthe impact of each component. It is shown that the coverage probability is over-estimated with unbounded or single-slope models in high density regimes. This is

3.2. Asymptotic Behavior of Ultra-Densification 35

100 101 102 103 104 105 106

BS density λb (BSs/km2)

0

0.2

0.4

0.6

0.8

1

Cov

erag

e pr

obab

ility

AnalyticalSimulation

Full load

λu=20

λu=200 λ

u=2000

Figure 3.4: Effect of λu on coverage probability under bounded dual-slope model.

SIR

Cov

erag

e Pr

obab

ility

BS Density

1

0Zero convergence

Unity convergence

Unity-zero interval saturation

(b) Dual-slope path loss

(a) Bounded path loss

(c) Partial load (inactive BS)

Baseline

(b) + (c)

(a) + (b) + (c)

Figure 3.5: Impacts of (a) bounded path loss, (b) dual-slope path loss, and (c) partial loadmodels on asymptotic SIR coverage probability as BS density increases

because those models either exaggerate the received power inside the bounded re-gion or underestimate the interference in the near-field. The inaccuracy of pathloss models may mislead the prediction of asymptotic behavior. For instance, thecoverage probability will converge to 1 with unbounded models but to a value be-tween 0 and 1 (assume λu > 0) when applying a bounded model. Consistent with

36 Key Results

100

102

104

106

10−10

10−5

100

105

1010

BS density λb (BSs/km

2)

AS

E (

bp

s/k

m2)

Bounded dual−slope (BD)

Lower bound of BD

Upper bound of BD

Unbounded single−slope (US)

Unbounded dual−slope (UD)

(a) ASE when network is fully loaded

100

102

104

106

10−1

100

101

102

103

BS density λb (BSs/km

2)

AS

E (

bp

s/k

m2)

Bounded dual−slope (BD)

Lower bound of BD

Upper bound of BD

Unbounded single−slope (US)

Unbounded dual−slope (UD)

(b) ASE when λu = 200

Figure 3.6: ASE bounds and comparison with previous models.

the result in Proposition ?? of Paper C, the converged value will decrease as UEdensity grows and finally falls to zero in the full load case as shown in Fig. 3.4. Thisis because the signal is limited by the bound effect and the overall performance willbe dominated by the interference which depends on UE density.

Figure 3.6 depicts the scaling of ASE with regard to BS density. We observethat the ASE first increase with BS density and then converge to a constant whichaligns with our theoretical analysis. The constant is larger than 0 in partially loadednetwork and becomes 0 when the network is full load (λu → ∞). Also, the per-formance is overestimated with unbounded or single-slope models in ultra-denseregions. This is because the simplified models either exaggerate the desired powerinside the bounded region or underestimate the interference in the near-field. Bycomparing Fig. 3.6 with Fig. 3.4, we find that a trade-off occurs between UE cov-erage probability (reliability) and network ASE (rate) during BS densification. Infull load case, network ASE keeps increasing before λb = 104 despite the declin-ing coverage probability. For partially loaded network with λu = 200, the trade-offexists approximately between λb = 101.5 to λb = 103.

Summary

In this section, we investigate the asymptotic behavior of ultra-densification of basestations. With more sorphisticated propagation and load models, we find that theasymptotic behavior of ultra-densification is different from what was known withsimpler assumptions. Both UE coverage probability and ASE converge to a non-zero value unless the UE density goes to infinity. Our results also suggest that thedensification cannot always improve the individual UE performance or boost thenetwork throughput.

3.3. BS Cooperation in UDNs 37

Active UE

Active BS

Cooperating BS

Inactive BS

Figure 3.7: Joint transmission in a homogeneous UDN.

3.3 BS Cooperation in UDNs

In this section, we will investigate the performance of BS cooperation in homoge-neous and heterogeneous UDN. More details and results can be found in Papers Vand VI.

3.3.1 Homogeneous UDN scenario

In previous studies, e.g., [32] [56], the network topology is composed by BS-centricVoronoi cells which represent the BS coverage areas as shown in Fig. 2.1. However,since there exists massive empty cells in a UDN, BS-centric topology may lead toan ineffective map partition. Concerning this, we propose to reverse the roles ofBS and UE in the topology and utilize UE-centric Voronoi cells for modeling thecooperation in a UDN, as shown in Fig. 3.7. The BSs inside a UE-centric Voronoicell is closer to the particular UE than to other UEs. Moreover, based on the UDNdensity assumption, we can approximate the kth closest BS inside the Voronoi cellas the kth closest BS in the entire network when k λb

λu. Similar with in previous

section, a bounded dual-slope propagation model and partially loaded network areconsidered.

Cooperation Schemes

We assume UE i in the network is jointly served by the set of N closest BSs inits own Voronoi cell as shown in Fig. 3.7, denoted by Ci =

⋃Nj=1 BSj . Within the

38 Key Results

cooperation set, all the BSs jointly transmit the same message to the UE withnon-coherent JT or coherent JT.

In non-coherent JT, the receivers apply open-loop joint processing scheme asin [73], where signals from different transmitters are added by power summation.The desired signal power for non-coherent JT is given by

SNJi =

∑x∈Ci

|hx,i|2`(dx,i) , (3.8)

where dx,i and hx,i denote the distance and channel between BS x and UE i,hx,i ∼ CN (0, 1).

In coherent JT, we assume the CSI is available at the BS side. In this case,BSs can design precoder to adjust the phase shift of the channel and amplify thecorresponding channel gain. The MRT precoder is considered which is describe inChapter 2.2.3. The desired signal power for coherent JT is thus

SCJi =

∣∣∣∣∣∑x∈Ci

hx,iwx,i

√`(dx,i)

∣∣∣∣∣2

. (3.9)

Therefore, the SIR of UE i with two cooperation schemes, γNJi and γCJ

i are shownas follows:

γNJi = SNJ

i∑x∈ΦC\Ci

|hx,i|2`(dx,i)(3.10)

γCJi = SCJ

i∑x∈ΦC\Ci

|hx,i|2`(dx,i), (3.11)

where ΦC is the set of active BSs in the whole network.

