ALMA MATER STUDIORUM - UNIVERSITA DI BOLOGNA ARCES - ADVANCED RESEARCH CENTER ON ELECTRONIC SYSTEMS Low Correlated Sequenes & Arhitetures and Algorithms for Analog-to-Information Converters: Theory, Design, Implementation and Appliations A THESIS PRESENTED BY Salvador Javier Haboba SUPERVISORS Professor Riardo Rovatti Professor Gianlua Setti Professor Sergio Callegari COORDINATOR Professor Claudio Fiegna DOCTORATE ON INFORMATION TECHNOLOGY JANUARY 2010 - DECEMBER 2012 XXV CYCLE - ING-INF/01 ABSTRACT Most eletroni systems an be desribed in a very simplied way as an assemblage of analog and digital omponents put all together in order to perform a ertain funtion. Nowadays, there is an inreasing tendeny to redue the analog omponents, and to replae them by operations performed in the digital domain. This tendeny has led to the emergene of new eletroni systems that are more exible, heaper and robust. However, no matter the amount of digital proess implemented, there will be always an analog part to be sorted out and thus, the step of onverting digital signals into analog signals and vie versa annot be avoided. This onversion an be more or less omplex depending on the harateristis of the signals. Thus, even if it is desirable to replae funtions arried out by analog omponents by digital proesses, it is equally important to do so in a way that simplies the onversion from digital to analog signals and vie versa. In the present thesis, we have study strategies based on inreasing the amount of proessing in the digital domain in suh a way that the implementation of analog hardware stages an be simplied. To this aim, we have proposed the use of very low quantized signals, i.e. 1-bit, for the aquisition and for the generation of partiular lasses of signals. More speially, on one hand, we have proposed a method for the generation of sets of binary sequenes to be used in multiple-input multiple-output ative sensing appliations, suh as radar, sonar and medial imaging. The generated sets of sequenes have very low auto- and ross-orrelation sidelobes, a desired property for this kind of appliations, providing performane metris far better than those from other families of binary sequenes, and a omparable performane ii to that of multibit approahes. The advantage of using binary sequenes is, for instane, the simpliation of the implementation of the transmitters always present in these appliations. On the other hand, we have proposed a new arhiteture for an analog to digital onverter. This arhiteture an be viewed as an extension of the funtionalities of a lassial Delta-Sigma onverter whih, by taking 1-bit measurements at a rate muh bigger than that of the signal bandwidth, produes a signal estimate with an auray that depends on the ratio between the sampling rate and the signal bandwidth. In our ase, relying on the struture of the signal of interest, and assuming that its information ontent is muh smaller than its bandwidth, we are able to produe a signal estimate that depends on the ratio between the sampling rate and the information ontent of the signal. ACKNOWLEDGMENTS The possibility to perform this thesis was thanks to the Erasmus Mundus sholarship that naned my stage at the University of Bologna. I would like to thanks authorities from Erasmus for giving me this opportunity. I would also want to thanks authorities and researhers from the Advane Researh Center on Eletroni Systems (ARCES) for aepting me as part of this big teem. Thanks to my supervisors Riardo Rovatti, Gianlua Setti and Sergio Callegari for guiding me during these years, for sharing with me their knowledge, and for introduing me on the amazing eld of ompressing sensing and on many other topis. I would like to thanks my mates: Carlos, Mauro, Salvatore, Fabio and Valerio, with whom I shared most of the time, sometimes disussing about our researhes and sometimes just hatting about life. Coming from a different ountry, life in Italy and espeially in Bologna was easier and even wonderful thanks to all the friends I made. They supported me and enouraged me during difult periods, and I share with them speial moments that made of Bologna one of the best ities to live. Thanks to Vitor, Silvia, Mart´n, Ferha, Gonzalo, Carlos, Mariano Yanina, Sandra y Laura (las hirusas), Maria Jose y Sof´a (las hirusinas), the Celti friends (Mark, Aaron, JB, Kyle, Valentina, Cristina Chapinara), Sahar and Nasrin. Thanks also to my friend Demian for the ideas we shared through the distane, for understanding my absene and for some advies in this work. I would like to thank my family: my parents and my sisters for their onstant iv support and enouragement. Finally, I would like to speially thank Vitoria, for her invaluable support during all the time we spent in Bologna, La Plata and everywhere, for her advies, enouragement, for being my partner and for making me enjoy life together. Y si trabajás al pedo y estás haiendo algo nuevo, adelante ... Charly Garia CONTENTS 1. : : : : :: : : : : : : : : : :: : : : : : : : : : :: : : : 1 1.1 Sets of Low Correlated Sequenes . . . . . . . . . . . . . . . . . 3 1.2 Analog to Information Converters . . . . . . . . . . . . . . . . . 4 1.3 Overview and Main Contributions . . . . . . . . . . . . . . . . . 5 Introdution Part I 2. 7 Low Correlated Sequenes : : : : :: : : : : : : : : : :: : : : 9 2.1 Introdution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 Integrated Sidelobe Level Problem 2.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . 11 2.3 Calulation of the ross-orrelation terms in the ISL . . . . . . . . 14 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes : : : : 21 3.1 Legendre Sequenes . . . . . . . . . . . . . . . . . . . . . . . . 21 3.2 Sets of RLS minimizing ISL . . . . . . . . . . . . . . . . . . . . 31 3.3 Numerial results . . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.4 Conlusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 Part II 4. Contents vi Analog to Information Conversion 41 Signal Models : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 43 4.1 Introdution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 4.2 Sparse Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 5. Compressive Sensing : : : : : : : : : : : : : : : : : : : : : : : : : : 47 5.1 Introdution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 5.2 The Restrited Isometry Property . . . . . . . . . . . . . . . . . . 48 5.3 CS Reonstrution Algorithms . . . . . . . . . . . . . . . . . . . 50 5.3.1 Optimization Based Reonstrution Algorithms . . . . . . 50 5.3.2 Support-Guessing Reonstrution Algorithms . . . . . . . 53 5.4 Analog-to-Information Converters . . . . . . . . . . . . . . . . . 54 6. 5.4.1 Random-Modulation-Pre-Integration RMPI . . . . . . . 55 5.4.2 Random Sampling RSAM . . . . . . . . . . . . . . . . 57 5.4.3 1-bit Compressive Sensing - 1bRMPI . . . . . . . . . . . 59 RADS Converter : : : : : : : : : : : : : : : : : : : : : : : : : : : : 61 6.1 Introdution and Motivation . . . . . . . . . . . . . . . . . . . . 61 6.1.1 Preliminaries: Delta-Sigma Modulation . . . . . . . . . . 62 6.2 RADS Converter arhiteture . . . . . . . . . . . . . . . . . . . . 67 6.3 Deoding and reonstrution . . . . . . . . . . . . . . . . . . . . 72 6.3.1 Reonstrution Signal to Noise Ratio Estimation . . . . . 74 6.3.2 Numerial Experiments . . . . . . . . . . . . . . . . . . 75 6.4 FCOSAMP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 6.4.1 Numerial Experiments . . . . . . . . . . . . . . . . . . 86 Contents vii 6.5 Time Domain Analysis . . . . . . . . . . . . . . . . . . . . . . . 93 6.5.1 = modulator time-domain analysis 6.5.2 RADS Converter time-domain analysis . . . . . . . . . . 101 6.5.3 Spae Dimension Analysis . . . . . . . . . . . . . . . . . 102 6.5.4 L1-norm Minimization . . . . . . . . . . . . . . . . . . . 105 6.5.5 Numerial Experiments . . . . . . . . . . . . . . . . . . 105 . . . . . . . . . . . 93 6.6 Hardware Implementation . . . . . . . . . . . . . . . . . . . . . 108 6.6.1 Measurement Setup . . . . . . . . . . . . . . . . . . . . . 109 6.6.2 Measurements and Validation . . . . . . . . . . . . . . . 111 6.7 Conlusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 7. Conlusions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 117 LIST OF FIGURES 3.1 Plots of ISL for M = 2 as a funtion of f1 and f2 : top: 3D-view, bottom: iso-ISL lines . . . . . . . . . . . . . . . . . . . . . . . . 29 3.2 Integrated Sidelobe Level as a funtion of sequene length. In blue, RLS with rotations minimizing asymptoti ISL; in blak, asymptoti value. . . . . . . . . . . . . . . . . . . . . . . . . . . 30 3.3 Integrated Sidelobe Level as a funtion of sequene length. In blue, RLS with an arbitrary rotation; in blak, asymptoti value. . 30 3.4 Integrated Sidelobe Level as a funtion of sequene length. Comparison of an RLS with an arbitrary rotation and RLS with rotations minimizing asymptoti ISL. . . . . . . . . . . . . . . . . . 31 3.5 Comparison between ORLS, with Q.CAN algorithm when quantization is imposed for (a) M = 4, (b) M = 8 and () M = 12, and x value of N = 1033. The performane metri as a funtion of the number of quantization steps. The performane of Q.CAN exeeds the performane of the binary ORLS for Q grater than 13 for M = 4, 18 for M = 4 and 22 for M = 12. . . . . . . . . . . 35 3.6 Comparison between ORLS, random sequenes, Gold sequenes and Q.CAN sequenes with binary quantization for (a) M = 2, (b) M = 3 and () M = 4. The solid horizontal line at 1 identies the theoretial maximum performane while the dashed horizontal line marks the asymptoti performane ahieved by RLS. . . . 36 List of Figures x 3.7 Comparison between ORLS, ORBS and Q.CAN sequenes for (a) M = 2, (b) M = 3 and () M = 4. The solid horizontal line at 1 identies the theoretial maximum performane while the dashed horizontal line marks the asymptoti performane ahieved by RLS. 38 5.1 Blok sheme of an RMPI enoder. The samples of the input signal are multiplied by M different random sequenes and aumulated up to time N . The aumulated values are then quantized by a b bit AD onverter. . . . . . . . . . . . . . . . . . . . . . . . 55 5.2 Blok sheme of an random sampling enoder. The samples are taken at random positions in time, over a predened grid. . . . . . 58 6.1 A time-domain blok diagram of a rst order = modulator. . . 63 6.2 A z -domain linear model of a rst order = modulator. . . . . . 64 6.3 Blok sheme of a RADS Converter. The input signal is multiplied by a random sequene and fed into a = onverter made of a 1-bit ADC and a loop lter in harge of noise shaping. . . . . 67 6.4 Frequeny oupany at the different point of the system. From top to bottom: spetrum of the input sparse signal in the Fourier domain x; spetrum of the modulating signal p; spetrum of the modulated signal y as a sum of different shifts of the modulating signal; spetrum of the output signal z with the addition of the quantization noise shaped by the NT F of the = modulator; remaining spetrum after low-pass ltering. . . . . . . . . . . . . 70 6.5 Deoding and reonstrution sheme for RADS Converter. The 1-bit input signal is rst ltered, deimated, and then proessed by a ompressive sensing reonstrution algorithm. . . . . . . . . 72 List of Figures xi 6.6 Effet of the reonstrution lter bandwidth for a RADS Converter with a third order = modulator for different sparsity levels. On top: RSNR as a funtion of lter bandwidth R; bottom: P SR as a funtion of R. For every ombination of K and N there is an optimal value for R. . . . . . . . . . . . . . . . . . . . . . . 77 6.7 Simulation of the performane of RADS Converter using the optimal reonstrution lter bandwidth for different sparsity levels. RSNR as a funtion of the oversampling ratio M=N . The support was orretly reovered 100% of the time. The solid line represents the simulation result, while the dashed line the theoretial result from eq. (6.8). . . . . . . . . . . . . . . . . . . . . . . . . 79 6.8 Simulation of the performane of the RADS Converter using the optimal reonstrution lter bandwidth for different sparsity levels. The input signal has an intrinsi SNR of 30 dB. RSNR as a funtion of the oversampling ratio M=N . The support was orretly reovered 100% of the time. . . . . . . . . . . . . . . . . . 80 6.9 Simulation of the performane of RADS Converter using the FCoSaMP algorithm in the reonstrution for different sparsity levels. On top: RSNR as a funtion of the oversampling ratio M=N ; bottom: P SR as a funtion of M=N . The large RSNR that is ahieved translates into a ompression fator of up to 15 times with respet to Nyquist based aquisition . . . . . . . . . . 87 6.10 Simulation of the performane of RADS Converter using the FCoSaMP algorithm for reonstrution, for different sparsity levels. The input signal has an intrinsi SNR of 30 dB. On top: RSNR as a funtion of the oversampling ratio M=N ; bottom: P SR as a funtion of M=N . The enoding proess and the reonstrution algorithm show to be robust against strong noise ondition. 89 List of Figures xii 6.11 Simulation of the performane of RADS Converter by using the FCoSaMP algorithm for reonstrution, for different sparsity levels. The input signal is sparse in a random basis. On top: RSNR as a funtion of the oversampling ratio M=N ; bottom: P SR as a funtion of M=N . The proposed arhiteture shows to work independently of the sparsity basis provided it is spread on the time axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 6.12 Comparison of the performane of RADS Converter deoded with FCoSaMP with 1bRMP I deoded with the RSS and BIHT algorithms. RSNR as a funtion of the oversampling ratio M=N for x signal length N = 1024 and sparsity level of K = 10. The same amount of 1-bit samples are onsidered for every ase. . . . 92 6.13 First order = modulator shemati diagram. . . . . . . . . . . 94 6.14 Lth -order = modulator shemati diagram. . . . . . . . . . . . 96 6.15 Monte Carlo integration of the hyper-volume of the solution spae for RADS Converter and for = onverter. On top: volume in linear sale as a funtion of the number of measurements M ; bottom: volume in logarithmi sale as a funtion of M . The enoding performed by RADS is more effetive in reduing the size of the solution spae. . . . . . . . . . . . . . . . . . . . . . . 104 6.16 Performane omparison of FCoSaMP and L1min. RSNR as a funtion of oversampling ratio M=N for an 8-sparse signal enoded with RADS onverter. . . . . . . . . . . . . . . . . . . . . 106 6.17 Reonstrution algorithm exeution time for FCoSaMP and L1min. Relationship between the time taken by L1min and FCoSaMP as a funtion of the oversampling ratio M=N . . . . . . . . . . . . . . 107 6.18 Simplied shemati diagram of the hardware implementation of the RADS Converter. . . . . . . . . . . . . . . . . . . . . . . . . 108 6.19 Piture of the hardware implementation of the RADS Converter. . 108 List of Figures xiii 6.20 Piture of the RADS Converter onneted to a Spartan 6 FPGA development kit. . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 6.21 Measurement setup for the evaluation of the hardware implementation of the RADS Converter. . . . . . . . . . . . . . . . . . . . 110 6.22 Aquisition of an analog signal with RADS Converter. The input signal is 8-sparse in a random basis and with a Nyquist rate half the sampling frequeny. On top: the syntheti signal superimposed to the reonstruted signal for the whole aquisition window; on bottom: a zoom-in of the same aquired signal. . . . . . . 112 6.23 Spetrum of the input signal aquired by RADS Converter. The spetrum has a full oupany for frequenies up to 2:5MHz . . . 113 6.24 Performane of the hardware implementation of the RADS Converter by using the FCoSaMP algorithm for reonstrution, for different sparsity levels. RSNR as a funtion of the oversampling ratio M=N . The support reovery was always orret for the 10 measurements. . . . . . . . . . . . . . . . . . . . . . . . . 114 1. INTRODUCTION Although most modern eletronis systems are omposed by a ombination of analog and digital omponents, the rapidly evolving apabilities of digital eletronis are shifting every funtion (before) handled in the analog domain into the digital domain. The advanes in integrated iruits design have enabled the reation of digital proessing systems that are more exible, heaper and robust than their analog ounterparts. This has lead to one of the most signiant development during the last deades of eletroni systems design: replaing analog omponents to perform their operation in the digital domain. However, for these systems to interfae with the real world, onversions between analog signals and digital signals are required. Analog-to-Digital (AD) and Digital-to-Analog (DA) onverters are the responsible of that onversion. Most AD and DA onverters rely on the Nyquist-Shannon sampling theorem that determines how any signal an be exatly reovered from a set of uniformly spaed samples taken at a rate of at least twie the highest frequeny present in the signal of interest. Nyquist-Shannon sampling theorem imposes a requirement on the time domain to the problem of how to represent an analog signal by a series of samples without any lost of information. However, in order to proess and store samples in a digital system, we must be able to represent eah sample using a nite number of bits, and hene the measurements will typially be subjet to the unavoidable quantization error. By inreasing the number of bits of the measurements, the 1. Introdution 2 quantization proess an be negleted, or hidden with respet to proesses present in the system suh as thermal noise. The main drawbak of this approah, is that the ost of inreasing the number of bits for Nyquist based AD and DA onverters require a huge amount of analog hardware. As an example, ash AD onverters exponentially inrease its hardware omplexity with the number of resolution bits, and beomes impratial at resolutions over 8 bits due to the large number of omparators required. There is a different approah for AD and DA onversion that it is not based on the Nyquist-Shannon sampling theorem. Delta-Sigma onverters rely on the utilization of a very small amount of bits to quantize signals (1-bit quantization is the most typial value used). Delta-Sigma onverters ahieve this by trading-off resolution with the sampling frequeny. These onverters oversample the signal by a large fator with respet to its bandwidth and, by a ltering proessing (analog or digital) they are able to obtain a nal signal represented with an auray muh bigger than the one used in the sampling proess. Among other advantages (low power, low ost) with respet to other onverter arhitetures, at the heart of the Delta-Sigma is the simpliation on the quantization stage that allows the onverter to operate with no linearity degradation. However, Delta-Sigma onverters only allows to efiently operate with signals with a redued spetra oupany, due to the high oversampling ratio needed to obtain the desired preision. The main motivation of this dissertation is to study simpliation strategies for the implementation of analog hardware stages present in most mixed systems, by inreasing the amount of proessing in the digital domain. We aomplished this by proposing the utilization of very low quantized signals, i.e. 1-bit, for the aquisition and for the generation of partiular lasses of signals. Following this approah we have arhived performanes similar (or even better) than those obtained through multibit approahes. We rst present the use of Legendre sequenes (1-bit sequenes) for the gener- 1. Introdution 3 ation of sets of sequenes with good auto- and ross-orrelation properties. These Sets of Low Correlated Sequenes an be used in MIMO (Multiple Input Multiple Output) ative sensing systems ahieving a signiant improvement with respet to other sets of binary sequenes, and a similar performane to the one ahieved by multibit approahes. Seondly, we present a new arhiteture for an Analog to Digital onverter (or more preisely, an Analog to Information Converter) that, based on a DeltaSigma onverter, produes a stream of 1-bit measurements, and ahieves a reonstrution performane proportional to the signal information ontent instead of that of the signal bandwidth. 