THE KLUWER INTERNATIONAL SERIES IN ENGINEERING AND COMPUTER SCIENCE WIRELESS OFDM SYSTEMS How to make them work? edited by Marc Engels IMEC, Belgium KLUWER ACADEMIC PUBLISHERS NEW YORK, BOSTON, DORDRECHT, LONDON, MOSCOW eBook ISBN: 0-306-47685-1 Print ISBN: 1-4020-7116-7 2002 Kluwer Academic Publishers New York, Boston, Dordrecht, London, Moscow Print 2002 Kluwer Academic Publishers All rights reserved No part of this eBook may be reproduced or transmitted in any form or by any means, electronic, mechanical, recording, or otherwise, without written consent from the Publisher Created in the United States of America Visit Kluwer Online at: http://kluweronline.com and Kluwer's eBookstore at: http://ebooks.kluweronline.com Dordrecht In memory of my father who died on 5 January 2002 How to make them work? Contents List of Figures List of Tables Preface Contributors Acknowledgements ix xv xvii xix xxiii Introduction 1.1 A connected world emerges 1.2 Wireless OFDM: the next technology wave 1.3 Wireless OFDM systems 1.4 Structure of the book Understanding the indoor environment 2.1 Introduction 2.2 Propagation losses 2.3 Multipath propagation 2.4 Time variant channels 2.5 Conclusions 1 1 3 5 7 11 11 12 17 26 30 The OFDM Principle 3.1 The OFDM principle 3.2 The OFDM system model 33 33 38 viii OFDM Systems 3.3 3.4 3.5 3.6 What if the channel is time-variant? OFDM receiver performance Coding: an essential ingredient Summary When people agree on OFDM 4.1 4.2 4.3 WLAN standards HIPERLAN/2 Differences between HIPERLAN/2 and IEEE 802.11a Beating the wireless channel 5.1 5.2 5.3 5.4 Introduction Channel models and characteristics One-Dimensional Channel Estimators Two-Dimensional Channel estimators. Avoiding a tower of Babel 6.1 6.2 6.3 6.4 Introduction Effects of out of sync transmission Timing synchronisation Frequency synchronisation Living with a real radio 7.1 7.2 7.3 7.4 7.5 Introduction How the front-end impairs the OFDM modem A system simulation tool Analysis and simulation of the main front-end effects Conclusions Putting it all together 8.1 8.2 8.3 8.4 Introduction The basedband signal processing ASIC The discrete system set-up Learning from results Abbreviations Variables Notation Index 41 45 48 50 53 53 54 73 75 75 76 80 90 95 95 96 100 106 113 113 115 122 127 149 151 151 155 171 178 191 195 199 201 How to make them work? List of Figures Figure 1.1. World-wide number of Internet users 1 Figure 1.2. World-wide number of mobile phones 2 Figure 1.3. Wireless Internet Technologies 3 Figure 1.4. Broad-band channel response 4 Figure 1.5. Single frequency network 5 Figure 2.1. Typical interference in 2.4 GHz ISM-band coming from a microwave oven (a) or a Bluetooth hopper (b) 17 Figure 2.2. multipath propagation situation 18 Figure 2.3. Ray-tracing example (reflections in a corner) 22 Figure 2.4. Floorplan for ray tracing example 1 24 Figure 2.5. Power delay profile and baseband frequency response for ray tracing example 1 24 Figure 2.6. Floorplan for ray tracing example 2 25 Figure 2.7. Power delay profile and baseband frequency response for ray tracing example 2 25 Figure 2.8. Time correlation for different models of mobility 29 Figure 2.9. Doppler spectra for different mobility models 30 Figure 3.1. Subdivision of the bandwidth into subbands 33 Figure 3.2. multicarrier modulation 34 Figure 3.3. Spectrum of an OFDM signal 36 Figure 3.4. OFDM demodulation 37 Figure 3.5. Cyclic Prefix 37 Figure 3.6. Discrete-time baseband equivalent model of an OFDM system 38 Figure 3.7. Discrete-frequency representation for the Doppler multipath channel and an OFDM receiver 44 Figure 3.8. OFDM performance for AWGN 45 x OFDM Systems Figure 3.9. OFDM-QPSK performance in a multi-path channel versus signal-to-noise ratio for various speeds of the mobile, subcarriers, 47 Figure 3.10. Forward error coding for OFDM 49 Figure 3.11. Coded OFDM performance in a multipath channel 50 Figure 4.1: HIPERLAN/2 protocol stack in the AP 55 Figure 4.2. Segmentation and reassembly operation 56 Figure 4.3. MAC frame structure 57 Figure 4.4. Broadcast PDU train with preamble 61 Figure 4.5. Downlink PDU train with preamble 61 Figure 4.6. Uplink PDU train with short preamble 62 Figure 4.7. Direct link PDU train with preamble 62 Figure 4.8. PHY layer reference configuration 63 Figure 4.10. Scrambler block diagram 65 Figure 4.12. Channel coder block diagram 65 Figure 4.13. Mother convolutional encoder block diagram 65 Figure 4.14. Mapping of data and pilot carriers 68 Figure 4.15. Payload section consisting of several OFDM symbols with CP 68 Figure 4.16. PHY burst format 69 Figure 4.17. Broadcast burst preamble structure 69 Figure 4.18. Overview of different PHY bursts: (a) broadcast burst, (b) Downlink burst, (c) Uplink burst with short preamble, (d) Uplink burst with long preamble, (e) Direct link burst 71 Figure 4.19. Transmit spectral mask 72 Figure 5.1. The OFDM system model 77 Figure 5.2. The "OFDM Channel" is a set of parallel Gaussian channels 78 Figure 5.3. Resampling a non-sample-spaced channel extends the channel length 79 Figure 5.4. Early and Late Synchronisation result in a longer channel and can lead to violation of the Cyclic Prefix condition 79 Figure 5.5. The ML estimator enables low complexity and time-frequency interpretation 85 Figure 5.6. Comb Spectrum for 85 Figure 5.7. FFT-based approaches outperform the SVD-based approaches by an order of magnitude for spectral shaping systems 86 Figure 5.8. Performance of ML and approximate LMMSE estimators 87 Figure 5.9. Simplified Time-Frequency grids in OFDM 91 Figure 6.1. Effect of early and late synchronisation 97 Figure 6.2. Principle of the Schmidl and Cox auto-correlation based timing synchronisation circuit 101 How to make them work? xi Figure 6.3. A sequence of inverted training symbols followed by a sequence of identical training symbols give a more accurate timing acquisition. 102 Figure 6.4. Principle of a cross-correlation frame synchronizer 103 Figure 6.5. Principle of a frame synchroniser based on the cyclic prefix.. 104 Figure 6.6. Transforming a training sequence into a cyclic prefix 105 Figure 6.7. Timing estimation accuracy in function of SNR, with the number of training sequences Mas a parameter 106 Figure 6.8. Principle of Moose auto-correlation based frequency synchronisation circuit 107 Figure 6.9. Frame/frequency synchroniser based on the cyclic prefix 109 Figure 6.10. Frequency estimation accuracy in function of SNR, with the number of training sequences M as a parameter 110 Figure 7.1. Simplified schematic of an OFDM transceiver 114 Figure 7.2. I/Q demodulation with I/Q imbalance 117 Figure 7.3. The implementation loss versus the difference between the power of an additional Gaussian noise source and the channel noise power resulting in a given BER is described by a unique curve 119 Figure 7.4. Implementation loss IL versus the noise power of an additional Gaussian noise source, for powers ranging from 25 dBc to 45dBc. This plot relates to 64 QAM transmission at BER of and 120 Figure 7.5. BER curve for coded 64QAM, with coding rate on an AWGN channel without any front-end effect 121 Figure 7.6. Schematic of the full link model as modeled in MATLAB for this study 123 Figure 7.7. BER curves for 52 non-zero 64QAM-modulated subcarriers with a coding rate of 3/4, showing a 3dB implementation loss due to corruption of the long training symbol not corrected in the equalizer. 124 Figure 7.8. The BER curve for uncoded 64QAM resulting from simulation matches the theoretical one 126 Figure 7.9. The optimal clipping level depends on the word-length of the transmitted symbols 128 Figure 7.10. BER curves, showing the implementation loss due to clipping and quantization noise 129 Figure 7.12. Effect of phase noise on 1 OFDM symbol . Legend: * transmitted symbol, received symbol 133 Figure 7.13. Effect of phase noise on all OFDM symbols in the same burst. Legend: * transmitted symbol, received symbol 133 Figure 7.14. Phase noise spectrum 134 Figure 7.15. Effect of phase noise on BER for uncoded 64QAM 137 xii OFDM Systems Figure 7.16. Domains of clipping after clipping operation (on magnitude of I and Q separately) and after clipping operation (on magnitude of ) 139 Figure 7.17. For an input signal clipped at and with +6dBm average power, the power amplifier is driven up to the limit of saturation for 141 Figure 8.7. Robust timing acquisition relies on preamble auto- correlation in combination with signal power monitoring 162 Figure 8.8. The carrier frequency offset estimate feeds a phase accumulator and a CORDIC to limit CFO on the signal entering the FFT 163 Figure 7.19. The IL is 0.5 dB for at ... 142 Figure 7.20. ADC clock jitter effects are included before conversion by an ideal ADC 143 Figure 7.22. SIR versus synchronization location for three test cases: no analog channel select filter, channel select filters A and B with similar inband ripple (no adjacent channels). Only filter A shows fairly high SIR for a large synchronization range 145 Figure 7.23. Influence of the filter impulse response on the BER performances for the 3 filters already considered in Figure 7.22, with synchronization on the sample that shows maximum SIR (uncoded 64QAM in AWGN channel). As expected, Filter B shows poor BER performances 145 Figure 7.24. SIR performance of the filter architecture A in dispersive channels: the SIR is quite insensitive to the multipath channel impulse response 146 Figure 7.25. SIR performance of the filter architecture B in dispersive channels: this architecture provides similar channel selection performances, but shows much more sensitivity to the multipath channel impulse response 146 Figure 7.26. BER performances evaluation for the example in a Gaussian channel 148 Figure 8.1. A wireless webcam scenario was chosen as application scenario. 152 Figure 8 2. Detailed wireless webcam communication scenario 153 Figure 8.3. The design goals for the ASIC are divided upon concept, implementation, and methodology to make sure that we arrive at an implementable solution 155 Figure 8.4. Partitioning of the ASIC 157 Figure 8.5. A radix 2-2 scheme significantly reduces the arithmetic complexity of the fast Fourier transform 159 Figure 8.6. The symbol-based samplereordering (SSR) unit essentially allows a set of intra-symbol data transfer operations based on a generic architecture 160 How to make them work? xiii Figure 8.9. The Festival equalizer reveals a low-cost solution with feed- forward channel estimation and feedback decision-directed tracking. 163 Figure 8.10. The Carnival equalizer requires an interpolator and divider in addition to the Festival equalizer since modulation schemes up to 64- QAM need to be processed 164 Figure 8.11. Simple, reference symbol-based channel estimation reveals poor noise-influenced results. The interpolator improves the channel estimate S/N by 2.5 to 3 dB 165 Figure 8.12. The impulse response is truncated and interpolated using a fully programmable transformation matrix operation 165 Figure 8.13. Clock offset is tracked by guard interval correlation and averaging over multiple OFDM symbols 166 Figure 8.14. The receiver and transmitter token flow exploits a closed loop token scheme 168 Figure 8.15. The object-oriented desing flow starts from C++ and ends in a conventional HDL-based design flow 170 Figure 8.16. 5 GHz WLAN demonstration setup excluding the power amplifier board 171 Figure 8.17. Software protocol stack for the webcam application (Windows) and for file transfer or test purposes (Linux) 172 Figure 8.18. The FPGA contains a central processing unit (CPU) that coordinates both data transfers (RX and TX) and signalling information (serial protocol, MPI) 173 Figure 8.19. Digital baseband board. Different versions exist for the Festival and Carnival ASICs because they are not pin-compatible 175 Figure 8.20. Baseband signal processing 175 Figure 8.22. Architecture of the digital IF implemented in an FPGA 176 Figure 8.23. The integrated front-end has a superheterodyne architecture. The discrete board set-up uses a similar architecture 177 Figure 8.24. Carnival ( CMOS, left) outperforms Festival ( CMOS, right) at the cost of an 30 % area increase only 179 Figure 8.25. Non-compliant burst (lower) and preamble (upper) format in Festival and Carnival ASICs 183 Figure 8.27. Proposed automatic gain control architecture 185 Figure 8.28. Signal-to-noise ratio and estimation error of the quantized ADC output 185 Figure 8.29. Generic transmitter architecture 187 How to make them work? List of Tables Table 2.1. Channel characteristics for set-up 1 and set-up 2 26 Table 4.1. Number of OFDM symbols per transport channel excluding physical layer preambles in function of the modulation and code rate R 62 Table 4.2. Mode dependent parameters 64 Table 4.3. OFDM parameters 67 Table 4.4. HIPERLAN/2 carrier frequencies and EIRP 72 Table 4.5. Comparison between HIPERLAN/2 and IEEE 802.11a 73 Table 7.1. Implementation Loss on uncoded 64QAM at a BER of due to quantizing and clipping for variable word-length b and clipping at 129 Table 7.2. Implementation Loss on uncoded 64QAM at a BER of due to I/Q imbalance 131 Table 7.4. Simulated Implementation Loss on uncoded 64QAM at a BER of due to magnitude clipping, for different normalized clipping factors (normalized to ) 139 Table 8.1. Both Festival and Carnival shall be highly flexible, programmable ASICs 154 Table 8.2. Carnival outperforms Festival with respect to spectral efficiency and energy efficiency at a moderate increase in area despite a significantly higher complexity 180 Table 8.3. There is an explosion in code size during top-down refinement steps 180 Table 8.4. A fair reuse percentage between the two designs was obtained despite significant algorithmic changes 181 How to make them work? Preface Wireless Local Area Networks (WLANs) experience a growing popularity recently. Where WLANs were primarily used for niche applications in the past, they are now deployed as wireless extensions to computer networks. The increase of the datarates from 2 Mbps up to 11 Mbps for roughly a constant price has played a major role in this breakthrough. As a consequence, an even greater success can be envisioned for the more recent OFDM-based WLAN standards in the 5 GHz band, which offer up to 54 Mbps. At IMEC we have realised this potential already several years ago and have established a successful research programme on OFDM- based WLAN. In 1995, we started our research on wireless OFDM in the frame of a co- operation project with SAIT, a Belgian telecom company. The goal of the project was to establish a robust network for industrial environments. This resulted in a first OFDM chip, supporting QPSK, for wireless networking at the end of the project (1999). 1999 was also the start of an intense co-operation with National Semiconductor Inc., which resulted in a second generation ASIC in 2000. This OFDM processor supports up to QAM-64 and has a more optimal channel estimation algorithm. Meanwhile, we have set-up a co-operation program, which currently includes more then 10 partners. In the program we expanded our activities from the digital baseband signal processing towards the 5GHz front-end and the MAC implementation. We also defined a roadmap to realise WLAN systems with an indoor range up to 100 meters and a capacity beyond 100 Mbps. The first goal requires for techniques like adaptive loading and turbo xviii OFDM Systems coding. For increasing the network capacity, multiple antenna techniques play an essential role. To let a wider audience profit from our long-time experience with implementing WLAN OFDM systems, we also developed a tutorial, which has been delivered several times with great success. Recently, however, the interest for wireless OFDM is spreading in new application domains with a rapid increase of the interested audience as a consequence. For instance, wireless access standards are adopting OFDM based solutions, it was proposed for wireless personal area networks and it is heavily promoted as a candidate for next generation mobile networks. This led us to the idea to put our tutorial material down into a manuscript. The result of this effort is in front of you. We hope that you enjoy reading it and that it is useful in your professional work. Marc Engels How to make them work? Contributors Boris Come is the leader of the architecture design team in the mixed- signal and RF applications (MIRA) group of IMEC. Boris was born in France in 1974. He graduated from the National Engineering School of Electronics in Toulouse, ENSEEIHT, in June 1998. As part of his studies, he performed two internships: the first one was a three-month summerstay in ESA-ESTEC in 1997; the second one a four month internship, from March to June 1998, at IMEC. After graduating, he started working in the MIRA group at IMEC. His main research interests are in the design of mixed-signal and RF front-ends for digital telecom system. For the past 2 years, he has been working on system specification and architecture definition for HIPERLAN/2 and IEEE 802.1la compliant WLAN transceivers. His main focus was the RF module for which a single package solution is targetted. Luc Deneire received the Engineering degree in Electronics from the University of Liege (Belgium) in 1988, the Engineering degree in Telecommunications from the University of Louvain-La-Neuve in 1994 and the Ph.D. degree in Signal Processing at Eurecom, Sophia-Antipolis, France in 1998. During this time, he was a Marie Curie Fellow. In 1999, he was consultant for Texas Instruments, Villeneuve-Loubet, France, for the UMTS base-station signal processing requirements. Since late 1999, he is a senior researcher at IMEC, the largest European independent research institute in Microelectronics. He is working on the signal processing algorithms involved in wireless communications, specifically for third generation mobile network, Wireless LANs and Wireless Personal Area Networks. His main interests are blind and semi-blind equalisation and channel estimation, modulation theory, multiple access schemes, smart antennas and link adaptation. He is the author of more than 40 conference and journal papers. xx OFDM Systems Wolfgang Eberle received the M.S. degree in Electrical Engineering from Saarland University, Saarbruecken, Germany, in 1996 with specialization in microwave engineering and telecommunication networks. He joined the Wireless Systems Group of IMEC in 1997 working on algorithm development and digital VLSI architecture design for OFDM- based wireless LAN modems. In 2000, he joined the Mixed-Signal and RF Applications Group of IMEC where he now focusses on system-level mixed- signal aspects including digital compensation of receiver nonidealities, power-efficient transmitters, and design methodologies, applied to wireless LANs. He is also working towards the Ph.D. degree in Electrical Engineering at the Katholieke Universiteit Leuven, Belgium. Marc Engels is co-founder of LoraNet, a new company in fixed wireless access, and responsible for research and product development. The company will focus on sub 11 GHz systems that operate under non line-of-sight conditions. Technologies involved are Orthogonal Frequency Division Multiplexing (OFDM) and Space division multiple access (SDMA). Before, Marc Engels was the director of the wireless department at IMEC, focussed on the implementation of telecommunication systems on a chip. For these systems, he overlooked research on the DSP processing, the mixed-signal RF front-end and the software protocols. He was also active in design methods and tools for implementing multi-disciplinary systems. Under his supervision, several systems have been realised, including a 54 Mbps WLAN terminal, a GPS-GLONASS receiver, a DECT-GSM dual mode phone, a cable modem, etc. Previously, Marc performed research at the Katholieke Universiteit Leuven, Belgium, Stanford University, CA, USA, and the Royal Military School, Brussels, Belgium. Marc Engels received the engineering degree (1988) and the Ph.D. (1993), both from the Katholieke Universiteit Leuven, Belgium. Marc Engels is a visiting professor of telecom system design at the Katholieke Universiteit Leuven and of embedded system design at the University of Lugano, Switzerland. He is an active member of the KVIV telecommunications society and URSI, secretary of the IEEE Benelux chapter on vehicular technology and telecommunications and member of the board of directors of SITEL. He is currently an associated editor for the Wirelss Personal Communications journal and was associated editor of IEEE transactions on VLSI in 1999-2000. Bert Gyselinckx is heading the Wireless Systems group of IMEC. This group performed projects in the fields of WLAN, broadband satellite communication, navigation systems and cellular communication. His main research interests are in spread-spectrum, wireless communications and VLSI systems. Bert received the M.S. degree in Electrical Engineering from How to make them work? xxi the Rijksuniversiteit Gent, Belgium, in 1992 and the the M.S. degree in Air and Space Electronics from the Ecole Nationale Superieure de l'Aeronautique et de l'Espace, Toulouse, France, in 1993. Previously he worked for the Research and Development group of Siemens in Munich, Germany. Jean-Paul Linnartz is a Department Head with the Natuurkundig Laboratorium (Nat.Lab.) of Philips Research, at Eindhoven, The Netherlands. Here, as a Principal Scientist he studied the protection of audio and video, in particular through the technology of electronic watermarking. In 1992-1993,he was an Assistant Professor at the University of California at Berkeley, where he worked on random access for wireless networks. In 1993, he was the first to use the name Multi-Carrier CDMA in one of the first papers on the combination of OFDM with CDMA. In 1991, he received his Ph.D. cum laude on multi-user mobile radio networks from Delft University of Technology, The Netherlands. He has twenty (pending) patents in the field of electronic watermarking, copy protection and radio communications. He authored over 100 papers, he is founding Editor-in- Chief of "Wireless Communication, The Interactive Multimedia CD-ROM", and he has been guest editor for two special journal issues on Multi-Carrier Modulation. Reto Ness received the engineering degree in Electrical Engineering from the University of Karlsruhe, Germany and from the Ecole Nationale Suprieure dElectronique et de Radiolectricit de Grenoble, France, in 1999. He carried out his thesis in the Wireless Systems group at IMEC,, Belgium, where he focussed on narrow-band interference cancellation in OFDM-based WLANs. Currently, he works in the development department of Tenovis GmbH & Co. KG, Germany. Steven Thoen was born in Leuven, Belgium, in 1974. He received the diploma of electrical engineering from the Catholic University of Leuven, Leuven, Belgium in 1997. In October 1997 he joined IMEC where he is currently finishing the Ph.D. degree from the Catholic University of Leuven, Leuven, Belgium. During this period, he spent 6 weeks as a visiting researcher at the Information Systems Lab, Stanford University, Palo Alto, California, USA. His current research interests lie in the area of digital communication theory including multiple antenna systems, OFDM modulation, adaptive modulation and wireless systems. He has authored several papers and one patent on these topics. Jan Tubbax is a PhD student at the K.U.Leuven, Who performs its research in the Wireless Systems (WISE) group of IMEC. The subject of his xxii OFDM Systems Ph. D. research is the design of a high-performance, low-cost wireless LAN system Jan Tubbax received his M.S. degree as Electrical Engineer, telecommunications in 1998 from the Department of Electrical Engineering (ESAT) at the Katholieke Universiteit Leuven, Belgium. The subject of his M.S. thesis was a study on the performance of protocols for wireless and mobile communications. Patrick Vandenameele is the chief systems architect for Resonext Communications, a fab-less semiconductor company based in San Jose, CA, USA, developing and marketing end-to-end two-chip Wireless LAN solutions. Patrick is responsible for the specifications and architectures of both the PHY and MAC functions. Also, he leads the companys systems engineering team located in Leuven, Belgium. Before joining Resonext Communications, he was a researcher in the Wireless Systems group at IMEC. His research, funded by an IWT scholarship, resulted in low- complexity detection algorithms for OFDM/SDMA, including solutions to real-world problems such as channel estimation, synchronization, power control and the integration in a multiple-access protocol. Patrick received the engineering degree (1996) and the Ph.D. degree (2000), both from the Katholieke Universiteit Leuven, Belgium. His thesis, entitled Space Division Multiple Access for Wireless LANs, was published by Kluwer Academic Publishers. During his studies, he did internships at ST-Microelectronics, Crolles, France; Sirius, Montpellier France; ENST, Paris, France; KTH- Electrum, Kista, Sweden; and the Smart Antennas Research Group at Stanford University, CA, USA. Liesbet Van der Perre received the M.Sc. degree in Electrical Engineering from the K.U.Leuven, Belgium, in 1992. She performed her M.Sc. thesis research at the ENST in Paris, France. She received the Ph.D. degree in Electrical engineering from the K.U.Leuven in 1997. Currently, she is the director of IMECs wireless program. Her work focuses on system design and digital modems for high-speed wireless communications. She was a system architect in IMECs OFDM ASICs development, and the leader of the turbo coding team. Also, she is a part-time professor at the University of Antwerp, Belgium. How to make them work? xxiii Acknowledgements This book was only possible with the help and support of many people. In the first place, I like to thank all the authors that contributed to the various chapters. I am particularly indebted to Liesbet Van der Perre, who also contributed to the concept of the book. The material in the chapters is the result of the wireless program at IMEC. I am grateful to all the people that worked with me in this program during the last 7 years and realised these excellent results. Finally, a word of thanks is due to my wife Els and my three daughters Heleen, Laura and Hanne for their patience and support. Marc Engels Chapter 1 Introduction The world goes wireless! Marc Engels IMEC 1.1 A CONNECTED WORLD EMERGES About a decade ago, the telecommunication infrastructure was targeted towards fixed analogue telephony, with support for voice and narrow band data communication. In 10 years, the digitalisation of the communication infrastructure, the support for broadband access at home and the success of mobile phones have changed this situation completely. End-to-end digitalisation of the telecom network was achieved with the introduction of the Integrated Services Digital Network (ISDN) [1]. 2 Chapter 1 However, market success was limited due to the lack of interesting services. This changed dramatically with the fast take-up of the Internet by business and residential customers in the 90s (see Figure 1.1). The success of the Internet rapidly created a market for data pipes to individual users. The 128 kbps of ISDN was no longer considered sufficient. Hence, Asymmetrical Digital Subscriber Line (ADSL) [2] and cable modem [3] technologies were developed to increase the datarate to several hundreds kbps, with a theoretical maximum of 10 Mbps. Very high speed Digital Subscriber Line (VDSL) technology [4] is currently investigated. It will further increase the datarate up to 52 Mbps. In the same time frame, digital mobile communication experienced an extraordinary success: the number of mobile phones doubled every two years (see Figure 1.2). Europe was a forerunner in this evolution because of the unified GSM (Groupe Speciale Mobile) standard [5]. In the U.S., on the contrary, a multitude of standards was deployed, IS-136 [6], based on Time- Division Multiple Access (TDMA) and IS-95 [7], based on Code Division Multiple Access (CDMA). The logical next step would be to bring these two trends together and create the wireless Internet. The wireless Internet will not be a single technology, but rather a collection of different systems (see Figure 1.3). Third generation mobile systems [8] will be the wide area network (WAN) that provides up to 2 Mbps to stationary users, while offering up to 384 kbps towards fast moving terminals. Research for fourth generation systems, that will provide at least 10 Mbps, has already started [9]. For indoor wireless local area networks (WLAN), standards (HIPERLAN/2 in Europe [3]; IEEE 802.11 in US [11]) are available that offer up to 54 Mbps. Also for the Introduction 3 wireless local loop (WLL), the wireless first mile solution, standards with similar performances are under construction (HIPERMAN in Europe and IEEE 802.16a in US). Further generations of these WLAN and WLL standards with datarates beyond 100 Mbps can be anticipated. Extension of the wireless Internet to a large set of personal devices is foreseen with personal area networks (PAN). An example is the current bluetooth system [12] that offers a peak data rate of 721 kbps. 1.2 WIRELESS OFDM: THE NEXT TECHNOLOGY WAVE The systems that constitute the wireless Internet operate in largely different environments. Some can still be considered as narrow-band (e.g. bluetooth) while other, like WLANs, have broadband characteristics. For some systems the performance is limited by the signal-to-noise ratio (SNR) and for others the interference from other cells is the limiting factor. To adapt to these environmental aspects the wireless systems apply various air interfaces. They can be grouped in three main technology waves that consecutively appeared and became popular. Early wireless technologies consisted of TDMA-based single carrier communication. This technology is ideally suited for narrow-band single-cell communication. Therefore it was adopted in second generation mobile phones (e.g. GSM [5]), cordless telephony (e.g. DECT [13]) and more recently in first generation personal area networks (e.g. bluetooth [12]). 4 Chapter 1 When a limited spectral efficiency is acceptable, constant modulus schemes, like Gaussian Minimum Shift Keying (GMSK)[14], are the most cost effective. To increase spectral efficiency, M-ary Phase Shift Keying (M- PSK) or M-ary Quadrature Amplitude Modulation (M-QAM) schemes are used. The performance of these schemes is limited by the received signal-to- noise ratio. In a cellular system, however, the performance is not limited by the signal-to-noise ratio but rather by the interference from neighbouring cells, called co-channel interference. For a TDMA system this implies that frequencies are only being re-used in cells which are sufficiently far from each other, reducing the overall network bandwidthefficiency. An alternative is to use systems that can cope with the co-channel interference. Code division multiple access, based on spread spectrum communication, has this property. As a consequence, it became a second technology wave that was adopted in second and third generation cellular mobile communication systems (e.g. IS-95 [7]and W-CDMA [8]) and in wireless local area networks (e.g. IEEE 802.11b [11]). In a multipath environment (e.g. indoor wireless networks) broadband communication will result is a frequency selective channel response, as shown in Figure 1.4. In such a situation, the above technologies do not work optimally and a modulation scheme is needed that is better suited to the environment. For this reason, Orthogonal Frequency Division Multiplexing (OFDM) became very popular recently. It is, for instance, used in the HIPERLAN/2 [3] and IEEE 802.11a [11] WLAN standards as well as in the draft IEEE 802.16a WLL standard [15]. As a consequence, it can be considered as the next technology wave. For instance, it is, potentially Introduction 5 combined with CDMA, one of the strong candidates for generation cellular systems. 1.3 WIRELESS OFDM SYSTEMS Although the principle of OFDM communication has been around for several decades, it was only in the last decade that it started to be used in commercial systems. The most important wireless applications that make use of OFDM are Digital Audio Broadcasting (DAB), Digital Video Broadcasting (DVB), wireless local area networks (WLAN), and more recently wireless local loop (WLL). The DAB system was originally developed by the Eureka 147 project [16]. It is a novel audio broadcasting system [17] intended to supersede the existing analog Amplitude and Frequency Modulation (AM, FM) systems. It is rugged, yet highly spectrum and power efficient. It has been designed for terrestrial and satellite as well as for hybrid and mixed delivery. DAB is seen as the future of radio as it makes more efficient use of crowded airwaves and provides CD-quality sound that is noticeably better than an FM analog broadcast. DAB broadcasts are virtually immune for interference and fading (i.e. programs are not suddenly lost when the car passes through a tunnel or under power lines). One of the principal advantages of switching to DAB is that a single frequency (called a "Multiplex") can carry up to six stereo or 12 mono services or any combination in between. DAB also allows going 6 Chapter 1 beyond audio and using some of the "Multiplex" capacity to transmit data that is not related to programming, such as financial data and digital video. DAB makes use of an OFDM transmission scheme with differential QPSK (Quadrature Phase Shift Keying) modulation. Four different modes are specified to cope with different multipath delays (up to a difference in propagation distance of 73,8 km). All of them transmit a signal with a bandwidth of approximately 1.54 MHz. One of the key benefits of using OFDM for DAB is the ability to establish single frequency networks (Figure 1.5). In a single frequency network several geographically separated transmitters, send the same signal simultaneously. As a consequence, any receiver receives the sum of these transmitted signals each with a different propagation delay. As long as the differences between these propagation delays are substantially smaller than the guard interval, the OFDM system will work with neglectible performance degradation. The DVB system [19] is very similar to the DAB standard but is intended for broadcasting of digital television signals. MPEG-2 source compression is used to reduce a standard television channel to approximately 3 Mbps and a high definition TV channel to 20 Mbps. Because of the higher datarates, the DVB system uses a 8 Mhz bandwidth. The subcarriers in the OFDM signal are also modulated with a higher order QAM constellation, with up to 64 points. Third generation WLAN systems are intended to offer high datarates in the 5 GHz frequency band. The standardization is taking place simultaneously in the U.S., Japan and Europe. The IEEE is working on the IEEE 802.11a standard for the U.S. The Multimedia Mobile Access Communication Systems (MMAC) group [14] of the Japanese Association of Radio Industries and Businesses (ARIB) establishes the HiSWANa standard. The European Telecommunication Standards Institute (ETSI) has a workgroup on Broadband Radio Access Networks (BRAN) that develops the HIPERLAN/2 standard [3]. All three standards have very similar physical layers optimized for indoor environments with a large amount of multipath with limited delays. The communication is based on OFDM in a 20 MHz bandwidth. Per subcarrier, the modulation scheme ranges from BPSK (Binary Phase Shift Keying) up to 64-QAM. Together with a variable error- coding rate, this allows the datarate to be adapted from 6Mbit/s to 54Mbit/s, depending on the propagation channel conditions. Wireless local loops provide high speed Internet access and multimedia services to fixed users. They are a competitive technology to VDSL and cable modems. Licensed frequency bands are reserved for WLL systems: 2.4 GHz in the US and 3.5 GHz in Europe. Up till now, a large variety of systems exists, keeping the production volumes low and the cost of the equipment high. To remedy this situation, which prevents success in the Introduction 7 market, a worldwide standardization effort was started. In the US, the IEEE 802.16 committee is working towards a standard medium access control (MAC) standard that can work with multiple physical layers. It defines a physical layer for bands above 11 GHz, one for the licensed bands between 2 GHz and 11 GHz (IEEE 802.16a), and one for the unlicensed bands between 2 GHz and 11 GHz (IEEE 802.16b). At the same time, ETSI BRAN has established workgroups for WLL systems below 11 GHz (HIPERMAN) and above 11 GHz (HIPERACCESS). The activities of both standardization committees are being aligned to come up with compatible systems in both standards. The sub 11 GHz standards are focussing on non line-of-sight situations that experience multipath propagation with relative long delays (up to 20 ms). They support a large variety of bandwidths with a maximum of 28 MHz, resulting in a maximum datarate of more than 100 Mbps. One of the supported modulation schemes is based on OFDM. Per subcarrier, QPSK, 16-QAM or 64-QAM is used. Also multiple error-coding rates are supporting. Modulation and error-coding rate are adapted depending on the propagation channel conditions. 1.4 STRUCTURE OF THE BOOK Because of the growing popularity of OFDM, an increasing number of engineers are involved in designing OFDM transceivers. Although the theory of OFDM is well understood, implementation aspects of OFDM systems are seldom discussed. This book fills this gap and gives a comprehensive overview of the implementation of OFDM systems. The book capitalises on the large experience of the authors with the implementation of OFDM base WLAN systems. After this first introductory chapter, we discuss the indoor propagation channel. Knowledge of the propagation properties is essential for several aspects of the receiver design (e.g. channel estimation methods). As a consequence, the first task in any wireless system design is to establish an accurate channel model. Next, we review the OFDM basics in chapter 3. This will result in a system model that will serve as a reference in the remainder of the book. Besides an overview of well-known theory, the chapter contains also some new material on the introduction of Doppler effects in this system model. The fourth chapter introduces the WLAN standards. Most material in this chapter deals with physical and medium access control layers of the HIPERLAN/2 standard. However, also a comparison with the IEEE 802.11a and the Japanese HiSWANa standards is included. 8 Chapter 1 The next two chapters are devoted to baseband implementation challenges. In chapter 5 we present several channel estimation algorithms. We show that large performance differences exist between the estimation methods. In chapter 6, we examine the various synchronisation problems and solutions for OFDM modems. An OFDM transceiver does not only consist of a baseband circuit but also needs a radio that translates the signal to and from its carrier frequency. Several effects of this radio part have a considerable influence on the performance of the OFDM system. In chapter 7, these effects are analysed in detail. Finally, in chapter 8, we put everything together and show some practical implementations of OFDM systems. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] I. Dorros, ISDN, IEEE Communications Magazine., Vol. 19, March 1981, pp. 16-19. D.L. Waring, J.W.Lechleider, T.R.Hsing, Digital Subscriber Line Technology Facilitates a Graceful Transition from Copper to Fiber, IEEE Communications Magazine, Vol. 29, March 1991, pp. 96-103. D. Fellows, D. Jones, DOCSIS cable modem technology, IEEE Communications Magazine, Vol. 39, issue 3, March 2001, pp. 202-209. J.M.Cioffi, V.Oksman, J.