MAY 2005 VOLUME53 NUMBER 5 ITPRED (ISSN1053-587X) PAPERS MultichannelSignalProcessingApplications StatisticalResolutionLimitsandtheComplexifiedCramér–RaoBound.. ... ... ... .... ..... ... ... ... .. S.T.Smith 1597 Covariance,Subspace,andIntrinsicCramér–RaoBounds.. ... .... ... ... ... ...... ... ... ... ... ... .. S.T.Smith 1610 OptimalDimensionalityReductionofSensorDatainMultisensorEstimationFusion. ... ... ... ... ... ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ..Y.Zhu,E.Song,J.Zhou,andZ.You 1631 Fourth-OrderBlindIdentificationofUnderdeterminedMixturesofSources(FOBIUM) .. ... ... ... ... ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... ...A.Ferréol,L.Albera,andP.Chevalier 1640 PenaltyFunction-BasedJointDiagonalizationApproachforConvolutiveBlindSeparationofNonstationarySources.... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... ...W.Wang,S.Sanei,andJ.A.Chambers 1654 AnEmpiricalBayesEstimatorforIn-ScaleAdaptiveFiltering . .... ... ... ... ...... ... ... ... ... ... P.J.Gendron 1670 MethodsofSensorArrayandMultichannelProcessing RobustMinimumVarianceBeamforming . ... ... ... ... ... ....... ... ... ... ... ... ..R.G.LorenzandS.P.Boyd 1684 BlindSpatialSignatureEstimationviaTime-VaryingUserPowerLoadingandParallelFactorAnalysis... ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... .. Y.Rong,S.A.Vorobyov,A.B.Gershman,andN.D.Sidiropoulos 1697 SourceLocalizationbySpatiallyDistributedElectronicNosesforAdvectionandDiffusion .. ... ... ... ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... ... .J.Matthes,L.Gröll,andH.B.Keller 1711 SignalandSystemModeling DampedandDelayedSinusoidalModelforTransientSignals.. .... ... ...... ... ... ...R.BoyerandK.Abed-Meraim 1720 FilterDesignandTheory Robust FilteringforUncertain2-DContinuousSystems .. ...... .... .. S.Xu,J.Lam,Y.Zou,Z.Lin,andW.Paszke 1731 TheoryandDesignofMultirateSensorArrays. ... ... ... ... ..... ..... ... ... ... ... ..O.S.JahromiandP.Aarabi 1739 ArmletsandBalancedMultiwavelets:FlippingFilterConstruction .. ... ... ... ... ..... .... ... ... ... ... ..J.Lian 1754 OptimizationofTwo-DimensionalIIRFiltersWithNonseparableandSeparableDenominator ... ...... ...B.Dumitrescu 1768 ANewApproachforEstimationofStatisticallyMatchedWavelet... .... ..... ... ..A.Gupta,S.D.Joshi,andS.Prasad 1778 PhaseletsofFramelets ... ... ... ... ... ... ... ... ... ..... ..... ... ... ... ... ... ... ... ... .. R.A.Gopinath 1794 (ContentsContinuedonBackCover) TLFeBOOK (ContentsContinuedfromFrontCover) NonlinearSignalProcessing DirectProjectionDecodingAlgorithmforSigma-DeltaModulatedSignals .. ... ...... ... ...I.WiemerandW.Schwarz 1807 AlgorithmsandApplications TestingforStochasticIndependence:ApplicationtoBlindSourceSeparation ... .... ..... ... .. C.-J.KuandT.L.Fine 1815 NeuralNetworksMethodsandApplications NonlinearAdaptivePredictionofComplex-ValuedSignalsbyComplex-ValuedPRNN .... ...S.L.GohandD.P.Mandic 1827 SignalProcessingforCommunications EqualizationWithOversamplinginMultiuserCDMASystems. .... ... ...... ... ... . B.VrceljandP.P.Vaidyanathan 1837 Noise-PredictiveDecision-FeedbackDetectionforMultiple-InputMultiple-OutputChannels. ... ... ... ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... .D.W.WatersandJ.R.Barry 1852 BlindEqualizationforCorrelatedInputSymbols:ABussgangApproach.... ... ... ... ... ... ... ... ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... ... G.Panci,S.Colonnese,P.Campisi,andG.Scarano 1860 LowNoiseReversibleMDCT(RMDCT)andItsApplicationinProgressive-to-LosslessEmbeddedAudioCoding. .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... ... .J.Li 1870 SignalTransmission,Propagation,andRecovery ConvolutionalCodesUsingFinite-FieldWavelets:Time-VaryingCodesandMore... ... ... ... ... ... ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ...F.Fekri,M.Sartipi,R.M.Mersereau,andR.W.Schafer 1881 BERSensitivitytoMistiminginUltra-WidebandImpulseRadios—PartII:FadingChannels . ... ... ... ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ..Z.TianandG.B.Giannakis 1897 DesignandAnalysisofFeedforwardSymbolTimingEstimatorsBasedontheConditionalMaximumLikelihoodPrinciple .. ... ... .... ... ... ... ... ... ... .... ..... ... ... ... .... ... ... ... ... ... ... .Y.-C.WuandE.Serpedin 1908 CORRESPONDENCE MethodsofSensorArrayandMultichannelProcessing ABayesianApproachtoArrayGeometryDesign. . ... ... ... .... ..... .... ... ... ... ... .Ü.OktelandR.L.Moses 1919 ComputationofSpectralandRootMUSICThroughRealPolynomialRooting. .. ... ... .... ..... ... ... ... . J.Selva 1923 SignalDetectionandEstimation EstimateofAliasingErrorforNon-SmoothSignalsPrefilteredbyQuasi-ProjectionsIntoShift-InvariantSpaces .. .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... .. W.Chen,B.Han,andR.