Flexible Signal Processing Algorithms For Wireless Communications
Flexible Signal Processing Algorithms For Wireless Communications
Flexible Signal Processing Algorithms For Wireless Communications
, Virginia Polytechnic Institute and State University (1996) Submitted to the Department of Electrical Engineering and Computer Science in partial ful llment of the requirements for the degree of Doctor of Philosophy in Electrical Engineering at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY June 2000 c Massachusetts Institute of Technology 2000. All rights reserved.
Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Department of Electrical Engineering and Computer Science May 10, 2000 Certi ed by. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . John V. Guttag Professor and Department Head, Electrical Engineering and Computer Science Thesis Supervisor Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Arthur C. Smith Chairman, Department Committee on Graduate Students
Abstract
Acknowledgments
I wish to thank my advisor, John Guttag, for his encouragement, ideas and support in bringing this work to fruition. I am grateful to all of my present and former colleagues in the Software Devices and Systems Group for their help and ideas throughout the course of my work. My work builds upon the work of many others, including Vanu Bose and Mike Ismert. I am particularly grateful to John Ankcorn for his helpful insights and feedback on many di erent aspects of this work, and even for all of his questions that led me to think about parts of this work in new ways. Some of this work was done in conjunction with David Karger and Rudi Seitz, who have helped me to learn a little bit more about how computer scientists think and solve problems. I hope that they have also learned a little bit about signal processing. I would also like to acknowledge the support of the National Science Foundation through a graduate research fellowship that has made much of my graduate studies possible. Finally, I am forever grateful to my loving wife and children for their understanding and support during these years that we have spent at MIT, and especially to my Lord Jesus Christ who has blessed me with such a wonderful family and so many rewarding opportunities.
Contents
1 Introduction 2 Flexible Design of Wireless Systems
2.1 Signal processing in the Physical Layer . . . . . 2.2 E ects of Mobility on the Wireless Channel . . 2.2.1 Propagation characteristics and capacity 2.2.2 Dynamic conditions . . . . . . . . . . . . 2.3 E ects of Diverse Communications Services . . 2.4 Design of Wireless Communications Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15 23
23 25 25 27 28 28 32 32 33 34 35 40 42 43 46 47 48 51 52 54 54 56 58 59 59
3.1 Key Elements for Flexible Algorithm Design . . . . . . 3.1.1 Flexibility to support function and performance 3.1.2 Understanding explicit data dependencies . . . 3.1.3 Techniques for e ciency with exibility . . . . . 3.2 Related Work . . . . . . . . . . . . . . . . . . . . . . .
31
4.1 Overview: The Digital Modulator . . . . . . . . . . . . . . 4.2 Conventional Digital Waveform Generation . . . . . . . . . 4.2.1 Conventional techniques for digital modulation . . . 4.2.2 Direct digital synthesis of a sinusoidal carrier signal 4.3 Direct Waveform Synthesis for Modulation . . . . . . . . . 4.3.1 A new approach to modulation: Direct mapping . . 4.3.2 Decomposing tables to reduce memory . . . . . . . 4.3.3 Extensions of direct waveform synthesis . . . . . . . 4.4 Performance and Resource Trade-o s . . . . . . . . . . . . 4.4.1 Performance comparison . . . . . . . . . . . . . . . 4.4.2 Memory versus computation . . . . . . . . . . . . . 4.5 Empirical Performance Evaluation . . . . . . . . . . . . . . 4.6 Implementation of a DWS Modulator . . . . . . . . . . . . 4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
39
CONTENTS
5.1 Overview: Channel Separation . . . . . . . . . . . . . . . 5.2 Conventional Approaches to Channel Separation . . . . . 5.3 A New Approach to Frequency Translation . . . . . . . . 5.3.1 Implementation of a narrowband ltering system 5.4 A New Approach to Bandwidth Reduction . . . . . . . . 5.4.1 Decoupling lter length and SNR . . . . . . . . . 5.4.2 Random sub-sampling . . . . . . . . . . . . . . . 5.4.3 Model for analysis . . . . . . . . . . . . . . . . . 5.4.4 Analytical evaluation of the sequence transformer 5.5 Evaluation of Random Sub-sampling . . . . . . . . . . . 5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 Overview: Symbol Detection . . . . . . . . . . . . . 6.2 Conventional Approach to Detection . . . . . . . . 6.2.1 The design of pulses for digital modulation . 6.2.2 The matched lter detector . . . . . . . . . 6.3 E cient Detection for Error Control . . . . . . . . 6.3.1 A general framework for detection . . . . . . 6.4 Detection for Controlled Error Probability . . . . . 6.4.1 Data-e cient decisions . . . . . . . . . . . . 6.4.2 A generalized threshold test . . . . . . . . . 6.4.3 Detection using multiple tests . . . . . . . . 6.5 Performance of a two-test detector . . . . . . . . . 6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
64 68 71 73 73 74 76 78 89 92 96
97 99 100 102 105 105 108 108 111 112 116 122
97
7.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 Investigation of Additional Physical Layer Functions 7.2.2 Flexible System Design . . . . . . . . . . . . . . . . . 7.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . .
123
List of Figures
2-1 A wireless communications system. . . . . . . . . . . . . . . . . . . . 3-1 Performance pro les for (a) a conventional algorithm and (b) anytime algorithm (adapted from Zilberstein, 1996]). . . . . . . . . . . . . . . 4-1 Stages of processing within the Channel Encoder . . . . . . . . . . . 4-2 Typical modulation techniques for encoding bits into a continuous waveform: (a) amplitude modulation, (b) frequency modulation, and (c) phase modulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3 Two symbol constellations used to map blocks of four bits into complex symbols for PAM, commonly known as (a) 16-QAM and (b) 16-PSK. 4-4 Conventional QAM modulator . . . . . . . . . . . . . . . . . . . . . . 4-5 Typical sin(x)=x pulse shape for a bandwidth-e cient system, where x = t=T and T is the symbol interval. . . . . . . . . . . . . . . . . . 4-6 Proposed QAM modulator . . . . . . . . . . . . . . . . . . . . . . . . 4-7 Section of a synthesized 4-PAM waveform with Ns = 8 samples per symbol interval. Individual table entries are indicated by boxes. . . . 4-8 QAM modulator using parallel look-ups to reduce table size. . . . . . 4-9 Plots showing memory requirements versus number of symbol periods per table entry for (a) 8-PAM and (b) 16-QAM waveform synthesis with K = 6, Ns = 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-10 Graphical user interface for an implementation of a direct waveform synthesis digital modulator. . . . . . . . . . . . . . . . . . . . . . . . 4-11 Received signal constellation diagrams for (a) 8-PSK under relatively high SNR (b) 8-PSK under low SNR and (c) 4-PSK under low SNR. . 5-1 Required processing steps in a digital receiver. . . . . . . . . . . . . . 5-2 Typical processing for narrowband channel selection. . . . . . . . . . 5-3 Conceptual steps used for conventional approach to narrowband channel selection: (a) frequency translation, (b) bandwidth reduction, and (c) sample rate reduction. . . . . . . . . . . . . . . . . . . . . . . . . 9 24 36 39 41 44 45 48 49 50 52 57 60 61 63 66 67
10
LIST OF FIGURES
5-4 Block diagram showing frequency translation of desired signal before ltering and decimation. . . . . . . . . . . . . . . . . . . . . . . . . . 68 5-5 Block diagram showing (a) frequency translation of desired signal before ltering and decimation, and (b) new approach that uses a composite lter to reduce computation. . . . . . . . . . . . . . . . . . . . 71 5-6 Illustration of how we decouple the length of the lter input region from the number of samples that it contains. . . . . . . . . . . . . . 75 5-7 Diagram showing conventional FIR lter and model for approximating output samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 5-8 Distortion due to discarding input samples according to random coin ips. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 5-9 Model for analysis of error variance due to random sub-sampling. . . 82 5-10 Random sub-sampling using a transformed sequence for sample selection. 85 5-11 Output error variance relative to desired signal power versus proportion of input samples used for several cases of output signal relative bandwidth. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 5-12 Comparison of Case I and Case II results for three values of var(xn). 95 6-1 Diagram of the channel decoder showing division of the symbol detector into synchronization and detection steps. . . . . . . . . . . . . . . . . 98 6-2 Plot of bit-error rate versus SNR for a two level PAM system. . . . . 99 6-3 Representation of (a) an isolated pulse satisfying the zero ISI condition, (b) multiple scaled and shifted pulses and (c) the noise-free composite waveform. (In the plots, T = Tb, the symbol interval.) . . . . . . . . . 101 6-4 Cascade of transmit pulse-shaping lter and receive lter whose combined response satis es the Nyquist Criterion for zero ISI. . . . . . . 103 6-5 Projection of noisy received vector onto the line connecting the two possible transmitted points. . . . . . . . . . . . . . . . . . . . . . . . 104 6-6 Conditional PDFs for H=0 and H=1. . . . . . . . . . . . . . . . . . . 105 6-7 Footprints of symbols within the sample sequence. . . . . . . . . . . . 107 6-8 Conditional PDFs for Un conditioned on H = 1 for several values of n, where n1 < n2 < n3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 6-9 Conditional PDFs for H = 0 and H = 1 and the three decision regions de ned by: (un < (;T )), (;T < uN < T ), and (T < uN ). . . . . . . . 112 6-10 Plot of the six di erent regions for the di erent combination of outcomes of the two tests in the S1 -S2 plane. . . . . . . . . . . . . . . . . 115 6-11 Plots of raised cosine and rectangular pulses (a) time domain and (b) cumulative energy in sorted samples. . . . . . . . . . . . . . . . . . . 117 6-12 Plot of BER versus number of samples used in decision for various levels of SNR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
LIST OF FIGURES
11
6-13 Modi ed performance curve for binary detection using minimum number of samples to achieve bounded BER. . . . . . . . . . . . . . . . . 119 6-14 Lowest threshold value for potential values of ki. . . . . . . . . . . . 120 6-15 Expected number of samples required for each bit decision using a two-test detector. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
12
LIST OF FIGURES
List of Tables
4.1 Required operations for conventional QAM modulator . . . . . . . . . 4.2 Required operations for proposed DWS modulator . . . . . . . . . . . 4.3 Results of software implementations of conventional QAM modulator and DWS modulator. . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 55 58
13
14
LIST OF TABLES
Chapter 1 Introduction
Communications- anytime, anywhere. This mantra is often repeated as designers and producers of communications networks advertise new capabilities and services available because of recent technological developments. Indeed, with the proper equipment, we can communicate voice, data and images anywhere in the world (or out of it). An important part of this universal communications network will, of course, be the portions that provide mobility and access to the global infrastructure without wires. Today, it is this wireless portion of the system that seems to lag behind expectations. To be sure, wireless communications systems have changed signi cantly in the past ten years. Most notable, of course, is the sheer number of people that use wireless communications services today. This trend of increased usage is expected to continue and has generated considerable activity in the area of communications system design. Another clear trend has been the type of information that these wireless systems convey: virtually all new wireless systems are digital communications systems unlike earlier broadcast and cellular systems, newer systems communicate using digital data encoded into radio waves. Even analog source information such as voice and images are digitized and then transmitted using digital formats. The reasons for this shift to digital wireless systems are several. Initially, the shift was made in cellular telephone systems to improve system capacity through more e cient usage of the limited radio frequency (RF) spectrum. This shift has also made available more advanced features, better performance, and more security for users. Another reason for the shift in future systems, however, will be the fact that almost all information communicated through such systems will already be digital, both between individual users and between computers. This shift toward digitizing all information for transmission does not mean that all data should be treated as equivalent. Another trend in future wireless communications should be di erentiated support for heterogeneous tra c, such as voice, data
15
16
CHAPTER 1. INTRODUCTION
and video. Di erent types of data tra c will have varying requirements for transmission through the communications system, including di erent requirements for latency, error performance and overall data rate. At the same time that usage becomes more widespread, users will also desire better performance: not only higher data rates for future applications, but also better reliability and coverage. In fact, users will want these systems to work as well as conventional wire-line systems, just without the wires. The desire for wireless connectivity will include not only increased total area of coverage, but also the ability to transparently move within zones, maintaining reliable connectivity using lightweight, low-power communication devices. This will lead to a wider range of operating conditions within the wireless channel due to mobility and the requirement for communications services in diverse environments. In order to meets these demands and ful ll the vision of universal connectivity, wireless systems of the future will have to provide a level of exible and e cient service not seen in current systems. These systems will have to carefully manage limited resources such as spectrum and power as they provide services and performance far superior to any available today. Furthermore, this increased exibility will have to extend to those layers of the system that interface with the wireless channel: the physical layer. The algorithms that perform the processing in this layer are directly impacted by the dynamic conditions and increased demands for e ciency and performance implied by the trends above. These algorithms will have to provide exibility in the way that they accomplish their work under changing conditions in the wireless channel, yet they will have to do this in a way that provides e cient use of power exceeding the best systems of today. In this work, we demonstrate that it is possible to design algorithms that provide both exibility and e ciency to meet the challenges that lie ahead. In today's wireless communications systems, the physical layer processing is typically implemented as a static design, providing an abstract interface to the upper layers of the system as simply a bit transmission medium with some level of uncertainty at the destination. Fixed hardware implementations reinforce this view of the physical layer as immutable. In this work, we consider the implication of recent technological advances that make it possible for large portions of the physical layer processing to be implemented in software Mitola, 1995]. This shift leads us to consider designing communications systems in which the physical layer implementation can be signi cantly modi ed in response to changes in the operating environment or the needs of the end applica-
17 tions and users. A major conclusion of this thesis is that through careful design, we can produce exible signal processing algorithms that enable the overall system to e ciently respond to such changes. An important part of this work has been the implementation of many of the resulting algorithms in a prototype wireless system. These implementations have helped not only to validate the results, but have also helped to shape and guide the research itself by providing a sense of the important problems and opportunities for further investigation. Together, these goals of understanding how to best design and implement a wireless communication system in which the physical layer can be controlled in response to dynamic conditions and requirements has led to a number of signi cant results. The original contributions can be classi ed into two separate categories: 1. Novel signal processing algorithms that provide both exibility and improved e ciency relative to conventional techniques. 2. A new approach to designing exible signal processing algorithms. We describe these two areas of contribution in more detail in the sections that follow.
18
CHAPTER 1. INTRODUCTION
{ The frequency shifting lter provides exible and e cient frequency trans-
lation with a computational complexity proportional to the output sample rate of the channel lter. The computation relative to conventional techniques is reduced by a factor equal to the ratio between input and output bandwidths of the channel separator, which can be several hundred or more in modern wideband receivers. { The channel separation process is computed using random sub-sampling of discrete waveforms to enable e cient ltering. This new new approach uses only a subset of the available input samples while maintaining controlled signal-to-noise levels at the output. The computational complexity is again decoupled from the input sample rate, and depends instead on the required output signal quality.
threshold matched lter detector that can provide a more e ective balance between computation and the output con dence levels needed for e cient overall system performance. Computational reductions of ve or ten fold are demonstrated relative to a full matched lter detector, depending on the required output quality and input noise levels.
In the design of a signal processing algorithm, a typical goal is to produce an algorithm that can compute a speci c result using some minimum amount of computation. From a systems perspective, however, it is appropriate to design algorithms that can help provide e ciency in the context of the entire communications system. This approach is especially important in the case of systems that will be expected to provide di erent services in a wide range of operating conditions. Several common themes emerged from this work that have proven useful in producing e cient, exible algorithms. These themes can be viewed as a general approach to the development of e ective signal processing algorithms in a larger communications system. The steps of this approach are: (1) Identify speci c modes of exibility that are useful in providing overall system e ciency. (2) Identify explicit relationships between input and output data samples for each processing function, and (3) E ciently develop algorithms using the results of (1) and (2) through the removal of unnecessary intermediate processing steps and the use of other tech-
19 niques borrowed from other disciplines, e.g. approximate and randomized algorithms. We must identify speci c modes of exibility needed for the overall communications system to adapt to changing conditions. We can also typically improve performance by eliminating types of exibility that are unnecessary for e cient global adaptation. An example of this is the matched lter detector described above. Because of mobility and dynamic channel conditions, the receiver often experiences varying levels of signalto-noise ratio (SNR). In the context of the entire system it is desirable for the detector to produce constant quality output (measured as bit-error rate) in the presence of variable input signal quality. In Chapter 6, we demonstrate a detector that can do this e ciently by reducing its computation under favorable SNR conditions. We need a more explicit understanding of the input to output data relationships for speci c functions. To achieve the goal of exibility with e ciency, it is important to understand which speci c input samples are relevant to the computation of a particular output value, as well as their relative importance. In the channel separation process discussed in Chapter 5, this understanding leads us to decouple the length of a lter response from the number of samples that are used to compute the output value. In the case of the detector from Chapter 6, this knowledge enables us identify the set of samples that contain information about speci c bits to be estimated, and to to preferentially process those input samples that contribute most to a useful result. The nal development step combines the understanding of the data relationships and the targeted modes of exibility. In some cases, the role of a particular processing function within the overall system enables a designer to specify a minimum level of quality that is suitable for e cient global operation under current conditions. An approach that provides the ability to simply approximate a more optimal computation in a exible way might then lead to more e cient algorithm. Because of the inherent statistical properties of noisy signals, we found that it is useful in such cases to apply techniques that provide statistical, rather than deterministic, performance guarantees as a way to achieve more e cient resource usage for desired levels of output quality. This is true for the detector described above and for the random sub-sampling scheme that approximates the output of the channel separation lter using statistical techniques. In other cases, where approximations to ideal output values are not appropriate, it is often possible to use the explicit data relationships to provide improved e ciency. In Chapters 4 and 5, we describe digital modulation and frequency translation approaches that produce results identical to conventional approaches while using signi cantly less computation. In these cases, understanding data relationships enabled the removal of unnecessary computation steps and the creation of more e cient algorithms without sacri cing any useful modes of exibility.