Cooperation Gains

The UE SE R is selected as the performance metrics, which is given by:

R = log2(1 + γ). (3.12)

We set the UE SE under single association Ro as our baseline to measure the gainof cooperation. The SE gain Gc is defined as the ratio between SE difference ∆Rand Ro as follows:

Gc = ∆RRo

= RJ −Ro

Ro, (3.13)

where RJ and Ro are the UE SE with and without BS cooperation, respectively.


0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 1502

3

4

5

6

7

8

Critical distance Rc (m)

Use

r S

E (

bps/

Hz)

Co−JT, N=2Co−JT, N=5No cooperationNon−JT, N=2Non−JT, N=5

No gain for nonJT

Higher gain for larger Rc

Little gain for small Rc,more cooperation BSs

may be even worse

Figure 3.8: UE performance with different critical distances when path loss expo-nents=[2,4], λb = 4000/km2, λu = 200/km2.

Numerical Results

We present the performance of both cooperation schemes in terms of critical dis-tance in Fig. 3.8. To our surprise, non-coherent JT has even worse performance thansingle BS association. In non-coherent JT, the desired power summation grows di-minishingly because the transmitting distance gets longer, but the interference canbe considered as almost linearly increasing with the number of cooperating BSs. Asa result, the increment in the desired signal cannot compensate the faster growinginterference which leads to no gain. Thus, we will neglect non-coherent JT in furtherdiscussions. Regarding coherent JT, we find that the cooperation gain grows alongwith critical distance. When critical distance is short, more cooperating BSs makeSE even worse; when the critical distance is large, the cooperation gain increaseswith the number of cooperating BS.

The impact of active UE and BS densities on cooperation is shown in Fig. 3.9.First, as active UE density increases, cooperation gain keeps growing while SE dropsbecause the baseline gets lower and more interference is generated. This trend isthe same as the effect of critical distance aforementioned since increasing criticaldistance is equivalent with increasing UE density in the near-field. Regarding BSdensity, the cooperation gain can be divided into three regions: first decreases loga-rithmically, then decreases in a lower speed, and finally starts to increase after theBS density reaches a threshold. In region 1, the cooperation gain will decrease sincethe baseline will increase logarithmically with λb [26]. In region 2, the decreasingspeed of the gain slows down because UEs start to enter the bounded region. Finallyin region 3, the probability of d1 < Rb is quite high and Ro tends to be a constant.Therefore, the cooperation gain increases via further densification.

40 Key Results

50 100 150 200 250 300 350 400 450 5002

3

4

5

6

7

8

UE density λu (active users/km2)

Use

r S

E (

bps/

Hz)

50 100 150 200 250 300 350 400 450 5000.1

0.15

0.2

0.25

0.3

Cooperation gain

CoJTNo cooperationCooperation gain

(a) Cooperation gain with different UEdensities

103

104

105

106

0

2

4

6

8

10

12

BS density λb (BSs/km2)

Use

r S

E (

bps/

Hz)

103

104

105

1060.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3

Cooperation gain

No cooperationCoJTCooperation gain

Region 1 Region 2 Region 3

(b) Cooperation gain with different BSdensities.

Figure 3.9: Cooperation gain when Rc = 70m, N = 5.

Wired fronthaul

c-RRH 1 c-RRH 2 c-RRH 3

UE 1 UE 2 UE 3 UE 4

w-RRH 1

Wireless access link

BBU Pool

Interference channels

RRH 3

Figure 3.10: An example of PW-CRAN.

3.3.2 Heterogeneous UDN scenarioThe concept of C-RAN emerged as a new promising architectures for achievingnetwork densification. One of the challenges in the C-RAN architecture is capacityexpansion after an initial network deployment. In the context of C-RAN, one mayconsider installing supplementary RRHs with additional fiber fronthaul links to theexisting architecture. However, installation of optical cables brings about immenseexpense in general [74], and will be cumbersome when the geographic location isinappropriate for fiber placement. This dilemma motivates us to consider a newarchitecture called a partly wireless C-RAN (PW-CRAN) shown in Fig. 3.10. InPW-CRAN, additional radio nodes are wirelessly connected to the adjacent RRHs


0 1 2 3 4 5 6 7 8 9 1010

12

14

16

18

20

22

24

26

28

τtotal

(ms)

Avera

ge s

um

SE

for

schedule

d U

Es (b

ps/H

z)

Global precodingPartially independent operationCompletely independent operation

Figure 3.11: Average sum SE for scheduled UEs vs. the total delay τtotal.

when fiber-based fronthauling to the BBU pool is not available. For brevity, wecall the conventional RRH as c-RRH and the extra wireless node as w-RRH. Theheterogeneous deployment enables a UE to served by either a c-RRH or a w-RRHdepending on cell coverage areas, which may lead to UE performance discrimina-tions. The UEs associated with c-RRH may outperform the others due to the jointlyprocessing driven from the centralized architecture. Such geographic unfairness willresult in fluctuant UE performance thus severely affect the QoS of the UEs. Withthis regard, the w-RRHs can join the coordination via the wireless link.

The main obstacle to the joint processing in the PW-CRAN architecture is thedelay τtotal = τfronthaul + τwireless, because the back and forth communications be-tween the BBU pool and a w-RRH need to go through the wireless link in additionto the existing wired fronthaul. Therefore, we compare two schemes, namely oper-ating with or without cooperation between c-RRHs and w-RRHs. The definition ofthe two options are given as follows:

• Global Precoding (GP): A global precoding strategy in which not only c-RRHs but w-RRHs forward their estimated CSI to the BBU pool, therebydetermining global precoding matrix.