-Q.Jia 1927 RobustSuper-ExponentialMethodsforDeflationaryBlindSourceSeparationofInstantaneousMixtures.. ... ... .... .. ... ... .... ... ... ... ... ... ... ... ... ... ... ... .... ... ... .. M.Kawamoto,K.Kohno,andY.Inouye 1933 SignalProcessingforCommunications IterativeDecodingofWrappedSpace-TimeCodes.... ... ... .... .... ..... ... ... ... ... ..A.SezginandH.Boche 1937 ANNOUNCEMENTS CallforPapers—2006IEEEInternationalSymposiumonBiomedicalEngineering:FromNanotoMacro . ... ...... .... 1942 CallforPapers—IEEETRANSACTIONSONSPEECHANDAUDIOPROCESSINGSpecialIssueonProgressinRichTranscription .. ... ... .... ... ... ... ... ... ... ... ... ... .... ..... .... ... ... ... ... ... ... ... ... ... ... ... .... 1943 CallforPapers—IEEETRANSACTIONSONINFORMATIONFORENSICSANDSECURITY . ... ... ... ..... .... ... ... .... 1944 CallforPapers—IEEETRANSACTIONSONSIGNALPROCESSINGSpecialIssueonGenomicSignalProcessing . ..... ..... 1945 EDICS—Editor’sInformationClassificationScheme.. ... ... .... ... ... ... .... ..... ... ... ... ... ... ... .... 1946 InformationforAuthors.. ... ... ... ... ... ... ... ... ... .... ...... ... ... ... ... ... ... ... ... ... ... .... 1947 TLFeBOOK IEEETRANSACTIONSONSIGNALPROCESSING,VOL.53,NO.5,MAY2005 1597 Statistical Resolution Limits and the Complexified Cramér–Rao Bound Steven Thomas Smith, Senior Member, IEEE Abstract—Arrayresolution limitsand accuracy boundson the I. INTRODUCTION multitudeofsignalparameters(e.g.,azimuth,elevation,Doppler, RESOLUTIONandaccurateestimationofsignalsandtheir range, cross-range, depth, frequency, chirp, polarization, ampli- tude,phase,etc.)estimatedbyarrayprocessingalgorithmsarees- parametersarecentralcapabilitiesofadaptivesensorarray sentialtoolsintheevaluationofsystemperformance.Thecasein processing.TheCramér–Raobound(CRB)[29],[34],[44],[54] which the complex amplitudes of the signals are unknown is of providesthebestaccuracyachievablebyanyunbiasedestimator particularpracticalinterest.Acomputationallyefficientformula- of the signal parameters and therefore provides a fundamental tionofthesebounds(fromtheperspectiveofderivationsandanal- physical limit on system accuracy. In this paper, deterministic ysis)ispresentedforthecaseofdeterministicandunknownsignal CRBs are derived for a general signal model with unknown amplitudes. A new derivation is given using the unknown com- plexsignalparametersandtheircomplexconjugates.Thenewfor- signalamplitudes,andnewclosed-formexpressionsforstatis- mulaisreadilyapplicabletoobtainingeithersymbolicornumer- ticalresolutionlimitsofarraysarederivedfromthesebounds. icalsolutionstoestimationboundsforaverywideclassofprob- The statistical resolution limit is defined as the source sepa- lems encountered in adaptive sensor array processing. This for- ration that equals its own CRB, providing an algorithm-inde- mulaisshowntoyieldseveralofthestandardCramér–Raoresults pendent resolution bound. Multiparameter CRBs for adaptive for array processing, along with new results of fundamental in- sensorarrayprocessingproblemsarewellknown[8],[10],[27], terest.Specifically,anewclosed-formexpressionforthestatistical [40], [43], [45], [55]. In this paper, a new formula and new resolutionlimitofanapertureforanyasymptoticallyunbiasedsu- perresolutionalgorithm(e.g.,MUSIC,ESPRIT)isprovided.The derivationforCRBsinvolvingacomplexparameterspace[56] statisticalresolutionlimitisdefinedasthesourceseparationthat areshowntobebothverygeneralandveryconvenient. equalsitsownCramér–Raobound,providinganalgorithm-inde- Thenewformula[see(29)]fordeterministicCRBshasbeen pendent bound on the resolution of any high-resolution method. usedinawidevarietyofapplicationstoobtainbothnumerical It is shown thatthe statistical resolution limit ofan array orco- herentintegrationwindowisabout12 SNR 1 4 relativetothe andanalyticalCRBresultsquicklyandtransparently.Whatdis- tinguishestheCRBresultsinthispaperfromthegreatmajority Fourier resolution limit of 2 radians (large number of arrayelements).Thatis,thehighestachievableresolutionispro- of existing derivations is a “complexified” approach to the portionaltothereciprocalofthefourthrootofthesignal-to-noise generalproblemofestimatingmultiplesignalparametersfrom ratio(SNR),incontrasttothesquare-root(SNR 1 2)dependence multiple signals in (proper [26]) complex noise, each signal of standard accuracy bounds. These theoretical results are con- withanunknownamplitudeandphase.Whereasthe“realified” sistentwithpreviouslypublishedboundsforspecificsuperresolu- approach to estimating complex parameters involves their tion algorithms derived by other methods. It is also shown that real and imaginary components, the complexified approach thepotentialresolutionimprovementobtainedbyseparatingtwo uses the complex parameters and their complex conjugates collineararrays(syntheticultra-wideband),eachwithafixedaper- ture wavelengthsby wavelengths(assumedlarge),isapprox- directly, oftentimes resulting in a greatly simplified analysis. imately( )1 