简介:Considersolvinganoverdeterminedsystemoflinearalgebraicequationsbyboththeleastsquaresmethod(LS)andthetotalleastsquaresmethod(TLS).Extensivepublishedcomputationalevidenceshowsthatwhentheoriginalsystemisconsistent.oneoftenobtainsmoreaccuratesolutionsbyusingtheTLSmethodratherthantheLSmethod.ThesenumericalobservationscontrastwithexistinganalyticperturbationtheoriesfortheLSandTLSmethodswhichshowthattheupperboundsfortheLSsolutionarealwayssmallerthanthecorrespondingupperboundsfortheTLSsolutions.InthispaperwederiveanewupperboundfortheTLSsolutionandindicatewhentheTLSmethodcanbemoreaccuratethantheLSmethod.Manyappliedproblemsinsignalprocessingleadtooverdeterminedsystemsoflinearequationswherethematrixandrighthandsidearedeterminedbytheexperimentalobservations(usuallyintheformofalimeseries).Itoftenhappensthatasthenumberofcolumnsofthematrixbecomeslarger,thera
简介:Datafittingisanextensivelyemployedmodelingtoolingeometricdesign.Withtheadventofthebigdataera,thedatasetstobefittedaremadelargerandlarger,leadingtomoreandmoreleast-squaresfittingsystemswithsingularcoefficientmatrices.LSPIA(least-squaresprogressiveiterativeapproximation)isanefficientiterativemethodfortheleast-squaresfitting.However,theconvergenceofLSPIAforthesingularleast-squaresfittingsystemsremainsasanopenproblem.Inthispaper,theauthorsshowedthatLSPIAforthesingularleast-squaresfittingsystemsisconvergent.Moreover,inaspecialcase,LSPIAconvergestotheMoore-Penrose(M-P)pseudo-inversesolutiontotheleast-squaresfittingresultofthedataset.ThispropertymakesLSPIA,aniterativemethodwithcleargeometricmeanings,robustingeometricmodelingapplications.Inaddition,theauthorsdiscussedsomeimplementationdetailofLSPIA,andpresentedanexampletovalidatetheconvergenceofLSPIAforthesingularleast-squaresfittingsystems.
简介:1.IntroductionThepurposeofthispaperistostudytheleastsquaresproblemofthematrixequationF~PGwithrespecttoPcSa,i.e.(PI)R\qIIF--PGll,whereF,GERnxmandG/0.Where11’11denotestheFrobeniusnorm,andSa~{XeS'fX20},S'={XER'''IX=X'}.Problem(PI)wasfirstformulatedbyAll...
简介:AnegativecurvaturemethodisappliedtononlinearleastsquaresproblemswithindefiniteHessianapproximationmatrices.Withthespecialstructureofthemethod,anewswitchisproposedtoformahybridmethod.Numericalexperimentsshowthatthismethodisfeasibleandeffectiveforzero-residual,small-residualandlarge-residualproblems.
简介:Inthispaper,wepresentsomeiterativemethodsforsolvinglthorderautoregressivemodels,proveglobalconvergenceforl=1case,andthenumericalresultsofnewalgorithmsseemtobemoreefficientthantheonesofCochrane-Orcuttiterativemethod.
简介:TheGalerkinandleast-squaresmethodsaretwoclassesofthemostpopularKrylovsubspacemethOdsforsolvinglargelinearsystemsofequations.Unfortunately,boththemethodsmaysufferfromseriousbreakdownsofthesametype:InabreakdownsituationtheGalerkinmethodisunabletocalculateanapproximatesolution,whiletheleast-squaresmethod,althoughdoesnotreallybreakdown,isunsucessfulinreducingthenormofitsresidual.Inthispaperwefrstestablishaunifiedtheoremwhichgivesarelationshipbetweenbreakdownsinthetwometh-ods.Wefurtherillustratetheoreticallyandexperimentallythatifthecoefficientmatrixofalienarsystemisofhighdefectivenesswiththeassociatedeigenvalueslessthan1,thentherestart-edGalerkinandleast-squaresmethodswillbeingreatrisksofcompletebreakdowns.Itappearsthatourfindingsmayhelptounderstandphenomenaobservedpracticallyandtoderivetreat-mentsforbreakdownsofthistype.
简介:Weproveconvergenceforameshfreefirst-ordersystemleastsquares(FOSLS)partitionofunityfiniteelementmethod(PUFEM).Essentially,byvirtueofthepartitionofunity,localapproximationgivesrisetoglobalapproximationinH(div)∩H(curl).TheFOSLSformulationyieldslocalaposteriorierrorestimatestoguidethejudiciousallotmentofnewdegreesoffreedomtoenrichtheinitialpointsetinameshfreedis-cretization.Preliminarynumericalresultsareprovidedandremainingchallengesarediscussed.