• Partially Independent Operation (PIO): A partially independent strategy inwhich all c-RRHs deliver their CSI to the BBU pool and generate precodingmatrix. Contrarily, w-RRHs separately communicate with their correspondingUEs.

42 Key Results

10.90.80.70.60.50.40.30.20.1010.5

11

11.5

12

12.5

13

13.5

14

14.5

15

15.5

16

α

Avera

ge s

um

SE

for

schedule

d U

Es (

bps/H

z)

GP, τtotal

=4ms

GP, τtotal

=6ms

GP, τtotal

=8ms

PIO, τtotal

=4ms

PIO, τtotal

=6ms

PIO, τtotal

=8ms

Figure 3.12: Average sum SE for scheduled UEs vs. the ratio α of the wireless link delayto the total delay for varying τtotal.

Numerical Results

In Fig. 3.11, we present the average sum SE for scheduled UEs as a function oftotal delay to evaluate the impact of delay on both cooperation strategies. Theperformance of completely independent operation scenario is also measured as abenchmark, in which all c-RRHs and w-RRHs work separately without joint pre-coding. It is shown that the average sum SE for scheduled UEs of GP decreasesmore dramatically with the increase in the total delay compared to that of PIO.This is due to the fact that the outdated CSI degrades the accuracy of GP morethan PIO as the total delay increases. We can observe in Fig. 3.11 that PIO becomessuperior to GP after τtotal = 4ms. Furthermore, even completely independent oper-ation outperforms GP for a large τtotal, e.g., greater than 7ms. The result indicatesthat it is beneficial to adopt PIO rather than GP if the total delay is greater thana specific value, which has a negative impact on the coordination.

Fig. 3.12 shows the average sum SE for scheduled UEs versus the ratio of thedelay on wireless link to the total delay, defined as α = τtotal−τc

τtotal, e.g., α = 0.7 in

Fig. 3.11. It is observed that if the total delay τtotal is high, GP is inferior to PIOeven with small value of α. This is because the increase in the total delay leadsto relatively larger additional delay on the extra wireless link, thus harming theperformance of global cooperation.

Summary

We evaluate BS cooperation schemes in two UDN scenarios. In a homogeneousUDN, non-coherent JT even weakens the UE performance because the cooperative

3.4. Initial Access in mmWave UDNs 43

...

Cell search Data transmission

BS beam 1 tx

UE beam m rx

1 scan cycle

Transmission frame

...

k scan cycles,k=0,1,2...

t

BS beam 2 tx

BS beam NBS tx

Random Access

...BS

beam 1 rx

t

BS beam 2 rx

BS beam NBS rx

UE beam k tx

Figure 3.13: Illustration of transmission frame under random beamforming.

power cannot compensate for the additional interference. With the knowledge ofCSI, coherent JT can provide SE gain depending on environmental characteris-tics. In a heterogeneous UDN, there exists a range of delay in which the w-RRHsshould not be involved in the BS coordination. If the amount of additional wirelesslink delay is large, the delayed CSI in GP exerts more adverse influence than theadditional interference in PIO.

3.4 Initial Access in mmWave UDNs

In this section, we study the initial access design in mmWave dense networks. Exist-ing initial access schemes focus on finding ’the best quality’ for data transmission.However, in many newly developed applications comprising wake-up radios, thesynchronization may take longer than the actual data transmission, leading to apoor latency performance. The large amount of BSs in a UDN offers massive con-necting possibilities if there is no restriction on finding the best link quality. Forthis reason, we propose to apply random beamforming as an alternative to reducethe cell-search latency, in which the BSs (and UEs) focus their antenna patternstoward a randomly pick direction. The details of the analysis and more results ofthe work can be found in Papers VII, VIII and IX.

3.4.1 Random Beamforming based Initial Access

We assume the system time is divided into two phases within each coherence inter-val: 1) an initial access period comprised of a cell search phase and a random accessphase, 2) a data transmission period with beam alignment. Figure 3.13 illustratesthe whole transmission frame.

44 Key Results

The cell-search period comprises of several mini-slots. In each mini-slot, everyBS randomly picks a direction without repetition out of NBS candidates and sendspilot to UEs. A scan cycle is defined as the period within which one BS finishessending cell-search pilots to all NBS directions, i.e. NBS mini-slots. In each scancycle, the UE antenna points to a random direction out of NUE, and the BS coversall non-overlapping NBS directions. Different from the existing schemes in whichthe search period is fixed [62], the cell-search period of random beamforming can bedynamically adjusted. Once the UE received a pilot signal that meets a predefinedSINR threshold, it is associated to the corresponding BS.

Given a successful cell search phase, the UE acquires the direction where itreceives the strongest signal. In the following random access phase, the UE initiatesthe connection to its desired serving BS by transmitting RA preambles to thatdirection. The BS will scan for the presence of the RA preamble and will also learnthe BF direction at the BS side. If cell search fails, the UE will skip the randomaccess phase and repeat the cell search in the next frame. In real systems, the UEpicks the RA preamble from a certain number of orthogonal preamble sequences.Since the main focus of this paper is random beamforming based cell search phase,the impact of random access performance is left in our future work. Therefore, wemake the following two assumptions in our work:

• There is no RA preamble collision for all UEs, i.e. the BS can detect all thepreambles.

• The random access phase and cell search phase share the same SINR thresh-old.

Based on these two assumptions, we have simultaneous cell search and randomaccess success or failure.

We will evaluate three performance metrics given a cell-search budget of Ncmini-slots, the detection failure probability Pf (Nc), the expected initial access la-tency DI(Nc), and the total data transmission delay DT (Nc), formally defined inDefinition ?? to ?? of Paper G. The blockage path loss model, analog beamforming,and 3GPP NR frame are applied in this section. In addition to the simple modelwhere only mainlobe gain and LOS links are considered, we study the effect of NLOSpaths and sidelobe gain as well by setting: αL ≤ αN <∞ and 0 < ε < 1. Leverag-ing tools from stochastic geometry, the expression of detection failure probability,expected initial access latency and total data transmission latency are present inProposition ?? to ?? of Paper G.

