LEAST-MEAN-SQUARE ADAPTIVE FILTERS LEAST-MEAN-SQUARE ADAPTIVE FILTERS Edited by S. Haykin and B. Widrow A JOHN WILEY & SONS, INC. PUBLICATION Thisbookisprintedonacid-freepaper. Copyrightq2003byJohnWiley&SonsInc.Allrightsreserved. PublishedsimultaneouslyinCanada. Nopartofthispublicationmaybereproduced,storedinaretrievalsystemortransmittedinanyformorby anymeans,electronic,mechanical,photocopying,recording,scanningorotherwise,exceptaspermitted underSections107or108ofthe1976UnitedStatesCopyrightAct,withouteitherthepriorwritten permissionofthePublisher,orauthorizationthroughpaymentoftheappropriateper-copyfeetothe CopyrightClearanceCenter,222RosewoodDrive,Danvers,MA01923,(978)750-8400,fax(978)750- 4744.RequeststothePublisherforpermissionshouldbeaddressedtothePermissionsDepartment,John Wiley&Sons,Inc.,111RiverStreet,Hoboken,NewJersey07030,(201)748-6011,fax(201)748-6008, E-Mail:[email protected]. Fororderingandcustomerservice,call1-800-CALL-WILEY. LibraryofCongressCataloging-in-PublicationData: Least-mean-squareadaptivefilters/editedbyS.HaykinandB.Widrow p.cm. Includesbibliographicalreferencesandindex. ISBN0-471-21570-8(cloth) 1. Adaptivefilters—Designandconstruction—Mathematics.2.Leastsquares.I.Widrow, Bernard,1929-II.Haykin,Simon,1931- TK7872.F5L43 2003 621.38150324—dc21 2003041161 PrintedintheUnitedStatesofAmerica. 10987654321 Thisbook isdedicatedto Bernard Widrowforinventingthe LMS filter and investigatingits theoryand applications Simon Haykin CONTENTS Contributors ix Introduction: The LMS Filter (Algorithm) xi Simon Haykin 1. On the Efficiency ofAdaptive Algorithms 1 Bernard Widrow and Max Kamenetsky 2. Traveling-Wave Model of Long LMS Filters 35 Hans J. Butterweck 3. EnergyConservation and the Learning AbilityofLMS Adaptive Filters 79 Ali H.Sayed and V. H.Nascimento 4. On the Robustness of LMS Filters 105 Babak Hassibi 5. Dimension Analysis for Least-Mean-Square Algorithms 145 Iven M. Y. Mareels, John Homer,and Robert R. Bitmead 6. Control ofLMS-Type Adaptive Filters 175 Eberhard Ha¨nslerand Gerhard Uwe Schmidt 7. Affine Projection Algorithms 241 Steven L. Gay 8. Proportionate Adaptation: New Paradigms inAdaptive Filters 293 Zhe Chen, Simon Haykin,and StevenL. Gay 9. Steady-State Dynamic Weight Behaviorin (N)LMS Adaptive Filters 335 A. A. (Louis) Beex and James R. Zeidler vii viii CONTENTS 10. Error Whitening WienerFilters:Theory and Algorithms 445 Jose C. Principe, YadunandanaN. Rao, andDeniz Erdogmus Index 491 CONTRIBUTORS A. A. (LOUIS) BEEX, Systems Group—DSP Research Laboratory, The Bradley Department of Electrical and Computer Engineering, Virginia Tech, Blacksburg, VA 24061-0111 ROBERT R. BITMEAD, Department of Mechanical and Aerospace Engineering, University of California, San Diego, 9500 Gilman Drive, La Jolla, CA 92093- 0411 HANS BUTTERWECK, Technische Universiteit Eindhoven, Faculteit Elektrotech- niek, EH 5.29, Postbus 513, 5600MB Eindhoven,Netherlands ZHE CHEN, Department of Electrical and Computer Engineering, CRL 102, McMaster University, 1280 Main Street West, Hamilton, Ontario, Canada L8S 4K1 DENIZ ERDOGMUS, Computational NeuroEngineering Laboratory, EB 451, Building 33, UniversityofFlorida, Gainesville, FL32611 STEVEN L. GAY, Acoustics and Speech Research Department, Bell Labs, Room 2D-531, 600 MountainAve.,Murray Hill, NJ 07974 PROF. DR.-ING. EBERHARD HA¨NSLER, Institute of Communication Technology, Darmstadt University of Technology, Merckstrasse 25, D-64283 Darmstadt, Germany BABAK HASSIBI, Department of Electrical Engineering, 1200 East California Blvd., M/C 136-93, CaliforniaInstitute ofTechnology,Pasadena, CA91101 SIMON HAYKIN, Department of Electrical and Computer Engineering, McMaster University, 1280 Main Street West, Hamilton,Ontario, Canada L8S 4K1 JOHN HOMER, School of Computer Science and Electrical Engineering, The UniversityofQueensland, Brisbane 4072 MAX KAMENETSKY, Stanford University, David Packard Electrical Engineering, 350 Serra Mall, Room 263, Stanford, CA 94305-9510 IVEN M.Y.MAREELS, DepartmentofElectricalandElectronicEngineering,The UniversityofMelbourne,Melbourne Vic3010 ix x CONTRIBUTORS V.H.NASCIMENTO, DepartmentofElectronicSystemsEngineering,Universityof Sa˜o Paulo, Brazil JOSE C. PRINCIPE, Computational NeuroEngineering Laboratory, EB 451, Building 33, UniversityofFlorida, Gainesville, FL32611 YADUNANDANA N. RAO, Computational NeuroEngineering Laboratory, EB 451, Building 33, UniversityofFlorida, Gainesville, FL32611 ALIH.SAYED, DepartmentofElectricalEngineering,Room44-123AEngineering IV Bldg, UniversityofCalifornia, Los Angeles, CA 90095-1594 GERHARD UWE SCHMIDT, Institute of Communication Technology, Darmstadt UniversityofTechnology,Merckstrasse25, D-64283Darmstadt, Germany BERNARD