Coordinated Transmission for Wireless Interference Networks HAMED FARHADI Doctoral Thesis in Telecommunications Stockholm, Sweden 2014 TRITA-EE 2014:064 ISSN 1653-5146 ISBN 978-91-7595-391-5 KTH, School of Electrical Engineering Department of Communication Theory SE-100 44 Stockholm SWEDEN Akademisk avhandling som med tillst˚ and av Kungl Tekniska h¨ ogskolan framl¨ agges till offentlig granskning f¨ or avl¨ aggande av teknologie doktorsexamen i telekommunikation fredagen den 19 December 2014 klockan 13.15 i h¨ orsal F3, Lindstedtsv¨agen 26, Stockholm, Sweden. © 2014 Hamed Farhadi, unless otherwise noted. Tryck: Universitetsservice US AB Abstract Wireless interference networks refer to communication systems in which multiple source–destination pairs share the same transmission medium, and each source’s transmission interferes with the reception at non-intended destinations. Optimizing the transmission of each source–destination pair is interrelated with that of the other pairs, and characterizing the performance limits of these networks is a challenging task. Solving the problem of managing the interference and data communications for these networks would potentially make it possible to apply solutions to several existing and emerging communication systems. Wireless devices can carefully coordinate the use of scarce radio resources in order to deal effectively with interference and establish successful communications. In order to enable coordinated transmission, terminals must usually have a certain level of knowledge about the propagation environment; that is, channel state information (CSI). In practice, however, no CSI is a priori available at terminals (transmitters and receivers), and proper channel training mechanisms (such as pilotbased channel training and channel state feedback) should be employed to acquire CSI. This requires each terminal to share available radio resources between channel training and data transmissions. Allocating more resources for channel training leads to an accurate CSI estimation, and consequently, a precise coordination. However, it leaves fewer resources for data transmissions. This creates the need to investigate optimum resource allocation. This thesis investigates an informationtheoretic approach towards the performance analysis of interference networks, and employs signal processing techniques to design transmission schemes for achieving these limits in the following scenarios. First, the smallest interference network with two single-input single-output (SISO) source–destination pairs is considered. A fixed-rate transmission is desired between each source–destination pair. Transmission schemes based on point-topoint codes are developed. The transmissions may not always attain successful communication, which means that outage events may be declared. The outage probability is quantified and the ǫ-outage achievable rate region is characterized. Next, a multi-user SISO interference network is studied. A pilot-assisted ergodic interference alignment (PAEIA) scheme is proposed to conduct channel training, channel state feedback, and data communications. The performance limits are evaluated, and optimum radio resource allocation problems are investigated. The analysis is extended to multi-cell wireless interference networks. A low-complexity pilot-assisted opportunistic user scheduling (PAOUS) scheme is proposed. The proposed scheme includes channel training, one-bit feedback transmission, user scheduling and data transmissions. The achievable rate region is computed, and the optimum number of cells that should be active simultaneously is determined. A multi-user MIMO interference network is also studied. Here, each source sends multiple data streams; specifically, the same number as the degrees of freedom of the network. Distributed transceiver design and power control algorithms are proposed that only require local CSI at terminals. Sammanfattning Tradl¨ ˙ osa interferensn¨atverk ¨ ar kommunikationssystem d¨ ar flera par av k¨allor och destinationer delar pa˙ samma ¨ overf¨ oringsmedium och varje k¨allas s¨andning st¨or mottagningen vid icke-avsedda destinationer. Optimering av data¨ overf¨oringen f¨or varje par av k¨allor och destinationer samverkar med optimeringen f¨or de andra paren, och att karakterisera prestandagr¨ anserna f¨or dessa n¨ atverk ¨ar d¨ arf¨or en utmanande uppgift. Om problemet med hantering av st¨orningar och datakommunikation f¨or dessa n¨ atverk l¨oses, kan dessa l¨osningar potentiellt till¨ ampas pa˙ flera befintliga och kommande kommunikationssystem. Tradl¨ ˙ osa apparater kan noggrant samordna anv¨andningen av begr¨ ansade radioresurser f¨or att effektivt hantera st¨ orningar och skapa tillf¨ orlitlig kommunikation. F¨or att m¨ojligg¨ ora samordnad ¨ overf¨ oring maste ˙ terminalerna oftast ha en viss niva˙ av kunskap om utbredningsmilj¨on, dvs. kanaltillstandsinformation ˙ (CSI). I praktiken finns dock a priori ingen CSI pa˙ terminalerna och l¨ampliga kanaltr¨aningsmekanismer (t.ex. pilotbaserad kanaltr¨aning och kanaltillst˙andsaterkoppling) ˙ b¨ or anv¨andas f¨or att f¨orv¨arva CSI. Detta inneb¨ar att varje terminal maste ˙ dela tillg¨angliga radioresurser mellan kanaltr¨aning och data¨ overf¨oring. Ju fler resurser som tilldelats f¨or kanaltr¨aningen, desto noggrannare blir CSI-uppskattningen och d¨ armed blir samordningen b¨ attre. Dock ¨ ar f¨ arre resurser kvar f¨ or data¨ overf¨oring. Effektiva metoder f¨or att genomf¨ ora resursf¨ ordelningen maste ˙ d¨ arf¨or unders¨ okas. En informationsteoretisk strategi f¨ or prestandaanalys av interferenssn¨ atverk utreds och signalbehandlingstekniker anv¨ands f¨ or att utforma ¨overf¨oringssystem som uppnar ˙ dessa gr¨anser. Denna avhandling behandlar samordnad ¨overf¨oring i f¨oljande scenarier. F¨orst unders¨ oks det minsta interferensn¨atverket, bestaende ˙ av tva˙ single-input single-output-par (SISO-par) med k¨allor och destinationer. En ¨overf¨oringsmetod med fast datatakt ¨ onskas mellan varje par av k¨allor och destinationer, givet ett bivillkor pa˙ anv¨and effekt. ¨ overf¨ oringssystem baserade pa˙ punkt-till-punkt Gaussiska koder utvecklas d¨ arf¨ or. ¨ overf¨ oringarna uppnar ˙ inte alltid tillf¨ orlitlig kommunikation, och avbrott kan d¨ arf¨ or f¨ orekomma. Avbrottssannolikheten kvantifieras och ǫavbrottsdatataktsregionen karakt¨ariseras. D¨arefter studeras ett fleranv¨andarinterferensn¨atverk med SISO-noder. Ett pilotassisterat system med ergodisk interferensuppr¨atning (PAEIA) f¨ oreslas ˙ f¨ or att genomf¨ ora kanaltr¨aning, kanaltillstands ˙ aterko˙ ppling och datakommunikation. Prestandagr¨anser utv¨ arderas, och problemet med optimal radioresursallokering studeras. Ett fleranv¨andarinterferensn¨atverk med multiple-input multiple-output-noder (MIMO-noder) studeras ocksa. ˙ H¨ ar s¨ander varje k¨alla lika manga ˙ datastr¨ ommar som tillats ˙ av frihetsgraderna av n¨ atverket. Distribuerad design av mottagare och s¨andare samt effektstyrningsalgoritmer som enbart kr¨ aver lokal CSI vid terminalerna f¨ oreslas. ˙ Analysen utvidgas ocksa˙ till ett fleranv¨andarinterferensn¨atverk med flera celler. Ett pilotassisterat lagkomplexitetssyst˙ em f¨or opportunistisk anv¨andarschemal¨ aggning f¨oreslas. ˙ Det f¨oreslagna systemet omfattar kanaltr¨aning, en-bitsaterkoppling, ˙ anv¨andarschemal¨ aggning samt data¨ overf¨oring. Den uppnabara ˙ datataktsregionen ber¨aknas och det optimala antalet celler som skall vara aktiva samtidigt best¨ ams. Acknowledgments It is a pleasant task to express my thanks to all those who contributed in many ways to the success of my PhD study and made it an unforgettable experience for me. Foremost, I owe my deepest gratitude to my advisor Prof. Mikael Skoglund. I am grateful to Mikael for welcoming me to the Department of Communication Theory, and for giving me the freedom to pursue my research interests. His openness to ideas, insightful suggestions, and supports enriched my PhD research. I also wish to thank my co-advisor Prof. Chao Wang for his valuable helps and supports through different stages of my research and for being a good friend of me all these years. It was a great pleasure for me to work with Mikael and Chao. I would like to thank Prof. Vahid Tarokh at Harvard University for giving me the opportunity to visit his research group. It has been an honor to work with him. I really appreciate his kind hospitality during my stay at Cambridge. This research visit has broadened my view on research, tremendously. I wish to thank Dr. Jinfeng Du at MIT for all his valuable helps and cares during the period of my stay at Cambridge. I am also thankful to Dr. Mohsen Farmahini and Alireza Mehrtash for all their kind helps during this visit. I am also grateful to all friends at SEAS who made my time enjoyable and rewarding. Karin Demin and Kathleen Masse helped me a lot in the administrative issues of this visit. I wish to acknowledge the John and Karin Engblom foundation, Knut and Alice Wallenbergs foundation, Ericsson Research foundation, Qualcomm, and the European School of Antennas for the financial support of the trips for the research visit and the conferences that I participated. Swedish Foundation for Strategic Research is acknowledged for the financial support of my PhD study. I would like to take the opportunity to thank Prof. David Gesbert for acting as the opponent for this thesis. I also thank Prof. Erik Larsson, Prof. Tommy Svensson, and Prof. Jeong Woo Cho for acting on the grading committee, and Prof. Joakim Jaldén for doing the quality review of the thesis. My sincere thanks also goes to my colleagues whom I had valuable discussions and collaborations. In particular, I am grateful to Prof. Lars Rasmussen for sharing his valuable international academic career experiences and being a source of inspiration for good academic practice. I also wish to thank Prof. Carlo Fischione for collaborations on the problems of mutual interest. I would like to thank Dr. Per Zetterberg for providing a fantastic wireless test-bed network for the real-time evaluation of algorithms proposed in this thesis and for his valuable insights on the practical considerations of wireless system design. Also, I am thankful to Nima Najari Moghadam for test-bed implementation and real-time measurements of some algorithms proposed in this thesis. It was a pleasure to work with Nima and Per. I gratefully acknowledge the discussions with my colleagues from KTH, Link¨oping University, and Ericsson Research within the RAMCOORAN project cooperation. Prof. Michail Matthaiou and Prof. Mats Bengtsson are acknowledged for valuable comments on my research. I wish to thank Dr. Majid Nasiri Khormuji for being al- viii ways available for fruitful discussions and research collaborations; Peter Larsson for sharing his valuable industrial research experiences; Dr. Mohammadreza Gholami for sharing his experiences; and Dr. Themistoklis Charalambous for discussions on the problems of common interest. Dr. Nafiseh Shariati is acknowledged for being always available for great conversations on various topics. I acknowledge the great discussions with Prof. Tobias Oechtering, Frédéric Gilbert Gabry, and Efthymios Stathakis regards our collaborative efforts on the implementation of interactive teaching. I am also very thankful to Dr. Ali Zaidi, Hadi Ghauch, and Dr. Nicolas Schrammar for our scientific collaborations. I wish to thank Nan Li, Hadi Ghauch, Farshad Naghibi, and especially Chao and Mikael for helping me proofread the thesis and for their valuable suggestions and feedbacks. I am thankful to Rasmus Brandt for his kind helps in editing the Swedish parts of this thesis. I also thank the computer support group for providing reliable resources, and Raine Tiivel and Dora S¨ oderberg for all their kind helps in administrative issues. I wish to thank all my past and present colleagues at the Department of Communication Theory and the Department of Signal Processing for making a friendly working environment. Especial thanks go to my officemate Iqbal Hussain. It was so nice to share the office with Iqbal all these years. I enjoyed after work Thai dinner events with Kittipong and Hieu! I would like to take this opportunity to acknowledge all individuals who have inspired or have encouraged me to do research. In particular, I am thankful to my teachers at KTH, University of Tehran, and Iran University of Science and Technology from whom I’ve learned a lot during years of study. My friends Maksym Girnyk, Karina, Farshad Naghibi, Serveh Shalmashi, Euhana Ghadimi, Somayeh Salimi, Elaheh Jafari, Nafiseh Shariati, Amirpasha Shi¨ razinia, Alla, Arash Owrang, Ehsan Olfat, Majid Gerami, and Ghazaleh Panahandeh in various ways made living in the beautiful city of Stockholm an enjoyable and memorable experience for me. I offer my warmest thanks to my parents for all their endless love and supports. My father was my most inspiring teacher and his caring relationships with students deeply influenced my view on teaching and education. I wish to thank my brothers for all their encouragements. Last but not least, I would like to thank my beloved wife Maryam for all the happiness she brought to my life and all her patience during my work on this thesis. If it was not her patience, I wouldn’t be able to work on this thesis. Hamed Farhadi Stockholm, November 2014 Contents Contents ix List of Figures xiii List of Notations xv List of Acronyms xvii 1 Introduction 1.1 Wireless Network Operations Overview . . . . . . . . . . . . . . 1.2 Thesis Scope and Contributions . . . . . . . . . . . . . . . . . . . 1 5 6 2 Background 2.1 Wireless Interference Networks . . . . . . . . . . . . . . . . . . 2.2 Two-user Interference Networks . . . . . . . . . . . . . . . . . . 2.3 K-user (K > 2) Interference Networks . . . . . . . . . . . . . . 2.3.1 Achievable Degrees of Freedom Region . . . . . . . . . . 2.3.2 Interference Alignment for MIMO Interference Networks 2.3.3 Interference Alignment for SISO Interference Networks . 2.3.4 Ergodic Interference Alignment . . . . . . . . . . . . . . 2.4 Wireless Interference Networks with Imperfect CSI . . . . . . . . . . . . . . . 21 21 22 23 25 26 28 31 32 3 Two-user Interference Networks: Point-to-Point Codes 3.1 Two-user SISO Interference Network . . . . . . . . . . . . . . 3.2 Orthogonal Transmission Scheme . . . . . . . . . . . . . . . . 3.3 Non-Orthogonal Transmission Schemes . . . . . . . . . . . . . 3.3.1 Direct Decoding at Both Receivers . . . . . . . . . . . 3.3.2 Successive Interference Cancellation at Both Receivers 3.3.3 Successive Interference Cancellation at One Receiver . 3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.A The Proof of Proposition 3.2.1 . . . . . . . . . . . . . . . . 3.B The proof of Proposition 3.3.1 . . . . . . . . . . . . . . . . 3.C The proof of Corollary 3.3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 36 38 40 40 45 51 56 57 58 60 ix . . . . . . . . . . x Contents 3.D The Proof of Proposition 3.3.2 . . . . . . . . . . . . . . . . . . 3.E The Proof of Proposition 3.3.3 . . . . . . . . . . . . . . . . . . 61 62 4 K-user SISO Interference Networks: Pilot-assisted Interference Alignment 4.1 Multi-user SISO Interference Network . . . . . . . . . . . . . . . 4.2 Pilot-assisted Ergodic Interference Alignment . . . . . . . . . . . 4.2.1 Pilot Transmission Phase . . . . . . . . . . . . . . . . . . 4.2.2 Feedback Transmission Phase . . . . . . . . . . . . . . . . 4.2.3 Data Transmission Phase . . . . . . . . . . . . . . . . . . 4.3 Analog Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Achievable Rate Region . . . . . . . . . . . . . . . . . . . 4.3.2 The Optimum Power Allocation . . . . . . . . . . . . . . 4.3.3 Achievable Degrees of Freedom Region . . . . . . . . . . . 4.3.4 Numerical Evaluation . . . . . . . . . . . . . . . . . . . . 4.4 Digital Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 Power Control Problem . . . . . . . . . . . . . . . . . . . 4.4.2 Throughput Maximization Problem . . . . . . . . . . . . 4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.A The Proof of Proposition 4.3.2 . . . . . . . . . . . . . . . . . . 4.B The Proof of Theorem 4.4.3 . . . . . . . . . . . . . . . . . . . 4.C The Proof of Theorem 4.4.4 . . . . . . . . . . . . . . . . . . . 65 66 67 67 67 69 70 70 72 73 74 76 78 86 96 98 98 100 5 K-user MIMO Interference Networks: Transceiver Design and Power Control 5.1 Multi-user MIMO Interference Network . . . . . . . . . . . . . . 5.1.1 Transmitter Structure . . . . . . . . . . . . . . . . . . . . 5.1.2 Receiver Structure . . . . . . . . . . . . . . . . . . . . . . 5.2 Transceiver Design and Power Control . . . . . . . . . . . . . . . 5.2.1 CSI Acquisition, Transceiver Design, and Power Control . 5.2.2 Distributed Power Control . . . . . . . . . . . . . . . . . . 5.3 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . 5.4 Test-bed Implementation . . . . . . . . . . . . . . . . . . . . . . 5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.A The Proof of Theorem 5.2.1 . . . . . . . . . . . . . . . . . . . . 5.B The Proof of Theorem 5.2.2 . . . . . . . . . . . . . . . . . . . . 101 102 102 103 103 104 107 118 123 124 126 126 6 Multi-cell Interference Networks: Pilot-assisted Opportunistic User Scheduling 6.1 Multi-cell Interference Network . . . . . . . . . . . . . . . . . . . 6.2 Pilot-assisted Opportunistic User Scheduling Scheme . . . . . . . 6.2.1 Pilot Transmission Phase . . . . . . . . . . . . . . . . . . 6.2.2 Feedback Transmission and User Selection Phase . . . . . 6.2.3 Data Transmission Phase . . . . . . . . . . . . . . . . . . 129 131 131 131 132 133 Contents 6.3 Achievable Rate Region . . . . . . . . . . . 6.3.1 Achievable Total Degrees of Freedom 6.3.2 Numerical Evaluation . . . . . . . . 6.4 Summary . . . . . . . . . . . . . . . . . . . 6.A The Proof of Theorem 6.3.1 . . . . . . . . xi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 135 136 137 139 7 Conclusion 7.1 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 141 143 Bibliography 147 List of Figures 1.1 Operations in a wireless communication network. . . . . . . . . . . . 5 2.1 2.2 2.3 K-user wireless interference network . . . . . . . . . . . . . . . . . . Interference management schemes . . . . . . . . . . . . . . . . . . . Interference alignment in a three-user MIMO interference network . 22 24 28 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 Two-user interference network. . . . . . . . . . . . . . . . . . . . . . Solution of the power control problem for the NOT1 scheme. . . . . The outage probability of the NOT1 scheme . . . . . . . . . . . . . . The ǫ-outage achievable rate region of the NOT1 scheme . . . . . . . Solution of the power control problem for the NOT2 scheme . . . . . The outage probability of the NOT2 scheme . . . . . . . . . . . . . . The ǫ-outage achievable rate region of the NOT2 scheme . . . . . . . Solution of the power control problem for the NOT3 scheme. . . . . The outage probability of the NOT3 scheme in a symmetric network The outage probability of the NOT3 scheme in a asymmetric network The ǫ-outage achievable rate region of the NOT3 scheme . . . . . . . 36 42 43 45 48 50 52 54 55 56 57 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10 4.11 4.12 Transmitted symbols’ format in a K-user interference network. . The optimum power allocation factor. . . . . . . . . . . . . . . . The achievable rate of the PAEIA and TDMA schemes. . . . . . The achievable sum-rate versus the number of users. . . . . . . . Feasibility probability versus transmission rate of each user . . . Average transmission power versus transmission rate of each user Feedback bits trade-off in a three-user interference network . . . Throughput versus power . . . . . . . . . . . . . . . . . . . . . . Feedback bits allocation trade-off . . . . . . . . . . . . . . . . . . Throughput of a K-user interference network. . . . . . . . . . . . Delay-limited throughput of a three-user network . . . . . . . . . Delay-limited throughput of a K-user interference network . . . . . . . . . . . . . . . 66 75 76 77 82 85 86 92 93 94 95 97 5.1 The structure of a transmitter and receiver pair . . . . . . . . . . . . 102 xiii . . . . . . . . . . . . xiv List of Figures 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.10 5.11 CSI acquisition, transceiver design, and power control procedure . . Feasibility probability versus transmission rate. . . . . . . . . . . . . Feasibility probability in a network with maximum power constraint. The mutual information of the source-destination pairs versus iterations Transmission power of a source versus number of iterations . . . . . The mutual information of the source-destination pairs versus iterations The computed transmission power versus the number of iterations . The mutual information of the different streams versus iterations . . Transmission powers versus the number of iterations . . . . . . . . . Computed power versus the number of iterations . . . . . . . . . . . 104 110 112 119 120 121 122 123 124 125 6.1 6.2 6.3 6.4 6.5 Schematic representation of different phases of the PAOUS scheme Transmitted symbols within one fading block. . . . . . . . . . . . . The achievable sum-rate versus power. . . . . . . . . . . . . . . . . The achievable sum-rate versus the decision threshold. . . . . . . . The achievable sum-rate versus the power allocation factor. . . . . 130 133 136 137 138 . . . . . List of Notations |x| CN (m, σ 2 ) f ′ (x) ∅ ν i [X] ≻ E[X] λmax (X) N (m, σ 2 ) X−1 ℑ[x] span(X) Nf fX Q(x) ∆ rank(X) ℜ[x] ρ(X) Tr[X] XT X∗ Absolute value of x Complex Gaussian distribution with mean m and variance σ 2 Derivative of function f (x) Empty set Eigenvector corresponding to the ith lowest eigenvalue of matrix X Element-wise vector inequality Element-wise strict vector inequality Expectation of random variable X The largest absolute eigenvalue of matrix X. Gaussian distribution with mean m and variance σ 2 Inverse of matrix X Imaginary part of x Linear span of columns of matrix X Number of feedback bits Probability density function of random variable X Q-function Quantization step-size Rank of matrix X Real part of x Spectral radius of matrix X Trace of matrix X Transpose of matrix X Transpose conjugate of matrix X xv List of Acronyms AWGN CoMP CSI CSIR CSIT DoF DSM EIA EIA-PC EIA-RA FDD FDMA KKT LICQ M2M MMSE MIMO OT PAEIA PAEIA-US PAOUS pdf QoS RHS SNR SIC SINR SISO TDD TDMA USRP WBAN Additive white Gaussian noise Coordinated multi point Channel state information Channel state information at receiver Channel state information at transmitter Degrees of freedom Discrete superposition model Ergodic interference alignment Ergodic interference alignment with power control Ergodic interference alignment with rate adaptation Frequency-division duplex Frequency-division-multiple-access Karush-Kuhn-Tucker Linear independent constraint qualification Machine-to-machine communications Minimum mean square error Multiple-input multiple-output Orthogonal transmission Pilot-assisted ergodic interference alignment Pilot-assisted ergodic interference alignment with user selection Pilot-assisted opportunistic user scheduling Probability density function Quality of service Right hand side Signal-to-noise ratio Successive interference cancelation Signal-to-interference-plus-noise ratio Single-input single-output Time-division duplex Time-division-multiple-access Universal software radio peripheral Wireless body area networks xvii Chapter 1 Introduction A NETWORKED SOCIETY in which everyone and everything, everywhere, have the potential to benefit from connectivity, will shape the future of human life [Sac03]. Wireless communication systems are expected to play an important role in the development of the networked society. Fore example, wireless networks will provide ubiquitous access to information, while wireless health monitoring systems will decentralize medical treatments. Vehicular communication networks will help make future transportation systems more efficient and safe, and networked control systems will take advantage of global information fusion in order to make intelligent decisions. The trend towards massive usage of wireless technology has led to an explosion in the number of connected devices, and tremendous growth in wireless traffic volume [Eri12]. However, the scarcity of radio resources and the inherent characteristics of the wireless transmission medium make handling communication in these conditions a formidable task. The radio spectrum is scarce and is considered to be among the most expensive natural resources [Her85, ZKAQ01]. In addition, the energy budget of mobile terminals is restricted and there are serious concerns regarding the vast energy consumption of wireless communication systems [CHA+ 11, DCG+ 13]. Consequently, spectral and energy efficient design is essential for emerging wireless technologies. Wireless transmission is generally subject to two phenomena: fading and interference [PL]. The former is a consequence of reflectors scattered in the environment surrounding a transmitter and a receiver, such that the receiver observes a superposition of multiple copies of the transmitted signal. The superposition of the signals can be either constructive or destructive depending on the phase shift and the attenuation of received signals from different paths. The randomness of fading may degrade communication quality. Several effective techniques have been devised in recent decades to overcome the adverse effects of random fading. For instance, multiple-antenna transmission techniques have been proposed to realize spatial diversity and to improve the performance of wireless systems [TJC99, JSO02, GSDs+ 03, LS03]. Another, even more challenging, obstacle in the operation of wireless networks 1 2 Introduction is interference. Because of the broadcast nature of wireless transmission medium, a user can interfere with the communication of any other user. This occurs in some existing and emerging wireless communication scenarios, including inter-cell interference in cellular networks; ad-hoc networks; interference between neighboring access points in wireless local area networks; inter-user interference between neighboring devices in machine-to-machine (M2M) communication scenarios; interference between implanted medical sensors in wireless body area networks (WBAN); and interference between licensed and unlicensed users in cognitive radio networks. These communication systems, usually composed of multiple information sources and destinations, can be modelled as an important class of wireless networks: interference networks. In an interference network, each source intends to communicate with its dedicated destination and all sources share the same transmission medium. Inter-user interference makes communication of different users interrelated, and if not properly managed, it can severely degrade communication quality. For instance, in a two-user interference network in which two sources intend to communicate with their respective destinations simultaneously in the same frequency band [Ahl74], the transmission power of each source affects the signal detection, not only at its desired destination, but also at the other destination. Therefore, optimizing the performance of each source-destination pair becomes interrelated with that of the other pair, and it becomes challenging to characterize the performance limits [AC78,HK81,ETW08,JLD08,LJ08,MK09]. The situation becomes even more complicated, when the number of users exceeds two. Wireless devices must carefully coordinate the use of the radio resources in order to effectively manage interference and enable successful communication [GKGI07]. The term coordination has been widely used in various disciplines, including computer science, sociology, management science, systems theory, and economics [MC90]. The broad definition of coordination according to the American Heritage Dictionary [Dic81] is “the act of working together harmoniously.” In the context of wireless interference networks, this implies that the terminals work together in order to design their transmit signals in such a way that the interference can be managed properly. Although, they may have conflicts of interest in using the available radio resources, they intend to cooperate with each other in order to achieve a reasonably good performance. Conventional interference management strategies (such as time division multiple access (TDMA) or frequency division multiple access (FDMA)) tend to divide the available radio resources and orthogonalize the transmissions of different source– destination pairs. This requirement causes the subspaces of different interference signals to be orthogonal to those of the desired signal at each destination, and also orthogonal to each other. Interference is avoided at the cost of low spectral efficiency. However, it has been shown recently that transmitters can shape their transmitted signals in an appropriate domain, such as space (by using multiple antennas), Introduction 3 time (by coding across time-varying channels), or frequency (by coding across different carriers in frequency-selective channels) [MAMK08, CJ08, MGMAK14], or they may perform opportunistic scheduling in dense networks [JwJS09, YSJP13], in order to manage interference and conduct communication more efficiently. For instance, the interference alignment concept [MAMK08, CJ08] reveals that, with proper coordinated transmission signaling design, different interference signals can be aligned together, such that more radio resources can be assigned to the desired transmission. Consider a multi-user interference network with more than two source–destination pairs. At each destination, the interference signals can be aligned so that so that up to half of the signal space can be left to its desired signal [CJ08]. This means that each user may achieve half of the interference-free transmission rate, regardless of how many interferers exist. Moreover, in dense communication networks in which a massive number of users should be served, an opportunistic scheduling scheme can be deployed to schedule only those users that experience less interference. This scheme can take advantage of the crowd of users to manage the interference [JwJS09, YSJP13]. However, the realization of such coordinated transmission schemes can be considerably more challenging than in the case of the conventional orthogonal transmission strategies. For instance, global channel state information (CSI) must usually be perfectly available at all sources and destinations to conduct the coordinated actions (for example, beamforming, scheduling, and resources allocation). Acquiring such perfect global CSI is clearly a difficult problem in time-varying environments. In practice, no CSI is a priori available at terminals and proper channel learning schemes thus must be applied. The destinations can obtain an estimation of the CSI through a pilot-based channel training scheme in which each source allocates a portion of the total transmission time and energy for transmitting pilot symbols, and the rest for data transmission. The quality of estimated CSI determines how accurately the coordinated transmission scheme can be conducted and therefore affects the performance of each decoder. In general, each transmitter shares its radio resources between pilot transmission and data transmission. More accurate channel estimation can be obtained by allocating more resources for pilots. This implies that fewer resources are left for data transmission. In such scenarios, the achievable performance of the network needs to be analyzed carefully and the optimum radio resource allocation is an important problem to be investigated. One possible way for other terminals to obtain the CSI is to require the destination to share its estimated channel knowledge with others via channel state feedback signals. [NJGV09, BT09, KV10, AH12] showed that as long as the capacity of the feedback channel is sufficiently large – such that the CSI regarding the whole network obtained by each terminal is accurate enough – interference alignment can be realized as if perfect global CSI is available. Clearly, this requirement is not usually practical. In most existing systems, the capacity of feedback channels would be strictly limited. Each terminal can attain only erroneous global CSI (for example, the quantized global CSI if digital feedback is used [BT09, KV10, KG13], or noisy global CSI if analog feedback is deployed [AH12]) or the CSI regarding only a part 4 Introduction of the network (local CSI, for example [GCJ08]). Therefore, it may not be possible to perform interference alignment perfectly. Ideally, the interference signals at each destination should be aligned together in the same subspace, which is distinguishable from the subspace for its desired signals, so that they can be completely canceled. However, the limited CSI at each terminal means that it may no longer be straightforward to perfectly separate these two subspaces. In other words, some non-negligible interference would leak into the desired signal subspace and it would not be possible to eliminate it. The communication performance is certainly affected by such interference leakage. Similar to the above-mentioned case of two-user interference networks, the transmission power of each source influences the signal detection at all destinations. Optimizing the performance of all source–destination pairs is interrelated and is a challenging problem. The objective of this thesis is to investigate the performance limits of wireless interference networks and to design efficient coordinated transmission schemes to achieve these limits. These schemes include employing channel training and channel feedback to acquire CSI, and using the estimation of CSI to coordinate data transmission. We also address the optimum radio resource allocation for channel training and data communication, and study the existing tradeoff on this matter. In our designs, we mainly consider two types of networks in terms of the quality of service (QoS) requirements. For the first type, each source must communicate with its destination at a fixed transmit data rate. A power control problem [Zan92, JBS04, FM93a, Yat95, RFLT98, SB04, CHLT08, TCS11, Ngu09]) is studied in order to properly assign transmission power (normally the minimum) to each source in order to guarantee the transmission’s success. For the second type, each source’s transmission power is fixed. We investigated a throughput maximization problem that performs rate adaptation in order to maximize network throughput. Since interference management and power control (or rate adaptation) are generally highly intertwined in the context of wireless interference networks, our aim is to efficiently address joint design of transmission strategy and power control (or rate adaption). Below, we first provide an overview of the operation of wireless communication systems, and then discuss in more detail how the abovementioned objectives are investigated in the thesis. 1.1. Wireless Network Operations Overview 5 CH4 CH5 CH4 Power Control Rate Adaptation CH4 CH5 CH6 CH3 CH4 CH6 CH5 CH3 CH4 CH5 CH6 Channel training Channel feedback Beamforming and Filtering Data transmission Scheduling CH6 Figure 1.1: Operations in a wireless communication network. 1.1 Wireless Network Operations Overview In wireless communication systems, propagation environment is continuously changing and channels are time-varying. As mentioned in the previous section, coordinated communication techniques (such as interference alignment, coordinated multi-point (CoMP) transmission, and opportunistic scheduling) usually require certain CSI to be known at terminals in order to adapt their transmission and conduct communication [LHL+ 08, MLES13]. Therefore, a mechanism is required to acquire CSI at terminals and adapt transmission strategy, accordingly. An illustrative representation of this mechanism is shown in Fig. 1.1. In order to estimate the local CSI at each terminal, a channel training scheme can be applied (see e.g. [HH03,CJKR10,KJC11,ALH12,KRB+ 13,FKS14,Sha14]). One possible way for the other terminals to acquire such estimated CSI is to share this channel knowledge to the other terminals via channel state feedback schemes (see e.g. [LHL+ 08, BT09,FWS11,ALH12]). The terminals can then use this estimated CSI to adapt the appropriate transmission strategy. For instance, in multiple-antenna systems, terminals can compute their beamforming and filtering [LHL+ 08], while in single-antenna systems they may perform opportunistic transmission (e.g. using the ergodic interference alignment scheme [NGJV12], or user scheduling) in order to manage the interference. Moreover, depending on the system requirements, the transmitters may apply power control (see e.g. [SB06,CHLT08,FWS14b]) or rate adaptation [FWS11] in order to minimize transmission power or maximize throughput, respectively. In large wireless interference networks with many users, a user scheduling can be applied to opportunistically serve a subset of the users that experience the acceptable channel condition [LTYF03, SN07a, SN07b, SBM09, JwJS09, FGS14]. The decisions can be made based on available CSI and the terminals then conduct data transmission. In this thesis, these items are investigated for wireless interference networks. 6 Introduction Table 1.1: Thesis outline. Chapter Chapter Chapter Chapter 1.2 3 4 5 6 Network size 2×2 K ×K K ×K K × KN Antennas SISO SISO MIMO SISO CSI global global local local CSIR perfect noisy perfect noisy CSIT perfect quantized perfect quantized Thesis Scope and Contributions This thesis investigates coordinated transmission schemes for wireless interference networks and their performance limits. The thesis is presented in six chapters. A summary of the system considered in each chapter and the underlying assumptions is shown in Table 1.1. Below, we briefly present the contents along with the contributions in each chapter. Chapter 2 This chapter is a review of results, concepts, and definitions that are required for the presentation of the materials in the following chapters. The presentation starts from the definition of wireless interference networks in this thesis. We then briefly review the main research results on seeking the capacity region (i.e., the largest rate region in which reliable communication is possible) of the two-user interference networks. For networks with more than two users, we introduce the concept of interference alignment and some techniques to realize it. The contents of this chapter are in part based on • [Far12] H. Farhadi, “Interference alignment and power control for wireless interference networks,” Licentiate Thesis, KTH Royal Institute of Technology, Stockholm, Sweden, Sep. 2012. • [MFZ+ 14] N. N. Moghadam, H. Farhadi, P. Zetterberg, M. Khormuji and M. Skoglund, “Interference alignment: Practical challenges and test-bed implementation,” book chapter in Contemporary Issues in Wireless Communications, INTECH Open Access Publisher, Nov. 2014. 1.2. Thesis Scope and Contributions 7 Chapter 3 In this chapter, we study a two-user single-input single-output (SISO) interference network in which each source communicates at a given fixed rate to the intended destination. Background The optimal solution to the throughput maximization problem and the power control problem for two-user interference networks in general case is unknown. Indeed, these have been the subject of extensive research. For instance, references [Tun08, WT11, RV11, AB12, EK12] have studied the throughput maximization problem for a two-user fading interference network, in which they deployed the Han-Kobayashi coding scheme [HK81]. However, the recent development of capacity achieving codes for point-to-point communications has made it attractive to use such codes for network communications (even though they might be sub-optimal in these scenarios). Applying point-to-point code between each transmitter–receiver pair is important from a practical viewpoint as the design and operations of these codes are relatively less complexity that those designed for multi-user scenarios. For fading interference networks, the throughput maximization problem – subject to using Gaussian point-to-point codes – has been addressed in [BGT11] and the capacity region has been characterized. The power control problem has been studied only for a class of fading interference networks in which transmitters simultaneously send their messages and each receiver decodes its message by treating the interference as noise (see e.g. [CHLT08, SWB06] and references therein). Various iterative power control algorithms have been proposed (see e.g. [Zan92,Yat95,CHLT08]) that assign power to the transmitters so that every transmitter–receiver pair meets a desired signal-to-interference-plus-noise ratio (SINR). This keeps the instantaneous mutual information within each fading block at a value larger than the desired transmission rate. However, only a few studies have been conducted on the feasibility of the solutions for the power control problems. Specifically, for each fading block in which channel gains remain constant, [CHLT08, Zen92, HC00a, MDEK10] have shown whether the power control problem has feasible solutions. Indeed, regardless of how large the powers are, for some channel gains, the power control problem may not have any solution. When there is no power constraint, a feasibility criteria was derived in [Zen92], [HC00a]. It is possible to use this criteria to determine whether, for specific channel gains, the power control problem has any solution to support transmission at desired rates. Nevertheless, the solutions may require large powers. Reference [MDEK10] has extended the results to the case in which there are linear power constraints. For a fading channel, when there is a short-term power constraint in the system, due to the random nature of fading, for some channel realizations there is no choice of power control that satisfy power constraint and can guarantee successful communication at the desired rate. In this case, we define the power control problem to be infeasible, and the system to be in 8 Introduction outage. At a given transmission rate, the probability that an outage event occurs is defined as outage probability. In many communication systems, a small value of outage probability is tolerable. In these cases an ǫ-outage achievable rate region is defined as the supremum of transmission rates for which the outage probability is less than ǫ [CTB99, VH94, TV05]. The outage probability and the ǫ-outage achievable rate region for random continuously distributed channels is unknown (e.g. Rayleigh fading). Contributions Chapter 3 considers a two-user Rayleigh block-fading interference network. Each transmitter utilizes a fixed-rate point-to-point Gaussian code to communicate with its dedicated receiver. Perfect CSI is globally available at all terminals. Each transmitter is subject to a short-term power constraint. We consider five different transmission schemes. When the two transmitter-receiver pairs are orthogonally activated, inter-user interference can be completely eliminated, possibly at the cost of spectral inefficiency. When both users non-orthogonally access the available channel, inter-user interference must be taken into account during the decoding process. This leads to the four following schemes: (1) both receivers directly decode their desired messages by simply treating interference as noise; (2) both receivers conduct successive interference cancellation (SIC); (3) the first receiver performs direct decoding and the second receiver performs SIC; and (4) the first receiver performs SIC and the second receiver performs direct decoding. For each of these five schemes, we start by finding the solution to the power control problem. Next, we derive a lower bound and an upper bound on the outage probability, as functions of channel statistics, desired transmission rates, and power constraints. These results are then used to find an outer bound and an inner bound on the ǫ-outage achievable rate region. The contributions of this chapter are based on • [FWS14a] H. Farhadi, C. Wang, M. Skoglund, “Delay-limited constant-rate transmission over fading interference channels using point-to-point Gaussian codes,” submitted to IEEE Trans. Commun., 2014. • [Far12] H. Farhadi, “Interference alignment and power control for wireless interference networks,” Licentiate Thesis, KTH Royal Institute of Technology, Stockholm, Sweden, Sep. 2012. 