Numerical Results

In this part, we focus on Nc = NBS, i.e. the random cell-search lasts for 1 scancycle. For a fair comparison, we set the number of SS blocks in one SS burst to16, 32, and 64 for random beamforming, iterative search and exhaustive search, so


10−5

10−4

10−3

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

BS density

Dete

ction failu

re p

robabili

ty

RB SBP sim+analysis (1)



RB ULA

IS ULA

ES ULA

Figure 3.14: BS density effect on failure probability. (1) refers ε = 0, αN =∞, (2) refersε = 0, αL ≤ αN <∞, and (3) refers 0 ≤ ε ≤ 1, αN =∞. ("RB", "IS", and "ES" stand forrandom beamforming, iterative search and exhaustive search)

that every scheme can complete one scan cycle within one transmission frame. Thebeamwidth in the first stage of iterative search is set to θBS1 = 90°.

Fig. 3.14 shows the cell-search performance against BS density λ. It is shown thatthe effect of NLOS path on failure probability is negligible due to the overlappingof curves (1) and (2). Nevertheless, the sidelobe gain plays an important role whichcauses a notable decrease of detection failure probability. In general, the detectionfailure probability reduces with the BS density for all schemes. The gap betweenrandom beamforming and other schemes diminishes rapidly as BS density grows. Indense regime where BS density approaches 10−3/m2, all the probabilities convergeto 0 gradually.

Fig. 3.15 compares the the expected initial access latency of three schemes overdifferent BS density regimes. When BS density is low, the initial access latencyof exhaustive search is the shortest even though it takes the longest cell-searchperiod. The reason is that it consumes the least number of overhead frames beforetransmission due to the low failure probability. The initial access latency tendsto converge to a constant, which is one ss burst period, as the failure probabilityconverges to 0. On the other hand, the initial access latency of random beamformingkeeps reducing as more BSs are deployed. When the BS density exceeds 150/km2,the expected initial access latency of random beamforming becomes lower thanthat of exhaustive search and iterative search schemes. The main reason is theavailability of more candidate BSs to register the UE. Noting that the best possiblevalues for the failure probability and normalized expected latency (number of mini-

46 Key Results

10−5

10−4

10−3

100

101

102

103

BS density

Expecte

d IA

late

ncy [m

s]

RB

IS

ES

Figure 3.15: BS density effect on IA delay (ULA model).

103

104

105

106

107

108

109

100

101

102

103

104

Packet size L [bits]

Tota

l data

tra

nsm

issio

n late

ncy [m

s]

RB

IS

ES

λ = 10−4

λ = 10−3

Figure 3.16: Total data transmission latency vs packet size.

slot) are 0 and 1, respectively. By employing random beamforming, we can get closeto those values in dense mmWave networks.

The reduced latency in initial access with random beamforming is based on thesacrifice of finding the ’best’ BS. Apparently, such scheme is not suitable for anykind of data transmission. For instance, when the data traffic is video streamingwith huge data package size, it may be better to take some extra time in the initial


access process and find the ’best’ BS. We present the result of total transmissiondelay in Fig. 3.16. When the BS density is not large enough (λ = 10−4), otherschemes always outperform random beamforming due to its high failure probabilityand low achievable rate. However in dense regime, as the available BSs increase, weachieve a lower total delay for relative shorter packet sizes with initial access underrandom beamforming. Thus the most favorable scenario for initial access underrandom beamforming is transmitting short packets in dense networks.

SummaryBy applying random beamforming to inital access in mmWave networks, we findthat it appears to be very efficient from signaling and computational perspectives.Meanwhile, the proposed scheme can provide a very good performance in termsof both control plane and data plane, especially in dense BS deployment scenariosand large antenna regimes. It is interesting to see the possibility of reducing thecomplexity of schemes as the network is getting more complicated.

Chapter 4

Conclusion and Future Work

"In literature and in life we ultimately pursue, not conclusions, butbeginnings."

Sam Tanenhaus

The upcoming data tsunami poses a tremendous challenge to the wireless indus-try and academia. In this thesis, we aimed to provide some insights from networkdensification perspective. To this end, we have carried out a study on the per-formance analysis and operational guidelines of a UDN. In this chapter, we firstprovide our concluding remarks related to the research questions stated in Chap-ter 1.4. Then, we pose some opening problems and highlight the potential researchdirections for future work.

4.1 Concluding Remarks

We started from discussing the benefit of densely deploying BSs in the network.Although proven to be the most effective method in the past, it is unclear if theconclusion will hold for the future scenario. Thus we proposed the first high-levelresearch question:

• HQ 1: How much benefit can we get from network densification?

To answer this question, we first investigated the resource trade-offs in networkprovisioning for both current macro-cellular network and future ultra-dense net-work. We compare the effectiveness of densification with the other two elementswhich can improve network capacity: spectrum expansion and multi-antenna sys-tems. In current network deployment, densification outperforms the other two op-tions. The linear capacity gain made network densification an efficient solution fromboth performance and cost perspectives. However, the gain of densification dimin-ishes as BS density keeps increasing and finally exceeds UE density which formsa UDN. In a UDN, acquiring more spectrum or installing multi-antennas are pre-ferred than further densification. Furthermore, the importance of specifying UE

49

50 Conclusion and Future Work

demand is highlighted. Our results reveal that spectrum has a significant impacton achieving extreme high individual requirements while densification plays a moreimportant role in accommodating massive UEs.