-J.Werner, T. Pollet, P.M.P. Spruyt, J.S. Chow, K.S. Jacobsen, Very-high-speed digital subscriber lines, IEEE Communications Magazine, Vol. 37, Issue 4, April 1999, pp. 72-79. M.Rahnema, Overview of the GSM system and protocol architecture, IEEE Communications magazine, Vol. 31, Issue 4, April 1993, pp. 92-100. N.R. Sollenberger, N. Seshadri, R. Cox, The evolution of IS-136 TDMA for third-generation wireless services, IEEE Personal Communications, Vol. 6, Issue 3, June 1999, pp. 8-18. D.N. Knisely, S. Kumar, S. Laha, S. Nanda, Evolution of wireless data services: IS-95 to cdma2000, IEEE Communications Magazine, Vol. 36, Issue 10, Oct. 1998, pp. 140-149. T. Ojanper, R. Prasad, WCDMA: Towards IP Mobility and Mobile Internet, Artech House, 2001. A. Bria, F. Gessler, O. Queseth, R. Stridh, M. Unbehaun, J. Wu, J. Zander, M. Flament, Wireless Infrastructures: sceanrios and Research Challenges, IEEE Personal Communications, Vol. 8, No. 6, December 2001, pp. 25-31. ETSI TS 101 475, Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Physical (PHY) layer. R. van Nee, G. Awater, M. Morikura, H. Takanashi, M. Webster, New high- rate wireless LAN standards, IEEE Communication Magazine, Vol. 32, No. 12, December 1999, pp. 82-88. B. Chatschik, An overview of the Bluetooth wireless technology, IEEE Communication Magazine, Vol. 39, Issue 12, December 2001, pp. 86-94. Introduction 9 [13] [14] [15] [16] [17] [18] [19] [20] W.H.W. Tuttlebee, Cordless personal communications, IEEE Communications Magazine, Vol. 30, Issue 12, Dec. 1992, pp. 42-53. K. Feher, Wireless Digital Communications: Modulation and Spread Spectrum Applications, Prentice-Hall, 1995. IEEE 802.16, http://grouper.ieee.org/groups/802/16/index.html The Eureka 147 consortium, http://www.eurekadab.org/index.html Alard M., Lasalle R., Principles of Modulation and Channel Coding for Digital Broadcasting for Mobile Receivers, EBU Review, August 1987,pp. 168-190. Le Floch B., Halbert-Lassalle R., Castelain D., Digital Sound Broadcasting to Mobile Receivers, IEEE Trans. On Consumer Electronics, Vol. 35, No. 3, August 89. Sari H., Karam G., Jeanclaude I., Transmission Techniques for Digital Terrestrial TV Broadcasting, IEEE Communication Magazine, February 1995, pp. 100-109. Multimedia Mobile Access Communication (MMAC) Systems, http://www.arib.or.jp/mmac/e/index.htm Chapter 2 Understanding the indoor environment Radio propagation models Liesbet Van der Perre, Reto Ness, Steven Thoen, Patrick Vandenameele, Marc Engels IMEC 2.1 INTRODUCTION (S)He who does not know the channel, can never be a good radio engineer, a wise man said [1]. Indeed, a good understanding of the propagation problem, is a key factor to success in designing wireless communication systems. Therefore, we dedicate this chapter to radio propagation modelling, specifically for systems based on OFDM- transmission. Traditionally, radio engineers like to dispose over channel models, which they can use in their system conception. Such models necessarily form a trade-off between the following major criteria: 1. On the one hand, the channel model should reflect as good as possible the physical propagation phenomena, in order to serve as a reliable base in the system research and design. 2. On the other hand, for system simulations and design considerations, one wants to dispose over a model which is as simple as possible. 3. Finally, in wireless transceiver design, not primarily an accurate model for one specific geometric situation, but rather a comprehensive set of channel responses (including typical and worst case situations) is needed. As such, it is evident that propagation modelling should be performed keeping in mind the application for which it is intended [2] [4]. Correspondingly to the above criteria, this chapter aims to introduce basic radio propagation and channel models for OFDMbased systems, without the 12 Chapter 2 objective of providing a complete survey of propagation mechanisms, which forms a scientific discipline by itself. The focus in this chapter is on indoor wireless propagation, as this environment is the main object of recent OFDM-based standards and system developments. We discuss three major effects: 1. Propagation losses: radio waves travelling through the ether undergo severe attenuation, as a consequence of what are generally called large- scale path losses. When doing communication system simulations in general and analysing a modems performance in particular, the average receive level is often used as a reference value. It is therefore important to get an idea of what the typical values are one can expect to find for this parameter. Section 2.2 discusses the basic theory and practice of assessing propagation losses. 2. Time dispersion: radio waves often face multiple reflections due to the presence of all kinds of obstacles (walls, furniture...), leading to multipath propagation. As a consequence, fading effects are experienced at the receiver. Focussing on broadband indoor communication, this fading can be characterised as frequency selective and time invariant. Section 2.3 introduces fading terminology, applies it for broadband indoor communication, and presents some typical channel responses generated by simple ray tracing. 3. Doppler effects: a wireless channel can never be guaranteed to be time- invariant, as persons and objects in the environment can always move. As a consequence, Doppler effects also need to be taken into account. Section 2.4 introduces basic definitions for time varying channels. The possible disastrous effects on OFDM-transmission are warned for. A new model for time-variant channel effects in indoor environments is proposed, where the reflectors in stead of the communicating terminals are moving. 2.2 PROPAGATION LOSSES Monitoring the budget is definitely a motto in wireless communication system design: not only does the severe competition force one to come with cost-effective solutions, also the link budget asks for careful accountancy. The link budget is the resulting signal to noise ratio at the receiver for a given transmit power. A significant part of the link budget is made up by the propagation loss, which is also generally called large-scale path loss. From the overall link budget, achievable Quality of Service (QoS) parameters (such as net data rate and error rate) can be predicted, as elaborated in [13]. The propagation loss can be exactly predicted in a free space situation, which we introduce as a reference case in paragraph 2.2.1. For severe Understanding the indoor environment 13 multipath propagation no exact analytical formula is available. Therefore an experimentally determined expression is proposed in paragraph 2.2.2. It is of practical use for statistical analyses, provided that appropriate parameters are available, Some guidelines for finding more accurate predictions in specific situations are given in paragraph 2.2.3. The other major term in the link budget is the noise and interference that is experienced by the receiver. They are discussed in paragraph 2.2.4. 2.2.1 Free space propagation When James Clerk Maxwell in 1864 described the Laws of Electrodynamics and formulated the famous equations named after him, experimental physics were not ready to confirm these phenomena with practical experiments. In 1887 Heinrich Hertz succeeded in validating Maxwells ideas by impressive experiments: the proof of propagation of electromagnetic waves in free space was a fact [11] [12], Consequently different radio engineering pioneers contributed to the effective realisation of information transfer by radio propagation pioneers. Little could they foresee the impact of their work on society more than a century later! From the basic radio equation, the signal loss in free space can be calculated as given by following formula: Where and are respectively the received and transmitted power, and the gains of the transmitting and the receiving antenna, the wavelength, and d the link distance. It is important to remark that, even in a completely free space, the power decays with the square of the distance. As a simple example, we can assume isotropic antennas and an operating frequency of 5 GHz. For distances varying between 10 and 50 meter, we find that the free space loss varies between 65 dB and 80 dB. Although the simplicity and exactness of the basic radio equation is very attractive for a radio engineer designing a system, one should be very aware of its limited validity. Indeed, the term free space needs to be taken literally. In practical systems it applies when the Line of Sight (LOS) (see section 2.3) contribution to the received signal is principally dominant. Such situations occur when e.g. very directive antennas are used in satellite communications. Clearly, as OFDM is primarily used because it allows coping elegantly with multipath propagation, we can not rely on free space conditions in the given context. 14 Chapter 2 2.2.2 Statistical estimation Also in environments where multipath propagation is an essential part to the communication, we need to estimate the path loss, being the attenuation that an electromagnetic wave experiences when propagating over the air. The path loss is usually expressed in dB Where S is the path loss of a path of d meter, and S typically also depends on the environment E. Clearly, expression (2.2) gives a pure statistical prediction, and should not be used for determining the precise signal strength in a specific location. Following the wish to be able to dispose over equations for the large- scale signal loss which are similarly simple as in free space, researchers have tried to find simple analytical expressions which match channel measurements. A typical example of an approximate expression for the overall signal loss in an indoor environment is given in following equation [6]: Where the parameters a and b depend on the operating frequency, the geometry of the environment, and on whether there is a LOS component or not (i.e. Obstructed line of sight or OBS situation). Physically, the parameter a can be interpreted as the exponent of decay of the received power with the distance, while and b are curve-fitting parameters. As a typical example, for an OBS situation at frequencies around 5GHz, parameters giving a good correspondence to measurements are: dB, and Obviously, the decay of the received power with the link distance is even more severe than in free space! For distances between 10 and 50 meter this gives S between 80 dB and 105 dB. As a consequence, if one wish to design OFDM-based wireless indoor communication with a range of some dozens of meters, a serious challenge is involved to devise receivers that can work at a low SNR. It is clear that, due to the increased losses by transmission through walls and reflections, the signal losses in an OBS-position depend much stronger on the specific geometry and the link distance than in a free space or LOS situation. From the frequency dependencies, the following conclusions can be drawn if wireless indoor communication systems below or above 5GHz are considered: Understanding the indoor environment 15 Since lower frequencies get less attenuated, they allow to cover a larger range with a single basestation. OFDM-systems complying with a standard for the 2.4 GHz range, can benefit from this effect. For significantly higher frequencies, obstacles behave as isolators. Consequently, the range at which a signal can be received decreases drastically. On the other hand, the reuse of frequencies is augmented. This feature can be exploited advantageously for setting-up high-capacity pico-cellular networks. Also for outdoor communication, similar expressions have been found to provide a good approximation of actually measured losses. The exponent of decay of the power can vary from 2 (for free space) to 5, depending on the environment (rural, urban,..). 2.2.3 Deterministic models For some applications, the information that is output by typical statistical models for propagation losses is not sufficient to perform system analysis and simulations. In such cases, a deterministic or also called geometrical model is often preferred. Thereto, the exact path loss is estimated from the transmissions and reflections encountered on the path between transmitter and receiver. In [7], a rule of thumb is given for the losses by transmission through walls, dividing walls into three categories with corresponding losses: light (1.5 dB), medium (5 dB), and heavy (10 dB). In [8], a more rigorous approach to calculate reflection and transmission coefficients of walls is given. It is quite evident that these depend on the thickness and the material of the wall, and the frequency of the electromagnetic waves. The derivation of exact expressions is however not straightforward, since the characteristics of materials are often frequency dependent in a difficult to reveal manner. Recently, freeware also became available [9], which allows calculating interesting parameters for deterministic path loss determination. 2.2.4 Noise and interference Successful communication depends on how good a receiver can retrieve the information out of the received signals. Even if there is very little attenuation and distortion on the channel, reception is far from perfect because of the noise that is inevitably superposed on the signals. The term noise is used for all undesired signals that add up to the actual information. For most communication systems, the signal-to-noise ratio is crucial for the quality of the link. 16 Chapter 2 We distinct noise of natural origin on the one hand, and generated by humans on the other hand. We call the latter category interference. 2.2.4.1 Natural noise An important source of natural noise is thermal noise. Thermal noise is caused by electrons, whose mobility is influenced by temperature. Therefore, it is physically impossible to build a completely noise-free electrical system, although one can try to minimise the influence of thermal noise during system design. Since the thermal movements of electrons are random in nature, so are the noise signals resulting from them. Thermal noise has a uniform frequency content, which we call white, and a temperature- dependent spectral density in which k is the Boltzmann constant en T the temperature in Kelvin degrees. Other sources of natural noise that are relevant for wireless links are among others atmospheric disturbances and radiation of the sun and other elements in our milky way and beyond. While we can notice these effects on outdoor radio channels, for example when listening to the radio during a thunderstorm, they have no perceptible influence on indoor communication. 2.2.4.2 Interference Human activity also generates a lot of noise in the ether. A major contribution comes from other transmitters emitting radiation in the same frequency band. Spectrum regularisation aims to prevent this kind of interference. Also, electrical equipment (as well residential as industrial) is seldom radiation-free. Recently, OFDM-systems are considered for increasing the data-rate of wireless systems in the unlicensed 2.4 GHz ISM-band. Strong narrowband interference, due for instance to the leakage of microwave ovens or interfering wireless communication systems operating in the same band, severely hampers the performance of these systems To simulate two major types of narrowband interference we developed a flexible complex baseband model for the interference i(t), given by [14]: With this model, two types of interference where modelled: Understanding the indoor environment 17 a) Microwave oven emission: The carrier frequency is set to the design frequency of the magnetron in the baseband ( Hz for a baseband center frequency of 2.45 GHz). A(t) is a rectangular waveform with a frequency of 50 Hz to model the switching of the magnetron due to its power supply. We constructed to match the results of broadband measurements [17] in the time and frequency domain. The spectrogram of a typical interfering signal generated by this model is given in Figure 2. 1a. b) Frequency hopping interference: The amplitude being constant we apply frequency shift keying, with a continuous phase, to modulate the carrier using a pseudo-random BPSK sequence with a symbol period of For we use a pseudo-random hopping sequence of frequencies spaced by 1 MHz within a bandwidth ranging from10 to 10 MHz. The hopping rate is 1600 Hz, as used in the Bluetooth system. Figure 2.1b shows the spectrogram of a typical interfering signal generated by this model. Analysis has shown that interference mitigation is crucial for systems operating in the 2.4 GHz band [14]. Fortunately, OFDM enables some nice solutions to cope with the specific interference in the frequency domain. 2.3 MULTIPATH PROPAGATION In most practical wireless communication systems, the radio propagation is far more complex than in a free space situation. In the presence of the earth, natural obstacles, buildings, furniture, etc., a radio signal travels via both the direct path and (possibly numerous) other paths from transmitter to receiver. As a consequence, the radio channel suffers from multipath conditions, leading to typical fading phenomena. This section focuses on situations 18 Chapter 2 where the time variance is very slow in comparison to the transmission rate, and can thus be neglected in the analysis. First the basic mechanisms of multipath channels are described in 2.3.1. Next, in paragraph 2.3.2, the terminology for describing a multipath channel is introduced. In paragraph 2.3.3 a simplified ray tracing approach to model multipath channels is presented. For two typical indoor examples, graphical results are included. 2.3.1 An intuitive understanding Most real radio channels are established in circumstances that differ substantially from free space. For example in an indoor environment, radio waves face multiple reflections due to the presence of all kinds of obstacles (walls, furniture...). Electromagnetic waves propagating in such an environment are partially transmitted through and reflected by these obstacles. As a consequence, waves emitted by the transmitter arrive at the receiver antenna over multiple paths, which is called multipath propagation. Figure 3.1 shows a simple multipath situation. The complete set of propagation paths between transmitter and receiver forms the multipath channel. Each path can be characterised by three parameters: delay, attenuation and phase shift. The path delay depends on the path length and on the speed at which a wave is propagating in the different media along the path. Attenuation and phase shift can be subdivided into two different components. One is due to the free space propagation and depends on the path length. The other one is caused by reflection and shadowing and depends on the reflection and transmission coefficients. These coefficients are functions of material properties, such as relative dielectric constant, permeability, conductivity, thickness, etc. Understanding the indoor environment 19 The effect of multipath propagation on a radio channel impulse response depends on the path-length difference(s) relatively to the wavelength. Consider the exemplary case study of 2 paths, each with their own delay, attenuation, and phase shift. The paths interfere destructively if they arrive with a phase difference of 180, which extinguishes the overall received signal. In case the path length difference is an entire multiple of the wavelength, the paths add constructively. Resultant, the received signal strength varies heavily with frequency, place, and in case of changes in the multipath composition also time. These variations are generally called fading characteristics. A distinction is made between Line of sight (LOS) and Obstructed line of sight (OBS) set-ups. In a LOS set-up, there is a line of sight between transmitter and receiver. The corresponding path is typically the least attenuated one and the one with the shortest delay. In an OBS set-up, there is no line of sight between transmitter and receiver, so that all incoming waves at the receiver antenna have been reflected or shadowed at least once. This is the more hostile case, since rarely one particular path dominates like the line of sight would do. From a communication point of view, a multipath channel can be characterised as a producer of echoes. A whole set of echoes with different amplitudes and phase shifts arrives at the receiver antenna, with each echo arriving at a different instant. This causes dispersion of the information signal in time. If no appropriate measures are taken, the different signal components arriving at the receiver antenna interfere with each other, which can lead to inter-symbol interference (ISI). 2.3.2 Multipath Terminology To describe multipath, a extensive set of mathematical terminology has been established. The most important concepts are introduced in the next sections. 2.3.2.1 Impulse response The complex impulse response h(t) of a time-invariant multipath channel with N paths can be modelled as 20 Chapter 2 Where k is the path index, is the path gain, is the path phase shift and is the path delay. The above model describes a static channel. For OFDM-based wireless indoor networks, the multipath parameters can indeed be assumed to be quasi-static, since they change very slowly with respect to the considered data rates. The frequency response H(f) of the channel is the Fourier transform of the impulse response. 2.3.2.2 Power Delay Profile The power delay profile (PDP) P(t) is defined as the squared absolute value of the channel impulse response. It gives the time distribution of the received signal power from a transmitted over the channel. The power delay profile represents the relative received power in function of excess delay with respect to the first received path. Power delay profiles are found by averaging instantaneous power delay profile measurements. The PDP of a channel is on the average exponentially decaying in an indoor RF channel, which can be explained by the path loss that increases with the length of the path. From the PDP, the instantaneous frequency response of the channel can be calculated. 2.3.2.3 RMS Delay Spread The RMS delay spread which can be calculated from the power delay profile, is most often used to quantify the time dispersion on the channel [4]. It is given by: Together with (2.7) this expression can be simplified to Understanding the indoor environment 21 The value of is commonly used to give a rough indication of the maximum data rate that can reliably be supported by the channel, when no special measures, such as equalization, are taken. The following rule of thumb is often applied for the length of the channel impulse response 2.3.2.4 Coherence Bandwidth The auto-correlation function of the channel frequency response is given by For the class of channels with an exponential delay profile, the auto- correlation can be computed as a statistical expectation. For a received signal with unity local-mean power, we find Jakes et al. [18] have shown that, for the channel amplitudes, The coherence bandwidth gives a measure for the statistical average bandwidth over which the channel characteristics are correlated. is defined as the value of for which the auto-correlation function of the channel frequency response has decreased with 3 dB, or: 22 Chapter 2 So, for the channel with exponential delay spread How the fading on the radio channel affects the communication over a channel is determined by the relation between the bandwidth of the information signal and the coherence bandwidth of the channel. If the channel response can be considered constant over the information bandwidth the fading is named flat. On the other hand, frequency selective fading conditions are encountered when the information bandwidth is significantly larger than the coherence bandwidth of the channel. Specifically because it enables elegant equalisation in frequency selective fading, OFDM is often preferred in these conditions. Therefore, it is a facilitator of broadband communications in harsh reflective environments. 2.3.3 Ray-Tracing Modelling Different modelling techniques have been developed in the quest of characterising multipath channels. The simplified approach that is presented here is based on the modelling of wave propagation by two-dimensional ray- tracing algorithms [3]. The ray-tracing model, some simulation examples, and finally an interpretation of the simulation results will be presented. 2.3.3.1 A simplified approach A rather simple ray-tracing approach is followed (leading to relatively fast simulation times), not comprising an accurate characterisation of the environment. This method has not been designed for precise channel predictions in specific geometrical situations (e.g. no diffraction effects are taken into account). Still, the results show to produce typical fading characteristics that correspond very well with measurements reported in literature [6], and with channel responses considered in the BRAN- standardisation [10]. Understanding the indoor environment 23 Figure 2.3 gives a very simple example of how the ray tracing works. Rays from a starting point to a target point are obtained by finding the direct path and the paths to all mirror images of the target point The floor plan of the set-up, including walls and objects, is modeled as a set of straight lines. Each line can be assigned different reflection and transmission coefficients. Further input parameters are the positions of the receiver and the transmitter antenna respectively. Both antennas are considered isotropic. This setup being given, all possible paths from the transmitter antenna to the receiver antenna are calculated by a software tool, and the corresponding path delays, attenuation and phase shifts are stored in a result file. The computation of a path is aborted, if its attenuation exceeds a certain threshold, e.g. 30 dB, with respect to the attenuation of the first received path. The results of the ray-tracing simulation are used to parameterise the mathematical channel model given by equation (2.6) in order to calculate the corresponding impulse response. Since the channel model is used for baseband simulations, the computation is performed for the desired centre frequency, i.e. phase shifts are calculated for the corresponding wavelength. Finally the impulse response is down-sampled to the desired sample rate using a raised cosine filter to avoid aliasing. 2.3.3.2 Simulation examples Figure 2.4 through Figure 2.8 show the results of ray-tracing examples for two different indoor set-ups. They were generated in the context of the design of an OFDM-transceiver for 5GHz WLANs. The first set-up (Figure 2.4) represents a typical home environment, where in this specific simulation the transmitter and the receiver are located in the same room, but the direct path between them is obstructed. The second set-up (Figure 2.6) corresponds to an office-like environment, with the basestation installed in the corridor and a terminal in one of the offices. The power delay profiles are given for a bandwidth of 80 MHz and 4 times over-sampling. Gain and phase of the baseband frequency responses are given for a bandwidth of 80 MHz, and a centre frequency at 5.25 GHz. The characteristics of the different channels, in terms of average delay delay spread and coherence bandwidth of the channel are summarised in Table 2.1. 24 Chapter 2 Understanding the indoor environment 25 26 Chapter 2 2.3.3.3 Interpretation of results A considerably large number of ray-tracing simulations for home-like and office-like environments have been carried out. We analysed scenarios with a single room containing various objects and with different geometries of multiple rooms. The following practical conclusions can be drawn from these results: The Power Delay Profiles on the average show an exponentially decaying behaviour with increasing delay, while we can also observe typical clusters of paths. Both these effects correspond to the channel models that were considered in ETSI-BRAN [10]. For typical indoor environments, the RMS delay spread is of the order of 10 to 50 nsec. (even for large settings). In ETSI-BRAN, some of the channels correspond to outdoor channels, which explains the larger delay spreads reported there. Two main categories of channels can be distinguished: LOS-situations mostly show a dominant component, the OBS delay profiles are less regular. For an environment with numerous reflections (walls and/or objects), the exact geometry of the building and location of the objects does not determine the kind of channel responses (dips) you can get in the environment. By moving transmitter or receiver over a very small distance (in the order of ) in one specific environment, you can get extremely different channel responses (going from almost flat to dips up to 40db!). The overall scale of the environment has an influence on the channel response, namely a scaling on the This effect can be calculated without needing new simulations. 2.4 TIME VARIANT CHANNELS The indoor propagation channel is often considered as a static environment, because of the relatively low mobility speed and the easy analysis according to this assumption. However, following the motto never trust a wireless channel, the time variance of the channel and its effect on OFDM transmission should be assessed. The focus is here on WLANs, but if Understanding the indoor environment 27 OFDM is considered for outdoor and even mobile systems, Doppler effects may ask for specific correction measures. 2.4.1 Terminology To describe the time variations a channel, a mathematical model is needed. The most important terminology is introduced in the next paragraphs. 2.4.1.1 Coherence Time The time correlation describes the correlation between two channel responses at different time instances and is defined as Generally, as the time between two channel responses increases, the correlation declines. The coherence time is defined as the time for which the correlation decreases by 3 dB. When the time variation of the channel is considered, a categorisation of fading channels can be made, relating the data rate to the coherence time of the channel. If the coherence time is very large compared to the transmitted signalling period, the channel essentially stays constant during the sampling period. This is referred to as slow fading. If the signalling period becomes comparable to the coherence time, the transmission system undergoes fast fading. In case of OFDM transmission or more generally systems making use of a block-based equaliser, a quasi-static approach is only robust if the channel remains stable during the entire OFDM symbol. The conventional classification then takes a more meaningful significance if the OFDM- symbol or block duration is used as a classification criterion, in stead of the signalling or sampling time. 2.4.1.2 Doppler spectrum By taking the Fourier transform of the time correlation function, the Doppler spectrum is obtained: 28 Chapter 2 The Doppler bandwidth is defined as the 3dB bandwidth of the Doppler spectrum. An interesting relation between the coherence time and the Doppler frequency is given by: 2.4.2 What is on the move in the indoor environment? In a typical mobile multipath channel, the Doppler effects are mostly modelled by the so-called Jakes spectrum [18]. This model is based on the assumptions that the receiver is moving at velocity v, and the angles of arrival of the multipath components are uniformly distributed. Defining the time correlation of the channel is then given by where stands for the zero-th order Bessel function of the first kind. The corresponding Jakes spectrum is given by: This model has also widely been used for WLANs mainly due to its simplicity, although typically neither transmitter nor receiver is moving in such scenarios. The approach results however in an extreme over-estimation of the time variance of such channels. Therefore, a new and more appropriate model for the time variance of indoor channels is proposed in [15]. It is based on a stationary transmitter and receiver, but reflectors that are moving at random speeds. These assumptions reflect the situation where objects or persons in the surroundings are moving, rather then the communicating terminals themselves. The resulting time correlation is then given by: Understanding the indoor environment 29 where a is the fraction of paths that are time-invariant. Compared to equation (2.17), the Bessel function is now squared. This is due to the fact that a movement of a scatterer has double the impact of a movement of the receiver. Unfortunately, there is no closed form expression for the above integral. Figure 2.8 shows the time correlation for different instantiations of the new model on the one hand, and the Jakes model on the other hand, for a carrier frequency The thick solid line gives the Jakes model. The worst case for moving reflectors, plotted in dashed line, represents the time correlation when all scatterers are assumed to move at the maximum speed We also show the situation in which 90% of the paths is static, which is a realistic value in indoor environments. This case is given in solid line. Clearly, the correlation degrades only slowly with time as not all reflections are dynamic. The corresponding Doppler spectra for all models are shown in Figure 2.9. The classical Jakes spectrum has a bandwidth of and a very high probability of high Doppler frequencies. In all other cases where 30 Chapter 2 movement of scatterers is considered, the bandwidth of the fading process is equal to The latter models give rise to a more peaky spectrum, matching indoor propagation measurements [19] [20] [21]. 2.5 CONCLUSIONS An exact knowledge of the problem, is often the most crucial step towards finding a solution. Therefore, when making plans for getting wireless OFDM systems to work, it is sound practice to start with an assessment of the wireless transmission channel. This chapter introduced the basic propagation effects to take into account. The focus is mainly on the indoor environment. The propagation is shown to be subject to rich multipath reflection, and slow time variation. Understanding the indoor environment 31 REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] Prasad, private communication L. Van der Perre, Modelling and simulation of the HF channel with the objective of testing intelligent narrowband and wideband modems. Ph. D. Thesis K.U.Leuven 1997 P. Vandenameele, L. Van der Perre, M. Engels, SDMA for OFDM-based WLANs, Kluwer Academic Publishers, 2001 T.S. Rappaport, Wireless Communications, Principles and Practice. Prentice Hall, 1996 H. Hashemi, The Indoor Radio Propagation Channel, Proceedings of the IEEE, Vol. 81, No 7, July 1993 G. J. M. Janssen et al., Wideband Indoor Channel Measurements and BER Analysis of Frequency Selective Multipath Channels at 2.4, 4.75 and 11.5 GHz, IEEE Transactions on Communications, vol. 44, no. 10, p. 1273. J. Doble, Introduction to Radio Propagation for Fixed and Mobile Communications, Artech House Publishers, 1996 C.D. Taylor, S.J. Gutierrez, S.L. Langdon, K.L. Murphy, and W.A. Walton, Measurement of RF Propagation into Concrete Structures over the Frequency Range 100 MHz to 3 GHz, in Wireless Personal Communications, Advances in Coverage and Capacity, Kluwer Academic Publishers http://komatsu.linux.student.kuleuven.ac.be/3dom/, K.U.Leuven software, June 1999 http://www.etsi.org/Bran P. Simon, Proof, Rhythm of the Saints, 1990 R. Cuyvers, B. Van der Herten, and al., Van tamtam naar virtuele realiteit, Davidsfonds, 1995 J.C. Proakis, Digital communications, Mc Graw-Hill, 1989 R. Ness, Steven Thoen, Liesbet Van der Perre, Bert Gyselinckx and Marc Engels, Narrowband interference mitigation in OFDM-based WLANs, workshop on multicarrier, S. Thoen, High-speed OFDM based wireless local area networks, PhD thesis KULeuven, May 2002 J.A.C. Bingham, Multicarrier Modulation for Data Transmission: An Idea Whose Time has Come, IEEE Comm. Magazine, vol. 28, pp. 5-14, May 1990. A.M. Saleh et al., A Statistical Model for Indoor Multipath Propagation, IEEE Selected Areas of Comm., vol. CSA-5, no. 2, pp. 1384-87, July 1991. W.C. Jakes, Microwave Mobile Communications, IEEE press, 1993 S.J. Howard and K. Pahlavan, Doppler spread measurements of the indoor radio channels, Electronic Letters, 26(2):107-109, January 1990 H. Hashemi, M. McGuire, and D. Tholl, Measurements and modelling of temporal variations of the indoor propagation channel, IEEE journal on Vehicular Technology, 43(3):733-737, August 1994 H. Hashemi A study on temporal variations of the indoor propagation channel, in PIMRC, pp 127-134, 1994 [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] Chapter 3 The OFDM Principle Divide et Impera Reto Ness, Jean-Paul Linnartz, Liesbet Van der Perre, Marc Engels IMEC 3.1 THE OFDM PRINCIPLE 3.1.1 Multicarrier modulation In single carrier modulation, data is sent serially over the channel by modulating one single carrier at a baud rate of R symbols per second. The data symbol period is then 1/R. In a multipath fading channel, the time dispersion can be significant compared to the symbol period, which results in inter symbol interference (ISI). A complex equaliser is then needed to compensate for the channel distortion. 