20
CHAPTER 1. INTRODUCTION
21 reduction. The rst algorithm presented is a composite digital lter that performs part of the frequency translation step as the channel lter output is computed. This results in a signi cant reduction in the computation required for the frequency translation step of channel separation. We then present an approach to the bandwidth reduction step that is designed to balance the output precision needs of the overall system with the desire to minimize computational complexity. The key idea is to compute the output of the channel lter using only a subset of the available input samples. We describe algorithms for choosing this subset of input samples in a non-deterministic manner to prevent narrowband aliasing while carefully controlling the signal-to-noise ratio at the output of the channel lter. One of the signi cant results of Chapter 5 is that the computational complexity of the channel separation function need not depend on the input sample rate. Instead, we demonstrate that the channel separation function can be implemented with a complexity that depends on the output sample rate and the level of interfering signals are present in the wideband input signal at the time the algorithm is run. Another essential function of a digital receiver is detection, where an estimate is made of the original transmitted data based on the samples of the received signal. In Chapter 6, we describe a new approach to detection, where the goal will be to produce estimates with a bounded probability of error using minimum computation. We present an algorithm that can produce such estimates with signi cantly reduced computation relative to conventional detection techniques. These performance gains result from two distinct improvements, both of which involve the evaluation of only a subset of the relevant samples for each data symbol estimate. For typical pulse shapes used to transmit data, there is an uneven distribution of signal energy over the duration of the pulse. We demonstrate that a signi cant reduction in computation is possible with little e ect on error probability by preferentially analyzing only those samples that contain the most signal energy. Further performance gains are then achieved through the introduction of a conditional test that enables the detection process to terminate early under favorable conditions, resulting is a detection algorithm that has a statistical running time. Finally, Chapter 7 summarizes the main contributions and themes of this work and indicates some directions for future work. A review and discussion of related work is presented in various places throughout the thesis.
22
CHAPTER 1. INTRODUCTION
24
Source
Source Encoder
Source Decoder
Error Encoder
Error Decoder
Digital Modulator
Wireless Channel
Signal Isolation
Demodulator
Transmitter
Receiver
Figure 2-1: A wireless communications system. isolating the signal components corresponding to the desired transmitted signal, the demodulator recovers the digital data from these received waveforms. In the gure we see that the higher layer functions, such as source and error correction coding and decoding are separated from the signal processing functions. In the earliest analog wireless communications systems, these signal processing functions encoded analog source signals directly into RF signals, so the source and error coding stages were not present. As systems were developed to communicate digital information, signal processing was still performed on continuous signals in the analog domain. In modern systems, however, signi cant portions of the physical layer processing are now performed using digital signal processing (DSP) techniques. Digital signal processing still involves signals, but they are now sampled and quantized representations of continuous waveforms. The processing of these discretetime signals is performed as numerical operations on sequences of samples. Signal processing has now become a computational problem and systems are designed with ever greater portions of the \physical layer" implemented in the digital domain due to its relative advantages in cost, performance and exibility Frerking, 1994]. When we consider the di erent layers of the wireless communications system in Figure 2-1, we see that the physical layer is typically the largest consumer of resources in the system. For example, several studies have found that wireless network adapters
25
for personal digital assistants often consume about as much power as the host device itself Lorch and Smith, 1998, Stemm and Katz, 1997]. Within the wireless network adapter, almost all of this power will be consumed in the physical layer implementation. This becomes clear when we consider that the upper layers of the protocol stack might require only a few operations per bit in an e cient implementation. In the physical layer, however, the signal processing might require tens of thousands of arithmetic operations to communicate a single bit from source to destination Mitola, 1995].
The physical channel only conveys continuous-valued, analog signals. The design of an e cient communications system will therefore require an understanding of how
26
such signals behave in the wireless channel, not just the abstract view of digital communications as seen at the higher layers of a communications protocol stack. At a fundamental level, the wireless channel is not binary: data are not simply either \received" or \lost". Rather, the relevant phenomena are often smoothly varying and communications is often a matter of degree, of varying levels of con dence. There is a qualitative di erence between the traditional interface presented to the upper layers of the system and the actual limitations due to the properties of the channel. One important e ect seen in the wireless channel is the attenuation of the signal as it moves from transmitter to receiver. The strength of the received signal relative to a xed level of uncertainty present, called noise, is important because it determines the capacity of the channel to convey information. Shannon's equation for channel capacity shows that the ability of a system to reliably communicate information through a xed-bandwidth (W Hz) channel depends on the ratio of the power of the received signal (P ) to the power (N ) of the noise (which is treated as a random signal) Cover and Thomas, 1990]: P C = W log2 1 + N bits=second (2.1) From this we see that the theoretical limit on the capacity (C ) of such communications is not xed, but rather changes with the received signal strength. In free space, the signal attenuation is between transmitter and receiver is proportional to 1=d2, where d is the distance between transmitter and receiver. In a terrestrial system, the attenuation can be much more severe because of e ect of the ground, signal attenuation is often assumed to be proportional to 1=d4, although measurements show that the exponent ranges from 2 to 6 in di erent environments Rappaport, 1996]. In a wireless system the strength of the signal sensed by the receiver can vary signi cantly with range, often by many orders of magnitude. Another e ect of propagation is the attenuation due to objects in the environment. In an indoor environment, the signal passes through walls and oors, causing signi cant reduction in signal strength, often by several orders of magnitude. In an outdoor environment, foliage, rain, contours of the landscape, etc. can cause attenuation of signals traveling from transmitter to receive Rappaport, 1991]. The wireless channel is also a shared medium in which signals are combined, leading to several types of interference. In one case, signals re ect o objects like buildings, walls, trucks, etc., producing multi-path reception. In this situation, multiple copies of the signal combine additively at the receiver. The separate copies of the signal experience di erent propagation delay and attenuation and cause self-interference. The e ect is particularly severe when the time di erential between successive arrivals equals or exceeds the time interval at which successive data symbols are encoded in
27
the RF signal Lee and Messerschmitt, 1994]. Co-channel interference often results from the geographical re-use of the same RF band in multiple locations (e.g. cellular telephone systems). This scheme is used to increase system capacity and relies on careful allocations of di erent frequency bands to limit interference. It is di cult or impossible for a receiver to separate co-channel signals without some additional techniques such as \smart" antennas or spread-spectrum approaches J. C. Liberti and Rappaport, 1999]. Since techniques to isolate signals are imperfect, other transmitters in close proximity can also cause adjacent channel interference. Transmitters that share a channel are typically required to use di erent frequency bands or di erent time periods. However, the ability of a receiver to remove interfering signals is limited by unintentional emissions of transmitters using adjacent frequency bands and the desire to minimize the complexity of the processing at the receiver. All real receiver implementations su er from imperfect signal separation techniques that are a compromise to reduce processing complexity. All of these propagation e ects must be taken into account in the design of a wireless communications system. It is generally the case that operating closer to theoretical limits on capacity in (2.1) requires an implementation with signi cantly more complex processing. Thus, as conditions change the theoretical capacity of a wireless link changes, or conversely, the complexity required to provide a xed rate changes if attenuation or interference conditions improve, a exible system could provide better service with the same complexity or could provide the same service using a less complex processing solution. Similarly, when a particular system is designed to operate at a xed rate under di erent conditions, its xed capacity is an artifact of the implementation, not of the underlying theoretical limits. The environmental factors described above can seriously a ect the performance of a communications system, even in a relatively static situation. More signi cant, however, are the e ects of dynamically changing conditions. This is one of the primary di erences between a xed wireless system and a mobile wireless system. The motion of one or both ends of a transmission link during a communications session can change the propagation range and cause variations in signal levels at the receiver. Relative motion between the transmitter and receiver also causes dynamic multi-path interference conditions that depend on the frequencies of the signals, ranges, relative speed of motion, etc. In large systems, motion during communications can also result in hand-o between di erent areas of coverage. This technique is used to manage capacity and provide increased coverage, but can also result in rapid changes in the attenuation and interference that mobile users experience as they move
28
between coverage areas. Mobility during operation makes the dynamic nature of the wireless channel more severe. Dynamic propagation characteristics a ect the strength and properties of the signal sensed by the receiver, and this has a direct e ect on the theoretical capacity of a wireless communications system, as well as implications for the complexity of the processing required in the receiver.
29
In particular, we are interested in those functions of the communications system such as medium access control, error control, etc. that have a direct impact on the physical layer signal processing. One of the rst steps in the design of a system is to identify the types of end-toend services that will be provided, including types of tra c, access patterns, length of sessions, continuous or intermittent connectivity. The cost of providing these services will depend heavily on the anticipated range of channel conditions and the available resources, such as RF spectrum. In the past, large scale wireless systems were designed to carry homogeneous trafc, o ering uniform service to all users channel conditions could be treated as static by designing to \worst-care" conditions. This approach enabled a static allocation of shared system resources to users and allowed the system to be optimized for a single type of functionality. The complexity of the system was reduced since there was no provision to re-allocate shared resources after the initial design. Future systems will need to support a dynamic mix of heterogenous tra c. There will be no \common case" for which the resource allocation can be optimized and hence any static allocation scheme will involve accepting inherent ine ciencies. The end-to-end principle indicates that resource allocations can best occur when the requirements of an individual user are known, which will be at the time that service is requested. Even in current proposals for future systems, there is an acknowledgement that dynamic allocation of bandwidth to users based on requested services can improve e ciency and should be used Grant et al., 2000]. Dynamic allocation of system resources will have a signi cant impact on physical layer processing, providing the potential to take advantage of the speci c channel conditions and service requirements of individual users in order to more e ciently allocate resources to each. We have seen that dynamic channel conditions can result in \unused capacity" when conditions are better than those for which a static design was made. Recovering this capacity though the use of exible physical layer designs will provide a much-needed improvement in overall system e ciency. On a smaller scale, each mobile device within a larger wireless communications network needs to use its own resources e ciently. The end user in a wireless system has the most information about desired types of service, available resources, and local channel conditions. In future systems, the unpredictable needs of users and changing conditions will make it impossible to optimize for a single type of service. In order to support diverse services, systems of the future will need some scheme to e ciently support dynamic types of tra c and conditions. This must extend to the physical layer, which contains most of the processing that depends on channel conditions and medium access control methods. In the remaining chapters of this thesis
30
we describe a general approach for exible signal processing algorithm development, as well as a number of speci c algorithms that will enable more e cient use of both systems resources (such as RF spectrum) and local resources (such as computation and power).
32
In this chapter, we present some of the major themes that have been signi cant in the algorithm development work that comprises this thesis. These themes can viewed as general approach to DSP algorithm design that has resulted in number of useful algorithms for exible wireless systems. They also highlight some approaches to algorithm design that are relatively uncommon in the design of DSP algorithms. These kinds of approaches are more common in the eld of computer science, particularly in the areas of computer system and computer algorithm design.
33
It is also helpful to understand more precisely how providing exibility within an algorithm can force us to give up e ciency in processing. One way to understand this is to realize that e ciency is often gained when multiple abstract processing steps can be combined for implementation. The principle is widely used in di erent areas of computer science, from computer algorithm design to compiler and computer language design. Providing exible operation a ects performance because it introduces partitions that limit our ability to combine processing steps. To see this, consider a hypothetical processing system that consists of a series of abstract stages. Many factors in uence where we place partitions in an actual implementation we note a few that are relevant to our work. We might need to place partitions in the actual implementation at points where: we need to expose an intermediate result between two stages, or subsequent processing depends on information not known at design time (late binding or re-binding), or there is a need to reduce excessive complexity that would occur with a larger composition of processing. We will see examples of each of these partition decisions as we examine the di erent DSP functions in our wireless communications system. This common step of determining speci c desired modes of exibility is an important rst step of the algorithm design process. As we examine each function in our system, we try to compose functions where possible, but retain modularity to provide desired modes of exibility. We do not have any general rules, other than to say that we determine the desired modes of exibility in each stage by considering the basic functions of each stage in light of the exibility desired of the overall system. A second common theme in our work has been the importance of understanding the relationship between a particular output sample of a processing function and the input samples that are required to compute it. It is important to identify this explicit relationship for several reasons: it allows us to compose processing functions, as described in the previous section, and it also allows us to develop techniques to compute approximate results in appropriate situations. It seems rather obvious that we would need to understand how each output depends on individual input samples as we design signal processing algorithms, but this relationship is not always clear at the outset. DSP algorithms are often developed
34
using analytical techniques based on frequency domain representations of the signals and processing systems, and these often require the assumption that the signals are of in nite duration. These design techniques tend to lead to a \stream-based" view of processing. In this view, the input-output relationship for a particular processing stage is characterized under the assumption that input and output will be in nite sequences of samples that represent continuous signals. This stream-based view tends to blur the relationship between individual elements of the output sequence and the speci c elements in the input sequence on which each output value depends. This viewpoint of data processing is adequate in situations where processing is repetitive and no conditional behavior is required. It does not provide much insight, however, in situations where it might be useful to specify di erent types of processing under di erent situations. This is precisely the type of processing that we would like to consider in our search for exible algorithms. \One size ts all" processing might be well matched to an approach where static conditions exist or are assumed, but makes it di cult to exploit opportunities to reduce computation when favorable conditions allow. The nal step of our approach uses the results of the rst two steps to design algorithms that can e ciently provide the exibility we desire. This step has taken two general forms in our work. In some cases, a clear understanding of input-to-output data relationships has led to the conclusion that current approaches using multiple processing steps can be modi ed to use fewer steps, yielding improved e ciency. In Chapter 4 we describe direct waveform synthesis, a technique for synthesizing digital modulation waveforms using a single mapping step implemented with a look-up table. This technique provides signi cant computational advantages over conventional approaches as it removes unnecessary intermediate steps that provide no useful exibility to the overall system. The new approach provides a useful form of exibility in its use of resources through the ability to use memory to reduce computational requirements. In Chapter 5 we demonstrate a similar improvement for the frequency translation step of a wideband receiver. In this case, two steps of frequency translation and digital ltering are combined to reduce computational complexity. The resulting algorithm retains the ability to provide ne control of frequency translation necessary in a digital receiver system. In other signal processing functions, we demonstrate that improved e ciency can be achieved by designing algorithms that produce approximate results with reduced computation when possible. In the random sub-sampling scheme for channel ltering in Chapter 5 and in the novel detection algorithm presented in Chapter 6 we are able
35
to improve e ciency by exploiting techniques that provide only statistical guarantees of required computation. Many signal processing functions have inherent statistical behavior because they involve random signals. In real-time DSP systems, however, the algorithms themselves are usually deterministic: they perform pre-determined processing on signals with some assumed statistical properties. This processing will produce a result that will have some statistical guarantee of performance, not because of the algorithm itself, but because of the random properties of the input signals. Deterministic algorithms are used because they ensure zero computational variance, allowing implementations to provide performance guarantees in real-time systems so that processing will always be completed by a processing deadline Winograd et al., 1996]. Unlike DSP, computer algorithms often have conditional behavior that depends on the actual input data. This leads to algorithms that, unlike traditional DSP algorithms, have statistical running-time performance guarantees. We have identi ed a number of such cases where these techniques can provide improved performance. The matched lter detector we present in Chapter 6 reduces computation by providing a possibility of early termination and therefore does not have a deterministic running time. In Chapter 5, our sub-sampling scheme that uses randomness in processing to break up patterns and allow reduced computation. The use of randomness means that we will only know the expected amount of computation required to provide some desired level of output quality.
Approximate Processing
Work by Zilberstein Zilberstein, 1993] and Zilberstein and Russell Zilberstein, 1996] addresses the problem of a system operating in a real-time environment where the system is required to perform some type of deliberation prior to performing an action. In particular, their work addresses the case where the time required to select an optimal action degrades the system's overall utility, requiring the trade-o of decision quality for deliberation cost. This work focuses on the use of anytime algorithms that allow a variable execution time to be speci ed to provide a time/quality trade-o .