The diminishing gain appearing in a UDN makes us curious to know if thereexists a terminal on the way of densification. Such uncertainty leads to the study onthe asymptotic behavior of densification in the next part of the thesis. In order tocapture the characteristics of a UDN, we applied a bounded dual-slope propagationand partial load model. Numerical results show that the coverage probability of theUE neither drops to zero as some pessimistic predictions in references nor goes toinfinity as we wish. Instead, it will converge to a finite value considering a certainUE density. Similarly, the ASE will converge as well under the same assumptions.Meanwhile, we find a rate-reliability trade-off during densification process whichshould draw attentions in network provisioning.

To summarize, BS densification is still an important option to achieve the futurehigh data traffic demand. Nevertheless, the benefit is diminishing in terms of bothsubstituting other resources and system performance. Therefore, the MNOs needto make effective as well as efficient decision on the combination of densificationwith other alternatives, e.g. acquiring more bandwidth and multi-antenna systems.

Next we shifted our focus towards the operations in a UDN. As indicated in theliterature, the operation in traditional cellular networks may need to be adaptedin a UDN scenario, e.g. interference management, backhauling design, handoverschemes. As complementary, we turned our attention into BS cooperation and initialaccess design, thereby addressing the second high-level question:

• HQ 2: How should certain operations be adjusted in a UDN scenario?

We first studied cooperative transmissions in a homogeneous UDN. By compar-ing schemes of non-coherent JT and coherent JT, we find that cooperation is notbeneficial without CSI in a UDN. With the CSI information, cooperation is favor-able in UDN scenarios with longer critical distance and smaller path loss exponents.Moreover, employing cooperation is more effective in high UE density regions. Onthe contrary, high BS density is not helpful until it reaches the threshold and BSsfall into the bounded region. Next, we considered a heterogeneous PW-CRAN sce-nario which is characterized by extra w-RRHs connected through existing c-RRHs.Two cooperation strategies are evaluated namely global precoding and partially in-dependent operation. We investigated the impact of extra delay caused by w-RRHsand revealed that there exists a range of delay in which w-RRHs should not beincluded in the cooperation. The reason is that if the amount of additional wirelesslink delay is substantial, the delayed CSI in global precoding exerts more adverseinfluence than the additional interference in partially independent operation.

At last, we investigated the performance of initial cell-search of mmWave net-works under random beamforming. Comparisons are made between our methodand two widely-used schemes in both control and data plane. Our results showedthat random beamforming, though being very efficient from signaling and compu-tational perspectives, can provide sufficient detection and latency performance in

4.2. Future Works 51

control plane, especially in dense BS deployment scenarios. From data plan perspec-tive, our scheme with random beamforming outperforms the others for short packettransmission in high density regimes. Therefore, it may be unnecessary to developcomplex algorithms for cell search process in future dense mmWave networks.

In conclusion, similar as other operations, BS cooperation and initial accessdesign may need reconsideration for a UDN environment. On one hand, the in-terference nature and heterogeneous deployment in a UDN make traditional coop-eration schemes less powerful. Thus, more complex schemes need to be designedto improve the system performance. On the other hand, the high BS density in ammWave UDN enables the UEs to accomplish the initial access in a simple mannerwith shorter delay while keeping a high data rate. Therefore, densification may alsohelp to reduce the complexity for certain operations.

4.2 Future Works

Despite the contributions made by this thesis and previous works, there are issuesstill remain to be addressed in this research area. In this section, we provide thefollowing suggestions as extension of our work and future research directions:

• Extension of current work: The resource tradeoffs in Chapter 3.1 is eval-uated in a qualitative manner, i.e. the inter-substitution property among thethree elements. It might be more intuitive to combine with the cost of eachelement, e.g. the price of different BSs as well as installing them and spec-trum auction price, then make a quantitative analysis. Our work in partlywireless C-RAN was based on a simple example with limited number of BSsand UEs. A large scale analysis is needed for better understanding the criteriafor successful cooperation in such scenarios.

• Learning based initial access design: As shown in the thesis, randombeamforming is effective for dense mmWave networks in an average sense.Nevertheless, each specific cell may benefit from different beamwidths and/orbeam patterns. The recent success of deep learning underpins new and pow-erful tools that may help manage such situations. It is possible to developa learning based initial access scheme serving every BS in the mmWave net-work. Although random beamforming achieved sufficient performance withlow complexity, a learning algorithm may generate individual search patternsfor each BS that further reduce the search latency and increase the transmis-sion data rate.

• Ultra-reliable communications in UDN: Ultra-reliable communicationis becoming demanding in 5G and beyond. Diversity schemes are utilized toimprove the reliability of link. Time diversity is not valid when the tolerablelatency is shorter than the channel coherence time or when the reliability re-quirements are very stringent. Meanwhile, frequency diversity may not scalewith the number of devices. In this case, spatial diversity seems to be the

52 Conclusion and Future Work

only solution. In a UDN, the idle BSs can provide multi-connectivity for theUEs, thereby ensuring high-reliable communication. However, some funda-mental questions remain to be addressed. For instance, what is the optimalnumber of links needed to ensure a given reliability target? How to deal withsynchronization, non-reciprocity issues, and other imperfections?

• Security of UDN: In wireless transmissions, anyone within the coverage areaof an open, unencrypted wireless network can "sniff", or capture and record,the traffic, gain unauthorized access to internal network resources as well asto the internet. Such security breaches have become important concerns forboth enterprise and home networks. Such security and privacy issues becomesmore serious for UDN due to the features: dynamic network topology, complexnetwork components, high UE diversity, high BS density and etc. In a UDNscenario, if a single node is compromised, a great number of nodes may becomeinsecure as a result. Thus, the investigation of attacks and their impact onUDN should be investigated. Afterwards, the effective countermeasures canbe designed to mitigate attacks based on the understanding of attacks.

Bibliography

[1] Ericsson, “Ericsson mobility report,” Nov 2017. [Online]. Avail-able: https://www.ericsson.com/assets/local/mobility-report/documents/2017/ericsson-mobility-report-november-2017.pdf

[2] Cisco, “Visual networking index: Global mobile data traffic fore-cast update, 2014-2019,” Feb 2015, White Paper. [Online]. Avail-able: http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visual-networking-index-vni/white_paper_c11-520862.pdf

[3] “5G systems,” White Paper, Ericsson, January 2015.