34 Chapter 3 The basic idea of multicarrier modulation was introduced and patented in the mid 60's by Chang [1]: the available bandwidth W is divided into a number of subbands, commonly called subcarriers, each of width The subdivision of the bandwidth is illustrated in Figure 3.1, where arrows represent the different subcarriers. Instead of transmitting the data symbols in a serial way, at a baud rate R, a multicarrier transmitter partitions the data stream into blocks of data symbols that are transmitted in parallel by modulating the carriers. The symbol duration for a multicarrier scheme is In its most general form (see Figure 3.2), the multicarrier signal can be written as a set of modulated carriers: where is the data symbol modulating the subcarrier in the signalling interval. is the waveform for the subcarrier. The symbol duration can be made long compared to the maximum excess delay of the channel (see previous chapter), or by choosing sufficiently high. At the same time the bandwidth of the subbands can be made small compared to the coherence bandwidth of the channel The subbands then experience flat fading, which reduces equalisation to a single complex multiplication per carrier. Increasing thus reduces the ISI and simplifies the equaliser into a single multiplication (remark that the number of multiplications is The OFDM Principle 35 proportional with but the rate at which they have to be calculated is reverse proportional with However, the performance in time variant channels is degraded by long symbols. If the coherence time of the channel is small compared to the channel frequency response changes significantly during the transmission of one symbol and a reliable detection of the transmitted information becomes impossible. As a consequence, the coherence time of the channel defines an upper bound for the number of subcarriers. Together with the condition for flat fading within the subbands a reasonable range for can be derived as 3.1.2 Orthogonal Frequency Division Multiplexing To assure a high spectral efficiency, the subchannel waveforms must have overlapping transmit spectra. They need to be orthogonal for enabling simple separation of these overlapping subchannels at the receiver. Multicarrier modulations that fulfil these conditions are called orthogonal frequency division multiplex (OFDM) systems. A general set of orthogonal waveforms, is given by: with Here is the subcarrier frequency and is the lowest frequency used The spacing between the adjacent subcarriers equals Since the waveform is restricted to the time window the intercarrier spacing must also satisfy The windowing results in a convolution with otherwise otherwise 36 Chapter 3 in the frequency domain. As a consequence, the different subbands overlap (as shown in Figure 3.3). Although the subchannels overlap, they do not interfere with each other at Indeed, they are orthogonal, or: The demodulation is based on this orthogonality of the subcariers and consists of a bank of matched filters that implement the relation A schematic view of such a demodulator is shown Figure 3.4. The implementation of an OFDM system that consists of oscillators in the transmitter and a bank of matched filters in the receiver is becoming very complex for a large number of subcarriers. However, as Weinstein and Ebert pointed out [2], an IDFT and DFT operation can replace the baseband modulator and the bank of matched filters respectively (if is a power of two). In addition to being much cheaper, such implementation does not suffer from the inaccuracies associated with an analogue oscillator bank. The OFDM Principle 37 3.1.3 Cyclic Prefix Passing the signal through a time-dispersive channel causes inter symbol interference (ISI). In an OFDM system, it also makes that the orthogonality of the subcarriers is lost, resulting in inter carrier interference (ICI). To overcome these problems, Peled and Ruiz [3] introduced the cyclic prefix (CP). A cyclic prefix is a copy of the last part of the OFDM symbol that is prepended to the transmitted symbol (see Figure 3.5) and removed at the receiver before the demodulation. The cyclic prefix should be at least as long as the significant part of the impulse response experienced by the transmitted signal. This way the benefit of the cyclic prefix is twofold. First, it avoids ISI because it acts as a guard space between successive symbols. Second, it also converts the linear convolution with the channel impulse response into a cyclic convolution. As a cyclic convolution in the time domain translates into a scalar multiplication in the frequency domain, the subcarriers remain orthogonal and there is no ICI. The length of the cyclic prefix should be made longer than the experienced impulse response to avoid ISI and ICI. However, the transmitted 38 Chapter 3 energy increases with the length of the cyclic prefix. The SNR loss due to the insertion of the CP is given by where denotes the length of the cyclic prefix and is the length of the transmitted symbol. Also the the number of symbols per second that are transmitted per Hz of bandwidth, decreases to In a digital implementation, is a multiple of the basic sample period Because of the loss of SNR and efficiency, the cyclic prefix should not be made longer than strictly necessary. When making equal to the length of the impulse response, the relative length of the cyclic prefix is typically small, so that the ISI and ICI-free transmission motivates the small SNR loss. However, when selecting the length of the cyclic prefix, the following issues should also be taken into account: Filter responses may add to the overall impulse response that should be compensated for by the guard interval. A part of the guard interval needs to be reserved for synchronisation margins. Not only is the time acquisition never guaranteed to be perfect, the effect of a clock offset between transmitter and receiver may still significantly increase the deviation. 3.2 THE OFDM SYSTEM MODEL The OFDM Principle 39 In the previous section, we introduced an OFDM system with as basic ingredients the DFT and IDFT operations and the cyclic prefix. The discrete-time baseband equivalent model of such an OFDM system is given Figure 3.6. In the transmitter, the incoming data stream is grouped in blocks of data symbols. These groups are called OFDM symbols and can be represented by a vector Next, an IDFT is performed on each data symbol block, and a cyclic prefix of length is added. The resulting complex baseband discrete time signal of the OFDM-symbol can be written as where n is the discrete time index. The complete time signal s(n) is given by the concatenation of all OFDM symbols that are transmitted In general, the received signal is the sum of a linear convolution with the discrete channel impulse response h(n) and additive white Gaussian noise n(n). For this, we implicitly assume that the channel fading is slow enough to consider it constant during one OFDM symbol. In addition, we assume that the transmitter and receiver are perfectly synchronised. Based on the fact that the cyclic prefix is sufficiently long to accommodate the channel impulse response, or for and we can then write: In the receiver the incoming sequence r(n) is split into blocks and the cyclic prefix associated with each block is removed. This results in a vector with The received data symbol is obtained by performing a DFT on this vector. Thus is given by 40 Chapter 3 By substituting r(n) with equation (3.10), (3.11) can be written as Substituting with equation (3.8), then yields the following result: where is the sample of the DFT of . Since n(n) is white Gaussian noise, is also white Gaussian noise. Because for all we can let run from 0 to instead of Additional swapping of the two inner sums and reordering yields The first part of this expression consists of an IDFT operation nested in a DFT operation. The inner sum is the sample of the DFT of or The equation hence translates into The OFDM Principle 41 This equation demonstrates that the received data symbol on each subcarrier k equals the data symbol that was transmitted on that subcarrier, multiplied by the corresponding frequency-domain channel coefficient in addition to the transformed noise contribution From the received data symbols the transmitted data symbols can be estimated using a single tap equaliser followed by a slicer. In the equaliser, the receiver divides each received data symbol by its corresponding channel coefficient. The result of this step is a soft estimate The slicer rounds this soft estimate towards the nearest symbol in the modulation alphabet, called the hard estimate For a more compact notation, a matrix equivalent is often used. For a single OFDM symbol, it equals where denotes the Hadamard (i.e. element-wise) product, DIAG(H) is the diagonal matrix with the elements of H, and When considering M OFDM symbols, we can define the following matrix notation: where and 3.3 WHAT IF THE CHANNEL IS TIME-VARIANT? The OFDM system model of previous section was derived for a time- invariant channel. However, as explained in chapter 2, even with fixed transmitters and receivers, a channel will be (slowly) time-variant. These time variations erode the orthogonality of the subcarriers and cause ICI or "FFT leakage." To model such behaviour a more complex system model will be needed. In this section we derive such a model. Although the mathematics are more sophisticated, it will appear soon that a relatively simple matrix channel model remains. 42 Chapter 3 Classic multipath channel models involve a collection of reflected waves. Each wave has its particular path delay complex-valued path gain and phase shift When the antenna is in motion, the amplitudes and path delays of the individual reflections will remain fairly constant, but the phase shifts experience a Doppler effect. In fact, for a signal at the k-th subcarrier, with the maximum Doppler shift and the angle of arrival. Without any loss of generality, we take and consider any non-zero initial phase to be taken into account by the phase of If, for ease of notation we focus on the initial signalling interval the received joint OFDM signal equals Note that the time delays are not necessarily an integer multiple of the sampling period, so here we prefer a continuous-time representation. Next, we force the above expression into the following form, with a complex received amplitude at the k-th subcarrier that is time varying, This is possible if we define the time-varying channel amplitude as [9] As described in the previous section, the OFDM receiver takes samples at sample intervals of and performs a DFT. The k-th output of the DFT is then found as We observe that The OFDM Principle 43 Now let's fill in r(t) in this equation. This results in From here we will work towards a matrix channel model. To this end, we take a Taylor expansion of the time varying amplitude, namely Here denotes the q-th derivative of the amplitude with respect to time at the k-th subcarrier and at instant We extend the notation of the vector H as follows: denotes the q-th derivatives of H. Remark that as defined in the previous section. Further, we introduce the ICI leakage constants defined as In fact, these constants describe the signal transfer over the q-th derivative of the amplitude at subcarrier k to the receive subcarrier. This allows us to rewrite as follows: A practical receiver typically removes the effect of a phase shift due to a time delay, so we simplify the above expression by taking So, the received signal can be written compactly in discrete frequency domain as, where the q-th order ICI crosstalk matrix equals 44 Chapter 3 Figure 3.7 depicts the channel and receiver in the discrete frequency domain. A frame of user symbols is offered to the system. The output vector after the DFT in the receiver is denoted by In a conventional system, W represents the equaliser, or automatic gain control per subcarrier. To adequately receive signals over a mobile channel, W also needs to compensate for ICI. The figure shows that one can simulate the radio link as a vector channel without explicitly executing the DFT operation. and are (correlated) Gaussian random vectors and is a fixed matrix, which can be implemented using a butterfly structure. If the Doppler spread is much smaller than the frequency resolution of the DFT grid we may restrict our analysis to zero and first order effects In particular, for For integer this reduces to the Kronecker delta function This is just a confirmation that subcarriers (with non-fading amplitude) are orthogonal, resulting in For we see that, if we assume integer the first-order derivatives leak according to The OFDM Principle 45 3.4 OFDM RECEIVER PERFORMANCE It has been shown in the previous sections that in an OFDM system, the received data symbol transmitted in the signalling interval on the subcarrier is given by the corresponding transmitted symbol, multiplied with the channel frequency response sampled at the subcarrier frequency plus noise. If we have an ideal linear time-invariant (LTI) frequency non-dispersive AWGN channel, this translates to a parallel set of AWGN channels, with equal SNR. As a consequence, the performance will be identical with single carrier modulation over AWGN, except for the SNR loss due to the cyclic prefix. As an example, Figure 3.8 shows the BER versus plot for an OFDM-QPSK system with carriers and various cyclic prefix lengths. 46 Chapter 3 For QPSK, the probability of a symbol error or SER in function of the effective SNR per bit is given by In the previous chapter it was shown that for indoor multipath communication the amplitude spectrum of the channel taps is approximately Rayleigh distributed. Therefore, we can assume that the effective SNR per carrier are exponentially distributed according to The probability of bit error can then be expressed as Substituting equations (3.31) and (3.32) in this formula results in the following expression The resulting SER versus graph is shown in Figure 3.9. It can be observed that the SER is only slowly decaying with increasing Intuitively, this can be understood by considering the significant differences in the signal-to-noise ratios between different carriers. The dips in the frequency response of typical indoor multipath channels, for instance, can reach up to 30 dB. These low SNR values cause extremely high bit error probabilities on the bad carriers which dominate the average bit error rate. Mobility also has an effect on the SER. Let's for the sake of simplicity assume that ICI affects the receiver performance in the same manner as the noise. We use the model of the previous section to estimate the average signal to interference-plus-noise ratio (SINR) for OFDM with Doppler. The intercarrier interference power on the carrier can be expressed as: The OFDM Principle 47 In a Rayleigh channel, the derivatives are zero-mean complex jointly Gaussian for any k and q. The covariance of and can be expressed for certain specific cases of the channel model. For instance for a uniform angle of arrival and an exponential delay profile, we can extend the correlation function of Chapter 2 into [9], Roughly speaking with So the ICI reduces slowly with increasing subcarrier separation. Relatively many subcarriers make a significant contribution to the ICI 48 Chapter 3 The expected signal-to-noise ratio, to be inserted in equation (3.34), becomes 3.5 CODING: AN ESSENTIAL INGREDIENT We indicated in the previous section that OFDM has a limited performance for a frequency selective channel, due to the dominance of carriers with low SNR. Remark that this is different from conventional single-carrier systems, where the SER or BER is determined by the average SNR in the transmission bandwidth. If this bandwidth is large compared to the coherence bandwidth and if perfect channel equalisation is used, the performance of a single-carrier system in a frequency selective environment significantly outperforms plain OFDM. To remedy this problem, any practical system uses either channel coding or adaptive loading. In adaptive loading, the modulation order and/or power of every carrier is adapted according to its SNR. As a consequence, the bad carriers, which dominate the performance of plain OFDM, are no longer used and a dramatic improvement of BER performance is experienced. In addition this benefit comes without reduction in the transmitted datarate. The main disadvantage of adaptive loading is the need for feedback of channel estimations, which limit it applicability in fast fading environments. Because detailed discussion of adaptive loading are beyond the scope of this book, we refer the user to the relevant literature [4] [5] [6]. The alternative way of improving performance is the use of an error- coding scheme over the carriers [7] (see Figure 3.10). Various error-coding methods can be applied on the incoming bit stream: block codes, like Reed- Solomon codes and convolutional codes are the most common ones. Also a concatenation of a block coder, an interleaver and a convolutional code is often used. It has the advantage of mitigating the output burst errors that are typical for convolutional Viterbi decoders. More recently, also trellis coded modulation, which operates on symbols in stead of bits, and Turbo codes have been proposed. The OFDM Principle 49 The coded bit- or symbol stream is next applied to an interleaver structure. In its most general form, the interleaver has both a time and frequency component. The frequency interleaver puts consecutive bits on uncorrelated carriers, such that they experience independent channel attenuations. The time interleaver separates consecutive bits over different OFDM symbols. As a consequence, for a fast fading channel these bits experience different channel attenuations. However, for slow fading channels the time interleaver has no effect and can be eliminated. The effect of block coding on an OFDM system can be illustrated by means of a simple analysis. Assume that we have an (n,k,t) code, where n is the number of total symbols per block and k is the number of source symbols. As a consequence, the code rate is k/n. The number of symbol errors that can be corrected with the code is t. In the remainder we take the block size equal to the DFT size or Based on equation (3.34), the probability of symbol error on a single carrier in a Rayleigh fading channel can be calculated. Assuming that all carriers experience independent fading, the probability of having symbol errors in an OFDM symbol can be easily calculated as: The resulting coded probability of symbol error is given by: 50 Chapter 3 The result of this expression for various values of t is plotted in Figure 3.11. Remark that the coding overhead had to be taken into account in the calculation of The SER curves show a much steeper descent, i.e., closer to the desired waterfall-like shape. With increasing t, the curves shift to the left of the figure (better performance). For large average the curve converges to straight line on a log-log plot. The slopes of the curve, or diversity order, equals -t, which is a dramatic increase with respect to the uncoded OFDM case This can be explained by the fact that the frequency diversity has now been exploited. 3.6 SUMMARY In this chapter we introduced the basic OFDM principle: a data stream is split into parallel lower rate data streams that are modulated on separate subcarriers. Practical implementations of OFDM systems use the IDFT operation for this modulation. Although this parallellization already reduces the impact of ISI, it can be completely removed by means of a cyclic prefix. For long, implementing an OFDM scheme was prohibitive complex. The OFDM Principle 51 Practical interest only appeared in the 90s due to the progress in digital signal processing and microelectronics [8]. In an AWGN channel, OFDM performs identical to a single carrier link. In a frequency selective channel, the performance is limited due to the impact of "bad" carriers. As a consequence, adaptive loading or error coding are essential ingredients for an OFDM system in such an environment. In a mobile environment, channel fading causes inter carrier interference, which leads to an increased amount of symbol errors. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] R.W. Chang, Orthogonal Frequency Division Multiplexing, U.S. Patent 3,488,445, filed 1966, issued Jan. 1970. S.B. Weinstein and P.M. Ebert, "Data transmission by frequency-division multiplexing using the discrete Fourier transform", IEEE Trans. on Communinations, COM-19(5):628-634, October 1971. A. Peled and A. Ruiz, "Frequency domain data transmission using reduced computational complexity algorithms", Proc. of the IEEE Int. Conf. Acoustics, Speech and Signal Processing, Denver, CO, 1980, pp. 964-967. Liesbet Van der Perre, Steven Thoen, Patrick Vandenameele, Bert Gyselinckx, Marc Engels., Adaptive loading strategy for a high speed OFDM-based WLAN, Globecom 98 Syndney, Australia, pp. 1936-1940, Nov. 1998. P.S. Chow et al., A Practical Discrete Multitone Transceiver Loading Algorithm for Data Transmission over Spectrally Shaped Channels, IEEE Transactions on Communications, vol. 43, pp. 773-775, Apr. 1995. R. Fischer and J. Huber, A New Loading Algorithm for Discrete Multitone Transmission, IEEE Proc. Globecom 96, London, England, pp. 724-728, Nov. 1996. W.Y. Zou and Y. Wu, "COFDM: An Overview", IEEE Trans. on Broadcasting, 41(l):l-8, March 1995. J.A.C. Bingham, Multicarrier Modulation for Data Transmission: An Idea Whose Time has Come, IEEE Communications Magazine, vol. 28, pp. 5- 14, May 1990. A. Gorokhov and J.P.M.G. Linnartz, Robust OFDM receivers for dispersive time varying channels: equalisation and channel estimation, ICC 2002, April 28- May 2, 2002, New York, Session OFDM-2 Chapter 4 When people agree on OFDM Wireless OFDM standards Bert Gyselinckx IMEC 4.1 WLAN STANDARDS Standards are essential in the market success that WLAN is currently experiencing. Especially the 802.11 series of standards from the IEEE (Institute of Electrical and Electronics Engineers) played a key role. The first generation of the IEEE 802.11 standard was finalized by the end of 1996. It provides a wireless Ethernet capability at a rate of 1 or 2 Mbps in the 2.4 GHz ISM (Industrial Scientific Medical) band. The standard supports various physical layers, e.g. direct sequence spread spectrum (DSSS) and frequency hopped spread spectrum (FHSS), under a common medium access control (MAC) protocol. Due to their relative low data rate, the market for these first generation WLAN products was limited to niche applications where mobility is an essential feature, like warehouses and industrial automation. In 1999, the second generation IEEE 802.11b standard became available. It increase the datarate for WLAN in the 2.4 GHz ISM band to 5.5 and 11 Mbps. This is achieved by supporting a new physical layer, i.e. complementary code keying (CCK) DSSS, under the standard IEEE 802.11 MAC layer. Because of its higher datarates, the IEEE 802.11b experienced a major succes for office WLAN applications. However, in order to provide high quality streaming video or fast Internet access even higher datarates are needed. This is especially crucial for opening up the residential WLAN market. Motivated by these commercial perspectives and by opening of new licence-exempt spectrum in the 5 GHz band, third generation WLANs are underway in the U.S., Japan and Europe. 54 Chapter 4 The IEEE is working on the IEEE 802.1la standard [15] for the U.S. The Multimedia Mobile Access Communication Systems (MMAC) group [14] of the Japanese Association of Radio Industries and Businesses (ARIB) establishes the HiSWANa standard. The European Telecommunication Standards Institute (ETSI) has a workgroup on Broadband Radio Access Networks (BRAN) that develops the HIPERLAN/2 standard [3]. All three standards have very similar physical layers based on OFDM with provisions for link adaptation. By varying the modulation and the coding schemes, the datarate can be varied from 6Mbit/s to 54Mbit/s. Their MACs, however, are very different. While IEEE 802.1la builds further on the distributed carrier sense multiple access with collision avoidance (CSMA/CA) Ethernet protocol of the previous standards, HIPERLAN/2 is based on a centrally controlled time sharing network. HiSWANa is the newest of the three standards and combines aspects of both MACs. In this chapter we will have a closer look into the OFDM based standards for WLAN. Our main focus will be on the European HIPERLAN/2 standard. 4.2 HIPERLAN/2 The HIPERLAN/2 standard was developed in the ETSI Project BRAN (Broadband Radio Access Networks) Error! Reference source not found.. The intention of the HIPERLAN/2 standard is to provide up to 54 Mbps multimedia communications between different broadband core networks and portable terminals. As a consequence, the standard comprises mechanisms to support services that are bounded by specific time delays to achieve an acceptable Quality of Service (QoS). To guarantee these QoS requirements, HIPERLAN/2 is based on a cellular networking topology. An access point that covers a certain geographical area controls a radio cell. A terminal communicates with other terminals or with the core network via this access point. This communication flow matches nicely with a typical business application scenario, where a portable computer gets services over a fixed corporate network infrastructure. In a home application scenario, however, the wireless network is supposed to interconnect consumer devices directly. Therefore, the HIPERLAN/2 standard also has a direct mode, where mobile terminals, controlled by the same access point, can directly exchange data. The HIPERLAN/2 standard defines the physical layer (PHY), the data link control layer (DLC), and several convergence layers (CL) for both the access point (AP) as well as the mobile terminal (MT). A schematic drawing of the HIPERLAN/2 protocol stack in the AP is shown in Figure 4.1. In the figure, the DLC layer is further detailed with the radio link control (RLC), medium access control (MAC) and error control (EC) functions. The higher When people agree on OFDM 55 layer protocols, which are located on top of the CL, are beyond the scope of the HIPERLAN/2 standard. The HIPERLAN/2 basic protocol stack at the MT side and its functions are similar to these of the AP. At the terminal side only a single RLC and MAC function is needed. The functionality of this MAC and RLC entity differs slightly from that of the AP, whereas the EC functions are symmetrical. The functionality of the PHY, DLC and CL layers is explained in more detail in the following sections. 4.2.1 Convergence layer (CL) The HIPERLAN/2 system can be used with a variety of different high level protocols. This is achieved by a convergence layer (CL) that interfaces between the high level protocol and the HIPERLAN/2 specific DLC layer. The standard defines CLs for ATM, IEEE1394 or firewire, UMTS and Ethernet. The CL has two main functions. The first is to translate the service requests from the higher layers to services offered by the DLC. The second is to reformat the different data formats coming from the high layers to data units with fixed size that can be treated by the DLC and vice versa. The process that does this is known as segmentation and reassembly (SAR) and is shown in Figure 4.2. The variable length data blocks coming from the higher layers, known as service data units (SDUs), are segmented into chunks of 48 bytes. These 48 bytes are then combined with additional header 56 Chapter 4 information (tag and flag). These new packets, the segmentation and reassembly protocol data units (SAR PDU) are then sent to the DLC for further processing. 4.2.2 Data link control layer (DLC) The DLC layer constitutes the logical link between HIPERLAN/2 devices. The DLC includes functions for medium access and transmission as well as terminal/user and connection handling, as shown in Figure 4.1. For this purpose the DLC consists of the following sublayers: 1. Error Control (EC) sublayer. 2. Medium Access Control (MAC) sublayer. 3. Radio Link Control (RLC) sublayer In the next sections we will describe the functionality of these three sublayers. When people agree on OFDM 57 4.2.2.1 Error Control (EC) sublayer This layer is responsible for detection and recovery from transmission errors on the radio link. HIPERLAN/2 foresees three different types of error control modes. 1. Acknowledged mode: in this mode a packet that is not received successfully and acknowledged is retransmitted. A cyclic redundancy check (CRC) of length 16 or 24 allows detecting whether the packet is received correctly. 2. Repetition mode: in this mode the packets are repeated in order to increase the reliability. 3. Unacknowledged mode: in this mode the packets are directly delivered to the higher layers. This mode is the least reliable of the three, but has the lowest latency. 4.2.2.2 MediumAccess Control (MAC) sublayer. The MAC layer controls the access of information to the physical layer and therefore the radio link. In HIPERLAN/2 the access point (AP) is in full control of the structure of the MAC frame. It will determine at which point in time the different mobile terminals will receive and can send their information. The MAC protocol is based on a dynamic time division multiple access / time division duplex (dynamic TDMA/TDD) scheme. Dynamic means that the AP can distribute radio resources to uplink and downlink of different users within one frame depending on the bandwidth needs of the different users. To see how this can be done we will examine the basic MAC frame structure shown in Figure 4.3. Each MAC frame has a duration of 2ms. It comprises transport channels for broadcast control (BCH), frame control (FCH), access control (ACH), downlink data (DL), uplink data (UL) and random access (RCH). It also 58 Chapter 4 foresees an optional direct link (DiL) channel that allows direct communication between mobile terminals. In all cases the allocation of the resources is performed by the AP. Let us examine these transport channels in somewhat more detail. The broadcast channel (BCH) is sent from the access point to all mobile terminals at the beginning of each MAC frame. It contains channel information concerning the entire radio cell such as: the AP transmission power, the expected AP reception power, pointers to and information about the different other transport channels in the frame. The BCH has a fixed length of 15 bytes. The frame channel (FCH) is sent from the access point to all mobile terminals. It conveys information that describes the structure of the air interface of the MAC frame. This is done by resource grants (RG) that define how many channels will be foreseen for uplink, downlink and direct link and which PHY modes that will be used in each of them. The access feedback channel (ACH) is sent from the access point to all mobile terminals. It is used to inform the terminals that have sent a request during the random access (RCH) period of the previous MAC frame about the results of their access attempts. The downlink and the uplink phases convey information to and from different terminals. This information is sent in long channels (LCH) and short channels (SCH). The long channels consist out of 54 bytes and contain the so-called user protocol data units (U-PDUs) which are used to convey user information. The short channels are 9 bytes long and are used to convey control information. Because the BCH, FCH, ACH and RCH are essential for the MAC operation they are always transmitted in the most robust physical mode, i.e. BPSK with code rate . The SCH and LCH are transmitted in a mode determined by the AP that depends on the quality of the channel. 4.2.2.3 Radio Link Control (RLC) sublayer The RLC sublayer performs the association control function (ACF), the radio resources control (RRC) and the DLC user connection control (DUCC). 4.2.2.3.1 The association control function (ACF) The association control functions primary goal is to associate the MT with the AP. For this purpose, the AP is periodically transmitting a beacon. The MT wanting to start a communication waits till it receives this beacon. If it does not receive the beacon after a certain time, it will send out a message itself, asking the AP to send the beacon. Upon receipt of the beacon, the MT will check whether it is allowed to associate with the When people agree on OFDM 59 network and whether the AP it received the beacon from has the right convergence layer. If all this is successful, the MT will request a MAC ID from the AP. Next, the MT and the AP negotiate the link capability to check things such as the version of the protocol, the highest modulation, the frequency bands, and the authentication and encryption algorithms that are supported by both. If encryption is negotiated, the MT and the AP will exchange their public Diffie-Hellman keys in order to calculate the encryption keys. This key is calculated with the DES or the Triple DES algorithm [6], [7] and is refreshed regularly in order to guarantee secrecy. The encryption ensures that the authentication can take place in encrypted mode. If the authentication succeeds, the MT is allowed to access the network and the association procedure will continue. Otherwise, the MT shall be rejected and the DLC connection between MT and AP will be terminated. The MT may also terminate the access attempt if the AP authentication fails. After successful association, the MT can request for a dedicated control channel (DCCH) that it uses to set-up DLC user connections. The MT can request multiple DLC user connections, each connection having its own unique support for QoS. Disassociation terminates the association between a MT and an AP. There are two types of disassociation: explicit and implicit. The explicit disassociation is requested by the MT or the AP. The implicit disassociation is used if the MT and the AP have lost the ability to communicate with each other for a certain period. In both cases, the AP will release the resources allocated to the MT. 4.2.2.3.2 The radio resources control (RRC) The RRC is responsible for the surveillance and the efficient use of the available radio resources. The first service to support RRC is handover. A network handover is carried out when an associated MT moves from one AP to another or when the link budget of another AP becomes more favorable. In this case, the MT notifies both the current and the target AP that it will perform handover. The target AP can then contact the old AP directly through the fixed network in order to get the information which is required to continue communication such as the MAC ID and the encryption key. Due to changes in the communication link, it can occur that the MT looses the connection to its current AP before it can connect to the new AP. In this case, the MT will have to set up a communication with the new AP starting from association on. A second service to support RRC is dynamic frequency selection (DFS). DFS assures that HIPERLAN/2 systems make equal use of the available 60 Chapter 4 frequencies (see section 4.2.3.1.7) under the consideration of avoiding interference of other devices in the same spectrum. This interference may originate from other HIPERLAN/2 systems in the neighborhood or from other systems working in the same frequency band. In order to perform DFS, the AP will collect received signal strength (RSS) measurements of the entire HIPERLAN/2 frequency band. For this purpose, the AP can make measurements itself or it can instruct associated MTs to make measurements and to send them back. Based on these measurements, the AP will determine to stay at the current frequency or to move to a new frequency. In the latter case the AP will broadcast a message to all associated MTs to change their frequencies. A third service to support RRC is transmission power control (TPC). In the uplink this means that every MT will transmit at the lowest possible power which still allows good reception by the AP. For the downlink, the power control is implementation specific. There are just a few rules that avoid interoperability problems and some spectrum regulatory requirements. A fourth service to support RRC is the so-called MT Alive function. This function is used to check whether an associated MT and an AP can still communicate with each other. For this purpose, the associated MT periodically sends out a MT Alive signal. If the AP does not get this message in the specified period it will explicitly solicit the MT to send this signal. If the MT fails to do so, the AP it will disassociate the MT and reuse its MAC ID and other radio resources. A fifth service to support RRC is power saving. This function allows the MTs to go to sleep mode in which the power consumption is greatly reduced because a large part of the receiver can be switched off. The MT can choose a sleeping time varying from 2 to MAC frames. In sleep mode the MT will only listen to the broadcast channel with a period determined by the sleeping time. At this moment, the MT will either wake up because the AP or the MT has new data to send or the MT will go back to sleep for another sleeping time period. 4.2.2.3.3 The DLC user connection control (DUCC) This control function is responsible for setting up, maintaining, re- negotiating and closing a DLC user connection (DUC) at the DLC layer. Both the MT and the AP can initiate the set up of the DUCC by a simple request. This request can either be acknowledged or rejected by the other side. The moment the DUCC was accepted, actual traffic transmission in the user plane can start. When people agree on OFDM 61 4.2.2.4 Mapping of the MAC frame to the PHY frame The different transport channels created by the MAC are combined in so- called PDU trains, which are forwarded to the PHY layer. There they are converted to PHY bursts, as we will see in the PHY operation (see section 4.2.3). HIPERLAN/2 distinguishes between 6 different PDU trains. A couple of them are described below. 1. Broadcast PDU train. This typically consists out of BCH-FCH-ACH and is transmitted by the access point at the beginning of each MAC frame. 2. FCH-and-ACH PDU train. This train is only used by an AP that makes use of multiple antenna sectors. We will not consider this burst further in this document. 3. Downlink PDU train. This consists of a set of SCHs and LCHs transmitted from the AP to a MT, as shown in Figure 4.5. Because a MT can maximally receive one downlink PDU train per MAC frame, sometimes multiple MAC frames are required to send all data from AP to MT over several downlink PDU trains. 4. Uplink PDU train with short preamble. This consists of a set of SCHs and LCHs transmitted from the MT to the AP, as shown in Figure 4.6. Because a MT can maximally transmit one uplink PDU train per MAC frame, sometimes multiple MAC frames are required to send all data from MT to AP over several uplink PDU trains. In addition to the SCHs and LCHs the MT may also make an access attempt with the RCH. 62 Chapter 4 5. Uplink PDU train with long preamble. This one is identical to the previous one except for the preamble. The BCH will announce which preamble has to be used in the uplink train. 6. Direct link PDU train. This PDU train consists of all LCHs and SCHs belonging to the same pair of source and destination MAC Ids, as shown in Figure 4.7. A set of SCHs and LCHs is granted for each DLCC by one RG. All corresponding DLCCs shall be grouped in a single PDU train. In Table 4.1 an overview can be found of how many OFDM symbols are needed to transmit the different transport channels depending on the code rate and the modulation that is being used in the physical layer. 4.2.3 Physical layer (PHY) For the purpose of elaborating the specification of physical layer functions, a reference configuration of the transmission chain is used as When people agree on OFDM 63 shown in Figure 4.8. It should be noted that only the transmission part is specified. The receiver is specified only via the overall performance requirements. This allows different manufacturers to develop their own intellectual property for algorithms and implementations. The PHY layer of HIPERLAN/2 offers information transfer services to the DLC of HIPERLAN/2. For this purpose, it provides functions to map the output of the DLC, the so called DLC PDU trains (see section 4.2.2.4), onto PHY bursts. These are appropriate for transmitting and receiving management and user information between an AP and a MT in the centralized mode or between two MTs in the direct mode. This includes the following functional entities in the transmitter: Configuring the transmission bit rate by choosing appropriate PHY mode based on the link adaptation mechanism, described in section 4.2.3.1. Scrambling the PDU train content described in section 4.2.3.1.1. Encoding the scrambled bits according to the forward error correction set during PHY layer configuration, described in section 4.2.3.1.2. Interleaving the encoded bits at the transmitter by using the appropriate interleaving scheme for the selected PHY mode described in section 4.2.3.1.3. Sub-carrier modulation by mapping the interleaved bits into modulation constellation points described in section 4.2.3.1.4. Producing the complex base-band signal by OFDM modulation described in section 4.2.3.1.5. Inserting pilot sub-carriers, appending appropriate preamble to the corresponding PDU train at the transmitter and building the PHY burst, described in section 4.2.3.1.6. Performing radio transmission by modulating the radio frequency carrier with the complex base-band signal at transmitter described in section 4.2.3.1.7. 64 Chapter 4 4.2.3.1 Different PHY modes The PHY layer of HIPERLAN/2 is based on the Orthogonal Frequency Division Multiplexing (OFDM) scheme. In a typical indoor communication scenario, the channel will vary as a result of varying multipath conditions or interference levels. In order to improve the radio link capability a multi-rate PHY layer is applied, where the appropriate mode will be selected by a link adaptation scheme. The data rate can be varied from 6 to 54 Mbps by using various signal alphabets for modulating the OFDM sub-carriers and by applying different puncturing patterns to a mother convolutional code. BPSK, QPSK, 16QAM are used as mandatory modulation formats, whereas 64QAM is optional for both AP and MT. The mode dependent parameters are listed in Table 4.2. 