1.1 Sets of Low Correlated Sequenes The design of sequenes sets with low aperiodi auto- and ross-orrelations is present in many elds of engineering and plays an important role in many appliations suh as radar, sonar, ommuniations, medial imaging and other ative sensing appliations. The task of designing sets of sequenes with presribed orrelation properties is a partiular ase of the general problem of waveform synthesis that is often a key point in establishing the performane of transmission, synhronization, or ative sensing systems [1, 2℄. Good auto-orrelation properties means that any sequene in the set is nearly unorrelated with its own shifted version while good ross-orrelation means that any member of the sequenes set is nearly unorrelated with any other members at any shift. A ommonly used metri of the goodness of the orrelation is the Integrated Sidelobe Level (ISL), being good set of sequenes those having a low ISL value. Although many state-of-the art algorithms were proposed for the minimization of the ISL [3, 4, 2, 5, 6℄, their performane is largely impaired when quantization 1. Introdution 4 is taken into onsideration. However, implementation onstraints strongly favors disrete-valued signals, possibly enforing quantization to an extremely limited number of levels. What we propose here is a proedure to onstrut sets of antipodal sequenes with extremely low ISL. The resulting performane largely exeeds that of lassial methods for the diret generation of low-ISL sets of sequenes. 1.2 Analog to Information Converters Analog to Digital onversion is one of the most important operations in signal proessing. It maps a ontinuous-time and real-value signal into a disrete sequene of disrete values. Classial sampling methods rely on the hypothesis that the analog signal to be aquired is band-limited, and the Nyquist-Shannon theorem states the minimum distane between samples (or Nyquist rate) needed to uniquely desribe the analog signal by its samples. While the assumption of bandlimited signals is of broad appliation, many natural signals when represented in a proper basis, orrespond to vetors in whih many omponents have a small value, or represent a small fration of the total energy. This harateristi alled sparsity is usually exploited to represent the signal with a muh smaller amount of data, and loser to the signal information ontent. A novel sampling paradigm that goes against the ommon approah in data aquisition has emerged in the last years and is alled Compressive Sensing (CS ) [7, 8, 9℄. CS theory asserts that one an reover ertain signals and images from far fewer samples or measurements than those used by traditional methods. This is possible due to the fat that many natural signals are sparse or ompressible, and, by measuring in a partiular way, it is possible to aquire the omplete information ontent of those kind signals. Analog-to-Information onverters relies on this idea, to measure the informa- 1. Introdution 5 tion ontent of the signal instead of measuring the omplete redundant data available for a partiular measurement domain. Following this approah, we have proposed a novel arhiteture for an Analogto-Information onverter that allows a simple hardware implementation for the aquisition of large bandwidth signals that are sparse over a variety of supports. 1.3 Overview and Main Contributions This thesis is mainly onerned on how signal proessing tehniques an be applied to real hardware appliations and help to redue the omplexity of its implementation. This work is divided into two parts, the rst part takle the problem of sequenes synthesis, and how a proper design of simple antipodal signals an ahieve a performane similar to that obtained using multibit sequenes for ative sensing appliations. The main ontributions of this part are: an analysis of the degradation of state-of-the-art algorithms for sequenes synthesis when quantization is imposed; a method based on generating funtions for the alulation of the rossorrelation omponents of the ISL of a set of sequenes; a proedure to onstrut sets of antipodal sequenes with extremely low ISL; an analytial expression for the asymptoti ISL of sets of rotated Legendre sequenes. The seond part of the thesis onerns about the implementation of a Analogto-Information onverter that produes a stream of 1-bit measurements and a nal resolution after reonstrution that is proportional to the information ontent of the signal. The main ontribution of this part are: 1. Introdution 6 a new arhiteture for ompressive sensing that produes 1-bit measurements; a new reonstrution algorithm for the proposed arhiteture that exploits not only the sparsity hypothesis but also the hardware arhiteture of the aquisition system; a theoretial analysis of the apabilities of a Delta-Sigma modulator to extrat the information ontent of a signal, that is later extended for the analysis of the proposed arhiteture; a hardware implementation of the proposed arhiteture, and a measurement setup to validate the theoretial analysis. We onlude this thesis work with a summary of our ndings in Chapter 7. Part I LOW CORRELATED SEQUENCES 2. INTEGRATED SIDELOBE LEVEL PROBLEM 2.1 Introdution The design of sequenes sets with good orrelation properties is present in many elds of engineering suh as radar, sonar, ommuniations, medial imaging and so on. Ative sensing appliations, have been greatly beneed by the use of multiple-input multiple-output (MIMO) systems. This kind of systems, transmit orthogonal waveforms via its antennas allowing to ahieve a great inrease virtual aperture. As an example, traditional phased-array radar system only transmits a single waveform trough its antennas. However, by the use of MIMO radar system a large inrease in parameter identiability [10℄, detetion performane [11℄, and resolution [12℄ an be ahieved. Besides orthogonality, good auto- and ross-orrelation properties of the transmitted waveforms are also often required [13, 14, 15℄. Good auto-orrelation properties means that any sequene in the set is nearly unorrelated with its own shifted version while good ross-orrelation means that any member of the sequenes set is nearly unorrelated with any other members at any shift. The design of a set of signals with small auto-orrelation sidelobes and small ross-orrelation between sequenes at any time delay ensure that the reeiver mathed lter an extrat the desired information while attenuating undesired signals. 2. Integrated Sidelobe Level Problem 10 A ommonly used metri of the goodness of the orrelation is the Integrated Sidelobe Level (ISL). The ISL of a set of M sequenes eah of N (possibly omplex) symbols that we will indiate with x(jp) with j = 0; : : : ; N 1 and p = 0; : : : ; M 1 is dened as ISL = X1 M 1M 1 X X N X1 2 jXx(p)x(p) (k)j + jXx(p)x(q) (k)j2 p=0 k = N +1 p=0 q =0 k = N +1 k 6=0 p6=q M X1 N (2.1) where Xx p x p (k) = minfNX k;N g 1 ( ) ( ) j =maxf0; kg x(jp) xj +(pk) k = N + 1:::N 1 k = N + 1:::N 1 is is the auto-orrelation of the sequene x(p) , and Xx p x q (k) = minfNX k;N g 1 ( ) ( ) j =maxf0; kg x(jp) xj +(qk) is the ross-orrelation between the sequenes x(p) and x(q) . Good set of sequenes are those having a low ISL value. Due to the strong interest in the design of sequenes with low ISL value, many algorithms have been suggested for its minimization [4, 2, 5, 6, 1℄. Suh a problem may be far from trivial when onstraints are introdued. For example, reeption may have to be stopped after a ertain time thus spoiling the adoption of periodi signals and leading to the onsideration of lipped or aperiodi orrelations. Further to that, implementation strongly favors disrete-valued signals, possibly 2. Integrated Sidelobe Level Problem 11 enforing quantization to an extremely limited number of levels. This latter onstraint, in partiular, is known to make optimization-based methods hard to apply sine ontinuous-optimization must either undergo quantization or be simply disarded in favor of almost exhaustive sans. Within this senario, starting from the lassial problem of designing an antipodal sequene with a low Integrated Sidelobe Level (ISL) [16℄ we address its generalization to sequene sets, for whih lobes are onsidered both for autoorrelation and for ross-orrelations. 2.2 Problem Formulation Given M and N , and based on (2.1) the general problem is that of nding the sequene set minimizing the ISL. Commonly, a further unimodularity onstraint is put on the sequenes thus requiring that jx(jp) j = 1 for p = 0; : : : ; M 1 and j = 0; : : : ; N 1. Suh a onstraint is appliation-driven in that it eases the implementation of the transmitters managing the eletrial signals orresponding to the sequene symbols. In fat, (p) in this ase one may set x(jp) = eij , where i is the imaginary unit, with j(p) 2 ( ; ℄ and design the set of phase sequenes fj(p)gNj=01 for p = 0; : : : ; M 1 that an be simply transmitted by onstant-envelope modulations. Given this onstraint, it is known that the ISL annot be dereased below its lower bound [17℄ ISLmin = N 2 M (M 1) so that the effetiveness of any approah an be measured in normalized terms by min ISL = ISL better approahes featuring an loser to 1. 2. Integrated Sidelobe Level Problem 12 It is well known that sets of unimodular sequenes with extremely high effetiveness an be obtained by the appliation of algorithms [3℄ that are extensions of those devised to minimize ISL in the single sequene ase (M = 1) [16℄. Yet, when those algorithms meet the even more implementationfriendly onstraint of antipodal sequenes, i.e. x(jp) = 1 for p = 0; : : : ; M 1 and j = 0; : : : ; N 1 , their effetiveness is largely impaired. Atually, the antipodal problem is reognized as being muh more difult: a known effet of the impossibility of addressing it with the tools of ontinuous optimization and the need of resorting to enumeration-based disrete optimization tehniques. In the following we onentrate on antipodal sequenes. Under suh an assumption, the partiular ase M = 1 in whih only autoorrelation terms appear, has attrated a lot of attention by itself. This led to a onspiuous literature analyzing more than a family of sequenes for whih ISL or the equivalent Merit Fator MF = N 2 =ISL an be omputed analytially at least in the asymptoti ase N ! 1 (see, e.g., [18, 19, 20, 21, 22℄). Beyond that a list of best known sequenes [23℄ is available for N up to 304. Our purpose is to develop an analytial expression that may drive optimization in some partiular difult ases, most notably when the antipodal onstrain (xpj = 1) is imposed. To failitate the disussion, denote the sum of squares orresponding to the auto-orrelation terms as Xx p x p ( ) ( ) 1 N X = k = k N +1 jX x(p) x(p) (k)j2 6=0 and the sum of squares orresponding to the ross-orrelation terms as (2.2) 2. Xx p x q ( ) ( ) Integrated Sidelobe Level Problem 1 N X = k = N +1 jX x(p) x(q ) 13 (k)j2 p 6= q (2.3) so that ISL = X1 M =0 Xx p x p ( ) ( ) p + X1 M X1 M p =0 q=0 p= 6 q Xx p x q ( ) (2.4) ( ) A general method for the alulation of Xx(p) x(p) of any sequenes of odd length is presented in [19, 24℄. This method hinges on generating funtions and writes orrelations as proper sums of their values on the unit irle in the omplex plane. The method works well when we have analytial insights on the generating funtions. Extending the ideas of [19℄, in setion 2.3 we devise a general method for the alulation of Xx(p) x(q) in (2.3) of any pair of real sequenes of odd length and thus, together with the result in [19, 24℄, the ISL for a set of sequenes. In setion 3.1 we use this method to obtain an asymptoti expression for the ISL value of a set formed by different rotations of Legendre sequenes. Finally, in setion 3.2 we propose an optimization proedure based on the latter expression where we nd the optimal rotations that minimize the ISL for any sequenes length N. Throughout this hapter we use the following asymptoti notation: We say that two sequenes aN and bN are asymptotially equivalent, aN a lim =1 !1 b N N a N N is asymptotially bounded by bN , aN = O(b ) iff N b N iff 2. Integrated Sidelobe Level Problem 9M > 0 and 9N o 2.3 14 ja j M jb j N 8N > N N o Calulation of the ross-orrelation terms in the ISL Let a0 ; a1 ; : : : ; aN 1 and b0 ; b1 ; : : : ; bN 1 be two real sequenes of length N, we want to obtain an expression for Xab . If we dene the generating funtions of the two sequenes as Qa (z ) = X1 N j =0 aj z Qb (z ) = j X1 N j =0 bj z j we have that 1 N X Qa (z )Qb (z ) = k = N +1 Xab (k)z k and thus N 1 X 2 jQa (z)Qb (z)j = k Now, set j = e N 2 i j X1 N j =0 Hene, if we dene = N X1 N +1 l= N +1 Xab (k)Xab (l)z k+l and note that for k; l = N + 1; : : : ; N j k+l = 8 <N 0 : if l + k = N; 0; N otherwise 1, 2. Integrated Sidelobe Level Problem 15 1 X S0 = jQa (j )Qb (j )j2 N =0 X1 j N =N k = N +1 X 2 (k)+N 2X N 1 ab k =1 1 X Xab (k)Xab (k N )+N k = N +1 Xab (k)Xab (k+N ) and (for N odd) 1 X S 00 = jQ N =0 N 1 X a ( )Q ( )j2 j b j =N k = N +1 Xab2 (k)+ N j 2X N 1 k =1 1 X Xab (k)Xab (k N ) N k = N +1 Xab (k)Xab (k+N ) we an express Xab (i.e. the sum of squares of ross-orrelations as in (2.3)) as Xab X1 N = k = N +1 Xab2 (k) = S 0 + S 00 2N To ompute S 00 we use the Lagrange interpolation polynomials to alulate the values of Qa ( j ) from Qa (k ) for j; k = 0; : : : ; N 1. In this speial ase the data points (k ) oinide with the omplex roots of unity and, for N odd, the Lagrange base polynomials simply redue to N2 j +kk [25, p. 89℄. Then Qa ( j ) = 2 X1 N k N k=0 j + k Qa (k ) (2.5) By substituting (2.5) into S 00 and developing the produt jQa ( j )Qb ( j )j2 we get 2. Integrated Sidelobe Level Problem 16 X1 N S 00 = " 1 k N X 1 j + k k =0 N 1 X 1 1 1 ) 1 l N X + Qa (l ) l l =0 j # N 1 X k l Qb (k ) Qb (l ) + + j k l l =0 j k =0 16 NX1 NX1 NX1 NX1 Q ( )Q ( )Q ( )Q ( ) N 4 k =0 l =0 k =0 l =0 a k a l b k b l N 1 X k l l k j + k j + l j + k j + l j =0 N 4 j =0 Qa (k 16 1 1 1 1 2 2 2 = 2 2 2 2 2 1 1 1 2 1 2 2 2 1 2 1 1 2 2 1 2 in whih we may exploit the fat that j = 1=j to write S 00 = 16 N4 X1 N X1 N X1 N X1 N k1 k k Qa (k )Qa (l )Qb (k )Qb (l =0 l1 =0 k2 =0 l2 =0 N X1 j 1 =0 j 2 1 1 + 1 j k1 j + l j 1 2 1 + k2 2 ) j (2.6) j + l 2 Let us dene now the innermost sum of (2.6) as W (k1 ; l1 ; k2 ; l2 ) = X1 N j with =0 j 1 + j k1 j + l j 1 1 + k2 j j + l 2 = X1 N j =0 fk ;l ;k ;l (j ) 1 1 2 2 z2 fp;q;r;s(z ) = (z + p)(z + q )(z + r )(z + s) Depending on p; q; r; s, the rational funtion fp;q;r;s(z ) an be transformed into a spei sum of simple rational parts. Eah of these rational parts an be summed 2. Integrated Sidelobe Level Problem 17 separately. This path is fully developed in [19℄ and we here exploit the results therein. In partiular we have that A) for 0 p < N 1 1 2 1 4 2 W (p; p; p; p) = N + N 16 3 3 2 p B ) for 0 p 6= q < N W (p; p; p; q ) = W (p; p; q; p) = W (p; q; p; p) = q + p 1 2 W (q; p; p; p) = N 8 p(q p )2 C ) for 0 p 6= q 6= r < N W (p; p; q; r) = W (p; p; r; q ) = W (p; q; p; r) = W (p; r; p; q ) = W (p; q; r; p) = W (p; r; q; p) = W (q; p; r; p) = W (r; p; q; p) = W (q; r; p; p) = 1 1 1 W (r; q; p; p) = N 2 4 q p r p D) for 0 p 6= q < N W (p; p; q; q ) = W (p; q; p; q ) = W (p; q; q; p) = E ) for 0 p 6= q 6= r 6= s < N 1N2 1 2 ( )2 p q 2. Integrated Sidelobe Level Problem 18 W (p; q; r; s) = 0 Taking into aount all the above ases we may write S 00 = 16 ( + + + Æ) N4 , where the terms , , , and Æ orrespond to the ontributions of the ases A, B, C and D respetively. For the ases inluded in A) we have that X1 1 1 2 2 4 2 = 16 3 N + 3 N =0 jQ ( )Q ( )j N a p b (2.7) p p for the ases in B ) we have 1 X1 = N2 8 =0 N p;q p q 6= ( q + p p (q p )2 (2.8) h 2p jQa (p )j2 Qb (p )Qb (q )+ p q jQa (p )j2 Qb (q )Qb (p )+ 2 Qa (p )Q (q ) jQb (p )j2 + p a q p Qa (q )Qa (p ) jQb (p )j2 for C ) we have ) i 2. 1 = N2 4 Integrated Sidelobe Level Problem ( 1 N X =0 6= 6= p;q;r p q r 1 )( ( q h p p ) r 19 (2.9) p q jQa (p )j2 Qb (q )Qb (r )+ p r jQa (p )j2 Qb (r )Qb (q )+ 2p Qa (p )Qa (q )Qb (p )Qb (r )+ 2 Qa (p )Q (r )Qb (p )Q (q )+ p a b p r Qa (p )Qa (q )Qb (r )Qb (p )+ Q ( )Q ( )Q ( )Q ( )+ p q a p a r b q b p q r Qa (q )Qa (p )Qb (r )Qb (p )+ Q ( )Q ( )Q ( )Q ( )+ r q a r a p b q b p q p Qa (q )Qa (r ) jQb (p )j2 + r p Qa (r )Qa (q ) jQb (p )j2 ) i and for D) ( 1 h jQ ( )Q ( )j2 + 1 X1 Æ = N2 2 2 =0 ( ) N p;q p q 6= p q p q a p b (2.10) q ) i 2 Qa (p )Q (q )Qb (p )Q (q ) + p q Qa (p )Q (q )Qb (q )Q (p ) p a b a b Summarizing, we an write the sum of squares orresponding to rossorrelations terms of the ISL as 2. Xab Integrated Sidelobe Level Problem 20 X1 1 = 2N jQ ( )Q ( )j2 + N164 ( + + + Æ) =0 N a j b j j where the quantities , , , Æ are dened in (2.7), (2.8), (2.9), (2.10). With the method presented above in onjuntion with the method presented in [19℄, we an have an analytial expression for the ISL for any set of real sequenes of odd length. The omputation of the above equations seems to be hard at a rst look, but in a number of ases, in partiular for sequenes from differene sets [24℄ may lead to signiant results. In the following, we use this method to evaluate the asymptoti trend of the ISL of a set of sequenes made up by different Rotations of a Legendre Sequene (RLS set) when N grows to innity. 3. INTEGRATED SIDELOBE LEVEL OF SETS OF ROTATED LEGENDRE SEQUENCES 3.1 Legendre Sequenes The Legendre Sequene (LS) `0 ; : : : ; `N 1 exists for any prime N and is dened as `0 = 1 `j = 8 < 1 : (mod N ) 1 if j is a nonsquare (mod N ) if j is a square A LS may be ylially rotated ta positions to the left to obtain a Rotated Legendre Sequene (RLS) aj dened as aj = `j +ta (mod N ) = `j +faN (mod N ) with fa = ta =N 2 [0; 1℄. The asymptoti value of Xaa for the family of RLS was alulated in [18℄ and [19℄ 1 noting that the asymptoti value of the modulus of the generating funtion of the LS (jQ` (j )j) is independent of j , yielding The rst ontribution relies on a Postulate of Mathematial Ergodiity to arrive at a result whih is formally proved by the seond. 