2,incontrasttotheresolutionimprovementof (See Appendix A for background and algebraic definitions of forafullaperture.Exactclosed-formresultsfortheseprob- realified and complexified vector spaces.) The complexified lemswiththeirasymptoticapproximationsarepresented. approach and the efficient expressions for CRBs it yields are Index Terms—Adaptive arrays, adaptive estimation, adaptive describedbyYauandBresler[56]andvandenBos[53].New signal processing, amplitude estimation, direction-of-arrival esti- results employing the complexified approach have begun to mation, error analysis, ESPRIT, estimation, Fisher information, appear [20], [40]. The derivation of the complexified CRB image resolution, maximum likelihood estimation, MUSIC, pa- presented here is greatly simplified from previous work, and rameter estimation, parameter space methods, phase estimation, the resulting formula is more general than Yau and Bresler’s radarresolution,signalresolution,spectralanalysis,superresolu- [56, Th. 2] because a more general steering vector model tion,ultra-widebandprocessing. and parameter space are considered, as well as an arbitrary interference-plus-noisecovariance. TheCRBformulaeareappliedtocomputingnewclosed-form expressions for the statistical resolution limit of both a single Manuscript received January 5, 2004; revised May 25, 2004. This work uniformlineararray(ULA)[see(61)]andasplit-apertureused was supported by the United States Air Force under Air Force Contract in synthetic ultra-wideband processing [see (73)]. The case of F19628-00-C-0002.Opinions,interpretations,conclusions,andrecommenda- a single data snapshot or nonfluctuating (deterministic) target tionsarethoseoftheauthorandarenotnecessarilyendorsedbytheUnited States Government. The associate editor coordinating the review of this model is considered. Synthetic ultra-wideband processing manuscriptandapprovingitforpublicationwasDr.FulvioGini. [1], [5], [6], [22], [40], [58] refers to a family of methods TheauthoriswiththeLincolnLaboratory,MassacusettsInstituteofTech- using model-based interpolation/extrapolation across widely nology,Lexington,MA02420USA(e-mail:[email protected]). DigitalObjectIdentifier10.1109/TSP.2005.845426 separated subbands/apertures, usually over large fractional 1053-587X/$20.00©2005IEEE TLFeBOOK 1598 IEEETRANSACTIONSONSIGNALPROCESSING,VOL.53,NO.5,MAY2005 bandwidths. Because the statistical resolution limit is based where is a proper complex [26] Gaussian interference-plus- ontheCRB,itrepresentsahighSNRboundontheresolution noisevectorwith -by- covariancematrix achievable by any asymptotically unbiased superresolution method,suchasMUSIC[24],[30],[35],[46],[48]orESPRIT (2) [32], [48]. CRB analysis of the source separation to bound isacomplex -by- “steering”matrixwhosecolumnsdefine the resolution has been used by Clark [4], Smith [40], and the responses of signals impinging upon the sensor, Ying et al. [57] in a mathematically equivalent treatment. It isthe -vectorofrealsignalparameters,and is shown that the statistical resolution limit of an aperture is isthe(unknown,deterministic)complex -vectorofampli- about SNR rad relative to the Fourier resolution limit tudescorrespondingtothe signals.Throughoutthepaper,it rad (large number of array elements). That is, the statistical resolution limit is proportional to the reciprocal of is assumed that the unknown amplitudes are fixed (determin- thefourthrootofthesignal-to-noiseratio(SNR),incontrastto istic), although the extension to the stochastic case is straight- the square-root SNR dependence of standard accuracy forward.AsexplainedinAppendixB,deterministicCRBresults bounds. For the application of array processing, this provides donotdependonthelackofknowledgeofthecovariancema- the best possible fraction of the Rayleigh resolution limit trix . Note that this signal model is identical to that of Yau achievablebyanysuperresolutionmethod. and Bresler’s [56] in the frequently encountered case, where The single-aperture results are comparable in many ways ( ,i.e., parametersper with the deterministic resolution limit found by Lee [23], the eachofthe signals).Theprobabilitydistributionof is stochastic CRB case explored by Swingler [49]–[52], and the resolvability detectors considered by Shahram and Mi- lanfar [37]–[39]. The analysis of ULA resolution is applied (3) to the problem of computing resolution limits for synthetic andthelog-likelihoodfunctionis ultra-wideband processing, in which widely separated aper- tures(eitherinthefrequencyorspatialdomains)arecoherently (4) combinedusingamodel-basedproceduretoprovideenhanced resolution.Itisshownthatsyntheticultra-widebandprocessing Bounds on the estimation error of the real parameters , does indeed improve the resolution above that obtained by which are typically composed of a collection of angles of using