4.2.3.1.1 Data scrambler The content of each PDU train ( bits) from the DLC shall be scrambled with a length-127 scrambler. The scrambler uses the generator polynomial S (x) that is illustrated in Figure 4.9 and defined by The same scrambler is used to scramble transmit data and to descramble receive data. All PDU trains belonging to a MAC frame are transmitted with the same initial state for scrambling. The initialization sequence depends on the type of burst that is transmitted. For the details of this mechanism we refer to the standard document [3]. When people agree on OFDM 65 4.2.3.1.2 FEC Coder A channel encoder unit encodes the scrambled PDU train of bits. The encoder block diagram is shown in Figure 4.10. It consists of four blocks: code termination, encoding, code rate independent puncturing (P1), and code rate dependent puncturing (P2). The code termination, encoding, and puncturing P1 is performed in a specific way for different types of PDU trains. For the details of this mechanism we refer to the standard document [3]. 66 Chapter 4 The convolutional encoder block diagram is shown in Figure 4.11. It has a code rate 1/2 with 64 states. The generator polynomials of the mother code are for the X output and for the Y output. The encoder is set to zero state before the encoding process. The P1 puncturing assures that an encoded PDU train fits precisely into an integer number of OFDM symbols. The puncturing P2 has as role to provide code rates of 9/16 and 3/4 by puncturing the output of P1. 4.2.3.1.3 Data interleaving All encoded data bits are interleaved by a block interleaver with a block size corresponding to the number of bits in a single OFDM symbol, The interleaver is defined by a two step permutation. The first ensures that adjacent coded bits are mapped onto nonadjacent sub-carriers. The second permutation ensures that adjacent coded bits are mapped alternately onto less and more significant bits of the constellation to avoid long runs of low reliability bits. We denote by the index of the coded bit before the first permutation; i is the index after the first and before the second permutation and j is the index after the second permutation, just prior to modulation mapping. The first permutation is then defined by the rule where function floor(.) denotes the largest integer not exceeding the parameter, and mod is the integer modulo operator. The second permutation is defined by the rule where the value of s is determined by the number of coded bits per sub- carrier, according to: 4.2.3.1.4 Signal constellations and mapping Depending on the PHY mode selected for data transmission, the OFDM sub-carriers are modulated with BPSK, QPSK, 16QAM or 64QAM. The When people agree on OFDM 67 interleaved binary serial input data is divided into groups of (1, 2, 4 or 6) bits and converted into complex numbers representing BPSK, QPSK, 16QAM or 64QAM constellation points. The conversion is performed according to Gray coded constellation mappings. 4.2.3.1.5 OFDM modulation The stream of complex valued sub-carrier modulation symbols at the output of the mapper is divided into groups of complex numbers. Each group is transmitted in an OFDM symbol. All data OFDM symbols contain data in data carriers and reference information in pilot carriers. For data there are carriers and for pilots carriers in each symbol. Thus, each symbol is constituted by a set of carriers and transmitted with a duration This symbol interval consists of two parts: a useful symbol with duration and a cyclic prefix with duration The cyclic prefix is a copy of the last samples of the symbol part sent in front of the symbol part. The length of the useful symbol part is equal to 64 samples and its duration is For the cyclic prefix length there are two possible values in the HIPERLAN/2 system: mandatory 800 ns and optional 400 ns. The 800 ns guard interval is sufficient to allow good performance in channels with a delay spread up to 250 ns. The 400 ns guard interval can be used for communication in small indoor environments. Numerical values for the OFDM parameters are given in Table 4.3. The reference signal transmitted in the pilot carriers is defined as and is generated with the polynomial S(x) used for data scrambling: This scrambler is initialized at the beginning of each PDU train. 68 Chapter 4 The mapping from data and pilot complex symbols into the sub-carrier frequencies is shown in Figure 4.12. Here, stands for complex data symbol i transmitted at OFDM symbol n. The resulting OFDM symbols are extended with a cyclic prefix and concatenated to constitute the baseband PDU train, called payload. The structure of the payload section is illustrated in Figure 4.13. It consists of variable number of OFDM symbols required to transmit the PDU train payload. 4.2.3.1.6 PHY burst formatter The HIPERLAN/2 system distinguishes between five different kinds of PHY bursts 1. Broadcast burst 2. Downlink burst 3. Uplink burst with short preamble 4. Uplink burst with long preamble 5. Direct link burst (optional) The PDU trains delivered by the DLC are mapped onto the PHY bursts depending on the type of link and service that is required (see section 4.2.2.4). Independently of the burst type, each burst consists of two sections: preamble and payload. Each burst is started with a preamble section, which is followed by a payload section, The basic structure of a PHY burst is illustrated in Figure 4.14. When people agree on OFDM 69 Each MAC frame start with a broadcast burst that will be used for automatic gain control and channel estimation, as well as frequency and frame timing synchronization [13]. Therefore we will discuss the preamble of the broadcast burst more in detail. The broadcast burst consists of a preamble of length and a payload section of length The structure of the broadcast burst preamble is illustrated in Figure 4.15. It is composed of three sections: Section 1, Section 2 and Section 3. Section 1 consists of 5 specific short OFDM symbols that are denoted A and IA in Figure 4.15. The term "short OFDM symbols" refers to their length of 16 samples instead of a regular OFDM symbol of 64 samples. The first 4 short OFDM symbols in section 1 (A, IA, A, IA) constitute a regular OFDM symbol consisting of 12 loaded sub-carriers and given by the frequency-domain sequence SA. The last short symbol in section 1 (IA) is a repetition of the preceding 16 time-domain samples. SA is defined as Section 2 consists of 5 specific short OFDM symbols that are denoted B and IB in figure 12. The first 4 short OFDM symbols in section 2 (B, B, B, B) constitute a regular OFDM symbol consisting of 12 loaded sub-carriers ( and ) given by the frequency-domain sequence SB. The last short symbol in section 2 (IB) is a sign-inverted copy of the preceding short symbol B, i.e. SB is defined as 70 Chapter 4 Section 3 consists of two OFDM symbols (C) of normal length preceded by a cyclic prefix (CP). All the 52 sub-carriers are in use and they are modulated by the elements of a frequency-domain sequence SC. The cyclic prefix CP is a copy of the 32 last samples of the C symbols and is thus double in length compared to the cyclic prefix of a normal data symbol. SC equals Concatenating the above-described preamble with the data payload forms the broadcast burst. The resulting broadcast burst is illustrated in Figure 4.16. This figure further also shows the format of the other PHY bursts. 4.2.3.1.7 Radio transmission The complex digital data leaving the burst formatter have to be converted to the analog domain and upconverted to a radio frequency. The nominal frequencies for HIPERLAN/2 are allocated in two frequency bands. A lower frequency band from 5150 MHz to 5350 MHz and an upper frequency band between 5470 MHz and 5725 MHz. The nominal carrier frequency corresponds to its carrier number, which is defined as: The nominal carrier frequencies are spaced 20 MHz apart. All transmissions shall be centered on one of the nominal carrier frequencies. The center frequencies in Europe are shown in Table 4.4. When people agree on OFDM 71 72 Chapter 4 For each of the different subcarriers the transmitted signal has to fall within the spectral mask defined in Figure 4.17. When people agree on OFDM 73 4.3 DIFFERENCES BETWEEN HIPERLAN/2 AND IEEE 802.11A At the same time that the HIPERLAN/2 standard was drafted in Europe, the IEEE 802.11a standard saw the light in the US. Thanks to a harmonisation process, the physical layers for the two standards are very similar [10], [11], [12], [13]. The protocols are however completely different. 802.11a uses a distributed MAC based on carrier sense multiple access with collision avoidance (CSMA/CA). HIPERLAN/2 uses a centralised scheduled MAC. In Table 4.5, a comparison between of the 802.11a and the HIPERLAN/2 standards is provided. In Japan the Multimedia Mobile Access Communication (MMAC) Organization [14] is also working on a 5 GHz WLAN standard. Three ad- hoc sub committees are active in this field. The wireless home-link committee is investigating a wireless 1394 solution. The 5 GHz Ethernet workgroup is focussed on indoor PC networks and adopted the IEEE 802.11a standard. The 5GHz high speed wireless access sub committee looks both in indoor and access networks and supports the HIPERLAN/2 standard. Further harmonisation between the two standards is pursued. A major complication for the success of these standard in Japan is the fact that only 100 MHz, from 5.15 to 5.25 GHz, is available. This results in only 4 carrier frequencies. 74 Chapter 4 REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] http://www.etsi.org/technicalactiv/hiperlan2.htm#Overview/, overview of ETSI HIPERLAN/2. HIPERLAN/2 Global Forum, http://www.hiperlan2.com/web/. ETSI TS 101 475, Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Physical (PHY) layer. ETSI TS 101 761-1, Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Data Link Control (DLC) layer, Part 1: Basic Data Transport Functions. ETSI TS 101 761-2, Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Data Link Control (DLC) layer, Part 2: Radio Link Control (RLC) sublayer. US National Bureau of Standards, Data Encryption Standard, Federal Information Processing Standard (FIPS) Publication 46-2, December 1993. US National Bureau of Standards, Guidelines for Implementing and Using the Data Encryption Standard, Federal Information Processing Standard (FIPS) Publication 74, April 1981, http://www.itl.nist.gov/div987/pubs/fip74htm. ETSI TS 101 493-1, Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Packet based Convergence Layer; Part 1: Common Part. ETSI TS 101 493-2, Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Packet based Convergence Layer; Part 2: Ethernet Service Specific Convergence Sublayer (SSCS). Martin Johnsson, HIPERLAN/2 The Broadband Radio Transmission Technology Operating in the 5 GHz Frequency Band, version 1, HIPERLAN/2 global forum. Richard van Nee, A new OFDM standard for high rate wireless LAN in the 5GHz band, Proceedings Vehicular Technology Conference, Volume 1, pp. 258-262, September 19-22, 1999. Neeli R. Prasad and Harold Teunissen, A state-of-the-art of HIPERLAN/2, Proceedings Vehicular Technology Conference, Volume 5, pp. 2661-2666, September 19-22, 1999. Jamshid Khun-Jush, Peter Schramm, Udo Wachsmann, and Fabian Wenger, Structure and Performance of the HIPERLAN/2 Physical Layer, Proceedings Vehicular Technology Conference, Volume 5, pp. 2667-2671, September 19- 22, 1999. Multimedia Mobile Access Communication (MMAC) Systems, http://www.arib.or.jp/mmac/e/index.htm IEEE Std 802.11a, Wireless LAN medium access control (MAC) and physical layer (PHY) specifications: High-speed physical layer in the 5GHz band, September 1999. Chapter 5 Beating the wireless channel Channel estimation and equalisation Luc Deneire IMEC 5.1 INTRODUCTION To achieve high data rates with good performance in current OFDM systems for Wireless Local Area Networks, we use coherent detection. The coherent detector relies on the knowledge of phase and amplitude variations that are present on each (flat fading) subcarrier, i.e. on the knowledge of the channel. Channel estimation can be done in various ways: with or without the help of a parametric model, with the use of frequency and/or time correlation properties of the channel, blind or training based, adaptive or not. Non parametric estimation determines the quantities of interest without relying on a specific channel model, whereas parametric estimation relies on a specific channel model, determines the parameters of this model and infers the quantities of interest (for example the frequency response). Parametric channel estimation usually offers better performance, as the number of quantities to estimate is smaller, but potentially suffer from model mismatch problems. Time and frequency correlations are specific properties of the wireless channel. Straight-forward non-parametric estimators do not take this property into account, whereas more advanced estimators explicitly or implicitly take profit of it. Training-based estimation techniques are common in most communication systems, where the sender emits some known signal, to achieve synchronisation and channel estimation (or equaliser training). Blind estimation on the other hand only relies on the properties of the signals (their 76 Chapter 5 statistical properties, like cyclo-stationarity or their deterministic properties, like their so-called Finite Alphabet property or constant amplitude property for GSM). Although blind estimation is seldom used in practical OFDM systems, we provide some references to it [15][16][17][18]. Adaptive channel estimation is needed whenever the channel is varying rapidly in time. Most channel estimators are derived in the non-adaptive frame (one usually talks about an acquisition phase) and further extended to the adaptive case (by merely extending the block algorithm or by performing specific parameter tracking). For the wireless OFDM channels, this chapter provides an overview of the major estimation techniques, along with a discussion on their characteristics in terms of performance and complexity. 5.2 CHANNEL MODELS AND CHARACTERISTICS This section recapitulates the system model of chapter 3 for wireless OFDM that is valid when the cyclic prefix condition (i.e. the channel length is shorter than the cyclic prefix) holds. 5.2.1.1 Notations To ease mathematical derivations, this chapter relies on a vector/matrix description of the signals. The following notations apply: normal letters represent scalar quantities, boldface letters represent vectors and boldface capitals are matrices. Slanted (respectively roman) letters indicate time (respectively frequency) domain quantities. and respectively mean transpose, conjugate transpose and Moore-Penrose pseudo-inverse of X. If X is full column rank, then is the orthogonal projection onto the space spanned by the columns of X. 5.2.1.2 Transmission model and training set-up OFDM modulation consists in multiplexing QAM data symbols over a large number of orthogonal carriers. To this end, the QAM symbols of an OFDM symbol are passed through an Inverse Fast Fourier Transform (IFFT). In the presence of a time dispersive channel, a Cyclic Prefix (CP) is prepended to each OFDM symbol to preserve orthogonality between carriers and eliminate Inter Symbol Interference (ISI). We consider a single user / single channel communication set-up (see Figure 5.1), with OFDM modulation, described by Beating the wireless channel 77 where denotes the Hadamard (i.e. element-wise) product of the columns of X with H. For a single OFDM symbol: An OFDM symbol is written as where denotes the number of carriers and m is a time (i.e. OFDM symbol) index. The latter is often omitted for clarity. After performing an IFFT and inserting a cyclic prefix of samples, the transmitted signal is transformed into For a channel response where the expression is the FFT of the channel response. Previous equations further take the additive (possibly coloured) Gaussian noise into account. Next, equalisation is a complex division on each carrier. Another classical view of this peculiar channel model is that the OFDM channel can be viewed as a set of parallel Gaussian channels (a complex gain followed by an Additive White Gaussian Noise), like sketched in Figure 5.2. Two types of training, Pilot Symbol Assisted Modulation (PSAM) and spectral shaping systems are considered. In classical training based estimation, all components of x are known. Spectral based systems use a minor modification of the classical training, zeroing a small number of carriers (named zero carriers) at the edges and in the middle of the utilised band. PSAM on the other hand bases it's channel estimation on a small fraction of the carriers, usually evenly spaced on the whole band, and possibly on varying positions from one OFDM symbol to the next. received vector is, after prefix removal and FFT, In this 78 Chapter 5 5.2.1.3 Non sample spaced channel, time synchronisation and channel estimation performance. The above channel model assumes sample spaced channels. A sample spaced channel has all delayed impulses of its channel impulse response at integer multiples of the system sampling interval. Hence, the continuous Fourier transform of the channel impulse response is a channel frequency response with non-zero values at multiples of the systems sampling rate. Due to this particular frequency response, and for the usual case of a sample rate which is (possibly a multiple of) the symbol rate, the samples of the channel frequency response coincide exactly with the DFT of the channel impulse response. In the non-sample-spaced channel, the channel is actually resampled in the receiver sampling process. As a consequence, there is no exact correspondence between the channel frequency response and the DFT of the sampled channel impulse response. Actually, the resampling process results in an extension of the equivalent sample-spaced channel impulse response (sketched on the right part of Figure 5.3). This leads to the need for a longer guard interval to ease synchronisation. It can also cause performance loss of the channel estimator, if it is based on a limited channel length assumption. Beating the wireless channel 79 Furthermore, the Cyclic Prefix Condition also relies on a correct frame (timing) synchronisation. In the case of early synchronisation 80 Chapter 5 (synchronisation tick is given earlier than should be), the effect is simply that the measured channel begins with some zeros, and hence there is some additional guard interval needed to fulfil the Cyclic Prefix Condition. In the late synchronisation case, the Cyclic Prefix Condition is violated. Hence, in the system design, the maximum synchronisation error should be specified and taken into account in the length of the Cyclic Prefix. The effects of late and early synchronisation are illustrated in Figure 5.4. 5.3 ONE-DIMENSIONAL CHANNEL ESTIMATORS The structure of OFDM signalling allows a channel estimator to use both time and frequency correlation. Such a two-dimensional estimator structure is generally too complex for a practical implementation. To reduce the complexity, separating the use of time and frequency correlation has been proposed [1], still with prohibitive complexity. To pave the way towards low complexity estimators, we will only use the frequency correlation of the channel in the estimation. Two different types of block-oriented channel estimators for OFDM are discussed. The first one relies on a linear minimum mean-squared error (LMMSE) estimator using only frequency correlation and on low-rank approximation theory to achieve low complexity. This method was proposed by Edfors et al. [2]. The second estimator relies on the sample-spaced channel model and proposes a Maximum Likelihood estimator in this framework. Apart from optimality (under the channel model hypothesis), this method, somewhat surprisingly, leads to a low complexity implementation and was proposed by Deneire et al. [3]. Next, we discuss a third estimator that goes a step further in exploiting the channel model. It relies on the non sample-spaced multipath channel model, employing the ESPRIT (estimation of signal parameters by rotational invariance techniques) method to do the initial multipath time delays acquisition and using an interpath interference cancellation delay locked loop to track the channel multipath time delays. This estimator was proposed by B. Yang et al. [10]. As a reference for these estimators we first introduce the straightforward Least Squares (LS) estimator. 5.3.1.1 The LS estimator The first and simplest channel estimator one can imagine consists simply in dividing the received signal by the symbols that have been actually sent (and that are supposed to be known). This estimator is usually known as the Beating the wireless channel 81 Least Squares estimator, and can be written as (where the division sign means element-wise division of y by x): The main advantage of this estimator is its simplicity: one division per carrier. The main disadvantage is it's poor performance, due to the use of an oversimplified channel model. Indeed, the frequency and time correlation of the channel are not taken into account in the LS estimator, as it is based on the parallel Gaussian channel model sketched in Figure 5.2. The frequency correlation, closely linked to the short channel impulse response of the channel, can be used both in the non-sample spaced channel model (leading to LMMSE and parametric multipath-based estimators) and in the sample spaced channel model (leading to the Maximum Likelihood estimator). The time correlation (linked to the slow variation of the channel), can also be used in the same frames, although it is usually not used as such, due to the delay that would be introduced by needing several OFDM symbols before estimation (and hence detection). Time correlation can be implicitly used by decision-feedback mechanisms. 5.3.1.2 The LMMSE estimator The Linear Minimum Mean Squared Error channel estimator tries to minimise the mean squared error between the actual and estimated channels, obtained by a linear transformation applied to The standard estimation theory tell us that this estimator is : where is the correlation matrix of the channel, is the noise power, DIAG(x) is the diagonal matrix with elements equal to the vector x taken in argument and CONJ(x) is the conjugate of it's argument x. The LMMSE estimator proposed is relying on a complex matrix multiplication, which implies a very high complexity compared to the two DFTs and divisions needed for the normal OFDM modem. The low complexity approximate LMMSE estimator relies on the fact that the channel correlation matrix is nearly rank deficient, due it's frequency correlation. Hence, one can apply rank reduction and only use the most significant part of Optimal rank reduction is achieved by using the singular value decomposition (SVD) [5]. The SVD of the channel correlation matrix can 82 Chapter 5 be written as where is a diagonal matrix with the singular values on it's diagonal and U is the unitary matrix formed by the singular vectors. It can then be shown [3] that the best rank-p estimator of the LMMSE estimator is given by: where is a rank-p diagonal matrix whose p first elements are equal to and the last elements are zeros. Factor is a constellation dependent constant. Based on this estimator, one can reduce the complexity from complex operations to operations. Hence, the complexity decrease is a function of the frequency correlation of the channel. Further refinements in the complexity analysis are given in the following section, devoted to the ML estimator. 5.3.1.3 The Maximum Likelihood OFDM channel estimator The frequency correlation implicitly used by the low complexity MMSE estimator is linked to the finite delay spread of the channel. This has been recognised by Rayleigh and Jones [6], who estimate the channel from a subset of the carriers, but restrict them to regularly spaced pilots (in frequency). Based on the sample-spaced deterministic model of the channel, we derive the associated Maximum Likelihood estimator, show that it can be interpreted as a transformation from frequency domain to time domain and back to frequency (like for the LMMSE). The actual estimation is done in the time domain, where the number of parameters (i.e. the channel length) is small. The estimator is obtained by minimising a quadratic criterion, which, combined with the small number of parameters, leads to a low complexity algorithm. As such, we have obtained an exact low complexity solution and extended it to Pilot Symbol Assisted Modulation (PSAM). Beating the wireless channel 83 5.3.1.3.1 Reduced order model Since the time domain channel h has a finite length (in a well-designed OFDM system smaller than the prefix length) these parallel channels feature correlated attenuations. Considering, without loss of generality, that x equals the received signal can be written as : where F is a FFT matrix. y is a Gaussian random variable with mean and covariance matrix However, the signal part of y is contained only in the space spanned by its mean. Separating the signal subspace from the noise only subspace, the received signal can be rewritten as (with partitioning of the F matrix): Relying on this, the reduced space signal is defined as where v is a zero mean Gaussian noise with covariance matrix If then where denotes an identity matrix of size The ML estimator is then given by [4]: where denotes the orthogonal projection on the column-space of As the channel estimator is the cascade of an IFFT, a weighting matrix and an FFT. This is equivalent to going from the frequency domain to the time domain, force the time channel estimator to be of length ( denoting the estimated channel length) and going back to the frequency domain. 5.3.1.3.2 Extensions to PSAM and spectral shaping systems The above derivation assumes that all carriers are present. However, in the case of Pilot Symbol Assisted Modulation and spectral shaping systems, not all symbols in x are known, and only a subset of measured carriers 84 Chapter 5 can be used. Only this part of the signal (noted ) will be used and equation (5.9) becomes where measured pilots have been grouped together and defined accordingly. The ML estimator for spectral shaping systems is (only the measured carriers are estimated, as they are the only ones carrying data) and, for PSAM, it is the whole channel is estimated). For these estimators a similar time-frequency interpretation is possible, with the following modifications : The initial IFFT is partial, as only part of the carriers are measured. The non-trivial part of the channel impulse response is weighted by 5.3.1.3.3 Combination of PSAM and Decision-Feedback (DF) The classical ML solution can be applied to a combination of PSAM and decision-feedback. Indeed, suppose we use the pilot symbols along with decisions taken on the other carriers, then remains valid, with a given which leads to Hence, if the designer can afford the increment in complexity, combination of PSAM and decision-feedback is desirable. Indeed, Figure 5.8 shows that a difference in performance of 2- 3 dB can be expected between an all-pilot system (which is equivalent to combined PSAM/DF if decision errors are neglected) and a PSAM system with 8 pilot carriers. 5.3.1.3.4 Complexity The complexity of the ML estimator is significantly lower than the complexity of the LMMSE estimator, both for spectral shaping and PSAM systems. This low complexity relies on the time-frequency interpretation and on the partial (I)FFTs. Spectral shaping systems By construction, is a low rank matrix (of rank ). Taking its hermiticity into account, it can be written as where V is a matrix of size that can be precomputed. Hence, the complexity for computing the ML estimator is reduced to complex Beating the wireless channel 85 multiplications. The complexity reduction for this ML and for the LMMSE estimator are illustrated below. Further complexity reduction can be obtained by using the time- frequency interpretation. Indeed, the projection operation can be expressed by the cascade of two partial FFTs, weighted by a matrix (if all carriers are used as pilots, it is an identity matrix). With a radix-4 implementation of the FFT, the complete estimator would require complex multiplications. Furthermore, some additional complexity gain can be achieved by using FFT pruning or transform decomposition [7]. Such techniques lead to a significant gain for the Fourier Transforms. However, the last term due to the weighting matrix remains unchanged. PSAM When using Pilot Symbol Assisted Modulation, a comb spectrum (Figure 5.6) has to be measured, and only the teeth of this comb are used for the FFTs. This particular case has been studied by He and Torkelson [8]. In this case, the DFT can be computed with complex multiplications, which represents a large gain when the number of carriers is large. 86 Chapter 5 Figure 5.7 shows complexity evaluations of the four algorithmic approaches: the SVD-based approach; the frequency-time approach with plain FFTs ; the frequency-time approach with FFT pruning like Sorensen in [7]; the frequency-time approach with FFT optimised for a comb spectrum (He and Torkelson [8]); Note that the complexity gains are larger for a large number of carriers. The complexity for FFT-based solutions is much lower than for the SVD- based approach, both for spectral shaping and PSAM systems. Futhermore, simulations show that the ML algorithm can work with a significantly smaller than the LMMSE, which results in a still larger gain than appears in Figure 5.7. For a relatively large number of pilot carriers, the main contribution to the complexity is due to the weighting matrix (see the curves in Figure 5.7). However, for pure PSAM with regularly spaced pilot carriers, it can easily be shown that the weighting matrix is proportional to the identity matrix, and complexity is even lower. This special case of our algorithm is the frequency correlation part of the algorithm developed by Rayleigh and Jones [6]. Beating the wireless channel 87 5.3.1.3.5 Simulation results To evaluate the performances of the ML estimator, and compare it with the LMMSE algorithm, we simulated a spectral shaping system and a PSAM-based system in an indoor radio channel. An 64-carriers OFDM scheme, with zero carriers as in HIPERLAN/2 and a 256-carriers PSAM system are simulated. The data rate is 20 Msamples/second over the air (i.e. including the cyclic prefix) with a carrier frequency of 5.6 GHz. The simulation are based on a collection of 120 indoor office-like channels obtained by ray-tracing, with channel lenghts in the order of 4 to 6 taps. For the spectral shaping system, the Bit Error Rate (BER) is simulated for both LMMSE and ML estimators and for ranging from 4 to 16. BER based on exact channel knowledge and raw measurements are evaluated for comparisons. Simulation results (Figure 5.8) clearly show that the LMMSE suffers from a threshold effect at high SNR, as reported in [3]. To obtain similar performances for both algorithms, must be 2 to 4 times larger for LMMSE than for ML. For the PSAM-based system, similar conclusions hold. 88 Chapter 5 5.3.1.4 A Parametric approach based on ESPRIT. The previous method relies on a very simple parametric model of the channel, namely that the channel has a small number of multipath that are located (in time) at the OFDM system sampling instants. In a real system however, the location of the multipath will not coincide with the system sampling instants. This leads to a parametric multipath channel model. When the channel correlation matrix is constructed based on this parametric channel model, the signal subspace dimension of the correlation matrix can be reduced further than for the ML estimator. Accordingly, the channel estimator performance can be improved. Moreover, the high-speed data transmission in wireless communications potentially results in a sparse multipath fading channel, this sparsity can be used to lower the complexity of the estimator. B. Yang [10] proposed an improved channel estimation method for OFDM transmission over the sparse multipath fading channels using pilot subcarriers. The channel estimator is derived to estimate the parameters of the channel model, which include the time delays, gains, and phases of the paths. The number of paths is determined by an information theoretic criterion (the Minimum Description Length or MDL criterion). Then the initial multipath time delays are determined by Estimation of Signal Parameters by Rotational In variance techniques (ESPRIT) [11]. Furthermore, to be able to track the (slowly) time-varying channel, they propose an interpath interference cancellation (IPIC) delay locked loop (DLL) to track the channel multipath time delays and rely on a MMSE approach to estimate the frequency response of the channel. 5.3.1.4.1 System assumptions The OFDM system under consideration is a PSAM system, with evenly spaced pilot carriers, at fixed positions and repeated at each OFDM symbol (generalisations to other pilot patterns are possible). The pilot pattern is chosen so that the sampling rates (in frequency and time domain) allow the estimation of the channel (sampling rates will be function of the maximum delay spread and maximum Doppler spread). 5.3.1.4.2 The Acquisition Phase The first step consists in determining the number of multiple paths present in the channel. To this end, the Minimum Description Length method is used. It tries to minimise a criterion which is a function of the channel (basically via a log likelihood function) and which is penalised by the number of parameters to be estimated (in order to "balance" between the log likelihood, which is a monotonous function, and the number of parameters). Without going into the details of the derivations (see [10]), the Beating the wireless channel 89 method relies on the logarithm of the geometric mean of the singular values of the channel correlation matrix. This is already an order of complexity higher than previous methods (here the SVD has to be computed online, unlike the two other methods). Initial multipath time delays are estimated by the ESPRIT [11] method. This method relies on the eigendecomposition of a matrix which is formed with the L (L being the number of multipaths estimated here above) eigenvectors associated with the largest eigenvalues of the channel correlation matrix. 5.3.1.4.3 Tracking phase As the ESPRIT method is computationally complex, and as the channel is slowly varying, the time delays can be tracked by a classical Delay Locked Loop (DLL). However, much like in W-CDMA systems, Inter Path Interference (IPI) dominates the DLL performance. Hence, an IPI Cancelling method is needed. The main idea of IPIC is, based on the knowledge of the channel at the previous symbol time, to subtract the IPI from the received signal. Indeed, the received signal can be written as: Hence, to estimate the delay of path one can use the cleaned-up received signal: This signal is then given to the "classic" DLL. Once the delay parameters are known, the channel estimation phase relies on an adapted version of the LMMSE estimator, with the additional benefit that, through the knowledge of delay parameter, the signal subspace dimension of the correlation matrix is known. This knowledge enables low- complexity channel estimation with no performance penalty (like in approximate LMMSE, due to the approximation, or in the ML, due to the more simple channel model). 90 Chapter 5 5.3.1.4.4 Simulation results The authors considered a 1024 carrier OFDM system, with 16 QAM modulation on the 901 used carriers. The system occupies 5 MHz in the 2.4 Ghz frequency band. The guard interval consists of 64 samples and there are 32 evenly space pilot carriers (in fact 29, due to the 124 zero carriers that perform spectral shaping). The sample period is (hence the symbol duration is 0.205 ms). The channel is the "Vehicular A" channel specified by the ETSI for the evaluation of UMTS, consisting of 6 paths with a maximum delay of (i.e. about 12 sample periods) and the maximum Doppler frequency is set to be 100 Hz. The most important simulation results are that, in this particular scenario, the proposed estimator performs about 4 dB better than the LMMSE estimator in a 3 paths channel, and this improvement goes up when more channels are present. 5.4 TWO-DIMENSIONAL CHANNEL ESTIMATORS. Previous methods only take advantage of the frequency correlation of the channel. But, as the channel is slowly varying, one can also benefit of the rather strong time correlation. The aforementioned estimators can be extended in the time direction. After a brief discussion on the pilot patterns that can be used in 2-D channel estimation, we will briefly point out the most relevant methods for the wireless channel, again stressing the complexity issues. 5.4.1.1 Pilot patterns When using pilot tones to perform 1-D channel estimation, the former methods use evenly spaced pilots. This choice can be theoretically justified by comparing the MMSE of the estimator for different set of pilots, as is pointed out by Negi and Cioffi in [12]. They also show that, from a point of view of performance, properly chosen pilot grids for the 2-D case could yield some advantage. In the case of 2-D channel estimation, the most natural choice is to use a rectangular time-frequency grid, like sketched in Figure 5.9, although other choices, like the triangular one in the same figure, are also possible. This particular issue has been tackled by Garcia et al in [13], where they compare the BER for different patterns, assuming a 2- D Wiener filtering and the same number of pilots. They come to the conclusion that the hexagonal pilot pattern provides better performance than rectangular pilot patters and triangular pilot patterns. Beating the wireless channel 91 5.4.1.2 The 2-D estimators. The best 2-D linear estimator is obtained by 2-D Wiener filtering. Stacking the Least-Squares estimates of the channel at the pilot location in a vector the LMMSE estimator can be written as: where is the cross-correlation matrix between the LS estimator and the LMMSE estimator and is the autocorrelation matrix of the 92 Chapter 5 LS estimator. This estimator, however, has a high computational complexity. To lower this complexity, separable filters have been applied instead of a 2D finite impulse response filter. Using this technique, the estimation is first performed in the frequency direction with a 1-D filter and then in the time direction. Similar low-complexity techniques as for the 1-D estimators can then be applied and are analysed in [14]. These estimators are relying on prior knowledge of some channel parameters (SNR, Doppler Frequency, etc.) and appear not to be robust against Doppler spread. To alleviate these shortcomings, Geoffrey Li [9] proposes a robust 2-D estimator. This estimator is based on the observation that the LMSSE filter can be shown to be a 2-D FFT, a (small) 2-D filter and a 2-D IFFT. This cascade accounts for the lower complexity of the estimator, while the increased robustness is due to the fact that the reduced number of parameters is less depending on the Doppler frequency and on the delays than the original set of parameters. Li reports a dramatic performance improvement for a Doppler frequency of 200 Hz, compared to a classical Decision-Directed channel estimator. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] P. Hoher, TCM on frequency-selective land-mobile fading channels, in Proc. Tirrenia Int. Workshop Digital Communications, Tirrenia, Italy, Sept. 1991, pp. 317328. O. Edfors, M. Sandell, J.J. van de Beek, S. K. Wilson, and P. O. Borjesson.``OFDM Channel Estimation by Singular Value Decomposition,".IEEE Trans. on Communications, 46(7):931--939, July 1998. Deneire, L.; Vandenameele, P.; van der Perre, L.; Gyselinckx, B.; Engels, M. "A low complexity ML channel estimator for OFDM" Communications, 2001. ICC 2001. IEEE International Conference on , Page(s): 1461 -1465 J. J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O.Borjesson, On channel estimation in OFDM systems, in Proc. IEEE Vehicular Technology Conf., vol. 2, Chicago, IL, July 1995, pp. 815819. L. L. Scharf, Statistical Signal Processing: Detection, Estimation, and Time Series Analysis. Reading, MA: Addison-Wesley, 1991. G.G. Rayleigh and V.K. Jones.