1 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes Xaa N2 2 3 4 f a 1 + 8 f 2 1 2 2 a 22 (3.1) We follow the same path as in [19℄ but for the alulation of the ross-orrelations terms of the ISL Xab [26℄. To proeed, remember that the generating funtion of the LS is p 1 + ` N if j 6= 0 and N = 1 (mod 4) p (3.2) Q ( ) = 1 + i` N if j 6= 0 and N = 3 (mod 3) > > > :1 if j = 0 Moreover, if we denote by Q ( ) the generating funtion of the RLS a = 8 > > > < ` j j j a `j +ta (mod N ) , then j j Qa (j ) = j ta Q` (j ) Assume now that the two sequenes aj and bj are obtained by rotating `j by, respetively, ta and tb positions to the left. We may ompute S 0 as S0 = X1 N j =0 j ta Q` (j )tjb Q` (j )2 = X1 N j =0 jQ ( )j4 ` j from (3.2) we know immediately that jQ` (j )j4 N 2 , then S 0 N 3 . Let us now ompute the asymptoti values of , , and Æ in (2.7), (2.8), (2.9), (2.10) for any pair of RLS. For in (2.7) we have 1 1 2 1 4 2 = N + N S0 N 7 16 3 3 48 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes For in (2.8) ( 1 X1 = N2 8 =0 N p;q p q 6= q + p p (q p )2 h 2p jQ` (p )j2 tpb q Q` (p )Q` (q )+ p q jQ` (p )j2 tqb p Q` (q )Q` (p )+ 2 ta Q` (p )Q (q ) jQ` (p )j2 + p p q ` q p tqa p Q` (q )Q` (p ) jQ` (p )j2 ( X1 1 2 8N =0 N p;q p q 6= q + p p (q p )2 N 2 `p`q 2p tpb q ) i + N 2` ` b + t p q p q q p N 2 `p `q 2p tpa q + N 2 `p `q p q tqa p ( X1 1 4 = 8N =0 N p;q p q 6= (1 `p `q p q )2 ) tpb +1q + ptb +2q + 1p tb q + b+ t p q 1 +2 + tpa +1 q + p ta p q ta q + ta p q ) 23 3. = 18 N 4 Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes X1 N k = +1 k 6=0 (X (k) + X (N `` `` 24 k)) N tkb +1 + tkb +2 + 1k tb + b + a+2 + a+1 + 1 a + (1 )2 k t t k t k k t k ta k Note that X``(k) + X`` (N k) is the periodi orrelation [24℄ of the LS. Then, from [18℄ and [27℄ we know that jX`` (k) + X`` (N k)j 3 for Legendre P 1 1 = O(N 2) (see (3.3) and (3.6) sequenes. Then, using the fat that N k =1 j1 k j2 below and set t = 0), and using the triangle inequality we get that = O(N 6 ). For the alulation of in (2.9), following the same steps we did for we have 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 1 N4 4 1 N X ( (1 =0 p;q;r p q r 6= 6= p `q `r q )(1 X1 N r ) p r tqb r + ta +1 tb +1 p q q tbr + tpb +1r tpa +1 q + p r p q + tb tb ta +1 ta p tbr tpa +1 q + p r p q + p q p r + tb ta p r p q = 14 N 4 p ( = +1 6= 6=0 u;v N u v X``(v + ta p r q r + u) + X`` (N (v (1 v )(1 u) ta p q q r u)) u ta v tb + u tua v + v u tav For Æ in (2.10) we have = O(N 6) ) u tub v u tbv + tub +1 tva +1 + tua +1 tvb +1 + u tb tva +1 + tua +1 v tb + v ta u tb + and again we have that 25 ) v + 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 26 ( h X1 1 1 2 Æ= N jQ ( )Q ( )j2 + 2 2 =0 ( ) N p p;q p q 6= X1 1 4 2N =0 N ( q p;q p q 6= X1 = N4 N k = X1 N k ` p ` q 2p tpa q Q` (p )Q` (q )tpb q Q` (p )Q` (q )+ ta p q p q = 12 N 4 p q q =1 k N p + (1 ta tb q p q + 1 )2 ta q p tb k ta k + 1 (1 )2 k + k ta tb k k +tb ) p + 1 (1 )2 k + k ta +1 6=0 ) i Q` (p )Q` (q )tqb p Q` (q )Q` (p ) ta +tb +tb (N jkj) (N jkj) Larger values of the summand are those for k lose to 1, whih make the denominator lose to zero and numerator N for some onstant (for k lose to N 1, the denominator beomes also lose to zero but the numerator is O(1)). Exploiting this and using the small angle approximation for the omplex exponential, we may write Æ N 1 X k + ta tb + 1 ta +tb 5 k k N 422 k2 N k =1 (3.3) To ontinue, we reall the denition of the Dilogarithm funtion and its series expansion [28℄ valid for jz j 1 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes Z Li2 (z) = 0 1 ln(1 zt) t dt = 1 k X z =1 k k 27 (3.4) 2 Taking the real part of (3.4) and evaluating on the unit irle gives [28, eq. (8.7)℄ Re Li2 i e = Re ( 1 ik X e k ) k2 =1 = 16 2 14 jj (2 jj) Exploiting (3.5) and onentrating on the rst period 0 Re where [℄1 = (1 X ) k2 k =1 t k = 2 1 6 t N 1 1 t N (3.5) 1 we obtain t N 1 (3.6) (mod 1). Hene, sine we know that Æ is real ( 1 1 1 t +t t +t 7 Æ N 1 + 4 6+6 N 1 N 1 ) 1 t t 1 t t 6 N 1 N 1 n = 14 N 2 21 [ f f ℄1 (1 [ f f ℄1 ) o [f f ℄1 (1 [f f ℄1 ) n = 14 N 2 21 [f + f ℄1 (1 [f + f ℄1) o [f f ℄1 (1 [f f ℄1 ) a b a b a b a b where we have dened fa = b a a b a b a ta N a b b a a b a b b and fb = tb N . Then, exploiting the symmetries of a 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 28 quadrati form of a modulus funtion we have for 0 fa ; fb 1 [f + f ℄1 (1 [f + f ℄1 ) = 14 [f f ℄1 (1 [f f ℄1 ) = 41 so that a b a b a b a b 2 1 jf + f 1j 2 2 1 jf f j 2 a b a b " 2 1 1 7 Æ N jf + f 1j 2 + jf 4 a b 2 1 fj 2 a # b Based on the above we are now interested in omputing the asymptoti value of 1 N2 Xab 1 1 16 0 00 3 = 2N 3 (S + S ) 2N 3 N + N 4 ( + + + Æ) 2 2 2 1 1 3 + 2 jf + f 1j 2 + 2 jf f j 2 a b a b (3.7) Going bak to our original problem for alulation of the ISL value of a set of M sequenes x(jp) with j = 0; : : : ; N 1 and p = 0; : : : ; M 1, where eah x(p) is made by a different rotation fp of a LS (RLS set), replaing (3.1) and (3.7) into (2.4) we nally have that ISL X1 2 4 f N2 =0 3 M p p X1 M X1 M =0 p 1 + 8 f 2 p 2 + 2 jf + f =0 3 p q p q 6= q 1 2 + 2 2 1 1j 2 + 2 jf p 2 1 fj 2 q (3.8) 3. Fig. 3.1: Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes Plots of ISL for iso-ISL lines M = 2 as a funtion of 29 1 and f2 : top: 3D-view, bottom: f As an example, Figure 3.1 reports the 3D and ontour plot of the right-hand side of (3.8) for M = 2. Diret visual inspetion of that Figure onrms that minima exists and an be easily identied. In the next setion we will exploit this result where an optimization proedure is developed to nd the optimal rotations that minimize the ISL for any sequenes length N . 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 30 Rotations minimizing asymptotic ISL 15.5 RLS Asymptotic value 15 ISL / N 2 14.5 14 13.5 13 12.5 12 0 Fig. 3.2: 200 400 600 N 800 1000 1200 Integrated Sidelobe Level as a funtion of sequene length. In blue, RLS with rotations minimizing asymptoti ISL; in blak, asymptoti value. Arbitrary Rotation 15.5 RLS Asymptotic value 15 ISL / N 2 14.5 14 13.5 13 12.5 12 0 Fig. 3.3: 200 400 600 N 800 1000 1200 Integrated Sidelobe Level as a funtion of sequene length. In blue, RLS with an arbitrary rotation; in blak, asymptoti value. 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 31 Comparison of an arbitrary and minimizing rotation 15.5 Rotation with minimum asymptotic ISL Arbitrary rotation 15 ISL / N 2 14.5 14 13.5 13 12.5 12 0 Fig. 3.4: 200 400 600 N 800 1000 1200 Integrated Sidelobe Level as a funtion of sequene length. Comparison of an RLS with an arbitrary rotation and RLS with rotations minimizing asymptoti ISL. As another example, in Figure 3.2, 3.3, 3.4 we plot the ISL for M funtion of the sequene length N . = 4 as a In Figure 3.2 the values of rotation are those that minimize the asymptoti ISL, while in ase Figure 3.3 we use an arbitrary rotation. In both ases we an see that the trend of the plots is in agreement with the asymptoti value alulated. In Figure 3.4, we plot together both urves to show that the one that ahieves the minimum asymptoti value of ISL, also ahieves the minimum ISL value for sequenes length greater than approximately 20. For different hoies of rotations and different number of sequenes (M ), the behavior is the same than presented. 3.2 Sets of RLS minimizing ISL The key idea [29℄ is to set x(jp) 0; : : : ; M 1. = `(jfp ) for properly hosen rotations fp, p = 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 32 Sine only N values for eah fp lead to distint rotations, a omplete san N requires only M trials that, though far from the exponential explosion of a full san (that would entail 2M N trials) may soon beome prohibitive. To ope with larger values of N we may resort to the asymptoti analysis made in the previous setion. From equation (3.8) we see that asymptoti ISL is invariant if we hange fp into 1 fp for any p. Therefore, by assuming f0 f1 fM 1 1=2 one may resolve all absolute values and easily ompute the rotation values for whih ISL = 0. This yields fp 2p + 1 (3.9) 4M that result in a minimum attainable ISL = N 2 M (M 1) + 16 and thus, in a fp = performane gure RLS = 6M (M 1) 6M (M 1) + 1 (3.10) indiating that, for large N , the performane of a set of RLS should be within 8% of the maximum possible, approahing it very rapidly as M inreases. Based on these asymptoti onsiderations it is easy to devise a muh faster san that drastially redues the number of trials by onsidering for the j -th ro+1 . Sine the length of tation only a narrow interval of possible values around 24pM suh an interval may be dereased as N inreases, the resulting searh burden goes N from M trials to (N )M with (N ) a funtion rapidly approahing a onstant as N inreases (experimentally we veried (N ) ' 20 for N larger than 200). The results of suh a san yields the Optimum RLS set (ORLS) whose performane is ompared with that of other known algorithms or sequene families in the following Setion. 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 3.3 33 Numerial results Beyond ORLSs that exist for every prime N , we onsider Random sequenes, that exist for any N and are generated by assigning xj(p) = 1 with uniform probability and independently for eah p = 0; : : : ; M 1 and j = 0; : : : ; N 1. For eah N and M we generate 104 sets of sequenes and reord the best ahieved performane. Gold sequenes, that exist when N = 2q 1 for some integer q and are obtained from the well known maximumlength sequenes to maintain low orrelation and simultaneously be able to produe sets of sequenes with relatively large ardinality. Though they are produed by linear-feedbak shift registers, Gold sequenes are designed to enjoy the same properties of random sequenes. For eah N we draw 103 NM M-tuples of Gold sequenes at random from those available, and we reord the least ISL. Q.CAN sequenes, that exists for any N and are obtained by the CAN algorithm desribed in [3℄ when quantization is applied at the end of the iterative proedure. For the ase of 1-bit quantization the option of leaving the algorithm operate with ontinuous phases and quantize only the nal result, has been disarded , after experimentally verifying that it was leading to poorer performane. In this ase, quantization was applied at every step of the iterative proedure. Optimally Rotated Best Sequenes (ORBS) that leverage on the fat that for eah N up to 304 one or more sequenes are reognized as the stateof-the art solution to ISL minimization problem for M = 1 (some of them are known to be the true optimum solutions, some others are only the best known solutions). For eah of those sequenes, we build a set of M sequenes by trying all the possible relative rotations and seleting the set of rotations yielding the minimum ISL. 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 34 In gure 3.5 we evaluate how state-of-the-art algorithm for the synthesis of ultra low-ISL sequenes is affeted when quantization is imposed. Note how the performane of Q.CAN is hardly impaired for low quantization depth, and reah the performane of ORLS when the number of quantization levels is greater than 13 for M = 4, 18 for M = 4 and 22 for M = 12 for the onsidered ases. For different hoies of M and N we have seen a similar trend, and that the performane of Q.CAN is only better than that of ORLS when the number of quantization bits of grater than 4, with an inreasing trend as M inreases. This shows a great advantage in the use of ORLS. 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 35 M=4 0.98 0.96 0.94 0.92 ε ORLS Q.CAN 0.9 0.88 0.86 0.84 0.82 5 10 15 20 25 Number of qantization steps (Q) 30 M=8 1 0.98 0.96 ε ORLS Q.CAN 0.94 0.92 5 10 15 20 25 Number of qantization steps (Q) 30 M=12 1 0.99 ε 0.98 ORLS Q.CAN 0.97 0.96 0.95 0.94 Fig. 3.5: 5 10 15 20 25 Number of qantization steps (Q) 30 Comparison between ORLS, with Q.CAN algorithm when quantization is imposed for (a) M = 4, (b) M = 8 and () M = 12, and x value of N = 1033. The performane metri as a funtion of the number of quantization steps. The performane of Q.CAN exeeds the performane of the binary ORLS for Q grater than 13 for M = 4, 18 for M = 4 and 22 for M = 12. 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 36 M=2 0.95 0.9 0.85 ORLS Q.CAN Random Gold S. ε 0.8 0.75 0.7 0.65 0.6 0.55 0 500 1000 N 1500 2000 M=3 1 0.95 0.9 ORLS Q.CAN Random Gold S. ε 0.85 0.8 0.75 0.7 0.65 0 500 1000 N 1500 2000 M=4 1 0.95 ORLS Q.CAN Random Gold S. ε 0.9 0.85 0.8 0.75 0 Fig. 3.6: 500 1000 N 1500 2000 Comparison between ORLS, random sequenes, Gold sequenes and Q.CAN sequenes with binary quantization for (a) M = 2, (b) M = 3 and () M = 4. The solid horizontal line at 1 identies the theoretial maximum performane while the dashed horizontal line marks the asymptoti performane ahieved by RLS. 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 37 Figure 3.6 ompares the performane of ORLS with that of random, Gold and Q.CAN sequenes with binary quantization for M = 2, M = 3 and M = 4. Note how ORLS learly outperform the other tehniques for all reasonably large N (say for N > 100) also revealing a distint improving trends approahing the theoretial limit as N inreases. On the ontrary the performane of random, Gold, and Q.CAN sequenes exhibits a lear dereasing trend. Aording to expetations, sine Gold sequenes are designed to mimi a random behavior, the orresponding performanes follow an analogous trend. Finally, though insufient to reah ORLS, the optimization impliit in the onstrution of Q.CAN sequenes make the orresponding performane learly superior to that of random-like sequenes. In Figure 3.7 we ompare eah ORLS with the orresponding ORBS and with Q.CAN sequenes with binary quantization for M = 2, M = 3 and M = 4. Again, ORLS perform uniformly better than ORBS for sufiently large N ; additionally ORBS do not exhibit a denite improvement with respet to Q.CAN at least for M > 2. This shows that the good performane of the proposed ORLS is only partially due to the exploitation of sequenes that feature a good autoorrelation properties but also hinges on a strutural property of Legendre Sequenes. 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 38 M=2 0.95 0.9 0.85 ε 0.8 0.75 0.7 0.65 ORLS Q.CAN ORBS 0.6 0.55 0 50 100 150 N 200 250 300 M=3 1 0.95 0.9 ε 0.85 0.8 0.75 ORLS Q.CAN ORBS 0.7 0.65 0 50 100 150 N 200 250 300 M=4 1 0.95 ε 0.9 0.85 0.8 0.75 0 Fig. 3.7: ORLS Q.CAN ORBS 50 100 150 N 200 250 300 Comparison between ORLS, ORBS and Q.CAN sequenes for (a) M = 2, (b) M = 3 and () M = 4. The solid horizontal line at 1 identies the theoretial maximum performane while the dashed horizontal line marks the asymptoti performane ahieved by RLS. 3. Integrated Sidelobe Level of Sets of Rotated Legendre Sequenes 3.4 39 Conlusion We apply a method based on generating funtions, whih has already been proposed for the alulation of the ISL of a sequene, to the alulation of the rossorrelation omponents of the ISL of a set of sequenes. The apparent omplexity of the resulting expressions an be takled in the asymptoti onditions for sequenes whose generating funtion has a relatively simple trend. Sine this is the ase of Legendre sequenes, we are able to derive an analytial expression for the asymptoti ISL of sets of rotated Legendre sequenes. Based on the later result, we propose a simple proedure to onstrut sets of antipodal sequenes with extremely low ISL. Eah sequene in the set is a different rotation of the Legendre Sequene of the same length. Optimal rotations are found by an exhaustive san whose omplexity is greatly redued by exploiting the asymptoti result yielding a general expression for the trend of the ISL of sets of innitely long sequenes. The resulting performane largely exeeds that of lassial methods for the diret generation of low-ISL sets of antipodal sequenes. The method we propose also outperforms a well-known algorithm able to generate extremely-low ISL sets of unimodular ontinuous-phase sequenes, whih is nevertheless impaired by the strong quantization needed to satisfy antipodality onstraint. Part II ANALOG TO INFORMATION CONVERSION 4. SIGNAL MODELS 4.1 Introdution The lassial aquisition approah based on the Nyquist-Shannon theorem states that for any analog band-limited signal, all its information ontent an be aquired by taking uniformed distributed samples at a rate that doubles the signal bandwidth. While this is one of the fundamentals theorems of Signal Proessing, by taking advantage on ertain strutures of the signal, a muh lever aquisition strategy an be develop in order to redue the number of measurements and still aquire its full information ontent. In order to exploit the peuliarities of a given lass of signal, we must be able to properly represent those signals of interest with aurate models. This models are useful to inorporate previous knowledge of a given lass of signal, and to distinguish them from other lasses of maybe no interest. Many lasses of signals, espeially when representing physial signals, an be modeled to have a linear struture, i.e., if we sum two signals that belongs to that lass, the new signal will also belong to the same lass. We will treat signals as real-valued funtions having domains that are either ontinuous or disrete. In the ase of a disrete signals, we an simply view them as vetors in N -dimensional Eulidean spae R N . For bandlimited analog signals with no frequeny omponents above N=2, or Nyquist rate equal to N , we will also represent them as vetors with dimension 4. Signal Models 44 equal to its Nyquist rate. Both representation are equivalent in the sense that one an pass to another with standard tehniques (sin interpolation). Note that the spae dimension N of both kinds of signals desribed above denes the degrees of freedom they have. In partiular, although analog signals an be more efiently represented by other representations, any analog bandlimited signal has at most N degrees of freedom, and we have hoose this model in order to be able to diretly ompared with Nyquist-based aquisition. Let denotes the N N matrix with olumns given by the set f i gN . If i=1 N the vetors in this set are linear independent, then they span a basis in R , and any vetor in this spaes has a unique representation as a linear ombination of the elements of that basis. For any x 2 R N there exist s 2 R N suh that x = s = N X =1 si i i For analog signals, note that this representation is equivalent to: x(t) = N X =1 si (t)i i , where the set of ontinuous time waveforms f (t)i gN i=1 are the sin-interpolated N signals obtains from the vetors in f i gi=1 (or equivalently, the vetor in f i gN i=1 N are form by taking samples of the waveform in f (t)i gi=1 at a rate N ). 4.2 Sparse Signals With the models given above, we are able to represent any linear signal (disrete or analog) of dimensionality equal to N , and with N degrees of freedom. However, many natural signal that are found in real situations have a smaller number of 4. Signal Models 45 degrees of freedom with respet to its dimensionality. In other words, not all possible vetors in R N represents valid signals for a given lass. Many natural signals an be expressed as a linear ombination of only just a few vetors from a given basis. This lass of signals are alled to be sparse signals, sine only a small amount of its oefients, when represented on that basis, are different from zero. The information ontent of this lass of signals is onentrated only on the values of the non-zero omponents and on the position of those omponents. For an N dimensional vetor a = (a0 ; : : : ; an 1 )> we dene the support of a as supp (a) = fj = 0; : : : ; n 1ja 6= 0g j , its sparsity spar (a) (sometimes indiated as L0 norm) as the ardinality of supp (a), and its usual p-norm as kak = n 1 X p j =0 ja j !1=p j p We will assume that a suitable basis exists whose vetors are the olumns of the N N matrix , and that the signal of interest is K -sparse, whih means that for any instane of x there is an N -dimensional vetor s suh that x = s and spar (s) K . Although the sparse model given above is of broad interest, it is difult to nd real life signals to be truly sparse. However, many natural signals an be very well approximated by sparse models. This lasses of signals are alled to be ompressible signals, and an be approximated by setting the smallest omponents to zero and keeping the biggest K . In the following, we will treat ompressible signals and sparse signals as to have a simple sparse representation. The error produed by this approximation 4. Signal Models 46 will be onsidered as if the sparse signal would have an intrinsi noise independently of the soure where it is generated. 