Professional Documents
Culture Documents
Thesis Adaptive Antennas For CDMA WIRELESS Networks
Thesis Adaptive Antennas For CDMA WIRELESS Networks
A DISSERTATION SUBMITTED TO THE DEPARTMENT OF ELECTRICAL ENGINEERING AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
ii
I certify that I have read this thesis and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy.
I certify that I have read this thesis and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy.
Thomas Kailath
I certify that I have read this thesis and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy.
Donald Cox
iii
Abstract
Wireless cellular communication is witnessing a rapid growth in markets, technology, and range of services. A major current thrust for cellular communication systems is improved economics through enhanced coverage early in the life cycle of a network and high spectrum efciency later in the life cycle. An attractive approach for economical, spectrally efcient, and high quality digital cellular and personal communication services (PCS) is the use of spread spectrum modulation with code division multiple access (CDMA) technology. Yet another very promising dimension for improving performance of all types of cellular networks is the use of antenna arrays at the base station. This thesis explores techniques for integrating these two dimensions. We study the use of multiple antennas at the base station and the associated advanced signal processing in CDMA wireless networks. Our focus is on the mobile to base or reverse link. We begin with a space-time cellular channel model for spread spectrum networks. We then describe the appropriate signal and interference models. We then propose a Beamformer-RAKE receiver structure for exploiting multiple antennas at the base station. An estimator for the beamforming weight vector in the presence of angle and Doppler spreads is derived using a code ltering approach and its tracking performance is established. The above Beamformer-RAKE structure is then applied to a specic M-ary orthogonal modulation and noncoherent RAKE combining receiver used in an existing interim CDMA standard. We propose an overall antenna array base station receiver architecture for this iv
system and study the system performance in terms of BER and power control loop performance. The effect of angle spread, number of antennas, multiple resolved paths and Doppler spread are studied.
Acknowledgments
irst and foremost, I am thankful to God, the most gracious most merciful for helping me nish this dissertation. It is my belief in him that helped me persevere at times
when it seemed impossible to go on. Many individuals have profoundly inuenced me during my graduate studies at Stanford, and it is a pleasure to acknowledge their guidance and support. I would like to begin by expressing sincere thanks to my advisor, Professor Arogyaswami Paulraj, for his support for the past three years. Professor Paulraj has also provided me with generous nancial support, both in terms of research assistantships and support for attendance at numerous technical meetings. I wish to express my appreciation to members of my orals and reading committees: Professors Thomas Kailath, John Ciof, Donald Cox, and Leonard Tyler for taking the time to read and critique my thesis. Professor Cox was especially helpful in this regard. A special thanks to Professor Kailath not only for recommending me to Professor Paulraj but also for his continuous support that I felt throughout my studies at Stanford. I truly appreciate the support of Professor Abbas El Gamal, during my rst two years at Stanford and for his encouragement and for being a sincere friend when I really needed one. Our groups Secretary Christine Linke was instrumental in getting things done. I am grateful to her for being such a good listener on those many occasions when I just needed to talk. My association with Professor Paulrajs and Professor Kailaths groups has been also a source of invaluable experience and friendship for me. I would particularly like to thank vi
(in alphabetical order): Dr. Hamid Aghajan, Suhas Diggavi, Khalid El Awady, Alper Erdogan, Derek Gerlach, Bijit Halder, Babak Hassibi, Babak Khalaj, Jen-Wei Liang, Tushar Moorti, Boon Ng, Gregory Raleigh, Sumeet Sandhu, Dr. Ali Sayed, Shilpa Talwar, and Dr. Mats Viberg. The stimulating discussions that we had during Tuesday and Friday afternoon group meetings helped me shape and redene my understanding of the eld, and had a direct inuence on this thesis. Suhas Diggavi, Babak Khalaj, Gregory Raleigh, and Dr. Mats Viberg were especially helpful in this regard. I have also enjoyed the friendship of many others within the Information Systems Lab. In particular, I would like to thank Barry Andrews, Michael Grant, Paul Dankoski, and Vincent K. Jones. I have very fond memories of our lively discussions and the good times we shared. Outside the Lab, I have also cherished the company of several friends throughout my stay at Stanford. In particular, I would like to thank Hisham Abdelhamid, Naofal Al Dhahir, Walid Azzam, Jalel Azaiez, Khalid El Awady, Yaser Haddarah, Youssef Ismail, Sadok Kallel, Wael Lut, Aladdin Nassar, and Mazhar Islam Raja. In addition, I shared unforgettable hours of fun, intellect, and spiritual uplift (not to forget the Friday dinners!) with many members (too numerous to mention) of the Islamic Society at Stanford. Our Friday meeting was always an event to look forward to after a long week. I am indebted to my mother and my late father for teaching me the importance of hard work and perseverance and for instilling in me the condence that I could succeed at whatever I chose to do. Finally, I would like to dedicate this work to my wife Fatma and to my three kids Ahmed, Sumayyah, and Omar. I want to thank Fatma for her patience, love, her desire that I succeed, for putting up with me, and for the many sacrices she has had to make over the past seven years. I also wish to thank Ahmed, Sumayyah, and little Omar for their smiles, their reckless enjoyment of life, and for always reminding me that the most important work that I will ever do is within the walls of my own home!.
vii
Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv vi xii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii 1 Introduction 1.1 Cellular Systems and Standards 1.1.1 1.1.2 1.1.3 1.2 1.3 1.4 . . . . . . . . . . . . . . . . . . . . European Global System for Mobile (GSM) . . . . . . . . . . North American TDMA Digital Cellular (IS-54) . . . . . . . . North American CDMA Digital Cellular (IS-95) . . . . . . . . 1 3 5 7 8 9 10 12 13 14 16 19 25 27 27
2 Description and Modeling of Wireless Channels 2.1 Fundamentals of Radio Propagation . . . . . . . . . . . . . . . . . . 2.1.1 2.1.2 2.1.3 2.2 Propagation Path Loss . . . . . . . . . . . . . . . . . . . . . Multipath Fast Fading . . . . . . . . . . . . . . . . . . . . . Log-Normal Slow Fading . . . . . . . . . . . . . . . . . . .
Vector Channel Modeling . . . . . . . . . . . . . . . . . . . Path Amplitudes and Power-Delay Prole . . . . . . . . . . . Macro Cells . . . . . . . . . . . . . . . . . . . . . . . . . . Micro Cells . . . . . . . . . . . . . . . . . . . . . . . . . . Pico or Indoor Cells . . . . . . . . . . . . . . . . . . . . . .
31 41 42 43 44 44 45 46 48 49 54 55 57 58 60 65 67 69 72 75 82 83 84 85 87 92
2.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3 Adaptive Beamforming with Antenna Arrays 3.1 Adaptive Beamforming Techniques . . . . . . . . . . . . . . . . . . . 3.1.1 3.1.2 3.1.3 3.2 Direction-Finding Based Beamforming . . . . . . . . . . . . . Beamforming Based on Training-Signals . . . . . . . . . . . . Signal-Structure-Based Beamforming . . . . . . . . . . . . . CDMA Signal Models . . . . . . . . . . . . . . . . . . . . . Code-Filtering Approach for Adaptive Beamforming . . . . . . Space-Time Matched Filter . . . . . . . . . . . . . . . . . . . Beamformer-RAKE Receiver Structure . . . . . . . . . . . . . A Simulation Example . . . . . . . . . . . . . . . . . . . . .
3.3
3.4 3.5
4 Beamforming for Time-Variant Channels 4.1 Recursive Adaptive Beamforming . . . . . . . . . . . . . . . . . . . 4.1.1 4.1.2 4.1.3 Recursive Estimation of the Channel Vector . . . . . . . . . . Time-Update of Covariance Estimates . . . . . . . . . . . . . Algorithm Summary . . . . . . . . . . . . . . . . . . . . . . ix
4.2 4.3
95 99 100
CDMA Forward Link . . . . . . . . . . . . . . . . . . . . . . . . . 102 CDMA Reverse Link . . . . . . . . . . . . . . . . . . . . . . . . . . 104 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 108
6 CDMA Base Station Receiver with Antenna Arrays 6.1 6.2 6.3
Received Signal Vector Model . . . . . . . . . . . . . . . . . . . . . 110 Receiver Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 Signal Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 6.3.1 6.3.2 6.3.3 Noise Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 118 Self and Multiple Access Interference Analysis . . . . . . . . . 119 Decision Statistics . . . . . . . . . . . . . . . . . . . . . . . 122 Low Doppler Frequency . . . . . . . . . . . . . . . . . . . . 124 High Doppler Frequency . . . . . . . . . . . . . . . . . . . . 126
6.4
6.5 6.6
CDMA Reverse Link Open Loop Power Control . . . . . . . . . . . . 140 CDMA Reverse Link Closed Loop Power Control . . . . . . . . . . . 141 Closed Loop Power Control Model . . . . . . . . . . . . . . . . . . . 142 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
154
Fading Correlation for Scalar Channels . . . . . . . . . . . . . . . . . 157 Fading Correlation for Vector Channels . . . . . . . . . . . . . . . . . 158 162 165
xi
List of Tables
3.1 3.2 4.1 4.2 4.3 4.4 6.1 Optimum weight vector for SINR, ML, and MMSE performance measures Estimated multipath parameters . . . . . . . . . . . . . . . . . . . . . . Power recursion for estimating principal eigenvector . . . . . . . . . . . Inverse square root time-update . . . . . . . . . . . . . . . . . . . . . . Beamforming algorithm summary . . . . . . . . . . . . . . . . . . . . . Floating point operations count . . . . . . . . . . . . . . . . . . . . . . Percent reduction in capacity at Pb = 102 and Pb = 103 for high Doppler frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 54 73 86 89 91 93
xii
List of Figures
1.1 1.2 1.3 1.4 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 2.10 2.11 2.12 2.13 2.14 3.1 3.2 If I had a cellular telephone . . . . . . . . . . . . . . . . . . . . . . . Overview of a cellular telephone network . . . . . . . . . . . . . . . . GSM frame and time-slot structure . . . . . . . . . . . . . . . . . . . IS-54 frame and time-slot structure . . . . . . . . . . . . . . . . . . . Illustration of the wireless propagation environment. . . . . . . . . . . Direct and indirect paths on a at-terrain environment. . . . . . . . . . Magnitude of a Rayleigh fading channel: fd = 80 Hz and Ts = 104.2s. Fast fading amplitude correlation vs. delay (in wavelengths traveled). . A Wireless communication system employing antenna arrays. . . . . . A one-dimensional array manifold. . . . . . . . . . . . . . . . . . . . The uniform linear array scenario. . . . . . . . . . . . . . . . . . . . An illustration of the vector channel. . . . . . . . . . . . . . . . . . . Model geometry. . . . . . . . . . . . . . . . . . . . . . . . . . . . . Spatial envelope correlation vs. antenna spacing: mean AOA = 0 . . Spatial envelope correlation vs. antenna spacing: mean AOA = 30 . Space-Time Fading: fd = 50 Hz, Space-Time Fading: fd = 50 Hz, Space-Time Fading: fd = 50 Hz, = 0 , and Ts = 208.3s. . . . . . . = 3 , and Ts = 208.3s. . . . . . . = 40 , and Ts = 208.3s. . . . . . 2 4 5 7 15 16 23 24 28 29 31 32 35 36 37 38 39 40 47 56
Block diagram of an adaptive array system . . . . . . . . . . . . . . . Constant modulus property restoral beamformer . . . . . . . . . . . . xiii
3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 4.1 4.2 4.3 4.4 4.5 5.1 5.2 5.3 6.1 6.2 6.3 6.4 6.5 6.6
Mobile transmitter block diagram . . . . . . . . . . . . . . . . . . . . CDMA receiver-beamformer block diagram . . . . . . . . . . . . . . RAKE receiver block diagram . . . . . . . . . . . . . . . . . . . . . Multipath CDMA RAKE receiver . . . . . . . . . . . . . . . . . . . Space-Time Matched Filter . . . . . . . . . . . . . . . . . . . . . . . Beamformer-RAKE receiver structure . . . . . . . . . . . . . . . . . Estimated multipath prole . . . . . . . . . . . . . . . . . . . . . . . = 0 . . . . . . . .
59 63 65 66 69 70 74
Balanced DQPSK Beamformer-RAKE receiver with incoherent combining 75 Pe for balanced DQPSK with incoherent combining and 77 78 79 81 82 94 95 96 97 98
Balanced QPSK Beamformer-RAKE receiver with coherent combining . Pe for balanced QPSK with coherent combining and = 0 . . . . . . . . . . . .
Pe for balanced QPSK with coherent combining and large Mismatch loss as a function of f d and Mismatch loss as a function of f d and
CDMA IS-95 forward link waveform generation . . . . . . . . . . . . 102 CDMA IS-95 reverse link waveform generation . . . . . . . . . . . . 105 CDMA Base station receiver: IS-95 approach . . . . . . . . . . . . . 106 CDMA Base station receiver: beamforming approach . . . . . . . . . 109 Base station receiver block diagram . . . . . . . . . . . . . . . . . . 113 Correlators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 Optimum Beamforming and Incoherent RAKE . . . . . . . . . . . . . 117 Simulation scenario . . . . . . . . . . . . . . . . . . . . . . . . . . 129 I-channel: rst antenna interference distribution . . . . . . . . . . . . 130 xiv
Q-channel: rst antenna interference distribution . . . . . . . . . . . . 131 Pb for f d = 5 Hz and closed loop power control. . . . . . . . . . . . . 132
(n) PDF of z1 for n = h at high f d . . . . . . . . . . . . . . . . . . . . . 133
= 0 , and power control. . . . . . . . . . . . . . 134 = 3 , and power control. . . . . . . . . . . . . . 135 = 60 , and power control. . . . . . . . . . . . . 136
Feedback Power Control Model . . . . . . . . . . . . . . . . . . . . 142 Power Control Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 145 Power-controlled received signal vs. simulated Rayleigh fading: fd = 5 Hz, K = 5, L = 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 Power-controlled received signal vs. simulated Rayleigh fading: fd = 100 Hz, K = 5, L = 4. . . . . . . . . . . . . . . . . . . . . . . . . . 147 Power control error vs. number of paths: K = 5. . . . . . . . . . . . . 148 Power control error vs. power step size: L = 2,K = 5. . . . . . . . . . 149 Power control error vs. loop delay: L = 2,K = 5. . . . . . . . . . . . 150 Power control error vs. forward link error rate: L = 2, K = 5. . . . . . 151 Power control error vs. angle spread: L = 4, and K = 5. . . . . . . . . 152 Power control error vs. angle spread: L = 4, and K = 9. . . . . . . . . 153
7.4
xv
Chapter 1 Introduction
The realization of wireless communications providing high-speed and high-quality information exchange between two portable terminals that might be located anywhere in the world is the new communications challenge for the next decade. The great popularity of cordless phones, cellular phones, radio paging, and other emerging portable communication technologies demonstrates a great demand for such services. For example, from 1990 to 1994, the number of cellular telephone users in the US has grown from 5.1 million to 23.3 million subscribers, and by the year 2000 it is projected that the number will nearly double to 46.9 million subscribers. What has emerged from the world wide research and development activity in this area is the need for new technology advances to meet such an explosive growth. These include: New techniques to improve the quality and spectral efciency of communication over wireless channels. Better techniques for sharing the limited spectrum to accommodate different wireless services. New signal processing techniques to implement various functions of the cellular system. 1
Chapter 1. Introduction
Figure 1.1: If I had a cellular telephone ... (Herman c Jim Unger [1]. Reprinted with permission of Universal Press Syndicate. All rights reserved) The physical limitations of the wireless channel present a fundamental technical challenge for reliable communications. The channel is susceptible to time-variant noise, interference, and multipath. Moreover, radio spectrum is now a limited resource, and even with the recent increase in the spectrum allocation for wireless services, it will be stretched out to its capacity limit to accommodate various current and emerging wireless services. Also, limitation in the power and size of the communication and computing device is another major design consideration. Most personal communications and wireless services devices are meant to be carried in a briefcase or a pocket and, therefore, must be small and lightweight, which translates to a low power requirement, since small batteries must be used.
Chapter 1. Introduction
However, many of the signal processing techniques used for efcient spectral utilization demand signicant processing power, precluding the use of low power devices. Continuing advances in VLSI and integrated circuit technology for low power applications will provide a partial solution to this problem. However, placing most of the signal processing burden on xed location sites with large power resources has and will continue to be the common trend in wireless systems design. Cellular communication systems involve two radio links: the reverse link from the mobile to the base station, and the forward link from the base station to the mobile. In this thesis, we study the use of multiple antennas at the base station and the associated advanced signal processing for CDMA wireless systems and the effect of such processing on system performance measures such as capacity, range, and mobile transmit power. In this work we will focus on the reverse (mobile to base) link.
Chapter 1. Introduction
T1/Wave MTSO
Figure 1.2: Overview of a cellular telephone network the boundary between two cells, its communication channel is switched, or handed off, to the base station in the new cell. The shape of each cell is determined by the power footprint of the transmitting base station. An illustration of a cellular telephone network is shown in Figure 1.2. In the US, cellular telephone service is provided mostly by the AMPS (Advanced Mobile Phone Service) system [2-4] which evolved from extensive research at Bell Laboratories in the 1970s. In this system analog voice signals are frequency modulated onto carriers in the 800 MHz band (824-849 for the downlink and 869-894 for uplink). To prevent signicant adjacent-channel interference, each user channel is allotted a 30 kHz bandwidth. Only a certain set of the carrier frequencies is available in a given cell, and neighboring cells must use different sets of carriers. This method of frequency reuse results in reduced cochannel interference, but the resulting system capacity is only one call per 210 kHz [4]. The increasing demand for cellular telephone service has pushed this technology to its capacity limits
Chapter 1. Introduction
Frame = 4.6155 ms 1 2 3 4 5 6 7 8
T 3
Data 26
F 1
Train 26
F 1
Data 26
T 3
Guard 8.25
Slot = 577 s T: Tail bits F: Flag Train: Equalizer training sequence Guard: Guard time interval
Figure 1.3: GSM frame and time-slot structure and resulted in proposals for a second generation of cellular systems, employing digital modulation techniques and alternative multiple access techniques. A brief description of several of these techniques follows.
Chapter 1. Introduction
The bands allocated to GSM are 890-915 MHz for downlink and 935-960 MHz for uplink. The band in either direction is divided into 124 frequency channels, each with carriers spaced 200 kHz apart. Each cell site in a GSM system has a xed number of frequency channels (two way) ranging from only one to usually not more than 15. In GSM, an FDMA/TDMA (frequency division multiple access/time division multiple access) radio channel structure is used. Each FDMA frequency channel supports a multiple user signal format. Each 200 kHz FDMA channel uses an aggregate bit rate of 270.833 kbits/sec, carried over the radio interface using GMSK (Gaussian-ltered minimum-shift keying) modulation with a bandwidth-time product of 0.3. The use of a transmitted bit rate as high as 270 kbits/sec requires the implementation of adaptive equalization techniques to deal with channel multipath, and GSM specications require that equipment be built to accommodate RMS delay spread up to 16 s. The GSM standard provides for the use of slow frequencyhopping as a means of reducing other user interference, though no GSM systems currently implements this technique. The 270-kbits/sec data stream in each FDMA channel is divided into 8 xed-assignment TDMA channels or time slots termed logical channels. Each slot is 577 s, which corresponds to the transmission time for about 156.25 bits, though only 148 bits are actually transmitted in each slot. The remaining time, 8.25 bits time duration or about 30.5 s, is guard time in which no signal is transmitted to prevent overlapping of signal bursts arriving at a base station from different mobile terminals. Figure 1.3 shows the GSM frame and time-slot structure. The logical channels are organized into a hierarchical frame structure that provides each mobile terminal with a two-way trafc channel and a separate two-way control channel. The numbering of time slots is offset between two directions on the downlink to prevent a mobile terminal from transmitting and receiving at the same time. More details on the GSM frame hierarchy and network issues can be found in [4-7].
Chapter 1. Introduction
Frame = 40 ms 1 2 3 4 5 6
Mobile to Base G R Data Synch 6 6 16 28 Data 122 CCH DVCC 12 12 Data 122
Slot = 6.67 ms Base to Mobile Synch 28 CCH 12 Data 130 DVCC 12 Data 130 RSVD 12
DVCC: Digital verification color code RSVD: Reserved for future use CCH: control channel
Chapter 1. Introduction
AMPS, with carriers spaced 30 kHz apart. This was done so as to allow conversion of individual analog channels to digital operation. In the initial phase of TDMA implementation, the mobile phones are dual-mode devices, capable of operating on either AMPS analog channels or TDMA digital channels, and cellular operators are required to continue supporting AMPS users as digital service is introduced. The IS-54 standard species the radio channel modulation as /4-shift DQPSK (differentially encoded phase shift keying), to be implemented with square-root raised cosine ltering and rolloff parameter 0.35. The principal advantage of this modulation scheme is bandwidth efciency. The channel transmission rate is 48.6 kbits/sec; and with a channel spacing of 30 kHz, this yields channel utilization of 1.62 bits/sec/Hz, a 20% improvement over GSM. The principal disadvantage of a linear modulation scheme is a lower efciency of linear transmitter power ampliers as compared to nonlinear ampliers that can be used with constant envelope modulation. This is reected in the size and weight of the hand set. In IS-54, each 30-kHz digital channel has a transmission rate of 48.6 kbits/sec. The 48.6kbits/sec stream is divided into 6 TDMA channels of 8.1 kbits/sec each. The IS-54 time slot and frame format, as shown in Figure 1.4, is much simpler than that of the GSM standard. The 40 ms frame is composed of six 6.67-ms time slots corresponding to 324 bits each. Although needed in only a few specic places for 48.8 kbits/sec transmission rate, a delay spread equalizer is required in the IS-54 standard.
Chapter 1. Introduction
forward and reverse links actually use separate carrier frequencies, spaced 45 MHz apart.) Clearly, this design does not lend itself to selective channel replacement of analog AMPS, unlike the TDMA IS-54 system. Instead, large blocks of channels will be replaced at one time by the CDMA system [5]. Therefore, the IS-95 standard species a dual-mode design of the mobile station so that the CDMA digital cellular units can be still used in areas with existing analog cellular coverage. As one of the goals of this thesis is to apply advanced array signal processing techniques derived therein to CDMA systems, a full description of the IS-95 CDMA system will be given later in Chapter 5.
Chapter 1. Introduction
10
For example, when a network is rst deployed, capacity is not an issue (especially in rural areas). As the system matures, more and more users will request wireless services and the need for more capacity develops. In this later stage, the gain due to spatial processing with antenna arrays can facilitate a denser use of the available bandwidth and, therefore, an increase in system capacity. In the early stages of system deployment, service providers strive to minimize the initial deployment cost by covering a given service area with the least number of base stations. The gain provided by spatial processing with antenna arrays can be used to help meet this coverage goal. This gain can also be used to improve the signal quality itself, which leads to low outage probability, better voice quality, and the ability to support higher data rates for wireless data services. Other possible benets of spatial processing with antenna arrays include the ability to reduce the loss in performance when low power terminals co-exist with high power terminals.