``Multivariate Modulation and Coding for Wireless Communication,".em IEEE Journal on Special Areas in Communications}, 17(5):851--860, May 1999. H. V. Sorensen and C.S. Burrus, ``Efficient computation of the DFT with only a subset of input or output points," IEEE Trans. on Signal Processing, vol. 41, no. 3, pp. 1184--1200, march 1993 S. He and M. Torkelson, ``Computing partial dft for comb spectrum evaluation," IEEE Signal Processing Letters, vol. 3, no. 6, pp. 173--175, june 1996. Ye (Geoffrey) Li "Pilot-symbol-aided channel estimation for OFDM in wireless systems", IEEE Transactions on Vehicular Technology, vol. 49, no. 4, july 2000 pp., 1207-1215; Beating the wireless channel 93 [10] [11] [12] [13] [14] [15] [16] [17] [18] Baoguo Yang, Khaled Ben Letaief, Roger S. Cheng, and Zhigang Cao, " Channel Estimation for OFDM Transmission in Multipath Fading Channels Based on Parametric Channel Modeling"; IEEE Transactions on Communications, vol. 49, no. 3, March 2001 pp. 467-479 R. Roy and T. Kailath, ESPRITEstimation of signal parameters via rotational invariance techniques, IEEE Trans. Acoust., Speech, Signal Processing, vol. 37, pp. 984995, July 1989. Rohit Negi and John Cioffi, "Pilot Tone Selection for Channel Estimation in a Mobile OFDM System," IEEE Transactions on Consumer Electronic, vol. 44, no. 3, August 1998, pp. 1122-1129. M.J. Fernandez-Getino Garcia, J.M. Paez-Borallo and S. Zazo, "Pilot Patterns for Channel Estimation in OFDM," Electronic Letters, vol. 36, no 12, pp. 1049- 1059, June 2000. Ove Edfors, "Low Complexity algorithms in digital receivers", PhD Thesis, Lulea University of Technology, 1996. K. Abed-Meraim, W. Qiu, and Y. Hua, Blind system identification, Proc. IEEE, vol. 85, pp. 13101322, Aug. 1997. A. Scaglione, G. B. Giannakis, and S. Barbarossa, Redundant filterbank precoders and equalizers Part II: Blind channel estimation, synchronization, and direct estimation, IEEE Trans. Signal Processing, vol. 47, pp. 20072022, July 1999. B. Muquet, M. de Courville, P. Duhamel, and V. Buenac, A subspace based blind and semi-blind channel identification method for OFDM systems, in Proc. IEEE-SP Workshop on Signal Proc. Advances in Wireless Comm., Annapolis, MD, May 912,1999, pp. 170173. R. W. Heath and G. B. Giannakis, Exploiting input cyclostationarity for blind channel identification in OFDM systems, IEEE Trans. Signal Processing, vol. 47, pp. 848856, Mar. 1999 Chapter 6 Avoiding a tower of Babel Synchronization Luc Deneire IMEC 6.1 INTRODUCTION Before an OFDM receiver can really demodulate data, it has to synchronise itself in time and frequency with the transmitter. In this chapter the effects of carrier frequency offset, symbol timing misalignment and clock offset on the performance of the OFDM receiver will be analysed. Classical solutions to the estimation of these parameters will be presented as well as some recent developments. The first section of this chapter describes the effects of symbol timing, sampling clock frequency and carrier frequency offsets. This part will follow the approach of the "classical" papers in OFDM synchronisation, like the papers by Pollet et al. [1] and Speth et al. [2]. A symbol timing offset (i.e. a timing offset that is a multiple of the sample duration) will add additional Gaussian noise, with variance proportional to the symbol offset, to the received signal. Moreover, as depicted in previous chapter, there is potentially an additional noise due to the truncation of the channel impulse response. A carrier frequency offset creates Inter Carrier Interference (ICI) and also induces a time-variant rotation of the symbol constellation on each subcarrier. A clock frequency offset produces a symbol rotation that is proportional to the subcarrier index. The second section describes the major synchronisation algorithms, categorising them in methods based on the auto-correlation of the input signal, on the cross-correlation of the input signal with the received signal and on the use of the cyclic prefix. 96 Chapter 6 6.2 EFFECTS OF OUT OF SYNC TRANSMISSION In this section, we first discuss the effect of symbol timing and sampling timing offsets. Next, we discuss the effect of a frequency offset on the carrier or sampling clock. 6.2.1 Symbol Timing Offset Before performing the channel estimation, equalisation and demodulation, the correct timing for the OFDM symbol has to be acquired. Perfect synchronisation is achieved if the data block selected for the FFT processor corresponds exactly to the transmitted IFFT block The most important effect of a timing misalignment is the violation of the cyclic prefix condition, as well as additional noise due to the estimation of a shortened channel. On top of these effects, there is additional ICI and ISI. 6.2.1.1 Violation of the cyclic prefix condition The effect of a mismatch in symbol timing is different for early and late synchronisation. In the case of late synchronisation, i.e. when the synchronisation tick is later than the perfect synchronisation point, a block of received data is selected that contains part of the cyclic prefix of the next symbol. This breaks the orthogonality of the subcarriers and leads to ICI. As a consequence, late synchronisation should be avoided in all cases. Early synchronisation, on the other hand, has no major effects if the cyclic prefix is sufficient long. Indeed, the effect of early synchronisation is simply that the measured channel begins with some zeros. To avoid ICI, the guard interval should be sufficiently long to fulfil the Cyclic Prefix Condition for this extended channel. Hence, in the system design, the maximum synchronisation error should be specified and taken into account in the length of the Cyclic Prefix. 6.2.1.2 Effect on the channel estimation performance. Moreover, for late synchronisation, part of the channel response can fall out of the channel estimation window (especially for the parameter-based estimation methods). Hence, there will be an additional noise due to the estimation error, which can be easily computed as Avoiding a tower of Babel 97 where is the number of data carriers. 6.2.1.3 Effect of a symbol timing offset on a subcarrier. Following Speth et al. [2], we assume a late timing offset of p.T samples, and express the received data symbols for carrier k and symbol m. For an AWGN channel, with a guard interval the vector of received signals can be written as Demodulation of this vector via the FFT yields [2]: According to this equation the effect of the symbol timing offset consists of three terms: a phase rotation, an ICI contribution and an ISI term. The 98 Chapter 6 phase rotation is proportional to the subcarrier index k. In addition, the received signal is attenuated. The ICI is due to the loss of orthogonality. The ISI term, on the other hand, occurs because of the inclusion of the next symbol in the received channel. Extending this analysis early synchronisation in case of a cyclic prefix the received vector for the symbol becomes Performing the same analysis as above learns that the ICI and ISI terms in (6.2) vanish. Also the attenuation factor disappears. As a consequence, the only remaining effect is the phase rotation, which can be handled by the equaliser. The extension to the multipath propagation case is easy to do. Indeed, as long as the sum of the offset and the maximum delay spread (expressed in samples) is smaller than the guard interval, the only effect will be the phase rotation, which can be handled by the equaliser. If the offset is larger, the attenuation term applied to the symbols appears to be negligible, whereas the effect of the ICI and ISI can be modelled as an additive white Gaussian noise whose variance is approximately linear with the offset p (see [2]). 6.2.2 Sample Timing Offset The analysis that was carried out here above relies on an integer offset (with respect to the sampling time). For an additional constant fractional offset, the only effect is a phase shift in the received signal, which can also be handled by the equaliser. Timing jitter, however, translates into phase noise, which has a detrimental effect on performance. The effect of phase noise is treated in the next chapter. 6.2.3 Carrier and Sampling Clock Frequency Offset OFDM is much more sensitive to carrier frequency offset (CFO) and sampling clock frequency offset (ScFO) than single carrier modulation (see [3]). Under these effects the received samples experience a phase rotation given by where is the CFO. The ScFO is defined as where T is the perfect and T the actual sampling interval. The received signal under the effect of this phase rotation of the samples can be expressed as: Avoiding a tower of Babel 99 From this equation, we can conclude that the carrier frequency and sampling frequency offsets create a loss of orthogonality between the carriers, resulting in ICI. In addition to this ICI, we can identify two other effects of CFO and ScFO: an OFDM symbol window drift, mainly due to the ScFO and a subcarrier symbol rotation, due to both CFO and ScFO. 6.2.3.1 OFDM symbol timing drift In the absence of sampling clock synchronisation, i.e. with a sampling clock offset of the frame sampling instant will drift, for an observation time of l OFDM symbols, of samples. In the case of HIPERLAN/2, the standard assumes 20 ppm clocks between transmitter and receiver: maximum mismatch is This results in a complete sample shift every 25000 samples, i.e. 312.5 OFDM symbols. Hence, for a burst of about 1000 OFDM symbols one can undergo a shift of 4 samples. Note that for other systems where the number of carriers can be much higher (e.g. 8192), like in Digital Video Broadcasting, this effect can be more severe. 6.2.3.2 Subcarrier symbol rotation Translating equation (6.3) in the discrete time domain yields the following expression [2]: denotes the OFDM symbol time, excluding the guard time. The attenuation due tio the symbol rotation equals and is neglectable in tracking mode. If we evaluate this formula for a specific carrier k and determine the phase increment from one OFDM symbol to another, we obtain: 100 Chapter 6 A constant phase rotation from one symbol to another is produced by the carrier frequency offset. The sampling clock offset, on the other hand, produces a phase rotation that grows with the subcarrier index. Considering a similar test case as above, and 64 carriers, the maximum phase difference between carriers on one OFDM symbol for is 6.2.3.3 Implementation loss due to ICI and attenuation of the received signal. Following [3], the combined effect of the attenuation and ICI generated noise leads to an implementation loss D (i.e. the extra SNR needed to have constant BER performance) of: where denotes the SNR at the receiver while B denotes the total bandwidth of the received signal. This equation tells us essentially that the loss grows linearly with the ratio of the frequency offset over the subcarrier spacing in a log-log graph. Moreover, to have an acceptable degradation for an HIPERLAN/2 system, the (residual) frequency offset should be about two orders of magnitude smaller than the subcarrier spacing. Note that [2] derives slightly different results, incorporating the multipath channel coefficients, but ignoring the effect of attenuation and with a different approach. 6.3 TIMING SYNCHRONISATION Timing synchronisation techniques can be roughly divided in auto- correlation methods on an OFDM symbol, cross-correlation methods on an OFDM symbol, and methods based on the cyclic prefix. Avoiding a tower of Babel 101 6.3.1 Methods based on auto-correlation of an OFDM symbol The most famous auto-correlation method is probably the so-called Schmidl and Cox method [4]. Their symbol timing recovery searches for a training symbol that has two identical halves in the time domain. This search is performed by scanning the received signal, making the auto- correlation of two parts that correspond to half the length of an IFFT and searching for a plateau at the output of the auto-correlation circuit. This scheme is sketched in Figure 6.2. The timing metric of this synchronizer, in function of the signal r, is : 102 Chapter 6 The main advantage of this metric is its robustness against carrier frequency offset, as well as its robustness against fading (due to the "AGC" present in the denominator of the metric. Moreover, the signal r can be generated by a classical OFDM scheme, in which the two halves of the training symbol are made identical by transmitting a pseudo noise (PN) sequence on the even frequencies, while zeros are used on the odd frequencies [4]. The main disadvantage of the Schmidl and Cox method is the plateau- like timing metric, which gives rise to a relatively high uncertainty on the starting time of the symbol. A simple solution to this problem, which has been used in IMEC's solution (which is documented in chapter 8), is to resort to a sequence of inverted training sequences, which, after auto-correlation, give rise to a saw- tooth signal (Figure 6.3). From this saw-tooth, one can not distinguish the two halves of the training symbol, hence, a classical sequence of identical training sequences is appended, which gives rise to the classical plateau-like metric. The combination of these two sequences then gives a more accurate timing. The principle of this method is sketched in Figure 6.3. Note that the first part of this scheme has been proposed in [5], where a performance analysis is presented. Moreover, this method enables the use of special constant amplitude sequences (know as CAZAC sequences [11]), or of PN sequences, like in [7]. The big advantage of these methods, on top of their accuracy, is that they can base themselves on one bit quantisation of the received signal. This has two major benefits, first, the complexity of the digital part is very low, but perhaps most important, one can design specific synchronisation circuits with 1 bit analog to digital converters, and hence lower the power consumption significantly while waiting for an incoming burst. Avoiding a tower of Babel 103 6.3.2 Methods based on cross-correlation of the received signal. Following the idea to use special synchronisation sequence, like PN sequences, it is rather natural to implement the renowned matched filter approach, which was developed for CDMA systems. In this method, the receiver performs a correlation between the received signal, at different epochs, and a local copy of the emitted synchronisation sequence. This method is illustrated in Figure 6.4. Starting from this simple principle, a large number of cross-correlation methods can be found in the classical CDMA books. We will not detail the properties of these synchronisation algorithms, but it suffices to know that they can be very accurate, at the expense of a large computational load. Their robustness to multipath will depend on the used sequence. They require that the CFO has been compensated accurately before the signal is applied to the cross-correlation time synchronisers. Pursuing the similarity between CDMA and OFDM, as far as synchronisation is concerned, Tufvesson et al. [6] propose the use of a pilot signal, which is superimposed on the OFDM signal. The principle of this approach is to use a PN-based preamble for acquisition of the timing, followed by a PN sequence, whose length can be a multiple of the OFDM symbol length, which is superimposed on the OFDM signal. Hence, there is no wasted bandwidth for the tracking. Nevertheless, the accuracy can very high when long PN sequences are used. 104 Chapter 6 6.3.3 Methods based on the cyclic prefix. As was highlighted in the previous sections, the synchronisation calls for the use of additional information, in the form of additional symbols or pilots. One way to circumvent this loss of information is to use the cyclic prefix, which is present in most of OFDM systems, to derive the timing information. A simple scheme is illustrated in Figure 6.5. The above scheme corresponds to the optimal estimator in case of an AWGN channel. Indeed, when expressing the received signal under this condition and searching for the Maximum Likelihood estimation of the time delay, one finds the following solution [12]: Avoiding a tower of Babel 105 This estimator has a several advantages: it has a high spectral efficiency by not using pilots; it is based on a relatively simple expression and hence results in a reasonable implementation complexity; it can perform simultaneous time and frequency offset estimation. The main disadvantage of the estimator is that in presence of severe multipath its performance is reduced. Indeed, when multipath is present the range on which the signal is cyclic becomes smaller (going to zero when the channel is as long as the cyclic prefix). 6.3.4 Methods based on a training sequence considered as a cyclic prefix. A natural extension to the method mentioned above is to used several replicas of the same signal coming at regular intervals, that is, a training sequence. The principle of transforming a training sequence into a cyclic prefix is illustrated in Figure 6.6, along with the modification in the frame length with respect to the number of carriers. Notice that to be able to produce the training sequence, one has to put specific values on specific carriers. In a similar way to above one can derive the ML estimator [13], which is a multidimensional generalisation of (6.8). Indeed, instead of a correlation of a cyclic prefix with the corresponding part of the useful data, we can now use multiple correlations between subsequent training sequences. Figure 6.7 shows the estimation accuracy in function of the SNR with the number of used training sequences M as a parameter. The curve for corresponds to the classical cyclic prefix based method. As can be seen from the graph, the 106 Chapter 6 use of 3 training sequences results in an order of magnitude gain in estimation accuracy at low SNR, or an equivalent SNR gain of 6.4 FREQUENCY SYNCHRONISATION Frequency synchronisation can be divided in auto-correlation methods, and methods based on the cyclic prefix. Like for timing synchronisation, the cyclic prefix could also be replaced by a training sequence. 6.4.1 Methods based on auto-correlation of an OFDM symbol The first auto-correlation method was the so-called Moose method [14]. In this method a preamble is used that consists of two identical OFDM signals without a guard interval in between them. The method relies on the Avoiding a tower of Babel 107 fact that these two OFDM signals are related by the following expression, assuming a frequency offset and noiseless transmission: From this expression, one can easily show (see Appendix of [14]) that the Maximum Likelihood estimate of the frequency offset is given by: Intuitively, this result corresponds to the fact that the angle of is equal to (see Figure 6.8). For small frequency offsets, the tangent can be approximated by its argument, and it is than straightforward to show that the estimate of the frequency offset is conditionally (on the OFDM symbol and the frequency offset) unbiased. Hence, this estimator is optimal and it's variance is given by: 108 Chapter 6 Note that in the presence of multipath and provided that the cyclic prefix condition is fulfilled, the estimator is still ML, as relation (6.9) still holds. The main disadvantage of this method is the limitation to a frequency offset that is smaller than half the inter-carrier frequency spacing. This disadvantage is alleviated by Schmidl and Cox [4], by using a special sequence of two OFDM symbols that consists of identical halves. Lets consider the case where the frequency offset is larger than half the inter-carrier frequency spacing. If we apply an algorithm, similar to the Moose algorithm on the half OFDM symbols, we can write the frequency offset as a multiple of half the intercarrier spacing z plus a fraction The Moose algorithm on the half OFDM symbols can correct the frequency offset due to The two half OFDM symbols are generated as follows: before the IFFT, a PN sequence is generated and given as input on the even subcarriers, while zeros are given as input on the odd subcarriers. To also estimate z, a second OFDM symbol is generated such that the two OFDM symbols are differentially encoded (on the even subcarriers), that is, for an even subcarrier After the two OFDM symbols are corrected by the frequency offset due to the PN sequence will be shifted by 2z positions at the output. Moreover, even if the two OFDM symbols were identical (at the sender), there would still be a constant phase shift on all carriers. Hence, in a similar manner as for the timing synchronisation, one can find the shift by the following optimisation: Note that, as the core of Schmidl and Cox's algorithm is the same as the Moose algorithm, they share the same optimality properties, and hence the same performance (in terms of error variance) and are both unbiased. 6.4.2 Methods based on the cyclic prefix. As was highlighted in the previous sections, the synchronisation calls for the use of additional information, in the form of additional symbols or pilots. Avoiding a tower of Babel 109 The same approach to reduce this overhead as was used for the timing synchroniser, i.e. the utilisation of the cyclic prefix, can also be applied for the frequency estimation. This will result in the approach that is illustrated in Figure 6.9 and which is equivalent to ML estimation. As for the timing estimator based on the cyclic prefix, the main advantages of this estimator are it's simplicity and the absence of pilots. Again, it is less robust than auto-correlation methods against multipath. 6.4.3 Methods based on a training sequence considered as a cyclic prefix. Similar to the approach for timing synchronisation, we can use a training sequence, i.e. replicas of the same signal coming at regular intervals, in stead of the cyclic prefix. The principle of transforming a training sequence into a cyclic prefix remains identical to before and was illustrated in Figure 6.6. Again, we can derive a multidimensional generalisation of the cyclix prefix frequency estimator above. Indeed, instead of a correlation of a cyclic prefix with the corresponding part of the useful data, we can now use multiple correlations between subsequent training sequences. Figure 6.10 shows the frequency estimation accuracy in function of the SNR with the number of used training sequences M as a parameter. The curve for M=2 corresponds to the classical cyclic prefix based method. These results are very similar to the ones obtained for timing estimation. Indeed, the use of 3 training sequences achieves an order of magnitude gain in estimation accuracy at low SNR, or an equivalent SNR gain of 110 Chapter 6 REFERENCES [1] [2] [3] [4] [5] [6] Thierry Pollet and Miguel Peelers, "Synchronization with DMT Modulation", IEEE Communications Magazine, April 1999, pp. 80 -86. Michael Speth, Stefan A. Fechtel, Gunnar Fock and Heinrich Meyer, "Optimum Receiver Design for Wireless Broad-Band Systems Using OFDM -- Part I" IEEE Transactions on Communications, Vol. 47, No 11, November 1999, pp. 1668 - 1677. Thierry Pollet, Mark Van Bladel and Marc Moeneclaey, "BER Sensitiity of OFDM Systems to Carrier Frequency Offset and Wiener Phase Noise", IEEE Transaction on Communications, vol. 43, No 2/3/4, February/March/April 1995, pp.l91-193. Tim Schmidl and Don Cox, "Robust Frequency and Timing Synchronization for OFDM". IEEE Transactions on Communications, vol. 45, No 12, December 1997, pp. 1613-1621. H. Minn, M. Zeng, and V. K. Bhargava,"On Timing Offset Estimation for OFDM Systems", IEEE Communication letters, Vol. 4, No 7, July 2000, pp. 242-244 F. Tufvesson, M. Faulkner, P. Hoeher and O. Edfors, "OFDM Time and Frequency Synchronization by Spread Spectrum Pilot Technique"m 8 IEEE Communication Theory Mini Conference, ICC'99, Vancouver, Canada, June 1999, pp.116-119. Avoiding a tower of Babel 111 [7] [8] [9] [10] [11] [12] [13] [14] F. Tufvesson and O. Edfors, "Preamble-based Time and Frequency Synchronization for OFDM Systems", Technical Report, part of the PhD Thesis of F. Tufvesson, University of Lund, 2000. Baoguo Yang, Khaled Ben Letaief, Roger S. Cheng, and Zhigang Cao "Timing Recovery for OFDM Transmission", IEEE Journal On Selected Areas In Communications, Vol. 18, No. 11, November 2000, pp-2278--2291 M. Speth, F. Classen, and H. Meyr, Frame synchronization of OFDM systems in frequency selective fading channels, in Proc. VTC97, pp. 1807-1811. Uwe Lambrette, Michael Speth and Heinrich Meyr, "OFDM Burst Frequency Synchronization by Single Carrier Training Data", IEEE Communications Letters, Vol.1, No.2, March 1997, pp. 46-48. A. Milewski, "Periodic sequences with optimal properties for channel estimation and fast start-up equalization", IBM J. Res. Develop., vol 27, no 5, Sept. 1983, pp. 426-431. J.-J. van de Beek, M. Sandell, and P. O. Brjesson, ML estimation of time and frequency offset in OFDM systems, IEEE Trans. Signal Processing, Vol. 45, pp. 1800-1805, July 1997. Luc Deneire , Bert Gyselinckx, Marc Engels "Training Sequence vs. Cyclic Prefix. A new look on Single Carrier Communications" IEEE Communication Letters Vol.5, No7, pp. 292-294, July 2001. Paul H. Moose, "A Technique for Orthogonal Frequency Division Multiplexing Frequency Offset Correction", IEEE Transaction on Communications, Vol. 42, No. 10, October 1994, pp.2908-2914. Chapter 7 Living with a real radio Impact of front-end effects Boris Come, Jan Tubbax 7.1 INTRODUCTION Signals generated by the digital modem at the transmit side of a telecommunication system must be transported wired or wireless to the receive side of this system with minimum deterioration of the signal quality. Local Authorities (e.g. the FCC in the US) that set boundary conditions in terms of transmitted power, unintentionally radiated powers, and correct use of the available spectrum, regulate this transmission. The mixed-signal front- end transforms the original signal so that it can be transported without violating these rules in the allocated frequency band. As a result, mixed- signal front-ends hardly ever differ from the schematic in Figure 7.1: the digitally modulated signal is first converted into an analog signal, then up- converted in one or several steps to the assigned RF frequency and finally amplified so that it can reach the receiver with sufficient power. The first up- conversion is generally combined with I/Q modulation, which can be performed either before or after the Digital-to-Analog Converter(s). Each operation is followed by a filtering operation to avoid unintentional transmission in adjacent frequency bands. At the receive side, the signal undergoes similar operations in reverse order, from the receiver input (antenna or cable connector) to the Analog-to-Digital Converter(s). The first operation always consists of the separation of the wanted signal from blockers and signals in adjacent frequency bands (filtering). Power alignment between the transmitted and received signal is generally controlled at both sides: at the transmitter by the Power Control function to IMEC 114 Chapter 7 avoid transmitting excess power, and at the receiver by the Automatic Gain Control (AGC) to fine compensate for the channel losses. Given the high data rates with required low bit error rates, and given the nature of the OFDM signal, a conservative analysis of the front-end requirements for systems compliant with [1] and [2] lead to severe, over- dimensioned specifications. Such a design would never meet the market requirements for consumer applications i.e. low-cost and low-power consumption. To optimize the design at system level, the interaction between the front-end and the OFDM digital modem must be better understood. This makes it impossible for the analog front-end designers alone to set the front- end specifications. The front-end effects on the link performances are difficult to analyze as they address mixed-signal issues. Considering for example the phase noise on the local oscillator port of a RF mixer, or the nonlinear behavior of a power amplifier, how is the RF signal corrupted and how does it affect the decision block output of the decoder? More generally, how to relate the specification of a mixed-signal front-end effect to an Implementation Loss (IL) or Bit Error Rate (BER) degradation? To answer such questions for the dominant front-end effects, we will follow two approaches in this chapter: in Section 7.2, we will qualitatively explain the BER performance degradation and simplify complex interactions between the front-end and the digital modem by additional Gaussian noise source. This for example Living with a real radio 115 to derive an initial operating point for the front-end architecture definition, or to define for each front-end effect, a relevant range of values for which system simulations should be performed. in Section 7.4, we will go more into details for each front-end induced effect, giving either a theoretical analysis or references to literature, and presenting simulation results. These simulation results, mainly BER curves, required a complete end-to-end model of the OFDM link, that will be introduced first in Section 7.3. 7.2 HOW THE FRONT-END IMPAIRS THE OFDM MODEM 7.2.1 An overview of the front-end impact on the OFDM signal The principle of OFDM is the parallel transmission of QAM-modulated subcarriers using frequency division multiplexing: the time domain OFDM symbols at the output of the digital modem are generated from the QAM- modulated subcarriers through an IFFT, then parallel-to-series converted by a multiplexer on I and Q paths. The baseband time domain OFDM- modulated symbols are then passed to the transmitting front-end, of which the principle of operation has been previously discussed. The RF signal is then sent through the channel and down-converted to baseband at the receiver side by the receive front-end. After time-domain synchronization, the received symbols are series-to-parallel converted and fed to the input of a FFT that generates back the transmitted symbols, shaped by the channel response and corrupted by front-end effects. Before QAM-detection, a simple single-tap equalizer in the frequency domain corrects for the fading channel. The original OFDM-modulated signal will be affected by the transmit and receive mixed-signal front-end by several non-ideal processing steps: limited word-length of the signal paths in the digital domain, especially in the Digital-to-Analog and Analog-to-digital Converters (DAC and ADC); phase and gain mismatches between the I and Q paths in the I/Q (de)modulator, leading to cross talk between the I and Q signals and distortion of the constellation ; phase noise on the local oscillator port of the mixers, causing a rotation of the constellation and subcarrier interference as the power on each subcarrier is spread out onto all other subcarriers ; nonlinear distortion as the signal is amplified all along the transmit and receive chains, generating harmonics (of minor influence as they can be 116 Chapter 7 filtered), and intermodulation products falling on top of neighboring subcarriers; clock jitter on the DAC and ADC inaccuracy in gain control at the receive side addition of the filter impulse response to the channel impulse response, adding subcarrier-dependent signal-to-noise ratio variations and delay spread to the signal path. In this chapter, we will analyze the impact on the BER performances of the OFDM link for the above listed front-end non-idealities, either at the transmitting or at the receiving side (it will be pointed out in section 7.4 that it makes little or no difference). This list is not exhaustive, and many other front-end effects such as image rejection in mixing stages, blockers saturating the analog circuitry, interferers mixing with the wanted signal, etc... should be well understood and characterized at system level to prevent performances losses. Those are however strongly architecture dependent, which would bring the analysis beyond the scope of this chapter. 7.2.2 Expected degradation of the digital modem BER performances We will show in the following that by making simplistic assumptions one can easily predict the impact of some of the front-end effects on the BER performances. A rigorous theoretical analysis or modeling of the front-end effects on the performances of an OFDM link is out of the scope of this section, but will be further addressed in section 7.4. Implementation Loss and BER degradation: Given a reference BER- versus-SNR curve and simulated or measured BER-versus-SNR curve, the Implementation Loss (IL) is the difference in SNR between these two curves at a specified BER. Conversely, the BER degradation is the difference in BER between these two curves at a specified SNR. Effect of quantization and clipping: One major problem related to the use of OFDM is known to be the large peak-to-average power ratio of the time-domain signal at the output of the IFFT in the transmitter. Many authors rather use the crest factor, which is the square root of the peak-to- average power ratio. If no special measures are undertaken, this large crest factor would lead to very inefficient implementations in terms of word- lengths in the digital modem and data converters, and of dynamic range in the analog front-end. A method that limits the corrupting influence of large crest factor, rather than preventing the effect itself, is the clipping of the signals coming out of the IFFT, operation performed in the digital domain. This straightforward approach is well documented in the literature [3] and it has been shown that it results in acceptable implementation losses. The impact of digital clipping will be studied in greater details in Section 7.4.1, Living with a real radio 117 together with the effect of limited word-lengths of the signal representation in the digital domain, another effect that limits the signal-to-noise ratio at the front-end input. One of the important results worth mentioning here is that quantizing and clipping each generate additive Gaussian noise that are uncorrelated except for their powers. For a given word-length of the samples at the output of the IFFT, digitally limiting the crest factor of the signal to a lower value increases the clipping noise while at the same time it reduces the quantization noise. Effect of I/Q imbalance: I/Q imbalance can also be seen as a source of Gaussian noise. I/Q imbalance results from gain mismatches between the I and Q paths and non-perfect quadrature generation in the I/Q (de)modulator. This effect might be critical in direct down-conversion receivers where the I/Q demodulator operates at 5 GHz. It can be shown that a phase or amplitude mismatch between the I and Q paths in the I/Q (de)modulator has the same effect as a parasitic up-/down-conversion by a tone at the negative frequency of the LO frequency (for details see section 7.4.3). Hence it is often specified in terms of Negative Frequency Rejection (NFR), expressed in dBc. This is illustrated in Figure 7.2, where the negative frequency components of the OFDM signal (here considered as an interferer in the image band) are down-converted to DC on top of the wanted signal (positive frequency components of the same OFDM signal). Because the LO parasitic component reflects the I/Q imbalance, it is NFR dB lower than the LO wanted component. Hence the parasitic down-converted signal is NFR dB lower than the wanted down-converted signal. Furthermore, as this parasitic signal consists of the negative frequency components of an OFDM signal, its amplitude follows a Gaussian distribution. As a result, it can be considered as a Gaussian additive noise perturbation, with a SNR equal to the NFR. 118 Chapter 7 Effect of phase noise: Phase noise and carrier frequency offset in OFDM systems have been extensively described in the literature [4] [5] [6] [7]. Carrier frequency offset is specified in the WLAN standards [1][2] to a maximum of 20 ppm. This impacts the crystal oscillator selection. It necessitates frequency compensation in the digital modem receiver [8]. Due to phase noise on the LO, the signal on each subcarrier is shaped accordingly to the LO Power Density Function (PDF) and corrupts the neighboring subcarriers. In [4], the disturbance introduced by a local oscillator (LO) with phase noise is studied in great details at the digital receiver: the author differentiates the Foreign Noise Contribution (FNC) from the Own Noise Contribution (ONC). The ONC is the contribution of one noisy subcarrier to its own noise component. The ONC, which is identical for all subcarriers, can be considered due to the mixing operation, as a multiplicative noise source constant over one OFDM symbol. As a result, the phase of the ONC causes a common rotation of the whole constellation, whereas the magnitude of the ONC scales the whole constellation, just as an amplitude modulation. However, when the LO is used as a switching signal in a mixing operation (as in a Gilbert mixer, which is the assumption for the rest of this text), it drives the LO port of the mixer into saturation. In that case, amplitude modulation on the LO signal is not seen and do not affect the BER performances anymore. The FNC is the contribution of noisy subcarriers to the noise on one subcarrier. OFDM consists of the parallel transmission of a large number of subcarriers that are independantly QAM-modulated. Hence, according to the central limit theorem, the disturbance caused by the superimposed contributions of all corrupted subcarriers has a Gaussian distribution. As a consequence, the disturbance of an OFDM link due to a noisy VCO is two-fold. Due to close-in phase noise (ONC) the whole constellation in a symbol is rotated; and due to higher frequency components of the phase noise (FNC), the SNR on each subcarrier is limited by inter-subcarrier interference. For an OFDM link characterized by a channel bandwidth B and a number of subcarriers the constellation rotation can be digitally estimated as the mean phase deviation during the symbol period and compensated for on a per-symbol basis (Section 7.4.4). On the other hand, the degradation induced by the FNC, modeled as a Gaussian noise source, can not be compensated. However, its negative impact can be predicted by approximating the noise power as the integral of the phase noise PDF over the channel bandwidth, except for the sub-band that has already been accounted for. Effect of inaccurate gain control at the receiver: Inaccurate power estimation will induce an inappropriate decision of the Automatic Gain Control (AGC) algorithm. Due to process variations, tolerances of the bias and supply voltages, temperature, and other external parameters, the switching gains in the RF, IF and baseband variable gain amplifiers (VGA) Living with a real radio 119 will vary from their nominal values, resulting in inaccurate actions of the AGC. Large inaccuracies in the AGC can lead to saturation of the analog circuitry. This can be avoided by setting reasonable tolerances on the power estimation accuracy and on the gain of the analog blocks. However, inaccuracy of the gain control can not be completely avoided without pushing these constraints to non-realistic values. As a consequence a residual miss-alignment of the signal power with the ADC dynamic range will exist. Too much gain in the receiving front-end will saturate the ADC and clip the signal; insufficient gain will result in higher quantization noise. Hence the effect of inaccurate gain control at the receiver can be studied the same way as for quantizing and clipping, developed at the beginning of this section. Conclusion: The effect of clipping and quantization, I/Q imbalance, phase noise on the local oscillator (at least for it foreign noise contribution), and inaccurate gain control in the receiver can be seen as additive Gaussian noise source. Their power can be estimated and, as a result, the Implementation Loss IL introduced by these front-end non-idealities can be estimated as follows: from the front-end characteristics, the power of these equivalent Gaussian noise sources is computed, 120 Chapter 7 from the theoretical BER-versus-SNR curves for a Gaussian channel, the signal to noise ratio to obtain a given BER is extracted, these two noise sources are combined, and the equivalent global signal- to-noise ratio computed, the difference is the implementation loss IL at the selected BER. The curve IL versus is unique. It is plotted on Figure of As an example, an I/Q imbalance of 32 dB Negative Frequency Rejection will result in an implementation loss of 0.7 dB at and 0.9 dB at Conversely, an implementation loss smaller than 1dB at requires that the LO phase noise power integrated over the OFDM signal bandwidth is at least of 30.5 dBc. This approach holds as well for coded simulation: a fairly good approximation of the implementation loss can be derived without lengthy 7.3 together with, as an example, the values of for uncoded BPSK, QPSK, 16QAM and 64QAM. On Figure 7.4, the implementation loss due to Gaussian noise sources with power ranging from 25 dBc to 45dBc for 64 QAM transmission at BER Living with a real radio 121 BER simulations The same example applied to I/Q imbalance and resulting in 32dB NFR, is illustrate here: the simulated BER curve for 64QAM coded (convolutional coding is applied as in [1][2]), is plotted on Figure 7.5; the nominal signal-to-noise ratios to get a BER of is 18.2 dB ; the addition of two Gaussian noise source with powers of 18.39 dBc and 32 dBc is equivalent to a Gaussian noise source with a power of 18.20 dBc; hence, the implementation loss due to such I/Q imbalance on coded 64QAM is 0.19dB. Similarly, at a BER of the implementation loss will be 0.23 dB. Note that the curve in Figure 7.3 is still valid, only the values of account for the coding added to the signal! The above discussion is only an intuitive approach to complex front-end effects on OFDM data transmission. However, this approach can quickly guide the architecture designer to a first operating point with a given budget in terms of implementation loss distributed over the complete chain. This initial study must be checked and refine with extensive BER simulations including more detailed models for all the front-end non-idealities. Such simulations are detailed in Section 7.4. 