5. COMPRESSIVE SENSING 5.1 Introdution The newly introdued paradigm of Compressive Sensing (CS ) [7, 8, 9℄ exploits speial signal features to extrat its information ontent with a smaller amount of samples (or measurements in the general ase) with respet to aquisition based on the Nyquist-Shannon sampling theorem. Aording to the sampling theorem, we an perfetly reonstrut any bandlimited signal by its samples provided that the sampling rate exeeds twie the maximum frequeny in the bandlimited signal. However, as we have seen before, the information ontent of some lasses of signals is onentrated in only few oefients for a given representation. Taking advantage on the knowledge of the struture of the signal, more sophistiated sampling methods an be developed in order to redue the number of samples neessary to reonstrut the signal. Compressive sensing theory exploits the sparsity representation in order to redue well below the number of measurements stated by the Nyquist-Shannon theorem, and still be able to perfetly reonstrut the original signal. Reduing the number of measurements has noteworthy advantages. It an redue the hardware omplexity, storage apaity, power onsumption, hannel bandwidth, et. In the ompressive sensing framework, few nonadaptive linear measurements of the signal are taken, i.e. projetions of the signal over vetors of a given basis. 5. Compressive Sensing 48 Based on these projetions, by means of a non-linear algorithm, it is possible to reover the signal. To make the disussion more onrete, onsider the general ase where the signal x 2 R N is measured through M inner produts of the form: y = x + e where y 2 R M is the measured vetor, is an M N measurement matrix, and e 2 R M is a vetor representing measurement noise. In general, given M < N , the matrix represents a dimensionality redution, i.e., it maps a vetor in R N into a vetor in R M . Under this ondition, there are innite different signals x that satisfy the above equation given the measurements y. At this point there are two main questions to be done: a) Under what onditions the appliation of the matrix preserves the information of the signal x? How it is reovered the original signal x from the redued set of measurements y? We will try to answer these question in the following setions. 5.2 The Restrited Isometry Property To partially answer the rst question, lets rst write the vetor x as x = s , and y = s + e = s + e Relaying on the a-priori knowledge that spar (s) K , it is possible to dene 5. Compressive Sensing 49 a subset of R N ontaining all the interesting instanes of x. Then, the aquisition mehanism should map this subset into the measurement spae R M quasibijetively in a sense that will be made more preise in the following. One of the most striking, and useful, fats that appear at this point is that, when sparsity is one of the priors, if an be thought of as a realization of a random matrix with independent entries drawn aording to a variety of distributions, then mapping by means of provides, with high probability, the needed quasi-bijetion. More formally, we say that a matrix is a restrited isometry [30℄ when there is a onstant 0 ÆK < 1 suh that (1 ÆK ) ksk22 ksk22 (1 + Æk ) ksk22 whenever spar (s) K . Hene, even if the dimensionality M of the o-domain of a restrited isometry is less than the dimensionality N of its domain, the mapping of K -sparse vetors leaves lengths substantially unaltered. If is made of independent random entries haraterized by a sub-Gaussian distribution then, with an overwhelming probability, the matrix is a restrited isometry with a onstant Æ provided that [31, 32, 33℄ M CK log(N=K ) (5.1) where C is some onstant depending on eah instane. If is a restrited isometry, one that supp (s) is known, we may restrit to that domain and obtain an injetive mapping. If the measurements in y addition N ally enode information on whih of the K possible supports must be hosen, the overall mapping an be reversed to yield the whole s. This is why a onstant ingredient in the reipes for all ompressive sensing arhitetures is randomness as a mean of apturing information that is known to 5. Compressive Sensing 50 be sparse. What is usually done is to overlook the fat that theory puts onditions on the statistial struture of and design a system in whih is random and hopefully transfers its beneial properties to = . An important side-effet of this assumption (widely veried in pratie) is that one does not design the aquisition matrix depending on the spei but relies on randomness to impliitly san all possible sparsity bases. 5.3 CS Reonstrution Algorithms One that a mapping allowing reonstrution has been devised, its inversion must be obtained by algorithmi means every time a measurement vetor omes in. Though reonstrution mehanisms may be designed jointly with the arhitetures produing the measurements, they are lassially addressed as separate omponents of the overall aquisition system. Their development and analysis is a ourishing eld that has reently produed strong and general results and taxonomies [34℄. We will here onentrate on the most frequently adopted methods, and note that those tehniques fall in one of two ategories: optimization-based reonstrution [30, 35, 36, 37, 38, 39℄ and iterative support-guessing reonstrution [40, 41, 42, 43, 44, 45℄. Both types of tehnique are ommonly devised and set up in the noiseless and idealized ase (i.e., for e = 0 and neither quantization nor saturation) and are proved (or simply seen) to work in more realisti settings. 5.3.1 Optimization Based Reonstrution Algorithms The key fat behind optimization-based methods is that, among all the possible ounterimages s of the vetor y = s the one that we are looking for is the most 5. Compressive Sensing 51 sparse, i.e., the one for whih spar (s) is minimum. Sine we usually have spar (s) K N this assumption is sensible. Moreover, it leads to some beautiful results on the possibility of reovering s by means of simple optimization problems [30℄. More formally, it an be shown that, if is a restrited isometry with onstant p Æ 2 1 then the ^s solution of the optimization problem min k^sk1 s:t: k^s is suh that for some onstant C > 0. k^s sk2 y k2 (5.2) C Hene, if we use to bound the maximum magnitude of the disturbanes involved in the measurement proess (for instane by setting it proportional to the variane of the noise plus that of the quantization error) we an guarantee that the reonstrution error vanishes when disturbanes go to zero. Though not impossible, the straightforward appliation of the above result, depends on a reliable estimation of the parameter that quanties the maximum foreseeable deviation between the unperturbed measurement and its atual value in presene of a mixture of known (e.g., quantization) and unknown (e.g., noise) disturbanes. It is therefore quite ommon to substitute k^s yk2 with ^s = y by impliitly assuming that the system is working in a relative low-disturbane regime that allows to assume ' 0. Within this approximation, it is onvenient to reexpress the resulting optimization problem within the framework of linear programming by dening u = (1; : : : ; 1)> and by introduing the auxiliary unknown 5. Compressive Sensing 52 vetor w = (w0 ; : : : ; wn 1)> to write min s:t: u> w s y w wsw ^ = 0 ^ (5.3) where vetor inequalities are meant to hold omponent-wise. The equality onstraints in (5.3) an be adjusted to ope with spei features of a given arhiteture or to take into aount quantization or saturation. In partiular, due to quantization, we know that the true value of the j -th measurement is somewhere in the interval [yj yj=2; yj + yj=2℄ with yj being the value known to the algorithm and yj the orresponding quantization step. Hene, in presene of a oarse quantization, it is sensible to substitute the equality onstraints ^s = y in (5.3) with y y=2 ^s y + y=2, where y = (y0; : : : ; ym 1)>. Though it surely models the aquisition proedure with greater auray, this adjustment does not neessarily lead to improvements and is ommonly employed only when one may expet the various yj to be substantially different one from the other. It is interesting to note that optimization-based reonstrution algorithms work without any knowledge of the exat value of K further to that impliit in the number of measurements that must be enough to allow reonstrution. This may be a plus in situations where K annot be exatly determined in advane. Regrettably, this positive feature is balaned by the fat that, in general, linear programming solution is omputationally more expensive that other kinds of iterative reonstrution. 5. 5.3.2 Compressive Sensing 53 Support-Guessing Reonstrution Algorithms As far an iterative support-guessing reonstrution is onerned, note that, if supp (s) were known we ould drop the olumns in that are surely multiplied by 0 and the orresponding entries in s to obtain an M K matrix supp(s) and a K -dimensional vetor ssupp(s) for whih y = supp(s) ssupp(s) . Sine M > K , this is an overonstrained problem that may be effetively (even optimally in ase of Gaussian disturbanes) inverted by using the Moore-Penrose pseudo-inverse ysupp(s) and omputing ssupp(s) = ysupp(s)y. Iterative support-guessing methods are, in general, proedures that alternate a rough, non-neessarily sparse, solution of y = s from whih an estimate of supp (s) is inferred (for example by thresholding on the magnitudes of the omponents of the temporary solution) that is then exploited in a pseudo-inverse-based step rening the value. Though more sophistiated alternatives exists, a referene algorithm within this lass is CoSaMP [40℄ that has some denite advantages. First, it works for matries that are restrited isometries and, if K is known and the isometry onstant Æ2K for vetors with 2K non-zero omponents an be bounded by Æ2K 0:025, then, given a tolerane > 0, the reonstruted vetor ^ satises k s0 k2 k^s sk2 C max ; p + k k2 K where s0 is the vetor that an be obtained by s by setting to zero its K=2 largest entries. The resulting algorithm is provably fast and, beyond the above formal guarantee on its performane, it is usually extremely stable and effetive in reovering the original signal. These favorable properties are paid with the additional assumption that the sparsity of s is known and that the isometry onstant Æ2K must be quite low. In analogy to what happens for optimization-based reonstrution, CoSaMP 5. Compressive Sensing 54 an be tailored to spei arhitetures. This an be done, for example, if it is known that errors in the magnitudes of the entries of s are orrelated by an impliit ltering in the aquisition sheme. Suh an effet an be exploited by inserting a ltering step when passing from support-guessing to pseudo-inversion. 5.4 Analog-to-Information Converters From the two previous setions, we get that to dene a ompressive sensing system we need to desribe two stages enoder: a hardware system performing some mixed analog-digital opera- deoder: an algorithm that takes the inoming bits and, based on the knowl- tions on the inoming signal to produe a stream of bits. The mixed analogdigital operations are modeled as instane of a random matrix linking the signal samples to the measurements whose quantization yields the stream of bits transferred from the enoder to the deoder; edge of , reonstruts the original signal. In this setion we will disuss various strategies for designing systems for aquiring ompressive measurements of real-world signals. Note that, in pratial implementation, we do not want to ommuniate to the deoder and thus most often exploit pseudo-random generators with a ommon initialization to yield matries that an be simultaneously known at both stages. Saturation and quantization are unavoidable in the signal path sine the ommuniation between the two stages happens along a digital hannel thus implying an ADC blok with a nite range (we will assume [ V max ; V max℄ for a ertain V max ) and a nite number of levels. In the following we will onsider the number B of bits generated by the enoder orresponding to the aquisition of the input signal over a given time interval. This is atually a bit budget sine it may be partitioned into digital words of 5. Compressive Sensing 55 different depths orresponding to different measurements. Additionally, in many appliations the total number of bits is onstrained, whih suggests a tradeoff between the number of measurements and the number of bits per measurement. 5.4.1 Random-Modulation-Pre-Integration RMPI This is probably the most straightforward implementation of ompressive sensing onepts [46℄. clock N RNG + N-1 S k=0 b-bit ADC y0 b RNG N-1 xk Fig. 5.1: + x(t) S k=0 b-bit ADC yM-1 b Blok sheme of an RMPI enoder. The samples of the input signal are multiplied by M different random sequenes and aumulated up to time N . The aumulated values are then quantized by a b bit AD onverter. With referene to Figure 5.1 the samples of the inoming signal xk are multiplied by the quantities j;k for a given j and then fed into an aumulation stage to yields the value of the j -th measurement yj that is then quantized by an b-bit ADC and aggregated with all the other quantized measurements into the stream of bits that is passed to the deoding stage. 5. Compressive Sensing 56 The implementation of the analog bloks preeding the ADC offers several options. The struture of the multiplier depends on the quantities j;k : some lassial approahes adopt Gaussian random variables (Gaussian RMPI) and fore the deployment of omplete four-quadrant analog multipliers, while more aggressive approahes suggest to onstrain j;k 2 f 1; +1g (antipodal RMPI) so that multipliation an be implemented by simple swithing. The aumulation stage may be implemented either as a ontinuous time integrator or as a swithed apaitor subiruit that impliitly mathes the disretetime operation of the multiplier. In any ase, the output of the aumulating devie will be subjet to saturation. Referring to a disrete-time implementation, where allegedly yj and relying on the following assumptions: x and = P =01 N k j;k xk , are independent stohasti proesses; the j;k are independent and identially distributed (either Gaussian or binary antipodal) random variables, with zero mean and unity variane; the energy of x in the aumulation time window is onstant; the random variable yj will onverge to a normal random variable independently of the input signal x. Given the above observation, the measurements y obtained with an RMPI arp hiteture will have a range that is potentially N -times larger than that of x (e.g., 3 around the signal average). When omparing an RMPI solution with a diret appliation of a Nyquist based AD onverter, and onsidering a uniform quantizer in both ases, in order to maintain the same amount of quantization error the number of bits needs to be inreased for an RMPI implementation. Moreover, sine a normal distribution is not limited, wherever the input range of the ADC is set, there is an unavoidable non-zero probability that yj falls out of the ADC onversion range. 5. Compressive Sensing 57 On one hand, RMPI arhiteture allows to redue the number of measurements for the aquisition of a given lass of signals with respet to lassial Nyquist based sampling. On the other hand, in order to obtain a given performane in terms of reonstrution error, the number of bits needed to enode eah of the measurement would be bigger than for Nyquist based aquisition. This suggests a tradeoff between the number of measurements M and the number of bits per measurement b. RMPI arhiteture presents a diret implementation of the ompressive sensing onepts developed in this setion. However, as it has been shown, some design onsideration are needed to be taken. More preisely, the hoie of a proper AD onverter is of ruial importane in order to obtain a given performane. Moreover, RMPI arhiteture requires the used of a huge amount of iruitry (ontinuous-time or disrete-time analog multiply-and-aumulate bloks, multibit AD onverters) leading to an expensive system implementation in terms of ost, power onsumption, and design effort. 5.4.2 Random Sampling RSAM In lassial aquisition systems, samples of the signal are taken regularly on the time axis at a given rate (usually not less than the Nyquist rate). Compressive sensing arhitetures relying on random sampling avoid this regularity to produe a number of measurements that, on the average, are less than those produed by Nyquist sampling, while still allowing the reonstrution of the whole signal thanks to sparsity and other priors. 5. Compressive Sensing 58 Dkj RNG clock DkDk b-bit x(t) Fig. 5.2: xkj ADC b yj Blok sheme of an random sampling enoder. The samples are taken at random positions in time, over a predened grid. In priniple, sampling instants an happen anywhere along the time axis. Yet, a straightforward implementation hooses them among regularly spaed time points that an be seleted by digital means. The result is shematized in Figure 5.2 where a bakward ounter is pseudo-randomly re-loaded eah time it reahes zeros, triggering onversion. Grid spaing, and thus lok rate, depends on the resolution with whih one wants to plae the sampling instants and thus may be expeted to be larger that Nyquist rate. To translate the above blok sheme into formulas , say that suh the lok identies a vetor x0 = (x00 ; : : : ; x0N 1 )> that oversamples a bandlimited x(t) by a fator with respet to x = (x0 : : : ; xN 1 )> ontaining the Nyquist samples. The two vetors x0 and x are linked by x0 = Ax, being A an upsampling matrix. With this, the M N matrix is nothing but the produt = PA, where P is the random sampling matrix dened by the M time instants k0 < k1 < < kM 1 at whih the ounter reahes 0 as in Pj;k 8 < = :1 0 if k = kj otherwise 5. Compressive Sensing 59 The resulting sampling follows a so-alled renewal-proess in whih all the inter-measurement intervals kj = kj +1 kj are drawn as independent integer random variables exponentially distributed in the interval [kmin; 1℄. The minimum inter-measurement gap kmin 1 depends on the speed of the ADC, whih must be ready for a new onversion eah time a measurement is taken so that, by inreasing kmin we loosen the onstraints on the ADC implementation. The exponential trend is then tuned to have an average inter-measurement N so that (at least for large N ) we expet an average of M measuregap equal to M ments. Eah of these measurements is ommonly quantized by means of a b-bit ADC to yield the bit stream passed to the deoder. RSAM is only subjet to the stati saturation due to the nite input range of the onversion stage. This poses no problem sine it an be takled at design time by simply resaling the signal input range as in onventional aquisition systems. 