Small cell networks and massive MIMO for radio-overfiber based indoor communications Wang, Q.

Transcription

1 Small cell networks and massive MIMO for radio-overfiber based indoor communications Wang, Q. Published: 18/05/2016 Document Version Publisher s PDF, also known as Version of Record (includes final page, issue and volume numbers) Please check the document version of this publication: A submitted manuscript is the author's version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website. The final author version and the galley proof are versions of the publication after peer review. The final published version features the final layout of the paper including the volume, issue and page numbers. Link to publication Citation for published version (APA): Wang, Q. (2016). Small cell networks and massive MIMO for radio-over-fiber based indoor communications Eindhoven: Technische Universiteit Eindhoven General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. Users may download and print one copy of any publication from the public portal for the purpose of private study or research. You may not further distribute the material or use it for any profit-making activity or commercial gain You may freely distribute the URL identifying the publication in the public portal? Take down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. Download date: 23. Dec. 2017

2 Small Cell Networks and Massive MIMO for Radio-over-Fiber Based Indoor Communications PROEFSCHRIFT ter verkrijging van de graad van doctor aan de Technische Universiteit Eindhoven, op gezag van de rector magnificus prof.dr.ir. F.P.T. Baaijens, voor een commissie aangewezen door het College voor Promoties, in het openbaar te verdedigen op woensdag 18 mei 2016 om 16:00 uur door Qing Wang geboren te Shaanxi, China

3 Dit proefschrift is goedgekeurd door de promotoren en de samenstelling van de promotiecommissie is als volgt: voorzitter: prof.dr.ir. A.B. Smolders 1 e promotor: prof.dr.ir. S. Heemstra de Groot 2 e promotor: prof.dr.ir. I.G.M.M. Niemegeers (TUD) leden: prof.dr.ir. A. Lo (Huwawei Technologies R&D Sweden) prof.dr. L. Munoz (University of Cantabria) prof.dr.ir. P.G.M. Baltus prof.ir. A.M.J. Koonen adviseur(s): dr. S. Pollin (KU Leuven) Het onderzoek of ontwerp dat in dit proefschrift wordt beschreven is uitgevoerd in overeenstemming met de TU/e Gedragscode Wetenschapsbeoefening.

4 A catalogue record is available from the Eindhoven University of Technology Library. Title: Small Cell Networks and Massive MIMO for Radio-over-Fiber Based Indoor Communications Author: Qing Wang Eindhoven University of Technology, 2016 ISBN: NUR 959 Keywords: small cell networks / massive MIMO / radio-over-fiber / WLAN Copyright 2016 by Qing Wang All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means without the prior written consent of the author.

5

6 I don't know anything, but I do know that everything is interesting if you go into it deeply enough. by Richard Feynman

7

8 Abstract Wireless technologies and applications have experienced rapid developments over the past years. Emerging applications like cloud storage and computing, virtual reality, and internet of things, have shown great prospects that excite the whole mobile industry. These new applications impose extreme requirements, namely ultra-high data rates, ultralow latency and response times, and very low energy consumption. These requirements drive the design of the next generation wireless networks, commonly referred to as 5G. Small cell networks (SCN) and massive MIMO are both promising techniques for satisfying such demands. SCN refers to a high-density cellular network where the cell radius is small, e.g., around 10 m, which improves the coverage and facilitates higher spatial reuse of the available spectrum. Massive MIMO scales up traditional MIMO systems to a much higher order to improve the spectral efficiency. This thesis investigates ways to significantly improve the indoor wireless network capacity using SCN and massive MIMO based on a radio-over-fiber (RoF) architecture. Not only do RoF systems improve the scalability, flexibility, and efficiency of the wireless networks, but they also fulfil the high requirements of SCN and massive MIMO. For SCN, two problems are addressed. One is the hidden node (HN) problem that occurs in the uplink media access in distributed antenna systems (DAS) when using the IEEE CSMA/CA (carrier sense multiple access with collision avoidance) protocol. We propose the hard- and soft-switching mechanisms to alleviate this problem by avoiding packet collisions, while not requiring changes to the IEEE standard. The other is the human shadowing problem in the 60 GHz band. 60 GHz communication is attractive because of the wide spectrum available and the huge bandwidth of the channels compared to the microwave versions of IEEE However, the signals are susceptible to blocking. Thus, maintaining continuous connectivity is an issue. We propose to use DAS to mitigate the shadowing problem. We address the problems of radio access unit (RAU) positioning, beamforming strategy (selective and blanket transmission), and trade-offs between densifying RAUs and increasing antenna array sizes. We find that selective transmission performs slightly poorer than blanket transmission but has lower implementation complexity. To provide good connectivity, at least two RAUs per room are needed. In addition, higher RAU density leads to better connectivity and also reduce the total number of antennas required. For massive MIMO, we focus on three problems. First, we investigate the practical performance of massive MIMO in indoor environments, which is based on the analysis of MIMO channel measurements. The results show that the performance for real channels is

9 poorer than a theoretical approach predicts, implying that more antennas are needed than in a theoretical approach to achieve good performance. Second, we perform a comparative evaluation of centralized (CAS) and distributed antenna system (DAS) architectures. We investigate their maximum achievable capacity as well as their improvement over traditional MIMO configurations. It turns out that DAS is substantially better. For example, assume a traditional MIMO system with 8 antennas and a massive MIMO system with 64 antennas, by applying zero-forcing precoding, the CAS architecture provides around twice the capacity of traditional MIMO but a DAS with 16 RAUs provides around 12 times the capacity gain. The third problem is the overhead caused by the feedback of channel state information. We particularly focus on a DAS architecture for massive MIMO. By utilizing the fact that the path loss from a given user to different RAUs may vary, we propose feedback bit allocation algorithms that utilize the feedback bits more efficiently and therefore successfully reduce the feedback overhead. For achieving 90% of the data rate provided by perfect channel knowledge, the feedback reduction is more than 20% in comparison with the IEEE n/ac scheme. The performance of SCN and distributed massive MIMO is compared in the 2.4 GHz and 60 GHz bands, which are both used for indoor communications. We find that distributed massive MIMO provides a much higher capacity than SCN in the 2.4 GHz band. For example, with a typical configuration, the capacity of distributed massive MIMO is more than twice that of SCN, and this is even better when more users are supported simultaneously. However, in the 60 GHz band, their performance is similar. We conclude that, since SCN is easier to implement than distributed massive MIMO, it is more suitable for the 60 GHz band than distributed massive MIMO. Finally, we also point out the remaining issues regarding the implementation of SCN and massive MIMO, as well as the interesting directions we foresee based on our studies.

10 Samenvatting De laatste jaren is er een enorme ontwikkeling geweest van draadloze netwerken en toepassingen hiervan. Opslag in de cloud, applicaties in de cloud, virtual reality en het internet-of-things zijn voorbeelden waar draadloze technologie een bijzondere rol speelt en die de hele mobiele communicatieindustrie tot bloei brengen en perspectieven bieden. Deze nieuwe toepassingen stellen bijzonder hoge eisen aan de draadloze netwerken, met name zeer hoge data snelheden, zeer kleine vertragingstijden en een zeer laag energieverbruik zijn nodig. Deze eisen hebben geleid tot de ontwikkeling van een nieuwe generatie draadloze netwerken, bekend onder de naam 5G. Twee veelbelovende technieken om aan deze eisen te voldoen zijn Small Cell Networks (SCN) and massive MIMO. SCNs zijn cellulaire netwerken met een hoge dichtheid van kleine cellen, typisch met afmetingen van de orde van 10 m, die erop gericht zijn de netwerk dekkingsgraad te verbeteren en het ruimtelijk hergebruik van het radiospectrum te verhogen. Massive MIMO is een evolutie van de MIMO technologie waarbij het aantal antennes zeer groot wordt, typisch honderd of meer, en die de efficientie van het spectrum gebruik sprongachtig verbetert. Het gebruik van SCN en massive MIMO om de capaciteit van WLANs in gebouwen drastisch te verbeteren is het onderwerp van dit proefschrift. Hierbij baseren we ons op een Radio-over-Fiber (RoF) architectuur om de Access Point in een centrale locatie te verbinden met de Radio Access Units (RAUs), d.w.z. de antenne locaties in de verschillende ruimtes van een gebouw. De keuze voor RoF is ingegeven door het gemak waarmee RoF de schaalbaarheid, de flexibiliteit en de efficientie van SCN en massive MIMO ondersteunt. In dit proefschrift onderzoeken we in eerste instantie hoe twee problemen die optreden bij het gebruik van SCN kunnen opgelost worden. Het eerste probleem is het hidden-node (HN) probleem, dat zich voordoet wanneer in een Distributed Antenne Systeem (DAS) het IEEE CSMA/CA MAC protocol toegepast wordt. We hebben twee oplossingen ontwikkeld, hard-switching en soft-switching, die ervoor zorgen dat botsingen tussen frames komende uit twee verschillende cellen vermeden worden. Onze oplossingen zijn compatibel met de IEEE standaard. Het tweede probleem is de human-shadowing, d.w.z. het blokkeren van een radiosignaal door de aanwezigheid van personen, wanneer gebruik gemaakt wordt van de 60 GHz mmgolf band. Deze frequentieband is attractief en nodig om te voldoen aan de vraag naar bandbreedte die niet gedekt wordt door de IEEE microgolf banden. Om een continue connectiviteit te garanderen hebben we een oplossing ontwikkeld die gebruik maakt van een DAS architectuur voor het netwerk. De vragen die hierbij rijzen zijn: hoeveel gedistribueerde RAUs

11 hebben we nodig, en hoe doen we beamforming. Verder moest er gekeken worden naar de trade-off tussen het aantal RAUs en het aantal antennes per RAU. Wat de beamforming betreft hebben we twee strategieën ontwikkeld: selective transmission en blanket transmission, waarbij de tweede beter presteert maar de eerste eenvoudiger is om te implemeteren. We vonden o.a. dat, om een continue connectiviteit te garanderen, tenminste twee RAUs per kamer nodig zijn en dat hoe dichter het netwerk van RAUs is, hoe beter de connectiviteit gegarandeerd wordt en hoe kleiner het totale aantal antennes mag zijn. Voor massive MIMO zijn er drie onderwerpen aan de orde. De eerste vraag is hoe massive MIMO zich gedraagt in een gebouw. We doen dit op basis van metingen van de MIMO kanalen. We vonden dat de prestaties duidelijker slechter zijn dan wat verwacht kan worden wanneer we werken met theoretische modellen. Dit leidt ertoe dat meer antennes nodig zijn dan in een theoretische benadering. Het tweede onderwerp is een vergelijkende studie van de prestaties van DAS en een Centralized Antenna Systeem (CAS). We onderzochten hierbij de maximale capaciteit en de verbetering ten opzichte van een traditioneel MIMO systeem. Het blijkt dat een DAS architectuur beduidend beter is. Voor een traditioneel MIMO systeem met 8 antennes en een massive MIMO systeem met 64 antennes en zero-forcing precoding, ontdekten we bijvoorbeeld dat met een CAS architectuur, massive MIMO tweemaal zoveel capaciteit heeft als traditioneel MIMO. Bij een DAS architectuur daarentegen, wordt dit 12 maal meer. Het derde onderwerp is de grote overhead die veroorzaakt kan worden door de noodzakelijk feedback van de Channel State Information in massive MIMO. We richten ons hierbij op de DAS architectuur en ontwikkelden een feedback-bit allocation algoritme, waarbij de overhead drastisch gereduceerd kan worden door gebruik te maken van het feit dat de path loss tussen een gebruiker en de RAUs sterke verschillen vertoont. We slagen erin bij een reductie van 20% van de overhead, zoals die voorkomt bij gebruik van IEEE n/ac, toch nog 90% van de capaciteit van een systeem met perfecte kanaalinformatie te halen. Verder hebben we de capaciteit van SCN en massive MIMO onderzocht en vergeleken voor 2.4 GHz en 60 GHz. We komen to de constatatie dat in een typisch scenario, voor 2.4 GHz, DAS massive MIMO veel beter presteert dan SCN, wel meer dan tweemaal zoveel en nog meer wanneer het systeem simultane transmissies van meerdere gebruikers ondersteunt. In de 60 GHz band ligt het anders. Daar zijn de prestaties vergelijkbaar en is SCN te verkiezen wegens de eenvoudiger implementatie. Tenslotte geven we ook een overzicht van wat er verder onderzocht moet worden om, op basis van onze bevindingen, de toepassing van SCN en MIMO in combinatie met ROF succesvol te maken.

12 Contents Chapter 1 Introduction WLAN and WiFi Small Cell Networks and Massive MIMO Radio-over-Fiber Networks Analog and Digitized RoF Architecture of A-RoF Links RoF-based WLAN Configurations of RoF-based SCN and massive MIMO Centralized and Distributed Antenna Systems Configurations of SCN and Massive MIMO Research Problems Contributions and Organization of the Thesis Chapter 2 Switching Mechanisms for Mitigating Hidden Node Problem in DAS Introduction Hidden Node Problem in DAS Solutions to the HN Problem Switching Mechanisms for Avoiding Inter-cell Collisions Hard Switching Soft Switching Downlink Traffic in Hard-Switching and Soft-Switching Comparative Performance Analysis Impact of the Number of STAs Impact of the Number of Cells Conclusions and Future Work Chapter 3 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs Introduction Shadowing Mitigation Techniques Distributed Antenna System for Shadowing Mitigation GHz DAS and Beamforming... 31

13 3.4.1 Optimizing the Placement of the RAUs Beamforming Strategies Channel Simulation Performance Evaluation Optimized RAU Positions Blanket and Selective Transmission Strategies More RAUs or Larger Antenna Array? Conclusions and Future Work Chapter 4 Massive MIMO in Indoor WLANs: Background and Insights Introduction Massive MIMO What is Massive MIMO? Potential Challenges Basics of Massive MIMO MU-MIMO Benefit of Large Antenna Arrays A Review of the Analysis of Massive MIMO Impact of the Propagation Channel Channel Measurements Capacity Analysis Impact of Channel Conditions Impact of Antenna Array Size on Sum-rates of the Precoding Techniques Conclusions and Future Work Chapter 5 Massive MIMO with CAS and DAS Architectures Introduction Literature Review Objectives and Methodologies System Model Signal Model Sum-rate Capacity Channel Models Ray-tracing Simulation Hybrid Channel Model Performance Analysis Massive MIMO versus Traditional MIMO... 68

14 5.6.2 Centralized versus Distributed Massive MIMO Maximum Average Sum-rate of Centralized and Distributed Massive MIMO Maximum Average Sum-rate versus the Total Number of Antennas Conclusions and Future Work Chapter 6 Optimizing CSI Feedback for Distributed Massive MIMO Systems Introduction Literature on CSI Compression Techniques for Massive MIMO System Model CSI Feedback Scheme and Feedback Bit Allocation Problem IEEE n/ac Channel Training Feedback Bit Allocation Problem Feedback Bit Allocation Adaptive allocation Equal allocation Sum-rate Model with Imperfect CSI Performance Analysis Conclusions and Future Work Chapter 7 Performance Comparison of Small Cell Networks and Distributed Massive MIMO Introduction Literature System Model Channel Models GHz GHz Unified Sum-rate Model Beamforming Transmission Power Optimization Sum-rate Analysis GHz GHz Conclusions and Future Work Chapter 8 Conclusions and Future Work Conclusions Future Work

15 Appendix A. Ray-tracing Channel Simulation 121 A.1 The Procedures of Ray-tracing Channel Simulation A.2 Extension from SISO Ray-tracing Channel to Arbitrary Antenna Array A.2.1 Ray-tracing Channel Model A.2.2 Construct Channel for Arbitrary Antenna Array A.3 Examples Bibliography 129 Acronyms 141 Notation 143 Publications 147 Acknowledgements 151 Curriculum Vitae 153

16 1 Chapter 1 Introduction Over the past decades, wireless technologies and applications have experienced great developments. Billions of devices, like cell phones, tablets, laptops and sensors, are now connected to the internet, and the number continues to grow. At the same time, more and more applications such as cloud storage and computing, virtual reality, and internet-ofthings (IoT), are emerging [1] [2]. The growth of wireless devices and applications demand higher and higher network capacity, and pose more stringent requirements on the latency and reliability of the wireless connections [1]. More advanced wireless technologies are needed to fulfill these challenging requirements. Since people spend most of their time indoors, e.g., in offices, homes, shopping malls and airport terminals, most of the mobile traffic originates and terminates indoors. It is reported in [3] that the indoor traffic accounts for as much as 70 to 80% of all mobile traffic in the US in This highlights the importance of indoor wireless networks. Currently, WiFi (Wireless Fidelity) and LTE (Long Term Evolution), are the two dominant broadband wireless access technologies. WiFi is based on the IEEE standards, which is the most popular wireless access technology for WLANs due to the low cost of the devices and the infrastructure. LTE, commonly known as 4G, is a standard designed for large commercial cellular networks. Although both WiFi and LTE are used for indoor communications, WiFi is more widely deployed nowadays. In fact, most mobile traffic goes through WiFi. For example, in the US in 2014, the WiFi traffic volume was 3 times that of the cellular network [3]. The research in this thesis is mainly based on the framework of the IEEE standards. In this thesis, we focus on improving the indoor wireless network capacity. The two major techniques considered are small cell networks (SCN) and massive MIMO [4]. We integrate these two techniques with radio-over-fiber (RoF) systems. On the one hand, RoF systems improve the scalability, flexibility, and efficiency of the wireless networks, making it an attractive solution by itself [5]. On the other hand, SCN and massive MIMO impose high requirements on the network infrastructure, making RoF a necessary choice. In the remainder of this chapter, we first review the WiFi technologies in Section 1.1. In Section 1.2, we introduce SCN and massive MIMO, and motivate the utilization of RoF systems. Section 1.3 explains the structure of RoF systems for indoor wireless networks. Section 1.4 presents the different configurations of RoF systems for SCN and massive MIMO. We summarize the research problems investigated in this thesis in Section 1.5, and our contributions in Section 1.6.

17 2 Introduction 1.1 WLAN and WiFi WiFi or the IEEE standard family is a set of media access control (MAC) and physical layer (PHY) specifications for implementing wireless communications in the 2.4, 5, 60 GHz and sub-ghz unlicensed frequency bands, for WLAN communications. The standard has gone through several amendments to integrate new technologies. An overview of the evolution of WiFi is given in Table 1.1, summarizing the changes including the frequency bands, the new key techniques, and the maximum data rates [6] [7] [8]. Table 1.1 IEEE standards. Year Version Frequency band (GHz) New technologies Max PHY data rate (Mbps) b 2.4 CCK a 5 OFDM g 2.4 OFDM n 2.4, 5 SU-MIMO, 4 SS ac 5 MU-MIMO, 8 SS, channel 6933 bonding ad 60 SU-Beamforming * ah Below 1 GHz 347 OFDM: Orthogonal Frequency-Division Multiplexing CCK: Complementary Code Keying SU: Single User MU: Multiple User MIMO: Multiple-Input Multiple-Output SS: Spatial Stream * The standard is expected to be finalized and arrive in The target of the standard is to support lower data rates, longer range and low power communications. IEEE ac and IEEE ad are the most recent improvements. Their two major features are wider spectrum and MIMO spatial multiplexing. Let us summarize the major amendments below: - IEEE ac introduces two new channel widths, 80 MHz and 160 MHz, in addition to 40 MHz supported in IEEE n and 20 MHz in the previous versions. In some countries, 160 MHz of contiguous spectrum is not available, so ac introduces two forms of 160 MHz channels: a single 160 MHz block, and an MHz channel that combines two non-contiguous 80 MHz channels [9].

18 1.2 Small Cell Networks and Massive MIMO 3 - IEEE ad employs the 60 GHz unlicensed band, in which 5 GHz of contiguous spectrum is globally available, and up to 9 GHz in Europe [6]. This suggests that the data rates of gigabits per second can be achieved even with low spectral efficiency modulation schemes. A major problem of 60 GHz signal is the high path loss, thus beamforming is needed to enhance the received signal power. - IEEE ac specifies up to eight spatial streams, compared to IEEE n s four spatial streams at the access point (AP). The extra spatial streams can be used to transmit data to multiple users at the same time (MU-MIMO). With such an ability, IEEE ac speeds up networks even more than SU- MIMO in IEEE n [9]. Presently, IEEE ax and IEEE ay are under development. They aim to reach not only higher data rates but also better energy efficiency. Specifically, IEEE ax targets to improve the per user throughput by 5 to 10 times (depending on technology and scenario), especially for dense deployment scenarios with frequency bands between 1 GHz and 6 GHz [10]. IEEE ay will extend the IEEE ad standard to reach a maximum throughput of at least 20 Gbps, while maintaining or improving the power efficiency per mobile station (STA), and extending the coverage range [11]. 1.2 Small Cell Networks and Massive MIMO Small cell network (SCN) and massive MIMO have been recognized as promising technologies for fulfilling the objectives of future high data rate wireless networks [4]. They are both examined in this thesis. Let us now give a brief introduction to them. Fig. 1.1 illustrates the differences between SCN, massive MIMO and traditional WLANs. AP Ominidirectional transmission STA Narrow beams (a) Traditional WLAN cell (b) Small cell network (c) Massive MIMO Fig. 1.1 Illustration of traditional WLAN, SCN and Massive MIMO.

19 4 Introduction SCN corresponds to a high-density cellular network where the cell radius is small, e.g., only a few meters [12]. Traditional WLAN APs typically have a coverage radius of tens of meters. However, large cell sizes may result in a high likelihood of poor coverage and low data rates, particularly at the cell boundaries. Especially for indoor environments, the walls and other large obstacles may strongly attenuate the radio signals, deteriorating the user experience of the WLAN service. By reducing the cell size, a number of benefits can be obtained. First, each AP covers a small area, so the APs are close to the users resulting in low path loss. As a result, higher data rates can be achieved for a given transmission power, or conversely, the transmission power can be lowered to achieve the same data rates as in traditional WLANs. Second, when the cell size is small, the number of users per cell is smaller, thus each STA can get more radio resources such as spectrum and time slots. Third, the APs can achieve a higher spatial reuse. Especially the low transmission power may help to reduce inter-cell interference. Higher spatial reuse means more data can be sent at the same time using the same spectrum in different small cells, so a higher network capacity can be achieved. Finally, SCN is necessary for 60 GHz since the signal is prone to high path loss and obstructions while the transmission power is low due to the circuit limitations [13] [14]. Massive MIMO is an extension of traditional MIMO systems by using many more antennas and supporting higher order spatial multiplexing. In IEEE n/ac and LTE, which we refer to as traditional MIMO systems, only up to 8 antennas are supported at the APs or base stations (BSs). In massive MIMO, the number of AP antennas is not strictly limited, and is usually assumed to be around several tens or hundreds [15]. Large antenna arrays enable very directional transmissions or receptions at the APs, i.e., the AP focuses narrow beams onto the STAs it is communicating with. The sharp beamforming can reduce both the inter-cell and intra-cell interference, and increases the energy efficiency with high beamforming gain [16]. In contrast, traditional WLAN and SCN APs use a small number of antennas, in particular for the low frequencies (2.4 and 5 GHz), thus the radio signals are broadcast in all directions, causing interference with adjacent cells. However, both SCN and massive MIMO come with some drawbacks that we summarize below: - SCN: o o The large number of APs and their wired connections lead to high installation and maintenance costs. The potential increase of the traffic of individual APs as well as the aggregate traffic load will demand higher capacity of the data distribution network which transports data between the APs and the Internet. At the same time, the larger number of APs makes the packet distribution network rather complex.

20 1.3 Radio-over-Fiber Networks 5 o The control and management of the network also become more complex. Small cells cause more frequent handover between the cells. This imposes processing loads and delays on the control and management plane to provide seamless connectivity. o The closer distance between the cells may cause strong inter-cell interference between neighboring cells. To reduce or avoid this, it requires coordination and transmission power control across the APs, and possibly a dynamic assignment of frequency bands to the APs. - Massive MIMO: o A high capacity data distribution network is required due to the high aggregate traffic load from the many simultaneously served STAs. o High computational complexity: the signal processing (precoding, channel estimation, symbol detection, encoding/decoding, etc.), and, the control and management of many STAs, require the AP to be capable of handling a high computational load [17]. As a result, the power demands and thermal dissipation of the APs will also be high. o The physical size of the AP will be large due to the large antenna arrays, e.g., a 2.4 GHz square antenna array of 64 antennas is around 0.5 m by 0.5 m (assuming half-wavelength spacing between the antennas). To obtain better channel conditions, the antenna spacing needs to be larger than half-wavelength, making the physical size of the antenna arrays even bigger [18]. Very large antenna arrays may be difficult to fit in indoor scenarios. These challenges can be effectively dealt with by radio-over-fiber (RoF) technology. 1.3 Radio-over-Fiber Networks The network architectures of traditional WLAN [19], and RoF-based indoor WLAN are illustrated in Fig. 1.2 and Fig. 1.3, respectively. The architecture of the RoF network is developed by of the Dutch IOP GenCom MEANS project [20]. In a traditional WLAN, the router distributes data packets to the APs, and collect uplink data frames from the APs, through the Ethernet. The APs use the IEEE wireless access interfaces to communicate with the STAs. The APs have both PHY and MAC layer functions, including the transmission, reception, and processing of the radio signals, as well as the media access protocols [19].

