ebook img

Complex Valued Nonlinear Adaptive Filters: Noncircularity, Widely Linear and Neural Models PDF

327 Pages·2009·14.716 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Complex Valued Nonlinear Adaptive Filters: Noncircularity, Widely Linear and Neural Models

Complex Valued Nonlinear Adaptive Filters Complex Valued Nonlinear Adaptive Filters: Noncircularity, Widely Linear and Neural Models Danilo P. Mandic and Vanessa Su Lee Goh © 2009 John Wiley & Sons, Ltd. ISBN: 978-0-470-06635-5 www.it-ebooks.info Complex Valued Nonlinear Adaptive Filters Noncircularity, Widely Linear and Neural Models DaniloP.Mandic ImperialCollegeLondon,UK VanessaSuLeeGoh ShellEP,Europe www.it-ebooks.info Thiseditionfirstpublished2009 © 2009,JohnWiley&Sons,Ltd Registeredoffice JohnWiley&SonsLtd,TheAtrium,SouthernGate,Chichester,WestSussex,PO198SQ,UnitedKingdom Fordetailsofourglobaleditorialoffices,forcustomerservicesandforinformationabouthowtoapplyfor permissiontoreusethecopyrightmaterialinthisbookpleaseseeourwebsiteatwww.wiley.com. TherightoftheauthortobeidentifiedastheauthorofthisworkhasbeenassertedinaccordancewiththeCopyright, DesignsandPatentsAct1988. Allrightsreserved.Nopartofthispublicationmaybereproduced,storedinaretrievalsystem,ortransmitted,inany formorbyanymeans,electronic,mechanical,photocopying,recordingorotherwise,exceptaspermittedbytheUK Copyright,DesignsandPatentsAct1988,withoutthepriorpermissionofthepublisher. Wileyalsopublishesitsbooksinavarietyofelectronicformats.Somecontentthatappearsinprintmaynotbe availableinelectronicbooks. Designationsusedbycompaniestodistinguishtheirproductsareoftenclaimedastrademarks.Allbrandnamesand productnamesusedinthisbookaretradenames,servicemarks,trademarksorregisteredtrademarksoftheir respectiveowners.Thepublisherisnotassociatedwithanyproductorvendormentionedinthisbook.This publicationisdesignedtoprovideaccurateandauthoritativeinformationinregardtothesubjectmattercovered.Itis soldontheunderstandingthatthepublisherisnotengagedinrenderingprofessionalservices.Ifprofessionaladvice orotherexpertassistanceisrequired,theservicesofacompetentprofessionalshouldbesought. LibraryofCongressCataloging-in-PublicationData Mandic,DaniloP. Complexvaluednonlinearadaptivefilters:noncircularity,widelylinear,andneuralmodels/byDaniloP. Mandic,VanessaSuLeeGoh,ShellEP,Europe. p.cm. Includesbibliographicalreferencesandindex. ISBN978-0-470-06635-5(cloth) 1. Functionsofcomplexvariables.2. Adaptivefilters–Mathematicalmodels.3. Filters(Mathematics) 4. Nonlineartheories.5. Neuralnetworks(Computerscience)I. Goh,VanessaSuLee.II. Holland,Shell. III. Title. TA347.C64.M362009 621.382’2–dc22 2009001965 AcataloguerecordforthisbookisavailablefromtheBritishLibrary. ISBN:978-0-470-06635-5 Typesetin10/12ptTimesbyThomsonDigital,Noida,India PrintedinGreatBritainbyCPIAntonyRowe,Chippenham,Wiltshire www.it-ebooks.info The real voyage of discovery consists not in seeking new landscapes but in having new eyes Marcel Proust www.it-ebooks.info Contents Preface xiii Acknowledgements xvii 1 TheMagicofComplexNumbers 1 1.1 HistoryofComplexNumbers 2 1.1.1 HypercomplexNumbers 7 1.2 HistoryofMathematicalNotation 8 1.3 DevelopmentofComplexValuedAdaptiveSignalProcessing 9 2 WhySignalProcessingintheComplexDomain? 