5.4.3 1-bit Compressive Sensing - 1bRMPI Given a total bit budget B , the trade-off between the number of measurements M and the number of bits b = B=M spent to enode eah of them is a lassial theme in signal aquisition and oding and applies also to CS arhitetures. Among other issues, it may help oping with the unavoidable saturation of the ADC sine the extreme solution b = 1 identies the ADC with a pure saturation entered in 0, thus ompletely eliminating the problem. In partiular, RMPI systems may be optimized in eah partiular setting to see how muh information in our original signal an be inserted into B bits [47℄ and is possible to think that eah measurement is represented by a single bit enoding its sign [48, 49℄. There are several benets to the 1-bit CS tehnique. Given that the quantizer an be implemented as a simple omparator that merely tests if a measurement is 5. Compressive Sensing 60 above or below zero, an efient hardware quantizer an be built to operate at high speeds. Furthermore, 1-bit quantizers do not suffer from dynami range issues nor linearity problems inherent of the implementation of a multibit AD onverter. Sine signs give no hint on the magnitude of the involved signals, the problem in (5.2), with = 0, with y = sign(^ ) and where the sign() operator applied omponent-wise, is reast into [48℄ min k^k1 Æ ^ 0 s:t: yk^ k2 = 1 (5.4) where Æ stands for omponent-wise produt1 and the seond, unit-energy onstraint is introdued as a sale-xing prior. This approah is referred in [48℄ as 1-bit CS. The above optimization problem is a non-onvex problem and must be addressed by speialized algorithms. Two state-of-the-art algorithms were presented in the lasts years to address this problem. The Restrited Step Shrinkage [50℄ that will be indiated here as RSS, and the Binary Iterative Hard Thresholding [51℄, indiated here as BIHT. These algorithm are proved to ahieve a higher average reovery SNR, and are an order of magnitude faster than other previous proposed algorithms in [48℄ and [49℄. Regrettably, even with BIHT, typial performane of an 1bRMPI arhiteture are largely inferior with respet to multibit RMPI or RSAM solutions for omparable bit budgets. 1 so that this result in a set of M omponent-wise inequalities 6. 6.1 RADS CONVERTER Introdution and Motivation As it has been shown in the previous setions, there is a trade-off between the number of measurements needed to uniquely identify a given lass of signals, and the number of bits that is neessary to represent them in order to obtain a given preision [47℄. On one hand, under ertain assumptions on the signal struture, ompressive sensing theory allows to redue the number of measurement by inreasing the hardware arhiteture omplexity. On the other hand, Delta-Sigma onverters allows to redued the number of bits per measurement, even to the extreme ase of only 1-bit, by inreasing the number of measurements and mixing time enoding information. The fundamental question is: is it possible to ombine the advantages of both theories in one single devie that allows to redue the total number of bits in a measurement, and simplify the hardware system implementation? The answer to this question is YES, and it is what we have alled The RADS Converter [52, 53, 54℄. In this setion we will introdue the RADS Converter whih onstitutes the main ontribution of this thesis. We will start by desribing its hardware arhiteture, and modeling the operations performed by the arhiteture in the frequeny domain. This model will lead to an intuitive understanding of its working priniple, and will give some insight on how the deoding stage an be efiently 6. RADS Converter 62 implemented. We will next introdue a time domain analysis that, starting by the analysis of a = modulator, and followed by the analysis of the whole RADS Converter arhiteture, will lead to a deeper understanding of apabilities of the system. We will lose this setion by presenting a set of measurements performed on an off-the-shelf implementation of the RADS Converter that onstitutes a proof of onept of the proposed arhiteture, and we disuss how it an be efiently implemented on a single silion devie. In order to evaluate the performane of the onverter, we have extensively appealed to numerial simulations. Performane is evaluated by mathing the reonstruted vetor ^s with the original vetor s and using two merit gures: the Probability of Support Reonstrution (PSR) and the Reonstrution Signal-toNoise Ratio (RSNR), i.e., PSR = Pr supp (s) suppminf g 5 (^s) s = " !# " k xk22 k sk22 ARSNR(dB)=E dB ks ^sk2 =E dB kx x^k2 2 2 !# where the thresholded support is onventionally dened as supp (a) = fj = 0; : : : ; n 1jjja j g j Probabilities and expetations were estimated using Monte Carlo simulations for whih statistis was gather after 5000 trials. 6.1.1 Preliminaries: Delta-Sigma Modulation In this subsetion we will make a short review of the main onepts that applies to Delta-Sigma (=) modulators. 6. RADS Converter 63 Let model a basi 1st -order = modulator struture as in Figure 6.1 where the blok [Q℄ represents a general quantizer and the blok [D℄ represents a one time-step delay. yn + + - vn + Q zn D D Fig. 6.1: A time-domain blok diagram of a rst order = modulator. The input sequene y feeds the = modulator that produes a lower resolution output sequene z at every time step n. The quantization stage of the modulator is usually implemented with a very low resolution quantizer. Single bit quantizers are the most ommon option for the implementation of this kind of onverters, sine it is partiularly appealing for hardware implementations. The quantizer takes the form of a omparator to zero, an extremely inexpensive and fast hardware devie. Furthermore, 1-bit quantizers do not suffer from dynami range issues (the sign of the measurement remains valid even if the quantizer saturates). Though beneial, 1-bit quantization is a very non-linear operation that makes difult to obtain simple models for the operation of the onverter. In order to provide an insight into the operation of the modulator, the analysis is usually takled in the z -domain [55, 56, 57, 58℄, for whih the quantizer has been replaed by it linear model as shown in Figure 6.2. 6. RADS Converter 64 E(z) Y(z) + + - V(z) + + Z(z) + z z Fig. 6.2: A z -domain linear model of a rst order = modulator. From the diagram we an write V (z ) = z 1 V (z ) + X (z ) z 1 Z (z ) Thus Z (z ) = V (z ) + E (z ) = z 1 V (z ) + X (z ) z 1 Z (z ) + E (z ) and rearranging we get Z (z ) = X (z ) + (1 z 1 )E (z ) (6.1) Equation (6.1) an be written in the general form Z (z ) = ST F (z )X (z ) + NT F (z )E (z ) (6.2) where the ST F refers to the Signal Transfer Funtion, that in this ase is unity, and the NT F refers to the Noise Transfer Funtion and is equal to NT F =1 z 1 (6.3) 6. RADS Converter 65 Equation (6.2) is the basi equation = modulator, and shows how the output an be expressed as a sum of a term aounting for the signal, and a term aounting for the quantization noise. For the ase presented above, the NT F has learly a high pass response, whih suppresses the quantization noise near d, and amplies it out of the signal band. This is the so alled noise shaping apabilities of the = modulators. By replaing z in equation (6.3) by ei2f =M , where M is the sampling frequeny, the power spetral density (PSD) of the output noise is found to be Sq (f ) = 2(sin(f=M ))2 Se (f ) , where Se (f ) is the PSD of the quantization noise of the internal quantizer of the onverter. Consider a signal bandwidth of B Hertz, and approximate Se (f ) = 2e2rms =M . By integrating Se (f ) in the signal band, we get that the in-band noise, i.e., the quantization noise present in the signal band, an be approximated as qrms = erms p 3 M B 3=2 (6.4) As it an be seen from equation (6.4), the in band noise dereases with inreasing the oversampling ratio, i.e., the ratio between the sampling frequeny and the signal bandwidth. In oder to inrease resolution, by replaing the quantizer stage in the blok diagram of Figure 6.1 by a new opy 1st -order = modulator, we will get a seond order = modulator. This proedure an be ontinued to obtain an Lth order = modulator. By extending the analysis we have made for the 1st -order, we an get a basi expression for the NT F of an Lth -order = modulator as 6. RADS Converter NT F = (1 66 z 1 )L (6.5) By integrating the above equation in the signal band, we get that the power of quantization noise of an Lth -order = modulator is L M qrms = erms p 2L + 1 B (L+ 12 ) (6.6) The equation given above is an approximation for the alulation of the inband quantization noise of an = modulator. This approximation does not take into onsideration quantizer overload thus inreasing the total power of quantization noise. Moreover, for higher order modulators, it is possible to hange the shape of the NT F to produe different behavior. However, for the sake of onreteness, it is enough the analysis made so far. 6. 6.2 RADS Converter 67 RADS Converter arhiteture The RAndom Delta-Sigma (RADS ) Converter illustrated in Figure 6.3 is nothing but a onventional = onverter whose input signal is pre-multiplied by a random sequene of symbols. RADS Converter exploits the noise shaping apabilities of Delta Sigma (=) strutures and produe a number of measurements (M N ) eah oarsely quantized (atually with only 1 bit). The use of RADS Converter with a proper exploitation of sparsity gives as a result a substantial ompression in the number of aquired bits with respet to lassial aquisition or to simply = modulation. The simpliity of the arhiteture also allows to operate at very high frequenies, making possible, for example, to aquire frequeny sparse signals that are spread over a large bandwidth with a very high resolution. clock M N pn RNG Fig. 6.3: xn + t=n/M x(t) yn 1-bit DS zn Blok sheme of a RADS Converter. The input signal is multiplied by a random sequene and fed into a = onverter made of a 1-bit ADC and a loop lter in harge of noise shaping. The loop lter and the nonlinear dynamis of the produe a progressive enoding of widening windows from the original signals so that there is no oneto-one relation between single bits and projetions. On one hand, suh a tehnique has the desired effet to allow squeezing amplitude information into a sequene of sign informations. On the other hand, its nonlinearity avoids the writing of a simple linear model linking the signal samples xn with the bits produed by the enoder zn . Although, this is formally true, it 6. RADS Converter 68 will be neessary to waive the detailed modeling of the operations to onentrate on its high-level funtionality of oversampling onverter with noise-shaping abilities. In doing so, we will obtain a model that an be effetively plugged into reonstrution algorithms. Without any lost of generality, we may fous on a normalized aquisition time of one seond, and model the signal x(t) to be sampled at the Nyquist rate N by olleting xk = x Nk for k = 0; : : : ; N 1. Clearly x 2 R N and x is sparse if there is an N N matrix suh that x = s for some vetor s 2 R N in whih at most K << N omponents are non-zero. Given that the analog waveform x(t) orresponding to the samples in x is sampled at frequeny M , that in general is larger than N , dening the oversampled signal x0n = x Mn for n = 0; : : : ; M 1 we an link the two vetors x0 and x by a linear operation x0 = Ax, being A an upsampling matrix that onsiders the omponents of x as the Nyquist samples of a bandlimited signal. Hene, = operations do not apply to the original omponents of the vetor but to a vetor oversampled by a fator M=N . The sin-interpolation matrix A 2 R M N is dened as: Aj;k = sin N M 1 (j 1) (k 1) 1 j = 1; : : : ; M k = 1; : : : ; N = M we have that A = I . Note that sine we are dealing with 1-bit measurements we have M = B , so and for the ase of N oversampling does not imply an inrement in the total number of bits. With referene to Figure 6.3, the samples in x0 are multiplied by a Nyquist- rate random sequene p1 ; p2 ; : : : ; pN . Applying a further linear operator indiated with the symbol P whih is dened by 6. RADS Converter 8 <p 69 =k (6.7) if j 6= k , therefore, the input of the is the vetor Px0 = PAx. The binary output of the = at time n an be expressed as the sum of the orPj;k = : d MN e 0 j if j responding input sample and a term aounting for the quantization noise whih spetral prole is ditated by the Noise Transfer Funtion (NTF) of the onverter loop [55℄. Hene z = PAx + where aounts for the quantization noise introdued by the = onverter. Conventional = approahes have P equal to the identity and exploit this onstrution by noting that low-pass ltering z is equivalent to low-pass ltering PAx + = Ax + and thus invert upsampling to reover x with an error equal to the low-pass ltering of , a term that an be made very small by playing with the NTF, i.e., making it as high-pass as possible given other implementation onstraints. In our ase, the matrix P is designed to introdue spreading in order to allow that higher frequeny omponents of the upsampled signal enter the baseband range in whih the bits in z are proessed. This alias normally prevents signal reonstrution. Yet, sparsity an be exploited to ounter alias and allow the aquisition of signal omponents that would otherwise fall out of the reah of the range (or, onversely, allow smaller oversampling to aquire the same signal). Figure 6.4 shows the spetrum at different points of the system. For simpliity, assume that the spetrum of pn is at in the interval ( N; N ) and negligible outside that interval. Note that y is now band limited to 23 N and that, depending on the value of the sampling frequeny M , the replias of the spetrum will alias on the disrete time signal yn. 6. RADS Converter 70 |X f)| K ( -M Fig. 6.4: -N -N/2 0 N/2 N M f Frequeny oupany at the different point of the system. From top to bottom: spetrum of the input sparse signal in the Fourier domain x; spetrum of the modulating signal p; spetrum of the modulated signal y as a sum of different shifts of the modulating signal; spetrum of the output signal z with the addition of the quantization noise shaped by the N T F of the = modulator; remaining spetrum after low-pass ltering. Given that we are multiplying the input signal by a pseudorandom sequene, with a very high probability and independent of the sparsity basis, the resulting signal after the modulation will be spread over a large bandwidth. Furthermore, sine the rate of hange of the pseudorandom sequene is equal to the Nyquist rate of the input signal, there will be always a ontribution of every omponent of 6. RADS Converter the original signal into the low part portion of the bandwidth. 71 6. 6.3 RADS Converter 72 Deoding and reonstrution In order to reonstrut the original signal x(t) form the 1-bit samples zn , sparsity is only one of the two priors we have, the other is the high-pass nature of the disturbane introdued by the = modulator. This further piee of information allows us to remove the biggest amount of energy of the quantization noise, while leaving enough information to reonstrut the original x(t) in the low-pass portion of the spetrum. Note that while signal energy dereases linearly as the band shrinks around DC, disturbane energy dereases polynomially thanks to the NTF of the modulator [55℄. The blok diagram shown in Figure 6.5 is used to reover the original signal x(t) from the 1-bit samples zn . The left-hand part of the blok diagram is a lowpass lter that removes the biggest ontribution of the quantization noise, and it is followed by a deimation operation that removes redundant samples. zn H M R zbn CS Alg x n Fig. 6.5: Deoding and reonstrution sheme for RADS Converter. The 1-bit input signal is rst ltered, deimated, and then proessed by a ompressive sensing reonstrution algorithm. Consider a low-pass lter with a utoff frequeny R=2. Depending on the ratio R=N , and onsidering a perfet lter with a deimation operation that leaves only R signiant samples, the remaining samples form a system of linear equations, that for R=N < N (whih is the most ommon ase) is undetermined. This system of equations an efiently be solved by the right-hand part of the diagram that represents any of the CS reonstrution algorithm we have seen in setion 5.3 in the previous hapter. 6. RADS Converter 73 Note that the band between R=2 and R=2 ontains the ontribution of all possible shifts of the spetrum of pn determined by the frequeny values present in x(t), as it is shown at the bottom of Figure 6.4. If spreading were not applied before = modulation, only a portion of the signal would enter in suh a band. To determine the value of the utoff frequeny of the lter R=2, it is desirable to take R as small as possible whih ontributes to remove the quantization noise produed by the = onverter. On the other hand, the signal obtained after ltering should ontain enough information for the reovery of the original sparse signal. In other words, the number of signiant omponents of the ltered signal must be large enough to guarantee that the CS reonstrution algorithm an reover the original signal with a high probability of suess while removing as muh noise as possible. The orret hoie of the bandwidth R will determine the system performane. To model the ltering proess we use an l-order FIR lter (l M ) and arrange its oefients h1 ; h2 ; : : : ; hl as the rows of a matrix H of M M elements. Hj;k 8 <h =: 0 =k+i 1 if j 6= k + i 1 if j i As an example with M 2 H = 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 j; k = 1; : : : ; M i = 1; : : : ; l: = 8 and h = [h1; h2; h3℄ h1 h2 h1 h3 h2 h1 h3 h2 h1 h3 h2 h1 h3 h2 h1 h3 h2 h1 h3 h2 h1 3 7 7 7 7 7 7 7 7: 7 7 7 7 7 7 5 6. RADS Converter 74 Enoding the downsampling operator in the matrix Dj;k 8 < = :1 0 = if j 6= if j k R k M R M j = 1; : : : ; R k = 1; : : : ; M we may nally link the sparse vetor s with the ltered and downsampled measurement z# as z# = DHPAs + DH: Dene the matrix = DHPA and the noise vetor e = DH to have z# = s + e, where 2 R N xR with R < M . This reasts the lassial Compressive Sensing problem presented in setion 5.1 of the previous hapter, and an be efiently solved with a greedy algorithm or an L1-norm minimization to nd the sparse vetor s. 6.3.1 Reonstrution Signal to Noise Ratio Estimation In this subsetion, we estimate the performane in terms of RSNR ahieved by the RADS Converter. Sine no other errors are modeled in the previous analysis, quantization noise limits the performane of the reonstrution algorithm and of the whole arhiteture. It is possible alulate the total power of quantization noise from equation (6.6) onsidering the remaining bandwidth determined by the lter utoff frequeny. On top of that, we have payed a prie when we deided to have a ontribution of every possible omponent of the signal in the low pass portion of the band, i.e. we have spread the energy of every single omponent over the whole bandwidth of the original signal. Given that signal energy is onserved as it pass trough the random multipliation (we have multiplied by a 1 sequene), the 6. RADS Converter 75 magnitude of the signal that remains at the end of the reonstrution algorithm will be inversely proportional to the original signal bandwidth. Finally, we an estimate the RSNR as 0 RSNR = 20log10 1 kxk2 Nerms p2LL+1 M R (L+ 12 ) A (6.8) Note that it is possible to ahieve a signiant improvement with respet to lassial = onversion by making R as small as possible, given that sine the oversampling ratio in the above expression is alulated with respet to the lter bandwidth instead the signal bandwidth. 