您好,欢迎访问三七文档
LINEARSYSTEMSOLVERS:SPARSEITERATIVEMETHODSHenkA.VanderVorstMathematicalInstituteUniversityofUtrechtUtrecht,theNetherlandse-mail:vorst@math.ruu.nlTonyF.Chan1DepartmentofMathematicsUniversityofCaliforniaatLosAngelesLosAngeles,CAe-mail:chan@math.ucla.eduABSTRACTInthischapterwewillpresentanoverviewofanumberofrelatediterativemethodsforthesolutionoflinearsystemsofequations.Thesemethodsareso-calledKrylovprojectiontypemethodsandtheyincludepopularmethodsasConjugateGradients,Bi-ConjugateGradients,LSQRandGMRES.Wewillsketchhowthesemethodscanbederivedfromsimplebasiciterationformulas,andhowtheyareinterrelated.Iterativeschemesareusuallyconsideredasanalternativeforthesolu-tionoflinearsparsesystems,likethosearisingin,e.g., niteelementor nitedi erenceapproximationof(systemsof)partialdi erentialequations.Thestructureoftheoperatorsplaysnoexplicitroleinanyoftheseschemes,andtheoperatormaybegivenevenasaruleorasubroutine.Althoughthesemethodsseemtobealmosttriviallyparallellizableat rstglance,thisissometimesapointofconcernbecauseoftheinnerproductsinvolved.Wewillconsiderthispointinsomedetail.Iterativemethodsareusuallyappliedincombinationwithso-calledpreconditioningoperatorsinordertofurtherimproveconvergenceproperties.Thisaspectwillreceivemoreattentioninaseparatechapterinthesamevolume.1TheworkofthisauthorwaspartiallysupportedbytheNationalScienceFoundationundercontractASC92-01266,theArmyResearchO ceundercon-tractsDAAL03-91-G-0150andsubcontactfromU.Tenn.DAAL03-91-C-0047,andONRundercontractONR-N00014-92-J-1890.1IterativemethodsversusdirectmethodsThepresentstateoftheartinnumericalmethodsisthatdirectmethodscanbeusedasblackboxes.Thisisbyfarnotthecaseforiterativemethods,atleastnotifwedonotknowaboutspeci cpropertiesofthematrixofthelinearsystemtobesolved.Andeventhenitisnotrivialmattertodecidewhentostoptheiterationprocessandtoobtainareasonableestimateoftheapproximationerrorintheresult.Therefore,westartwithsomeveryglobalanalysisofthecomplexityofGaussianeliminationandaniterativesolutionmethodforamodelproblem.Thismayserveasamotivationforthefurtherstudyofiterativemethods.Gaussianeliminationleadsto ll-in,thismakesthemethodoftenexpensive.Usuallylargesparsematricesarerelatedtosomegridornetwork.Ina3Dsituationthisleadstypicallytoabandwidth n23(=m2andm3=n,1=mthegridsize).Thenumberof opsisthentypicallyO(nm4) n213(GolubandVanLoan,1989)(Note:weassumethatthesparsitystructureisnotparticularlyregularsothatspecialcomputationalvariantscannotbeemployed).Ifonehastosolvemanysystemswithdi erentright-handsides,thenthecostsforsolvingeachsystemwillvaryliken53.Ifweassumethatthematrixissymmetricpositivede nitethentheConjugateGradient(CG)iterationmethod(whichwewilldescribeinmoredetaillateron)canbeusedsafely.TheerrorreductionperiterationstepofCGis p 1p +1,with =kAk2kA 1k2(GolubandvanLoan,1989).Fordiscretizedsecondorderpde’s,overgridswithgridsize1mwetypicallysee m2.Hence,forsuch3Dproblemswehavethat n23.Foranerrorreductionof wemusthavethat0@1 1p 1+1p 1Aj (1 2p )j e 2jp ;wherejisthenumberofiterationstepstobecarriedout.Hence,itfollowsforjthat)j log 2p log 2n13:Ifweassumethenumberof opsperiterationtobe fn(fstandsforthenumberofnonzerosperrowofthematrixandtheoverheadperunknownintroducedbytheiterativescheme)thenthenumberof opsnecessarytoachieveanerrorreductionby is fn43log .Conclusion:Ifwehavetosolveonesystematatime,thenforlargen,orsmallf,ormodest ,Iterativemethodsmaybepreferable.Ifwehavetosolvemanysimilarsystemswithdi erentright-handside,andifweassumetheirnumbertobesolargethatthecostsforconstructingthedecompositionofAisrelativelysmallpersystem,thenitseemslikelythatfor2Dproblemsdirectmethodsmaybemoree cient,whereasfor3Dproblemsthisisstilldoubtful,sincethe opscountforadirectsolutionmethodvariesliken73,andthenumberof opsfortheiterativesolver(forthemodelsituation)variesliken43.Example:TheabovegivenargumentsarequitenicelyillustratedbyobservationsmadebySimon(Simon,1989).Heexpectsthatbytheendofthiscenturywewillhavetosolverepeatedlylinearproblemswithsome5 109unknowns.Forwhathebelievestobeamodelproblematthattime,hehasestimatedtheCPUtimerequiredbythemosteconomicdirectmethod,availableatpresent,as520;040years,providedthatthecomputationcanbecarriedoutataspeedof1TFLOP.Ontheotherhand,heestimatestheCPUtimeforpreconditionedconjugategradients,assumingstillaprocessingspeedof1TFLOPS,as575seconds.Thoughweshouldnottakeitforgrantedthatinparticularaverysparsepreconditioningpartcanbecarriedoutatthathighprocessingspeed(forthedirectsolverthisismorelikely),weseethatthedi erencesinCPUtimerequirementsaregigantic,indeed(wewillcometothispointinmoredetail).Alsotherequirementsformemoryspacefortheiterativemethodsaretypicallysmallerbyordersofmagnitude.Thisisoftentheargumentfortheusageofiterativemethodsin2Dsituations,when opcountsforbothclassesofmethodsaremoreorlesscomparable.Finallyitshouldbenotedthatiterativemethodscanexploitgoodinitialguesses,e.g.,intimedependentproblems.Thepreconditionercanoftenbechosentoadapttothemachinearchitecture.Remarks: Forspecialproblemsspecialmethodsmaybefaster:multigrid,fastpoi
本文标题:Linear system solvers Sparse iterative methods
链接地址:https://www.777doc.com/doc-3275422 .html