13 2.1 SomeExamplesofComplexValuedSignalProcessing 13 2.1.1 DualityBetweenSignalRepresentationsinRandC 18 2.2 ModellinginCisNotOnlyConvenientButAlsoNatural 19 2.3 WhyComplexModellingofRealValuedProcesses? 20 2.3.1 PhaseInformationinImaging 20 2.3.2 ModellingofDirectionalProcesses 22 2.4 ExploitingthePhaseInformation 23 2.4.1 SynchronisationofRealValuedProcesses 24 2.4.2 AdaptiveFilteringbyIncorporatingPhaseInformation 25 2.5 OtherApplicationsofComplexDomainProcessingofRealValuedSignals 26 2.6 AdditionalBenefitsofComplexDomainProcessing 29 3 AdaptiveFilteringArchitectures 33 3.1 LinearandNonlinearStochasticModels 34 3.2 LinearandNonlinearAdaptiveFilteringArchitectures 35 3.2.1 FeedforwardNeuralNetworks 36 3.2.2 RecurrentNeuralNetworks 37 3.2.3 NeuralNetworksandPolynomialFilters 38 3.3 StateSpaceRepresentationandCanonicalForms 39 www.it-ebooks.info viii Contents 4 ComplexNonlinearActivationFunctions 43 4.1 PropertiesofComplexFunctions 43 4.1.1 SingularitiesofComplexFunctions 45 4.2 UniversalFunctionApproximation 46 4.2.1 UniversalApproximationinR 47 4.3 NonlinearActivationFunctionsforComplexNeuralNetworks 48 4.3.1 Split-complexApproach 49 4.3.2 FullyComplexNonlinearActivationFunctions 51 4.4 GeneralisedSplittingActivationFunctions(GSAF) 53 4.4.1 TheCliffordNeuron 53 4.5 Summary:ChoiceoftheComplexActivationFunction 54 5 ElementsofCRCalculus 55 5.1 ContinuousComplexFunctions 56 5.2 TheCauchy–RiemannEquations 56 5.3 GeneralisedDerivativesofFunctionsofComplexVariable 57 5.3.1 CRCalculus 59 5.3.2 LinkbetweenR-andC-derivatives 60 5.4 CR-derivativesofCostFunctions 62 5.4.1 TheComplexGradient 62 5.4.2 TheComplexHessian 64 5.4.3 TheComplexJacobianandComplexDifferential 64 5.4.4 GradientofaCostFunction 65 6 ComplexValuedAdaptiveFilters 69 6.1 AdaptiveFilteringConfigurations 70 6.2 TheComplexLeastMeanSquareAlgorithm 73 6.2.1 ConvergenceoftheCLMSAlgorithm 75 6.3 NonlinearFeedforwardComplexAdaptiveFilters 80 6.3.1 FullyComplexNonlinearAdaptiveFilters 80 6.3.2 DerivationofCNGDusingCRcalculus 82 6.3.3 Split-complexApproach 83 6.3.4 DualUnivariateAdaptiveFilteringApproach(DUAF) 84 6.4 NormalisationofLearningAlgorithms 85 6.5 PerformanceofFeedforwardNonlinearAdaptiveFilters 87 6.6 Summary:ChoiceofaNonlinearAdaptiveFilter 89 7 AdaptiveFilterswithFeedback 91 7.1 TrainingofIIRAdaptiveFilters 92 7.1.1 CoefficientUpdateforLinearAdaptiveIIRFilters 93 7.1.2 TrainingofIIRfilterswithReducedComputational Complexity 96 www.it-ebooks.info Contents ix 7.2 NonlinearAdaptiveIIRFilters:RecurrentPerceptron 97 7.3 TrainingofRecurrentNeuralNetworks 99 7.3.1 OtherLearningAlgorithmsandComputationalComplexity 102 7.4 SimulationExamples 102 8 FilterswithanAdaptiveStepsize 107 8.1 BenvenisteTypeVariableStepsizeAlgorithms 108 8.2 ComplexValuedGNGDAlgorithms 110 8.2.1 ComplexGNGDforNonlinearFilters(CFANNGD) 