21 6 Introduction Floor Building AP External network /Internet Router Ethernet cables Fig. 1.2 Architecture of traditional WLAN. Floor Building RAU External network /Internet Router Central Station Optical fibers Fig. 1.3 Architecture of RoF-based WLAN. RoF refers to a technology whereby a radio signal is modulated onto light and transmitted over an optical fiber link to facilitate wireless access [21] [22]. In short, radio signals are carried over optical fibers. As a consequence, the computational resources that are located at all APs can now be centralized into the Central Station (CS). At the same time, optical fibers are used to deliver radio signals between the CS and the antennas. The conversion of the signals between the optical and the electrical domains is performed at both the CS and the antennas. This function is achieved by the radio-optical interfaces. We denote Radio Access Unit (RAU) as the unique component that integrates the radio-optical interfaces and the antennas. The functions of the RAUs are very simple, i.e., just electro/optical (E/O) conversion. In summary, the radio-optical interface and the optical fiber network play the role of radio signal transportation in both directions between the CS and the remote antennas.

22 1.3 Radio-over-Fiber Networks Analog and Digitized RoF Depending on how the radio signals are transported in the optical fibers, RoF systems can be categorized into analog and digitized RoF (A- and D-RoF) [23]. - In A-RoF, the radio signals are directly modulated onto optical carriers, e.g., by modulating a DFB laser (Distributed Feedback Laser) with the radio signals. - In D-RoF, the radio signals are first quantized through an analog-to-digital converter (ADC), encapsulated in packets, and then transported through the optical network with dedicated communication protocols. For example, the Common Public Radio Interface (CPRI) is a widely used standard for D-RoF [24]. The advantages and disadvantages of A-RoF and D-RoF are summarized in Table 1.2. The most suitable choice depends on practical considerations such as cost and performance requirements. But currently, mobile operators are in favor of D-RoF [23], because the RoF interface is standardized, e.g., CPRI. Nevertheless, A-RoF is still very attractive for indoor communications from the economical perspective [23] [22]. For that reason, the work in this thesis focuses on A-RoF. Table 1.2 Comparison of A-RoF and D-RoF. Advantages A-RoF Transparent to radio signals Lower backhaul capacity demand Lower costs D-RoF Standardized optical transportation protocol (e.g., CPRI) No radio signal impairments Disadvantages Radio signals may be impaired due to optical channel distortion Not standardized High backhaul capacity demands Architecture of A-RoF Links In Fig. 1.4, we present the generic architecture of an A-RoF link (downlink) [23] [25]. The uplink is similar. The radio signal is power amplified, and then it directly modulates a DFB laser. Upon reception at the RAU, the analog modulated optical carrier is photodetected using a PIN diode implemented with a TIA. The recovered radio signal is then amplified and radiated into the air. The transportation of radio signals in the optical fiber channel is independent of the modulation and coding formats of the radio signals, i.e., an A-RoF link is transparent to the radio signals.

23 8 Introduction RAU RF signal PA Analog optical transmitter (DFB laser) Analog fiber link Analog optical receiver (PIN+TIA) RF BPF PA PA: Power Amplifier TIA: Trans-Impedance Amplifier BPF: Band-Pass Filters Fig. 1.4 Generic architecture of A-RoF link. Although the radio signals may be distorted in the optical fiber due to the non-linearity of the optical channels, the distortion can be significantly mitigated by pre-distortion techniques [26]. Thus, we can assume the optical channels are ideal. The antennas can be linked to the CS via several possible optical network topologies: tree, star, bus, ring, etc. [23] [27] [28]. It is shown in [27] that a multilevel-ring topology has better reliability than others, and ring-stars topology has the best scalability. In this thesis, the optical network topology is not a subject of investigation, and we will further assume a star topology for the sake of simplicity RoF-based WLAN The architecture of an RoF-based WLAN is illustrated in Fig The CS hosts a number of APs, which are connected to the RAUs through an arbitrary optical network topology (represented by a cloud). The association between the APs and the RAUs can be fixed or dynamic. Each RAU has an antenna array, or a single antenna. The placement of the RAUs can be done in the same way as radio planning in traditional WLANs [29]. External network /Internet Central Station AP AP AP AP Radio-Optical Interface Optical link Optical Network RAU RAU RAU Wireless channel STA STA Fig. 1.5 Architecture of RoF-based WLAN. Let us now introduce the benefits of RoF systems: - Since the RAUs only perform simple signal conversions, they can in principle be inexpensive such that high-density deployments should be economically feasible.

24 1.4 Configurations of RoF-based SCN and massive MIMO 9 - Unlike APs in traditional WLAN, the CS has less restrictions regarding form factor, heat dissipation, etc., thus multiple powerful processors can be installed in the CS. - Optical fibers provide tremendous capacity and have low attenuation over distance, especially compared with copper cables. Thus optical fibers are very suitable for the signal or data distributions. In addition, an optical fiber can carry multiple wideband radio signals, which reduces the number of wires, lowering the installation complexity. - RoF links are transparent to radio signals, so there is no need to change the RAUs if there are upgrades of radio technologies. In addition, centralization of the processing components in the CS also makes upgrades easier. Thus the RoF network is expected to be cost-effective in the long-term. However in this thesis, we will not address the economics of the solutions we propose. We conclude that RoF facilitates SCN and massive MIMO. 1.4 Configurations of RoF-based SCN and massive MIMO Centralized and Distributed Antenna Systems There are mainly two antenna system architectures, namely the centralized antenna system (CAS) and the distributed antenna systems (DAS). In CAS, all the antennas are colocated, and typically the spacing between the antennas is of the order of the carrier frequency wavelength. In DAS, in contrast, the antennas are divided into multiple smaller arrays, and these arrays are separated far apart, usually at distances on the order of meters. The RoF system we consider is essentially a DAS. DAS was proposed to overcome the coverage problem in indoor environments due to the high attenuation of walls [30]. Initially, coax cables were used to connect the distributed antennas to the APs. Later with the development of RoF techniques, optical fiber took the place of coax cables [31], allowing radio signals to be transported over large distances while maintaining high signal quality. In addition, DAS was at first only used as a MISO system, not supporting spatial multiplexing [32]. Single-user and multi-user MIMO spatial multiplexing (SU- and MU-MIMO) are proposed for DAS in [33]. The general advantages of DAS over CAS are twofold: - The received signal strength is improved due to the smaller average path loss, as the signals travel between the multiple RAUs surrounding the STAs with short propagation distances, which we refer to as macro-diversity. - The large spatial separation between the small arrays have independent spatial paths, i.e., better spatial diversity. This channel independency is especially beneficial for MIMO spatial multiplexing, which turns into a higher capacity gain. We term this type of diversity as micro-diversity.

25 10 Introduction Therefore, in principle, it should be beneficial to use DAS instead of CAS Configurations of SCN and Massive MIMO The architecture of RoF systems for SCN and massive MIMO are intrinsically the same, as shown in Fig However, there is much flexibility in the CS. The CS can be configured to form different systems, depending on how the RAUs are utilized and how the radio signals are processed at the CS. In this thesis, we will consider three configurations. We illustrate them in Fig (a) The SCN case assumes that each RAU is associated with one AP in the CS. Each small cell is served by an AP. (b) The SCN-DAS case assumes a smaller number of APs than in the SCN case, and each AP is associated with a small number of RAUs forming a DAS. The DASs can improve coverage, and can also be used as distributed MIMO for spatial multiplexing. (c) The distributed massive MIMO case assumes a very large number of RAUs coordinated by a single AP. Multiple APs can be placed in a CS for very large areas, i.e., multiple distributed massive MIMO systems. It is not possible to determine which configuration is the best in general terms. The choice should be made depending on the wireless communication standards for MAC and PHY, and traffic load. We give more specific explanations in the following. SCN uses a large number of APs, which creates a lot of redundancy. However, this may cause handovers to occur very often due to the small cell sizes. Actually, if the traffic load is relative low, that many APs may not be necessary. Therefore, the SCN-DAS configuration is more cost-effective since a smaller number of APs can be used. Depending on the wireless standard employed, the SCN-DAS configuration may utilize the DASs differently. For example, IEEE a/b/g only support SISO, so the DAS is mainly used for improving coverage taking advantage of spatial diversity [34]. IEEE n allows SU-MIMO, thus the DAS can be used to perform MIMO spatial multiplexing. But for SU-MIMO/SISO, it has been shown that using one of the RAUs is more beneficial since it reduces the intercell interference [35]. IEEE ac allows MU-MIMO, and the best performance is achieved by activating all the RAUs in the DASs [35]. The distributed massive MIMO configuration is the most complex form as the AP uses many RAUs in a coordinated fashion to form a massive MIMO system, which we will discuss more in the later chapters. In principle, this high-degree of cooperation may alleviate the inter-cell interference problem in SCN. However, it is not known yet whether it will perform better than the other configurations. We will discuss that in Chapter 7.

26 1.4 Configurations of RoF-based SCN and massive MIMO 11 RAU AP AP 2 AP AP (a) SCN AP AP AP (b) SCN-DAS AP (c) Distributed massive MIMO Fig. 1.6 Configurations of RoF system for SCN and massive MIMO.

27 12 Introduction 1.5 Research Problems To facilitate the realization of SCN and massive MIMO with RoF, a number of problems need to be solved. In this thesis, we will investigate some of them. The IEEE standards can be directly applied in SCN. Especially, it is possible to implement the RoF-based SCN using the commercially available IEEE chipsets. However, the IEEE standards did not take into account techniques like RoF and DAS. Therefore, some issues may occur. Moreover, RoF systems insert an optical medium into the IEEE PHY layer. It is not clear whether or to what extent the DAS architecture and the optical network affect the mechanisms defined in the IEEE GHz communication is attractive because of the wide spectrum available and the huge bandwidth of the channels compared to the microwave versions of IEEE However, the signals are susceptible to blockings. Thus maintaining continuous connectivity is an issue. The IEEE ad standard has taken into account the blockage problem and proposed adaptive beamforming as a solution. However, the specific mechanisms for beamforming, the antenna system architecture, etc., have not been specified. Massive MIMO is a relatively new technique despite the fact that it can be interpreted as an extension of traditional MIMO systems. The research on massive MIMO is still going on, thus a number of questions are not yet answered conclusively. What is the practical performance? How much gain can massive MIMO offer in comparison with traditional MIMO? How to deal with the large form factors of the large antenna arrays? How to obtain the channel knowledge for beamforming? How to reduce signal processing complexity and signaling overheads introduced by a large number of antennas? Both SCN and massive MIMO can be implemented using the same RoF architecture. The question is which one is performing better and under what circumstances, given that there are multiple frequency bands available for IEEE , with their own distinct channel characteristics. To summarize, the specific problems we address in this thesis are the following. - How to overcome the performance degradation due to hidden node problem in SCN-DAS? - How to deal with the shadowing problem in 60 GHz? - What is the theoretical and practical performance of distributed massive MIMO? - How to optimize the channel state information feedback? - How does the performance of distributed massive MIMO compare with SCN in both 2.4 and 60 GHz bands?

28 1.6 Contributions and Organization of the Thesis Contributions and Organization of the Thesis The main objective of the thesis is to investigate ways to significantly increase the indoor wireless network capacity using SCN and massive MIMO, based on RoF architecture. Chapter 2 studies the application of IEEE in RoF-based SCN with DAS. We identify the hidden node (HN) problem in the uplink transmissions due to the failure of the CSMA/CA (carrier sense multiple access with collision avoidance) protocol. We propose two switching techniques that alleviate this problem significantly by avoiding packet collisions, while not requiring changes to the IEEE standard. Chapter 3 discusses 60 GHz indoor communications (IEEE ad), and concentrates on the shadowing problem caused by objects, in particular human bodies. 60 GHz signals are easily blocked by human bodies and objects, which is a severe problem in indoor environment. We propose to use DAS to mitigate the shadowing problem. We address the problems of RAU positioning, beamforming strategy, and trade-offs between densifying RAUs and increasing antenna array sizes. Chapter 4 presents the fundamentals of massive MIMO, and a first performance analysis in indoor environments to verify the potential of massive MIMO. The analysis is based on real channel measurements. We evaluate the performance of massive MIMO, and the behavior of different beamforming techniques in real propagation channels. Chapter 5 moves to the architectural perspective of massive MIMO, in particular, the CAS and DAS architectures. We present mathematical modeling and numerical results that answer a number of questions, in particular: (1) how much gain can be obtained by massive MIMO compared to traditional MIMO; (2) how much more capacity DAS provides than CAS; (3) what is the maximum capacity that CAS and DAS can provide, and how large the capacity differences are. Chapter 6 discusses the problem of channel state information (CSI) feedback for distributed massive MIMO, based on the framework of the IEEE n/ac standard. Accurate CSI is critical for attaining good performance in MIMO systems, but the CSI feedback overhead is very large for massive MIMO due to the large number of antennas, and therefore the many channels that need to be considered. We present solutions to feedback bit allocations that take advantage of the path loss differences between the distributed RAUs, to more efficiently utilize the CSI feedback bits. Chapter 7 compares the capacity of RoF-based SCN and distributed massive MIMO in both the microwave (2.4 GHz) and millimeter-wave (60 GHz) WiFi frequency bands. We present a unified model of the average sum-rate of the two configurations. Our findings show that, for 2.4 GHz, massive MIMO performs much better than SCN; but for 60 GHz, the difference is marginal.

29 14 Introduction Chapter 8 draws conclusions from our work and presents future research problems that ought to be solved to bring the technologies we focused on in this thesis closer to the market. Chapter 1 Introduction SCN Chapter 2 Switching Mechanisms for Mitigating Hidden Node Problem in DAS Chapter 3 Distributed Antenna System for Shadowing Mitigation in 60 GHz WLAN Chapter 7 Performance Comparison of Small Cell Networks and Distributed Massive MIMO Massive MIMO Chapter 4 Massive MIMO in Indoor WLANs: Background and Insights Chapter 5 Massive MIMO with CAS and DAS Architectures Chapter 6 Optimizing CSI Feedback for Distributed Massive MIMO Systems Chapter 8 Conclusions and Future Work Fig. 1.7 Structure of the thesis.

30 15 Chapter 2 Switching Mechanisms for Mitigating Hidden Node Problem in DAS 2.1 Introduction In this chapter, we consider applying the standard IEEE DCF MAC in SCN-DAS. In Chapter 1, we saw that DAS helps improve the received signal power, both at the AP and the STA, so as to enable higher data rates. Conversely, the transmission power can be reduced for a given data rate, which is especially beneficial to battery-powered STAs. However, due to the spatial distribution of the RAUs, there is a high likelihood that STAs cannot sense each other s transmissions, e.g., due to attenuations by walls or low transmission power, because they are out of the reception range of each other. This causes the IEEE DCF (Distributed Coordination Function) to fail in the uplink media access, which is termed the hidden node (HN) problem. The carrier sensing failure allows STAs to transmit simultaneously, causing collisions at the AP. As a result, the network throughput may be significantly degraded. Since IEEE DCF is widely implemented in practice (because of its simplicity [36]), this problem is important. In this chapter, we propose two switching mechanisms to alleviate this problem: hard-switching and softswitching. They are based on the observation that an AP simply combines the signals from all RAUs, while in fact the frames from the hidden STAs are received by different RAUs. The rest of the chapter is organized as follows. Section 2.4 introduces the HN problem in DAS. Section 2.3 reviews the existing solutions to the HN problem in DAS. The switching mechanisms we propose are described in Section 2.4, and evaluated by means of simulations in Section 2.5. Section 2.6 draws the conclusions. 2.2 Hidden Node Problem in DAS The DCF MAC is a carrier sense multiple access with collision avoidance (CSMA/CA) protocol. But, to achieve good performance with CSMA/CA, it is necessary for the STAs to be able to sense each other s transmissions. The carrier sensing is based on the examination of the received signal power of the medium. When the received signal power is above a threshold, an STA will infer that someone else is transmitting and will back off to avoid interference. Thus, if a pair of STAs cannot successfully detect each other s transmissions, they will consider the medium to be available and transmit at the same time, which causes a collision. This problem is called the HN problem [37], which is

31 16 Switching Mechanisms for Mitigating Hidden Node Problem in DAS illustrated in Fig Since STA B and C are out of each other s carrier sensing range, they transmit to STA A at the same time. The HN problem is especially severe when the uplink traffic load is high, resulting in throughput degradation. collision Frame B A C Fig. 2.1 Hidden node problem in WiFi. In a DAS WLAN, the HN problem may also happen for two reasons. - First, the STAs served by a particular AP are so far away that the distances between them are larger than the carrier sensing range. - Second, in case of small cells, the STAs use much less transmission power, which makes it even harder to sense the signals from other cells. In order to illustrate this, let us look at the example in Fig. 2.2, where an AP is used to serve several rooms. Room 1 and 2 are adjacent to each other, while 1 and 3 are separated by a large distance. For each room, four sample STA positions at the cell boundaries are given in the figure. Room 1 Room 2 Room 3 Room m 1 RAU m collision Access Point Hidden STA pair Frame path RoF link STA Fig. 2.2 An access point using a DAS with four RAUs.

32 2.3 Solutions to the HN Problem 17 Assume the IEEE n channel model B (2.4 GHz) and a typical 10 dbm transmission power [38]. We can easily calculate that the STAs in Room 1 are out of the carrier sensing range of the STAs in Room 4. So STA 4 and STA 9 are a pair of hidden STAs, and so are the more distant STAs. In addition, as low as -5 dbm transmission power is needed for an STA to transmit to its nearest RAU for the highest data rate (54 Mbps). If such transmission power is used, STA 4 will be hidden from STA 6, 7 and 8, and the same for farther STA pairs. 2.3 Solutions to the HN Problem The usual way to solve the HN problem in WiFi is the RTS/CTS (Request to Send/Clear to Send) mechanism. This mechanism uses an RTS frame to contend for the channel and then reserve the channel for data transmission, so as to avoid collision with the data frames of other STAs. RTS/CTS itself is based on the basic DCF. Thus the RTS/CTS mechanism also encounters the HN problem. But the cost of RTS frame collisions is smaller than for data frames since the RTS frames are much shorter. A study of RTS/CTS for uplink transmission in DAS has been presented in [39], which shows that RTS/CTS provides almost a constant throughput, while the basic DCF throughput degrades fast with the number of hidden STAs. Another choice is to use successive interference cancellation (SIC), a multi-user detection technique, to decode data from collided frames. This is proposed in [40] for the HN problem in DAS. However, as the numerical results indicate, there is a small chance of decoding the collided frames successfully. Specifically, with transmission power control, there is only a 20% probability of getting more than 10% capacity gain for a two-rau DAS. The reason is that the successful decoding requires the signal strength of the collided frames to be significantly different. This means that the distances between the STAs and the RAUs should be sufficiently different. However, the typical RAU-STA distance is only tens of meters or less in indoor networks, hence the path loss differences are not enough. In addition, the frames must be encoded with appropriate data rates that can allow SIC decoding. This is, however, not guaranteed since the STAs choose the data rate based on the interference-free assumption. For these reasons, SIC does not perform well in IEEE WLANs. 2.4 Switching Mechanisms for Avoiding Inter-cell Collisions Even though the cause of the HN problem in DAS is similar to the HN problem in WLANs with one antenna location, we observe an essential difference: the frame collisions happen not in the wireless link between the STA and RAU but at the AP where the signals from different RAUs arriving via the optical fibers. This inspires us to solve the problem by finding ways to avoid a blind combination of the signals from different RAUs. To that

33 18 Switching Mechanisms for Mitigating Hidden Node Problem in DAS end, we propose two cell switching schemes: hard switching, and soft switching. Briefly, hard switching employs a time division multiplexing scheme of a given AP among the associated RAUs, while soft switching performs a selective reception of simultaneous uplink signals from the RAUs. These switching mechanisms are independent from the MAC protocols, thus they can be combined with the basic CSMA/CA (using data frames for channel contention) and the RTS/CTS protocols. Let us illustrate this using the diagram of RoF-based WLAN in Fig. 2.3, where we focus on a single AP. The multiplexer controls the association between the AP and the RAUs. The proposed mechanisms are implemented at the multiplexer. Clearly, if the multiplexer simply adds the incoming signals from the E/O ports, the AP cannot decode the signals from different RAUs (one RAU covers one cell). However, before being combined, the frames are transported interference-freely through the fibers. Access Point Radio-Optical Interface E/O Fiber RAU MAC PHY Multiplexer E/O E/O RAU RAU Central Station Fig. 2.3 RoF-based WLAN Hard Switching Hard switching employs a time division multiplexing scheme for sharing the APs among the cells served by that AP. Specifically, an AP sweeps the associated RAUs periodically. For each time interval, one AP serves one cell by switching to the RAU of that cell. The other cells can be considered as being temporarily disconnected from the AP. As a consequence, the channel contention is confined to a single cell. Since the STAs in the same cell are mostly within the carrier sensing range of each other, CSMA/CA is very likely to function well. For SCN, the cell size is small so the number of STAs per cell is also small, thus the channel contention intensity is low [41]. In order to be compatible with the IEEE standard, we make the sweeping interval equal to the beacon interval. The communication pattern of hard switching, for the case of an AP with two RAUs, is illustrated in Fig The beacon interval is the period between beacon frames. Beacon frames are sent by the AP periodically. A beacon interval initiates a new cycle of both uplink and downlink communication between the AP and the

34 2.4 Switching Mechanisms for Avoiding Inter-cell Collisions 19 STAs. The beacon frames are also sent using the CSMA/CA mechanism. Therefore, if the AP switches to another cell at the end of the current beacon interval, the communication in the current cell should have completed, i.e., the switching will not interrupt the ongoing communication. After sending a new beacon frame to another cell, a new communication cycle will be initiated. At the same time, the other cells that do not get beacon frames will run the IEEE MAC default procedure for the case that a number of beacon frames are lost. If an STA misses some beacon frames, it will roam to another AP (if available) or trigger a new association process [42]. The condition for triggering the roaming procedure depends on the number of consecutively lost beacon frames, which is a manufacture-specific parameter. So, if the number of cells exceeds the minimum value of this parameter, some STAs would be disconnected from their current AP. This is a limitation of hard switching. For example, if the number of beacons for assessing the connectivity of a specific AP is 4 (see [43]), then a single AP can be connected to maximally 4 RAUs. To avoid that problem, it is necessary to restrict the number of RAUs connected to a single AP not to exceed the maximal allowable number of missing beacon frames. Cell 1 Cell 2 Beacon Interval Beacon Beacon lost Data Frame AP Switch to RAU 1 Time Switch to RAU 2 Switch to RAU 1 Fig. 2.4 Illustration of frame exchange pattern in hard switching. The hard switching mechanism is analogous to TDMA. Thus, when the beacon interval is large, it may introduce unacceptable latency, since the next medium access opportunity of an STA arrives after several beacon intervals. For traffic types like data transfer, the tolerable latency is high. For video streaming, buffering can to a certain extent cope with the delay jitter and latency. Therefore, for such cases, the drawback of hard switching is not severe. For inter-active real-time applications such as voice calls and video conferencing, the latency requirements are stringent. We expect that user scheduling can solve the problem, the same way as in standard TDMA systems [44]. In addition, the beacon intervals can be shortened to reduce the latency. For example, with 4 cells sharing a single

35 20 Switching Mechanisms for Mitigating Hidden Node Problem in DAS AP, we can reduce the beacon interval to 25 ms, in such a way that the total sweeping cycle time equals to the common 100 ms beacon interval Soft Switching Since the colliding frames are received by different RAUs, we would be able to avoid collisions altogether if we decoded the input signals separately. In principle, all the signals received at the RAUs can be separately processed by the AP. But usually an AP can only process a single uplink frame at a time. So, processing all the incoming frames at the same time, either in parallel or sequentially, is not technically feasible. Moreover, the transmitters require immediate acknowledgements (ACKs) of the frames. Some transmissions may have timed out if the ACKs are sent after all the colliding signals are decoded. Therefore, we propose a selective reception scheme which selects one of the frames from the RAUs. We call this soft switching. To this end, we have two possible approaches. First, we can buffer the incoming signals from all RAUs and then only select one of them for decoding. This requires the multiplexer to be capable of buffering the radio signals from the RAUs. Second, we can implement a busy indicator for every RoF link between the CS and the RAUs to indicate which RAUs have signals. The busy assessment can be based on energy detection of the electrical signal of RoF links after O/E conversion, the same as in carrier sensing. Therefore, the same implementations of carrier sensing in current WiFi circuits can be used. If multiple RoF links are detected to be active, then the multiplexer switches to one of the RAUs. The selection can be random, e.g., with equal probability. If fairness should be considered due to the different traffic load in the cells, a random selection with different probabilities can be used. However, this case is not considered in our analysis. An example of soft switching for the case of two cells sharing one AP, is illustrated in Fig Consequently, the AP can receive collision-free frames. Soft switching is done on the time scale of the frame lengths. So for all the cells, the connectivity is not interrupted, which is an advantage with respect to hard-switching. All the STAs associated with the same AP will contend for the channel at the same time. For the contending STAs, if they succeed in the contentions inside their cells, their frames still have a probability of not being received. The frames discarded by the APs will not be ACKed to the STAs, and the STAs may need to re-transmit the discarded frames. So, the collision effect is still present in this approach. However, the collision probability is reduced by a factor of 1 N RAU, where N RAU is the number of RAUs from which the frames are sent simultaneously. For a small number of RAUs, we expect that the frame losses due to collisions can be significantly reduced. For a large number of RAUs, the performance may be poor.