Chapter 1. Introduction
11
both the temporal and spatial structure of the received multipath signal to maximize performance. In Chapter 4, we develop a recursive algorithm for estimating the vector channel and the optimum beamforming weights. We then study its tracking performance under realistic channel conditions. Chapters 5 and 6 describe an application of the above space-time processing techniques to the current CDMA standard IS-95. Chapter 5 gives a brief description of the proposed IS-95 system, with more emphasis on describing the reverse link modem. In Chapter 6, we present an overall base station receiver structure with antenna arrays and describe its building blocks. We then study the performance of the proposed receiver in terms of the uncoded bit error rate (BER) as a function of loading (number of users) under different channel conditions. In our analysis, we assume perfect channel estimates and do not include tracking errors. As we mentioned earlier, closed-loop power control is a key element of the current implementation of the IS-95 system. In Chapter 7, we describe a method for closed-loop control. We also present a simulation study to characterize the performance of closed-loop power control under different operating scenarios. Chapter 8 contains concluding remarks and a summary of the thesis. In Appendix A, we derive the fading correlation for both scalar and vector multipath channels. In Appendix B, we derive the probability distributions used in the analysis in Chapter 6. Several of the main ideas in this thesis have already been published by the author in the course of his research. The work on the vector channel model has been partially published in [9]. The idea of space-time processing and Beamformer-RAKE is presented in [10, 11]. The performance enhancement in CDMA systems due to the use of antenna arrays has been published in [12-14]. The recursive algorithm for estimating and tracking the channel vector is presented in [15, 16]. The proposed base station receiver architecture and its performance analysis are presented in [17-19]. Results on the power control performance are presented in [20].
Chapter 1. Introduction
12
13
14
15
Dominant Reflector
Base Station
Dominant Reflector
Figure 2.1: Illustration of the wireless propagation environment. In general, the mobile radio propagation in such environments is characterized by three partially separable effects: path loss, multipath fading, and shadowing. Path loss is a function of the distance between the mobile and the base-station. As discussed earlier, multiple signal reections arrive at the base-station each with its own phase, which causes destructive and constructive interference. The resulting variations in the signal amplitude, called multipath fading, vary over distances proportional to the signal wavelength; thus, this type of signal fading is referred to as fast fading. When the number of multipath components is large, the central limit theorem can be invoked to model the fast fading by a ltered complex Gaussian process. Multipath fast fading is described by its envelope fading (at fading), Doppler spread (time-selective fading), time-delay spread (frequency selective fading), and angle spread (space selective fading). In addition to the multipath fast fading, the LOS and reected paths may be attenuated by large obstructions such as large buildings and hills that are positioned between the mobile
16
` d ht r ` r d
Figure 2.2: Direct and indirect paths on a at-terrain environment. and base station. This type of fading, which is called shadowing, varies with distances that are proportional to the sizes of the buildings (or the obstructions in general), and is thus referred to as slow fading. Most empirical studies show that the variations in signal level due to the slow fading follow a log-normal distribution.
hr
Consider the propagation of a radio wave in a at-terrain environment. The transmitted signal may reach the receiving antenna in several ways: Through a direct LOS path. Through an indirect path consisting of the radio wave reected by the ground. Through an indirect path consisting of a surface wave.
17
Let be the time delay between the direct LOS path and the reected path. If we assume that is much smaller than the inverse bandwidth of the transmitted signal B1 and if we neglect the effect of surface wave attenuation1, then the received power Pr is [21] Pr = Pt Gt Gr 2d
2
1 + Re j
(2.1.2)
where Gt and Gr are the gains of the transmitting and receiving antennas, respectively (note that we assumed that Gt Gr is the same for both the direct LOS path and the reected path), is the wavelength, R is the reection coefcient of the ground, and ference between the direct LOS path and the reected path = = 2(r + r d) ` ` 2 h t + hr d
2
1/2
+1
h t hr d
1/2
(2.1.3)
+1
where ht and hr are the heights of the transmitting and receiving antennas, respectively. Equation (2.1.2) has been shown to agree very closely with the measurements in [22]. The ground reection coefcient R is given by R= where 2 r cos / Z= cos2
r
(2.1.4)
(2.1.5)
and
is the dielectric constant of the ground, which for earth or road surfaces is approxi-
mately that of a pure dielectric ( r = 15). For large d, we have 4ht hr d (2.1.6)
Also, in this case the grazing angle 0 , and therefore R 1. Moreover, R tends to 1 for frequencies above 100 MHz and incidence angles less than 10 , irrespective of the
1
This is a valid approximation for antennas located more than a few wavelengths above the ground.
18
(2.1.7)
Thus, in the asymptotic limit of large d, the received power falls off inversely with d4 . In [22], plots of (2.1.2) as a function of distance illustrate this asymptotic behavior. Up to a certain critical distance dc , the wave experiences constructive and destructive interference of the two rays. At distance dc , the nal maxima is reached, after which the signal power falls off inversely with d4 . If we average out the local maxima and minima, the resulting average power loss can be approximated by dividing the power loss curve into two regions. For d < dc , the average power fall off with distance corresponds to free space (i.e. d2 ). For d > dc , the average power fall off with distance is approximated by the fourth power law in (2.1.7). In practice, the attenuation of radio signals is greater than that of free space and less than that of free space with a perfectly reecting ground. Measurements [24] show that for frequencies around 100 MHz, the attenuation versus distance curve has a slope comparable to that of free space with additional loss depending on the environment. For example for a suburban environment around 100 MHz, the attenuation curve follows free space with an additional 8.5 dB loss. In [25], it was found that at a carrier frequency of 910 MHz in an urban environment, the attenuation loss in dB has a steeper slope than that of free space but not as steep as with a fourth power law. Measurements at 800 MHz [26, 27] also show two distinctly different propagation loss slopes before and after the critical distance dc . For d < dc , the slope is slightly less than 2 while for d > dc the slope is close to 4. The two-ray model described above is a simple model that characterizes signal propagation in isolated areas with few reectors such as rural roads or highways. It requires information only about antenna heights. A more complicated model for urban area transmissions is the dielectric canyon or ten-rays model developed by Amitay [28]. This model assumes rectilinear streets (such as in downtown Manhattan) with buildings along both sides of the
19
street and transmitter and receiver well below the tops of the buildings. The building-lined streets act as a dielectric canyon to the propagating signal. The power fall off in both the dielectric canyon model and urban measurements [22, 29, 30] is proportional to d 2 , even at relatively large distances. Moreover, the fall off exponent is relatively insensitive to the transmitter height, as long as the transmitter is signicantly below the building skyline. This fall off with distance squared is due to the dominance of the wall-reected rays, which decay as d2 , over the combination of the LOS and groundreected rays (two-path model above), which decays as d4 .
where s(t) is the complex baseband signal with bandwidth B, f is the carrier frequency, and o is an arbitrary initial phase. Without loss of generality we will assume that o is zero. If we assume that the mobile is moving at speed v and there is no direct LOS, and if we ignore the receiver additive white Gaussian (AWGN) noise, the corresponding received signal at the base station is the sum of all multipath components [21, 31]
L
y(t) = A
i=1
(2.1.9)
where A includes the effects of distance loss and antenna gains (here we assumed that the propagation distance spread r = max ri min ri is much less than the propagation distance
i i
ri for all i and, therefore, that the attenuation with distance is the same for each component. This is true in general if all reectors are within the vicinity of the mobile). For the ith multipath component, R2 is the fraction of the incoming power in the ith path, i = ri /c is the i
20
multipath delay where ri is the propagation distance and c is the speed of light, and fd cos i is the Doppler shift where fd = v/ is the maximum Doppler shift and i is the direction of the ith scatterer with respect to the mobile velocity vector. The Doppler spread fm is given by 2 f d . These parameters vary with time. As noted in [21], the process y(t) is widesense stationary with respect to ensemble averages. It is not stationary with respect to time averages, however, and thus nonergodic. But the difference between time and ensemble averages decreases as the number of paths L becomes large; thus the statistical properties will be computed on the basis of ensemble averages. Throughout the rest of this section, we will assume that A = 1. If we assume that the multipath delay spread dened as
T = max i min i
i i
is a narrowband signal, then s(t i ) s(t o ) where o [min i , max i ] [31]. Then, we can rewrite (2.1.9) as
L
y(t) s(t o )
i=1
Ri e ji (t) e j2 f t
(2.1.11)
where i (t) = 2( f d cos i t f i ). The phases i (t) modulo 2 can be modeled as i.i.d. random variables uniformly distributed over [0, 2] [32]. Analytical results based on this assumptions agree with measurement results in [32]. The equivalent lowpass received signal is
L
y(t) s(t o )
i=1
Ri e ji (t)
(2.1.12)
In addition to the time delay, the received signal in (2.1.12) differs from the original transmitted signal by the complex scale factor in the parentheses. Let
L
(t) =
i=1
(2.1.13)
21
Since y (t) is the response of an equivalent lowpass channel to the lowpass signal s(t), it fol lows that the equivalent lowpass channel is described by the time-variant impulse response [33] h(t; ) = ( o )(t)e j(t) (2.1.14)
If we assume that the Ri are i.i.d. and independent of the i , then the rst and second moments of (t) are E{(t)} = 0 E{(t) (t + )} =
i
If we also assume that, in addition to being i.i.d., the Ri have bounded variance, then (t) will approach a complex Gaussian random variable as the number of scatterers L becomes large [34]. In this case (t) has a Rayleigh distribution [33] f () = 2 2 exp 2 , 2 0 (2.1.17)
where 2 = E{2 }. Therefore the variation of the received signal envelope is Rayleigh, which has also been conrmed by measurements [35, 36]. Now, if the direct LOS is not obstructed, then (t) will have a Rician distribution [33] f () = 2 + 2 2 exp 2 2 Io 2 , 2 0 (2.1.18)
where 2 is the average power in the direct LOS and In (.) is the modied Bessel function of the n-th order [37]. Note that the real and imaginary parts of (t) are independent,
L 2 0
E{Re{(t)}Im{(t)}} =
i=1
= 0
Note: A more general distribution for the fast fading amplitude is given by the Nakagami distribution [32] f () = 2mm 2m1 m2 exp 2 (m) 2
(2.1.20)
22
where 2 = E{2 } and m = 4 /E{(2 2 )2 } and (.) is the gamma function. This distribution is general in the sense that none of the above assumptions has to hold. When the scattering process generates merely diffuse wave eld, then m 1 [32] and the Nakagami distribution in (2.1.20) is identical to the Rayleigh distribution in (2.1.17). When a direct component is present, the Nakagami distribution approximates the Rice distribution in (2.1.18) with m > 1. It remains now to determine the time-frequency correlation behavior of (, t), where (, t) is given in (2.1.13). In order to do that, let f T () be the probability density function of the time delay i , where f T () is nonzero for 0 < and zero otherwise. In Appendix A, it is shown that the time-frequency correlation of (, t) and ( + is given by ( , ) = E{(, t) ( + = Jo (d ) FT ( j ) , t + )} (2.1.21) (2.1.22) , t + )
where Jn (.) is the Bessel funcion of the rst kind of order n [37] and FT ( j ) is the Fourier transform of f T (). Interpretation of some measured data [21] indicates that the path time delay i can be modeled as an exponential random variable. That is f T () = 1 exp T T , 0 (2.1.23)
(2.1.25)
To a good approximation [21, 23, 38], the envelope correlation ( , ) is equal to the squared magnitude of the complex signal correlation, i.e.. Cov{(, t) ( + , t + )} ( , ) = Var{(, t) }Var{( + , t + )} (2.1.26)
23
10
-5
-1 0
-1 5
(2.1.27)
Figure 2.3 shows the magnitude for a simulated multipath Rayleigh fading channel. The mobile is assumed to be moving such that f d = 80 Hz. The transmitted signal is a balanced QPSK signal with symbol rate of 9600 symbols/sec. Figure 2.4 shows the amplitude timecorrelation (0, ) = Jo (2 f d ) as a function of fd (which is the number of wavelengths traveled by the mobile over sec). As the signal bandwidth B increases so that T B1, the approximation s(t i ) s(t o ) for o [min i , max i ] is no longer valid. Then, the received signal is a sum
i i
of copies of the original signal, each is delayed in time by i and phase shifted by i . For wideband signals, the channel response can be approximated using Turins model [36] if the incoming paths form subpath clusters. In this model, paths that are approximately the same
24
1.0
0.8
Amplitude Correlation ()
0.6
0.4
0.2
0.0
-0.2
-0.4 0 1 2 3 4 5
Figure 2.4: Fast fading amplitude correlation vs. delay (in wavelengths traveled). length ( |i j | < B1 ) are not resolvable at the receiver. Thus, they are combined into a single path. If we assume a nite number of resolvable paths, then the received signal can be written as y(t) =
l=1 L
(2.1.28)
where L in this case represents the number of resolvable paths or subpath clusters, and l , l , and l are the amplitude, phase, and delay of each resolvable path. The complex gains l (t) = l (t)e jl (t) are independent complex Gaussian processes. Again, the equivalent lowpass channel is described by the time-variant impulse response [33]
L
h(t; ) =
l=1
( l )l (t)e jl (t)
(2.1.29)
25
As the mobile moves in the environment, the number and position of scatterers contributing to the received signal will change and, therefore, the time delay l , the amplitude fade l , the phase l , and the number of resolvable paths L will also change with time and, therefore, in addition to specifying their rst order statistics, one also needs to specify their time correlation properties. The time correlation properties for the complex path gain l (t) are also described by the correlation function in (2.1.22). As suggested by measurements [39], the time delays l are characterized by a Poisson or modied Poisson process. More details on characterizing the above parameters for mobile radio and indoor wireless channels can be found in [39-42].
is the amplitude of the wave entering this obstruction and si is the wave amplitude after the
26
obstruction, then si = si1 exp (ai ri ) Then, it follows that the signal leaving the nth obstruction is given by
n
(2.1.30)
sn = so exp
i=1
ai ri
(2.1.31)
tion, then as the number of obstructions gets large enough (n ) we can use the central limit theorem to show that x=
i=1 n
ai ri
(2.1.32)
is approximately Gaussian. Therefore, S = 10 log10 s will have a Gaussian distribution with mean s and standard deviation s . However, we should note that in a typical urban propagation scenario, only few obstructions affect the signal propagation. The value of s depends on the environment and varies from 4 to 12 dB [25, 45-47]. The correlation behavior of the slow fading process is not known in general. However, measurement data in [47] suggest that S(t) can be modeled as a rst order Markov process with autocorrelation
2 S () = s exp
v|| Xc
(2.1.33)
where v is the mobile velocity and Xc is the decorrelation distance which is a function of the surrounding obstruction sizes and layout. Values of Xc under different measurement conditions are reported in [47]. However, in a dense propagation environment such as New York city where building sizes and heights vary considerably, the shadowing process is highly non-stationary and an independent increments process model is more suitable [48].
27
we will use the so-called narrowband data model for array signal processing to obtain the model for the received signal vector at the array. This model inherently assumes that as the signal wavefront propagates across the array, the envelope of the signal remains essentially
28
Antenna 1
Antenna K
Figure 2.5: A Wireless communication system employing antenna arrays. unchanged. The term narrowband is used here since the assumption of a slowly varying signal envelope is most often satised when either the signals or the antennas have a bandwidth that is small relative to the carrier frequency f . However, this assumption can also be satised by wideband signals, as is the case in CDMA, provided that the frequency response of each antenna is approximately at over the signal bandwidth, and provided that the propagation time across the array is small compared to the inverse bandwidth of the signal B1 . Under these assumptions, we can write the complex base-band representation of a real narrowband received signal at the kth antenna as [54] xk (t) = Hk ()e jk s(t) + nk (t) (2.2.1)
where s(t) is the complex baseband transmitted signal, Hk () is the frequency response of the kth antenna, k is the propagation time delay, nk (t) is the additive white Gaussian noise (AWGN), and = 2 f is the carrier frequency. We can write (2.2.1) as x(t) = v( )s(t) + n(t) (2.2.2)
29
v ()
Array Manifold
Figure 2.6: A one-dimensional array manifold. where x(t) is the array output vector and n(t) is the additive white Gaussian noise vector. The vector v( ) C K1 is the array response vector given by v( ) = H1 ()e j1 H2 ()e j2 H K ()e j K and is a function of the parameter vector
T
(2.2.3)
the mobile in some coordinate system, the signal carrier frequency, polarization angles, etc. If there are p < K elements (different parameters) in v( ), then will trace a p-dimensional surface in C K as is varied over the parameter space. This surface is referred to as the array
A = {v( ) :
(2.2.4)
where denotes the set of all possible parameter vectors. Throughout this thesis, we will assume that we have identical antennas in the array. In this case Hk () will be the same for all antennas. Also, all antennas will have the same
30
response in any given direction. In this case, the array response vector is parameterized by the angular carrier frequency and the time delays 1 , 2 , , k , which can be shown to be a function of the mobile position with respect to each antenna. To simplify the discussion we also assume that the mobile is in the far-eld and that the mobile and the antenna array are in the same plane2 . Under the above assumptions, the parameter vector contains only the azimuth and the array manifold A is a one-dimensional rope winding through C K as illustrated in Figure 2.6. In this case, by redening the signal s(t) (by denoting H()s(t) as s(t) or by including H() within the channel response), then v() = e j1 e j2 e j K
a a a
(2.2.5)
a a where the time delays 1 , 2 , , a are now with respect to a given reference point and are K
functions of the array geometry and . For example, for a uniform linear array (ULA) of identical sensors as illustrated in Figure 2.7, taking the 1st antenna as a reference point, the
a time delay k is given by a k = (k 1)
d sin , c
k = 1 K
(2.2.6)
where is the angle between the arriving signal and the normal to the array, and d is the spacing between sensors. In this case the array response vector v() is given by 1 j2d sin / e j4d sin / v() = e . . . . j2(K1)d sin / e
(2.2.7)
2 These assumptions may be relaxed at the expense of complicating the array response vector model. In the case of near-eld and 3-D arrays, three parameters are required to dene the parameter vector which are azimuth, elevation and range.
31
Mobile
2 d
32
Mobile v
Dominant Reflector #3
Figure 2.8: An illustration of the vector channel. dominant reectors. For the vector channel, the angle of arrival of each reected wave with respect to the base station coordinate system is determined by the physical position of the dominant reector with respect to the base. An illustration of the vector multipath channel is shown in Figure 2.8. Let s(t) be the complex baseband transmitted signal. The complex baseband received signal vector at the base station antenna array can be written as
L
x(t) =
i=1
(2.2.8)
where i is the direction of the local scatterer with respect to the mobile velocity vector and i is the angle of arrival of the ith signal path which is also the angular position of the ith dominant reector (or local scatterer) with respect to the base station coordinate system. v(i ) is the K 1 array response vector for signals arriving in the ith wavefront and is dened by (2.2.5). R2 is the fraction of the incoming power in the ith path. i In general, signals arrive at the receive antennas mainly from one given direction. For
33
example, in rural or suburban areas, a high base station antenna array typically has a lineof-sight path to the mobile, with local scattering around the mobile generating signals that arrive mainly within a given range of angles or beamwidth. We assume that all signals from the mobile arrive at the base station antenna array uniformly within of arrival [50]. That is, we assume that 1 + i 2 f (i ) = 0 otherwise The value of of the mean angle
+ ,
(2.2.9)
is called the angle spread (around the mean angle of arrival). We may point
out that, although this assumption is not based on any measured data, according to [55], analytical results obtained based on this assumption were shown to be consistent with measurement data in [21]. Assuming that s(t) is a narrowband signal (T << B1 ), then s(t i ) s(t o ) where o [min i , max i ] and we can write the lowpass received signal vector as
i i L
x(t) s(t o )
i=1
v(i ) Ri e ji (t)
(2.2.10)
where i = 2( f d cos i t f i ). As before, the i modulo 2 are assumed to be i.i.d. over [0, 2]. We dene the complex channel vector a(t) as
L
a(t) =
i=1
(2.2.11)
Similar to the scalar channel case, the lowpass vector channel is described by the timevariant impulse response h(t; ) = ( o ) a(t) (2.2.12)
For a large number of incoming paths, the complex channel vector will approach a zero mean complex Gaussian random vector, that is a(t) N (0, Ra ) (2.2.13)
34
and is a function of the angular frequency , the mean angle of arrival , the angle spread , and the spacing between sensors and their geometry. Most of the adaptive antenna arrays processing techniques, which are the focus of this work, depend on the correlation of the received signals at the array. The resulting correlation matrices will play a key role in this work, making it worthwhile to characterize and study their structure, how this structure expresses and describes the propagation environment and signals propagating in that environment, and what structure is naturally present (induced by stationarity assumptions, for example). Therefore, we need to study the spacetime-frequency correlation behavior of the vector channel. In Appendix A.2, it is shown that the space-time-frequency correlation matrix of a(1 , t) and a(2 , t + ) is given by
Ra (1 , 2 , ) = E{a(1 , t)a (2 , t + )}
= Jo (d ) FT ( j ) Rs where Rs is the array spatial correlation matrix dened as 1 Rs = 2
+ +
(2.2.15) (2.2.16)
v(1 , )v (2 , )d
(2.2.17)
and v(, ) is dened in (2.2.5). Similar to Ra , Rs is a function of the angular frequencies 1 and 2 , the mean angle of arrival , the angle spread , and the spacing between sensors and their geometry. For 1 = 2 = we will write Ra as Ra (, ). In Appendix A.2, it is shown that the real and imaginary parts of Rs (m, n), the signal correlation between the mth and nth antenna elements, are given by Re{Rs (m, n)} = Jo (zmn ) + 2 Im{Rs (m, n)} = 2
l=0 l=1
(2.2.18)
35
rm mn dmn /2
Sensor m
dmn /2
r mn
Sensor n
rn
dmn
Reference Point
Figure 2.9: Model geometry. where m,n is the mean angle of arrival measured with respect to the normal to the line joining the two sensors as shown in Figure 2.9 and zm,n and m,n are dened in (A.2.13).For 1 = 2 = and = 0, we have Ra = Rs = Ra , mn = 0, and zm,n = dm,n /c where dm,n is the distance between the mth and nth sensors. Figures 2.10 and 2.11 show the spatial envelope correlation s = |Rs (m, n)| for = 0 and = 30 , respectively, and for various angle spreads (2.2.18) and (2.2.19). Figure 2.10 shows that, as based on the expressions in
tion occurs at a larger antenna spacing. Specically, the rst zero occurs for d/ 30/ . When the signal arrives from a direction other than broadside, as in Figure 2.11, the antenna spacing for low correlation increases and the envelope correlation is never zero for all values of = 0 and < 180 ( s is zero when Re{Rs (m, n)} and Im{Rs (m, n)} have zero
crossings at exactly the same spacing). The space-time array correlation matrix Ra (, ) provides a full characterization of the
36
1.2
0.8
0.6
= 0 = 5 = 15 = 60 =0
0.4
0.2
Figure 2.10: Spatial envelope correlation vs. antenna spacing: mean AOA = 0 dynamics of the channel vector a(t). First, we rewrite Ra (, ) as follows
Ra (, ) = () Rs
(2.2.20)
where () is the time correlation part, which under the assumptions stated above, was shown to be Jo (d ). As time progresses, the channel vector a(t) amplitude and direction in C K will change. While the time correlation part of Ra describes how fast a(t) changes, with the rate of change being proportional to d , the spatial correlation part Rs characterizes the complexity of change of a(t) (in terms of the dimension of the subspace spanned by a(t)) in C K . This complexity is proportional to the angle spread . By (2.2.11), a(t) is a linear combination of the path vectors v(1 ), v(2 ), , v( L ), and hence it tends to lie in the subspace spanned by them. For example, if the angle spread is small, all path vectors
37
1.2
0.8
= 0 =5 = 15 = 60
0.6
0.4
= 30
0.2
Figure 2.11: Spatial envelope correlation vs. antenna spacing: mean AOA = 30 will point approximately to the same direction. Therefore, the uctuations in a(t) will be mostly in magnitude only and for a large percentage of time it will be in a certain preferred direction. In this case Rs will have off-diagonal correlation terms that are a large fraction of the diagonal entries. In this case, we have nonspace selective fading. For a large angle spread, the path vectors will span the full space and therefore the a(t) does not exhibit any preferred direction and the changes of a(t) in C K will be in both amplitude and direction. The off-diagonal entries in Rs will become very small compared to the diagonal elements. In this case, we have space selective fading.
38
4 2
0 2 4 6 8 0 2 4
Space (d / ) 6
8 0 50 100 150
200
A measure of the complexity of a(t) variability can be obtained by considering the eigenvalue spectrum of Rs . Each eigenvalue is a measure of how strongly a(t) points in the corresponding eigendirection in C K . If the eigenvalues are all approximately equal, then a(t) spans the full space. If only few of the eigenvalues are large compared to the others, the variability of a(t) will tend to be conned to the subspace spanned by the corresponding eigendirections. To demonstrate the effect of angle spread and Doppler spread on the channel vector dynamics, the above channel model was simulated. The mobile was assumed to be moving at 37.5 mph which corresponds to a maximum Doppler shift of 50 Hz at 900 MHz carrier frequency. The transmitted signal is a balanced QPSK signal with a symbol rate of 4800
39
10 0 2 4 6
Space (d / )
8 0 50
100
150
200
symbols/sec. We considered a base station with a ULA with aperture size 8. Figures 2.12, 2.13, and 2.14 plot the received signal level across the array over a period of 200 symbols for angle spreads of 0 , 3 , and 40 respectively. For the zero angle spread case in Figure 2.12, we notice that the signal level is the same across the array. This is because, with zero angle spread, all paths contributing to the received signal at the array arrive from the same direction and have the same relative phase at different points across the array and therefore will add up (either constructively or destructively) in the same way. In this case, we have timeselective fading only. For angle spread = 3 in Figure 2.13, we notice that, for any given
instant of time, the signal level across the array will vary. In this case, different paths contributing to the received signal arrive from different directions and their relative phases will
40
0 5 10 15 20 0 2 4 6
Space (d / )
8 0 50
100
150
200
be different, and, therefore they will add up either constructively or destructively at each point across the array depending on the relative phase relationship. As increases, as in
Figure 2.14, the signal level across the array will change more rapidly. In this case, we have space-time selective fading. As in the scalar channel case, for wideband signals, the vector channel response can be approximated using Turins subpath model described in the previous Section. In this case, the equivalent lowpass vector channel is described by the time-variant impulse response
L
h(t; ) =
l=1
( l )al (t)
(2.2.21)
where L is the number of resolvable paths or subpath clusters. The lth channel vector al (t),
41
as dened in (2.2.11), is itself a linear combination of a large number of path vectors that are approximately the same length. Each channel vector al (t) is characterized by its mean angle of arrival l and angle spread Rs (
l , l ). l
Under the same assumptions above, we can show that the channel vectors for
l , l )
if l = k, otherwise
(2.2.22)
and
, R ),
2 l N (s , s ) l = 1 L
(2.2.23) (2.2.24)
= [1 2 L ]
and s is the area mean and s varies between 4-12 dB depending on the degree of shadowing. In general, the average power output of the channel is not constant with time delay. The average propagation loss increases with time delay resulting in a reduction in average path power as path delay l increases. The effect is accounted for by the deterministic power delay prole p(l ). Therefore, we rewrite the time-variant impulse response of the vector channel as h(t; ) =
l=1 L l
( l )al (t)
(2.2.25)
where
= Sl p(l ).
l
For the vector channel impulse response dened above, we have al (t) and
modeled
as random variables. As such we dene the power delay prole of the vector channel P ()
42
as
(2.2.26) (2.2.27)
=
l=1
K l ( l )
where K = E{a (t)al (t)} and l = E{ l }. Here, we used the fact that al (t) and ak (t) are l independent for l = k. By estimating P () for any given channel, one can estimate the rms delay spread of the channel rms as follows. If we model P () as a probability distribution by normalizing it with
0
P ()d P ()d
P ()d P ()d
2
(2.2.28)
L 2 l=1 2 l l L l=1 l
L l=1 2 l l L l=1 l
(2.2.29)
Analysis of measurement data [5] shows that for some propagation environments, P () can be approximated by a decaying exponential of the form P0 e/rms where P0 is the total output power of the channel.
43
44
100-120 s. The number of peaks in the impulse response will depend on the specic terrain. The fast fading is well described by the Nakagami fading model with m ranging between 1 and 15. For rural or open areas, very few buildings or scatterers exist and therefore we have either low or no shadowing at all. The delay spread is on the order of 0.1-0.3 s. The impulse response in this case typically consists of one peak. The fast fading is also well described by the Nakagami fading model with m up to 100.
45
micro cells. If a LOS path is present, the measured impulse response consists of a strong rst path and few smaller components due to wall reections within 50-100 ns. In non-LOS propagation, the measured impulse response will include many paths that are incident from all directions and within 100-150 ns.
2.4 Summary
The rst half of Chapter 2 is a brief overview of radio channel modeling and describes some problems associated with a wireless radio channel. The models described therein are for scalar channels. In the second half of Chapter 2 we develop a statistical vector channel model based on the physical propagation environment and its statistical properties. Such a model is necessary in order to analyze and design adaptive beamforming techniques with antenna arrays at the base station that will be introduced in the next chapter.
47
x1 (t)
sensor 1
w1 w2
x2 (t)
sensor 2
y(t)
array output
xK(t)
sensor K
wK
Mobile
Available Informations
Figure 3.1: Block diagram of an adaptive array system Several algorithms have been proposed in the array signal processing literature to design adaptive beamformers to separate multiple co-channel signals based on the availability of prior spatial or temporal information. The traditional spatial algorithms combine high resolution direction-nding (DF) techniques such as MUSIC, ESPRIT, and WSF (weighted subspace tting) [68-71] with optimum beamforming to estimate the signal waveforms [66, 72]. However, these techniques require certain assumptions on the number of signal wavefronts arriving at the base station and on the complexity of the propagation environment, which restricts their applicability in a wireless mobile communications setting. Other techniques use reference or training signals to nd the optimum adaptive beamformer [73-75]. In the recent past, several property-restoral techniques have been developed that exploit the temporal or spectral structure of communication signals, while assuming no prior spatial information. These techniques take advantage of signal properties such as constant modulus (CM)[76-78, 67], discrete alphabet [79-82], self-coherence [83], and high order statistical properties [84].
48
In this chapter we will review some of the previous techniques for adaptive beamforming. Then, we will derive a novel technique for estimating the vector channel and the corresponding adaptive beamformer for CDMA wireless systems. In this technique, we perform code-ltering at each antenna for each user in the system. We exploit the eigenstructure of the pre- and post-correlation array covariance matrices to estimate the channel vector and derive the corresponding adaptive beamformer. We also extend this technique to the case where we have multipath propagation. The resulting overall receiver structure is called Beamformer-RAKE. Our approach is a blind technique in the sense that it does not require any training signals, although it assumes the perfect knowledge of the spreading code for each user. Also, it does not require any assumptions on the signal propagation and is therefore suitable for different propagation settings.
(3.1.1)
si (t)ai
(3.1.2)
represents the composite of the N cochannel interferers and n(t) = nc (t) + jns (t) is the additive noise vector, which is modeled as a complex Gaussian random vector with zero
49
(3.1.3)
2 where I is the K K identity matrix and n is the antenna noise variance. Equation 3.1.3
implies that the noise is spatially white. The noise vectors nc (t) and ns (t) are both lowpass white Gaussian random processes. In addition, the noise vector n(t) is assumed to be temporally white, i.e
2 E{n(t1 )n (t2 )} = n I(t1 t2 )
(3.1.4)
This modeling of the noise vector does not take into account the effect of low pass ltering. This ltering will be considered in the performance analysis results presented in Chapter 6. Let wo represent the desired weight vector for linearly combining the array outputs. From Figure 3.1, the output of the beamformer y(t) is given by y(t) = w x(t) o (3.1.5)
where (.) denotes the Hermitian transpose. As we mentioned above, there are several techniques to estimate wo . A brief review of some of these techniques follows.
contribute to the received signal will essentially arrive from the same direction. In such case the channel vector as dened in (2.2.11) becomes
L
ai v(i )
l=1
(3.1.6)
where i is the angular position of the ith source with respect the base station. In this case, we rewrite the received signal vector x(t) as
N
x(t) = so (t)v(o ) +
i=1
si (t)v(i ) + n(t)
(3.1.7)
50
where si (t) = i (t) si (t). Let i2 = E{si (t)s (t)}. A reasonable assumption to make is that i the signals so (t), s1 (t), , s N (t) are uncorrelated. All DF-based techniques use the array output x(t) and the knowledge of the array manifold A to get an estimate of the directions of arrival (DOA) o , 1 , , N and the corresponding estimates of the array response vectors v(o ), v(1 ), , v( N ). A number of high resolution techniques such as MUSIC, ESPRIT, and WSF can be used to estimate the DOAs. A reasonable strategy is to nd the best weight vector to optimally combine the array outputs under some suitable criterion. Minimization of mean squared error (MMSE), maximization of signal to interference-plus-noise ratio (SINR), and maximum likelihood (ML) have been widely used as optimization criteria, and in all these cases the optimal weight vector turns out to be a function of signal strengths of the desired and undesired signals, their directions of arrival, and their covariances. Before we proceed to derive the optimal weight vector under any of those criteria, we rst dene some of the statistical quantities for the signal vector model in (3.1.7). Let u(t) = i(t) + n(t) be the total undesired signal vector. We dene the array covariance Rxx and the undesired signal vector covariance Ruu as Rxx = E{x(t)x (t)} = Ruu = E{u(t)u (t)} = Maximization of SINR Beamformer The weight vector wo can be optimized by maximizing the SINR at the output of the beamformer. First, we rewrite (3.1.7) as x(t) = so (t)v(o ) + u(t) This gives the output SINR to be (SI N R)o =
2 E{|w so (t)v(o )|2 } o |w v(o )|2 o o = E{|w u(t)|2 } w Ruu wo o o i=o N i=1 N 2 i2 v(i )v (i ) + n I
(3.1.8) (3.1.9)
2 i2 v(i )v (i ) + n I
(3.1.10)
(3.1.11)
51
= (SI N R)max with the equality achieved for w S I N R = R1 v(o ) uu where is any nonzero complex constant.
(3.1.15)
Maximum Likelihood Beamformer The derivation of the maximum likelihood optimum beamformer requires the assumption that the composite of the interferers i(t) is Gaussian so the total undesired signal vector u(t) = i(t) + n(t) will have a multivariate Gaussian distribution. We dene the likelihood function of the input signal vector as
L (x(t)) = fU (x(t)) =
(3.1.16)
The ML optimum beamformer is obtained by solving for the estimate of s(t) that maximizes (3.1.16) or equivalently its logarithm. Taking the partial derivative of log (L (x(t)) and setting the result to zero yields 0= log (L (x(t)) = 2v (o )R1 x(t) + 2s(t)v (o )R1 v(o ) uu uu s(t) R1 v(o ) x(t) uu s(t) = = w L x(t) M 1 v( ) v (o )Ruu o Thus, the ML optimum weight vector has the form wM L = R1 v(o ) uu v (o )R1 v(o ) uu (3.1.19)
(3.1.17)
It immediately follows that the estimate of s(t) that maximizes log (L (x(t)) is given by (3.1.18)
52
Comparing (3.1.19) with (3.1.15), we notice that the ML beamformer also maximizes the SINR. MMSE Beamformer The MSE beamformer minimizes the error e(t) between the beamformer output w x(t) o and the desired signal s(t). Thus e(t) = s(t) w x(t) o and the mean squared error (wo ) is given by
2 (wo ) = E{|e(t)|2 } = o Re{2w rxs w Rxx wo } o o
(3.1.20)
(3.1.21)
where rxs = E{x(t)so (t)}. Taking the partial derivative of (wo ) with respect to wo and set ting the result to zero yields 0= or w M S E = R1 rxs xx (3.1.23) (wo ) = 2(Rxx wo rxs ) wo (3.1.22)
This is often referred to as the optimum Wiener solution [85]. With the assumption that
2 the signals so (t), s1 (t), , s N (t) are uncorrelated, rxs = o v(o ) and the use of the matrix
(3.1.24)
In this case, we see that the MMSE beamformer wM S E is a scalar multiple of the SINR beamformer wS I N R in (3.1.15). This is only natural, since for Gaussian measurements, the MMSE and the ML criterion lead to the same estimator [87]. Given S snapshots of the received signal vector at the array x(1), x(2), , x(S) and the estimates of the array response vectors V( ) = [v(o ), v(1 ), , v( N )], it was shown in
53
[72] that the maximum likelihood (ML) estimates of Rxx and Ruu are given by 1 Rxx = S
S l=1
xl x l
(3.1.25) (3.1.26)
where o is the ML estimate of the desired signal power and is given by the (1,1) element 2 of Rss = V ( )(Rxx n I)V ( ) 2 and V ( ) = V ( )V( )
1
(3.1.27)
V ( )
(3.1.28)
is the pseduo-inverse of V( ). Furthermore, n is the ML estimate of the noise variance 2 and is given by n = 2 where P = I V( ) V ( )V( ) V 1 Tr P Rxx V K N 1
1
(3.1.29)
of V( ) and Tr(.) denotes the trace operator. Based on the above covariance and array response vector estimates, the optimum weight vector estimates according to the different performance measures discussed above are summarized in Table 3.1. While DF-based beamforming techniques are analytically more tractable, they suffer from several drawbacks that may limit their applicability in a wireless setting. First, all techniques start with a DOA estimation step that involves an eigen-decomposition and one or more multidimensional, non-linear optimizations, which may be a difcult and time consuming task. Also, the DOA step requires knowledge of the array manifold and is very sensitive to errors in this knowledge. The presence of multipath in urban or suburban environments, where the condition of zero or relatively small angle spread does not hold,
makes knowledge of the array manifold unreliable. Moreover, a key assumption in all DOA
54
ML
MMSE
o 2 uu R1 v(o ) ( )R1 v( ) 1 + v o uu o
Table 3.1: Optimum weight vector for SINR, ML, and MMSE performance measures estimation techniques is that the number of signal wavefronts including cochannel interference signals must be less than the number of antennas in the array, a fact that restricts their applicability in a wireless setting.
55
where rxd = E{x(t)d(t)}. Therefore, given the snapshots of the received signal at the array x(1), x(2), , x(S) and the corresponding reference signal d(1), d(2), , d(S), the estimate of the beamformer weight vector is xx w M S E = R1 rxd where Rxx is given by (3.1.25) and rxd is similarly dened as rxd = 1 S
S
(3.1.31)
xl d(l)
l=1
(3.1.32)
The weight vector estimate can be calculated by a number of different techniques such as LMS [88] or Direct Matrix Inversion (DMI) [85]. The technique of training-signals based beamforming was studied for use with the digital mobile radio system IS-54 in [74]. In this approach, the base station uses the synchronization sequence in each block as a reference signal and uses DMI as in (3.1.31) for weight acquisition in each block. The use of training signals requires prior carrier and symbol recovery, which is made difcult by the presence of co-channel interference. In addition, sending a training signal along with the information signal reduces the payload efciency of the system. Also, training-signals based beamforming techniques are not applicable in systems where training or synchronization signals are not available such as in the IS-95 CDMA standard.
56
FM Signal 1 s1 (t)
Beamformer w
^ s1 (t)
s2 (t) FM Signal 2
property restored
Figure 3.2: Constant modulus property restoral beamformer that in the CMA algorithms is the nite alphabet (FA) property of digital signals. In addition to the low modulus variation, this approach uses the digital modulation and channel coding structure, and the base station processor attempts to construct the beamformer by tting the FA model to the underlying data model described in (3.1.1). Another class of signalstructure based beamforming is based on spectral self-coherence. Most communication signals are correlated with frequency-shifted and possibly conjugated versions of themselves for certain discrete values of frequency shift. This spectral self-coherence is commonly induced by periodic gating, mixing, and multiplexing operations at the transmitter. As an example of the property-restoral beamforming techniques, a brief description of the CMA-based beamforming follows. Because of multipath propagation and the presence of interference, the received signal at the array output will exhibit amplitude uctuation although the transmitted signal has a constant envelope. As shown in Figure 3.2, the objective of the CMA beamformer is to restore the array output y(t) to a constant envelope signal on the average. This can be accomplished by choosing the weight vector w to minimize a cost function J(y(t)) that provides a measure of the amplitude uctuation. An example of such
57
a cost function is the mean squared amplitude uctuation (t) = E y(t) y(t) |y(t)|
2
(3.1.33)
Other cost functions may be used. Note that the noise tolerance and the convergence behavior of the CMA will depend on the choice of the cost function. The update equation of the weight vector w is [76, 77] w(t + 1) = w(t) e(t) x(t) y(t) e(t) = y(t) |y(t)| (3.1.34) (3.1.35)
where is the step constant. Equation (3.1.34) is very similar to the LMS algorithm. However, the CMA beamformer does not require any reference signal and uses the array output y(t) itself to generate the reference signal. The signal-structure based beamforming techniques are very robust against different propagation conditions. No knowledge of the array manifold or DOAs is required. However, a drawback that limits the applicability of signal structure-based beamforming methods for wireless applications is that the convergence and capture characteristics are still not well understood.
58
The different beamforming techniques discussed in the previous section are not suitable for CDMA systems for the following reasons. First, as we mentioned above, all users in a CDMA wireless system are cochannel and their number may easily exceed the number of antennas. Moreover, due to multipath propagation and the fact that each transmission path may contain direct, reected and diffracted paths at different time delays, the array manifold may be poorly dened. Therefore DF-based beamforming techniques are not applicable. Also, no training or reference signals are present in the mobile to base link [8]. Therefore, reference-signal based techniques cannot be used. In this section we propose a new spacetime processing framework for adaptive beamforming with antenna arrays in CDMA. This approach may be seen as a signal-structure based technique since, as will be seen later, it exploits both temporal and spatial structure of the received signal
where Pi is the ith users transmitted power per dimension (i.e. the total transmitted power is 2Pi ), c is the angular carrier frequency, bi (t) is the differentially encoded bit of duration Tb , and ciI (t) and ciQ (t) are the in-phase and quadrature spreading codes used by the ith user,
59
Information bits
differential encoder
bi (t)
a (I)
cos c t
si (t)
a (Q)
sin c t
Figure 3.3: Mobile transmitter block diagram respectively. In our analysis, the spreading code ciI (t) is represented by ciI (t) =
h= I ci,h p(t hTc )
(3.2.2)
where ci,h are assumed to be i.i.d. random variables taking values 1 with equal probability, Tc is the chip duration, and p(t) is the chip pulse shape. p(t) can be any time-limited waveform. Here we assume that p(t) is rectangular although our results can be easily extended for any time-limited waveform. The processing gain is dened as G = Tb / Tc . We assume the multipath vector channel model dened in (2.2.21). We also assume that the channel parameters vary slowly as compared to the bit duration Tb , so that they are constant over several bit durations. After down-converting to baseband, we can write the complex baseband received signal vector from the i-th user at the K element antenna array as xi (t) =
l=1 L
(3.2.3)
where
Pl,i is the average power per dimension from the ith user in the lth multipath com-
ponent, ci (t) = ciI (t) + jciQ (t), l,i = c l,i, and L is the number of multipath components. The total received signal at the cell site is the sum of all the users signals plus noise and is given by
N L
x(t) =
i=1 l=1
(3.2.4)
60
(3.2.5)
Pi bi (t i )ci (t i )e ji ai
(3.2.6)
is the composite of multiple access interference (MAI). In our approach, the antenna outputs are correlated with c1 (t 1 ) to yield one sample vector per bit. The post-correlation signal vector for the nth bit is given by 1 y1 (n) = Tb
nT+1 (n1)T+1
x(t)c (t 1 )dt 1
(3.2.7) (3.2.8)
= 2 Tb P1 b1 (n)e j1 a1 + i1 + n1 where 1 n1 = Tb
nT+1 (n1)T+1
n(t)c (t 1 ) dt 1
(3.2.9)
61
i1 =
i=2
Pi I1,i e ji ai
(3.2.10)
is the undesired component due to the multiple access interference, where I1,i is dened as 1 I1,i = Tb
nT+1 (n1)T+1
bi (t i )ci (t i )c (t 1 ) dt 1
(3.2.11)
In Chapter 6, we show that n1 is a zero mean complex Gaussian vector with covariance
2 2n Tc I and that I1,i , i = 2 N are i.i.d with zero mean and variance 4Tc 1 . Using (3.2.5)
and (3.2.6), we can show that the short term covariance of the pre-correlation signal vector x(t) is given by Rxx = E{x(t)x (t)} = 2 P1 a1 a + 1
N i=2 2 2Pi ai a + n I i
(3.2.12)
Similarly, using (3.2.8), (3.2.9), and (3.2.10), we can show that the short term covariance of the post-correlation signal vector y1 (n) is given by 1 R yy,1 = E{y1 (n)y (n)} = 2G P1 a1 a + 1 1 2Tc
N i=2 2 2Pi ai a + n I i
(3.2.13)
Let u1 = i1 + n1 be the total undesired component due to thermal noise and multiple access interference. Then, the short term interference plus noise covariance is Ruu,1 = E{u1 u } 1
N
=
i=2
2 2Pi ai a + n I i
(3.2.14)
Note that we can rewrite both Rxx and R yy,1 as Rxx = 2P1 a1 a + Ruu,1 1 R yy,1 = 2G P1a1 a + Ruu,1 1 = 2(G 1)P1 a1 a + Rxx 1
1
In this analysis we included the effect of ideal lowpass ltering after down conversion to baseband.
62
We can easily see there is a processing gain, equal to G, in power for the desired user as compared to the interference plus noise. Assuming we have adequate data to obtain near asymptotic covariance estimates Rxx and R yy,1, we can estimate the channel vector a1 as the generalized eigenvector corresponding to the largest eigenvalue of the Hermitian-denite matrix pencil Ryy,1 Rxx [90]. In particular, we solve the following generalized eigenvalue problem to estimate a1 : R yy,1 a1 = Rxx a1 Similarly, the interference-plus-noise covariance Ruu,1 can be estimated as Ruu,1 = G 1 Rxx R yy,1 G1 G (3.2.19) (3.2.18)
In the previous section, it was shown that the beamformer that will maximize the SINR has the form w1 = R1 a1 uu,1 (3.2.20)
Since is arbitrary and does not affect the SINR, we will set = 1. Therefore, given the weight vector w1 , the beamformer output corresponding to the nth bit is z1 (n) = w y1 (n) = 2 Tb P1 b1 (n)e j1 w a1 + w i1 + w n1 1 1 1 1 and the corresponding signal to interference-plus-noise ratio (SINR) 1 is 1 = 2G P1 a R1 a1 1 uu (3.2.22) (3.2.21)
A more computationally attractive approach to nd the weight vector w1 that will maximize the SINR is to solve for w1 directly (i.e. no intermediate step to estimate a1 ). This can be done as follows. First, we rewrite the post-correlation signal vector y1 as y1 (n) = s1 (n) + i1 + n1 (3.2.23)
where s1 (n) = 2 Tb P1 e j1 a1 . We can write the SINR at the output of the beamformer as 1 (w) = w Rss,1w w Ruu,1w (3.2.24)
63
LPF
c1 (t- 1 ) w1*
j t
LPF
c1 (t- 1 ) w2*
w* y1
j t
Beamformer output
LPF
c1 (t- 1 ) wK*
j t
x
Ruu= G-1 Rxx - G-1 Ryy
G 1
y1
Figure 3.4: CDMA receiver-beamformer block diagram where Rss,1 is the covariance of s1 (n) signal component dened as Rss = 1 E{s1 (n)s (n)} 1 2Tc (3.2.25)
The goal is to nd w1 that will maximize 1 (w), that is w1 = max 1 (w) = max
w w
w Rss,1w w Ruu,1w
(3.2.26)
w Rss,1w1 1 w Ruu,1w1 1
(3.2.27)
(3.2.28)
64
It is clear from (3.2.28) that the pair (, w1 ) is the solution corresponding to the largest eigenvalue of the generalized eigenvalue problem Rss,1w = Ruu,1w Now, we note that R yy,1 = Rss,1 + Ruu,1 and the problem (Rss,1 + Ruu,1 )w = Ruu,1w is equivalent to Rss,1w = ( 1)Ruu,1w (3.2.32) (3.2.31) (3.2.30) (3.2.29)
Therefore, we can nd the weight vector w1 directly by solving for the eigenvector corresponding to the largest eigenvalue of the generalized eigenvalue problem R yy,1 w = Ruu,1w (3.2.33)
The proposed code-ltering approach for adaptive beamforming can be summarized as follows Perform code ltering at each array element for each user in the cell. Given the pre- and post-correlation vectors samples x(1), x(2), , x(S) and y(1), y(2), , y(S), estimate the array covariance matrices Rxx and R yy,1, assuming that we have enough samples (at least 2K uncorrelated snapshots [91]) to obtain quasiasymptotic covariance estimates. Estimate the channel vector a1 as the generalized eigenvector corresponding to the largest generalized eigenvalue of the matrix pair (Ryy,1 , Rxx ). Estimate the interference-plus-noise covariance Ruu,1 according to (3.2.19). The optimum weight vector can be found according to (3.2.20).
65
input
decision
channel sounder
66
input
1
~* c (t)
2
~* c (t)
Combiner
decision
L
~* c (t)
Figure 3.6: Multipath CDMA RAKE receiver The RAKE receiver can be simplied by implementing only a small number of taps in the chip-rate matched lter, as shown in Figure 3.6. A scanning correlator identies the delays at which the most signicant peaks occur, and only at those time delays is the received signal despread and combined. The performance of this structure is optimum for the case of discrete multipath when the number of paths equals to the number of correlators. In the case of a diffuse multipath, this structure performs signicantly worse than the RAKE receiver in Figure 3.5, because then a signicant fraction of the transmitted energy is lost unless a large number of correlators is used [96]. Thus, in a conventional multipath receiver for CDMA, temporal code ltering is used to resolve paths that are separated in time by more than a chip interval Tc and to estimate the multipath structure of the channel (in terms of the time delay and the complex path strength of each resolvable path). Then, the signal is passed through a RAKE correlator that is matched to the channel output and that coherently/incoherently combines correlators outputs. In this way the temporal structure of the multipath received signal is exploited
67
efciently. We refer to this as a 1-D or temporal RAKE since only temporal structure is exploited. In general, different multipath components arrive at the receiver not only with different time delays, but also from different directions and therefore each multipath component has a different channel vector al,1 . A single antenna receiver can not exploit the spatial structure of the received signal. With antenna arrays, the spatial dimension can be used to efciently resolve and combine different multipath components. A new approach that we describe below is to identify both the temporal and spatial structure of the individual paths arriving at the receiver and then construct a space-time receiver that we call a Beamformer-RAKE. This receiver is matched to the desired signal while maximally rejecting the interfering signals.
(3.3.1)
In order to derive the optimal space-time matched lter for the received signal, we consider the likelihood function of the received signal, conditioned on the knowledge of all param eters. First, let sl (t) = Pl b(t l )c(t l )e jl al . The likelihood function of the received multipath signal can be written via the Cameron-Martin formula [97] as
2 (b(t))/n
(3.3.2)
(b(t)) is dened as
L
l=1
s (t)x(t)dt l
L l=1
sl (t) dt
(3.3.3)
the likelihood function. Assuming that the user sends M information bits, the rst integral
68
yields
L
l=1
s (t)x(t)dt = l =
L l=1 M
Pl e jl
L
b(n)
n=1 l=1
Pl e jl zl (n)
where z1 (n), z2 (n), , z L (n) are the matched lter outputs synchronously sampled with respect to each path signal and zl (n) = The second integral in
nTb +l (n1)Tb +l
c (t l )a x(t) dt , l
l = 1, , L
(3.3.6)
Therefore, the sufcient statistic for the detection of bits b(n), n = 1 M, is zl (n), l = 1 L, obtained by a linear operation on the received signal vector x(t). Equation (3.3.6) means that the sufcient statistic is obtained by passing the noisy received signal vector through a spatial matched lter A = [a1 a2 a L ] , which is equivalent to conventional beamforming [91], followed by a bank of L matched lters. The lth matched lter is matched to the code waveform in the lth path c(t l ). In general, we can use a more selective spatial matched lter. This is important if the noise is not spatially white and if interference is present. In this case, the lth sufcient statistic is given by zl (n) =
nTb +l (n1)Tb +l
c (t l )w x(t) dt , l
l = 1, , L
(3.3.7)
where the spatial matched lter, or the beamformer weight vector, wl for the lth path is chosen so that w x(t) is a good estimate of the signal received in the lth path. In the previous l section, we showed that wl is given by wl = R1 al uu,l where Ruu,l is the covariance matrix of the sum of all of the undesired signals. (3.3.8)
69
x(t)
beamformer * w1
~* c (t- 1 )
z1(n)
x(t)
beamformer * w2
~* c (t- 2 )
z2(n)
x(t)
beamformer * wL
~* c (t-L )
zL(n)
70
Spatial
Temporal
RAKE Combiner
Beamformer
Mobile "Beamformer-RAKE"
Figure 3.8: Beamformer-RAKE receiver structure with additive white Gaussian noise, with orthogonal codes or codes with low crosscorrelations they can be used in CDMA multiple access systems [8, 98]. However, in this case the receiver structure, although simple to implement, is not optimum and other structures based on multiuser detection have been proposed [99-102]. In order to construct the Beamformer-RAKE receiver, we use the code ltering approach described in the previous section for each resolvable multipath component as follows. First, we recall the total received signal vector as
N L
x(t) =
i=1 l=1
(3.3.9)
Without loss of generality, we consider the 1st user and assume that the time delays l,1 , l = 1 L are perfectly known. Later in this section, we will show how to use the code ltering approach to estimate the time delays l,1. For the nth bit, the post-correlation signal vector for the lth multipath component is given by 1 yl,1 (n) = Tb
nT+l,1 (n1)T+l,1
(3.3.10) (3.3.11)
71
where
nT+l,1 1 nl,1 = n(t)c (t l,1 ) dt 1 Tb (n1)T+l,1 is the undesired component due to thermal noise and L
(3.3.12)
il,1 =
k=1 k=l
Pk,1 Il,1,k,1e
jk,1
ak,1 +
i=2 k=1
(3.3.13)
is the undesired component due to multiple access interference (MAI) plus self interference (SI), where Il,1,k,i is dened as 1 Il,1,k,i = Tb
nT+1 (n1)T+1
(3.3.14)
Let ul,1 = il,1 + nl,1 be the total undesired (all array signals other than 1st users lth path signal) component. We can easily show that Rxx , R yy,l,1, and Ruu,l,1 are given by Rxx = 2Pl,1al,1a + Ruu,l,1 l,1 R yy,l,1 = 2G Pl,1al,1a + Ruu,l,1 l,1
L N L
(3.3.15) (3.3.16)
2 2Pl,ial,ia + n I l,i
Ruu,l,1 =
k=1 k=l
2Pk,1ak,1a + k,1
(3.3.17)
i=2 l=1
Hence, we can use the same approach developed in the previous section to nd wl,1 as wl,1 = R1 al,1 uu,l,1 (3.3.18)
Therefore, the corresponding beamformer output, or decision statistic, for the lth path is zl,1 (n) = w yl,1 (n) = 2 Tb Pl,1 b1 (n)e jl,1 w al,1 + w ul,1 l,1 l,1 l,1 and the corresponding path signal to interference-plus-noise ratio is l,1 = 2G Pl,1 a R1 al,1 l,1 uu,l,1 (3.3.20) (3.3.19)
It remains now to show how to estimate the time delays l,1 , l = 1 L. We recall that in order to nd the channel vector al,1 we solve for the principal eigenvector of the matrix pencil R yy,l,1al,1 Rxx al,1 and the corresponding eigenvalue max is a good estimate
72
for the average signal power Pl,1 in that path. Based on this observation, we can estimate the path amplitudes and time delays by computing the maximum eigenvalue max () of the matrix pencil R yy,1 ()al,1 Rxxal,1 as changes over the range of possible time delays [min , max ]. The values of that correspond to the local maxima of max () are good estimates of the time delays l,1 , l = 1 L. The value of each of those local maxima is also a good estimate of the corresponding average path strength. In practice, each local maximum is compared to a preset threshold to determine whether this local maximum corresponds to a signal path or not and also to determine the number of existing signal paths L that will be used in the subsequent detection stage. The threshold level needs to be optimized in order to maximize system performance [94]. If the threshold level is smaller than the optimum level, then some time delays corresponding to noise-plus-interference only will be taken as time delays corresponding to resovable paths which will lead to performance degradation [33]. On the other hand, if the threshold level is larger than the optimum level, the performance will also degrade as thresholding becomes more likely to reject some time delays corresponding to resolvable paths as well as time delays corresponding to noise-plus-interference only.
73
18.7 dB. Also, the channel vector al,1 is 1 j2d sin l,1 / e al,1 = l,1 e j4d sin l,1 / j6d sin l,1 / e e j8d sin l,1 /
(3.3.21)
where l,1 is the average path strength. Although this simulation scenario is simple and does not fully simulate the vector channel model described in Chapter 2, it enables us to see the ability of the approach to estimate different parameters of the multipath vector channel accurately where other approaches would fail. Path # l,1 l,1 l,1 1 90 0 1 2 45 4Tc 0.65 3 135 8Tc 0.8
89.43 1.28 42.71 1.88 137.88 2.35 1 0.66 0.029 0.81 0.027 Table 3.2: Estimated multipath parameters
For the rst user, for each multipath component we considered the values of the angle of arrival l,1 , relative time delay l,1 , and average path strength l,1 as shown in Table 3.2. The pre and post-correlation signal vectors x and yl,1 corresponding to 10 transmitted bits were used to estimate the array covariances Rxx and R yy,l,1. Using Rxx and R yy,l,1, we estimated the channel vector al,1 . From this estimate, the angle of arrival l,1 can be estimated as l,1 = sin1 al,1 (2) 2d (3.3.22)
where al,1 (2) is the second element in al,1 . The mean and standard deviation of the estimated angles of arrival over 500 runs are also shown in Table 3.2. The Table also shows
74
0.6
0.4
0.2
Figure 3.9: Estimated multipath prole the mean and standard deviation of the estimated average path strength. These numbers show that with the code-ltering approach, we are able to estimate the channel vector and the multipath parameters accurately. Note that under this scenario with 25 3 different arriving paths, estimating different parameters for each arriving signal would be impossible if we use any of the above mentioned DF techniques. In addition, the multipath delay prole for the 1st user was estimated using the approach described above. The estimated prole is shown in Figure 3.9. As we can see, the peaks of the estimated prole occur at values of that correspond to the actual delays.
75
x(t)
beamformer * w1
~* c (t- 1 )
DPSK demod.
x(t)
beamformer * w2
~* c (t- 2 )
DPSK demod.
decision device
x(t)
beamformer * wL
~* c (t-L )
DPSK demod.
76
(3.4.1)
b =
l=1
l,1 = 2G
l=1
(3.4.2)
and Cl =
1 L1l 2L 1 k l! k=0
(3.4.3)
The probability of error derived above is a conditional probability since it depends on b . Therefore, to get the average bit error probability, we need to average this conditional probability over the statistics of b . First, let us assume that the average signal power per path is the same for all paths. That is, we assume Pl,1 = P, l = 1 L. If the total undesired signal vector ul,1 can be modeled as a spatially white Gaussian random vector2 , then Ruu,l,1 = 2 I where 2 is the total interference-plus-noise power. In this case we have 2G P b = 2
L
|al,1 |2
l=1
(3.4.4)
In Chapter 6 we show that for zero angle spread, i.e. function (pdf) with 2L degrees of freedom f b () =
L1 e/ K ( K ) L (L 1)!
(3.4.5)
where = 2G P/2 is the average SINR per path per antenna and K is the number of an tennas in the array. In this case, we can easily show that the average bit error probability Pe is given by Pe =
0
2
Pb () f b () d
(3.4.6)
Simulation results in Chapter 6 show that this assumption holds if the code length G is large and if the total number of path signals N L is large, with a uniform angle of arrival distribution.
77
10-3
10-4
10-5
10-6
= 0
Pe =
l=0
Cl 2L1 (L 1)!( K ) L 2
e
0 L1
K+1 K
L1+l d K 1 + K
l
(3.4.7) (3.4.8)
1 22L1 (L 1)!(1 + K ) L
Cl (L + l 1)!
l=0
Also, in Chapter 6 we show that for large angle spread b has a 2 pdf with 2K L degrees of freedom K L1 f b () = e/ K L (K L 1)! ( )
(3.4.9)
Therefore, we can easily show that in this case the average bit error probability Pe is
78
10-1 10-2 10-3 Pe , Probability of Bit Error 10-4 10-5 10-6 10-7 10-8 10-9 10-10 5 10 15 20 25 30 L, Average Antenna SINR/bit, dB K =1 K=2 K=4
Figure 3.12: Pe for balanced DQPSK with incoherent combining and large given by Pe = =
l=0 0 L1
Pb () f b () d
(3.4.10) e
0 L1 l=0
+1
Cl 2L1 (K L 1)!( ) K L 2
K L1+l d
l
(3.4.11) (3.4.12)
1 = 2L1 2 (K L 1)!(1 + ) K L
Cl (K L + l 1)! 1+
The probability of bit error performance of the Beamformer-RAKE receiver with balanced DQPSK and incoherent combining is shown in Figures 3.11 and 3.12 for angle spread = 0 and large respectively. The probability of bit error Pe in (3.4.8) and (3.4.12) is plotted as a function of the average antenna SINR/bit, dened as L, for K = 1, 2, and 4.
79
x(t)
beamformer * w1
~* c (t- 1 )
P1 e
-j1
decision device
x(t)
beamformer * w2
~* c (t- 2 )
P2 e
-j2
x(t)
beamformer * wL
~* c (t-L )
PL e
-jL
Figure 3.13: Balanced QPSK Beamformer-RAKE receiver with coherent combining We assumed that L = 2. For the zero angle spread case in Figure 3.11, we notice the reduction in the required antenna SINR to achieve certain probability of bit error performance as the number of antennas increases. This reduction is proportional to the number of sensors and is due to the beamforming operation only. This is due to the fact that with zero angle spread, the same path signal will undergo the same fading at all antennas and, therefore, the array will not provide any space diversity. For the case of large angle spread, the signal fading at different antennas (that are sufciently separated) will be independent. Therefore, in addition to the beamforming gain, the array will provide a space diversity gain. The gain due to space diversity can be seen from the Pe plot in Figure 3.12. As we can see, with 2 antennas, an additional gain of 5.5 dB due to space diversity can be obtained. Figure 3.13 shows a Beamformer-RAKE receiver structure for balanced QPSK spreading with maximal ratio diversity combining and coherent detection. After the space-time matched lter, each path signal is weighted by the complex (conjugate) amplitude of that path Pl,1 e jl,1 . The weighted path signals are then added together and fed to a decision
80
device. For coherent detection with maximal ratio combining, in addition to the beamformer weight vector wl,1 the receiver needs also an estimate of the path amplitude Pl,1 and the
phase e jl,1 . The path amplitudes can be estimated using the approach described above. The phase information e jl,1 is lost in the process of computing R yy,l,1. Therefore, for coherent detection we assume that a pilot sequence is available. In this case, the method described in [89] can be used to estimate the phase e jl,1 . Analysis of CDMA communication systems with QPSK spreading are reported in [33, 89, 106]. The bit error probability as a function of b is [33] 1 Pb (b ) = erfc b 2 2 erfc(x) =
x
(3.4.13)
where
et dt
2
(3.4.14)
and b is given by (3.4.2). If we make the same assumption that the total undesired component is white then for zero angle spread b has the probability density function in (3.4.5) and the average bit error probability Pe is [33] Pe =
0
Pb () f b () d L1 1 e/ K L (L 1)! ( K )
(3.4.15) ex dx d
2
=
0
1 = = where 1 =
K . 1+ K
e
0
x2 0 L L1 l=0
x2
L1 e/ K d dx ( K ) L (L 1)! L1+l l 1 + 1 2
l
1 1 2
For large angle spread, b has the pdf in (3.4.9). In this case, the aver-
Pb () f b () d K L1 1 e/ ( ) K L (K L 1)!
(3.4.19) ex dx d
2
=
0
(3.4.20)
81
10-2
10-3
10-4
10-5
10-6
10-7
Figure 3.14: Pe for balanced QPSK with coherent combining and 1 = = where 2 =
. 1+ x2 x2
= 0
Pe
e
0
1 2 2
0 K L K L1 l=0
K L1 e/ d dx K L (K L 1)! ( ) KL 1 + l l 1 + 2 2
l
(3.4.21) (3.4.22)
The probability of bit error performance of the Beamformer-RAKE receiver with balanced QPSK and coherent combining with L = 2 is shown in Figures 3.14 and 3.15. We also plot Pe for both angle spread = 0 and large and as a function of the average antenna SINR/bit L. These gures also show the performance improvement due to both beam forming gain and space diversity gain provided by the array.
82
10-1 10-2 10-3 Pe , Probability of Bit Error 10-4 10-5 10-6 10-7 10-8 10-9 10-10 5 10 15 20 25 30 L, Average Antenna SINR/bit, dB K =1 K=2 K=4
Figure 3.15: Pe for balanced QPSK with coherent combining and large
3.5 Summary
In this chapter, we reviewed different beamforming techniques and the conditions under which they are applicable. We also discussed the reasons why those techniques are not applicable for wireless CDMA systems. We then introduced our space-time approach for estimating the channel vector and optimum beamforming weight vector. We derived a spacetime receiver that we call a Beamformer-RAKE to exploit both temporal and spatial structure of the received multipath signal to maximize performance. In Chapter 4, we will look at techniques to implement the beamforming approach described above.
83
84
antenna array processing techniques and some numerical methods have already been proposed [107-109]. Some methods take into account the fact that the array covariance matrix varies by a rank one update from one time step to the next in order to speed up computational time [90, 110]. In addition to this fact, in the method that we will describe shortly, we also make use of the inherent time correlation structure to further speed up computation.
and that the channel vector al,1 can be also estimated as the principal eigenvector of the generalized eigenvalue problem R yy,l,1al,1 = Rxxal,1 (4.1.3)
85
The constant does not affect the beamformer output SINR and, therefore, we will not consider its value. Therefore, to nd the optimum weight vector at each sampling instant we need an estimate for the channel vector al,1 (n) and the undesired signal covariance Ruu,l,1.
R yy,l,1cm |R yy,l,1cm |
(4.1.4)
(4.1.5) (4.1.6)
lim cm = al,1
where max is the maximum eigenvalue of R yy,l,1. Since Rxx can be any arbitrary positive denite matrix, the idea is to decompose Rxx into the product of two matrices Rxx = R/2R1/2 xx xx (4.1.8)
where R1/2 is called the square root of the matrix Rxx . Using (4.1.3) and (4.1.4), we have xx R/2R yy,l,1R1/2 R1/2 al,1 = R1/2 al,1 xx xx xx xx (4.1.9)
Dening the matrix Bl,1 = R/2R yy,l,1R1/2 and the vector el,1 = R1/2 ak,1 , we can rewrite xx xx xx (4.1.9) as Bl,1 el,1 = el,1 (4.1.10)
86
Initialization: 1. co (n) = el,1 (n 1) 2. c1 (n) = Bl,1 (n)co (n)/|Bl,1 (n)co (n)| 3. m = 1 Computation: 1. while (|Bl,1 (n)cm (n) Bl,1 (n)cm1 (n)|/|Bl,1 (n)cm (n)| ) do
f g
Table 4.1: Power recursion for estimating principal eigenvector and hence the recursion dened in (4.1.4) can be used to estimate (max , el,1,max ). Then, the array response vector al,1 is given by al,1 = R1/2 el,1,max xx (4.1.11)
As we mentioned earlier, the array covariance matrices are functions of time and need to be estimated for every symbol n from samples of the pre- and post-correlation signal vectors x(n) and y yy,l,1 (n), respectively, where y yy,l,1(n) is dened as 1 yl,1 (n) = yl,1 (n) (4.1.12) 2Ts Therefore, for every symbol the estimates R yy,l,1 (n) and Rxx (n) are used in place of Ryy,l,1 and Rxx.
87
The above recursion for estimating the channel vector al,1 for each symbol can be summarized as in Table 4.1. The convergence behavior of the above recursion depends on the initial guess of the eigenvector co (n) and on the ratio |max /2 | where 2 is the second largest eigenvalue [90]. Due to the inherent time correlation of the channel vector, the previously estimated eigenvector el,1 (n 1) is a close estimate to el,1 (n) and, therefore, is used as the initial guess to speed up convergence. For example, for a mobile moving at 75 mph and symbol rate of 4800 symbols per second, the time correlation between the elements of the channel vector of two consecutive symbols is 0.995. Also, the ratio |max /2 | is lower bounded by the average SINR per path1 . In the current CDMA IS-95 standard, the closed loop power control holds the SINR at the output of the RAKE combiner such that 1% FER ( frame error rate) is achieved. This translates to average SINR per path of 10.5 dB, 5.7 dB, and 2.8 dB for 1, 2, and 4 independent paths used for combining, respectively [111]. In general, the lower the average SINR per path the more the number of iterations required for convergence.
(4.1.13)
In fact, if we assume that the total undesired signal is spatially white, we can easily show that |max /2 | = 1 + SINR.
88
R yy,l,1 (n) = R yy,l,1 (n 1) + yl,1 (n) (n) yl,1 (4.1.14) G 1 x(n)x (n) yl,1 (n) (n) (4.1.15) Ruu,l,1(n) = Ruu,l,1 (n 1) + yl,1 G1 G1 where is a positive constant between 0 and 1 and is called the forgetting factor. The use of the forgetting factor < 1 is used to ensure that the data in the distant past are forgotten in order to afford the possibility of statistical variation of the observed received signal vector as the mobile moves in the propagation environment. Simulation results presented at the end of this chapter show that a typical value of should range between 0.8 and 1. xx However, we need the time update for the inverse square root matrix R1/2 (n) and not Rxx (n). In order to nd the time update for the inverse square root, consider a K K Hermitian positive denite matrix Q dened as Q = P + pp (4.1.16)
where P is a K K Hermitian positive denite matrix and p is a K 1 vector. Our goal is to nd Q1/2 given P1/2 and p. Using the matrix inversion lemma [86], we can show that Q1 is given by Q1 = (P + pp )1 P pp P 1 + 1 p P1 p P1/2P/2 pp P1/2 P/2 = 1 P1/2 P/2 2 1 + 1 p P1/2 P/2p = 1 P1 2 Let f = 1/2 P/2p and = 1 + f f. Then, we can rewrite Q1 as Q1 = 1 P 1 P1/2 ff P/2 1/2 P1/2 ff P/2 1/2 = 1/2 P1/2 P/2 1/2 1 = 1/2 P1/2 I ff P/2 1/2 1 I ff = (I ff ) (I ff ) (4.1.20) (4.1.21) (4.1.22)
1 1
89
Therefore, it follows from (4.1.8) and (4.1.24) that the inverse square root matrix Q1/2 given P1/2 and p is given by Q1/2 = 1/2 P1/2 (I ff ) (4.1.25)
This is summarized in the SR UPDATE procedure shown in Table 4.2. Therefore we can xx write the time-update for the inverse square root R1/2 (n) as xx xx R1/2 (n) = SR UPDATE(R1/2 (n 1), x(n), ) (4.1.26)
Similarly, for the undesired signal vector we need the time-update of the matrix inverse uu,l,1 R1 (n) not Ruu,l,1(n). This time-update can be obtained by using the time-update expres sion for Ruu,l,1(n) in (4.1.15) and the matrix inversion lemma [86]. First we note that the time-update of Ruu,l,1(n) in (4.1.15) is a rank-2 update. We start by rewriting Ruu,l,1(n) as Ruu,l,1(n) = Kl,1 (n) 1 yl,1 (n) (n) yl,1 G1 (4.1.27)
90
Hence, using the matrix inversion lemma, we can easily show that the time-update of the uu,l,1 matrix R1 (n) is given by the two-step update K1 (n) l,1 uu,l,1 = R1 (n 1) 2
1
y yl,1 1 K1 (n) l,1 (n) (n)K1 (n) l,1 l,1 1 (n)K1 (n) (n) G 1 1 G1 yl,1 yl,1 l,1
(4.1.30)
uu,l,1 uu,l,1 The above time-update for R1 (n) involves subtraction. While R1 (n) is theoretically Hermitian positive denite (being the inverse of a Hermitian positive denite matrix), accumulation of numerical and quantization errors with time-update may lead to the loss of Hermitian positive deniteness. This can be avoided if we update the inverse square root uu,l,1 R1/2 (n) and since R1 (n) = R1/2 (n)R/2 (n), it remains Hermitian positive denite uu,l,1 uu,l,1 uu,l,1 regardless of any numerical or quantization errors in the inverse square root. Even in the presence of those errors, the numerical conditioning of the inverse square root R1/2 (n) is uu,l,1 generally much better than that of R1 [112]. Hence, using (4.1.27) and (4.1.27) we can uu,l,1 write the two-step time-update for R1/2 (n) as uu,l,1 K1/2 (n) = SR UPDATE(R1/2 (n 1), a1 x, ) l,1 uu,l,1 R1/2 (n) = SR UPDATE(K1/2 (n), a2 yl,1 , 1) uu,l,1 l,1 where a1 =
G G1
(4.1.31) (4.1.32)
and a2 =
j . G1
As we mentioned in Chapter 2, the dynamics of the channel vector depend on the speed of the mobile v, or equivalently the maximum Doppler shift fd , and the angle spread in
the corresponding received signal vector. Therefore, as we will see later in this chapter, we should expect that the value of that will give the best tracking performance will be a function of the channel dynamics, i.e. a function of ( fd , ) and hence should be a function
91
Initialize = a small positive constant, 1/2 Rxx (0) = R1/2 (0) = 1/2 I, uu,l,1 R yy,l,1 (0) = I, el,1 (0) = 1,
New Data: get x(n) and yl,1 (n). Time-Update: 1. if n = 1, 2 (n) = 1 else | (n 2) l,1 (n 1)| a a (n) = l,1 | l,1 (n 2)|| l,1 (n 1)| a a 1/2 1/2 2. Rxx (n) = SR UPDATE(Rxx (n 1), x(n), (n)) 1/2 3. Kl,1 (n) = SR UPDATE(R1/2 (n 1), a1 x(n), (n)) uu,l,1 1/2 (n) = SR UPDATE(K1/2 (n), a2 yl,1 (n), 1) 4. R
uu,l,1 l,1
5. R yy,l,1(n) = (n)R yy,l,1 (n 1) + yl,1 (n) (n) yl,1 1/2 /2 6. Bl,1 (n) = Rxx (n)R yy,l,1 (n)Rxx (n) Weight Computation: 7. el,1 (n) = POWER(Bl,1 (n), el,1 (n 1)) xx 8. al,1 (n) = R/2 (n)el,1 (n) 9. wl,1 (n) = R1/2 (n)R/2 (n) l,1 (n) a uu,l,1 uu,l,1
92
of time (n). In general, the value of (n) needs to be optimized to satisfy two competing requirements. The value of (n) should be as close as possible to 1 so that the effective number of samples used in estimating the array covariances is large enough to obtain quasiasymptotic covariance estimates and better noise averaging. On the other-hand the value of (n) should be small enough to track the channel dynamics. Finding an expression for (n) that will give the best tracking performance for al,1 (n) (which is estimated as the principal generalized eigenvector of Rxx (n) and R yy,l,1 (n)) might not be feasible. However, certain approximations can be made to nd a close estimate for (n). One such approximation is to take (n) to be equal to the normalized inner product l,1 (n) between last two estimates al,1 (n 1) and al,1 (n 2) dened as l,1 (n) = a | (n 2) l,1 (n 1)| al,1 | l,1 (n 2)|| l,1 (n 1)| a a (4.1.33)
which is a good measure of how fast the channel vector is changing with time.
xx where is a small positive constant. Correspondingly, the initial value for R1/2 (0) and
93
Step # 1 2 3 4 5+6 7 8 9
Table 4.4: Floating point operations count R1/2 (0) uu,l,1 xx R1/2 (0) = R1/2 (0) = 1/2I uu,l,1 (4.1.35)
The initial value of the principal eigenvector el,1 (0) is set to 1, where 1 is a K 1 vector whose elements are all equal to 1. This choice ensures that the initial guess used in the power recursion has a component in the direction of el,1 (1) and therefore will ensure convergence.
Table 4.4 gives the number of oating point operations (ops) for each step of the algorithm shown in Table 4.3. From this Table we can easily see that the overall complexity of the algorithm is O (K 2 ). The number of ops is taken to be the number of multiplication operation involved. The reason is that the instruction set of many of the current generation digital signal processors, such as the TMS320C30 by Texas Instruments, include a multiply-and-accumulate instruction that will implement calculations such as a b + c in one operation. For the sake of simplicity, we made no distinction between real and complex numbers. For instance, the product of a complex K 1 vector and a scalar (either real or complex) will be said to require K (complex) ops. In calculating the number of ops required for step 5 and 6,
94
0.0 -0.2 -0.4 -0.6 Mismatch Loss (dB) -0.8 -1.0 -1.2 -1.4 -1.6 -1.8 -2.0 0 20 40 60 80 1 00 Maximum Doppler Frequency fd (Hz) = 0 = 5 = 10 = 20 = 40 = 60 Forgetting Factor = 0.98
for = 0.98
xx we considered the rank-1 update structure for both R1/2 (n) and R yy,l,1 (n) in (4.1.14) and (4.1.26). If we were to directly implement steps 5 and 6, then the number of ops required is 2K 3 + 2K 2 . By considering the rank-1 update structure, the number of ops required is 11K 2 + 2K, which implies a reduction in the required number of ops for K 5. For step 7, the number of required ops will depend on the number of iterations required for con vergence. The use of el,1 (n 1) as an initial guess for el,1 (n) will minimize the number of iterations required and, therefore, will speed up convergence.
95
-0.4
-0.6 = 0 =5 = 10 = 20 = 40 = 60
-0.8
-1.0
-1.2 0 20 40 60 80 1 00
for = 0.95
96
Angle Spread = 5o
-0.1 -0.2
of = 90 and average SNR = 18dB. The processing gain G is 128. We assumed balanced DQPSK spreading with the same information bits in both the in-phase and quadrature channels. The bit rate was assumed to be 9.6 kbps. In each simulation run, a block of 100 bits was generated and differentially encoded. The differentially encoded data was spread in the in-phase and quadrature channels using two different spreading codes. The channel was generated using the model described in Chapter 2. We considered different cases with different values of angle spread , maximum doppler frequency fd , and forgetting factor . The algorithm in Table 4.3 was used to estimate the channel vector al,1 . To study the performance of the algorithm, we considered the mean beamformer output
0 20 40
cy f d quen r F re e oppl mD ximu a
60
80
0.8 0
1 00
M
= 5
97
1.0 5
1.0 0
Forgetting Factor
0.9 5
0.9 0 = 0 = 5 = 10 = 20 = 40 = 60
0.8 5
0.8 0
Figure 4.4: Optimum forgetting factor as a function of fd and SINR normalized to the maximum output SINR (output SINR when al,1 = al,1, i.e. no errors in estimating the channel vector). This mean normalized SINR is given by = E 20 log10 | al,1| al,1 | l,1||al,1 | a (4.2.1)
This quantity represents the loss in the average beamformer output SINR due to errors in the channel vector estimate al,1 and is therefore called the mismatch loss. Some of the simulation results are shown in Figures 4.1-4.5. Figure 4.1 shows the mismatch loss as function of f d and for = 0.98. We notice that for low doppler frequency,
the loss at low angle spread is higher than the loss at high angle spread, which can be explained by the reduction in space diversity gain provided by the array. At high Doppler frequency, the situation is reversed. The reason for this is that at high Doppler frequency and
98
0.0
-0.1
-0.2
-0.3
-0.4
-0.5
= 0 = 5 = 10 = 20 = 40 = 60
high angle spread, the channel variations are much faster than the rate at which array sample covariances are updated (which is a function of the forgetting factor ). In this case more and more incorrect samples will be used in estimating the array sample covariances which will lead to errors in the estimated channel vector. On the other hand, at high Doppler frequency and low angle spread, the channel vector tends to uctuate only in magnitude, and the uctuations in direction (phase) are very small and hence the errors in the estimated array response vector will be reected mainly in its magnitude, which will not affect the beamformer output SINR. Figure 4.2 is similar to Figure 4.1 but with forgetting factor = 0.95. We notice that there is a slight increase in the mismatch loss at low Doppler frequencies while at high Doppler frequencies we notice a reduction in the mismatch loss.
99
As we mentioned earlier, the forgetting factor that will give the best tracking performance is a function of fd and . With this optimum value of the forgetting factor , the = 5 . The optimum forgetting factor for
mismatch loss will be minimized. This is clearly illustrated in Figure 4.3 which shows the mismatch loss as a function of f d and for different values of fd and
4.3 Summary
In this chapter, we presented an adaptive algorithm for estimating the channel vector and the corresponding weight vector. The algorithm is based on the code-ltering approach presented in the previous chapter. The algorithm consists of two basic building blocks. The rst is a recursive procedure POWER for tracking and estimating the channel vector. The second block is a procedure SR UPDATE for time-updating the inverse square roots xx R1/2 (n) and R1/2 (n). The overall algorithm complexity is O (K 2 ). Simulation results uu,l,1 show that with optimum time-update (i.e. if we use optimum forgetting factor), we can track the channel vector within 0.5 dB of the true value even in severe propagation environments.
100
101
simultaneously, there is no need to avoid co-channel interference by requiring different frequency usage in neighboring cells. In fact the concept of a cell in terms of frequency planning is no longer necessary. The CDMA IS-95 standard uses time and path diversity to mitigate the effect of frequency selective multipath fading. Time diversity is obtained by the use of forward error correction (FEC) and interleaving. Path diversity is inherently provided by the CDMA approach by spreading the signal over a wide bandwidth (1.25 MHz). Such a signal with wide bandwidth will resolve the multipath components and, thus, provide the receiver with several independently fading signal paths. This path diversity is exploited by the use of a RAKE receiver to combine different multipath components. The use of a wideband signal for transmission may be also viewed as just another method for obtaining frequency diversity [33]. In a CDMA system every user is a source of interference to every other user, which makes mobile station transmit power control a key element in its current implementation. The current CDMA standard uses several power control techniques (open loop and closed loop power control) to optimize the system performance. In addition, the base station uses three sectored antennas (each covers 120 of the azimuth) to reduce the multiple access interference and thereby increase system capacity. Field test results of the IS-95 cellular system are reported in [118, 111]. This test was conducted in August 1993 in the San Diego area and included four base stations for a total of eight sectors and 86 class I mobile units. Capacity estimates from this eld test show that the IS-95 radio capacity/sector is 11.4 times better than that of AMPS. This translates to Erlang capacity/sector 16.4 times higher than that of AMPS. However, we should note that in [119], in commenting on a similar previous eld test experiment, Cox cites several potential capacity-reducing factors that were not exercised in this experiment. In the following sections we will briey describe the different signal processing functions used for forward and reverse link modems.
102
Interleaver +
+
Baseband Filter Q(t) sin( c t)
ci (t) User Long Code Mask Long Code Generator Wp (t) Pilot Walsh Code
(Q)
103
encoder used has a constraint length k = 9 and a code rate r of one-half. This will bring the data rate to 19.2 kbits/sec. The convolutional encoding is followed by interleaving over a 20-msec interval for burst error protection due to fast fading in the radio channel. The 19 kbits/sec output of the interleaver is then modied by the use of the so-called long code, which serves as a privacy mask. The modied stream is then encoded for spread-spectrum transmission using binary orthogonal Walsh codes [33] of dimension 64. This will produce 64-fold spreading of the data stream, resulting in a transmission rate of 1.2288 Mchips/sec. The structure of the Walsh code provides 64 orthogonal sequences, and one of the 64 sequences is assigned to a mobile unit during call setup. In this way, 64 orthogonal channels can be established on the forward link by the CDMA encoding on the same carrier frequency. After Walsh encoding , the spread data stream is separated into I and Q streams, each of which is modied by a unique short code of length 32768. The resulting spread spectrum stream is carried over the air interface with ltered QPSK modulation. All signals transmitted from a base station in a particular CDMA radio channel share the same set of 64 Walsh codes and the same pair of short codes. However, signals from different base stations are distinguished by time offsets from the basic short code which allows CDMA signals from each base station to be uniquely identied. Different signals transmitted from a given base station in a particular CDMA radio channel are distinguished at the mobile receiver by the orthogonal Walsh code. The orthogonality of the Walsh codes provides near perfect isolation between the multiple signals transmitted by the same base station. However, the presence of multipath propagation will partially destroy this orthogonality. An important aspect of the forward link waveform generated at the base station is the use of a pilot signal that is transmitted by each base station and is used as a coherent carrier reference for demodulation by all mobile receivers. The pilot is transmitted a relatively higher power level (approximately 20% of total forward link power budget [113, 120]) than other
104
types of signals which allows for tracking of the carrier phase. The pilot channel signal is unmodulated by information and uses the zeros Walsh function. Thus the pilot signal simply consists of the quadrature pair of short codes. The mobile receiver can obtain synchronization with the nearest base station without prior knowledge of the identity of the base station by searching out the entire length of the short code. The strongest signals time offset corresponds to the time offset of the short code of the base station to which the mobile has the best propagation channel (which is often the nearest base station). After synchronization, the pilot signal is used as a coherent carrier phase reference for demodulation of the other signals from this base station. The mobile receiver uses a limited number of correlators in a RAKE structure for demodulating the signals from the base station. The receiver uses the pilot signal to estimate the phase and the amplitude for each of the tracked multipath components and to coherently combine them.
105
Speech Coder
I-Channel Short Code (I) a (t) I BaseBand Filter
cos(ct) I(t)
Interleaver
S/P J bits
+
BaseBand Filter Q(t) sin(ct)
Power Control
a (t)
Q-Channel Short Code
(Q)
Figure 5.2: CDMA IS-95 reverse link waveform generation are used to select one of 64 different orthogonal Walsh functions for transmission. At the output of the Walsh modulator, the chip rate is 307.2 kchips/sec. The nal signal processing elements perform the direct-sequence spreading functions. First, the modulation symbols , or Walsh functions, are spread by using the mobile-specic long code at a rate of 1.2288 Mchips/sec, i.e., 256 chips per modulation symbol. The data stream is then split into I and Q streams where it is modied with the short code pair. The resulting spread spectrum signal is then carried over the air interface with a ltered O-QPSK (Offset-QPSK) modulation. Note that the use of the Walsh function in generating the reverse link waveform is different from that on the forward link. On the forward link, the Walsh function is determined by the channel that is assigned to each mobile while on the reverse link the Walsh function is determined by the information being transmitted. The use of Walsh function modulation on the reverse link is a simple way of obtaining 64-ary modulation. This is the best way of providing a high quality link over a fading channel with low SINR where a pilot phase reference cannot be provided [89]. Figure 5.3 shows a block diagram for the base station receiver. The base station uses
106
1 2 Searchers
6 Antenna Outputs
Figure 5.3: CDMA Base station receiver: IS-95 approach dual antenna diversity in each sector to provide path diversity. The antennas are widely separated such that independent fading at each antenna is obtained. For any given mobile, this structure implements a noncoherent RAKE receiver with a limited number of correlators for demodulating the L strongest multipath components received on all base station antennas. Noncoherent reception is used since, as we mentioned earlier, a pilot signal is not available on the reverse link. Time delay information of the L strongest multipath components is obtained by the front-end searcher. The front-end searcher has a larger number of scanning correlators that continuously scans all six antenna outputs for multipath signals received from the mobile. Tracked multipath components are rank ordered based on their energy and the L strongest are used for combining (current implementation of cell modem ASICs searches for 8 paths and uses the strongest 4 for combining [8, 111]).
107
5.3 Summary
This chapter provides a brief overview of the CDMA IS-95 cellular standard. We briey described different signal processing functions used for both the forward and reverse link modems. In the next chapter, we will apply the adaptive beamforming and BeamformerRAKE ideas presented in Chapter 3 to CDMA wireless systems that are based on the IS-95 standard and propose an antenna array-based base station receiver architecture and study its performance.
109
Searcher
Beamformer-RAKE Receiver
Figure 6.1: CDMA Base station receiver: beamforming approach station receiver does not make any distinction as to whether different paths came from the same antenna or different antennas. Also, signals received at two or more different antennas due to the same multipath component wavefront are considered different multipath signals. Clearly, this approach does not utilize any spatial structure in the received multipath signal. This spatial structure can be utilized by the use of the Beamformer-RAKE receiver. In this case, the base station will have an antenna array of K elements in each sector. The proposed base station receiver structure is similar to that of the IS-95 system except that the front-end searcher now searches for vector multipath components (by using the approach outlined in Chapter 3). In addition, each multipath demodulator will be preceded by a beamformer. A simplied overview of the Beamformer-RAKE base station receiver is shown in Figure 6.1. In this chapter, we will present an overall base station receiver architecture for CDMA with M-ary orthogonal modulation based on the Beamformer-RAKE structure and study its performance in terms of the uncoded bit error rate as a function of number of users and
110
number of antennas. In this analysis, we shall make use of the power control performance results that will be presented later in Chapter 7. We start by developing a model for the received signal vector at the base station array. Then, we will describe the proposed receiver structure and its different signal processing blocks. This will be followed by a detailed performance analysis for the proposed receiver and numerical and simulation results. The analysis in this chapter makes use of analysis results for DS/CDMA with M-ary orthogonal modulation but without antenna arrays that appeared in [121-125].
where Pi is the transmitted power per symbol per dimension, Tw is the symbol period, c is the carrier angular frequency, To = Tc /2 is the time offset between the I and Q channels, and nally i is a Bernoulli random variable that models the voice activity of the ith user (we assume that a user will be on with probability and will be off with probability 1 ).
111
W (h) (t) is the hth orthogonal Walsh function, h = 1, , M. Let the processing gain be G = Tw / Tc . For simplicity of notation, we shall denote the product of the users PN code and the I or Q channel PN code as ci( I ) (t) = ci (t)a ( I ) (t) and ci( Q) (t) = ci (t)a ( Q) (t)
To simplify our analysis, the PN codes ci( I ) (t) and ci( Q) (t) are represented by [125, 106] ci( I ) (t) =
r= r= ( Q) (I where ci,r) and ci,r are assumed to be i.i.d. random variables taking values 1 with equal (I ci,r) p(t kTc ) ( Q) ci,r p(t kTc )
(6.1.2) (6.1.3)
ci( Q) (t) =
probability, and p(t) is the chip pulse shape, which can be any time-limited waveform. Here we assume that p(t) is rectangular although our results can be easily extended for any timelimited waveform. We assume the multipath vector channel model in (2.2.25). To simplify our analysis, we assume that we have a constant deterministic power-delay prole and that the log-normal slow fading is the same for all multipath components. We also assume that the channel parameters vary slowly as compared to the symbol duration Tw so that they are constant over several symbol durations. Therefore, after downconverting to baseband, we can write the K 1 complex baseband received signal vector for the i-th user as
Li
xi (t) =
Si Pi i
l=1
(6.1.4)
where Si represents the log-normal shadowing experienced by the ith user, l,i = c l,i and Li is the number of multipath component for the ith user, al,i is the K 1 channel vector of the base station antenna array to signals in the l-th path from the ith mobile, and ci(h) (t) is dened as ci(h) (t) = W (h) (t)ci( I ) (t) + jW (h) (t To )ci( Q) (t To ) (6.1.5)
112
Let N be the number of cochannel mobiles. The total received signal at the cell site is the sum of all users signals plus noise and is given by
N Li
x(t) =
i=1 l=1
(6.1.6)
The vector n(t) = nc (t) + jns (t) is the K 1 additive Gaussian noise vector with zero mean and covariance
2 E{n(t1 )n (t2 )} = n I (t1 t2 ) 2 where n is the noise variance per antenna.
(6.1.7)
beamforming weight vector wl,i for the lth multipath component of the ith mobile from the pre-correlation and post-correlation array covariances Rxx and R yy,l,i 1 using the code ltering approach derived in Chapter 3.
As we have seen earlier, all matrices and vectors are functions of time. However, for simplicity of notation, we will drop the time index in the rest of this chapter.
z1
(1)
data
z1
Optimum Beamforming and Incoherent RAKE Combining
(M)
E T Threshold
z
w , w , ... , w
1 2 L
Select Index of Maximum
weight vectors
113
(1)
a (Q)
LPF
zk,1 (m)
(1)
+ +
W
1
a (I)
cos( c t)
M
c1(t)
sin(c t)
a (Q)
LPF
+ W
1
a (I)
W
M
zk,1 (m)
(M)
(1)
115
(h) For each multipath component, we have M different post-correlation signal vectors yl,i , (h) h = 1, , M. The vectors yl,i are fed to an optimum beamformer. The outputs of the (h) L1 beamformers for the hth Walsh function w yl,i are then fed into an incoherent RAKE l,i
combiner. The output of the incoherent RAKE combiner zi(h) is the decision variable for the hth Walsh function. The beamformer and the incoherent RAKE combiner for the hth Walsh function are shown in Figure 6.4. In order to update the post-correlation array covariance Ryy,l,i (that will be used in es(h) timating al,i and wl,i), the receiver needs the post-correlation vector yl,i corresponding to
the true transmitted Walsh symbol W (h) (t). However, at this stage the receiver has no prior
(n) knowledge of which post-correlation vector yl,i is the right one. Here, the receiver relies on
the inherent correlation of the multipath vector channel and the assumption that the channel remains almost constant over several symbol periods. In this case, the receiver uses a delayed update of Ryy,l,i (and hence delayed estimation of the channel vector and the optimum beamforming weight vector). This is done by using the decision on the current Walsh
( h) symbol h to select the post-correlation vector yl,i to update R yy,l,i and obtain the optimum
weight vector wl,i. This weight vector wl,i will be used for beamforming for the next symbol. The decision variables zi(1), , zi(M) at the output of the incoherent RAKE are then fed to an M-ary decoder, deinterleaver, and Viterbi convolutional decoder. Without loss of generality let us assume that the 1st user is the desired user and let k,1 be the time delay of the kth tracked multipath which is assumed to be estimated perfectly and k = 1, , L1 . Then,
(n) we can write the post-correlation signal vector yl,1 for the kth tracked multipath component
if if
116
where
(n) dk,1 = 2 Tw S1 P1 e jk,1 ak,1 (n) (n) (n) (n) uk,1 = mk,1 + sk,1 + nk,1
(6.2.4) (6.2.5)
(n) (n) and dk,1 is the desired signal vector, mk,1 is the multiple access interference (MAI) signal (n) vector, sk,1 is the self-interference (SI) signal vector due to other multipath components of (n) the 1st user, and nk,1 is due to the AWGN. Let Ai = Si Pi i . Also, for the kth tracked
multipath let the optimum beamforming weights determined using the previously estimated Walsh symbol be wk,1 . For an equal gain combining incoherent RAKE, the nth decision variable of the 1st user corresponding to the nth Walsh symbol is given by [33]
(n) z1 = L1 k=1 (n) zk,1 = L1 k=1 (n) |w yk,1 |2 k,1
n = 1, , M
(6.2.6)
(n) Now, to select which post-correlation signal vector yl,1 should be used in estimating the
post-correlation array covariance Ryy,l,i, a hard decision is made on which Walsh symbol was transmitted h = arg max
n=1,,M (n) z1
(6.2.7)
However, for the data a symbol-by-symbol M-ary decoder is used [89]. Both approaches yield exactly the same decisions for the M-ary symbol and both are optimal (i.e. a maximum likelihood rule) for an AWGN channel. Since the MAI is not necessarily Gaussian, this decision rule is actually not optimal. However, when the number of cochannel users is large, the multiple access interference can be modeled as Gaussian noise and therefore this decision rule can be used. The primary reason for using the symbol-by-symbol approach for the data is to provide improved performance with error-correcting codes by using soft decision decoding [33, 89]. Note also that we cannot use the output after the convolutional decoding and deinterleaving to select the post-correlation signal vector. The reason for this is that we will have to wait for a decision to be made on the current symbol and convolutionally encode and
117
(n) z1,1
|.|
Time Align 1
* w1,1
z2,1
(n)
|.|
* w2,1
Time Align 2
z1(n)
(n) z L,1
|.|
Time Align L
* wL,1
Figure 6.4: Optimum Beamforming and Incoherent RAKE interleave again. By the time this process is over (which is at least twice the time of one frame of bits), the channel would have changed and the estimated channel vector and the channel vector of the new symbol will be quite different. This will lead to a degradation in the beamformer output SINR.
118
1 (n) nk,1 = Tw =
(6.3.1) (6.3.2)
(n), (n), (n), (n), where nk,1 I I , nk,1 Q Q, nk,1 I Q, and nk,1 Q I are dened as
k,1 +Tw k,1 k,1 +Tw k,1 k,1 +Tw k,1 k,1 +Tw k,1
nc (t)ci( I ) (t k,1 )W (n) (t k,1 ) dt ns (t)ci( Q) (t To k,1 )W (n) (t To k,1 ) dt nc (t)ci( Q) (t To k,1 )W (n) (t To k,1 ) dt ns (t)ci( I ) (t k,1 )W (n) (t k,1 ) dt
1 (n), nk,1 I I = Tw 1 = Tw 1 = Tw
(n), We can easily show that nk,1 I I is a Gaussian random vector with zero mean and covariance
2 n I. 2
(n), (n), (n), Similarly, we can show that nk,1 Q Q, nk,1 I Q, and nk,1 Q I are all uncorrelated zero mean
(n) Gaussian random vectors with the same covariance. Hence nk,1 is a zero mean Gaussian 2 random vector with covariance 2n I.
119
1 Tw
L1
(6.3.10)
= A1
l=1 l=k
(6.3.11)
L1
= A1
l=1 l=k
(n), I I (n), Q (n), I Q (n), Q Ik,1,l,1 + Ik,1,l,1Q + j Ik,1,l,1 + Ik,1,l,1I e jl,1 al,1
(6.3.12)
(n), I (n), Q (n), I (n), Q where Ik,1,l,iI , Ik,1,l,i Q, Ik,1,l,iQ, and Ik,1,l,i I are dened as
1 (n), I Ik,1,l,iI = Tw
(n) (n), Q Ik,1,l,i Q
( W (t k,1 )c1I ) (t k,1 ) dt k,1 +Tw 1 = W (h) (t To l,i )ci( Q) (t To l,i ) Tw k,1 ( W (n) (t To k,1 )a1Q) (t To k,1 ) dt k,1 +Tw 1 = W (h) (t l,i )ci( I ) (t l,i ) Tw k,1 ( W (n) (t To k,1 )c1Q) (t To k,1 ) dt k,1 +Tw 1 = W (h) (t To l,i )ci( Q) (t To l,i ) Tw k,1 ( W (n) (t k,1 )c1I ) (t k,1 ) dt
(6.3.14)
(n), I Ik,1,l,iQ
(6.3.15)
(n), Q Ik,1,l,i I
(6.3.16)
=
i=2 l=1 N Li
1 Tw
(6.3.17) (6.3.18)
=
i=2 l=1 Li N
=
i=2 l=1
Ai
(n), Q (n), I (n), Q (n), I Ik,1,l,iI + Ik,1,l,i Q + j Ik,1,l,iQ + Ik,1,l,i I e jl,i al,i . (6.3.19)
120
Let W
(h)
(t
l,i )ai( I ) (t
l,i ) =
r=
(6.3.20)
(I (I (I where qi,r) = Wr(h) ci,r). It follows from (6.1.2) and (6.1.3) that qi,r), r = , , is an
i.i.d. binary random sequence taking values 1 with equal probability. Hence, it follows
(n), I (n), I that Ik,1,l,iI is zero mean. Using Equation (6.3.20), we can rewrite Ik,1,l,iI as (n), I Ik,1,l,iI
1 = Tw 1 = Tw
G1 b=0 G1 b=0
(I) q1,b
r=
(6.3.21)
where k,1,l,i = l,i k,1 modulo-Tc and R p (s) is the partial auto correlation of the chip waveform dened as
s
R p (s) =
o
p(t) p(t + Tc s) dt
0 s Tc
(6.3.22)
For rectangular pulses, R p (s) = s. For asynchronous networks, a reasonable assumption is that k,1,l,i are independent and uniformly distributed over [0, Tc ]. Let
(I) (I Fb = qi,b1 R p (k,1,l,i ) + qi,b) R p (Tc k,1,l,i )
(6.3.23)
Using the results in [106], we can show that {Fb }b=0,,G1 are independent random variables. Hence Var
(n), I Ik,1,l,iI G1
1 = Tw = = 1 Tw
Var {Fb }
b=0 G1
1 Tc2 2 = Tc 3
(n), Q (n), I (n), Q (n), I Similarly, we can also show that Ik,1,l,iI ,Ik,1,l,i Q, Ik,1,l,iQ, and Ik,1,l,i I are all zero mean uncor-
121
(n), I Remark: In deriving the variance of Ik,1,l,iI , we used the assumption that the chip pulse
shape is rectangular. In reality the channel is bandlimited (due to low pass ltering following the down converter) and the received signal cannot be a square wave. Under the condition that the same amount of energy is received regardless of the channel used, the received signal through a bandlimited channel will have a higher peak value, resulting in a higher level of MAI interference due to larger uctuations. In [124, 106] it was shown that if the bandlimited channel has ideal low pass lter characteristics with a bandwidth B = 1/ Tc , then, we would have
(n), I Var{Ik,1,l,iI } = Tc (n) Similarly, covariance of nk,1 will then be (n) 2 Var nk,1 = 2Tc n I
(6.3.28)
(6.3.29)
2
(n) (n) (n) The total interference vector ik,1 = mk,1 + sk,1 is modeled as a zero mean complex Gaus(n) (n) (n) sian random vector with covariance Ik,1 = E{ik,1 ik,1 }. Although, this assumption does not
always hold for CDMA analysis, it was shown in [106] that it is valid for large G. Moreover, simulation results presented later in this chapter show that for large N L, if we assume that the angles of arrival of the multipath components are uniformly distributed over the sector,
(n) the total interference vector ik,1 will be spatially white. In this case (n) Ik,1 = 2Tc 2 I I
(6.3.30)
where 2 is given by I
N
2 I
=C
i=2
(6.3.31)
and C is a constant equal to 2 for a bandlimited channel and 4 for a rectangular pulse shape. 3 For the remainder of our analysis we will assume the case of a bandlimited channel, i.e.
122
(n) C = 2. The covariance of the interference-plus-noise vector uk,1 is then given by (n) Ruu,k,1 = 2 I 2 where 2 = 2Tc ( 2 + n ). I
(6.3.32)
sumption that the MAI is spatially white, the optimum beamforming weights can be shown to be wk,1 = ak,1 (6.3.33)
where is some arbitrary constant (that does not change the beamformer output SINR). For simplicity of the analysis, we set = 1/ a ak,1. k,1
(n) Dene the beamformer output for the k-th multipath component of the 1st user w yk,1 k,1 (n) as Uk,1 (n) (n) Uk,1 = 2 A1 Tw |ak,1 |2 e jk,1 + w uk,1 k,1 (n) = w uk,1 k,1
for n = h , for n = h
(6.3.34) (6.3.35)
where |ak,1 | =
(n) (n) a ak,1. We can easily show that Vk,1 = w uk,1 is a zero mean complex k,1 k,1
Gaussian random variable with variance 2 . For simplicity of notation, let L1 = L. Then, the decision variables for the rst user are
(n) z1 = L l=1 (n) From [33], and conditioned on A1 and al,1 , l = 1 L, we can show that for n = h, z1 has (n) |Ul,1 |2
, n = 1, , M
(6.3.36)
E = 4 A2 Tw 1
l=1
|al,1|2
(6.3.37)
123
(n) The non-centrality parameter E is the symbol energy. For n = h, z1 has a 2 distribution
with 2L degrees of freedom. Therefore we can write the conditional probability density
(n) function of z1 as
f z(n) (z|s ) =
1
1 2 1
z 2 s
L1 2
s 2 +z 2
I L1
4s z 2
n = h, n=h
(6.3.38)
2L (L)
L1 z/ 2
where IL (.) is the modied Bessel function of the L-th order dened earlier in Section 2.1.2, (.) is the Gamma function, and s =
E
2
(6.3.39)
Remark: In this analysis, we used the assumption that the channel vector remains constant over two symbol periods. Also, we assumed that ak,1 is estimated perfectly. In reality the channel is time varying and the array covariances are estimated using few samples. This will lead to errors in the estimated channel vector and hence a reduction in the symbol energy s , as shown in Chapter 4. Therefore, the analysis results obtained here can be regarded as an upper bound on the system performance.
124
PM (s ) = 1 Pc
(2) (1) (3) (1) (M) (1) = 1 P(z1 < z1 , z1 < z1 , , z1 < z1 )
= 1
0
M1
and
(2) (1) P(z1 < z|z1 = z) = z 0
f z(2) (x)dx
1
(6.4.4) 1 z l! 2
l
= 1 ez/
L1
2
(6.4.5)
l=0
The symbol error probability and the corresponding bit error probability derived above are conditional probabilities and are functions of s , the symbol energy to interference plus noise ratio. s itself is a function of the channel vectors for the multipath components received from the 1st user a1,1 , , a L,1, shadowing and path loss S1 , and the 1st user transmitted power P1 . Also note that because of power control (both open loop and closed loop), P1 , S1 , and a1,1 , , a L,1 are generally dependent variables. The dependency among these variables is in general a function of the maximum Doppler shift fd of the rst user. From the simulation results in Chapter 7 and the results in [126, 127], a reasonable assumption is that the combination of open loop and closed loop power control is perfect in eliminating the slow fading due to shadowing and path loss. Based on the mobile speed (or fd ), we consider the following different cases.
125
the case of ideal power control, s is a xed quantity and is given by s = L K (6.4.7)
where is symbol energy to interference plus noise ratio per path per antenna. Also, the den
(n) sity function of z1 for n = h given in (6.3.38) becomes an unconditional density. Therefore,
1e
L1 2
z/ 2
L1 l=0
1 z l! 2
M1
1 2
z 2 s
s 2 +z 2
I L1 2
s z dz 2
(6.4.8)
However, due to the delay in the control loop, nite step size by which the mobile can increase or decrease its power, and errors on the downlink, power control can not be ideal (see Chapter 7). Therefore the symbol error probability obtained above needs to be averaged over the probability density function of s , which is not known. However, an approximation to the bit error probability can be obtained as follows. First, let Cv denote the coefcient of variation of s , dened as Var{s } Cv = E{s }
(6.4.9)
The average symbol error probability, denoted by PM , is an expectation of a real function of a random variable, namely s , so that PM = E { PM (s )} =
0
PM () f s () d
(6.4.10)
where PM (s ) is given by (6.4.8) and f s () is the probability density function of s . The objective here is to avoid carrying out the integration in computing (6.4.10), which requires the knowledge of fs (). Here, we use the results of [128] which are outlined below to obtain PM . PM (s ) is expanded in terms of a Taylors series expansion, so that PM (s ) = PM (s ) + (s s )P M (s ) + (s s )2 P 2! M (s ) + + (s s )n n PM (s ) + Rn n! (6.4.11)
126
and Rn is a remainder that vanishes as n becomes large. Taking expectations of both sides of (6.4.11) and ignoring terms beyond the second order term yields PM PM (s ) +
2 P 2
M (s )
(6.4.12)
where is the standard deviation of . We can easily see that also represents the power control error. However, in order to evaluate (6.4.12), we need the second derivative of PM (s ), which is very difcult to obtain. Instead, we expand PM (s ) in terms of central differences [37]; then taking expectation and ignoring terms beyond the second order term yields
2 PM (s + h) + PM (s h) 2 PM (s ) (6.4.13) PM PM (s ) + 2 h2 Choosing h = 3 is shown to give good accuracy [128]. The results in [125] show that
this approximation is accurate for low coefcient of variations Cv . Therefore, a reasonable approximation of the symbol error probability is 2 1 1 PM PM (s ) + PM (s + 3 ) + PM (s 3 ) 3 6 6 (6.4.14)
where s is the mean symbol energy to interference-plus-noise ratio and is the power control error. Then, the corresponding bit error probability for small Cv (less than 0.3 [125]) is 2 J1 Pb = J PM 2 1 (6.4.15)
s =
l=1
|al,1 |2 .
(6.4.16)
127
Small angle spread - For zero (or relatively small) angle spread , the channel vector of the lth multipath component can be expressed as al,1 l,1vl,1 (6.4.17)
where l,1 is a zero mean complex Gaussian random variable and for a ULA vl,1 is a Vandermonde vector [91] given by vl,1 = [1 e j sin l,1 D/ e j sin l,1 D(K1)/]T (6.4.18)
In this case, we can show that s has a 2 distribution with 2L degrees of freedom. That is f s () = L1 e/(K ) L ( K ) (L 1)!
(6.4.19)
Hence, it can be shown that the corresponding unconditional probability density func(n) tion of z1 for h = n is (see Appendix B)
f z(n) (z) =
1
(6.4.20)
Large angle spread - For large angle spread, the elements of al,1 becomes uncorreL
tion with 2 degrees of freedom. Therefore, s is distributed as a 2 random variable with 2K L degrees of freedom f s () = K L1 e/ KL ( ) (K L 1)! (6.4.21)
(K1)L
Rl
l=0
(6.4.22)
128
where Rl = (K 1)L 1 2 (1 + )) K Ll l ( 1+
(K1)Ll
1 1+
(6.4.23)
li |uli |2
(6.4.24)
where u11 u LK are i.i.d. zero mean complex Gaussian random variables and li = li , l,i
i=1,,K
are the eigenvalues of Rs,l,1, the spatial correlation matrix of the 1st
mobiles lth multipath component dened in (2.2.17). Let {li}l=1L,i=1K be equal to {i }i=1K L. Also, we assume that the li s are distinct (this is true if the angles of arrival are sufciently different). Then, s is distributed as [33]
KL
f s () =
i=1
i /i e i k = 1, , LK
(6.4.25)
where k =
LK
i=1 i=k
k , k i
(6.4.26)
(see Appendix)
KL
f z(n) (z) =
1
i
i=1
ez/( (1+i )) ( 2 (1 + i ))
2
g(z)
(6.4.27)
where z L2 2 g(z) = 2(L1) ez/ (L 2)! and denotes the convolution operation. (6.4.28)
(n) Using the unconditional pdf of z1 for h = n in Equations (6.4.20), (6.4.22), and (6.4.27),
1e
z/ 2
L1 l=0
1 z l! 2
M1
f z(1) (z)dz
1
(6.4.29)
and the corresponding average bit error probability is given by Equation (6.4.15).
129
d=8
were assumed uniform over [0, 60]. The results of 10000 post-correlation signal vectors were used to estimate the statistics of the MAI signal vector. Figures 6.6 and 6.7 show the empirical PDF of both the I and Q component of the MAI at the rst antenna. From both gures we can see the validity of the Gaussian approximation. Also, the spatial correlation
130
0.08 0.07 0.06 0.05 PDF 0.04 0.03 0.02 0.01 0.00 -20 -10 0 Interference Signal Level 10 20 Simulations Normal Fitting
Figure 6.6: I-channel: rst antenna interference distribution (n) matrix of the MAI vector Ruu,k,1 was estimated as (n) Ruu,k,1 = 102 100.43 0.62 + 0.29i 2.63 + 0.04i 2.12 + 0.03i 2.04 + 1.13i 0.62 0.29i 100.39 1.98 0.06i 0.90 + 0.09i 1.02 0.27i 2.63 0.04i 1.98 + 0.06i 100.55 0.39 + 0.11i 1.11 1.15i 2.12 0.03i 0.90 0.09i 0.39 0.11i 100.55 1.56 + 0.19i 2.04 0.13i 1.02 + 0.27i 1.11 + 1.15i 1.56 0.19i 100.27 The Frobenius norm [90] of the error e
F
(n) = Ruu,k,1 I
also shows that the MAI can be assumed to be spatially white. Next, we study the system bit error probability. For the low fd case, we used the power
131
0.08 0.07 0.06 0.05 PDF 0.04 0.03 0.02 0.01 0.00 -20 -10 0 Interference Signal Level 10 20 Simulations Normal Fitting
Figure 6.7: Q-channel: rst antenna interference distribution control simulation results presented in Chapter 7. In this simulation, we assumed that symbol rate is 4800 symbols/sec and that the mobile can increase/decrease its transmit power by 0.5 dB at a time and that the power control command was sent every 1.25 msec. We assumed zero additional loop delay and a forward link error rate of 0.05. Figure 6.8 shows the bit error probability for the case of ideal power control and for the case of power control with f d = 5 Hz. For the ideal power control case, the probability of error was computed using (6.4.6) and (6.4.8). For the power control case with fd = 5 Hz, the approximation in (6.4.14) was used. The resulting Pb is plotted for L = 4 and K = 1, K = 3, and K = 5. Note that this is independent of the angle spread since with fd = 5 Hz, the fading is slow enough to
132
10-1
10-3 K = 1, Cv = 0.084 K = 3, Cv = 0.077 10-4 K = 5, Cv = 0.077 Power Control, fd = 5 Hz Ideal Power Control 10-5 50 100 N, Number of Users 150 200
Figure 6.8: Pb for f d = 5 Hz and closed loop power control. be tracked by the power control loop for all values of control simulation results presented in the next chapter. If we assume that the required bit error rate is 103 , then for K = 1 the maximum number of users is 29 for ideal power control and 28 for power control with fd = 5 Hz. For K = 3, these numbers go up to 85 users and 82 respectively. This shows the improved performance due to beamforming. For the high f d case, as mentioned earlier, the distribution of s , and hence the distri(n) bution of z1 for n = h, depends on the angle spread (n) . Figure 6.9 shows the pdf of z1
and with ideal power control. From this gure we can see that
the higher the angle spread is, the closer to the ideal power control case the pdf becomes.
133
0.06
0.05
0.04
0.03
K = 5, L= 4, s = 13.82 dB
0.02
0.01
0.00 0 20 40 60
(1)
80
100
Decision Variable z1
This can be explained as follows. At zero angle spread, the received signal in any multipath component will experience the same fading at all antennas and the antenna array will not provide any space diversity for this multipath component. As the angle spread increases, the signal fading at different antennas becomes more and more uncorrelated which leads to less variations in the RAKE output. Figures 6.10, 6.11, and 6.12 show the bit error probability for = 0 , = 3 , and = 60 for a high maximum Doppler frequency fd (high enough such that the statistics of the received signal after power control remain the same as those of the multipath fast fading). For = 0 we can see that for Pb = 103 the maximum number of allowable mobile reduces to (compared to the perfect power control case) 10, 29, and 55 for K = 1, 3,
134
10-1
10-2
10-3
10-4
and 5, respectively, which corresponds to a 65% reduction in system capacity. This capacity reduction is due to the multipath fading which was not eliminated by the closed loop power control. With a single antenna, the statistics of s does not depend on the angle spread. Therefore, the maximum number of allowable mobiles for K = 1 is the same at 10 mobiles per cell for any value of angle spread. However, for angle spread = 3 this number goes up to 44 and 90 mobiles for K = 3 and K = 5, respectively. This is due to the additional diversity gain provided by the array. For = 60 , these number goes even higher to 58 and 110, which
is, again, due to the space diversity gain provided at large angle spreads. The results of Figures 6.10, 6.11, and 6.12 are summarized in Table 6.1. This table gives the % reduction
135
10-1
10-2
10-3
10-4
in capacity at Pb = 102 and Pb = 103 relative to the case with ideal power control. We make the following observation. For a xed angle spread, increasing the number of antennas will provide more space diversity gain. Similarly, for a xed number of antennas, increasing the angle spread will provide more space diversity gain.
136
10-1
10-2
10-3
10-4
K 1 3 5
Reduction at Pb = 102 =0 =3 = 60 50% 50% 50% 50% 34.6% 19.6% 50% 26.5% 13.8%
Reduction at Pb = 103 =0 =3 = 60 65.5% 65.5% 65.5% 65.5% 48.8% 31.8% 65.5% 36.9% 21.0%
Table 6.1: Percent reduction in capacity at Pb = 102 and Pb = 103 for high Doppler frequency
137
6.6 Summary
In this chapter, we proposed an antenna array-based base station receiver architecture for wireless CDMA systems. The receiver structure is based on the Beamformer-RAKE idea derived earlier in Chapter 3. We also studied the performance of the proposed base station receiver. In our performance analysis we used some of the results in the next chapter on power control performance. The average uncoded bit error probability was evaluated as a function of the number of mobiles, number of antennas, and angle spread for different power control scenarios. An improvement in the system performance that is proportional to the number of antennas is observed. Additional improvement is obtained due to space diversity gain at high angle spread.
138
139
the performance of this mobile is acceptable, but interference to all other mobiles that are sharing the same channel is increased, and may result in unacceptable performance to other mobiles. In fading radio environments, the system capacity is in general limited by multipath fading, shadowing and path loss, and multiple access interference. The system capacity can be improved by using several techniques including diversity combining, forward error correcting codes (FEC) and interleaving, and power control [126, 129]. Two different control mechanisms may be used for power control which are open loop power control (OLPC) and closed loop power control (CLPC). In OLPC, disparities caused mostly by different path loss and shadowing effects are adjusted individually by each mobile by controlling its transmitted power according to the received forward link signal level. OLPC will be affected by the multipath fast fading dynamics [130] since the forward and reverse link propagation losses and multipath fast fading are not symmetric, particularly when their center frequencies are widely separated. Errors or inaccuracy in open loop power control will cause uctuations of the signal to interference ratios, the effects of which cannot be corrected by coding and interleaving. Therefore, OLPC alone is not enough and a more rapidly acting closed loop power control (CLPC) must be used. In CLPC, the base station monitors the received signal power from each user and compares it to a preset threshold. Depending on the comparison result, a one bit up/down command is sent on the forward link to the user at a rate higher than the rate of multipath fading [115]. The current IS-95 CDMA system, uses a combination of OLPC and CLPC for power control on the reverse link. Previous work on power control in CDMA includes the work in [127, 131, 132]. The work in [127, 131] considered two approaches for CLPC: CLPC based on signal strength and CLPC based on signal to interference-plus-noise ratio SINR. CLPC based on SINR may results in improved system performance over CLPC based on signal strength [131]. However, previous studies considered antenna diversity at the base station as a source of path
140
diversity only. With the adaptive beamforming approach proposed in the previous chapter, the signal statistics of each multipath component at the output of the beamformer will not follow a Rayleigh distribution and will be affected by the presence of CLPC and other system parameters such as angle spread. In this chapter we will study the performance of signal strength-based CLPC of a single cell CDMA system.
141
command must be high enough to permit tracking the Rayleigh fading on the reverse link. It is important that the latency in estimating the power control signal and the transmission process be kept small so that the channel conditions will not change signicantly before the control bit can be received and acted upon. The system controller residing at the MTSO provides each cell site controller with a value of the Pd (or SINRd ) to be used for each individual mobile based on the error rate performance of that mobile. This level is passed to the channel controller where it is used to determine whether to command a particular mobile to increase or decrease its transmitted power. This overall mechanism is called closed loop power control.
142
Desired Level d (dB) Variation Due to Channel and Receiver Processing x i (dB) Control Command Decision (dB) 1 Return Channel Error 1 Transmit Power P (dB) i Tp Integrator Step Size p
ei
kTp
Loop Delay
143
power is estimated using the algorithm described below and compared to a preset threshold d. Based on the error e j = p j + x j d, a power control bit is generated and sent to the mobile requesting an increase or decrease by p.
However, although the base station makes the decision to increase or decrease the received signal power, the up/down command must still be transmitted on the down link to the mobile so that it may increase or decrease its transmit power level accordingly. If this command is received in error, the opposite action will take place. Moreover, the power control bit is sent unprotected, since the usual long delay due to coding and interleaving is inconsistent with a fast acting power control. Therefore, to include the effects of the forward link errors on the power control performance, we will assume that the power control bit will be received incorrectly with probability , where is the probability of error for the forward link. Errors on the forward link are assumed to be independent of those on the reverse link. This is a reasonable assumption since the multipath fast fading processes on both links are independent. In addition, this model accounts for extra loop delay kTp which accounts for the two-way signal propagation delay and the time delay involved in generating, transmitting, and executing the power control command. We can summarize the closed loop power control model described above as follows p j+1 = p j p jk Q (e jk ) 1 if eik 0 Q (e jk ) = 1 if e < 0 ik 1 with probability 1 jk = 1 with probability eik = pik + xik d (7.3.1) (7.3.2)
(7.3.3) (7.3.4)
The random variable represents the forward link errors. We can rewrite (7.3.1) in terms of the received signal power E j = x j + p j as E j+1 = E j p jk Q (e jk ) + x j+1 x j (7.3.5)
144
This equation is the control equation of the rst-order discrete-time feedback control loop shown in Figure 7.1. The quantity x j+1 x j represents the incremental change in the propagation loss. There is no obvious way to solve (7.3.5) in its current form. A simplifying assumption could be used by considering the time-continuous limit and using Fokker-Planck techniques [134]. However, it is more useful to simulate (7.3.5), and from this we can draw several conclusions about the CLPC performance. In the next section, we will present some of our simulation results. It remains to be shows how to estimate the received signal power at the base station. Because of the orthogonality of the M-ary signals used for modulation on the reverse link,
(1) (2) on the average at a relatively high SINR the maximum of the decision variables z1 , z1 , (M) , z1 corresponds to the signal power plus interference-plus-noise power. The remaining
M 1 variables correspond to interference-plus-noise power only. A reasonable assumption to make is that the fast fading remains almost constant over Tp symbols in any power control sampling period. Therefore, if we make a hard decision to select the maximum of
(1) (2) (M) z1 , z1 , , z1 , then we can estimate the short term interference-plus-noise and signal
(n) z1 (i)
(7.3.6)
1 ET = Tp
Tp i=1
n=1:M
(7.3.7)
(1) (2) (M) The decision variables z1 , z1 , , z1 are also passed to a symbol-by-symbol M-ary
decision device followed by a deinterleaver and a Viterbi convolutional decoder. The receiver uses the output bit sequence and other information from the convolutional decoder (such as the branch metrics) to get an estimate of the frame error rate (FER). Based on this estimated FER, the receiver selects the appropriate threshold level to be used. Figure 7.2 shows a block diagram for the power control algorithm.
145
J = log2 M
Measure FER
2 =
1 Tp(M-1) T
Tp
(n)
M z
(n)
(j)
ET
2
Threshold
ET =
1 max Tp j=1 M
{ z 1 (j) }
is zero.
146
15 Signal Level (dB) 10 5 Power-Controlled Signal 0 -5 0 200 400 600 800 1000 Symbol Period Tw Multipath Fast Fading
1.0 0.9 0.8 Pr {Signal Level Abscissa} 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 -4 -2 0 Normalized Signal Level (dB) 2 4 Power-Controlled Signal Level Multipath Fast Fading
Figure 7.3: Power-controlled received signal vs. simulated Rayleigh fading: fd = 5 Hz, K = 5, L = 4.
147
15 Signal Level (dB) 10 Power-Controlled Signal 5 0 -5 0 200 400 600 800 1000 Symbol Period Tw Multipath Fast Fading
1.0 0.9 0.8 Pr {Signal Level Abscissa} 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 -4 -2 0 Normalized Signal Level (dB) 2 4 Power-Controlled Signal Level Multipath Fast Fading
Figure 7.4: Power-controlled received signal vs. simulated Rayleigh fading: fd = 100 Hz, K = 5, L = 4.
148
Figure 7.5: Power control error vs. number of paths: K = 5. First, we study the effect of closed loop power control on the received signal level statistics. Figures 7.3 and 7.4 show the RAKE output received signal level waveform and the estimated distribution versus the simulated multipath fast fading for f d = 5 Hz and f d = 100 Hz, respectively. From these two gures we can easily see that closed loop power control eliminates most of the channel variations due to fast fading at low fd , while at high f d the received signal statistics after power control remain almost the same as that of the simulated multipath fast fading with only average power control. This result is consistent with the results in [127]. In addition, we may note that in [127, 115], it was suggested that for high fd the signal level distribution can be approximated by a log-normal distribution.
149
p = 0.25 dB p = 0.5 dB p = 1 dB
Figure 7.6: Power control error vs. power step size: L = 2,K = 5. Figure 7.5 shows the effect of the number of RAKE branches L on the power control error E . It can be seen from this gure that E is reduced with higher diversity order. We can also note that with high diversity order and low fd , CLPC can eliminate most of the channel variation. Figure 7.6 shows the effect of step size E for p on the power control error p and p = 0.25, 0.5, and 1 dB. We can see that E is lower with large step size
high f d , while at low f d , a small step size provides more precise control. The reason is that at high f d the fading rate is too high and a large step size is necessary to track the fading. On the other hand, at low fd the fading rate is slow enough to allow the control loop to track the fading with a small step size. Figure 7.7 shows the effect of loop delay kTp for
150
Figure 7.7: Power control error vs. loop delay: L = 2,K = 5. k = 1, 2, and 3. We can see from the gure that the longer the loop delay is, the poorer E . Figure 7.8 shows the effect of return channel error rate on E and it suggests the CLPC is robust against return channel errors. Again, these results are consistent with the results in [127, 132]. Next we look at the effect of angle spread and number of antennas in the array. As in the previous chapter, we assumed that the array size is 8 and the sensors are placed as a ULA. In Figures 7.9 and 7.10 we plot the power control error E against f d for different values of for K = 5 and 9. From these gures we can make the following observations. With zero angle spread , the number of antennas has no effect on the CLPC error E . This is due
151
= 0.05 = 0.1
Figure 7.8: Power control error vs. forward link error rate: L = 2, K = 5. to the fact that with zero angle spread , the received signal in any multipath component
will have the same fading at each antenna. Thus, the antenna array will not provide any space diversity for this multipath component. Also, for a given number of antennas K, as the angle spread increases, the gain due to space diversity increases which will lead to a reduction in the power control error E (i.e. better CLPC performance).
152
3 =0 =5 = 10 = 40 = 60
7.5 Summary
In this chapter, we studied the closed loop power control performance of the reverse link in wireless DS/CDMA through a combination of analysis and discrete event simulation. The closed loop power control mechanism is modeled as a rst-order discrete time feedback control loop. Based on this model, we simulated the closed loop power control on the communication link from a CDMA mobile to the base station. Several conclusions can be drawn from our simulation results. Our simulation results show that with high diversity order (either path or space) and low Doppler frequencies ( relative to the power control rate),
153
3 =0 =5 = 10 = 40 = 60
Figure 7.10: Power control error vs. angle spread: L = 4, and K = 9. closed loop power control can eliminate most of the channel variations, and with small angle spread adaptive beamforming with antenna arrays has a little effect on the power control error.
Chapter 8 Conclusions
The main focus of this thesis has been the development and application of advanced array signal processing techniques to CDMA wireless systems that have a practical implementation complexity and achieve high performance levels. This work has been prompted by the current thrust in wireless communication technology to look for new approaches and technologies to improve spectrum efciency and to be able to support the projected capacity demands with the introduction of new personal communication services.
Chapter 8. Conclusions
155
In Chapter 3, we start by reviewing general techniques for adaptive beamforming and discuss the reasons why such techniques are not suitable for CDMA wireless mobile systems. We then present a new space-time code-ltering approach for channel vector estimation and optimum beamforming. This approach exploits both the temporal and spatial information in the received signal. A key result in this thesis is the derivation of the spacetime matched lter receiver for multipath signals over AWGN channels. This receiver consists of a front-end spatial matched lter, or beamformer, followed by a temporal matched lter. Based on the space-time matched lter, we construct a Beamformer-RAKE receiver for CDMA signals. Performance analysis results of the Beamformer-RAKE receiver show improved bit error rate performance due to the exploitation of the spatial dimension in the received signal. In Chapter 4, we consider the time-variant nature of the wireless channel and develop a recursive algorithm for tracking the channel vector and estimating the beamformer weight vector. The algorithm has a reasonable computational complexity of O (K2 ) per path per user. Simulation results show that the proposed algorithm can track the channel vector closely even in severe propagation conditions. Chapters 6 and 7 look at the application of the array signal processing techniques derived in the previous chapters to the existing CDMA IS-95 cellular standard which we briey describe in Chapter 5. In Chapter 6 we propose an overall base station receiver architecture based on the Beamformer-RAKE structure and describe different signal processing functions. We then study the performance of the proposed base station receiver in terms of the uncoded bit error probability as a function of loading (number of users) and number of antennas. Under the propagation conditions considered, the analysis results show a performance improvement in terms of the number of users that can be supported due to the use of antenna arrays and the associated signal processing. These analysis results make use of the results of Chapter 7, where we look at modeling and performance of closed loop power control. Many conclusions can be drawn from the simulation results described in Chapter 7. The most important result is that beamforming reduces the power control error when
Chapter 8. Conclusions
156
(, t) =
i=1
Ri e ji (t)
(A.1.1)
where i = d cos i t i and the i modulo 2 are assumed to be i.i.d. and uniform over [0, 2]. We assume that the angular position of the ith scatterer i is uniformly distributed over [0, 2]. We also assume that the time delays i i are i.i.d. with probability density function f T (), where f T () is nonzero for 0 < and zero otherwise. Then, the timefrequency correlation of (, t) can be derived as follows: (1 , 2 , t, t + ) = E{(1 , t) (2 , t + )}
L L
(A.1.2) (A.1.3)
= E
i=1 l=1
The average will vanish unless i = l. In this case i (1 , t) i (2 , t + ) = d cos i i , where = 1 2 . Therefore (1 , 2 , t, t + ) = ( , ) 157 (A.1.4)
158
(1 , 2 , t, t + ) =
i
i )
(A.1.5)
E{R2 } represents the average fraction of incoming power in the ith path, which can be rewriti ten as E{R2 } = 2 f (i ) f T ()di di i (A.1.6)
where 2 is the total radiated from the mobile, which will be assumed to equal 1, and the term f (i ) f T (i )di di represents the average fraction of incoming power within di of angle i and within di of the time i . If we assume that L is large (i.e. L ), then we can express the sum in (A.1.5) with integrals, independent of i, 1 ( , ) = 2 1 = 2
2 0 2 0 0
e j(d cos
) 0
f T ()dd e j
e j(d cos )d
0
f T ()d
= Jo (d )
e j
f T ()d
= Jo (d ) FT ( j ) where FT (s) is the characteristic function of the time delay dened as FT (s) =
0
es f T () d
(A.1.11)
We notice that FT ( j ) is also the Fourier transform of the probability density function f T ().
a(, t) =
i=1
(A.2.1)
159
In addition to the assumptions used in the preceding Appendix, we also use the assumption from (2.2.9) that the angle of arrival of the ith path i is uniformly distributed over [ + , + ] where is the mean angle of arrival and 2 is the angle spread. Then, we can write Ra , the space-time-frequency correlation matrix of the channel vector, as
Ra (1 , 2 , ) = E{a(1 , t)a (2 , t + )}
=
i
i )
The average fraction of incoming power in the ith path E{R2 } is now redened as i E{R2 } = f (i ) f T () f (i )di di di i (A.2.4)
where the term f (i ) f T () f (i )di di di now represents the average fraction of incoming power form all paths that arrive within di of angle i within di of angle i and within di of the time i . Again, as L , then we can express the sum in (A.2.3) with integrals, independent of i,
Ra (1 , 2 , ) =
1 2
2 0
0 +
e j(d cos
1 v(1 , )v (2 , )d 2 + = Jo (d ) FT ( j ) Rs
where Rs is dened in ( 2.2.17). To evaluate the matrix Rs , we consider the model geometry in Figure 2.9 and evaluate the (m, n) element of Rs as follows Rs (m, n) = 1 2
+ +
vm (1 , )v (2 , )d n
(A.2.7)
vm (1 , ) is the phase shift for a plane wave propagating with angular frequency 1 between the mth sensor and the reference point. Similarly, vn (2 , ) is the phase shift for a plane wave propagating with angular frequency 2 between the nth sensor and the reference point.
160
where c is the speed of light and rm and rn are given by rm = rn dmn r sin mn + dmn sin(r + mn ) mn 2 dmn r = dmn sin(r + mn ) sin mn mn 2 (A.2.9) (A.2.10)
Let = (1 rm 2 rn )/c and dene z1 = z2 1 + 2 1 2 r dmn + dmn cos r mn 2c 2c 1 2 r dmn sin r = mn 2c z1 . zmn (A.2.11) (A.2.12)
We can easily show that = z1 sin mn + z2 cos mn = zmn sin(mn + mn ) where zmn = Therefore, we have Rs (m, n) = = = 1 2 1 2 1 2
+ + + + + +
z2 + z2 1 2
and
mn = cos1
(A.2.13)
where = mn + mn. Now, by making use of the well-known series representation cos(zmn sin x) = Jo (zmn ) + 2 sin(zmn sin x) = 2 we can integrate (A.2.16) to get Re{Rs (m, n)} = Jo (zmn ) + 2 Im{Rs (m, n)} = 2
l=0 l=1 l=0 l=1
(A.2.17) (A.2.18)
(A.2.19)
161
1 2j
(B.0.1)
f z(n) (z) =
1
f z(n) (z|s ) f s () d
1
1 = 2j 1 = 2j
ds
For small angle spread, fs () is given by Equation (6.4.19) and f z(n) (z) =
1
1 2j 1 = 2j
163
(B.0.3)
For large angle spread, fs () is given by Equation (6.4.21) and f z(n) (z) =
1
1 2j 1 = 2j
(B.0.4)
( 2 s + 1)(K1)L ( 2 (1 + )s + 1) K L
(K1)L
Rl s + 1/( 2 (1 + )
K Ll
(B.0.5)
s=
1 (2 (1+ ))
(K1)Ll
1 1+
(B.0.6)
(K1)L
Rl
l=0
(B.0.7)
For other values of angle spread, rst we need to derive the density function for s itself. We have s =
l=1 L
|al,1 |2
(B.0.8)
Let al,1 = R/2 ul where Rs,l,1 is the K K spatial correlation matrix of the array for s,l,1 al,1 and ul is a K 1 zero mean complex Gaussian random vector with covariance matrix I. Since Rs,l,1 is Hermitian, we can rewrite Rs,l,1 as Rs,l,1 = Ul l U l (B.0.9)
164
where Ul is orthogonal and l is a diagonal matrix of the eigenvalues of Rs,l,1. Let l = diag{l1 l2 l K }. Then we can write |al,1 |2 as |al,1 |2 = u Ul l U ul l l l = u l ul
K
=
i=1
li |uli |2
(B.0.10)
where ul = U ul is also a zero mean complex Gaussian random vector with covaril ance I. Then, we can rewrite s as
L K
s =
l=1 i=1
li |uli |2
(B.0.11)
where li = li and, therefore, s has the density function in Equation (6.4.25). It follows that f z(n) (z) =
1
1 2j
KL
esz ( 2 s + 1) L 1 2j ( 2 s
2
KL
i
l=i sz
2s + 1 2 (1 + i )s + 1
ds
=
l=i KL
i i
i=1 KL
= =
i=1
Bi e
i 1 + i
zl 2l l!
(B.0.12)
where Bi =
i i 2
i + 1 i
L2
(B.0.13)
Bibliography
[1] J. Unger, Herman, The Third Treasury. Kansas City: Andrews and McMeel Inc, 1982. [2] Bell Laboratories, Special Issue on Advanced Mobile Phone Service, Bell Syst. Tech. Journal, vol. 58, pp. 1269, January 1979. [3] F. H. Blecher, Advanced Mobile Phone Service, IEEE Trans. Veh. Technol., vol. VT-29, pp. 238244, 1980. [4] W. C. Lee, Mobile Cellular Telecommunications: Analog and Digital Systems. New York: McGraw-Hill, 1995. [5] K. Pahlavan and A. H. Levesque, Wireless Information Networks. New York: John Wiley and Sons, 1995. [6] T. Haug, Overview of GSM: Philosophy and Results, Int. J. Wireless Inf. Networks, vol. 1, pp. 716, 1994. [7] M. Rahnema, Overview of the GSM System and Protocol Architecture, IEEE Commun. Mag., vol. 31-(4), pp. 92100, April 1993. [8] Qualcomm Inc., Wideband Spread Spectrum Digital Cellular System, April 1992. Proposed EIA/TIA Interim Standard. 165
. Bibliography
166
[9] G. Raleigh, S. N. Diggavi, A. F. Naguib, and A. Paulraj, Characterization of Fast Fading Vector Channels for Multi-Antenna Communication Systems, in Proc. 27th Asilomar Conference on Signals, Systems and Computers, vol. II, (Pacic Grove, CA), November 1994. [10] B. Suard, A. Naguib, G. Xu, and A. Paulraj, Performance Analysis of CDMA Mobile Communication Systems using Antenna Arrays, in Proc. ICASSP93, vol. VI, (Minneapolis, MN), pp. 153156, April 1993. [11] A. Naguib and A. Paulraj, Performance of CDMA Cellular Networks with BaseStation Antenna Arrays, in Proc. International Zurich Seminar on Digital Communications, (Zurich, Switzerland), pp. 87100, March 1994. [12] A. F. Naguib, A. Paulraj, and T. Kailath, Capacity Improvement with Base-Station Antenna Arrays in Cellular CDMA, IEEE Trans. Veh. Tech., vol. VT-43(3), pp. 691 698, August 1994. [13] A. Naguib and A. Paulraj, Performance of CDMA Cellular Networks with BaseStation Antenna Arrays: The Downlink, in Proc. ICC94, vol. II, (New Orleans, Louisiana), pp. 795799, May 1994. [14] A. F. Naguib, A. Paulraj, and T. Kailath, Capacity Improvement with Base-Station Antenna Arrays in Cellular CDMA, in Proc. 27th Asilomar Conference on Signals, Systems and Computers, vol. II, (Pacic Grove, CA), pp. 14371441, November 1993. [15] A. F. Naguib and A. Paulraj, Recursive Adaptive Beamforming for Wireless CDMA, in Proc. ICC95, vol. III, (Seattle,WA), pp. 15151519, June 1995. [16] A. F. Naguib and A. Paulraj, Recursive Adaptive Beamforming for Wireless CDMA. submitted to IEEE Commu. Transactions, 1995.
. Bibliography
167
[17] A. F. Naguib and A. Paulraj, A Base-Station Antenna Array Receiver for Cellular DS/CDMA with M-ary Orthogonal Modulation, in Proc. 27th Asilomar Conference on Signals, Systems and Computers, vol. II, (Pacic Grove, CA), November 1994. [18] A. F. Naguib and A. Paulraj, Performance of Cellular CDMA with M-ary Orthogonal Modulation and Cell Site Antenna Arrays, in Proc. ICC95, vol. II, (Seattle,WA), pp. 697702, June 1995. [19] A. F. Naguib and A. Paulraj, Performance of Cellular CDMA with M-ary Orthogonal Modulation and Cell Site Antenna Arrays. to appear in IEEE J. Select. Areas Commun., 1995. [20] A. F. Naguib and A. Paulraj, Power Control in Wireless CDMA: Performance with Cell Site Antenna Arrays, in Proc. GLOBECOM95, vol. I, (Singapore), pp. 225 229, 1995. [21] W. C. Jakes, Microwave Mobile Communications. New York: John Wiley and Sons, 1974. [22] A. J. Rustako, N. Amitay, G. J. Owen, and R. S. Roman, Radio Propagation at Microwave frequencies for Line-of-Sight Microcellular Mobile and Personal Communications, IEEE Trans. Veh. Technol., vol. VT-40(1), pp. 203210, Feb 1991. [23] M. D. Yacoub, Foundations of Mobile Radio Engineering. Boca Raton, FL: CRC Press, 1993. [24] Y. Okumura, E. Ohmori, T. Kawano, and K. Fukuda, Field Strength and its Variability in VHF and UHF Land-Mobile Radio Service, Rev. Elec. Commun. Lab., vol. 16, pp. 825873, September-October 1968.
. Bibliography
168
[25] D. C. Cox and R. P. Leck, Delay Spread and Path loss Correlation for 910-MHz Urban Mobile Radio Propagation, IEEE Trans. Veh. Tech., vol. VT-4, pp. 340344, November 1977. [26] F. Niu and H. L. Bertoni, Path Loss and Cell Coverage of Urban Microcells in HighRise Building Environments, in Proc. GLOBCOM93, vol. I, pp. 266267, 1993. [27] H. L. Bertoni, UHF Propagation Prediction for Wireless Personal Communications, IEEE Proceedings, vol. 82(9), pp. 13331359, September 1994. [28] N. Amitay, Modeling and Computer Simulation of Radio Wave Propagation in Lineal Line-of-Sight Microcells, IEEE Trans. Veh. Tech., vol. VT-41, pp. 337342, November 1992. [29] J. F. Wagen, Signal Strength Measurements at 881 MHz for Urban Microcells in Downtown Tampa, in Proc. GLOBECOM91, vol. III, pp. 131311317, 1991. [30] R. J. C. Bultitude and G. K. Bedal, Propagation Characteristics on Microcellular Urban Mobile Systems, IEEE J. Selected Areas Commun., vol. JSAC-7(1), pp. 31 39, January 1989. [31] R. S. Kennedy, Fading Dispersive Communication Channels. New York: John Wiley and Sons, 1969. [32] W. R. Braun and U. Dersch, A Physical Mobile Radio Channel Model, IEEE Trans. Veh. Tech., vol. VT-40(2), pp. 472482, May 1991. [33] J. G. Proakis, Digital Communications. McGraw-Hill, second ed., 1989. [34] D. Williams, Probability with Martingales. Cambridge, England: Cambridge Mathematical Textbooks, 1991.
. Bibliography
169
[35] D. C. Cox, 910 MHz Urban Mobile Radio Channels: Multipath Characteristics in New York City, IEEE Trans. Commun., vol. COM-21(11), pp. 11881194, November 1973. [36] G. Turin, Introduction to Spread-Spectrum Antimultipath Techniques and their Application to Urban Digital Radio, Proc. of IEEE, vol. 68, pp. 328353, March 1980. [37] E. Kreyszig, Advanced Engineering Mathematics. New York: John Wiley, 1983. [38] S. Ogose, K. Murota, and K. Hirade, A Transmitter Diversity for MSK with TwoBit Differential Detection , IEEE Trans. Veh. Technol., vol. VT-33(1), pp. 3743, February 1984. [39] H. Hashemi, Simulation of the Urban Radio Propagation, IEEE Trans. Veh. Tech., vol. VT-28(3), pp. 213225, Aug 1979. [40] G. L. Turin, A Statistical Model of Urban Multipath Propagation, IEEE Trans. Veh. Tech., vol. VT-21, pp. 19, February 1972. [41] H. Suzuki, A Statistical Model for Urban Radio Propagation, IEEE Trans. Commun, vol. COM-25(7), pp. 673680, July 1977. [42] H. Hashemi, The Indoor Radio Propagation Channel, Proc. IEEE, pp. 943968, 1993. [43] F. Ikegami, S. Takeuchi, and S. Yoshida, Theoretical Prediction of Mean Field Strength for Urban Mobile Radio, IEEE Trans. Antennas Propagat., vol. AP-39(3), pp. 299302, March 1991. [44] M. Hata, Empirical Formula for Propagation Loss in Land Mobile Radio Services, IEEE Trans. Veh. Tech., vol. VT-29(3), pp. 317325, August 1980.
. Bibliography
170
[45] S. C. Swales, M. A. Beach, D. J. Edwards, and J. P. McGeehn, The Performance Enhancement of Multibeam Adaptive Base Station Antennas for Cellular Land Mobile Radio Systems, IEEE Trans. Veh. Tech., vol. VT-39(1), pp. 5667, February 1990. [46] A. J. Goldsmith and L. J. Greenstein, A Measurement-Based Model for Predicting Coverage Areas of Urban Microcells, IEEE J. Select. Areas Commun., vol. JSAC11(7), pp. 10131023, September 1993. [47] M. Gudmundson, Correlation Model for Shadow Fading in Mobile Radio Systems, Electronics Letters, vol. 27-(23), pp. 21452146, November 1991. [48] D. Cox, Private Communications, 1995. [49] J. D. Parsons and A. M. D. Turkumani, Characterization of Mobile Radio Signals: Model Description, IEE Proceedings-I, vol. 138(6), pp. 549556, December 1991. [50] J. Salz and J. H. Winters, Effect of Fading Correlation on Adaptive Arrays in Digital Mobile Radio , IEEE Trans. Veh. Tech., vol. VT-43(4), pp. 10491057, November 1994. [51] R. Price and J. P.E. Green, A Communication Technique for Multipath Channels, Proc. IRE, vol. 46, pp. 555570, March 1958. [52] G. Turin, A Statistical Model of Urban Multipath Propagation, IEEE Trans. Veh. Tech., vol. VT-21(1), pp. 19, February 1980. [53] A. A. M. Saleh and R. A. Valenzuela, A Statistical Model for Indoor Multipath Propagation, IEEE J. Select. Areas Commun., vol. JSAC-5(2), pp. 128137, February 1987. [54] D. H. Johnson and D. E. Dudgeon, Array Signal Processing. Englewood Cliffs: Prentice Hall, 1993.
. Bibliography
171
[55] J. Winters, Private Communications, 1995. [56] D. C. Cox and R. P. Leck, Distributions of Multipath Delay Spread and Average Excess Delay for 910 MHz Urban Mobile Radio Paths, IEEE Trans. Antennas and Prop., vol. AP-23, pp. 206213, March 1975. [57] D. C. Cox and R. P. Leck, Correlation Bandwidth and Delay Spread Multipath Propagation Statistics for 910 MHz Urban Mobile Radio Channels, IEEE Trans. Commun., vol. COM-23(11), pp. 12711280, November 1975. [58] T. S. Rappaport and S. Y. Seidel, 900 MHz Multipath Propagation Measurements for U.S. Digital Cellular Telephony, IEEE Trans. Veh. Tech., vol. VT-39(2), pp. 794 807, February 1990. [59] D. J. Goodman, Trends in Cellular and Cordless Communications, IEEE Communications Magazine, vol. 29, pp. 3140, June 1991. [60] P. Balaban and J. Salz, Optimum Diversity Combining and Equalization in Digital Data Transmission with Application to Cellular Mobile Radio, IEEE Trans. Veh. Tech., vol. VT-40(2), pp. 342354, May 1991. [61] P. Balaban and J. Salz, Optimum Diversity Combining and Equalization in Data Transmission with Application to Cellular Mobile Radio - Part I: Theoretical Considerations, IEEE Trans. Commun., vol. COM-40(5), pp. 885894, May 1992. [62] P. Balaban and J. Salz, Optimum Diversity Combining and Equalization in Data Transmission with Application to Cellular Mobile Radio - Part II:Numerical results, IEEE Trans. Commun., vol. COM-40(5), pp. 895907, May 1992. [63] J. H. Winters, Optimum Combining in Digital Mobile Radio with Cochannel Interference, IEEE J. Select. Areas Commun., vol. JSAC-2(4), pp. 528539, July 1984.
. Bibliography
172
[64] J. H. Winters, Optimum Combining for Indoor Radio Systems with Multiple Users, IEEE Trans. Commun., vol. COM-35(11), pp. 12221230, November 1987. [65] J. H. Winters, On the Capacity of Radio Communication Systems with Diversity in a Rayleigh Fading Environment, IEEE J. Select. Areas Commun., vol. JSAC-5(5), pp. 871878, June 1987. [66] S. Anderson, M. Millnert, M. Viberg, and B. Wahlberg, An Adaptive Array for Mobile Communication Systems, IEEE Trans. Veh. Tech., vol. VT-40(1), pp. 230236, February 1991. [67] T. Ohgane, T. Shimura, N. Matsuzawa, and H. Sasaoka, An Implementation of a CMA Adaptive Array for High Speed GMSK Transmission in Mobile Communications, IEEE Trans. Veh. Technol., vol. VT-42(3), pp. 282288, August 1993. [68] R. O. Schmidt, A Signal Subspace Approach to Multiple Emitter Location and Spectral Estimation. PhD thesis, Stanford University, Stanford, CA94305, 1981. [69] A. Paulraj, R. Roy, and T. Kailath, Estimation of Signal Parameters by Rotational Invariance Techniques (ESPRIT), in Proc. of 19th Asilomar Conference on Circuits, Systems and Comp., 1985. [70] B. Ottersten and M. Viberg, Analysis of Subspace Fitting Based Methods for Sensor Array Processing, in Proc. ICASSP89, vol. IV, (Glasgow, Scotland), pp. 2807 2810, 1989. [71] M. Viberg and B. Ottersten, Sensor Array Processing Based on Subspace Fitting, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-39(5), pp. 11101121, May 1991.
. Bibliography
173
[72] B. Ottersten, R. Roy, and T. Kailath, Signal Waveform Estimation in Sensor Array Processing, in Proc. 23rd Asilomar Conference on Signals, Systems and Computers, vol. II, (Pacic Grove, CA), pp. 787791, November 1989. [73] J. H. Winters, Signal Acquisition and Tracking with Adaptive Arrays in Wireless Systems, in Proc. 43rd Vehicular Technology Conf., vol. I, pp. 8588, November 1993. [74] J. H. Winters, Signal Acquisition and Tracking with Adaptive Arrays in Digital Mobile Radio System IS-54 with Flat Fading, IEEE Trans. Veh. Tech., vol. VT-42(4), pp. 377384, July 1993. [75] J. H. Winters, J. Salz, and R. D. Gitlin, The Impact of Antenna Diversity on the Capacity of Wireless Communication Systems, IEEE Trans. Commun., vol. COM41(4), pp. 17401751, April 1994. [76] R. Gooch and J. Lundell, The CMA Array: An Adaptive Beamformer for Constant Modulus Signals, in Proc. ICASSP86, (Tokyo), pp. 25232526, April 1986. [77] B. G. Agee, Blind Separation and Capture of Communication Signals Using Multitarget Constant Modulus Beamformer, in Proc. IEEE MILCOM89, vol. I, pp. 340 346, 1989. [78] S. Mayrargue, Spatial Equalization of a Radio-Mobile Channel Without Beamforming Using Constant Modulus Algorithm (CMA), in Proc. ICASSP93, vol. III, pp. 344347, 1993. [79] R. Gooch and B. Sublett, Joint Spatial and Temporal Equalization in a DecisionDirected Adaptive Antenna System, in Proc. 23rd Asilomar Conference on Signals, Systems and Computers, vol. I, (Pacic Grove, CA), pp. 255259, November 1989.
. Bibliography
174
[80] A. Swindlehurst, S. Daas, and J. Yang, Analysis of Decision Directed Beamformer. submitted to IEEE Trans. Acoust., Speech, Signal Processing, 1993. [81] S. Talwar, A. Paulraj, and M. Viberg, Reception of multiple co-channel digital signals using antenna arrays with applications to pcs, in Proc. ICC94, vol. II, pp. 700 794, 1994. [82] S. Talwar, M. Viberg, and A. Paulraj, Blind Estimation of Multiple Co-Channel Digital Signals Using an Antenna Array, IEEE Signal Processing Letters, vol. (1)-2, pp. 2931, Feb 1994. [83] B. Agee, A. V. Shell, and W. A. Gardner, Spectral Self-Coherence Restoral: A New Approach to Blind Adaptive Signal Extraction Using Antenna Arrays, Proc. IEEE, vol. 78, pp. 753767, April 1990. [84] L. Tong, Y. Inouye, and R. Liu, Waveform-Preserving Blind Estimation of Multiple Independent Sources, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP41(7), pp. 24612470, July 1993. [85] R. A. Monzingo and T. W. Miller, Introduction to Adaptive Arrays. New York: John Wiley and Sons, 1980. [86] T. Kailath, Linear Systems. Englewood Cliffs, N.J.: Prentice-Hall, 1981. [87] H. L. Van Trees, Detection, Estimation, and Modulation Theory, vol. I. New York, NY: John Wiley and Sons, 1969. [88] B. Widrow and S. Stern, Adaptive Signal Processing. Prentice-Hall, 1985. [89] A. J. Viterbi, Principles of Spread Spectrum Multiple Access Communications. Reading, MA: Addison-Wesley, 1995. Englewood Cliffs, NJ:
. Bibliography
175
[90] G. H. Golub and C. F. V. Loan, Matrix Computations. Baltimore and London: The John Hopkins University Press, second edition ed., 1989. [91] S. U. Pillai, Array Signal Processing. New York, NY: Springer Verlag, 1989. [92] M. Kavehrad and P. McLane, Performance of Low Complexity Channel Coding and Diversity for Spread Spectrum in Indoor Wireless Communications, AT&T Technical Journal, vol. 64, pp. 19271964, October 1985. [93] U. Grob, A. L. Welti, Z. Zollinger, R. Kung, and H. Kaufmann, Micro-Cellular Direct-Sequence Spread-Spectrum Radio System Using N-path RAKE receiver, IEEE J. Select. Areas Commun., vol. JSAC-8(5), pp. 772780, May 1990. [94] R. E. Kane, K. S. Gong, and R. R. Kurth, Performance of a RAKE Demodulator with Pre-Decision Multipath Thresholding , in Proc. MILCOM90, pp. 10251029, May 1990. [95] A. Higashi and T. Matsumoto, Combined Adaptive RAKE diversity (ARD) and Coding for DPSK DS/CDMA mobile Radio, IEEE J. Select. Areas Commun., vol. JSAC-11(7), pp. 10761084, September 1993. [96] L. F. Chang, Dispersive Fading Effects in CDMA Radio Systems, in Proceedings of First Int. Conf. on Universal Personal Communications, (Dallas, TX), September 1992. [97] H. V. Poor, An Introduction to Signal Detection and Estimation. New York, NY: Springer Verlag, 1988. [98] C. Kchao and G. L. St ber, Analysis of Direct-Sequence Spread-Spectrum Cellular u Radio System, IEEE Trans. Commun, vol. COM-41(10), pp. 15071516, October 1993.
. Bibliography
176
[99] S. Verd , Minimum Probability of Error for Asynchronous Gaussian Multipleu Access Channels, IEEE Trans. Inform. Theory, vol. 32(1), pp. 8596, January 1986. [100] S. Verd , Multiuser Detection, in Advances in Signal Processing (H. V. Poor and u J. B. Thomas, eds.), vol. 2:Signal Detection, pp. 369409, Greenwich,CT: JAI Press, 1992. [101] H. V. Poor, Signal Processing for Wideband Communications, IEEE Information Theory Society Letter, vol. 42(2), pp. 110, June 1992. [102] A. Duel-Hallen, J. M. Holtzman, and Z. Zvonar, Multiuser Detection for CDMA Systems, IEEE Personal Communication Magazine, vol. 2(2), pp. 4658, April 1995. [103] G. Turin, The Effects of Multipath and Fading on the Performance of DirectSequence CDMA Systems, IEEE J. Select. Areas Commun., vol. 2, pp. 597603, July 1984. [104] M. Kavehrad and B. Ramamurthi, Direct-Sequence Spread Spectrum with DPSK Modulation and Diversity for Indoor Wireless Communications, IEEE Trans. Commun., vol. COM-35(2), pp. 224236, February 1987. [105] M. M. J. Wang and L. B. Milstein, Predetection Diversity for CDMA Indoor Radio Communications, in Virginia Tech. Symposium on Wireless Personal Commun., (Blackburg, VA), pp. 13.113.10, June 1992. [106] D. J. Torrieri, Performance of Direct-Sequence Systems with Long Pseudonoise Sequences, IEEE J. Select. Areas Commun., vol. JSAC-10(4), pp. 770781, May 1992. [107] V. U. Reddy, B. E. Gardt, and T. Kailath, Least Squares Type Algorithm for Adaptive Implementation Pisarenkos Harmonic Retrieval Method, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-30(3), pp. 399405, March 1982.
. Bibliography
177
[108] D. R. Fuhrmann, Adaptive MUSIC, in Proc.SPIE Conf., (San Diego, CA), 1987. [109] I. Karasalo, Estimating Covariance Matrix by Signal Subspace Averaging, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-34(1), pp. 812, January 1986. [110] J. R. Bunch and C. P. Nielsen, Updating the Singular Value Decomposition, Numerische Math., vol. 31, pp. 111129, 1978. [111] R. Padovani, Reverse Link Performance of IS-95 Based Cellular Systems, IEEE Personal Communications Magazine, vol. 1, pp. 2834, Third Quarter 1994. [112] S. Haykin, Adaptive Filter Theory. Englewood Cliffs, NJ: Prentice Hall, second edition ed., 1991. [113] K. S. Gilhousen, I. M. Jacobs, R. Padovani, A. Viterbi, L. A. Weaver, and C. Wheatly, On the Capacity of a Cellular CDMA System, IEEE Trans. Veh. Tech., vol. VT40(2), pp. 303312, May 1991. [114] A. M. Viterbi and A. J. Viterbi, Erlang Capacity of a Power Controlled CDMA System, IEEE J. Select. Areas Commun., vol. 11-(6), pp. 892900, August 1993. [115] A. M. Viterbi, A. J. Viterbi, and E. Zehavi, Performance of Power-Controlled Wideband Terrestrial Digital Communication, IEEE Trans. Commun., vol. 41-(4), pp. 559569, April 1993. [116] D. J. Goodman and S. X. Wei, Efciency of Upacket Reservation Multiple Access., IEEE Transactions on Vehicular Technology, vol. 40-(1), 1991. [117] W. Wong, C.-E. W. Sundberg, and N. Seshadri, Shared Time Division Duplexing: an Approach to Low-Delay High-Quality Wireless Digital Speech Communications, IEEE Transactions on Vehicular Technology, vol. 43-(4), pp. 934944, November 1994.
. Bibliography
178
[118] R. Padovani, The Capacity of CDMA Cellular: Reverse Link Field Test Results, in Proc. International Zurich Seminar on Digital Communications, (Zurich, Switzerland), pp. 5666, March 1994. [119] D. C. Cox, Wireless Network Access for Personal Communications, IEEE Communications Magaine, vol. 30(12), pp. 96115, December 1992. [120] L. M. A. Jalloul and K. Rohani, CDMA Forward Link Capacity in a Flat Fading Channel, in Virginia Tech.s Fifth Symposium on Wireless Personal communications, (Virginia Tech, Blacksburg, Virginia), pp. 7.17.11, 1995. [121] A. J. Viterbi, Very Low rate Convolutional Code for Maximum Theoretical Performance of Spread-Spectrum Multiple Access Channels, IEEE J. Select. Areas Commun., vol. JSAC-8(40), pp. 641649, May 1990. [122] K. I. Kim, On the Error Probability of a DS/SSMA with a Noncoherent M-ary Orthogonal Modulation, in Proc. VTC92, vol. I, (Denver, Colorado), pp. 482485, 1992. [123] Q. Bi, Performance Analysis of a CDMA Cellular System, in Proc. VTC92, vol. I, (Denver, Colorado), pp. 43486, 1992. [124] Q. Bi, Performance Analysis of a CDMA Cellular System in the Multipath Fading Environment, in Proc. IEEE Int. Conf. Personal, Indoor and Mobile Radio Commun. PIMRC92, vol. I, (Boston, MA), pp. 108111, 1992. [125] L. M. A. Jalloul and J. M. Holtzman, Performance Analysis of DS/CDMA with Noncoherent M-ary Orthogonal Modulation in Multipath Fading Channels, IEEE J. Select. Areas Commun., vol. JSAC-12(5), pp. 862870, September 1994.
. Bibliography
179
[126] A. Jalali and P. Mermelstein, Effect of Diversity, Power Control, and Bandwidth on the Capacity of Microcellular CDMA Systems, IEEE J. Select. Areas. Commun., vol. JSAC-12(5), pp. 952961, June 1994. [127] S. Ariyavisitakul and L. F. Chang, Signal and Interference Statistics of a CDMA System with Feedback Power Control, IEEE Trans. Commun., vol. COM-41(11), pp. 16261634, November 1993. [128] J. M. Holtzman, A Simple, Accurate Method to Calculate Spread-Spectrum Multiple Access Error Probabilities, IEEE Trans. Commun., vol. COM-40(3), pp. 461 464, March 1992. [129] L. F. Chang and N. Sollenberger, Comparison of Two Interleaving Techniques for CDMA Radio Communications Systems, in Proc. VT92, vol. I, pp. 275278, 1992. [130] T. Fujii and S. Kozono, Received Signal Level Characteristics with Adaptive Transmitter Power Control in Mobile Communications, Electronics and Communications in Japan, Part 1, vol. 73-(9), pp. 7685, 1990. [131] S. Ariyavisitakul, Signal and Interference Statistics of a CDMA System with Feedback Power Control-Part II, IEEE Trans. Commun., vol. COM-42(2/3/4), pp. 597 605, February/March/April 1994. [132] A. DAvella, D. Marizza, and L. Moreno, Power Control in CDMA Systems: Performance Evaluation and System Design Implication, in Intern. Conf. on Universal Personal Commun., (San Diego), pp. 7377, 1994. [133] Qualcomm Inc., An overview of the Application of Code Division Multiple Access (CDMA) to Digital Cellular Systems and Personal Cellular Networks, Tech. Rep. EX60-10010, Qualcomm, Inc., May 1992.
. Bibliography
180
[134] L. Popken and W. Kriedte, Statistical Description of Non-Coherent Automatic Gain Control, International Journal of Satellite Communications, vol. 11, pp. 8186, March-April 1992.