112 8.3 SimulationExamples 113 9 FilterswithanAdaptiveAmplitudeofNonlinearity 119 9.1 DynamicalRangeReduction 119 9.2 FIRAdaptiveFilterswithanAdaptiveNonlinearity 121 9.3 RecurrentNeuralNetworkswithTrainableAmplitudeofActivation Functions 122 9.4 SimulationResults 124 10 Data-reusingAlgorithmsforComplexValuedAdaptiveFilters 129 10.1 TheData-reusingComplexValuedLeastMeanSquare(DRCLMS) Algorithm 129 10.2 Data-reusingComplexNonlinearAdaptiveFilters 131 10.2.1 ConvergenceAnalysis 132 10.3 Data-reusingAlgorithmsforComplexRNNs 134 11 ComplexMappingsandMo¨biusTransformations 137 11.1 MatrixRepresentationofaComplexNumber 137 11.2 TheMo¨biusTransformation 140 11.3 ActivationFunctionsandMo¨biusTransformations 142 11.4 All-passSystemsasMo¨biusTransformations 146 11.5 FractionalDelayFilters 147 12 AugmentedComplexStatistics 151 12.1 ComplexRandomVariables(CRV) 152 12.1.1 ComplexCircularity 153 12.1.2 TheMultivariateComplexNormalDistribution 154 12.1.3 MomentsofComplexRandomVariables(CRV) 157 12.2 ComplexCircularRandomVariables 158 12.3 ComplexSignals 159 12.3.1 WideSenseStationarity,Multicorrelations,andMultispectra 160 12.3.2 StrictCircularityandHigher-orderStatistics 161 12.4 Second-orderCharacterisationofComplexSignals 161 12.4.1 AugmentedStatisticsofComplexSignals 161 12.4.2 Second-orderComplexCircularity 164 www.it-ebooks.info x Contents 13 WidelyLinearEstimationandAugmentedCLMS(ACLMS) 169 13.1 MinimumMeanSquareError(MMSE)EstimationinC 169 13.1.1 WidelyLinearModellinginC 171 13.2 ComplexWhiteNoise 172 13.3 AutoregressiveModellinginC 173 13.3.1 WidelyLinearAutoregressiveModellinginC 174 13.3.2 QuantifyingBenefitsofWidelyLinearEstimation 174 13.4 TheAugmentedComplexLMS(ACLMS)Algorithm 175 13.5 AdaptivePredictionBasedonACLMS 178 13.5.1 WindForecastingUsingAugmentedStatistics 180 14 DualityBetweenComplexValuedandRealValuedFilters 183 14.1 ADualChannelRealValuedAdaptiveFilter 184 14.2 DualityBetweenRealandComplexValuedFilters 186 14.2.1 OperationofStandardComplexAdaptiveFilters 186 14.2.2 OperationofWidelyLinearComplexFilters 187 14.3 Simulations 188 15 WidelyLinearFilterswithFeedback 191 15.1 TheWidelyLinearARMA(WL-ARMA)Model 192 15.2 WidelyLinearAdaptiveFilterswithFeedback 192 15.2.1 WidelyLinearAdaptiveIIRFilters 195 15.2.2 AugmentedRecurrentPerceptronLearningRule 196 15.3 TheAugmentedComplexValuedRTRL(ACRTRL)Algorithm 197 15.4 TheAugmentedKalmanFilterAlgorithmforRNNs 198 15.4.1 EKFBasedTrainingofComplexRNNs 200 15.5 AugmentedComplexUnscentedKalmanFilter(ACUKF) 200 15.5.1 StateSpaceEquationsfortheComplexUnscentedKalman Filter 201 15.5.2 ACUKFBasedTrainingofComplexRNNs 202 15.6 SimulationExamples 203 16 CollaborativeAdaptiveFiltering 207 16.1 ParametricSignalModalityCharacterisation 207 16.2 StandardHybridFilteringinR 209 16.3 TrackingtheLinear/NonlinearNatureofComplexValuedSignals 210 16.3.1 SignalModalityCharacterisationinC 211 16.4 SplitvsFullyComplexSignalNatures 214 16.5 OnlineAssessmentoftheNatureofWindSignal 216 16.5.1 EffectsofAveragingonSignalNonlinearity 216 16.6 