36 2.5 Comparative Performance Analysis 21 Beacon Interval Cell 1 Collisions Beacon Data Frame Cell 2 AP Switch to RAU 1 Switch to RAU 1 Switch to RAU 2 Time Fig. 2.5 Illustration of frame exchange pattern in soft switching Downlink Traffic in Hard-Switching and Soft-Switching In this chapter, we mainly focus on the uplink, however, downlink communications may also take place during the beacon intervals. For hard switching, one AP is completely dedicated to a single cell at a time. So the downlink frames for the other cells need to wait until the AP is available to serve those cells again. In contrast, in soft switching, the downlink is not affected since the switching mechanism only functions when there is uplink traffic. Therefore, hard-switching may impact the downlink traffic, e.g., increase the latency. One way to solve this problem is to decrease the length of the beacon intervals, such that a sweeping cycle takes a shorter time. The length of the beacon intervals can then be optimized based on the requirement of the applications. However, the optimization is not considered in this thesis. 2.5 Comparative Performance Analysis Given the above explanation, we now turn to the evaluation of the performance of the proposed mechanisms. The objective is to evaluate the saturation throughput [45]. The IEEE g standard is considered here. The uplink access mechanism of the other standard versions (for frequencies below 6 GHz) are the same but with some variations of the parameters, e.g., the preamble length is longer in IEEE n/ac. Therefore, the result for these other versions of the standard would be similar to the ones we show here. The IEEE g standard employs two media access schemes: basic DCF CSMA/CA and RTS/CTS [3]. We remind the reader that the basic DCF mechanism uses data frames for channel contention, and the RTS/CTS mechanism uses an RTS frame for channel contention which is also based on DCF. These mechanisms can co-exist with the proposed switching mechanisms. Thus the possible access modes are:

37 22 Switching Mechanisms for Mitigating Hidden Node Problem in DAS Basic DCF (DCF) DCF with RTS/CTS (RTS/CTS) Hard switching with basic DCF (HS-DCF) Hard switching with RTS/CTS (HS-RTS/CTS) Soft switching with basic DCF (SS-DCF) Soft switching with RTS/CTS (SS-RTS/CTS). It is clear that parameters like frame size, contention window size, number of cells, and modulation and coding schemes (MCSs), all can influence the throughput performance. To investigate all these would require a very large simulation effort. Instead we have limited ourselves to an analysis that concentrates on typical values of the parameters. These are taken from [46] [39] and listed in Table 2.1. We only use MCS 7 (corresponding to a data rate of 54 Mbps) since the DAS can guarantee high SNRs due to the short propagation distances. In addition, we assume that all collision-free frames are correctly received by the AP, and the colliding frames cannot be decoded successfully. We assume STAs in the same cell are within each other s carrier sensing range. The STAs in different cells are hidden from each other, which is the worst case scenario. Table 2.1 Simulation parameters. MAC header 34 bytes RTS size 44 bytes CTS size 38 bytes ACK size 38 bytes PHY header 192 bits PHY preamble 20 μs Data frame bit rate 54 Mbps (216 bits/symbol) Control frame bit rate 24 Mbps (96 bits/symbol) Slot time 20 μs SIFS 10 μs Min contention window size 15 Max contention window size 255 Retry limit 4 The simulations are done in MATLAB, using a continuous-time simulation method to simulate the CSMA/CA protocol [47]. For each case, a total transmission of 10,000 packets to the AP is simulated. The average throughput is defined as the number of bits successfully delivered per unit time [48]. The packets for which the number of re-transmissions exceeds the maximal retry limit are discarded, and they do not contribute to the throughput. But the total time also takes into account the time consumed by the unsuccessful transmissions.

38 2.5 Comparative Performance Analysis 23 We analyze the impacts of the variables including MAC protocol data unit (MPDU) size, number of STAs, and number of cells. The small to large MPDU sizes considered are, similar as [39], 500, 1500 and 2500 bytes. The frame formats and the calculation of the time length of all frames are defined in IEEE g [49]. Note that in the standard, the RTS/CTS mechanism is enabled when the packet size exceeds a threshold. The default RTS/CTS threshold is 128 Bytes [50], and it can also be adjusted in practice by the vendors [36]. Note also that, for analyzing the basic DCF, the RTS/CTS mechanism is disabled Impact of the Number of STAs The first discussion is about the influence of the number of STAs. For a larger number of STAs, more STAs are likely to transmit simultaneously, which results in more HN-caused collisions. We assume a simple case where two cells are served by a single AP, and the number of STAs in each cell is the same. A larger number of RAUs is analyzed in the next subsection. The simulation results are given in Fig. 2.6 for different MPDU sizes. Fig. 2.6 Throughput versus the number of STAs in each cell for different MPDU sizes. In general, the switching schemes provide a significant improvement upon the basic DCF and RTS/CTS. In hard switching, HS-DCF performs better than HS-RTS/CTS with a small number of STAs and worse when the number increases. This is because, when there are a small number of STAs, the RTS and CTS frames preceding the data frame transmission add more overhead than basic DCF. But when the contention is more intensive, RTS/CTS reduces the collisions, thus providing a higher throughput. For soft switching, SS-RTS/CTS performs better than SS-DCF, which is consistent with our expectations. When increasing the MPDU size, the performance of basic DCF-based mechanisms degrades faster with the number of STAs since the collision probability increases with longer frames. Although RTS/CTS based mechanisms (including HS-RTS/CTS and SS- RTS/CTS) add some overhead, the throughput drops less rapidly than basic DCF-based mechanisms (including HS-DCF and SS-DCF).

39 24 Switching Mechanisms for Mitigating Hidden Node Problem in DAS Overall, hard switching performs better than soft switching in terms of the saturation throughput. The benefit of the switching mechanisms is more noticeable for large MPDU sizes. For less than around 10 STAs per cell, DCF is better used in hard-switching; and for more than 10 STAs per cell, RTS/CTS is a better choice. As an example, for the medium MPDU size of 1500 bytes and 10 STAs per cell, hard-switching improves the throughput by about 10 times Impact of the Number of Cells Except for the hard-switching based modes, the other access modes all experience intercell collisions. Apparently, the number of cells has a critical impact on the performance of these mechanisms. Therefore, we simulated the saturation throughput performance for different numbers of RAUs. The number of STAs in each cell is fixed at 10. The results are shown in Fig Fig. 2.7 Throughput versus the number of RAUs for different MPDU sizes. It can be observed that hard switching provides constant throughput since there are no collisions due to the HN problem, even when the number of RAUs increases. In contrast, the performance of all the other mechanisms degrades. Soft-switching based mechanisms also significantly improve the throughput in comparison to the basic DCF and RTS/CTS mechanisms. The improvement is larger for longer frame sizes as more air time is used for data transmissions. However, the throughput of soft switching degrades fast. This suggests that it cannot support many RAUs, as many packets are discarded due to the selective reception. Therefore, for multiple RAUs, hard switching is especially better than soft switching. If the number of RAUs is small, e.g., less than 4, SS-RTS/CTS can also be a good choice as it still provides much better performance than RTS/CTS.

40 2.6 Conclusions and Future Work Conclusions and Future Work This chapter proposes hard switching and soft switching mechanisms to alleviate the HN problem in SCN-DAS. They can be applied to the current IEEE standard. To realize the switching mechanisms, additional functions are needed in the CS. Specifically, for hard switching, the AP should be able to switch between the RAUs (see, e.g., [51]); for soft switching, the AP needs to be able to buffer the signals from all the RAUs (see, e.g., [52]). However, no changes are required at the STAs. The simulation results show that both schemes offer significant improvements of the basic DCF and RTS/CTS mechanisms. Hard switching generally performs better than soft switching in terms of the saturation throughput. In particular, for a large number of RAUs, hard switching can provide constant throughput while the performance of soft switching quickly degrades with the number of RAUs. For a small number of RAUs, e.g., less than 4, soft switching with RTS/CTS also performs well, thus should also be a good choice. While not using the switching mechanisms, the RTS/CTS mechanism provides much higher throughput than the basic DCF, thus we suggest to enable the RTS/CTS in the uplink media access in a DAS. There are a number of points that we did not address, but should be investigated to produce a more accurate view for realistic settings and scenarios. First, in our simulations, we have assumed that the number of active STAs are identical in all the cells, while they may be strongly different in some scenarios. Second, we have assumed STAs in different cells are hidden from each other, which may not always be the case as the coverage area of adjacent cells can overlap, and hence some STAs may be visible to each other. Taking into account of these points should provide more realistic results.

41 26 Switching Mechanisms for Mitigating Hidden Node Problem in DAS

42 27 Chapter 3 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs 3.1 Introduction The main motivation for using the 60 GHz unlicensed mmwave band is the wide spectrum available. However, the signals can be easily blocked by objects like human bodies, furniture and walls. Thus, only line-of-sight (LOS) communication is possible between the transmitter and receiver. For static objects like walls, radio planning can be used to avoid blocking. However, mobile objects like humans create more severe problems as blocking occurs randomly and the attenuation is high [53]. For instance, in the presence of one to five people in a large laboratory room, the connection is blocked as frequently as every 5 seconds, and the average duration of the outage is around 300 ms, as reported in [53]. Clearly, this will disturb the communication especially for real-time applications. In addition, the attenuation is typically more than 20 db [53]. In Fig. 3.1, we show an example to illustrate the effect of human shadowing, which we simulated using ray-tracing. In this scenario, the receivers are at 1 m above the ground and the transmitter is at 3 m height. We can see that the human bodies roughly cause more than 15 db power loss than LOS condition. To deal with this, the straightforward solution would be to use higher transmission power, which however may not be practical due to the considerations of the power regulations and potential health consequences, as well as energy consumption of the devices. Another solution is to use a large antenna array. However, the compensation of the power loss of tens of db needs several tens of antennas, which will increase the device complexity and the form factor. In this chapter, we propose to use spatial diversity with distributed antenna arrays, i.e., DAS, to overcome the human shadowing problem. In order to do this, we need to answer three important questions: (1) The positioning of the RAUs: the position of the RAUs affects the LOS visibility of the RAUs to the STAs, as well as the path loss. (2) The beamforming strategy: given multiple RAUs, there are different ways to use them, e.g., just select one of the RAUs or use all the RAUs. The performance may vary for different beamforming strategies.

43 28 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs (3) The trade-off between using more RAUs and larger antenna array sizes: using more RAUs each with a small antenna array, or using fewer RAUs but larger antenna arrays, are both possible to provide acceptable performance, thus there should be a trade-off. In this chapter we set out to answer these questions. Shadow 6 m Human Transmitter 6 m Fig. 3.1 Example of human shadowing simulated by 3D ray-tracing. The remainder of the chapter is organized as follows. Section 3.2 reviews existing techniques in the literature for human shadowing mitigation. Section 3.3 elaborates the motivations for proposing DAS. Section 3.4 presents the RAU position optimization algorithms, and the signal models for 60 GHz DAS beamforming. Section 3.5 introduces the realistic channel simulation based on 3D ray-tracing. Section 3.6 presents the numerical results, and the analyses answer the above questions. The conclusions are drawn in Section Shadowing Mitigation Techniques To overcome the human shadowing problem, several approaches have been proposed in the literature. Let us summarize them below. Reflective paths: Except for the LOS path, there may be a number of reflective paths between the transmitter and receiver induced by surfaces of walls and furniture. The transmitter and the receiver can align their beams to the direction of the reflective paths when the LOS path is blocked. This technique is, in principle, simple since it only requires the devices to be capable of adapting the beamforming directions. The problem is that the

44 3.3 Distributed Antenna System for Shadowing Mitigation 29 reflective paths typically have a higher power loss than the LOS path. The reflection loss depends on the materials of the reflective surfaces. Typical attenuations are around 15 to 20 db [54], but may be lower for some materials [55]. If the power loss is sufficiently low, beamforming can be used to compensate for that loss. But for higher losses, beamforming may not be able to provide enough link strength. In addition, in some scenarios, no strong reflective paths can be found. For these reasons, in [56] it is proposed to place metal reflectors on the walls. It is claimed that the connectivity can be significantly increased by deploying two reflectors at optimal positions for a certain pair of transmitter and receiver positions in a 2D plane. However, results are not known for the more realistic case where devices are at arbitrary positions in a 3D environment. Relaying: Relaying can also be helpful as reported in [57] and [58]. It can be done in the link layer (decode the received packet and then forward it) or PHY layer (simplify amplify the received signal and then forward it). Two-hop link-layer relaying is already standardized in IEEE ad. However, relaying is feasible with the provision that there are other devices adjacent willing to relay the packets, and strong connections exist in all the hops. Thus, in low STA density scenarios it does not work well. Then dedicated relaying devices are needed [57]. In addition, relaying consumes more network resources since for transmitting the same data more air time and energy is used [59]. Coordination for establishing the multi-hop links also introduces some overhead [60]. The relaying algorithms, e.g., amplify-and-forward which works on the PHY layer, may introduce noise to the signals [61]. Deploy more APs: A denser deployment of APs can also alleviate the shadowing problem [62] [63]. The connectivity with multiple APs is discussed in [62]. For providing continuous connectivity, it is necessary to hand over the sessions between APs. Therefore, [63] proposes to use a central controller to select the best visible AP to take over the session if blockage happens. In addition, the blockage may be temporary, thus the connection should be re-established with the former AP when the blockage disappears. These procedures results in large overhead and latency, which is also mentioned in [63]. 3.3 Distributed Antenna System for Shadowing Mitigation The common point of relaying and deploying more APs is to use additional antenna arrays to obtain spatial diversity. In order to maximize the spatial diversity gain, it is necessary to place the antenna arrays far apart so as to get de-correlated shadowing fading [64]. This is essentially the idea of DAS. The DAS architecture can be combined with RoF. The signals from the RAUs can be coherently added at the receiver. Adaptively adjusting the beamforming coefficients of the RAUs (including power and direction) can handle any physical channel changes including the human blockage. We illustrate the principle in Fig. 3.2.

45 30 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs AP RoF link AP RAU RAU Human (a) LOS situation STA STA (b) Human shadowing situation Fig. 3.2 Illustration of DAS for mitigating human shadowing. In comparison with the aforementioned techniques, DAS has several advantages: - The signals (after beamforming processing) are fed to the RAUs through RoF links, so there are no costs of radio resources. But in relaying, the coordination information packets need to be exchanged via wireless links between the transmitter, receiver and the relays. In addition, the data packet is forwarded through several hops. These procedures may consume more resources including air time and energy, than a single-hop communication. - The connection between an STA to the AP is not interrupted when an obstacle blocks the LOS connection with the RAU, since other RAUs with LOS connection to the same STA can take over quickly. Only the tuning of the beamforming coefficients based on the current channel state is required. This consumes much less time than a handover above the physical layer, which we can see from the IEEE ad standard [65]. For handover between APs, several procedures are required, including discovering the APs, beam alignment, association, authentication, etc. - DAS offers alternative LOS paths instead of NLOS paths which provides much higher received signal power. This is unlike using reflective paths which may degrade the signal strength, or relaying which deteriorates the signal quality. Despite the advantages, DAS requires RoF connections between the APs and the RAUs. If more antennas are used, the signal distribution network will be more complex. For example, for A-RoF, one antenna may need one optical wavelength [66]. Thus more optical wavelengths are required when more antennas are used, which also means more lasers and optical fibers are needed. It is necessary to use fewer antennas, and also to optimize the system to provide sufficient received signal strength over the coverage area. For that purpose, we need to answer a number of questions: (1) where to place the RAUs (2) what

46 GHz DAS and Beamforming 31 beamforming strategy should be used (3) how many RAUs and how many antennas per RAU are needed. We discuss these questions in detail in the following sections GHz DAS and Beamforming The architecture of the RoF-based WLANs has been presented in Chapter 1. As mentioned earlier, 60 GHz signals are confined by the walls, so at least one AP is needed for each room. This means multiple RAUs should be installed per room to form a DAS, which is illustrated in Fig For example, AP 1 serves Room 1 with two RAUs, forming a DAS. Central Station AP 1 RAU External network /Internet AP 2 AP 3 AP x Radio-Optical Interface Optical link Optical Network RAU RAU Room 1 RAU Fig. 3.3 A RoF based DAS for 60 GHz WLAN Optimizing the Placement of the RAUs Consider that M RAUs are employed in a DAS for a particular scenario. The first important step is to optimize the locations of the RAUs. In general, this problem is formulated as x arg max g( x) (3.1) x where x contains the coordinates of the M RAUs, is set of all the possible positions, and g( x ) is the 5-percentile received signal power [67], which is usually used for network planning. The received signal power is calculated by summing the received signal power from all the RAUs directly. To simplify the problem, we make the following assumptions:

47 32 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs - We assume the RAUs are installed on the ceiling, which is common in practice. It is also known that a higher position gives a better chance of having the LOS connection [62]. - The ceiling surface is divided into a square lattice with a mesh size of d which is set to 2 m in the later numerical analysis. The RAUs are placed at the center of the squares. As a result, we have a limited number of candidate RAU positions, and an exhaustive search can be done to find the optimal solution. - The optimization is done assuming that no humans are present. This simplifies the optimization. In reality there may be a random number of humans in a room at random positions. The channel response is simulated with ray-tracing which is introduced in Section 3.5. Then, the optimum RAU positions are found through an exhaustive search. The beamforming strategies introduced next will be based on these optimized RAU positions Beamforming Strategies Beamforming with multiple RAUs can actually be done in two ways: blanket transmission and selective transmission. In blanket transmission, all the RAUs are active. In contrast, only the RAU with the highest channel gain is used in selective transmission. Consider a single AP with a DAS serving a single STA at a time. The DAS has M RAUs and each has N RAU antennas that are configured as a rectangular planar array, with halfwavelength spacing in both dimensions. Every STA has N STA antenna elements with the same array construction. See Fig Ceiling plane x 1 x 2 x 4 x 3 RAU antenna array z x y Half a wavelength STA antenna array STA plane Fig. 3.4 Illustration of the RAU and STA antenna arrays.

48 GHz DAS and Beamforming 33 The frequency bandwidth is B and it is divided into L OFDM subcarriers. We assume that the total power is equally allocated to the subcarriers, i.e., p Pt L where is the transmission power for each subcarrier, and P t is the total transmission power. Perfect channel knowledge is assumed to be available throughout the analysis for the sake of simplicity. We did not investigate the effects of imperfect channel knowledge. This is still an open question Blanket Transmission The complete channel matrix of the is l -th subcarrier between the M RAUs and the STA G G G l, l1, G l M (3.2) which is an NSTA MNRAU matrix. G lm is the channel matrix between the m -th RAU and the STA. The received signal is y G w xp n (3.3) 1/2 l l l l l where x l is the symbol transmitted on the l -th subcarrier with unit average power, i.e. 2 E( xl ) 1. n l is the N 1 STA i.i.d. Gaussian noise vector with zero mean and unit variance elements. w is the transmit beamforming vector for all the antennas, which satisfies H w w 1. l l The decision variable after receiver beamforming is then given by r c y c G w p x c n G p x c n (3.4) H H 1/2 H 1/ 2 H l l l l l l l l l l, cw l l l H where c is the receiver beamforming vector satisfying c c 1. l, H l l l G cw H l l p c G w is the effective channel gain after beamforming. Then, we can obtain the noise power of the l -th subcarrier as The SNR of this subcarrier is E( ) H 2 2 cl n l (3.5) l 2 pgl, cw SNR l,blanket ( clw l) (3.6) 2

49 34 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs To achieve the highest SNR, we need to find the beamforming vectors, w l and l, to maximize the effective channel gain Gl, cw. This is called dominant eigenmode transmission or eigenmode beamforming [68], which we explain in the following. Assuming perfect channel knowledge, let us apply singular value decomposition (SVD) to the channel matrix G l G U DV (3.7) l where U l and V l are unitary matrices of size NSTA NSTA and MNRAU MNRAU, respectively, and D l is a diagonal matrix with the diagonal elements, denoted by j, jfor j 1,,min NSTA, MNRAU, being the sorted singular values in descending order, and the other elements are zero. Then, the optimum beamforming vectors are derived as l l H l c u, w v (3.8) l l,1 l l,1 where u l,1 and v l,1 are the first column of U l and be obtained as where max{,, } l,max l,1 l,min( NSTA, MNRAU ). V l. Finally, the maximum SNR can 2 lmax, l p (3.9) 2 Due to the use of OFDM, the subcarrier SNRs should be mapped to the link SNR, which is usually performed with the exponential effective SNR mapping (EESM) [69], given by the following expression c eff L1 1 l ln exp( ) (3.10) L l0 where is a parameter determined by the modulation and coding scheme (MCS). The values of can be found in [70] for different MCSs. The typical value of ranges from 1 to 35 [71]. Thus it is necessary to select according to the MCS applied. To simplify the analysis, we set 2, which is a typical value for QPSK modulation (For different coding schemes, is around 2 [72]).

50 GHz DAS and Beamforming Selective Transmission In selective transmission, the RAU with the best channel condition is used to transmit signals with full power. For example, if the m -th RAU is chosen, the received symbol of the -th subcarrier after beamforming is given by l So the SNR of the received signal is r c G w p x c n G p x c n (3.11) H 1/2 H 1/2 lm lm lm lm l lm l lm,, cw l H lm l SNR lm Glm, cw, p 2 2 (3.12) The optimum beamforming is performed by using the following beamforming vectors c u, w v (3.13) lm lm,1 lm lm,1 where u lm,1 and v lm,1 are the first columns of U lm and V lm, which are derived from the H SVD of H lm, i.e. Glm UlmDlmV lm. Then, the maximum SNR of the l -th subcarrier using the m -th RAU is derived as The effective SNR is thus expressed as eff, m lm 2 lmmax (3.14) 2 L1 1 lm ln exp( ) (3.15) L l0 Then, the RAU with the highest effective SNR is selected, i.e. the effective SNR of selective transmission is given by tive max m m, m1,, M (3.16) eff, selec eff, From the above formulation, we can see that blanket transmission can achieve a higher beamforming gain since MN RAU antennas are active. In contrast, selective transmission only uses the N RAU antennas of the strongest RAU. Selective transmission has several advantages. First, it has a lower beamforming complexity due to the smaller antenna array [73]. Second, since the radio signal only needs to be delivered to one RAU, which has a small number of antennas, the RoF network could be simpler, e.g., simpler optical hardware. Hence, selective transmission is attractive from the practical point of view. It is thus

51 36 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs interesting to quantify the performance difference between blanket and selective transmission to help justify the choice. 3.5 Channel Simulation In order to get realistic results, the channel is simulated using a 3D ray-tracing software called Radio Propagation Simulator (RPS) [74]. The details of the simulation methodology, data extraction, and post processing are given in Appendix A. To account for different scenarios, we consider three room sizes: 6 m 6 m 3 m (living room), 10 m 6 m 3 m (office room), and 10 m 10 m 4 m (conference room). Similar scenarios are defined in the IEEE ad channel models [75]. This is illustrated in Fig The locations of the STAs is assumed to be 1 m above the floor, so this is where the signals sent by the RAUs are measured. 10 m Conference room 10 m Living room Office room 6 m Candidate RAU positions Fig. 3.5 Simulation scenarios. For all the rooms, the left side wall is the exterior wall which is made of concrete, and there is a large window of size 4 m 1 m in the center. The other three sides of the walls are interior walls covered with plasterboard. A door is located on the right side wall with dimensions 1.5 m (width) 2 m (height). The door is assumed to be closed in the simulations. Humans are modeled as absorptive rectangular prisms with the same height (1.75 m) and width (0.5 m), facing in random directions, and located at random positions. In each RPS simulation, the humans are in a random position drawn from a uniform distribution. For

52 3.6 Performance Evaluation 37 every room scenario, we take 20 RPS simulations, which we observe is sufficient for the statistical analyses. 3.6 Performance Evaluation This section presents the analysis results in the following order. First, we give the optimized positions of the RAUs for the different scenarios. Then, we compare the received SNRs while applying blanket and selective transmission strategies in different scenarios and with different number of RAUs, but with a fixed density of humans in the room. Finally, we analyze the relationship between the outage probability and the number of RAUs and the antenna array sizes, assuming different populations. The outage probability is defined as the probability that the effective SNR falls below a certain threshold over the coverage area [76]. The simulation parameters are given in Table 3.1. The transmission power and noise figure are typical values for state-of-the-art electrical components for 60 GHz [13]. Table 3.1 Simulation parameters. Carrier frequency 60 GHz Bandwidth 2 GHz Number of subcarriers 512 Total transmission power 10 dbm Noise figure 10 db Noise power 174 dbm/hz Number of RAUs Variable from 1 to 4 Number of antennas per RAU N =4 RAU Number of antennas per STA N =4 STA Optimized RAU Positions The locations of the RAUs are optimized using the algorithm given in Section For M =1 ~ 4, the optimal positions for different scenarios are shown in Fig. 3.6, Fig. 3.7, and Fig. 3.8, respectively. The solid dots are the optimal RAU positions. For M 1, the optimum position is the center of the ceiling. However, for M 2, it is necessary to re-run our optimization algorithm since the rooms may have different sizes and shapes, which results in different optimal RAU locations.