each aperture individually. It is also shown that the arrival, Doppler frequencies, ranges, frequencies, damping potential resolution improvement obtained by separating two factors, etc., are to be computed. If the unknown complex collinear arrays, each with a fixed aperture wavelengths by amplitudes are in fact known, then the Cramér–Rao theorem wavelengths(assumedlarge),isapproximately , [29], [34], [54] asserts that the covariance of any unbiased in contrast to the resolution improvement of for a full estimatorofthe (real-valued)unknownparametersisbounded aperture. belowbytheinverseoftheFIM: Itshouldbekeptinmindthattheaccuracyandresolutionre- sultsdescribedinthispaperareallderivedfromCRBs.There- (5) fore,theyinheritthepropertiesofsuchbounds:Theyarelocal (inthesensethatforsymmetricmatrices and , if (notglobal)boundsanddonotreflectthelargeerrorsthatoccur andonlyif ispositivesemidefinite),where atlowerSNR’s,wheretheambiguityfunctionhashighsidelobe levels. (6) Thepaper’sbodyisorganizedintothreesections.SectionII containsthedescriptionofthecomplexifiedFisherinformation istheFIM,and matrix(FIM)andthecorrespondingcomplexifiedCRBs.These bounds are applied to the key example of pole model estima- (7) tioninSectionIII.Thepolemodelaccuracyboundsobtainedin SectionIIIarethenappliedtotheproblemofcomputingreso- (8) lution limits inSectionIV, which includesa treatment ofboth thesingle-apertureandsplit-aperturecases,aswellasacompar- are the Jacobian and Hessian matrices (1-by- and -by- isonoftheresolutionimprovementbetweenthetwo.Thereare matrices, respectively) of derivatives of with respect to the alsoappendicesdefiningthealgebraicconceptofcomplexified elementsoftheparametervector .Recallthatif isan vectorspaces,derivingthecomplexifiedFIM,andapplyingthe arbitrarymappingfrom to ,thentheJacobianmatrixand, complexifiedCRBtoasimpleangle-angleestimationproblem. consequently,theFIMtransformcontravariantly,i.e., (thechainrule).Therefore II. COMPLEXIFIEDCRB (9) A. Preliminaries Themeasurementvector fromasensorarraycontaining where is the Jacobian matrix of the elementsisgivenbytheequation transformation. Furthermore, if there are real nuisance pa- rameters thatmustbeestimated(such (1) asthesignalamplitudes)alongwiththedesiredparameters , TLFeBOOK SMITH:STATISTICALRESOLUTIONLIMITSANDTHECOMPLEXIFIEDCRAMÉR-RAOBOUND 1599 thentheCramér–Raolowerboundontheestimationerrorof brieflydiscussed.Thereal-valuedFIM ispositivedefi- isgivenbytheinverseoftheSchurcomplementof inthe nitewithrespecttothemetric ,i.e.,alleigenvaluesof full FIM the matrix pencil are positive. By (12), (10) the metrics ThisCRBontheerrorcovariance is (11) (14) which increases the CRB due to the lack of knowledge of . (RecallthattheSchurcomplementofthesquarematrix inthe are equivalent. Therefore, the complexified FIM blockmatrix is .Theinverse of (13) is positive definite with respect to the metric oftheSchurcomplementisdenotedhereas .)Note , i.e., the that the CRB depends on the unknown parameter vector generalized eigenvalues of the matrix pencil .Equation(11)isinvarianttoarbitrarylineartransformations ofthenuisanceparameters ,where isaninvertible (whichequaltheeigenvaluesof ) matrixthatmaypossiblyhavecomplex-valuedelements.By(9), areallpositive.Comparethesefactspertainingtotheeffectof thechangeofvariables inducesthefollowingtrans- lineartransformationsonestimationaccuracywiththeproper- formationsin : and ; tiesofgeneralcomplexrandomvariables[36];seealsoKrantz therefore, for an arbitrary [19]forageneraltreatmentoffunctionsofcomplexvariables. invertiblematrix . Note that the change of variables in (12) and the chain rule Consider the unknown complex amplitude vector definesderivativeswithrespectto and [19]: Re Im appearingin(1).Traditionally,CRB results are derived in terms of the real and imaginary parts and .However,duetotheinvarianceof(11),theCRBremains (15) unchangedifthecomplexifiedvariables (16) (12) Thesederivativesareconsistentwiththedesiredformulae areused.Becausethemostcomputationallyconvenientformof (17) thelog-likelihoodfunctiondependsonthecomplexifiedparam- eters and andnottherealparameters and ,acomplexi- Asanaside,if isafunctionofseveralcomplexvariables, fiedapproachisgenerallypreferred[53],[56].By(9),theFIM isanalyticifandonlyif ,i.e.,theCauchy-Riemann oftherealandimaginarycomponents and isrelatedtothe equations.Becausethelog-likelihoodfunctionisreal-valuedev- complexifiedcoordinates and bytheequation erywhereandthereforedependson , isnotanalytic,andin- deed,thetheoryofanalyticfunctionsisirrelevantfortheresults ofthispaper. B. ComplexifiedFIMandCRB (13) TheFIM (Equation (13) is equivalent to Theorem 1 of Yau and Bresler (18) [56]). The log-likelihood function of many array processing problems[e.g.,(4)]dependslinearlyon and (i.e.,sesquilin- willbeusedwith(11)todeterminetheCRBonthecovarianceof earlyon ),oftentimesgreatlysimplifyingthederivationofthe foranyunbiasedestimator .Thedetailsarestraightforward complexifiedFIM of(13).Notethatin andprovidedinAppendixB,asisthereasonthedeterministic contrast to the complex Hessian matrix encountered in CRBisnotaffectedbytheinclusionoftheknownorunknown thestudyofplurisubharmonicfunctions[19],thecomplexified covariancematrix intheparametervector.ThefullFIMis FIMrepresentstheexpectationofthefullHessianmatrixwith respect to all the elements of the complexified vector . However, the diagonal blocks and usually vanish. (19) Therefore, only the cross terms typically appear in the resultingCRBs. Some additional nonstandard properties of complexified where is an -by- complex matrix (the FIM’s that do not appear elsewhere in the literature will be negative expectation of the complex Hessian matrix TLFeBOOK 1600 IEEETRANSACTIONSONSIGNALPROCESSING,VOL.53,NO.5,MAY2005 [19]), Re isa -by- realmatrix, In the simplest case of a single source and a uniform linear array, , and letting is a -by- complex matrix, and the be the array center’s phase, (23)–(28) reduce subscripted notation “ ” denotes the substitution of the first to SNR , SNR , yielding derivative with respect to wherever the th index of is thewell-knownresult required(seeAppendixBforspecificexamples).Theelements ofthematrices and aregivenby rad (31) SNR Re thelement (20) throw (21) which is independent of the phase center . It also is in- structive to compare the deterministic CRB of (29) to the where stochastic CRB [2], [14], [49], [50], where the signal am- plitudes are random with unknown covariance matrix (22) , and the output probability distribution of (1) is with co- isthe -by- complexmatrixofthederivativeofthesteering variance matrix , and data matrix matrix withrespectto . . In this case, the full FIM is given by Applying(11)to(19),thecomplexifiedCRBisobtained: . This is an appropriate data model for multiple data snapshots with fluctuating target -by- (23) amplitudes.Matveyevetal.[25]provideadetailedcomparison ofthedeterministicandstochasticCRBs. where -by- (24) The bounds shown in (29) must be evaluated for a partic- Re -by- ularchoiceofamplitudevector andparametervector .These vectorsaresetatsomenominalvaluesofinterest.Themagni- (25) tude of the vector may be set to achieve a given array-level Re -by- (26) signal-to-noiseratio(SNR)viatheequation -by- (27) SNR (32) -by- (28) where is the covariance matrix assumed for the noise The somewhat obvious notation for a subscripted “ ” is ex- model. Note that . In the specific plainedinthepreviousparagraph,andtheformulaefor , , caseconsideredinSectionIVoftwosourceswithidentical(but and are repeatedfor convenience. Equations(24)–(28) may unknown) amplitudes, the array SNR when the sources coa- becombinedtoobtaintheexpression lesceisgivenbySNR SNR ,whichisobtainedfrom Re , SNR , and , and SNR denotestheelement-levelSNR. (29) The standard application of computing bounds on azimuth andelevationmeasurementsisprovidedinAppendixC. In the frequently encountered case where the signal model is ,(29)isreadilyseen III. EXAMPLE:POLEMODELESTIMATION tocorrespondto Considertheproblemofmodelingfrequencydatausingaset ofdampedexponentials: Re (30) (33) where“ ”and“ ”denotetheSchur–Hadamard(component- wise)andKronecker(tensor)products,respectively, denotes where the are unknown complex numbers, and the are the -by- matrixhavingunityelements,and,byabuseofno- samples in frequency. This may also be viewed as a time-se- tation, [in riesmodelingproblemofsinusoidsincomplexnoise[43],[45], (30)only].ThisissimplyYauandBresler’s[56]resultforthe [48], [49], [52]; however, the mathematically equivalent fre- case . The matrix may be interpreted as an array of quency-domainversion used by Cuomo et al. [5], [6] for syn- “crossdifferenceSNR”terms;thematrix maybeinterpreted theticultra-widebandprocessingisconsideredhere.Steedlyand astheextralossofaccuracyduetotheunknownamplitudes.Ap- Moses[43]computedeterministicCRBsforthecomplexpoles plicationofthiscomplexifiedapproachisstraightforwardfora andtheunknownamplitudesusingthestandardreal/imaginary largeclassofapplications;compare(29)and(30)tootherdeter- approach. In this section, the complexified approach will be ministicCRBformulaereportedintheliterature[8],[25],[27], usedtoobtainformulaefortheidentical CRboundsofsignif- [43], [45]–[47], [55]. If is an unbiased estimator of , the icantlygreatersimplicity.Theseformulaewillthenbeapplied CRB holds. tocomputingclosed-formresolutionlimitsinSectionIV. TLFeBOOK SMITH:STATISTICALRESOLUTIONLIMITSANDTHECOMPLEXIFIEDCRAMÉR-RAOBOUND 1601 Equation(33)canbeexpressedcompactlyusing(1)andthe itshouldbeverifiedthat(39)–(46)areconsistentwiththeinter- steeringmatrix pretationof givenin(21)and(22).Toseetherelativesim- plicityachievedbythecomplexifiedCRBs,compare(38)–(46) totheequivalent,real/imaginarybounds[43,pp.1306–1309]. InSectionIV,theseCRBsontheestimatesofthepoles