36
Output Quality
Resource Usage
(a)
(b)
Figure 3-1: Performance pro les for (a) a conventional algorithm and (b) anytime algorithm (adapted from Zilberstein, 1996]). The di erence between these algorithms and conventional algorithms is clearly seen in a performance pro le that quantify the speci c time/quality tradeo for each. In gure 3-1, for example, we see a conventional algorithm in (a) where the output quality remains zero until the algorithm nishes computing the complete result. In (b), however, we see an algorithm for which the output quality gradually increases from zero to maximum as more computation time is expended. While performance pro les can be found for any algorithm, only algorithms with pro les that facilitate the trade-o of output quality for computational resource usage are useful for approximate processing. Work by Nawab et. al. Nawab et al., 1997] brings the concepts of approximate processing to the area of digital signal processing. They note that DSP seems to be a good application of the concepts of approximate processing because there exists a rich set of tools for quantifying the performance of DSP algorithms. These tools distinguish DSP from some other types of computational problems in that there are very precise ways to quantify and compare the output of DSP algorithms. Additionally, there has been a great deal of analysis on resource requirements for DSP systems such as arithmetic complexity and memory usage. Their work presents several speci c DSP algorithms developed with an eye toward application in approximate processing systems. They show, for example, that the fast-Fourier transform (FFT) has a natural incremental re nement structure that allows the algorithm output quality (in terms of probability of signal detection) to be improved by evaluating additional stages of the FFT computation. This concept of incremental re nement is seen as a key idea, since re nement of the quality of a DSP algorithm output thorough additional computation ts quite well with approximate processing ideas and many traditional DSP algorithms already display a natural incremental re nement property.
37
A general approach for developing ASP algorithms is found in work by Winograd Winograd, 1997]. The approach is based on the idea of a decomposition of either the input signal or the processing system into multiple components. In one case, a partial result can then be computed by using only a subset of the decomposed signal elements passed through the processing system, alternatively, the original input signal can be passed through selected portions of the decomposed system to develop an output signal with the desired degree of accuracy or precision. Our work in thesis this examines the use of some of these techniques in the context of the physical layer processing in a wireless system. Some of the results that we present in later chapters began with the idea of decomposing the input signal to produce an approximation of the output for a speci c function. We also provide some analytical tools that are helpful in understanding the quantitative e ects of some of these decompositions, as well as some ideas about which of the decompositions techniques proposed in Winograd, 1997] might be most useful. Other relevant results by Ludwig Ludwig, 1997] demonstrate a technique to implement a digital lter that uses a lter structure with variable order to reduce computation in a frequency selective lter. This work demonstrates reduced computation relative to a xed-order lter while maintaining a minimum ratio between passband and stop-band power at the lter output. In this work we also develop a technique to approximate the output of an frequency selective lter, but treat it as an approximation of the input samples stream through sub-sampling instead of approximating the lter itself.
38
ent algorithms to determine which is fastest. This is an example of a system that can improve its global performance in di erent computational environments by adapting to changing conditions. In our work, the goal it to enable this type of system design on a larger scale in a wireless communications system where adapting signal processing algorithms to di erent conditions or resource constraints would help to improve overall system e ciency.
Source
Source Encoder
Error Coder
Digital Modulator
Channel Encoder
40
ple Wicker, 1995, Berlecamp et al., 1987], and will not be further discussed here. Instead, we focus on the second processing step of the channel encoder: the digital modulator. The modulator performs the processing required to transform the data into a form appropriate for transmission through a particular physical medium or channel. After examining the speci c functions required in this modulation step, we review some of the conventional approaches used to perform the required processing in the transmitter. The remainder of the chapter will then be devoted to a presentation of a novel technique for digital modulation that extends some of the ideas of DDS to more complex digital modulation waveforms. We describe how this technique, which we term direct waveform synthesis (DWS), enables the creation of an e cient digital modulator appropriate for many di erent types of wireless systems.
41
(a)
(b)
(c)
"1"
"0" T
"1"
Figure 4-2: Typical modulation techniques for encoding bits into a continuous waveform: (a) amplitude modulation, (b) frequency modulation, and (c) phase modulation. In a digital wireless system, it is useful to think of the modulation process as encoding the digital data into a high frequency signal. In Figure 4-2 we show several common techniques for encoding digital information into a continuous waveform. These various techniques modify, or modulate, a high frequency carrier signal according to the two possible values of the bits of the original data in Figure 4-2, the modi cation is made to either the amplitude, the frequency, or the phase of the sinusoidal carrier signal, respectively. There are also modulation techniques that can modify the carrier waveform to produce more than two distinct patterns. These techniques might use multiple values for amplitude or combine modi cations to both amplitude and phase to create a larger set of distinct patterns to encode data. These approaches can then encode larger alphabets than just zero or one. Instead, they can encode B bits as a single symbol into a section of the carrier waveform by using 2B distinguishable modi cations of the carrier sinusoid. In such a system, the task of the
42
receiver is to distinguish among the 2B di erent possible forms of the signal in order to determine which symbol was transmitted in each speci c time interval. Another point to note regarding Figure 4-2 is that, in each case, the encoding of successive symbols is separated into disjoint time intervals. This is not true in general, and there are very good reasons why the sections of the encoded carrier waveform corresponding to continuous symbols often overlap signi cantly. We examine the implications of these overlapping segments more in the sections to come. Many di erent modulation techniques have been devised for encoding a sequence of input data symbols into a waveform that can be transmitted through a wireless RF channel. Phase shift keying (PSK), frequency shift keying (FSK) and quadrature amplitude modulation (QAM) are just few of the more common techniques. The wireless environment can vary signi cantly for di erent systems, and modulation techniques are designed to provide di erent sets of trade-o s between system performance and resource consumption. Some modulation formats can be optimized for high datarate, power-constrained applications, while others might enable relatively simple and inexpensive receivers Lee and Messerschmitt, 1994, Proakis, 1995]. A exible technique that can support many di erent modulation techniques is clearly useful. If we desire a system that can provide e cient operation in a wide range of environments while supporting di erent applications, it is evident that we may need to change between di erent modulation techniques as we encode data for wireless transmission.
43
s(t) =
1 X
k=;1
ak p(t ; kTb )
(4.1)
Here s(t) is the resulting continuous waveform as a function of time, t. The summing index k indexes the elements in the sequence of symbols, ak , that encode the original data bits. The constant Tb (another term for symbol is baud) is the time interval between consecutive shifts of the pulse as it encodes the data symbols. The sum in (4.1) is over multiple shifted and scaled pulses because the duration of the pulse can, in general, exceed the shift interval. This situation often occurs in systems that are designed to be e cient in their use of RF spectrum a pulse that has a compact
44
Im
1 1 0 0 11 00 11 00 11 00 1 0 11 00 1 0 1 0 1 0 11 00 1 0 1 0 11 00 1 0 1 0 11 00 11 00
Re
Re
1 11 0 00 1 0 1 0 1 0
1 0 1 0 1 0
(a)
(b)
Figure 4-3: Two symbol constellations used to map blocks of four bits into complex symbols for PAM, commonly known as (a) 16-QAM and (b) 16-PSK. frequency-domain representation will tend to have a long time-domain representation. It is helpful to note here that using a set of complex-valued symbols will result in a complex-valued s(t). Of course, a wireless channel can only convey one-dimensional signals: the voltage signal coupled to the antenna for transmission is a real-valued function of time. The complex-valued signals are therefore converted to real-valued signals during the translation to higher frequencies prior to transmission. Because we are speci cally interested in generating discrete sequences, we also de ne a discrete form of (4.1) in which sn is a sequence of uniformly spaced samples of s(t): sn = s(nTs), where Ts is the time interval between samples:
sn =
1 X
k=;1
ak p n ; kNs]
(4.2)
The discrete pulse p n] is a discrete version of the pulse p(t) and the time shift has now become a simple shift in the time index by kNs, where Ns is the (integer) number of samples in the output sequence occurring in each symbol interval, Tb = NsTs. In (4.2) weighted and shifted pulses are combined by simple addition to form the output sequence that will eventually be converted to an analog waveform and transmitted through the channel. The limits on the summation in (4.2) indicate that each output sample may, in theory, depend on an in nite number of symbols. In practice, though, pulse shapes are used that have a nite duration and often simply
45
Figure 4-4: Conventional QAM modulator correspond to truncated versions of much longer pulses that are attractive for other reasons (such as spectral properties) Frerking, 1994, Lee and Messerschmitt, 1994]. It is normal, however, for pulses to overlap even after truncation because they are often still longer in duration than the symbol interval. A typical algorithm for digital modulation is illustrated in Figure 4-4. Here we see that there are several steps in the production of the nal output sequence. The rst step is to map the input bits to a sequence of complex-valued symbols. Next, this symbol sequence is passed through a set of digital lters called pulse-shaping lters which multiply the real and imaginary components of the input symbol values by the sequence p n] representing the desired pulse shape and summing the result as in (4.2). A nal step is often included: the discrete waveform is translated to a higher frequency before conversion to an analog signal. This translation to an intermediate frequency (IF) allows the complex-valued sequence to be transformed to a real-valued sequence prior to conversion to the analog domain Lee and Messerschmitt, 1994]. This operation is equivalent to multiplication of the real and imaginary components by a cosine and sine waveforms with frequency fc (the carrier frequency), as shown in Figure 4-4:
2 3 1 X sIF n] = Real 4ej2 fcnTs ak p n ; kNs]5 (4.3) k=;1 1 1 X X = cos(2 fcnTs) ak r p n ; kNs] ; sin(2 fcnTs) ak ip n ; kNs]
k=;1 k=;1
We assume that p n] is a real-valued pulse and that ak = ak r + jak i are the real and imaginary components of the symbol values.
46
So we see that the conventional approach is characterized by a series of mapping steps: bit to symbols, symbols to discrete waveforms, and baseband waveforms to IF waveforms. We will refer back to this series of mappings as we examine new techniques to generate digital waveforms that provide both more exibility and more computational e ciency. To understand how we can develop new techniques to produce channel waveforms, it is helpful to examine a waveform generation technique known as direct digital synthesis (DDS), which has become increasingly popular in recent years. This technique is well understood and widely used in the design of digital communications systems. In this technique, a sinusoidal signal is synthesized from precomputed values instead of by computing the values of the samples as they are needed. This technique can take several di erent forms, depending on the amount of exibility needed in the waveform to be synthesized. Because a sinusoid is periodic, the single-frequency form of this technique e ciently generates a sinusoid of xed frequency: we simply pre-compute the samples for a single period of the desired sinusoid and then cycle through and continuously output these values to produce a continuous sinusoidal sequence. One requirement for this technique is that we use an integral number of samples per cycle of the sinusoid. In this case, Tc = kTs , the period of the carrier signal is a integral multiple of the sample interval, Ts. This scheme provides no exibility as to the frequency of the sinusoid generated, but it is computationally very e cient. A more powerful scheme, the phase-accumulator technique, uses a table of precomputed values and a phase accumulator to generate sinusoidal sequences of di erent frequencies. This approach uses a circular bu er containing a large number of samples, say 2N , of a single period of a sinusoid. A sinusoidal sequence with period of the form Tc = 2N Ts=k can be generated by outputting every kth sample in the circular bu er. The phase accumulator refers to an accumulator that stores the index value of the current sample. This value corresponds to the phase value of the sinusoid that is being generated. This technique allows a system to avoid computing (and re-computing) in realtime the values of the sinusoidal sequence, sn = cos(2 fcnTs), where fc is the desired carrier frequency, by using the bu er of pre-computed samples. These samples of a transcendental function are particularly di cult to generate in digital logic, and other alternatives to this look-up table approach include approximation with Taylor series or use of the cordic algorithm Frerking, 1994]. The phase accumulator technique is more exible than the single-frequency technique because it can generate sinusoids with di erent frequencies. This additional
47
exibility, however, comes at the cost of reduced performance: we must compute a new index value for each output sample. Although DDS techniques are relatively straight forward, they can provide signi cant computational savings in a digital transmitter relative to a direct-computation approach. Relative to using a high-precision analog oscillator, DDS has the advantage of using simple hardware components it can even be implemented in software.
48
-4T
-3T
-2T
-T
2T
3T
4T
Figure 4-5: Typical sin(x)=x pulse shape for a bandwidth-e cient system, where x = t=T and T is the symbol interval. technique to allow the synthesis of digital waveforms at passband, thereby removing the need to translate the baseband waveform to IF, and also to allow the mapping of multiple symbols at a time, thereby providing more exibility in the computationmemory trade-o .
49
Compute Index
Look-up Table
in Welborn, 1999]. In this work our results will be derived in terms of an arbitrary K. In any case, the pulses used to transmit consecutive symbols will overlap in time because truncation to a single symbol period spreads too much energy. From (4.2) we see that we need sum only over the symbols for which the shifted version of p n] is non-zero: X sn = ak p n ; kNs] (4.4)
k: p n;kNs]6=0
From this result we begin to see how we can pre-compute the output samples sn. Each output sample depends on K input symbols, or BK bits when we use a constellation of 2B symbols. During a single symbol interval, Ns output samples all depend on the same K input symbols, that is, the output waveform in that interval is completely determined by K input symbols. To produce our direct mapping, it is now clear that we can map a sequence of BK input bits to Ns output samples. In this way, we can produce a look-up table that contains all possible symbol-length sequences of output samples by computing these sequences for all possible 2BK sequences of input bits. Once we have produced this table, we can use it to e ciently synthesize a continuous sequence that corresponds to the discrete-time digital waveform. This technique is shown in Figure 4-6. The steps of this direct waveform synthesis technique in its simplest form are summarized here:
50
Figure 4-7: Section of a synthesized 4-PAM waveform with Ns = 8 samples per symbol interval. Individual table entries are indicated by boxes. Step 0: Compute entries of look-up table for all 2BK possible input bit sequences. Prepend bit sequence to be transmitted with B (K ; 1) dummy bits to initialize modulator. Set k = 0. Step 1: Compute index for table entry using bits Bk +1 through B (k +K ). Output Ns samples of the discrete output waveform, sn. Step 2: Set k = k + 1, goto Step 1.
Figure 4-7 shows a typical section of a synthesized waveform where each table entry (indicated by boxes) contains eight samples of a discrete waveform. It is important to note that this algorithm does not produce an approximation to the waveform produced using the conventional technique shown in Figure 4-4. Both schemes produce exactly the same output sequence when they use the same pulse shape and associated parameters. Even the conventional technique must use an approximation to any in nite-length pulse in order to implement the digital pulse shaping lter. Although they are separate table entries, the segments of output samples produced using the DWS technique will always \ t together" to form a smooth, continuous sequence because of the process of using a sliding window of BK input bits.
51
52
11 00 11 00
11 00 11 00 1 0 11 00 11 00 1 0
Delay Delay
Lookup Table
Figure 4-8: QAM modulator using parallel look-ups to reduce table size.
53 (4.5)
X
k: p n;kNs]6=0
where ak = mk ej k is the magnitude-phase form of the complex symbol value. We will assume that the pulse shape p n] is real-valued in (4.5), although we can still use this technique for a complex-valued p n] if we modify the equations properly. In order to synthesize the digital waveform at passband as above, we combine the cosine term and the pulse sequence p n] in (4.5). If we restrict ourselves to using an IF carrier frequency that has an integral number of periods in one symbol interval, fc = m=(NsTs) for some integer m, then we can combine the frequency translation into the pulse shape by de ning pIF (t) = cos(2 fct)p(t). The time-shifted versions of the discrete form of this pulse now become:
sIF n] =
(4.7) To synthesize a waveform at passband, we simply use the same technique as before, but we compute the entries in the look-up table using the composite pulse sequence pIF k n] = cos(2 fcnTs + k )p n].
k: p n;kNs]6=0
k: p n;kNs]6=0
mk pIF k n ; kNs]
54
more options to trade-o between memory and computation in a digital waveform synthesizer. Using this generalization, each table entry now contains M Ns samples and there are 2B (K +M ;1) entries if we use a single table. The total memory space required for B K M; M Ns W bytes the general case using a decomposition is now: D 2 D and the cost of using the table decomposition to reduce memory is still the same: we perform D ; 1 additions for each output sample. One case in which this technique of storing multiple symbol intervals in each table entry is useful is in =4-di erential-quadrature PSK ( =4-DQPSK), which is used in digital cellular radio systems Wiesler and Jondral, 1998]. In this modulation format, two di erent sets of four symbols (with a 45o rotational o set) are used to alternately encode the data. In this situation, the modulator needs to maintain some state to determine which constellation should be used for encoding, and to use two tables for the di erent cases. If the DWS technique is used to produce the waveform, blocks of four bits could be mapped directly to segments that are two symbol-periods in length, and there would be no need to maintain any state or multiple tables.
( + 1)
55
IF Carrier Genera- 2 array-fetches + increment DDS phase accumution lator Translation to IF 2 multiplies + 1 addition
8 > B ; bit shift < > modulo operation : add symbol value
56
Conventional approach:
Index computation: 1 million ( bit-shift + power-of-two modulo operation + addition) per second Table Look-up: 8 million array-fetches/second + 8 million increments/second
Clearly the DWS technique required much fewer operations. Although there are several di erent types of operations involved, there is about a factor of 25 di erence between the total counts for the two: about 20 million operations per second compared to 500+ million operations per second for the conventional approach. Even this higher number seems modest by today's standards, but when we consider scaling to a system that modulates billions of symbols per second, the di erences become more substantial. The trade-o s provided between use of memory and computation through the use of DWS are fairly straightforward. In the most general case, the total amount of memory B K M; required is D 2 D M Ns W bytes and the computation is essentially D ;1 addition operations per output sample. Because of the many parameters in the above equation, we have included a few plots that illustrate the di erent trade-o s available for a typical set of parameters. In Figure 4-9 we show the total memory requirements for both a 8-PAM system and a 16-QAM system. These plots show the trade-o between memory usage and computation using the table decomposition described above to generate sample sequences for multiple (M ) symbols. The computation required in each case shown in the plots is determined by the value of D used to decompose the table (the computation is D ; 1 additions per output sample). In each of these two plots, the most important point is that there are a variety of operating points (indicated by the points where D divides K + M ; 1 on the various lines), allowing the system to choose between the use of memory and computation according to other constraints that might exist in the system.