53 38 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs Fig. 3.6 Optimized RAU positions for living room scenario. Fig. 3.7 Optimized RAU positions for office room scenario. Fig. 3.8 Optimized RAU positions for conference room scenario Blanket and Selective Transmission Strategies Next we present the received SNR for the two transmission strategies based on the optimal placement of the RAUs. We consider the case of 10 humans (blocking objects) in each scenario. For comparison purposes, we also give the results when no humans are present, thus all receiver positions are within LOS to all the RAUs. The former is indicate by NLOS, and the latter by LOS, in the figures. The SNR cumulative distribution func-

54 3.6 Performance Evaluation 39 tion (CDF) for the three scenarios are shown in Fig. 3.9, Fig. 3.10, and Fig. 3.11, respectively. Note that for M =1, blanket and selective transmission have the same results, so their CDF curves are identical. It can be clearly seen that, due to human shadowing, there is a high probability of low SNRs, which entails the long and high tails in the CDF curves than in LOS situation. This is especially obvious for M =1. For larger M, the SNRs increase thus the CDF shifts to the right and approaches the LOS case. In particular, at M =4, the CDF in NLOS and LOS are almost the same for both blanket and selective transmission. This suggests that the human shadowing is almost completely mitigated by using DAS. Blanket transmission generally performs better than selective transmission. However, the difference, which is around 1 db maximum, is marginal in all the evaluated scenarios. Thus, it is justifiable to choose selective transmission due to its advantages and comparable performance. However, it is still noteworthy to mention that the comparison assumes the same total transmission power for consistent comparison. If the total transmission power is not the same, e.g., it depends on the number of active antennas (which is not fair), the above results will not apply. In addition, the conference scenario has a larger area size and the ceiling is higher. Therefore, with the same number of humans, the shadowing is less severe than in the living room and office room scenarios. Fig. 3.9 Distribution of the received signal SNR in living room scenario.

55 40 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs Fig Distribution of the received signal SNR in office room scenario. Fig Distribution of the received signal SNR in conference room scenario.

56 3.6 Performance Evaluation More RAUs or Larger Antenna Array? As the number of humans increases, the human shadowing problem becomes more severe. To cope with that, we can use more RAUs as we understand from the above analysis. But we could also increase the size of the antenna arrays in the RAUs. Therefore, using a smaller number of RAUs each with a larger antenna array, or a larger number of RAUs each with a smaller antenna array, could both be satisfactory. Thus there may be a tradeoff between the two choices. Despite the performance perspective, the cost and complexity of the network should also be considered. For example, installing an additional RAU incurs more wiring costs, while increasing the antenna array size requires a higher capacity of the RoF links. The marginal cost of an additional RAU and of an additional antenna are likely to be different. Therefore there is also a trade-off between the performance and the choice of the configuration (i.e., the combination of the number of RAUs and the number of antennas per RAU). This should be investigated when one is considering the economics of the different solutions, which is beyond the scope of this thesis. To study the trade-off, we analyze the outage performance of the different choices with respect to both M and N RAU. The outage probability is defined as the probability that the effective SNR falls below a certain threshold over the coverage area [76]. For simplicity, we limit ourselves to the office scenario and selective transmission. However, the room dimensions and the expected population may vary in practice, then it is necessary to repeat the analysis here. The outage SNR threshold is set to 10 db which is required for MCS 17 (QPSK with a code rate ¾ which is the highest MCS using QPSK), see [77] and [78]. The average outage probabilities with 95% confidence intervals for different number of people, are shown in Fig LOS refers to the case where that are no humans in the room. We consider two antenna array sizes at the STA, i.e., N STA =4 and 16, which correspond to a small and a large antenna array, both feasible in practice [79]. M ranges from 1 to 4 while the number of antennas per RAU is set to N RAU =4, 9, 16, and 25. The dashed line indicates the 5% outage probability. We can see that, in the LOS situation, one RAU is sufficient to achieve a low outage probability for both N STA =4 and 16, although a different antenna array size at the RAU is needed. However, when a number of humans in present in the office, one RAU cannot provide a good coverage (lower than 5% outage probability). We see from the figures that, at least two RAUs are needed, and it also depends on the density of people and the STA antenna array size. For example, for the case of 5 people, using two RAUs are sufficient, and the required N RAU is 25 and 9 for N STA =4 and 16, respectively. In addition, for all cases, we observe a significant improvement from N RAU =1 to 2.

57 42 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs (a) N STA =4 (b) N STA =16 Fig Outage probability for different number of humans. In fact, to achieve a certain outage probability, we may use different combinations of M and N RAU. However, with a larger M, the required total number of antennas is much less. For example, for the case of 10 humans and N STA =4, we can either use ( M =3, NRAU 25 ) or ( M =4, N RAU =9). For this example, the total number of antennas is 75 for M =3, but only 36 for M =4, less than half. In reality, the expected density of people, the STA antenna array size, and scenario size, may vary. The procedures of the analysis in this subsection should be re-done to find the optimal configuration of M and N RAU to achieve a particular expected outage probability. Specifically, the optimization is done in two steps: first, optimize the positions of the RAUs using the algorithm in Section 3.4.1;

58 3.7 Conclusions and Future Work 43 second, find the configuration that meets the outage probability requirement and uses the smallest number of antennas. The cost of the RAUs could be integrated into this optimization process. 3.7 Conclusions and Future Work In this chapter, we have proposed a DAS architecture to mitigate the human shadowing effect in 60 GHz WLANs. We have provided mathematical models and conducted simulations to seek answers to several questions such as the optimal placement of the RAUs, the beamforming strategy, and the tradeoffs between the number of RAUs and the size of the antenna arrays. We discovered that blanket transmission strategy, although utilizing more antennas simultaneously, offers a marginally superior performance than selective transmission. It is thus justified to use selective transmission in practice as it requires a lower infrastructural complexity. Increasing the number of RAUs and increasing the antenna array size per RAU both can improve the outage performance. A practical choice will need to consider both the cost and the performance. It is thus necessary to balance between installing more RAUs and using larger antenna arrays. We find that with only one RAU per room, it is not possible to resolve human shadowing, thus it is not an acceptable choice. Therefore, using more RAUs are necessary. We see that the largest improvement gradient is from one to two RAUs per room. Using two RAUs per room is satisfactory for a low population density, e.g., 5 people per room. However, it is hard for a higher population density and when the STA has a small antenna array. Using more RAUs can reduce the total amount of antennas required. The most cost-effective solution would require an optimization based on the particular scenario, population, and STA antenna array size, to obtain the expected performance and minimize the total number of antennas. In future research, one needs to consider non-uniform STA distributions so as to better optimize the location of the RAUs. In our analyses, we have used the optimal eigenmode beamforming algorithm, while practical beamforming may only use a number of discrete beam directions [80]. The effect of this should be considered in future research.

59 44 Distributed Antenna System for Mitigation of Shadowing in 60 GHz WLANs

60 45 Chapter 4 Massive MIMO in Indoor WLANs: Background and Insights 4.1 Introduction Massive MIMO is another promising technique for improving the wireless network capacity, which we will focus on in the following chapters. The objectives of this chapter are two-fold. First, we give a brief introduction to massive MIMO: the basic concepts, the advantages and the challenges. Second, we present some first performance analysis of massive MIMO in indoor environments based on measured channel data. Specifically, we have compared the channel capacity in measured channel and ideal channel, and analyzed the relation between the performance of different precoding techniques and the antenna array size. The analyses provide some insights on the behavior of massive MIMO in indoor channels. The remainder of the chapter is organized as follows. We first introduce massive MIMO in Section 4.2, and present the theoretical perspectives of massive MIMO in Section 4.3. Section 4.4 presents the channel measurement in an indoor scenario. Based on the measurement results, we analyze the capacity performance in Section 4.5. The conclusions are drawn in Section Massive MIMO What is Massive MIMO? The concept of massive MIMO originates from [81], where we find the first discussions on the benefits of using large antenna arrays in cellular networks. Later, massive MIMO was usually considered as a system that uses antenna arrays with a few hundred antennas, simultaneously serving tens of terminals within the same time-frequency resource [15]. However, how many antennas should be involved to call it massive is not clear. An argument in [82] is that the number of antennas needed depends on the degrees of freedom the channel offers, SNR, and path loss. The point of view behind this definition is that, in massive MIMO, simple linear precoders like maximum ratio transmission (MRT) and zero-forcing (ZF), should be able to get close to the ultimate performance limit provided by an infinite number of antennas. However, [83] suggests that there should be no strict requirements on the relation between the number of AP antennas and the number of user

61 46 Massive MIMO in Indoor WLANs: Background and Insights antennas, since the relation depends on a variety of conditions, e.g., the system performance metric, the propagation environment, and the coherence block length (product of coherence time and bandwidth). In this thesis, we employ the definition from [83]: massive MIMO is a system with unconventionally many active antenna elements at the APs that each can serve an unconventionally number of STAs. In contrast, conventional MIMO systems use only up to 8 antennas at the AP and the STAs, e.g., WiFi and LTE [15]. Another important aspect of massive MIMO is the physical size of the antenna arrays. Intuitively, the more antennas are used, the larger the physical size of the antenna array is. In practice, the space to place the antenna array is always limited. The common spacing between the antenna elements is half a wavelength, which is sufficient for getting independent channels between adjacent antennas in rich scattering environment [68]. Packing more antennas in a fixed volume is not helpful for getting more spatial degrees of freedom [84], or higher power gain [85]. Therefore, the size of the antenna array, i.e., the number of antennas, is limited as well. In addition, the physical size of the antenna array depends on the carrier frequency for the same number of antennas and configuration. For higher frequencies, the wavelength is shorter, thus the physical size of the antenna array can be smaller. The antennas can be configured geometrically in different forms, e.g., linear, planar, and volumetric. For example, for 2.4 GHz, the wavelength is around 12 cm. A 64- element uniform planar array with half-wavelength spacing has a size of around 0.5 m by 0.5 m. But a uniform linear array has a length of roughly 4 m, which may be too large for indoor environments. For 60 GHz, the wavelength is only 0.5 cm, the corresponding physical sizes are 4 cm by 4 cm and 16 cm for uniform planar and linear arrays, respectively. Therefore, for high frequencies, more antennas can be used Potential Massive MIMO has a number of attractive features [82]. The most important ones are the potential to significantly increase both the spectral and energy efficiency. The spectral efficiency can be increased by enabling high-order MU-MIMO spatial multiplexing. The energy efficiency can be improved through coherent signal combining. These benefits are valid for both uplink and downlink transmissions. It is expected that massive MIMO can improve the spectral efficiency 10 times and simultaneously improve the energy efficiency 100 times [15] [86]. This potential makes massive MIMO an important technology for achieving the 5G targets in spectral and energy efficiency [87]. In this thesis, we mainly focus on the spectral efficiency. Energy efficiency is another broad topic for which there are many studies, and one can refer to [88] and the references therein Challenges There are three major challenges one has to deal with in the design of massive MIMO systems: complexity, channel knowledge and the limitation of the propagation environment. Let us discuss them.

62 4.3 Basics of Massive MIMO 47 Complexity: It is apparent that the use of a large number of antennas results in high complexity. The complexity potentially exists in many domains such as downlink precoding and uplink signal combining, symbol detection, transmission power optimization, antenna array design, etc. [89]. For example, the computational complexity of linear ZF precoding 2 increases at least quadratically with the number of antennas, i.e., OM ( ), where M is the number of massive MIMO antennas [16]. More detailed modeling of other precoders can be found in [90]. For that reason, linear precoding techniques such as MRT and ZF, are commonly accepted as practical candidates as they have relatively low complexity. Nevertheless, techniques need to be developed to further reduce their complexities to make them feasible for even larger antenna arrays [16] [90]. Channel knowledge: Channel state information at the AP is necessary for massive MIMO to perform well. The challenge is that the overhead of channel knowledge acquisition is large, especially for a large number of users. In practice, there are additional constraints such as the channel coherence time, and the latency caused by the channel acquisition process. Taking advantage of the channel reciprocity and employing TDD, the channel estimation can be done through uplink pilot training [15]. The overhead of this approach is only related to the number of STAs simultaneously served, but not the number of antennas at the AP. Therefore, there is no limitation on the number of AP antennas in TDD channel reciprocity based channel training. However, the number of orthogonal pilots is limited and they have to be reused in adjacent cells. This results in inter-cell interference which limits the performance of massive MIMO in cellular networks [89]. In addition, TDD requires calibration of the radio frequency (RF) chains of the receiver and the transmitter, which are in general not reciprocal [91]. Moreover, the calibration procedure also induces overhead. Propagation environment: Massive MIMO works better under rich-scattering environments, which however may not be the case in practice. For example, the number of scatters may not be large enough to support a large number of spatial streams. Therefore, measures need to be taken to overcome that, e.g., by employing a distributed antenna system (DAS) which is discussed in this thesis in Chapter Basics of Massive MIMO In this section, we provide a theoretical view on the potential of capacity improvement with massive MIMO. We use sum-rate as the measure of the PHY layer performance (in bps/hz), i.e., the sum spectral efficiency of the spatial streams to all STAs [68]. Sum-rate is commonly used in evaluating the performance of MIMO systems. A higher sum-rate results in a higher total data rate from the AP to the STAs for the downlink for a given bandwidth, therefore a higher network throughput (in bps). The network throughput is calculated taking into account the protocol overheads, e.g., the PHY and MAC protocol overheads, which are referred to as the PHY and MAC throughput (see [81], [92], [93]

63 48 Massive MIMO in Indoor WLANs: Background and Insights and [48]). The per-sta data rate and throughput can also be found to be used in the literature. Per-STA data rate or throughput is a performance measure for the links of the individual STAs, which is more commonly used when evaluating the fairness and qualityof-service (QoS) of the relevant algorithms, e.g., user scheduling algorithms [94]. Let us start with an introduction to the general MU-MIMO system MU-MIMO Consider a single cell WLAN, where the AP is equipped with M antennas that simultaneously serve K single-antenna STAs. Let us focus on the downlink and assume perfect channel knowledge, then the narrow-band received signal is given by [16] where 0 is the average SNR, H y 0 H xn (4.1) H[ h1 h K is the MIMO channel matrix where h k is the channel vector for the k - 2 th STA with h M for all K, k x is the transmitted signal vector with unit total transmission power, i.e., 2 E( x ) 1 and n is the i.i.d. complex Gaussian noise vector with unit variances. The sum-rate capacity, which can be achieved with the non-linear dirty paper coding (DPC), is known to be [95] Csum, DPC k, k 1,, K log2 IM 0 H max log det( HD H ) 1/2 H 1/2 max, 1, 2det( 0 ) k k, K log IK D H HD (4.2) where k is the power allocated to the k -th STA, and D dia g{ 1,, K } is the power allocation matrix, subject to the constraint 1. The power optimization can be done k by the iterative waterfilling algorithm [96]. As we can see, for a given 0 which is determined by the transmission power, the channel matrix H is the fundamental element limiting the achievable capacity. k

64 4.3 Basics of Massive MIMO Benefit of Large Antenna Arrays Assume that the channel vectors of the STAs are mutually orthogonal, i.e., the most favorable propagation condition, then we have H H MI K H. From the above capacity model, we derive the optimal power allocation k 1/ K for all. The sum-rate under favorable propagation conditions is then given by [16] C log (1 M ) Klog (1 M ) (4.3) K 0 0 sum 2 2 k K K As we can see from this equation, we can get K times the spectral efficiency gain as well as M times the power gain. We refer to this case as the interference-free (IF) condition [95], and denote the capacity by C sum,if in the rest of the chapter. The favorable propagation is an optimistic assumption. Nevertheless, if the AP antenna array size is much larger than the number of STAs, i.e., M >> K, then the column vectors of H become asymptotically orthogonal by the law of large numbers [86]. The favorable propagation condition can therefore be approximated, and we may be able to achieve the interference-free capacity. Moreover, using large antenna arrays also allows linear beamforming techniques to achieve the optimal capacity. Let us specify that. With linear beamforming in the downlink, the received signal model is given by 0 H y= H Ws+n (4.4) K where W [ w1,, w K ] is the linear precoding matrix, and s is the information symbol intended for the K STAs. Note that equal power allocation is assumed here. The SINR of the k -th user is k SINR The sum-rate capacity is then given by k kk 0 H hk w K 0 H hk w K 2 k 2 k' 1 (4.5) C sum, linear precoder 2 k 1 K log (1 SINR k ) (4.6)

65 50 Massive MIMO in Indoor WLANs: Background and Insights Let us take maximum ratio transmission (MRT) and zero-forcing (ZF) precoders for example. The beamforming matrix is obtained as follows. First, we calculate the non-normalized beamforming matrices of the MRT and ZF F MRT H H FZF H( H H) 1 (4.7) Then, the normalized beamforming vector w k is obtained by w k f f k k (4.8) where f k is the k -th column of F precoder, such that w k has unit power. Substituting into Eq. (4.5) and (4.6), we can calculate the sum-rate of the linear precoders. Assume favorable propagation is achieved when M >> K, then we have 1 WMRT WZF H. As a result, we obtain Csum, MRT Csum, ZF Csum, IF. That is, linear M precoding can potentially achieve the optimal performance, which is an attractive benefit of using large antenna arrays. These benefits of massive MIMO rely on the channel matrix H to become well-conditioned as M increases. However, the condition of the channel matrix does not solely depend on the number of antennas. Instead, there are many other factors involved, including the configuration of the antenna array (e.g., linear or planar), propagation environment (e.g., indoor or outdoor), position of the user devices, and carrier frequency. In particular, the antenna array aperture size determines the angular resolution of the antenna array. Therefore, the physical size of the massive MIMO antenna array increases with M. Comprehensive investigation of these factors are necessary, for which there are a number of reports in the literature which we review in the following section A Review of the Analysis of Massive MIMO Most studies on massive MIMO assume the i.i.d. Rayleigh fading channel model, e.g., [81], [97] and [98]. However, the i.i.d. channel is optimistic as it assumes the propagation environment offers rich scattering, i.e., the scatters are uniformly distributed in all directions around both the transmitter and receiver. Real channels usually have a limited number of scatters clustered in certain directions [68] [99]. A number of recent reports are based on measured channel data in different environments and frequencies, and with different antenna array configurations, which are summarized in Table 4.1. The general question answered in these analyses is whether the channel vectors between the STAs become orthogonal as the number of AP antennas increases, as predicted by theoretical w k

66 4.3 Basics of Massive MIMO 51 results. On the one hand, the real channel based analyses confirm that large antenna arrays can improve channel orthogonality. On the other hand, the improvement gradually vanishes beyond a certain number of antennas. An explanation is that the channel offers limited spatial degrees of freedom [100]. Moreover, the antenna array configuration is found to have significant impact. For example, with the same number of antennas, a linear array performs much better than cylindrical array [95]. Table 4.1 Literature on massive MIMO analysis based on measured channel (below 6 GHz). Reference Environment Frequency band Antenna array configuration, type and size [100] outdoor (urban) 2.6 GHz CAS, synthetic linear and cylindrical, 128, [95] outdoor (urban) 2.6 GHz CAS, synthetic circular and synthetic linear, 112 [17] indoor 2.4 GHz CAS, uniform planar array, Impact of the Propagation Channel One may wonder why the channel is influenced by the aforementioned factors, including the propagation environment and the antenna array configuration. To explain that, we employ the Kronecker narrow-band channel model [68] where the elements of H R H R (4.9) 1/2 1/2 T w R H w follows i.i.d. (0,1), and, R T and R R denote the deterministic transmit and receive correlation matrices, respectively. This model is not always valid since it assumes the transmit and receive correlation matrices to be separable [68], but it is easy to understand. The optimum case is that H consists of i.i.d. elements since then the channel has full rank [101]. As we can see, in order to have i.i.d. elements in H, both the transmit and the receive correlation matrices must be identity matrices [102]. This requires there are a large number of scatters surrounding both the transmitter and receiver antennas for co-located antennas, or, alternatively, the antennas are widely separated so they experience independent fading [103]. Real environments usually have limited number of scatters clustered in certain directions [99]. It is likely that the channels are correlated if the spacing between the antennas is small, e.g., half-wavelength. When the receiver antennas are far apart, e.g., they are at separate STAs, the receive correlation matrix is an identity matrix as the channels between the STAs are independent. However, the channels between the transmit antennas may still be correlated because of insufficient antenna separation.

67 52 Massive MIMO in Indoor WLANs: Background and Insights In addition, if the transmit and receive antennas are in LOS to each other, a LOS component need to be added to the above model, i.e., HHLOS H NLOS, where H LOS is the LOS component, and H NLOS is the NLOS component which can be modeled by the above Kronecker model [104]. H LOS is deterministic and is only related to the geometrical position of the transmit and receive antennas. To get full rank of H LOS, the STAs have to be separated far enough to be resolvable in the angular domain [103]. Thus when the STAs are close to each other, they cannot be separated, which will result in high channel correlations [104]. In summary, the above theoretical explanations have motivated the use of massive MIMO, and we also saw that the real propagation channel may limit the performance. In the following sections, we show some numerical analysis based on channel measurements in indoor environment. The objective is to investigate the sum-rate performance in real channels in comparison with the ideal i.i.d. channel to obtain some insights. 4.4 Channel Measurements The measurements were conducted in an office room (of size 6 m 4.7 m 3.2 m) of the Potentiaal building of the Eindhoven University of Technology. See the graphical illustration in Fig The setup is intended to simulate a WLAN in a typical office scenario. Note that the propagation channel depends on all the objects surrounding the antennas, not only the objects in the figure. This figure merely shows the dimensions of the scenario, and the position of the transmitter and the receiver antennas. LOS position VNA window VNA NLOS position Vector Network Analyzer Tx Array 32 1 desk 1 2 drawer 5~ ~12 3 cabinet 4 Fig. 4.1 Top-down view of the measurement scenario. The transmit antenna array, which is assumed to belong to a single AP, is placed on the left side wall at 1.7 m height. It consists of a (virtual) uniform linear array (ULA) of 32

68 4.5 Capacity Analysis 53 antenna elements with half-wavelength spacing at 2.5 GHz (around 6 cm). So the antenna array is around 2 m, which is significantly larger than the traditional MIMO arrays. The linear array configuration offers a large aperture size which determines the spatial resolution of the MIMO system [105]. The antenna element of the large antenna array was a patch antenna that had a quasi-omni directional pattern in the half sphere. The receive antenna had an omni-directional pattern in the horizontal plane, which was vertically placed in the measurements. The receive antenna positions are divided into four groups: LOS-colocated, NLOScolocated, LOS-distributed and NLOS-distributed. In total, there are 14 measurement positions, which are shown in Fig. 4.1 with their indices. Among them, 1, 3, 58, and 13 are LOS positions, and the others are NLOS positions which are blocked by objects like the metal cabinet beside the wall and metal drawers under the desks. The receiver positions 5through 8 and 9through 12 are separated by a distance of half a wavelength (6 cm), representing the colocated STAs, while the others have much larger separations. To synthetize different channel conditions, we can select from the measured STA positions. For example, the group of 5 through 8 STA positions are LOS-colocated. The channel measurements were performed in the frequency band from 2.4 GHz to 2.5 GHz (100 MHz). An HP 8753C Vector Network Analyzer (VNA) was used to measure the channel transfer function by sweeping the entire band with a frequency step of 0.5 MHz. As a result, we obtained channels over the 100 MHz band (201 frequency points). The channel measurements were done using a synthetic approach, i.e., by moving a single antenna element along the 32 positions while keeping the single receive antenna fixed. This approach is also used in other studies like [100] and [95] for channel measurements of 112 or 128-element antenna arrays. Since the channels of the entire antenna array needed to be measured at different times, the physical environment must be static. The interaction between the channel and mobile objects is hard to characterize for the synthetic approach of channel measurements. Therefore, the channel measurements were conducted in a weekend, all the objects were kept stable and no humans were moving in the measurement area. The synthetic approach of channel measurement is valid for static or a quasi-static environment, which is in fact the case of indoor scenarios. The complexity and cost for building a massive MIMO testbed that can conduct simultaneous channel measurements is very high. At this moment, only a few are available, for which details can be found in the survey in [106]. 4.5 Capacity Analysis The channel measurements provide data on the channel matrix H, which is then used in the subsequent analysis. Note that the raw channel data also contains the path loss, so the

69 54 Massive MIMO in Indoor WLANs: Background and Insights channel vectors are first normalized to have h k M [68]. For i.i.d. channel of each antenna configuration, 2000 Monte Carlo simulations are done to produce the result. In the following analysis, we will consider different antenna array sizes M. This is achieved by reusing the measured channel data. Specifically, every M adjacent antennas from the 32 measured positions is treated as an M -element antenna array. Therefore, we can have 32 M 1 possible choices, and all of them are analyzed. The aperture size of the array is given by L 0.06M (m). K In order to focus on the spatial multiplexing gain, we assume 0, such that each M user has the same received SNR. As we can see from Eq. (4.3), under the favorable propagation condition, we have the interference-free sum-rate Csum, IF Klog 2(1 ), which is a constant for a specific K and Impact of Channel Conditions For this analysis, we only consider DPC, which gives the upper bound capacity. We fix K = 4, and vary M. The SNR is = 20 db. The interference-free sum-rate is bps/hz. The average sum-rates with standard deviations versus antenna array size in different channel conditions are plotted in Fig Fig. 4.2 Average sum-rate capacity versus the number of antennas with DPC

70 4.5 Capacity Analysis 55 For co-located STAs, the difference with the i.i.d. channel is quite large. As explained earlier, this is because the small spatial separation of the STAs creates high channel correlations. On the contrary, for distributed STAs, the performance is very similar to the i.i.d. channel case Impact of Antenna Array Size on Sum-rates of the Precoding Techniques The linear precoding techniques are usually employed in practice because of their low computational complexity. An interesting question is how they perform with respect to the antenna array size. We show the results in Fig The sum-rates of the precoders are indicated by the ratio with respect to the interference-free sum-rate. For the analysis here, we randomly choose K from all the 14 measured positions. The SNR is = 20 db. The interference-free sum-rates for K = 4 and 8 are bps/hz and bps/hz, respectively. We observe that the sum-rate capacity experiences a significant increase with M for all precoders. Overall, MRT performs much poorer than ZF, i.e., it has more stringent requirements on the channel orthogonality between the STAs. Therefore, many more antennas are needed to perform as well as ZF. In contrast, ZF can closely approximate DPC, and is able to achieve more than 80% of interference-free capacity with a relatively small M. However, many more antennas are needed in the measured channel than in the i.i.d. channel for the same performance. For example, to achieve 80% interference-free sumrate for K =8 with ZF, 14 antennas ( L =0.84 m) are needed under the i.i.d. channel assumption, while 22 antennas ( L =1.32 m) are needed for the realistic measured channel case, nearly two times as much. Therefore, a large number of antennas and hence large physical size of massive MIMO antenna array is potentially required, which might be a problem for indoor scenarios. In addition, the improvement with M quickly disappears, as the curves almost level off for large M. Hence, to achieve the ideal favorable propagation condition is difficult. It is more reasonable to make a trade-off between the antenna array size and the performance. For example, for K =8 with ZF, the gain is marginal after M =24, and therefore M =24 may be a good choice. Another question that one may ask is why using more antennas with simple precoding instead of using conventional MIMO with sophisticated precoding. We stress that using large antenna arrays not only improve the spectral efficiency but also the energy efficiency. The energy efficiency improvement is proportional to M as we can see from Eq. (4.3), i.e., it is mainly related to the number of antennas not the precoding. Regarding spectral efficiency, the improvement with an increasing number of antennas is also attractive. For example, in the measured channel with K =8, M =8 for DPC offers the same capacity as M 13 for ZF, i.e., ZF only needs 5 more antennas. Note that DPC is too complex to implement in practice, but ZF can be realized much more easily [17].