简介:NearlyorthogonalLatinsquaresareusefulforconductingexperimentseliminatingheterogeneityintwodirectionsandusingdifferentinterventionseachateachlevel.Inthispaper,someconstructionsofmutuallynearlyorthogonalLatinsquaresareprovided.Itisprovedthatthereexist3MNOLS(2m)ifandonlyifm≥3ndthereexist4MNOLS(2m)ifandonlyifm≥4withsomepossibleexceptions.
简介:SeveralARMAmodelingapproachesareaddressed.Inthesemethodsonlypartofacorrelationsequenceisemployedforestimatingparameters.Itissatisfying,ifthegivencorrelationsequenceisofrealARMA,sinceanARMAprocesscanbecompletelydeterminedbypartofitscorrelationse-quence.Butforthecaseofameasuredcorrelationsequencethewholesequencemaybeusedtore-ducetheeffectoferroronmodelparameterestimation.Inaddition,thesemethodsnowdonotguar-anteeanonnegativespectralestimate.Inviewoftheabove-mentionedfact,aconstrainedleastsquaresfittingtechniqueisproposedwhichutilizesthewholemeasuredcorrelationsequenceandguar-anteesanonnegativespectralestimate.
简介:TheLeastSquaresSupportVectorMachines(LS-SVM)isanimprovementtotheSVM.CombinedtheLS-SVMwiththeMulti-ResolutionAnalysis(MRA),thisletterproposestheMulti-resolutionLS-SVM(MLS-SVM).TheproposedalgorithmhasthesametheoreticalframeworkasMRAbutwithbetterapproximationability.AtafixedscaleMLS-SVMisaclassicalLS-SVM,butMLS-SVMcangraduallyapproximatethetargetfunctionatdifferentscales.Inexperiments,theMLS-SVMisusedfornonlinearsystemidentification,andachievesbetteridentificationaccuracy.
简介:Inthispaperwepresentanonmonotonetrustregionmethodfornonlinearleastsquaresproblemswithzero-residualandproveitsconvergenceproperties.Theextensivenumericalresultsarereportedwhichshowthatthenonmonotonetrustregionmethodisgenerallysuperiortotheusualtrustregionmethod.
简介:TwoLatinsquaresofordervarer-orthogonaliftheirsuperpositionproducesexactlyrdistinctorderedpairs.Thetwosquaresaresaidtober-orthogonalidempotentLatinsquaresanddenotedbyr-MOILS(v)iftheyareallidempotent.Inthispaper,weshowthatforanyintegerv≥28,thereexistsanr-MOILS(v)ifandonlyifr∈[v,v~2]\{v+1,v~2-1}.
简介:HASM(当模特儿的高精确性表面)技术基于表面的基本理论,它被证明了在表面适合改进插值精确性。然而,在以前的研究的不可分的反复的解决方案在计算和巨大的存储器用法导致了高时间的复杂性以便把这种技术放进申请变得困难,特别为大规模数据集。在学习,一个创新模型(HASM广告)根据数据调整理论根据顺序的最少的广场被开发。顺序的分割在这种技术被采用,以便线性方程能被划分成在顺序要处理,时间的复杂性在计算极大地减少了的组。实验显示HASM广告技术在精确性超过传统的空间插值方法。另外,交叉验证结果为土壤PH性质的空间插值证明一样的结论,数据在江西省取样了。而且,HASM广告技术显著地减少计算复杂性并且在计算减少记忆用法,这在学习被表明。
简介:Thefollwingsituationinusingthemethodofleastsquarestosolveproblemsoftenoccurs.Aftermexperimentscompletedandasolutionofleastsquaresobtained,the(m+1)-thexperimentismadefurtherinordertoimprovetheresults.Amethodofalgebraicoperationofspecialmatricesinvoledintheproblemisgivenisthispaperforobtaininganewsolutionforthem+1experimentsbasedupontheoldsolutionfortheprimarymexperiments.Thismethodisvalidformoregeneralmatrices.
简介:WeextendtheobliqueprojectionmethodgivenbyY.Saadtosolvethegeneralizedleastsquaresproblem.Thecorrespondingobliqueprojectionoperatorispresentedandtheconvergencetheoremsareproved.SomenecessaryandsufficientconditionsforcomputingthesolutionortheminimumN-normsolutionofthemin||Ax-b||M2havebeenproposedaswell.