6.3.2 Numerial Experiments In this subsetion, we present the results of a set of numerial experiments designed to verify and validate the RADS Converter arhiteture. All the simulated points showed in the plots are the mean value over more than 5000 simulations where a new signal was generated with a random support in every trial. The 1-bit enoding was made using third order = modulator designed with delsig [59℄ toolbox for Matlab [60℄. For all the simulations of this setion we have xed a set of parameters that illustrates the most signiant ases. The number of samples was always xed to M = 2048 independently of the time sale used. We have onsidered an input signal that is K -sparse in the Fourier domain , i.e. it is onstruted of up to K different periodi tones, x = Fs where Fj;k = real n e 2 ( 1)(k 1) N j o j; k = 1; : : : ; N 6. RADS Converter 76 and the value of N is varied aross the simulation. In order to show deferents behaviors of the system, we have simulated a set of different sparsity values of K (K = 4, 12, 20 and 28). The power of the input signal was kept onstant along all values of K , whih implies a derease in the value of every single omponent as the sparsity value inreases. The = modulator was hosen to be a third order modulator, whih is a typial onguration for this kind of onverters and due to the fat that in higher order modulators instabilities are more frequent to happen in the loop lter [55℄. The CS reonstrution algorithm at the end of the hain is CoSaMP, and the number of iteration is xed to 200. In the rst set of simulation we have estimated the RSNR and the P SR as a funtion of the utoff frequeny of the reonstrution lter. The results are plot in Figure 6.6. 6. RADS Converter 77 Effect of the reconstruction filter bandwidth 60 K =4 K =12 K =20 K =28 50 RSNR [dB] 40 30 20 10 0 50 100 150 200 250 300 Filter Cutoff frequency (R) 350 1 PSR 0.8 0.6 0.4 K =4 K =12 K =20 K =28 0.2 0 Fig. 6.6: 50 100 150 200 250 300 Filter Cutoff frequency (R) 350 Effet of the reonstrution lter bandwidth for a RADS Converter with a third order = modulator for different sparsity levels. On top: RS N R as a funtion of lter bandwidth R; bottom: P S R as a funtion of R. For every ombination of K and N there is an optimal value for R. As it is shown in the plots, for small values of R, it is not possible to reonstrut the original signal. This is due to the fat that only a small amount of information is left after ltering, and it is not possible to distinguish whih are the 6. RADS Converter 78 omponents present in the signal. In different words, there is a large probability that more than one signal ould be a good andidate solution with this redued set of measurements. As we inrease the value of R, the probability of reonstrution jumps from almost 0 to almost 1, for small values of K . This behavior is onsistent for different values of R at every K we have simulated. At this point, there is enough information to distinguish whih are the omponents present in the original signal from the low pass portion of the mixed signal. The reonstrution error is limited by the quantization noise that is left in this portion. As we ontinue inreasing the lter bandwidth, there is a dereased in the performane in terms of RSNR, as well as P SR. The deterioration in the RSNR an be easily explained due to the fat that a larger bandwidth produes an inrement in the power of the quantization noise (see eq. (6.8)) , i.e., the residual noise energy is large ompared to the signal energy. On the other hand, for large values of R, as we inrease K there is a derease in performane in terms of P SR. As we have the same amount of quantization noise power for a xed R, inreasing the value of K redues the power of every single omponent present in the input signal, making it harder for the reonstrution algorithm to identify those omponents in a noisy environment (the magnitude of the noise is omparable with the magnitude of the signal). This fat illustrates the existing trade-off in the seletion of the R parameter. As we inrease the value of R in order to obtain a better performane in terms of P SR, there is a detriment in terms of RSNR. The optimal value of R will be the smallest value that produes a P SR near to one, and this value is a funtion of K as well as of N . We have extensively studied the optimal value of R trough an empirial approah using numerial simulation. We have found that to obtain a P SR 0:99 then N R 1:4Klog K + 6 + 30 (6.9) 6. RADS Converter 79 . This equation is in aordane with equation (5.1) presented in setion 5.2 of the previous hapter. However, in the following setion we will see how a further exploitation of the RADS arhiteture will lead to an improvement in both gures of merits. In Figure 6.7, we show the simulation results for the optimal value of R given by equation (6.9) as we vary the value of N . We have also added the urves given by equation (6.8) in order to ompare the simulation result with the predited theoretial RSNR. Rads performance for optimal filter bandwidth 70 60 RSNR [dB] 50 40 30 20 K =4 K =12 K =20 K =28 10 0 1 Fig. 6.7: 2 3 4 5 6 Oversampling ratio (M/N) 7 8 Simulation of the performane of RADS Converter using the optimal reonstrution lter bandwidth for different sparsity levels. RS N R as a funtion of the oversampling ratio M =N . The support was orretly reovered 100% of the time. The solid line represents the simulation result, while the dashed line the theoretial result from eq. (6.8). The estimated RSNR follows the behavior of the simulated system, in terms of variation of parameters K and N (the same ours with M , not shown). The differenes are due to the linear model used in the approximation of the in-band noise of the = onverter (whih a very non-linear system), and the non-ideal 6. RADS Converter 80 behavior of the lters used for the reonstrution. Clearly, the expression in equation (6.8) an be used as a design guideline. Note also that by taking just 1-bit measurements at Nyquist rate we an get resolutions of up to 52dB , obtaining a ompression rate of about 8 times with respet to Nyquist sampling for the same resolution. Finally, in Figure 6.8 we have simulated the same setting as before. In this ase, we have added some intrinsi noise to the original signal (i.i.d. additive Gaussian noise) to get an input SNR of 30 dB. Rads performance for optimal filter bandwidth 50 RSNR [dB] 40 30 20 K =4 K =12 K =20 K =28 10 0 1 Fig. 6.8: 2 3 4 5 6 Oversampling ratio (M/N) 7 8 Simulation of the performane of the RADS Converter using the optimal reonstrution lter bandwidth for different sparsity levels. The input signal has an intrinsi S N R of 30 dB. RS N R as a funtion of the oversampling ratio M =N . The support was orretly reovered 100% of the time. As we an see in the plots, the performane of the onverter is limited by the intrinsi noise present in the signal, even if there is a sort of denoising for the smallest values of K . In this setion we have shown how the RADS Converter an be employed 6. RADS Converter 81 to aquire analog-sparse signals with a total number of bits muh smaller than Nyquist multi-resolution analog-to-digital onverters. Simulation results have shown that the proposed arhiteture ollets the neessary information to suessfully reonstrut sparse signals. In the next setion we will see how we an exploit the peuliarities of the aquisition strategy to produe an improved estimate of the signal in terms of auray and probability of suessful reonstrution over different sparsity onditions. This further exploitation will derive in an algorithm that we have alled FCoSaMP whih is free of the parameters that ompromise both gures of merit. 6. RADS Converter 6.4 82 FCOSAMP In general, signal reonstrution for a CS aquisition sheme an be split into two parts: support reovery, i.e. the identiation of the loation of the nonzero omponents, and amplitude estimation over that support. Consider rst the situation in whih the support is already known. If the olumns of the measurement matrix indexed by the loation of the nonzero omponents form a full-rank matrix, the natural approah is to reonstrut the signal by least squares, and the approximation error will be only limited by the power magnitude of the noise introdued by the measurement proess. On the other hand, in the general ase where the support is not known, most algorithms an be ensured to work based on the RIP of the measurement matrix as stated in the previous setion. For some matrix onstrution with entries that are Gaussian or sub-Gaussian, the RIP is satised with overwhelming probability if the number of measurements is bigger than a multiple of the signal sparsity M CK log(N=K ). If the number of measurements falls below a ertain minimum number, the probability of suessful reonstrution hange from a very high to a very poor one (see e.g. [61℄). This phenomena, in terms of ompressive sensing, is the so alled phase transition effet. As we have seen before, thanks to spreading, every non-zero entry in s implies a waveform whose energy an be deteted at pratially any frequeny inluding those where the quantization error is redued by the =. Hene, to remove the quantization noise it is desirable to take only a small bandwidth around zero where the high-pass nature of the disturbane has only a small ontribution. On the other hand, the onsidered signal should ontain enough information for the reovery of the original sparse signal. This trade-off ts partiularly well into algorithms iterating an elementary step that estimates supp (s) and then alulates the orresponding non-null entries. In these algorithms, it is possible to low-pass lter (and deimate to remove redundant samples) the input vetor at eah iteration. Doing so, as the reon- 6. RADS Converter 83 strution proeeds, its renement happens with values that are progressively less affeted by disturbanes sine, while signal energy dereases linearly as the band shrinks around DC, disturbane energy dereases polynomially thanks to the NTF. The nature of the arhiteture allows to develop an iterative algorithm that reovers the support with a very high probability, sine we start the estimation proess with a big number of measurements, and redues the quantization noise to the minimum possible depending on the sparsity level. The algorithm we propose to exploit this intuition is reported in Algorithm 1 and will be referred as FCoSaMP in the following. We an model the lter proess as the appliation of an l-order FIR lter (l m) and arrange its impulse response oefients h1 ; h2 ; : : : ; hl as the rows of a matrix H(m) of m m elements, where m is the length of the sequene to be ltered. (m) = Hj;k 8 <h =k+i 1 :0 if j 6= k + i 1 i if j j; k = 1; : : : ; m i = 1; : : : ; l Depending on the lter utoff frequeny, the ltered sequene an be dei mated by a fator d. We an model this operator in the matrix D(d;m) of md m elements D(d;m) = j;k 8 < 1 :0 = if j 6= if j k d k d j = 1; : : : ; md k = 1; : : : ; m By writing the number of measurements as M = 2Kd0 d1 : : : dJ 1 with dj being a small downsampling fator (typially 2 or 3) and J being the total number of downsampling steps of the algorithm, at the j -th iteration the outer loop lters the signal and downsample it by a fator dj to redue quantization noise. In our 6. RADS Converter 84 implementation, low-pass ltering was obtained by sin frequeny proles with lobes mathed with the subsampling ratio. Downsampling ontinues until the number of available samples is 2K sine this is the minimum information needed to disriminate between two different K -sparse vetors. The inner loop is performed a xed number of times and is based on CoSaMP to iteratively produe an improved estimation of s by least squares over a redued support made of the support of the previous iteration plus the support of the largest omponents of the residuals of the previous iteration. 6. RADS Converter 85 Algorithm 1 Reonstrut x from 1-bit vetor z Complex onjugate transpose of . y Pseudoinverse of . y = ( ) 1 . wjK Set to zero all w but the K biggest omponent. supp(w) Indexes of the nonzero omponents of w. dj Downsampling ratios suh that M = 2K d0 d1 : : : dJ 1 Require: Sampling matrix , 1-bit vetor z, sparsity level K . m M s ( (0; : : : ; 0)T for j = 1; : : : ; J 1 do z ( D(dj ;m) H(m) z ( D(dj ;m) H(m) s m bm=d for i = 1; : : : ; I do w( v T ( supp(wj ) [ supp(sj ) b(T ) ( (; T )ys b(f1; : : : ; N gnT ) ( (0; : : : ; 0) s ( bj v ( z s end for T ( supp(sj ) b(T ) ( (; T )y s b(f1; : : : ; N gnT ) ( (0; : : : ; 0) s ( bj end for x^ ( s v(z j K K T K K T K Intuitively, the high probability of orret support reovery omes from the fat that we estimate it under large noise ondition, but large number of measurements. One the support is identied (at every iteration the support estimation is improved), the bandwidth is dereased in order to redue the quantization noise. The key fat is to note that the signal energy dereases linearly when the frequeny dereases, while noise energy dereases polynomially thanks to the = 6. RADS Converter 86 noise shaping properties [55℄. This ombination of ltering and estimation, has the benet of reovering the signal with a very high probability of suess, while reduing the quantization noise to the minimum. 6.4.1 Numerial Experiments The simulations run in this setion share the same set of parameters and onguration as the simulations draw in Setion 6.3.2. In the rst experiment (Figure 6.9), we have run the same enoding as that in Figure 6.7, and we have made the deoding with FCoSaMP. Note the inrement in terms of RSNR of at least 30 dB in all the ases. It is also important to note the large RSNR that is ahieved espeially for very sparse signals. Note that for M/N = 1 we are just taking 1-bit measurement at Nyquist rate and obtaining a RSNR of up to 90dB , whih translates into a ompression fator of about 15 times with respet to Nyquist sampling. 6. RADS Converter 87 Rads performance with FCoSaMP 100 RSNR [dB] 80 60 K =4 K =12 K =20 K =28 K =36 K =44 40 20 0 1 2 3 4 5 6 Oversampling ratio (M/N) 7 8 1 PSR 0.8 0.6 K =4 K =12 K =20 K =28 K =36 K =44 0.4 0.2 0 1 Fig. 6.9: 2 3 4 5 6 Oversampling ratio (M/N) 7 8 Simulation of the performane of RADS Converter using the FCoSaMP algorithm in the reonstrution for different sparsity levels. On top: RS N R as a funtion of the oversampling ratio M =N ; bottom: P S R as a funtion of M =N . The large RSN R that is ahieved translates into a ompression fator of up to 15 times with respet to Nyquist based aquisition Another interesting fat is that the support reovery was always orret for K < 36, while it was substantially less than 100% only for K 36 when low oversampling ratios are onsidered. This is mainly due to the large bandwidth that 6. RADS Converter 88 remains at the end of the algorithm, i.e., to the residual noise energy that is large with respet to the signal energy therefore large noise energy ompared with the signal energy. In the seond experiment, we have added intrinsi noise to the signal by adding i.i.d. Gaussian noise to get an input SNR of 30 dB. The results are shown in Figure 6.10. 6. RADS Converter 89 Rads performance with FCoSaMP 50 RSNR [dB] 40 30 20 K =4 K =12 K =20 K =28 K =36 K =44 10 0 1 2 3 4 5 6 Oversampling ratio (M/N) 7 8 1 PSR 0.8 0.6 K =4 K =12 K =20 K =28 K =36 K =44 0.4 0.2 0 1 Fig. 6.10: 2 3 4 5 6 Oversampling ratio (M/N) 7 8 Simulation of the performane of RADS Converter using the FCoSaMP algorithm for reonstrution, for different sparsity levels. The input signal has an intrinsi SN R of 30 dB. On top: RS N R as a funtion of the oversampling ratio M =N ; bottom: P S R as a funtion of M =N . The enoding proess and the reonstrution algorithm show to be robust against strong noise ondition. As in the experiment presented in previous setion, the performane of the onverter is limited by the intrinsi noise of 30 dB. However the denoising effet is less evident in this simulation. This is due to the fat that as we lter and 6. RADS Converter 90 deimate iteratively, given the not ideal behavior of the lters, part of the noise is aliased into the low part of the band and added to the total noise energy at the end of the algorithm. In spite of this, it is shown that the enoding proess of the RADS Converter, as well as the behavior of reonstrution algorithm are robust against strong noise ondition showing a behavior in terms of P SR similar to that of the previous simulation. To avoid possible biases due to the hoie of a partiular sparsity basis, in the third experiment we have hanged the sparsity basis and we have simulated the aquisition of a signal that is sparse along a random basis obtained by orthonormalizing a matrix with Gaussian independent entries with zero average. The results are shown in Figure 6.11. Comparing this results with those in Figure 6.9, note that there is a slight differene in terms of RSNR. This is due to the fat that the former may ontain some omponents that when looked in the time domain onentrate most of its energy in small time intervals. In other words, the signal energy is not uniformly distributed along the time axis, making many of the samples taken by the RADS Converter useless or without information. In the extreme ase, when all the energy is onentrated in a small period of time ompared with the time-window used for the proessing, RADS Converter will fail to deode this kind of signals. 6. RADS Converter 91 Rads performance with FCoSaMP 100 RSNR [dB] 80 60 K =4 K =12 K =20 K =28 K =36 K =44 40 20 0 1 2 3 4 5 6 Oversampling ratio (M/N) 7 8 1 PSR 0.8 0.6 K =4 K =12 K =20 K =28 K =36 K =44 0.4 0.2 0 1 Fig. 6.11: 2 3 4 5 6 Oversampling ratio (M/N) 7 8 Simulation of the performane of RADS Converter by using the FCoSaMP algorithm for reonstrution, for different sparsity levels. The input signal is sparse in a random basis. On top: RS N R as a funtion of the oversampling ratio M =N ; bottom: P S R as a funtion of M =N . The proposed arhiteture shows to work independently of the sparsity basis provided it is spread on the time axis. Finally, in the last experiment we have ompared the performane ahieved by our system with two state of the art 1-bit ompressive sensing algorithms, i.e., 6. RADS Converter 92 the Restrited Step Shrinkage (RSS) [62℄ and Binary Iterative Hard Thresholding (BIHT) [51℄, that are generi shemes working on measurement matries with nie theoretial properties (Figure 6.12). Comparison of RADS, RSS and BIHT 90 80 RADS−FCoSaMP 1bRMPI−RSS 1bRMPI−BIHT RSNR [dB] 70 60 50 40 30 20 1 Fig. 6.12: 1.2 1.4 1.6 1.8 Oversampling ratio (M/N) 2 Comparison of the performane of RADS Converter deoded with FCoSaMP with 1bRM P I deoded with the RSS and BIHT algorithms. RS N R as a funtion of the oversampling ratio M =N for x signal length N = 1024 and sparsity level of K = 10. The same amount of 1-bit samples are onsidered for every ase. In every trial we have simulated the aquisition of a signal that is 10-sparse along a random basis. Independently of the arhiteture, the same amount of 1-bit samples z are onsidered as input for the reonstrution algorithm. In all ases, the RADS sheme was able to perfetly reonstrut the support of the original signal and, as shown in Figure 6.12, it ahieves an RSNR largely superior to that of the referenes. 6. 6.5 RADS Converter 93 Time Domain Analysis Up to now, we have made a high level analysis based on frequeny domain assumptions of the funtionality of the RADS onverter, and we have evaluated two different reonstrution algorithms that produe different results in terms of the seleted performane metris. It seems that in the studied ases, the ahieved performane is limited by the seleted reonstrution algorithm, and not by the aquisition arhiteture itself. The main question we want to answer in this setion is: what is the maximum ahievable performane of the RADS onverter (independently of the reonstrution algorithm)? To answer this question we annot perform only the high level analysis we have made so far, but we need a deeper understanding of the enoding proess. For this purpose, we will make a time-domain analysis of the onverter, starting by a time-domain analysis of single 1st -order = modulator, then generalizing it to a Lst -order = modulator, and nally analyzing the whole RADS Converter arhiteture. In addition, we will also show how to exploit the time-domain analysis made for the RADS onverter in order to reonstrut the original signal from the one bit measurements. 