[4] “The 1000x data challenge,” White Paper, Qualcomm Incorporated, June 2013.

[5] M. Dohler, R. W. Heath, A. Lozano, C. B. Papadias, and R. A. Valenzuela,“Is the phy layer dead?” IEEE Communications Magazine, vol. 49, no. 4, pp.159–165, April 2011.

[6] F. Rusek, D. Persson, B. K. Lau, E. Larsson, T. Marzetta, O. Edfors, andF. Tufvesson, “Scaling up MIMO: Opportunities and challenges with very largearrays,” IEEE Signal Processing Magazine, vol. 30, no. 1, pp. 40–60, Jan 2013.

[7] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive mimo fornext generation wireless systems,” IEEE Communications Magazine, vol. 52,no. 2, pp. 186–195, February 2014.

[8] N. Bhushan, J. Li, D. Malladi, R. Gilmore, D. Brenner, A. Damnjanovic,R. T. Sukhavasi, C. Patel, and S. Geirhofer, “Network densification: the domi-nant theme for wireless evolution into 5G,” IEEE Communications Magazine,vol. 52, no. 2, February 2014.

[9] D. Lôpez-Pèrez, M. Ding, H. Claussen, and A. H. Jafari, “Towards 1 gbps/uein cellular systems: Understanding ultra-dense small cell deployments,” IEEECommunications Surveys Tutorials, vol. 17, no. 4, pp. 2078–2101, Fourthquar-ter 2015.

[10] Arraycomm. Cooper’s law. [Online]. Available: http://www.arraycomm.com/technology/coopers-law/

53

https://www.ericsson.com/assets/local/mobility-report/documents/2017/ericsson-mobility-report-november-2017.pdf

https://www.ericsson.com/assets/local/mobility-report/documents/2017/ericsson-mobility-report-november-2017.pdf

http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visual-networking-index-vni/white_paper_c11-520862.pdf

http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visual-networking-index-vni/white_paper_c11-520862.pdf

http://www.arraycomm.com/technology/coopers-law/

http://www.arraycomm.com/technology/coopers-law/

54 Bibliography

[11] V. Chandrasekhar, J. G. Andrews, and A. Gatherer, “Femtocell networks: asurvey,” IEEE Communications Magazine, vol. 46, no. 9, pp. 59–67, September2008.

[12] R. Iyer, J. Parker, and P. Sood, “Intelligent networking for digital cellular sys-tems and the wireless world,” in Proc. GLOBECOM ’90: IEEE Global Telecom-munications Conf. and Exhibition, Dec. 1990, pp. 475–479 vol.1.

[13] J. G. Andrews, “Seven ways that hetnets are a cellular paradigm shift,” IEEECommunications Magazine, vol. 51, no. 3, pp. 136–144, March 2013.

[14] J. G. Andrews, H. Claussen, M. Dohler, S. Rangan, and M. C. Reed, “Femto-cells: Past, present, and future,” IEEE Journal on Selected Areas in Commu-nications, vol. 30, no. 3, pp. 497–508, April 2012.

[15] R. Heath, S. Peters, Y. Wang, and J. Zhang, “A current perspective on dis-tributed antenna systems for the downlink of cellular systems,” IEEE Com-munications Magazine, vol. 51, no. 4, pp. 161–167, April 2013.

[16] M. Kamel, W. Hamouda, and A. Youssef, “Ultra-dense networks: A survey,”IEEE Communications Surveys Tutorials, vol. 18, no. 4, pp. 2522–2545, 2016.

[17] C. S. Chen, V. M. Nguyen, and L. Thomas, “On small cell network deployment:A comparative study of random and grid topologies,” in 2012 IEEE VehicularTechnology Conference (VTC Fall), Sept 2012, pp. 1–5.

[18] M. Ding, D. Lopez-Perez, G. Mao, P. Wang, and Z. Lin, “Will the area spectralefficiency monotonically grow as small cells go dense?” in Proc. IEEE GlobalCommunications Conf. (GLOBECOM), Dec. 2015, pp. 1–7.

[19] R. Baldemair, T. Irnich, K. Balachandran, E. Dahlman, G. Mildh, Y. Selén,S. Parkvall, M. Meyer, and A. Osseiran, “Ultra-dense networks in millimeter-wave frequencies,” IEEE Communications Magazine, vol. 53, no. 1, pp. 202–208, Jan. 2015.

[20] Nokia, “Ultra dense network (UDN) white paper,” 2016.

[21] J. Park, S. Kim, and J. Zander, “Tractable resource management with up-link decoupled millimeter-wave overlay in ultra-dense cellular networks,” IEEETransactions on Wireless Communications, vol. 15, no. 6, pp. 4362–4379, Jun.2016.

[22] J. Zander, “Beyond the ultra-dense barrier: Paradigm shifts on the road beyond1000x wireless capacity,” IEEE Wireless Communications, vol. 24, no. 3, pp.96–102, Jun. 2017.

[23] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage andrate in cellular networks,” IEEE Transactions on Communications, vol. 59,no. 11, pp. 3122–3134, November 2011.

Bibliography 55

[24] X. Zhang and J. G. Andrews, “Downlink cellular network analysis with multi-slope path loss models,” IEEE Transactions on Communications, vol. 63, no. 5,pp. 1881–1894, May 2015.

[25] J. Liu, M. Sheng, L. Liu, and J. Li, “Effect of densification on cellular networkperformance with bounded pathloss model,” IEEE Communications Letters,vol. PP, no. 99, pp. 1–1, 2016.

[26] J. Park, S. L. Kim, and J. Zander, “Asymptotic behavior of ultra-dense cellularnetworks and its economic impact,” in 2014 IEEE Global CommunicationsConference, Dec 2014, pp. 4941–4946.