71 56 Massive MIMO in Indoor WLANs: Background and Insights (a) K = 4 (b) K = 8 Fig. 4.3 Average sum-rates of different precoding techniques versus the number of antennas. 4.6 Conclusions and Future Work This chapter introduced the basics of massive MIMO, needed for the rest of the thesis, and also presents performance analysis results in indoor environment. We have evaluated

72 4.6 Conclusions and Future Work 57 the impact of channel condition, antenna array size, and precoding techniques, in order to provide insights in the expected behavior of massive MIMO. The results show that real indoor channels generally offer, to a noticeable extent, less spatial degrees of freedom than the ideal i.i.d channel, implying that lower capacity is achieved for the same antenna configuration. Regarding precoding, ZF requires much less antennas than MRT thus should be more suitable for practical use. A large number of antennas is needed to achieve a high capacity gain, however, the sum-rate saturates quickly with an increasing number of AP antennas. Moreover, for getting the same capacity, many more antennas are needed in real channel condition than when i.i.d. channels are assumed. We conclude therefore that the physical size of massive MIMO antenna array is possibly large, especially in the microwave frequency bands. This may create a noticeable problem for indoor WLANs. Note that our measurements were conducted in an office scenario with a linear antenna array placed on a wall. It is possible that if we consider a significantly different room size, a different antenna array configuration, or a different placement of the antenna array (e.g., the ceiling), the results may be different. For getting a more comprehensive understanding of the realistic performance of massive MIMO, all these factors deserve more investigations in future research.

73 58 Massive MIMO in Indoor WLANs: Background and Insights

74 59 Chapter 5 Massive MIMO with CAS and DAS Architectures 5.1 Introduction In this chapter, we discuss massive MIMO in a CAS and DAS architecture, which we refer to as centralized and distributed massive MIMO, respectively. Most recent studies about massive MIMO only consider the centralized solution [107]. A major motivation of centralized massive MIMO is that, in cellular networks, the large antenna arrays can be conveniently installed on the existing base stations. However, for WLANs, given the size of the antenna array for microwave frequencies, a CAS may be impractical. On the other hand, a DAS, although requiring a higher-cost installation, can benefit from the smaller path loss and the independent fading of the channels. This chapter is organized as follows. Section 5.2 reviews the literature on massive MIMO. Section 5.3 presents the objectives of this chapter and the methodologies used. Section 5.4 provides the system model and the mathematical formulation of the sum-rates for different precoding techniques. Section 5.5 introduces the channel modeling methods, and Section 5.6 provides the analysis. Section 5.7 concludes this chapter. 5.2 Literature Review Analyses of centralized massive MIMO systems are abundant. [108] and [100] report some results about the channel correlations and the sum-rates based on measured channels. [17] introduces the Argos testbed that demonstrates the feasibility of massive MIMO with a 64-element antenna array, using ZF and MRT for precoding. The viability of distributed massive MIMO is discussed in [109], and it shows that a higher average data rate can be obtained than the centralized solution. Moreover, [82] shows how many antennas are needed to achieve a certain percentile of the ultimate sum-rate performance limit. However, these studies mainly look at the improvements of scaling up the antenna array size for a specific numbers of users. On the other hand, it is also interesting to know, for a given massive MIMO system, what the optimal number of users is. This question has not been fully answered yet. In addition, the definition of optimal may depend on the specific performance metric. Lately, [98] [97] reported mathematical derivations about how many users and pilots should be scheduled in centralized massive MIMO to achieve

75 60 Massive MIMO with CAS and DAS Architectures the maximum throughput. It is shown that the optimal number of users has the following relationship: K TcWc /2, where T c is the channel coherence time, Wc the coherence bandwidth, and the pilot reuse factor. The very recent work in [110], [88] and [111] discusses the energy efficiency of massive MIMO. Many parameters, including transmission power and circuit power consumption, e.g., for linear processing, site cooling, etc., determine the energy efficiency. This kind of analysis is rather complex, and the results may not be conclusive since technologies change over time and the implementations are subject to the choice of the manufacturers. A general conclusion is that there is an optimal number of users, which is of the same order as the number of AP antennas. This is interesting as it is contrary to the common belief that the number of AP antennas should be many times larger than the number of co-scheduled users. However, these analyses have some shortcomings: the channels are assumed to follow an i.i.d. Rayleigh distribution; only CAS is considered in [110], [88] and [111]; [98] and [97] assumes the transmission power is controlled in such a way that the path loss differences between users are compensated. In addition, all of them focus on cellular network setups, which by nature is different from indoor WLAN. Specifically, in cellular networks, the cell size is much larger, the transmission power is usually also much higher, the physical channel characteristics are much different for outdoor environments, and the frequency bands are also different. This could lead to different results. Without doubt, valuable insights have been obtained. However, considering indoor network parameters, distributed massive MIMO, non-i.i.d. Rayleigh channel, other power control algorithms, additional analysis is required. 5.3 Objectives and Methodologies The objectives of this chapter are two-fold. The first is to know how much gain can be obtained by centralized and distributed massive MIMO in comparison with traditional MIMO. The second is to find the maximal capacity achievable by CAS and DAS. Moreover, in DAS, the antennas can be distributed over different number of RAUs. The question is what level of distribution is optimal? The answers to these questions depend on the following factors: the number of antennas, the precoding technique, the propagation channel characteristics, and, the number and the placement of the RAUs. In particular, the propagation channel is a critical factor for the performance of a MIMO system, thus deserving special attention. Ray-tracing is employed in our analysis to accurately model the channels in indoor environment [112]. Raytracing, however, may underestimate the multipath effect due to the simplification of the simulation scenario. For example, we usually only consider the main objects like the walls, floors and ceiling, and neglect the others objects such as furniture. With that consideration in mind, we also employ a hybrid channel model that assumes i.i.d. Rayleigh distribution for small scale fading but keeps the average channel gain the same as the one obtained by ray-tracing. By comparing the results of the two channel models, we can

76 5.4 System Model 61 better understand the impact of the small scale fading on the performance. In addition, for both CAS and DAS, the position of the RAUs is optimized based on the coverage performance [67], which will be specified in Section System Model Centralized and distributed massive MIMO are illustrated in Fig For centralized massive MIMO, the antennas are co-located, therefore, RoF links are not necessary. On the other hand, distributed massive MIMO requires optical fibers to connect the RAUs with the AP. We denote the number of RAUs, the number of antennas per RAU, and the number of co-scheduled single-antenna STAs by M, N and K, respectively. So the total number of AP antennas is MN. For convenience, centralized massive MIMO is taken as a special case of distributed massive MIMO, where M =1, and consequently N equals the total number of antennas. 1 STA 1 Access Point Wireless Channel N STA K (a) Centralized massive MIMO 1 N RAU 1 STA 1 Access Point Optical fiber RAU 2 Wireless Channel RAU M STA K (b) Distributed massive MIMO Fig. 5.1 Architecture of centralized and distributed massive MIMO. Before introducing the signal model, the following assumptions are made: User scheduling is not applied. User scheduling could improve the MIMO performance but would require channel knowledge of all candidate STAs. Obtaining this

77 62 Massive MIMO with CAS and DAS Architectures would create a high overhead. However, massive MIMO in principle can achieve the optimal performance without sophisticated user scheduling due to its capability of high spatial resolution [15], which we have explained in Chapter 4. We assume that the STAs are uniformly distributed over the entire coverage area. Certain cases where users are highly clustered, like in a lecture hall, may affect the results, but are not considered in this chapter. We assume a sum power constraint (SPC), i.e., the total transmission power to the STAs at the AP is limited. In addition, equal power allocation to the STAs is applied. This makes the problem formulation much easier, without significantly distorting the results [113]. We will verify in Chapter 7 that the sum-rate with the stricter constraint, per antenna power constraint (PAPC), yields similar results as SPC when transmission power optimization is applied. More details about optimizations under SPC and PAPC can be found in [114]. We assume perfect channel knowledge at the AP in order to simplify the analysis. In Chapter 6, we will focus on the channel feedback problem, and establish the effect on the performance of imperfect channel knowledge Signal Model The signal model for block-fading narrow-band channels is given in the following. Let us denote the full channel matrix by G=[ g1,, g K ], where g k, k =1,, K, is the channel vector of the k -th STA. The channel coefficients consist of both large scale fading and small scale fading. Specifically, T T g = 1 k 1k, k h h T, Mk Mk (5.1) where mk is the large scale fading and h mk is the small scale fading vector between the m -th RAU and the k -th STA, respectively. The received signal vector at the STAs is expressed by y = G H x n (5.2) where x is the symbol vector after precoding, n is the Gaussian noise vector with i.i.d. elements following the distribution (0, 2 2 ), and is the noise power Sum-rate Capacity Let us formulate the sum-rate capacity for the different precoding techniques, DPC, ZF and MRT, that were introduced in Chapter 4. The sum-rate with interference-free (IF) assumption, which determines the upper bound, is also taken into account for reference. The mathematical formulation is elaborated below. However, in contrast to Chapter 4, we

78 5.4 System Model 63 assume that the total transmission power is given by p P0 MN, where P 0 is a predefined constant. And the transmission power is equally allocated to the K STAs. Therefore, the total transmission power is scaled down by MN to offset the power gain from precoding, i.e., the transmission power is preserved by MN times. In this way, the received SNR of the STAs is kept the same for the same antenna configuration. Therefore, for CAS, the capacity improvement is solely because of the spatial multiplexing gain. For DAS, both spatial multiplexing gain and power gain can be achieved, and the same amount of power is preserved as in CAS. The interesting point of this power assumption is that we can achieve a significant capacity gain while saving the transmission power MN times. Note that transmission power, including the dissipation at the power amplifiers, accounts for the major part of the power consumption in a wireless network [115]. This power scaling can also be applied to the uplink transmissions, see, e.g., [116] and [117]. Therefore, a power saving at the mobile devices can also be achieved Interference-free Assumption If we assume that the channels of the STAs are mutually orthogonal, there is no inter-user interference. Then the data rate for a random STA, say the k -th, is given by [16] where R p g = log 1 k IF, k is the noise power. The sum-rate is given by 2 (5.3) C sum,if = K RIF, k (5.4) k =1 Given that interference-free is an ideal situation, if the precoders provide a similar performance as interference-free, we say that the precoder achieves a close-to-optimal performance. However, this requires MN >> K, which may not be practical as the physical size of the antenna array is always limited. For all the precoders, the average sum-rate is denoted by C sum, where the average is taken over the random channel realizations due to the random position of the STAs. Since the K STAs are selected independently, the average sum-rate of interference-free is Csum, IF KCIF, singleuser, where C IF,singleuser is the average data rate when there is only a single user. Therefore, for a fixed K, C sum, IF is a constant DPC Precoding DPC can achieve the optimal sum-rate which is given by

79 64 Massive MIMO with CAS and DAS Architectures 1 H Csum,DPC max log 2det I GPG 2 (5.5) P where P is a diagonal matrix representing the power allocation to the STAs, i.e., P =diag{ p1,, p K}. In general, the power allocation to the STAs for DPC is obtained K by the water-filling algorithm [96], subject to the constraint p k 1 k P t. In the high SNR regime (much larger than 0 db), it can be approximated by [118] C P H = log deti GG K (5.6) total sum,dpc 2 2 i.e., with equal power allocations to the STAs. Therefore, we use the equal power allocation in this chapter as the power used in the analysis ensures SNRs at the STAs in the high SNR regime Linear Precoding Beamforming with the linear precoders, ZF and MRT, has the following general received signal model 1/2 y= G H WP sn (5.7) where s is the information symbol, and W is the beamforming matrix. Therefore, the SINR of the k -th STA is given by SINR = k jk g w H 2 k k k H 2 2 k j p j g w p (5.8) The beamforming matrix W is obtained through the following procedure. First, the nonnormalized precoding matrices of MRT and ZF are calculated as follows F = G (5.9) MRT H 1 F G G G (5.10) ZF = ( ) Then, the columns are normalized to get unit-power beamforming vectors. Specifically, in the final beamforming matrices, W ZF and W MRT, the columns are obtained by H wk = fk f k. Since for ZF, gi w j =0 for i j, the first term of the denominator of H 2 2 Eq. (5.8) will be zero, then we have SINR g w p. k k k k Finally, the sum-rate for linear precoding is calculated by

80 5.5 Channel Models 65 C 5.5 Channel Models sum,zf/mrt K log 2 k,zf/mrt (5.11) k =1 = (1 SINR ) Let us now present the ray-tracing simulation method and the hybrid channel model used in the subsequent analysis. The ray-tracing simulation is used to characterize the propagation channel more realistically. The hybrid channel model is used to describe the ideal rich scattering environment, i.e., the small scale fading is i.i.d. Rayleigh distributed. However, we keep the average channel gain between any RAU and any STA to be the same in these two channel simulation methods. Thus the two approaches result in the same received SNR Ray-tracing Simulation The ray-tracing simulations are done with the software tool RPS [74]. The details of the simulation methodology are presented in Appendix A. The selected scenario is based on one of the floors of the Flux building of the Eindhoven University of Technology, which is illustrated in Fig The main objects considered in the ray-tracing simulations are the walls, floor and ceiling. The walls are plastered, and ceilings and floors are of concrete, typical materials of large surfaces of office buildings. As we can see, the STAs can be in both LOS and NLOS of the RAUs. In particular, the central area is mostly in LOS of a number of RAUs, thus a high signal strength is obtained but high channel correlation may also be caused. To improve the coverage with a limited number of RAUs, the positions of the RAUs are optimized through the following procedures. First, we divide the ceiling plane into a lattice with 5 m grids where the candidate RAU positions are located. Then, the group of RAU positions that provides the highest 5-percentile received signal power over the entire coverage area is selected. A greedy-search algorithm is used to find the optimal positions. The basic principle of greedy search is to iteratively add one RAU to maximize the objective function (see [67] for more details). Another reason for choosing this algorithm is that the optimization result is deterministic and therefore reproducible. As an example, Fig. 5.2 shows the cases of 4 and 16 RAUs, of which the optimal positions are indicated by the dots. The color map shows the received signal strength when assuming 0 dbm transmission power at each RAU.

81 66 Massive MIMO with CAS and DAS Architectures (a) M= Hybrid Channel Model (b) M=16 Fig. 5.2 Simulation scenario. The i.i.d. channel creates the ideal conditions for MIMO spatial multiplexing [100], hence is a benchmark. To be consistent with the ray-tracing channels, we have to force the received signal power between any pair of STA and RAU at the subcarriers to be the same. Therefore, we derive the the ideal channels through the following procedure. First, let us denote the average channel gain between the m -th RAU and the k -th STA obtained by ray-tracing channel simulation as mk, which is calculated by the following formula L g RT l 2 mk [] F RT l =1 m k = (5.12) NL where F is the Frobenius norm. g RT mk [] l is the channel vector on the l -th subcarrier, and l =1,, L (RT is short for ray-tracing). As a result, the small scale fading

82 5.6 Performance Analysis 67 h []= l g [] l will have unit average power over all the N antennas and RT RT RT mk mk mk subcarriers. Note that this average channel gain is obtained from an instantaneous channel, which is different from the common large scale fading factor derived from the long-term average. L Next, we denote the small scale fading of the i.i.d. Rayleigh channel as elements follow i.i.d. (0,1). It is then normalized by h IID mk [] l, where the h l NL IID IID mk []= h [] L mk 2 h IID mk [] l F l =1 l (5.13) to obtain unit average power over all the antennas and subcarriers. Finally, the i.i.d. Rayleigh MIMO channel used for the analysis is derived by g h (5.14) IID RT IID mk []= l mk mk [] l As a result, the ray-tracing channels g RT mk [] l and i.i.d. Rayleigh channels g IID mk [] l will have the same average power gain. 5.6 Performance Analysis The numerical results are obtained based on the above models using the parameters listed in Table 5.1. In all the simulations, the transmission power P 0 is fixed at 20 dbm for all antenna configurations, to be consistent for the comparisons. We will first discuss how the sum-rate improves with respect to the antenna array size for the CAS architecture, particularly when compared with a traditional MIMO configuration. Next, we analyze how much improvements DAS can offer with different distribution level M, in comparison to CAS ( M 1 ). Then, we turn to find out the maximum average sum-rate of centralized and distributed massive MIMO systems with different configurations. Finally, we show the maximum average sum-rate of centralized and distributed massive MIMO with different values of MN. Table 5.1 Simulation parameters. Frequency 2.4 GHz Bandwidth 20 MHz Number of subcarriers 64 Background noise -174 dbm/hz Noise figure 10 db

83 68 Massive MIMO with CAS and DAS Architectures Transmission power Scenario =20 dbm 100 m40 mm (STAs at 1 m height) Massive MIMO versus Traditional MIMO For the analysis in this subsection, we assume CAS for both massive MIMO and traditional MIMO. For massive MIMO, we assume N ranges from 8 to 64, and for traditional MIMO, N =8. We consider a constant K =8, which is the maximum for traditional MIMO. The average sum-rate C sum versus MN are plotted in Fig As expected, the sum-rate increases with MN and tends to approximate the upper bound with interference-free assumption. In comparison with traditional MIMO, a significant capacity improvement of up to 4 times (ZF in ray-tracing channel) is achievable. But the exact amount of improvement depends on the precoder and the channel. ZF achieves a similar performance as DPC when MN is a few times greater than K. ZF provides much better performance than MRT. MRT sum-rate also improves with MN but quite slowly. Therefore, for the same performance, MRT requires many more antennas. For example, to achieve a 20 bps/hz sum-rate in the ray-tracing channel, ZF and MRT require 16 and 112 antennas respectively. The ray-tracing channel offers a smaller number of spatial degrees of freedom, so we see lower sum-rates than for the i.i.d. channel. To achieve the same spatial multiplexing gain, a much larger antenna array is thus needed. For example, to achieve 40 bps/hz with ZF, less than 16 antennas are needed for the i.i.d. channel case, but 80 antennas are needed in ray-tracing channel. This raises the question whether one should approach the optimal performance for a small number of STAs by using so many antennas at the AP.

84 5.6 Performance Analysis 69 Fig. 5.3 Average sum-rate versus the number of antennas with CAS Centralized versus Distributed Massive MIMO We now investigate how much gain can be obtained by using a DAS architecture. Let us analyze this by keeping MN fixed and vary the number of RAUs M. The average sumrate versus M is shown in Fig. 5.4, where we have assumed MN 64, and M =1, 4, 16, and 64. We see that the improvement with M is significant. However, the improvement levels off when M is beyond 16. Therefore, it is not beneficial to deploy a high density of RAUs. The improvement by DAS with respect to centralized massive MIMO is around 2 times for all precoders. The major reason for the improvement of DAS is due to the increase of received signal strength instead of the more independent small scale fading, as the sum-rate for the ray-tracing channels and the i.i.d. channels is similar. This means that determining the number and the location of the RAUs can be based on the received signal power criteria. However, the analyses in this and the previous subsection consider the case that the massive MIMO antennas are many more than the number of STAs. Next, let us find the maximum sum-rate for a given antenna system configuration by varying the number of STAs.

85 70 Massive MIMO with CAS and DAS Architectures Fig. 5.4 Average sum-rate versus the number of RAUs Maximum Average Sum-rate of Centralized and Distributed Massive MIMO For a given antenna system configuration, i.e., M, N and MN are fixed, we can find the optimal K that gives the highest average sum-rate, which we formulate as the following optimization problem: C sum K1,, MN Csum max C ( K) (5.15) For the following analysis, we consider a constant MN 64, and assume M = 1, 4, 16, and 64. The traditional MIMO configuration is also simulated. For simplicity, only ZF and MRT are considered as they are more likely to be used in practice than DPC. We also define the ratio of improvement over traditional MIMO as C / C (5.16) precoder sum,precoder traditional MIMO Fig. 5.5 and Fig. 5.6 show the results of C sum, ZF and C sum, MRT versus K, respectively. The optimal K s for different cases are also indicated. The exact values of the optimal K, maximum sum-rates, and maximum gain of ZF and MRT can be found in Table 5.2 and Table 5.3, respectively. We can see from the figures that the optimal K depends on the precoder and the channel characteristics. For traditional MIMO, the optimal K is not 8 which we used in the last two subsections. For massive MIMO, ZF has an optimal K smaller than MN, but MRT has the optimal K MN (see the tables). Therefore, for getting higher capacity gain, we

86 5.6 Performance Analysis 71 should schedule a large number of STAs of the same order of the number of AP antennas. A small K or a high MN / K ratio, is optimal only for centralized massive MIMO in channels that have limited spatial degrees of freedom (which restricts the multiplexing gain), i.e., the case of the ray-tracing channel. A larger M yields a higher sum-rate. But the difference between M =16 and 64 is small. Therefore, for achieving maximum average sum-rate, it is not necessary to deploy the antennas with an extremely high distribution level such as M =64. When M becomes larger, the channels are more independent across the antennas, so the average sum-rate is closer to the i.i.d. channel. However, both large scale fading and small scale fading play a role in determining the optimal number of STAs. We observe from the result that, K is very different for ray-tracing channels than for i.i.d. channels for small M (e.g., 1 and 4), but similar for large M (e.g., 16 and 64). The reason is that, for larger M the channels between the RAUs become more independent, while for small M the co-located antennas at the RAU experience strong channel correlation. Therefore, to optimize the performance in practical systems, it is necessary to consider both the large and small scale fading characteristics of the channel when each RAU has a large antenna array. Due to the protocol overhead such as channel feedback, the optimal number of STAs may be smaller than the ones we obtained. We can understand this using a simple formulation K of the PHY layer throughput: Cthroughput (1 CSI ) Csum [102], where CSI is the time Tc overhead of CSI acquisition, and T c is the channel coherence time. Therefore, when K increases, the overhead of CSI will also increase, making the optimal K smaller.

87 72 Massive MIMO with CAS and DAS Architectures Fig. 5.5 Average sum-rates versus the number of STAs for ZF. Fig. 5.6 Average sum-rates versus the number of STAs for MRT.

88 5.6 Performance Analysis 73 Table 5.2 Maximum sum-rates of ZF. IID (M=1,N=8) (M=1,N=64) (M=4,N=16) (M=16,N=4) (M=64,N=1) Optimal K Max sum-rate Gain RT Optimal K Max sum-rate Gain Table 5.3 Maximum sum-rates of MRT. IID (M=1,N=8) (M=1,N=64) (M=4,N=16) (M=16,N=4) (M=64,N=1) Optimal K Max sum-rate Gain RT Optimal K Max sum-rate Gain Maximum Average Sum-rate versus the Total Number of Antennas Finaly, we investigate how the maximum average achievable sum-rates by centralized and distributed massive MIMO depend on the total number of antennas MN. We consider MN values ranging from 8 to 128. For distributed massive MIMO, we select M =16. M =16 has been shown to provide close-to-optimal average sum-rate (see Fig. 5.4), and is more practical than M =64 which results in a very high density of RAUs. The maximum average sum-rates are given in Fig. 5.7 and Fig. 5.8 for ZF and MRT, respectively. We can see that there is a quasi-linear relationship between the maximum capacity and MN. Distributed massive MIMO performs significantly better than the centralized case. Their difference is even larger for ray-tracing channels. This advantage of massive MIMO is more apparent particularly for the ray-tracing channels. For example, DAS offers 4 times more capacity than CAS for MN 64. The gain is higher for a larger MN. Therefore, DAS is strongly motivated for massive MIMO. Since distributed massive MIMO provides higher spatial multiplexing gain and power gain, it requires significantly less antennas for the same capacity. Some examples are shown in the figures, indicating that only a fraction of the antennas in a CAS are required for DAS, for both ray-tracing channels and i.i.d. channels.