( + 1)
57
1e+07 D=2
10000
1000
100 1 2 3 4 5 6 7 8 9 10 Length of output sequence in look-up table (in M = number of symbol intervals) (b) 16-QAM with B=4, K=6, Ns = 10, W=2 1e+08 11
1e+07
D=3
D=4
100000
10000
D=8
1000
100 1 2 3 4 5 6 7 8 9 10 Length of output sequence in look-up table (in M = number of symbol intervals) 11
Figure 4-9: Plots showing memory requirements versus number of symbol periods per table entry for (a) 8-PAM and (b) 16-QAM waveform synthesis with K = 6, Ns = 10.
58
Algorithm Transmit Rate Table Look-up 18.9 Mbit/sec (512 Kbytes, D=1) Direct Computation 0.3 Mbit/sec 16-QAM (K=16) Table Look-up 28.3 Mbit/sec (2048 Kbytes, D=4) Direct Computation 1.2 Mbit/sec 16-QAM (K=10) Table Look-up 36.8 Mbit/sec (16384 Kbytes, D=2) Direct Computation 1.9 Mbit/sec 16-QAM (K=4) Table Look-up 72.2 Mbit/sec (512 Kbytes, D=1) Direct Computation 4.6 Mbit/sec 64-QAM (K=8) Table Look-up 38.7 Mbit/sec (128 Kbytes, D=4) Direct Computation 3.6 Mbit/sec Table 4.3: Results of software implementations of conventional QAM modulator and DWS modulator.
59
4.7 Summary
Direct waveform synthesis provides a signi cant performance improvement over conventional approaches to digital modulation. This gain comes from its ability to directly synthesize waveforms using a simple table look-up based on the input bits to the modulator. We have also presented a technique that provides a exible trade-o
60
Figure 4-10: Graphical user interface for an implementation of a direct waveform synthesis digital modulator.
4.7. SUMMARY
61
(a)
(b)
(c) Figure 4-11: Received signal constellation diagrams for (a) 8-PSK under relatively high SNR (b) 8-PSK under low SNR and (c) 4-PSK under low SNR.
62
between computation and memory usage for DWS by allowing the mapping to be implemented using a number of smaller tables instead of a single large table. In the context of a wireless communications system, the digital modulator does not need to provide much exibility during operation. Any change to the modulation format will require coordination between receiver and transmitter, and such changes will therefore tend to be infrequent. DWS provides a good balance of e cient performance, exible implementation properties and the ability to be recon gured if channel conditions or system requirements so dictate.
User
64
separation and symbol detection. Our rst goal is to develop a more fundamental understanding of the functionality required by these two processing steps in the digital domain. We then use this understanding to develop algorithms that provide both exibility and improved e ciency over conventional techniques in a wide range of situations. The channel decoder is traditionally divided into the speci c steps of channel separation and detection because the wireless channel is a shared medium that combines many signals. The distortion caused by the channel takes on several forms: additive interfering signals, random noise, delayed versions of the same signal, etc. We therefore model the input to a digital receiver as a sum of an indexed sequence, sn, representing the signal of interest, with I interfering signals from other nearby transmitters and a random noise component, nn:
rn = sn +
I X i=1
si n + nn
(5.1)
The overall process of channel decoding requires the recovery of the transmitted data from this received signal. Algorithms for this type of analysis are sensitive to interference in the input signal, but often perform well in the presence of only additive random noise. We therefore divide the processing of the received wideband signal into two stages: rst, the channel separation stage removes interfering signals, transforming the input sequence into another sequence that is simply a noisy version of the original signal, from which the detector produces an estimate of the original data encoded by the digital modulator. In the next few sections we discuss the problems of channel separation in more detail and describe current techniques used to solve these problems. We then present several new techniques that enable us to perform the functions required in a channel separator in manner that provides both exibility and greater e ciency than conventional approaches. We discuss the problem of detection, the second stage of processing, in the next chapter.
65
detector to recover the original data sequence from the received signal. Stated another way, the channel separator should produce a sequence containing components in the received signal corresponding to the signal of interest, while removing the interference. To facilitate separation, it is common to restrict, through regulation, the emissions of potential interfering transmitters, so that any interfering signals will occur in disjoint frequency bands in the RF spectrum. This technique for providing multiple access to the physical medium is known as frequency division multiple access or FDMA. Other techniques such as time division multiple access (TDMA) and code division multiple access (CDMA) can also be used to allow multiple transmitters to share the same band of frequencies, but are not addressed here. In FDMA, separating the desired signal from potential adjacent channel interferers is equivalent to extracting from the received sequence only those components that occur within the band of frequencies corresponding to the signal of interest. This separation is accomplished through the use of digital lters that pass signal components in the desired band of frequencies, the passband, and attenuate signal components outside of this band, in the region known as the stopband. In a wideband receiver the use of digital lters to perform channel separation can lead to an implementation with very high computational complexity. For a receiver in which the channel separation is performed in the digital domain, the work required to extract individual channels from the input of a wideband receiver provides a rst-order estimate of the computational resources required for the entire receiver Mitola, 1995, Wepman, 1995]. The reasons for this high computational burden are several. First, a wideband digital receiver will necessarily have a high input sample rate to adequately represent the wideband input signal. In addition, to cleanly extract a narrow band of frequencies, the receiver must accurately resolve those frequencies at the boundary of the band of interest. These two conditions combine to produce a situation that can require large amounts of processing to perform channel separation using conventional techniques. Before addressing this problem, however, we examine the basic steps that are widely used to perform the entire channel separation procedure using digital lters. In Figure 5-2 we see the conceptual steps that are often used to perform channel separation. The results of these same steps on a hypothetical wideband signal (represented in the frequency domain) are shown in Figure 5-3. First, there is a frequency translation step. Here the wideband signal is shifted in frequency to bring the band of interest into the passband of a digital lter and the interfering signals into the stopband. This translation is represented in Figure 5-2 by a multiplication of the received sequence and a complex sinusoidal sequence. The magnitude of the frequency shift is determined by the frequency of the complex sinusoidal sequence. The second step in this process is bandwidth reduction. Here the lter produces
66
Narrowband Output
an output sequence in which the components at di erent frequencies are attenuated or ampli ed according to the frequency response of the lter. Because this output sequence does not contain any signi cant components outside the band of interest, it is now possible to represent the desired signal using a lower sample rate without signi cantly distorting those components within the band of interest. If there were signi cant signal components present outside the band of interest, reducing the sample rate would cause aliasing of those components into the band of interest Oppenheim and Schafer, 1989]. Hence a nal step of sample-rate reduction, or decimation, reduces the rate of samples in the sequence by a constant factor D, the decimation rate. This rate reduction is also important because it reduces the processing load in subsequent stages. Much of the work in this chapter was initially motivated by the desire to develop e cient techniques to perform the steps of the channel separation process in a wideband digital receiver implemented in software as part of the SpectrumWare project. One of the early goals of this project was to implement such a receiver that could extract a 30 kHz channel from a 25 mega-sample per second input sample stream. It soon became clear that conventional approaches to frequency translation and ltering would result in a design that would require too much computation to run in real-time on the 200 Mhz personal computers used in the implementation. Many of the results presented in the remainder of this chapter have helped produce an e cient software implementation of a wideband receiver, which will be described in more detail in a later section. In the next section, however, we rst examine the speci c steps of the channel separation process in more detail and then review some of the techniques that have been developed to perform them.
67
Frequency
Frequency
Frequency
Frequency/D
Figure 5-3: Conceptual steps used for conventional approach to narrowband channel selection: (a) frequency translation, (b) bandwidth reduction, and (c) sample rate reduction.
68
Figure 5-4: Block diagram showing frequency translation of desired signal before ltering and decimation.
yn =
M X m=0
hmxn;m =
M X m=0
hm rn;me;j2
fc (n;m)Ts
(5.2)
where sn = e;j2 fcnTs is a complex sinusoidal sequence with frequency fc (the original carrier frequency) and Ts is the interval between samples. Cascaded frequency translation and ltering is a very common approach and is seen in many digital receiver implementations Frerking, 1994, Mitola, 1995]. It provides the exibility of modifying the amount of frequency translation without re-
69
designing the entire digital lter. In particular, the technique of DDS is often used to provide precise and exible frequency translation. One drawback of this approach is that the frequency translation stage requires a complex multiplication be performed for every input sample prior to ltering (in addition to generating the complex sequence sn). This computational load can become quite high, especially as we consider the design of receivers with wider input bandwidths: the wider input bandwidth requires proportionally higher sample rates and higher computation for frequency translation. For ltering, M + 1 multiply-accumulate operations are required for each output sample, where the output rate is Rout = Rin=D. An example of a wideband digital receiver with typical values for these parameters helps to make these relationships more concrete: A wideband receiver for a cellular telephone base station needs to access 12:5 MHz of spectrum, so we will use Rin = 30 M samples/second. We assume that a narrowband voice channel of 30 kHz would require a lter with about 1800 coe cients Zangi and Koilpillai, 1999] and we will use a decimation factor of D = 600 to produce a complex-valued output sequence with Rout = 50 k samples/second. Using these numbers, the computation required each second to generate the output sequence for a single narrowband voice channel is 50 000 1800 = 90 million complex real multiply-accumulate operations (for ltering), plus at least 30 million complex real multiplications for frequency translation. Many techniques have been used to make this computationally-intensive ltering task more manageable. We now examine a few of these techniques and how each improves e ciency over the conventional approach. One common technique is to use a cascade of multiple FIR lters that perform the bandwidth reduction and sample-rate reduction in several stages. Using this approach, a lower average number of operations per output sample can be achieved Frerking, 1994, Orfanidis, 1996]. The basis for this improvement is that the each of the multiple stages (except the last) computes an intermediate result that is used in computing multiple output samples in order to amortize the computational costs. Another approach, the lter bank, is appropriate when multiple independent narrowband channels are to be extracted from the same wideband input sequence simultaneously. If each of the desired output channels has identical bandwidth and response (just di erent center frequencies) then techniques exist that can exploit the special relationship between the multiple sets of lter coe cients. These techniques can compute the multiple output sequences at a lower average cost than multiple, independent single channel lters Crochiere and Rabiner, 1981, Zangi and Koilpillai, 1999]. This
70
approach is similar to the recursive decomposition approach used by the fast Fourier transform, which can be viewed as a bank of uniformly-spaced frequency-selective lters. A third approach, often used in dedicated digital ltering hardware, is a special lter structure known as a cascade integrator-comb (CIC) lter or Hogenauer lter Baines, 1995, Hogenauer, 1981]. This technique uses a special lter structure with cascaded stages of accumulators and combs that can implement a bandpass lter using no multiplication operations. This approach is e ective where multiple addition operations are more economical than a single multiplication. All of these approaches have several characteristics in common: Each approach statically speci es the lter passband and stopband. In each approach, lter complexity is proportional to input sample rate as receivers are designed with wider input bandwidths, the cost of extracting a constant-width channel increases as well. Finally, each approach needs a larger number of input samples, relative to the direct approach of (5.2), to compute a particular output sample. In a sense, they are less e cient in their use of each input sample relative to the single highorder lter whose coe cients are optimized to provide a desired lter response. This increased input-output dependence is ameliorated by the fact that each input sample is used in the computation of multiple output samples. In the remainder of this chapter, we develop an alternative approach to channel separation that scales more e ciently with the input bandwidth (or Rin ) of a receiver. This requires that we decouple the some of e ects that unnecessarily lead to this computational dependence on Rin . In the next section, we present an new algorithm that performs frequency translation with computational complexity proportional to the output sample rate. Following this, we describe how to separate the bandwidth reduction step into two independent dimensions: interference rejection and SNR improvement. Decoupling can reduce the computational complexity of the bandwidth reduction step, permitting less work to achieve the desired result. We present a technique that uses random sub-sampling of the input sequence to achieve such a reduction. We also present experimental results that demonstrate the algorithm's e ectiveness in performing channel separation in a wideband digital receiver.
71
hn
sn rn yn cn
D
Figure 5-5: Block diagram showing (a) frequency translation of desired signal before ltering and decimation, and (b) new approach that uses a composite lter to reduce computation.
hm ej2 fcmTs
m=0
m=0
72
put sequence (at the lower output sample rate) to place the desired output signal at a center frequency of zero at complex baseband. Although the idea of combining frequency translation with an FIR lter is not new, previous approached required the use of multiple sets of lter coe cients, thereby restricting it usefulness Frerking, 1994]. Our approach combines only the time-invariant part of the translation with the lter coe cients and factors the time-varying portion to accomplish arbitrary frequency translation using a single set of compound coe cients. This technique has several advantages over the multi-stage technique. The most obvious is that we can dramatically reduce the amount of computation required for frequency translation. The generation of the complex exponential sequence and the frequency translation take place at the lower sample rate, Rout , and we have therefore reduced this portion of the computation by a factor of D. We are required to precompute the compound lter coe cients cn one time, but this is trivial when we consider the potential savings as the lter might be used to generate thousand or millions of output samples. We still retain the ability to ne-tune the channel lter to the desired signal because we can vary the frequency after the lter to track any small changes in carrier frequency. The largest part of the frequency translation, the part that we combined with the lter coe cients in (5.3), does not typically need to be changed rapidly, so the increased e ciency of the frequency translating lter comes at the price of a mode of exibility that was unnecessary. Another advantage of this technique has to with the type of multiplication operations that are performed at each point in the processing. In the conventional approach, the received samples are real-valued and when we multiply by the complex sinusoidal sequence, sn, we have a complex-valued output sequence. (This is indicated by the double line connecting the two processing blocks in gure 5-5(a).) The lter coe cients in the conventional approach are real-valued, which is reasonable since this is typically a low-pass lter with a response that has even symmetry. The multiplications in the lter in gure 5-5(a) are thus complex real. If we were to use complex-valued coe cients for the lter hn we would need to perform complex complex multiplications and this would more than double the computation required. In the new approach shown in part (b), the lter coe cients are already complexvalued, so that we must perform real complex multiplications, as before. If we choose to design the lter cm = hm ej2 fcmTs using a complex-valued hn lter, however, we could conceivably reduce the length of the lter by a factor of two while still satisfying the same passband and stopband speci cations Ochi and Kambayashi, 1988, Komodromos et al., 1995]. This complex-valued design would not result in any additional run-time computation (we would still perform real complex multiplications to evaluate the lter output), but we would get up to a 2 reduction in computation due to the reduced length of the lter sequence, cn.
73
In the case of the example wideband base station receiver described in Section 5.2, this new technique allows the complexity of the frequency translation to be reduced by a factor of 600, from 30 million to only 50 thousand complex real multiplies. In addition, the use of a complex-valued design for the initial lter hn might reduce the complexity of the ltering by as much as a factor of two, from 90 million to 45 million multiplies. The overall reduction in computation is thus over 60% in this particular application. The disadvantage of this algorithm relative the conventional approach is that it requires that the composite lter coe cients be computed before ltering, or recomputed to tune the lter to a signi cantly di erent carrier frequency. In a system with su cient memory to store multiple lter de nitions, selecting a di erent carrier frequency for the lter would simply require using a di erent set of pre-computed lter coe cients. The frequency translation techniques described above were implemented in a number of wireless applications developed in the SpectrumWare Project. One such application was a narrowband demodulator for the analog AMPS cellular telephone system Bose et al., 1999]. The composite lter technique was essential in this implementation, enabling it to perform real-time separation of a 30 kHz channel from a 10 MHz wideband input, in addition to FM demodulation and audio processing. The same technique was used in a number of other applications, including a study of e ective ways to scale the channel separation applications to a multi-processor platform demonstrated linear performance improvements in a multi-processor system through the use of an architecture that separated data management and control functions from the signal processing functions Vasconcellos, 1999]. This work demonstrated that a 4-processor Pentium III personal computer could perform channel separation for up to 32 narrowband AMPS channels simultaneously.
74
required to improve or maintain the output SNR. In a receiver where we extract a narrowband signal from a wideband sample stream, it is often necessary to specify a sharp transition between passband and stopband to reject adjacent channel interference. This requirement often leads to a large number of coe cients in the resulting FIR lter, a direct implementation of which will have a computational complexity proportional to the input sample rate Jackson, 1989]. A second consequence of using a high-order FIR lter is that we can get a signi cant improvement in the SNR of the output signal relative to the input signal. In fact, for every factor of two by which the narrowband signal is initially over-sampled, we can improve the SNR by 3 dB if appropriate ltering is performed to remove out-of-band noise Wepman, 1995]. This means that the ability to reject adjacent channels is related to the length (in time) of the impulse response of the channel lter, while the improvement in SNR due to oversampling and ltering (as well as the amount of computation required) depends on the number of input samples used to compute each output sample. In practice it is the rst e ect, the rejection of potential interferers, that dominates the lter design the choice of lter length then determines the number of input samples used, not the requirement for some minimum output SNR. In order to decouple these two e ects, we present an approach in which a narrowband lter is designed with an impulse response that satis es the ltering requirements, but then we compute the lter output values using only a subset of the available input samples. In particular, we will use a lter with a su ciently long time response to provide the sharp transition we desire, using only as many samples within that response interval as are necessary to produce or maintain the required output SNR. These ideas are shown more clearly in Figure 5-6(a)-(c). In part (a) of this gure we see a case where the number of samples in the input region of the lter is determined by the length of this region (the length of the lter response). In (b), we see where an increase in the length of the lter response will lead to a larger number of samples used when evaluating the lter output. Alternatively, in part (c), we see that we can decouple the number of samples used from the length of the response by using only a subset of the samples in the region corresponding to the lter input region. In the following sections we analyze the usefulness of this approach and present several techniques that can be used to determine how such a subset of samples can be determined for a channel selection system.