WIDROW, Stanford University, David Packard Electrical Engineering, 350 Serra Mall, Room 273, Stanford, CA 94305-9510 JAMES R. ZEIDLER, Department of Electrical and Computer Engineering, UniversityofCalifornia, San Diego,La Jolla, CA92092 INTRODUCTION: THE LMS FILTER (ALGORITHM) SIMON HAYKIN The earliest work on adaptive filters may be traced back to the late 1950s, during which time a number of researchers were working independently on theories and applications of such filters. From this early work, the least-mean-square ðLMSÞ algorithm emerged as a simple, yet effective, algorithm for the design of adaptive transversal(tapped-delay-line) filters. TheLMSalgorithmwasdevisedbyWidrowandHoffin1959intheirstudyofa pattern-recognition machine known as the adaptive linear element, commonly referred to as the Adaline [1, 2]. The LMS algorithm is a stochastic gradient algorithminthatititerateseachtapweightofthetransversalfilterinthedirectionof theinstantaneousgradientofthesquarederrorsignalwithrespecttothetapweight inquestion. Let ww^ðnÞ denote the tap-weight vector of the LMS filter, computed at iteration (time step) n. The adaptive operation of the filter is completely described by the recursiveequation (assumingcomplex data) ww^ðnþ1Þ¼ww^ðnÞþmuðnÞ½dðnÞ(cid:1)ww^HðnÞuðnÞ(cid:2)*; ð1Þ whereuðnÞisthetap-inputvector,dðnÞisthedesiredresponse,andmisthestep-size parameter.Thequantityenclosedinsquarebracketsistheerrorsignal.Theasterisk denotescomplexconjugation,andthesuperscriptHdenotesHermitiantransposition (i.e., ordinarytransposition combinedwith complexconjugation). Equation (1) is testimony to the simplicity of the LMS filter. This simplicity, coupledwithdesirablepropertiesoftheLMSfilter(discussedinthechaptersofthis book)andpracticalapplications[3,4],hasmadetheLMSfilteranditsvariantsan importantpartoftheadaptivesignalprocessingkitoftools,notjustforthepast40 yearsbutformanyyearstocome.Simplyput,theLMSfilterhaswithstoodthetest oftime. AlthoughtheLMSfilterisverysimpleincomputationalterms,itsmathematical analysis is profoundly complicated because of its stochastic and nonlinear nature. Indeed, despite the extensive effort that has been expended in the literature to xi xii INTRODUCTION:THELMSFILTER(ALGORITHM) analyze the LMS filter, we still do not have a direct mathematical theory for its stability and steady-state performance, and probably we never will. Nevertheless, we do have a good understanding of its behavior in a stationary as well as a nonstationary environment, asdemonstrated inthe chapters ofthis book. The stochastic nature of the LMS filter manifests itself in the fact that in a stationaryenvironment,andundertheassumptionofasmallstep-sizeparameter,the filterexecutesaformofBrownianmotion.Specifically,thesmallstep-sizetheoryof the LMS filter is almost exactly described by the discrete-time version of the Langevinequation1 [3]: Dn ðnÞ¼n ðnþ1Þ(cid:1)n ðnÞ k k k ð2Þ ¼(cid:1)mln ðnÞþfðnÞ; k¼1;2;...;M; k k k whichisnaturallysplitintotwoparts:adampingforce(cid:1)mln ðnÞandastochastic k k force fðnÞ.Theterms used herein are defined as follows: k M ¼order (i.e., number of taps) of the transversal filter around which the LMS filter is built l ¼ktheigenvalueofthecorrelationmatrixoftheinputvectoruðnÞ,which k is denoted by R fðnÞ¼kth component of the vector (cid:1)mQHuðnÞe*ðnÞ k o Q¼unitary matrix whose M columns constitute an orthogonal set of eigerivectorsassociatedwiththeeigenvaluesofthecorrelationmatrixR e ðnÞ¼optimum error signal produced by the corresponding Wiener filter o driven by the input vectoruðnÞand the desired response dðnÞ ToillustratethevalidityofEq.(2)asthedescriptionofsmallstep-sizetheoryof theLMSfilter,wepresenttheresultsofacomputerexperimentonaclassicexample of adaptive equalization. The example involves an unknown linear channel whose impulseresponse is described bythe raised cosine [3] 8 (cid:3) (cid:1) (cid:2)(cid:4) <1 2p 1þcos ðn(cid:1)2Þ ; n¼1;2;3; h ¼ 2 W ð3Þ n : 0; otherwise wheretheparameterW controlstheamountofamplitudedistortionproducedbythe channel, with the distortion increasing with W. Equivalently, the parameter W controlstheeigenvaluespread(i.e.,theratioofthelargesteigenvaiuetothesmallest eigenvalue) of the correlation matrix of the tap inputs of the equalizer, with the eigenvalue spread increasing with W. The equalizer has M ¼11 taps. Figure 1 presentsthelearningcurvesoftheequalizertrainedusingtheLMSalgorithmwith the step-size parameter m¼0:0075 and varying W. Each learning curve was obtainedbyaveragingthesquaredvalueoftheerrorsignaleðnÞversusthenumberof iterations n over an ensemble of 100 independent trials of the experiment. The 1TheLangevinequationisthe“engineer’sversion”ofstochasticdifferential(difference)equations.