1.2. Thesis Scope and Contributions 9 Chapter 4 In this chapter, we study a time-varying K-user SISO interference network. We assume that no CSI is a priori available at terminals, and devise a coordinated schemes to acquire CSI and also to conduct communication. Background The performance limits of K-user interference networks have attracted a great deal of interest for decades; for example, the capacity region of the two-user interference networks has been the subject of extensive research. Although various inner bounds and outer bounds on the capacity region of the two-user interference network have been proposed, the exact capacity region is still unknown in general [EK11]. Extension of the results to general K-user interference networks is even more complicated. It has been shown that when channels are time-varying, applying an interference management technique called interference alignment [MAMK08, CJ08], at a given signal-to-noise-ratio (SNR) the sum-rate K 2 log(SNR) + o(log(SNR)) can be achieved, where limSNR→∞ o(log(SNR))/SNR = 0 [CJ08]. This achievable sumrate linearly scales with the number of users and is substantially higher than the sum-rate log(SNR) + o(log(SNR)) achieved by the conventional TDMA and FDMA schemes. Furthermore, an ergodic interference alignment scheme has been developed in [NGJV12] so that, when channel gains are symmetrically distributed, the 2 sum-rate K 2 E[log(1 + 2|h| SNR)] is achievable under an ergodic setting. Such a result implies that the interference network in time-varying fading environments may not be interference-limited at any SNR. The outstanding performance promised by the aforementioned schemes is based on underlying assumptions that the CSI is perfectly known at all terminals, and an asymptotically long delay in transmission (due, for example, to the symbol extension technique in [CJ08], or the channel pairing technique in [NGJV12]) can be tolerated. Regarding the delay issue, it has been shown through studying the delay-rate tradeoff for the ergodic interference alignment scheme [KWG10,MM12,JAP12] that the delay can be reduced by sacrificing transmission data rate. In general, however, an asymptotically long delay in transmission is still required in order to exhibit the advantage of the ergodic interference alignment scheme. It is not known whether the scheme is capable of providing performance gains over orthogonal transmissions under finite delay constraint. Moreover, in practice no CSI is a priori available at terminals. They can deploy a pilot-based channel training scheme to acquire an estimate of CSI at receivers, and a channel state feedback scheme to obtain an estimate of CSI at transmitters. Channel training schemes for point-to-point communication systems have been studied extensively (see, e.g., [HH03, BLM03, LSD04, ZCLB07, Sha14]). However, in multiuser scenarios they are even more challenging and the underlying performance limits of pilot-based channel training for these systems are less known. Channel training schemes for broadcast channels have been studied in [CJKR10, KJC11], 10 Introduction for multiple-access systems in [HKD11], and for certain interference networks in [ALH12, MGL13, FKS14]. [ALH12] studied channel training for multiple antenna interference networks and also the impact of the allocated time for channel training on the system performance. Transmission power for pilot transmission was assumed to be the same as that for data transmission. In general, however, they can certainly be different. A more accurate channel estimation can be obtained by allocating more power for pilot transmission, which implies that less power is left for data transmission. The interesting problem of optimum power allocation to pilot symbols and data symbols in point-to-point communication scenarios was investigated in [HH03]. In interference networks, finding the solution to this optimum power allocation problem is even more important because the quality of CSI estimation not only affects the performance of each decoder, but also determines how accurately the interference alignment can be performed. It is also important to characterize the performance limits of the pilot-assisted channel training schemes for interference networks. Regarding acquiring CSI at the transmitter-side, several references (e.g. [BT09, KV10, FWS11, NGJV12, RG12, LK12, KMLL12, KLC11, NWHC12, CY14]) have investigated cases in which each destination provides only the quantized version of its incoming channel gains to the other terminals through channel state feedback. It has been shown that when the number of quantization bits is proportional to log (SNR), the achievable rate of interference alignment with perfect CSI at the high-SNR regime can be preserved [BT09]. However, the capacity of feedback channels is limited in practice and terminals may not be able to attain a sufficiently accurate CSI estimation. Applying interference alignment based on the imperfect CSI, inter-user interference can be only partially eliminated so that some residual interference remains at each destination. This residual interference, if not appropriately managed, will degrade the system performance. On the other hand, instead of exploiting the available (even imperfect) CSI to solely conduct interference alignment, the sources can also adapt their transmission strategies; for example, by controlling transmission data rate or power, to compensate for the aforementioned performance loss and fulfil service requirements. Contributions In this chapter, we propose a pilot-assisted ergodic interference alignment scheme. This scheme deploys pilot-based channel training in order to acquire an estimate of CSI at destinations. Each destination obtains a noisy estimate of its local CSI and sends an un-quantized (quantized) version of the estimated CSI to the other terminals via analog (digital) feedback signals. The ergodic interference alignment scheme is then applied to perform data transmissions. Each transmitter shares available radio resources between pilot transmission and data transmission phases. If analog feedback is deployed, we compute an achievable rate region and investigate the optimum power allocation between channel training and data transmission phases. We also address the scenario in which digital feedback is used to send quan- 1.2. Thesis Scope and Contributions 11 tized CSI with limited resolution. Two problems are studied in this case. The first is a power control problem in which each user wishes to successfully transmit information at a fixed rate using minimum power. We use a power control scheme that adapts transmission power values such that the mutual information corresponding to each source–destination pair is always larger than the transmission rate, which means that transmitted codewords can be successfully decoded at the desired destination. We then study a throughput maximization problem in which each source has a fixed transmission power value and the network throughput is desired to be maximized [FWS11]. Since each source only knows quantized CSI, it is not aware of the exact value of mutual information between itself and its intended destination. Therefore, for some channel realizations, the mutual information may fall below the transmission rate and communications fail, which leads to an outage event. The outage probability can be used to quantify throughput as a measure of the amount of information that can be successfully transmitted. We propose a rate adaptation scheme to maximize network throughput. As noted earlier, most existing works on ergodic interference alignment assume that the system can tolerate asymptotically long delays. To understand how the considered schemes perform in realistic situations, we extend our results to communication systems with finite delay constraint, and quantify network throughput in these delay-constrained systems. The contributions of this chapter are based on • [FWS14c] H. Farhadi, C. Wang, and M. Skoglund, “Interference alignment with limited feedback: Power control and rate adaptation,” submitted to IEEE Trans. Wireless Comm., 2014. • [FWS11] H. Farhadi, C. Wang, and M. Skoglund, “On the throughput of wireless interference networks with limited feedback,” in Proc. IEEE Int. Symp. Info. Theory (ISIT’11), St. Petersburg, Russia, Jul. 2011. • [FWS12] H. Farhadi, C. Wang, and M. Skoglund, “Power control in wireless interference networks with limited feedback,” in Proc. IEEE Int. Symp. Wireless Comm. Sys. (ISWCS’12), Paris, France, Aug. 2012. • [FKWS13] H. Farhadi, M. N. Khormuji, C. Wang, and M. Skoglund, “Ergodic interference alignment with noisy channel state information,” in Proc. IEEE Int. Symp. Info. Theory (ISIT’13), Istanbul, Turkey, Jul. 2013. • [FKS14] H. Farhadi, M. N. Khormuji, and M. Skoglund, “Pilot-assisted ergodic interference alignment for wireless networks,” in Proc. IEEE Int. Conf. Acoustics, Speech and Signal Processing (ICASSP’14), Florence, Italy, May 2014, [Best Student Paper in Signal Processing for Communications and Networking]. • [Far12] H. Farhadi, “Interference alignment and power control for wireless interference networks,” Licentiate Thesis, KTH Royal Institute of Technology, Stockholm, Sweden, Sep. 2012. 12 Introduction Chapter 5 In Chapter 5, we study a K-user multiple-input multiple-output (MIMO) interference network. Each source intends to send multiple independent data streams to its corresponding destination where the number of data streams coincides with the achievable DoF of the network. Each data stream is encoded at a fixed data rate while different streams can have different rates. Only local CSI (that is, the knowledge related to the channels directly connected to a terminal) is available at each terminal. We propose iterative algorithms to perform distributed power control and transceiver design. Transmitter beamforming matrices and receiver filtering matrices are designed to maximize the SINR at each receiver corresponding to each stream. Power control is conducted to assign the minimum power to each encoded data stream to guarantee successful communication. Background Characterizing the performance limits of K-user MIMO interference networks and designing efficient interference management schemes to achieve the limits have both generated a great deal of interest among researchers. As already noted, it is a challenge to find the capacity region of interference networks even with only two source-destination. Instead, degrees of freedom (DoF)– defined as the asymptotic scaling factor of the capacity with respect to the logarithm of SNR– has been investigated recently [CJ08,MAMK08,Jaf11]. Intuitively, DoF represents the number of independent streams that can be transmitted interference-free in the asymptotically high-SNR regime. The total achievable DoF of the conventional interference management schemes (for example, TDMA or FDMA) in which the transmissions of different source-destination pairs are orthogonalized, does not increase with the number of users. However, as we already mentioned the interference alignment concept [MAMK08, CJ08] reveals that, with proper transceiver design, different interference signals at each destination can be aligned together, such that more radio resources can be assigned to the desired transmission. In certain cases (such as a three-user MIMO interference network), applying interference alignment can achieve a total DoF equal to half of the total achievable DoF in the interferencefree network. To do so, the sources perform linear beamforming to simultaneously transmit multiple independent streams in such a way that, at each destination, interference signals align and span only half of the available signal space (in the spatial domain). Consequently, the interference can be eliminated with linear zero-forcing filter at each destination [CJ08]. Thus, a total DoF proportional to the number of users in the network can be achieved. This idea has inspired several linear beamforming and filtering design solutions for interference networks. For instance, it has been shown that the beamforming and filtering matrices can be further optimized to achieve a larger sum-rate at the finite-SNR regime [SPLL10]. Also, for MIMO interference networks with finite alphabet channel inputs, a linear beamforming design is proposed in [WXG+ 13] to maximize sum-rate. 1.2. Thesis Scope and Contributions 13 Although interference alignment can achieve a larger DoF than that achieved by orthogonal transmission strategies, several challenges must be addressed to enable the deployment of this technique in future wireless networks [APH13, MET13]. To compute transmitter beamforming matrices and receiver filtering matrices, global CSI must usually be perfectly known at all terminals [CJ08, SPLL10]. Acquiring such channel knowledge is clearly a challenging problem in practice. In most cases, it is more convenient for each terminal to obtain only the CSI corresponding to the links directly connected to it (that is, the local CSI). For instance, the local CSI can be the exact channel matrices corresponding to the local links or some function of these matrices. Accordingly, distributed interference alignment have been proposed in the literature (e.g., [GCJ11, PH11]). Consider a MIMO interference network in which each user may send multiple independent data streams. An iterative algorithm for distributed interference alignment, referred to as the DIA algorithm, is proposed in [GCJ11]. Since only local CSI is available at terminals, interference signals cannot be perfectly aligned at each receiver. Therefore, zero-forcing filtering cannot perfectly eliminate the interference and some leakage interference remains at the receivers. This algorithm iteratively minimizes the power of the leakage interference. The solution computed by the DIA algorithm achieves the DoF of the network in the high-SNR regime. However, practical systems have finite SNR and the solutions intended for the high-SNR regime may not be efficient in this case. Therefore, another iterative algorithm is proposed in [GCJ11] to maximize the signal-to-leakage interference-plus-noise ratio for each transmitted data stream. This algorithm, which is referred to as the Max-SINR algorithm, achieves a larger sum-rate than the DIA algorithm at the finite-SNR regime while achieving the same DoF at high SNR. In practical wireless environment, channel variations mean that the equivalent SINR corresponding to each transmitted data stream changes over time. Thus, adaptive transmission is required for reliable communication. Depending on the objectives and the constraints, two different types of adaptive schemes can be applied. In a class of communication systems where the maximum throughput is desired, similar to the one considered for the DIA algorithm and Max-SINR algorithm in [GCJ11], an adaptive coding and modulation scheme is required to adjust the data rate based on channel state [GC98]. In such systems, transmission powers might be fixed or can be adapted (for instance, by using an algorithm similar to the one proposed in [QZH09]) to maximize the throughput. In another group of applications, such as voice/video communications, or control over communication networks, a fixed-rate data transmission is desired instead [WGM07,CHLT08]. Therefore, a power control problem should be solved. Several iterative power control algorithms have been proposed for the uplink and downlink transmissions of cellular systems (see e.g. [Zan92, FM93b, Yat95, FJC12]). [Yat95] introduced the family of standard power control problems and showed the convergence of the corresponding iterative power control algorithm. Furthermore, joint beamforming design and power control has been studied in some MIMO communication systems (such as the MIMO downlink channel) where the power control problem is intertwined with the 14 Introduction transceiver design problem [RFLT98,SB04,CHLT08,TCS11,HZB+ 11,HZB+ 12]. In interference networks, the SINR corresponding to each source-destination pair depends on the transmission powers of all sources. The solution of the power control problem for each user is coupled to that for the other existing users. This dependency leads to conflicting goals. Specifically, when each source tries to increase the transmission power to compensate interference at its intended destination, it also increases the interference to the other destinations. Ignoring this dependency in signaling design may cause unnecessarily power demand or unsatisfactory communication quality. With regard to MIMO interference networks, directly applying conventional power control strategies when multiple source-destination pairs are non-orthogonally activated is highly likely to be infeasible or lead to large power requirements. Thus, to guarantee successful fixed-rate transmission for each sourcedestination pair, power control is conventionally carried out when the transmissions of different sources are orthogonalized. According to the recent intuitions from the interference alignment concept, this may not be a spectrally efficient transmission strategy as shown in [Far12, FWS13]. Contributions In this chapter, we aim to address the problem of transceiver design and power control for MIMO interference networks by further taking into account the achievable DoF of the network. Specifically, we consider a network where each source sends multiple data streams: the same number as the corresponding DoF achieved by interference alignment. We propose two iterative algorithms that compute transmitter beamforming matrices and receiver filtering matrices to maximize the SINR for each stream, and allocate the minimum powers to realize the desired fixed-rate communications. In both algorithms, the required power values are computed in a distributed fashion at each destination and the associated source is informed via a feedback link. In the first algorithm, the exact value of the computed power is sent, while only a one-bit feedback signal is transmitted in the second algorithm, via feedback. The proposed algorithms can provide reliable communication when multiple streams are transmitted, as each is encoded with potentially different rates. This is particulary useful in wireless networks in which each user intends to communicate multiple multimedia data with diverse contents, each with possibly different QoS requirements. Numerical evaluations confirm that these algorithms require substantially smaller power values compared to the conventional orthogonal transmission strategies. These algorithms are implemented on KTH’s four-multi test-bed, which consists of three source-destination pairs of USRP-based terminals [MFZS14, MFZ+ 14]. The experimental measurements in indoor environment also confirm the promised performance of the proposed algorithms. The contributions of this chapter are based on the following publications. • [FWS14b] H. Farhadi, C. Wang, and M. Skoglund, “Distributed transceiver design and power control for wireless MIMO interference networks,” accepted 1.2. Thesis Scope and Contributions 15 for publication in IEEE Trans. Wireless Commun., Oct. 2014. • [FWS13] H. Farhadi, C. Wang, and M. Skoglund, “Distributed interference alignment and power control for wireless MIMO interference networks,” in Proc. IEEE Wireless Commun. and Networking Conf. (WCNC’13), Shanghai, China, Apr. 2013. • [FZF+ 13] H. Farhadi, A. Zaidi, C. Wang, and M. Skoglund, “Distributed interference alignment and power control for wireless MIMO interference networks with noisy channel state information,” in Proc. Int. Black Sea Conf. Commun. and Networking (BlackSeaCom’13), Batumi, Georgia, Jul. 2013. • [Far12] H. Farhadi, “Interference alignment and power control for wireless interference networks,” Licentiate Thesis, KTH Royal Institute of Technology, Stockholm, Sweden, Sep. 2012. • [MFZ+ 14] N. N. Moghadam, H. Farhadi, P. Zetterberg, M. Khormuji and M. Skoglund, “Interference alignment: Practical challenges and test-bed implementation,” book chapter in Contemporary Issues in Wireless Communications, INTECH Open Access Publisher, Nov. 2014. 16 Introduction Chapter 6 In Chapter 6, we consider a multi-cell interference network with multiple cells, each of which has a base station and multiple mobile terminals. Each base station communicates to mobile users in the corresponding cell. We assume that no CSI is a priori available at terminals. We propose a low-complexity scheme to conduct channel training and data communication in these networks. Background It has been predicted that one of the most typical scenarios in 5G communications systems will be to support an exponentially increasing demand for data rate, in ultra-dense deployments. Such communication scenarios are characterized by a high data rate requirement that needs to be sustained, irrespective of the harsh urban propagation conditions [OBB+ 14, JMZ+ 14]. Moreover, the relatively high user density in such settings implies that channel training and feedback overhead is a major challenge. Consequently, spectrally efficient transmission techniques with low-overhead are greatly desired. In order to enhance spectral efficiency, the time-varying characteristics of wireless transmission medium can be effectively exploited. It has been shown that opportunistic transmission can benefit from the time variations of propagation environment and enhance system performance. These schemes schedule a subset of the users depending on the instantaneous CSI. Several opportunistic transmission schemes have been developed in the literature; these include opportunistic scheduling [XCS01, LTYF03, TW08, SBM09], opportunistic beamforming [VTL02], random beamforming [SH05], and opportunistic interference alignment [PDLC08, JNPS12, YSJS14, LGLL14]. The early opportunistic schemes were mainly designed to exploit multi-user diversity in single-cell communication scenarios (e.g. [XCS01, VTL02,LTYF03,SH05,TW08]). Studies have shown that opportunistic transmission schemes can also mitigate inter-cell interference, thereby achieving multiplexing gain in multi-cell communication scenarios (e.g. [PDLC08, JNPS12, YSJS14, LGLL14]). However, the aforementioned schemes exhibit several characteristics that hinder their application in dense cellular deployments. For instance, the proposed schemes in [XCS01, LTYF03, PDLC08] require CSI to be perfectly known at mobile terminals and base stations. while the schemes proposed in [VTL02,SH05,TW08,SBM09, JNPS12, YSJS14, LGLL14] require only finite rate feedback to acquire CSI at base stations, they also need CSI to be a priori known at mobile terminals. This effectively limits their scalability since the overhead required for channel estimation at mobile terminals may degrade the expected delivered performance gains. In practice, CSI is not a priori available at mobile stations and they may only obtain imperfect CSI via channel training schemes. This has a twofold impact: Firstly, base stations need to allocate part of their radio resources for channel training, which means that fewer resources will be available for data transmission; Secondly, the performance of opportunistic transmission schemes degrades as a consequence 1.2. Thesis Scope and Contributions 17 of imperfect scheduling and erroneous decoding at mobile terminals. Less is known about how to perform opportunistic transmission when no a priori CSI is available at terminals, and the performance limits of networks in such scenarios is a high priority for investigation. Contributions We consider a dense cellular communication scenario in which one base station in each cell serves a large number of mobile terminals, with no a priori CSI available. We propose a pilot-assisted opportunistic user scheduling (PAOUS) scheme consisting of low complexity channel training and one-bit feedback transmission. We compute the achievable rate region for the proposed scheme and characterize the achievable DoF region. Our results reveal that, in a multi-cell network with B base stations and a coherence time T , the achievable sum-rate increases as the number of mobile terminals scales and the total DoF Bopt (1 − Bopt /T ) is achievable, given that the number of mobile terminals in each cell scales is proportional to SNR. This result indicates that, to maximize the achievable sum DoF only a subset of base stations should be activated, where the optimum number of active base stations is Bopt = min {B, T /2}. The contribution of this chapter is based on • [FGS14] H. Farhadi, H. Ghauch, and M. Skoglund, “Pilot-assisted opportunistic user scheduling for wireless multi-cell networks,” submitted to IEEE Int. Conf. Commun. (ICC’15), London, UK, Oct. 2014. Chapter 7 In the last chapter, we summarize our contributions in the thesis, and discuss potential directions for future research. 18 Introduction Contributions Outside the Scope of the Thesis In addition to the contributions listed above, the author of this thesis has also contributed to some other related works, which are published in the papers listed below. For consistency of the thesis structure, these are not included in the thesis. Test-bed Implementation of Transceiver Design and Power Control Algorithm We implemented our proposed iterative transceiver design and power control algorithm (presented in Chapter 5) on the KTH four-multi test-bed (see [Zet, ZM12, Zet14] for more details about this test-bed). The test-bed is composed of three base stations and three mobile stations. All base stations were transmitting simultaneously and on the same frequency band. The baseband processing at the terminals was implemented on universal software radio peripheral (USRP) platforms. Each terminal has two antennas, each of which is connected to one dedicated USRP. The indoor measurements reveal that, the proposed algorithm can achieve at least 4 dB reduction in transmission power in 90% of the experiments compared to the case where MaxSINR algorithm was implemented. The power saving gains as high as 13 dB was also observed in 10% of the measurements. This implementation and the measurement results are presented in the following publications: • [MFZS14] N. N. Moghadam, H. Farhadi, P. Zetterberg and M. Skoglund, “Test-bed implementation of iterative interference alignment and power control for wireless MIMO interference networks,” in Proc. IEEE Int. Workshop on Signal Proc. Advances in Wireless Commun. (SPAWC’14), Toronto, Canada, June 2014. • [MFZ+ 14] N. N. Moghadam, H. Farhadi, P. Zetterberg, M. Khormuji and M. Skoglund, “Interference alignment: Practical challenges and test-bed implementation,” book chapter in Contemporary Issues in Wireless Communications, INTECH Open Access Publisher, Nov. 2014. Multi-user Relay Networks We have studied the achievable sum DoF of a class of multi-user SISO relay networks. In these networks, the communications between K unconnected source– destination pairs are provided by a large number of half-duplex relays. When the number of relays is sufficiently large, we show that the sum DoF of this network is K. This can be achieved through the combination of spectrally efficient relaying and interference alignment. This result implies that allowing only distributed processing and half-duplex operation can provide performance that is similar to permitting joint processing and full-duplex operation in wireless relay networks at a high-SNR regime. This material has been published in: 1.2. Thesis Scope and Contributions 19 • [WFS12] C. Wang, H. Farhadi, M. Skoglund, “Achieving the degrees of freedom of wireless multi-user relay networks,” IEEE Trans. Commun., vol. 60, no. 9, pp. 2612-2622, Sept. 2012. • [WFS10] C. Wang, H. Farhadi, and M. Skoglund, “On the degrees of freedom of parallel relay networks,” in Proc. IEEE Global Commun. Conf. (GLOBECOM’10), Miami, USA, Mar. 2010. Cognitive Interference Network We have investigated the throughput of a K-user cognitive fading interference network. Specifically, we have considered a cognitive radio network consisting of one primary and multiple secondary source–destination pairs. The secondary sources have non-causal knowledge of the message of the primary user. We have found a tuple of achievable rates by utilizing the discrete superposition model (DSM), which is a simplified deterministic channel model. The coding scheme devised for the DSM can be translated into a coding scheme for the additive white Gaussian noise (AWGN) channel model, where the rate achieved in the AWGN channel model is at most a constant gap below the one achieved in the DSM. We have also derived the average throughput of the secondary users under Rayleigh fading environments. Our results show that the sum-throughput of the proposed scheme increases with the number of secondary pairs when the interference is weak. This material has been published in: • [SFRS12] N. Schrammar, H. Farhadi, L. K. Rasmussen, M. Skoglund, “Average throughput in AWGN cognitive fading interference channel with multiple secondary pairs,” in Proc. Int. Conf. Cognitive Radio Oriented Wireless Net. and Comm. (CROWNCOM’12), Stockholm, Sweden, Jun. 2012. Copyright Notice As specified in Section 1.2, parts of the material presented in this thesis are partly verbatim, based on the thesis author’s joint works, which have previously been published or submitted to conferences and journals held by or sponsored by the Institute of Electrical and Electronics Engineer (IEEE). The IEEE holds the copyright of the published papers and will hold the copyright of the submitted papers if they are accepted. Materials (such as figures, graphs, tables, or textual material) have been reused in this thesis with permission. Chapter 2 Background I N this chapter we will review some concepts, definitions and results which are required for the presentation of the material in the following chapters. First, we will describe wireless interference networks and some of the current research challenges. Next, we will briefly review the results on the capacity region characterization of the two-user interference networks. We will explain the concept of interference alignment for larger networks with more than two source-destination pairs. Also, we will present some interference alignment techniques and will review the results on interference alignment with imperfect/partial CSI. 2.1 Wireless Interference Networks The wireless interference network is a model for communication systems composed of multiple sources and destinations. Each source intends to communicate with its dedicated destination and all sources share the same transmission medium. Because of the broadcast nature of wireless medium, each destination also overhears the signals from the unintended sources. Hence, each destination observes a noisy combination of the transmitted signals from the desired and undesired sources, weighted by the corresponding channel gains. Figure 2.1 shows a K-user wireless interference network with sources and destinations denoted as Sk and Dk (k ∈ {1, 2, ..., K}), respectively. Many practical wireless communication scenarios can be modelled as in Figure 2.1. Examples include cellular networks, device-to-device communication systems, ad-hoc networks, wireless local area networks, and cognitive radio networks. An increasing demand for wireless data traffic in the future has been forecasted. For instance, in cellular networks an exponential data traffic growth has been reported by Ericsson [Eri12]. As a consequence, current wireless networks will expand, more wireless infrastructures will be deployed, and more wireless devices will operate in such networks. This will lead to an increasing demand for radio resources such as spectrum and energy. However, the radio spectrum is scarce and is considered as one of the most expensive natural resources. Also, there are serious concerns regarding vast energy consumption and the energy budget of mobile terminals is 21 22 Background z1 h11 S1 + D1 h2 h K1 1 z2 2 S2 h1 h22 + D2 hK h 2K h1 K 2 SK hKK zK + DK Figure 2.1: K-user wireless interference network restricted due to the limited battery storage capacity. Thus, spectral and energy efficient design is essentially required for emerging wireless systems. Nevertheless, finding the optimum transmission schemes and characterizing the best performance in wireless networks is in general a challenging problem. To efficiently utilize radio resources, proper interference management and resource allocation techniques are required. Sufficient CSI knowledge at each terminal would be important. Such knowledge can be obtained through coordinations among users, e.g. in the form of pilot transmission and channel estimation at the receivers, and feedback from destinations to the sources. In practice, perfect coordination may be difficult to be guaranteed, due to different reasons such as limited resources for channel training, and limited bandwidth and noise/delay in the feedback channels. Thus, designing transmission schemes with imperfect or only partial CSI is required. In the next sections, we will briefly review some of the key transmission techniques for wireless interference networks. 2.2 Two-user Interference Networks The basic interference network is composed of two source-destination pairs. Characterizing the capacity region (the closure of the set of rate vectors for which jointly reliable communications are possible with independent sources [Car78]) of this network has been the subject of research for many years. The two-user interference network was first studied by Ahlswede, who established basic inner and outer bounds on the capacity region [Ahl74]. Some achievable rates and upper bounds have been further proposed in later literatures [EK11]. However, except in some special scenarios such as when the inter-user interference is very strong [Sat81, CE87], the capacity region in general case is still unknown. 2.3. K-user (K > 2) Interference Networks 23 Despite the intuition that interference always degrades the network’s performance, it has been shown that in certain cases the capacity region does not shrink due to interference. For example, Carleial showed that in Gaussian interference networks when the interference is very strong, each destination can first decode the message of the unintended source and subtract it from the received signal before decoding its own message [Car75]. In this way, the capacity region would not be affected by interference. The scheme was extended to the strong interference scenario and the capacity region was established by Sato [Sat81]. Costa and El Gamal further generalized this result to the discrete memoryless interference channel model [CE87]. When the interference between users is moderate or weak, destinations may not be able to decode the message of the interfering sources. Characterizing the capacity region is more challenging compared to the strong interference scenario and the very strong interference scenario. Carleial applied the superposition coding technique, which was originally developed for broadcast channels by Cover in [Cov72], to the two-user interference network. An inner bound of the capacity region was established through data splitting at the sources and successive decoding at destinations [Car78]. This inner bound was further improved via joint decoding and coded time sharing by Han and Kobayashi [HK81]. An equivalent characterization with a reduced set of inequalities was presented in [CGG08]. Some outer bounds to the capacity region have also been derived. For instance, a genie-based outer bound was presented by Kramer in [Kra04]. Etkin, Tse, and Wang used a variant of this genie-based outer bound and the Han-Kobayashi inner bound to establish the capacity region of the two-user Gaussian interference network within one-bit [ETW08]. Also, incorporating a tight outer bound, the sum capacity of the two-user Gaussian interference network with weak interference has been derived in [MK09], [SKC09], [AV09]. It has been shown that in a weak interference regime, where the channel gains between undesired source-destination pairs are below certain thresholds, using Gaussian codebooks and performing decoding by treating the interference as noise can achieve the sum capacity. 2.3 K-user (K > 2) Interference Networks Applying the above techniques developed for two-user interference networks to larger networks is not straightforward. Three major approaches to dealing with interference in multi-user interference networks are displayed in Figure 2.2. In Figure 2.2 (a) all sources simultaneously transmit in the same frequency band. Each source applies single-user coding techniques. At each destination, the desired signal cannot be distinguished from interference signals. Hence, the destination performs decoding by directly treating the interference signals as noise. In the low-SNR region, the level of interference may be limited by proper power control techniques. However, when SNR is high, inter-user interference would be dominant. Power control alone does not suffice to manage the interference and this transmission strategy 24 Background H11 S1 D1 H21 H31 H12 H22 S2 D2 H32 H13 H23 S3 D3 H33 (a) H11 S1 D1 H21 H31 H12 H22 S2 D2 H32 H13 H23 S3 D3 H33 (b) signal subspace H11 S1 interference subspace D1 H21 H31 H12 S2 H22 D2 H32 H13 H23 S3 D3 H33 (c) Figure 2.2: Transmission schemes in three-user interference networks: (a) nonorthogonal transmission and decdoing by treating interference as noise, (b) orthogonal transmission, and (c) interference alignment. 2.3. K-user (K > 2) Interference Networks 25 may not lead to a good performance. To avoid interference at the destinations, the conventional approach is to orthogonalize the transmissions of different users. Each source-destination pair has access to only a portion of the available channel, as shown in Figure 2.2 (b). Although signal reception at each destination does not directly suffer from inter-user interference, this scheme may not be spectrally efficient. This is because at each destination the interference signals span a large dimension of the received signal space, since they are unnecessarily orthogonal to each other. Clearly, if at each destination the dimension of the subspace occupied by only the interference signals can be reduced, a larger interference-free subspace would be left for desired transmission. In fact, this can be realized using a new technique called interference alignment [MAMK08]. Specifically, interference alignment for interference networks refers to “a construction of signals in such a manner that they cast overlapping shadows at the receivers where they constitute interference while they remain distinguishable at the intended receivers where they are desired” [CJ08]. In general, two conditions should be satisfied. The first is to align interference signals at the same subspace, termed interference subspace. The second is that the subspace left for the desired signal, called desired subspace, should be independent from the interference subspace. Both conditions are essential to interference alignment techniques. An illustrative representation of this concept is shown in Figure 2.2 (c). Interference alignment can be performed in different domains such as space (across multiple antennas [MAMK08, CJ08]), time (exploiting propagation delays [MJS12, MAT10] or coding across time-varying channels [CJ08, NJGV09]), frequency (coding across different carriers in frequency-selective channels [CJ08]), and code (aligning interference in signal levels [MGMAK14]). For different system models with different assumptions on the available CSI, different interference alignment techniques have been developed in the literature. In this section, we briefly review some of them. 2.3.1 Achievable Degrees of Freedom Region Consider a K-user interference network. Source Sk (k ∈ {1, 2, ..., K}) intends to send an independent message wk ∈ Wk to its destination, where Wk denotes the corresponding message set. The message wk is encoded to a codeword of length n. k| , where |Wk | denotes the cardiThus, the corresponding code rate is Rk = log |W n nality of Wk . The rate tuple (R1 , R2 , ..., RK ) is said to be achievable if a sequence of codebooks exists, such that the probability that each destination incorrectly decodes its message, can be arbitrarily small, by choosing long enough codewords. The capacity region of the network is the closure of the set of all achievable rates. In Gaussian interference networks where the noise is additive white Gaussian, the capacity region depends on the transmission powers of sources, the noise powers and channel gains. Since the exact capacity region is difficult to find, as a starting point one can use the degrees of freedom (DoF) region to characterize/approximate the capacity/achievable rate region in the high-SNR region (where interference is 26 Background the dominant phenomenon that degrades system performance). The achievable DoF region is defined as follows Rk D = (d1 , ..., dK ) ∈ R+ |∃(R1 , ..., RK ) ∈ C(p), dk = lim , k ∈ {1, ..., K} , p→∞ log p (2.1) where C(p) denotes the capacity region, and p is the transmission power of each source. The DoF can be seen as the pre-log factor of the achievable rate and the DoF region describes how the capacity region expands as transmission power increases. Another more ‘practical’ interpretation of DoF per user, dk , is the number of interference-free signaling dimensions. 