will . . . (34) beusedtoderiveboundsonthefundamentalresolutionachiev- . . . . . . ablebyanaperture,i.e.,thesmallestpoleseparationmeasurable byanyunbiasedestimatorofthepoles. Because the unknowns are complex numbers , thetreatmentofSectionIImustbeslightlymodifiedtohandle IV. STATISTICALRESOLUTIONLIMITS thiscase.Thereisachoice:Usethecomplexifiedapproachfor The CRB quantifies a lower bound on estimation error, but and directly,orusethemorecommonparameterizations itdoesnotdirectlyindicatethebestresolutionachievablebyan oftherealandimaginarypartsorthemoduliandphasesofthe unbiasedestimator.Nevertheless,theCRBcanbeusedtodefine poles .Inthisexample,the steering an absolute limit on resolution. The minimum requirement to matrix dependsonlyonthecomplexpolevalues (andnot resolvetwosignalsis theircomplexconjugates ),andderivativeswithrespectto and ,whichareeasilyobtainedfrom(15)and(16),are standarddeviation sourceseparation (47) (35) ofsourceseparation (36) Thestatisticalresolutionlimitisdefinedasthesourceseparation atwhichequalityin(47)isachieved,i.e.,thesourceseparation whichmaybeextractedfromtherealandimaginarypartsofthe thatequalsitsownCRB.Becausethisconceptofresolutionis derivativesof withrespectto .Therefore,itwillbesimplest based on the CRB, it has the advantage of bounding the reso- tochoosetheparameterization lutionofallsuper-resolutionmethodsintheregimeswherethe CRBholds:highSNRs,unbiasedestimators,andnomodeling (37) orsignalmismatch.Inthissense,thestatisticalresolutionlimit providesthebest-caseresolutionboundforanyalgorithm.Due i.e., the standard real/imaginary approach, and proceed as de- totheso-calledSNRthresholdeffect[17],[31]causedbyesti- scribedinSectionIIusingahybridoftherealandcomplexified mationoutliersinthethresholdregion,theresolutionofspecific approachestocomputeboundsontheunknowncomplexpoles algorithmstypicallyexceedsthestatisticalresolutionlimit,al- withtheunknowncomplexamplitudes . thoughtheirqualitativebehaviorsmaybesimilar.Bystatistical Given(37),theCRBforanyunbiasedestimator of from considerations alone, Swingler [51] concludes that the practi- (23)–(28)isgivenby callyachievableresolutionlimitisaboutonetenthtoonefourth of the Rayleigh resolution limit. Furthermore, mismatch and modelingerrorscangreatlydegraderesolution[11],[17],[31], butnoneoftheseeffectsaretakenintoaccountbythestatistical (38) resolutionlimit. where Re (39) The CRB of the source separation has been used to bound resolutioninpreviousworkbyClark[4],Smith[40],andYing (40) etal.[57],whoprovideamathematicallyequivalenttreatment of(47).TheCRBofthesourceaccuraciesthemselves(notthe Re (41) CRB of the difference) has been used by Lee [23], Swingler (42) [49]–[52],Steedlyetal.[42],andseveralresearchersintheop- tics community (see the references cited in the survey by den (43) DekkerandvandenBos[7],especiallythepaperbyHelstrom [16],aswellasLucy[21]andBettensetal.[3]).Aswillbeseen, (44) theresultsderivedinthispaperarequalitativelyconsistentwith manyotherpublishedresults. (45) Equation (47) is an appealing definition of resolution in a practical sense. In the regime where the sources are resolved, thcolumnof (46) the estimation accuracy is bounded by the CRB. If the poles are too close, they cannot be resolved because the two spec- Notethatthetensorproductsin(40)and(44)appearintheoppo- tral peaks of each signal coalesce into one; this fact is behind siteorderthanthatin(30)becauseofthedifferentlexicographic much of the analysis of the resolution limits for several algo- orderofthe parametervectorin(37).Becausethe thcolumn rithms. In this case, the two poles are estimated to lie at the of dependsonlyon and ,thetensor same location somewhere between the two true locations, the maybeefficientlyrepresentedusingan -by- matrix; difference between the poles is estimated to be zero, and the TLFeBOOK 1602 IEEETRANSACTIONSONSIGNALPROCESSING,VOL.53,NO.5,MAY2005 root-mean-squareerror(RMSE)oftheestimateddifferenceap- Equation(57)isVanTrees’sproperty5ofCRBsofnonrandom proximatelyequalsthepoleseparation. variables[54,sec.2.4]. Symbolicalgebrapackages,suchasMapleorMathematica, A. Single-ApertureResolutionLimit areusefulforcomputingtheLaurentseries Theresolutionlimitfortwosignalssampledoveranequally Var (58) spacedapertureof pointswillbecomputed.Assumethatsig- nals are modeled by the poles and ( small),thateachsignalhasidenticalbutunknownamplitudes, SNR and that the noise background is white . For this ex- (59) ampleproblem,(38)–(46)canbeexpressedcompactlyusingthe Dirichletkernel for small, where SNR SNR is the array SNR whenthepolescoalesce,andSNR istheelement-levelSNR. As discussed above, it is desired to compute the separation (48) thatyieldsatleastastandarddeviationof ,i.e., Because of the special form of (33), it will be convenient to