section 2 digital signal processing - semantic scholarthe field of digital signal processing grew...

Digital Signal Processing

Chapter 1 Digital Signal Processing Research Program

Chapter 2

Chapter 3

Advanced Telecommunications and SignalProcessing Program

Combined Source and Channel Coding forHigh-Definition Television

303

Section 2

304 RLE Progress Report Number 139

Chapter 1. Digital Signal Processing Research Program


Academic and Research Staff

Professor Alan V. Oppenheim, Professor Arthur B. Baggeroer, Professor Anantha P. Chandrakasan, Pro-fessor Gregory W. Wornell, Dr. Steven H. Isabelle, Giovanni Aliberti

Visiting Scientists and Research Affiliates

Dr. Bernard Gold, Dr. S. Hamid Nawab,1 Dr. James C. Preisig, Dr. Ehud Weinstein 2

Graduate Students

Chalee Asavathiratham, Richard J. Barron, Soosan Beheshti, John R. Buck, Brian Chen, Trym H. Eggen,Christoforos N. Hadjicostis, Warren M. Lam, J. Nicholas Laneman, Li Lee, Haralabos C. Papadopoulos,Jeffrey T. Ludwig, James M. Ooi, Wendi B. Rabiner, Matthew J. Secor, Alan Seefeldt, Andrew C. Singer,Shawn M. Verbout, Kathleen E. Wage, Alex Che-Wei Wang

Technical and Support Staff

Darla J. Chupp, Janice M. Zaganjori

1.1 Introduction

The field of digital signal processing grew out of theflexibility afforded by the use of digital computers inimplementing signal processing algorithms andsystems. It has since broadened into the use of avariety of both digital and analog technologies,spanning a broad range of applications, band-widths, and realizations. RLE's Digital Signal Pro-cessing Group carries out research on algorithmsfor signal processing and their applications.Current application areas of interest include signalenhancement and active noise cancellation;speech, audio and underwater acoustic signal pro-cessing; advanced beamforming for radar andsonar systems; and signal processing and codingfor wireless and broadband multiuser communica-tion networks.

In some of our recent work, we have developednew methods for signal enhancement and noisecancellation with single or multisensor measure-ments. We have also been developing newmethods for representing and analyzing fractalsignals. This class of signals arises in a widevariety of physical environments and also haspotential in problems involving signal design. Weare also exploring potential uses of nonlinear dyna-

mics and chaos theory of signal design and anal-ysis. Another research emphasis is on structuringalgorithms for approximate processing and succes-sive refinement.

In other research, we are investigating applicationsof signal ano array processing to ocean and struc-tural acoustics and geophysics. These problemsrequire the combination of digital signal processingtools with a knowledge of wave propagation todevelop systems for short time spectral analysis,wavenumber spectrum estimation, source localiza-tion, and matched field processing. We emphasizethe use of real-world data from laboratory and fieldexperiments such as the Heard Island Experimentfor Acoustic Monitoring of Global Warming andseveral Arctic acoustic experiments conducted onthe polar ice cap.

A major application focus of the group involvessignal processing and coding for wireless multiusersystems and broadband communication networks.Specific interests include commercial and militarymobile radio networks, wireless local area networksand personal communication systems, digital audioand television broadcast systems, and multimedianetworks. Along with a number of other directions,we are currently exploring new code-division

1 Associate Professor, Boston University, College of Engineering, Boston, Massachusetts.

2 Department of Electrical Engineering, Systems Division, Faculty of Engineering, Tel-Aviv University, Israel; adjunct scientist, Depart-ment of Applied Ocean Physics and Engineering, Woods Hole Oceanographic Institution, Woods Hole, Massachusetts.

305


multiple-access (CDMA) strategies, new techniquesfor exploiting antenna arrays in wireless systems,and new methods for modeling and management oftraffic in high-speed packet-switched networks.

Much of our work involves close collaboration withthe Woods Hole Oceanographic Institution, MITLincoln Laboratory, and a number of high tech-nology companies in the Boston Area.

1.2 Model-Based Signal Enhancement

Sponsors

Lockheed Sanders, Inc.Contract BZ4962

U.S. Army Research LaboratoryGrant QK-8819

U.S. Navy - Office of Naval ResearchGrant N00014-93-1-0686

Project Staff

Richard J. Barron, Professor Alan V. Oppenheim

A common signal processing task is the estimationof an information-bearing signal from a distortedversion of the waveform. Traditional solutions, suchas the Wiener filter, assume a particular stochasticdescription for the signal source and channel dis-tortion, and minimize an error criterion accordingly.In this project, we explore the use of signal featuresbased on a model of the signal source as a methodof signal estimation. A signal is often classified bythe source from which it originates; speech ornatural images are examples of certain classes ofsignal. Often there exist models of a particularsignal source described by a few key parameters,or features, that capture much of the behavior ofthe signal. This feature information derived fromthe clean signal is often available to the processorthat is enhancing the corrupted waveform. Wehave derived algorithms that exploit signal featureinformation to enhance signals from a variety ofsources.

1.3 Multipass Receivers for

Spread-Signature CDMA Systems

Sponsors

National Science FoundationGrant MIP95-02885

U.S. Navy - Office of Naval ResearchGrant N00014-93-1-0686Grant N00014-95-1-0834Grant N00014-96-1-0930

Project Staff

Soosan Beheshti, Professor Gregory W. Wornell

Spread-signature code-division multiple-access(CDMA) systems were recently introduced as anattractive alternative to conventional CDMA systemsfor use in time-varying multipath environments.Using long signatures in an overlapped manner forsuccessive symbols, spread-signature CDMA canachieve a substantial temporal diversity benefit.Furthermore, the broadband nature of the signa-tures allows an additional spectral diversity benefitto be simultaneously realized.