Almost all lter design strategies are based on the assumption that a digital lter will use a contiguous block of samples as input, as shown in Figure 5-6 parts (a) and (b). Algorithms to design optimal lters for channel selection applications are no excep-
75
(a)
Filter
Filter (b)
Filter (c)
Figure 5-6: Illustration of how we decouple the length of the lter input region from the number of samples that it contains.
76
tion, and often the optimal design is viewed as the lowest-order lter (smallest number of multiplications) that will provide a desired level of stopband attenuation and passband ripple, regardless of SNR considerations Steiglitz et al., 1992, Jackson, 1989]. We will approach the design of such a lter by using a conventional lter design, and then discard some of the terms in the summation in order to reduce computation while maintaining an acceptable output signal quality. We will treat this discarding of terms as a removal of some of the samples in the input sample stream and we will try to determine which samples and how many samples can be safely discarded. Any time that we simply discard some of the samples in a sample stream we will cause distortion to the signal represented by the sequence. The object of the work presented in this section is to determine whether we can minimize this distortion e ect within the narrow band of frequencies that will be passed through the channel lter. The approach presented here is based on ideas from the area of randomized signal processing presented in Bilinskis and Mikelsons, 1992]. However, in this work we use random choice in a di erent way. We introduce the randomness while processing a stream of uniformly spaced samples, as opposed to introducing the randomness while performing quantization or sampling of the original analog signals. The goal of the channel lter is to remove adjacent channels from the wideband signal so that sample rate can be reduced without causing the aliasing of other interfering signals into band of interest. This is accomplished by designing the lter to reject all potential interfering signals, and only then reducing the sample rate in order to reduce the computational load in subsequent stages. For the remainder of this chapter, we assume that the input to the channel lter is a sequence of uniformlyspaced real-valued samples of a signal with bandwidth W0. We wish to generate an output sequence that contains only those components of this signal that lie within a certain narrow frequency band WN << W0 . This is often accomplished using a decimating FIR lter that passes only the band of frequencies we desire. In such a lter, the values of the output sequence are computed from the input using discrete-time convolution:
yn =
M X m=0
hm rn;m
(5.4)
Here hm is the length-(M + 1) sequence whose elements are the coe cients of the order-M FIR lter. The input sequence rn is assumed to be in nite and n is the time index for the sample rn. The order, M , of the lter has been chosen to su ciently attenuate all out-of-band signal components so that, after ltering, we can compress
77
our representation of the output signal by the decimation factor of D. This need to resolve and remove out-of-band components is the primary factor that drives the determination of the minimum length of the lter response. The compression of the output representation of the decimating FIR lter is accomplished as we compute the output samples only for times n = kD. Each of these output samples will, of course, require M + 1 multiplications and M additions to compute. Note that the lter order M is not chosen speci cally to provide some required 0 level of output SNR in the channel lter. Our idea is to produce output samples yn that only approximate the yn to the extent that they still provide the desired level of SNR at the output while requiring fewer operations to compute. We will compute these approximate samples by only partially evaluating the summation shown in (5.4) for each sample: X 0 yn = hm rn;m (5.5) where the selection set Sn f0 1 ::: M g is the subset of the indices of the lter coe cients used to compute yn. We would like to nd a way to choose this subset that will adequately approximate the original output sequence yn while using the smallest amount of computation, that is, using the smallest expected number of terms in each sum, E fjSnjg. Our investigation of this problem will begin with the development of some tools to provide a quantitative understanding of the e ect of discarding input samples. We rst introduce a new model that allows us to analyze the e ect of discarding di erent sets of input samples. We also present an expression that represents the distortion caused by this operation of discarding samples. Using these tools, we then evaluate an algorithm that allows us to discard samples while bounding the distortion measured at the lter output. This is equivalent to reducing the computation required in channel separation while maintaining some minimum level of SNR at the lter output. The technique that we present has two forms. The rst makes no assumptions about the input signal: it simply discard samples randomly to reduce computation. The other form demonstrates that we can use knowledge about the input signal (not actual samples, but rather in terms of the expected distribution of energy at di erent frequencies) to further reduce computation for the same level of output distortion. This approach to reducing computation can be generalized in a number of ways, but in this work we restrict our consideration to a scheme that induces distortion that has a at spectrum, that is, a type of distortion in which the error at each point in the sequence is uncorrelated with the error at other sample points. This white distortion is often easier to deal with in subsequent processing stages, such as a detector.
m2Sn
78
rn hn zn rn
yn
y n hn
Sample Selector
(b)
Figure 5-7: Diagram showing conventional FIR lter and model for approximating output samples.
hmrn;m zn;m =
hmrn;m zn;m
(5.6)
Before we proceed with a complete analysis of this model for sub-sampling and its e ects on the lter output, it is helpful to present a short example of one simple random sub-sampling scheme. We perform a simple analysis of this example case to build some intuition for sub-sampling before analyzing a more general rule for discarding samples in the sections that follow.
79
One simple way to pick the values of zn in Figure 5-7 is to use a biased coin that gives probability p of retaining a sample:
p zn = 0 with probability p (5.7) otherwise Here the samples that are retained are multiplied by the constant 1=p. To understand the e ect on the lter output of this discarding of samples, we use standard results from signal processing that describe the e ect of passing a random signal through a linear lter like the one in Figure 5-7. Before showing these results, however, we need a few de nitions. We de ne the autocorrelation sequence (ACS) of a real-valued random sequence sn as the expectation: Rs n m] = E fsnsm g (5.8) A random sequence sn is wide-sense stationary (WSS) if: 1. the expected value, E fsng is independent of time, and
2. the ACS can be written as a function only of the di erence, k = m ; n, between the samples in the expectation: Rs n m] = Rs k] = E fsnsn+k g. When a sequence is WSS, we can also de ne the power spectrum density (PSD) as the discrete time Fourier transform (DTFT) of the ACS:
1 X
k=;1
Rs k]e;j
(5.9)
The subscript s in both Rs k] and Ss( ) refer to the original sequence sn. We also note that whereas Rs k] is a discrete sequence, Ss( ) is a continuous function in the frequency domain. We use as the frequency domain parameter (as opposed to !) to indicate that this is the transform of a discrete sequence and is therefore periodic in the frequency domain with period 2 . The sequence of lter coe cients hn is a nite-length deterministic sequence, and we will write its DTFT as:
H ( ) = DTFTfhng =
1 X
k=;1
hk e;j
(5.10)
When a WSS random sequence is passed through a digital lter such as hn, its output is also a WSS random sequence, and we can write the PSD of the output sequence
80
in terms of in the input PSD and the lter. For the lter shown in Figure 5-7(a), the output PSD is Oppenheim and Schafer, 1989]:
(5.11)
+1 X Sz ( ) = Rz = 1 ;1 +2 ( ; 2 l) (5.13) p k=;1 l=;1 The PSD of the input to the lter in Figure 5-7(b) is the PSD of the product rnzn, which is the periodic convolution of Sr ( ) and Sz ( ) Oppenheim and Schafer, 1989]: 1 Z S ( )S ( ; )d Sfrzg( ) = 2 (5.14) r z 2 If we substitute from (5.13) and carry out the convolution we get:
For the case of our simple coin- ipping sample selector, we can now analyze the e ect of discarding samples. The ACS and PSD of the selection sequence zn de ned in (5.7) are: (1 p Rz k] = E fznzn+k g = 1 k = 0 (5.12) k 6= 0
+1 X
k]e;j k
X 1 ; p Z S ( )d + +1 Z S ( ) ( ; + 2 l)d (5.15) Sfrzg( ) = 2 p r r 2 l=;1 2 The second term in the right-hand side of this result reduces to simply Sr ( ) because of the sifting property of the integration with the impulses and the periodic spectrum of the PSD. This PSD at the lter output can now be written as simply the original lter output from (5.11) plus a second additive term:
Sy0 ( ) = jH (
)j2S
r(
) + jH (
)j2
1 ; p Z S ( )d 2 p 2 r
(5.16)
This result in (5.16) helps us to understand the e ect of discarding samples according to simple biased coin ips. The rst term in (5.16) is equal to the PSD of the original output signal when no samples were discarded. The second term is additive and represents the distortion caused by discarding some samples. Note that when p = 1 (all samples are used) the distortion is zero and the distortion increases as p decreases. In Figure 5-8, the distortion is plotted as a function of the probability p of retaining each sample for typical values of hn and Sr ( ). The results shown in this gure are developed more fully in the next few sections, but we can see that as p decreases from one to near zero the level of distortion increases to levels that exceed the signal
81
-5
-10
-15
-20
-25
-30 0 0.2 0.4 0.6 Fraction of Input Samples Used (p) 0.8 1
Figure 5-8: Distortion due to discarding input samples according to random coin ips. of interest (at 0 dB in the plot, the power of the distortion equals the power of the signal itself). The example of choosing the values of zn using a biased coin helped us to gain some intuition about the e ect of discarding input samples as we compute the output of a narrowband channel lter. Discarding samples led to additive distortion in (5.16) whose PSD increased in magnitude as samples were discarded. In order to understand how to choose the selection sequence zn in a manner that will allow us to more carefully control the induced distortion, we now provide a more general analysis the e ect of discarding samples. 0 We rst de ne the error between the approximate lter output yn and the original output sequence: 0 en = yn ; yn (5.17) We would like this error sequence, en, to have zero mean (to provide an unbiased approximation) and, for a given choice of the sequence zn , we would like to compute its variance, i.e., the mean-squared error (MSE) of the distorted output sequence relative to the original output: var(en) = E fe2 g n (5.18)
82
+ -
dn hn
en
Figure 5-9: Model for analysis of error variance due to random sub-sampling. In Figure 5-9, we have combined the two schemes from Figure 5-7 to produce the error sequence for the purpose of analysis. Before using this combined model to derive the relationship between the input sequences and the output variance, we need to state a few assumptions that will simplify the derivation. We again assume that the input sequence rn is wide-sense stationary. We also assume that the zn are chosen independently of the values of rn. This allows us to fully realize the computational savings of discarding some input samples without examination, as well as allowing us to pre-compute zn. It is important that the sequence zn have a non-zero mean we will see later that this determines the amplitude of the desired signal in the output sequence. Without loss of generality, we assume that E fzng = 1. This speci c choice prevents problems with scaling factors later but does not limit our choice of sequences, as long as we scale them appropriately. We also will de ne vn = zn ; 1 to simplify the notation in our analysis (so E fvng = 0). We can now write the ACS for the lter input sequence, dn = rnzn ; rn (the distortion sequence), as:
(5.19)
Our model for approximating the ltering operation has several desirable characteristics. First, if the length of the lter (M + 1 coe cients) is greater than the decimation factor, then some input sample will be required in the computation of multiple output values. When this is the case, our model will ensure that these samples are used in every such computation or in none of them. This is useful in a real implementation where much of the cost of the computation is retrieving a sample from memory, not in performing the actual arithmetic operation (hm rn;m). In a sense, we are approximating the input sequence as opposed to approximating the lter. The selection set Sn can be viewed as the set of lter coe cients that are
83
used to compute each output. If this set were the same for every n, it would simply de ne a new lter that is an approximation of the original lter. This approximation approach could be evaluated using standard lter response techniques and the analysis of such an approach is not part of this work.
Problem Statement
Using the model shown in Figure 5-9 and the de nitions above, we now state our problem more precisely: Find the sequence zn that minimizes the expected amount of computation required to approximate the lter output while ensuring that the error variance is less than or equal to a bound B : min Prfzn 6= 0g] such that E fe2 g B n (5.20)
To analyze the e ects of zn on the variance of the error sequence, we rst write the power spectrum density of the output of the lter in Figure 5-9. This output PSD can be written in terms of the input PSD and the frequency response of the lter, similar to (5.11): Se( ) = jH ( )j2 Sd( ) (5.21) The variance of en can be written as the inverse DTFT of this PSD evaluated at k = 0: 1 Z S ( )e;j k d 1 Z S ( )d var(en) = Re 0] = 2 = (5.22) e e 2 k=0 2 2 Writing this variance in terms of the input sequence PSD: Z var(en) = 21 jH ( )j2 Sd( )d (5.23) 2 The PSD of the input sequence, Sd( ), represents the distribution in frequency of the expected distortion (the squared di erence, d2 ) caused by discarding samples n according to the sequence zn . To reduce the variance at the output we would ideally like this distortion to occur at frequencies for which the amplitude of H ( ) in (5.23)is small: the stopband of the lter. If dn is a white sequence then Sd ( ) will be constant for all (since dn must also be zero-mean). This implies, from (5.23), that the output variance will be simply proportional to the input variance, E fd2 g, and that n
84
the proportionality factor will depend only on the response of the bandpass lter, H ( ). If we substitute in (5.23) using the de nition of the DTFT of Rd k],
(5.24)
Exchanging the order of summation and integration, we get: 1 X 1 Z jH ( )j2 e;j k d var(en) = Rd k] 2 2 k=;1
(5.25)
The expression in the square brackets above is the inverse DTFT of the squared response of the lter. This can be written as the ACS of the deterministic, nitelength sequence of lter coe cients hn, which we will call ch k]:
ch k ] =
1 X
m=;1
(5.26)
We can now combine all of these results to show that the output error variance is simply the sum of a product of three sequences: var(en ) =
1 X
k=;1
Rr k]Rv k]c k] =
1 X
k=;1
Rr k](Rz k] ; 1)ch k]
(5.27)
To check this result, consider the case of zero output distortion: choosing zn = 1 for all n results in Rv k] = 0 for all k, giving zero error. This result in (5.27) is signi cant for several reasons. First, it shows that the error variance of our approximation scheme depends on the sequence vn (and hence zn) only through its ACS. Second, it shows that the error variance depends linearly on all three of the key parts of the system: the ACS of the input sequence, the ACS of the selection sequence and the coe cients of the channel selection lter.
85
wn
gn
xn
Sequence Transform zn
Offline rn
yn hn
Figure 5-10: Random sub-sampling using a transformed sequence for sample selection. 2. It allows us to discard as many samples as possible, that is to maximize the probability that zn = 0, subject to property (1) above. 3. It produces a distortion with uncorrelated error values at each point in the sequence. We now identify two separate cases as we try to decide which samples to discard. In (5.27), we saw that the induced error variance depended only on the ACS of the received wideband sequence and the ACS of the channel lter. Although it is conceivable that in some cases we may have a good idea of the spectral distribution of the received signal (and therefore its ACS), we may not always have this information. We therefore identify two cases as we try to nd a good choice for the selection sequence, zn . Although it is well known how to generate a random sequence with a desired ACS (e.g. by generating \shaped noise", see Stark and Woods, 1986]), we found no prior work on how to directly generate such a sequence with a relatively high probability that zn = 0. Instead, we will start with a candidate selection sequence, xn, that has a desirable ACS and perform a sequence transformation to convert it to a sequence zn that has more zero elements and an ACS that remains \close" to that of xn , i.e. Rz k] Rx k]. This sequence transformation approach is shown in Figure 5-10, which also depicts the creation of the initial sequence xn by the ltering of a white noise sequence wn with the shaping lter gn. We will describe how the lter gn is chosen in a later section. In terms of our transformation scheme, the error variance of (5.27) can be written as two components: one component due to choice of the original sequence xn (the rst term below), and another to the transforming e ect that introduces more zeros
86
and produces zn :
We can perform a similar decomposition to that of (5.19) which gave us ACS for the distortion sequence dn. This decomposition also has two terms: the rst term is the portion of the autocorrelation due to the initial choice of xn, the second term corresponds to the part of the distortion due to the transformation process:
(5.29)
Because we want the distortion to be uncorrelated at each sample (white), we would like Rd k] to be non-zero only for k = 0. To achieve this goal requires that for all k 6= 0: (1) we choose the sequence xn to make the rst term in the right-hand side of (5.29) zero, and (2) we choose the transformation such that (Rz k] ; Rx k]) = 0 in the second term. We now explain these two steps in more detail. We will rst describe the way that our sequence transformer is designed, and then discuss the best way to choose the initial sequence xn. Given a sequence xn with ACS Rx k], we want to produce a sequence zn with Rz k] Rx k] while increasing the number of zero elements, zn = 0. Furthermore, we saw in (5.29) that we must have Rz k] ; Rx k] = 0 for k 6= 0 while minimizing the di erence Rz 0] ; Rx 0] to minimize the error variance due to the choice of transformation. We now present a transformation that will change some of the elements of the sequence to zero while ensuring that the di erence sequence, dn will be white, that is, it will satisfy E fdndn+k g = 0 for k 6= 0. This may not be the optimal transformer, since there may be some sequences that allow even less computation for the same error variance, producing a non-white dn. In addition, we would like our sequence transformer to be able to transform the sequence one element at at a time: each element of the transformed sequence, zn, will depend on only the corresponding element of the candidate sequence, xn. To design this transform scheme, we begin with a general rule for transforming the individual elements of xn:
(5.30)
This rule requires that we de ne two mappings: f : < ! < and q : < ! 0 1]. In
87
choosing f and q to accomplish our goals, we rst compute the ACS of the transformed sequence zn . For the case k 6= 0 we have
E fznzn+k g =
=
Z1Z1
Z;1 Z;1 1 1
;1 ;1
znzn+k p(zn zn+k )dzndzn+k f (xn)q(xn)f (xn+k )q(xn+k )p(xn xn+k )dxndxn+k (5.31)
To ensure that E fznzn+k g = E fxn xn+k g for k 6= 0, we now choose f (xn) = xn=q(xn), giving:
E fznzn+k g =
Z1Z1
;1 ;1
(5.32)
non-zero output distortion as seen in the second term of (5.28). We wish to nd the function q that minimizes the expected computation (minimizes Prfzn 6= 0g) for a given bounded output distortion, var(en).