2.3.2 Interference Alignment for MIMO Interference Networks In this section, we show how to align interference signals at each destination in the spatial domain, through the scheme proposed in [CJ08]. Consider a network with three source-destination pairs (K = 3). Each terminal is equipped with M antennas. For presentation simplicity here we assume M to be even. The transmission scheme when M is odd is also provided in [CJ08]. Let Sk (k ∈ {1, 2, 3}) send vector xk . Then, the channel output at destination Dk is yk = Hk1 x1 + Hk2 x2 + Hk3 x3 + zk , (2.2) where xl is an M ×1 transmitted signal vector of source Sl , Hkl is the M ×M channel matrix between Sl and Dk , and zk is an M × 1 AWGN noise vector. The channels are time-invariant and do not change during the transmission. The goal is to show that the achievable DoF for each source-destination pair is M/2. Sk (k ∈ {1, 2, 3}) transmits M/2 independent codeword streams, denoted as xik (i ∈ {1, 2, ..., M/2}), by modulating vectors vik as follows: M/2 xk = X xik vik = Vk xk , (2.3) i=1 M/2 where xk = [x1k x2k · · · xk ]T , and Vk is the beamforming matrix of Sk . Therefore, according to (2.2) the received signal of Dk can be represented as follows yk = Hk1 V1 x1 + Hk2 V2 x2 + Hk3 V3 x3 + zk . (2.4) Each destination tries to recover the desired message from the received signal. The receiver can perform linear filtering as follows yk = U∗k yk , (2.5) and decode its desired messages. There are two interference signals at each destination. The interference signals at each destination will be aligned if we can design 2.3. K-user (K > 2) Interference Networks 27 the beamforming matrices V1 , V2 and V3 such that they satisfy the following conditions: Alignment at D1 : Alignment at D2 : span(H12 V2 ) = span(H13 V3 ), H21 V1 = H23 V3 , Alignment at D3 : H31 V1 = H32 V2 , (2.6) where span(A) is the space spanned by the column vectors of matrix A. Thus, the interference signals only occupy an M/2-dimensional subspace and the first requirement of interference alignment is satisfied. Since the elements of matrices Hkl (∀ k, l ∈ {1, 2, 3}) are randomly chosen from continuous distributions, then the latter has a full rank of M , almost surely. Therefore, the above set of equations can be re-written as follows: span(V1 ) = V2 = V3 = span(EV1 ), FV1 , GV1 , (2.7) where E −1 −1 , H−1 31 H32 H12 H13 H23 H21 , F G , H−1 32 H31 , , H−1 23 H21 . (2.8) This problem has different solutions. One of the solutions is V1 = [e1 , e2 , ..., eM/2 ], V2 V3 = F[e1 , e2 , ..., eM/2 ], = G[e1 , e2 , ..., eM/2 ], (2.9) where ei (i ∈ {1, ..., M }) are the eigenvectors of matrix E. It can be verified that the solution in (2.9) satisfies all the alignment conditions in (2.6). To retrieve the desired message from the received signal, the second condition for interference alignment, which requires the desired signal subspace and the interference subspace to be linearly independent, must be satisfied. This requirement can be fulfilled if the following conditions are satisfied: rank ([H11 V1 , H12 V2 ]) = M rank ([H22 V2 , H21 V1 ]) = M rank ([H33 V3 , H31 V1 ]) = M. (2.10) It has been shown in [CJ08] that these conditions are almost surely satisfied. Therefore, sum DoF 3M/2 is achievable in this network. Figure 2.3 shows an illustrative 28 Background H13 V3 V1 H11 S1 D1 H12 V2 H11 V1 H21 H31 H22 V2 H12 V2 H22 S2 D2 H21 V1 H23 V3 H32 H13 H31 V1 H23 V3 S3 D3 H33 H32 V2 H33 V3 Figure 2.3: Interference alignment in a three-user MIMO interference network with two antennas (M = 2) at each terminal. example of the solution of this scheme when M = 2. In this network each source can transmit one stream performing interference alignment. The receivers can retrieve the desired message from the received signal by zero-forcing filtering. If two conditions of the interference alignment are fulfilled, the transmitter-side beamforming matrices and the receiver-side filtering matrices satisfy the following conditions: U∗k Hkj Vj = rank(U∗k Hkk Vk ) = 0, ∀j 6= k : j, k ∈ {1, 2, 3}, M , ∀k ∈ {1, 2, 3}. 2 (2.11) It is clear that the solution requires global CSI to be available at each terminal. Specifically, matrix E is required for the calculation of the solution in (2.9) and also finding the zero-forcing receiver at each destination. 2.3.3 Interference Alignment for SISO Interference Networks When there is only one antenna at each terminal, interference signals cannot be aligned in the spatial domain. The techniques mentioned in the previous sections cannot be applied. However, it has been shown that in time-varying or frequencyselective fading environments, interference alignment is still possible. We use an example to reveal the basic idea of interference alignment for time-varying SISO interference networks. Consider a three-user interference network (K = 3). The received signal at Dk (k ∈ {1, 2, 3}) is: yk (t) = hk1 (t)x1 (t) + hk2 (t)x2 (t) + hk3 (t)x3 (t) + zk (t), (2.12) 2.3. K-user (K > 2) Interference Networks 29 where xl (t) is the transmit symbol of Sl at time instant t, and hkl (t) is the channel coefficient between Sl and Dk at time instant t. The channel coherence time is assumed to be one time slot (channel gains remains constant within one time slot, but change independently across different time slots). Global CSI is perfectly known at all terminals. Since each terminal has only one antenna, at each time slot there are not enough spatial dimensions to separate interference subspace with desired signal subspace. This problem can be resolved using the symbol extension technique proposed in [CJ08]. We define the q symbols transmitted over q time slots by Sk as a vector: xk (t) , [xk (q(t − 1) + 1), xk (q(t − 1) + 2), ..., xk (qt)]T . (2.13) Similarly, denote the q symbols received over q time slots by Dk as a vector: yk (t) , [yk (q(t − 1) + 1), yk (q(t − 1) + 2), ..., yk (qt)]T . (2.14) The noise vector zk (t) is the similar expansion of the noise over q time slots. Thus, the received signal at Dk can be expressed as follows yk (t) = Hk1 (t)x1 (t) + Hk2 (t)x2 (t) + Hk3 (t)x3 (t) + zk (t). (2.15) In this equation Hkl (t) (k, l ∈ {1, ..., K}) is a diagonal extended channel matrix defined as follows hkl (q(t − 1) + 1) 0 ··· 0 0 hkl (q(t − 1) + 2) · · · 0 Hkl (t) , (2.16) .. .. .. .. . . . . . 0 0 · · · hkl (qt) This is called the extended interference channel model where each destination has a q-dimensional received signal space. The goal of interference alignment design is to align all the interference signals at every destination within one half of the total received signal space, leaving the other half interference-free for the desired signal. Let q = 2n + 1, where n is a positive constant. The source S1 encodes its message to m n + 1 independent data streams xm 1 (t) (m = 1, 2, ..., n + 1). Each data stream x1 (t) m is sent along a q × 1 vector v1 . Therefore, x1 (t) can be represented as follows: x1 (t) = n+1 X m xm 1 (t)v1 = V1 x1 , (2.17) m=1 (t)]T and V1 , [v11 , v21 , · · · , vn+1 ]. Similarly, S2 where x1 , [x11 (t), x21 (t), · · · , xn+1 1 1 and S3 encode their messages to n independent data streams as follows: x2 (t) = x3 (t) = n X m=1 n X m xm 2 (t)v2 = V2 x2 , m=1 m xm 3 (t)v3 = V3 x3 . (2.18) 30 Background Therefore, the received signal at Dk can be expressed in the following matrix format yk (t) = 3 X Hki (t)Vi xi + zk (t). (2.19) i=1 As we mentioned in Section 2.3 two conditions should be satisfied to realize interference alignment. First, the interference signals should be aligned at each destination such that interference occupies a subspace with dimensions less than the total dimensions of the available signal space. The second condition is that the interference and desired signal subspaces should be independent. To obtain the (n+ 1)-dimensional interference-free desired signal subspace from the received (2n + 1)dimensional signal y1 (t), the number of dimensions of the interference subspace must not be larger than n. This can be achieved by aligning the interference signals from S2 and S3 as follows H12 (t)V2 = H13 (t)V3 . (2.20) To have an n-dimensional interference-free subspace at D2 , the interference signals from S1 and S3 must be aligned as follows span(H23 (t)V3 ) ⊂ span(H21 (t)V1 ). (2.21) Similarly, the alignment condition at D3 is span(H32 (t)V2 ) ⊂ span(H31 (t)V1 ). (2.22) This set of equations have more than one solution. One of them is V1 = [u, Tu, ..., Tn u], V2 V3 n−1 = H−1 u], 32 (t)H31 (t)[u, Tu, ..., T −1 2 n = H23 (t)H21 (t)[Tu, T u, ..., T u], (2.23) −1 −1 where T = H12 (t)H−1 21 (t)H23 (t)H32 (t)H31 (t)H13 (t) and u is a (2n + 1) × 1 all-one vector. To check the second condition, it has been shown in [CJ08] that the columns of matrix [H11 (t)V1 , H12 (t)V2 ] (2.24) are linearly independent, with probability one. Thus, the desired signal and interference subspaces at D1 can be separated, almost surely. The same results hold for the following matrices corresponding to the other destinations [H22 (t)V2 , H21 (t)V1 ], (2.25) [H33 (t)V3 , H31 (t)V1 ]. (2.26) 2.3. K-user (K > 2) Interference Networks 31 Consequently, S1 , S2 and S3 can transmit n + 1, n and n independent messages, n+1 n n respectively over 2n + 1 channel uses. Thus, the DoF tuple 2n+1 , 2n+1 , 2n+1 is achievable in the asymptotic case where the transmission power goes to infinity. As n → ∞, the DoF tuple 12 , 21 , 12 is achievable. Every user can achieve half of its interference-free DoF (i.e., DoF one, since only one user is present). This result generalizes to the K-user case, where each user is still able to achieve 1/2 DoF. 2.3.4 Ergodic Interference Alignment The interference alignment scheme described in the previous section promises that each user can achieve 1/2 DoF only as SNR goes to infinity. For a time-varying fading environment and when the fading coefficients follow symmetric distributions, an ergodic interference alignment scheme is proposed in [NJGV09] to align the undesired signals and eliminate interference at any SNR. We briefly review the idea of the ergodic interference alignment in this part, and will provide more details about this scheme in Chapter 4. Consider a K-user SISO interference network. Assume that Sk (k ∈ {1, ..., K}) transmits signal xk at time slot t. Then, the received signal at Dk is yk (t) = hkk xk + K X hkl xl + zk (t), (2.27) l=1,l6=k where hkl is the channel fading coefficient between Sl and Dk and is perfectly known at all terminals. Let channel matrix at time t be denoted as follows h11 h12 ··· h1K h22 ··· h2K h21 . H(t) , (2.28) .. .. .. .. . . . . hK1 hK2 ··· hKK Then, the complement channel matrix to H(t) can be defined as follows h11 −h12 ··· −h1K h22 ··· −h2K −h21 . H(tc ) , .. .. .. .. . . . . −hK1 −hK2 ··· hKK (2.29) Assume that in time slots t and tc , the channel matrices are complement to each other. If the sources transmit the same signals in time slot tc , the received signal at Dk is: yk (tc ) = hkk xk − K X l=1,l6=k hkl xl + zk (tc ). (2.30) 32 Background If Dk adds the two signals it received in time slots t and tc , we have yk = yk (t) + yk (tc ) = 2hkk xk + zk (t) + zk (tc ). (2.31) Therefore, an interference-free channel can be provided at a price of two channel uses. We will refer to this as the channel pairing technique throughout the thesis. For symmetric fading distributions (e.g. Rayleigh fading distribution), the complement channel matrices H(t) and H(tc ) occur with the same probability. Since the channel gains are drawn from a continuous distribution, the probability that each individual channel matrix happens is equal to zero. To handle this issue, the channel pairing can be performed based on quantized channel gains. Thus, if the time duration of interest is long enough, for every channel matrix, a corresponding complement channel matrix can be found, almost surely [NJGV09]. To take this advantage, let each source transmit an independent codeword to its destination in every time slot. If in one time slot, the channel matrix is complement to that in a previous slot, all sources repeat their codewords. In this way, if the time duration of interest is long enough, it is as if that the sources use half of the total time slots to send independent signals to their destination, and the transmissions do not experience inter-user interference. This means the average rates Rk = 21 E[log(1 + 2|hkk |2 SNR)] (k ∈ {1, 2, ..., K}) are achievable. Each user can achieve half its interference-free data rate at any SNR, no matter how many users exist in the network. This rate is archived at the cost of asymptotic delay assuming that perfect CSI is globally available at terminals. The data transmission employing this technique for delaylimited systems for which no CSI is a priori available at terminals is investigated in Chapter 4. 2.4 Wireless Interference Networks with Imperfect CSI To achieve the outstanding performance promised by the aforementioned schemes, in general perfect global CSI is required at all terminals. In practice, no CSI is a priori available at terminals. An estimate of CSI can be acquired through a pilotbased channel training scheme in which each source allocates a portion of the total radio resources for transmitting pilot symbols and the rest for data transmission. The pilot-based channel training schemes have been investigated in the literature for several multiple antenna communication systems including point-to-point communication systems (see e.g. [HH03, KRB+ 13, SWB14]), and multi-user communication systems (e.g. broadcast channel [CJKR10,KJC11], uplink systems [HKD11], interference channel [XYMN11, ALH12, MGL14]). In [ALH12], the impact of the allocated time for channel training on the performance of interference alignment for multiple antenna systems has been addressed. It has been assumed that transmission power for pilot transmission is the same as the one for data transmission. In general, they can be different. A more accurate channel estimation can be obtain by allocating more power for pilot transmission which implies a lower power is left for data transmission. The interesting problem 2.4. Wireless Interference Networks with Imperfect CSI 33 of optimum power allocation to pilot symbols and data symbols in point to point communication scenarios has been investigated in [HH03]. In multi-user interference networks, finding the optimum power allocation to pilot symbols and data symbols is even more important because of the fact that the quality of CSI estimation not only affects the performance of each decoder, but also determines how accurately the interference alignment can be performed. In Chapter 4, we will investigate the performance limits of the PAEIA scheme for multi-user interference networks, and will study this resource allocation problem. Also, similar problem is addressed in Chapter 6 for the PAOUS scheme proposed for multi-cell interference networks. After the estimation of CSI, the destinations can transmit either un-quantized CSI using analog feedback or quantized CSI employing digital feedback to inform the other terminals about the CSI. In [AH12] an analog channel state feedback scheme for interference alignment is proposed. Destinations directly transmit the channel coefficients as un-coded quadrature and amplitude modulated symbols. It has been shown that no loss in achievable DoF is incurred, if the transmission power of the analog feedback signals scales similar to the actual transmission power of the sources. Digital channel state feedback strategies for the interference alignment scheme mentioned in Section 2.3.3 are provided in [BT09, KV10, FWS11, NGJV12, RG12, LK12,KMLL12,KLC11,NWHC12,CY14]. It has been shown that the same DoF as when perfect CSI is available can be obtained, provided that the feedback signals’ rate is proportional to log P , where P is the transmission power of each source. This result is also observed in multiple-antenna cases in [KV10]. A continuous tradeoff between the scaling of feedback rate and achievable DoF has also been demonstrated. This feedback strategy is further improved in [KMLL12] by introducing proper filtering before quantizing the channel gains. Clearly, sufficiently accurate CSI estimates can still guarantee the performance of interference alignment. However, in practice the radio resources (e.g. bandwidth and power) available for feedback channels are limited so that terminals may not be able to attain CSI estimates of sufficient quality. The issue will be considered in this thesis. Specifically, in Chapter 4, we will study the performance of the ergodic interference alignment scheme, when the quantization resolution is strictly limited. Chapter 3 Two-user Interference Networks: Point-to-Point Codes T HIS chapter studies a two-user interference network where each transmitter intends to communicate to the corresponding receiver at a fixed rate. The channels follow a block-fading model with Rayleigh distributed amplitudes. We assume that perfect CSI is globally available at all terminals. Each transmitter is subject to a short-term power constraint. We consider the use of a point-topoint Gaussian code to conduct communication between each transmitter-receiver pair, and investigate five different transmission schemes. When the two transmitterreceiver pairs are orthogonally activated, inter-user interference can be completely eliminated, with the possible price of spectral inefficiency. When both users nonorthogonally access the available channel, inter-user interference must be taken into account at the decoding process. This leads to four schemes: 1) both receivers directly decode their desired messages by simply treating interference as noise; 2) both receivers conduct successive interference cancellation (SIC); 3) the first receiver performs direct decoding and the second receiver performs SIC; 4) the first receiver performs SIC and the second receiver performs direct decoding. For each of these five schemes, we find the solution of the power control problem. For some channel realizations this problem may not have a feasible solution. In these cases, an outage event will be declared. We compute a lower bound and an upper bound on the outage probability, as functions of channel statistics, desired transmission rates, and power constraints. These results are then used to find an outer bound and an inner bound on the ǫ-outage achievable rate region. The structure of this chapter is as follows. In Section 3.1, we present the two-user interference network model and define the performance metrics which will be investigated in this chapter. Section 3.2 addresses the orthogonal transmission scheme. The non-orthogonal transmission schemes are described and their performance limits are analyzed in Section 3.3. Finally, Section 3.4 summarizes this chapter. 35 36 Two-user Interference Networks: Point-to-Point Codes H m1 E1 {x1 (t)}nt=1 P1 {z1 (t)}nt=1 n √ p1 x1 (t) t=1 + h11 {y1 (t)}nt=1 H m ˆ1 D1 1 h2 H {z2 (t)}nt=1 E2 {x2 (t)}nt=1 P2 h1 2 m2 n √ p2 x2 (t) t=1 h22 + {y2 (t)}nt=1 H m ˆ2 D2 Figure 3.1: Two-user interference network. 3.1 Two-user SISO Interference Network We consider a single-antenna fading interference network with two transmitterreceiver pairs as shown in Figure 3.1. The channels follow a Rayleigh block-fading model, in which the channel gains remain constant within a coherent interval (the time slots in which the considered communications occur and their number is denoted as n), and independently change across different intervals. We denote the fading coefficient of the link between the source Sk (k ∈ {1, 2}) and the destination Dk (direct link) as hkk ∼ CN (0, σS2 ), and denote that of the link between Sk and Dl (l ∈ {1, 2}, l 6= k) (interference link) as hlk ∼ CN (0, σI2 ). The parameters σS2 and σI2 are the variances of the direct links and the interference links, respectively. These generally can have different values and their ratio is denoted as ρ , σS2 /σI2 . (3.1) The channel gains are mutually independent. We denote the network channel matrix as H whose element on the ith row and the jth column is hij . In this chapter, we assume that perfect channel knowledge to be globally available, i.e. H is known at every terminal. During one fading block, each source sends one message to its destination. The message from Sk , mk , is encoded using a point-to-point Gaussian codebook with fixed data rate Rk (bits/channel use). Definition 3.1 (Encoders). The source Sk (k ∈ {1, 2}) has an encoding function, Ek : Mk → Cn , that maps its message mk, which is independently and uniformly chosen from the set Mk = 1, 2, ..., 2nRk , into a length-n codeword {xk (t)}nt=1 . The codeword satisfies the power constraint n 1X 2 |xk (t)| ≤ 1. n t=1 (3.2) Each encoder is concatenated with a power controller as shown in Figure 3.1. 3.1. Two-user SISO Interference Network 37 Definition 3.2 (Power controllers). The power controller associated with Sk (k ∈ n {1, 2}) applies a function, Pk : Cn → Cn , that scales the codeword {xk (t)}t=1 ac√ n cording to the channel gains to pk xk (t) t=1 , where pk denotes the average transmission power of Sk . The assigned power obeys short-term constraint, i.e. within each fading block it satisfies pk ≤ pmax,k , (3.3) where pmax,k is the maximum transmission power of Sk . Definition 3.3 (Decoders). Each receiver has a decoding function, Dk : Cn → Mk , that maps its observed length-n channel output {yk (t)}nt=1 to an estimate m ˆ k of the transmitted message mk . We address the power control problem for different transmission schemes in the considered network. Specifically, in each fading block, transmitters seek proper transmission powers which meet the short-term power constraints in (3.3) and also guarantee successful communications for both transmitter-receiver pairs. In the following, we provide some definitions and performance metrics regarding the considered system. Definition 3.4 (Achievable rate region). Within each fading block, for a given transmission power vector p (p = [p1 p2 ]T ) and channel matrix H, applying a transmission scheme ‘A’, the average probability of error is defined as n o ˆ 1, M ˆ 2 ) 6= (M1 , M2 ) , (3.4) Pe(n) = Pr (M ˆ k (k ∈ {1, 2}) denote a randomly transmitted message and the where Mk and M corresponding decoded message, respectively. If there exist channel encoding and (n) decoding functions such that limn→∞ Pe = 0, then we say that a rate pair (R1 , R2 ) is achievable. We denote the corresponding achievable rate region as C A (p, H). Definition 3.5 (Feasible transmission scheme). If the achievable rate region corresponding to a power vector p, where 0 p pmax , and pmax = [pmax,1 pmax,2 ]T , includes the required transmission rates of the network, i.e. (R1 , R2 ) ∈ C A (p, H), then the power control problem for a transmission scheme ‘A’ has a solution, and we say that the transmission scheme ‘A’ is feasible. We define the set of these vectors as follows A PH ((R1 , R2 ) , pmax ) , p : 0 p pmax , (R1 , R2 ) ∈ C A (p, H) . (3.5) A Using ∅ to denote an empty set, PH 6= ∅ means that the transmission scheme ‘A’ is feasible. In fact, due to the random nature of fading, for some channel realizations, a transmission scheme ‘A’ may not be feasible. In such cases, the messages cannot be successfully transmitted, and the system is said to be in outage. We define the outage probability as follows. 38 Two-user Interference Networks: Point-to-Point Codes Definition 3.6 (Outage probability). The outage probability of a transmission scheme ‘A’ is defined as follows A A Pout ((R1 , R2 ) , pmax ) , Pr PH ((R1 , R2 ) , pmax ) = ∅ . (3.6) Some wireless applications can tolerate a certain amount of outage probability. Thus, it is desirable to characterize the rate region for which an outage probability less than the maximally tolerable outage probability can be attained. In the following we provide the definition of a related performance measure which will be investigated for the considered system in this chapter [TV05]. Definition 3.7 (ǫ-outage achievable rate region). The ǫ-outage achievable rate region of a transmission scheme ‘A’ is A CǫA (pmax ) , (R1 , R2 ) : Pout ((R1 , R2 ) , pmax ) ≤ ǫ , (3.7) where ǫ is the maximum outage probability that a specific application can tolerate. In the following sections, we study the performance of different transmission schemes. Our analysis starts from the orthogonal transmission scheme. 3.2 Orthogonal Transmission Scheme In wireless networks, inter-user interference may significantly degrade the communication system’s performance. As discussed in Chapter 2, one intuitive solution to deal with interference is to orthogonalize different users’ operations. Since each user has access to only a fraction of the available channel, this can provide interferencefree communication for each user. We term this transmission scheme orthogonal transmission (OT) throughout the chapter. Using δ (0 < δ < 1) to denote the channel-sharing factor, the fractions of the total channel used by the first and the second transmitters are δ and (1 − δ), respectively. In a fading block, the achievable rate region C OT (p, H) includes the rate pairs (R1 , R2 ) which satisfy |h11 |2 p1 (3.8) R1 ≤ δ log2 1 + δN0 |h22 |2 p2 R2 ≤ (1 − δ) log2 1 + , (3.9) (1 − δ)N0 where N0 is the noise power at each receiver. The set of solutions of the corresponding power control problem is OT PH ((R1 , R2 ) , pmax ) = {p : nT p pmax }, (3.10) (1 − δ)N T δN 0 0 R2 /(1−δ) R1 /δ 2 −1 . nT = 2 −1 2 2 |h11 | |h22 | (3.11) in which 3.2. Orthogonal Transmission Scheme 39 Clearly, the minimum required transmission powers of the OT scheme are δN 0 pOT = 2R1 /δ − 1 (3.12) 1 |h11 |2 (1 − δ)N 0 . (3.13) pOT = 2R2 /(1−δ) − 1 2 |h22 |2 This is similar to the channel inversion method in single-user point-to-point communication systems [GV97]. The solutions in (3.12) and (3.13), however, may violate the maximum power constraints in (3.3). Thus, with a certain probability, the scheme is infeasible and outage events occur. The following proposition characterizes the outage probability. Proposition 3.2.1. The outage probability of the OT scheme is OT Pout ((R1 , R2 ) , pmax ) = 1 − e − N0 σ2 S (1−δ) δ +(2R2 /(1−δ) −1) p (2R1 /δ −1) pmax,1 max,2 Proof. See Appendix 3.A. . (3.14) The channel sharing factor δ can be carefully selected to minimize the outage OT probability. It can be shown that Pout ((R1 , R2 ) , pmax ) is a convex function of δ. The optimum choice of δ, denoted as δopt , can be found by solving the following equation: pmax,1 R2 /(1−δopt ) R1 ln 2 R2 ln 2 1− −1 − 2R1 /δopt 1− 2 −1 = 0. pmax,2 (1−δopt) δopt For instance, if pmax,1 = pmax,2 and R1 = R2 , then δopt = 0.5 is the solution. The result in Proposition 3.2.1 can be used to obtain the boundary of the ǫ-outage achievable rate region CǫOT (pmax ) by solving OT Pout ((R1 , R2 ) , pmax ) = ǫ, for (R1 , R2 ). The solution is ( R1 = t δ × R2 = (1 − δ) log2 1 − 2t/δ − 1 1−δ where pmax,2 pmax,1 (3.15) − pmax,2 1−δ × σ2 pmax,1 × S ln(1 − ǫ) . 0 ≤ t ≤ δ log2 1 − δ N0 2 σS N0 , ln(1 − ǫ) (3.16) (3.17) Each (R1 , R2 ) corresponding to one particular t shows one point on the boundary of CǫOT (pmax ). As mentioned earlier, the OT scheme eliminates inter-user interference with the possible price of spectral inefficiency. In fact, depending on the values of σS2 , σI2 and pmax , permitting both transmitters to send messages non-orthogonally may outperform the OT scheme. In the next section, we will focus on non-orthogonal transmission schemes. 40 Two-user Interference Networks: Point-to-Point Codes 3.3 Non-Orthogonal Transmission Schemes We permit the two transmitters to transmit non-orthogonally. The source Sk (k ∈ {1, 2}) sends a unit-power codeword xk as described in Section 3.1. The channel outputs at time t (t ∈ {1, ..., n}) are √ √ y1 (t) = p1 h11 x1 (t) + p2 h12 x2 (t) + z1 (t), √ √ y2 (t) = p1 h21 x1 (t) + p2 h22 x2 (t) + z2 (t), (3.18) where zk (t) is the AWGN with power N0 . Each destination may either directly decode its intended message by treating interference as noise, or it may first decode the message of the unintended source and next decode its intended message after removing the interference. Thus, depending on the decoding strategy, we have four different transmission schemes. We study their power control strategies and performance in what follows. 3.3.1 Direct Decoding at Both Receivers In this part, we require each receiver to directly decode its desired message by treating the interference as noise. We term this transmission scheme NOT1 . For a fading block, the achievable rate region C NOT1 (p, H) includes the rate pairs (R1 , R2 ) which satisfy |h11 |2 p1 , (3.19) R1 ≤ log2 1 + |h12 |2 p2 + N0 |h22 |2 p2 R2 ≤ log2 1 + . (3.20) |h21 |2 p1 + N0 After some mathematical manipulations, we can present these conditions as the following power constraints 2R1 − 1 p1 ≥ |h12 |2 p2 + N0 , (3.21) 2 |h11 | 2R2 − 1 |h21 |2 p1 + N0 . (3.22) p2 ≥ |h22 |2 The constraints in (3.21) and (3.22) can be shown in the matrix format p DS FS p + nS , where DS = " nS = 0 2R1 − 1 0 2R2 − 1 2 R1 N0 −1 |h11 |2 2 # , FS = R2 (3.23) " 0 |h21 |2 |h22 |2 N0 −1 |h22 |2 T . |h12 |2 |h11 |2 0 # , (3.24) 3.3. Non-Orthogonal Transmission Schemes 41 The matrix DS is related to the transmission rates, while the matrix FS depends only on the channel gains. Power Control Solution The positive element-wise minimum transmission power vector among those which satisfy the constraint in (3.23) - if there is any such vector - is [CHLT08] pNOT1 = (I − DS FS )−1 nS , (3.25) where I is the 2 × 2 identity matrix. Therefore, the minimum required powers are |h21 |2 N0 2R1 − 1 |h 2 + l | 11 1 = pNOT , (3.26) 1 |h21 |2 (1 − l) |h12 |2 N0 2R2 − 1 |h 2 + l | 22 1 , (3.27) pNOT = 2 |h12 |2 (1 − l) |h12 |2 |h21 |2 where l = 2R1 − 1 2R2 − 1 |h 2 2 . For some channel realizations, there is no 11 | |h22 | positive power vector that satisfies the constraints in (3.3) and (3.23). In these cases, the power control problem does not have any feasible solution, and transmissions will cause outage events. In the following, we will investigate the outage probability. Outage Probability Analysis For a given channel H, the set of feasible solutions of the power control problem is NOT1 PH ((R1 , R2 ), pmax ) , {p : p DS FS p+nS , 0 p pmax } . (3.28) Figure 3.2 illustrates this set when there is at least one vector p that satisfies the constraints in (3.3) and (3.23). The red region in the figure indicates the powers which satisfy the maximum power constraint in (3.3), and the green region illustrates the powers which satisfy the inequality in (3.23). The intersection of the NOT1 two regions indicates the powers within the set PH ((R1 , R2 ) , pmax ). For some channel realizations, there is no positive power vector which satisfies (3.23), or such a positive power vector exists but does not satisfy the maximum power constraints NOT1 in (3.3). In this case, PH = ∅, i.e. the power control problem is infeasible, and an outage event occurs. We aim to characterize the outage probability. For this purpose, we first present the following lemma to provide a necessary and sufficient condition for the existence of a positive vector that satisfies (3.23). Lemma 3.1. There exists at least one positive vector p (p ≻ 0) that satisfies the inequality in (3.23), if and only if λmax (DS FS ) < 1, where λmax (DS FS ) denotes the largest magnitude of the eigenvalues of matrix DS FS given in (3.24). Proof. The proof is similar to that of Theorem 5 in [HC00b]. 42 Two-user Interference Networks: Point-to-Point Codes p2 pmax,2 1 pNOT 2 nS,2 nS,1 1 pNOT 1 pmax,1 p1 Figure 3.2: Solution of the power control problem for the NOT1 scheme. Now, we use the above result to characterize a lower bound and an upper bound NOT1 on the outage probability of the NOT1 scheme denoted as Pout ((R1 , R2 ) , pmax ). Proposition 3.3.1. The outage probability of the NOT1 scheme is bounded as o o n n NOT1 NOT1 NOT1 NOT1 NOT1 (3.29) . ≤ Pout ((R1,R2) , pmax) ≤ min 1, 2−PF,1 −PF,2 1−min PF,1 ,PF,2 In these equations γβk σS4 −(bk +ρak ) βk ak σS4 k ak ρ ak ρ(1−γ)−γb γ e e + E1 ρ(1−γ) 1−γ γ 4 −bk βk γσS e ak ρ ak ρ(1−γ) γ + e E (a ρ) − E , k ∈ {1, 2}, (3.30) 1 k 1 ρ(1 − γ)2 γ NOT1 PF,k , − where The function E1 (x) , γ , ak , bk , βk , R∞ x Proof. See Appendix 3.B. e−t t dt ρ2 / 2R1 − 1 2R2 − 1 , N0 / σS2 pmax,k , N0 2Rk − 1 / σS2 pmax,k , ρ (ak ρ−bk ) e . σS4 (3.31) denotes exponential integral as defined in [AS64]. 3.3. Non-Orthogonal Transmission Schemes 100 ldbc ldbc ldbc ldbc 43 bcld ldbc ldbc ldbc bcld bc ld 10−1 bc ld bc ld bc bc A Pout ld bc ldbc 10−2 bc bc bc ld 10−3 ld ld 10−4 0 10 NOT1 Pout , NOT1 Pout,u , NOT1 Pout,l , NOT1 Pout , NOT1 Pout,u , NOT1 Pout,l , OT Pout 20 ld ρ = 30 ρ = 30 ρ = 30 ρ = 40 ρ = 40 ρ = 40 dB dB dB dB dB dB bc bc bc bc bc bc bc bcbc bcbc bc bc ld ld ld ldld ld ld ld ld ld ld ld 30 40 50 ld 60 ldld 70 SNRmax (dB) A Figure 3.3: Pout versus SNRmax , R1 = R2 = 5 (bits/channel use). In general, deriving the closed-form of outage probability for the NOT1 scheme is involved. However, in the following two special cases, the exact value of this probability can be found. The first case is when one transmitter has sufficiently large maximum transmission power. Specifically, if pmax,1 → ∞, then condition 1 < pmax,1 in (3.79) almost surely holds and comparing (3.79) and (3.81) in pNOT 1 Appendix 3.B we can see that NOT1 NOT1 Pout ((R1 , R2 ) , pmax ) = 1 − PF,2 . (3.32) Similarly, if pmax,2 → ∞, then NOT1 NOT1 Pout ((R1 , R2 ) , pmax ) = 1 − PF,1 . (3.33) The next case is when one of the transmission rates is sufficiently small. Specifically, 1 1 if R1 → 0, then according to (3.26) pNOT → 0, and condition pNOT < pmax,1 in 1 1 (3.79) almost surely holds. Comparing (3.79) and (3.81) in Appendix 3.B it can be concluded that the outage probability is that in (3.32). Also, it can be shown that if R2 → 0, then the exact outage probability is the probability given in (3.33). In other cases, we can verify the tightness of the bounds through numerical evaluation. 44 Two-user Interference Networks: Point-to-Point Codes NOT1 NOT1 Figure 3.3 shows Pout,l ((R1 , R2 ) , pmax ) and Pout,u ((R1 , R2 ) , pmax ) given in Proposition 3.3.1 versus SNRmax (SNRmax = pmax,1 /N0 = pmax,2 /N0 ) for different values of ρ. Also, the simulation results of the exact outage probability NOT1 OT Pout ((R1 , R2 ) , pmax ) and the value of Pout ((R1 , R2 ) , pmax ), computed according to Proposition 3.2.1, are shown for comparison. In this example, we set σS2 = 1 and R1 = R2 = 5 (bits/channel use). We can observe that in a certain range of SNRmax , NOT1 outperforms OT by attaining lower outage probabilities. The range increases when ρ becomes larger. However, when SNRmax is selected to be sufficiently large, the outage probability is no longer sensitive to the change of power constraints, i.e. an error floor exists. In this region, the OT scheme can attain better performance in terms of outage probability. The reason is that at high SNRmax , the constraint in (3.23) is the dominant constraint which is required to be satisfied for successful communication. Since according to Lemma 3.1, whether this constraint is satisfied does not depend on pmax , in fact, it only depends on the channel gains and the rates. Thus, the outage probability saturates at certain level as SNRmax increases. This figure also shows that as ρ increases, the NOT1 scheme attains a lower outage probability at asymptotically high SNRmax . This is because as ρ increases, the interference links become statistically weaker compared to the desired links and the outage probability decreases. We drive the outage probability at asymptotically high SNRmax to investigate how it depends on ρ. Corollary 3.3.1. For the network presented in (3.18), the outage probability of the NOT1 scheme at asymptotically high SNRmax is ( γ ln(γ) γ + (γ−1) 1 − γ−1 2 γ 6= 1 NOT1 . (3.34) Pout ((R1 , R2 ), ∞) = 0.5 γ=1 Proof. See Appendix 3.C. It is worth mentioning that, although Proposition 3.3.1 provides a lower bound on the outage probability, Corollary 3.3.1 gives the exact value of the outage probability at high SNRmax . This result shows that at this regime, the outage probability only depends on parameter γ, which is a function of R1 , R2 , and the channel parameter ρ. The ǫ-outage Achievable Rate Region We can use the lower bound in Proposition 3.3.1 to obtain an outer bound on NOT1 NOT1 CǫNOT1 (pmax ) denoted as Cǫ,out (pmax ) by solving Pout,l ((R1 , R2 ) , pmax ) = ǫ for (R1 , R2 ). This equation has different solutions for (R1 , R2 ); each of them denotes NOT1 one point on the boundary of Cǫ,out (pmax ). Similarly, we can characterize an inner NOT1 NOT1 (pmax ) denoted as Cǫ,in bound on Cǫ (pmax ) by solving NOT1 Pout,u ((R1 , R2 ) , pmax ) = ǫ (3.35) 3.3. Non-Orthogonal Transmission Schemes 45 11 NOT1 Cǫ,out , NOT1 Cǫ,in , NOT1 Cǫ,out , NOT1 Cǫ,in , NOT1 Cǫ,out , NOT1 Cǫ,in , CǫOT 10 9 R2 (bits/channel use) 8 7 ρ = 20 ρ = 20 ρ = 30 ρ = 30 ρ = 40 ρ = 40 dB dB dB dB dB dB 6 5 4 3 2 1 0 0 1 2 3 4 5 6 7 8 9 10 11 R1 (bits/channel use) Figure 3.4: The inner and the outer bounds on the ǫ-outage achievable rate region of the NOT1 scheme and ǫ-outage achievable rate region of the OT scheme (ǫ = 0.001, SNRmax = 50 dB). for (R1 , R2 ). Figure 3.4 shows the inner bounds and the outer bounds on the ǫoutage achievable rate regions of the NOT1 scheme for different values of ρ. We also plot the ǫ-outage achievable rate regions of the OT scheme for comparison. In this example, we set ǫ = 0.001, and SNRmax = 50 dB. This figure shows that for these parameters, if ρ is sufficiently large, then the achievable rate region of the NOT1 scheme is larger than that of the OT scheme. Although the outage probability of the NOT1 is small when the interference links are relatively weak compared to the desired links (i.e. when ρ is large), it is large when the interference links become relatively stronger than the desired links (i.e. when ρ is small). Therefore, for such channels, decoding the desired message by treating interference as noise may not be the best decoding strategy to apply. In the next section, we consider another scheme which is suitable for such scenarios. 