takederivativeswithrespecttothepoles andthen SNR (60) transformthemtoderivativeswithrespecttotherealandimag- Solving(60)forthis (assumedsmall)determinesthestatis- inarypartsofthepolesvia(35)and(36).Expressing(38)–(46) ticalresolutionlimit intermsoftheexample,itisfoundthat (61) Re (49) SNR for large (62) SNR (50) Dividing(62)bythestandardFourierresolutionlimit rad yieldsausefulrule-of-thumbfortheabsoluteresolutionlimit: (51) (52) resolution re.Fourierres. (63) SNR (53) Inaddition,notethat(61)impliesthatmorethanthreesensors arerequiredtosuperresolvethetwocomplexpoles. Re (54) Equation(61)indicatesthattodoubletheresolutionusinga fixedaperture,theSNRmustincreaseby12dB,ortoincrease where , and for some unknown the resolution tenfold, the SNR must increase by 40 dB. This propertyisshowninFig.1,whichfortheexampleof complex parameter . In the assumed case of unit-variance shows the RMSEs and the CRBs of the modulus of the dif- whitenoise, SNR ,whereSNR istheelement-level ference of the estimated poles versus the pole separation. The SNR, and is an unknown phase. Applying (32) when the “ringing”patternobservedinFig.1whenthepolesareresolved polescoalesce ,SNR SNR .Tosimplify occurs as the sources move in and out of the nulls and peaks thecomputations,aphase-centeredapertureisused. of their respective array responses. The error is lowest when Asdiscussedabove,aboundisdesiredfortheRMSEofthe onepoleisnearanulloftheotherarray’sresponseandhighest difference whenitisnearapeak.Equation(61)isverifiedbyMonteCarlo complexified (55) simulations (Figs. 2 and 4) of the maximum-likelihood (ML) methodforpoleestimationwithunknownamplitudes.Fig.2il- or real (56) lustrates an example with 50 samplesand SNRof 48 dB. The maximum-likelihoodmethodachievestheCRBwhenthepoles are resolved but breaks away from the CRB curve at the pre- where , and , dicted statistical resolution limit of about 0.08 of the Fourier as above. Therefore, the FIM must be computed in terms of resolution rad. The resolution performance of both the the variable or .Becausereal/imaginarycoordinatesare MUSICandESPRITalgorithmsisalsoshown,usingadatama- used in (38)–(46), the parameter vector from (56) is used, trixconstructedfromaslidingwindowof samples. resultingin Both of these superresolution methods achieve resolutions of about0.3–0.4oftheFourierresolution rad,withESPRIT achieving a slightly better resolution than MUSIC due to the (57) invariance properties of pole estimation that ESPRIT exploits. TLFeBOOK SMITH:STATISTICALRESOLUTIONLIMITSANDTHECOMPLEXIFIEDCRAMÉR-RAOBOUND 1603 Fig.1. Cramér-RaoboundofthepoledifferenceforN = 50andseveral SNRs,itsasymptotecomputedusing(60),andtheresolutionlimitscomputed using(61)expressedusingastandardFourierresolutioncell(FRC)of2(cid:25)=N (cid:25) 1=8rad.NotethathalvingtheresolutionrequiresincreasingtheSNRby12dB. Inthisexample,thestandarddeviationofthedifferencebetweenpolelocations Fig. 2. Full aperture predicted versus measured resolution with two poles. iscomputedassumingtwosignalswithpolesat1ande ((cid:1)(cid:30)(cid:17)separation) TheRMSEofthedifferenceofthepoleestimatesversuspoleseparationfor andwithequalbutunknownamplitudes.The“ringing”patternobservedresults themaximum-likelihood(ML)method(reddashedcurvewith(cid:14)s),theMUSIC fromthesourcesmovinginandoutofthenullsandpeaksoftheirrespective algorithm(purpledashedcurvewith+s),theESPRITalgorithm(bluedashed arrayresponses. curvewithxs),andtheCRB(darkgreensolidcurve)areallshownusingthe unitofaFourierresolutioncell(FRC)2(cid:25)=N (cid:25) 1=8rad,whereN = 50, and SNR = 48dB in this example. For both the MUSIC and ESPRIT Beneath their resolution limits, MUSIC and ESPRIT provide algorithms,aslidingwindowofN=2 = 25samplesisusedtogeneratethe the identical estimate for the two different poles. Hence, the datamatrices.Themaximum-likelihoodmethodfailstoresolvethetwopoles at the resolution predicted by (63) and breaks away from the CRB at this RMSEofthedifferenceequalsthepoleseparationoftheunre- point;thestatisticalresolutionlimit(separation=RMSE)isthedashedblack solvedpoles.Incontrast,beneaththeresolvabilityoftheMLal- linetotheupperright.TheSNRthresholdbehaviorofboththeMUSICand gorithm,therandombehaviorofpeaksintheeight-dimensional ESPRIT algorithms is observed, resulting in an achieved resolution greater thanthestatisticalresolutionlimitforbothofthesemethods.Asexpected,the log-likelihoodsurface(magnitudeandphasefortwopolesand resolutionoftheESPRITmethodisslightlygreaterthanthatoftheMUSIC twocomplexamplitudes)yieldsanearlyflatRMSEcurvewhen methodbecauseESPRITexploitsinvarianceinformationnotusedbyMUSIC. thepolesareunresolved. AMonteCarlosimulationwith1000trialsisusedtocomputetheRMSEs.The “ringing”patternobservedisthesourcesmovinginandoutofthenullsand These resolution results are qualitatively consistent with peaksoftheirrespectivearrayresponses. manyresolutionboundsreportedintheliterature[40].Lee[23, Eqs.