[27] Y. Yang and K. W. Sung, “Tradeoff between spectrum and densification forachieving target user throughput,” in 2015 IEEE 81st Vehicular TechnologyConference (VTC Spring), May 2015, pp. 1–6.

[28] C. Li, J. Zhang, M. Haenggi, and K. B. Letaief, “User-centric intercell in-terference nulling for downlink small cell networks,” IEEE Transactions onCommunications, vol. 63, no. 4, pp. 1419–1431, Apr. 2015.

[29] M. Polignano, P. Mogensen, P. Fotiadis, L. Chavarria, I. Viering, and P. Zanier,“The inter-cell interference dilemma in dense outdoor small cell deployment,”in Proc. IEEE 79th Vehicular Technology Conf. (VTC Spring), May 2014, pp.1–5.

[30] J. Huang, H. Zhang, W. Xu, and H. Zhang, “Grouping based inter-cell in-terference coordination in LTE-A dense small-cell networks,” in Microwave,Antenna, Propagation and EMC Technologies for Wireless Communications(MAPE), 2013 IEEE 5th International Symposium on, Oct 2013, pp. 78–83.

[31] B. Soret, K. I. Pedersen, N. T. K. Jørgensen, and V. Fernández-López, “In-terference coordination for dense wireless networks,” IEEE CommunicationsMagazine, vol. 53, no. 1, pp. 102–109, Jan. 2015.

[32] Y.-P. Zhang, S. Feng, P. Zhang, L. Xia, Y. chun Wu, and X. Ren, “Inter-cellinterference management in LTE-A small-cell networks,” in Vehicular Tech-nology Conference (VTC Spring), 2013 IEEE 77th, June 2013, pp. 1–6.

[33] M. Coldrey, J. E. Berg, L. Manholm, C. Larsson, and J. Hansryd, “Non-line-of-sight small cell backhauling using microwave technology,” IEEE Communi-cations Magazine, vol. 51, no. 9, pp. 78–84, September 2013.

[34] D. C. Chen, T. Q. S. Quek, and M. Kountouris, “Backhauling in heterogeneouscellular networks: Modeling and tradeoffs,” IEEE Transactions on WirelessCommunications, vol. 14, no. 6, pp. 3194–3206, Jun. 2015.

56 Bibliography

[35] P. Kela, M. Costa, J. Turkka, K. Leppänen, and R. Jäntti, “Flexible backhaul-ing with massive MIMO for ultra-dense networks,” IEEE Access, vol. 4, pp.9625–9634, 2016.

[36] B. Li, D. Zhu, and P. Liang, “Small cell in-band wireless backhaul in massiveMIMO systems: A cooperation of next-generation techniques,” IEEE Transac-tions on Wireless Communications, vol. 14, no. 12, pp. 7057–7069, Dec. 2015.

[37] X. Ge, H. Cheng, M. Guizani, and T. Han, “5G wireless backhaul networks:challenges and research advances,” IEEE Network, vol. 28, no. 6, pp. 6–11,Nov. 2014.

[38] Y. Hong, X. Xu, M. Tao, J. Li, and T. Svensson, “Cross-tier handover analysesin small cell networks: A stochastic geometry approach,” in Proc. IEEE Int.Conf. Communications (ICC), Jun. 2015, pp. 3429–3434.

[39] H. Leem, J. Kim, D. K. Sung, Y. Yi, and B. Kim, “A novel handover schemeto support small-cell users in a HetNet environment,” in Proc. IEEE WirelessCommunications and Networking Conf. (WCNC), Mar. 2015, pp. 1978–1983.

[40] J. Park, S. Y. Jung, S. Kim, M. Bennis, and M. Debbah, “User-centric mobilitymanagement in ultra-dense cellular networks under spatio-temporal dynam-ics,” in Proc. IEEE Global Communications Conf. (GLOBECOM), Dec. 2016,pp. 1–6.

[41] R. Arshad, H. ElSawy, S. Sorour, T. Y. Al-Naffouri, and M. Alouini, “Handovermanagement in dense cellular networks: A stochastic geometry approach,” inProc. IEEE Int. Conf. Communications (ICC), May 2016, pp. 1–7.

[42] E. Demarchou, C. Psomas, and I. Krikidis, “Mobility management in ultra-dense networks: Handover skipping techniques,” IEEE Access, vol. 6, pp.11 921–11 930, 2018.

[43] I. Hwang, B. Song, and S. Soliman, “A holistic view on hyper-dense hetero-geneous and small cell networks,” IEEE Communications Magazine, vol. 51,no. 6, pp. 20–27, June 2013.

[44] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling andanalysis of k-tier downlink heterogeneous cellular networks,” IEEE Journal onSelected Areas in Communications, vol. 30, no. 3, pp. 550–560, April 2012.

[45] M. Giordani, M. Mezzavilla, and M. Zorzi, “Initial access in 5G mmWavecellular networks,” IEEE Communications Magazine, vol. 54, no. 11, pp. 40–47, Nov. 2016.

[46] C. N. Barati, S. A. Hosseini, M. Mezzavilla, T. Korakis, S. S. Panwar, S. Ran-gan, and M. Zorzi, “Initial access in millimeter wave cellular systems,” IEEETransactions on Wireless Communications, vol. 15, no. 12, pp. 7926–7940, Dec.2016.

Bibliography 57

[47] V. Desai, L. Krzymien, P. Sartori, W. Xiao, A. Soong, and A. Alkhateeb, “Ini-tial beamforming for mmwave communications,” in Proc. Systems and Com-puters 2014 48th Asilomar Conf. Signals, Nov. 2014, pp. 1926–1930.

[48] H. R. Varian, Intermediate Microeconomics: A modern approach, 8th ed. W.W. Norton & Company, 2010.