89 74 Massive MIMO with CAS and DAS Architectures With the given configurations, MRT provides one-third of the capacity of ZF. Note also that MRT achives the maximum sum-rate by scheduling many more STAs, thus the per STA data rate is lower. MRT beamforming processing can be distributed, which may be an advantage in practical implementations. However, since the signal processing in RoF networks is done centrally at the CS, this advantage is not so appealing. Hence ZF is the better choice for RoF-based massive MIMO systems. We remind the reader that we have assumed the total power is preserved by a factor of MN in the analyses. Despite the multiplexing gains of massive MIMO, we also get an MN times transmission power gain. Although the analyses only consider the downlink, the power gain in the uplink is similar as demonstrated in [111]. Therefore, both spectral efficiency and energy efficiency can have a linear increase with MN, for the optimized number of STAs. Fig. 5.7 Maximum average sum-rates of ZF.

90 5.7 Conclusions and Future Work 75 Fig. 5.8 Maximum average sum-rates of MRT. 5.7 Conclusions and Future Work In this chapter, we have analyzed the average sum-rate performance of CAS and DAS architectures for massive MIMO in indoor WLANs. We have investigated the impact of antenna array configuration, the precoding technique, and the propagation channel, on the performance. The analyses include two parts. First, we have analyzed how much capacity improvement can be achieved by centralized and distributed massive MIMO in comparison with the traditional MIMO configuration for a fixed number of STAs. The results show that, in realistic channels simulated by ray-tracing, the capacity can be significantly improved by centralized massive MIMO due to a finer spatial resolution. Several times more capacity can be offered by using DAS due to the higher received signal strength. Second, we have investigated the maximum capacity by optimizing the number of co-scheduled STAs. It was shown that the optimal number of STAs depends on the precoder and the channel characteristics. For a specific antenna system configuration, scheduling a small number of STAs, as is usually assumed in massive MIMO research, is not always optimal. Our analyses show that, when DAS is used or when the channel has rich scattering, higher order spatial multiplexing is possible, and scheduling a large number of STAs can achieve higher capacity while using the same amount of energy. Therefore, it is better to consider an optimized number of co-scheduled STAs instead of a constant one. As observed from our analyses, the optimization of the number of STAs should consider the antenna system configuration, the propagation channel, as well as the precoding technique.

91 76 Massive MIMO with CAS and DAS Architectures Since distributed massive MIMO offers much superior performance, especially in the realistic channel we simulated by 3D ray-tracing, it is preferred in indoor WLAN. The RAUs that are equipped with small antenna arrays can fit more easily to the indoor environment, so DAS also mitigates the large form factor problem in centralized massive MIMO. If each RAU has a small number of antennas, i.e., a high density of RAUs, the channels between the RAUs are to a large extent independent, so the determination of the position of the RAUs can be simply based on the path loss. However, when each RAU has a quite large antenna array, e.g., more than 16, the small scale fading correlation between the antennas at an RAU also significantly affects the system behavior. In such cases, it is necessary to optimize the number of RAUs and the location of them considering both the path loss and small fading characteristics for real channels. It would be interesting to consider the following for future research. First, other frequency bands of WiFi, e.g., the 5 GHz band, should be analyzed. Second, we have considered perfect channel knowledge in the analyses, while the channel knowledge is imperfect in practice. In particular, the channel sounding signal may be corrupted by noise and interference. Imperfect channel knowledge not only affects the capacity gain but also the power gain [85], therefore need to be taken into account in this sort of analysis. Third, characterizing the indoor MIMO channel for large antenna arrays is also important for optimizing the design and deployment of massive MIMO systems. Fourth, the cases that the RAUs are non-identical, i.e., the RAUs may have different number of antennas; user distribution is not uniform, i.e., users may be spatially clustered, are also interesting to investigate. Finally, in order to get a better understanding of the relationships between the parameters, e.g., the power and capacity gain with the number of antennas, deriving analytical models should be very helpful.

92 77 Chapter 6 Optimizing CSI Feedback for Distributed Massive MIMO Systems 6.1 Introduction Channel State Information (CSI) at the transmitter is critical for the performance of MIMO systems [119]. The typical approach to obtain CSI is to employ a closed-loop scheme: the transmitter transmits a known signal to the receiver, the receiver estimates the channel and then feeds back the quantized CSI to the transmitter. The acquired CSI is only valid for a short period, equal to the channel coherence time which ranges from tens to hundreds of milliseconds in indoor environment [120] [121]. The channel estimation and transmission of the CSI packets consume time and energy, thus it should be done efficiently. It is obvious that this problem becomes more critical when the number of antennas is very large as many more channel coefficients need to be fed back, which is the case of massive MIMO. For example, assuming a channel coherence time of 100 ms, 100 AP antennas serving 20 single-antenna STAs simultaneously, the channel feedback in IEEE n/ac channel training scheme takes more than 13% of the air time 1. This chapter focuses on CSI feedback of distributed massive MIMO. The specific problem we need to solve is how to quantize the CSI with a limited number of feedback bits while still providing good performance, e.g., the sum-rate [122]. In distributed massive MIMO, the large scale fading of the various RAUs may be rather different, so each RAU contributes a different amount of power to the desired signal for a given STA. As a consequence, the channels of different RAUs may need different levels of accuracy, i.e., different quantization resolutions. Moreover, the channels of sufficiently distant RAUs may not need to be fed back at all. We therefore propose two feedback bit allocation schemes based on the large scale fading of the RAUs, which are denoted as adaptive allocation and equal allocation. The adaptive allocation scheme assigns an un-equal number of quanti- 1 Consider the channel of each subcarrier is quantized by 8 bits, the feedback packets are transmitted using the highest data rate (64-QAM with 5/6 coding rate), the symbol duration is 4 s, the channel feedback takes / ms, excluding the overhead of MAC and PHY layer headers and control frames.

93 78 Optimizing CSI Feedback for Distributed Massive MIMO Systems zation bits to the RAUs according to their large scale fading to a given STA. Equal allocation allocates an equal number of quantization bits to a dominant subset of the RAUs that have the least large scale fading. In such a way, the feedback bits are utilized more efficiently, and thus less feedback bits are needed for achieving the same performance. Both schemes can be formulated as convex optimization problems to achieve a low computational complexity. The sum-rate performance is evaluated for both schemes in combination with ZF precoding in distributed massive MIMO, which will be elaborated in Section 6.6. The remainder of the chapter is organized as follows. Section 6.2 reviews the related work on channel feedback compression techniques. Section 6.3 describes the distributed massive MIMO system model. Section 6.4 presents the channel training protocol and the formulation of the feedback bit allocation problem. In Section 6.5, we propose two feedback bit allocation algorithms. Section 6.6 introduces the sum-rate model of ZF precoding in distributed massive MIMO taking into account of imperfect channel knowledge. Section 6.7 provides the simulation results, and Section 6.8 concludes the chapter. 6.2 Literature on CSI Compression Techniques for Massive MIMO The MIMO channel feedback can be reduced by exploiting the channel correlations in the frequency, time, and space dimensions [119]. In particular, for massive MIMO, the colocation of the antennas may cause high channel correlations between the neighboring antennas in the spatial dimension. There are a number of papers that focus on channel feedback reduction in centralized massive MIMO systems, e.g., [123], [124], [125] and [126], which mainly exploit the spatial channel correlation. In [123], compressive sensing (or sparse sampling) is introduced to take advantage of the channel sparsity in spatial domain. In [124], joint spatial division and multiplexing is proposed, under the assumption that users are clustered so the channels between the users in the same cluster are correlated. In [125], joint compression in both spatial and temporal dimensions is considered. In [126], grouping of adjacent antennas and mapping the channel of the grouped antennas into a single representative value is proposed. However, indoor distributed massive MIMO systems may not benefit from the techniques based on spatial correlation. First, the antennas distributed at different RAUs are independent due to their large spatial separations. Second, in indoor environment, the antennas at the RAUs can be separated by a small distance (in the order of a number of wavelengths) to de-correlate the channels, due to the rich-scattering [127]. Note that low channel correlation is beneficial for achieving a higher spatial multiplexing gain.

94 6.3 System Model System Model Consider a distributed massive MIMO system (see Fig. 5.1) that consists of M RAUs, each equipped with N antennas, and serving K single-antenna STAs simultaneously. When the number of active STAs is larger than K, the other STAs will be scheduled in other time slots. For the subsequent modeling and analysis, we make the following general assumptions: - Each STA is served by only those RAUs for which CSI feedback is provided. In Section 6.4, we will introduce the feedback bit allocation algorithms. Due to the optimization in the feedback bit allocation, the channel of some distant RAUs are not fed back. - The channel estimation error is negligible in comparison with the channel quantization error. This is justified by the fact that, the serving RAUs are close to the STAs (the last assumption), so the received training signal power is high. - The large scale fading of the active RAUs is known at the STAs. The large scale fading can be estimated at the STAs based on the channel training signals. In addition, the large scale fading changes much slower than small scale fading, therefore it can be tracked much more easily [128]. 6.4 CSI Feedback Scheme and Feedback Bit Allocation Problem This section introduces the channel training and feedback for distributed massive MIMO based on the framework of IEEE n/ac. We first introduce the IEEE n/ac channel training protocol, and then formulate the feedback bit allocation problem IEEE n/ac Channel Training AP Pilot Polling Polling Beamformed frame STA 1 CSI feedback STA 2 CSI feedback Fig. 6.1 Multi-user channel training procedure in IEEE n/ac. Time

95 80 Optimizing CSI Feedback for Distributed Massive MIMO Systems The channel training protocol defined in IEEE n/ac is illustrated in Fig The reader can also refer to [7] and the references therein for more details. Briefly, the channel acquisition procedure is the following: the AP first broadcasts a training signal using all the antennas, then the STAs estimate and quantize the channels, and finally the AP polls the STAs sequentially to acquire the CSI feedback packets. We assume that the CSI feedback frames are received error-free. This is a reasonable assumption since the frames are received via polling, i.e., collision-free. In addition, the dense deployment of the RAUs guarantees high received signal SNRs, e.g., the frames are received from the RAUs near to the STAs. Note that, the CSI feedback packets cannot be received simultaneously, as there is no uplink multi-user multiple access in IEEE n/ac. This results in an inefficient channel utilization in the CSI feedback procedure. The size of the CSI feedback packet increases with the amount of feedback bits. So more feedback bits will cause higher overhead, as well as more energy consumption and higher latency. IEEE n defines three CSI feedback types, i.e., the types of content in the feedback packets: plain CSI feedback, non-compressed beamforming weights, and compressed beamforming weights. In IEEE ac, only the last one is supported because the industry has not been able to converge on one particular choice. In plain CSI feedback, the STAs quantize the estimated CSI with the same number of bits, e.g., 8 bits for each real and imaginary component. Since only scalar uniform quantization (SUQ) is needed after channel estimation, the computational complexity of plain CSI feedback is low. The amount of feedback overhead (in terms of the amount of CSI quantization bits) per subcarrier can be calculated by OCSI where b is the number of quantization bits of each subcarrier. MNb (6.1) In the feedback using non-compressed beamforming weights, the STAs compute the beamforming weights and then send them to the AP (any beamforming algorithm can be used). The overhead of this method is identical to that of plain CSI feedback [7]. In compressed beamforming weights feedback scheme, the standard proposes to use the Givens rotations to convert the unitary matrix V in the SVD of the channel matrix to a number of angles. The angles are quantized and sent to the AP. The quantized V is used for transmitter precoding at the AP. For each single-antenna STA, the amount of feedback is given by [7] where OCompressed 2( MN1)( b' 1) (6.2) b ' is the number of quantization bits for the angles.

96 6.4 CSI Feedback Scheme and Feedback Bit Allocation Problem 81 Therefore, the amount of overhead reduction by compressed feedback with respect to the plain CSI feedback is 1 O Compressed 1 (1 1 ) b CSI 2( ' 1) (6.3) O MN b As we can see from this equation, the overhead reduction is higher for smaller MN, given a specific b and b '. For massive MIMO, MN is large, so the overhead reduction can be approximated by 2( ' 1) (6.4) b Compressed 1 O b 1 OCSI Therefore, for a traditional MIMO system, the feedback reduction is significant. As shown in [129], to achieve similar quantization distortion as plain CSI feedback with b =14 (i.e., 7 bits for the real component and 7 bits for the imaginary component), the required b ' is 7 bits. For example, when MN =4, the reduction is around 14%. However, for massive MIMO, the number of bits used in compressed feedback is even larger than the plain CSI feedback (observe from Eq. (6.4)). In addition, the computational complexity of SVD is 2 2 of the order of O(min{ mn, m n} ) for an m n matrix [130], which may be very high in the case of massive MIMO. Hence, the plain CSI feedback is better than the compressed feedback scheme for massive MIMO when using IEEE n/ac. In this chapter, we consider using the plain CSI feedback scheme. The allocation algorithms will be introduced in the subsequent sections. However, our feedback bit allocation algorithms either allocate non-equal number of feedback bits to different antennas, or only feedback a selected subset of antennas, which is different from the original IEEE n/ac scheme. Since the channel training protocol, training signal, and the format of the feedback packets, are already standardized in IEEE n/ac, our scheme does not require major amendments to the standard. Specifically, the minor amendments that is needed to support our scheme, is to allow multiple quantization levels, e.g., from b =2, 4, to 16, in every CSI feedback packet; and support selective feedback for a subset of the antennas Feedback Bit Allocation Problem We start with the formulation of the quantization distortion. The quantization distortion function is a performance measure of a quantization scheme, the lower it is the better [131]. Similar to its time-domain formulation given in [122], we define the quantization distortion function for distributed massive MIMO narrow-band channel as

97 82 Optimizing CSI Feedback for Distributed Massive MIMO Systems 2 ˆ ˆ 2 k E k k E( mnk mnk ) mn, mn, D h h h h D (6.5) where the expectation is taken over the random vector h k, and D mnk is the quantization distortion at the mn -th antennas of the k -th STA. The second equality follows due to the independency of the channels between the antennas. From now on, we will omit the generic STA index to simplify the expressions. The quantization distortion at the mn -th antenna is given by [131] k mnk Qmn /21 ( i1) mn mn Dmn ( Bmn, mn ) 2 i mn f( ) d i mn i0 2 mn 2 ( Q 1 ) ( ) ( Qmn 2 ) mn /2 mn f d (6.6) where B is the number of bits for quantizing the channel of the mn -th antenna ( B 2 for mn the real and the imaginary components), mn Q 2 B mn mn is the number of quantization steps, mn is the quantization step size, and / mn mn f( ) 1 e [132]. The constraint of limited feedback bits can be written as 2 B B. mn, mn tot Clearly, the quantization distortion function depends not only on the number of bits, but also on the channel variance, and the choice of the quantization step size. Given and B mn, finding the optimal mn requires a numerical search. In order to simplify this, we employ the rule-of-thumb design given in [131]: the overload point is equal to 4 times the standard deviation of the Gaussian random input, such that the overload probability (i.e., the probability of the input being out of the quantization range) is negligible. The quantization step size is then given by mn Bmn 2 mn mn 2 (6.7) Bmn 2 When the number of quantization bits B mn is large, the distortion function in Eq. (6.6) is simplified into [131] mn

98 6.5 Feedback Bit Allocation B 2 2 mn mn Dmn (6.8) 3 This large B assumption is valid since most of the feedback bits are allocated to a small m number of RAUs close to the STA in the our optimization algorithms to be introduced in the next section. With the above results, we can formulate the quantization distortion as a function of the feedback bits 16 16N D( B 2 (6.9) 2Bmn 2 2Bm 2 ) 2 mn m mn, 3 3 m where the second equality is due to the fact that he antennas at the same RAU have the same large scale fading. B m is the number of bits for each antenna at the m -th RAU, and B [ B,, B ] T M. So the limited feedback constraint is expressed as 1 B tot ' Bm Btot (6.10) m 2N Since the bit allocations are based on large scale fading, it suggests a low frequency of updating the bit allocations, as the large scale fading varies slowly in indoor scenarios [128]. The de-correlation distance of large scale fading is of the order of 1 to 2 m [133], and may be as large as 5.3 m [128]. Therefore, for a walking speed around 1 m/s, the update frequency is around once per second or less. For users that are not moving, e.g., are seated, the large scale fading is almost static, so updates can be much less frequent. Finally, we formulate the feedback bit allocation problem as minimize D( B) s.t. Btot Bm 2N m (6.11) B, m 1,,M M are the variables need to be opti- where D( B ) is given by Eq. (6.9), mized. 6.5 Feedback Bit Allocation For solving the feedback bit allocation problem, we propose two schemes: m Adaptive allocation, in which the bits are allocated by directly solving Eq. (6.11). Equal allocation, in which the number of feedback bits is the same for all the selected RAUs.

99 84 Optimizing CSI Feedback for Distributed Massive MIMO Systems In the equal allocation scheme, the problem is to find a dominant RAU subset. It simplifies the original multi-variable optimization problem to a single-variable optimization problem, which in general has a lower complexity Adaptive allocation The distortion function is convex (which can be easily proved by its second-order derivative [134]), so the feedback bit allocation problem can be solved by convex optimization techniques. The only issue is that B m should be integer, however, convex optimization algorithms usually apply to continuous variables. So we have to find the solution in two steps. First, we use convex optimization techniques to get the continuous-value solutions. Then, we use a floor function to get the nearest integer solution that fulfills the total feedback bit constraint. Generally, the complexity of the convex optimization does not exceed a polynomial of the problem dimensions [134]. We can also apply the Karush Kuhn Tucker (KKT) conditions to get an analytical solution. This turns the original problem into a water-filling problem for which there are many efficient solvers [135]. We show the derivations in the following. The Lagrangian function is constructed as [134] So the KKT conditions are 16 N vb B (6.12) 3 2B 2 ' 2 m m ( m tot ) m m 16NN Bm 3 B B 0 m m ' tot 2 2B 1 ln 2 2 m m v 0 By solving the above equations, one can verify that the bit allocation solution is (6.13) Bm log 2 m (6.14) where ( x) max( x,0), and is the water level chosen such that Bm Btot 2N is satisfied. Typically, water-filling algorithms have a complexity of the order of OM ( ) [136]. As we can see from Eq. (6.14), since the optimal is a constant, for larger m, B is larger given log2 0. So more bits will be allocated to stronger RAUs. If m 2 m 0 m log, no feedback bits will be allocated to that RAU. m

100 6.5 Feedback Bit Allocation Equal allocation Intuitively, the RAUs with the strongest channel should be active as they contribute the most to the power of the desired signals. As a simplified approach, we then equally allocate the feedback bits to the set of RAUs that have the strongest channel gains. The original problem can then be simplified to finding the optimum number of strong RAUs. Equal bit allocation is also considered in multi-cell cellular networks described in [132]. The difference is that, in previous works, only up to 3 or 5 RAUs are considered, so an equal number of bits can be allocated to all the RAUs. But there may be multiple tens of RAUs in distributed massive MIMO, hence, allocating the feedback bits equally to all RAUs is not feasible. Therefore, only an optimum subset is selected in our scheme. The equal allocation algorithm is given below: Step 1: Sort the path-loss coefficients in descending order such that ( m) ( m 1), m 1,,M M, where ( m) is the RAU index before sorting. Step 2: Use exhaustive search to obtain M arg min D( B ), M' 1,, M (6.15) M ' Btot where B ( m), for m1,,m M, and B ( m) 0, for m M' 1, 2 NM ',M M. Step 3: Return the optimum number of RAUs M, and the bit allocations: B m Btot 2NM, for the selected RAU set { ( m) m 1,, M }, and B 0 for the other RAUs. The number of search steps in Step 2 is M. In fact, D has a convex shape, so the search can be stopped at the M 11 step, since if D at M 1 1 is larger than M, we can infer M is optimal. The allocation algorithms are given for narrow-band channels. Since the channel is highly correlated in the frequency dimension, for wide-band channels, adjacent subcarriers can be grouped and then a single channel matrix is fed back for each group [7]. This reduces the overhead multiple times. Groupings of every two subcarriers and four subcarriers are permitted by the IEEE n/ac standard, reducing the overhead to a half and a quarter, respectively [7]. For different subcarrier groups, the bit allocation solution is the same since the large scale fading is the same. However, the bit allocation schemes aim at minimizing the quantization distortion, which is not directly related to the data rate. The achievable data rates also depend on other parameters, in particular, the channels of all the co-scheduled STAs and the precoding techniques. Although we will see that adaptive allocation achieves a lower distortion than m

101 86 Optimizing CSI Feedback for Distributed Massive MIMO Systems equal allocation, it is not directly observable how much higher the sum-rate is. In the next section, we formulate the STA sum-rates with the imperfect CSI obtained from the feedback process. 6.6 Sum-rate Model with Imperfect CSI Let S [ S ] denote the relation between the M RAUs and the K STAs, where Smk mk M K 1 means the m -th RAU is active for serving the k -th STA (i.e., the k -th STA feeds back the CSI of the m -th RAU), and Smk 0 means the m -th RAU is not active. We further define the following notation: k { { m Smk 1, m 1,, M } is the set of RAUs serving the k -th STA; c k {1,, M }\ k is the set of RAUs not serving the k -th STA; k k k k is the set of all active RAUs. Let us denote the full channel matrix by where hk is the channel vector of the k -th STA, H[ h,, h K ] [ h ] (6.16) 1 K mk mk MNK mk is the large scale fading and h mk is the small scale fading vector of the m -th RAU and the k -th STA. The small scale fading follows an i.i.d. Rayleigh distribution, which is an acceptable assumption for small antenna arrays due to the rich scattering in indoor environments. With the channel quantization error due to limited feedback bits, we can represent the narrow-band channel matrix as H=H+E ˆ (6.17) where Ĥ is the quantized feedback channel matrix, and E is the quantization error matrix. The received signal model at STA k is given by where H y ˆ ˆ k h H k w k ps k k h k w j p js j n k (6.18) desired signal j k interference due to channel error w ˆ k is the ZF beamforming vector, p k is the transmission power (normalized by noise power),

102 6.7 Performance Analysis 87 s k is the data symbol and n k is the noise following (0,1). ZF has an acceptable computational complexity [17], and has been shown to perform well for massive MIMO in Chapter 5, so it is considered for our analysis. The beamforming vectors are calculated, based on the quantized channel matrix, denoted by H ˆ ( ). The rows of H corresponding the non-active RAUs (the RAUs in c ) are eliminated, thus H ˆ ( ) is of size N RAU K. The details of calculating the ZF precoder are omitted here but can be found in Chapter 5.4. Then, the instantaneous data rate of the k -th STA is given by Rˆ log 1 k H h wˆ p k k k 2 2 H 1 h ˆ k w j p j jk 2 (6.19) If perfect CSI is assumed to be known at the AP, the beamforming vectors are calculated based on H ( ). So, the second term of Eq. (6.18) will be zero due to the orthogonal property of ZF precoding. Therefore, Eq. (6.19) can be simplified into H k 2 k k k 2 R log 1 h w p (6.20) where wk is the beamforming vector obtained based on perfect CSI. The sum-rate is calculated by Csum k Rk, and C ˆ ˆ sum k Rk for perfect and imperfect CSI, respectively. The transmission power is given by M pk Pt for k 1,, K (6.21) K where P t is the transmission power limit of each RAU. That is, the total transmission power is equally allocated to the STAs. 6.7 Performance Analysis In this section we investigate the sum-rate performance with respect to the feedback bit allocation methods, via simulations in MATLAB. The two feedback bit allocation schemes we proposed, i.e., adaptive and equal allocation, are examined. For comparison, we also show the case of perfect CSI, and the IEEE scheme where an equal number of feedback bits are allocated to all antennas. Furthermore, we also look at the influence of the DAS distribution level, by considering three values for the number of

103 88 Optimizing CSI Feedback for Distributed Massive MIMO Systems RAUs M = 4, 16 and 64, for the same total number of antennas MN =64. A larger M provides a higher degrees of freedom for the proposed bit allocation algorithms. The simulation parameters are listed in Table 6.1. The values of the parameters are typical for the analysis of IEEE WLAN, see e.g., [137] and [138]. For each case, 1000 groups of STAs (each group consists of K STAs that are uniformly random distributed over the coverage area), and 100 random small scale fading channel matrices for each STA group, are simulated. Table 6.1 Simulation Parameters. Scenario Number of RAUs 4, 16, 64 Total number of antennas 64 Number of STAs 16 Positions of RAUs Uniformly placed at 4 m height (ceiling) Positions of STAs Uniformly random distributed over 1 m plane above the floor Frequency band 2.4 GHz Bandwidth 20 MHz Noise power 174 dbm/hz Noise figure 10 db Transmission power 10 mw/rau Path loss model IEEE n/ac Model B (open space and office environment) [38] Small scale fading i.i.d. Rayleigh The average sum-rate versus the total number of feedback bits per subcarrier B tot for the different bit allocation schemes are shown in Figure 6.4. The error bars indicate the standard deviations of the sum-rate. The annotations show the amount of feedback overhead of adaptive allocation with respect to the IEEE scheme, for achieving 90% average sum-rate with perfect CSI.