3.3.2 Successive Interference Cancellation at Both Receivers If each receiver knows the codebooks of both transmitters, it can perform SIC by decoding the message of the interfering transmitter, removing the interference, and 46 Two-user Interference Networks: Point-to-Point Codes then decoding the message of the desired transmitter. We term this transmission scheme NOT2 . The receivers would be able to decode the messages of the interfering transmitters in a fading block, if the following conditions are satisfied: |h21 |2 p1 R1 ≤ log2 1 + , (3.36) |h22 |2 p2 + N0 |h12 |2 p2 R2 ≤ log2 1 + . (3.37) |h11 |2 p1 + N0 The equivalent power constraints are p1 p2 2R1 − 1 |h21 |2 ≥ 2R2 − 1 |h12 |2 ≥ |h22 |2 p2 + N0 , (3.38) |h11 |2 p1 + N0 . (3.39) The corresponding power constraint in matrix format is p DI FI p + nI , (3.40) where DI , " nI , 2R1 − 1 0 2R1 − 1 0 R2 2 −1 N0 |h21 |2 # , FI , 2R2 − 1 " |h22 |2 |h21 |2 0 |h11 |2 |h12 |2 N0 |h12 |2 T 0 . # , (3.41) The matrix DI depends on the rates and the matrix FI is a function of only the channel gains. After interference cancellation, each receiver can successfully decode its desired message if the following conditions are satisfied: |h11 |2 p1 R1 ≤ log2 1 + , (3.42) N0 |h22 |2 p2 R2 ≤ log2 1 + . (3.43) N0 Therefore, the transmission powers should also satisfy the following condition p nF , (3.44) where T nF , [nF,1 nF,2 ] = 2 R1 N0 −1 |h11 |2 2 R2 N0 −1 |h22 |2 T . (3.45) 3.3. Non-Orthogonal Transmission Schemes 47 Power Control Solution The positive element-wise minimum transmission power vector among those satisfying the constraints in (3.40) and (3.44) - if there is any such vector - can be found as described in the following. Let’s define pI , (I − DI FI )−1 nI where pI , [pI,1 pI,2 ]T . Depending on the channel gains and transmission rates, the minimum required powers can be found according to one of the four cases that will be mentioned in the following. Figure 3.5 shows four plots each corresponding to one possible case. The red region denotes powers which satisfy the power constraint in (3.3); the green region shows the powers which satisfy (3.40); the blue region illustrates the powers which satisfy the constraint in (3.44). The intersection of these regions shows powers that can provide successful transmission. There are four different possible cases to allow this to happen. In each case, we find the minimum required powers as follows. Case 1 If nF pI pmax , as shown in Figure 3.5 (a), the minimum power solution of the power control problem is pNOT2 = pI . (3.46) This solution is marked by a small circle in the figure. Case 2 If the following conditions are satisfied nF,2 pI,1 ≤ nF,1 ≤ pmax,1 , ≤ 2R1 2R2 − 1 |hN120|2 ≤ pmax,2 , (3.47) as shown in Figure 3.5 (b), the minimum required powers are N0 |h11 |2 N0 . −1 |h12 |2 2 pNOT 1 = nF,1 = 2R1 − 1 (3.48) 2 pNOT 2 = 2R1 2R2 (3.49) The solution is shown by a small circle in the figure. Case 3 If nF,1 ≤ 2R2 2R1 − 1 N0 |h21 |2 ≤ pmax,1 pI,2 ≤ nF,2 ≤ pmax,2 , (3.50) as shown in Figure 3.5 (c), the minimum required powers are N0 |h21 |2 N0 . −1 |h22 |2 2 pNOT 1 = 2R2 2R1 − 1 (3.51) 2 pNOT 2 = nF,2 = 2R2 (3.52) 48 Two-user Interference Networks: Point-to-Point Codes p2 p2 pmax,2 pmax,2 pI,2 pI,2 nF,2 nF,2 nF,1 pI,1 pmax,1 p1 pI,1 nF,1 (a) Case 1 pmax,1 p1 (b) Case 2 p2 p2 pmax,2 pmax,2 nF,2 nF,2 pI,2 pI,2 nF,1 pI,1 pmax,1 p1 (c) Case 3 pI,1 nF,1 pmax,1 p1 (d) Case 4 Figure 3.5: Solution of the power control problem for the NOT2 scheme. In these plots, the red region denotes powers satisfying the maximum power constraint in (3.3); the green region shows the powers satisfying the constraint in (3.40); the blue region illustrates the powers which satisfy the constraint in (3.44). The intersection of these regions shows powers which can provide successful transmission. The small circle in each plot indicates the minimum required powers in each case. This solution is shown in the figure by a small circle. Case 4 If the following conditions satisfy pI,1 ≤ nF,1 ≤ pmax,1 2 11 | R2 −1 +1 ≤ nF,2 2R1−1 |hN120|2 |h |h22 |2 2 3.3. Non-Orthogonal Transmission Schemes nF,2 ≤ N0 1 |h11 |2 49 |h21 |2 1 − |h21 2 | |h22 |2 , (3.53) as shown in Figure 3.5 (d), the minimum required powers are N0 |h11 |2 N0 . −1 |h22 |2 2 pNOT 1 = 2R1 − 1 (3.54) 2 pNOT 2 = 2R2 (3.55) This solution is marked in the figure by a small circle. Outage Probability Analysis The set of feasible solutions of the power control problem for the NOT2 scheme is NOT2 PH ((R1 , R2 ) , pmax ) = {p : p DI FI p + nI , nF p pmax } . (3.56) This set is illustrated in Figure 3.5 as the intersection of the green region, the red region, and the blue region. The red region in this figure denotes powers which satisfy the constraint in (3.3); the green region shows the powers satisfying (3.40); the blue region illustrates the powers which satisfy the constraint in (3.44). Therefore, if these three regions have no common intersection region, then NOT2 PH ((R1 , R2 ) , pmax ) = ∅ and an outage event happens. In the following corollary, we provide a lower bound on the outage probability of the NOT2 scheme. Proposition 3.3.2. For the network described in (3.18), the outage probability of the NOT2 scheme is bounded as follows: n o NOT2 NOT2 NOT2 NOT2 1 − min PF,D,1 , PF,D,2 , PF,I ≤ Pout ((R1 , R2 ) , pmax ) . (3.57) In this equation a (1−γ ′ )−γ ′ρb k k ak γ ′ ρβk′ σS4 − ρbkρ+ak βk′ ak σS4 ργ ′ e E e + 1 ′ ′ ′ (1−γ ) 1−γ ργ a (1−γ ′ ) ′ ′ 4 −bk k β ργ σS e ak ak + k − E E e ργ ′ 1 1 ′ 2 ′ (1 − γ ) ρ ργ R R 2 1 (2 −1) + (2 −1) N − 0 NOT2 PF,D,k =− NOT2 PF,I =e σ2 S pmax,1 pmax,2 k ∈ {1, 2}, (3.58) where γ ′ = 1/ ρ2 2R1 −1 2R2 −1 , ak = N0 / pmax,k σS2 , bk = N0 2Rk − 1 / pmax,k σS2 , 1 (ak −ρbk )/ρ βk′ = e . ρσS4 (3.59) 50 Two-user Interference Networks: Point-to-Point Codes 100 rSbC rSbC rSbC rSbC rSbC bCrS 10−1 bCrSbC bC rSbC bC 10−2 bC bC bCbC bC A Pout rS bCbC bC bC bC rS 10−3 rS rS bC bC rS 10−4 rS 10−5 0 10 NOT2 Pout , NOT2 Pout,l , NOT2 Pout , NOT2 Pout,l , OT Pout 20 ρ = −30 ρ = −30 ρ = −40 ρ = −40 30 dB dB dB dB rS rS rS rS 40 50 rS rS 60 rS 70 SNRmax (dB) A Figure 3.6: Pout versus SNRmax , R1 = R2 = 5 (bits/channel use). Proof. See Appendix 3.D. NOT2 Figure 3.6 plots the lower bound on Pout versus SNRmax for different values NOT2 OT of ρ. Also, Pout and the simulation results of the exact values of Pout are shown for comparison. This shows that in a certain range of SNRmax , NOT2 outperforms the OT scheme. We can also see that the outage probability of NOT2 saturates at high SNRmax . Thus, if SNRmax is larger than a certain value, OT attains a lower outage probability and outperforms NOT2 . This is similar to the behaviour of the NOT1 scheme. The reason is that, at high SNRmax , the dominant factor leading to an outage event is that there is no positive power vector to satisfy (3.40) rather than the violation of the power constraints. However, in contrast to the NOT1 scheme, the performance of the NOT2 scheme improves as ρ decreases. The reason is that for the smaller values of ρ, the interference links are most likely stronger than the desired links. Thus, it becomes easier to decode the message of the interfering transmitter and remove the interference. Even at high SNRmax , the outage event may occur due to the fact that no positive power vector may exist to satisfy (3.40). The following corollary characterizes the outage probability of the NOT2 scheme at asymptotically high SNRmax . 3.3. Non-Orthogonal Transmission Schemes 51 Corollary 3.3.2. For the network presented in (3.18), the outage probability of the NOT2 scheme with asymptotically high SNRmax is ( ′ ′ ′ ) ′ 1 − γ ′γ−1 + γ(γ ′ln(γ NOT2 −1)2 γ 6= 1 Pout ((R1 , R2 ), ∞) = , (3.60) 0.5 γ′ = 1 where γ ′ , 1/ ρ2 2R1 − 1 2R2 − 1 and ρ is given in (3.1). Proof. The proof is similar to that of Corollary 3.3.1. This result confirms our observation in Figure 3.6, regarding the saturation of the feasibility probability at high SNRmax . The ǫ-outage Achievable Rate Region NOT2 We can characterize an outer bound on CǫNOT2 (pmax ) denoted as Cǫ,out (pmax ) using the lower bound given in Proposition 3.3.2 by solving NOT2 Pout,l ((R1 , R2 ) , pmax ) = ǫ, (3.61) for (R1 , R2 ). Each solution of this equation for (R1 , R2 ) denotes one point on the NOT2 boundary of Cǫ,out (pmax ). The outer bound on CǫNOT2 (pmax ), for different values of ρ, and the ǫ-outage achievable rate region of the OT scheme are shown in Figure 3.7. In this particular example, we set ǫ = 0.001 and SNRmax = 50 dB. It can be seen that, when ρ has a small value, NOT2 achieves a larger rate region compared to OT. Indeed, non-orthogonal transmission and SIC at both receivers is beneficial in this case. If only one of the receivers observes strong interference, SIC at both receivers may not be the best decoding strategy. Instead, SIC can be employed at the receiver which observes strong interference and direct decoding by treating the interference as noise can be performed at the other receiver to achieve a better performance. We investigate this scheme in more detail in the next part. 3.3.3 Successive Interference Cancellation at One Receiver The receivers may implement different decoding strategies. For instance, the first receiver can perform SIC, while the second receiver decodes its message directly by treating the interference as noise. We refer to this scheme as NOT3 . The following conditions should be satisfied for a successful transmission in a fading block: |h12 |2 p2 , (3.62) R2 ≤ log 1 + |h11 |2 p1 + N0 |h11 |2 p1 , (3.63) R1 ≤ log 1 + N0 |h22 |2 p2 R2 ≤ log 1 + . (3.64) |h21 |2 p1 + N0 52 Two-user Interference Networks: Point-to-Point Codes 10 NOT2 Cǫ,out , ρ = −20 dB NOT2 Cǫ,out , ρ = −30 dB NOT2 Cǫ,out , ρ = −40 dB OT Cǫ 9 R2 (bits/channel use) 8 7 6 5 4 3 2 1 0 0 1 2 3 4 5 6 7 8 9 10 R1 (bits/channel use) Figure 3.7: The outer bounds on the ǫ-outage achievable rate region (ǫ = 0.001, SNRmax = 50 dB). Therefore, the powers should satisfy p nR , (3.65) where nR = 2R2−1 N0 max N0 |h11 |2 2 2 R1 2R1 |h11 | +|h21 | (2 −1) , |h12 |2 |h22 |2 |h11 |2 2R1−1 . (3.66) Power Control Solution The minimum powers which satisfy the constraint in (3.65) are 3 pNOT 1 = 3 pNOT 2 = N0 , |h11 |2 ( ) N0 2R1 N0 |h11 |2+|h21 |2 2R1−1 R2 . , 2 −1 max |h12 |2 |h22 |2 |h11 |2 2R1−1 (3.67) (3.68) 3.3. Non-Orthogonal Transmission Schemes 53 Based on the same argument, we can show that the minimum required powers for the successful transmission of the NOT4 scheme (performing SIC at the second receiver, while the first receiver directly decodes its message by treating the interference as noise) are ( ) N0 2R2 N0 |h22 |2+|h12 |2 2R2−1 NOT4 R1 , p1 = 2 −1 max |h21 |2 |h11 |2 |h22 |2 4 pNOT = 2R2−1 2 N0 . |h22 |2 (3.69) However, for some channel realizations, the calculated powers in (3.67) and (3.68) for NOT3 or those in (3.69) and (3.69) for NOT4 may violate the power constraint in (3.3); thus, an outage event may happen. In the next part, we investigate the outage probability in more details. Outage Probability Analysis The set of the feasible solutions of the power control problem for the NOT3 scheme is NOT3 PH ((R1 , R2 ) , pmax ) , {p : nR p pmax } . (3.70) This set is illustrated in Figure 3.8 as the intersection region of the red and the blue regions. The red region shows powers which satisfy the maximum power constraint in (3.3) and the blue region denotes the powers which satisfy the constraint in (3.65). If the red and the blue regions do not overlap, then an outage event occurs. The outage probability is characterized as follows. Proposition 3.3.3. For the network presented in (3.18), the outage probability of the NOT3 scheme is bounded as follows o o n n NOT3 NOT3 NOT3 NOT3 NOT3 . ≤ Pout ((R1 , R2 ) , pmax ) ≤ min 1,2−PF,1 −PF,2 1−min PF,1 ,PF,2 (3.71) In these equations e−(2 −1)N0 /(pmax,1 σS ) , c1 b2 c1ρb2 c1 b 2 NOT3 PF,2 = e−b2 − e−(c1 +1)b2 ρ , e E1 ρ ρ where c1 = 2R1 − 1 and b2 = N0 2R2 − 1 /pmax,2 σS2 . NOT3 PF,1 = Proof. See Appendix 3.E. R1 2 (3.72) (3.73) 54 Two-user Interference Networks: Point-to-Point Codes p2 pmax,2 nR,2 nR,1 pmax,1 p1 Figure 3.8: Solution of the power control problem for the NOT3 scheme. Similarly, we can determine upper and lower bounds on the outage probability of the NOT4 scheme as mentioned in the following proposition. Proposition 3.3.4. For the network presented in (3.18), the outage probability of the NOT4 scheme is bounded as n o n o NOT4 NOT4 NOT4 NOT4 NOT4 1−min PF,1 , PF,2 ≤ Pout ((R1 , R2 ),pmax ) ≤ min 1,2−PF,1 −PF,2 . (3.74) In these equations, c2 b 1 c2 b1 c2ρb1 NOT4 e−(c2 +1)b1 ρ , e E1 PF,1 = e−b1 − ρ ρ R2 2 P NOT4 = e−(2 −1)N0 /(pmax,2 σS ) , F,2 where c2 = 2R2 − 1, and b1 = N0 2R1 − 1 /pmax,1 σS2 . (3.75) (3.76) NOT3 Figure 3.9 shows the lower bound and the upper bound on Pout versus SNRmax for different values of ρ. The figure shows that in this specific setting the OT scheme outperforms NOT3 . To obtain a better evaluation of this scheme, the performance of NOT3 and OT have been shown for a sample asymmetric network in Figure 3.10. In the considered network, the users have different rates (R1 = 7, R2 = 2). It can be seen that in this case NOT3 outperforms OT. 3.3. Non-Orthogonal Transmission Schemes 100 uTbCrS uTbCrS uTbCrS uTbCrS uTbCrS uTbCrS uTbC 55 uTbC uT rS uT bC uT rS bC rS uT rS 10−1 bC uT rS bC rS 10−2 A Pout rS rS bC 10−3 bC bC uT uT 10−4 uT 10−5 0 10 NOT3 Pout,l , NOT3 Pout , NOT3 Pout,u , NOT3 Pout,l , NOT3 Pout , NOT3 Pout,u , NOT3 Pout,l , NOT3 Pout , NOT3 Pout,u , OT Pout 20 uT rS ρ = 0 dB ρ = 0 dB ρ = 0 dB ρ = 10 dB ρ = 10 dB ρ = 10 dB ρ = 20 dB ρ = 20 dB ρ = 20 dB 30 bCbC uT rS bC rS bC rS rS 40 50 60 70 SNRmax A Figure 3.9: Pout versus SNRmax , R1 = R2 = 5 (bits/channel use). The ǫ-outage Achievable Rate Region Similar to the approach that we have considered for the other schemes, the lower bound on the outage probability in Proposition 3.3.3 can be used to obtain an NOT3 outer bound on CǫNOT3 (pmax ), denoted as Cǫ,out (pmax ), by solving the equation NOT3 Pout,l ((R1 , R2 ) , pmax ) = ǫ for (R1 , R2 ). This equation may have many solutions for (R1 , R2 ); each of them denotes one point on the boundary of the outer bound NOT3 region. Similarly, an inner bound on CǫNOT3 (pmax ), denoted as Cǫ,in (pmax ), can NOT3 be found by solving Pout,u ((R1 , R2 ) , pmax ) = ǫ for (R1 , R2 ). Also, the results of Proposition 3.3.4 can be used to find an inner bound and an outer bound on CǫNOT4 (pmax ). Figure 3.11 shows the inner bound and the outer bound on the ǫ-outage achievable rate regions of the NOT3 scheme and the ǫ-outage achievable rate region of the OT scheme when ρ = 0 dB. In this example, we set ǫ = 0.001 and SNRmax = 50 dB. The figure shows that, the achievable rate region of none of these schemes is strictly larger that the other. In fact, each of them can achieve certain rates that may not be achievable by the other scheme. Similar observations can be obtained for the NOT4 scheme. 56 Two-user Interference Networks: Point-to-Point Codes 100 rS rS rS rS rS rS rS rS 10−1 rS A Pout rS 10−2 rS rS rS 10−3 rS 10−4 0 10 rS OT Pout NOT3 Pout,u , ρ = 0 dB 20 30 40 50 60 70 SNRmax A Figure 3.10: Pout versus SNRmax , R1 = 7, R2 = 2 (bits/channel use). 3.4 Summary In this chapter, we have studied transmission schemes for fixed-rate transmission over a two-user Rayleigh block-fading interference network. The proposed schemes use a point-to-point encoder at each transmitter in concatenation with a power controller. The power controllers adjust transmission powers during each fading block, considering a short-term individual power constraint, to successfully transmit at desired rates. We have considered orthogonal and non-orthogonal transmission schemes. In the latter case, we have investigated different decoding schemes in which each receiver either directly decodes its desired message by treating the interference as noise, or performs SIC. For each of these schemes, we have found the solution of the power control problem to assign the minimum required power to each transmitter. Since the power control problem for some channel realizations may not have any feasible solution, transmission may be in outage. We have computed lower bounds and upper bounds on the probability of such events for these schemes. Using these results, we have characterized an inner bound and an outer bound on the ǫ-outage achievable rate region for each transmission scheme. 3.A. The Proof of Proposition 3.2.1 57 8 CǫOT NOT3 Cǫ,out NOT3 Cǫ,in 7 R2 (bits/channel use) 6 5 4 3 2 1 0 0 1 2 3 4 5 6 7 8 R1 (bits/channel use) Figure 3.11: The inner bound and the outer bound on the ǫ-outage achievable rate region. (ǫ = 0.001, ρ = 0 dB, SNRmax= 50 dB). 3.A The Proof of Proposition 3.2.1 The outage probability of the orthogonal transmission scheme can be found as follows OT OT Pout ((R1 , R2 ), pmax ) = 1−Pr PH ((R1 , R2 ), pmax ) 6= ∅ δN 0 < pmax,1 , = 1 − Pr 2R1 /δ − 1 |h11 |2 (1 − δ)N 0 R2 /(1−δ) 2 −1 < pmax,2 |h22 |2 δN (a) 0 2 R1 /δ < |h11 | = 1 − Pr 2 −1 pmax,1 (1 − δ)N 0 R2 /(1−δ) 2 ×Pr 2 −1 < |h22 | pmax,2 N (1−δ) − 20 (2R1 /δ −1) p δ +(2R2 /(1−δ) −1) p (b) σ max,1 max,2 = 1−e S (3.77) 58 Two-user Interference Networks: Point-to-Point Codes where (a) follows the independence of the channel gains, and (b) follows the fact that Xk = |hkk |2 (k ∈ {1, 2}) has an exponential distribution with a cumulative distribution function (cdf) as follows: 2 FXk (x) = 1 − e−x/σS , x > 0. 3.B (3.78) The proof of Proposition 3.3.1 NOT1 The probability Pout ((R1 , R2 ) , pmax ) is n o NOT1 NOT1 Pout ((R1 , R2 ),pmax ) = 1−Pr PH ((R1 , R2 ),pmax ) 6= ∅ = 1−Pr λmax (DS FS ) < 1, pNOT1≺pmax , (a) (3.79) 1 1 where pNOT and pNOT are the minimum required transmission powers for the 1 2 NOT1 scheme calculated in (3.26) and (3.27), and λmax (DS FS ) is the largest magnitude of the eigenvalues of matrix DS FS given in (3.24). The equality (a) follows NOT1 the definition in (3.28) and Lemma 3.1. Defining the probabilities PF,1 and NOT2 PF,1 as o n NOT1 1 PF,1 , Pr λmax (DS FS ) < 1, pNOT < pmax,1 1 n o NOT1 1 PF,2 , Pr λmax (DS FS ) < 1, pNOT < p , max,2 2 (3.80) (3.81) we can bound the feasibility probability as follows (a) NOT1 NOT1 PF,1 +PF,2 −1 ≤ Pr λmax (DS FS )<1, pNOT1 ≺ pmax n o (b) NOT1 NOT1 ≤ min PF,1 , PF,2 , (3.82) n T o T where the equality (a) follows the fact that Pr {A B} = 1 − Pr A B = 1 − S Pr A B ≥ 1 − Pr A − Pr B = Pr {A} + Pr T {B} − 1, where A and B are random events; and the equality (b) follows Pr {A B} ≤ Pr {A}. The probability NOT1 PF,1 can be further simplified as follows NOT1 PF,1 21 |2 |h |2 |h |2 γ N0 2R1−1 |h +l 2 |h11 | 12 21 , < , <p = Pr max,1 |h11 |2 |h22 |2 ρ2 |h21 |2 (1 − l) (3.83) 3.B. The proof of Proposition 3.3.1 59 2 2 2 2 |h12 | |h21 | ρ ρ where γ = (2R1 −1)(2 R2 −1) , and l = |h |2 |h |2 × γ . After some mathematical ma11 22 nipulations we have ) ( 2 2 |h22 | 2 2 NOT1 2 2 ρ PF,1 = Pr 2 |h11 | −b1 σS > |h12 | +a1 σS γ |h21 | |h21 |2 |h12 |2+a1 σS2 o n γ < 2 = Pr b1 σS2 < |h11 |2 ×Pr 0< |h |2 |h |2−b σ 2 ρ 11 1 S 22 = where Q = Pr 0 < e−b1 × Q (3.84) 2 |h21 |2 (|h12 |2 +a1 σS ) 2 |h22 |2 (|h11 |2 −b1 σS ) < γ ρ2 , a1 = N0 , 2p σS max,1 and b1 = N0 (2R1 −1) . 2p σS max,1 To calculate Q, we derive the probability density function (pdf) of the random |h21 |2 (|h12 |2 +a1 σ2 ) variable M , |h |2 |h |2 −b σ2S . Let Xkl , |hkl |2 and Ykk , |hkk |2 (∀k, l ∈ {1, 2}, 1 S) 22 ( 11 k 6= l), we have 1 −x/σI2 x ≥ 0, e σI2 2 1 fYkk (y) = 2 e−y/σS y ≥ 0. σS fXkl (x) = (3.85) (3.86) Let |h21 |2 X21 = , |h22 |2 Y22 X12 + a1 σS2 |h12 |2 + a1 σS2 = , , 2 2 |h11 | − b1 σS Y11 − b1 σS2 Z21 , Z12 using the fact that the pdf of a random variable Z = we can show that the pdf of Z21 and Z12 are X Y R +∞ is fZ (z)= −∞ |y|fXY (zy, y)dy, ρ , z ∈ R, (1+ρz)2 1+ρz β1 σS4 β1 a1 σS4 e−a1( z ) , z ∈ R, fZ12 (z)= + (1+ρz)z (1+ρz)2 fZ21 (z)= where ρ = 2 σS σI2 and β1 = ρ (a1 ρ−b1 ) . 4e σS (3.87) (3.88) (3.89) The pdf of the random variable 2 |h12 | + a1 σS2 = Z12 Z21 M, |h22 |2 |h11 |2 − b1 σS2 2 |h21 | (3.90) 60 Two-user Interference Networks: Point-to-Point Codes R +∞ 1 fZ12 Z21 (t, m cumulative districan be derived as fM (m) = −∞ |t| t )dt. Since the Rm bution function (cdf) of the random variable M is FM (m) = x=−∞ fM (x)dx, we have Z m Z ∞ β1 σS4 ρ(a1 t2+(1+a1ρ)t) −(a1 ρ+a1 t) FM (m)= e dtdx 2 2 x=−∞ t=0 (1 + ρxt) (ρ + t) 1 −ρa1 β1 σS4 =− − a E (a ρ) e 1 1 1 1 − ρ2 m ρ a1 (1−ρ2 m) β1 ρmσS4 β1 a1 σS4 a1 ρm . − e × E (a ρ) − E − 1 1 1 1 − ρ2 m (1 − ρ2 m)2 ρm (3.91) Therefore, Q = FM (γ/ρ2 ) − FM (0) a1 ρ(1−γ) a1 ρ γβ1 σS4 −ρa1 β1 a1 σS4 e γ + e E1 =− ρ(1 − γ) 1−γ γ 4 a1 ρ(1−γ) a1 ρ β1 γσS γ e E1 (a1 ρ)−E1 + ρ(1 − γ)2 γ (3.92) NOT1 NOT1 By plugging (3.92) in (3.84), we can find PF,1 . It is possible to find PF,2 with similar approach. By plugging these in (3.82), the lower bound and the upper NOT1 bound on Pout ((R1 , R2 ) , pmax ) can be obtained. This completes the proof. 3.C The proof of Corollary 3.3.1 We have (a) NOT1 (b) NOT1 lim P NOT1 = 1− lim PF,2 =1− lim PF,1 pmax,1 ,pmax,2 →∞ out pmax,2→∞ pmax,1 →∞ NOT1 = 1− lim PF,1 = 1−(A+B +C), a1,b1 →0 (3.93) NOT1 where (a) and (b) follows the equality in (3.79), and the definitions of PF,1 and NOT1 PF,2 in (3.80) and (3.81), respectively. The values A, B and C can be calculated as follows γ γβk σS4 −(bk +ρak ) e =− ak ,bk →0 ρ(1 − γ) 1−γ ak ρ(1−γ)−γbk (a) βk ak σS4 ak ρ γ B = lim e = 0 E1 ak ,bk →0 1 − γ γ a1 ρ a1 ρ(1−γ) βk γσS4 e−bk γ e E1 (a1 ρ)−E1 C = lim a1 ,b1 →0 ρ(1 − γ)2 γ A = lim − (3.94) (3.95) 3.D. The Proof of Proposition 3.3.2 61 a1 ρ γ (1+a1q+· · · ) lim E1 (a1 ρ) − E1 = (1 − γ)2 a1 →0 γ (c) γ a1 ρ = lim E1 (a1 ρ) − E1 (1 − γ)2 a1 →0 γ Z ∞ −t Z ∞ −t ! e γ e (d) lim dt − dt = a1 ρ (1 − γ)2 a1 →0 t t a1 ρ γ Z aγ1 ρ −t γ e = lim dt (1 − γ)2 a1 →0 a1 ρ t Z aγ1 ρ (1 − t + t2 /2 − · · · ) γ (e) lim dt = 2 (1 − γ) a1 →0 a1 ρ t γ =− ln γ (1 − γ)2 (b) (3.96) where q = (1 − γ)ρ/γ. The equalities (a) and (c) follow the fact that limx→∞ x · E1 (x) = 0; (b) follows the Taylor series expansion of ea1 ρ(1−γ)/γ ; (d) follows the definition of E1 (x); and (e) follows the Taylor series expansion of e−t . Therefore, lim pmax,1 ,pmax,2 →∞ 3.D NOT1 Pout =1− γ γ ln γ. + γ − 1 (γ − 1)2 (3.97) The Proof of Proposition 3.3.2 We have NOT2 Pout ((R1 , R2 ),pmax ) ≥ 1 − Pr{λmax (DI FI ) < 1, pI pmax ,nF pmax } . (3.98) Defining PF,I , PF,D,k , Pr{nF pmax } Pr{λmax (DI FI ) < 1, pI,k ≤ pmax,k } k ∈ {1, 2}, (3.99) (3.100) we can show that Pr {λmax (DI FI ) < 1, pI pmax , nF pmax } (a) ≤ min {PF,I , PF,D,1 , PF,D,2 } . T T The inequality (a) follows Pr {A B C} ≤ Pr {A}. We have PF,I =Pr {nF pmax } N0 N0 R2 −1 < p , 2 < p =Pr 2R1 −1 max,1 max,2 |h11 |2 |h22 |2 ( ) ( ) N0 2R1 −1 N0 2R2 −1 2 2 =Pr |h11 | < ×Pr |h22 | < pmax,1 pmax,2 (3.101) 62 Two-user Interference Networks: Point-to-Point Codes =e =e −(2R1 −1) − N0 σ2 S N0 σ2 pmax,1 S ×e −(2R2 −1) (2R1 −1) + (2R2 −1) pmax,1 pmax,2 N0 σ2 pmax,2 S (3.102) and PF,D,1 = Pr {λmax (DI FI ) < 1, pI,1 ≤ pmax,1 } ( |h11 |2 |h22 |2 = Pr 2R1 − 1 2R2 − 1 < 1, |h12 |2 |h21 |2 |h22 |2 N0 R2 R1 − 1 1 + 2 − 1 2 |h21 |2 |h12 |2 < p max,1 2 2 11 | |h22 | 1 − (2R1 − 1) (2R2 − 1) |h |h12 |2 |h21 |2 ) ( |h21 |2 − b1 σS2 |h12 |2 2 2 2 = Pr a1 σS + |h11 | |h22 | < (2R1 − 1) (2R2 − 1) ( ) |h22 |2 |h11 |2 +a1 σS2 2 2 ′ = Pr b1 < |h21 | ×Pr 0 < <ρ γ , |h12 |2 (|h21 |2 −b1 σS2 ) (3.103) where a1 = N0 / σS2 pmax,1 , and b1 = N0 2R1−1 /σS2 pmax,1 . Comparing the equality in (3.103) and (3.84), we can obtain (3.58) by substituting σS2 , 1/ρ, and σI2 , instead of σI2 , σS2 , and ρ in (3.30), respectively. 3.E The Proof of Proposition 3.3.3 NOT3 The probability Pout ((R1 , R2 ) , pmax ) can be simplified as n o NOT3 NOT3 Pout ((R1 , R2 ) , pmax )= 1−Pr PH ((R1 , R2 ) , pmax ) 6= ∅ o n NOT3 3 = 1−Pr pNOT <p , p <p max,1 2 max,2 . 1 Defining the probabilities we have NOT3 PF,1 NOT3 PF,1 , NOT3 PF,2 , (3.104) NOT3 PF,2 and as follows n o 3 Pr pNOT < p max,1 1 o n 3 Pr pNOT < pmax,2 , 2 o n (a) NOT3 NOT3 NOT3 3 2 − PF,1 − PF,2 ≤ Pr pNOT < p , p < p max,1 max,2 1 2 o n (b) NOT3 NOT3 , ≤ min PF,1 , PF,2 (3.105) (3.106) (3.107) 3.E. The Proof of Proposition 3.3.3 63 where the equality (a) follows the fact that \ n \ o n [ o Pr A B = 1−Pr A B = 1 − Pr A B ≥ 1−Pr A −Pr B = Pr {A}+Pr {B}−1, where A and B are random events; and the equality (b) follows Pr {A We have o n NOT3 3 PF,1 = Pr pNOT < pmax,1 1 N0 2 R1 2 −1 = Pr |h11 | > pmax,1 N0 R1 − −1) 2 (2 = e pmax,1 σS , T (3.108) B} ≤ Pr {A}. (3.109) and NOT3 PF,2 = Pr 0 < M < 1/ bσS2 (3.110) o n R N0 1 |h21 |2 1 2R1 − 1 + 1 , and b = 2R2 − 1 pmax,2 where M , max |h212 |2 , |h22 2. σS |2 |h11 |2 Therefore, we require to find the cdf of random variable M . We define ! 2 |h21 | X , c+1 , |h11 |2 1 Y , (3.111) 2, |h22 | where c = 2R1 − 1 . It can be shown that fX (x) = fY (y) = ρc x ≥ 1, (c − aρ + ρx)2 1 − yσ12 S y ≥ 0. e 2 y σS2 (3.112) (3.113) We define Z , XY , and we have Z ∞ Z ∞ z − x ρcx 1 σ2 z S dx. dx = e fX (x)fY (3.114) fZ (z) = 2z2 2 (c − ρ + ρx) |x| x σ S 1 −∞ The cdf of Z is FZ (z) = Z z t=−∞ fZ (t)dt = e − 1 σ2 z S c c 2 − 2 e ρσS z E1 σS zρ c ρσS2 z . (3.115) 64 Two-user Interference Networks: Point-to-Point Codes Now, defining W , c+1 , |h12 |2 we have FW (w) = e − c+1 2 σ w I w > 0. (3.116) Therefore, for the random variable M = max{Z, W }, since Z and W are independent, we have FM (m) = FZ (m)FW (m). (3.117) Then, NOT3 PF,2 = Pr 0 < M < 1/ bσS2 = FM 1/ bσS2 − FM (0) bc bc bc −b ρ e−(c+1)bρ . = e − e E1 ρ ρ (3.118) Chapter 4 K-user SISO Interference Networks: Pilot-assisted Interference Alignment I N the previous chapter we studied transmission schemes for the smallest (twouser) interference networks. We have assumed that global CSI is a priori available at terminals (sources and destinations). These transmission schemes may not be effective for larger networks; furthermore, usually no a priori CSI is available at terminals. Conducting coordinated transmission in such networks when no a priori CSI is available at terminals is a difficult task. In this chapter, we study K-user (K > 2) SISO interference network with time-varying channels and no a priori CSI available at terminals. We propose the pilot-assisted ergodic interference alignment (PAEIA) transmission scheme. The proposed scheme consists of three phases: pilot transmission, feedback transmission, and data transmission phase. We consider orthogonal pilot transmission and minimum mean square error (MMSE) channel estimation in the pilot transmission phase. We study analog and digital feedback schemes for channel state information feedback. In the data transmission phase, an ergodic interference alignment (EIA) is applied. In addition, transmitters perform rate adaptation or power control based on the feedback signal to adapt the transmission strategy according to the channel state. We evaluate the performance limits of the proposed scheme, and investigate radio resource allocation problems. We first consider networks with analog feedback, and derive an achievable rate region. Moreover, we find the optimum power allocation for pilot transmission and data transmission. To gain insight into the performance of the system at high-SNR regime, we derive the achievable DoF region. Next, we study networks with digital feedback. We investigate two problems: first, we study a power control problem, and we propose a power control scheme that adapts transmission powers such that the mutual information corresponding to each source-destination pair is always larger than the transmission rate, and thus, transmitted codewords can be successfully decoded at the desired destination. Next, we study a throughput maximization problem for communication systems in which the transmission powers are fixed and cannot be adjusted. We propose a rate adaptation scheme to maximize network 65 66 K-user SISO Interference Networks: Pilot-assisted Interference Alignment KTτ = αT Td = (1 − α)T Tτ z1 h11 S1 + D1 h2 h K1 1 z2 2 S2 h1 h22 + D2 hK h 2K h1 K 2 SK hKK zK + DK Figure 4.1: Transmitted symbols within one fading block in a K-user interference channel. The crosshatched red slot, the plain green slot, and the blue angle lined slots denote no transmission, pilot symbols, and data symbols, respectively. throughput. This chapter is organized as follows. Section 4.1 describes the considered multi-user SISO interference network. We present the pilot-assisted ergodic interference alignment scheme in Section 4.2. Section 4.3 presents the achievable performance when analog feedback is deployed. In this section, an achievable rate region is presented, and the achievable degrees of freedom region is computed. We also compute the optimum power allocation between channel training phase and data transmission phase. For networks with digital feedbacks, we present power control algorithm and rate adaptation scheme in Section 4.4. Finally, we summarize this chapter in Section 4.5. 4.1 Multi-user SISO Interference Network We consider an interference network composed of K single-antenna source–destination pairs, as shown in Figure 4.1. The sources and the destinations are denoted by Sl and Dk (k, l ∈ {1, 2, ..., K}), respectively. The channels are time-varying and the channel gain from Sl to Dk at time t is denoted as htkl . The channels follow block fading model in which the channel gains are constant over one fading block. nT +i At fading block n, we have hnT (i = 1, ..., T − 1), where T is channel cokl = hkl herence time. The channel gains are ergodic time-varying and have independent and identical distribution across different fading blocks. The channel gains are independently drawn from a zero-mean unit-variance complex Gaussian distribution, i.e. hnT kl ∼ CN (0, 1). 4.2. Pilot-assisted Ergodic Interference Alignment 4.2 67 Pilot-assisted Ergodic Interference Alignment As shown in Figure 4.1, transmission is performed in three phases: pilot transmission, feedback transmission and data transmission phase. The pilot transmission phase and the data transmission phase are conducted within each fading block and have the duration of αT and (1 − α)T , respectively. The channel sharing factor α (K/T ≤ α ≤ 1) is a design parameter. A frequency-division duplex (FDD) transmission is considered for pilot and data transmission from sources to destinations and feedback transmission form destinations to sources. 4.2.1 Pilot Transmission Phase Channel training is performed in an orthogonal fashion in which the training period is divided into K equal time slots (each has the duration of Tτ = αT /K), as shown in Figure 4.1. Each destination estimates the gain of the corresponding direct link and interference links. Let source Sl (l ∈ {1, ..., K}), at fading block n, sends Tτ known pilot symbols with power Pτ as follows p i Xτ,l = Pτ i = nT + (l − 1)Tτ + 1, ..., nT + lTτ , then the received signals at Dk is i Yτ,k = p i Pτ hnT kl + Zk . (4.1) The MMSE estimate of the channel gain between Sl and Dk is obtained as follows ˜ nT = h kl Pτ N0 + Tτ Pτ nTX +lTτ i Yτ,k . (4.2) i=nT +(l−1)Tτ +1 The following equation holds nT ˜ nT hnT kl = hkl + εkl , (4.3) nT ˜ nT where εnT kl is the channel estimation error. The random variables hkl and εkl 2 are independent zero mean Gaussian distributed with variances 1 − σε and σε2 , respectively, where σε2 = 1 . 1 + Tτ Pτ /N0 (4.4) It is clear that the variance of channel estimation error decays by allocating more time or power to channel training. 4.2.2 Feedback Transmission Phase As we have discussed in the previous part, each destination can acquire a noisy estimate of the local CSI through a pilot-based channel training scheme. The destinations then can send the estimated CSI to the other terminals via channel state 68 K-user SISO Interference Networks: Pilot-assisted Interference Alignment feedbacks. They can transmit either un-quantized CSI (analog feedback) or quantized CSI (digital feedback) via feedback channels. The feedback signal correspond˜ t can be denoted as follows ing to the channel gain h kl ˆ t , f (h ˜ t ), h kl kl (4.5) where the function f (.) is the feedback signal transmit function. We consider orthogonal feedback channels and there is no interference in the transmission of feedback signals. In the scenarios that analog feedback is applied, we assume that f (x) = x, i.e. destinations send the estimated channel gains to the other terminals. In the scenarios that digital feedback is deployed, the estimated channel gains are first quantized, and then the index of the quantized value is sent over the feedback channel. To be able to analytically investigate the problem and gain insight on the performance of interference networks with limited feedback, we consider a uniform quantization scheme in this chapter. It can be conjectured that using more sophisticated quantization schemes (such as the quantizers mentioned in [GN98]) may lead to an even better performance. We deploy a two-dimensional vector quantizer to quantize each complex-valued channel gain. The channel gains are unbounded, but the number of quantization regions is limited. To resolve this problem, only channel gains within a bounded region will be quantized. Specifically, we only quantize those channel gains for which the corresponding channel matrix Ht (the element on the ith row and the jth column is denoted as htij ) belongs to the following set H = H ∈ CK×K |hmin < |Re [hkk ]| , |Im [hkk ]| < hmax . (4.6) The constants hmin and hmax are quantizer design parameters, and in Section 4.4.2 we will discuss how to select these parameters. If the channel matrix does not belong to this set, no transmission occurs during this fading block. We refer to this event as channel outage event, and channel outage probability is defined as follows: out Pch , Pr {H ∈ / H} . (4.7) The complex plane is divided into multiple equal-sized (∆ × ∆) square regions. Each of these regions is called a quantization cell, and ∆ is termed quantization step-size. The quantizer maps channel coefficients within a quantization cell to the centroid of the corresponding cell, i.e. the quantized value. We represent the quantization of ˜t . ˆ t , and the corresponding quantization error is denoted as δ t = h ˆt − h ˜ t as h h kl kl kl kl kl The number of bits associated to the quantization of direct links and interference links can be different in general. Therefore, we consider two types of quantizers at each destination with possibly different resolutions. A direct link quantizer is associated to the direct link and uses 2NI bits, and an interference link quantizer is associated to the interference links and uses 2NII bits for the quantization of each interference link’s gain. The step size of the direct (interference) link quantizers is ∆I = (hmax − hmin ) /2NI −1 ∆II = (hmax − hmin ) /2NII −1 . 4.2. Pilot-assisted Ergodic Interference Alignment 69 Before data transmission starts, each destination broadcasts the estimated channel gains (when analog feedback is deployed), or sends Nf = 2NI + 2(K − 1)NII (4.8) bits through feedback channels to all other terminals (when there is digital feedback). The feedback channels are in general subject to errors. Since we intend to investigate the impact of quantized CSI in the case of digital feedback and channel estimation error in the case of analog feedback on network performance, we assume that feedback channels are error-free. 4.2.3 Data Transmission Phase For data transmission, we consider a multiplexed coding scheme similar to the one proposed in [GV97], where there are multiple codebooks each associated with a specific channel state. For a given channel state, Sk encodes message mk to a ′ Td ′ length N ′ Td codeword {Xki }N i=1 , where N is the number of fading blocks with the same channel state and Td = (1 − α)T is the duration of data transmission within each block. There is a power constraint E |Xk |2 < Pd . In fading block n, Sk sends (n+1)T {Xki }i=nT +KTτ +1 during Td data transmission time slots. The channel output at Dk is i i Yd,k = hnT kk Xk + K X i i hnT kl Xl + Zk , i = nT + KTτ + 1, ..., (n + 1)T (4.9) l=1,l6=k where Zki ∼ CN (0, 1). We apply the EIA scheme proposed in [NGJV12], but we assume that only the estimated channel gains are available at terminals. Thus, if the quantized channel gains at fading blocks n and np (np > n) satisfy ˆ np T = h ˆ nT h kk kk np T ˆ ˆ nT , (∀k, l ∈ {1, 2, ..., K}, k 6= l), hkl = −h kl (4.10) then Sk at fading block np retransmits the signal which was transmitted at fading n T +i block n, i.e. Xk p = XknT +i (i = KTτ + 1, ..., KTτ + Td ). To avoid measure zero events, the channel pairing in (4.10) in the analog feedback case can be performed based on a quantized version of the estimated channel gain using sufficiently fine quantization [NGJV12]. It can be shown that if channel gains have symmetric distribution and long delay can be tolerated, with high probability the complement channels can be found. The destination Dk receives the following signals nT +i nT +i Yd,k = hnT + kk Xk K X nT +i hnT +ZknT +i kl Xl (4.11) l=1,l6=k K X np T +i np T n T n T +i nT +i Yd,k =hkk Xk + hklp XlnT +i+Zk p . l=1,l6=k (4.12) 70 K-user SISO Interference Networks: Pilot-assisted Interference Alignment Then, it combines the received signals and forms the following signal nT +i n T +i Y d,k = YknT +i + Yk p ˜ nT + εnT + εnp T + δ nT + δ np T = 2h XknT +i kk kk kk kk kk + K X np T n T n T +i nT εnT + δkl + δklp XlnT +i+ ZknT +i+Zk p . (4.13) kl +εkl l=1,l6=k The receiver decodes its message after receiving all N ′ segments of the transmitted codeword. Each source transmits at power P . The transmission power of pilot symbols (Pτ ) and the one for data symbols (Pd ) can be different in general. Let Pd = βP , where 0 ≤ β ≤ 1/(1 − α) is a power allocation factor. A large value of β implies that a large power is allocated to data transmission and less power is left for pilot transmission. Because of energy conservation, we have αT Pτ /K + (1 − α)T Pd = T P. (4.14) Therefore, Pτ = K ((1 − (1 − α) β)/α) P . Clearly, there is a trade-off between the power allocation for channel training and the one for data transmission. 4.3 Analog Feedback In this part, we evaluate the performance of the proposed scheme in the previous section when analog feedback is deployed. We also study the optimum power allocation between channel training and data transmission. 4.3.1 Achievable Rate Region In this section, we present an achievable rate region. Proposition 4.3.1. In the K-user interference network presented in Section 4.1, a rate tuple (R1 , R2 , ..., RK ) is achievable, where Rk = i 1−α h e , ∀k ∈ {1, 2, ..., K} E I Xk ; Y d,k | H 2 (4.15) e is the estimated and random variable Y d,k is given in (4.13). In this equation H channel matrix. Proof. The proof follows that of Theorem 2 in [NGJV12]. The difference is that in [NGJV12] the EIA scheme is applied based on the assumption that each destination has perfect knowledge of its incoming channel gains, but here only imperfect estimations of the channel gains are available. 4.3. Analog Feedback 71 We next present a closed-form inner bound on the achievable rate region in (4.15). Proposition 4.3.2. An inner bound on the achievable rate region in (4.15) is " !# 2 hkk Pd 2 ˜ 1−α Rk ≥ ∀k ∈ {1, 2, ..., K}. (4.16) Eh˜ kk log 1 + 2 1 + (Kσε2 ) Pd e in (4.15) can be lower bounded as Proof. The term I Xk ; Y k | H e (a) e − h Xk | H, e Yk I Xk ; Y k | H = h Xk | H (b) e Yk = h (Xk ) − h Xk | H, (c) bk | H, e Yk = h (Xk ) − h Xk − X (d) bk | H, e Yk = log 2πePd − h Xk − X (e) ≥ log 2πePd − log 2πeσ 2 (4.17) where (a) follows the definition of the conditional mutual information; (b) holds since the transmitted codeword is chosen independent of the noisy CSI; (c) follows bk to be a function of H, e and Y k , which will be specified in the below; by defining X (d) follows by the assumption that Xk is a complex Gaussian random variable; (e) follows by [CT91, Theorem 8.6.5] that shows the entropy of random variables with bounded variance is upper bounded by that of a random variable with Gaussian distribution. To obtain a tight bound on the achievable rate in (4.17), we choose bk to be an MMSE estimate of Xk ; that is X i h ∗ e Yk E Xk · Y k | H, bk = h iY k X ∗ e Yk E Y k · Y k | H, ∗ e hkk Pd (4.18) = 2 Y k 1 + Kσε2 + 2 e hkk Pd which yields σ2 = 1+ Pd . ˜ kk |2 Pd 2|h (4.19) 1+(Kσε2 )Pd The details of the derivation of σ 2 are presented in Appendix 4.A. The proof is completed by substituting (4.19) in (4.17). 72 K-user SISO Interference Networks: Pilot-assisted Interference Alignment Theorem 4.3.1. In the considered K-user interference network in Section 4.1, the rate-tuple (R1 , R2 , ..., RK ) is achievable where " !# ˜ kk |2 P 1−α 2β|h Rk = , (4.20) Eh˜ kk log 1 + βKP 2 N0 + 1+(1−β(1−α))T P /N0 ˜ kk ∼ CN 0, T P (1−β(1−α))/N0 . and h 1+T P (1−β(1−α))/N0 Proof. The proof follows from Proposition 4.3.2 by using the fact that the estimation error of an MMSE estimator is uncorrelated with the estimated channel. The variance of the estimation error given in (4.4) can be simplified by substituting Tτ = αT and Pτ given in (4.14). This achievable rate region can be further simplified by computing the expectation in (4.20) and using equation (34) in [AG99] as presented in the next corollary. Corollary 4.3.1. The achievable rate Rk given in (4.20) can be simplified as follows Rk = where E1 (x) = R∞ 1 1−α log2 (e) exp (1/SNReq )E1 (1/SNReq ) 2 1 −xt dt, te SNReq = N02 (4.21) x > 0, and 2(1 − (1 − α)β)T βP 2 . + (1 − (1 − α)β)T P N0 + KβP N0 (4.22) We use this result in the next part to find the optimum power allocation between channel training phase and data transmission phase. 4.3.2 The Optimum Power Allocation If transmitters are capable to transmit at different powers during the channel training and the data transmission phases, then power can be allocated such that the achievable rate region in Corollary 4.3.1 be enlarged. Proposition 4.3.3. In the considered interference network, for a given α, the optimum power allocation between pilot transmission phase and data transmission phase is Pd,opt Pτ,opt = βopt P, = K ((1 − (1 − α) βopt )/α) P, (4.23) where βopt 1 1+ = 1−α s 1 + KP/(1 − α) 1 + P T /N0 !−1 . (4.24) 4.3. Analog Feedback 73 Proof. Since Rk in (4.21) is a monotonic increasing function of SNReq , it is sufficient to maximize SNReq . We can prove that SNReq is a strictly concave function of β. Therefore, a unique β that maximizes Rk can be found by solving KKT conditions [BV04]. Corollary 4.3.2. If P ≫ 1, then the optimum power allocation factor is approximately as follows βopt ≈ 1/(1 − α) p . 1 + KN0 /T (1 − α) (4.25) This can be further simplified for sufficiently large or sufficiently small networks: ( p T /((1 − α)KN0 ) K≫T . (4.26) βopt ≈ 1/(1 − α) T ≫K This result shows that at high-power regime, in sufficiently large networks, βopt depends on T , α, and K; while in sufficiently small networks, βopt is approximately equal to its maximum possible value which is 1/(1 − α). It can be seen that, in sufficiently large networks, βopt increases as T increases, thus, by increasing T more power should be allocated for data transmission. Also, βopt decays as the size of the network K increases. This implies that as the size of the network increases, less power should be allocated for data transmission, and instead channel training most be performed more accurately. In all cases, βopt decays as α decreases. The reason is that the allocated power for pilot transmission should be increased to compensate the rate loss due to the shorter period of pilot transmission. 4.3.3 Achievable Degrees of Freedom Region In this section, investigate the performance of the proposed scheme at high-SNR regime. The following theorem presents the achievable DoF region. Theorem 4.3.2. In the K-user interference network with coherence time T , if no CSI is a priori available at terminals, then the DoF region (d1 , d2 , ..., dK ) is achievable where ( 1 1− K K<T 2 T dk = . (4.27) 0 K≥T Proof. Using (4.20), we have dk = lim Rk /log P P →∞ ˜ kk |2 P (a) 1 − α 2β|h Eh˜ kk lim log1+ = P →∞ 2 1 + T P βKP 1+ N (1−β(1−α)) 0 1−α , = 2 /log P (4.28) 74 K-user SISO Interference Networks: Pilot-assisted Interference Alignment where (a) follows the monotone convergence theorem [MW12]. To maximize the achievable DoF in (4.28) we set α equal to its minimum possible value which is K/T . According to Theorem 4.3.2, if T ≫ K, then the achievable DoF by the EIA scheme with perfect CSI can be preserved. However, the achievable DoF for each user decays by increasing the number of users. The achievable total DoF depends on both the number of the users and the achievable DoF for each user. A specific number of users maximizes the achievable total DoF. Thus, we select only a subset of users called active users to transmit. Theorem 4.3.3. In the K-user interference network with ergodic block fading Gaussian channel and coherence time T , the achievable total DoF is K dΣ = 12 Kopt 1 − Topt , (4.29) where Kopt is the number of the active users Kopt = min T2 , K . (4.30) Proof. Let K ′ < T denote the number of the active users, then according to Theorem 4.3.2 we have ′ dΣ = K X i=1 2 ′ 1 dk = 12 K ′ 1− KT = − 2T K ′ − T2 + T 8. We can observe that dΣ is maximized when K ′ = min K, T2 . To maximize the total DoF (and the network throughput at high SNR), in large networks (K > T /2), Theorem 4.3.3 suggests to first apply a user selection scheme, and then perform channel training and interference alignment only within the subset of active users. Since the network is symmetric, a random user selection is sufficient. In addition, this theorem crystallizes the dependency of the optimum number of active users to be selected on the coherence time of channel. 