122 Chapter 7 7.3 A SYSTEM SIMULATION TOOL To accurately quantify the impact of front-end non-idealities and to extract optimal front-end specifications, each design choice should be assessed from a system point of view, through the evaluation of the BER performances of the complete OFDM link. In the following, we briefly describe such a simulation model. Complete description of this tool can be found in [9] and [10]. 7.3.1 Baseband model for the OFDM link This model has been developed in MATLAB. It comprises baseband models for a digital baseband OFDM modem, a generic front-end and multipath channel responses. These channel responses, which correspond to an indoor environment, are either loaded from a set of simulation results obtained from a ray-tracing simulation tool [11], or generated at simulation time according to channel models that are described in [12]. The simulation model follows the structure of Figure 7.1. The spectrum of the OFDM signal in [1] and [2] occupies 16.8MHz in a 20MHz channel, and the digital modem produces I, Q signals sampled at 20MHz. To avoid high sample rates when simulating this band-pass signal at RF frequency of 5.2GHz, simulations are conducted at baseband using the complex low-pass equivalent representation. Yet, in the simulation over-sampling by 4 is added in the interface between the modem and the front-end so that nonlinear distortion due to clipping can be evidenced. Clipping is performed digitally on the I and Q paths, right after the time-domain signal generation. Note that over-sampling can also be present in hardware implementations, where the I/Q modulation is performed digitally. We will also show in Section 7.4.5 that introducing further clipping on the modulus of the up-sampled signal helps improving the system power efficiency. The analog front-end model consists of three blocks: an I/Q modulator adding I/Q imbalance, a mixer with a local oscillator signal defined by its phase noise power density function, and a power amplifier (PA) exhibiting a cubic non-linearity. The power amplifier is implemented with ideal power control that keeps the PA average RF input power constant, to make fair comparisons between the power transfer functions. In this chapter the front-end non-idealities are considered at the transmitting side only. Nonetheless, simulations show that they produce similar BER degradation if placed at the receiving side, even when considering multipath channels. Although only briefly mentioned here, all these non-idealities are fully implemented both at the transmitting and receiving side of the link. A schematic of the full link model is shown on Living with a real radio 123 Figure 7.6. Details on the models and simulation results will be given in the following section, where all these effects are investigated separately. 7.3.2 Considerations on the equalizer Channel equalization is performed at the receiver on the basis of the received Long Training Symbol (LTS): the LTS is a predefined OFDM- symbol prepended to the transmitted payload. At the receive side, the equalizer estimates the channel coefficients from the received LTS. As the thermal noise from the channel (white in a Gaussian channel, colored in multipath channels) as well as front-end non-idealities corrupt the payload and the LTS, the channel estimation will be impaired and the channel equalization imperfect. The current model used for the equalizer corresponds to a "worst-case" implementation, as none of these corruptive effects is compensated for; it typically results in an additional IL of 3dB at compared to simulations where perfect channel knowledge at the receiver is assumed (Figure 7.7). An actual implementation of the equalizer is expected to operate within these two limits. 124 Chapter 7 7.3.3 Validity of the model Uncoded OFDM: we validated this model by comparing theoretical BER curves and BER curves resulting from simulations. This comparison is illustrated here for uncoded BPSK and 64QAM modulations. Bit error probability for uncoded BPSK [13] is Bit error probability for uncoded 64-QAM [13] equals In these formulas Q(x) is defined as Living with a real radio 125 where erfc(x) is the MATLAB built-in Complementary Error Function. The received energy per bit versus noise spectral density can be expressed as with; is the bandwidth of the modulated signal; is the uncoded BPSK data-rate, is the uncoded 64QAM data-rate; Error! Objects cannot be created from editing field codes. is a correcting factor introduced to take into account the actual signal bandwidth (the zero-carriers and the pilot-carriers of the OFDM symbol do not convey data). For uncoded BPSK this results in: Similar derivations for uncoded 64QAM OFDM give In Figure 7.8, the theoretical BER curve versus SNR for uncoded 64QAM is plotted together with the BER curve resulting from a simulation. The OFDM link model used for this simulation was implemented with up- sampling by 4 at the transmitter and down-sampling by 4 at the receiver, plus all other non-idealities in the transmit and receive front-end as described previously, but set to negligible values. Furthermore, perfect equalization at the receiver is assumed. The two curves match by better than 0.1dB in SNR for BER ranging from to The theoretical and simulation curves for BPSK match even better. This validates the models in the system simulation tool. Remark: when assessing the performances of different modulation techniques, another correcting factor is often taken into account to compare 126 Chapter 7 their bit error probability versus For OFDM, is then corrected by a factor where is the length of the cyclic prefix, and the number of subcarriers to reflect the fact that part of the energy of a transmitted OFDM symbol is lost in the cyclic prefix. Coded OFDM: The coding gain when using convolutional codes can only be estimated. BER simulations have been performed including convolutional coding at the transmitter for BPSK with a coding rate of 1/2, and Viterbi soft decoding (with constraint length as in [1][2]) at the receiver. It resulted in a BER of at 0.55dB SNR or, equivalently equals 4.53 dB. For BPSK and Soft Decision Viterbi decoding the expected result for a BER of are [13] and coding The expected with coding is therefore 4.5 dB at to be compared with 4.53 dB we read from our simulations, which validates the model for coded OFDM. Living with a real radio 127 7.4 ANALYSIS AND SIMULATION OF THE MAIN FRONT-END EFFECTS In Section 7.2, we have shown that the effect of clipping and quantization, I/Q imbalance, phase noise on the local oscillator (at least for it foreign noise contribution) and inaccurate gain control in the receiver can be seen as additive Gaussian noise sources. We have seen as well that their power can be estimated and that, as a result, the implementation loss introduced by these front-end non-idealities can be estimated using a single reference curve, given in Figure 7.3. In this section, we will go more into details for each front-end induced effect, giving either a theoretical analysis or references to literature, and present simulation results. For most front-end effects we will present simulation results for 52 non- zero 64QAM-modulated subcarriers (uncoded or with a coding rate of 3/4) as this modulation is the most demanding case in the standards [1][2]. Furthermore, mostly Gaussian channels are considered here, as front-end effects are easier to analyze and predict in an Additive White Gaussian Noise channel (in short: AWGN or Gaussian channel) than in a multipath channel. 7.4.1 Word-length of the transmitted symbols and optimal clipping level The word-length b of the symbol at the output of the transmit digital modem has a major impact both on implementation cost and performance limitation. As b decreases, the power consumption and the complexity of the DACs decreases at the expense of the quantization noise hence the BER performances. However, b can be limited with acceptable performance degradation, due to the limited signal-to-noise ratio (SNR) on the channel. Furthermore, the digital clipping operation after the IFFT (see Section 7.3) limits the Crest- factor of the signal, enhancing for a given word-length b the average-signal- to-quantization-noise power ratio. Still, it also introduces an additional noise source: clipping noise. It is therefore a joint-optimization process on b and the normalized clipping level (ratio of the clipping level to the RMS amplitude of the time-domain signal). The global signal-to-noise ratio after quantization and hard-clipping operations on a modulated OFDM signal has been derived in [3]: 128 Chapter 7 where: Hence, for a given word-length b, an optimal clipping level can be derived, as it is a trade-off between two noise sources: lowering the clipping level enhances the clipping noise while at the same time reducing the quantization noise. In Figure 7.9, the SNR of an OFDM signal after clipping and quantizing is plotted versus for word-lengths b ranging from 6 to 9 bits: digital clipping at is close to the optimal clipping level for all word- lengths. Hence b can now be optimized separately. Figure 7.10 shows results from BER simulations considering only the digital modem and a Gaussian channel. Table 7.1 shows a comparison between the implementation losses obtained from simulations and expected from the derivations in Section 7.2.2. Living with a real radio 129 The IL due to the finite word-length b is less than 0.15 dB at for For smaller word-length, the implementation loss increases rapidly. Taking into account implementation cost of large word-length systems, 8-bit word-length with digital clipping at is an optimum. 7.4.2 Wordlength of the received symbols In the receiver a similar analysis can be performed for the word-length of the ADC(s). Here, we have to take into account the imperfect receiver gain setting at the input of the ADC as well as the Crest factor regrowth in the transmission channel. 130 Chapter 7 Crest factor regrowth: After extensive simulations in multipath channels, the crest factor regrowth due to frequency fading never exceeded 150%. Hence, the number of bits in the ADC should be slightly larger than the number of bits in the DAC of the transmitter. Imperfect receiver gain setting: We first assume for both the DAC at the transmitter and the ADC at the receiver 8-bit quantization with nominal clipping at Clipping at the transmitter is intentional as already discussed, whereas further clipping at the receiver should be avoided. The automatic gain control (AGC) is responsible for this task. It consists of a power estimator that provides input to the gain controller (control logic) that sets appropriate gains in the receiver front-end. If the signal at the input of the ADC is well positioned then it will be responsible for an implementation loss of 0.11 dB, just as the DAC in the previous section. If the receive front-end has too much gain, the Crest-factor of the signal will be reduced due to clipping in the ADC. The clipping noise power is thus enhanced, while the quantization noise power is lowered, but the total noise power increases as shown on Figure 7.9. The operating point shifts on the x- axis towards low Crest-factors, and the global SNR goes down from its maximum, following the curve corresponding to the ADC word-length. Conversely, when the receiver front-end has too low gain, the quantization noise is enhanced, the clipping noise is reduced, but again the global SNR moves out from its maximum. Using the results of Section 7.2.2, we can perform a straightforward sensitivity analysis on the implementation loss due to the gain control accuracy. We study the effect on uncoded 64QAM modulation, for which we target a maximum implementation loss due to the ADC quantization and clipping of 0.5 dB at a BER of From Figure 7.3, we conclude that 0.5dB IL corresponds to For uncoded 64QAM, needs to be 24.6 dB to get a BER of so that the SNR resulting from clipping and quantizing (the additional Gaussian noise source) should be dB. From Figure 7.9, we read that clipping and quantizing results in 34.2 dB SNR for a word-length of 8 bits with clipping at or at These clipping factors, normalized to and converted in dB, correspond to the maximum and minimum (power) gain mismatches +1.18/-3.35 dB that can be tolerated to get an implementation loss smaller that 0.5dB at a BER of These requirements are quite severe, as gain mismatches can raise both from implementation, but as well from power estimation errors in the AGC, unavoidable due to the time-limited estimation process. A straightforward solution to cope with limited AGC accuracy is to increase the number of bits in the ADC. Choosing for a 10-bit ADC solves Living with a real radio 131 the problem, as this would give a safety margin of +3dB/-3dB on the required power gain accuracy at the input of the ADC. As a result, a possible architecture choice to combat excess implementation loss due to crest factor regrowth and imperfect receiver gain setting could be: for the transmitter, 8-bit DACs, of which the 2 MSB are for signal amplitudes above RMS amplitude (that is clipping at ), for the receiver, 10-bit ADCs. 7.4.3 Imbalance between I and Q paths I/Q imbalance results from two effects: gain mismatch between the I and the Q paths and non-perfect quadrature generation ( and respectively). The effect of a mismatch between the I and the Q paths has already been qualitatively described in Section 7.2.2: it has the same effect as a parasitic up-or down-conversion by a tone at the negative frequency of the LO frequency. Equation (7.24) shows this effect and quantifies it. The power ratio between the positive and the negative tones of the local oscillator is referred to as the Negative Frequency Rejection (NFR) in dBc. It relates NFR and by: I/Q imbalance on is modeled at baseband with: Table 7.2 shows a comparison between the implementation losses obtained from simulations and expected from the derivations in Section 7.2.2. 132 Chapter 7 These results show a discrepancy between the IL obtained from BER simulations and from the derivation in Section 7.2.2. Furthermore, this error is increasing with the I/Q imbalance. However, for the kind of implementation losses we target (around 1 dB at a BER of ) the derivations of Section 7.2.2 still give acceptable results. These simulations were repeated with the multipath (MP) channel library from the system simulation tool presented in Section 7.3.2. A negative frequency rejection of 32 dB in AWGN channels gives 0.19 dB implementation loss for coded 64QAM (Section 7.2.2); this resulted in comparable IL for about 80 % of the MP channels in our library. The NFR specification for coded 64QAM should be increased to 35 dB to cover 86 % of the MP channels and to 39 dB to cover 98 % of them. As the nominal NFR of 32 dB is already at the edge of today's ICs, these few dBs difference in NFR specification for operation in MP channels will have a large influence on the front-end cost! 7.4.4 Phase noise Phase noise has been studied extensively in literature [4] [6]. It is shown that phase noise has two effects on an OFDM symbol: the Own Noise Contribution (ONC) and the Foreign Noise Contribution (FNC). The ONC comes from the low frequency part of the phase noise. It results in an identical rotation of all subcarriers, which is therefore also called the Common Phase Error (CPE). This CPE can be estimated and thus corrected. The FNC comes from the high frequency phase noise contributions. On a subcarrier it gives rise to interference from all other subcarriers, also called Inter Carrier Interference or ICI. This can be represented as an additional Gaussian noise source, which cannot be corrected. Since in our case the inter-carrier-spacing is quite large with respect to the phase noise bandwidth the ONC is dominant over the FNC and the FNC can be neglected. This means that every OFDM symbol is rotated over the CPE. Moreover since the variations in phase noise will be a lot slower than the period of an OFDM symbol. Therefore this CPE can be considered constant over the duration of one OFDM symbol. This effect is shown in Figure 7.11 and Figure 7.12. Figure 7.11 shows the effect of phase noise on one OFDM symbol: we clearly see a rotation of the entire constellation. On Figure 7.12, the constellation of all symbols in a same burst are superimposed: the rotation angle varies from symbol to symbol. Living with a real radio 133 We want to compute the BER degradation caused by the phase noise as Therefore we have to calculate the effect of the phase noise on the BER of 64QAM as 134 Chapter 7 In this expression the effect of a phase rotation on the BER performance of 64-QAM is [14]: with the symbol-to-noise power and Next we need to derive the distribution function of the phase rotations Living with a real radio 135 From literature and measurements [4] we know that the phase noise spectrum can be represented by a piece-wise linear function (Figure 7.13): a flat level close to the carrier frequency representing the PLL floor (-78.2 dBc/Hz in our case), a flat low level representing the system noise floor (-118.2 dBc/Hz), and a transition in between with a slope of -20dB/decade (according to Leesons model). This results in an integrated phase noise power of 32 dBc, integrated over the signal bandwidth. Let be the frequency domain representation of the phase noise then or with K the number of frequency points in the phase noise spectrum specification. The positive part of the phase noise spectrum is represented by the negative part by and the DC-component by We can write as with the amplitude specification of the phase noise (as it is shown in Figure 7.13) and is the phase of that frequency component. Substituting in (7.30) leads to in which From here on, we omit the index n from because the introduction of makes the dependency on n implicit. Since the real and imaginary parts of the phase noise are As stated in Section 7.2.2, we assume that there is no amplitude modulation due to the phase noise. Since then should be zero. To construct a real phase the negative part of its spectrum should be chosen to be the complex conjugate of the positive part. This means resulting in 136 Chapter 7 This is equivalent to Filling in (7.32) in (7.33) leads to The Central Limit Theorem tells us that the distribution of a sum of a large number of random variables is Gaussian, with a mean equal to the sum of the means and a variance equal to the sum of the variances. So, the real part of the phase noise has a Gaussian distribution. Its mean and variance are Note that we only need the integrated phase noise power to describe the phase noise distribution. This shows that the shape of the phase Living with a real radio 137 noise spectrum has no impact, only the integrated phase noise (as long as the FNC is negligible). The probability distribution of the phase rotation due to the phase noise is thus This derivation shows that we can assess the BER degradation directly from the phase noise specifications through equation (7.27). We verified the analytical expressions through simulations and they match perfectly. Expressed in terms of implementation loss, we find an IL of 3 dB at as shown in Figure 7.14. As we stated earlier, the phase rotation of an OFDM symbol is the same for all carriers in an OFDM symbol and can thus be estimated and corrected. We followed a decision-directed approach: we estimate the phase rotation as the average angle between the received symbols and the hard decisions of the symbols. We apply the estimated phase rotation to the OFDM symbol and redo the decision. This approach eliminates the phase rotation caused by 138 Chapter 7 the ONC. The only degradation that is left is caused by the Gaussian noise contribution of FNC, which in our case results in a BER shift of about 0.2 dB (Figure 7.14). It is clear from this discussion that compensating the Common Phase Error greatly diminishes the BER degradation by the phase noise. This permits us to allow more phase noise, up to the point where the introduced FNC becomes too large. In WLAN with 64 sub-carriers the ONC is the dominant effect. In other applications, such as digital broadcasting, the inter-carrier spacing is a lot smaller and thus is the FNC the dominating effect. 7.4.5 Limited linearity The use of a linear Power Amplifier (PA) is mandatory due to non- constant envelope signals. For simplicity, the PA here is assumed class A with back off. The output range limitation, due to hard clipping when the PA is driven into its saturation region, can be anticipated to cause a strong limitation on the BER performance. As a consequence, the saturation of the PA should be avoided by all means. Also because this would produce strong out-of-band radiation, require a long relaxation time to bring the PA again out of saturation and corrupt a large number of samples. A straightforward solution to prevent saturation of the PA is digital clipping at baseband of high signal amplitudes. This approach has already been presented as a necessary step to limit the word-length of the digital signal path and to limit the number of bits in the DAC. In addition it is also possible to add clipping on the magnitude of the complex signal to reduce even further the signal Crest factor. As sketched in Figure 7.15, this operation adds only a limited noise power, as the probability of clipping magnitudes above is low. This second clipping operation can be implemented at low cost if the transmitter architecture includes a digital IF (with digital I/Q modulation); if not, it requires a considerable amount of DSP, hence increase the system cost and power consumption. During the system design this must be traded- off against the PA implementation cost. In the following of this text, we assume magnitude clipping on such that the PA is never driven into saturation. Before considering a specific PA model for simulation, the impact of magnitude clipping on alone has been checked. This would correspond to a PA model with limited output power and ideally linearized (implemented for example with an ideal predistortion scheme). The simulated implementation losses at a BER of are reported in Table 7.3. We can conclude that magnitude clipping allows to reduce the clipping level compared to performing the operation on I and Q separately. Living with a real radio 139 PA amplitude transfer function model: The PA linearity is a key parameter as it reflects the distortion introduced on non-constant envelope signals, and is closely related to power consumption. It will be quantified in the following by the PA input-referred 3rd order intercept point IIP3 or 1dB- compression point for a given average RF input power The amplitude transfer function of the PA, assuming that it is kept out of saturation, equals: where is the limit of the saturation region, G the linear gain, and quantify the non-linearity of the device. The non-linearity coefficient can be expressed as a function of IIP3, or 140 Chapter 7 As a consequence, IIP3 and are related by the well-know equation Real-life systems do not present only cubic non-linearities. However in- band distortion is mainly introduced by cubic distortions. Hence the specifications derived for cubic non-linearities are still valid for PAs with more complex nonlinear behavior. Indeed, the distortion created by higher orders non-linearities generate mainly out-of-band components that can be filtered out. One remark is that in general the relation between IIP3 and becomes Therefore, it is best to characterize the non- linearity of the PA by its to make sure that the average signal power is far enough from the PA compression region. PA model for in-band distortion: the PA model based on the amplitude transfer function described above can not be used as such with a lowpass representation of the signal, as at baseband one can only compute in-band distortion. Instead, the model for baseband simulation is based on two functions, the AM-AM and AM-PM conversion functions, f and g respectively. If one applies to the non-linearity, this results in Combining with (7.38) results in: The PA model for baseband simulation, computing in-band distortion only, is therefore: Note: When f(t) and g(t) are functions of only, as in (7.26), the non- linearity is referred as envelope (band-pass) non-linearity [20]. This guaranties that the PA can be linearized with predistortion techniques. Figure 7.16 shows, for various values of the back-off between the 1dB- compression point and the average input power the amplitude transfer function curves of a PA described by (7.38). On the same plot are reported a constant input power (here ) and the corresponding peak power for a clipping factor From this figure, it appears that the PA will only remain out of saturation for a back-off greater than +8.4 dB; this corresponds to a rather high back-off, that may set stringent constraints on the PA design. To enforce the condition that the PA remains out of saturation even for a clipping level lower than the provided by baseband clipping on I and Q separately Living with a real radio 141 would be required. As proposed before, it is best to add clipping on the magnitude of keeping on I and Q the optimal clipping level of derived in Section 7.4.1. This technique was applied to all the simulations we present in this section: first baseband clipping at on I and Q separately, then additional magnitude clipping at for simulations where BER simulation results for PA non-linearity: the PA transfer function is modeled at baseband by (7.41). Figure 7.17 shows that for low implementation losses and without predistortion, the PA can be operated with only 5.4 dB back-off between and (15dB between and IIP3). This can be translated in terms of system power efficiency into 13%, assuming a class A PA with 50% intrinsic power efficiency. Moreover, it appears from uncoded simulations that, at low BER, the degradation is for a large extent introduced by the hard clipping, and not by the gain compression in itself. Indeed, with ideal predistortion applied as a linearization technique, the IL at is for 70% due to hard clipping (PA output power saturation), and for 30% due to soft clipping (gain compression of the PA). Thus, predistortion applied as a linearization will not help 142 Chapter 7 significantly with respect to the system power efficiency versus BER trade- off. However, predistortion might be considered to reduce out-of-band radiated powers (inter-modulation products and harmonics). 7.4.6 Clock jitter on the ADC An ADC processes a continuous time, continuous amplitude signal x(t) to produce a sampled, quantified representation of this signal x[n]. The sampling instant, ideally suffers from clock jitter, which can be modeled by a stochastic variable As a result, the actual sampling instant should be written as where is the value of at In a simulation model of an ADC it is impractical to add clock jitter directly, for this would require a large over-sampling rate. Therefore, the effect of clock jitter on the signal x[n] will be added first. Next, is fed through an ideal ADC. This way, it is still possible to perform baseband simulations at the nominal sampling rate As illustrated on Figure 7.18, the sampling error is computed as a linear interpolation of the input signal x[n] at the sampling instant where follows a Gaussian distribution with a zero mean and a variance This can be formalized as Living with a real radio 143 The error between the interpolated value y[n] and the actual value of seems quite large on Figure 7.18. However, because the simulation set-up includes up-sampling by 4 (Section 7.3.1) this error is significantly reduced. Simulation conducted for an ADC clock jitter variance of 1.4 ns on a clock frequency of 80 MHz (that is 11.2 % of the ADC sampling period) resulted in an implementation loss smaller than 0.5 dB at It shows that the link is quite robust with respect to clock jitter, as could be expected due to the high correlation between the signal sampled without clock jitter and the signal sampled with clock jitter. 7.4.7 Filter design in OFDM systems The last front-end induced degradation we want to discuss here relates to filters in the signal path, either analog or digital. The frequency domain performances of filters relate to cascade analysis and appropriate frequency planning. In-band ripple will result in subcarrier dependent SNR reduction. However, this can be combated for the transmitter filter chain by pre- 144 Chapter 7 compensation, that can be directly implemented on the OFDM symbols before the IFFT. At the receiver, inband ripple from the filter chain has no impact on the SNR, as signal and noise will be shaped in the same way (provided that the noise floor of the system is not reached). Limited stop- band performances will have similar effects as for any communication system: in-band SNR reduction, in this case subcarrier dependent as well, due to adjacent channels or interferers falling back in the desired signal band after mixing operations, nonlinear distortion, or aliasing in the ADC. More specific to OFDM is the performance degradation that results from the extension of the impulse response of the channel due to the insertion of the filters in the signal path. As demodulation in an OFDM receiver is based on frequency domain block processing, strong performance degradation will possibly occur because of the front-end filters' and multipath channels' impulse response. This performance degradation, due to intersymbol (ISI) and intercarrier (ICI) interference, is prevented by inserting a cyclic prefix (CP) with a minimal length equal to the significant part of the equivalent baseband channel impulse response [15]. The response of the channel in an indoor WLAN system is typically short (<500ns in HIPERLAN/2 channel models) but in combination with the transmit/receive filters, the length of the total impulse response might exceed the cyclic prefix length. This effect can be reduced to a certain extent by appropriate time synchronization at the receiver, as described in [16] [17] [18] where the optimal synchronization that minimizes ISI and ICI is explained. A simulation tool has been described in [19] that efficiently analyses the impact of a filter chain on the amount of ISI and ICI injected in the system. This tool, which does not require lengthy BER simulations, plots for a user- defined filter chain the Signal-to-Interference-Ratio (SIR) versus the synchronization location. Examples of such curves extracted with this tool are plotted on Figure 7.19 and Figure 7.20. A good design choice to design the filter chain in an OFDM link is to make sure that the amount of ISI and ICI injected is low enough on a synchronization range of a few samples (e.g. SIR greater than 35dB for at least 5 samples). Living with a real radio 145 146 Chapter 7 Living with a real radio 147 These curves prove to be even more interesting when considering the additional impact on the delay spread introduced by multipath channels: some architecture choices may be more robust with respect to ISI and ICI, as exemplified in Figure 7.21 and Figure 7.22. It is worth mentioning here that such curves do not guaranty low BER: high-enough SIR is only a necessary condition for correct reception, but it is still possible that one or two data carriers are in a dip of the multipath channel. 7.4.8 A front-end design example Specifying a front-end is highly application-specific: type of environment (indoor or outdoor, office-like or home-like), maximum distance for transmission, type of data to be transmitted and corresponding Quality Of Service (video streaming, data transfer, etc.), targeted product cost, etc. This issue, which is especially sensitive in a multipath environment, as mentioned in section 7.4.3, can not be addressed here, as it would require the definition of a complete business model. However, it would be the main input for decision on cost-performance trade-offs... Instead, the simulation results for all front-end non-idealities considered so far individually are now combined to as a set of specification for a super- heterodyne transceiver with a first IF at 880 MHz, a second digital IF at 20 MHz and digital I/Q (de-)modulation. This transceiver consists of a single- package RF module integrated in a BiCMOS process, and IF and baseband modules implemented as a discrete board design on PCB (see chapter 8). The specifications for the single-package RF module are: Transmitter: DAC word-length: 8 bits, with digital clipping at baseband at ; linearity: back-off between the cascade input and average input power ; phase noise on the LO port of the RF mixer: total integrated phase noise ; I/Q imbalance: not relevant as digital I/Q (de-)modulation. Receiver: ADC word-length: 10 bits, assuming that the 4 MSBs are used to code signal amplitudes above the signal RMS value (worst case for quantization noise) ; linearity: back-off between the cascade input P1dB and average input power (a 3dB margin has been included to allow for possible Crest factor re-growth in the channel) ; phase noise on the LO port of the RF mixer: total integrated phase noise ; I/Q imbalance: not relevant as digital I/Q (de-)modulation ; 148 Chapter 7 ADC clock jitter: of at 80 MHz. BER / PER simulations considering the complete transmit and receive front-end models: The BER performances of an OFDM link including this transceiver have been evaluated. Two cases have been simulated: Case 1: the transmitted and received signal powers are such that the back-off between the cascaded P1dB and the average signal power are minimal. This back-off is therefore of 5.4 dB for the transmitter, and of 8.4 dB for the receiver ; Case 2: the transmitter does not transmit at full power, and the received signal strength is far away from the maximum specified level of operation. As a result, the back-off between the cascaded P1dB and the average signal powers have increased, so that the nonlinear behavior of the transmit and receive chains is negligible. The BER performance limitation is then due to the phase noise on the RF LO signals. Figure 7.23 plots the simulation results for various modes of the IEEE 802.11a [1] and HIPERLAN/2 [2] standards: Living with a real radio 149 the implementation loss for the 54 Mbit/s mode at 10-5 BER is 2.2dB when the transmitter operates at full power (case 1), and 1.6dB when the transmitter operates at reduced power (case 2); the implementation loss for the 48 Mbit/s mode at a BER of 10-5 equals 1.1 dB ; the implementation loss for the 36 Mbit/s mode with a BER of 10-5 is 0.4dB. 7.5 CONCLUSIONS In this chapter, an overview of the interaction between the mixed-signal front-end and the OFDM digital modem was presented. This interaction must be well understood to analyze the impact of a mixed-signal front-end on the link bit error rate performance and to set at system level the front-end specifications optimizing the design margins and the associated risks. We have shown that basic assumptions on the degradation of the OFDM signal due to some of the front-end non-idealities lead to a fairly good approximation of the implementation loss. When possible, this approach helps in setting a first operation point for more detailed investigations However, analytical derivations of the link degradation are unfortunately not always possible and often require over-simplified models. In such cases, we proposed an approach based on an end-to-end system simulation tool that outputs the BER versus signal-to-channel noise power ratio, for user-defined front-end models. Such a tool is also helpful to verify the results obtained with the first method, and it allows to take into account more complex models of the front-end non-idealities, as well as multipath channel effects. It is a powerful tool to get insight in the IL degradation along the complete link due to front-end non-idealities, and to trade-off IL among the different front-end non-idealities on the one hand, and IL in the digital part versus IL in the front-end on the other hand. As a driving application, the main mixed signal contributions of a super- heterodyne transceiver have been considered to specify the overall link performances. However, it was pointed at that this example lacks the definition of a business model that would set additional constraints on the cost-performance trade-offs in a changing multipath environment REFERENCES [1] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications: High-speed Physical Layer in the 5 GHZ Band, IEEE std P802.11a-1999. 