104 6.7 Performance Analysis 89 (a) M =4 (b) M =16

105 90 Optimizing CSI Feedback for Distributed Massive MIMO Systems (c) M =64 Fig. 6.2 Sum-rates for different number of RAUs and bit allocation schemes. It can be seen that, the average sum-rates of equal and adaptive allocations converge at small and large B tot. This is because, for a small B tot, the allocation methods lead to the same solution; for a large B tot, the quantization distortion is small, the sum-rate for all feedback bit allocation schemes are close to the perfect CSI case. Adaptive allocation schemes perform much better at intermediate B tot. Therefore, for small B tot, adaptive and equal allocations are equivalent. For a larger B tot, e.g., more than 128, adaptive allocation should be the better choice. In comparison with the IEEE scheme, the proposed bit allocation schemes significantly reduce the amount of feedback overhead, for achieving the same average sumrate. For equal allocation, the major gain is when the number of feedback bits is small or when target average sum-rate is low, i.e., the left end of the curves. The equal allocation and the IEEE schemes perform similarly at a large B tot, since equal allocation scheme assigns the feedback bits to all RAUs, the same as the IEEE scheme. It is more interesting for getting a high data rate, e.g., 90% of the average sum-rate with perfect CSI. For that target, equal allocation is not superior to the IEEE scheme. But, adaptive allocation requires only 80%, 57%, and 62.5% of the amount feedback bits as the IEEE scheme for M =4, 16, and 64, respectively. That is, the overhead reduction is around 20% to 43%. This suggests that adaptive allocation is the best of all the evaluated schemes, and is a promising solution for the feedback bit allocation in distributed massive MIMO.

106 6.8 Conclusions and Future Work 91 In addition, the feedback bit allocation schemes result in CSI feedback for only the closeby RAUs. This means that only a subset of RAUs needs to be trained for a particular STA. STAs that are close to each other see the same set of adjacent RAUs. Thus the channel sounding procedure can be further simplified by grouping the STAs, i.e., the training signal only need to be transmitted through these common neighboring RAUs. Then the signal processing load of the STAs, like channel estimation and quantization, can be reduced. 6.8 Conclusions and Future Work In this chapter, we have proposed two feedback bit allocation schemes: adaptive allocation and equal allocation, for indoor distributed massive MIMO. The feedback bit allocation schemes allocate the feedback bits according to the large scale fading between the RAUs and the STAs. They are then evaluated by comparing with the perfect CSI case and the IEEE scheme, in terms of the sum-rate with ZF precoding. The simulations have shown that both the adaptive and equal allocation schemes can significantly reduce the required number of feedback bits. Equal allocation is more advantageous to the IEEE scheme when the number of feedback bits is small, and is equivalent when it is large. Adaptive allocation is equivalent to equal allocation for a small number of feedback bits, but for most cases it performs better than equal allocation. Adaptive allocation can reduce the feedback bits by more than 20% compared to the IEEE scheme, for achieving 90% of the average sum-rate with perfect CSI. So, overall, adaptive allocation is superior to equal allocation and the IEEE scheme. The proposed feedback bit allocation schemes are based on the channel training mechanism in IEEE n/ac. To apply the ideas, minor amendments to the standard are required: a new format of CSI feedback packet that can support multiple quantization levels, and selective feedback for only a subset of the antennas. Future research should address the following issues. First, a limited range of quantization levels, that could simplify the implementation, should be considered. Our analyses did not restrict the maximum number of feedback bits for an antenna. However, it is not necessary to use a large number of feedback bits, e.g., 8 bits per real and imaginary component is sufficient as the quantization distortion is already very low. Second, the computational load of channel estimation and quantization will be high at the STAs, while in fact not all antennas contribute significantly to them. Moreover, the problem of training all antennas at the same time is not efficient when the number of antennas is very large. Therefore, the combination of user grouping and CSI feedback should be considered.

107 92 Optimizing CSI Feedback for Distributed Massive MIMO Systems

108 93 Chapter 7 Performance Comparison of Small Cell Networks and Distributed Massive MIMO 7.1 Introduction In the previous chapters, we have addressed different issues about SCN and distributed massive MIMO. Both SCN and distributed massive MIMO will significantly boost the network capacity. An important question is which of them will provide the higher capacity and by how much. In order to answer this question, we use the sum-rate (see Chapter 4), as the performance measure. We will assume that both SCN and distributed massive MIMO use the same RoF network architecture for carrying the RF signals between the RAUs and the central AP(s). The analysis will be conducted for both the 2.4 GHz (microwave) and the 60 GHz (mmwave) frequency bands as they experience rather different propagation characteristics. Their channels are modeled individually. We then present a unified signal model for the sum-rate analysis that can be used for both SCN and distributed massive MIMO, and for all the frequency bands. Based on this signal model, we formulate the sum-rate optimization under two power constraints: sum-power constraint (SPC) where the total transmission power is limited, and per-antenna power constraint (PAPC) where the transmission power from each antenna is limited. SPC is more straightforward and the transmission power optimization problem is easier to solve. However, each transmit antenna typically has its own power amplifier, and thus PAPC is more realistic. The remainder of this chapter is organized as follows. We first review the related works in Section 7.2. The system model for RoF-based SCN and distributed massive MIMO is introduced in Section 7.3. Next we present the channel models for 2.4 GHz and 60 GHz in Section 7.4. The signal model and the sum-rate optimization algorithms are elaborated in Section 7.5. Then we present the analysis results in Section 7.6, and draw the conclusions in Section Literature A number of recent papers, [139], [110] and [140], have presented performance comparisons of SCN and massive MIMO. They investigate the issue from different perspectives.

109 94 Ray-tracing Channel Simulation In addition, they all aim at outdoor cellular networks and frequencies below 6 GHz. Millimeter wave bands are not considered yet. In particular, [139] compares the energy efficiency of SCN and centralized massive MIMO systems using the same total number of antennas, and finds that SCN performs better. [110] analyzes the energy efficiency gain with respect to cell densification. It concludes that the energy efficiency improvement saturates with the cell density. However, by using larger antenna arrays (i.e., massive MIMO) in the cells, further improvement can be obtained. [140] evaluates the per-sta throughput in SCN and distributed massive MIMO using cellular network parameters. It is found that distributed massive MIMO provides more uniform per-sta throughput than SCN, or equivalently, fairer quality-of-service across the STAs. However, the paper considers MRT precoding only, which limits the insights. We have shown in previous chapters that MRT does not perform well in distributed massive MIMO indoors. 7.3 System Model Let us introduce the system-level model of RoF-based SCN and distributed massive MIMO. We illustrate this in Fig Assume that there are M RAUs in the RoF system, each equipped with and simultaneously serving K STAs that each with N RAU antennas, N STA antennas. The total number of STAs in the entire area may be larger than K. In such a case, they have to be grouped and served in different time slots using TDMA. For consistent comparisons, we consider the same RoF infrastructure is used for SCN and distributed massive MIMO. The same RoF infrastructure implies that the set of RAUs, frequency bands, and maximum transmission power is the same. For SCN, M APs are co-located at the CS. Each AP is connected with one RAU. We assume that an AP only serves one STA at a time. This setup is essentially the idea of SCN, i.e., each cell only needs to serve a small number of STAs when the cell area is small. In addition, when the traffic is denser in certain locations, load-balancing techniques can be used to equalize the load across the APs [141], e.g., by controlling the association of the STAs with the APs [142]. We assume an STA is associated to the AP that has the smallest path loss to it. For serving the K STAs, K APs as well as the K connected RAUs need to be active. When M K, the other M K APs do not transmit signals. The signals simultaneously transmitted from multiple RAUs may interfere with each other, so the transmission power of the antennas is optimized to maximize the average sum-rate; this is elaborated in Section For distributed massive MIMO, for the sake of consistency, we assume that the same set of RAUs as in the case of SCN, are active. This assumption not only guarantees that the same number of antennas is used, but also that the transmission power limitations are the same, in order to give a fair comparison.

110 7.4 Channel Models 95 External network /Internet Central Station AP 1 AP 2 AP 3 AP M Radio-Optical Interface Optical link Optical Network 1 N RAU RAU 1 RAU 2 RAU M Wireless channel STA K STA 1 (a) SCN External network /Internet Central Station AP Radio-Optical Interface Optical link Optical Network 1 N RAU RAU 1 RAU 2 RAU M Wireless channel STA K STA 1 (b) Distributed massive MIMO Fig. 7.1 RoF-based SCN and distributed massive MIMO. 7.4 Channel Models GHz For 2.4 GHz, we assume that each STA has a single antenna, i.e., NSTA 1. Let us denote the general narrow-band channel matrix of the active RAUs by G g ij (7.1) KNRAU K where g ij is the channel vector between the j -th STA and the i -th active RAU, and i, j 1,, K. The channel vectors consist of both large and small scale fading. Specifically, gij ijh ij, where ij is the large scale fading scalar, and h ij is the small scale fading vector. The specific models for large and small scale fading are arbitrary for the modeling in this chapter. In fact, due to rich scattering in indoor environments, placing the antennas one or two wavelengths apart is likely to be sufficient to get independent channels [18]. So we assume i.i.d. Rayleigh small scale fading channel for the antennas in the same RAU. The IEEE n/ac path loss model is used to characterize the path loss between the RAUs and the STAs. The IEEE n/ac path loss model is [38] PL FS( d) X( LOS ) d d PL(dB) PL ( d) X( ) 10nlog ( d / d ) d d FS NLOS 10 BP BP BP (7.2) where

111 96 Ray-tracing Channel Simulation PL FS( d ) is the free space path loss at a distance d, d BP is the breakpoint distance, n is the path loss exponent after the breakpoint distance, and X ( ) is the log-normal shadowing component with standard deviation of. The values of these parameters can be found in [38] GHz In contrast to 2.4 GHz, the spatial paths in 60 GHz channel are sparse, and the channel is dominated by the LOS path. The small scale fading variations have a minor impact on the channel [143]. In addition, the signals are easily blocked by objects. Therefore, in practice, the 60 GHz RAUs should be placed in LOS of the STAs. For those reasons, we only consider the direct path, i.e., the LOS path, in the 60 GHz channel model, and ignore the reflective paths or NLOS paths [144]. At the same time, we employ the IEEE ad path loss models [145]. Let us now derive the channel model. Let us denote the full channel matrix by G [ G1,, G ] T K, and Gk [ Gkm] [ kmh km] NSTA KN RAU, where km is the path loss between the m -th RAU and the k -th STA, and H km is the small scale fading matrix considering the LOS paths given by [103] 2 j d km H km NRAU NSTA e STA, km RAU, km H e e (7.3) where e is the unit spatial signature in the direction from the m -th RAU to the k -th STA expressed as [103] 1 2 e STA, km exp( j n ) N kx (7.4) STA NSTA e RAU, km exp( j n ) N kx (7.5) RAU NRAU 1 where n is the index of the antenna elements in the antenna arrays at the STAs or the RAUs,

112 7.4 Channel Models 97 xk xm is the unit direction vector given by k, where x m and x k are the positions xk xm of the RAU and the STA arrays, respectively, as shown in Fig k x n is the local coordinates of the n -th antenna element in the RAU or STA arrays. Ceiling plane x m RAU antenna array z x y x k Half a wavelength STA antenna array STA plane Fig. 7.2 Coordinates of the RAU and the STA and the antenna array configurations. The path loss model is given by [145] A 20 log10( f) 10 n ( d) X( ) d d PL(dB) A 20 log10( f) 10 n ( d) X( ) d d LOS LOS LOS BP NLOS NLOS NLOS BP (7.6) where A LOS and A NLOS are constants given in [145], d is the distance between the transmitter and the receiver, f is the carrier frequency, X ( ) is the log-normal shadowing component with standard deviation, d BP is the breakpoint distance. For the following channel modeling, we first make these assumptions: - Each STA has only a single radio-frequency (RF) chain, which is the usual case due to the consideration of the costs and power consumption of the RF chains [146]. Thus the 60 GHz STAs cannot receive multiple spatial streams. In fact,

113 98 Ray-tracing Channel Simulation the small antenna arrays at the STAs cannot achieve much multiplexing gain due to the poor scattering in the channel [147]. - Each STA steer its beam to the strongest RAUs where most of the received signal energy comes from. The other RAUs that are far away contribute with significantly less signal power. In addition, the receive beams of the STAs can be adaptively controlled by tuning the phases of the signal transmitted at each antenna via a network of analog phase shifters, implemented in the RF domain [146]. - There are MN RAU RF chains at the CS, therefore, the CS can perform fully digital beamforming for all the antennas. As a result, we allow the CS (the APs at the CS) to perform MU-MIMO spatial multiplexing. With such assumptions, we can have the following received signal model where H C is the receive beamforming matrix at the STAs, and G ', of size K KN RAUs The receive beamforming matrix is given by H yc GxnG' xn (7.7), is the effective channel matrix after receive beamforming. c C H 1 c H K H (7.8) where c k is the receive beamforming vector at the k -th STA. c k is given by, where m is the index of the strongest RAU to the k -th STA. c k e H STA, km We then have G' [ g ' 1,, ' ] T H T H T g K, where g' k ( ck Gk) [ ck kmh km] 1 KN. RAU c H H k km km km N km N 2 j dkm H RAU STA e erau, km N m m 2 j dkm H H RAU STA e e e STA, km STA, kmerau, km N m m (7.9) We can see that G ' H has the same size as the channel matrix G for 2.4 GHz. They are used later for getting the unified sum-rate model.

114 7.5 Unified Sum-rate Model Unified Sum-rate Model In this section, we present a unified signal model for both SCN and distributed massive MIMO. We assume that perfect channel state information (CSI) is available at the AP. The perfect CSI assumption mainly serves as the basis for finding the upper bound of the achievable data rate. The sum-rate model will be derived using G to represent the channel matrix for the case of 2.4 GHz, but we can replace it by G ' H for the 60 GHz analysis. The same approach is also applicable to the precoding techniques and the power allocation algorithms. The received signal model for both SCN and distributed massive MIMO is given in a general linear beamforming form [68] where W is the beamforming matrix, diag p1,, pk H 1/2 yg WP sn (7.10) P is a diagonal matrix containing the power allocations to the STAs, s is the information symbol, and n is the noise vector following 2 0, I where 2 is the noise power. W and P will be defined in the following subsections Beamforming The beamforming matrices for SCN and distributed massive MIMO are given in the following. We remind the reader that, for a given STA in SCN, only the antennas of the single RAU belonging to its associated AP, is used for transmitting signals to that STA. This is different from distributed massive MIMO where all the active antennas are used to transmit to all STAs. Let us derive the beamforming matrix W for SCN and distributed massive MIMO. SCN: We apply MRT at each RAU, such that the beamforming vector of the k -th STA is given by wkk gkk / g kk, where g kk is the channel between the k -th STA and its serving RAU. T T T T Then we can write wk 1, kk, 0 w 0 k N KkN, where 0 is the zero vector. 1 K W w,, w is used in Eq. (7.10). Applying MRT at the RAUs can reduce inter-cell

115 100 Ray-tracing Channel Simulation interference as signals are more concentrated on the target STAs and less directed to the other cells. Distributed massive MIMO: W w1,, w K is obtained based on the selected precoding techniques. In this chapter, we consider only ZF. The other linear precoder, MRT, has been employed in the previous chapters but it performs much poorer than ZF. The precoding matrix of ZF is obtained through the following procedure. First, calculate the non-normalized beamforming matrix H 1 F G G G (7.11) ZF ( ) Then obtain the normalized beamforming vectors by w f / f, k 1,, K, where f k is the k -th column of F. k k k H Then we define the effective power gain matrix as Q, in which Qij gi w j represents the power gain of the signal for the j -th STA to the i -th STA [148]. We further define a normalized power gain matrix Q where Let Qij / Qii if j i Q ij (7.12) 0 if j i DQ diag Q11,..., Q KK be the diagonal matrix representing the direct link channel gains, such that, the desired signal power vector can be given by and the total interference and noise power is given by the vector As a result, the SINR for the k -th STA is given by IN p D P (7.13) S Q Q p Q D Pp (7.14) n 2 SINR k p p k (7.15) S, k IN, The sum-rate capacity can then be derived as C ( WP, ) log (1SINR ) (7.16) sum 2 k 1 K which is a function of the beamforming matrix W and the power allocation matrix P. k

116 7.5 Unified Sum-rate Model Transmission Power Optimization Next, we consider different power constraints and optimize the power allocations to maximize the sum-rate. The optimization is done for a given beamforming matrix W that was introduced earlier for SCN and distributed massive MIMO. As a result, the optimization problem is about finding the optimum power allocation matrix to maximize C ( P ). sum The transmission power is limited by the power amplifiers in the RF chains. The maximum output power has to be kept within a limited level, in order to keep the amplifier in the linear operation range [149]. We consider two power constraints: sum power constraint (SPC), and per-antenna power constraint (PAPC) [150]. SPC assumes that the total transmission power from all the active antennas (in the downlink) is limited. PAPC limits the transmission power for each antenna, which should be done when each antenna is fed with an individual RF chain [15]. SPC is therefore a more relaxed constraint than PAPC, and it mainly serves as a theoretical benchmark. PAPC is a more realistic constraint, as usually the antennas are fed with individual RF chains. The sum-rate optimization problems with SPC and PAPC are formulated as follows: For SPC P maximize C s. t. K p k k 1 P sum max ( P) (7.17) p k is the transmission power allocated for the k -th STA, and max P is the maxi- where mum sum transmission power. For PAPC maximize C ( P) sum K 2 max n wnk pk k 1 RAU (7.18) s.t. p p, for n1,, KN R max where p is the maximum transmission power at the antennas, n is the index of the antennas from all the active RAUs, and KN RAU is the total number of antennas. In addition, to allow a fair comparison between the two power constraints, we set max max KN p P. RAU

117 102 Ray-tracing Channel Simulation It can be seen that, the constraints are linear functions of the transmission powers, thus the constraints are convex. The objective function, however, is not a convex function of the transmission powers (see Eq. (7.15)). Hence, the optimization problem is non-convex and is intrinsically intractable [134]. Fortunately, at high SINR regime, the optimization problem can be turned into a standard Geometrical Programming (GP) problem which can be transformed into a convex form with a logarithmic change of all the variables [151]. A high SINR is actually required in practice to get a high SINR in order to achieve high data rates, so we claim that this approximation is reasonable. Nevertheless, it may not always be valid. This approximation implicitly favors higher SINRs but neglects low or zero SINRs. However, sometimes, assigning zero transmission power to some STAs offers a higher sum-rate, like the waterfilling algorithm does. Let us briefly summarize the process of transforming the original optimization problem into a GP problem. More details can be found in [151] and [152]. In the high SINR regime (SINR>>0 db), the data rate log 1 SINR R can be approximated by log2 SINR k. So the following objective functions are equivalent k 2 k maximize Csum k Rk (7.19) (7.20) maximize log (SINR ) log ( SINR ) k 2 k 2 k k minimize 1/SINR k k (7.21) The last one is the new objective function after the approximation. By substituting the SINR equations Eq. (7.15) into this function, we can easily figure out that it is a posynomial function of the transmission powers [151]. Then, the original sum-rate optimization problem is transformed into a GP problem whose objective function is K minimize 1 p p (7.22) k 1 1 IN, k S, k This GP problem can be solved by convex optimization techniques. Although the function in Eq. (7.22) is not a standard posynomial function, it can be processed by a convex optimization toolbox to get numerical results. For our analysis, we use CVX, a MATLAB package for specifying and solving convex programs [153]. In addition, the sum-rate optimization with ZF beamforming can actually be largely simplified since inter-user interference is non-existent (provided the perfect CSI assumption), i.e., pin p n. Thus, the objective function can be rewritten as

118 7.6 Sum-rate Analysis Sum-rate Analysis K minimize 1 k 1 1 n, k S, k p p (7.23) Let us now present the numerical results of the sum-rate analysis based on simulations in MATLAB. The goal of our analysis is to compare the average sum-rate performance of SCN and distributed massive MIMO. We present the results from three perspectives: (1) The sum-rate versus the number of STAs, which shows the performance behavior when more STAs are simultaneously served. More STAs being served at the same time corresponds a higher network traffic load. (2) The sum-rate versus the number of antennas per RAU to investigate the impact of the antenna array sizes. This analysis will give guidelines on the choice of the number of antennas in the RAUs. (3) The sum-rate versus the per-antenna power limit to examine the effect of the transmission power. This gives us insight in how much transmission power is needed for SCN and distributed massive MIMO. The simulation parameters are listed in Table 7.1 and Table 7.2 for 2.4 GHz and 60 GHz, respectively. The parameters are typical for indoor WLAN, e.g., see [137] and [138] for 2.4 GHz, and [154], [145], and [155] for 60 GHz. The scenarios for 2.4 GHz and 60 GHz are given in Fig. 7.3, where the square areas are of size 10 m 10 m, representing the average room size. To capture the effect of inter-cell interference in SCN, we can use a smaller scenario size for 60 GHz band than for the 2.4 GHz case. This is because 60 GHz signal has a much higher path loss than 2.4 GHz. So we reduce the area size of 60 GHz to 1/4 that of the 2.4 GHz case. The total number of antennas is smaller, thus we can simplify the computational complexity of the optimization algorithms in the analysis, without distorting the results. We will give the results of 2.4 GHz first, followed by 60 GHz. The results show different patterns in the two frequency bands. Hence, when we discuss the 60 GHz case, their differences will also be discussed. In all the figures, the average sum-rates are shown, together with the 95% confidence intervals. Table 7.1 Simulation Parameters for 2.4 GHz. Scenario Inter-RAU distance Positions of RAUs Positions of STAs Frequency band 10 m ( M =64 in the scenario) Placed on a uniform grid with 10 m spacing, at 4 m height (ceiling) Uniformly random in each cell, over 1m plane above the floor 2.4 GHz

119 104 Ray-tracing Channel Simulation Bandwidth Noise power Path loss model Small scale fading 20 MHz 174 dbm/hz IEEE n/ac Model B (open space and office environment) [38] i.i.d. Rayleigh Table 7.2 Simulation parameters for 60 GHz. Scenario Inter-RAU distance 5 m ( M =64 in the scenario) Positions of RAUs Placed on a uniform grid with 10 m spacing, at 4 m height (ceiling) Positions of STAs Uniformly random in each cell, over 1m plane above the floor Frequency band 60 GHz Bandwidth 2 GHz Noise power 174 dbm/hz Path loss model IEEE ad: cubicle environment [145] A LOS =32.5, n LOS =2, LOS =0; A NLOS =44.2, n NLOS =1.8, NLOS =1.5; d BP =5 m. Small scale fading Direct path only (a) Simulation scenario for 2.4 GHz.

120 7.6 Sum-rate Analysis GHz (b) Simulation scenario for 60 GHz. Fig. 7.3 Simulation scenarios for 2.4 GHz and 60 GHz Average Sum-rate versus the Number of STAs We assume the power constraint: max p =20 dbm, which is the maximum transmission power at 2.4 GHz by European regulations [50]. In fact, 0 dbm transmission power in a single-input single-output (SISO) configuration can already achieve as high as 23 db SNR at 8 m (cell boundary). So this assumption results in an unrestrictive power constraint. We also assume the number of antennas per RAU is NRAU 1 (MRT is then not applicable in SCN in this analysis). The results are given in Fig Due to the inter-cell interference in SCN, the capacity is bounded. Therefore, it is not possible to serve many STAs simultaneously and provide high data rates to each of them at the same time. We see that, at around K =16, the average SINR per STA is already as low as 10 db. Therefore, in an SCN, simultaneous communication with more than 16 STAs is not feasible. Assume each room has a size of 10 m 10 m, this is equivalent to allowing only 1 active user every 4 rooms. For supporting more STAs simultaneously, it is necessary to apply frequency reuse so that all STAs can communicate with satisfactory data rates. For example, the 2.4 GHz ISM band has three 20 MHz channels, so roughly three times more users can be served at the same time. In contrast, the sum-rate of distributed massive MIMO increases almost linearly with the number of STAs. The average achievable data rate per-sta is higher. For example, at K

121 106 Ray-tracing Channel Simulation =16, the sum-rate of distributed massive MIMO is around 5 times that of SCN. This shows that, with the same spectrum and the same power constraint, many more STAs can be served simultaneously in distributed massive MIMO than SCN. This is a strong motivation to employ distributed massive MIMO, rather than SCN, for the 2.4 GHz band. It can be seen that the difference between using SPC and PAPC is negligible. The average sum-rates for SCN with SPC and PAPC constraints are similar as their curves overlap. For distributed massive MIMO, the performance difference is only noticeable when many STAs are served simultaneously. The PAPC constraint makes the actual transmission power for the STAs smaller than SPC, since the transmission powers of all the STAs need to be low enough to make sure the aggregated transmission power at every antenna is smaller than the maximum [156]. Fig. 7.4 Sum-rate versus the number of STAs for 2.4 GHz Average Sum-rates versus the Number of Antennas per RAU The performance of SCN and distributed massive MIMO can be improved with more antennas per RAU, which is analyzed in this section. Briefly, the more antennas deployed at the RAUs, the less the interference with the other cells in SCN as the signal radiation is more directional. The same effect can be obtained in distributed massive MIMO. In fact, the additional antennas should always improve the sum-rate performance obtained by the linear precoders [16]. For the following analysis, we fix K =16, and vary N RAU. If a larger K is assumed the observations below still hold. However, when K is large, SCN will perform poorly (see

122 7.6 Sum-rate Analysis 107 Fig. 7.4). This also implies a very low data rate per STA, which is not desirable. In addition, we use a different power assumption compared to the previous subsection: =100 mw/ N RAU, such that the total transmission power per RAU is limited to 20 dbm. The results are presented in Fig max p Fig. 7.5 Sum-rate versus the number of antennas per RAU for 2.4 GHz. Clearly, both SCN and distributed massive MIMO benefit from a larger N RAU. The improvement is actually more significant for SCN. The capacity is almost doubled when N RAU increases from 1 to 8. Distributed massive MIMO has a smaller but still noticeable gain. We can expect that the gain is a little higher if a larger N RAU is used, provided it is feasible. The limitation on N RAU may be because of the large form factor of the antenna arrays especially at low frequencies, the complexity of the signal distribution network, as well as the channel feedback overhead. However, the improvement is quite marginal beyond around N RAU =4. So for a given K, it is not necessary to have larger N RAU. Ultimately, distributed massive MIMO with ZF achieves around 4 times the average sumrate of SCN. Again, there is no significant difference between the outcomes when applying the SPC and PAPC in SCN. But, a small difference can be observed in distributed massive MIMO. The explaination is that when more antennas are used, less power is actually radiated. Note that, as inter-user interference is non-existent with the assumption of perfect CSI in ZF precoding, the performance is limited by the received signal power.