4.3.4 Numerical Evaluation In this section, we numerically evaluate the analytical results presented in the previous sections when analog feedback is deployed. Figure 4.2 shows the optimum power allocation factor βopt given in (4.24) as a function of the number of users for different values of T . We set P/N0 = 20 dB, and α = 0.1. It can be observed that βopt decays as the number of users increases. The intuition behind this observation is that in large networks the impact of residual interference due to imperfect interference alignment become more important, thus, it is recommended to allocate more power to pilot symbols in order to acquire CSI more accurately. Also, we can 4.3. Analog Feedback 75 1.4 bC 1.2 rS ut qp qp qp ut qp ut ut rS 1 qp qp ut = 100 = 1000 = 10000 = 100000 qp rS rS ut bC T T T T qp βopt 0.8 rS ut bC qp 0.6 bC rS ut qp 0.4 bC rS ut bC qp rS ut qp 0.2 bC rS ut bC rS bC 0 101 102 103 104 ut bC 105 bC rS rS bC 106 K Figure 4.2: The optimum value of power allocation factor (β) versus the number of users (K) for different coherence time (T ). observe that βopt increases by increasing T . This implies that as the channel coherence time increases, a larger power should be allocated for data transmission. It is clear from (4.4) that to preserve a given variance of the channel estimation error, a lower Pτ is required as T and consequently Tτ increases. Thus, a larger power can be devoted for data transmission. Figure 4.3 shows the achievable rate per user of the PAEIA scheme in both cases that power allocation is optimized (β = βopt ) and when there is no power optimization (β = 1). The network parameters are K = 40, T = 1000, and we set α = 0.04. We plot the achievable rate of the TDMA with pilot-based channel training scheme and that of the EIA scheme with perfect CSI for comparison. This figure shows that, for the given parameters, the PAEIA scheme can achieve almost the same (slightly less) DoF as the EIA scheme with perfect CSI. This confirms the result in Proposition 4.3.2 when K is sufficiently smaller than T . A large gap between the achievable rate of the EIA scheme and that of the TDMA scheme can be seen. Furthermore, 2 dB gain can be seen using the optimum power allocation compared to the case with uniform power allocation. Figure 4.4 shows the achievable sum-rate of the PAEIA scheme and that of the PAEIA with user selection (PAEIA-US) for different number of users. In this exam- 76 K-user SISO Interference Networks: Pilot-assisted Interference Alignment 7 rS Achievable rate [bits/channel use] 6 bC uT 5 PAEIA, β = βopt PAEIA, β = 1 EIA, perfect CSI TDMA rS bC rS bC rS bC 4 rS bC rS 3 bC rS bC 2 rS bC 1 rS bC rS bC 0 uT uT uT uT 0 5 10 15 uT uT uT uT uT 20 25 30 35 40 P/N0 [dB] Figure 4.3: The achievable rate of the PAEIA and TDMA schemes. ple, T = 100, P/N0 = 10 dB, and β = βopt . For the PAEIA scheme, we set α = K/T and for the PAEIA-US scheme we set α = Kopt /T , where Kopt is given in (4.30). It can be seen that the achievable sum-rate of the PAEIA scheme is maximized for a specific number of users. This observation coincides with Theorem 4.3.3. The intuition behind this result is that, by increasing the number of the active users, in one hand the number of the independent transmitted symbols increases, and on the other hand the achievable rate per each user decreases due to the less available resources for the channel training and consequently more interference. It can be seen that the PAEIA-US scheme outperforms the PAEIA scheme in large networks (K > T /2). 4.4 Digital Feedback Since the bandwidth of the feedback channel is in general limited, the destinations can compress the estimated CSI and send quantized CSI instead through digital feedback signals. Applying interference alignment based on the quantized CSI, inter-user interference can be only partially eliminated so that some residual interference remains at each destination. This residual interference, if not appropriately managed, will degrade the system performance. Hence, the sources should exploit 4.4. Digital Feedback 77 200 PAEIA-US, no a priori CSI PAEIA, no a priori CSI EIA, perfect CSI Achievable sum-rate [bits/channel use] 180 160 140 120 100 80 60 40 20 0 0 10 20 30 40 50 60 70 80 90 100 K Figure 4.4: The achievable sum-rate versus K. the quantized CSI not only to partially eliminate the interference, but also to adapt their transmission strategies, e.g. by adapting transmission rate or controlling transmission power, to fulfil certain service requirements, i.e. to increase throughput or to transmit successfully at given rates with minimum power, respectively. In this part, we investigate the impact of quantized CSI on the performance of the EIA and propose adaptive transmission schemes to enhance the performance of the EIA with digital feedback: we first study a power control problem in which each user desires to successfully transmit information at a fixed rate using the minimum transmission power. We propose a power control scheme which adapts transmission power values such that the mutual information corresponding to each sourcedestination pair is always larger than the transmission rate, and thus, transmitted codewords can be successfully decoded at the intended destination. Next, we study a throughput maximization problem for adaptive communication systems in which each user desires to maximize the throughput. In contrary to the previous case, transmission power values are fixed. Since each source only knows quantized CSI, it is not aware of the exact value of mutual information between itself and its intended destination. Therefore, in certain channel realizations, the mutual information may fall below the transmission rate and communications fails which leads to an outage event. The outage probability can be used to quantify 78 K-user SISO Interference Networks: Pilot-assisted Interference Alignment throughput as a measure of the amount of information that can be successfully transmitted. The impact of the transmission rate on throughput is twofold: On the one hand, increasing the rate tends to improve the throughput; but on the other hand, a higher rate leads to a higher outage probability which results in a throughput loss. It can be shown that, for given channel gains, there is a specific rate which maximizes network throughput. We propose a rate adaptation scheme to maximize network throughput. Moreover, the EIA scheme presented in Chapter 2 requires asymptotically long delay to achieve the promised performance. It has been shown in [JAP12] that the required delay can be reduced by sacrificing transmission rate, but still asymptotically long delay is needed to achieve the promised rates. The performance limit of the EIA scheme in delay-limited communication systems is in general unknown. In the following parts of this chapter, we extend our results on the throughput maximization problem to communication systems with delay constraints. In the case of analog feedback, as discussed in the first part of the chapter, there is a natural connection between estimating the CSI at receivers and sending analog feedback. I n this part we however look at digital feedback, and thus there is no natural way of connecting the CSI estimation at the receivers with a corresponding feedback mechanism. Since we wish to focus on the impact of the feedback quantization itself, and not the specific implementation of estimator algorithms at the receivers, we now assume that the receivers have perfect CSI a priori (εnT kl = 0). Hence no radio resources are explicitly used for channel training, while in practice the receivers would still need to get CSI information via a training phase. In addition, sources may have different transmission powers and pk denotes the transmission ˆ n and δ n to denote h ˆ nT and δ nT , power of Sk . For the sake of simplicity, we use h kl kl kl kl respectively. In the following, we first study the power control problem and next address the throughput maximization problem. 4.4.1 Power Control Problem In certain communication systems, fixed-rate transmission is desired [CHLT08]. In a time-varying environment, channel quality is changing over different blocks. Therefore, radio resources (e.g. power) can be adaptively allocated to support successful transmission. Consider the time-varying K-user interference network presented in Section 4.1 where transmission rates are fixed. We desire to allocate powers such that the mutual information corresponding to each source-destination pair remains larger than a certain level. Assume that the channels with block indices m and mp are complement. To study the power control problem, we assume that a complement channel corresponding to each channel realization can be found after sufficiently many fading blocks to occur. According to the transmission scheme presented in Section 4.2, we require each source to repeat the same codeword over blocks m and mp . Following the input-output relation in (4.13) and the assumption that perfect CSI is a priori available at destination, the SINR of the equivalent received signal at Dk 4.4. Digital Feedback 79 (k ∈ {1, 2, ..., K}) is SINRym k = 2 ˆm m m + δkkp pk 2hkk + δkk . P δ m + δ mp 2 p l 2+ K l=1,l6=k kl (4.31) kl This SINR value is random and depends on the quantization errors which are ≥ SINRmin unknown to the sources. This value is lower bounded as SINRym , where ym k k min SINRym can be computed as follows: k h i h i ˆ m 2 2 ˆ m + Im h ˆ m pk 2 h +∆ −2∆ h Re I I kk kk kk . (4.32) SINRmin = P ym K k 1 + ∆2II l=1,l6=k pl Therefore, the mutual information of the source-destination pair Sk − Dk can be min 1 lower bounded by 2 log2 1 + SINRym where the factor 21 comes from the fact k that two blocks are used to transmit one codeword. In order to guarantee successful transmission at the fixed rate Rk , the following condition should be satisfied: 1 log2 1 + SINRym > Rk . (4.33) k 2 Clearly, if the sources compute their transmission powers to meet the following condition, we can ensure that the inequality in (4.33) is satisfied: 1 m log 1 + SINRymin > Rk . (4.34) k 2 According to (4.32), the condition in (4.34) can be rewritten as a power constraint pk > Ik (p), where PK 22Rk − 1 1 + ∆2II l=1,l6=k pl Ik (p) , 2 (4.35) h i h i , ˆ m 2 − 2∆ m + m ˆ ˆ 2 h + ∆ h h Re Im I kk kk kk I and p , [p1 · · · pK ]T . Thus, we define feasibility set as follows: P EIA , {p|pmax p 0, p ≻ I(p)}, (4.36) where I(p) , [I1 (p) I2 (p) ... IK (p)]T is the interference function and Ik (p) is defined in (4.35). The notation a ≻ b (a b) means that every element of vector a is larger than (larger than or equal to) the corresponding element of vector b. The power control problem can be formulated as the following optimization problem: minimize s.t.p∈P EIA K X pl . (4.37) l=1 If P EIA 6= ∅, then this problem is said to be feasible, otherwise the problem has no solution and terminals stop transmission during the corresponding fading block. 80 K-user SISO Interference Networks: Pilot-assisted Interference Alignment Feasibility of the Power Control Problem In this part, we study the feasibility of the power control problem in (4.37). This problem has a feasible solution if the set in (4.36) is non-empty. The set P EIA given in (4.36) can be represented in the following equivalent form P EIA = {p|pmax p 0, p ≻ Ap + b}, where the element of the kth row and the lth column of the matrix A is 22Rk − 1 ∆2II A(k, l) , 2 h i h i , ˆ m 2 − 2∆ ˆ m + Im h ˆ m 2 h + ∆ h Re I kk I kk kk (4.38) (4.39) and the kth element of the vector b is 22Rk − 1 b(k, 1) , 2 h i h i . ˆ m 2 − 2∆ m + m ˆ ˆ 2 h + ∆ h h Re Im I kk kk kk I (4.40) We use this representation of the feasibility set to study the condition under which the problem has a feasible solution. Since this set depends on transmission rates, the number of quantization bits, and the number of users, the feasibility of the power control problem will be affected by all these parameters. The next theorem presents the condition under which this problem has a feasible solution. Theorem 4.4.1. Assume that quantized channel gains are given, and the maximum transmission power is asymptotically large (i.e., pmax → ∞). The power control problem in (4.37) is feasible (P EIA 6= ∅) if and only if λmax (A) < 1, where λmax (A) is the Perron-Frobenius eigenvalue of the matrix A, and the matrices A is defined in (4.39). Proof. Since it is assumed that pmax → ∞, we need to show that there is a positive power vector p satisfying p ≻ Ap+b if and only if λmax (A) < 1. For K > 2, the matrix A is a regular matrix (a matrix X is called a regular matrix, if all the entries of some power of the matrix X are positive) because all the entries of A2 are positive. Therefore, the Perron-Frobenius Theorem [PSS05] guarantees the existence of a positive eigenvalue λmax (A) and the corresponding positive right and left eigenvectors pr ≻ 0 and pl ≻ 0 which satisfy λmax (A)pr = Apr and λmax (A)pTl = pTl A, respectively. To prove the necessary condition, assuming p1 ∈ P EIA we have p1 ∈ P EIA (a) ⇒ p1 ≻ Ap1 + b ⇒ p1 − Ap1 ≻ 0 (b) (c) ⇒ pTl (p1 − Ap1 ) > 0 ⇒ (1 − λmax (A))pTl p1 > 0 ⇒ λmax (A) < 1, (d) (e) (4.41) 4.4. Digital Feedback 81 where (a) follows from the definition in (4.38); (b) follows from the fact that b ≻ 0; (c) follows from the positivity of the left eigenvector (pl ≻ 0); (d) follows from the characteristic of eigenvectors and (e) follows from the fact that pl ≻ 0 and p1 ≻ 0, and consequently pTl p1 > 0. To prove the sufficient condition, assuming λmax (A) < 1 we have (a) λmax (A)pr = Apr ⇒ p2 = Ap2 + b + a(1 − λmax (A))Apr − b (b) ⇒ p2 Ap2 + b (c) ⇒ p2 ∈ P IA ⇒ P IA 6= ∅, (4.42) where (a) holds if we set p2 = aλmax (A)pr ; (b) holds if we set a = max i b(i, 1) + ǫ, (1 − λmax (A))A(i, :)pr (4.43) where b(i, 1) is the ith element of b, A(i, :) is the ith row of A, and ǫ > 0 is a constant, then since λmax (A) < 1 we have a(1 − λmax (A))Apr − b ≻ 0; (c) follows from the definition in (4.38) and the assumption that pmax → ∞. Figure 4.5 shows the asymptotic feasibility probability, when NI = NII = 8 and pmax → ∞, versus rate for different number of users (K ∈ {3, 4, 5, 6}). This figure shows that the feasibility probability monotonically decreases as transmission rate increases. In a communication system, it is desired to have the feasibility probability larger than a certain threshold. These curves can be used to find the maximum transmission rate such that feasibility probability can be maintained at the desired level. Also, we can see that as the number of users increases the probability that the power control problem has a feasible solution decays. The reason is that by increasing the number of users, each user experience more severe interference, in addition the number of constraints that need to be satisfied for a feasible power control problem increases. In a fading block m, terminals stop transmission either if channel outage event occurs (i.e. Hm ∈ / H), or if the power control problem does not have a feasible solution i.e. P EIA = ∅ . Therefore, the probability that no transmission occurs in fading block m is PnoT X = = Pr{H ∈ / H} + Pr{H ∈ H} × Pr{P EIA = ∅} out out × Pr{P EIA = ∅}, Pch + 1 − Pch (4.44) out where Pch is defined in (4.7). The channel outage probability and the feasibility probability both are affected by the quantizer parameters hmin and hmax . A larger hmax reduces the channel outage probability, but also reduces the feasibility probability of the power control problem because of larger quantization regions. In 82 K-user SISO Interference Networks: Pilot-assisted Interference Alignment uTrSbC 100 uTrSbC uTrSbC uTrSbC uTrSbC uTrSbC uTrSbC uTrSbC uT rS bC uT rS bC uT rS bC uT rS bC uT rS bC uT rS bC uT rS bC uT rS bC uT rS bC uT rS bC uT Sr Cb uT rS bC uT rS bC uT rS bC uT rS uT bC K K K K uT Feasibility probability 10−1 rS bC =3 =4 =5 =6 uT rS bC uT rS bC uT rS uT bC rS uT rS bC 10−2 uT bC rS 10−3 rS 10−4 bC 0 1 2 3 4 5 uT rS 6 Rate per user (bits/channel use) Figure 4.5: Feasibility probability versus transmission rate of each user in a K-user interference network, NI = NII = 8. contrast, a larger hmin increases the channel outage probability, and increases the feasibility probability of the power control problem. There are specific values for m these parameters which minimize PnoT X . In the numerical evaluations, we select these parameters by searching for the values which minimize this probability. If the power control problem has feasible solutions, the next question to be answered is how to find the solution which is corresponding to the minimum required transmission power. In the next subsection, we present an iterative power control algorithm to address this question. Iterative Power Control Algorithm We present an iterative power control algorithm, shown in Algorithm 1, to solve the power control problem in (4.37). In each iteration, Dk (k ∈ {1, 2, ..., K}) comˆ m and the total transmission putes function Ik (p) given in (4.35) according to h kk powers of the other sources in the previous iteration. Next, it updates the transmission power of the corresponding source following Algorithm 1. As we will show in Section 4.4.1, this algorithm converges to the optimum solution if the power control problem in (4.37) has a feasible solution. 4.4. Digital Feedback 83 Algorithm 1 Iterative Power Control Initialize: p1 (0), ..., pK (0), maxitr for t = 1 : maxitr do for k = 1 : K do Sk updates its transmission power: P pl (t−1) (22Rk −1) 1+∆2II K l=1,l6=k pk (t) = Ik (p(t − 1)) = ˆ m 2 2 . ˆ m ]|+|Im[h ˆ m ]|) 2|hkk | +∆I −2∆I (|Re[h kk kk end for t=t+1 end for Convergence of the Power Control Algorithm There are two main questions regarding the convergence of Algorithm 1. The first question is whether the sequence of the computed powers by the algorithm converges to a fixed-point. If the answer to this question is positive, we need to answer whether the fixed-point is corresponding to the minimum required powers. In the following theorem we answer these questions. Theorem 4.4.2. If P EIA 6= ∅, for any initial power vector p(0), Algorithm 1 converges to a unique fixed-point p∗ which corresponds to the solution of the problem in (4.37). Proof. To provide the convergence proof of the proposed algorithm, we refer to a family of functions defined in [Yat95]. I′ (p) is called a standard interference function if for all vectors p, p′ 0, it satisfies the following conditions: 1) Positivity condition : I′ (p)≻0, 2) Monotonicity condition : I′ (p)I′ (p′ ) ∀ p p′ , 3) Scalability condition : αI′ (p)≻I(αp) ∀ α > 1. (4.45) We now show that the function I(p) given in (4.35) satisfies these conditions. For the simplicity of presentation, we rewrite Ik (p) as follows K X (4.46) pl , Ik (p) = Lk 1 + ∆2II l=1,l6=k where 22Rk − 1 Lk = h i h i > 0 ˆ m 2 2 − 2∆ ˆ kk + Im h ˆ kk 2 h + ∆ h Re I kk I is a constant. (4.47) 84 K-user SISO Interference Networks: Pilot-assisted Interference Alignment 1) Positivity condition: Ik (p) = Lk 1 + ∆2II K X l=1,l6=k pl ≥ L k > 0 (4.48) Thus, the interference function given in (4.35) satisfies the positivity condition. 2) Monotonicitycondition: PK PK If p p′ , then 1+∆2II l=1,l6=k pl ≥ 1+∆2II l=1,l6=k p′l and since Lk > 0 we have Ik (p) ≥ Ik (p′ ). Therefore, the monotonicity condition is satisfied. 3) Scalability condition: If α > 1, then K X pl Ik (αp) = Lk 1 + α∆2II < αLk 1 + ∆2II and the scalability condition is satisfied. l=1,l6=k K X l=1,l6=k pl = αIk (p), These conditions are satisfied for each k (k ∈ {1, ..., K}), and we can conclude that the function I(p) given in (4.35) is a standard interference function. Therefore, according to Theorem 2 in [Yat95] for any initial power vector p(0) Algorithm 1 is a standard power control algorithm and converges to a unique fixed-point p∗ . Lemma 1 in [Yat95] implies that this fixed-point corresponds to the solution with minimum powers. In this part, we numerically evaluate the performance of the proposed interference alignment and power control (EIA-PC) scheme. Figure 4.6 shows the required powers of the IA-PC scheme for each user, averaged over different channel realizations, in a K-user interference network (K ∈ {3, 5, 7}, NI = NII = 8) versus transmission rate of one user (R1 = R2 = R3 = R). Also, the performances of TDMA with power control denoted as TDMA-PC (with the same number of feedback bits), and the EIA scheme with perfect CSI are shown for comparison. It can be seen in Figure 4.6 that even with only limited feedback bits, applying the EIA-PC scheme outperforms the TDMA-PC scheme in the intermediate rate region. When the interference alignment scheme is performed based on perfect CSI, the required powers for transmission at a given rate do not increase by increasing the number of users. However, they considerably increase for the TDMA-PC scheme, especially at high transmission rates. If only limited feedback bits are available and EIA-PC is performed, at the low-rate regime, increasing the number of users does not significantly increase the required powers. However, if the transmission rate is high, the 4.4. Digital Feedback 85 uT 40 rS Tu uT uT rS uT uT 35 bC rS bC 30 uT bC bC bC rS rS bC uT bC bC bC rS bC rS uT bC bC rS rS bC Average power [dB] rS bC 25 bC uT rS uT 20 bC rS 15 uTrS bC bC K = 3, EIA-PC, quantized CSI K = 3, TDMA-PC, quantized CSI K = 5, EIA-PC, quantized CSI K = 5, TDMA-PC, quantized CSI K = 7, EIA-PC, quantized CSI K = 7, TDMA-PC, quantized CSI EIA-PC, perfect CSI bC rS uTrSbC bC bCuTrSbC uT bCuT rS rS rS bC uT 5 uT 0 bCrS −5 0 bC rSuTbC rSuTbC 10 rS uT 1 2 3 4 5 6 7 Rate [bits/channel use] Figure 4.6: Average transmission power versus rate of one user in a K-user interference channel, NI = NII = 8. performance is more severely affected by the residual interference; consequently, the required powers notably increase as the number of users increases. Figure 4.7 illustrates the performance of the EIA-PC scheme for different strategies to allocate feedback bits to direct link and interference links when the total number of feedback bits from each destination is fixed at Nf = 42. It can be seen that the best feedback allocation strategy depends on the transmission rates. This figure reveals that in the low-rate regime, allocating more bits to the quantization of the direct links is preferred, while in the high-rate regime, it is more efficient to allocate more bits to the quantization of the interference links instead. This is because when the desired transmission rate is low, the network operates in the noise-limited regime and accurate power control is more effective. However, to ensure successful transmission at high-rate regime, the users are required to transmit with large powers, and the network becomes interference-limited. Therefore, at this regime, it is preferred to more precisely eliminate the interference by allocating more feedback bits to the interference links. 86 K-user SISO Interference Networks: Pilot-assisted Interference Alignment 45 bCbC rS bC rS uT rS rS bC uT uT 35 Average total power [dB] uT bC 40 uT uT rS uT uT bC uT uT uT uT uT 30 uT rS bC 25 rS rS rS 20 rS rS rS bC bC 15 bC uT 10 bC rS bC 5 0 0 2 4 6 8 10 EIA-PC, perfect CSI EIA-PC, NI = 5, NII = 8 EIA-PC, NI = 7, NII = 7 EIA-PC, NI = 9, NII = 6 12 14 16 18 20 Sum-rate [bits/channel use] Figure 4.7: Different strategies for sharing total feedback bits (Nf = 42) among direct link quantizer and interference link quantizer in a three-user interference channel. 4.4.2 Throughput Maximization Problem Considering adaptive wireless communication systems, another group of systems, in contrast with the ones addressed in the previous section, may not have capability to adaptively adjust transmission powers and successful fixed-rate transmission is not their primary concern. Instead, they can adapt their transmission rate according to the channel conditions in order to maximize network throughput. In this section, we discuss this group of systems and address a throughput maximization problem. We propose a rate adaptation scheme to maximize network throughput. The transmission scheme mentioned in Section 4.2 is employed to remove interference at each destination. However, due to the limited resolution of the quantized CSI at sources, interference alignment cannot be performed accurately and some interference remains at destinations. The exact value of each source-destination pair’s mutual information is unknown at the corresponding source. Thus, for some channel realizations the instantaneous mutual information might become lower than the data transmission rate, and consequently the transmission fails. In addition, terminals stop transmission if channel gains do not belong to the set H defined in (4.6). Out- 4.4. Digital Feedback 87 age probability is defined as the probability that transmission fails because of either of these events [TV05]. The outage probability increases as we increase data transmission rate. Network throughput, defined as the expected rate of successful data transmission over the network, depends on both data transmission rate of each user and the corresponding outage probability. To maximize network throughput we use the fact that, for a given quantized CSI available at terminals, there is a specific data transmission rate for each source which maximizes the network throughput. In this section, we first compute a closed-form upper bound on the outage probability, and then we propose a rate adaptation scheme based on the quantized CSI at each source which maximizes the lower bound on the throughput. Outage Probability Analysis Terminals can have a successful communication only if the channel matrix belongs to H defined in (4.6) (otherwise no transmission occurs) and the mutual information is larger than the transmission rate. Otherwise, an outage event will be declared. Let channels with block indices m and mp belong to H and be complement according to (4.10). Requiring each source to transmit the same codeword over these blocks, the SINR of the equivalent received signal at Dk is given in (4.31). The mutual . If information between the source–destination pair Sk − Dk is 21 log2 1 + SINRym k the transmission rate Rkm is smaller than this mutual information, decoding error probability can be made arbitrary small by choosing proper code with sufficiently long codewords. Otherwise, the channel between the source–destination pair Sk −Dk is said to be in outage. The outage probability given that quantized CSI is available at terminals is 1 out,m out,m m ˆ m , (4.49) < R h log2 1 + SINRym Pkout,m , Pch + 1 − Pch × Pr k kk k 2 out,m where Pch is channel outage probability defined in (4.7). Finding the exact expression for the outage probability in (4.49) is a challenging task. Therefore, instead we present a closed-form upper bound on Pkout,m in the following theorem. m Theorem 4.4.3. If Rkm < Rmax,k , then the outage probability Pkout,m defined in (4.49) can be upper bounded as follows out,m out,m out,m Pkout,m ≤ Pch + 1 − Pch × Pup,k (4.50) where out,m Pup,k = m Rmax,k = 1 2, 1 + (rkm ) ˆ m |2 + ∆2 pk 12| h kk I 1 , log2 1 + PK 2 ∆2II i=1,i6=k pi + 6 (4.51) (4.52) 88 K-user SISO Interference Networks: Pilot-assisted Interference Alignment ˆ m |2 +∆2 pk 12|h kk I rkm s = out,m Pch 3 2Rm 2 k −1 7 4 90 ∆II K P i=1,i6=k p2i − + ∆2II 3 8 3 K P i=1,i6=k 2 2 7 ∆I + 90 ∆4 |hˆ m kk | 2 I 2Rm 2 k −1 2K 2 1 − (2Q (hmin ) − 2Q (hmax )) = pi − 2 , (4.53) p2k , (4.54) otherwise Pkout,m ≤ 1. out,m Proof. The probability Pch defined in (4.7) can be computed as follows out,m Pch = 1− K Y K Y k=1 l=1 = 1− K Y K Y m (Pr{hmin < |Re[hm kl ]| < hmax }×Pr{hmin < |Im[hkl ]| < hmax }) 2 (2Q(hmin)−2Q(hmax))2 = 1−(2Q(hmin )−2Q(hmax))2K (4.55) k=1 l=1 where Q(x) = Pr √1 2π R∞ x 2 e−u /2 du. For the other term in (4.49) we have ) ( ˆ m |2 pk 1 4|h m m m kk ˆ ˆ log2 1 + SINRym − 2 hkk , (4.56) < Rk hkk = Pr Y ≥ 2Rm k 2 2 k −1 ˆ | p 4|h ˆ m is given, is a constant where the quantity 2Rkkm k − 2, conditioned on that h kk 2 k −1 value. The random variable Y is m 2 K X Y = l=1,l6=k − − m m m 2 m 2 pl + Im δkl + δkl p Re δkl + δkl p h i h i ˆ m Re δ m + δ mp + Im h ˆ m Im δ m + δ mp pk 4 Re h kk kk kk kk kk kk m 22Rk − 1 m m mp 2 m 2 Re δkk + δkk + Im δkk + δkkp pk m 22Rk − 1 . (4.57) The mean and the variance of Y are (computed in Appendix 4.B) µY = σY2 = ∆2II 3 K X ∆2I pk , m 3 22Rk − 1 l=1,l6=k 2 7 8 ˆ m 2 4 ∆ + p2k h ∆ K I kk 3 90 I 7∆4II X 2 . pl + 2 m 90 22Rk − 1 l=1,l6=k pl − (4.58) (4.59) 4.4. Digital Feedback 89 2 Let X be a random variable with mean µX and variance σX . For any real value r > 0, the Cantelli inequality implies Pr {X − µX ≥ rσX } ≤ 1/ 1 + r2 [MR69]. ˆ m |2 pk / 22Rm k − 1 − 2, and applying the Setting X = Y and rkm σY + µY = 4|h kk Cantelli inequality lead to the value of rkm that is given in (4.53), and the upper bound on Pkout,m shown in (4.50). Theorem 4.4.3 clarifies how the upper-bound on the outage probability of each source–destination pair depends on different parameters such as the transmission rate of the source, the transmission powers of all sources, the number of users, the quantization resolutions, and the quantized direct link’s gain. We will use this result in Section 4.4.2 to design a rate adaptation scheme which maximizes the lower bound on network throughput. In the special case that quantization bits are asymptotically large, we have the following corollary of Theorem 4.4.3. ˆ m |2 pk , then high-resolution quantizers Corollary 4.4.1. If Rkm < 21 log2 1+2|h kk lead to lim ∆I ,∆II →0 Proof. If Rkm < 1 2 Pkout,m = 0. ˆ m |2pk , then log2 1+2|h kk (4.60) lim rkm = +∞ and the upper bound ∆I ,∆II →0 on Pkout,m approaches 0. Since, Pkout,m is lower bounded by zero, we can conclude that lim Pkout,m = 0. ∆I ,∆II →0 This result shows that when the quantizers are sufficiently fine, if the transmis2 sion rate Rkm is less than 21 log(1 + 2|hm kk | pk ), then reliable transmission is possible. 1 Indeed, the average rate of 2 Eh log2 (1 + 2|h|2 pk ) is achievable, where Eh [·] is the expectation over direct link’s gain. This coincides with the results in [NGJV12], where perfect CSI is available at terminals. Network Throughput Maximization For any pair of complement blocks m and mp , we define the throughput of source– destination pair Sk − Dk as follows Tkm , 1 − Pkout,m Rkm . (4.61) The network throughput T m can be represented as the summation over the throughput of the individual users, i.e., Tm = K X k=1 Tkm = K X 1 − Pkout,m Rkm . k=1 (4.62) 90 K-user SISO Interference Networks: Pilot-assisted Interference Alignment m m According to Theorem 4.4.3, for Rkm ≤ Rmax,k where Rmax,k is given in (4.52), out,m we can use the upper bound on Pk in (4.50) to find a lower bound on the throughput as follows Tm ≥ out,m 1 − Pch K X k=1 out,m 1 − Pup,k Rkm , (4.63) m while for Rkm > Rmax,k the lower bound on the throughput is the trivial value of zero. For each complement block pair, the lower bound (4.63) can be maximized by a proper rate adaptation. This network throughput maximization problem can be formulated as follows m Topt , max 0≤Rm ≤Rm i max,k i∈{1,...,K} K X out,m out,m 1 − Pch 1 − Pup,k Rkm . (4.64) k=1 In the following, we discuss the solution of this problem. For a given transmisout,m sion power vector p, since Pup,k only depends on variable Rkm , the optimization problem in (4.64) can be decomposed to individual optimization problems, each corresponding to one source, i.e., K X out,m m = 1 − Pch Topt k=1 maxm m 0≤Rk ≤Rmax,k out,m Rkm . 1 − Pup,k (4.65) out,m Substituting Pup,k given in (4.50), each source solves the following optimization problem (rkm )2 out,m m Tk,opt = 1 − Pch × maxm Rkm , (4.66) 0≤Rm ≤Rmax,k 1 + (rkm )2 k m where Rmax,k and rkm are given in (4.52) and (4.53), respectively. After certain mathematical manipulations and introducing an auxiliary optimization variable x, we have the following equivalent optimization problem: out,m m Tk,opt = 1 − Pch × max m 2R k −1 x=2 0≤x≤xmax , 0≤Rm k Ax2 + Bx + C m R , Dx2 + Ex + F k where A = − B= 2 3 ∆2II 3 K X i=1,i6=k 2 pi − 2 , ˆ m |2 + ∆2 − ∆II 12|h kk I 3 K X i=1,i6=k pi − 2 pk , (4.67) 4.4. Digital Feedback 91 1 ˆm 2 12|hkk | + ∆2I p2k , 9 2 K K 2 X X 7 4 ∆ ∆ p2i + − II pi − 2 , D= 90 II 3 i=1,i6=k i=1,i6=k K X 2 ˆ m |2 + ∆2 − ∆II E= 12|h pi − 2 pk , kk I 3 3 i=1,i6=k 8 ˆm 2 2 7 4 2 1 ˆm 2 2 2 12|hkk | + ∆I pk + |h | ∆I + ∆I pk , F = 9 3 kk 90 C= m xmax = 22Rmax,k − 1. (4.68) This problem is not convex. However, it can be proved that the feasible set of this problem satisfies the linear independent constraint qualification (LICQ) conditions [Hen92], and its duality gap is zero. Therefore, any pair of the primal and the dual optimal points of this problem must satisfy the Karush-Kuhn-Tucker (KKT) conditions [BV04]. Solving the KKT conditions, the necessary condition on the optimal solution (Rkm∗ ,x∗ ) is f (x∗ ) = ln(x∗ +1)+ (x∗ +1)−1 (Ax∗ 2 +Bx∗ +C)(Dx∗ 2 +Ex∗ +F ) = 0. (AE −DB)x∗ 2 +2(AF −DC)x∗ +(BF −EC) (4.69) The Newton method iteratively solves (4.69) as follows: x(i + 1) = x(i) − f (x(i)) , f ′ (x(i)) (4.70) where i is the iteration index and f ′ (x) is the derivative of f (x) with respect to x. The solution for the optimum rate is Rkm∗ = 21 log(1 + x∗ ), where x∗ is the convergence point of series {x(i)}∞ i=1 . In general, hmax and hmin are design parameters that can be optimized; a larger (smaller) value of hmax (hmin ) reduces the channel outage probability defined in (4.7); but, on the other hand it increases (decreses) the upper-bound on the outage probability given in (4.51) because of the larger (smaller) quantizer step-size. In the simulations, we numerically measure the throughput for different values of these parameters, and select the ones which maximize the network throughput. We numerically evaluate the throughput of the ergodic interference alignment and rate adaptation (EIA-RA) scheme. Figure 4.8 shows the lower bound on the expected throughput of a three-user interference network, where pk = p (k ∈ {1, 2, 3}) and NI = NII = N . As demonstrated in this figure, by increasing the number of quantization bits (N ), the throughput approaches the one when perfect CSI is available at each terminal. Also, it can be seen that for a given value of N , the throughput saturates at the high-SNR regime. The reason is that the residual interference power is proportional to the transmission power p and according to (4.31), as p increases, the SINR converges to a limited value. 92 K-user SISO Interference Networks: Pilot-assisted Interference Alignment 25 N = 3, EIA-RA N = 4, EIA-RA N = 5, EIA-RA N = 6, EIA-RA N = 7, EIA-RA N = 8, EIA-RA N = 9, EIA-RA N = 10, EIA-RA Perfect CSI, EIA Network throughput [bits/channel use] bC rS 20 r uT u 15 l ld * 10 * * ld ld *ld l l l *ldl u u u u ld * l u uT *u uT uT uT uT ld l uT *lulduT r r r r 5 *lduTlur luduT 0 *lduTrSbC −20 r r r *lurdTSbC *lduTrSbC −10 *rrSbC *lurlduTrS rS rS rS rS rS rS bC bC bC bC bC bC rS bC rS bC bC 0 10 20 30 40 SNR [dB] Figure 4.8: The lower bound on the throughput of a three-user interference network with limited feedback when the EIA-RA scheme is applied, NI = NII = N . For a fixed number of total feedback bits broadcasted by each destination (Nf = 30), the lower bound on the throughput is shown in Figure 4.9. We consider three different scenarios of feedback bits allocation similar to the ones mentioned for the power control problem in Figure 4.7. In the first one, we equally allocate bits to the direct and interference link quantizers (NI = 5, NII = 5), in the second scenario we allocate more bits to the quantizer dedicated to direct link (NI = 7, NII = 4), and finally in the third scenario we allocate more bits to the quantizer associated to the interference links (NI = 3, NII = 6). It can be seen in Figure 4.9 that at the low-SNR regime the second scenario outperforms the others, while at the high-SNR regime, allocating more bits to the quantization of the interference links provides a larger throughput. The intuition behind this result is that noise is the dominant factor which degrades network throughput in the low-SNR regime, and in this regime it is recommended to have more accurate information about the direct link’s gain to perform rate adaptation more accurately. On the other hand, at the high-SNR regime, interference instead of noise is the dominant phenomenon which degrades network throughput, thus, in this regime it is preferred to have more accurate information about the interference links to more precisely eliminate interference. This coincides with our observations in Figure 4.7 where we 4.4. Digital Feedback 93 9 rS Network throughput [bits/channel use] 8 bC uT 7 rS rS rS NI = 3, NII = 6, EIA-RA NI = 5, NII = 5, EIA-RA NI = 7, NII = 4, EIA-RA rS rS bC bC bC bC uT uT uT uT bC bC 6 rS bC 5 uT uT rS uT 4 bC uT 3 rS uT bC 2 rS uT bC 1 rS uTbC uT 0 rSbC −20 uTrSbC rS −10 0 10 20 30 40 SNR [dB] Figure 4.9: The lower bound on network throughput under different strategies for sharing the total feedback bits (Nf = NI + 2NII = 30) among direct link quantizer and interference link quantizer in a three-user interference network. studied feedback bits allocation in the power control problem. The lower bound on the throughput of K-user (K ∈ {3, 5, 10}) interference networks for different number of feedback bits is shown in Figure 4.10. Also, the performance of the TDMA scheme is shown assuming perfect CSI is available at the sources. This figure shows that the IA-RA scheme can outperform the TDMA scheme even with only limited feedback given that the number of feedback bits is sufficiently large. Specifically, for a given number of users and number of feedback bits, there is a specific SNR below which the EIA-RA scheme outperforms TDMA. This result together with our observation in Figure 4.6 confirm that, in a certain operating region, the interference alignment based transmission schemes (e.g. EIAPC and EIA-RA) can outperform the TDMA scheme even if the number of feedback bits are limited. Delay-Limited Throughput So far, we have studied the cases that the delay required for finding the complement fading blocks in the EIA scheme can be asymptotically long. But in many practical 94 K-user SISO Interference Networks: Pilot-assisted Interference Alignment 20 uT uT bC bC uT uT uT Network throughput [bits/channel use] 18 EIA-RA, N = 3, EIA-RA, N = 3, EIA-RA, N = 3, EIA-RA, N = 6, EIA-RA, N = 6, EIA-RA, N = 6, TDMA, K = 3 TDMA, K = 5 TDMA, K = 7 rS bC 16 uT rS 14 bC 12 uT rS 10 bC uT 8 K K K K K K =3 =5 =7 = 3 uT =5 =7 uT bC bC bC rS bC uT bC bC rS uT bC rS uT bC rS rS rS uT uT bC uT rS rS rS uT bC rS rS uT bC bC rS 6 uT rS uT uT bC 4 rS rS bC uT uT uT uT uT rS bC bC bC bC bC rS rS rS rS rS bC uT uT uT bC uTrSuT bC 2 bCbC 0 −20 bC uTuTbCrS uTuTbCbCrS rS −10 bC rS rS rS bC uT uTbCrS rS rS uTuTrS bC rS 0 10 20 30 40 SNR [dB] Figure 4.10: Throughput of a K-user interference network for different number of feedback bits. applications only limited communication delay can be tolerated. To incorporate this requirement in the considered rate adaptation scheme, in this subsection we consider that terminals can accept only a limited delay in terms of the number of fading blocks needed to perform EIA, denoted by T . Let m and mp to be the block indices of two subsequent complement channels. In the considered transmission scheme, the waiting time for the complement channel to occur may pass the acceptable delay, i.e. mp − m > T . In this case, communication fails and an outage event will be declared. We define delay outage probability as follows o n ˆm , (4.71) PDout,m , Pr mp − m > T |H ˆ m is the network quantized channel matrix at fading block m. The following where H proposition quantifies this probability. Theorem 4.4.4. The delay-limited outage probability is approximately PDout,m ≈ !T √ 2 2 (2hmax /(σ π))K − 12 PK PK |hˆ m | kl l=1 k=1 e σ , 1− 2KNf /2 (4.72) 4.4. Digital Feedback 95 Delay-limited throughput [bits/channel use] 6 N=3 N=4 N=5 N=6 5 4 3 2 1 0 100 101 102 103 104 105 106 107 T Figure 4.11: Delay-limited throughput of a three-user network versus the maximum acceptable delay (NI = NII = N ). where σ 2 is the variance of channel gains. Proof. See Appendix 4.C. Theorem 4.4.4 shows that, for a given T , PDout,m is an increasing function of Nf , and a decreasing function of hmax . This implies that a more accurate channel quantization increases the delay-limited outage probability in (4.71). In addition, it can be observed that limT →∞ PDout,m = 0 which coincides with the intuition. In the following, we define the delay-limited throughput of the network for transmission over blocks m and mp . Definition 4.1. The delay-limited throughput of the network, at blocks m and mp , is defined as follows TDm , 1 − PDout,m × T m , (4.73) where T m and PDout,m are defined in (4.61) and (4.71), respectively. Figure 4.11 shows the numerical expectation of the delay-limited throughput of a three-user network as a function of the maximum acceptable delay (T ) for different number of quantization bits NI = NII = N . In the numerical evaluations, we set 96 K-user SISO Interference Networks: Pilot-assisted Interference Alignment pk /N0 = 1000 [dB] (k ∈ {1, 2, 3}). It can be observed that when the maximum acceptable delay is large, a higher throughput can be achieved by a finer quantization. In contrast, when acceptable delay is limited to small values, coarse quantization should be applied instead to achieve a larger throughput. The reason is that using fine quantizer incurs a larger expected delay for a complement channel to happen which increases delay outage probability defined in (4.71). On the other hand, more accurate quantization leads to less residual interference, and less uncertainty about the actual mutual information value, and consequently outage probability defined in (4.49) decays. Depending on the value of the maximum acceptable delay, one of the mentioned phenomena is dominant. Specifically, in applications with strict delay constraints, outage events due to delay constraints are dominant and it is recommended to use coarse quantization, while in the applications which are robust to delay, residual interference due to the lack of perfect CSI is the main cause of the outage events and a fine quantization is suggested to be deployed instead. Figure 4.12 shows the delay-limited throughput of a K-user interference network as a function of the maximum acceptable delay for different number of users in the network. In the simulations, we set pk /N0 = 1000 [dB] (k ∈ {1, 2, 3}), and NI = NII = N = 3. It can be observed that for large values of T , the throughput increases by increasing the number of users. The trend is different for small values of T : the throughput decays as the number of users increases. This is because for small values of T , the outage events due to delay constraint are the dominant reason of failure in transmission. By increasing the size of the network, users are required to wait longer for the complement channels to happen; consequently, the delay outage probability defined in (4.72) increases and delay-limited throughput decays. 