In this research, computationally efficient multipassdemodulation and decoding algorithms are devel-oped for use in receivers with these systems.These algorithms efficiently suppress both inter-symbol and interuser (multiple-access) interferenceto achieve a substantial diversity benefit and goodnear-far resistance characteristics. Moreover, it isshown that relatively few iterations are required forconvergence to typical target bit-error rates.Several other aspects of the performance of thealgorithms are also explored.

1.4 Single Mode Excitation in theShallow Water Acoustic Channel UsingFeedback Control

Sponsor

U.S. Navy - Office of Naval ResearchGrant N00014-95-1-0362Grant N00014-93-1-0686

Project Staff

John R. Buck, Professor Alan V. Oppenheim

The shallow water acoustic channel supports far-field propagation in a discrete set of modes. Oceanexperiments have confirmed the modal nature ofacoustic propagation, but no experiment has suc-cessfully excited only one of the suite of mid-frequency propagating modes propagating in acoastal environment. The ability to excite a singlemode would be a powerful tool for investigatingshallow water ocean processes. A feedback controlalgorithm incorporating elements of adaptive esti-mation, underwater acoustics, array processing andcontrol theory to generate a high-fidelity singlemode is presented. This approach also yields acohesive framework for evaluating the feasibility ofgenerating a single mode with given array geom-


etries, noise characteristics and source power limi-tations. Simulations and laboratory waveguideexperiments indicate the proposed algorithm holdspromise for ocean experiments.

1.5 Coding and Modulation of AnalogData for Transmission over Broadcastand Fading Channels

Sponsors

National Defense Science and EngineeringFellowship

U.S. Air Force - Office of Scientific ResearchGrant F49620-96-1-0072


Project Staff

Brian Chen, Professor Gregory W. Wornell

In many communication applications, the informa-tion to be transmitted over the channel of interest isinherently analog (i.e., continuous-valued) in nature.A traditional digital approach for transmitting suchdata involves appropriately quantizing the sourcedata and encoding the quantized data using a suit-ably designed channel code so that the quantizeddata can be recovered with arbitrarily low probabilityof error. Shannon's source-channel separationtheorem is frequently invoked to argue that perfor-mance need not be sacrificed using such anapproach. However, for many important classes ofchannels that arise in practice, Shannon's theoremdoes not apply and in fact, performance is neces-sarily sacrificed using this digital approach. Such isthe case, for example, when the signal-to-noiseratio (SNR) is unknown at the transmitter, or equiv-alently, in broadcast scenarios where there are mul-tiple receivers with different SNRs, as well as inlow-delay systems operating in the presence oftime-selective fading due to multipath propagation.

In these kinds of settings, which arise in, forexample, a variety of wireless communicationsystems, separate source and channel coding isinherently suboptimum. Such digital approachesare inadequate because their performance dependscrucially on being able to choose the propernumber of quantization levels, which in turndepends on there being a specific target SNR.Motivated by these observations, in this research


we explore efficient analog coding strategies forscenarios precisely of this type.

The properties of chaotic dynamical systems makethem useful for channel coding for a variety of prac-tical communication applications. We have devel-oped novel analog error-correcting codes that arepotentially useful in such applications, examples ofwhich are communication over broadcast channelsand low-delay communication in time-varying fadingenvironments. These systematic analog codes aregenerated from iterations of a nonlinear state spacesystem governed by chaotic dynamics, with theanalog message embedded in the initial state.Within this class are practical codes having compu-tationally very efficient recursive receiver structuresand important performance advantages over con-ventional codes. We are in the process of devel-oping a method for generalizing and optimizingsuch codes. Indeed, we envision a general frame-work for developing a broader error-correctioncoding theory that encompasses both the theory ofmodern digital codes and classical analog modu-lation techniques.

1.6 Underwater AcousticCommunication

Sponsor

U.S. Navy - Office of Naval ResearchGrant N00014-95-1-0362Grant N00014-93-1-0686

Project Staff

Trym H. Eggen, Professor Arthur B. Baggeroer

Underwater acoustic coherent communication ispossible by conventional communication systemsonly when the underwater communication channelis especially simple. One problem is the frequencydispersion of the channel, and by using a specificlinear time variant model analysis of conventionalreceivers as well as development of a new receiverhas been carried out. The receivers have beentested on real data from the ocean, and emphasishas been on communication channels where theDoppler spread is more severe than the delayspread which is only a subset of all the existingunderwater communication channels. Some phys-ical conditions for these channels to exist havebeen derived, and examples from the real oceanhas been found. Many conventional receivers, asused in other communication areas as cellular radioand indoor wireless, are not well suited for this typeof channel.

307


1.7 Algebraic and ProbabilisticStructure in Fault-Tolerant Computation

Sponsors


U.S. Navy - Office of Naval ResearchGrant N00014-93-0686

Project Staff

Christoforos N. Hadjicostis, Professor George C.Verghese

The traditional approach towards fault-tolerant com-putation has been modular redundancy. Althoughuniversal and simple, modular redundancy is inher-ently expensive and inefficient in its use ofresources. Recently developed algorithm-basedfault tolerance (ABFT) techniques offer more effi-cient fault coverage, but their design is specific toeach application. A particular class of ABFT tech-niques involves the design of arithmetic codes thatprotect elementary computations. For the case ofcomputations that can be represented as operationsin a group, the recent doctoral thesis by Beckmann 3

has shown how to obtain a variety of useful resultsand systematic constructive procedures.