6.5.1 1 st = modulator time-domain analysis -order = modulator time-domain analysis Consider rst a disrete time 1st -order = modulator as in gure 6.13 with zero initial onditions, where the disrete sequene y feeds the modulator, that outputs the disrete sequene z , and where the internal state is dened by the state variable v at any time n. The blok alled [Q℄ represents a general quantizer and the blok alled [D℄ represents a one time-step delay. 6. yn + RADS Converter + - vn + 94 zn Q D D Fig. 6.13: First order = modulator shemati diagram. Following the signal path we an write the following equation at any time n: vn = yn zn 1 + vn 1 (6.10) and of ourse vn 1 = yn 1 zn 2 + vn 2 (6.11) Replaing (6.11) into (6.10) we get vn = yn zn 1 + yn 1 zn 2 + vn 2 Extending the same reasoning up to n = 1 and for v1 tions) we an write vn = yn zn 1 + yn 1 vn = yn + + y1 zn 1 zn 2 + yn 2 z1 = = 0 (zero initial ondi- zn 3 + vn 3 n X =1 i yi n 1 X =1 zi (6.12) i This equation relates the urrent state variable v at any time n with the whole history of input y and output z up to time n 1. RADS Converter 6. 95 On the other hand, sine we are using a 1-bit quantizer [Q℄, without loss of generality by enoding the value 0 10 for all positive inputs of the quantizer (inluding zero), and the value 0 10 for all negative values of the input, we have that vn zn 0 8n (6.13) Dening the vetors 2 y = 6 6 6 6 6 6 4 y1 : : : yn 3 2 7 7 7 7 v 7 7 5 6 6 6 6 6 6 4 ; = v1 : : : vn 3 2 7 7 7 7 z 7 7 5 6 6 6 6 6 6 4 ; = z1 : : : zn 3 7 7 7 7 7 7 5 and the matries 2 Z 2 1 0 0 6 6 1 1 0 6 6 1 1 1 6 = 666 6 6 6 4 : : : 1 1 1 = 6 6 6 6 6 6 6 6 6 6 6 4 : : : : 0 0 0 z2 0 0 0 z3 z1 : : : : : : 0 : : : 0 : : : 0 : : : : : : : : : zn 3 7 7 7 7 7 7 7 7 7 7 7 5 0 0 0 2 3 0 0 0 : : 0 6 7 6 1 0 0 : : 07 6 7 6 1 1 0 7 : : 07 6 7 ; = 666 : : 7 7 : : : : : : : 1 7 7 7 5 6 6 6 4 : : 1 1 1 ; : : : 0 : : : 0 : : : 0 : : : : : : : : : 0 3 7 7 7 7 7 7 7 7 7 7 7 5 6. RADS Converter 96 we an write v = y z (6.14) and Zv 0 (6.15) where the last inequality is omponent-wise. In this way, ombining equation (6.14) with equation (6.15), and given the measurements z we an dene a solution spae for any input y: Z yZ z This spae ontains all possible instanes of the input y that are solutions of the = modulation proess. Lth -order = modulator time-domain analysis Consider now a disrete time Lth -order = modulator as in gure 6.14 with zero initial onditions. yn wn L-order Loop Filter th Fig. 6.14: L vn Q zn -order = modulator shemati diagram. In this ase, the vetor w 2 R L denes the state vetor of the loop lter, while RADS Converter 6. 97 the state variable v is equal to the last omponent of the state vetor (w:;L ). The urrent state at time n an be omputed based on the previous states as wn = By 1 + Cz 1 + Aw n n n 1 where the matrix A 2 R LxL and the vetors B; C 2 R L ontain the oefients that determine the transfer funtion of the loop lter. The output of the loop lter is simply vn = wn;L and the output of the modulator is alulated as zn = sign(vn), where w:;L is the Lth omponent of the state vetor w. Analogously as proeeded with the 1st -order modulator, we an write wn wn = By 1 + Cz 1 + ABy 2 + ACz 2 + A(2) w n n = A(0) By 1 + + A( n n n n 2 2) By1 + A(0) Czn 2 + + A(n 2) Cz1 n 1 X n 1 X ( i 1) wn = B A yi + C A(i 1) zi i=1 i=1 sine w1 = 0 (zero initial onditions). We also have that, vn zn 0 Dening the vetors n 8n (6.16) 6. RADS Converter 2 2 y = 6 6 6 6 6 6 4 y1 : : : yn 1 3 7 7 7 0 7 v 7 7 5 ; = 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 w2;1 w2;2 : : w2;L w3 : : : wn 1 98 3 7 7 7 7 7 7 7 7 7 7 7 z 7 7 7 7 7 7 7 7 5 ; = 2 6 6 6 6 6 6 4 z1 : : : zn 1 and the matries 2 Z = 6 6 6 6 6 6 6 6 6 6 6 4 0 0 0 z2 0 0 0 z3 z1 : : : 0 0 0 : : : 0 : : : 0 : : : 0 : : : : : : : : : zn 1 Dene i = A(i) B Æ i = A(i) C then, 3 7 7 7 7 7 7 7 7 7 7 7 5 ; 3 7 7 7 7 7 7 5 6. 2 0 = 0 0 0 1 0 0 2 1 0 : : : n 2 n 3 n 4 6 6 6 6 6 6 6 6 6 6 6 4 2 0 = RADS Converter 0 99 : : : 0 : : : 0 : : : 0 : : : : : : : : : 0 7 7 7 7 7 7 7 7 7 7 7 5 : : : 0 6 6 Æ0 : : : 0 6 6 Æ1 Æ0 : : : 0 6 6 6 : : 6 6 : : 6 6 : : 4 Æn 2 Æn 3 Æn 4 : : : Æ0 Æ0 Æ1 Æ2 : : : 0 0 3 ; 3 7 7 7 7 7 7 7 7 7 7 7 5 We an now write equation (6.16) in matrix form as v0 = 0y 0z (6.17) In order to keep only the last omponent of the state vetor, dene the matrix K 2 R L(N 1)xN 1 as 2 K = 6 6 6 6 6 6 6 6 6 6 6 4 0 : : : 0; 1 0 0 : : : 0 0 0 : : : 0; 1 0 0 0 0 : : : 0; 1 0 0 : : : : : : : : : : : : : : : 0 0 0 : : : 0 : : : 0; 1 3 7 7 7 7 7 7 7 7 7 7 7 5 ; 6. RADS Converter 100 where 0 represents a row vetor of size L with all zeros in its inputs. Dening v as 2 v = 6 6 6 6 6 6 4 3 v1 : : : vn 1 7 7 7 7 7 7 5 to get v = K 0 y K 0z , and v = y z (6.18) where = K 0 and = K 0 . As in the 1st -order ase we have, Zv 0 (6.19) and ombining equation (6.18) with (6.19) to have Z yZ z (6.20) The spae dened by equation (6.20) ontains all possible instanes of the input y that are solution of the = modulation proess, given the measurements z. Now, we an highlight some observations for the above equation: 6. RADS Converter 101 The input signal denes a point in the multi-dimensional spae that is ontained in the solution spae dened by the set of equations in (6.20). Fixing the dimension of the input signal, as we add new measurements, every measurement will split the spae into two sub-spaes. Only one of those sub-spaes will ontain possible solutions. The minimum number of measurements needed to dene a losed region is equal to the dimension of the input signal plus one. Adding a new measurement, does not imply a redution in the solution spae. A smaller solution spae implies an estimation of the input signal with a bigger auray. In other words, the smaller the solution spae, the bigger the SNR of the estimated signal. 6.5.2 RADS Converter time-domain analysis We now have all we need to analyze the whole RADS onverter arhiteture. The set of inequalities Zy Zz dene the solution spae given by the = modulator. In the same way as above, to ompletely model the RADS Converter and the input signal itself, we an easily write y = DAx = DAs to have Z DA s Z z (6.21) where the matrix A is the upsampling operator and the matrix D represents the pre-modulation proess, as we have dened before. 6. RADS Converter 102 We have now a omplete desription of the modulation proess of the RADS Converter arhiteture in time domain. The main differenes between the RADS Converter and a = alone are two: rst, the pre-modulation inreases the probability that every new measurement modify the solution spae, inreasing in this way the auray in the estimation. Seondly, under the assumption that s is sparse in a given domain, it is possible redue the solution spae to only those andidates that satisfy this ondition, reduing even more the solution spae. Sine sparsity is not a dimensionality redution, it is not possible to know a priori whih are the diretions to look at, but as we proeed with the measurements, there will be many andidates to disard sine they are not sparse enough to be a possible solution. 6.5.3 Spae Dimension Analysis As we have seen before, it is possible to dene a monotoni dependene between the size of the solution spae and the SNR of the estimation of the input signal. As a measure of size, and onsidering that any point in the solution spae is a andidate with the same probability, it is possible to onsider the hyper-volume of that solution spae as a measure of preision of the estimation of the input signal. Regrettably, an analytial expression for the alulation of the hyper-volume in high-dimensional spaes is a difult task, and we need to resort to numerial integration. For that purpose, we will use Monte-Carlo integration [63, 64℄ in order to have an estimation of the hyper-volume of the solution spae as a funtion of the number of measurements. Monte Carlo integration is a tehnique for numerial integration that uses random numbers, and is partiularly useful for higher dimensional integrals. Informally, to estimate the volume of a given domain D, we have rst, to pik a simple domain E whose volume is easily alulated and whih D is ontained. Then, we generate a sequene of random points that fall within E, some of whih will also fall within D. Finally, we alulate the area of D as the area of E by the fration of points that fall between E. 6. RADS Converter 103 In this ase, we have set the ontainer volume as an hyper-ube of 1 1 1 1 and we have generated 500; 000 M random points with uniform distribution within this range for every point in the plot. We have plot in Figure 6.15 the hyper-volume of the solution spae as a funtion of the number of measurements for two ases: using a single = onverter (equation (6.20)); and using the RADS Converter arhiteture (equation (6.21)). As a referene, we have also plot a line with slope 1=2M . This line will our only when every ut of the spae produed by a new measurement divides the solution spae exatly into two equal parts. 6. RADS Converter 104 Monte Carlo Integration of the Solution Space 0.5 RADS Delta−Sigma 1/2M Volume 0.4 0.3 0.2 0.1 0 5 10 15 20 25 Number of measurements (M) 30 0 RADS Delta−Sigma −M log2(volume) −5 −10 −15 −20 −25 Fig. 6.15: 5 10 15 20 25 Number of measurements (M) 30 Monte Carlo integration of the hyper-volume of the solution spae for RADS Converter and for = onverter. On top: volume in linear sale as a funtion of the number of measurements M ; bottom: volume in logarithmi sale as a funtion of M . The enoding performed by RADS is more effetive in reduing the size of the solution spae. As an be observed in Figure 6.15 the differene in size of the solution spae obtained using the RADS Converter approah is orders of magnitude more onvenient than using lassial = modulation. This differene is muh more evident RADS Converter 6. 105 as the number of measurements is inreased. However, as the plot shows it is still possible to obtain an important improvement, sine the line indiating the optimal uts is far away from the one desribed by the RADS Converter. L1-norm Minimization 6.5.4 It was demonstrated above that the set of inequalities given by equation (6.21) dene the solution spae of the RADS modulation proess. This spae still ontains many possible input vetors s, but we are partiularly interested in the sparsest vetor that exist in this spae. In order to nd suh a vetor we an reast to a L1norm minimization, sine from the observed in the previous hapter, it enfores sparsity aross all possible solutions. We an write the following minimization problem s^ = argmin N X =1 js j i s.t. Z DA s Z z (6.22) i whih from now on we will all L1min. 6.5.5 Numerial Experiments In this setion we will show the results from a series of simulation we have run in order to evaluate the minimization problem presented in equation (6.22). We have setup the same onditions for the simulation in setion 6.3.2, exept that in this ase we have redued the number of measurements from 2048 to 1024. It was neessary to redue this number for the simulation to be omputational feasible, sine every measurement produe a new onstrain to be pass to the solver. The minimization problem was solved by the software plex [65℄. Figure 6.16 shows the performane ahieved by FCoSaMP ompared with that obtained using the minimization problem of equation (6.22), by xing the sparsity number to K = 8. 6. RADS Converter 106 As it is observed in the plot, L1min outperforms FCoSaMP by around 10 dB in the whole range. This behavior an be veried using different experimental setups not shown. Performance comparison 85 RSNR [dB] 80 75 70 65 FCoSaMP L1min 60 1 Fig. 6.16: 2 3 4 5 6 Oversampling ratio (M/N) 7 8 Performane omparison of FCoSaMP and L1min. RS N R as a funtion of oversampling ratio M =N for an 8-sparse signal enoded with RADS onverter. The main drawbak of this reonstrution algorithm is the running time needed to solve the minimization problem. In Figure 6.17 we have plotted the relationship between the average simulation time taken by L1min over the time taken by FCoSaMP. 6. RADS Converter 107 Running Time tL1min / tFCoSaMP 250 200 150 100 50 1 Fig. 6.17: 2 3 4 5 6 Oversampling ratio (M/N) 7 8 Reonstrution algorithm exeution time for FCoSaMP and L1min. Relationship between the time taken by L1min and FCoSaMP as a funtion of the oversampling ratio M =N . As it is shown in the plot, the time needed for L1min is between 50 and 250 times longer than that of FCoSaMP. As we inrease the oversampling ratio more equations enter into play, whih redue the searh spae of the minimization algorithm. However, this long reonstrution time make this algorithm only feasible in partiular ases. 6. 6.6 RADS Converter 108 Hardware Implementation In this setion we propose a hardware implementation of the RADS Converter in order to validate the ideas presented above by a real appliation. The implementation was made in a redued size PCB with off-the-shelf omponents. Some onstrains were imposed in the design of the board, sine the output of the = onverter must be a 1-bit output, but it is rather difult to nd a ommerial = onverter with this harateristi in the market (most onverters inlude the deimation lter as well). Figure 6.18 shows a simplied shemati diagram of the implemented arhiteture, and the aspets of the implemented board an be observed in Figure 6.19. signal_in x(t) V+ + clock_in - DS - V- data_out z + V_ref Fig. 6.18: RNG_in Simplied shemati diagram of the hardware implementation of the RADS Converter. Fig. 6.19: Piture of the hardware implementation of the RADS Converter. From the user point of view the onverter presents two inputs: a lok and a 6. RADS Converter 109 random sequene, and one output: the 1-bit RADS Converter output. The lok input and input for the random sequene must be synhronized, and the frequeny relationship must be the oversampling ratio determined by the appliation. The 1-bit output is synhronized with the input lok and must be read before the next rising edge of the inoming lok. The rst stage of the onverter is nothing but an amplier, whih funtions is to onvert the signal from single-ended to differential, and to adapt the signal input level to the = onverter level. After this stage, the signal is passed trough a ombination of swithes, that hange the polarity of the signal as it is ommanded by the RNG input. This proessing is equivalent to the multipliation stage showed in Figure 6.3. The last blok is a onventional = onverter, whih produes a 1-bit output stream. The hosen onverter was an AD7401A, from Analog Devies, whih is a 2nd order disrete time modulator with a maximum sampling rate of 20MSPS. Note that in an integrated implementation the whole arhiteture an be diretly implemented with a slight modiation of the rst stage of a disrete time = onverter. 6.6.1 Measurement Setup Figure 6.20 and Figure 6.21 show the measurement setup. The omplete setup is omposed by the RADS onverter board, a Spartan 6 Development Board responsible for to generating the pseudorandom sequene and to interfae it to a PC trough a USB port, a signal generator with GPIB interfae, a power supply, and a laptop for the ontrol and the aquisition of the measurements. 6. Fig. 6.20: Fig. 6.21: RADS Converter 110 Piture of the RADS Converter onneted to a Spartan 6 FPGA development kit. Measurement setup for the evaluation of the hardware implementation of the RADS Converter. The measurement proedure is desribed below: Set the number of measurements (M ), the Nyquist rate of the input signal 6. RADS Converter 111 to be generated (N ), the sparsity level (K ), and hoose a basis of sparsity for the signal. Generate the samples of the signal to be aquired in the PC with Matlab, and send them through the GPIB interfae to the signal generator. Start the aquisition with the RADS board, save the measurements temporally in the Spartan 6 development board, and transfer them to the PC trough USB. Proess the aquired samples in the PC with FCoSaMP and ompare the reonstruted signal with the synthetially generated signal. The proposed measurement setup is very exible and allows to exploit the whole spae of parameters of the aquisition proess. 6.6.2 Measurements and Validation We have made a series of measurements in order to validate the funtioning of RADS Converter. We have xed the sampling frequeny to 10MHz and we have vary the time window in order to hange the number of aquired measurements. As an example, Figure 6.6.2 shows a plot of an 8-sparse (in a random basis with a Nyquist rate of 5MHz ) syntheti signal generated in Matlab, and superimposed to it, it is the signal aquired with RADS Converter and reonstruted with FCoSaMP. The obtained RSNR was of 32dB . 6. RADS Converter 112 Example of Signal Aquisition with RADS Converter Input Voltage [V] 0.5 0 Original Signal Reconstructed Signal −0.5 0 50 100 time [useg.] 150 200 Input Voltage [V] 0.5 0 Original Signal Reconstructed Signal −0.5 10 Fig. 6.22: 12 14 16 time [useg.] 18 20 Aquisition of an analog signal with RADS Converter. The input signal is 8sparse in a random basis and with a Nyquist rate half the sampling frequeny. On top: the syntheti signal superimposed to the reonstruted signal for the whole aquisition window; on bottom: a zoom-in of the same aquired signal. Figure 6.6.2 shows the signal input spetrum. As it is shown, the spetrum oupany is of 2:5MHz , whih implies a Nyquist rate of 5MHz . With the bit budget utilized by RADS Converter in the aquisition of this signal, it would be obtained a maximum SNR of 12dB by the used of a lassial Nyquist onverter. 6. RADS Converter 113 Input Signal Spectrum Input Power Spectral Density [dB] 25 20 15 10 5 0 −5 −10 Fig. 6.23: −2 −1 0 1 frequency [MHz] 2 Spetrum of the input signal aquired by RADS Converter. The spetrum has a full oupany for frequenies up to 2:5M H z Another example using a sparse signal in the Fourier domain is presented in the 6.24. The plot shows the mean value over 10 measurements obtained by the RADS onverter board. 6. RADS Converter 114 Measurements over Rads converter 60 50 RSNR [dB] 40 30 K =4 K =8 K =12 K =16 K =24 K =32 20 10 0 1 Fig. 6.24: 2 3 4 5 6 Oversampling ratio (M/N) 7 8 Performane of the hardware implementation of the RADS Converter by using the FCoSaMP algorithm for reonstrution, for different sparsity levels. RS N R as a funtion of the oversampling ratio M =N . The support reovery was always orret for the 10 measurements. As an be observed, the trend is the same to that obtained in the simulation, but the performane arhived in terms of RSNR is inferior to the expeted by the simulation. These differenes an be due to, imperfetions in the utilized swithes (on/off resistane, swithing time, frequeny response), different soures of noise (power supply noise, noise introdued by the amplier, thermal noise, quantization noise in the signal generator) and most important, the bandwidth of the input stage of the utilized = onverter (the modulated signal that enter into the = onverter exeeds greatly the onverter speiation). In spite of this, the implementation of the onverter has shown that this arhiteture is promising as an analog-to-information onverter, signiantly reduing the total number of bits with respet to Nyquist based sampling for spei lasses of signals. 