4.5 Summary In this chapter, we have studied SISO interference networks. In the first part of the chapter, we looked at networks where no a priori CSI is available at the terminals. We proposed a pilot-assisted ergodic interference alignment scheme in order to obtain CSI and conduct data communications. We considered pilot-based channel training and investigated both analog and digital feedback cases. When there is analog feedback from destinations to sources, we computed an achievable rate region. Our study reveals that the total DoF Kopt (1−Kopt/T )/2 is achievable when the number of the active users is selected to be Kopt = min{T /2, K}. Thus, it can be recommended that, in large networks (K > T /2), to perform a user selection, and to apply interference alignment only within the set of the active users. In addition, we have derived the optimum problem allocation to the channel training and the data transmission. Our results show that to increase the achievable rate in large networks more power should be allocated to the channel training instead of the data transmission. In the second part of the chapter, we looked at using digital feedback to convey CSI to the transmitters. Here, we however assumed that the receivers have 4.5. Summary 97 Delay-limited throughput [bits/channel use] 1.4 K K K K 1.2 1 =3 =4 =5 =6 0.8 0.6 0.4 0.2 0 100 101 102 103 104 105 106 107 T Figure 4.12: Delay-limited throughput of a K-user interference network versus the maximum acceptable delay. perfect CSI. Two problems have been addressed. First, we have studied a power control problem to compute the minimum required power to guarantee that each source-destination pair can successfully communicate at the desired fixed rate. We proposed a power control algorithm, and investigated its convergence behavior. Next, we have addressed a throughput maximization problem when transmission powers are fixed. We derived an upper bound on the outage probability, and using that we proposed a rate adaptation scheme to maximize the lower bound on network throughput. Finally, we have studied the impact of the constraint on the maximum acceptable delay on the throughput for delay sensitive applications. Our study shows that with a proper power control or rate adaptation, interference alignment can outperform conventional orthogonal transmission schemes, even if only quantized CSI with limited resolution is available at the sources. For a given number of total feedback bits, we have investigated feedback bits allocation to the quantization of direct links or interference links. For the power control scheme, to decrease transmission power, our study reveals that for transmission at high rates more bits should be allocated to interference links; on the other hand, more bits should be allocated to direct links for transmission at low rates. For the rate adaptation scheme, our results recommend to allocate more bits to interference links at the 98 K-user SISO Interference Networks: Pilot-assisted Interference Alignment high-SNR regime, and to allocate more bits to direct links at the low-SNR regime. When there is a maximum acceptable delay, we have shown that if the maximum acceptable delay is large, the throughput increases as the number of feedback bits increase; However, for a small maximum acceptable delay, increasing the number of feedback bits leads to a decrease of throughput. Furthermore, we have seen that when the maximum acceptable delay is large, the throughput increases by increasing the number of users; while the throughput decays by increasing the number of users when the maximum acceptable delay is low. 4.A The Proof of Proposition 4.3.2 The variance in (4.19) can be derived as follows σ2 = = (a) = = (b) = (c) = = i h i2 ∗ h e Y k bk H, e Y k − E Xk − X bk H, b k Xk − X Xk − X i ∗ h e Yk bk H, b k Xk − X E Xk − X ∗ h i e Yk bk H, E Xk · Xk − X h ∗ i e Yk bk H, Pd − E Xk · X i h ˜ kk Xk Y k ∗ H, e Yk E Pd h Pd − 2 hkk Pd 1 + Kσε2 +2 ˜ 2 2 hkk Pd 2 ˜ Pd − 2 hkk Pd 1 + Kσε2 + 2 ˜ E 1+ Pd ˜ kk |2 Pd 2|h (4.74) 1+Kσε2 Pd where (a) follows from the orthogonality of the estimated signal to the estimation bk given in (4.18); error of the MMSE estimator; (b) follows by the substitution of X and (c) follows substituting Y k given in (4.13), and noting that Xk is mutually m independent of Zkm , Zk p and Xl (∀l ∈ {1, 2, ..., K}, l 6= k). 4.B The Proof of Theorem 4.4.3 In this part, we compute the mean and the variance of the random variable Y defined in (4.57). We exploit the property that the quantization error of a uniform quantizer quantizing a Gaussian random variable with variance σ 2 is uniformly distributed with an acceptable approximation when σ/∆ ≥ 1, where ∆ is the quantizer step size [SS77]. The parameter hmin is set to be zero in the throughput 4.B. The Proof of Theorem 4.4.3 99 maximization problem. Therefore, we assume uniform distribution for variables m m m m ], Im[δkl p ] ∼ U (−∆II /2, ∆II /2), where ∆II = hmax /2NII −1 Re[δkl ], Im[δkl p ], Re[δkl for NII bit quantization when the magnitude of the real and imaginary parts of ℜ ℑ the channel gains are limited to hmax . We define random variables gkl and gkl as follows: m m ℑ m ℜ m = Im[δkl + δkl p ] ∀k, l ∈ {1, 2, ..., K}, k 6= l. gkl = Re[δkl + δkl p ] , gkl (4.75) Assuming uniform distribution for the quantization errors, since fX+Y = fX ∗ fY , where ‘*’ is the convolution operation, the pdfs of these variables are given by: fgℜ (x) = fgℑ (x) = kl kl 1 (∆II − |x|) 0 < |x| < ∆II . ∆2II (4.76) The mean and the variance of these random variables are as follows: ℑ ∆2II ℜ ℑ ℜ = = var gkl = 0 , var gkl = E gkl E gkl . 6 (4.77) ℑ Also, we define the random variables sℜ kl and skl as follows: m p m sℜ kl = Re[δkl + δkl ] If Y = g(X), then n(y) fY (y) = X k=1 2 mp 2 m . , sℑ kl = Im[δkl + δkl ] 1/ g ′ (gk−1 (y)) .fX (gk−1 (y)), (4.78) (4.79) where n(y) is the number of solutions in x for the equation g(x) = y, and gk−1 (y) is the kth solution. Exploiting (4.76), the pdfs of these random variables are fsℜ (x) = fsℑ (x) = kl kl 1 1 √ − ∆II x ∆II 2 0 < x < ∆2II . (4.80) Also, the mean and the variance of these random variables are ℑ 7∆4II ℑ ∆2II , var sℜ . E sℜ kl = var skl = kl = E skl = 6 180 (4.81) According to (4.57), (4.81), and (4.77), we have the mean and the variance values which are given in (4.58). 100 K-user SISO Interference Networks: Pilot-assisted Interference Alignment 4.C The Proof of Theorem 4.4.4 The delay outage probability defined in (4.72) can be computed as follows o o n n ˆ m+T 6= H ˆ mp ˆ m+1 6= H ˆ mp , ..., H ˆ m = Pr H Pr mp − m > T |H oT n (a) ˆ m+1 6= H ˆ mp = Pr H (b) = 1− n K i io h n Y h h mi ˆm ˆ Pr Re hm+1 ∈ Re h kk − ∆I /2, Re hkk + ∆I /2 kk k=1 h i io h h mi ˆ ˆm × Pr Im hm+1 ∈ Im h kk − ∆I /2, Im hkk + ∆I /2 kk K Y K n h i io Y h h mi ˆ ˆm Pr Re hm+1 ∈ Re −h × kl − ∆II /2, Re −hkl + ∆II /2 kl k=1 l=1 l6=k h io i h h mi ˆ ˆ m + ∆II /2 ∈ Im − h − ∆ /2, Im − h × Pr Im hm+1 II kl kl kl n !T (4.82) where (a) follows the fact that channels are i.i.d; (b) follows the independence of the elements of channel matrix. We have hm+1 ∼ CN (0, σ 2 ), and hm+1 ∼ CN (0, σ 2 ), kk kl therefore n Pr Re hm+1 kk i io Z h h h mi ˆ m +∆I /2 = ∈ Re ˆ hkk −∆I /2, Re h kk ˆ m ]+∆I /2 Re[h kk ˆ m ]−∆I /2 Re[h kk x2 1 √ e− σ2 dx σ π 2 (Re[hkk ]) ∆I σ2 . ≈ √ e− σ π ˆm Similarly, we can compute the other items. We can conclude that n ˆm Pr mp − m > T |H o ! (K 2 −K) PK PK ˆ m 2 T 1 ∆K ∆ − | h | ≈ 1 − I √II K 2 e σ2 k=1 l=1 kl (σ π) !T √ 2 2 (2hmax /(σ π))K − 12 PK PK |hˆ m | kl l=1 k=1 e σ (4.83) = 1− 2KNf /2 where Nf = 2NI + 2(K − 1)NII . Chapter 5 K-user MIMO Interference Networks: Transceiver Design and Power Control I N the previous chapter we focused on SISO interference networks, where time variations of the channel have been exploited to partially eliminate the interference. In this chapter, we consider MIMO interference networks in which multiple antennas at terminals can be exploited to manage the interference. We aim to address the power control problem for these networks, by further taking into account the achievable DoF of the network. Specifically, we consider a network in which each source sends multiple data streams the number of which is the same as the corresponding DoF achieved by interference alignment. We propose two iterative algorithms that compute transmitter beamforming matrices and receiver filtering matrices to maximize the SINR for each stream, and allocate the minimum powers to realize the desired fixed-rate communications. In both algorithms, the required power values are computed in a distributed fashion at each destination and the associated source is informed via a feedback link. In the first algorithm, the exact value of the computed power is sent, and in the second algorithm only a one-bit feedback signal is transmitted via feedback. The proposed algorithms can provide reliable communication when multiple streams are transmitted each being encoded with potentially different rates. Numerical evaluations show that these algorithms require substantially smaller powers, when compared to the conventional orthogonal transmission strategies. Our test-bed implementation of these algorithms in a network consisting three source-destination pairs is reported in [MFZS14]. The experimental measurements in indoor environment confirm the promised performance of the proposed algorithms. The structure of this chapter is as follows. Section 5.1 presents the considered multi-user MIMO interference network. In Section 5.2, we elaborate on the iterative algorithms to conduct transceiver design and power control. Section 5.3 provides numerical performance evaluations. Finally, Section 5.5 summarizes this chapter. 101 102 K-user MIMO Interference Networks: Transceiver Design and Power Control mdkk Ekdk cdkk Pkdk m1k Ek1 q pdkk cdkk .. . .. . Sk xdkk = c1k Pk1 nS Vk p x1k = p1k c1k nD xk k yk k .. . .. . x1k yk1 y dkk Dkdk m ˆ dkk .. . Uk y 1k Dk1 Dk m ˆ 1k Figure 5.1: The structure of transmitter and receiver corresponding to the sourcedestination pair Sk − Dk . 5.1 Multi-user MIMO Interference Network We consider a MIMO interference network consisting of K source-destination pairs. The sources and the destinations are denoted as Sk and Dk (k ∈ {1, 2, ..., K}), and they are equipped with nSk and nD k antennas, respectively. Each source intends to communicate to the corresponding destination. The structure of transmitter and receiver corresponding to the source-destination pair Sk −Dk is shown in Figure 5.1, and will be described in more details in what follows. 5.1.1 Transmitter Structure The source Sk sends dk independent messages mdk (d ∈ {1, ..., dk }). The value of dk is selected according to the DoF region of the network characterized in [CJ08,GJ10, GCJ11] 1 . The encoder Ekd encodes mdk to a unit-power codeword cdk selected from d d a Gaussian codebook q with code rate Rk . Next, the power controller Pk scales this codeword to xdk = pdk cdk , where pdk is the power of transmitted signal and satisfies h iT a maximum power constraint, i.e. pdk < pmax . The dk × 1 vector xk = x1k , ..., xdkk denotes these scaled codewords. The transmitted signal of Sk is xk = Vk xk , h (5.1) i where Vk = v1k , ..., vdkk is the nSk × dk beamforming matrix, and vdk denotes the unit-norm beamforming vector corresponding to the dth transmitted codeword. 1 In this chapter, the number of streams sent by the sources is selected as the DoF value that can be maximally achieved through linear interference alignment. In fact, this value is unknown in the case of general MIMO interference channels. However, due to the rapid development of the interference alignment concept in recent years, the maximal achievable DoF for several network structures has been discovered. For instance, in three-user MIMO interference channels with M antennas at each terminal the available DoF has been characterized in [CJ08], and subsequently that of K-user interference channels in which each transmitter has M antennas and each receiver has N antennas has been characterized in [GJ10]. For networks with an arbitrary number of antennas at each terminal, the necessary condition for the achievability of a tuple of DoF has been discussed in [YTJK10], and the achievability can be numerically checked using the method proposed in [GCJ11]. The results presented in these references can be used to determine the number of streams sent from each source. 5.2. Transceiver Design and Power Control 5.1.2 103 Receiver Structure The received signal at the destination Dk is yk = Hkk Vk xk + K X Hkl Vl xl + zk , (5.2) l=1,l6=k where Hkl is the channel matrix corresponding to the link from Sl to Dk , and zk is additive noise. Throughout this chapter, the link from Sk to Dk is called desired link and those from Sl to Dk (l 6= k) are referred to as interference links. The channel gains are constant during the transmission of each codeword, but independently change across time. The channel gains follow a complex Gaussian distribution, i.e. hkl ∼ CN (0, σS2 InS ×nD ) where InS ×nD is a nSk × nD × nSk × nD identity matrix, l l k l k l and hkl denotes vector representation of the elements of matrix Hkl . The noise has complex Gaussian distribution, i.e. zk ∼ CN (0, N0 InD ) in which N0 is noise k D power at each destination, and InD is an nD k × nk identity matrix. Let Uk denote an k d nD k ×dk receiver filtering matrix with unit-norm column vectors uk (d ∈ {1, ..., dk }). h iT The filter output of Dk , yk = y 1k , ..., ydkk , is yk = U∗k yk , (5.3) where A∗ denotes the conjugate transpose of matrix A. When the sources are non-orthogonally activated, the filter output of Dk generally contains interference, i.e., yk = U∗k Hkk Vk xk + K X U∗k Hkl Vl xl + U∗k zk . (5.4) l=1,l6=k The decoder Dkd (d ∈ {1, ..., dk }) decodes the received signal y dk to an estimate of the transmitted message m ˆ dk by treating the remaining interference as noise. In the considered network, it is desired that source Sk reliably communicates dk independent streams with its intended destination Dk while stream d (d ∈ {1, ..., dk }) is encoded at a given fixed rate Rkd . This requires Vk , Uk , and pdk , to be properly designed. 5.2 Transceiver Design and Power Control In this section, we address transceiver design and power control which are performed before the actual data transmission starts. The transceiver design and power control are performed in an iterative fashion and each iteration occurs during one training slot. There are N training slots, and within each slot different tasks are performed as depicted in Figure 5.2. The quantity N is a design parameter and the accuracy of system design can be improved by setting N with a larger value. The computed 104 K-user MIMO Interference Networks: Transceiver Design and Power Control Training slot 1 forward training ... receivers optimization Training slot n power updating ... reverse training Training slot N transmitters optimization Figure 5.2: CSI acquisition, transceiver design, and power control. beamforming vector, filtering vector, and power corresponding to the lth stream of the source-destination pair Sk − Dk at the nth training slot are denoted as vlk (n), ulk (n), and plk (n), respectively. In this chapter, the term CSI is used to denote channel knowledge in the sense of the following discussion. To compute receiver filtering matrix and transmission q d power, in training slot n, the destination Dk needs to know pj (n − 1)Hkj vdj (n−1) (j ∈ {1, ..., K}, d ∈ {1, ..., dj }). These can be obtained using training sequences sent by the sources over forward channels (channels from sources to destinations). Similarly, to compute transmitter beamforming matrix, in training slot n, the source q Sk requires to know pdFj Hrkj udj (n) (j ∈ {1, ..., K}, d ∈ {1, ..., dj }), where Hrkj de- notes the channel matrix from Dj to Sk . These can be obtained at each source using training sequences transmitted over reverse channels (channels from destinations to sources). We assume that reverse channels are separated from forward channels in time via time-division duplexing (TDD), and the reciprocity assumption holds, i.e., Hrkl = H∗lk (∀l, k ∈ {1, 2, ..., K}). This is the channel knowledge required to conduct the transceiver design and power control. Since each terminal (source/destination) needs to know the knowledge related to local links only, we refer to this channel knowledge as local CSI in this chapter. 5.2.1 CSI Acquisition, Transceiver Design, and Power Control In the following, we briefly present the process of iterative CSI acquisition, transceiver design, and power control. Before the first training slot starts, Sk (k ∈ {1, 2, ..., K}) initializes power plk (0) = pmax (l ∈ {1, ..., dk }) and randomly selects unit-norm beamforming vector vlk (0). The nth (n ∈ {1, 2, ..., N }) training slot is composed of the following phases: 1) Forward training phase: During this phase, the destination Dk estimates 5.2. Transceiver Design and Power Control 105 interference-plus-noise covariance matrix Qlk (n) , dj K X X ∗ ∗ pdj (n − 1)Hkj vdj (n − 1) vdj (n − 1) (Hkj ) j=1 d=1 ∗ ∗ −plk (n − 1)Hkk vlk (n − 1) vlk (n − 1) (Hkk ) + N0 InD , k (5.5) and effective desired channel matrix l Hkk (n) , Hkk vlk (n − 1) (5.6) corresponding to the lth (l ∈ {1, ..., dk }) stream. To obtain these matrices, an orthogonal training PK scheme can be deployed. Let the length of the forward training phase be L × k=1 dk . Each source subsequently sends L pilot symbols for each data stream modulated with the corresponding beamforming vector. The other sources remain silent during thePpilot transmission of one source. Pj−1Specifically, the j−1 source S at time slot t = L d + (d − 1)L + 1, ..., L j i i=1 i=1 di + dL sends q pdj (n − 1)vdj (n − 1), where d ∈ {1, ..., dj }. The destination Dk receives yk (t) = q pdj (n − 1)Hkj vdj (n − 1) + zk (t), (5.7) where zk (t) ∼ CN (0, N0 InD ) is the receiver noise. If L ≫ 1, then according to the law of large numbers we have Pj−1 L d +dL q i=1 Xi 1 yk (t) ≈ pdj (n − 1)Hkj vdj (n − 1). (5.8) L Pj−1 t=L i=1 di +(d−1)L+1 Since the destination Dk (k ∈ {1, ..., K}) knows its noise power N0 , it can use the estimated quantities in (5.8) for different values of j ∈ {1, ..., K} and d ∈ {1, ..., dj } to compute the interference-plus-noise covariance matrix given in (5.5). Also, since Dk knows the transmitted power pdk (n − 1) of the corresponding source (in Section III.B.5 we will propose algorithms to allow each destination to compute the required power for the corresponding source, and inform the source about the computed value via a feedback signal), it can compute the effective desired channel matrix in (5.6) using the estimated values in (5.8). In [MFZS14], a similar channel training approach is used in a test-bed implementation of Algorithm 2 proposed in Section 5.2.2 and demonstrates good performance. Also, we have investigated the impact of channel estimation errors on system performance in [FZF+ 13, FKWS13]. Since channel training is not the concern of this chapter, we assume that perfect estimation of these matrices are available at terminals. The interested readers are referred to [KRB+ 13] for more thorough discussion on channel training design for MIMO systems. 106 K-user MIMO Interference Networks: Transceiver Design and Power Control 2) Receivers optimization phase: Using the interference-plus-noise covariance matrix in (5.5) and the effective desired channel matrix in (5.6) estimated in the forward training phase, the receiver filtering vector ulk (n), similar to the one for Max-SINR algorithm in [GCJ11], is selected to be an MMSE filter as follows ulk (n) = (Qlk (n))−1 Hkk vlk (n − 1) . k(Qlk (n))−1 Hkk vlk (n − 1)k2 (5.9) This filter maximizes the received SINR of the lth stream at destination Dk denoted as SINRlk . 3) Power updating phase: In this phase, the power values are updated using a distributed power control scheme that will be described in Section 5.2.2. In Section 5.2.2, we will first consider an iterative method. In each iteration, each destination Dk computes the required transmission powers of Sk , and then inform Sk about these values via feedback signals. We will prove the convergence of this approach. Since conducting an iterative scheme in the power updating phase may increase system complexity and communication delay, we next simplify the process by considering two methods that update powers only once, i.e., no iteration is performed. In the first method, each destination sends back the exact values to which the source transmission powers should be updated. In the second method, each destination uses only a one-bit feedback signal associated with each stream to inform the corresponding source to scale the transmission power by a certain factor. The convergence behaviour of the overall iterative process, considering the transceiver design and power control, can be confirmed by simulations shown in Section 5.3. 4) Reverse training phase: To acquire CSI at the sources, the destinations orthogonally broadcast training sequences. The destination Dk beamforms its training sequences with a fixed power pF uniformly allocated to different sequences, using r an nD k × dk matrix Vk (n − 1) = Uk (n). The source Sk (k ∈ {1, 2, ..., K}) estimates the reverse interference-plus-noise covariance matrix corresponding to the lth data stream Qr,l k (n) = dj K X X pF j=1 d=1 − dj ∗ r ∗ r,d Hrkj vr,d j (n − 1) vj (n − 1) (Hkj ) ∗ r ∗ pF r r,l Hkk vk (n − 1) vr,l , k (n − 1) (Hkk ) + N0 InS k dk (5.10) r where vr,d j (n − 1) denotes the dth column of matrix Vj (n − 1). In addition, Sk estimates the effective reverse desired channel matrix corresponding to the lth stream of the kth source-destination pair: r,l Hkk (n − 1) = Hrkk vr,l k (n − 1). (5.11) The similar channel training scheme as the one mentioned in the forward training phase can be deployed for reverse channel training as well to acquire Qr,l k (n) and r,l Hkk (n − 1). 5.2. Transceiver Design and Power Control 107 5) Transmitters optimization phase: In the reverse training phase, the source Sk applies an nSk × dk filtering matrix Urk (n) to its received signal. Similar to the receivers optimization phase, ur,l k (n) (∀l ∈ {1, ..., dk }) is chosen as follows to maximize the received SINR of the lth stream received at Sk ur,l k (n) = −1 (Qr,l Hkk vlk (n − 1) k (n)) −1 H vl (n − 1)k k(Qr,l kk k 2 k (n)) . (5.12) The source can compute these vectors using the effective reverse desired channel matrix in (5.11) and the interference-plus-noise covariance matrices in (5.10) estimated in the reverse training phase. Next, the source Sk sets Vk (n) = Urk (n) as its updated beamforming matrix in this training slot. An illustrative summary of the five phases above is shown in Figure 5.2. The whole process repeats in the subsequent training slots until the whole training is completed (i.e., after N training slots). In the following subsection, we will explain the principles of the distributed power control scheme, performed during the power update phase, in more details. 5.2.2 Distributed Power Control In the power updating phase of the nth training slot, the transmitter beamforming vectors and receiver filtering vectors at Sk and Dk are vlk (n − 1) and ulk (n) (k ∈ {1, 2, ..., K}, l ∈ {1, 2, ..., dk }), respectively. For the simplicity of presentation, we omit the slot index and let ulk , vlk , and plk denote ulk (n), vlk (n−1), and plk (n−1), respectively. Then, the SINR of the lth stream at Dk is 2 l ∗ uk Hkk vlk plk l , (5.13) SINRk = ϕlk (p) + N0 where ϕlk (p) = dj K X 2 2 X ∗ l ∗ uk Hkj vdj pdj − ulk Hkk vlk plk , (5.14) j=1 d=1 iT h P K and p = p11 , ..., pd11 , ..., p1K , ..., pdKK is a k=1 dk × 1 power vector. The mutual information corresponding to the lth stream of the source-destination pair Sk − Dk is log2 1 + SINRlk . To guarantee successful transmission, the following condition should be satisfied: log2 1 + SINRlk ≥ Rkl . (5.15) According to (5.13), (5.15) can be rewritten as a power constraint in the following format: plk ≥ Ikl (p), (5.16) 108 K-user MIMO Interference Networks: Transceiver Design and Power Control where l 2Rk − 1 ϕlk (p) + N0 . Ikl (p) = 2 l ∗ uk Hkk vlk (5.17) Therefore, all power constraints can be represented as p I(p), (5.18) where the operator denotes an element-wise vector inequality, and h iT dK 1 I(p) , I11 (p), ..., I1d1 (p), ..., IK (p), ..., IK (p) (5.19) is called interference function. Definition 5.1. For given transmitter beamforming matrices and receiver filtering matrices, the set of the positive power vectors, which satisfies (5.18) and the power constraint, is defined PH = {p|0 p pmax , p I(p)}, (5.20) where I(p) is the interference function defined in (5.19), and pmax is the maximum transmission power vector whose elements are set to be pmax . Therefore, the power control problem can be formulated as the following optimization problem: min s.t. p∈PH dk K X X plk . (5.21) k=1 l=1 In the remaining parts, we first discuss the existence of a feasible solution for this problem. Next, we present an iterative algorithm to find the solution of this problem. Finally, we study the convergence of the presented algorithm. Feasibility of the Power Control Problem The power control problem in (5.21) has a feasible solution when the set PH is nonempty. We can rewrite the condition in (5.18) as a linear power constraint in matrix form as follows: p DFp + n, (5.22) K matrix F is the normalized gain matrix. Its where the k=1 dk k=1 dk × P P k−1 j−1 elements on the m=1 dm + l th row and the m=1 dm + d th column (k, j ∈ P P K 5.2. Transceiver Design and Power Control 109 {1, ..., K}, l ∈ {1, ..., dk }, d ∈ {1, ..., dj }) are F k−1 X dm + l, j−1 X ! dm + d m=1 m=1 |(ul )∗ H vd |2 kj j k |(ul )∗ Hkk vl |2 k = j 6= k k 2 |(ulk )∗ Hkk vd k| l )∗ H l |2 |(u v k kk k j = k, d 6= l . (5.23) 0 j = k, d = l P K The k=1 dk matrix D is a diagonal matrix and its elements on the k=1 dk × P P k−1 j−1 d + l th row and the d + d th column (k, j ∈ {1, ..., K}, l ∈ m m m=1 m=1 P K {1, ..., dk }, d ∈ {1, ..., dj }) are D k−1 X dm + l, m=1 m=1 The P k−1 m=1 j−1 X ! dm + d = ( l 2Rk − 1 j = k, l = d 0 otherwise . (5.24) dm + l th element (k ∈ {1, ..., K}, l ∈ {1, ..., dk }) of the vector n is n k−1 X m=1 dm + l, 1 ! = l 2Rk − 1 N0 |(ulk )∗ Hkk vlk |2 . (5.25) Therefore, we can represent the set PH defined in (5.20) as follows: PH = {p|0 p pmax , p DFp + n}. (5.26) Since this set depends on the channel matrices, the maximum transmission powers, the transmission rates, the transmitter beamforming matrices, and the receiver filtering matrices, all these parameters affect the feasibility of the problem in (5.21). The following theorem characterizes the condition under which the power control problem is feasible. Theorem 5.2.1. Assume that the transmitter beamforming matrices and the receiver filtering matrices for the scheme considered in Section 5.2.1 are given as Vk and Uk (∀k ∈ {1, 2, ..., K}), and the maximum transmission power is asymptotically large (i.e., pmax → ∞). The power control problem in (5.21) is feasible (PH 6= ∅) if and only if λmax (DF) < 1, where λmax (DF) is the Perron-Frobenius eigenvalue of the matrix DF, and the matrices D and F are defined in (5.23) and (5.24), respectively. Proof. See Appendix 5.A. For given channel matrices, beamforming matrices, and receiver filtering matrices, the condition given in Theorem 5.2.1 determines whether the power control problem has feasible solutions. For random channel matrices, with a certain probability, the power control problem has solutions. This probability is referred to as 110 K-user MIMO Interference Networks: Transceiver Design and Power Control uTbCrS 100 uTbCrS rSbC uT rS uTbC rS rS bC uT rS bC uT rS bC rS uT bC rS bC uT rS Feasibility Probability 10−1 bC rS uT bC uT 10−2 bC M M M M rS 10−3 bC uT =2 =4 =6 =8 uT uT 10−4 0 2 4 6 8 10 12 R [bits/channel use] Figure 5.3: Feasibility probability versus transmission rate in a three-user interference network with M antennas at each terminal. feasibility probability. In wireless communication systems, the feasibility probability should be larger than a certain threshold. The value of this threshold depends on the reliability requirements of the target application. In sensitive applications such as industrial control systems, this threshold can be high; while for applications such as voice transmission, a relatively low threshold can be accepted. Figure 5.3 shows the asymptotic feasibility probability, when pmax → ∞, versus transmission rate for different number of antennas at the terminals. We consider a three-user interference network and assume that each terminal is equipped with M antennas (M ∈ {2, 4, 6, 8}). The achievable DoF of each source-destination pair is M/2, and each source intends to send M/2 independent data streams each with rate R. It can be observed that as the number of antennas increases the feasibility probability decays. The reason for this observation is that in this simulation set up, as the number of antennas increases we also increase the number of transmitted streams in each source-destination link. This leads to a larger inter-stream interference and consequently a lower feasibility probability. Remark 5.1. The necessary and sufficient condition for the feasibility of the considered power control problem given in Theorem 5.2.1 is valid in the asymptoti- 5.2. Transceiver Design and Power Control 111 cally high-SNR regime. If the power budget is finite, the condition λmax (DF) < 1 is only a sufficient condition but not a necessary condition. In other words, for finite values of pmax , if the considered power control problem is feasible, then λmax (DF) < 1. However, the reverse statement does not apply. Figure 5.4 shows the feasibility probability versus rate for different values of pmax . In this example, we consider a three-user interference network where each terminal is equipped with two antennas, and each source sends one data stream. It can be observed that for a fixed transmission rate, when the power budget decreases the feasibility probability decreases. The feasibility probability under unlimited power budget is the same as the probability that the condition λmax (DF) < 1 is satisfied, and serves as the upper bound of the case under finite power limit. Clearly, if the power budget level is relatively large compared to the transmission rate demand, the feasibility probability is close to the upper bound. The effect of limited power budget is small. Thus if a system has a large power budget, with a high probability the condition given by Theorem 5.2.1 can be used to correctly predict whether it is possible to guarantee successful communication at the demanded fixed rate. Note that a large power budget does not necessarily mean a large power consumption. The convergence behavior of our proposed algorithms shows that the lowest level of power consumption can be reached, which exhibits the advantages of the proposed algorithms. If the power control problem has feasible solutions, the next question to be answered is how to find the solution corresponding to the minimum required transmission power. In the next subsection, we present an iterative power control scheme to address this question. Iterative Power Control Consider an iterative power control where the maximum number of iterations is set to be T . Recall that the transmitter beamforming matrices and the receiver filtering matrices are Vk and Uk (k ∈ {1, ..., K}), respectively. We use plk [i] (l ∈ {1, ..., dk }) to denote the computed power at the ith iteration. In training slot n, before the first iteration starts, Sk sets its initial power values plk [0] = plk (n − 1). At the ith iteration (i = 1, 2, ..., T ), Dk computes Ikl (p[i − 1]) using (5.17). Next, it sends these values to Sk . The source then updates its power values as follows: plk [i] = min{Ikl (p[i − 1]) , pmax }. (5.27) After the last iteration, the source Sk updates the power of the lth data stream at the nth training slot as plk (n) = plk [T ]. Convergence The iterative power control algorithm in (5.27) generates a sequence of vectors p[1], p[2], ..., p[T ]. In this part, we answer two questions regarding the convergence of this sequence. The first is whether this sequence converges to a fixed-point p∗ 112 K-user MIMO Interference Networks: Transceiver Design and Power Control *uTrSbC 100 *uTrSbC uTrSbC * uTrSbC uTrSbC * bCuTrS * rSbC uT rSbC rSbC bC bC rS uT bC rS uT * rS uT Feasibility Probability * 10−1 uT * bC 10−2 rS uT * pmax /N0 pmax /N0 pmax /N0 pmax /N0 pmax /N0 →∞ = 50 dB = 40 dB = 30 dB = 20 dB * uT 10−3 0 2 4 6 8 10 12 R [bits/channel use] Figure 5.4: Feasibility probability versus transmission rate in a three-user interference network with two antennas at each terminal and maximum transmission power pmax . (i.e., p∗ = I(p∗ )) for sufficiently large T . If yes, the second question is whether fixed-point p∗ reflects the minimum required transmission powers. The answers to these questions are summarized in the following theorem. Theorem 5.2.2. If the problem in (5.21) is feasible (PH 6= ∅), for any initial power vector p[0], the recursive equation in (5.27) generates a sequence of vectors which converges to a unique fixed-point p∗ . The fixed-point p∗ corresponds to the solution of the problem in (5.21). Thus the proposed iterative power control scheme in the power updating phase can attain the solution of the power control problem (5.21). Proof. See Appendix 5.B. In the distributed implementation of the iterative power control scheme, the corresponding convergence rate directly affects the radio resources required for training. Therefore, in the next part we study the convergence rate of the above scheme. 5.2. Transceiver Design and Power Control 113 Convergence Rate In this part, we find the convergence rate of the iterative power control scheme. This can be used to determine the minimum number of iterations which is required to have the computed powers sufficiently close to the solution of the power control problem. For this propose, we investigate geometrical convergence defined as follows [CHLT08]. Definition 5.2. A sequence a[1], a[2], ... is said to geometrically converge to a∗ at rate α (a smaller rate is corresponding to faster convergence) if there exist nonnegative constants A and α (α < 1) such that ka[n] − a∗ k ≤ Aαn ∀n ∈ {1, 2, ...}, (5.28) where kxk denotes a general norm of the vector x. Let weighted maximum norm of a vector x = [x1 , ..., xm ]T be defined as follows: xj , kxkw = max (5.29) ∞ j wj where w = [w1 , ..., wm ] is an element-wise positive vector. This definition induces a matrix norm [HJ85] kAkw ∞ = max x6=0 kAxkw ∞ . kxkw ∞ (5.30) The following lemma characterizes the connection between the weighted maximum norm and the Perron-Frobenius norm of a non-negative square matrix. Lemma 5.1. [BT89, Corollary 6.1] For a non-negative square matrix A, there exists a vector w 0 such that kAkw ∞ < 1 if and only if λmax (A) < 1. Theorem 5.2.3. If λmax (DF) < 1 and pmax is asymptotically large, then the convergence rate of the power control algorithm in (5.27) with respect to the weighted maximum norm defined in (5.29) is α = kDFkw ∞. Proof. kp[i] − p∗ kw ∞ = = = (a) ≤ ≤ kI(p[i − 1]) − I(p∗ )kw ∞ k(DFp[i − 1] + n) − (DFp∗ + n)kw ∞ kDF(p[i − 1] − p∗ )kw ∞ ∗ w kDFkw ∞ kp[i − 1] − p k∞ i ∗ w (kDFkw ∞ ) kp[0] − p k∞ , (5.31) where (a) follows from Theorem 5.6.2 in [HJ85]. Since λmax (DF) < 1, according to Lemma 5.1 there exists a vector w allowing us to have kDFkw ∞ < 1. Consequently, 114 K-user MIMO Interference Networks: Transceiver Design and Power Control the upper-bound on kp[i] − p∗ kw ∞ converges to zero as i → ∞. This result together with the fact that norm of a vector is always non-negative implies the convergence of p[i] to p∗ . The convergence rate of the scheme according to Definition 5.28 is α = kDFkw ∞. If the matrix DF is irreducible (which it is likely to be since the channel matrices are random), using Proposition 6.6 in [BT89] it can be shown that for the positive right Perron-Frobenius eigenvector of DF denoted as v we have kDFkv∞=λmax (DF), and thus the convergence rate is α = λmax (DF). In the case that all transmission rates are the same, i.e., Rkl = R (k ∈ {1, 2, ..., K}, l ∈ {1, 2, ..., dk }), the convergence rate is α = 2R − 1 λmax (F). (5.32) Using the bounds on Perron-Frobenius eigenvalue given in [HJ85], the convergence rate can be bounded as follows αmin < α < αmax , (5.33) where αmin αmax 2 dj ul ∗ H vd K X k X kj j = (2R − 1) min min 2 − 1, ∗ k l l j=1 d=1 ul H v kk k k 2 dj ul ∗ H vd K XX kj j k = (2R − 1) max max 2 − 1. ∗ k l l j=1 d=1 ul H v kk k k (5.34) It can be seen that by decreasing the transmission rate the convergence rate exponentially decreases and consequently the sequence converges faster. This result can be used to find the minimum number of iterations required to have the calculated powers sufficiently close to the solution of power control problem (p∗ ). It can be shown that if T ≥ log(ǫ)/ log(λmax (DF)), where ǫ < 1 is a constant, then kp[T ] − p∗ kv∞ /kp[0] − p∗ kv∞ ≤ ǫ. Therefore, the minimum number of required iterations (Tmin ) to have kp[Tmin ] − p∗ kv∞ /kp[0] − p∗ kv∞ ≤ ǫ is Tmin = log(ǫ)/ log 2R − 1 λmax (F) . (5.35) Clearly, Tmin decreases by decreasing transmission rate R. Iterative Transceiver Design and Power Control The presented power control scheme requires iterative computations to be performed during the power updating phase of each training slot. In each iteration, 5.2. Transceiver Design and Power Control 115 Algorithm 2 Transceiver Design and Power Control Initialize: Vk (0), plk (0) = pmax (k ∈ {1, ..., K}, l ∈ {1, 2, ..., dk }). n=1 repeat Update receiver filtering matrix: ulk (n) = (Qlk (n))−1 Hkk vlk (n − 1) k(Qlk (n))−1 Hkk vlk (n − 1)k2 where Qlk (n) given in (5.5) is obtained in the forward training phase. Update transmission power: l ∗ 2Rk − 1 ulk (n) Qlk (n)ulk (n) . , p plk (n) = min max ∗ | ulk (n) Hkk vlk (n − 1)|2 Set reverse beamforming matrix: Vrk (n − 1) = Uk (n). Update transmitter beamforming matrix: ur,l k (n) = −1 r (Qr,l Hkk vr,l k (n)) k (n − 1) −1 Hr vr,l (n − 1)k k(Qr,l 2 k (n)) kk k , where Qr,l k (n) given in (5.10) is obtained in the reverse training phase. Set beamforming matrix: Vk (n) = Urk (n). Increase the training slot index n = n + 1. until n = N updated interference-plus-noise covariance matrices, effective desired channel matrices, reverse interference-plus-noise covariance matrices, and effective reverse desired channel matrices need to be acquired, in order to update power. This can increase the complexity of the transmission scheme. Nevertheless, this scheme can be simplified such that during the power updating phase, each source updates its transmission powers only once (the number of iterations set to be one). Although we do not prove the convergence of the new approach, it can be confirmed by extensive numerical evaluations, as shown in Section 5.3. Now, we present the whole process of the iterative transceiver design and power control in each training slot. As mentioned above, there are five phases. The first two and the last two phases are dedicated for distributed design of transmitter beamforming and receiver filtering, and the third one is used to find the minimum powers to support successful communication. The complete procedure is conducted in an iterative fashion across the N training slots. Two algorithms are proposed, shown in Algorithm 2 and Algorithm 3. Both of them set the iteration number 116 K-user MIMO Interference Networks: Transceiver Design and Power Control Algorithm 3 Transceiver Design and Power Control: one-bit feedback signal for each power value Initialize: Vk (0), plk (0) = pmax (k ∈ {1, ..., K}, l ∈ {1, 2, ..., dk }). Set 0 < γ < 1, n = 1. repeat Update receiver filtering matrix: ulk (n) = (Qlk (n))−1 Hkk vlk (n − 1) , k(Qlk (n))−1 Hkk vlk (n − 1)k2 where Qlk (n) given in (5.5) is obtained in the forward training phase. Update transmission l power: ∗ R 2 k −1 ((ulk (n)) Qlk (n)ulk (n)) l if pk (n − 1) < , then |(ulk (n))∗ Hkk vlk (n−1) |2 plk (n) = min γ −1 plk (n − 1), pmax else plk (n) = min γplk (n − 1), pmax end if Set the reverse beamforming matrix: Vrk (n − 1) = Uk (n). Update transmitter beamforming matrix: ur,l k (n) = −1 (Qr,l Hkk vlk (n − 1) k (n)) −1 H vl (n − 1)k k(Qr,l kk k 2 k (n)) , where Qr,l k (n) given in (5.10) is obtained in the reverse training phase. Set the beamforming matrix: Vk (n) = Urk (n). Increase the training slot index: n = n + 1. until n = N of power control in each power updating phase as one. The difference between them lies in different requirements regarding the feedback channels which are used to send the computed power values from each destination to the corresponding source. In Algorithm 2, during a training slot each destination first updates its receiver filtering matrix and computes the required transmission powers of the corresponding source. Next, it feeds the values of the computed powers back to the corresponding source. Each source updates its transmission powers accordingly. This algorithm requires a perfect feedback channel from each destination to the corresponding source. A simplified updating procedure is proposed in Algorithm 3, and requires only a one-bit feedback signal corresponding to each data stream. Thus, each destination informs the corresponding source to either increase or decrease the transmission powers. The updated power is the scaled version of the power in the previous training slot. Depending on the value of feedback signal, the scaling factor is either γ 5.2. Transceiver Design and Power Control 117 or γ −1 . The design parameter γ (0 < γ < 1) controls the convergence speed of the algorithm and the accuracy of the computed solution. On one hand, decreasing this parameter increases the convergence speed of the algorithm. On the other hand, decreasing this parameter to small values may lead to large fluctuations in the computed powers and consequently large fluctuations of mutual information around the desired rates in the steady state (i.e. when the training slot index n is large). Remark 5.2. Although in this chapter the number of streams delivered between the sources and destinations is selected to be the maximal DoF value of the networks, the system model is not limited by this assumption. The proposed algorithms can be readily applied to scenarios in which the number of streams is smaller than the available DoF (i.e. any value within the achievable DoF region). Complexity Analysis To investigate the complexity of the proposed algorithms, we follow reference [SYY12] and use the number of complex multiplication operations as the complexity criterion to evaluate the proposed algorithms. For the sake of simplicity we consider D S S a symmetric network where dk = d, nD k = n , and nk = n , ∀k ∈ {1, ..., K}. In other words, d codeword streams are delivered between each source-destination pair. The sources have the same number of nS antennas and the destinations have the same number of nD antennas. During the nth training slot, the main computation operations that each destination performs are listed in the following: • The destination Dk needs to compute Qlk (n) q for each l ∈ {1, ..., d}. AsD sume that Dk estimates the n × 1 matrix plj (n − 1)Hkj vlj (n − 1) using the method presented in Section III.A. Then, it can compute matrix ∗ ∗ plj (n − 1)Hkj vlj (n − 1) vlj (n − 1) (Hkj ) (5.36) 2 by a matrix multiplication of nD ×1 and 1×nD matrices, with nD complexvalued multiplications. The interference-plus-noise covariance matrix corresponding to the lth stream can be computed as follows ∗ ∗ Qlk (n) = Qk (n) − plk (n − 1)Hkk vlk (n − 1) vlk (n − 1) (Hkk ) , (5.37) where Qk (n) , K X d X j=1 l=1 ∗ ∗ plj (n − 1)Hkj vlj (n − 1) vlj (n − 1) (Hkj ) + N0 InD . (5.38) Therefore, the total complexity (i.e. the number of multiplications) of computing the interference-plus-noise covariance matrices at each destination is 2 equal to that required for the computation of Qk (n), which is Kd nD . 