(62)and (63)]hypothesizes anecessaryconditionforthe SNR , which is a good match to the theoretical result resolvability of signals (deterministicCRB assumption) and SNR . Steedly and Moses [43] (deterministic assumption) deduces a dependence of SNR or SNR for . numerically compute the CRB for the angle separation of two Swingler [49]–[52] (stochastic assumption) uses an approxi- polesontheunitcircleandillustrateaplot(Fig.5onp.1312) mationofthestochasticCRBtoobtainresolutionbounds.For thatisqualitativelysimilartoFig.1.Thisresultisalsoderived large , high SNR, and a single data snapshot, Swingler’s by Kaveh and Lee [17, Eq. (5.118)] (stochastic case), using approximation [51, Eq. (3)] implies that the reso- asymptotic properties of the Wishart distribution. Stoica and lution (normalized by the Fourier resolution of ) is at Nehorai [46], [47] (deterministic and stochastic assumptions) least SNR SNR .Uptotheconstant illustrate numerical evidence that the standard deviation is multiple of 1.7, this is comparable to the statistical resolution proportional to the inverse of the separation as in (60). Stoica limit of SNR given in (63). Ying et al. [57] define a and Söderström (stochastic case) observe in [48, p. 1842] that resolutionlimitintheir(5)thatismathematicallyequivalentto the covariance elements become large when two frequencies thestatisticalresolutionlimitdefinedin(47),andtheyillustrate are close but do not quantify this increase. Zhang [59, Fig. 1] resolution proportional to SNR in their Fig. 1, as well illustrates numerical results for the probability of resolution provide as an alternate geometric interpretation of this limit. thatshowaclosematchtothepredictedSNR dependence. ShahramandMilanfar[37]–[39]useadetectiontheoreticcon- Incontrasttothisbodyofwork,Kosarev[18]usesaninfor- ceptofresolvabilitytoarriveataresolutionlimitthatdepends mation-theoreticapproachtoarriveatasuperresolutionbound onSNR .LeeandWengrovitz[24](stochasticassumption) thatdependslogarithmicallyontheSNR. showintheirEq.(91)thattheresolutionthresholdforMUSIC Ifthemagnitudesofthetwopolesarefixedontheunitcircle, isproportionaltoSNR .Farina[9,Sec.5.7]confirmsthere- a similar analysis shows that the resolution is also determined sultofGabriel[12]thattheachievablesuperresolutionisabout by(61)iftheresolutionlimitismultipliedby . TLFeBOOK 1604 IEEETRANSACTIONSONSIGNALPROCESSING,VOL.53,NO.5,MAY2005 B. Split-ApertureResolutionLimit Split-aperturearraysrefertothecasewhereinsteadofhaving a full aperture, there are two apertures separated by a large gap. The analysis of the preceding section can be modified to compute the achievable resolution of split-aperture arrays with so-called synthetic ultra-wideband coherent processing [5], [6], [40], [58]. CRBs on source direction of arrivals are well known for arbitrary array geometries [13], [28]. The CRB of the source separation for a split-aperture is consid- ered here. Let the synthetic ultra-wideband coherent aperture consist of samples (as above), but with the middle samples removed(for convenience, assume thatthe difference isevensothatthetwowidelyseparatedapertureshave equal length). The frequency samples occur at the samples , deletedband , , . The results of the pre- vioussectioncarryoveriftheDirichletkernel of(48)is replacedwiththemodifiedDirichletkernel Fig.3. BandgapCRBforB = 25,M = 500(yieldingN = 550),and several SNRs expressed using a standard Fourier resolution cell (FRC) of 2(cid:25)=N (cid:25)0:011rad.Theasymptotesfrom(71)andtheresolutionlimitsfrom (73)arealsoshown.The“ringing”patternobservedisthesourcesmovingin andoutofthenullsandpeaksoftheirrespectivearrayresponses. (64) Intermsofthesubaperturebandwidth andthebandsepara- tion ,(65)becomes ComputingtheLaurentseriesofthevarianceforsmall Var Var SNR SNR (65) (71) (72) where SNR (66) (for large, fixed), where . Applying(47)to(65)[or(71)]yieldsthestatisticalresolution limitforasplit-bandaperture (67) (68) SNR (73) (74) SNR (radians, large, fixed). (69) Fig.3illustratestheCRBandresolutionachievableforesti- matingtwopolepositionswhosesignalamplitudesareassumed SNR SNR isthearraySNRwhenthepoles equal(with36dBarraySNR)butunknownfortwofrequency coalesce, and SNR is the element-level SNR. Note that this bandsof25sampleseachseparatedby100samples.Fig.4com- equationissymmetricin and andthatitreducesto(59)if parestheperformanceofthemaximum-likelihoodestimatorto (asdesired).Theultrawidebandwidth ismoreconve- thecomputeCRBs.Notethatattheresolutionlimit,theRMSE nientlyparameterizedintermsofthetwoseparatedbandwidths ofthemaximum-likelihoodestimatebreaksawayfromthereso- andthebandgap as lutionboundbecausethepolesarenolongerresolvedandthere- fore cannot be estimated individually as the CRB analysis as- (70) sumes. Comparing Figs. 1 and 3, it is seen that splitting and TLFeBOOK