150 Chapter 7 ETSI TS 101 475 V1.2.1A, April 2000: Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Physical (PHY) layer D.J.G. Mestdagh, P.M.P. Spruyt, et. al., "Effect of amplitude clipping in DMT- ADSL transceivers", Electronic Letters, Vol. 29, No 15, pp. 1354-1355. Influence of RF oscillators on an OFDM signal - C. Muschallik - IEEE Transactions on Consumer Electronics, Vol. 41 No. 3 pp. 592-603, August 1995 Effect of carrier phase jitter on single-carrier and multi-carrier QAM systems - T. Pollet, I. Jeanclaude, H.Sari - International Conference on Communications, Vol. 2 pp. 1046 -1050, IEEE 1995 Analysis of the Effects of Phase-noise in Orthogonal Frequency Division Multiplex (OFDM), systems - P. Robertson, S. Kaiser - International Conference on Communications, Vol. 3 pp. 1652 -1657, IEEE 1995 Phase noise and sub-carrier spacing effects on the performance of an OFDM communication system - A. Garcia Armada, M. Calvo, IEEE Communications Letters, Vol. 2 Issue 1 pp. 11-13, January 1998 The effect of frequency errors in OFDM - J.H. Stott - Research and development department technical resources, report BBC RD 1995/15 (http://www.bbc.co.uk/rd/pubs/reports/), BBC 1995 B. Come, R. Ness, S.Donnay, L. Van der Perre, P. Wambacq, M. Engels, and I. Bolsens "Impact of front-end non-idealities on Bit Error Rate performances of WLAN-OFDM transceivers", IEEE Radio and Wireless Conference (RAWCON 2000), pp. 91-94, Sep 10-13, 2000, Denver, Colorado. B. Cme, et al., "Impact of front-end non-idealities on Bit Error Rate performances of WLAN-OFDM transceivers", Microwave Journal, Vol. 44, No. 2, pp. 126-140, February 2001. P. Vandenameele, L. Van der Perre, et. al. , "An SDMA Algorithm for High- Speed Wireless LAN", Globecom 98, pp. 189-194, Nov. 1998. Channel models for HIPERLAN/2 in different indoor scenarios, ETSI EP BRAN Meeting #3, Document 3ERI085B, March 30, 1998 B. Sklar, "Digital Communications, Fundamentals and applications", Prentice- Hall International, 1988 Digital Communication Techniques: Signal Design and Detection, Marvin K. Simon, Sami M. Hinedi, William C. Lindsey, Prentice Hall, 1995 R. Van Nee, R. Prasad, OFDM Wireless Multimedia Communications, Artech House Publishers, 2000. S. Muller et al., Analysis of frame- and frequency synchronizer for (bursty) OFDM, CTMC Globecom, pp.201-206, November 1998. T. Pollet, M. Peeters, Synchronization with DMT Modulation, IEEE Communications Magazine, pp. 80-86 vol. 37 Issue 4, April 1999. G. Malmgren, Impact of Carrier Frequency Offset, Doppler Spread and Time Synchronization errors in OFDM based Single Frequency Networks, GTC Globecom, pp. 729-733 vol.1, 1996. B. Debaillie, B. Come, W.Eberle, S. Donnay, M. Engels, I. Bolsens, "Impact of front-end filters on Bit Error Rate performances in WLAN-OFDM transceivers", IEEE Radio and Wireless Conference (RAWCON), pp. 193-196, Boston, Massachusetts, USA, August 2001. M.C. Jeruchim, P. Balaban, K.S. Shanmugan, Simulation of communication systems, Plenum Press, New York & London Section 2.11.5: Block Models for Memoryless Nonlinearities, pp. 144-152 [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] Chapter 8 Putting it all together From theory to a working solution Wolfgang Eberle IMEC 8.1 INTRODUCTION The previous chapters presented different challenges and solutions for OFDM-based systems in the context of wireless local area networks. To arrive at an actual implementation, we need to combine all this knowledge. A case study of such an implementation is described in this chapter. We will first define an application scenario for the system that will guide the detailed specifications. These specifications serve as input to the architecture selection phase. We opted for a discrete system set-up with off- the-shelf components around an application specific integrated circuit (ASIC) design for the baseband signal processing. Architecture and design of this baseband chip is discussed next. The evaluation of measurement results and design cost parameters leads to a discussion on future improvements. 8.1.1 Focus on an application scenario Imagine you want to build a wireless LAN transceiver right now. You have studied the previous chapters and you are familiar with the required ingredients. How would you start? The first question to ask is what requirements should be met. Three can be distinguished: the product should not disturb any other applications, the product should be compatible with products of other vendors, the client should buy the product. The first requirement is achieved by following frequency regulation. It limits the power at which transmission is allowed, puts restrictions on the antenna gain and defines a transmission spectrum mask. The second 152 Chapter 8 requirement is handled by means of standards (cfr. chapter 4). This is where most constraints come into play. Often regulation and standardisation are linked tightly together, e.g. by allocating certain frequency bands for specific standardised applications only. The last issue is part of the business plan. Where can we make a point in the market? To answer this question we need to come up with one or several application scenarios. We next try to tune the system for those scenarios within the constraints of standards and regulation. Remark that the definition of an application scenario can have a considerable impact on the system because it defines many performance criteria related to the quality of service (QoS). For instance, video streaming requires a different error handling than packet-based message transfers. In our case, we decided to design the wireless LAN prototype for a wireless webcam scenario, which was considered a sufficiently complex environment to study performance and system integration issues. Sufficiently complex means in this context that we will not only encounter problems at one level, e.g. the RF front-end design, but we will face all problems at all layers, be it protocol software, DSP hardware, or D/A and A/D conversion. Before we specify our requirements, let us first have a closer look at the webcam application (Figure 8 2). At the application level, standard available functionality is used, such as a webcam with a VideoForWindows software interface, or viewer software at the receiver side. A combination of a software stack and hardware (i.e. the AMCC board in the figure) is needed to bridge between the computer platform and the actual wireless transceiver. Remark that a computer platform is not a necessity: we could also use a thin Putting it all together 153 multimedia terminal. The wireless communication set-up is partitioned into the digital baseband functionality (digital board) and the front-end with the antenna. The digital baseband employs a configurable FPGA. Figure 8 2 contains already a first sketch of an architecture. Principally, we could have restricted the discussion to the data flow of the application. However, in practical design cases it is more common to take an architectural view to describe the application scenario. The final architecture is then derived by iterations between specification and architecture selection, going from general requirements to detailed specifications. 8.1.2 Design goals A wireless webcam can be employed both in a business or home environment. As a consequence, the following constraints come into mind: cost efficiency, easy maintainability, and compactness. These goals guide us to IC design for the integration capability, towards standard IC technologies for the cost, and into heterogeneous software/hardware solutions for easy maintenance and upgrading options. In the road towards integration, we first chose to focus on the digital baseband ASIC as a custom design and to design the remaining system based on off- the-shelf components. In the mean time, steps towards the integration of the RF front-end have been undertaken, based on the experience with the discrete set-up. 154 Chapter 8 For the initial specification, we decided to distinguish between hard and soft constraints. Whenever reasonable and already known (the development started before the standards where fixed), we tried to comply with the standard. Our own ideas on performance improvements are reflected in extended programmable parameter ranges or parameters instead of constants. Table 8.1 gives an overview of the most import parameters for the two ASICs that we realised, called Festival and Carnival. Table 8.1 does only contain the most important hard design parameters. Maximum clock rate, power consumption, and area were not constrained; instead, the goal was to minimise them. First investigations during the architecture selection revealed reasonable estimates such that constraints Putting it all together 155 could be defined and costs could be predicted. As a consequence, the initial specification was not complete and was developed during the prototype exploration. To reach an optimal solution for the physical layer and parts of the data link layer of a wireless transmission scheme, we must deal with key problems in three different fields (Figure 8.3): communications and signal processing, architecture design, and design methodology. Indeed, the previous chapters show that many different design factors influence the system performance. An efficient design methodology is needed to support the exploration of this vast search space. In addition, a smooth path from model towards implementation saves time and allows iterations in the exploration process. 8.2 THE BASEDBAND SIGNAL PROCESSING ASIC As explained in the previous section, we decided to focus on the digital baseband signal processing design first. Initial analysis [4] showed that an ASIC solution is required to reach the desired throughput and to keep the power consumption reasonably low at the same time. Consequently, we went for an ASIC design. So far, two ASICs were designed with different specifications and intentions. The first ASIC, Festival, was already specified in 1998 prior to the finalisation of the standardisation efforts in IEEE and ETSI. As a consequence, the specifications differed in quite some points from the later defined standards. For instance, it only supported BPSK and QPSK and it 156 Chapter 8 used a different preamble. However, the main target of the design was to prove that a scalable and still flexible OFDM ASIC design could be achieved [5], which by programming its parameters could come in the close vicinity of the standard specifications. The design of this ASIC was completed in the spring of 1999. A second version, Carnival, followed Festival in the spring of 2000 [8]. With the IEEE standard being finalised only in the fall of 1999, the design of the ASIC was already partially completed. We chose not to delay the design for full compliance but rather tried to include the major requirements, such as 64-QAM operation and full compliance with the OFDM symbol structure including the pilots. 8.2.1 The ASIC architectures Both ASICs implement the inner transmitter and receiver datapath (Figure 8.4) required for a high-speed, wireless OFDM system employing a half-duplex protocol suitable for standard-compliant time-division duplex operation. Hardware resources such as the Fast Fourier transform are shared between transmitter and receiver and various data reordering tasks are merged into a centralised unit (symbol reordering). A burst controller (BC) allows self-controlled processing of entire transmission and reception bursts, reducing the load of an external medium access control (MAC) or general-purpose processor. The transceiver only requires initial programming of parameters and triggering of MAC requests for transmission and reception and delivers status information through a dedicated BC interface. In transmission mode, payload data enters the ASIC on request, through a 6-bit parallel interface. The data enters the symbol mapper unit where bits are mapped onto either BPSK, QPSK, 16-QAM, or 64-QAM subcarrier constellations. A programmable number of zero carriers is introduced near DC and/or the Nyquist frequency. The zero carriers around DC allow to accommodate DC notch filtering, while the zero carriers around the Nyquist frequency reduces the requirements on the low pass filters. A BPSK pilot sequence is inserted either on a fixed subset of 4 carriers (according to the standards) or using a rotating pilot pattern with a period of 13 OFDM symbols. A complex value weights each subcarrier, allowing for transmitter pre-emphasis and phase pre-distortion. The mapper provides a sequential series of 64 carriers, for Festival also 128 or 256 carriers, to the IFFT, termed an OFDM symbol. The mapper also adds an entire programmable BPSK OFDM symbol serving as a reference sequence prior to the payload or inserts it periodically into the stream of OFDM symbols. The inverse FFT transforms the frequency domain constellation into a time-domain sequence. Scaling and digital hard clipping is performed at the Putting it all together 157 IFFT output to select a suitable peak-to-average power ratio (PAPR) and signal-to-quantization noise ratio. OFDM symbols are then passed to the symbol reordering unit (SSR) that inserts the acquisition preamble and the cyclic prefix. The SSR sends data sampled at the chip clock frequency through a 2x8-bit parallel I/Q interface to the external digital IF stage. Setting the ASIC clock frequency to 20 MHz results in a standard-compliant stream of OFDM symbols. In reception mode, data is provided from an external digital low-IF down-conversion stage in 2x10-bit format to the gain control and timing synchronisation stage. The preamble serves to estimate received signal strength, frame start, and carrier frequency offset (CFO). Before entering the FFT, the CFO on incoming samples is reduced to about +/- 4 kHz resulting in negligible inter-carrier leakage effects. Also, the guard interval is stripped off forming again plain OFDM symbols of 64, 128, or 256 subcarriers. The FFT translates them into the frequency domain where the SSR removes zero carriers, identifies pilot carriers and reference symbols. Payload-carrying sub-carriers are passed to the equaliser along with this extracted information. The equaliser performs an initial channel estimate, based on the BPSK reference symbol. In the Carnival ASIC, this estimate is improved by interpolation. With this estimate being available, the acquisition phase has finished and the data reception and tracking phase starts. During the tracking phase, received data is still being compensated by the time- domain CFO. The FFT timing is controlled and updated by a clock offset 158 Chapter 8 estimation and compensation algorithm. Fine frequency offset compensation is performed in the equaliser in a decision-directed averaging phase loop, updating the channel. Also, time variations of the channel are traced by means of the pilot scheme, where rotating pilots outperform fixed pilots. The equaliser divides the received constellation by its channel response per sub-carrier and provides, through the demapper, either hard decision, 2x3-bit soft decision, or 2x6-bit high resolution output to e.g. an external decoder or interleaver block. Table 8.1 describes the major programming parameters for the two ASICs. An OFDM symbol structure compliant with IEEE and ETSI standards can be achieved by choosing 64 carriers, 16 guard samples, 0 zero carriers near DC, 5 zero carriers near Nyquist, fixed pilot scheme, and a frequency diversity factor of 1. 8.2.2 Joint Algorithm and Architecture Design In this section, we will focus on the algorithms and architectures of the major signal processing parts of the OFDM transceiver. We start with the Fast Fourier Transform, move on to the centralised symbol reordering unit, address time-domain based burst acquisition and finally come to equalisation and tracking in the receiver. 8.2.2.1 Fast Fourier Transform The complex Fast Fourier Transform (FFT) is the heart of an OFDM system, converting frequency domain constellations to time domain and vice versa. The high peak-to-average power ratio (PAPR) of multi-carrier signals requires careful fixed-point exploration to maximise the performance/cost ratio. Wireless burst operation requires an FFT with low latency and power consumption. A pipelined complex FFT architecture (Figure 8.5) based on radix 2-2 decomposition [3] has been chosen since it achieves both the simplicity of butterflies from a radix-2 scheme and the low number of complex multipliers from a radix-4 scheme. Every other multiplier is replaced by rotator logic involving only multiplexing and sign inversion. Using simple butterflies and less multipliers also simplifies control and allows a straightforward design of a variable-length (i.e. 64, 128 or 256 subcarriers) FFT. The IFFT operation is obtained by conjugation of input and output signals. Putting it all together 159 The radix 2-2 scheme requires the minimum amount of memory locations. Memory is implemented as feedback register banks or dual-port RAMs (128- and 256-word banks only) distributed along the pipeline, starting with the maximum word-count according to a decimation-in- frequency scheme. We benefit from the fact that the wordlength through the FFT increases towards the output starting with a small input wordlength, saving 25 % memory in 64-carrier mode compared to decimation in time. Compared to a fixed-wordlength implementation, we achieve a reduction of 30 % in memory size from the fact that we start with 10 bits and end with 15 bits. We introduce a fixed scaling by 2 at every butterfly stage, so wordlength increases only at every full multiplier. To derive the 2 unknowns per multiplier, i.e. the post-multiplier datapath wordlength and the coefficient look-up-table (LUT) wordlength, we performed a parametric exhaustive search by simulation [17]. This search becomes feasible since we have reduced the unknown wordlengths to only 4 in the 64/128-carrier and 6 in the 256-carrier case for the entire FFT. Scaling and saturation at the output stage facilitate the implementation of digital hard Cartesian amplitude clipping in the transmitter. In previous chapter we have seen that the optimum performance requires a trade-off between clipping noise and quantization noise. The choice between 5-bit and 8-bit outputs offers dynamic ranges from 30 dB to 48 dB, while the choice of the clipping level results in different distortion and quantisation noise levels. 160 Chapter 8 Parameters are programmed according to the front-end, especially the power amplifier non-linearity, and the QoS requirements. There is a latency of one OFDM symbol between the input and the output. In addition, the final FFT implementation has a core delay of 10 clock cycles resulting from 1 pipeline stage per butterfly and 2 per complex multiplier. The FFT provides its output in bit-reversed order with post- compensation of that effect in the SSR. 8.2.2.2 Centralised symbol reordering unit OFDM symbols are meta-symbols compared to samples in conventional single-carrier systems. This inherent scalability makes OFDM powerful. However, to exploit this flexibility, architectures that support a discrete set of parameter choices are required. In a conventional distributed design process, the design would be first partitioned into modules and then optimised locally per module. This would result in a large number of distributed buffers that would be individually sized for worst case. Putting it all together 161 Therefore, we investigated if a number of sample-reordering tasks could be efficiently implemented with a central dual-port memory. To this end, we analysed, based on a high-level dataflow description, the data transfers between signal processing units, their intra-unit storage and the inter-unit buffering requirements to handle the multi-rate issues. The flexibility in the OFDM symbol leads to a large set of different I/O rates, requiring the insertion of buffers. More specifically, we encountered buffering issues due to bit-reversed reordering of the FFT output, removal of pilots and zero carriers, despreading, insertion of the programmable length cyclic prefix, and the preamble. We centralised the storage in a single unit (Figure 8.6) consisting of two single-port RAMs with memory arbiters and a set of address generators. Two address generators run in parallel, producing read and write addresses respectively. The RAM access mode is toggled after every OFDM symbol. This approach results in the minimum amount of memory, i.e. twice the number of sub-carriers, without additional latency. 8.2.2.3 Time-domain burst acquisition Wireless LAN systems are packet-based and minimise the preamble to increase the efficiency of the physical layer. They can not tolerate loss of initial data and are hence critically dependent on fast and accurate burst acquisition. At the same time, the received signal is distorted by a number of indoor channel and front-end effects. Receiver acquisition has to detect the incoming signal, adapt its signal power, achieve timing synchronisation, and compensate for the carrier frequency offset (CFO) introduced by local oscillator mismatches in transmit and receive front-ends. Fast acquisition prohibits the use of frequency domain signal processing for timing synchronisation and CFO estimation, popular in wire-bound systems with long acquisition preambles, like ADSL, or wireless broadcasting systems, like DAB and DVB. We have implemented a timing acquisition (Figure 8.7 a) based on a two-phase auto-correlation process (Figure 8.7 b) using a programmable BPSK time domain code sequence which is repeated according to a second meta-level sequence. ETSI and IEEE standardised a different preamble scheme, but the intrinsic signal processing concept can be maintained. Since the sliding window correlator only requires a 2 x 1-bit input, it is very robust against automatic gain control transients and results in a low cost, low power implementation. A parallel sliding window signal power estimation is used to validate the correlator results. Alternating bipolar correlation peaks during phase 1 determine the relative code sequence start, while the transition to phase 2 defines the absolute frame reference. The receiver only uses information on the codeword length and the meta-level sequence; the codeword itself is not known. Probabilities of false alarm and 162 Chapter 8 missing detection depend on the programmed numbers of peaks-to-detect in phase 1 and phase 2, respectively. Phase 3 starts when phase 2 has obtained enough confirmations and counts until the frame start. Carrier offset is estimated by means of a repeated sequence of length 64, 128, 256, or 512, which follows the frame start. For multipath immunity reasons the estimator is based on auto-correlation. A larger preamble size trades off higher noise suppression against a lower capture range. Carrier offset must be reduced to a fraction, usually 1-2%, of the sub-carrier spacing, to achieve negligible inter-carrier interference in the FFT. A single-operator sequential CORDIC converts the Cartesian estimate into a phase difference. A phase accumulator reproduces the evolution of the phase caused by the carrier offset. This phase is translated in Cartesian co-ordinates by means of a pipelined CORDIC stage (Figure 8.8). The CORDIC uses a constant input reference 1+j0 to provide a Cartesian output with a conversion accuracy independent of the highly amplitude-varying receive signal. Putting it all together 163 8.2.2.4 Frequency-domain channel estimation and tracking The received signal after the FFT is still affected by multipath fading and contains a remaining low carrier frequency offset. However, by a proper choice of the sub-carrier spacing relative to the coherence bandwidth, the FFT produces a highly oversampled channel response. The equaliser can exploit this in two ways. First, it requires only a single complex channel coefficient per sub-carrier, as the channel response is independent for the different sub-carriers. Secondly, the rank of the matrix is reduced, since high oversampling translates into correlated channel coefficients. Thus, we can apply filtering to suppress noise and interpolate a smoothed channel vector from a smaller set of coefficients. This has been implemented in the Carnival ASIC, because the initial reference symbol- based estimate was poor for the 16-QAM and 64-QAM cases. The Festival 164 Chapter 8 ASIC contains a simpler equaliser whose architecture we describe before the more complex one of the Carnival ASIC. The Festival equaliser (Figure 8.9) implements the basic one-tap frequency domain equalisation, consisting of a single complex multiplier with a coefficient memory to store the channel response [2]. The channel is estimated by multiplying received initial or periodic reference symbols with a known reference. A decision-directed loop estimates either individual sub- carrier phase error or average phase error based on QPSK slicing. The channel estimate is thus updated for phase only, tracking such effects as fine carrier frequency offset or, to a limited amount, clock offset. Gain control on I and Q parts, using a greatest common divider (GCD) algorithm [4], stabilises the loop and prevents amplitude drift. The Carnival equaliser (Figure 8.10) also uses the concept of a single complex operator with coefficient memory. 16-QAM and 64-QAM constellation schemes however require accurate amplitude correction, which is performed by a complex divider. In addition to initial and periodic reference symbols, to update part of the channel, a pilot pattern is sent with every symbol. The channel estimate obtained from a single reference symbol still contains a considerable MMSE error (Figure 8.11). A channel interpolator (Figure 8.12), consisting of an initial noisy stage with the CFO phase error update, is followed by a cascade of 4 blocks implementing the matrix operation: Putting it all together 165 S is a 64x9 programmable complex coefficient matrix. The first two stages transform the noisy channel estimate into an impulse response vector of length 9, effectively suppressing any noise beyond the tap. The last two stages reconstruct the full 64-tap frequency response from this truncated impulse response vector [2]. The first three stages employ full parallelism such that an interpolated channel tap is available after one OFDM symbol latency. 166 Chapter 8 Coefficient sets are stored in 9 RAMs next to a pre-programmed set in a look-up-table (LUT). The interpolator is also used during tracking, improving the channel estimate by 2.5 to 3 dB. Together with the rotating pilot scheme it is also able to suppress spurs, e.g. from the equaliser feedback loop, reducing error propagation. Clock offset between receiver and transmitter sampling clock does not only have an impact on the sub-carrier phase. Indeed over the burst length of 2 ms for HIPERLAN/2 or up to 5 ms for IEEE 802.11a, it can shift the actual OFDM symbol out of the FFT frame leading to a low signal-to- interference ratio. Maximum clock offset values according to the standards are as high as +/- 40 ppm of the 20 MHz system reference oscillator. The clock drift is estimated by correlating the cyclic prefix with its original counterpart in the same OFDM symbol (Figure 8.13 a). The correlation peaks are determined and averaged over more than 32 OFDM symbols to reduce the impact of noise. Compensation occurs by either dropping an entire sample from or adding one to the cyclic prefix (Figure 8.13 b), resembling a sigma-delta architecture. The shifting events are communicated to the equaliser to adapt the stored sub-carrier phases to the instantaneous sample shift. Putting it all together 167 8.2.3 Integration The previous section proposed a set of signal processing algorithms and architectures to solve individual problems. Next, we need to integrate them into a complete system. Ease of integration is essential to guarantee that the system can be designed with reasonable effort assuming limited support from electronic design automation (EDA) tools. Easy of integration means that we need to pay special attention to reducing the data transfer and storage costs between design units [1], to simplifying the system control and to ease the clocking strategy. Also the use of an EDA framework for system integration can be beneficial for the system design. 8.2.3.1 Communication and storage Wireless LAN transceivers require both high throughput and low latency, leaving limited space for sequential processing. The FFT processes about 1 Gops/s while the channel interpolating part of the equalizer needs 1.7 Gops/s and a memory bandwidth of 3.4 Gbps. Higher clock speeds could reduce the needed parallelism, but introduce more data caches to adapt between the different rates. However, this multi rate problem can be solved by either sharing a common memory or by a distributed memory approach depending on the local processing needs. Remark that the multi-rate problem is present anyway because of the flexibility that was introduced in the ASICs. For the FFT, a distributed memory architecture was found to be superior to a single memory running at higher clock speed with respect to its data transfer-related power consumption. On the other hand, a number of sample- reordering tasks were efficiently implemented with a central dual-port memory of minimum length in the SSR. Both solutions efficiently use the memory transfer bandwidth while maintaining a regular access pattern. The final on-chip datapath does not contain any caching beyond the minimum required by the signal format defined in the standards. This caching latency is two OFDM symbols for both receive and transmit path evenly divided on FFT processing and bit-reverse reordering. All design units contain their own local register banks to store programming parameters. This supports the IP (intellectual property) block concept, which is important for later reuse. It also significantly cuts interconnect dependencies between design units, a benefit for the layout stage. Multiple instantiations in case of common parameters have negligible cost. A single write address for the same parameter in all units and individual read addresses guarantee correct programming and easy verification. 168 Chapter 8 8.2.3.2 Token-based distributed control To stress the IP concept, a generic communication protocol is required between all design units. We implemented a scheme based on token semantics that follows the natural data flow through the transmit and the receive path (Figure 8.14). A closed token-loop scheme is used between the burst controller and the datapath. Tokens contain three types of information: meta-symbol start, burst state information (BSI), and dynamic datapath information (DDI). Tokens are not sent at the sampling rate, but at the rate of meta-symbols, i.e. at OFDM symbol rate. The BSI indicates reference symbols and the last symbol of a burst and is returned by the last unit in the datapath to indicate that an entire burst has been completely processed. DDI can be added to a token by any datapath block to transfer data-dependent information synchronously with the current symbol to another unit down the processing chain. For example, the clock offset estimation unit uses this to inform the equaliser in case of a FFT frame timing shift. The token scheme scales with multi-rate and simplifies also the design task, since a token arrival time window is defined instead of a discrete point in time, keeping detailed unit latency information locally. 8.2.3.3 Clocking strategy Low power operation is crucial for portable operation. The clock tree dominates the power consumption in synchronous systems. However, analysis of a typical receive scenario reveals that the receiver remains a considerable amount of time in listening mode searching for a receive signal. Reducing power dissipation on the average compared to the peak power consumption has been achieved by matching activation of units with the time windows they are effectively required from the networking protocol and Putting it all together 169 burst format point of view. Clock gating with a state-based activation was used to implement this behaviour. The burst controller and decentralised smart senders (Figure 8.15) control the clock generation. We also use clock gating to implement multi-rate interfaces between units. Transitions between units operating on different clocks are facilitated through retiming on a common inverted core clock, reducing the potential skew complexity from to O(n), with n being the number of clock signals. The ASICs are master for all datapath interfaces and provide on-chip generated clock signals. These clocks are generated locally to the other interface I/O signals to allow joint skew optimisation. 8.2.3.4 Object-oriented design methodology The demands of high-speed modems in terms of latency, area, and power dissipation are approaching the technological limits. Therefore, a joint optimisation of algorithm and architecture is required to come to a feasible solution. Many of the blocks in the OFDM architecture are flexible. To investigate reasonable parameter ranges and their interdependencies a fast high-level simulation model is required. On the other hand, the architecture must eventually be implemented in silicon, and, consequently, a smooth path from the high-level simulation model towards a synthesizable register transfer (RT) description is vital. In this process, architectural exploration must be maximally supported. We started with a high-level dataflow model in C++, using the OCAPI [15] hardware libraries. The OCAPI technology supports the gradual refinement of an object-oriented C++ model starting from behavioural code. Its application to the design of the OFDM transceiver consists of several phases: 1. 2. 3. 4. behavioural description of the algorithm using a set of class libraries to express data-flow semantics; design partitioning in which functionality is being grouped in larger entities to be mapped onto single hardware units; scheduling of the operations inside each entity to get a clock cycle-true description and formal mapping to finite-state machines (FSMs) and signal flow graphs (SFGs) resulting in a register-transfer description; automatic generation of synthesizable RT-level HDL code. Performance evaluation, algorithm selection, fixed-point refinement, and functional partitioning were performed on the C++ dataflow model. Object- oriented design gives the designer freedom to design generic classes that construct hardware from given user constraints. Inheritance and fully parametrizable, hierarchical instantiation are strong assets for a clean code database. The transceiver, for example, is instantiated twice and configured either as transmitter or receiver just at the top level. Internally, 170 Chapter 8 interconnection and scheduling are optimised for simulation speed or for hardware operation matching. Also, on-the-fly reconfiguration is possible during simulation. The C++ dataflow model was refined towards a C++ description based on integrated finite-state-machines and datapath (FSMD) blocks. It is important to start exploration of data transfer and storage issues already at the dataflow level [18], since this prevents frequent and time-consuming loop back between the FSMD and the dataflow design. Refinement includes mainly operator sharing and scheduling. VHDL RT-level code is generated automatically from the C++ FSMD description. An option for Verilog is available too. Both Festival and Carnival make use of existing native VHDL code. These units, developped prior to the C++ design, were modelled as abstract dataflow blocks to obtain a complete dataflow end to end link. Carnival also used native Verilog code, showing that a C++ entry-level approach can be integrated into a heterogeneous design flow. From RT on, a conventional standard cell design flow is followed with logic synthesis, floorplanning, and layout steps. Clock tree routing was performed at layout level and included into the back-annotation. During the design process, simulation-based verification is used extensively to track correctness of the design refinements. C++ based simulation is used during the system-level design phases, HDL-based simulation is used during the synthesis and back-end flow. Generated HDL, gate-level and back-annotated gate-level netlists were all verified against the same test vectors generated from the C++ dataflow model. Extraction of Putting it all together 171 simulation results from RT and gate-level simulation only requires synchronization of control token flow and dataflow at the top level of the design to match the different abstraction level. This was the only HDL code modification required to execute all testbenches. 8.3 THE DISCRETE SYSTEM SET-UP The first version of the demonstration set-up was implemented on a set of printed circuit boards (PCBs) (Figure 8.16). A PC platform was equipped with a digital PCI-based multi-function data acquisition card (AMCC S5933 board). Cables for data transfer and protocol connect to two custom designed boards: the baseband signal processing board and from there to the digital IF board. The baseband signal processing boards hosts the custom designed OFDM ASICs. Commercially available analog-to-digital conversion and digital-to-analog conversion evaluation boards were used to bridge to the radio front-end. The radio front-end itself was implemented in a stack of three boards: a synthesizer board, the IF board, and the RF board. The power amplifier was implemented on a separate board including the cooling facilities. 172 Chapter 8 8.3.1 Software and Protocols For the webcam application, we have chosen a PC as platform. In combination with different operating systems (MS Windows and Linux), this guarantees a fast development for the driver software and enables us to use off-the-shelf application software. An AMCC multi-function board connects the PCs PCI bus to the digital baseband board. The kernel drivers use direct memory access (DMA) to access the digital board resources through the AMCC board in a transparent way. The software protocol stack was implemented on two platforms (Figure 8.17): Windows 95/98 and Linux. The Windows software protocol stack could be developed faster due to available drivers, but it was also limited in its flexibility. The Windows version including a graphical user interface (GUI) was primarily used in the beginning phase when testing the interaction between the application and the FPGA-based hardware protocol stack. To eliminate the limitations, a Linux-based protocol stack was implemented: first a simple testbench with automated self-tests and then the actual webcam demo. The protocol code is subdivided between the FPGA and the PC-based kernel driver. The hardware MAC architecture on the FPGA (Figure 8.18) is centred around a CPU with customised instructions. The remaining FPGA resources are mainly used for different interfaces: short First-In-First-Out (FIFO) buffers for data traffic buffering between the AMCC card and the FPGA, a microprocessor programming interface (MPI) for programming of the ASIC and the convolutional encoder/decoder; a serial interface to control front-end resources (e.g. automatic gain control, power-up/down, carrier frequency selection) Putting it all together 173 The hardware MAC is responsible for the burst timing itself. Data framing or the start-up sequence including programming of the ASIC is accomplished by the hardware MAC without software MAC support. Instructions from the PC software MAC to the FPGA hardware MAC occur on basis of acknowledged messages and are available at different abstraction levels: a) Burst level Start communication: This command tells the FPGA to switch the system in transmit or receive mode. Whenever data arrives, the system starts processing it. b) Resource level Read from ASIC: This command tells the FPGA to read at a certain address in the baseband ASIC. Write to ASIC: This command tells the FPGA to write a certain value at a certain address in the baseband ASIC. Start timer: This command starts a timer in the FPGA. Serial communication: This command lets the FPGA communicate with the IF modulator/demodulator board. Timing-critical sequences are pre-programmed. They are called when a burst level message arrives and use the facilities of the lower resource level. The software MAC can be informed of time-out situations through watchdog functionality based on the FPGAs timers. 174 Chapter 8 From a networking point of view, the webcam application requires only a point-to-point protocol. Apart from the physical layer implementation, a limited set of data link control (DLC) functions was implemented, such as an automatic packet numbering. Provisions were taken to hook up the custom kernel functions to TCP/IP. The data to transmit is prepared in a buffer in the PC memory by the driver. The driver then writes commands (e.g. program festival, enable front- end, start loading FIFOs) to the OFDM baseband board. The FPGA interprets the commands and then reads the data from the PC memory and writes it to the TX-FIFO on the baseband board. The Festival ASIC is switched to TX-mode, reads the data from the TX-FIFO and sends out its digital I/Q signals to the IF modulator/demodulator board. The PC software driver allocates a buffer, where the received data will be put, in memory. The driver then sets up the OFDM Baseband board for RX mode. When the Festival ASIC receives its data, it writes the data to the RX FIFO on the OFDM Baseband board. The FPGA detects that there is data in the RX FIFO and writes it to the buffer in PC memory. 