118 K-user MIMO Interference Networks: Transceiver Design and Power Control • To compute the receiver vector ulk (n) corresponding to the lth stream, the des−1 tination needs to perform multiplication of the nD ×nD matrix (Qk (n)) and 2 the nD ×1 matrix Hkk vlk (n−1) with nD multiplication operations, one ma 3 trix inverse of complexity nD and one norm computation of nD multiplica 2 3 tion operations. Each destination needs to perform d nD + nD + nD complex-valued multiplications. • To compute power corresponding to each stream using the computed nD × nD interference-plus-noise covariance matrix Qlk (n) and the computed nD × 1 ∗ receiver filter ulk (n), the destination Dk can compute ulk (n) Qlk (n)ulk (n) 2 by performing nD + nD multiplication operations. Using the estimated nD × 1 effective desired channel matrix Hkk vlk (n − 1), and the computed ul (n) ∗ Hkk vl (n − nD × 1 receiver filter ulk (n), the receiver can compute k k 2 2 1) using nD + 1 multiplications. In total d nD + 2nD + 1 multiplication operations is needed to compute power values at each destination. 3 2 In total, d nD + (K + 2) nD + 3nD + 1 complex-valued multiplication are performed at each destination. Similarly, using the fact that there is no power control in the reverse training phase it can be shown that the totalnumber of multi3 2 plications needed at each source is d nS + (K + 1) nS + nS . Comparing the number of multiplications needed for power updating with the total multiplication required for each source-destination link, it is clear that the power control has low overhead to the complexity of the transmitter and receiver design. 5.3 Performance Evaluation In this section, we numerically evaluate the performance of Algorithm 2 and Algorithm 3. We also provide simulation results for TDMA to compare its performance with that of the proposed algorithms. In the orthogonal scheme, the sources are individually activated to avoid interference. Consequently, the transmitter beamforming matrix design and the receiver filtering matrix design for each source-destination pair is simplified to that for point-to-point MIMO systems. Thus, the solution of the power control algorithm for each interference-free channel can be computed by the water-filling approach. We consider two representative examples to evaluate the performance of the proposed algorithms. In the first example, we consider a three-user network in which each terminal is equipped with two antennas. When the orthogonal transmission is performed, only one source is active at a time, and the active source transmits two independent data streams. However, it has been known that the degrees of freedom (d1 , d2 , d3 ) = (1, 1, 1) is achievable in this network [CJ08]. Therefore, in our proposed algorithms, all sources are concurrently activated while each of them trans- 5.3. Performance Evaluation 119 4 bC rS bC bC 3.5 rS bC Mutual Information rSrS bC rS bC 3 user 3 bC rSbC bC bC bC bC rS rSbC rS rSbC rSbC rSbC rS bC rS 2.5 bC rS rS rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC bC rS user 2 bC rS bC 2 rSbC bC rS rSbC rS rSbC rSbC rSbC rSbC rSbC rSbC rSbC rS bC 1.5 rS user 1 bC rS 1 0.5 2 4 6 rS bC rS 8 10 rSbC 12 rSbC 14 16 18 20 Number of training slots Figure 5.5: Mutual information of source-destination pair Sk − Dk (Ik ) versus the number of training slots using Algorithm 2 (−) and Algorithm 3 with γ = 0.5 (◦) and γ = 0.4 (). mits one data stream. The rates are fixed as (R11 , R21 , R31 ) = (1, 2, 3) [bits/channel use], and plk (0)/N0 = pmax /N0 = 20 [dB]. The number of different channel realizations is set to 108 , and we evaluate Algorithm 3 for γ = 0.5 and γ = 0.4. Figure 5.5 displays the mutual information of each of the three source-destination pairs averaged over channel realizations. It can be seen that, using either of the proposed algorithms, the mutual information of each pair converges to the corresponding desired transmission rate as the number of training slots increases. Figure 5.6 shows the average computed powers in each training slot. We can see that by increasing the number of training slots computed power values by these algorithms converge. The computed power values by Algorithm 2 converge faster compared to the ones computed by Algorithm 3. The computed powers (relative to the noise power N0 ) for sources S1 , S2 , and S3 using TDMA scheme are 6.2 dB, 12 dB, and 19 dB, respectively. The proposed algorithms compared to TDMA converge to substantially lower power for each user. These two figures illustrate how the proposed algorithms reduce transmission powers to the minimum required levels while successful transmissions at given rates are still possible for all source-destination pairs. Also, it can be seen that in Algorithm 3, by decreasing the parameter γ the algorithm 120 K-user MIMO Interference Networks: Transceiver Design and Power Control 18 bC 16 rS bC bC rS bC 14 rS bC rS bC pk /N0 [dB] 12 bC rS bC user 3 rS rS bC bC 10 bC rS rS rS bC bC rS rSbC rSbC rSbC rSbC rSbC rSbC rS rS bC rSbC rS 4 rSbC rSbC rSbC rSbC rSbC rS rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC bCrS bCrS bCrS bCrS rSbC bC bC user 1 rS bC 2 rS 0 rSbC rSbC user 2 bC rSbC 6 rSbC rS 8 bC 2 4 6 rS bC rS 8 rSbC rSbC 10 rSbC rSbC 12 rSbC rSbC 14 rSbC 16 18 20 Number of training slots Figure 5.6: Computed pk /N0 versus the number of training slots using Algorithm 2 (−) and Algorithm 3 with γ = 0.5 (◦) and γ = 0.4 (). converges faster. However, very small values of γ may lead to large fluctuations in the computed powers and corresponding mutual information in the steady state (large number of training slots). This shows a trade-off between convergence speed of Algorithm 3 and accuracy of the computed solutions. Figure 5.7 and Figure 5.8 show simulation results of Algorithm 2 for similar set up as the one in the previous example. Computed mutual information and computed power versus the number of training slots are shown for different values of pmax /N0 (pmax /N0 ∈ {15 dB, 20 dB, 30 dB}). It can be seen that in each of these different settings, mutual information values converge to the desired rates, and the computed power of each user converges to a certain value. This shows that when the initial power is set to a value larger than the one required for reliable communication, the algorithm reduces power to the required level. However, further simulations with smaller values of pmax /N0 show that the mutual information may not converge to the desired rates. This coincides with the observation in Figure 5.4 and is a consequence of the fact that in the presence of power constraint the power control problem may not be always feasible. In the second example, we evaluate the performance of the proposed algorithms in a three-user network in which each terminal is equipped with four antennas. 5.3. Performance Evaluation 121 4.5 bC bC Mutual information [bits/channel use] 4 user 3 rS 3.5 bC rS bC 3* bC rS *bC * bC bC rS rS * * rS * rSbC * rSbC * rSbC * rSbC * rSbC * rSbC * rSbC rSbC * * rSbC * rSbC * rSbC * rSbC * rSbC * rSbC * rS 2.5 bC 2 user 2 rS bC * rS bC * * * rS bC rS * bC 1.5 0.5 rSbC * rSbC * rSbC * rSbC * rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC rSbC * * * * * * * * * * bC *rS bC *rSbC *rS bC *rSbC *rSbC *rSbC *rSbC *rSbC *rSbC rS * 1 rSbC * user 1 rS * 2 bC rS * bC rS * 4 rSbC * rSbC * 6 rSbC * bC *rS 8 bC *rS bC *rS 10 *rS 12 14 16 18 20 Number of training slots Figure 5.7: Mutual information of source-destination pair Sk − Dk (Ik ) versus the number of training slots using Algorithm 2 for different values of pmax /N0 : pmax /N0 = 30 dB (◦), pmax /N0 = 20 dB (), pmax /N0 = 15 dB (∗). Since the degrees of freedom (d1 , d2 , d3 ) = (2, 2, 2) are achievable [CJ08], each source transmits two independent data streams. We consider a scenario that each user encodes transmitted data streams with different transmission rates. This represents communication scenarios in which each user desires to send different type of data while the transmission of each of them should satisfy a specific level of quality of service. In this particular example, (R11 , R12 , R21 , R22 , R31 , R32 ) = (1, 2, 3, 4, 5, 6) [bits/channel use], and plk (0)/N0 = pmax /N0 = 20 [dB]. The number of different channel realizations is set to 108 . Figure 5.9 shows the mutual information corresponding to each data stream of the source S1 averaged over different channel realizations. It can be seen that the mutual information of each stream converges to the transmission rate. Therefore, each of the streams can be reliably transmitted. Figure 5.10 shows the computed power associated to each stream of S1 . The powers computed according to Algorithm 3 converge to approximately similar values as the ones by Algorithm 2. These numerical results indicate that the algorithms converge to the desired solution. It can be observed that each computed power compared to the corresponding mutual information in Figure 5.9 has a slower convergence. The reason is that 122 K-user MIMO Interference Networks: Transceiver Design and Power Control 35 30 bC 25 pk /N0 [dB] bC 20 bC user 3 bC rS 15 * 5 user 2 rS * rS 10 * bC bC *bCrS rS user 1 rS * rS * bC *bC *bC rS 0 2 bC bC bC bC bC bC bC bC bC bC bC bC bC bC bC * * * * * * * * * * * * * * * rS rS * bC * rS *bC rS rS rS rS rS bC * rS bC * rS bC * rS bC * * bC rS bC * * 4 rS rSbC * 6 rSbC * rS bCrS rS rSbC * 8 rS * rSbC * bCrS * rSbC * 10 rS bCrS * bCrS * rS bCrS * bCrS * 12 rS bCrS rS bCrS * * bCrS bCrS * * 14 rS bCrS * bCrS * rS rSbC * rSbC * 16 rS rSbC * rSbC * rS rSbC * rSbC * 18 rS bC rSbC * rSbC * rS * rSbC * 20 Number of training slots Figure 5.8: Computed pk /N0 versus the number of training slots using Algorithm 3 for different values of pmax /N0 : pmax /N0 = 30 dB (◦), pmax /N0 = 20 dB (), pmax /N0 = 15 dB (∗). multiple sets of power values can provide the desired mutual information for each stream. However, only one of them is corresponding to the lowest transmission power. Therefore, mutual information may converges to the desired value while transmission power values are still larger than the possible lowest values. Figure 5.11 compares the computed power values in two different scenarios: in the first scenario the number of transmitted streams is set according to the DoF of the network, while in the second scenario the number of transmitted streams is set to a number smaller than the DoF of the network. Specifically, we consider a three-user interference channel in which each terminal has four antennas. In the first scenario, each source sends two streams while in the second scenario each source transmits a single stream. The computed transmission power values for each stream are shown. This result shows that the algorithm converges even if the number of transmitted streams is set to a smaller value compared to the DoF of the network. It can be seen that by sacrificing the number of streams that are sent over the network the powers needed for transmission of certain number of streams can be reduced. The reason is that reducing the number of transmitted streams decreases the amount of leakage interference, thus, lower powers would be sufficient for successful communication. 5.4. Test-bed Implementation 123 Mutual Information [bits/channel use] 7 Algorithm Algorithm Algorithm Algorithm 6 3, 2, 3, 3, I11 I12 I11 I12 5 4 3 2 1 0 0 10 20 30 40 50 60 70 80 90 100 Number of training slots Figure 5.9: The mutual information of the lth stream l ∈ {1, 2} of the sourcedestination pair S1 − D1 (I1l ) versus the number of iterations. The mutual information values of the other source-destination pairs similarly converge to the corresponding transmission rates. 5.4 Test-bed Implementation Wireless test-beds (e.g. the ones based on USRP or WARP hardware) are powerful tools for the experimental verifications of novel interference management algorithms. In recent years, several experimental validations of the algorithms inspired by the interference alignment concept have been reported in the literature [APH10,GRS+ 11,ZM12,MSS+ 12,MAH+ 13,Zet14,MFZS14,MFZ+ 14]. We implemented the proposed algorithm in the previous sections on KTH four-multi testbed and verified its performance in realtime measurements [MFZS14]. KTH fourmulti is a USRP-based wireless communication test-bed consisting of three stationary terminals and three movable terminals, where each terminal is equipped with two antennas [Zet, ZM12]. The channel state feedback links are provided through separate Ethernet links. A three-level synchronization scheme is also applied to synchronize the signals in the time and frequency level. We performed measurements in an indoor office environment at KTH. The measurements indicate at least 4 dB reduction in transmission power in 90% of 124 K-user MIMO Interference Networks: Transceiver Design and Power Control 20 Algorithm Algorithm Algorithm Algorithm 18 16 2, 2, 3, 3, p11 p21 p11 p21 pl1 /N0 [dB] 14 12 10 8 6 4 2 0 10 20 30 40 50 60 70 80 90 100 Number of training slots Figure 5.10: The power of the lth stream (l ∈ {1, 2}) of source S1 (pl1 ) versus the number of iterations. The computed power values of the other sources also show similar convergence behavior. the experiments and at the same time a better bit-error-rate (BER) performance compared to the case where MaxSINR algorithm in [GCJ11] with no power control was implemented. The power saving gains as high as 13 dB was also observed in 10% of the measurements. The benefits of performing power control in the proposed transceiver design and power control algorithm is in fact two-fold. By decreasing the transmission power, while retaining the target SINR, not only less interference is received at each destination, but also the distortion noise due to transceiver hardware impairments (e.g. distortions due to the power amplifier nonlinearities) decreases. 5.5 Summary In this chapter we have addressed transceiver design and power control for MIMO interference networks to provide reliable communication at given rates. Each source intends to send possibly multiple independent data streams where the number of the streams is selected according to the DoF of the network. Each stream is encoded with a fixed data rate while different streams transmitted by one source can possibly 5.5. Summary 125 20 first scenario, p11 first scenario, p12 first scenario, p13 second scenario, p11 second scenario, p12 second scenario, p13 15 p1k /N0 [dB] 10 5 0 −5 −10 0 10 20 30 40 50 60 70 80 90 100 Number of training slots Figure 5.11: Computed p1k /N0 versus the number of training slots when each terminal has four antennas. In the first scenario each terminal sends two streams, and in the second scenario each terminal sends only a single stream. have different rates. We have proposed two distributed iterative algorithms which require only local CSI at each terminal. The transmitter beamforming matrices and the receiver filtering matrices are designed such that the SINR corresponding to each received data stream can be maximized. In addition, power control is performed to assign the minimum required power to each data stream while keeping the mutual information of the corresponding stream larger than the transmission rate. We have investigated the convergence behaviour of the power control scheme, and provided a condition that needs to be fulfilled to have feasible solutions. Numerical performance evaluations confirmed the convergence behavior and that the proposed algorithms required substantially lower powers compared to the conventional orthogonal transmission schemes. Experimental verifications of the proposed algorithm on KTH wireless test-bed confirmed a considerable performance gain compared to the transceiver design algorithms without power control. 126 5.A K-user MIMO Interference Networks: Transceiver Design and Power Control The Proof of Theorem 5.2.1 Since it is assumed that pmax → ∞, we need to show that there is a power vector p satisfying p DFp + n if and only if λmax (DF) < 1. The proof is inspired by the analysis in [HC00a] addressing the feasibility of power control in uplink transmission of the cellular networks. For K > 2, the matrix DF is a regular matrix (a stochastic matrix in which all the entries of some power of the matrix are positive is called regular matrix) because all the elements of (DF)2 are positive. Therefore, the Perron-Frobenius Theorem [PSS05] guarantees the existence of a positive eigenvalue λmax (DF) and the corresponding positive right and left eigenvectors pr and pl which satisfy λmax (DF)pr = DFpr and λmax (DF)pTl = pTl DF, respectively. To prove the necessary condition, assuming p1 ∈ PH we have p1 ∈ PH (a) ⇒ p1 DFp1 + n ⇒ p1 − DFp1 ≻ 0 (b) (c) ⇒ pTl (p1 − DFp1 ) > 0 ⇒ (1 − λmax (DF))pTl p1 > 0 ⇒ λmax (DF) < 1, (d) (e) (5.39) where (a) follows from the definition in (5.26); (b) follows from the fact that n ≻ 0; (c) follows from the positivity of the left eigenvector (pl ≻ 0); (d) follows from the characteristic of eigenvectors and (e) follows from the fact that pl ≻ 0 and p1 ≻ 0, and consequently pTl p1 > 0. To prove the sufficient condition, assuming λmax (DF) < 1 we have (a) λmax (DF)pr = DFpr ⇒ p2 = DFp2 + n + a(1 − λmax (DF))DFpr − n (b) ⇒ p2 DFp2 + n (c) ⇒ p2 ∈ PH ⇒ PH 6= ∅, (5.40) where (a) holds if we set p2 as p2 = aλmax (DF)pr ; (b) holds if we set a = n(i,1) maxi (1−λmax (DF))(DF)(i,:)p where n(i, 1) is the ith element of n and (DF)(i, :) is r the ith row of DF, then since λmax (DF) < 1 we have a(1 − λmax (DF))DFpr − n 0; (c) follows from the definition in (5.26) and the fact that pmax → ∞. 5.B The Proof of Theorem 5.2.2 According to Theorem 2 in [Yat95], if there is at least one positive vector p (p ≻ 0) that satisfies p I′ (p) and the function I′ (p) is a standard interference function, then for any initial power vector p[0] the sequence generated by the recursive equation p[i] = I′ (p[i − 1]) converges to a unique fixed-point p∗ . A function I′ (p) is 5.B. The Proof of Theorem 5.2.2 127 called a standard interference function if for all p, p′ 0, it satisfies the following conditions [Yat95]: 1) Positivity: I′ (p) ≻ 0; 2) Monotonicity: I′ (p) I′ (p′ ), ∀p p′ ; 3) Scalability: αI′ (p) ≻ I′ (αp), ∀α > 1. First, we prove that the function I(p) introduced in (5.18) is a standard interference function by showing that it satisfies the above three conditions. To simplify this verification, we rewrite Ikl (p) as Ikl (p) = L · ϕlk (p) + N0 , (5.41) where L = Rl 2 k −1 |(ulk )∗ Hkk vlk |2 > 0 is a constant. 1) To check the positivity condition we have: (a) Ikl (p) = L · ϕlk (p) + N0 ≥ LN0 > 0, (5.42) where (a) follows from the fact that ϕlk (p) ≥ 0. Therefore, the positivity condition is satisfied. 2) To verify the monotonicity condition, we assume p p′ : 2 2 ∗ ∗ (a) l p p′ ⇒ plj ulk Hkj vlj ≥ p′ j ulk Hkj vlj , ∀l, k, j (b) ⇒ ϕlk (p) ≥ ϕlk (p′ ), ∀l, k (c) ⇒ Ikl (p) ≥ Ikl (p′ ), ∀k, l ⇒ I(p) I(p′ ), (5.43) 2 where (a) follows from the fact that (ulk )∗ Hkj vlj ≥ 0; (b) follows from the definition of ϕlk (p) in (5.14); (c) follows from the definition in (5.17) and the fact that l ∗ (u ) Hkk vl 2 > 0. Therefore, the monotonicity condition is satisfied. k k 3) To check the scalability condition we assume α > 1, then we have Ikl (αp) = L · ϕlk (αp) + N0 = L · αϕlk (p) + N0 (5.44) < αL · ϕlk (p) + N0 = αIlk (p). This condition is valid for any k ∈ {1, 2, ..., K}. Thus, the scalability condition is also satisfied. Now, we can conclude that I(p) given in (5.18) is a standard interference function. Based on Theorem 7 in [Yat95], if I(p) is a standard inh iT terference function, then ˆI(p) = Iˆ1 (p), ..., Iˆd1 (p), ..., Iˆ1 (p), ..., IˆdK (p) , where 1 1 K K 128 K-user MIMO Interference Networks: Transceiver Design and Power Control Iˆkl (p) = min{Ikl (p), pmax }, is also a standard interference function. According to Theorem 2 in [Yat95], if the problem is feasible (PH 6= ∅), for any initial power vector p[0] the iterative algorithm in (5.27) converges to a unique fixed-point p∗ . Lemma 1 in [Yat95] implies that this fixed-point is the solution with the minimum required transmission power for each source. Chapter 6 Multi-cell Interference Networks: Pilot-assisted Opportunistic User Scheduling M ULTI-CELL interference networks are mathematical models for cellular communication systems. These networks consist of multiple cells each of which has one base station that communicates to multiple mobile terminals. The transmitted signals of each base station may interfere to the reception at the mobile terminals within neighboring cells. The number of mobile terminals in cellular systems has been rapidly increasing, and it has been predicted that ultra-dense networks is one of likely scenarios in the next generations of wireless systems (e.g. 5G). The impact of increasing the number of mobile terminals on future networks is twofold: in one hand, coordinated transmission becomes even more important for dealing with interference; and on the other hand, acquiring CSI at mobile terminals and base stations becomes even more difficult to handle and may cause a huge overhead to the system. Therefore, low-complexity coordinated transmission based on low-overhead CSI acquisition schemes must be developed. In order to obtain CSI and to perform data transmissions, we propose a pilot-assisted opportunistic user scheduling (PAOUS) scheme. The proposed scheme consists of a low-complexity channel training mechanism to acquire local CSI at mobile terminals, and a one-bit feedback scheme to enable scheduling at base stations. We compute the achievable rate region for the proposed scheme and characterize the achievable DoF region. The structure of this chapter is as follows. We present the multi-cell interference network in Section 6.1. In Section 6.2, we propose the pilot-assisted opportunistic user scheduling scheme. The achievable rate region and the achievable degrees of freedom region are characterized in Section 6.3. We provide numerical evaluations of the proposed scheme in Section 6.3. Finally, Section 6.4 summarizes this chapter. 129 130 Multi-cell Interference Networks: Pilot-assisted Opportunistic User Scheduling Figure 6.1: Schematic representation of different phases of the PAOUS scheme: (a) orthogonal pilot transmission; (b) feedback transmission and user scheduling; and (c) data transmission phase. 6.1. Multi-cell Interference Network 6.1 131 Multi-cell Interference Network Consider a downlink transmission in a cellular network of K neighbouring cells. In each cell, there is one single-antenna base station serving N single-antenna mobile terminals. The base station in the kth cell (k ∈ {1, . . . , K}) is denoted as BSk , and the mobile terminals are shown as MSkj (j ∈ {1, . . . , N }). Each base station intends to transmit independent messages to mobile terminals in the associated cell. The channel gain between BSk and MSpj (p ∈ {1, . . . , K}) is denoted as hpj,k . We consider the block-fading channel model with coherence time T , where channel gains are constant within one fading block, i.e. hpj,k (nT + t) = hpj,k (nT ) (t ∈ {1, . . . , T − 1}, n ∈ {1, 2, ...}), and change to independent values across subsequent blocks. Channel gains have zero-mean complex Gaussian distribution, i.e. hpj,k ∼ CN (0, 1), and are mutually independent across different users and cells. We assume that no a priori CSI is available at mobile terminals and base stations. 6.2 Pilot-assisted Opportunistic User Scheduling Scheme The proposed PAOUS scheme at each fading block is conducted in three subsequent phases: pilot transmission, feedback transmission and user scheduling, and data transmission phase as shown in Figure 6.1. Within each fading block, αT time slots are allocated to pilot transmission phase where each base station sends Tτ = αT /K pilot symbols. The remaining Td = (1 − α)T time slots are left for data transmission phase. The parameter α (0 < α < 1) is the channel sharing factor. A larger α implies that more channel uses are allocated to pilot transmission. The transmission power of pilot symbols is Pτ and the one for data symbols is Pd . These can be different in general as follows Pd Pτ = = βP, K ((1 − (1 − α) β)/α) P (6.1) (6.2) where P is the transmission power of each base station and β (0 ≤ β ≤ 1/(1 − α)) is a power allocation factor. This power allocation follows the energy conservation law αT Pτ /K + (1 − α)T Pd = T P. (6.3) Clearly, there is a trade-off between the power allocation for pilot transmission and the one for data transmission. In the following, we explain the aforementioned phases in more details. 6.2.1 Pilot Transmission Phase We consider a pilot-assisted channel training scheme to acquire an estimation of local CSI at each mobile terminal, i.e. channel gains between base stations and 132 Multi-cell Interference Networks: Pilot-assisted Opportunistic User Scheduling the corresponding mobile terminal. Channel training is performed in an orthogonal fashion in which the training period is divided into K equal training slots each of which has the duration of Tτ = αT /K. Each base station transmits Tτ pilot symbols during one training slot and remains silent during other time slots as shown in Figure 6.2. Then, each mobile terminal in the network estimates the gain of the corresponding link between the active base station and itself. Consider transmission at the nth fading block. The base station BSk (k ∈ {1, . . . , K}) sends Tτ known pilot symbols as follows p Xτk (t) = Pτ , t ∈ Tnk , (6.4) where Tnk = {nT + (k − 1)Tτ + 1, . . . , nT + kTτ }. Consequently, the received signals at mobile terminal MSpj (p ∈ {1, . . . , K}, j ∈ {1, . . . , N }) is p p Yτ,j (t) = Pτ hpj,k (nT ) + Zjp (t), t ∈ Tnk , where the receiver noise Zjp (t) has Gaussian distribution, i.e. Zjp (t) ∼ CN (0, N0 ). The mobile terminal performs an MMSE estimation of the channel gain hpj,k (nT ) as follows ˜ p (nT ) = h j,k Pτ N0 + Tτ Pτ nTX +kTτ p Yτ,j (t). (6.5) t=nT +(k−1)Tτ +1 The following equation holds hpj,k (nT ) = ˜ hpj,k (nT ) + εpj,k (nT ), (6.6) where εpj,k (nT ) denotes the corresponding channel estimation error. The random ˜ p (nT ) are independent zero-mean Gaussian distributed variables εpj,k (nT ) and h j,k with variances σε2 and 1 − σε2 , respectively, where σε2 = 1 . 1 + Tτ Pτ /N0 (6.7) At the end of the training phase, mobile terminal MSpj obtains the estimation ˜ p (nT ) (k ∈ {1, ..., K}). Then, this noisy estimation of of local channel gains, i.e. h j,k CSI is used to compute the feedback signal as described in the next part. 6.2.2 Feedback Transmission and User Selection Phase Each mobile terminal computes a measure based on the estimated strength of interference links and locally makes a decision whether this measure is below a certain threshold. Then, it sends a one-bit feedback signal to the associated base station. Specifically, in the nth fading block, MSpj computes δjp (nT ) that is defined as follows δjp (nT ) K X 2 p ˜ , hj,i (nT ) . i=1,i6=p (6.8) 6.2. Pilot-assisted Opportunistic User Scheduling Scheme KTτ = αT 133 Td = (1 − α)T Tτ BS1 BS2 BSK Figure 6.2: Transmitted symbols by base stations BSk (k ∈ {1, ..., K}) within one fading block. The crosshatched red slot, the plain green slot, and the blue angle lined slots denote no transmission, pilot symbols, and data symbols, respectively. Next, it sends the one-bit feedback signal fjp (nT ) defined as ( 1 δjp (nT ) ≤ ǫth p fj (nT ) , , 0 δjp (nT ) > ǫth (6.9) where ǫth (ǫth > 0) is a design parameter. The feedback channels are orthogonal to each other and they are assumed to be error-free. The base station BSk collects feedback signals from all mobile terminals within the corresponding cell, i.e. fjk (nT ) (j ∈ {1, ..., N }). A mobile terminal is called a candidate mobile terminal to be scheduled if the corresponding feedback signal is one. We define the set of candidate mobile terminals in the kth cell as follows Ak , ifik (nT ) = 1, i ∈ {1, ..., N } . (6.10) The number of candidate mobile terminals in the kth cell is ρk = |Ak |, where |A| is the cardinality of the set A. If ρk 6= 0, then BSk schedules a randomly selected mobile terminal from the set of the candidate mobile terminals Ak . The selected mobile terminal is denoted as MSkαk . Otherwise, no mobile terminal will be scheduled. The random scheduling ensures that all mobile terminals will be scheduled with the same probability. This implies that the proposed scheme indeed offers fairness. 6.2.3 Data Transmission Phase There are N buffers at each base station, and each of them stores messages that should be sent to a specific mobile terminal. In the data transmission phase, each base station communicates to the scheduled mobile terminal in the associated cell. Messages are encoded according to the multiplexed coding scheme similar to the 134 Multi-cell Interference Networks: Pilot-assisted Opportunistic User Scheduling one proposed in [GV97]. Corresponding to each mobile terminal, there are multiple codebooks each associated with a specific channel state. For a given channel state, BSk (k ∈ {1, 2, . . . , K}) selects message mkαkoindependently with a uniform distrin ′ ˜ ˜ k > 0 is the code rate, and bution from the set M = 1, 2, . . . , 2N Td Rk , where R N ′ is the number of fading blocks that spans one codeword. Then, it encodes the ′ Td k message mkαk to a length N ′ Td codeword {Xd,α (i)}N i=1 . Moreover, the codewords k must satisfy a power constraint h 2 i k < Pd . (6.11) E Xd,α k In fading block n, BSk sends n o(n+1)T k Xd,α (i) k i=nT +KTτ +1 during Td data transmission time slots. All base stations transmit simultaneously at the same frequency band. Consequently, the channel output at MSkαk is k Yd,α (i) k k (i)+ = hkαk ,k (nT )Xd,α k K X l (i) + Z k (i), hkαk ,l (nT )Xd,α k l=1,l6=k i = nT + KTτ + 1, . . . , (n + 1)T (6.12) where Z k (i) ∼ CN (0, 1). The mobile terminal collects all N ′ received signals, decodes the received codeword and estimates the transmitted message. 6.3 Achievable Rate Region In this section, we study the achievable rate region of the PAOUS scheme. Theorem 6.3.1. The pair BSk − MSki can achieve the following rate " !# 2 h P β ˜ Tτ k k γi Eh˜ log 1+ , Ri = 1−K T 1 + β (Kσε2 + ǫth ) P (6.13) where 1 , (6.14) 1 + KTτ ((1 − (1 − α) β) /α) P/N0 ! N 1 ǫth γik = 1− 1−F N σε2 γ K − 1, x2 , (6.15) F (x) , Γ(K − 1) R ˜ ∼ CN 0, 1 − σ 2 . The function Γ(z) , ∞ tz−1 e−t dt is the Gamma function, and h 0 R x z−1ε −t and γ(z, x) , 0 t e dt is the lower incomplete Gamma function. σε2 = 6.3. Achievable Rate Region 135 Proof. See Appendix 6.A. Corollary 6.3.1. The achievable sum-rate of the network RΣ , PK PN k=1 i=1 Rik is " !# 2 h P β ˜ Tτ RΣ =K 1−K γEh˜ log 1+ , T 1 + β (Kσε2 + ǫth ) P where N ǫth γ =1− 1−F . σε2 (6.16) This result can be used to compute the achievable DoF of the network. 6.3.1 Achievable Total Degrees of Freedom The achievable total DoF is characterized in the following theorem. Theorem 6.3.2. In a multi-cell network with K base stations and coherence time T , the achievable total DoF is dΣ = Kopt 1 − Kopt T , (6.17) where Kopt = min T 2,K , (6.18) if the number of mobile terminals (N ) scales proportional to SNR. Proof. We set ǫth = 1/P , and N ∝ P . Then, the achievable total DoF can be computed as dΣ = limP →∞ RΣ / log P , where RΣ is given in Corollary 6.3.1. Using the dominated convergence theorem [MW12], it can be shown that this limit is equal to K (1 − KTτ /T ). We select Tτ = 1 to maximize the achievable total DoF. It can be shown that when K > T2 , the number of active base stations that maximizes the total DoF is K ′ = T2 . To maximize the total DoF, or equivalently the network throughput at highSNR regime, in large networks (K > T /2), Theorem 6.3.2 recommends that turn on only a subset of base stations, and perform the proposed PAOUS scheme within the cells with active base station. Since the network is symmetric, a random base station selection works. In addition, this theorem crystalizes the dependency of the optimum number of active base stations on the channel coherence time. 136 Multi-cell Interference Networks: Pilot-assisted Opportunistic User Scheduling 45 uT 40 bC rS Sum-rate [bits/channel use] 35 N N N N =3 =5 = 10 = 100 rS rS rS bC rS 30 bC rS 25 bC rS bC 20 uT uT bC rS bC uT uT bC rS uT uT uT rS 5 uT bC bC rS 10 uT bC rS uT bC rS 15 uT bC rS uT bC uT 0 10 20 30 40 50 60 70 SNR [dB] Figure 6.3: Sum-rate versus power for different number of mobile terminals in each cell (N ). 6.3.2 Numerical Evaluation In this section, we numerically evaluate the performance of the PAOUS scheme in a three-cell network (K = 3). Figure 6.3 shows the achievable sum-rate versus SNR for different number of mobile terminals in each cell (N ). It can be seen that the sum-rate increases as N increases. The reason is that as the number of mobile terminals increases, it is more likely that the set of candidate mobile terminals in (6.10) be nonempty and mobile terminals be scheduled. Figure 6.4 illustrates the achievable sum-rate as a function of threshold ǫth defined in (6.9) for different SNR values. It can be observed that, for a given SNR, a specific ǫth maximizes sum-rate. The optimum ǫth decays as SNR increases. Increasing ǫth , on one hand increases the probability that a mobile terminal is scheduled in each cell, but on the other hand, the corresponding mutual information decays as a consequence of a larger interference. By increasing SNR interference becomes dominant and ǫth should be reduced in order to limit the level of interference and increase the achievable rate. Figure 6.5 shows the achievable sum-rate versus β for different values of N . It can be observed that for each value of N a specific β maximizes the sum-rate. 6.4. Summary 137 25 SNR = 10 SNR = 20 SNR = 30 SNR = 40 Sum-rate [bits/channel use] 20 dB dB dB dB 15 10 5 0 −60 −50 −40 −30 −20 −10 0 10 20 ǫth [dB] Figure 6.4: The achievable sum-rate versus ǫth for different SNR values. The optimum value of β increases as N increases. This implies that when there are large mobile terminals in the network more power should be allocated to data transmission instead of pilot transmission. 6.4 Summary In this chapter, we have investigated a typical scenario in 5G communication systems, where a large number of mobile terminals in a multi-cell network have to be served efficiently (e.g. with low training and feedback overhead), when no a priori CSI is available neither at mobile terminals nor at base stations. We proposed the PAOUS scheme, and we have shown that the proposed scheme is well-suited for such scenarios: it offers low-computational complexity, and requires only a one-bit feedback signal from mobile terminals to their respective base stations. Furthermore, we computed the achievable rate region for the proposed scheme. We have illustrated that the achievable sum-rate scales as the number of mobile terminals in each cell increases. Our results reveal that in a multi-cell network with K base stations, given that the number of mobile terminals properly scales with the SNR, the sum degrees of freedom Kopt (1 − Kopt /T ) is achievable, where Kopt = min {K, T /2} is the optimum number of the base stations that need to be activated in the network. 138 Multi-cell Interference Networks: Pilot-assisted Opportunistic User Scheduling 12 Sum-rate [bits/channel use] 10 8 6 4 N N N N 2 0 0 0.2 0.4 =5 = 10 = 50 = 100 0.6 0.8 1 1.2 1.4 β Figure 6.5: Sum-rate versus β for different number of users in each cell (N ). Finally, performance evaluations confirm that the proposed opportunistic transmission scheme can exploit the multi-user diversity to mitigate interference even with single-antenna terminals at mobile terminals and base stations, and thereby enhance the achievable sum-rate. 6.A. The Proof of Theorem 6.3.1 6.A 139 The Proof of Theorem 6.3.1 Assuming that MSkαk is scheduled, the mutual information between BSk and the selected mobile terminal MSkαk can be lower bounded as = = = (a) ≥ k k ˜ k ˜k hαk ,1 , . . . , h I Xd,α ; Yd,α αk ,K k k k k k k k ˜k ˜k ˜ ˜ hαk ,1 , . . . , h h Xd,α αk ,K − h Xd,αk hαk ,1 , . . . , hαk ,K , Yd,αk k k k k k k ˜k ˆ d,α ˜ h Xd,α , . . . , h , Y h − h X − X α ,K d,α α ,1 d,α k k k k k k k k k k k ˜ ˆ ˜ log (2πePd ) − h Xd,αk−X d,αk hαk ,1 , . . . , hαk ,K , Yd,αk log (2πePd ) − log 2πeσ 2 , (6.19) k ˜k , . . . , h ˜k ˆk , f h where X is a function of the received signal and αk ,1 αk ,K , Yd,αk d,αk k ˆk the estimated local CSI, and σ 2 is the variance of Xd,α − X d,αk . In this equation k (a) follows from the fact that the entropy of a random variable with a given variance is upper bounded with the entropy of a Gaussian distributed random variable. We ˆ k to be the MMSE estimate of X k as follows select X d,αk d,αk ˆk X d = = ∗ i h ˜ k ˜k ˆk Y k E X h , . . . , h α ,1 αk ,K d,αk d k k ∗ h i Yd,α k ˜ k k k k ˜ E Yd,αk Yd,αk hαk ,1 , . . . , h αk ,K ∗ k ˜k h Yd,α Pd αk ,k k . 