( 2 ) Z 1 x2 xn n p(x )dx = E 2 Rz 0] = E fzng = (5.33) n n q(xn) ;1 q (xn ) Now, since 0 q(xn ) 1, it is clear that Rz 0] Rx 0], with equality only in the case where q(xn) = 1, for all values of xn that occur with non-zero probability. This implies that anytime Prfq(xn) < 1g > 0, our sequence transformer will induce some
To simplify the problem of nding the best q, we assume that the xn are drawn from a discrete distribution: xn 2 X (and therefore zn 2 Z ). This simpli cation will allow us to solve a discrete optimization problem to nd the best choice for q(xn) and does not signi cantly a ect the eventual result. We use the notation xi to indicate an element of the set X and we refer to the probability mass function over these values using pi, that is Prfxn = xig = pi. We will also use the notation qi to refer to the values of function q(xn ), that is q(xn) = qi when xn = xi. So we now have f : X ! Z and q : X ! 0 1]. In order to ensure Rz k] = Rx k] for k 6= 0, we choose f (xi) = xi=qi , and (5.33) now becomes Rz 0] = E fx2 =qig when we use a discrete distribution for xn. i At this point, nding the best choice for the function qi can be written as an optimization problem. The goal is to choose the values of qi providing the highest probability of zn = 0 (lowest probability that zn 6= 0) given a bounded distortion for the transformer:
88
Problem Statement
Find qi to give min
"X
i
piqi
X pi x2 i qi i
(5.34)
Note that although the value we actually need to bound is Rz 0] ; Rx 0] = E fx2=qi g; i E fx2 g, since both terms in the di erence are positive and E fx2g is xed for the given i i sequence xn , we can equivalently bound the rst term alone.
Solution
We can assume that equality holds in the constraint above at any optimal solution otherwise we would be able to increase one of the qi to achieve equality, thereby reducing the sum to be minimized. This constraint is a hyperbolic surface in the qis and since all of the parameters are non-negative, this is a convex surface. The convexity implies that there exists a single global minimum point on the constraint surface. Any global solution to this constrained minimization in (5.34) will also satisfy pair-wise optimality for the individual pi s. That is, the same global solution will also satisfy: 2 p x2 min piqi + pj qj such that pixi + j j = ij (5.35) qi qj for some xed ij for all i 6= j . If this were not so, then we would be able to improve on the global optimum. Because of this pairwise optimality condition, we can solve the pairwise problem by eliminating qj through substitution in the above, setting the derivative equal to zero and nding the optimal qi (and qj by symmetry): (5.36) q = 1 jx j p jx j + p jx j
i ij i i i j j
These relationships allow us to show that the ratio qi=qj does not depend on qi=qj = jxij=jxj j. From here we can easily nd the global solution:
ij :
qi = jxij E fjxnjg
(5.37)
jxij > =E fjxnjg this would result in a value of qi that is greater than one, which
This result for the values of qi is not quite complete. If any of the xi are such that
would be unacceptable for a probability value. For such xi , the optimal solution is to choose qi = 1, so that such values will never be changed as we transform the se-
89
quence xn into zn. The intuition here is that these are the outlying values of xi in the distribution, and changing these values with non-zero probability causes more distortion than simply changing the values of smaller-magnitude xi with higher probability. Only those with smaller jxi j are subject to possible transformation according to (5.30) and in the limit as ! E fx2 g none of the xis will be subject to rounding. In this n case we will not be able to transform any elements to zero, but will have zn = xn 8n in order to satisfy the bound on distortion. To capture this e ect, we de ne the set SR as those xi that will be subject to being transformed with non-zero probability, that is, those xi such that jxij < =E fjxnjg. The nal rule for our sequence transformer is now: ( x with probability q x zn = 0 n=q(xn) with probability 1(;nq)(x ) (5.38)
n
E fjxnjg
if xi 2 SR = if xi 2 SR
(5.39)
The set SR consists of xi that are subject to rounding (those xi for which qi < 1) are those for which: SR : xi such that jxi j < E fjx jg (5.40) n
90
step approach, yet provide superior performance when used to select which samples should be discarded in our random sub-sampling approach. This being said, we now analyze the performance of our sequence transformation scheme. The performance of the sequence transformer can be measured by determining how much it is able to reduce the cost of computing the lter output for a given bound on error variance. This cost reduction is simply the probability that a particular element of the input sequence will be discarded: Prfzn = 0g = E f1 ; q(xn)g =
X
i
pi(1 ; qi )
(5.41)
Substituting the expression for qi from (5.39), we can write the complement of this, the amount of computation required: X X X Prfz 6= 0g = p q = E fjxnjg p jx j + p (5.42)
n i i i i: xi 2SR i i i: xi 2SR = i
We can use this result to compare the two cases that we identi ed earlier to see if having knowledge of the received wideband signal and the lter response can help to reduce computation.
1
1
91
Using this rule, we will always have xi < 1=E fjxnjg = 1, so the expected computation is: (5.46) Prfzn 6= 0g = 1 This case where the candidate sequence is chosen as xn = 1 for all n corresponds to the simple case of using a biased coin that we discussed earlier in this chapter, although is is now clear how we should pick the probability p = 1= 1 to achieve a desired level of error. We will perform further analysis on these results in the next section to provide more insight on the e ectiveness of this scheme for this case and the next.
1
(5.47) (5.48)
(5.49)
Using this rule, the expected computation from (5.42) is: Prfzn 6= 0g = E fjxnjg
2
! X
i: xi 2SR
pijxij +
X
i: xi 2SR =
pi
(5.50)
92
As we consider the above results, we will need to understand the e ect of the E fjxnjg term that appears in (5.50). We do so in the context of our approach which generates xn using an FIR lter to shape the spectrum of a white sequence, as in Figure 5-10. Here xn has an ACS that is determined by the response of the lter gn, which itself is determined from (5.28) in order to minimize . When we generate xn using an FIR lter, each sample of xn will be the weighted sum of a large number of independent samples of the white sequence wn. This will tend to produce samples of xn that have a Gaussian distribution. When we consider (5.50), we see that for a Gaussian distribution with mean E fxn g = 1 (because zn must also have unit mean and the transformer preserves the mean), E fjxnjg is strictly greater than one. Although E fjxnjg > 1, it is not much greater if the variance of xn is relatively small so that the negative tail the distribution of xn (the part a ected by the absolute value operation) has a small area. To get a qualitative feel for the relative performance of the two cases, we can now compare the results of (5.46) and (5.50). We consider the case of values of the bound B on the variance of the error in (5.20) for which most of the xi are subject to rounding, that is, xi 2 SR for most xi . In this case, the second summation in (5.50) will be very small and the rst summation will be approximately equal to E fjxnjg 1. We will then have Prfzn 6= 0g 1= 2. Furthermore, if we assume that in (5.48) is small, the we see that a non-zero variance for xn will lead to 2 > 1 and thus to an improvement in performance for case II relative to case I. To achieve this improved performance, however, we need to nd a general solution which allows us to determine an ACS for a candidate sequence xn with non-zero variance in addition to a small (or zero) resulting value for , the rst term in (5.28). This general solution is the goal of on-going work. In the next section, we present the result of computer simulations that validate the results of (5.46) and provide a more concrete comparison of the two di erent cases.
93
-25
-30 0.5
0.55
0.6
0.65
0.85
0.9
0.95
Figure 5-11: Output error variance relative to desired signal power versus proportion of input samples used for several cases of output signal relative bandwidth. input samples. The plot shows results for ve di erent values of ch 0], corresponding to di erent bandwidths for the desired signal relative to the wideband input. The results show that if we need to bound the distortion at some level, say ;15dB relative to the received signal, the approach of Case I would provide some reasonable reduction in computation. The results shown are for the base case where there are no interfering signals in adjacent channels and all of the signal energy is in the desired band. If any other signals are present in the wideband input signal, the respective curves would shift upward to re ect the increased value of Rr 0], the input signal variance. A contrast between the approach of Case I and an alternative, a uniform subsampling scheme, is their behavior in the presence of interference. The qualitative e ect of the random sub-sampling approach is to provide more uniform behavior, regardless of the location of any interfering signals relative to the desired signal. For example, consider the case where a single interfering signal is present in the stopband of the channel lter. For the random sub-sampling approach, this would result in an upward shift of the curves in Figure 5-11 by 3 dB when the interferer is independent of the desired signal and has equal power (so Rr 0] will be twice as much
94
as with no interferer present). This result does not depend on the location of the interferer within the stopband: the interfering signal could have a center frequency anywhere in the stopband and the result would be additive uncorrelated distortion at about -15 dB relative to the desired signal (using half of the samples with a relative bandwidth of 2% in Figure 5-11). If half of the samples are instead chosen uniformly, this is equivalent to decimation by a factor of two prior to ltering. The resulting e ect would depend on the relative frequencies of the desired signal and the interferer. The interfering signal might be aliased into the stopband of the lter, resulting in negligible distortion. On the other hand, it is possible that the interferer would be aliased into the passband of the lter by the decimation, and this would result in signi cant distortion at the lter output (zero dB relative to the desired signal). It is also interesting to note the di erence between the individual curves in Figure 5-11 for the di erent relative bandwidths of the output signal. These curves show that there is smaller error variance when the bandwidth of the output signal is more narrow relative to the input bandwidth. This observation leads to a more general and important conclusion about the results of (5.46) and (5.50), which is that the amount of computation required to separate a narrowband signal with a xed level of output distortion need not depend on the input bandwidth, but rather on output bandwidth and the amount of interference in the stopband. For example, consider the case of a narrowband lowpass channel separation lter. If we increase the input sample rate while the output sample rate is held constant, the sum of the coe cients hn will remain constant to provide constant gain at center of the passband (zero frequency). The number of coe cient in the lter, however, will increase in proportion to the increasing input bandwidth because its response must span a constant interval of time (and the sample interval Ts will decrease as the input bandwidth increases). The value of ch 0] is the sum of the squared coe cients, ch 0] = PM h2 , and this will decrease in inverse proportion to an increasing input sample n=0 n rate. From (5.43), we conclude that the computation, CNB , required to separate the narrowband channel will thus remain constant if a xed level of distortion (B ) is speci ed in the output: R CNB = PRfzn 6= 0g fLength of lterg Constant as R in increases (5.51) out Although the amount of computation required for the channel ltering does not depend directly on the input sample rate, this amount does depend on the amount of interference present in the stopband. From (5.43) we see that if more interfering signals are present, the variance of the input signal, Rr 0], will increase, resulting in an increasing amount of computation required to maintain a xed level of output
95
Figure 5-12: Comparison of Case I and Case II results for three values of var(xn ).
error variance, B . In Figure 5-12, we show a comparison of the performance for the two di erent cases. This plot shows a comparison of the results for cases I and II from (5.46) and (5.50) for several di erent values of Rx 0] under the assumption that = 0 and the output signal bandwidth is 10% of the input bandwidth. Here we see that using a candidate sequence with a non-zero variance can further reduce the proportion of samples required to produce an output signal with a bounded error variance. As in the previous case, this plot shows the base case where all of the received signal power is assumed to be in-band, any interfering signals present in the stopband of the lter would result in an upward shift in each of the curves. The curves show that at low levels of distortion (less than -15 dB), increasing the variance of the candidate sequence reduces the computation required to produce a xed quality output only to a point (when var(xn 2), beyond which a further increase in variance is not helpful.
96
5.6 Summary
In this chapter, we have addressed the two di erent steps of the channel selection process: frequency translation and bandwidth reduction through ltering. For frequency translation, the conventional approach provides excessive exibility in frequency translation, but at the cost of signi cant excess computation. We have demonstrated how to perform the same function using a composite lter followed by nal translation at the low output sample rate after ltering, this enabling the computation of the frequency translation step to be made proportional to the output sample rate. We have also describes an unnecessary coupling that exists between the length of the lter response and the number of input samples used when computing the output of a narrowband channel separating lter using conventional approaches. In order to provide adequate out-of-band rejection, it may incur unnecessary computation by processing excessive samples. Instead, we have demonstrated that we can independently control lter response and output SNR. We control output SNR by bounding the distortion caused by approximating the input signal through random sub-sampling. We then can perform bandwidth reduction using a relatively small number of input samples while maintaining the desired output SNR. This reduced number of samples reduces the number of memory accesses required to retrieve data, at the cost of potentially performing an additional multiplication for each input sample used. The two-step approach presented here utilizes a sequence transformer to produce white additive distortion (noise) at the output, although it is conceivable that a more general approach can be developed that provides greater computation reduction using non-white distortion.
98
Synchronization Channel Channel Separation Detection Symbol Detection Channel Decoder Error Decoder Source Decoder User
Figure 6-1: Diagram of the channel decoder showing division of the symbol detector into synchronization and detection steps. decide which is the most likely corresponding data symbol. In this chapter, we focus on the detection process under the assumption that the required synchronization has been accomplished. A more detailed discussion of techniques for synchronization in digital receivers is available in Meyr et al., 1997]. The output of the detection process is the sequence of symbols that the receiver believes was originally transmitted. This sequence is passed to the error decoder, which will either correct those errors it can, or simply detect errors and take other appropriate actions. It is in this context of protecting against errors in the receiver that the overall role of the detector in the system becomes clear. In a noisy environment, the detector will always produce some errors. One of the functions of the receiver as a whole is to control the level of errors through various mechanisms to ensure that end user sees only an acceptable level of errors. These mechanisms are not limited to the detector, or even the channel decoder, but include redundancy at higher layers in the communication system or other techniques such as protocols for retransmission. To design an e cient and exible system, we need to provide the required error performance in a way that e ciently uses the available system resources. When seen in this light, the detector does not necessarily need to produce an optimal estimate of the received symbols in every situation. Rather, it needs to recover symbols in such a way that the required error performance is e ciently achieved through the composite behavior of all the error control mechanisms. Figure 6-2 shows a typical performance curve for an optimal digital detector. In this plot, we see that the probability of bit errors depends on the SNR of the received signal. If a particular system is designed such that the detector needs to achieve a bit error rate (BER) of 10;4 at some worst-case SNR, say 8.5 dB SNR, then it will provide much better BER
99
0.01
0.001
1e-05
1e-06
1e-07
1e-08
Figure 6-2: Plot of bit-error rate versus SNR for a two level PAM system. if the SNR is higher. For example, if the system is experiencing an SNR of 11 dB then the optimal detector will produce a BER of 10;6. This extra performance will not hurt the overall system performance, but if it is achieved at the cost of unnecessary resource consumption, then it will not be the most e cient solution. In this chapter we will present an approach that enables the detector to e ciently provide a desired level of BER for a speci c SNR from current channel conditions. Before presenting these results, we rst review some conventional techniques to perform detection and try to develop an understanding of how this process can be modi ed to provide e cient and controllable performance trade-o s.
100
noise. We will consider this case in the context of a system that has limited RF spectrum available and needs to e ciently use this limited bandwidth to achieve the best possible data rate. As we consider some standard approaches to this problem, we will review how the system can use pulse shaping techniques to e ectively use its spectrum as well as the concept of the matched lter detector. In chapter 4 we discussed how to synthesize waveforms that encode sequences of symbols into a continuous waveform. We noted that each symbol can be encoded with a shaped pulse and then time-shifted pulses added together to form continuous waveform. At the same time we saw that these pulses often overlap. As we try to recover the values of original data symbols, we need to somehow separate the original pulse shapes that were added together. One common way to make this separation possible is to use pulses that satisfy the Nyquist Criterion for zero intersymbol interference (ISI) Lee and Messerschmitt, 1994]. This criterion ensures that for each encoded symbol there will be one instant in time when the contributions in the composite waveform from all adjacent pulses will be zero (although there will still be noise present in the received waveform). Such a pulse shape is shown in gure 6-3, part (a). In part (b) of this gure are a series of eight scaled and time shifted pulses that encode eight consecutive data symbols, and in part (c) is the composite waveform, the sum of the components in (b). To estimate a speci c symbol, the receiver can examine the waveform at that instant in time when only the desired corresponding pulse had a non-zero component (these instants are indicated by the black dots in the composite waveform). These locations in the continuous waveform are known as the ideal slicing points. This property makes it clear why the pulse has the distinctive shape with multiple zero-crossings: it is zero at time t = kTb for all k 6= 0 (Tb is the symbol, or baud, interval). In systems that use such pulses, the transmitter ensures that the waveform is generated in such a way that this zero ISI condition will be met. The receiver has the task of identifying these speci c points in time for each symbol. This fact indicates the importance in the receiver of the synchronization, or symbol timing recovery: the receiver needs to synchronize its time reference with the transmitter in order to examine the waveform at the correct locations. If the receiver is not precisely synchronized it will reduce the receiver's ability to determine the correct symbol in each interval. In practice the receiver will not be exactly synchronized, but it needs to get close enough to ensure that the small amount of ISI from the adjacent pulses does not cause excessive detection errors Lee and Messerschmitt, 1994]. As we examine the detection process in the following sections, we treat the received
101
Figure 6-3: Representation of (a) an isolated pulse satisfying the zero ISI condition, (b) multiple scaled and shifted pulses and (c) the noise-free composite waveform. (In the plots, T = Tb , the symbol interval.)