[49] S. Catreux, P. F. Driessen, and L. J. Greenstein, “Simulation results foran interference-limited multiple-input multiple-output cellular system,” IEEECommunications Letters, vol. 4, no. 11, pp. 334–336, Nov 2000.

[50] A. Ganz, C. M. Krishna, D. Tang, and Z. J. Haas, “On optimal design ofmultitier wireless cellular systems,” IEEE Communications Magazine, vol. 35,no. 2, pp. 88–93, Feb 1997.

[51] A. D. Wyner, “Shannon-theoretic approach to a gaussian cellular multiple-access channel,” IEEE Transactions on Information Theory, vol. 40, no. 6, pp.1713–1727, Nov 1994.

[52] J. Xu, J. Zhang, and J. G. Andrews, “On the accuracy of the wyner model incellular networks,” IEEE Transactions on Wireless Communications, vol. 10,no. 9, pp. 3098–3109, September 2011.

[53] S. N. Chiu, D. Stoyan, W. S. Kendall, and J. Mecke, Stochastic geometry andits applications, 3rd ed. Wiley, 2013.

[54] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for modeling,analysis, and design of multi-tier and cognitive cellular wireless networks: Asurvey,” IEEE Communications Surveys Tutorials, vol. 15, no. 3, pp. 996–1019,Third 2013.

[55] G. Alfano, M. Garetto, and E. Leonardi, “New insights into the stochasticgeometry analysis of dense CSMA networks,” in 2011 Proceedings IEEE IN-FOCOM, April 2011, pp. 2642–2650.

[56] S. M. Yu and S.-L. Kim, “Downlink capacity and base station density in cellularnetworks,” in Proc. Ad Hoc and Wireless Networks (WiOpt) 2013 11th Int.Symp. and Workshops Modeling and Optimization in Mobile, May 2013, pp.119–124.

[57] D. Stoyan, W. Kendall, and J. Mecke, Stochastic Geometry and Its Applica-tions. John Wiley and Sons, 1996.

[58] C. Li, J. Zhang, and K. Letaief, “Throughput and energy efficiency analysis ofsmall cell networks with multi-antenna base stations,” IEEE Transactions onWireless Communications, vol. 13, no. 5, pp. 2505–2517, May 2014.

58 Bibliography

[59] S. Lee and K. Huang, “Coverage and economy of cellular networks with manybase stations,” IEEE Communications Letters, vol. 16, no. 7, pp. 1038–1040,July 2012.

[60] B. Błaszczyszyn, M. K. Karray, and H. P. Keeler, “Wireless networks appearpoissonian due to strong shadowing,” IEEE Transactions on Wireless Com-munications, vol. 14, no. 8, pp. 4379–4390, Aug. 2015.

[61] T. Bai, A. Alkhateeb, and R. Heath, “Coverage and capacity of millimeter-wave cellular networks,” IEEE Communications Magazine, vol. 52, no. 9, pp.70–77, Sept. 2014.

[62] Y. Li, J. G. Andrews, F. Baccelli, T. D. Novlan, and C. J. Zhang, “Design andanalysis of initial access in millimeter wave cellular networks,” IEEE Transac-tions on Wireless Communications, vol. 16, no. 10, pp. 6409–6425, Oct. 2017.

[63] H. Inaltekin, M. Chiang, H. V. Poor, and S. B. Wicker, “On unbounded path-loss models: effects of singularity on wireless network performance,” IEEEJournal on Selected Areas in Communications, vol. 27, no. 7, pp. 1078–1092,September 2009.

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

[65] T. Bai, R. Vaze, and R. Heath, “Analysis of blockage effects on urban cellularnetworks,” IEEE Transactions on Wireless Communications, vol. 13, no. 9,pp. 5070–5083, Sept. 2014.

[66] T. Bai and R. Heath, “Coverage and rate analysis for millimeter wave cellularnetworks,” IEEE Transactions on Wireless Communications, vol. 14, no. 2,pp. 1100–1114, Feb. 2015.

[67] T. K. Y. Lo, “Maximum ratio transmission,” IEEE Transactions on Commu-nications, vol. 47, no. 10, pp. 1458–1461, Oct. 1999.

[68] A. M. Hunter, J. G. Andrews, and S. Weber, “Transmission capacity of ad hocnetworks with spatial diversity,” IEEE Transactions on Wireless Communica-tions, vol. 7, no. 12, pp. 5058–5071, December 2008.

[69] H. Shokri-Ghadikolaei, C. Fischione, G. Fodor, P. Popovski, and M. Zorzi,“Millimeter wave cellular networks: A MAC layer perspective,” IEEE Trans-actions on Communications, vol. 63, no. 10, pp. 3437–3458, Oct. 2015.

[70] M. Akdeniz, Y. Liu, M. Samimi, S. Sun, S. Rangan, T. Rappaport, andE. Erkip, “Millimeter wave channel modeling and cellular capacity evaluation,”IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1164–1179, Jun. 2014.

Bibliography 59

[71] M. Giordani, M. Polese, A. Roy, D. Castor, and M. Zorzi, “A tutorial onbeam management for 3GPP NR at mmWave frequencies,” arXiv preprint,arXiv:1804.01908.

[72] METIS deliverable D1.1, “Scenarios, requirements and KPIs for 5G mobileand wireless system,” April 2013.

[73] V. Garcia, Y. Zhou, and J. Shi, “Coordinated multipoint transmission in densecellular networks with user-centric adaptive clustering,” IEEE Transactions onWireless Communications, vol. 13, no. 8, pp. 4297–4308, Aug 2014.

[74] D. Wang, Y. Wang, R. Sun, and X. Zhang, “Robust c-ran precoder design forwireless fronthaul with imperfect channel state information,” in Proc. IEEEWireless Communications and Networking Conf. (WCNC), Mar. 2017, pp. 1–6.

performanceanalysisanddesigninsights1264028/fulltext01.pdf · samarbete i två udn-scenarier:...

Documents