2 1 + (Kσε + ǫth ) Pd (6.20) Therefore, the variance σ 2 in (6.19) is σ2 = (a) = = i ∗ ˜k , . . . , h ˜k ˆ k h ˆ k Xk − X Xdk − X αk ,1 αk ,K d d d i h ∗ ˜k ˆ dk ˜ hkαk ,1 , . . . , h E Xdk Xdk − X αk ,K E h 1+ P d 2 h˜ kα ,k Pd , (6.21) k 1+(Kσε2 +ǫth )Pd where (a) follows the orthogonality principle. Substituting the computed variance in (6.19), the lower bound on the mutual information can be computed. In addition, the probability that the mobile terminal MSki is scheduled is γik = N1 γ k , where γ k is the probability that one mobile terminal is scheduled in the kth cell. The probability 140 Multi-cell Interference Networks: Pilot-assisted Opportunistic User Scheduling γ k can be computed as ( k γ = Pr = 1− N [ k δi < ǫth i=1 N Y i=1 Pr δik ) = 1 − Pr (N \ i=1 δik > ǫth N (a) ǫth > ǫth = 1− 1−F , σε2 δk ) (6.22) where (a) follows the fact that the random variable σi2 has Chi-squared distribution ε with degrees of freedom 2(K − 1). The corresponding cumulative density function (CDF) is γ K − 1, x2 , (6.23) F (x) = Γ(K − 1) R∞ Rx where Γ(z) = 0 tz−1 e−t dt is the Gamma function, and γ(z, x) = 0 tz−1 e−t dt is the lower incomplete Gamma function. Substituting Pτ = K ((1 − (1 − α) β) /α) P in (6.7), σε2 given in (6.14) can be computed. Chapter 7 Conclusion 7.1 Concluding Remarks This thesis addresses coordinated transmission schemes in wireless interference networks and their performance limits. The focus has been put mainly on issues regarding channel training, CSI feedback, adaptive transmission, transceiver design, and user scheduling. These issues have been investigated in different wireless communication scenarios, summarized as follows. Fixed-rate transmission in two-user SISO interference networks has first been studied. This scenario can model several practical wireless networks. One obvious example is the real-time controlling signal delivery in machine-to-machine communication systems. To effectively minimize implementation complexity, point-topoint codes are employed at each user. Depending on whether the two sources are activated orthogonally and the decoding strategy at the destinations, five different transmission schemes have been considered. The power control problem for each scheme has been thoroughly analyzed. The inner and outer bounds of the ǫ-outage achievable rate region of these schemes have also been computed. The results have shown that the transmission strategy should be decided based on the relative statistical strength of interference links and direct links to enhance the system performance. The thesis has also considered K-user SISO interference networks with the timevarying fading environment. Due to the fact that in practice normally no CSI is a priori available at terminals (transmitters and receivers), a unified framework has been developed to assign radio resources for channel training and data transmission. Based on the framework, a pilot-assisted ergodic interference alignment scheme has been proposed to conduct coordinated transmission. Via achievable rate region analysis, it has been revealed that with coherence time T , the sum DoF Kopt (1 − Kopt /T )/2 is achievable when the number of active users is selected to be Kopt = min{T /2, K}. Thus, in large networks (K > T /2), it is recommended to first schedule a subset of T /2 users, then perform channel training and coordinated transmission (in this case, performing ergodic interference alignment) within the set 141 142 Conclusion of the scheduled users. Moreover, the optimum way of allocating power to channel training and data transmission has been found. The observations have implied that to increase the achievable rate, when the network is large or the channel coherence time is small, more power should be allocated to the channel training instead of the data transmission. The above results have revealed the inherent performance limits of wireless interference networks due to the intrinsic requirements for the radio resources dedicated for channel training. They can provide intuition helpful for the design of coordinated transmission schemes over interference networks with time-varying channels when no CSI is a priori available at terminals. In these K-user SISO interference networks, the case that only quantized CSI is available at sources has also been investigated. The coordinated transmission design has been established again based on the ergodic interference alignment concept and focused on two problems. An iterative power control algorithm has been proposed to find the smallest power consumption to guarantee a fixed-rate transmission. In addition, a rate adaptation scheme has been presented to maximize network throughput. For the latter case, performance in delay-sensitive systems has also been studied. The above results have shown that with proper transmission design, applying the ergodic interference alignment scheme can still provide performance enhancements over conventional orthogonal transmission strategies, even if only quantized CSI with limited resolution is available at the sources. If the total number of feedback bits is fixed, different feedback bit allocation approaches would have diverse impacts on system performance. Specifically, when the network is operating in noise-limited regime, more bits should be assigned to the direct links to more accurately conduct rate adaptation (or power control). On the other hand, in interference-dominant scenarios, more accurate quantization should be provided to interference links to realize better interference alignment. Furthermore, when acceptable delay is limited, the accuracy of channel quantization and the number of users within the network may affect the system performance in a way that is different from the case in systems without strict delay limits. These results can be used to properly choose transmission strategies and select design parameters according to different system requirements. Joint transceiver design and power control for K-user MIMO interference networks has been addressed to provide reliable communication at given rates. Equipped with multiple antennas, potentially multiple independent data streams with different data rates can be delivered between each user pair. Two distributed iterative algorithms, which require only local CSI at each terminal, have been proposed. Their performance advantages over conventional orthogonal transmission schemes in terms of power consumption have been confirmed by both computer simulations and practical hardware test-bed implementations. A typical scenario in the next generation of communication systems (i.e. 5G systems) has been studied, where a large number of users in a multi-cell network have to be served efficiently (e.g. with low training and feedback overhead), when no a priori CSI is available at mobile terminals and base stations. A pilot-assisted opportunistic user scheduling (PAOUS) scheme is proposed. It has been shown that 7.2. Future Work 143 the proposed scheme is well-suited for such scenarios: it offers low-computational complexity, and requires only a one-bit feedback signal from each mobile terminal to its respective base station. The achievable rate region for the proposed scheme is computed. It has been illustrated that the proposed scheme takes advantage of the crowd of mobile terminals as the achievable sum-rate scales by increasing the number of mobile terminals in each cell. The results reveal that in a multi-cell network with B base stations, given that the number of mobile terminals properly scale with the SNR, the sum degrees of freedom Bopt (1 − Bopt /T ) is achievable, where Bopt = min B, T2 is the optimum number of the base stations that need to be activated in the network. Finally, performance evaluations confirm that the proposed opportunistic transmission scheme can exploit multiple users in the network to mitigate interference even with single-antenna terminals and no a priori CSI available at mobile terminals and base stations, and thereby enhance the achievable sum-rate. 7.2 Future Work There are several possible directions to extend the results of this thesis to support the development of future wireless systems. In this part, we highlight some of these extensions for further study: Self-organizing Coordinated Networks By self-organizing coordinated networks we mean the systems in which multiple terminals without any a priori knowledge about the propagation environment establish coordination in order to conduct information transmissions. In these networks, part of the radio resources can be allocated for establishing coordination, and the rest can be used for coordinated transmissions. Several communication scenarios including ad-hoc networks, wireless sensor networks with random deployment of the sensors, and machine-to-machine communication networks in non-stationary environments are examples of such networks. The theoretical framework which has been developed in Chapter 4 and Chapter 6 for the design and analysis of coordinated transmission when no a priori CSI is available at terminals can be applied to design and to analyze transmission schemes for these networks. Multi-layer Coding for Coordinated Transmission In Chapter 4 and Chapter 6 we have considered systems in which noisy CSI is available at destinations. As we have shown, this degrades the performance of decoders. Developing coding schemes that are more robust to such imperfections is of high importance. To address this issue, a multi-layer coding technique can be designed for fading interference networks in which only noisy CSI is available. Specifically, the information bits can be encoded using a multi-layer coding scheme, an interference alignment scheme or a user scheduling can be applied to partially remove the 144 Conclusion interference, and successive decoding can be applied at receivers to decode different layers of encoded signal. An initial attempt for developing multi-layer coding technique for fading multiple-access channel with noisy CSI is reported in [AP14], where it has been shown that a larger rate region can be achieved compared to the single-layer coding scheme. Synchronization in Interference Networks The synchronization of terminals is essential for coordinated data transmission and coherent signal detection. In this thesis, we have assumed that terminals are perfectly synchronized. To realize this assumption, low-complexity synchronization schemes for the schemes proposed in this thesis should be developed. The performance limits of such synchronization schemes should be quantified. In addition, similar to channel training and feedback transmission, part of radio resources must be allocated for performing synchronization. Therefore, the optimum resource allocation for synchronization is also an important problem to be investigated. Channel Training and Feedback Design for Two-user Networks In Chapter 3, we have considered transmission schemes based on point-to-point coding when perfect global CSI is available at terminals. Design of low complexity channel training and channel state feedback schemes for these transmission schemes is a possible extension. Data-aided channel Training for Interference Networks In Chapter 4 and Chapter 6, we have studied a pilot-based channel training for interference networks. We have shown that allocating radio resources for channel training is a limiting factor of the investigated coordinated transmission schemes. It is an interesting problem to design more efficient channel training to enhance the performance of these schemes. For example, in packet-based communication systems, data packets have headers which are highly protected with error correcting codes. Terminals first decode the headers, and then proceed with decoding the payload data. Therefore, if the headers are successfully decoded, then they can be used as pilots for channel estimation purpose. This idea can be further extended to communication systems in which a high-rate data source is multiplexed with a low-rate data source. If the low-rate data is decoded successfully, then it can be used as a pilot for channel estimation and decoding of the high-rate data. The early results on this channel training scheme has been presented in [YT03]. This approach can be investigated to further enhance the channel training schemes devised in this thesis. 7.2. Future Work 145 Distributed Transceiver Design and Power Control for WLAN In Chapter 5, we proposed algorithms for distributed transceiver design and power control for MIMO systems. The proposed algorithms can provide multi-stream transmissions by each user at desired rates. The test-bed implementation of these algorithms confirmed that they can provide efficient and reliable communication in indoor environments [MFZS14, MFZ+ 14]. There have been recent developments in the standardizations of high-rate multi-stream data transmission schemes for wireless local area networks (WLANs) (such as IEEE 802.11ac standard). The aforementioned algorithms can be further investigated as possible solutions for multi-stream fixed-rate transmission in WLAN. Adaptive Channel Training and Channel State Feedback The real wireless channels can be modeled as temporally correlated channels in which the current channel realization is dependent on the previous channel realizations. This temporal correlation can be exploited in order to reduce the radio resources allocated for channel training and channel state feedback. For single-user MIMO communication systems, adaptive feedback schemes have been studied (see e.g. [KLC11]), and channel prediction algorithms are proposed (see e.g [WE06]). These solutions exploit the temporal correlations of the channel to reduce the overhead of channel training and feedback. For interference networks, developing such schemes is a more involved problem, however, a considerable performance gain can be predicted to be achieved. Coordinated Cluster Design The theoretical results in Chapter 4 and Chapter 6 for the performance limits of interference networks can be used for the design of future clustered coordinated networks. These networks are composed of multiple clusters, where within each cluster multiple users perform coordinated transmission. The results developed in these two chapters can be used to find the optimum size of the clusters and perform radio resource allocation between the clusters. Bibliography [AB12] C. Ak¸caba and H. B¨olcskei, “Diversity-multiplexing tradeoff in twouser fading interference channels,” IEEE Trans. Inf. Theory, vol. 58, no. 7, pp. 4462–4480, Jul. 2012. [AC78] A. A. Carleial, “Interference channels,” IEEE Trans. Inf. Theory, vol. 24, no. 1, pp. 60–70, Jan. 1978. [AG99] M. S. Alouini and A. J. Goldsmith, “Capacity of Rayleigh fading channels under different adaptive transmission and diversitycombining techniques,” IEEE Trans. Veh. Technol., vol. 48, no. 4, pp. 1165–1181, Jul. 1999. [AH12] O. E. Ayach and R. W. Heath, “Interference alignment with analog channel state feedback,” IEEE Trans. Wireless Commun., vol. 11, no. 2, pp. 626–636, Feb. 2012. [Ahl74] R. Ahlswede, “The capacity region of a channel with two senders and two receivers,” Ann. Prob., vol. 2, no. 5, pp. 805–814, 1974. [ALH12] O. E. Ayach, A. Lozano, and R. Heath, “On the overhead of interference alignment: Training, feedback, and cooperation,” IEEE Trans. Wireless Commun., vol. 11, no. 11, pp. 4192–4203, Nov. 2012. [AP14] J. F. A. Pastore, T. Koch, “A rate-splitting approach to fading multiple-access channels with imperfect channel-state information,” in Proc. Int. Zurich Seminar on Commun., Zurich, Switzerland, Feb. 2014. [APH10] O. E. Ayach, S. W. Peters, and R. W. Heath, “The feasibility of interference alignment over measured MIMO-OFDM channels,” IEEE Trans. Veh. Tech., vol. 59, no. 9, pp. 4309–4321, Nov. 2010. [APH13] O. E. Ayach, S. W. Peters, and R. J. Heath, “The practical challenges of interference alignment,” IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 35–42, Feb. 2013. [AS64] M. Abramowitz and I. A. Stegun, Handbook of mathematical functions. Courier Dover Publications, 1964. 147 148 Bibliography [AV09] V. S. Annapureddy and V. V. Veeravalli, “Gaussian interference networks: Sum capacity in the low-interference regime and new outer bounds on the capacity region,” IEEE Trans. Inf. Theory, vol. 55, no. 7, pp. 3032–3050, Jul. 2009. [BGT11] F. Baccelli, A. E. Gamal, and D. N. C. Tse, “Interference networks with point-to-point codes,” IEEE Trans. Inf. Theory, vol. 57, no. 5, pp. 2582–2596, May 2011. [BLM03] I. Barhumi, G. Leus, and M. Moonen, “Optimal training design for MIMO OFDM systems in mobile wireless channels,” IEEE Trans. Signal Process., vol. 51, no. 6, pp. 1615–1624, Jun. 2003. [BT89] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and distributed computation. Prentice-Hall, 1989. [BT09] H. B¨olcskei and I. J. Thukral, “Interference alignment with limited feedback,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT’09), Seoul, Korea, Jun. 2009. [BV04] S. Boyd and L. Vanderbenghe, Convex optimization. Univ. Press, 2004. [Car75] A. Carleial, “A case where interference does not reduce capacity,” IEEE Trans. Inf. Theory, vol. 21, no. 5, pp. 569–570, Sep. 1975. [Car78] ——, “Interference channels,” IEEE Trans. Inf. Theory, vol. 24, no. 1, pp. 60–70, Jan. 1978. [CE87] M. H. M. Costa and A. A. Elgamal, “The capacity region of the discrete memoryless interference channel with strong interference,” IEEE Trans. Inf. Theory, vol. 33, no. 5, pp. 710–711, Sep. 1987. [CGG08] H. F. Chong, H. K. Garg, and H. E. Gamal, “On the Han-Kobayashi region for the interference channel,” IEEE Trans. Inf. Theory, vol. 54, no. 7, pp. 3188–3195, Jul. 2008. [CHA+ 11] H. Congzheng, T. Harrold, S. Armour, I. Krikidis, S. Videv, P. M. Grant, H. Haas, J. S. Thompson, I. Ku, W. Cheng-Xiang, A. L. Tuan, M. R. Nakhai, Z. Jiayi, and L. Hanzo, “Green radio: radio techniques to enable energy-efficient wireless networks,” IEEE Commun. Mag., vol. 49, no. 6, pp. 46–54, Jun. 2011. [CHLT08] M. Chiang, P. Hande, T. Lan, and C. W. Tan, Power control in wireless cellular networks. Foundations and Trends in Networking, 2008, vol. 2, no. 4. Cambridge Bibliography 149 [CJ08] V. R. Cadambe and S. A. Jafar, “Interference alignment and degrees of freedom of the K-user interference channel,” IEEE Trans. Inf. Theory, vol. 54, no. 8, pp. 3425–3441, Aug. 2008. [CJKR10] G. Caire, N. Jindal, M. Kobayashi, and N. Ravindran, “Multiuser MIMO achievable rates with downlink training and channel state feedback,” IEEE Trans. Inf. Theory, vol. 56, no. 6, pp. 2845–2866, Jun. 2010. [Cov72] T. M. Cover, “Broadcast channels,” IEEE Trans. Inf. Theory, vol. 18, no. 1, pp. 2–14, Jan. 1972. [CT91] T. Cover and J. Thomas, Elements of information theory. 1991. [CTB99] G. Caire, G. Taricco, and E. Biglieri, “Optimum power control over fading channels,” IEEE Trans. Inf. Theory, vol. 45, pp. 1498–1489, Jul. 1999. [CY14] X. Chen and C. Yuen, “Performance analysis and optimization for interference alignment over MIMO interference channels with limited feedback,” IEEE Trans. Signal Process., vol. 62, no. 7, pp. 1785–1795, Apr. 2014. [DCG+ 13] F. Daquan, J. Chenzi, L. Gubong, L. J. J. Cimini, F. Gang, and G. Y. Li, “A survey of energy-efficient wireless communications,” IEEE Commun. Surveys Tuts., vol. 15, no. 1, pp. 167–178, Jan. 2013. [Dic81] American Heritage Dictionary. [EK11] A. Elgamal and Y. H. Kim, Network Information Theory. Cambridge University Press, Dec. 2011. [EK12] H. Ebrahimzad and A. K. Khandani, “On the optimum diversitymultiplexing tradeoff of the two-user Gaussian interference channel with Rayleigh fading,” IEEE Trans. Inf. Theory, vol. 58, no. 7, pp. 4481–4492, Jul. 2012. [Eri12] Ericsson. (2012) Traffic and market report. [Online]. Available: http://www.ericsson.com/res/docs/2012/traffic_and_ market_report_june_2012.pdf [ETW08] R. H. Etkin, D. N. C. Tse, and H. Wang, “Gaussian interference channel capacity to within one bit,” IEEE Trans. Inf. Theory, vol. 54, pp. 5534–5562, 2008. [Far12] H. Farhadi, “Interference alignment and power control for wireless interference networks,” Licentiate thesis, KTH Royal Institute of Technology, Communication Theory Department, Sep. 2012. Wiley, Boston: Houghton Mifflin, 1981. 150 Bibliography [FGS14] H. Farhadi, H. Ghauch, and M. Skoglund, “Pilot-assisted opportunistic user scheduling for wireless multi-cell networks,” in Submitted to IEEE Int. Conf. Commun. (ICC’15), London, UK, Oct. 2014. [FJC12] H. R. Feyzmahdavian, M. Johansson, and T. Charalambous, “Contractive interference functions and rates of convergence of distributed power control laws,” IEEE Trans. Wireless Commun., vol. 11, no. 12, pp. 4494–4502, Dec. 2012. [FKS14] H. Farhadi, M. N. Khormuji, and M. Skoglund, “Pilot-assisted ergodic interference alignment for wireless networks,” in Proc. IEEE Int. Conf. Acoustics, Speech and Signal Process. (ICASSP’14), Florence, Italy, May 2014, [Best Student Paper in Signal Processing for Communications and Networking]. [FKWS13] H. Farhadi, M. N. Khormuji, C. Wang, and M. Skoglund, “Ergodic interference alignment with noisy channel state information,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT’13), Istanbul, Turkey, Jul. 2013. [FM93a] G. J. Foschini and Z. Miljanic, “A simple distributed autonomous power control algorithm and its convergence,” IEEE Trans. Veh. Technol., vol. 42, no. 4, pp. 641–646, Nov. 1993. [FM93b] ——, “A simple distributed autonomous power control algorithm and its convergence,” IEEE Trans. Veh. Technol., vol. 42, no. 4, pp. 641– 646, Nov. 1993. [FWS11] H. Farhadi, C. Wang, and M. Skoglund, “On the throughput of wireless interference networks with limited feedback,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT’11), Saint Petersburg, Russia, Jul. 2011. [FWS12] ——, “Power control in wireless interference networks with limited feedback,” in Proc. Int. Symp. on Wireless Commun. Systems (ISWCS’12), Paris, France, Aug. 2012. [FWS13] ——, “Distributed interference alignment and power control for wireless MIMO interference networks,” in Proc. IEEE Wireless Commun. and Networking Conf. (WCNC’13), Shanghai, China, Apr. 2013. [FWS14a] ——, “Delay-limited constant-rate transmission over fading interference channels using point-to-point Gaussian codes,” Submitted to IEEE Trans. Commun., Dec. 2014. [FWS14b] ——, “Distributed transceiver design and power control for wireless MIMO interference networks,” Accepted for publication in IEEE Trans. Wireless Commun., Oct. 2014. Bibliography 151 [FWS14c] ——, “Interference alignment with limited feedback: Power control and rate adaptation,” Submitted to IEEE Trans. Wireless Commun., Feb. 2014. [FZF+ 13] H. Farhadi, A. Zaidi, C. Fischione, C. Wang, and M. Skoglund, “Distributed interference alignment and power control for wireless MIMO interference networks with noisy channel state information,” in Proc. Int. Black Sea Conf. Commun. and Networking (BlackSeaCom’13), Batumi, Georgia, Jul. 2013. [GC98] A. J. Goldsmith and S. G. Chua, “Adaptive coded modulation for fading channels,” IEEE Trans. Inf. Theory, vol. 46, no. 5, pp. 595– 602, May 1998. [GCJ08] K. Gomadam, V. R. Cadambe, and S. A. Jafar, “Approaching the capacity of wireless networks through distributed interference alignment,” in Proc. Global Telecommunications Conf. (GLOBECOM’08), New Orleans, USA, Dec. 2008, pp. 1–6. [GCJ11] ——, “A distributed numerical approach to interference alignment and applications to wireless interference networks,” IEEE Trans. Inf. Theory, vol. 57, no. 6, pp. 3309–3322, Jun. 2011. [GJ10] T. Gou and S. A. Jafar, “Degrees of freedom of the K user M × N MIMO interference channel,” IEEE Trans. Inf. Theory, vol. 56, no. 12, pp. 6040–6057, Dec. 2010. [GKGI07] D. Gesbert, S. G. Kiani, A. Gjendemsjo, and G. E. Ien, “Adaptation, coordination, and distributed resource allocation in interferencelimited wireless networks,” Proc. IEEE, vol. 95, no. 12, pp. 2393– 2409, Dec. 2007. [GN98] R. M. Gray and D. L. Neuhoff, “Quantization,” IEEE Trans. Inf. Theory, vol. 44, no. 6, pp. 2325–2383, Oct. 1998. [GRS+ 11] O. González, D. Ramírez, I. Santamaria, J. A. García-Naya, and L. Castedo, “Experimental validation of interference alignment techniques using a multiuser MIMO testbed,” in Proc. Int. ITG Workshop on Smart Antennas (WSA’11), Aachen, Germany, Feb. 2011, pp. 1–8. [GSDs+ 03] D. Gesbert, M. Shafi, S. Da-shan, P. J. Smith, and A. Naguib, “From theory to practice: an overview of MIMO space-time coded wireless systems,” IEEE J. Sel. Areas Commun., vol. 21, no. 3, pp. 281–302, Apr. 2003. 152 Bibliography [GV97] A. J. Goldsmith and P. P. Varaiya, “Capacity of fading channels with channel side information,” IEEE Trans. Inf. Theory, vol. 43, no. 6, pp. 1986–1992, Nov. 1997. [HC00a] J. D. Herdtner and E. K. P. Chong, “Analysis of a class of distributed asynchronous power control algorithms for cellular wireless systems,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, pp. 436–446, Mar. 2000. [HC00b] ——, “Analysis of a class of distributed asynchronus power control algorithms for cellular wireless systems,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, Mar. 2000. [Hen92] R. Heniron, “On constraint qualifications,” J. Optim. Theory Appl., vol. 72, no. 1, Jan. 1992. [Her85] C. A. J. Herter, “The electromagnetic spectrum: A critical natural resource,” J. Nat. Resources, vol. 25, Jul. 1985. [HH03] B. Hassibi and B. M. Hochwald, “How much training is needed in multiple-antenna wireless links?” IEEE Trans. Inf. Theory, vol. 49, no. 4, pp. 951–963, Apr. 2003. [HJ85] R. A. Horn and C. R. Johnson, Matrix analysis. versity Press, 1985. [HK81] T. Han and K. Kobayashi, “A new achievable rate region for the interference channel,” IEEE Trans. Inf. Theory, vol. 27, no. 1, pp. 49–60, Jan. 1981. [HKD11] J. Hoydis, M. Kobayashi, and M. Debbah, “Optimal channel training in uplink network MIMO systems,” IEEE Trans. Signal Process., vol. 59, no. 6, pp. 2824–2833, Jun. 2011. [HZB+ 11] Y. Huang, G. Zheng, M. Bengtsson, K. K. Wong, L. Yang, and B. Ottersten, “Distributed multicell beamforming design with limited intercell coordination,” IEEE Trans. Signal Process., vol. 59, no. 2, pp. 728–738, Feb. 2011. [HZB+ 12] ——, “Distributed multicell beamforming design approaching pareto boundary with max-min fairness,” IEEE Trans. Wireless Commun., vol. 11, no. 8, pp. 2921–2933, Aug. 2012. [Jaf11] S. A. Jafar, Interference alignment: A new look at signal dimensions in a communication network. Foundations and Trends® in Communications and Information Theory, Jun. 2011, vol. 7, no. 1. [JAP12] O. Johnson, M. Aldridge, and R. Piechocki, “Delay-rate tradeoff in ergodic interference alignment,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT’12), Cambridge, MA, USA, 2012, pp. 2626–2630. Cambridge Uni- Bibliography 153 [JBS04] E. A. Jorswieck, H. Boche, and A. Sezgin, “Delay-limited capacity and maximum throughput of spatially correlated multiple antenna systems under average and peak-power constraints,” in Proc. IEEE Inf. Theory Workshop, San Antonio, Texas, USA, Oct. 2004. [JLD08] E. A. Jorswieck, E. G. Larsson, and D. Danev, “Complete characterization of the pareto boundary for the MISO interference channel,” IEEE Trans. Signal Process., vol. 56, no. 10, pp. 5292–5296, Oct. 2008. [JMZ+ 14] V. Jungnickel, K. Manolakis, W. Zirwas, B. Panzner, V. Braun, M. Lossow, M. Sternad, R. Apelfr¨ojd, and T. Svensson, “The role of small cells, coordinated multipoint, and massive MIMO in 5G,” IEEE Commun. Mag., vol. 52, no. 5, pp. 44–51, May 2014. [JNPS12] B. C. Jung, G. Nat, D. Park, and W. Y. Shin, “Opportunistic interference mitigation achieves optimal degrees-of-freedom in wireless multi-cell uplink networks,” IEEE Trans. Commun., vol. 60, no. 7, pp. 1935–1944, Jul. 2012. [JSO02] G. J¨ongren, M. Skoglund, and B. Ottersten, “Combining beamforming and orthogonal space-time block coding,” IEEE Trans. Inf. Theory, vol. 48, no. 3, pp. 611–627, Mar. 2002. [JwJS09] C. Jeong-woo, M. Jeonghoon, and C. Song, “Joint network-wide opportunistic scheduling and power control in multi-cell networks,” IEEE Trans. Wireless Commun., vol. 8, no. 3, pp. 1520–1531, Mar. 2009. [KG13] P. D. Kerret and D. Gesbert, “CSI sharing strategies for transmitter cooperation in wireless networks,” IEEE Trans. Wireless Commun., vol. 20, no. 1, pp. 43–49, Feb. 2013. [KJC11] M. Kobayashi, N. Jindal, and G. Caire, “Training and feedback optimization for multiuser MIMO downlink,” IEEE Trans. Commun., vol. 59, no. 8, pp. 2228–2240, Aug. 2011. [KLC11] T. Kim, D. J. Love, and B. Clerckx, “MIMO systems with limited rate differential feedback in slowly varying channels,” IEEE Trans. Commun., vol. 59, no. 4, pp. 1175–1189, Apr. 2011. [KMLL12] J. S. Kim, S. H. Moon, S. R. Lee, and I. Lee, “A new channel quantization strategy for MIMO interference alignment with limited feedback,” IEEE Trans. Wireless Commun., vol. 11, no. 1, pp. 358–366, Jan. 2012. 154 Bibliography [Kra04] G. Kramer, “Outer bounds on the capacity of Gaussian interference channels,” IEEE Trans. Inf. Theory, vol. 50, no. 3, pp. 581–586, Mar. 2004. [KRB+ 13] D. Katselis, C. R. Rojas, M. Bengtsson, E. Bj¨ornson, X. Bombois, N. Shariati, M. Jansson, and H. Hjalmarsson, “Training sequence design for MIMO channels: an application-oriented approach,” EURASIP J. Wireless Commun. Net., vol. 2013, no. 245, pp. 1–22, 2013. [KV10] R. T. Krishnamachari and M. K. Varanasi, “Interference alignment under limited feedback for MIMO interference channels,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT’10), Jun. 2010. [KWG10] J. C. Koo, W. Wu, and J. T. Gill, “Delay-rate tradeoff for ergodic interference alignment in the Gaussian case,” in Proc. Ann. Allerton Conf. Commun., Control, and Computing, IL, USA, Sep. 2010. [LGLL14] H. Liu, H. Gao, W. Long, and T. Lv, “A novel scheme for downlink opportunistic interference alignment,” arxiv.org, 2014. [LHL+ 08] D. J. Love, R. W. Heath, V. K. N. Lau, D. Gesbert, B. D. Rao, and M. Andrews, “An overview of limited feedback in wireless communication systems,” IEEE J. Sel. Areas Commun., vol. 26, no. 8, pp. 1341–1365, Oct. 2008. [LJ08] E. Larsson and E. Jorswieck, “Competition versus cooperation on the MISO interference channel,” IEEE J. Sel. Areas Commun., vol. 26, no. 7, pp. 1059–1069, Sep. 2008. [LK12] H. H. Lee and Y. C. Ko, “Interference alignment with random vector quantization for MIMO interference channels,” in Proc. IEEE Vehicular Tech. Conf. (VTC’12), Québec City, Canada, 2012, pp. 1–5. [LS03] E. G. Larsson and P. Stoica, Space-Time Block Coding for Wireless Communications. Cambridge University Press, 2003. [LSD04] T. Lang, B. M. Sadler, and M. Dong, “Pilot-assisted wireless transmissions: general model, design criteria, and signal processing,” IEEE Signal Process. Mag., vol. 21, no. 6, pp. 12–25, Nov. 2004. [LTYF03] D. Liang, L. Teng, and H. Yih-Fang, “Opportunistic transmission scheduling for multiuser MIMO systems,” in Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Process. (ICASSP’03), Hong Kong, Hong Kong, Apr. 2003. Bibliography 155 [MAH+ 13] M. Mayer, G. Artner, G. Hannak, M. Lerch, and M. Guillaud, “Measurement based evaluation of interference alignment on the Vienna MIMO testbed,” in Proc. Int. Symp. Wireless Commun. Systems (ISWCS’13), Ilmenau, Germany, Aug. 2013, pp. 1–5. [MAMK08] M. A. Maddah-Ali, A. S. Motahari, and A. K. Khandani, “Communication over MIMO X channels: Interference alignment, decomposition, and performance analysis,” IEEE Trans. Inf. Theory, vol. 54, no. 8, pp. 3457–3470, Aug. 2008. [MAT10] M. A. Maddah-Ali and D. Tse, “Completely stale transmitter channel state information is still very useful,” in Ann. Allerton Conf. Commun., Control, and Computing (Allerton’10), Sep. 2010. [MC90] T. W. Malone and K. Crowston, “What is coordination theory and how can it help design cooperative work systems?” in Proc. ACM Conf. Computer-supported Cooperative Work. New York, NY, USA: ACM, 1990, pp. 357–370. [Online]. Available: http://doi.acm.org/ 10.1145/99332.99367 [MDEK10] H. Mahdavi-Doost, M. Ebrahimi, and A. K. Khandani, “Characterization of SINR region for interfering links with constrained power,” IEEE Trans. Inf. Theory, pp. 2816–2828, May 2010. [MET13] METIS, “Deliverable D3.1: positioning of multi-node/multi-antenna technologies,” Tech. Rep., Jul. 2013. [MFZ+ 14] N. N. Moghadam, H. Farhadi, P. Zetterberg, M. Khormuji, and M. Skoglund, Interference alignment: Practical challenges and testbed implementation, book chapter in Contemporary Issues in Wireless Communications. INTECH Open Access Publisher, Nov. 2014. [MFZS14] N. N. Moghadam, H. Farhadi, P. Zetterberg, and M. Skoglund, “Testbed implementation of iterative interference alignment and power control for wireless MIMO interference networks,” in Proc. IEEE Int. Workshop on Signal Process. Advances in Wireless Commun. (SPAWC’14), Toronto, Canada, Jun. 2014. [MGL13] R. K. Mungara, G. George, and A. Lozano, “Pilot-assisted interference alignment in time-selective fading channels,” in Proc. IEEE Global Commun. Conf. (GLOBECOM’13), Atlanta, GA, USA, Dec. 2013, pp. 3977–3982. [MGL14] ——, “Overhead and spectral efficiency of pilot-assisted interference alignment in time-selective fading channels,” IEEE Trans. Wireless Commun., vol. 13, no. 9, pp. 4884–4895, Sep. 2014. 156 Bibliography [MGMAK14] A. S. Motahari, S. O. Gharan, M. A. Maddah-Ali, and A. K. Khandani, “Real interference alignment: Exploiting the potential of single antenna systems,” IEEE Trans. Inf. Theory, vol. 60, no. 8, pp. 4799– 4810, Aug. 2014. [MJS12] H. Maleki, S. A. Jafar, and S. Shamai, “Retrospective interference alignment over interference networks,” IEEE J. Sel. Topics Signal Process., vol. 6, no. 3, pp. 228–240, Jun. 2012. [MK09] A. S. Motahari and A. K. Khandani, “Capacity bounds for the Gaussian interference channel,” IEEE Trans. Inf. Theory, vol. 55, pp. 620– 643, Feb. 2009. [MLES13] B. Makki, J. Li, T. Eriksson, and T. Svensson, “Coordinated multipoint joint transmission with partial channel information feedback,” in Proc. The European Wireless Conf. (EW’13), Guildford, UK, Apr. 2013, pp. 1–5. [MM12] H. Maier and R. Mathar, “Uniform and non-uniform delay-rate tradeoffs in partial ergodic interference alignment,” in Proc. IEEE Inf. Theory Workshop (ITW’12), Lausanne, Switzerland, 2012, pp. 672– 676. [MR69] C. L. Mallows and D. Richter, “Inequalities of Chebyshev type involving conditional expectations,” Ann. Math. Statist., vol. 40, no. 6, pp. 1922–1932, Dec. 1969. [MSS+ 12] J. W. Massey, J. Starr, L. Seogoo, L. Dongwook, A. Gerstlauer, and R. W. Heath, “Implementation of a real-time wireless interference alignment network,” in Proc. Asilomar Conf. Signals, Systems and Computers (ASILOMAR’12), Asilomar, CA, USA, Nov. 2012, pp. 104–108. [MW12] J. McDonald and N. A. Weiss, A course in real analysis, 2nd ed. Elsevier Inc., 2012. [NGJV12] B. Nazer, M. Gastpar, S. A. Jafar, and S. Vishwanath, “Ergodic interference alignment,” IEEE Trans. Inf. Theory, vol. 58, no. 10, pp. 6355–6371, Oct. 2012. [Ngu09] K. D. Nguyen, “Adaptive transmission for block-fading channels,” Ph.D. dissertation, University of South Australia, 2009. [NJGV09] B. Nazer, S. Jafar, M. Gaspar, and S. Vishwanath, “Ergodic interference alignment,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT’09), Seoul, Korea, Jun. 2009. Bibliography 157 [NWHC12] L. Namyoon, S. Wonjae, R. W. Heath, and B. Clerckx, “Interference alignment with limited feedback for two-cell interfering MIMOMAC,” in Proc. Int. Symp. Wireless Commun. Systems (ISWCS’12), Paris, France, Aug. 2012, pp. 566–570. [OBB+ 14] A. Osseiran, F. Boccardi, V. Braun, K. Kusume, P. Marsch, M. Maternia, O. Queseth, M. Schellmann, H. Schotten, H. Taoka, H. Tullberg, M. A. Uusitalo, B. Timus, and M. Fallgren, “Scenarios for 5G mobile and wireless communications: the vision of the METIS project,” IEEE Commun. Mag., vol. 52, no. 5, pp. 26–35, May 2014. [PDLC08] S. M. Perlaza, M. Debbah, S. Lasaulce, and J. M. Chaufray, “Opportunistic interference alignment in MIMO interference channels,” in Proc. IEEE Int. Symp. on Personal, Indoor and Mobile Radio Commun. (PIMRC’08), Cannes, France, Sep. 2008. [PH11] S. W. Peters and R. W. Heath, “Cooperative algorithms for MIMO interference channels,” IEEE Trans. Veh. Technol., vol. 60, no. 1, pp. 206–218, Jan. 2011. [PL] K. Pahlavan and A. H. Levesque, Wireless information networks. [PSS05] S. U. Pillai, T. Suel, and C. Seunghun, “The Perron-Frobenius theorem: some of its applications,” IEEE Signal Process. Mag., vol. 22, pp. 62–75, Mar. 2005. [QZH09] L. Qian, Y. J. Zhang, and J. W. Huang, “MAPEL: Achieving global optimality for a non-convex wireless power control problem,” IEEE Trans. Wireless Commun., vol. 8, no. 3, pp. 1553–1563, 2009. [RFLT98] F. Rashid-Farrokhi, K. J. R. Liu, and L. Tassiulas, “Transmit beamforming and power control for cellular wireless systems,” IEEE J. Sel. Areas Commun., vol. 16, no. 8, p. 1437–1450, Oct. 1998. [RG12] M. Rezaee and M. Guillaud, “Limited feedback for interference alignment in the K-user MIMO interference channel,” in Proc. IEEE Inf. Theory Workshop (ITW’12), Lausanne, Switzerland, 2012, pp. 667– 671. [RV11] A. Raja and P. Viswanath, “Diversity-multiplexing tradeoff of the two-user interference channel,” IEEE Trans. Inf. Theory, vol. 57, no. 9, pp. 5782–5793, Sep. 2011. [Sac03] J. S. Sachs. (2003) Readiness for the networked world. Centre for International Development at Harvard University. [Online]. Available: http://cyber.law.harvard.edu/readinessguide/ 158 Bibliography [Sat81] H. Sato, “The capacity of the Gaussian interference channel under strong interference,” IEEE Trans. Inf. Theory, vol. 27, pp. 786–788, Nov. 1981. [SB04] M. Schubert and H. Boche, “Solution of the multiuser downlink beamforming problem with individual SINR constraints,” IEEE Trans. Veh. Technol., vol. 53, no. 1, pp. 18–28, Jan. 2004. [SB06] ——, QoS-Based Resource Allocation and Transceiver Optimization. Foundations and Trends in Communications and Information Theory, 2006, vol. 2. [SBM09] M. A. Sadrabadi, A. Bayesteh, and E. Modiano, “Opportunistic scheduling in large-scale wireless networks,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT’09), Seoul, Korea, 2009. [SFRS12] N. Schrammar, H. Farhadi, L. K. Rasmussen, and M. Skoglund, “Average throughput in AWGN cognitive fading interference channel with multiple secondary pairs,” in Proc. Int. Conf. Cognitive Radio Oriented Wireless Networks (CROWNCOM ’12), 2012. [SH05] M. Sharif and B. Hassibi, “On the capacity of MIMO broadcast channels with partial side information,” IEEE Trans. Inf. Theory, vol. 51, no. 2, pp. 506–522, Feb. 2005. [Sha14] N. Shariati, “Robust transmit signal design and channel estimation for multiantenna systems,” Ph.D. dissertation, ACCESS Linnaues Centre, Signal Processing Laboratory, KTH Royal Institute of Technology, 2014. [SKC09] X. Shang, G. Kramer, and B. Chen, “A new outer bound and the noisy interference sum rate capacity for Gaussian interference channels,” IEEE Trans. Inf. Theory, vol. 55, no. 2, pp. 689–699, Feb. 2009. [SN07a] S. Sanayei and A. Nosratinia, “Opportunistic beamforming with limited feedback,” vol. 6, no. 8, pp. 2765–2771, Aug. 2007. [SN07b] ——, “Opportunistic downlink transmission with limited feedback,” IEEE Trans. Inf. Theory, vol. 53, no. 11, pp. 4363–4372, Nov. 2007. [SPLL10] H. Sung, S. H. Park, K. J. Lee, and I. Lee, “Linear precoder designs for K-user interference channels,” IEEE Trans. Wireless Commun., vol. 9, no. 1, pp. 291–301, Jan. 2010. [SS77] A. Sripad and D. Snyder, “A necessary and sufficient condition for quantization errors to be uniform and white,” IEEE Trans. Acoust., Speech, Signal Process., vol. 25, no. 5, pp. 442–448, Oct. 1977. Bibliography 159 [SWB06] S. Sta´ nczak, M. Wiczanowski, and H. Boche, Resource allocation in wireless networks: theory and applications. Springer, 2006. [SWB14] N. Shariati, J. Wang, and M. Bengtsson, “Robust training sequence design for correlated MIMO channel estimation,” IEEE Trans. Signal Process., vol. 62, no. 1, pp. 107–120, Jan. 2014. [SYY12] C. Sun, Y. Yang, and Y. Yuan, “Low complexity interference alignment algorithms for desired signal power maximization problem of MIMO channels,” EURASIP J. Adv. in Signal Process., vol. 2012, no. 137, 2012. [TCS11] C. W. Tan, M. Chiang, and R. Srikant, “Maximizing sum rate and minimizing MSE on multiuser downlink: Optimality, fast algorithms, and equivalence via max-min SINR,” IEEE Trans. Signal Process., vol. 59, no. 12, pp. 6127–6143, Dec. 2011. [TJC99] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space-time block codes from orthogonal designs,” IEEE Trans. Inf. Theory, vol. 45, no. 5, pp. 1456–1467, Jul. 1999. [Tun08] D. Tuninetti, “Gaussian fading interference channels: Power control,” in Proc. Asilomar Conf. Signals, Systems and Computers, CA., USA, Oct. 2008. [TV05] D. Tse and P. Viswanath, Fundamentals of wireless communication. Cambridge Univ. Press, 2005. [TW08] A. Tajer and X. Wang, “Opportunistic multi-antenna downlink transmission with finite-rate feedback,” in Proc. Ann. Allerton Conf. Commun., Control, and Computing, Sep. 2008. [VH94] S. Verdú and T. S. Han, “A general formula for channel capacity,” IEEE Trans. Inf. Theory, vol. 40, no. 4, pp. 1147–1157, Jul. 1994. [VTL02] P. Viswanath, D. N. C. Tse, and R. Laroia, “Opportunistic beamforming using dumb antennas,” IEEE Trans. Inf. Theory, vol. 48, no. 6, pp. 1277–1294, Aug. 2002. [WE06] I. Wong and B. Evans, “Exploiting spatio-temporal correlations in MIMO wireless channel prediction,” in Proc. IEEE Global Telecom. Conf. (GLOBECOM’06), San Francisco, CA, USA, Nov. 2006, pp. 1–5. [WFS10] C. Wang, H. Farhadi, and M. Skoglund, “On the degrees of freedom of parallel relay networks,” in Proc. IEEE Global Commun. Conf. (GLOBECOM’10), Miami, FL., USA, Dec. 2010. 160 Bibliography [WFS12] ——, “Achieving the degrees of freedom of wireless multi-user relay networks,” IEEE Trans. Commun., vol. 60, no. 9, pp. 2612–2622, Sep. 2012. [WGM07] X. Wang, G. B. Giannakis, and A. G. Marques, “A unified approach to QoS-guaranteed scheduling for channel-adaptive wireless networks,” Proc. IEEE, vol. 95, no. 12, pp. 2410–2431, Dec. 2007. [WT11] Y. Weng and D. Tuninetti, “Outage analysis of block-fading Gaussian interference channels,” IEEE Trans. Inf. Theory, vol. 57, no. 10, pp. 6487–6501, Oct. 2011. [WXG+ 13] Y. Wu, C. Xiao, X. Gao, J. D. Matyjas, and Z. Ding, “Linear precoder design for MIMO interference channels with finite-alphabet signaling,” IEEE Trans. Commun., vol. 61, no. 9, pp. 3766–3780, Sep. 2013. [XCS01] L. Xin, E. K. P. Chong, and N. B. Shroff, “Opportunistic transmission scheduling with resource-sharing constraints in wireless networks,” IEEE J. Sel. Areas Commun., vol. 19, no. 10, pp. 2053–2064, Oct. 2001. [XYMN11] B. Xie, L. Yang, H. Minn, and A. Nosratinia, “Interference alignment under training and feedback constraints,” in Proc. Global Telecom. Conf. (GLOBECOM’11), Dec. 2011, pp. 1–6. [Yat95] R. D. Yates, “A framework for uplink power control in cellular radio systems,” IEEE J. Sel. Areas Commun., vol. 13, no. 7, pp. 1341–1347, Sep. 1995. [YSJP13] H. J. Yang, W. Y. Shin, B. C. Jung, and A. Paulraj, “Opportunistic interference alignment for MIMO interfering multiple-access channels,” IEEE Trans. Wireless Commun., vol. 12, no. 5, pp. 2180–2192, May 2013. [YSJS14] H. J. Yang, W. Y. Shin, B. C. Jung, and C. Suh, “Opportunistic interference alignment for MIMO interfering broadcast channels,” in Proc. IEEE Int. Conf. Acoustics, Speech and Signal Process. (ICASSP’14), Florence, Italy, May 2014. [YT03] Z. Yang and L. Tong, “Exploiting protocol information for transmission over unknown fading channels,” in Proc. IEEE Int. Symp. Info. Theory (ISIT’03), Kanagawa, Japan, Jun. 2003, pp. 388–388. [YTJK10] C. M. Yetis, G. Tiangao, S. A. Jafar, and A. H. Kayran, “On feasibility of interference alignment in MIMO interference networks,” IEEE Trans. Signal Process., vol. 58, no. 9, pp. 4771–4782, Sep. 2010. Bibliography 161 [Zan92] J. Zander, “Distributed cochannel interference control in cellular radio systems interference,” IEEE Trans. Veh. Technol., vol. 41, no. 3, pp. 305–311, Aug. 1992. [ZCLB07] T. Zijian, R. C. Cannizzaro, G. Leus, and P. Banelli, “Pilot-assisted time-varying channel estimation for OFDM systems,” IEEE Trans. Signal Process., vol. 55, no. 5, pp. 2226–2238, May 2007. [Zen92] J. Zender, “Performance of optimum transmitter power control in cellular radio systems,” IEEE Trans. Veh. Technol., vol. 41, Feb. 1992. [Zet] P. Zetterberg, “The multi-antenna, multi-node, multi-band, multicell (four multi) framework: A user guide,” accessed June 20, 2013. [Zet14] ——, “Interference alignment (IA) and coordinated multi-point (CoMP) with IEEE802.11ac feedback compression: testbed results,” in Proc. IEEE Int. Conf. Acustics, Speech, and Signal Process. (ICASSP’14), Florence, Italy, May 2014. [ZKAQ01] J. Zander, S. Kim, M. Almgren, and O. Queseth, Radio Resource Management for Wireless Networks. Norwood, MA, USA: Artech House, Inc., 2001. [ZM12] P. Zetterberg and N. N. Moghadam, “An experimental investigation of SIMO, MIMO, interference-alignment (IA) and coordinated multi-point (CoMP),” in Proc. Int. Conf. Systems, Signals and Image Process. (IWSSIP’12), Vienna, Austria, Apr. 2012, pp. 211–216.
© Copyright 2024