123 108 Ray-tracing Channel Simulation Sum-rates versus the Power Constraint max Let us now analyze the impact of the maximum transmission power, i.e., p, on the max max sum-rate capacity (note that P KNRAU p ). We assume that K =16, and N RAU =4. As we can see from the analysis in the previous subsection, 4 antennas per RAU can already provide close-to-optimum performance. The per antenna maximal power varies from -20 to 10 dbm, but we do not explicitly limit the per RAU transmission power as we did in the last subsection. The results are given in Fig max p Fig. 7.6 Sum-rate versus the maximum transmission power of antennas for 2.4 GHz. We can see that the sum-rate generally increases with the per-antenna power limit for both SCN and distributed massive MIMO. However, for SCN, the sum-rate marginally increases and is tightly upper-bounded due to inter-cell interference. Increasing the transmission power beyond a certain value is not beneficial: max p -5 dbm is sufficient for both SPC and PAPC. For this reason, the average sum-rate stays flat with increasing maximum per-antenna transmission power. Distributed massive MIMO is a different case. As mentioned before, ZF is power limited, max so the sum-rate increases linearly with p. Therefore, a higher per-antenna transmission power, i.e., higher transmission power per-sta, is always helpful, not considering max interference to other neighboring network. At p =-5 dbm, distributed massive MIMO achieves twice the capacity of SCN. For larger, the improvement is even higher. max p

124 7.6 Sum-rate Analysis GHz Sum-rate versus the Number of STAs We assume max p =10 mw, which is the typical output power of 60 GHz power amplifiers [13]. The number of antennas per RAU is N RAU =4, and the number of antennas per STA N STA is =4, which yields an SNR of 16 db at the cell edge (around 4 m from the RAU). The average sum-rates C sum versus the number of STAs K is given in Fig Fig. 7.7 Sum-rate versus the number of STAs for 60 GHz. We can see that the sum-rate increases linearly with K for both SCN and massive MIMO. This is very different from the 2.4 GHz case. The major reason is that SCN is much less affected by inter-cell interference in the 60 GHz band. There are two reasons of the lower inter-cell interference. One is that the signals from the farther RAUs are much weaker than the nearest ones for a given STA. The other is that the directional beams at the RAUs and STAs reduce the amount of inter-cell interference. However, SCN provides lower sum-rate than distributed massive MIMO when K is large, e.g., 32. The explanation is that the inter-cell interference now becomes more noticeable. Nevertheless, the difference between SCN and distributed massive MIMO is marginal even for K =32. This means SCN can still support a high spatial reuse. Assuming each room has an average size of 10 m 10 m, this means more than 2 active STAs per room can be allowed using the same frequency band in SCN. This is a significant advantage of using the mmwave frequencies.

125 110 Ray-tracing Channel Simulation Sum-rate versus the Number of Antennas per RAU We can use larger antenna array sizes at the RAUs to get narrower beams which helps to achieve a higher beamforming gain and reduce inter-cell interference, resulting in improved performance. For this analysis, we only consider SPC because we observe from the analyses in the last subsection that PAPC provides a similar performance as SPC, while SPC makes the power allocation problem easier. We assume K =16 and 32, which correspond to averagely 1 and 2 STAs per room that are receiving simultaneously. Typically, we assume one AP per room [145], i.e., at most one STA is transmitting or receiving a frame per room at the same time. So we refer to K =16 as a moderate traffic load case, max and K =32 a high traffic load case. The transmission power is assumed to be p =10 mw/ N RAU, such that the total transmission power per AP is limited to 10 mw which is a typical value for 60 GHz [13]. The average sum-rate C sum versus the number of antennas per RAU N RAU is shown in Fig We can see that, for both SCN and massive MIMO, the sum-rates improve logarithmically with N RAU due to higher received signal power. This is because the Shannon capacity formula (see e.g., Eq. (7.16)) is a logarithm function of the received signal SINR. Although the increase slows down, a significant improvement can be obtained. It is likely that for NRAU 36 there is still further improvement of sum-rate. The reason is that 60 GHz communication is more power limited, thus increasing N RAU can improve the signal strength to achieve higher data rates. This is very different from the 2.4 GHz case for which the performance improvement with N saturates at a small number. RAU

126 7.6 Sum-rate Analysis 111 Fig. 7.8 Sum-rate versus the number of antennas per RAU for 60 GHz Impact of the Power Constraint Let us finally analyze the relationship between the average sum-rate and the transmission power limit. For 60 GHz band, the typical power limit is 10 dbm, a constraint imposed max by the power amplifier [13]. We limit the maximum per antenna power to p =0 to 30 dbm, but, we do not explicitly limit the per RAU transmission power. We consider, like in the previous section, two values of K, 16 and 32. We consider also two different antenna array sizes for the RAUs and STAs, i.e., N RAU = N STA =4, and N RAU = N STA =16,

127 112 Ray-tracing Channel Simulation which are both feasible in practice [157]. In Europe, the Effective Isotropic Radiated Power (EIRP) constraint is as high as 57 dbm [158]. The maximum EIRP using the above parameters is around 54 dbm, so the power assumption is still within the regulated range. The average sum-rate C sum versus the transmission power limit max p is given in Fig We already argued that the performance of SCN is restricted by inter-cell interference when the transmission power is high. In contrast, distributed massive MIMO offers more max than 25% more capacity when p >=30 dbm. The major difference is observed when there are a large number of active STAs and the antenna arrays at the STAs and RAUs are small, specifically, for the case that K =32 and N RAU = N STA =4. Note that, the 60 GHz band is mainly used for short-range communications in indoor, e.g., 10 m [60]. The highest MCS defined in the IEEE ad standard requires an SNR of around 24 db [159] [78]. So a higher transmission power is not necessary. For example, for K =16 and 32, with an average SINR of 24 db at each STA, the sum-rates are around 128 bps/hz max and 255 bps/hz, respectively. As we can see from the figures, the required p is less than 15 dbm. In this transmission power range, distributed massive MIMO does not provide much higher capacity than SCN. Therefore, we can conclude that SCN is the better choice than distributed massive MIMO for 60 GHz. Specifically, SCN can provide very similar sum-rate performance as distributed massive MIMO, however, distributed massive MIMO has much higher implementation complexity, e.g., due to channel training and precoding, as we argued in Chapter 1.

128 7.7 Conclusions and Future Work 113 Fig. 7.9 Sum-rate versus the maximum transmission power per antenna for 60 GHz. 7.7 Conclusions and Future Work In this chapter, we have analyzed the average sum-rate capacity of SCN and distributed massive MIMO in the 2.4 and the 60 GHz frequency bands. We have investigated the effect of the number of antennas used at the RAUs and the STAs, as well as the transmission power, on the average sum-rate.

129 114 Ray-tracing Channel Simulation For the 2.4 GHz band, the path loss is relatively small and the signals can propagate over long distances than 60 GHz signals. SCN is mainly limited by inter-cell interference, thus the sum-rate is bounded. Distributed massive MIMO performs better since the beamforming techniques avoid interference between the spatial streams to the STAs. We have shown that distributed massive MIMO can provide more than twice the average sum-rate of SCN. The improvement increases with the transmission power. Therefore, distributed massive MIMO is the better choice for the 2.4 GHz band. The results for the 60 GHz band are much different. The 60 GHz signals have a high path loss, and the communications are restricted to LOS. Hence, the inter-cell interference is not the dominant problem of SCN. In addition, for a given STA, most of the signal power comes from nearby antennas. The narrow beams of the RAUs and the STAs can also avoid the inter-cell interference. Therefore, we observed that distributed massive MIMO does not provide significantly better performance than SCN, except when the STAs and RAUs have small antenna arrays, and the transmission power is high. Due to the usual operating conditions of 60 GHz WLANs where the transmission power is low and the communication range is around a few meters, a high transmission power is not needed nor is feasible. For such realistic cases, SCN provides very similar average sum-rate performance as distributed massive MIMO. Therefore, SCN is the better choice for 60 GHz indoor networks, given the considerable implementation complexity of distributed massive MIMO. Further evaluations need to consider the following. First, the effect of imperfect channel knowledge, which degrades the performance of SCN and distributed massive MIMO, needs to be analyzed. Second, one needs to analyze the impact of interference from coexisting networks using the same frequency band, e.g., legacy WiFi networks. Third, frequency reuse should be investigated as proper frequency planning helps to achieve a higher capacity for SCN, but this has not been considered in our analyses. Fourth, in our analysis of SCN, we did not consider the use of SU-MIMO and MU-MIMO spatial multiplexing in each cell. SU-MIMO and MU-MIMO spatial multiplexing can improve the per-ap capacity, which may shrink the performance difference between SCN and distributed massive MIMO. The enhancement of SCN with SU-MIMO and MU-MIMO should be an interesting extension of our work. Finally, the other frequency bands of WiFi, e.g., the 5 GHz band, should also be examined.

130 115 Chapter 8 Work Conclusions and Future 8.1 Conclusions In this thesis, we have investigated the means to improve the network capacity of small cell networks (SCN) and massive MIMO in indoor wireless networks, based on an RoF architecture that connects a central station (CS) to multiple radio access units (RAU) via optical fibers. RoF allows a lot of flexibility in configuring the system in a cost-effective way, by concentrating all the system functionality above the physical layer in a central entity, the CS. We have addressed two problems in SCN. - First, we have identified the hidden node problem in the uplink media access of DAS when applying the IEEE MAC. Two mechanisms have been proposed to solve the problem: hard switching and soft switching. Hard switching employs a time-division multiplexing scheme for an AP to serve the multiple cells, while in soft switching the AP performs a selective reception of simultaneous uplink packets from the RAUs. The merit of these switching mechanisms is that no changes to the standards are required in order to implement them. The essence of our solutions is to avoid the direct combination of the RF signals from the RAUs at the APs. We have shown that both the switching mechanisms are effective, but hard-switching provides a higher throughput. Without these switching mechanisms, i.e., using the basic DCF, the uplink throughput drops close to zero when the user density is high. In contrast, we have demonstrated that hard switching improves the throughput by more than 10 times for a two RAU case with medium packet size, and performs even better for more RAUs and higher STA density. - Second, we have proposed to use DAS to mitigate the human shadowing problem for 60 GHz. We have considered two beamforming strategies: blanket transmission and selective transmission. Both show similar performance, but the selective strategy has a lower implementation complexity and hence is a better choice. In addition, we have investigated the trade-off between using more RAUs and using larger antenna arrays. Using more RAUs, each with a small antenna array, or using fewer RAUs equipped with larger antenna arrays, are both possible to provide good performance. This implies that, given a required outage probability, there is an optimal number of RAUs and number of antennas per RAU to minimize the total required number of antennas.

131 116 Ray-tracing Channel Simulation The optimal parameters depend on the population density, the antenna array size at the STAs, and the room dimensions. We have presented the method to determine these parameters. For massive MIMO in the 2.4 GHz band, we have addressed the following questions. - First, we have investigated the performance of massive MIMO in an indoor environment based on channel measurements. Theoretically, using larger antenna arrays can potentially de-correlate the MIMO channels between the users, thus achieving closeto-optimal capacity. However, our results showed that realistic channels offer a limited number of spatial degrees of freedom and thus provide lower capacity than the ideal case. Getting a high capacity gain requires the antenna array to have a large aperture size. Specifically, to achieve 80% of the optimal capacity for realistic channels, the antenna array size needs to be twice the size of an array for an ideal channel. Therefore, the physical size of a centralized massive MIMO antenna array may be too large. - Second, we have analyzed how much improvement over traditional MIMO can be obtained considering both centralized antenna system (CAS) and distributed antenna system (DAS) architectures. We have answered this question from two perspectives. First, we have analyzed the capacity gain with large antenna arrays for a constant number of users. In comparison with traditional MIMO systems, CAS improves the capacity more than 4 times as a result of improved multiplexing capability. Considering the same number of antennas in CAS, DAS can offer twice more capacity by improving the received signal strength, a benefit of placing the antennas closer to the STAs. Second, the maximum average sum-rate with a given total number of antennas, has also been analyzed for CAS and DAS. The maximum average sum-rate has been analyzed by optimizing the number of STAs simultaneously supported. With such optimization, massive MIMO offers even a higher sum-rate over traditional MIMO. In addition, a DAS is much more advantageous than CAS for realistic channels due to the compound effect of lower path loss and independent small scale fading. We have also found that the optimal number of users depends on the precoding technique and the antenna system configuration, and is generally in the same order of the number of antennas belonging to the APs. This is different from the classic view of massive MIMO in which case the number of antennas belonging to the APs should be about ten times larger than the number of simultaneously served users. The classic view is consistent with the case in which CAS is used and the channels have a high correlation. In summary, from both perspectives, DAS is more advantageous than CAS due to the much higher sum-rate it offers. In addition, DAS divides the antennas into small arrays, which solves the problem that very large antenna arrays in CAS indoor systems may have impractical physical size. - Third, we have proposed two CSI feedback bit allocation schemes for distributed massive MIMO to deal with the large CSI feedback overhead problem. We have argued that the feedback reduction techniques based on spatial correlation do not work

132 8.2 Future Work 117 well for indoor networks in the microwave frequency bands. The reason is, in an indoor environment, the rich scattering causes antennas that are separated by a small distance (in the order of a few wavelengths) to have independent channels. We have proposed to reduce the CSI feedback overhead by exploiting the strong differences in large scale fading between the RAUs to a given STA: a small number of feedback bits, or even no feedback are required for channels that exhibit a high large scale fading. We have shown that the overhead could be reduced by more than 20% in comparison with the IEEE scheme, for achieving 90% of the average sumrate with perfect CSI. In the final part of the thesis, we have compared the performance of SCN and distributed massive MIMO in the 2.4 GHz and 60 GHz bands. For 2.4 GHz, distributed massive MIMO offers much higher sum-rate than SCN which suffers from inter-cell interference. Therefore, from a performance point of view, distributed massive MIMO is the better choice for the 2.4 GHz band. In contrast, 60 GHz SCN is mainly constrained by the received signal power, and the signal is easily blocked by objects like walls. As a result, distributed massive MIMO and SCN perform similarly. Since the implementation of SCN is less complex than distributed massive MIMO, we conclude that SCN is the better choice for the 60 GHz band. 8.2 Future Work Based on the insights we have gained in this thesis on RoF-based SCN and massive MIMO for indoor communications, we identified a number of issues that need to be addressed in future research. Standardization of DAS: We have seen the benefits of DAS in indoor communications. In order for the beamforming mechanisms and CSI compression techniques to work in practice, the RAUs need to be identifiable for the AP and the STAs. This means that the format of the information of the DASs, e.g., the address of the RAUs, the number of antennas of the RAUs, should be standardized. This is not yet done in the IEEE standards. Spatial multiplexing in 60 GHz DAS: DAS can be used to improve the 60 GHz link connectivity by coordinating multiple RAUs. We have argued to use a single RAU selectively, which implies that there are always a number of RAUs not being used. In fact, these RAUs can be used to transmit to or receive from other STAs simultaneously, which offers a spatial multiplexing gain. More generally, 60 GHz DAS can be used for both spatial multiplexing and connectivity improvement. However, spatial multiplexing is not yet supported in the IEEE ad standard. But, the on-going IEEE ay standard will include spatial multiplexing. In the literature, we can find some reports about MIMO spatial multiplexing with 60 GHz CAS, e.g. [160] [161], but there is limited work for DAS. DAS spatial multiplexing uses multiple antenna arrays for beamforming to multiple

133 118 Ray-tracing Channel Simulation STAs, thus how to efficiently do the beamforming, i.e., with low latency and low computation, needs to be investigated. Massive MIMO with TDD based channel training in IEEE : The feedback scheme we considered in Chapter 6 is part of a closed-loop process. The limitation of this scheme is that, the CSI feedback overhead will increase linearly with both the AP antenna array size and the number of co-scheduled STAs. Therefore, for very large antenna arrays and very large number of STAs, the overhead may prohibit this scheme. This problem is similar to the FDD scheme in cellular networks [83]. On the contrary, the TDD scheme [89], i.e., uplink channel training utilizing the channel reciprocity feature, only increases with the number of co-scheduled STAs [83]. IEEE is a TDD system. However, uplink channel training is part of the IEEE n standard but was later removed in the IEEE ac version. If large antenna arrays, e.g., with hundreds of antennas, are used in IEEE , uplink channel training is necessary. The fundamental problem of uplink channel training is that the RF chains at the transmitter and the receiver for the uplink and downlink are not identical. Therefore, to be able to rely on channel reciprocity, the RF chains have to be calibrated. There has been much progress for calibration in MIMO systems recently, e.g., see [162] and [163]. Therefore, it is expected that uplink channel training will be feasible in the near future. The TDD training scheme in IEEE n can be used for massive MIMO. Since IEEE n only supports SU-MIMO, the TDD training scheme needs to be adapted to support MU-MIMO. Furthermore, since IEEE does not have an uplink multi-user transmission mechanism yet, the channel training frames have to be sent sequentially from the STAs, which is not efficient. This problem has to be solved as well. Energy efficiency in RoF-based massive MIMO: Allowing transportation of radio signal through optical fibers requires additional components that lead to high power consumption of the system [164]. Specifically, the lasers, photo-diodes, power amplifiers, and DAC/ADC components, increase the energy consumption in comparison with non- RoF systems. This implies that the average energy consumption per antenna is much higher in RoF, thus we either need to use the antennas more efficiently, or reduce the energy consumption of the RoF links. We have seen from the analysis in Chapter 7 that the necessary power for radio transmission is actually low due to the dense deployment of the RAUs. In [165], we can also see that the components in the RoF system account for the majority of the power consumption of the network. Therefore, ways to improve the energy efficiency of RoF links should be investigated in order to make RoF based massive MIMO feasible. Distributed beamforming in distributed massive MIMO: We have used centralized beamforming algorithms in our work, i.e., all the beamforming coefficients of the antennas are calculated centrally at the AP. The problem of centralized beamforming is that the complexity becomes prohibitive for very large numbers of AP antennas [166]. One possible solution is to use distributed beamforming algorithms [167], e.g., the beamforming

134 8.2 Future Work 119 coefficients could be calculated locally for each RAU. This type of beamforming algorithm is more widely studied in sensor networks where the sensors cooperate to transmit data [168]. We recommend to investigate distributed beamforming algorithms, to reduce the computational load of the APs. Coexistence of massive MIMO and other networks: The precoding techniques in massive MIMO deal primarily with interference between the spatial streams. The external sources of interference are usually ignored. For example, there may be traditional WiFi networks or peer-to-peer WiFi networks (or WiFi Direct [169] ) that co-exist with the massive MIMO network, but operate independently. The massive MIMO system has to share the spectrum and the space with these co-existing networks. These networks may induce interference that degrades the performance of massive MIMO system. In particular, massive MIMO systems use a low transmission power to attain higher energy efficiency. This means that the received signal power will be low, making the massive MIMO receivers more vulnerable to interference from other networks. The co-existence with other networks is therefore critical and needs to be studied.

135 120 Ray-tracing Channel Simulation

136 121 Appendix A. Ray-tracing Channel Simulation This appendix introduces procedures of using RPS to simulate indoor wireless channels. It mainly includes two steps. First, construct indoor scenarios in RPS and configure the parameters according to the need of the evaluations of wireless systems. Second, extract the channel data from RPS, and extend the channel response to antenna arrays of arbitrary size and configuration. We first present the procedures of ray-tracing simulations in Section A.1. In Section A.2, we introduce the basic ray-tracing channel model, and the algorithm for constructing the channel responses from the ray-tracing channels. We show some examples of the raytracing channels and the constructed channels in Section A.3. A.1 The Procedures of Ray-tracing Channel Simulation The flow chart of ray-tracing simulation and analysis are given in Fig. A. 1, which is selfexplanatory.

137 122 Bibliography Wireless system parameters Transmitter and receiver antenna positions Carrier frequency and bandwidth Transmission power Noise power RPS Scenario construction Scenario: walls, ceiling, floor, windows, humans Set dielectric properties to the materials Randomize mobile objects like humans RPS Simulation Configure ray-tracing parameters: diffraction, reflection, scattering, ray angular resolution, etc. Simulate and output raw data to txt files MATLAB Raw data processing Read raw data from RPS output files Convert to MATLAB readable data files Other processing Extend to arbitrary antenna array structure and size Transform to band-limited channels Numerical analysis Fig. A. 1 Flaw chart of ray-tracing channel simulation and post-processing. The electro-magnetic (EM) properties of the different materials used in RPS simulations are listed in Table A. 1 and Table A. 2 for 2.4 GHz and 60 GHz, respectively [54] [170]. Table A. 1 Relative permittivity of various materials at 2.4 GHz. Material Thickness(m) Real Imaginary Inner wall Plasterboard Ceiling Concrete Floor Concrete

138 A.2 Extension from SISO Ray-tracing Channel to Arbitrary Antenna Array 123 Table A. 2 Relative permittivity of various materials at 60 GHz. Material Thickness(m) Real Imaginary Exterior wall Concrete Interior wall Plasterboard Ceiling Concrete Ground Wood Window Soda-Borosilicate glass Door Wood Human body A.2 Extension from SISO Ray-tracing Channel to Arbitrary Antenna Array In order to simply the simulation and make the ray-tracing results reusable, we usually use a single antenna at the transmitter and receiver in the RPS simulations, and then extend the simulated channel to that of arbitrary antenna arrays [112]. We introduce the algorithm in the following. A.2.1 Ray-tracing Channel Model The 3D ray-tracing model for time-invariant channel is given by L1 al l t t, l r r, l l0 h() ( )( Ω Ω )( Ω Ω ) (A.1) where a l is the complex gain, l is the delay, Ω, [,,, ] T tl tl tl and Ω, [,,, ] T rl rl rl are the departure and arrival angles (elevation and azimuth) vectors of the l -th path; L is the number of paths. For a simple three-ray case, the ray-tracing model is illustrated in Fig. A. 2. The 0-th path is the LOS path or direct path, and the other paths are NLOS paths or reflective paths.

139 124 Bibliography Reflector 0 1 Transmitter 2 Receiver Fig. A. 2 Illustration of ray-tracing channel model. A.2.2 Construct Channel for Arbitrary Antenna Array The RPS ray-tracing can output all information of the paths in the exact form of Eq. (A.1). Considering that the scatters or reflectors are in the far-field of the transmitter and receivers, the elements in a co-located antenna array only have a phase difference for each ray [112]. Therefore, for arbitrary antenna arrays, the CIRs of all the antenna elements can be constructed by adding the array response to the channel of an SISO system, e.g., the channel between the first element of the transmitter and receiver antenna array. This method is only valid for co-located antenna arrays, so it has to be repeated for each RAU for DASs. The mathematical description is given below. Fig. A. 3 shows the positions of the antenna arrays at the RAU and the STA. Ceiling plane x 1 x 2 x 4 x 3 RAU antenna array z x y Half a wavelength STA antenna array STA plane Fig. A. 3 Illustration of the antenna arrays at the RAU and the STA. The antenna array channel impulse response (CIR) is given by

140 A.3 Examples 125 H() N N ae ( Ω ) e ( Ω ) ( ) (A.2) T l RAU STA l STA ST A, l RAU RA U, l l where N RAU and N STA are the number of antennas at the transmitter and receiver antenna array, and e RAU ( ΩRAU, l ) and e STA ( ΩSTA, l ) are the RAU and STA antenna array response vectors in the direction of the -th path. The antenna array response vectors are specifically given as l 1 2 e j (A.3) RAU ( ΩRAU ) exp ( ( RAU ) n ) RAU N k Ω x RAU NRAU e j (A.4) STA ( ΩSTA ) exp ( ( STA ) n ) STA N k Ω x STA N STA 1 where n is the index of the antenna elements of the antenna arrays at the STAs and the RAUs, k( Ω ) is the unit direction vector given by k [sin cos,sinsin, cos ] T, x is the coordinates of the antenna elements in their local coordinate system, and is the inner product. For band-limited channel, the obtained CIRs are first convolved with a pulse shaping filer ct which is typically a raised cosine filter, or a rectangular window of the symbol width. The output is then sampled at the symbol rate to get discrete channels. After that, the discrete band-limited CIRs are converted to frequency domain response through DFT. Finally, the obtained channel is used for the numerical analysis. A.3 Examples Now we give some examples of the simulation scenario and ray-tracing results. We assume the carrier frequency is 60 GHz and the signal bandwidth is 0.5 GHz. The transmit antenna array is a 2 by 2 planar array with half-wavelength spacing and the receiver has one isotropic antenna. The living room scenario used in Chapter 3 is given in Fig. A. 4. The ray-tracing result for an example STA position is given in Fig. A. 5, where the transmitter is equipped with an isotropic antenna. Fig. A. 6 shows the constructed channel response from the ray-tracing channel from RPS.

141 126 Bibliography Transmit antenna Human body Fig. A. 4 Simulation scenario in RPS. Transmitter Signal paths Receiver Fig. A. 5 Signal paths obtained from 3D ray-tracing.