In our research so far we have been able to gener-alize this work to the case of computations occur-ring in semigroups and semirings4 and to outline aprocedure that reflects such algebraically-basedABFT design into hardware. Currently, we areexploring extensions of our approach to sequencesof computations associated with the evolution ofdynamic systems in particular algebraic settings,such as linear systems over groups, or rings, orsemirings, or finite automata and discrete-eventsystems. Along these lines, we have obtained anilluminating characterization of all possible redun-dant linear time-invariant (LTI) state-space embed-dings of a given LTI state-space model. We alsointend in future work to fold probabilistic models forfailures and errors into the design and analysis ofABFT systems.

1.8 Multiscale Signal Processing withFractal Renewal Processes

Sponsors



Project Staff

Warren M. Lam, Professor Gregory W. Wornell

Point processes with fractal characteristics have apotentially important role to play in the modeling ofnumerous natural and man-made phenomena,ranging from the distribution of stars and planets inthe universe, to the occurrence of transmissionerrors in communication channels and traffic over anumber of packet-switched networks. However, incontrast to fractal waveforms, which have beenexplored in considerable depth,5 development ofefficient algorithms for synthesizing, analyzing, andprocessing fractal point processes has generallyproven difficult, largely due to the lack of an ade-quate mathematical framework. In this work, weintroduce a novel multiscale representation forfractal point processes and apply it to a number ofpractical signal processing problems involving suchpoint processes.

Our study of fractal point processes is focused pri-marily on an important subclass called fractalrenewal processes which possess a sense of sta-tionarity as well as self-similarity. Recently, wehave developed a multiscale representationwhereby a fractal renewal process is viewed as therandom mixture of a multiscale family of constituentPoisson processes. 6 Exploiting existing efficientalgorithms for Poisson processes, this frameworkhas appeared promising for the study of fractalrenewal processes. Indeed, based on the multi-

3 P.E. Beckmann, Fault-Tolerant Computation Using Algebraic Homomorphisms, RLE TR-580 (Cambridge: MIT Research Laboratoryof Electronics, 1993).

4 C.N. Hadjicostis, Fault-Tolerant Computation in Semigroups and Semirings, RLE TR-594 (Cambridge, MIT Research Laboratory ofElectronics, 1995).

5 G.W. Wornell, "Wavelet-based Representations for the Family of Fractal Processes," Proc. IEEE 81(10): 1428-1450 (1993).

6 W.M. Lam, and G.W. Wornell, "Multiscale Synthesis and Analysis of Fractal Renewal Processes," Proceedings of the Sixth IEEEDSP Workshop, Yosemite, California, October 1, 1994.


scale framework, we have developed an efficientalgorithm for synthesizing fractal renewal pro-cesses. In addition, we have successfully appliedthis framework to several practical signal pro-cessing problems including estimation of the fractaldimension of a point process and recovery of afractal renewal process from corrupted measure-ments. Application of this framework to other prac-tical problems is currently being investigated.

1.9 Estimation and Equalization of

Wireless Fading Channels

Sponsors

National Science FoundationGraduate Research FellowshipGrant MIP 95-02885


Project Staff

J. Nicholas Laneman, ProfessorWornell

Gregory

A central issue in the wireless communicationssetting is the problem of signal fading. Due to mul-tiple propagation paths, many copies of the trans-mitted signal arrive at the receiver antenna, eachwith a given attenuation level and phase shift.When the receiver antenna is set in motion, as isusually the case in such applications as cellulartelephony, the received power level fluctuates sincethe multipath components add constructively ordestructively. Because of this fading characteristic,wireless channels exhibit dramatically poorer bit-error performance than traditional additive whiteGaussian noise channels when using uncodedtransmissions.

Recent work7 has suggested a technique known asspread-response precoding for combating signalfading found in wireless links. The idea behind thissort of precoding is to distribute the energy of eachsymbol in time to achieve the average effect of thechannel rather than the instantaneous fade. A keyelement of the receiver is an equalizer whichessentially inverts the effect of the fading channel.


In our present work, we focus on a realisticequalizer structure derived from channel estimates.

We are evaluating a joint state and parameters esti-mator for the channel response based on Kalmanfiltering ideas and the Estimate Maximize (EM)algorithm. We are studying how to use estimatesof the channel or channel inverse to best equalizethe received signal, and how to take advantage ofthe special properties of the transmitted signal toallow the algorithm to perform blindly.

A flexible set of hardware has been assembled fordemonstrating a variety of these signal processingalgorithms indoors. Special purpose analog hard-ware has been built for modulating basebandsignals to radio frequencies and back. The labora-tory includes digital-to-analog (D/A) and analog todigital (A/D) converters for converting the basebandsignal into discrete-time. Finally, four digital signalprocessors (DSPS) are used to perform the pre-coding and equalization of the baseband signals.The high computational power of the DSPS allowus to implement complex algorithms in real-time.

1.10 Properties of ApproximateParks-McClellan Filters

Sponsors




Project Staff

Li Lee, Professor Alan V. Oppenheim

For digital signal processing applications with real-time or low-power constraints, it is often desirableto use algorithms whose output quality can beadjusted depending on the availability of resourcessuch as time or power. For this reason, recentlythere has been increased interest in approximatesignal processing algorithms whose intermediateresults represent successively better approxi-mations to the desired solution. We have observedempirically that each coefficient in aParks-McClellan filter converges to a steady state

7 G.W. Wornell, "Spread-Response Precoding for Communication over Fading Channels," IEEE Trans. Info. Theory 42(2): 488-501(1996); G.W. Wornell, "Spread-Signature CDMA: Efficient Multiuser Communication in the Presence of Fading," IEEE Trans. Info.Theory 41(5): 1418-1438 (1995).

309


value as the filter length increases. This suggeststhe possibility of obtaining filters that are nearoptimal while "re-using" filter coefficients fromshorter filters in the design of longer filters. In thecontext of approximate processing this then allowsa filtering operation to be done in stages. A paperdemonstrating this observation and examining someof its implications will be presented in ICASSP '97.

1.11 Distributed Signal Processing

Sponsors




Project Staff

Li Lee, Professor Alan V. Oppenheim

We have initiated a project directed at implementingsignal processing algorithms in a distributednetwork environment with unpredictable anddynamically changing resources. Our initialapproach is to define a hierarchical set of signalprocessing modules with rules for describing thehierarchy. Our focus is on the signal processingaspects rather than the network aspects and conse-quently we represent the network issues statis-tically. Work on this project this year has involveddefining the primitives and hierarchy and imple-menting a first stage simulation of the structure.

1.12 Approximate Signal Processing

Sponsors

Lockheed Sanders, Inc.Grant N00014-93-1-0686Contract BZ4962

U.S. Army Resarch LaboratoryGrant QK-8819


Project Staff

Jeffrey T. Ludwig, Dr. S. Hamid Nawab

It is increasingly important to structure signal pro-cessing algorithms and systems to allow for flexi-bility in trading off between the accuracy oroptimality of their results and their utilization of

resources such as time, power, bandwidth, andphysical space. Approximate signal processingalgorithms are designed to incorporate suchtradeoffs.

Our recent research indicates the enormous poten-tial of approximate signal processing algorithms.These results show progress toward the ultimateobjective of developing, within the context of signalprocessing and design, a more general and rig-orous framework for utilizing and expanding approx-imate processing concepts and methodologies.

We have successfully applied approximate pro-cessing concepts to the area of low-power signalprocessing. Techniques for reducing power con-sumption have become important due to thegrowing demand for portable multimedia devices.We have developed an approach to the design oflow-power frequency-selective digital filters basedon the concepts of adaptive filtering and approxi-mate processing. The technique uses a feedbackmechanism in conjunction with well-known imple-mentation structures for FIR and IIR digital filters.Our algorithm is designed to reduce the totalswitched capacitance by dynamically varying thefilter order based on signal statistics. A factor of 10reduction in power consumption over fixed-orderfilters has been demonstrated for the filtering ofspeech signals.

Our aim is to extend the development of formalstructures for using approximate processing con-cepts in designing novel signal processing algo-rithms to areas such as time-frequency analysis,adaptive beamforming, and image coding.

1.13 New Techniques forCommunication with Feedback

Sponsor


Project Staff

James M. Ooi, Professor Gregory W. Wornell

Many communication links are inherently bidirec-tional, supporting the two-way exchange of voice,video, and other data between a pair of users. Insuch scenarios, a natural feedback path exists foreach user's transmission, and as is well known thisfeedback path can generally be exploited toimprove overall system performance in a variety of


ways. Indeed, almost all duplex communicationlinks in widespread use today exploit the availabilityof feedback, usually via an automatic repeat-request (ARQ) or related protocol.

In a typical feedback communication system, thetransmitter sends data to a receiver over a noisyforward channel, and receives information aboutwhat the receiver actually observes via a feedbackchannel. The feedback path is often a relativelynoise-free link, particularly when the receiver isfeeding back information to the transmitter either atcomparatively low rate or high power. Powerasymmetry, specifically, arises naturally in anumber of existing applications, including, e.g.,Earthbound transmissions in satellite systems, andreverse (mobile-to-base) link transmission in a cel-lular mobile radio network. More generally, thenoise-free feedback channel model is a good onefor communication between portable, low-powertransmitters and stationary, high-power receivers ina host of emerging systems for providing wirelesspersonal communication services such as cellulartelephony, paging, wireless local area networks,and wireless private branch exchanges.

While it is well established that feedback does notincrease the capacity of discrete memoryless chan-nels, it is well known that noise-free feedback canbe used to substantially lower complexity andincrease reliability of communication in practice.

We have developed a powerful framework for low-complexity, high-reliability communication overchannels with feedback. We have used the frame-work to develop capacity-achieving coding schemesfor arbitrary discrete memoryless and finite-statechannels with noise-free feedback. We have alsodeveloped a universal communication scheme inwhich neither the transmitter nor the receiver needknow the channel statistics to communicate reliably.We are exploring how the framework can beapplied to coding for multiple-access channels aswell as what the framework tells us about coding forchannels with noisy feedback.


1.14 Analysis and Applications ofSystems Exhibiting StochasticResonance

Sponsors



Project Staff

Haralabos C. Papadopoulos, Professor Gregory W.Wornell

Stochastic resonance is a phenomenon encount-ered in certain nonlinear systems when driven bynoisy information-bearing signals. Specifically,increasing the input noise level in these systemsoften results in an enhancement of the information-bearing signal response, reflected for example, asoutput signal-to-noise ratio (SNR) enhancement, oras improved detection/estimation performance.Such systems are therefore appealing candidatesfor use in a variety of engineering contexts. Interms of signal analysis, such systems constitutepotentially useful models for natural phenomenasuch as the regularity of appearance of earth's iceages," as well as for detection mechanisms incertain species, such as predator sensing bycrayfish.9 In terms of signal synthesis, the inducedsignal enhancement renders them attractive in anumber of applications in signal communication andprocessing. In order to exploit stochastic reso-nance in such applications, there is a need for toolsto analyze these systems in the presence of variousforms and degrees of distortion.

One of the main directions of our research istowards the development of novel techniques foranalysis of dynamical systems exhibiting stochasticresonance, and considering their viability in varioussignal processing and communication contexts. Inaddition, the research explores the phenomenon ofstochastic resonance in the context of generalsignal processing problems which includes signaldetection, classification, and enhancement.

8 R. Benzi, A. Sutera, and A. Vulpiani, "The Mechanism of Stochastic Resonance," J. Phys. A14: L453- L457 (1981).

9 J.K. Douglass, L. Wilkens, E. Pantazelou, and F. Moss, "Noise Enhancement of Information Transfer in Crayfish Mechanoreceptorsby Stochastic Resonance," Nature 365: 337-340 (1993).

311


1.15 Modeling and Design ofApproximate Digital Signal Processorsand Approximate DSP Networks

Sponsors



Project Staff

Matthew J. Secor, Professor George C. Verghese,Professor Alan V. Oppenheim

This research investigates the area of approximatedigital signal processing, studying the interactionsamong, and collective performance of multipleapproximate processors or processes. Theapproach we take is to study networks of digitalsignal processing (DSP) modules whose parame-ters and functionality can be varied to adapt tochanges in specifications and constraints. Thespecifications that we focus on are ones thatprovide metrics or tolerances for various features ofinput and output quality (such as time andfrequency-resolution, quantization, probability oferror), thus allowing the individual DSP modules tocarry out what has come to be known as approxi-mate processing. The flexibility allowed by approxi-mate processing can be critical to accommodatingconstraints placed on a system comprised ofapproximate processors. The constraints of interestinvolve such resources as cost, time, power,memory, and inter-processor communication. Ourresearch includes modeling, the development ofresource allocation and scheduling schemes, andsimulation/testing. Another major area of ourresearch is the development of new DSP algorithmsand modification of existing DSP algorithms suchthat they exhibit incremental refinement propertiesand thus can be incorporated into larger approxi-mate processing systems.

1.16 Sinusoidal Analysis Synthesis

Sponsors




Project Staff

Alan Seefeldt, Professor Alan V. Oppenheim

In our research, the use of sinusoidal analysis/syn-thesis (SAS) for the enhancement of noise cor-rupted speech is being explored. SASapproximates a digital speech waveform as a finitesum of time varying sinusoidal tracks. For the pur-poses of enhancement, the idea is to extract fromthe spectrum of the corrupted speech sinusoidaltracks that correspond to the speech alone. Inorder to attain an upper bound on the performanceof SAS enhancement, the original uncorruptedspeech is used as an aid in this track extractionprocedure. Various processing techniques, such asspectral subtraction and amplitude smoothing, arethen applied to the extracted tracks to reduce anyremaining noise residual. The quality of speechenhanced with this SAS technique is being com-pared to that of previously developed enhancementprocedures.

1.17 Signal Processing andCommunication with Solitons

Sponsors




Project Staff

Andrew C. Singer, Professor Alan V. Oppenheim

Traditional signal processing algorithms rely heavilyon models that are inherently linear. Such modelsare attractive both for their mathematical tractabilityand their applicability to the rich class of signalsthat can be represented with Fourier methods.Nonlinear systems that support soliton solutionsshare many of the properties that make linearsystems attractive from an engineering standpoint.Although nonlinear, these systems are solvablethrough inverse scattering, a technique analogousto the Fourier transform for linear systems. Solitonsare eigenfunctions of these systems which satisfy anonlinear form of superposition and display richsignal dynamics as they interact. By using solitonsfor signal synthesis, the corresponding nonlinearsystems become specialized signal processorswhich are naturally suited to a number of complexsignal processing tasks. Specific analog circuitscan generate soliton signals and can be used as


natural multiplexers and demultiplexers in a numberof potential soliton-based wireless communicationapplications. These circuits play an important rolein investigating the effects of noise on solitonbehavior. Finally, the soliton signal dynamics canprovide a mechanism for decreasing transmittedsignal energy while enhancing signal detection andparameter estimation performance.

1.18 A Parametric Framework for

Non-Gaussian Signal Processing

Sponsors




Project Staff

Shawn M. Verbout, Professor Alan V. Oppenheim

Recent work has been aimed at developing aframework for analyzing and processing discrete-time non-Gaussian autoregressive (AR) randomsignals. As in the conventional linear-Gaussian ARtime series model, the signal equation consists oftwo components: a regression component and adriving noise component. In the newly proposedmodel, the regression parameters are taken to beconstant, but the parameters of the Gaussiandriving noise are subject to abrupt changes overtime that occur according to a finite-state Markovchain. Important problems in statistical signal pro-cessing that are being analyzed under this newmodel include (1) identification of unknown signalparameters, and (2) filtering of signals (with knownparameters) in additive noise.

The classical linear AR time series model has longbeen used for the statistical analysis of exper-imental observations. Though the AR model is arather restricted version of the general linear model,it has gained wide acceptance in disciplines suchas economics, biology, geophysics, and engineeringfor several reasons: (1) it is inherently simple, bothmathematically and computationally; (2) it iscapable of representing a wide range of correlationpatterns with a relatively small number of parame-ters; and (3) it is ideally suited to representing manynatural phenomena, which tend to be stronglytemporally correlated and thus exhibit sharp spec-tral peaks, particularly in low-frequency bands.


In many practical situations, the aspect of the con-ventional AR time series model that limits its perfor-mance is not the assumption that the underlyingprocess is autoregressive; rather, it is the assump-tion that the process is Gaussian. Although theGaussian model is adequate in certain cases, it isnot flexible enough to represent populations thatare, for example, heavy-tailed or strongly skewed.In fact, since the Gaussian pdf falls off sharply atvalues that are even moderately far from the mean,the Gaussian assumption is clearly inappropriatewhen the driving noise is characterized by grossfluctuations in amplitude, sudden random bursts, orfrequently occurring spikes or impulses. Thesekinds of driving inputs are typically encountered inproblems such as speech processing, explorationseismology, low-frequency communication systems,and underwater signal detection.

In recent research, we have considered a modifiedversion of the standard AR time series model inwhich the driving noise is no longer taken to bei.i.d. and Gaussian. Instead, we assume that thedriving noise is characterized, at each time step, byone of a finite set of underlying regimes, and that itswitches randomly from one regime to anotheraccording to a Markov chain with constant transitionprobabilities. At a given time step, the regime ofthe driving process is selected by a discreteMarkovian regime variable, and the input noisesample at that time is drawn from a Gaussian pdfwhose mean and variance are dependent on thecurrent value of the regime variable. This modelincludes the case in which the driving noise is ani.i.d. Gaussian sequence; clearly, this case isrepresented by letting the Markov chain have onlyone state. However, the model is much moregeneral than the conventional model in that it allowsfor a wide range of densities and temporal corre-lation structures for the input samples.

Although the proposed AR model will undoubtedlyyield a more accurate representation of many phys-ical systems, it does not enjoy the mathematicalsimplicity associated with the conventionalGaussian model. Moreover, widely used methodsthat are based on the Gaussianassumption-methods such as least squares (forparameter estimation) or Kalman filtering (for signalestimation)-can not be expected to yield optimal oreven near-optimal performance for many signalsthat are accurately characterized by the proposedmodel. Thus, the main challenges of this work liein the creation of new algorithms for signal andparameter estimation, and more generally in thedevelopment of a unifying parametric framework forhandling non-Gaussian signals.

313


1.19 Array Processing Techniques forBroadband Mode Estimation and ModalTomography

Sponsor


Project Staff

Kathleen E. Wage, Professor Arthur B. Baggeroer

This research focuses on advanced array pro-cessing techniques for underwater applicationssuch as ocean acoustic tomography. Specifically,the goal of this research is to develop a signal pro-cessing framework for estimating the normal modedecomposition of low-frequency, broadband recep-tions. Modal representations are useful in seismo-acoustic modeling since they are directly related toa solution of the wave equation and because modeamplitudes and phases contain valuable informationabout the source and the propagation medium.Modal signal processing is an important area ofresearch for several reasons. First, verification ofmode propagation and scattering theories requiresaccurate estimates of the field to be determinedfrom measurements made using mode-resolvingarrays. A second motivation is that acoustic tomog-raphy applications rely on very precise measure-ments of mode travel times. The AcousticThermometry of Ocean Climate (ATOC) project isan ongoing research project designed to demon-strate that tomographic systems can be used tomeasure ocean climate variability over ranges of3,000 to 10,000 km in the North Pacific. Incorpo-rating normal mode data from long-range propa-gation studies into an inversion for environmentalparameters requires a thorough understanding ofthe resolution of the underlying estimators. Thepurpose of this research is to develop and evaluatemethods of processing modal signals for projectssuch as ATOC.

Typically, acoustic modes are used for describingand analyzing the temporal/spatial structure ofnarrowband signals. The overall objective of thisresearch is to develop a general framework forbroadband modes and, using this framework, todevelop algorithms for detecting and analyzingbroadband modal signals. Specifically, there aretwo related signal processing problems that thisresearch will address. The first is the estimation ofa time series of modal excitation coefficients basedon measurements from a hydrophone array. Pre-vious work on mode estimation has concentratedprimarily on modal decompositions for narrowband

sources, often in the context of source localizationproblems. Studies of broadband sources, such asthose used for tomography, are rather limited. Thisresearch will define a framework for broadbandmode estimation which can be used to explore thetime/frequency resolution tradeoffs inherent in theprocessing of transient or non-stationary signals.Within this framework research efforts will focus ondeveloping robust data-adaptive methods for modalbeamforming, using concepts similar to those ofmatched field processing. Performance of conven-tional mode estimators is directly linked to how wellthe hydrophone array samples the water column.The intent of the robust techniques is to mitigatethe sensitivity of the estimator to single-sensor fail-ures and to aid in the design of shorter arrays withminimal loss in resolution capabilities. Mode esti-mation is closely related to adaptive beamformingand linear inverse theory, thus the results of thisresearch may be relevant to a broader class ofsignal processing problems.

The second issue this research will address is thedetection of broadband pulse arrivals in the acou-stic modes and the estimation of the associatedtravel times. Although numerous researchers haveproposed using modal group delay perturbations fortomographic inversions, very few have examinedthe signal processing required to determine thearrival times. An objective of the proposedresearch is to develop optimal receivers for themodal signals and to characterize their perfor-mance. The receiving strategies must account forthe dispersive nature of the ocean waveguide andextend to channels that have random coupling andfading due to internal waves.

1.20 Multiscale State-Space Algorithmsfor Processing 1/f Signals

Sponsors



Project Staff

Alex Che-Wei Wang, Professor Gregory W. Wornell

Natural landscapes, noise in electrical devices, andfluctuations in the stock market are among theextraordinary variety of phenomena that exhibitfractal structure. The prevalence of fractal geom-etry in nature indicates the value of algorithms forprocessing fractal signals.


The 1/f processes are an important class of fractalrandom processes. Due to the wide range of phe-nomena modeled as 1/f processes, many usefulapplications based on processing these signals canbe envisioned. Algorithms for predicting futurevalues of a 1/f signal given observations of theprocess over a finite time interval could have appli-cations in economic forecasting, for instance.

This research develops signal processing algo-rithms involving 1/f processes both as the signal ofinterest and as an obscuring noise process. Wedevelop a linear time-invariant (LTI) multiscalestate-space model for discrete 1/f processes. Thismodel leads naturally to computationally efficientalgorithms for processing 1/f signals with traditionallinear LTI state-space methods such as the Kalmanfilter and Kalman smoother.

1.21 Publications

1.21.1 Journal Articles

Buck, J.R., J.C. Preisig, M. Johnson, and J.Catipovic. "Single Mode Excitation in theShallow Water Acoustic Channel Using Feed-back Control." IEEE Oceanic Eng. Forth-coming.

Ludwig, J.H., S.H. Nawab, and A.P. Chandrakasan."Low Power Digital Filtering Using ApproximateProcessing." IEEE Solid-State Circuits (31)3:395-400 (1996).

Ooi, J.M., S.M. Verbout, J.T. Ludwig, and G.W.Wornell. "A Separation Theorem for PeriodicSharing Information Patterns in DecentralizedControl." IEEE Trans. Automat. Contr. Forth-coming.

Winograd, J., J. Ludwig, H. Nawab, A. Chandra-kasan, and A.V. Oppenheim. "ApproximateSignal Processing." VLSI Signal Process.Forthcoming.

Wornell, G.W. "Emerging Applications of MultirateSignal Processing and Wavelets in Digital Com-munications." Proc. IEEE Special Issue onApplications of Wavelets (84)4: 586-603 (1996).

Wornell, G.W. "Spread-Response Precoding forCommunication over Fading Channels." IEEETrans. Info. Theory (42)2: 488-501 (1996).


1.21.2 Conference Papers

Chen, B., and G.W. Wornell. "Efficient ChannelCoding for Analog Sources using ChaoticSystems." Proceedings of Globecom '96,London, England, November 18-22, 1996.

Isabelle, S.H., and G.W. Wornell. "Recursive Multi-user Equalization for CDMA Systems in FadingEnvironments." Proceedings of the AllertonConference on Communication, Contr., andComputing, Allerton, Illinois October 1996.

Lam, W., and G.W. Wornell. "Multiscale Analysis ofFractal Point Processes and Queues." Pro-ceedings of ICASSP '96, Atlanta, Georgia, May7-10, 1996.

Lee, L., and R.C. Rose. "Speaker NormalizationUsing Efficient Frequency Warping Procedures."Proceedings of ICASSP '96, Atlanta, Georgia,May 7-10, 1996.

Ludwig, J.T., S.H. Nawab, and A.P. Chandrakasan."Convergence Results on Adaptive ApproximateFiltering." Proceedings of the International Sym-posium on Optical Science, Engineering andInstrumentation, Denver, Colorado, August1996.

Narula, A., M.D. Trott, and G.W. Wornell."Information-theoretic Analysis of Multiple-Antenna Transmission Diversity for FadingChannels." Proceedings of the International Sym-posium on Information Theory and Applications,Victoria, Canada, September 1996.

Ooi, J.M., and G.W. Wornell. "DecentralizedControl of a Multiple Access Broadcast Channel:Performance Bounds." Proceedings of the Inter-national Conference Decentralized Control,Japan, December 1996.

Oppenheim, A.V., K. Cuomo, and R. Barron."Channel Equalization for Self-SynchronizingChaotic Systems." Proceedings of ICASSP '96,Atlanta, Georgia, May 7-10, 1996.

Papadopoulos, H.C., and G.W. Wornell. "A Classof Stochastic Resonance Systems for SignalProcessing Applications." Proceedings ofICASSP '96, Atlanta, Georgia, May 7-10, 1996.

Ramamurthy, M., S.H. Nawab, J.M. Winograd, andB.L. Evans. "Integrated Numeric and SymbolicSignal Processing Using a HeterogeneousDesign Environment." Proceedings of the Inter-national Symposium on Optical Science, Engi-

315


neering and Instrumentation, Denver, Colorado,August 1996.

Singer, A.C. "Detection and Estimation of SolitonSignals." Proceedings of ICASSP '96, Atlanta,Georgia, May 7-10, 1996.

Winograd, J.M., S.H. Nawab, and A.V. Oppenheim."FFT-Based Incremental Refinement of Subop-timal Detection." Proceedings of ICASSP '96,Atlanta, Georgia, May 7-10, 1996.

Winograd, J.M., J. Ludwig, S.H. Nawab, A.V.Oppenheim, and A. Chandrakasan. "FlexibleSystems for Digital Signal Processing." Pro-ceedings of the AAAI Fall Symposium on Flex-ible Computation, Cambridge, Massachusetts,November 9-11, 1996.

Wornell, G.W., and M.D. Trott. "Signal ProcessingTechniques for Efficient Use of Transmit Diver-sity in Wireless Communications." Proceedingsof ICASSP '96, Atlanta, Georgia, May 7-10,1996.

1.21.3 Technical Reports

Chen, B. Efficient Communication over AdditiveWhite Gaussian Noise and Intersymbol Interfer-ence Channels Using Chaotic Sequences. RLETR-598. Cambridge: MIT Research Laboratoryof Electronics, 1996.

Ooi, J.M., and G.W. Wornell. Fast Iterative Tech-niques for Feedback Channels RLE TR-613.Cambridge: MIT Research Laboratory of Elec-tronics, 1996.

Singer, A.C. Signal Processing and Communica-tion with Solitons. RLE TR-599. Cambridge:MIT Research Laboratory of Electronics, 1996.

Verbout, S.M., J.M. Ooi, J.T. Ludwig, and A.V.Oppenheim. Parameter Estimation for Auto-regressive Gaussian-Mixture Processes: TheEMAX Algorithm. RLE TR-611. Cambridge:MIT Research Laboratory of Electronics, 1996.

1.21.4 Books

Buck, J.R., M.M. Daniel, and A.C. Singer. Com-puter Explorations in Signals and Systems UsingMatlab. Upper Saddle River, New Jersey:Prentice Hall, 1997.

Oppenheim, A.V., A.S. Willsky, and S.H. Nawab.

Signals and Systems. 2d ed. Upper SaddleRiver, New Jersey: Prentice Hall, 1997.

1.21.5 Theses

Asavathiratham, C. Digital Audio Filters DesignUsing Warped Filters. M. Eng. thesis. Dept. ofElectr. Eng. and Comput. Sci., MIT, 1996.

Barron, R.J. Channel Equalization for Self-Synchronizing Chaotic Systems. M. Eng. thesis.Dept. of Electr. Eng. and Comput. Sci., MIT,1996.

Beheshti, S. Techniques for Enhancing the Perfor-mance of Communication Systems EmployingSpread-Response Precoding. M. Eng. thesis.Dept. of Electr. Eng. and Comput. Sci., MIT,1996.

Buck, J.R. Single Mode Excitation in the ShallowWater Acoustic Channel Using FeedbackControl. Ph.D. diss., Dept. of Electr. Eng. andComput. Sci., MIT, 1996.

Chen, B. Efficient Communication over AdditiveWhite Gaussian Noise and Intersymbol Interfer-ence Channels Using Chaotic Sequences. M.Eng. thesis. Dept. of Electr. Eng. and Comput.Sci., MIT, 1996.

Lee, L. A Frequency Warping Approach to SpeakerNormalization. M. Eng. thesis. Dept. of Electr.Eng. and Comput. Sci., 1996.

Singer, A.C. Signal Processing and Communica-tion with Solitons. Ph.D. diss. Dept. of Electr.Eng. and Comput. Sci., MIT, 1996.

1.22 Network-Driven Motion Estimationfor Wireless Video Terminals

Sponsors

National Science FoundationGraduate Fellowship

U.S. Army Research Laboratory/ARL AdvancedSensors Federated Lab ProgramContract DAAL01-96-2-0001

Project Staff

Professor Anantha P. Chandrakasan, Wendi B.Rabiner

Video is becoming an integral part of many portabledevices such as wireless cameras, personal com-



Figure 1. Wireless video terminals in a networked environment.

munication devices, and video cellular phones.Due to the massive amount of data contained invideo signals and the limited bandwidth of thewireless channel, developing compression tech-niques for these applications is extremely important.Conventional video systems use some form ofscene motion estimation/motion compensation atthe encoder to reduce the temporal correlationinherent in most video signals and hence achievehigh compression ratios. However, since mostmotion estimation algorithms require a large amountof computation, it is undesirable to use them inpower constrained applications, such as batteryoperated wireless video terminals. Minimizingpower dissipation is the key to maximizing batterylifetime and thus should be one of the driving forceswhen designing motion-estimation algorithms forportable video encoders.

Figure 1 shows a low-power wireless camera in anetworked environment. The goals for the designof this system include a long battery lifetime as wellas high video compression ratios. We have devel-oped a motion-estimation algorithm, termednetwork-driven motion estimation, which reducesthe power dissipation of wireless video devices in anetworked environment by exploiting the predict-ability of object motion. Since the location of anobject in the current frame can often be predictedaccurately from its location in previous frames, it ispossible to optimally partition the motion-estimationcomputation between battery operated portabledevices and high powered compute servers on thewired network.

In network driven motion estimation, a remote highpowered resource at the base-station (or on thewired network) predicts the motion vectors of thecurrent frame from the motion vectors of the pre-vious frames. The base-station then sends thesepredicted motion vectors to a portable videoencoder, where motion compensation proceeds asusual. Network-driven motion estimation adjuststhe coding algorithm based on the amount ofmotion in the sequence. This technique usesmotion prediction to code portions of the videosequence which contain a large amount of motionand conditional replenishment to code portions ofthe sequence which contain little scene motion.Network driven motion estimation achieves areduction in the number of operations performed atthe encoder for motion estimation by over twoorders of magnitude, while introducing minimaldegradation to the decoded video compared withconventional full search encoder-based motion esti-mation, as shown in figure 2.

Figure 2 also shows that, even though network-driven motion estimation and conditional replenish-ment require the same number of operations at theencoder, network driven motion estimation greatlyimproves the quality of the decoded images com-pared with conditional replenishment. Thusnetwork-driven motion estimation obtains the powerefficiency of conditional replenishment while main-taining the high quality reconstructed images ofencoder-based full search motion estimation.

317

Wired Network

Remote Server


Figure 2. Tennis sequence coded with a constant bitrate. (a) Encoder-based motion estimation, SNR = 27.3dB. (b) Network driven motion estimation, SNR = 25.4dB. (c) Conditional replenishment, SNR = 22.3 dB.

1.22.1 Publications

Rabiner, W.B., and A.P. Chandrakasan. "NetworkDriven Motion Estimation for Portable Video Ter-minals." Paper presented at the InternationalConference on Audio, Speech, and Signal Pro-cessing, Munich, Germany, April 21-24, 1997.

Rabiner, W.B. Network Driven Motion Estimationfor Wireless Video Terminals. S.M. thesis.Dept. of Electr. Eng. and Comput. Sci., MIT,1997.


section 2 digital signal processing - semantic scholarthe field of digital signal processing grew...

Documents