6. RADS Converter 6.7 115 Conlusion In this hapter we have introdued the RADS Converter, we have evaluated its performane through theoretial results, numerial simulations and a hardware implementation of the aquisition arhiteture. We have proposed a number of reonstrution algorithms among those we highlight the FCoSaMP and the L1norm minimization. The proposed arhiteture allows a simple hardware implementation for the aquisition of large bandwidth signals that are sparse over a variety of supports, obtaining a very high resolution after reonstrution. This ontrast with lassial sampling methods, where the resolution drastially dereases with the sampling frequeny. We have also evaluated numerially the quality of the algorithm to retrieve a orret support under different input signal ondition, obtaining a very high probability over a wide range of sparsity levels. Finally we have proposed a different approah for the study of the RADS Converter and for = modulators in general. Contrary to what is found in the literature for this kind of onverters, usually evaluated in the frequeny domain [55, 56, 57, 58℄, the proposed approah is based on a time-domain analysis. 7. CONCLUSIONS This thesis builds on the eld of signal proessing, and illustrates with two different appliations how, by inreasing the efforts in the digital domain, it is possible to redue the requirements for the implementation of analog hardware. Speially, we have foused on the analysis of the use of very oarse quantization, more preisely 1-bit quantization, with the aim of obtaining a simpliation in the implementation of both, analog to digital onverters, and digital to analog onverters. We have shown that a proper exploitation of binary quantization an lead to performanes that are similar, and sometimes even better, than those obtained using multibit approahes. In the rst part we have proposed the use of Legendre sequenes (binary sequenes) for the utilization in MIMO ative sensing systems. We have proposed the onstrution of set of sequenes, where eah of the sequenes in the set is made from a different rotation of the same Legendre sequene. We have found that optimal rotations exist, and that the set formed by this binary sequenes has a performane in terms of ISL beyond the one obtained by other sets of binary sequenes. We have also found that the performane obtained by our sequenes is omparable to state-of-the-art algorithms that produe real value sequenes, when quantization is imposed to them up to a ertain level of quantization depth. In order to obtain the optimal rotations, we have presented an analytial expression for the alulation of the ross-orrelation omponents of the ISL of a set of sequenes. This expression, put together with a previously obtained expression for the alulation of the ISL of a single sequene, allowed the reation 7. Conlusions 118 of a omplete expression for the ISL of a set of sequenes. Under asymptoti onditions, this expression an be used to alulate the ISL of sequenes whose generating funtion has a relatively simple trend. Sine this is the ase of Legendre sequenes, we were able to derive an analytial expression for the asymptoti ISL of sets of rotated Legendre sequenes. Suh an expression was exploited to drive the optimization proedure needed to onstrut small-ISL sets of antipodal sequenes of any sequene length with potential appliations to ommuniation and ative sensing systems. We have started the seond part of this thesis by introduing the models neessary to represent the lasses of signals of interest, i.e. sparse signals. We have shown how many high-dimensional signals atually have a limited number of degrees of freedom ompared to its dimensionality. These lasses of signals are known as sparse signals, whih are one of the main ingredients for the development of the ompressive sensing theory. In this part of the thesis we have dealt partiularly with the design and development of a hardware arhiteture for the implementation of a ompressive sensing system. Based on the motivation of this thesis work, one of the requisite we have impose for the implementation of suh a system, was that it must lead into a simple hardware/system implementation. In this way , we have introdued a new arhiteture for an Analog to Information onverter that was alled the RADS Converter. The proposed arhiteture is based on a well-known = onverter that produes 1-bit measurements of the inoming signal. Starting from a = onverter, a straightforward modiation of the input stage topology lead to the implementation of the RADS Converter arhiteture. The reonstrution performane obtained using the proposed onverter was found to depend on the signal information ontent, instead of depending on the signal bandwidth, as it is in the ase for a lassial = onverter. This results in the possibility of aquisition of large bandwidth signals that are sparse over a vari- 7. Conlusions 119 ety of supports, with an extremely high auray after being proessed. Based on ompressive sensing onepts, RADS Converter is able exploit the sparse signal struture to apture all its information ontent by taking single bit measurements. An important nding of this work, was that by exploiting the peuliarities of the aquisition strategy we were able to develop a new reonstrution algorithm that produes an improved estimate (with respet to general algorithms) of the signal in terms of auray and probability of suessful reonstrution. This suggest that, while most of the reonstrution algorithms for ompressive sensing are based on guaranties on the struture of the measurement matrix (RIP based algorithms), it is possible to get a prot by generating more lever algorithms that math with the aquisition arhiteture itself. The modeling of the RADS Converter in the frequeny domain has led to an intuitive understanding of the enoding proess, and has given light on how proeed to efiently reonstrut the input signal from the measurements. However, in order to get a deeper insight into the funtioning of the proposed onverter, we were able to develop a time-domain model of the operations performed to the signal in the enoding proess. With this aim we have raised an algebrai analysis of the spae determined by the measurements, and its redution as new measurements ome into onsideration. The study of the size of that spae, evidenes the differene between the RADS enoding and the = enoding, and allows the alulation/ estimation of the theoretial maximum limit that an be expeted by taking 1-bit measurements of any form. The different perspetive given by the time domain modeling of the enoding proess, has led to the proposal of a new reonstrution algorithm for the RADS Converter arhiteture. This algorithm is based on lassial ompressive sensing onepts that promotes sparsity through the minimization of the L1-norm. It has been demonstrated that the use of this algorithm an produe a better estimate of the signal than its frequeny-based ounterpart. However, the omplex task of minimizing the L1-norm over the huge amount of onstraints, makes this im- 7. Conlusions 120 provement be ahieved at the expense of an inrease in exeution time, making this appliation only feasible for ertain appliations. Besides the extensively numerial simulations performed during the development of this thesis to validate the results, we have implemented the RADS Converter arhiteture in a redued size PCB with off-the-shelf omponents. The implementation of the onverter has demonstrated that this arhiteture is promising as an analog to information onverter, signiantly reduing the total number of bits with respet to Nyquist based sampling, for spei lasses of signals. Although the performane attained by the hardware implementation differs from the one ahieved in simulations, we believe that a proper implementation of the RADS Converter in a speially designed integrated devie an lead to an inrease in the performane lose to the obtained in the simulation. BIBLIOGRAPHY [1℄ J. Oppermann and B. Vueti, Complex spreading sequenes with a wide range of orrelation properties, IEEE Transations on Communiations, vol. 45, no. 3, pp. 365 375, mar. 1997. [2℄ H. Deng, Polyphase ode design for orthogonal netted radar systems, IEEE Transations on Signal Proessing, vol. 52, no. 11, pp. 3126 3135, nov. 2004. [3℄ J. L. H. He, P. Stoia, Designing unimodular sequene sets with good orrelations - inluding an appliation to mimo radar, IEEE Transations on Signal Proessing, vol. 57, no. 11, pp. 43914405, Nov. 2009. [4℄ H. He, P. Stoia, and J. Li, Designing unimodular sequene sets with good orrelations-inluding an appliation to mimo radar, Signal Proessing, IEEE Transations on, vol. 57, no. 11, pp. 4391 4405, nov. 2009. [5℄ H. Khan, Y. Zhang, C. Ji, C. Stevens, D. Edwards, and D. O'Brien, Optimizing polyphase sequenes for orthogonal netted radar, IEEE Signal Proessing Letters, vol. 13, no. 10, pp. 589 592, ot. 2006. [6℄ C. Chen and P. Vaidyanathan, Mimo radar ambiguity properties and optimization using frequeny-hopping waveforms, Signal Proessing, IEEE Transations on, vol. 56, no. 12, pp. 59265936, 2008. [7℄ E. J. Candes, J. Romberg, and T. Tao, Robust unertainty priniples: Exat signal reonstrution from highly inomplete frequeny information, Information Theory, IEEE Transations on, vol. 52, no. 2, pp. 489509, 2006. Bibliography 122 [8℄ E. Candes and T. Tao, Near-optimal signal reovery from random projetions: Universal enoding strategies? Information Theory, IEEE Transations on, vol. 52, no. 12, pp. 5406 5425, de. 2006. [9℄ D. Donoho, Compressed sensing, Information Theory, IEEE Transations on, vol. 52, no. 4, pp. 1289 1306, april 2006. [10℄ J. Li, P. Stoia, L. Xu, and W. Roberts, On parameter identiability of mimo radar, Signal Proessing Letters, IEEE, vol. 14, no. 12, pp. 968971, 2007. [11℄ E. Fishler, A. Haimovih, R. Blum, L. Cimini Jr, D. Chizhik, and R. Valenzuela, Spatial diversity in radars-models and detetion performane, Signal Proessing, IEEE Transations on, vol. 54, no. 3, pp. 823838, 2006. [12℄ D. Bliss and K. Forsythe, Multiple-input multiple-output (mimo) radar and imaging: degrees of freedom and resolution, in Signals, Systems and Computers, 2003. Conferene Reord of the Thirty-Seventh Asilomar Conferene on, vol. 1. IEEE, 2003, pp. 5459. [13℄ M. Skolnik, Radar handbook, 2nd ed. 1990. [14℄ J. Li, P. Stoia, and X. Zheng, Signal synthesis and reeiver design for mimo radar imaging, Signal Proessing, IEEE Transations on, vol. 56, no. 8, pp. 39593968, 2008. [15℄ J. Li, L. Xu, P. Stoia, K. Forsythe, and D. Bliss, Range ompression and waveform optimization for mimo radar: a ramerrao bound based study, Signal Proessing, IEEE Transations on, vol. 56, no. 1, pp. 218232, 2008. [16℄ J. L. P. Stoia, H. He, New algorithms for designing unimodular sequenes with good orrelation properties, IEEE Transations on Signal Proessing, vol. 57, no. 4, pp. 14151425, Apr. 2009. [17℄ H. He, P. Stoia, and J. Li, On aperiodi-orrelation bounds, Signal Proessing Letters, IEEE, vol. 17, no. 3, pp. 253256, 2010. Bibliography 123 [18℄ M. Golay, The merit fator of legendre sequenes, IEEE Transations on Information Theory, vol. 29, no. 6, pp. 934936, Ot. 1983. [19℄ H. J. T. Hø holdt, Determination of the merit fator of legendre sequenes, IEEE Transations on Information Theory, vol. 34, no. 1, pp. 161164, Jan. 1988. [20℄ L. B. M. Antweiler, Merit fator of hu and frank sequenes, IEE Eletronis Letters, vol. 26, no. 25, pp. 20682070, De. 1990. [21℄ R. K. P.B. Rapaji, Merit fator based omparison of new polyphase sequenes, IEEE Communiations Letters, vol. 2, no. 10, pp. 269270, Ot. 1998. [22℄ J. Jedwab, A survey of the merit fator problem for binary sequenes, T. Helleseth et al, eds., Leture Notes in Computer Siene, Sequenes and their Appliations - Proeedings of SETA 2004, Springer-Verlag, vol. 3486, pp. 3055, jan. 2005. [23℄ Best known sequenes. [Online℄. Available: http://www2.hemistry.msu.edu/faulty/linedantus/merit fator reords.html [24℄ T. Høholdt, The merit fator problem for binary sequenes, in Applied Algebra, Algebrai Algorithms and Error-Correting Codes, ser. Leture Notes in Computer Siene. Springer Berlin / Heidelberg, 2006, vol. 3857, pp. 5159. [25℄ G. S. d. G. Polya, Aufgeben und Lehrs¨atze aus der Analyse II. Springer, 1925. Berlin: [26℄ S. J. Haboba, R. Rovatti, and G. Setti, Determination of the integrated sidelobe level of sets of rotated legendre sequenes, arXiv preprint arXiv:1012.3638, 2010. Bibliography 124 [27℄ G. G. S. W. Golomb, Signal Design for Good Correlation: For Wireless Communiation, Cryptography, and Radar. Cambridge University Press, 2005. [28℄ L. Maximon, The dilogarithm funtion for omplex argument, Proeedings of the Royal Soiety, part A: Mathematial, Physial & Engineering Sienes, vol. 459, pp. 28072819, 2003. [29℄ J. Haboba, R. Rovatti, and G. Setti, Integrated sidelobe level of sets of rotated legendre sequenes, in Aoustis, Speeh and Signal Proessing (ICASSP), 2011 IEEE International Conferene on. IEEE, 2011, pp. 2632 2635. [30℄ E. Candes and T. Tao, Deoding by linear programming, Information Theory, IEEE Transations on, vol. 51, no. 12, pp. 42034215, 2005. [31℄ R. Baraniuk, M. Davenport, R. DeVore, and M. Wakin, A simple proof of the restrited isometry property for random matries, Construtive Approximation, vol. 28, no. 3, pp. 253263, 2008. [32℄ S. Mendelson, A. Pajor, and N. Tomzak-Jaegermann, Uniform unertainty priniple for bernoulli and subgaussian ensembles, Construtive Approximation, vol. 28, no. 3, pp. 277289, 2008. [33℄ E. J. Candes, J. K. Romberg, and T. Tao, Stable signal reovery from inomplete and inaurate measurements, Communiations on pure and applied mathematis, vol. 59, no. 8, pp. 12071223, 2006. [34℄ A. Maleki and D. Donoho, Optimally tuned iterative reonstrution algorithms for ompressed sensing, Seleted Topis in Signal Proessing, IEEE Journal of, vol. 4, no. 2, pp. 330341, 2010. [35℄ S. S. Chen, D. L. Donoho, and M. A. Saunders, Atomi deomposition by basis pursuit, SIAM journal on sienti omputing, vol. 20, no. 1, pp. 3361, 1998. Bibliography 125 [36℄ D. L. Donoho, M. Elad, and V. N. Temlyakov, Stable reovery of sparse overomplete representations in the presene of noise, Information Theory, IEEE Transations on, vol. 52, no. 1, pp. 618, 2006. [37℄ J. A. Tropp, Just relax: Convex programming methods for identifying sparse signals in noise, Information Theory, IEEE Transations on, vol. 52, no. 3, pp. 10301051, 2006. [38℄ B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani, Least angle regression, The Annals of statistis, vol. 32, no. 2, pp. 407499, 2004. [39℄ D. Donoho and Y. Tsaig, Fast solution of l1-norm minimization problems when the solution may be sparse, Information Theory, IEEE Transations on, vol. 54, no. 11, pp. 4789 4812, nov. 2008. [40℄ D. Needell and J. Tropp, Cosamp: Iterative signal reovery from inomplete and inaurate samples, Applied and Computational Harmoni Analysis, vol. 26, no. 3, pp. 301 321, 2009. [41℄ D. Needell and R. Vershynin, Signal reovery from inomplete and inaurate measurements via regularized orthogonal mathing pursuit, Seleted Topis in Signal Proessing, IEEE Journal of, vol. 4, no. 2, pp. 310 316, april 2010. [42℄ J. Tropp and A. Gilbert, Signal reovery from random measurements via orthogonal mathing pursuit, Information Theory, IEEE Transations on, vol. 53, no. 12, pp. 4655 4666, de. 2007. [43℄ M. Fornasier and H. Rauhut, Iterative thresholding algorithms, Applied and Computational Harmoni Analysis, vol. 25, no. 2, pp. 187208, 2008. [44℄ T. Blumensath and M. E. Davies, Iterative thresholding for sparse approximations, Journal of Fourier Analysis and Appliations, vol. 14, no. 5, pp. 629654, 2008. Bibliography 126 [45℄ K. Herrity, A. Gilbert, and J. Tropp, Sparse approximation via iterative thresholding, in Aoustis, Speeh and Signal Proessing, 2006. ICASSP 2006 Proeedings. 2006 IEEE International Conferene on, vol. 3, may 2006, p. III. [46℄ J. N. Laska, S. Kirolos, M. F. Duarte, T. S. Ragheb, R. G. Baraniuk, and Y. Massoud, Theory and implementation of an analog-to-information onverter using random demodulation, in Ciruits and Systems, 2007. ISCAS 2007. IEEE International Symposium on. IEEE, 2007, pp. 19591962. [47℄ J. Laska and R. Baraniuk, Regime hange: Bit-depth versus measurementrate in ompressive sensing, Signal Proessing, IEEE Transations on, vol. 60, no. 7, pp. 34963505, 2012. [48℄ P. Boufounos and R. Baraniuk, 1-bit ompressive sensing, in Information Sienes and Systems, 2008. CISS 2008. 42nd Annual Conferene on. IEEE, 2008, pp. 1621. [49℄ P. Boufounos, Greedy sparse signal reonstrution from sign measurements, in Signals, Systems and Computers, 2009 Conferene Reord of the Forty-Third Asilomar Conferene on. IEEE, 2009, pp. 13051309. [50℄ J. Laska, Z. Wen, W. Yin, and R. Baraniuk, Trust, but verify: Fast and aurate signal reovery from 1-bit ompressive measurements, Signal Proessing, IEEE Transations on, vol. 59, no. 11, pp. 52895301, 2011. [51℄ L. Jaques, J. N. Laska, P. T. Boufounos, and R. G. Baraniuk, Robust 1-bit ompressive sensing via binary stable embeddings of sparse vetors, arXiv preprint arXiv:1104.3160, 2011. [52℄ J. Haboba, M. Mangia, R. Rovatti, and G. Setti, An arhiteture for 1-bit loalized ompressive sensing with appliations to eeg, in Biomedial Ciruits and Systems Conferene (BioCAS), 2011 IEEE. IEEE, 2011, pp. 137 140. Bibliography 127 [53℄ J. Haboba, M. Mangia, F. Pareshi, R. Rovatti, and G. Setti, A pragmati look at some ompressive sensing arhitetures with saturation and quantization, Emerging and Seleted Topis in Ciruits and Systems, IEEE Journal on, vol. 2, no. 3, pp. 443459, 2012. [54℄ J. Haboba, R. Rovatti, and G. Setti, Rads onverter: An approah for analog to information onversion, in Eletroni Ciruits and Systems (ICECS), 2012 IEEE International Conferene on. IEEE, 2012, pp. 4952. [55℄ R. Shreier and G. C. Temes, Understanding delta-sigma data onverters. IEEE press Pisataway, NJ, USA, 2005, vol. 74. [56℄ S. R. Norsworthy, R. Shreier, G. C. Temes et al., Delta-sigma data onverters: theory, design, and simulation. IEEE press New York, 1997, vol. 97. [57℄ J. C. Candy and G. C. Temes, Oversampling delta-sigma data onverters: theory, design, and simulation. IEEE press New York, 1992. [58℄ G. I. Bourdopoulos, A. Pnevmatikakis, V. Anastassopoulos, and T. L. Deliyannis, Delta-sigma modulators: modeling, design and appliations. World Sienti Publishing Company, 2003. [59℄ R. Shreier. (2009) Delta sigma toolbox. [Online℄. Available: www.mathworks.om/matlabentral/leexhange/19-delta-sigma-toolbox [60℄ The mathworks, in. http://www.mathworks.it/produts/matlab/ [Online℄. Available: [61℄ J. Tropp, J. Laska, M. Duarte, J. Romberg, and R. Baraniuk, Beyond nyquist: Efient sampling of sparse bandlimited signals, Information Theory, IEEE Transations on, vol. 56, no. 1, pp. 520 544, jan. 2010. [62℄ J. Laska, Z. Wen, W. Yin, and R. Baraniuk, Trust, but verify: Fast and aurate signal reovery from 1-bit ompressive measurements, Signal Proessing, IEEE Transations on, 2011, to appear. Bibliography 128 [63℄ M. E. Newman and G. T. Barkema, Monte arlo methods in statistial physis, Monte Carlo methods in statistial physis/MEJ Newman and GT Barkema. Oxford: Clarendon Press, 1999., vol. 1, 1999. [64℄ W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numerial Reipes with Soure Code CD-ROM 3rd Edition: The Art of Sienti Computing. Cambridge University Press, 2007. [65℄ Ibm ilog plex optimizer. [Online℄. Available: http://www01.ibm.om/software/integration/optimization/plex-optimizer/

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement