资源描述:
《pythagoras papadimitriou》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、ISSN1360-1725UMISTParallelSingularValueDecompositionviathePolarDecompositionN.J.HighamandP.PapadimitriouNumericalAnalysisReportNo.239October1993ManchesterCentreforComputationalMathematicsNumericalAnalysisReportsDEPARTMENTSOFMATHEMATICSReportsavailablefrom:And
2、overtheWorld-WideWebfromURLsDepartmentofMathematicshttp://www.ma.man.ac.uk/MCCM/MCCM.htmlUniversityofManchesterftp://ftp.ma.man.ac.uk/pub/narepManchesterM139PLEnglandParallelSingularValueDecompositionviathePolarDecompositionyNicholasJ.HighamPythagorasPapadim
3、itriouOctober30,1993AbstractAnewmethodisdescribedforcomputingthesingularvaluedecomposition(SVD).ItbeginsbycomputingthepolardecompositionandthencomputesthespectraldecompositionoftheHermitianpolarfactor.Themethodisparticularlyattractiveforsharedmemoryparallelco
4、mputerswitharelativelysmallnumberofprocessors,becausethepolardecompositioncanbecomputedecientlyonsuchmachinesusinganiterativemethoddevelopedrecentlybytheauthors.Thisiter-ativepolardecompositionmethodrequiresonlymatrixmultiplicationandmatrixinversionkernelsfo
5、ritsimplementationandisdesignedforfullrankmatrices;thustheproposedSVDmethodisintendedformatricesthatarenottooclosetobeingrank-decient.OntheKendallSquareKSR1virtualsharedmemorycomputerthenewmethodisuptosixtimesfasterthanaparallelizedversionoftheLAPACKSVDrouti
6、ne,dependingontheconditionnumberofthematrix.Keywords.singularvaluedecomposition,polardecomposition,numericalstability,LAPACK,level3BLAS,KendallSquareResearchKSR1computerAMSsubjectclassications.primary65F15DepartmentofMathematics,UniversityofManchester,Manch
7、ester,M139PL,England(na.nhigham@na-net.ornl.gov).TheworkofthisauthorwassupportedbyScienceandEngineer-ingResearchCouncilgrantGR/H52139,andbytheEECEspritBasicResearchActionProgramme,Project6634(APPARC).yDepartmentofMathematics,UniversityofManchester,Manchester,
8、M139PL,England(na.papadimitriou@na-net.ornl.gov).Currentaddress:DataInformationSystemsPLC,125Thes-salonikis,N.Philadelphia,Athens14243,Greece.ThisauthorwassupportedbyanSERCResearchStudent