CollaborativeFiltersforGeneralComplexSignals 217 16.6.1 HybridFiltersforNoncircularSignals 218 16.6.2 OnlineTestforComplexCircularity 220 www.it-ebooks.info Contents xi 17 AdaptiveFilteringBasedonEMD 221 17.1 TheEmpiricalModeDecompositionAlgorithm 222 17.1.1 EmpiricalModeDecompositionasaFixedPointIteration 223 17.1.2 ApplicationsofRealValuedEMD 224 17.1.3 UniquenessoftheDecomposition 225 17.2 ComplexExtensionsofEmpiricalModeDecomposition 226 17.2.1 ComplexEmpiricalModeDecomposition 227 17.2.2 RotationInvariantEmpiricalModeDecomposition(RIEMD) 228 17.2.3 BivariateEmpiricalModeDecomposition(BEMD) 228 17.3 AddressingtheProblemofUniqueness 230 17.4 ApplicationsofComplexExtensionsofEMD 230 18 ValidationofComplexRepresentations–IsThisWorthwhile? 233 18.1 SignalModalityCharacterisationinR 234 18.1.1 SurrogateDataMethods 235 18.1.2 TestStatistics:TheDVVMethod 237 18.2 TestingfortheValidityofComplexRepresentation 239 18.2.1 ComplexDelayVectorVarianceMethod(CDVV) 240 18.3 QuantifyingBenefitsofComplexValuedRepresentation 243 18.3.1 ProsandConsoftheComplexDVVMethod 244 AppendixA:SomeDistinctivePropertiesofCalculusinC 245 AppendixB:Liouville’sTheorem 251 AppendixC:HypercomplexandCliffordAlgebras 253 C.1 DefinitionsofAlgebraicNotionsofGroup,RingandField 253 C.2 DefinitionofaVectorSpace 254 C.3 HigherDimensionAlgebras 254 C.4 TheAlgebraofQuaternions 255 C.5 CliffordAlgebras 256 AppendixD:RealValuedActivationFunctions 257 D.1 LogisticSigmoidActivationFunction 257 D.2 HyperbolicTangentActivationFunction 258 AppendixE:ElementaryTranscendentalFunctions(ETF) 259 AppendixF:TheONotationandStandardVectorandMatrixDifferentiation 263 F.1 TheONotation 263 F.2 StandardVectorandMatrixDifferentiation 263 www.it-ebooks.info xii Contents AppendixG:NotionsFromLearningTheory 265 G.1 TypesofLearning 266 G.2 TheBias–VarianceDilemma 266 G.3 RecursiveandIterativeGradientEstimationTechniques 267 G.4 TransformationofInputData 267 AppendixH:NotionsfromApproximationTheory 269 AppendixI:TerminologyUsedintheFieldofNeuralNetworks 273 AppendixJ:ComplexValuedPipelinedRecurrentNeuralNetwork(CPRNN) 275 J.1 TheComplexRTRLAlgorithm(CRTRL)forCPRNN 275 J.1.1 LinearSubsectionWithinthePRNN 277 AppendixK:GradientAdaptiveStepSize(GASS)AlgorithmsinR 279 K.1 GradientAdaptiveStepsizeAlgorithmsBasedon∂J/∂μ 280 K.2 VariableStepsizeAlgorithmsBasedon∂J/∂ε 281 AppendixL:DerivationofPartialDerivativesfromChapter8 283 L.1 Derivationof∂e(k)/∂w (k) 283 n L.2 Derivationof∂e∗(k)/∂ε(k−1) 284 L.3 Derivationof∂w(k)/∂ε(k−1) 286 AppendixM:APosterioriLearning 287 M.1 APosterioriStrategiesinAdaptiveLearning 288 AppendixN:NotionsfromStabilityTheory 291 AppendixO:LinearRelaxation 293 O.1 VectorandMatrixNorms 293 O.2 RelaxationinLinearSystems 294 O.2.1 ConvergenceintheNormorStateSpace? 297 AppendixP:ContractionMappings,FixedPointIterationandFractals 299 P.1 HistoricalPerspective 303 P.2 MoreonConvergence:ModifiedContractionMapping 305 P.3 FractalsandMandelbrotSet 308 References 309 Index 321 www.it-ebooks.info

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.