102
pulses as if they are isolated pulses, that is, as if they are not part of a sequence of time-shifted pulses added together to form a continuous waveform. We will review a technique that is used to perform optimal detection on noisy versions of such pulses, in the sense that it will make decisions with minimum probability of error. This optimal detector is known as the matched lter, for reasons that we will describe later. The basic approach is to pass the received sequence through a digital lter and generate output samples that correspond to the ideal slicing points, as shown in gure 6-3(c). A model of this approach is shown in gure 6-4, part (a). This lter shows the transmit lter, an addition of noise (which we will treat as the channel distortion) and nally a receiver lter. The e ects of interfering signals are not shown because we assume that they are removed in the channel separation process discussed in the previous chapter. Although our discussion will focus on the detection for an isolated pulse, we will still be able to apply the results of this analysis to the case of a sequence of pulses because we will assume that it is the cascaded response of the transmit and receive lters that satis es the Nyquist Criterion. This means that we can look at the output of the receive lter and treat is as an isolated pulse in the sense that there will still be an instant in time for each symbol where there will be zero contribution from pulses encoding the adjacent symbols. This situation will be equivalent to that in gure 6-3(c) if we assume that the waveform shown is at the output of the receive lter Frerking, 1994]. As we begin our discussion of the detection problem in the next section, we will consider the case of only two symbols (i.e. 0 and 1). This makes the subsequent discussion more simple yet still captures the essential parts of our analysis. All of the results for the binary case of the detection problem can be generalized to case of M -symbol modulation with the appropriate modi cations. Given the assumptions described above, i.e., that there is correct synchronization and a binary constellation, we can consider the detection problem in a very simple form. We consider a length-N vector that takes on one of two possible values X 2 fs0 s1g. These two vectors represent the discrete-time waveforms that are transmitted through the channel when the transmitter sends either a zero or a one. The receiver has access to a noisy version of this vector: Y = X + Z, where Z is a vector of independent, identically distributed Gaussian random variables (RVs) with mean zero and each with variance 2. The receiver must decide between two hypotheses: either the transmitter sent a zero (H = 0) or a one (H = 1). We assume that the prior probabilities of the two hypotheses (0 or 1) are equal. If we wish to minimize the probability of making an incorrect decision, the opti-
103
Transmit Filter
Receive Filter
Figure 6-4: Cascade of transmit pulse-shaping lter and receive lter whose combined response satis es the Nyquist Criterion for zero ISI. mal solution is well known as the matched lter or correlation detector, and detailed derivations are available in many texts, for example Lee and Messerschmitt, 1994]. In this case of a discrete-time FIR implementation, we simply compute a weighted sum of the N observations and use a threshold test on this sum:
U=
N X i=1
b i Yi = b T Y
(6.1)
In the general N -dimensional case, b is the vector of weights that is determined as the di erence b = s1 ; so and bi are its elements. The PDFs of this RV U when conditioned on each of the two hypotheses are shown in gure 6-6. The threshold test 1 is made using the threshold = 2 bT (s0 + s1) and the particular value of the sum, ^ ^ ^ U = u. We decide H = 0 if u < and H = 1 if u > , where H is our estimate. The intuition here is that we project the received vector onto the line connecting the two possible vectors represented as points in N -space. We then choose as our ^ estimate H the point in the constellation that is closest to the projection of the received point onto the line (illustrated in gure 6-5 for 2-dimensional case). The conditional PDFs in gure 6-6 are the conditional distributions of the projection U along the line that passes through s1 and s0. In a binary system the minimum energy constellation is achieved by s1 = ;s0 .
104
11 00 11 00
s1
s0
1 0
Figure 6-5: Projection of noisy received vector onto the line connecting the two possible transmitted points. Without loss of generality, we will assume this is true for the rest of the discussion in order to simplify the notation, giving b = 2s1 and = 0. The probability of making an error, Pe, in our decision is always non-zero because of the Gaussian distributions. Due to symmetry and the equal prior probabilities of the hypotheses, this unconditional error probability is equal to the probability of error, conditional on either value of the transmitted bit: (6.2) Where the function Q(x) is the area under the tail of the Gaussian distribution given by Z1 1 ;y2 ! dy (6.3) Q(x) = p exp 2 x 2 This function Q(x) becomes very small as x becomes large, but is always non-zero. The result in (6.2) show us that probability of error depends on the ratio between the total energy in the pulse (bT b) and the variance of the random noise ( 2). In many implementations, the matched lter is used exactly as presented here. As already noted, this structure provides the solution to the detection problem with ^ ^ Pe = PrfH 6= H g = PrfH 6= H jH = 0g =
Z1
0
pT ! pU jH (ujH = 0)du = Q b b
105
pU jH (ujH = 0)
pU jH (ujH = 1)
bT s0
=0
bT s1
Figure 6-6: Conditional PDFs for H=0 and H=1. lowest possible probability of error for the conditions described above. This result can be generalized in a number of ways, including the case where the transmitted pulse is one of a set of 2b di erent pulses, so that b bits of information are communicated with each pulse transmitted, or the case where the prior probabilities of transmitted symbols are unequal.
106
The framework
Our framework begins with a very general model of the input to a digital signal processing function. This input is a sequence of uniformly spaced samples of analog waveform, with interval between samples of Ts seconds. The sample rate is therefore fs = 1=Ts Hz, and this rate determines the maximum width of frequency band that the input sequence can represent. For a general channel decoder, the timing of the individual sample instants for the sequence is independent of the timing of any data symbols encoded by the analog signal. We will refer to this as asynchronous sampling, because in general it assumes no synchronization between the data encoded by the waveform and the samples that represent the waveform. Our channel separation techniques in the previous chapter, for example, required no synchronization of the sampling process. We will discuss this synchronization aspect of the framework more in Chapter 7. The purpose of the framework is to help establish a more explicit relationship between input and output values. In a digital receiver, for example, the overall task is to produce an estimate of the original data symbol sequence encoded by the transmitter: f:::ak;1 ak ak+1 :::g. For each symbol to be estimated we therefore identify the ^ ^ ^ set of all input samples that contain any information about that particular symbol. This set is the footprint of the symbol in the input sequence, and is represented in gure 6-7. In general, the footprints for di erent symbols can overlap, often to a very high degree. This overlap is caused by interference in the channel, as well as possible overlap due to pulse-shaping and coding at the transmitter. Although there might be many samples in the footprint of a particular symbol, these samples are not all equally useful to the receiver as it performs detection. In the set of samples that contain non-zero information about a speci c symbol, ak , some samples may be very useful, others only slightly so. This idea will be important as we design a detection algorithm that can produce symbol estimates with some desired level of quality (probability of error) while e ciently using computational resources. The nal piece of our framework, then, is an understanding of the utility of each of the samples in the footprint, either in an absolute sense, or in a relative sense of which samples are more useful than others.
107
a
k
k+1
Figure 6-7: Footprints of symbols within the sample sequence. Use only some of input data and \completely" process those samples, that is we decompose the input data in time. The assumption here is that using less data allows approximation with less computation than using the full set of data. Use the entire set of data and perform \sub-optimal" processing to produce an approximation. The assumption here is that the sub-optimal result would require less computation than the optimal result, that is we decompose the system. In some cases, it is even possible to identify algorithms that produce a series of successive approximations as the computation proceeds. Use a combination of these two approaches, where a sub-optimal computation is performed using only part of the input data. As we seek to e ciently produce estimates of the transmitted data, it may sometimes be best to \squeeze" all of the relevant information out of each sample we use and only use as many as necessary. In other situations, it may be better to use all of the available data and nd a way to e ciently extract only the information that we need. In case of the matched lter detector, we see that the per-sample optimal processing is very simple the matched lter implementation requires only a multiplyaccumulate operation. This means that performing sub-optimal processing on the
108
entire data set is not likely to be much easier than performing optimal processing. On the other hand, if we consider using only some of the input samples, we can realize the savings of not retrieving those samples from memory for processing, in addition to the reduced processing.
Case I: When the allowable error probability of the decision is larger than the N -
109
This result says that to make the required decision using a minimum number of samples, we need only use the samples corresponding to the M largest magnitude weights in the weight vector b. Proof: To show this, we de ne a family of RVs similar to U , but where the summation is truncated at n terms and then scaled to give a unit-magnitude expected value:
Result: The smallest number of observations M that always allows a decision with probability of error Pe M Pe allow using a simple threshold test is the smallest M pT such that Pe allow Q b2M bM where bM is the vector of the M largest magnitude weights from b.
Un = T1
bk Yk = bn Yn bn bn k=1 bT bn n
T
n X
(6.4)
Here bn is the truncated column vector bn = b1 b2 ::: bn ]T . We also use the notation Yn to refer to similar truncations of the random vector Y. Each Un is a Gaussian RV with unit magnitude expected value (conditioned on each hypothesis): E fUnjH = 0g = ;1 and E fUnjH = 1g = 1 for all n 2 f1:::N g. Additionally, the two conditional variances for each Un are equal: varfUnjH = 0g = varfUnjH = 1g = bT b
2
n n
(6.5)
Because the value of the product bT bn is non-negative and non-decreasing as n n goes from 1 to N , we can see from (6.5) that the conditional PDFs, pUnjH (unjh) have non-increasing variances as n increases from 1 to N . This fact is represented in gure 6-8 which shows the conditional PDFs pUnjH (unjH = 1) for three di erent Un with n1 < n2 < n3 . Because of these non-increasing variances, when we make a decision about the original hypothesis using each of the Un and the same simple threshold test, we see that the probability of error for each decision (denoted by ^ Hn) decreases as more terms in the sum are computed: Pe n Pe n Pe n where n1 < n2 < n3 and the error probabilities are de ned as:
1 2 3
0q T 1 bb ^ Pe n = PrfHn 6= H g = Q @ n n A
2
(6.6)
This result shows that if we desire to minimize the number of observations used and bound the error probability, we can easily solve for the smallest M such that Pe M Pe allow as in (6.6) above. We can do even better, however, if instead of using the observations sequentially, we allow ourselves to choose a particular subset of m
110
pUn jH (un jH = 1)
3 3
pUn jH (un jH = 1)
2 2
pUn jH (un jH = 1)
1 1
un
Figure 6-8: Conditional PDFs for Un conditioned on H = 1 for several values of n, where n1 < n2 < n3. of the N observations. We de ne a more general family of RVs Um that are also scaled, weighted partial sums of the observations, but where the samples are re-ordered according to the set of ordering indices (i1 :::iN ):
m T X Um = bT1b bik Yik = bm Ym bT b m m k=1 m m
(6.7)
We use the notation bm to indicate a length-m truncated vector of the elements of b re-ordered according to (i1:::iN ) and similarly Ym for the observations. For this new family of RVs, Um , we still have the property that each conditional PDF is Gaussian with unit-magnitude conditional mean and that the conditional variances for each Um are equal: 2 varfUm jH = 0g = varfUm jH = 1g = T (6.8) Using this result, we can re-order the observations to improve the quality of our mobservation decision for any xed m. We choose the order in which the observations are used so that for each m (from 1 to N ) the value of bT bm is maximized. This m property is ensured if we choose the indices (i1 :::iN ) in a greedy manner so that the magnitudes of the elements of bm are in descending order: jbi j jbi j ::: jbiN j. Choosing this ordering ensures that any decisions made using the simple decision rule and only a subset of m observations will have the highest quality (lowest conditional
1 2
bm bm
111
variance and therefore probability of error) possible for each value of m. We can now describe the solution to our problem for the case of Pe allow > Pe N and where the decision is made using a single threshold test. This solution is to 1. order the observations according to the magnitudes of the weights in b, and 2. nd that M that is the smallest value of m that satis es Pe m Pe allow. We ^ then evaluate the partial sum for UM as in (6.7), deciding H = 0 if uM < and ^ = 1 if uM > , QED H In the previous section, we discussed making a decision with a desired probability of error in the case where Pe allow Pe N . We now present a second case.
Case II: This case occurs when the allowable error probability is smaller than the
minimum achievable with all N available samples, Pe allow < Pe N . If we use the simple decision rule presented above we will not always be able to make a decision with the desired probability of error, even if we use all N of the available observations. If we use a more complex decision rule, however, we can make a satisfactory decision some of the time. To see this, we simply recall that the probability of error was equal to the area under the tail of the PDF that was in the \wrong" decision region. To provide a decision rule with a lower probability of error, we simply need to evaluate all of the observations and then move the boundaries of the decision regions. For example, to determine the decision region R1 (where we choose p N ^ = 1), we place the boundary at the value of uN for which Q bT bN (uN +1) = H Pe allow. When we do this for both hypotheses, we nd that we now have three decision ^ ^ regions, as shown in gure 6-9: R0 for H = 0, R1 for H = 1, and RX for Cannot decide. These three regions are determined by the values uN = ;T and uN = T , where the threshold T is de ned by:
0 ;1 1 P T = @ Qq (T e allow) A ; 1
bN bN
(6.9)
When the value of UN is computed and we nd that uN is in either R0 or R1 , ^ ^ we can decide H = 0 or H = 0 with an acceptable probability of error. In the new third case, however, we could make a \best guess", but we will not be able to say that we have error probability less than Pe allow. Another result that can be seen from
112
R0 ( ) Rx ( ) R1
-1
;T
+T
un
Figure 6-9: Conditional PDFs for H = 0 and H = 1 and the three decision regions de ned by: (un < (;T )), (;T < uN < T ), and (T < uN ). gure 6-9 is that we can determine the probability that our computed value UN = uN will exceed the threshold and allow a decision with the desired error probability. This decision probability, PD , is equal to the area under the tails of the conditional PDF that exceeds the threshold T : n o P = Pr u 2 fR R g = Q T ; ; Q T + (6.10)
D N
1 0
where = E fUN jH = 1g is the conditional mean of the N -term sum. We can use the above results to produce a decision rule for Case II above: we compute the threshold T that provides an acceptable probability of error, then make decisions only when the N -term sum exceeds this value (i.e. in R0 or R1 ). This will ensure an acceptable decision in the largest number of cases.
113
For simplicity we design our algorithm only for the case when Pe allow Pe N (case I above). When this is true, we know from the earlier section that we can always make a decision with the desired probability using at most some number M N of the observations.
S=
where:
kf X i=1
bi Yi = bT Y = S1 + S2 bi Yi S2 =
kf X i=ki
(6.11) (6.12)
S1 =
ki X i=1
bi Yi
These two sums are independent when conditioned on the case that a one (or zero) was transmitted. Given these two partial sums, we perform an initial test after the rst ki steps using threshold T, and decide \0", \1" or continue. We now consider the e ect on the probability of error (conditioned on H = 1) of incorporating this initial test at step ki. We write the total conditional error probability produced using two tests, PejH =1 total , relative to the conditional probability of error using only the nal test at the end, PejH =1 kf :
(6.13)
We already know how to compute the rst term on the right-hand side. To compute the second, it is helpful to look at the regions in the S1 ; S2 plane shown in gure 6-10. The six di erent regions shown in this plane represent the six di erent outcomes possible on the two tests: Ra b means outcome a on the initial test and b on the nal
114
test. The only regions in this gure that a ect the PejH =1 term in (6.13) are R0 1 and R1 0 . Although the areas of these regions are equal in Figure 6-10, the volumes under the joint PDF for each region are not. Region R1 0 contains those cases that would result in a correct decision on the initial test and an incorrect decision on the nal test. In the situation where the nal test is not performed if the algorithm terminates, this would be an outcome that improves the error performance: it produces a correct decision where an incorrect decision would have occurred with the single test. The cases that fall into region R0 1 are the opposite, i.e., they cause an incorrect decision where a correct decision would have previously been made. The remaining four regions in gure 6-10 do not matter when we compute the relative error performance, PejH =1, since R1 1 and RX 1 both produce a correct answer in either situation, and R0 0 and RX 0 always produce an incorrect answer. Computing PejH =1 is now the di erence:
PejH =1 =
ZZ
R0 1
ZZ
R1 0
Here p(S1 S2jH = 1) is the conditional PDF for the two partial sums. This PDF is Gaussian and is centered over the point ( 1 2) where 1 = E fS1jH = 1g and 2 = E fS2 jH = 1g. Although this integral cannot be solved in closed form, we can evaluate the relative error probability ( PejH =1) numerically, and we can see the e ect of choosing di erent values for a threshold Tki for the initial test used at step ki. To design the best two-test detector, we know that there will be a nal zero-threshold test at some step kf between M and N (since we must always satisfy Pe total Pe allow). We can initially compute the Pe due to this nal test at each choice of kf . For each possible value of kf we then have some excess Pe to \spend" as we try to reduce the expected number of steps required for a decision. For each value of kf 2 fM:::N g we can nd the reduction in computation for spending our excess Pe at each possible step ki 2 f1:::(kf ; 1)g. We then choose the combination of ki and kf that minimizes the expected number of samples required for each decision:
(6.15)
This procedure allows us to design a detector that will use the minimum number of samples to produce a decision with Pe total = Pe allow using two threshold tests. If we wish to further improve performance (lower expected number of samples used for a given Pe allow , we might consider designing a test that uses threshold tests at more than two places in the computation of the weighted sum of observations. However,
115
S2 R 0,1 S 1 >T
R X,1
R 1,1
R 1,0
Figure 6-10: Plot of the six di erent regions for the di erent combination of outcomes of the two tests in the S1-S2 plane.
116
nding the best combination of step number and thresholds for each step to minimize the expected number of steps becomes more complex when more tests are used.
117
0.8
0.4
0.2
-0.2 -4 10 9 8 Cumulative Energy in Ordered Samples 7 6 5 4 3 2 1 0 0 10 20 (b) 30 40 50 60 Proportion of Samples (%) 70 80 90 100 -3 -2 (a) -1 0 Time 1 2 3 4
Figure 6-11: Plots of raised cosine and rectangular pulses (a) time domain and (b) cumulative energy in sorted samples.
118
1 0.01 0.0001 Bit Error Rate 1e-06 1e-08 1e-10 1e-12 1e-14 0 20
SNR=8 SNR=10
SNR=12
SNR=14
80
100
Figure 6-12: Plot of BER versus number of samples used in decision for various levels of SNR. a performance improvement by \spending" any excess error performance to reduce computation at the expense of a carefully-controlled increase in BER up to the desired maximum level. The design of a two-test detector for speci c set of performance speci cations proceeds in several steps. The two-test detector is designed for a speci c level of SNR and for a speci c maximum probability of error. Given these two values, we rst nd the step number for the nal threshold test, kf . This is chosen as the smallest number of samples that will always provide a error probability less than the maximum allowable, Pe kf Pe allow . Figure 6-13 shows the performance for a detection scheme using only a nal test at this lowest kf for a range of SNR values, assuming that a minimum error probability is Pe allow = 3 10;3. The unusual shape of the performance curve is due to the detection scheme using a di erent number of samples at di erent SNR levels. The performance curve jumps between members of a family of curves that would each be obtained using a xed subset of the available samples. A few of these curves are shown in Figure 6-13 as dashed lines for the cases of decisions based on using only the largest one, two, or three samples. This single-test scheme will ensure a satisfactory error probability, but does not yet provide the lowest computation. In Figure 6-13, the data point marked with a star on the modi ed performance curve represents a BER of 2:3 10;3 at a SNR of 6 dB and uses only seven of the 32 input samples. This BER is below the acceptable bound, so we can further reduce the expected amount of computation by adding a second test
119
0.1
0.0001
1e-05
Figure 6-13: Modi ed performance curve for binary detection using minimum number of samples to achieve bounded BER.
120
1e+06
100000
10000
1000
100
10
Figure 6-14: Lowest threshold value for potential values of ki. at some step ki < kf that uses a non-zero threshold. This initial test will allow some decisions to be made early to reduce the expected computation, while still ensuring that the total error probability is below the allowable level. Figure 6-14 shows the lowest threshold that can be used for each potential value of ki < kf . This information is used to determine which value for ki will yield the lowest expected amount of computation, which is shown in Figure 6-15. In the example shown in the gures, using an initial threshold test after the fourth data sample reduces the expected amount of computation to about 5.3 samples per decision, a reduction of about 25% below the result for the single, xed test after the seventh data sample. There are several points to note in these results that would be interesting areas for further work. The ability of this scheme to reduce computation depends on knowledge of the current SNR conditions. This is not unreasonable for a wireless system many current wide area wireless networks periodically transmit sequences of known data to enable the receiver to estimate channel conditions Rappaport, 1996]. The e ect of uncertainly in the SNR estimate has not been analyzed in this work, but would be important factor in an actual implementation. Equally important would be
121
6.5
5.5
Figure 6-15: Expected number of samples required for each bit decision using a twotest detector.
122
the scheme's performance under dynamic SNR conditions. This technique could be modi ed to use tests at di erent times and di erent threshold values in response to changes in SNR. This would add complexity to the implementation, but the potential savings in computation might be worth it.
6.6 Summary
This chapter describes an approach to data symbol detection that allows a wireless receiver to reduce computational complexity by computing decisions with a sub-optimal, but bounded, BER. The results show two e ective techniques for reducing computation. The rst technique is to approximate the optimal decision by using a partial evaluation of the matched lter sum. We showed that the best way to approximate the sum is to use evaluate the received samples using a greedy ordering according to the magnitude of the corresponding lter coe cients. We can compute how many samples should be evaluated in the partial sum as a function of the desired BER and the current SNR. The second technique is used to further reduce computation by introducing a second threshold test after only a portion of the approximate sum has been evaluated. This initial test can be designed to allow some of the decisions to be made early while ensuring that the acceptable level of BER is achieved. This second technique results in a run-time that is statistical, but which is lower than the original run-time that was achievable using a deterministic decision rule.
124
and the idea of the footprint of a desired result. The random sub-sampling techniques described in Chapter 5 recognize that an approximate result for a frequency selective lter needs to be computed from data that span the same interval of time, but need only use a subset of those samples in that interval. Similarly, the idea of a symbol footprint in Chapter 6 stems from the need to identify the precise set of samples that had information relevant to a speci c symbol estimate in a detector. In this case, an approximate result can be found using a subset of the samples in the footprint. This led to an investigation of which samples in the footprint were most useful and would therefore be used rst. In other situations, approximate results are not appropriate. In these cases, however, the same understanding of input-output data relationships led to algorithms that could produce the desired results using less computation by taking advantage of larger amounts of memory or removing intermediate processing steps. In all of these cases, it was also important to understand the role of each processing stage within the large communications system. It is this understanding that led to the conclusion that approximate results would be su cient in some cases, or that intermediate processing stages could be eliminated because they were unnecessary for overall system operation. One nal aspect of this work that has been instrumental in helping to guide the choice of problems and validate the results has been the implementation of the algorithms in a working wireless communications system. Some of these implementations were described in chapters 4 and 5 of this thesis. These implementation e orts helped to identify important issues for the design of new signal processing algorithms, such as the relative costs of memory access and computation, and have also led us to identify new areas for future investigation.
7.1 Contributions
This thesis has demonstrated that it is possible to design DSP algorithms for the physical layer of a wireless communications system that is itself designed to use exibility to improve overall system performance. These algorithms allow the system to take advantage of variable conditions in the wireless channel and changing system performance requirements to provide more e cient system operation. In this algorithm development work, we have achieved the following:
synthesis is a technique that e ciently synthesizes digitally modulated waveforms at IF. This technique provides a 20 ; 25 computation reduction relative to conventional techniques using a look-up table of pre-computed samples. We
7.1. CONTRIBUTIONS
125
also described table decomposition techniques that allow a exible trade-o between memory and computation required to synthesize di erent waveforms.
Conventional FIR lter design techniques assume that lters will operate on contiguous blocks of input samples. This assumption leads to an unnecessarily coupling between the length of the FIR lter and the number of input samples used. We have demonstrated a model for ltering that decouples the requirement for a long lter response from the number of input samples used, avoiding excessive computation. This decoupling is accomplished through algorithms that use only as many input samples as necessary to compute output signals with the desired SNR levels. Developed a Frequency-Translating Filter: Our technique shows how the function of arbitrary frequency translation can be combined with a decimating FIR lter through the design of a composite lter followed by a frequency shift at the lower output sample rate. We also described how the use of a complexvalued lter design can lead to an additional two-fold reduction in computation by reducing the number of coe cients required for the lter. Demonstrated the E ectiveness of Random Sub-sampling: This technique is used to realize the decoupling of bandwidth reduction and SNR control in the channel separation lter. Random sub-sampling provides an e ective way to approximate the input sample stream of a narrowband FIR lter, allowing us to control the trade-o between computation and output SNR. Using the analytical model we developed to understand this technique, we showed that amount of computation required to produce an output signal with desired SNR depends only on the output sample rate and the signal energy in the lter stopband, not on the lter input sample rate. This result allows us to increase the input bandwidth of the system without increasing the computation required to separate each channel or, conversely, to narrow the output bandwidth and reduce the required computation accordingly. Designed a Detector that Provides Controlled Error Probability: We demonstrate how to design a matched lter detector that uses multiple threshold tests to provide signi cant computational saving relative to a conventional implementation. The reduction in computation is due to two distinct e ects. First, we demonstrate that widely-used pulse shaping techniques result in a concentration of the useful signal energy in a small number of samples. Our technique identi es those most useful samples, providing a reduction in computation of ve-fold or more relative to evaluating the full matched lter sum.
126
7.3. CONCLUSIONS
127
7.3 Conclusions
The dynamic nature of the wireless operating environment and the variable performance requirements of wireless applications is in con ict with the traditional static implementations of the physical layer. The dynamic channel and variable performance requirements preclude e ective static optimization of the physical layer processing, and provide an opportunity for exible implementations to dramatically improve overall system performance. This presents an opportunity for exible algorithms that can enable the entire communication system to adapt to the changing conditions and demands. We believe that future wireless systems will have to incorporate this type of physical layer exibility if they are to provide the types of e cient communications services expected by the users of tomorrow. It is often assumed that exibility in an algorithm comes at the cost of reduced computational e ciency. In this thesis, however, we have demonstrated a suite of exible algorithms that take advantage of changing conditions and system demands to provide signi cantly improved performance relative to static designs. Although the results of this work are very encouraging, we believe that the algorithms presented are just a start, and that there are still many signi cant and interesting problems to be solved. Complex system requirements and technological advances have led to a merging of the elds of digital signal processing, theoretical computer science and communications system design. We believe that e ective designs for the systems of tomorrow will need to draw on results from all of these disciplines to meet the growing demand for communication services without wires.
128
Bibliography
Baines, 1995] Baines, R. (1995). The DSP Bottleneck. IEEE Communications Magazine, 33(5):46{54. Berlecamp et al., 1987] Berlecamp, E., Peile, R., and Pope, S. (1987). The application of error controls to communicatitons. IEEE Communiations Magazine. Bilinskis and Mikelsons, 1992] Bilinskis, I. and Mikelsons, A. (1992). Randomized Signal Processing. Prentice Hall. Bose, 1999] Bose, V. (1999). Design and Implementation of Software Radios using a General Purpose Processor. PhD thesis, Massachusetts Institute of Technology. Bose et al., 1999] Bose, V., Ismert, M., Welborn, M., and Guttag, J. (1999). Virtual Radios. IEEE JSAC issue on Software Radios. Cover and Thomas, 1990] Cover, T. and Thomas, J. (1990). Elements of Information Theory. John Wiley and Sons. Crochiere and Rabiner, 1981] Crochiere, R. and Rabiner, L. (1981). Interpolation and decimation of digital signals: a tutorial review. Proceedings of the IEEE. Frerking, 1994] Frerking, M. E. (1994). Digital Signal Processing in Communiations Systems. Van Nostrand Reinhold. Frigo and Johnson, 1997] Frigo, M. and Johnson, S. (1997). The fastest fourier transform in the west. Technical Report MIT-LCS-TR-728, MIT Laboratory for Computer Science. Grant et al., 2000] Grant, D., Randers, E., and Zvonar, Z. (2000). Data over cellular: A look at GPRS. Communications System Design. Hogenauer, 1981] Hogenauer, E. (1981). An economical class of digital lters for decimation and interpolation. IEEE Transactions on Acoustics, Speech and Signal Processing. 129
130
BIBLIOGRAPHY
Ismert, 1998] Ismert, M. (1998). Making Commodity PCs Fit for Signal Processing. In USENIX. USENIX. J. C. Liberti and Rappaport, 1999] J. C. Liberti, J. and Rappaport, T. S. (1999). Smart Antennas for Wireless Communications: IS-95 and Third Generation CDMA Applications. Prentice Hall. Jackson, 1989] Jackson, L. B. (1989). Digital Filters and Signal Processing. Kluwar Academic Publishers. Komodromos et al., 1995] Komodromos, M., Russell, S., and Tang, P. T. P. (1995). Design of FIR lters with complex desired frequency response using a generalize remez algorithm. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing. Lee and Messerschmitt, 1994] Lee, E. A. and Messerschmitt, D. G. (1994). Digital Communication. Kluwer Academic Publishers, 2nd edition. Lorch and Smith, 1998] Lorch, J. and Smith, A. (1998). Software strateies for portable computer energy management. IEEE Personal Communications. Ludwig, 1997] Ludwig, J. (1997). Low power digital ltering using adaptive processing processing. PhD thesis, Massachusetts Institute of Technology. Meyr et al., 1997] Meyr, H., Moeneclay, M., and Fechtel, S. A. (1997). Digital Communication Receivers: Synchronization, Channel Estimation, and Signal Processing. John Wiley and Sons. Mitola, 1995] Mitola, J. (1995). The Software Radio Architecture. IEEE Communications Magazine, 33(5):26{38. Nawab et al., 1997] Nawab, S. H., Oppenheim, A. V., Chandrakasan, A. P., Winograd, J. M., and Ludwig, J. T. (1997). Approximate signal processing. J. VLSI Sig. Proc. Ochi and Kambayashi, 1988] Ochi, H. and Kambayashi, N. (1988). Design of complex coe cient FIR digital lters using weighted approximation. In Proceeding of ISCAS. Oppenheim and Schafer, 1989] Oppenheim, A. V. and Schafer, R. W. (1989). Discrete-Time Signal Processing. Prentice Hall. Orfanidis, 1996] Orfanidis, S. (1996). Introduction to Signal Processing. PrenticeHall.
BIBLIOGRAPHY
131
Proakis, 1995] Proakis, J. G. (1995). Digital Communications. McGraw Hill. Rao, 2000] Rao, S. (2000). Architecture for abaptible wirelesss networks. Master's thesis, Massachusetts Institute of Technology. Rappaport, 1991] Rappaport, T. (1991). The wireless revolution. IEEE Communiations Magazine. Rappaport, 1996] Rappaport, T. S. (1996). Wireless Communications: Principles and Practice. Prentice Hall. Reed, 1998] Reed, J. H. (1998). Direct digital synthesis. In Tutorial notes from PMIRC 1998. Saltzer et al., 1981] Saltzer, J. H., Reed, D. P., and Clark, D. D. (1981). End-to-end arguments in system design. In Second International Conference on Distributed Computing Systems. Secor, 1996] Secor, M. (1996). Modelling and optimizing quality for networks of approximate processors. Master's thesis, Massachusetts Institute of Technology. Stallings, 1990] Stallings, W. (1990). Handbook of Computer-Communications standards, Vol. 2. Stallings/MacMillan. Stark and Woods, 1986] Stark, H. and Woods, J. (1986). Probability, Random Porcesses, and Estimation Theory for Engineers. Prentice-Hall, Inc. Steiglitz et al., 1992] Steiglitz, K., Parks, T. W., and Kaiser, J. F. (1992). METEOR: A constraint-based FIR lter design program. IEEE Transactions on Signal Processing. Stemm and Katz, 1997] Stemm, M. and Katz, R. (1997). Measuring and reducing energy consumptio of network interfaces in hand-held devices. IEICE Transactions on Fundamentals of Electronics, Communications, and Computer Science, Special Issue on Mobile Computing. Vasconcellos, 1999] Vasconcellos, B. (1999). Parallel signal processing for everyone. Master's thesis, Massachusetts Institute of Technology. Welborn, 1999] Welborn, M. (1999). Direct waveform synthesis for software radios. In Proceedings of WCNC. Wepman, 1995] Wepman, J. A. (1995). Analog-to-Digital Convertors and Their Applications in Radio Receivers. IEEE Communications Magazine, 33(5):39{45.
132
BIBLIOGRAPHY
Wicker, 1995] Wicker, S. (1995). Error Control Systems for Digital Communications and Storage. Prentice Hall. Wiesler and Jondral, 1998] Wiesler, A. and Jondral, F. (1998). Software radio structure for second generation mobile communication systems. In Proceedings of VTC. Winograd, 1997] Winograd, J. M. (1997). Incremental re nement structures for approximate signal processing. PhD thesis, Boston University. Winograd et al., 1996] Winograd, J. M., Nawab, J. T. L. S. H., Oppenheim, A. V., and Chandrakasan, A. P. (1996). Flexible systems for digital signal processing. In AAAI Fall Symposium on Flexible Computation. Zangi and Koilpillai, 1999] Zangi, K. and Koilpillai, R. D. (1999). Software radio issues in cellular basestations. IEEE JSAC issue on Software Radios. Zilberstein, 1993] Zilberstein, S. (1993). Operational Rationality through Compilation of Anytime Algorithms. PhD thesis, Computer Science Division, University of California at Berkeley. Zilberstein, 1996] Zilberstein, S. (1996). Optimal composition of real-time systems. Arti cial Intelligence.