8.3.2 Baseband signal processing The baseband signal processing part is implemented on a separate board (Figure 8.19). Around the central OFDM signal processing ASIC, a number of support processors are grouped (Figure 8.20): a convolutional encoder/decoder chip; the FPGA handling the data transfer between ASIC, FIFO, and PC and performing the control of the baseband board (see section 8.3.1); a transmit and a receive FIFO for data buffering between the FPGA and the ASIC. Once activated for transmit or receive mode, the ASIC communicates through a FIFO-based transmit and receive interface as a master with the PC in a slave position. In transmit mode, the ASIC requests data for transmission when ready. In receive mode, it passes the received data to the PC. To implement the tasks of the outer transmitter (receiver), the FPGA can redirect data through the convolutional encoder (decoder) or it can scramble (descramble) the date. The configuration code for the FPGA can be either loaded from an on-board EEPROM at start-up or programmed through a serial interface from the same PC. Towards the front-end, the ASIC provides digital I/Q interfaces. Either a pair of ADCs/DACs or a digital IF stage can be connected. Additional signals are provided to support analogue automatic gain control in the receiver and to control power-up and down in the front-end. Putting it all together 175 176 Chapter 8 The chip features an asynchronous microprocessor interface for programming. An additional 5-pin direct control interface allows the MAC to select one out of four operational modes (transmit, receive, programming, and sleep) and watch the status of those modes. Any intra-unit data bus can be monitored parallel and at full clock speed through an external test interface. For example, this bus can provide an adaptive loading extension or a decoder with the channel estimates. 8.3.3 Establishing the wireless connection The radio front-end is required to translate the baseband signals to and from the carrier frequency in the bands allocated for the WLAN application in the 5 GHz. The OFDM ASIC provides an I/Q interface to the radio front- end. Samples are provided and expected at the minimum, i.e. 20 MHz, sampling rate. In chapter 7, we have already discussed the specification for the design of the front-end. Two front-ends with different implementation details are briefly described. The first is a discrete superheterodyne front-end with an additional digital intermediate frequency stage for the final conversion to and from baseband. The second one is also based on the superheterodyne principle but is far more integrated using chip-package co-design [19]. It converts to and from baseband by subsampling. 8.3.3.1 Discrete superheterodyne front-end The discrete superheterodyne front-end (Figure 8.16) consists of three boards: a synthesizer board providing the fixed 430 MHz intermediate frequency and the tunable 5 GHz local oscillator signals for the mixers to the Putting it all together 177 IF board and the RF board respectively. The analog baseband signal is centered around 20 MHz and sampled at 80 MHz by a single DAC/ADC. A digital intermediate frequency conversion stage, implemented in a separate FPGA (Figure 8.21), translates this IF to baseband including the low- pass filtering. A digital IF reduces the imbalance between I and Q paths compared to an analog quadrature mixing stage. The power amplifier was placed on a separate board. 8.3.3.2 Towards an integrated front-end A first step towards integration was the packaging of the RF section and the IF band selection filter into an integrated RF module implemented in a mix of active BiCMOS devices on an Multi Chip Module (MCM-D) substrate (Figure 8.22). The use of MCM-D allows different trade-offs for the quality of passive components, especially for inductances, which led to a different frequency planning: the IF was chosen at 900 MHz, the IF at 140 MHz. ADC and DAC still sample at 80 MHz, working now in a subsampling mode. The baseband signal appears now at 60 MHz. An optional external RF local oscillator was foreseen. The IF section is not yet integrated. For heat dissipation reasons, the power amplifier remains 178 Chapter 8 outside the package. All controllable elements in the front-end, TX/RX switch, variable gain amplifier (VGA) gains, and LO frequencies are controlled by the FPGA on the baseband board (see section 8.3.1). 8.4 LEARNING FROM RESULTS Now that we are familiar with the system architecture and the dedicated baseband ASIC that both have been implemented, we can move on to evaluate this design effort. Our focus will be on the custom baseband design where we have most insight. We will have a look at its performance and the design process. The fact that two ASICs were developed gives an interesting opportunity to evaluate reuse questions, too. Of course, we have to ask what is missing to make this system fully compliant with the standard and discuss the relevance and the effort for these modifications. Having solved the key problem in OFDM baseband signal processing for wireless LANs, other issues become more apparent. We will see that the sensitivity of the remaining problems to the architecture choice is even higher and almost always require a joint design of front-end and baseband signal processing. Automatic gain control and power-efficient transmission will serve as examples. 8.4.1 Measurement results and performance comparison Our demonstrator goal was the implementation of a complete wireless webcam. This goal was achieved by a heterogenous approach combining the development of software from the application to the machine/kernel level, FPGA design, ASIC design, standard PCB design, and RF design. 8.4.1.1 Application Tests Both ASICs were tested in the experimental set-up. Tests started with wired connections at the digital level, followed by analogue baseband, IF, and finally RF connections. Then we moved on to wireless tests, employing commercial dipole antennas at the 430 MHz IF and finally the 5 GHz range. Two identical platforms were built for this purpose (except for the power amplifier). At all levels, full application tests with a webcam image transmission, video transmission, and file transfers were successfully run between two of these platforms over the air. Putting it all together 179 8.4.1.2 ASIC performance Both ASICs have been implemented in digital CMOS technologies: Festival in a 5LM Alcatel Microelectronics and Carnival in a 6LM National Semiconductor process (Figure 8.23). Both designs were pad-limited with 144 and 160 pads respectively. The nominal clock rate is specified up to 50 MHz for Festival and up to 20 MHz for Carnival. Both ICs use embedded SRAM for datapath and parameter storage, being 9 units in Festival and 19 in Carnival. A fair comparison at the same data rate and overhead between Festival and Carnival (Table 8.2) shows the superior spectral efficiency and energy efficiency of the latter at the cost of a moderate area increase of 30 %. The highly programmable equaliser occupies 63 % of the area in the 64-QAM chip compared to 10 % for the FFT. By fixing the coefficient set for the interpolation, the area of the equaliser in the Carnival can be reduced to below 50 % of the total area. Power consumption has been measured separately for 1.8V core and 3.3V I/O supply for the Carnival ASIC in typical transmit, receive, and programming scenarios. During transmission, 156 mW I/O and 43 mW core power consumption were observed. During reception, the much higher core activity dominates with 146 mW compared to a lower 66 mW I/O consumption due to less I/O switching. In programming mode, logic switching is zero but all clocks are enabled, leading to 35 mW I/O and 81 mW core consumption. 180 Chapter 8 8.4.1.3 Design experience Design of transceivers involves more and more system design trade-offs. The design of a specific component such as the digital baseband processor, cannot be seen in isolation of the system. Indeed, the modelling of the entire transceiver chain, including the front-end and parts of the higher protocol layers, becomes more and more important to optimise throughput and to minimise implementation loss. We have also encountered the commonly faced code explosion (Table 8.3) during model refinement, but we have not suffered from it during the refinement process. The scalability and the code generation capabilities of our C++ design approach has saved us from costly iterations involving code rewriting. Also, the use of object-oriented techniques helped us to reduce the amount of code required for the C++ system testbench. The majority of the 11,000 lines of system testbench code were reused for testing the transmitter with the receiver by applying different abstraction levels to both of them, i.e. an ideal transmitter with a fixed-point receiver to evaluate the receiver's implementation loss. The fact that the Carnival design was based on the Festival design, allows us to conclude on design reuse issues as well (Table 8.4). Except for design Putting it all together 181 units that required an entirely new algorithmic approach, e.g. the equaliser, reuse was fairly high. The reasons for this were: use of unified token flow for control purposes; consequent parameterised design; partitioning based on data transfer and storage costs, leading to less design dependencies and simpler constraints. The choice of a scalable multi-processor architecture with distributed control using token semantics allows to maintain a high degree of flexibility and programmability. A high code reuse percentage in the Carnival design proved the scalability. The object-oriented, FSMD-centric design approach using C++ has shown its strength at higher abstraction levels for system exploration and at FSMD level for HDL generation in a heterogeneous mixed-language flow. 8.4.2 Towards or beyond full standard compliance? The realisation of two digital baseband signal processing ASICs, achieving bit rates up the 54 Mb/s with moderate technology constraints and area costs, show the viability of cost-efficient deployment of broadband wireless indoor systems both for the consumer market and business applications. Spectrally efficient 64-QAM constellation puts high requirements on transceiver performance. We have shown that novel digital signal processing techniques such as an interpolating equaliser, rotating pilots, and guard-interval based clock offset estimation can cope with the multipath channel and analogue front-end 182 Chapter 8 impairments. On the other hand, we have not developped a fully HIPERLAN/2 or IEEE 802.11a compliant physical layer implementation. 8.4.2.1 Beyond the standard Instead, we came up with a number of add-on's that can improve the performance of a wireless LAN system: a simple frequency diversion scheme using spreading in the frequency domain was implemented; this scheme lends itself also to measuring the channel correlation; a facility in the symbol mapper to adapt amplitude and phase per carrier at the transmitter; this can be used for predistortion and accurate gain calibration in an adaptive loading scheme [11]; flexible number of carriers; additional carriers can be nulled out or normally unused outer carriers could be used to reduce the peak-to- average power ratio; as an option, entire OFDM symbols can be inserted as reference symbols into the burst instead of pilots. The insertion frequency can be chosen and e.g. result in the same overhead as a pilot-based approach; a rotating pilot scheme was added to improve the performance of the channel estimation during tracking compared to fixed pilots with the same overhead and negligible design cost. Similar schemes showed their benefits already for terrestrial OFDM-based DVB-T. 8.4.2.2 Towards the standard All these add-ons are implemented as options and can be disabled, too. If all special enhancements are disabled, we are quite close to the standard implementation. The remaining differences are: Preambles are different for ETSI and IEEE. Festival and Carnival ASICs are not compliant with any of them (Figure 8.24). This affects synchronisation and the initial channel estimation algorithms. However, similar principles were used such that the proposed algorithms and architectures (section 8.2.2.3 and 8.2.2.4) remain largely applicable. Carnival and Festival implement the inner transmitter/receiver only. Fortunately, the development of the corresponding compliant outer transmitter/receivers is straightforward since common interleaving, scrambling, and encoding/decoding algorithms are used. automatic gain control (AGC) has not been integrated yet into the synchronisation. Currently, AGC is implemented independently in the digital IF FPGA. This leads to a suboptimal signal frame detection. On the one hand, the preamble appears to be the only major modification. For the rest, only extensions are required. On the other hand, this means that Putting it all together 183 our current solution may be suboptimal in general. A new optimisation, including the extensions, may be required. 8.4.3 Optimizing at the system level an outlook Assume we modify the previously described architecture such that it fully meets the standard, how far are we away from the optimum? At least we can say that it is unlikely that we are at the optimum already, since we still used local optimisations in our architecture proposal. We tried to optimise the digital baseband ASIC for certain goals (see section 8.1.2), but we never did the same to the system architecture. This means that there must be still room for improvement. Certainly, we can improve individual components further, but losses will become clearly visible at the interfaces between the components of the system. One performance issue is the latency and memory requirements of the packet handling between physical layer and MAC. Another is finding the optimum position for the analogue/digital boundary between the analogue front-end and the digital signal processing. Two examples from the physical layer, automatic receiver gain control and power-efficient transmission are discussed to illustrate the need for a more global system optimisation. 8.4.3.1 Example 1: Automatic Gain Control The range of signal strength at the receive antenna over which a compliant WLAN receiver has to be able to detect the input signal and decode it, is defined by the maximum input power and the sensitivity level. For the worst case of the standards, we end up with 56 dB of gain variation. Over this range, we have to accommodate any signal strength and produce a signal at the input of the digital receiver with sufficient signal-to-noise ratio for our desired bit error probability. This is the task of the automatic gain control. A single analog or digital gain stage is not able to meet our requirements: 184 Chapter 8 no analogue gain control would require a very wide dynamic range and thus a large wordlength of the analog-to-digital converters. At sampling rates of 20 MHz or larger this translates into high power consumption. On the positive side we could achieve very good accuracy. no digital gain control would require very accurate and wide-range analogue variable gain amplifiers. This again increases cost without need. An additional constraint is the amount of time available in the preamble to estimate the actual signal strength and adapt the gain settings. Taking into account the needs of the other estimation processes (coarse frequency estimation and timing estimation), the part of the preamble that can be used for AGC and signal detection can be established. For HIPERLAN/2 this is equivalent of 4 short training sequences (STS), while the first 4 short symbols can be used in IEEE 802.11a. This gives us a budget of Non- linearities due to clipping or quantization at the ADC or, the limited accuracy of an analogue received signal strength indicator (RSSI), may require more than one estimation and gain adaptation step. Within these constraints, the gain distribution and the optimum boundary between analogue and digital functionality must be optimised. A typical gain control architecture for a superheterodyne front-end (Figure 8.25) applies a mix of both digital and analogue gain control. A switchable low noise amplifier (LNA) allows both a high IP3 in case of strong input signals and a low noise figure in case of weak ones. The IF VGA helps arranging the analogue signal optimally within the dynamic range of the ADC. Finally, the digital gain control allows us to fine-tune before the filtering and again after, when out-of-band interferers have been suppressed. The same mix exists for the estimation process. We could either implement a purely analogue RSSI or a digital signal strength estimator. Again, we have a trade-off between accuracy and implementation cost. Figure 8.25 shows a digital signal strength estimate before and after the IF band-select filter; the difference between those measurements is an indicator for the amount of out-of-band interference (including an aliasing error). The optimisation process iterates on a cascade analysis. Criteria are the cascaded noise figure and linearity requirements. Clipping and quantization in the ADC sampling process introduce noise and non-linearity. The gain settings of each variable gain element are the tuning parameters. They are specified through tuning ranges or discrete settings, e.g. for the switchable LNA. From the optimisation, we obtain the switching points, as a function of the input signal power, between different gain configurations. The process is successful if we can find a valid gain setting for the entire specified input signal strength range for a specific set of front-end constraints. Putting it all together 185 If the gain estimation is digital, additional constraints come from the estimation error due to the clipping non-linearity (Figure 8.26). For signals 186 Chapter 8 in strong saturation the estimation error increases quickly. For the signal-to- noise ratio, there is an optimal gain setting. 8.4.3.2 Example 2: Power-efficient transmission Did we encounter a problem with power-efficient transmission so far? Not in the digital baseband design since the core power consumption was dominated by the receive mode. Unfortunately, we have seen that the problem of power efficiency is linked to the power amplifier in the transmit front-end. This is due to the high peak-to-average power ratio of the OFDM transmission scheme. Traditionally, wireless systems were defined for smaller user bandwidth. Spectral efficiency was much less important and was consequently traded- off for less amplitude modulation and thus a smaller PAPR. Frequency modulation (FM) or GMSK (as used for the GSM mobile phone networks) have a more constant envelope. Wireless LANs however demand a large amount of signal bandwidth. Thus, we have to live with a non-constant envelope and improve architectures and circuits. The power amplifier in the transmitter represents the power consumption bottleneck in todays systems. Power-efficient amplifiers of e.g. class C, D, E, ... exist, but they can not be used to transmit non-constant envelope signals since their strong non-linearity introduces severe distortion of the wanted signal (constellation warping). Also, power leaks into adjacent bands increasing interference for other systems (spectrum regrowth). Conventionally, a rather large back-off is used for power amplifiers to operate them in their linear range, at the cost of a considerable reduction in power efficiency. Linear amplifiers of class A match the linearity requirements but fail concerning power efficiency. Basically, an improvement of the amplifier linearity will not improve the situation, especially if we also consider transmit power control. The amplifier should maintain its efficiency over a large output power range. Therefore, linearization techniques try to transform the signal to transmit into another or a set of other signals with different properties. Linearization reduces constraints on the power amplification stage at the cost of increased complexity, typically leading to architectures with multiple transmission paths. Rather than focussing on the power amplifier alone, we should focus on the entire transmitter. A generic architecture of a transmitter (Figure 8.27) is fed with data from an information source. The signal pre-processing performs the processing steps at the physical layer, including outer (e.g. coding) and inner modem functionality, on the data. The pre-processed information is passed to the analogue chain that performs upconversion steps and the main amplification (power amplifier). Putting it all together 187 Transmitter control and calibration can be fairly simple for narrow-band transmission, constant-envelope signals, or short range communication. When it comes to OFDM, however, a wide bandwidth (3 bands of more than 450 MHz), a power control range of 45 dB, and signals with a large PAPR must be handled efficiently. In addition, distortion must be avoided. In this case, operating points must be tuned towards the current requirements: ranging from the average output power requirement during a burst down to even an adaptation per sampling period. Our design goals can be summarised as follows: 1. power efficient amplification of non-constant envelope signals to enable a bandwidth-efficient modulation scheme like OFDM; 2. power efficient transmit power control to both reduce the interference for other systems and to reduce the power consumption, thus increasing the battery lifetime for wireless devices; 3. short response and transmit-receive TDD/TDMA turn-around time to reduce the protocol overhead in TCP/IP packet-based or ATM cell- based transmission. This has impact on power-up/down ramping and on calibration. An optimisation without a cost function is impossible. Even a simple architecture may meet some of these three goals partially, i.e. for a limited parameter range. This means that we do not need any improvement there. Thus, first, we have to define the cases (for example depending on output power and modulation scheme), where we need further improvement, how much, and what cost function to use. At low output powers of 15 dBm to 0 dBm (1mW) for example, the power dissipation of a class B power amplifier is negligible compared to the power dissipation of the physical layer (which is in the orders of several 100 mW). An improvement of the power efficiency of the PA alone in this case will not lead to a significant drop of the total power dissipation. 188 Chapter 8 This example clearly shows that we end up with a set of operation regions that require a different kind of improvement. This process is decision-driven and operates on a set of. tunable variables. Thus a mixed- signal architecture involving digital control elements seems the most feasible. The actual digital/analogue partitioning and the definition of the switching rules are the goals of the optimisation. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] F. Catthoor, S. Wuytack, E. De Greef, F. Franssen, L. Nachtergaele, H. De Man, System-level Transformations for low data transfer and storage, in Low Power CMOS Design, edited by A. Chandrakasan, R. Brodersen, IEEE Press, pp. 609-618, 1998. L. Deneire, P. Vandenameele, L. Van der Perre, M. Engels, B. Gyselinckx, A Low Complexity ML Channel Estimator for OFDM, in Proc. IEEE 2001 ICC, June 2001. A. M. Despain, Very Fast Fourier Transform Algorithms for Hardware Implementation, in IEEE Trans. Computers, vol. C-28, pp. 333-341, May 1979. W. Eberle, M. Badaroglu, V. Derudder, S. Thoen, P. Vandenameele, L. Van der Perre, M. Vergara, B. Gyselinckx, M. Engels, I. Bolsens, Flexible OFDM Transceiver for High-Speed WLANs, in Proc. IEEE 1999 VTC Fall, vol. 5, pp. 2677-2681, October 1999. W. Eberle, M. Badaroglu, V. Derudder, S. Thoen, P. Vandenameele, L. Van der Perre, M. Vergara, B. Gyselinckx, M. Engels, I. Bolsens, A Digital 80 Mb/s OFDM Transceiver IC for Wireless LAN in the 5 GHz Band, in Dig. Tech. Papers, IEEE 2000 SSCC, pp. 74-75, February 2000. W. Eberle, L. Deneire, H. De Man, B. Gyselinckx, M. Engels, Automatic Gain Control for OFDM-based Wireless Burst Receivers, in Proc. International OFDM Workshop, Hamburg, Germany, September 2000. W. Eberle, M. Badaroglu, V. Derudder, L. Van der Perre, M. Vergara, B. Gyselinckx, M. Engels, I. Bolsens, H. De Man, A Flexible OFDM Transceiver ASIC for High-Speed Wireless Local Networks, in Proc. Int. Conf. on Telecommunications, pp. 1122-1128, Acapulco, Mexico, May 2000. W. Eberle, V. Derudder, L. Van der Perre, G. Vanwijnsberghe, M. Vergara, L. Deneire, B. Gyselinckx, M. Engels, I. Bolsens, H. De Man, A Digital 72 Mb/s 64-QAM OFDM Transceiver for 5 GHz Wireless LAN in 0.18 mm CMOS, in Dig. Tech. Papers, IEEE 2001 ISSCC, pp. 336-337, February 2001. W. Eberle, V. Derudder, G. Vanwijnsberghe, M. Vergara, L. Deneire, L. Van der Perre, M. Engels, I. Bolsens, H. De Man, 80 Mb/s QPSK and 72 Mb/s 64- QAM, flexible and scalable digital OFDM Transceiver ASICs for wireless local area networks in the 5 GHz band, accepted for Journal of Solid-State Circuits, November 2001. ETSI TS 101 475 Technical Specification Broadband Radio Access Networks; HIPERLAN Type 2; Physical (PHY) Layer, ETSI, April 2000. R.F.H. Fischer, J.B. Huber, A new loading algorithm for discrete multitone transmission, in Proc. of Globecom 1996, pp. 724-728, 1996. Putting it all together 189 [12] [13] [14] [15] [16] [17] [18] [19] J. A. Huisken, et al., A Power-Efficient Single-Chip OFDM Demodulator and Channel Decoder for Multimedia Broadcasting, in IEEE Journal of Solid-State Circuits, vol. 33, no. 11, pp. 1793-8, November 1998. IEEE Std 802.11a Supplement to IEEE Std Part 11: WLAN MAC and PHY specifications: High-speed Physical Layer in the 5 GHz Band, IEEE, September 1999. C. Mandl, M. Bacher, G. Krampl, F. Kuttner, 0.35 mm COFDM Receiver Chip for DVB-T, in Dig. Tech. Papers, IEEE 2000 ISSCC, pp. 76-77, February 2000. P. Schaumont, S. Vernalde, L. Rijnders, M. Engels, I. Bolsens, A design environment for the design of complex high-speed ASICs, in Proc. Design Automation Conference, pp. 315-320, June 1998. D. Veithen, et al., A 70 Mb/s Variable-Rate DMT-based Modem for VDSL, in Dig. Tech. Papers, IEEE 1999 ISSCC, pp. 248-249, February 1999. M. Vergara, M. Strum, W. Eberle, B. Gyselinckx, A 195 kFFT/s 256-points High Performance FFT/IFFT Processor for OFDM Applications, in Proc. of SBT/IEEE Int. Telecommunications Symposium, vol. 1, pp. 273-278, 1998. D. Verkest, W. Eberle, P. Schaumont, C++ based System Design of a 72 Mb/s OFDM transceiver for wireless LAN, in Proc. IEEE 2001 CICC, pp. 433-439, May 2001. P. Wambacq, S. Donnay, P. Pieters, W. Diels, K. Vaesen, W. De Raedt, E. Beyne, M. Engels, I. Bolsens, Chip-package co-design of a 5 GHz RF Front- End for WLAN, in Dig. Tech. Papers, IEEE 2000 ISSCC, pp. 318-319, February 2000. Abbreviations ACH ADC ADSL AGC AM AP ARIB ARQ ASIC ASK ATM AWGN BC BCH BER BPSK BRAN BSI CC CDMA CFO CL CM CORDIC CP CPE CPFSK CPU CRC C-SAP Access feedback Channel Analog to Digital Converter Asymmetrical Digital Subscriber Line Automatic Gain Control Amplitude Modulation Access Point Association of Radio Industries and Businesses (Japan) Automatic Repeat reQuest Application Specific Integrated Circuit Amplitude Shift Keying Asynchronous Transfer Mode Additive White Gaussian Noise Burst Controller Broadcast Channel Bit Error Rate Binary Phase Shift Keying Broadband Radio Access Networks Burst State Information Central Controller Code Division Multiple Carrier Frequency Offset Convergence Layer Centralised Mode Coordinate Rotation Digital Computer Cyclic Prefix Common Phase Error Continuous Phase Frequency Shift Keying Central Processing Unit Cyclic Redundancy Check Control Service Access Point 192 OFDM Systems DAB DAC DC DCCH DDI DECT DES DFS DFT DiL DLC DLCC DM DMA DSSS DUC DUCC DVB EC EDA EEPROM ESPRIT ETSI FCC FCH FFT FH FHSS FIFO FM FNC FSM FSMD GCD GMSK GSM GUI ICI IDFT IEEE IF IFFT IL IP IP Digital Audio Broadcasting Digital to Analog Converter Direct Current Dedicated Control Channel Dynamic Datapath Information Digital European Cordless Telephony Data Encryption Standard Dynamic Frequency Selection Discrete Fourier Transform Direct Link Data Link Control DLC Connection Direct Mode Direct Memory Access Direct Sequence Spread Spectrum DLC User Connection DLC User Connection Control Digital Video Broadcasting Signal to interference ratio per bit Signal-to-noise-ratio per bit Error Control Electronic Design Automation Electrically Erasable Programmable Read Only Memory Estimation of Signal Parameters by Rotational Invariance Techniques European Telecommunication Standards Institute Federal Commission on Communications Frame CHannel Fast Fourier Transform Frequency Hopping Frequency Hopping Spread Spectrum First-In-First-Out (buffers) Frequency Modulation Foreign Noise Contribution Finite State Machine Finite State Machine with Datapath Greatest Common Divider Gaussian Minimum Shift Keying Groupe Special Mobile Graphical User Interface Inter Carrier Interference Inverse Discrete Fourier Transform Institute of Electrical and Electronics Engineers Intermediate Frequency Inverse Fast Fourier Transform Implementation Loss Intellectual Property Internet Protocol How to make them work? 193 I/Q ISDN ISI ISM LAN LCH LFSR LMMSE LNA LO LOS LS LSB LTI LTS LUT MAC MAC ID MCM MPEG MSB ML MP MPI MT NFR OBS OFDM ONC PA PAN PAPR PC PCB PCI PDF PDP PDU PM PN ppm PSAM PSK QAM QoS QPSK RCH RF In-Phase versus Quadrature Integrated Services Digital Network Inter Symbol Interference Industrial, Scientific and Medical Band Local Area Network Long transport CHannel Linear Feedback Shift Register Linear Minimum Mean-Squared Error Low Noise Amplifier Local Oscillator Line of Sight Least Squares Least Significant Bit Linear Time Invariant Long Training Symbol Look-Up Table Medium Access Control MAC Identifier Multi Chip Module Motion Picture Expert Group Most Significant Bit Maximum Likelyhood Multi-Path (propagation). Microprocessor Programming Interface Mobile Terminal Negative Frequency Rejection OBstructed line of Sight Orthogonal Frequency Division Multiplexing Own Noise Contribution Power Amplifier Personal Area Network Peak to Average Power Ratio Personal Computer Printed Circuit Board Peripheral Component Interconnect Power Density Function Power Delay Profile Protocol Data Unit Phase Modulation Pseudo Noise part per million Pilot Symbol Assisted Modulation Phase Shift Keying Quadrature Amplitude Modulation Quality of Service Quadrature Phase Shift Keying Random CHannel Radio Frequency 194 OFDM Systems RG RLC RMS RR RSS RSSI RT RX SAP SCH ScFO SDU SER SFG SINR SIR SNR SSR STS SVD TCP TDD TDMA TX U-SAP VCO VDSL VGA WAN WLAN WLL Resource Grant Radio Link Control Protocol Route Mean Square Resource Request Received Signal Strength Received Signal Strength Indicator Register Transfer Receive Service Access Point Short transport Channel Sample Clock Frequency Offset Service Data Unit Symbol Error Rate Signal Flow Graph Signal-to-Interference-plus-Noise Ratio Signal-to-interference Ratio Signal-to-Noise Ratio Symbol Reordening Unit Short Training Sequences Singular Value Decomposition Transmission Control Protocol Time Division Duplex Time Division Multiple Access Transmit User Service Access Point Voltage Controlled Oscillator Very high speed Digital Subscriber Line Variable Gain Amplifier Wide Area Network Wireless Local Area Network Wireless Local Loop How to make them work? Variables a a A(t) b b B c d f F G h(t) h H(f) H exponent of decay of the received power with the distance fraction of time-invariant paths gain of the path amplitude in function of time wordlength curve-fitting parameter bandwidth information bandwidth coherence bandwidth speed of light covariance matrix of vector n distance between transmitter and receiver frequency FFT matrix carrier frequency Doppler frequency frequency of the subcarrier phase noise bandwidth received energy per bit gain receiving antenna gain transmitting antenna gain complex channel impulse response q-th derivative of the channel response at the k-th subcarrier at time channel taps (vector) channel frequency response channel frequency coefficients (vector) least squares estimation of channel linear minimum mean squared error channel estimator 196 OFDM Systems i(t) IIP3 k k k k m n N P(t) R R S s(t) S(d) T T W X identity matrix of size interference input related order intercept point 1dB compression point carrier index path index overhead correction factor Boltzmann constant OFDM symbol index discrete time index noise matrix noise spectral density number of carriers length of the cyclic prefix (in samples) number of channel taps estimated number of channel taps number of paths noise vector for OFDM symbol power delay profile bit error probability Doppler spectrum average input power (of power amplifier) received power transmitted power data symbol rate data bit rate autocorrelation of the channel frequency response over autocorrelation of the amplitudes of the channel frequency response over autocorrelation of the channel response over correlation matrix of the channel interpolation matrix discrete time transmitted signal discrete time transmitted signal for the OFDM symbol transmitted signal path loss for a distance d. temperature transmitted symbol duration coherence time duration of the cyclic prefix sample period symbol duration receiver velocity maximum speed of scatterers time variant amplitude of channel bandwidth transmitted data symbol matrix (frequency domain) How to make them work? 197 Y transmitted data symbol vector for the OFDM symbol (frequency domain) transmitted data symbol vector for the OFDM symbol (time domain) transmitted data symbol (frequency domain) on the carrier in the OFDM symbol transmitted data symbol (time domain) on the carrier in the OFDM symbol saturation voltage received data symbol matrix (frequency domain) received data symbol (frequency domain) on the carrier in the OFDM symbol received data symbol (time domain) on the carrier in the OFDM symbol non-linearity factor of a PA device delta function subcarrier spacing carrier frequency offset phase mismatch (between I and Q) length of channel impulse response gain mismatch (between I and Q) phase noise phase in function of time frequency domain representation of phase noise symbol-to-noise power per bit symbol-to-noise power wavelength singular value normalized clipping level subchannel waveform phase shift of the path power efficiency variance of the signal average delay of a channel delay of the path maximum excess delay of a channel RMS delay spread of a channel ICI leakage constant q-th order ICI crosstalk matrix received data symbol vector for the OFDM symbol (frequency domain) received data symbol vector for the OFDM symbol (time domain) How to make them work? Notation DIAG(x) CONJ(x) q-th derivative of X transpose of matrix X conjugate transpose of matrix X Moore-Penrose pseudo-inverse of matrix X orthogonal projection onto the space spanned by the columns of matrix X diagonal matrix with the elements of vector x Hadamard (i.e. element-wise) product of X with Y conjugate of vector x How to make them work? Index access feedback channel 58 access point 54 acquisition 76 adaptive loading 48 ADC 130 ADSL 2 AM-AM conversion 140 AM-PM conversion 140 ARIB 6, 54 association control function 58 Association of Radio Industries and Businesses 6, 54 auto-correlation 101, 106 automatic gain control 118, 130, 183 AWGN 45 back-off 140 baud rate 33 block codes 48 bluetooth 3 Boltzmann constant 16 BRAN 6, 7, 54 Broadband Radio Access Networks 6, 54 broadcast channel 58 cable modem 2 carrier frequency 70 carrier frequency offset 98 carrier offset 162 carrier sense multiple access with collision avoidance 54, 73 CDMA 2 202 OFDM Systems channel estimation 164 channel frequency response 20 channel model 11 clipping 116, 128, 159 clock frequency offset 98 clock jitter 142 clock offset 166 co-channel interference 4 code division multiple access 2, 4 coded OFDM 48 coherence bandwidth 21 coherence time 27 coherent detection 75 common phase error 132 complementary code keying 53 convergence layer 55 convolutional codes 48 crest factor 116 crest factor regrowth 130 cross-correlation 103 cyclic prefix 37, 104, 109 DAB 5, 6 data link control 174 data link control layer 54 DECT 3 delay locked loop 89 deterministic path loss 15 DFT 36, 39 digital audio broadcasting 5 digital video broadcasting 5 direct mode 54 direct sequence spread spectrum 53 DLC user connection control 60 Doppler bandwidth 28 Doppler effect 12 Doppler shift 42 Doppler spectrum 27 DVB 5, 6 dynamic frequency selection 59 dynamic TDMA 57 equaliser 164 error control 54, 57 ESPRIT 88 estimation of signal parameters by rotational invariance techniques 88 ETSI 6, 7, 54 European Telecommunication Standards Institute 6, 54 excess delay 20 How to make them work? 203 fading 19 fast fading 27 Fast Fourier Transform 158 flat fading 22 foreign noise contribution 118, 132 frame channel 58 frame synchronisation 79 free space loss 13 frequency hopped spread spectrum 17, 53 frequency selective channel response 4 frequency selective fading 22 frequency synchronisation 106 front-end 176 gain mismatch 131 Gaussian minimum shift keying 4 GMSK 4, 186 GSM 2, 3 HIPERACCESS 7 HIPERLAN/2 2, 4, 6, 54 HIPERMAN 3, 7 HiSWANa 6, 54 I/Q imbalance 117, 131 ICI crosstalk matrix 43 ICI leakage constants 43 IDFT 36, 39 IEEE 802.11 2, 53 IEEE 802.1la 4, 6, 54, 73 IEEE 802.l1b 53 IEEE 802.16a 3, 4, 7 implementation loss 100, 116, 120 impulse response 19 inter carrier interference 37, 132 inter symbol interference 19, 33 interference 16 interleaver 49, 66 IS-136 2 IS-95 2, 4 ISDN 1 Jakes spectrum 28 least squares estimator 81 line of sight 19 linear minimum mean squared error estimator 81 link adaptation 64 link budget 12 204 OFDM Systems MAC 57 maximum likelihood estimator 82 medium access control 53, 57, 172 microwave oven 17 minimum description length criterion 88 MMAC 6, 54, 73 mobile terminal 54 multi chip module 177 Multimedia Mobile Access Communication Systems 6, 54 multipath propagation 12, 18 negative frequency rejection 117, 131 noise 15 non line-of-sight 7 non-sample-spaced channel 78 obstructed line of sight 14, 19 OFDM 4, 35, 114 OFDM symbol 39 orthogonal frequency division multiplexing 4, 35 own noise contribution 118, 132 PAN 3 path delay 18 path loss 12 peak-to-average power ratio 116, 186 personal area networks 3 phase noise 98, 118, 132 Phase Shift Keying 4 physical layer 54 pico-cellular networks 15 pilot symbol assisted modulation 77 PLL 135 power amplifier 122, 138, 186 power delay profile 20 preamble 69 probability of bit error 46 propagation loss 12 protocol data unit 56 PSK 4 QAM 4 Quadrature Amplitude Modulation 4 quadrature mismatch 131 quantization 128 radio link control 54 radio resources control 59 How to make them work? 205 random access channel 58 Rayleigh channel 47 ray-tracing 22 received signal strength indicator 184 Reed-Solomon codes 48 RMS delay spread 20 sample spaced channels 78 sample timing jitter 98 sample timing offset 98 scrambler 64 segmentation and reassembly 55 service data unit 55 signal-to-noise ratio 3, 15 single frequency networks 6 singular value decomposition 81 slow fading 27 SNR 3 spectral shaping 77 statistical path loss 14 subcarriers 34 symbol duration 34 symbol timimg drift 99 symbol timing 96 TDMA 2, 3 thermal noise 16 time-division multiple access 2 timing acquisition 161 timing synchronisation 100 training sequence 105, 109 transmission power control 60 trellis coded modulation 48 Turbo codes 48 VDSL 2 Viterbi decoder 48 WAN 2 W-CDMA 4 wide area network 2 Wiener filtering 91 wireless local area networks 2, 4, 5 wireless local loop 6 WLAN 2, 4, 5, 6, 53 WLL 3, 4, 5, 6
2016 - Using Modern Plant Trait Relationships Between Observed and Theoretical Maximum Stomatal Conductance and Vein Density To Examine Patterns of Plant Macroevolution